DSP Principles Algorithms and Applications Third Edition

77un/ Edition
DIGITAL SIGNAL PROCESSING

Principles, Algorithms, m l Applications
J o h n G. Proakis Dimitris G. M anolakis
Digital Signal Processing

Principles, Algorithms, and Applications
Third E dition
John G. Proakis
Northeastern U niversity
Dimitris G. Manolakis
Boston C ollege
PRENTICE-HALL INTERNATIONAL, INC.
This edition may be sold only in those countries to which it is consigned by Prentice-Hall International. It is not to be reexported and it is not for sale in the U .S.A ., Mexico, or Canada. 1996 by Prentice-Hall, Inc. Simon & Schuster/A Viacom Company U pper Saddle River, New Jersey 07458
All rights reserved. No part of this book may be reproduced, in any form or by any means, without permission in writing from the publisher.
The author and publisher of this book have used their best efforts in preparing this book. These efforts include the development, research, and testing of the theories and programs to determine their effectiveness. The author and publisher make no warranty of any kind, expressed or implied, with regard to these programs or the documentation contained in this book. The author and publisher shall not be liable in any event for incidental or consequential damages in connection with, or arising out of. the furnishing, performance, or use of these programs.
Printed in the United States of America 10 9 8 7 6 5
ISBN 0-13-3TM33fl-c l
Prentice-Hall International (U K ) Limited. L ondon Prentice-Hall of Australia Pty. Limited, Sydney Prentice-Hall Canada, Inc., Toronto Prentice-Hall Hispanoamericana. S.A., M exico Prentice-Hall of India Private Limited, N ew D elhi Prentice-Hall of Japan, Inc., T okyo Simon & Schuster Asia Pie, Ltd., Singapore Editora Prentice-Hall do Brasil, Ltda., R io de Janeiro Prentice-Hall, Inc, Upper Saddle River, N ew Jersey
Contents
PREFACE 1 INTRODUCTION
1.1
S ignals, S ystem s, and S ignal P ro cessin g 2 1.1.1 Basic Elements of a Digital Signal Processing System. 4 1.1.2 A dvantages of Digital over Analog Signal Processing, 5 C lassificatio n o f Signals 6 1.2.1 Multichannel and Multidimensional Signals. 7 1.2.2 Continuous-Time Versus Discrete-Time Signals. 8 1.2.3 Continuous-Valued Versus Discrete-Valued Signals. 10 1.2.4 Determ inistic Versus Random Signals, 11 T h e C o n c e p t o f F re q u e n c y in C o n tin u o u s -T im e an d D isc re te -T im e S ignals 14 1.3.1 Continuous-Time Sinusoidal Signals, 14 1.3.2 Discrete-Time Sinusoidal Signals. 16 1.3.3 Harmonically Related Complex Exponentials, 19 A n a lo g -to -D ig ita l an d D ig ita l-to -A n a lo g C o n v e rs io n 21 1.4.1 Sampling of Analog Signals, 23 1.4.2 The Sampling Theorem , 29 1.4.3 Q uantization of Continuous-Am plitude Signals, 33 1.4.4 Quantization of Sinusoidal Signals, 36 1.4.5 Coding of Quantized Samples, 38 1.4.6 Digital-to-Analog Conversion, 38 1.4.7 Analysis of Digital Signals and Systems Versus Discrete-Time Signals and Systems, 39 S u m m a ry a n d R e fe re n c e s 39
xiii
1
1.2
1.3
1.4
Problems
40
iii
iv
Contents
DISCRETE-TIME SIGNALS AND SYSTEMS

2.1 D isc rete-T im e S ignals 43 2.1.1 Some Elem entary Discrete-Time Signals, 45 2.1.2 Classification of Discrete-Time Signals, 47 2.1.3 Simple Manipulations of Discrete-Time Signals, 52 D isc re te -T im e S ystem s 56 2.2.1 Input-O utput Description of Systems, 56 2.2.2 Block Diagram Representation of Discrete-Time Systems, 59 2.2.3 Classification of Discrete-Time Systems, 62 2.2.4 Interconnection of Discrete-Tim e Systems, 70 A n alysis o f D isc re te -T im e L in e a r T im e -In v a ria n t S ystem s 72 2.3.1 Techniques for the Analysis of Linear Systems, 72 2.3.2 Resolution of a Discrete-Time Signal into Impulses, 74 2.3.3 Response of LTI Systems to A rbitrary Inputs: The Convolution Sum, 75 2.3.4 Properties of Convolution and the Interconnection of LTI Systems, 82 2.3.5 Causal Linear Tim e-Invariant Systems. 86 2.3.6 Stability of Linear Tim e-Invariant Systems, 87 2.3.7 Systems with Fim te-D uration and Infinite-Duration Impulse Response. 90 D isc rete-T im e System s D e s c rib e d by D iffe re n c e E q u a tio n s 91 2.4.1 Recursive and Nonrecursive Discrete-Tim e Systems, 92 2.4.2 Linear Time-Invariant Systems Characterized by Constant-Coefficient Difference Equations, 95 2.4.3 Solution of Linear Constant-Coefficient Difference Equations. 100 2.4.4 The Impulse Response of a Linear Tim e-Invariant Recursive System, 108 Im p le m e n ta tio n o f D isc re te -T im e S ystem s 111 2.5.1 Structures for the Realization of Linear Tim e-Invariant Systems, 111 2.5.2 Recursive and Nonrecursive Realizations of FIR Systems, 116 C o rre la tio n of D isc re te -T im e S ignals 118 2.6.1 Crosscorrelation and A utocorrelation Sequences, 120 2.6.2 Properties of the A utocorrelation and Crosscorrelation Sequences, 122 2.6.3 Correlation of Periodic Sequences, 124 2.6.4 Com putation of Correlation Sequences, 130 2.6.5 Input-O utput Correlation Sequences, 131 S u m m ary a n d R e fe re n c e s 134
43
2.2
2.3
2.4
2.5
2.6
2.7
Problems
135
Contents
THE Z-TRANSFORM AND ITS APPLICATION TO THE ANALYSIS OF LTI SYSTEMS

3.1 T h e r-T ra n sfo rm 151 3.1.1 The Direct ^-Transform. 152 3.1.2 The inverse : -Transform, 160 P ro p e rtie s o f th e ; -T ra n sfo rm 161
151
3.2 3.3
R a tio n a l c-T ran sfo rm s 172 3.3.1 Poles and Zeros, 172 3.3.2 Pole Location and Time-Domain Behavior for Causal Signals. 178 3.3.3 The System Function of a Linear Tim e-Invariant System. 181 In v e rs io n o f th e ^ -T ra n sfo rm 184 3.4.1 The Inverse ; -Transform by Contour Integration. 184 3.4.2 The Inverse ;-Transform by Power Series Expansion. 186 3.4.3 The Inverse c-Transform by Partial-Fraction Expansion. 188 3.4.4 Decomposition of Rational c-Transforms. 195 T h e O n e -sid e d ^ -T ra n sfo rm 197 3.5.1 Definition and Properties, 197 3.5.2 Solution of Difference Equations. 201 A n aly sis o f L in e a r T im e -In v a ria n t S ystem s in th e --D o m a in 3.6.1 Response of Systems with Rational System Functions. 203 3.6.2 Response of P ole-Z ero Systems with Nonzero Initial Conditions. 204 3.6.3 Transient and Steady-State Responses, 206 3.6.4 Causality and Stability. 208 3.6.5 P ole-Z ero Cancellations. 210 3.6.6 M ultiple-Order Poles and Stability. 211 3.6.7 The Schur-C ohn Stability Test, 213 3.6.8 Stability of Second-Order Systems. 215 S u m m ary an d R e fe re n c e s P ro b le m s 220 219 203
3.4
3.5
3.6
3.7
FREQUENCY ANALYSIS OF SIGNALS AND SYSTEMS

4.1 F re q u e n c y A n aly sis o f C o n tin u o u s-T im e Signals 230 4.1.1 The Fourier Series for Continuous-Time Periodic Signals. 232 4.1.2 Power Density Spectrum of Periodic Signals. 235 4.1.3 The Fourier Transform for Continuous-Time Aperiodic Signals, 240 4.1.4 Energy Density Spectrum of Aperiodic Signals. 243 F re q u e n c y A n aly sis o f D isc re te -T im e Signals 247 4.2.1 The Fourier Series for Discrete-Time Periodic Signals, 247
230
4.2
V)
Contents
Power Density Spectrum of Periodic Signals. 250 The Fourier Transform of Discrete-Time Aperiodic Signals. 253 Convergence of the Fourier Transform. 256 Energy Density Spectrum of Aperiodic Signals, 260 Relationship of the Fourier Transform to the i-Transform , 264 The Cepstrum, 265 The Fourier Transform of Signals with Poles on the Unit Circle, 267 4.2.9 The Sampling Theorem Revisited, 269 4.2.10 Frequency-Domain Classification of Signals: The Concept of Bandwidth, 279 4.2.11 The Frequency Ranges of Some N atural Signals. 282 4.2.12 Physical and M athematical Dualities. 282 4.3 P ro p e rtie s of th e F o u rie r T ra n s fo rm fo r D isc re te -T im e S ignals 286 4.3.1 Symmetry Properties of the Fourier Transform, 287 4.3.2 Fourier Transform Theorems and Properties, 294 F re q u e n c y -D o m a in C h a ra c te ristic s of L in e a r T im e -In v a ria n t S ystem s 305 4.4.1 Response to Complex Exponential and Sinusoidal Signals: The Frequency Response Function. 306 44.2 Steady-State and Transient Response to Sinusoidal Input Signals. 314 4.4.3 Steady-State Response to Periodic Input Signals, 315 4.4.4 Response to Aperiodic Input Signals. 316 4.4.5 Relationships Between the System Function and the Frequency Response Function. 319 4.4.6 Com putation of the Frequency Response Function. 321 4.4.7 Input-O utput Correlation Functions and Spectra, 325 4.4.8 Correlation Functions and Power Spectra for Random Input Signals. 327 L in e a r T im e -In v a ria n t S ystem s as F re q u e n c y -S e le c tiv e F ilters 330 Ideal Filter Characteristics, 331 4.5.1 4.5.2 Lowpass, Highpass, and Bandpass Filters, 333 4,5.3 Digital Resonators, 340 4.5.4 Notch Filters, 343 4.5.5 Comb Filters. 345 4.5.6 All-Pass Filters. 350 4.5.7 Digital Sinusoidal Oscillators, 352 In v e rse S y stem s an d D e c o n v o lu tio n 355 4.6.1 Invertibility of Linear Tim e-Invariant Systems, 356 4.6.2 Minimum-Phase. Maximum-Phase, and Mixed-Phase Systems. 359 4.6.3 System Identification and Deconvolution, 363 4.6.4 Hom om orphic Deconvolution. 365
4.2.2 4.2.3 4.2.4 4.2.5 4.2.6 4.2.7 4.2.8
4.4
4.5
4.6
Contents
vii S u m m ary a n d R e fe re n c e s P ro b le m s 368 367
4.7
THE DISCRETE FOURIER TRANSFORM: ITS PROPERTIES AND APPLICATIONS

5.1 F re q u e n c y D o m a in Sam pling: T h e D isc re te F o u rie r T ra n s fo rm 394 5.1.1 Frequency-Dom ain Sampling and Reconstruction of Discrete-Time Signals. 394 5.1.2 The Discrete Fourier Transform (DFT). 399 5.1.3 The D FT as a Linear Transform ation. 403 5.1.4 Relationship of the DFT to O ther Transforms, 407 P ro p e rtie s o f th e D F T 409 5.2.1 Periodicity. Linearity, and Symmetry Properties, 410 5.2.2 Multiplication of Two DFTs and Circular Convolution. 415 5.2.3 Additional DFT Properties, 421 L in e a r F ilte rin g M e th o d s B ased on th e D F T 5.3.1 Use of the DFT in Linear Filtering. 426 5.3.2 Filtering of Long Data Sequences. 430 F re q u e n c y A n aly sis o f S ignals U sing th e D F T S u m m ary an d R e fe re n c e s P ro b le m s 440 440 425
394
5.2
5.3
5.4 5.5
433
EFFICIENT COMPUTATION OF THE DFT: FAST FOURIER TRANSFORM ALGORITHMS

6.1 E fficien t C o m p u ta tio n of th e D F T : F F T A lg o rith m s 448 6.1.1 Direct Com putation of the DFT, 449 6.1.2 D ivide-and-Conquer Approach to Com putation of the DFT. 450 6.1.3 Radix-2 FFT Algorithms. 456 6.1.4 Radix-4 FFT Algorithms. 465 6.1.5 Split-Radix FFT Algorithms, 470 6.1.6 Im plem entation of FFT Algorithms. 473 A p p lic a tio n s o f F F T A lg o rith m s 475 6.2.1 Efficient Com putation of the D FT of Two Real Sequences. 475 6.2.2 Efficient Com putation of the D FT of a Z N -Point Real Sequence, 476 6.2.3 Use of the FFT Algorithm in Linear Filtering and Correlation, 477 A L in e a r F ilte rin g A p p ro a c h to C o m p u ta tio n o f th e D F T 6.3.1 The Goertzel Algorithm, 480 6.3.2 The Chirp-z Transform Algorithm, 482 479
448
6.2
6.3
viii 6.4
Contents
Q u a n tiz a tio n E ffects in the C o m p u ta tio n o f th e D F T 486 6.4.1 Quantization Errors in the Direct Com putation of the DFT. 487 6.4.2 Quantization Errors in FFT Algorithms. 489 S u m m ary an d R e fe re n c e s P ro b le m s 494 493
6.5
7 IMPLEMENTATION OF DISCRETE-TIME SYSTEMS

7.1 7.2 S tru c tu res fo r th e R e a liz a tio n o f D isc re te -T im e S ystem s S tru c tu res fo r F IR System s 502 7.2.1 Direcl-Form Structure, 503 7.2.2 Cascade-Form Structures. 504 7.2.3 Frequency-Sampling S tructures1. 506 7.2.4 Lattice Structure. 511 S tru c tu re s for IIR S ystem s 519 7.3.1 Direct-Form Structures. 519 7.3.2 Signal Flow Graphs and Transposed Structures. 521 7.3.3 Cascade-Form Structures, 526 7.3.4 Parallel-Form Structures. 529 7.3.5 Lattice and Lattice-Ladder Structures for IIR Systems, 531 S tate-S p a ce System A n aly sis a n d S tru c tu re s 539 7.4.1 State-Space Descriptions of Systems Characterized by Difference Equations. 540 7.4.2 Solution of the State-Space Equations. 543 7.4.3 Relationships Between Input-O utput and State-Space Descriptions, 545 7.4.4 State-Space Analysis in the z-Domain, 550 7.4.5 Additional State-Space Structures. 554 R e p re s e n ta tio n of N u m b e rs 556 7.5.1 Fixed-Point Representation of Numbers. 557 7.5.2 Binary Floating-Point R epresentation of Numbers. 561 7.5.3 E rrors Resulting from R ounding and Truncation. 564 Q u a n tiz a tio n of F ilte r C o e fficien ts 569 7.6.1 Analysis of Sensitivity to Quantization of Filter Coefficients. 569 7.6.2 Q uantization of Coefficients in FIR Filters. 578 7.7 R o u n d -O ff E ffects in D igital F ilte rs 582 7.7.1 Limit-Cycle Oscillations in Recursive Systems. 583 7.7.2 Scaling to Prevent Overflow, 588 7.7.3 Statistical Characterization of Q uantization Effects in Fixed-Point Realizations of Digital Filters. 590 S u m m ary a n d R e fe re n c e s P ro b le m s 600 598 500
500
7.8
Contents
ix
DESIGN OF DIGITAL FILTERS

8.1 G e n e ra l C o n s id e ra tio n s 614 8.1.1 Causality and Its Implications. 615 8.1.2 Characteristics of Practical Frequency-Selective Filters. 619 D e sig n o f F IR F ilters 620 8.2.1 Symmetric and Antisym m eiric FIR Filters, 620 8.2.2 Design of Linear-Phase FIR Filters Using Windows, 623 8.2.3 Design of Linear-Phase FIR Filters by the Frequency-Sampling M ethod, 630 8.2.4 Design of Optimum Equiripple Linear-Phase FIR Filters, 637 8.2.5 Design of FIR Differentiators, 652 8.2.6 Design of Hilbert Transformers, 657 8.2.7 Comparison of Design M ethods for Linear-Phase FIR Filters, 662 D esig n o f I I R F ilters F ro m A n a lo g F iiters 666 8.3.1 IIR Filter Design by Approxim ation of Derivatives. 667 8.3.2 IIR Filter Design by Impulse Invariance. 671 8.3.3 IIR Filter Design by the Bilinear Transform ation, 676 8.3.4 The M atched-; Transform ation, 681 8.3.5 Characteristics of Commonly Used Analog Filters. 681 8.3.6 Some Examples of Digital Filter Designs Based on the Bilinear Transform ation. 692 F re q u e n c y T ra n s fo rm a tio n s 692 8.4.1 Frequency Transform ations in the Analog Dom ain, 693 8.4.2 Frequency Transform ations in the Digital Dom ain. 698 D esig n o f D ig ital F ilters B a sed on L e a st-S q u a re s M e th o d 8.5.1 Pade Approxim ation Method, 701 8.5.2 Least-Squares Design Methods, 706 8.5.3 FIR Least-Squares Inverse (W iener) Filters, 711 8.5.4 Design of IIR Filters in the Frequency Dom ain, 719 S u m m ary an d R e fe re n c e s P ro b le m s 726 724 701
614
8.2
8.3
8.4
8.5
8.6
SAMPLING AND RECONSTRUCTION OF SIGNALS

9.1 S am p lin g o f B a n d p a ss S ignals 738 9.1.1 R epresentation of Bandpass Signals, 738 9.1.2 Sampling of Bandpass Signals, 742 9.1.3 Discrete-Time Processing of Continuous-Time Signals, 746 A n a lo g -to -D ig ita l C o n v e rsio n 748 9.2.1 Sample-and-Hold. 748 9.2.2 Quantization and Coding, 750 9.2.3 Analysis of Q uantization Errors, 753 9.2.4 Oversampling A /D Converters, 756
738
9.2
Contents
9.3
D ig ita l-to -A n a lo g C o n v e rsio n 763 9.3.1 Sample and Hold, 765 9.3.2 First-Order Hold. 768 9.3.3 Linear Interpolation with Delay, 771 9.3.4 Oversampling D/A Converters, 774 S u m m ary an d R e fe re n c e s P ro b le m s 775 774
9.4
10 MULTIRATE DIGITAL SIGNAL PROCESSING

10.1 10.2 10.3 10.4 10.5 In tro d u c tio n 783 784 787 790
782
D e c im a tio n by a F a c to r D In te rp o la tio n by a F a c to r /
S am p lin g R a te C o n v e rsio n by a R a tio n a l F a c to r I ID F iite r D esig n an d Im p le m e n ta tio n for S a m p lin g -R ate C o n v e rsio n 792 10.5.1 Direct-Form FIR Filter Structures, 793 10.5.2 Polyphase Filter Structures, 794 10.5.3 Time-Variant Filter Structures. 800
10.6 10.7
M u ltistag e Im p le m e n ta tio n o f S a m p lin g -R a te C o n v e rs io n
806
S a m p lin g -R a te C o n v e rsio n o f B a n d p a ss S ignals 810 10.7.1 Decim ation and Interpolation by Frequency Conversion, 812 10.7.2 M odulation-Free Method for Decimation and Interpolation. 814 S a m p lin g -R a te C o n v e rsio n by an A rb itra ry F a c to r 815 10.8.1 First-O rder Approxim ation, 816 10.8.2 Second-Order Approximation (Linear Interpolation). 819 A p p lic a tio n s o f M u ltira te Signal P ro c essin g 821 10.9.1 Design of Phase Shifters. 821 10.9.2 Interfacing of Digital Systems with Different Sampling Rates, 823 10.9.3 Im plem entation of Narrowband Lowpass Filters, 824 10.9.4 Im plem entation of Digital Filter Banks. 825 10.9.5 Subband Coding of Speech Signals, 831 10.9.6 Q uadrature M irror Filters. 833 10.9.7 Transmultiplexers. 841 10.9.8 Oversampling A/D and D /A Conversion, 843 S u m m ary an d R e fe re n c e s
P ro b le m s 846
10.8
10.9
10.10
844
Contents
xi
11 LINEAR PREDICTION AND OPTIMUM LINEAR FILTERS

11.1 In n o v a tio n s R e p re s e n ta tio n o f a S ta tio n a ry R a n d o m P ro c e ss 852 11.1.1 Rational Power Spectra. 854 11.1.2 Relationships Between the Filter Param eters and the Autocorrelation Sequence, 855 F o rw a rd an d B a ck w ard L in e a r P re d ictio n 857 11.2.1 Forw ard Linear Prediction, 857 11.2.2 Backward Linear Prediction, 860 11.2.3 The Optimum Reflection Coefficients for the Lattice Forward and Backward Predictors, 863 11.2.4 Relationship of an A R Process to Linear Prediction. 864 S o lu tio n o f th e N o rm al E q u a tio n s 864 11.3.1 The Levinson-Durbin Algorithm. 865 11.3.2 The Schiir Algorithm. 868 P ro p e rtie s o f th e L in e a r P re d ic tio n -E rro r F ilte rs 873
852
11.2
11.3
11.4 11.5
A R L attice an d A R M A L a ttic e -L a d d e r F ilters 876 11.5.1 AR LaLtice Structure. 877 11.5.2 A RM A Processes and Lattice-Ladder Filters. 878 W ie n e r F ilters fo r F ilterin g a n d P re d ictio n 880 11.6.1 FIR W iener Filter, 881 11.6.2 Orthogonality Principle in Linear M ean-Square Estimation, 884 11.6.3 IIR W iener Filter. 885 11.6.4 Noncausal Wiener Filter. 889 S u m m ary an d R e fe re n c e s P ro b le m s 892 890
11.6
11.7
12 POWER SPECTRUM ESTIMATION

12.1 E stim a tio n o f S p e c tra from F in ite -D u ra tio n O b s e rv a tio n s o f Signals 896 12.1.1 Com putation of the Energy Density Spectrum. 897 12.1.2 Estim ation of the Autocorrelation and Power Spectrum of Random Signals: The Periodogram. 902 12.1.3 The Use of the DFT in Power Spectrum Estim ation, 906 N o n p a ra m e tric M e th o d s fo r P o w er S p ectru m E s tim a tio n 908 12.2.1 The B artlett Method: Averaging Periodograms, 910 12.2.2 The Welch Method: Averaging Modified Periodogram s, 911 12.2.3 The Blackman and Tukey Method: Smoothing the Periodogram, 913 12.2.4 Perform ance Characteristics of N onparam etric Power Spectrum Estim ators, 916
896
12.2
xii
Contents
12.2.5 Com putational Requirem ents of Nonparam etric Power Spectrum Estimates, 919 12.3 P a ra m e tric M e th o d s fo r P o w er S p e c tru m E stim a tio n 920 12.3.1 Relationships Between the A utocorrelation and the Model Param eters, 923 12.3.2 The Y ule-W alker M ethod for the A R Model Param eters, 925 12.3.3 The Burg M ethod for the A R Model Param eters, 925 12.3.4 Unconstrained Least-Squares M ethod for the A R Model Param eters, 929 12.3.5 Sequential Estim ation M ethods for the A R Model Param eters, 930 12.3.6 Selection of A R Model O rder, 931 12.3.7 MA Model for Power Spectrum Estim ation, 933 12.3.8 A R M A Model for Power Spectrum Estim ation, 934 12.3.9 Some Experim ental Results, 936 M in im u m V a rian ce S p ectral E stim a tio n 942
12.4 12.5
E ig e n an aly sis A lg o rith m s fo r S p e c tru m E stim a tio n 946 12.5.1 Pisarenko Harm onic Decom position M ethod, 948 12.5.2 Eigen-decomposition of the A utocorrelation Matrix for Sinusoids in White Noise, 950 12.5.3 MUSIC Algorithm. 952 12.5.4 ESPR IT Algorithm, 953 12.5.5 O rder Selection Criteria. 955 12.5.6 Experim ental Results, 956 S u m m ary an d R e fe re n c e s P ro b le m s 960 959
12.6
A RANDOM SIGNALS CORRELATION FUNCTIONS, AND POWER SPECTRA B RANDOM NUMBER GENERATORS C TABLES OF TRANSITION COEFFICIENTS FOR THE DESIGN OF LINEAR-PHASE FIR FILTERS D LIST OF MATLAB FUNCTIONS REFERENCES AND BIBLIOGRAPHY INDEX
A1 B1
C1 D1 R1
11
L j_ Preface
T h is b o o k w as d e v e lo p e d b ased on o u r te ach in g o f u n d e rg ra d u a te and g ra d u a te level co u rse s in d ig ital signal p ro cessin g o v er th e p a s t several y ears. In this b o o k w e p re se n t th e fu n d a m e n ta ls o f d isc re te -tim e signals, system s, and m o d e rn d ig ital p ro cessin g a lg o rith m s an d a p p lic a tio n s fo r stu d e n ts in electrical e n g in e e r ing. c o m p u te r en g in eerin g , a n d c o m p u te r science. T h e b o o k is su itab le fo r e ith e r a o n e -se m e s te r o r a tw o -se m e ste r u n d e rg ra d u a te level c o u rse in d isc re te system s a n d dig ital signal p ro cessin g . It is also in te n d e d fo r use in a o n e -se m e s te r first-year g ra d u a te -le v e l co u rse in digital signal processing. It is a ssu m ed th a t th e s tu d e n t in electrical and c o m p u te r e n g in e e rin g has h ad u n d e rg ra d u a te c o u rses in a d v an ce d calculus (in clu d in g o rd in a ry d iffe re n tia l e q u a tio n s). an d lin ear sy stem s fo r c o n tin u o u s-tim e signals, including an in tro d u c tio n to th e L ap lace tran sfo rm . A lth o u g h the F o u rie r se ries a n d F o u rie r tra n sfo rm s of p e rio d ic an d a p e rio d ic signals a re d escrib ed in C h a p te r 4, we ex p ect th a t m any s tu d e n ts m ay have h ad th is m a te ria l in a p rio r course. A b ala n c e d co v erag e is p ro v id e d of b o th th e o ry an d p ra c tic a l ap p licatio n s. A larg e n u m b e r o f w ell d esigned p ro b le m s a re p ro v id e d to h e lp th e s tu d e n t in m a ste rin g th e su b ject m a tte r. A so lu tio n s m a n u a l is av ailab le fo r th e b en efit o f th e in stru c to r an d can be o b ta in e d fro m th e p u b lish er. T h e th ird e d itio n o f th e b o o k covers basically th e sa m e m a te ria l as th e se c o n d e d itio n , b u t is o rg an ized d ifferen tly . T h e m a jo r d ifferen ce is in th e o rd e r in w hich th e D F T a n d F F T alg o rith m s are co v ered . B a sed o n su g g estio n s m a d e by se v era l rev iew ers, w e n o w in tro d u c e th e D F T a n d d esc rib e its efficient c o m p u ta tio n im m e d ia te ly fo llo w ing o u r tr e a tm e n t of F o u rie r analysis. T his re o rg a n iz a tio n h as also allo w ed us to elim in a te re p e titio n o f so m e to p ics c o n cern in g th e D F T and its ap p licatio n s. In C h a p te r 1 w e d escrib e th e o p e ra tio n s in v o lv ed in th e an alo g -to -d ig ital c o n v ersio n o f an alo g signals. T h e p ro cess o f sa m p lin g a sin u so id is d escrib ed in so m e d e ta il an d th e p ro b le m o f aliasing is ex p lain ed . Signal q u a n tiz a tio n an d d ig ita l-to -a n a lo g co n v ersio n a re also d escrib ed in g e n e ra l term s, b u t th e analysis is p re s e n te d in su b s e q u e n t c h a p te rs. C h a p te r 2 is d e v o te d e n tire ly to th e c h a ra c te riz a tio n a n d analysis o f lin e a r tim e -in v a ria n t (sh ift-in v arian t) d isc re te -tim e system s a n d d isc re te -tim e signals in th e tim e d o m a in . T h e co n v o lu tio n sum is d e riv e d a n d system s a re categ o rized a c co rd in g to th e d u ra tio n of th e ir im p u lse re sp o n s e as a fin ite -d u ra tio n im p u lse xiii
xiv
Preface
re sp o n se (F IR ) an d as an in fin ite -d u ra tio n im pulse re sp o n se ( II R ) . L in e a r tim ein v a ria n t sy stem s c h a ra c te riz e d by d ifferen ce e q u a tio n s are p r e s e n te d an d th e so lu tio n o f d ifferen ce e q u a tio n s w ith initial c o n d itio n s is o b ta in e d . T h e c h a p te r co n clu d es w ith a tre a tm e n t o f d isc re te -tim e c o rre la tio n . T h e z -tra n sfo rm is in tro d u c e d in C h a p te r 3. B o th th e b ila te ra l an d th e u n ila te ra l z -tra n sfo rm s are p re se n te d , a n d m e th o d s fo r d e te rm in in g th e in v erse z -tra n sfo rm are d esc rib e d . U se o f the z -tra n s fo rm in the analysis o f lin ear tim ein v a ria n t sy stem s is illu stra te d , an d im p o rta n t p ro p e rtie s o f system s, su c h as c a u s a l ity a n d stab ility , a re re la te d to z-d o m ain ch aracteristics. C h a p te r 4 tr e a ts th e analysis o f signals and sy stem s in th e fre q u e n c y d o m ain . F o u rie r se ries an d th e F o u rie r tra n sfo rm a re p re s e n te d fo r b o th co n tin u o u s-tim e an d d isc rete-tim e signals. L in e a r tim e -in v a ria n t (L T I) d isc rete sy stem s are c h a r a c terized in th e fre q u e n c y d o m a in by th e ir freq u e n c y resp o n se fu n c tio n an d th e ir re sp o n se to p e rio d ic an d a p e rio d ic signals is d e te rm in e d . A n u m b e r of im p o rta n t ty p es o f d isc re te -tim e system s are d esc rib e d , in clu d in g re s o n a to rs , n o tc h filters, co m b filters, all-p ass filters, a n d o scillato rs. T h e desig n of a n u m b e r of sim ple F IR a n d IIR filters is also co n sid ered . In a d d itio n , th e stu d e n t is in tro d u c e d to th e co n c e p ts o f m in im u m -p h a se , m ix ed -p h ase, an d m a x im u m -p h a se system s an d to th e p ro b le m o f d e c o n v o lu tio n . T h e D F T . its p ro p e rtie s an d its a p p licatio n s, a re th e topics c o v e re d in C h a p te r 5. T w o m e th o d s a re d e sc rib e d fo r using th e D F T to p e rfo rm lin e a r filtering. T h e use o f th e D F T to p e rfo rm fre q u e n c y analysis o f signals is also d escrib ed . C h a p te r 6 co v ers th e efficient c o m p u ta tio n o f th e D F T . In c lu d e d in this c h a p te r are d e sc rip tio n s o f radix-2, ra d ix -4, a n d sp lit-ra d ix fast F o u rie r tra n sfo rm (F F T ) alg o rith m s, a n d a p p lic a tio n s o f th e F F T a lg o rith m s to th e c o m p u ta tio n o f c o n v o lu tio n a n d c o rre la tio n . T h e G o e rtz e l alg o rith m a n d the ch irp -z tra n sfo rm are in tro d u c e d as tw o m e th o d s fo r c o m p u tin g th e D F T using lin e a r filtering. C h a p te r 7 tre a ts th e re a liz a tio n o f I I R an d F IR system s. T h is tre a tm e n t in clu d es d irect-fo rm , cascad e, p a ra lle l, lattice, a n d la ttic e -la d d e r re a liz a tio n s. T h e c h a p te r in clu d es a tr e a tm e n t o f sta te -sp a c e analysis an d s tru c tu re s fo r d isc rete-tim e system s, an d ex am in es q u a n tiz a tio n effects in a d igital im p le m e n ta tio n o f F IR and I IR system s. T e c h n iq u e s fo r d esign o f digital F IR a n d IIR filters are p r e s e n te d in C h a p te r 8. T h e d esign te c h n iq u e s in clu d e b o th d irect design m e th o d s in d isc re te tim e an d m e th o d s involv in g th e co n v ersio n o f an a lo g filters in to digital filters by v ario u s tra n sfo rm a tio n s. A lso tre a te d in this c h a p te r is th e d esig n o f F I R a n d IIR filters by le a st-sq u a re s m e th o d s. C h a p te r 9 fo cu ses o n th e sam pling o f c o n tin u o u s-tim e sig n a ls a n d th e r e c o n s tru c tio n o f such signals fro m th e ir sam ples. In th is c h a p te r, w e d eriv e th e sam p lin g th e o re m fo r b a n d p a ss co n tin u o u s-tim e -sig n a ls an d th e n co v e r th e A /D an d D /A co n v ersio n te c h n iq u e s, including o v e rsam p lin g A /D a n d D /A co n v erters. C h a p te r 10 p ro v id e s an in d e p th tre a tm e n t o f sa m p lin g -ra te c o n v ersio n and its a p p lic a tio n s to m u ltira le d ig ital signal p ro cessin g . In a d d itio n to d escrib in g d e c im atio n a n d in te rp o la tio n by in te g e r facto rs, we p re s e n t a m e th o d o f sa m p lin g -rate
Preface
xv
co n v e rsio n by an a rb itra ry facto r. S ev eral a p p licatio n s to m u ltira te signal p ro c e ss ing a re p re s e n te d , in clu d in g th e im p le m e n ta tio n o f d igital filters, su b b a n d cod in g o f sp e ech sig n als, tra n sm u ltip le x in g , an d o v ersam p lin g A /D a n d D /A c o n v e rte rs. L in e a r p re d ic tio n an d o p tim u m lin e a r (W ien er) filters a re tr e a te d in C h a p te r 11. A lso in clu d ed in this c h a p te r are d escrip tio n s o f th e L e v in s o n -D u rb in alg o rith m a n d Schiir a lg o rith m fo r solving th e n o rm a l e q u a tio n s , as w ell as th e A R la ttic e a n d A R M A la ttic e -la d d e r filters. P o w e r sp e c tru m e stim a tio n is th e m ain to p ic of C h a p te r 12. O u r co v erag e in clu d es a d e s c rip tio n o f n o n p a ra m e tric an d m o d el-b ased (p a ra m e tric ) m e th o d s. A lso d e s c rib e d a re e ig e n -d e c o m p o sitio n -b a se d m e th o d s, in clu d in g M U S IC an d E S P R IT . A t N o r th e a s te r n U n iv ersity , w e h av e u se d th e first six c h a p te rs o f this b o o k fo r a o n e -se m e s te r (ju n io r level) c o u rse in d isc rete sy stem s a n d d ig ital signal p r o cessing. A o n e -s e m e s te r se n io r level c o u rse fo r stu d e n ts w h o h av e h a d p rio r e x p o su re to d isc rete sy stem s can u se th e m a te ria l in C h a p te rs 1 th ro u g h 4 for a q u ick rev iew a n d th e n p ro c e e d to co v er C h a p te r 5 th ro u g h 8. In a first-v ear g ra d u a te level c o u rse in digital signal p ro cessin g , th e first five c h a p te rs p ro v id e th e s tu d e n t w ith a goo d rev iew of d isc re te -tim e system s. T h e in stru c to r can m o v e q u ick ly th ro u g h m o st o f th is m aterial a n d th e n co v e r C h a p te rs 6 th ro u g h 9, fo llo w ed by e ith e r C h a p te rs 10 and 11 o r by C h a p te rs 11 an d 12. W e h a v e in c lu d e d m an y ex am p les th ro u g h o u t th e b o o k an d a p p ro x im a te ly 500 h o m e w o rk p ro b le m s. M an y o f th e h o m ew o rk p ro b le m s can b e so lv ed n u m e r ically on a c o m p u te r, using a so ftw are p ack ag e such as M A T L A B . T h e se p r o b lem s a re id e n tifie d by an asterisk . A p p e n d ix D co n tain s a list o f M A T L A B fu n c tio n s th a t th e s tu d e n t can use in solving th e se p ro b lem s. T h e in s tru c to r m ay also w ish to c o n s id e r th e u se o f a s u p p le m e n ta ry b o o k th a t c o n ta in s c o m p u te r b ased exercises, su c h as th e b o o k s Digilal Signal Processing Us ing M A T L A B (P.W .S. K e n t, 1996) by V. K. In g le a n d J. G . P ro a k is a n d C o m p u te r- B a s e d Exercises f o r S ignal P ro cessing Using M A T L A B (P re n tic e H all, 1994) by C. S. B u rru s e t al. T h e a u th o rs a re in d e b te d to th e ir m an y facu lty c o lleag u es w ho h av e p ro v id e d v alu ab le su g g e stio n s th ro u g h review s o f the first an d se co n d ed itio n s o f this b o o k . T h e se in clu d e D rs. W . E . A le x a n d e r, Y. B re sle r, J. D e lle r, V. Ingle, C. K eller, H . L e v -A ri, L. M e ra k o s , W. M ik h a e l, P. M o n ticcio lo , C. N ikias, M . S ch etzen , H . T ru ssell, S. W ilso n , a n d M. Z o lto w sk i. W e a re also in d e b te d to D r. R , P ric e fo r re c o m m e n d in g th e in clu sion o f sp lit-ra d ix F F T alg o rith m s a n d re la te d su g g estio n s. F in ally , w e w ish to ac k n o w le d g e th e su g g e stio n s an d c o m m e n ts o f m an y fo rm e r g ra d u a te s tu d e n ts , a n d especially th o se by A . L. K ok, J. L in an d S. S rin id h i w ho assisted in th e p r e p a r a tio n o f several illu stra tio n s an d th e so lu tio n s m an u al. J o h n G . P ro a k is D im itris G , M a n o lak is
Introduction
D ig ital signal p ro cessin g is an are a o f science a n d e n g in e e rin g th a t h a s d ev e lo p e d rap id ly o v e r th e p ast 30 y ears. T his rap id d e v e lo p m e n t is a resu lt o f th e signif ican t ad v an ce s in digital c o m p u te r tech n o lo g y an d in te g ra te d -c irc u it fab rica tio n . T h e digital c o m p u te rs an d asso ciated digital h ard w are of th re e d e c a d e s ago w ere relativ ely larg e an d ex p en siv e and, as a co n seq u en ce, th e ir use w as lim ited to g e n e ra l-p u rp o s e n o n -re a l-tim e (o ff-line) scientific c o m p u ta tio n s an d business a p p licatio n s. T h e ra p id d ev e lo p m e n ts in in te g ra te d -c irc u it te c h n o lo g y , sta rtin g with m ed iu m -scale in te g ra tio n (M S I) an d p ro g ressin g to large-scale in te g ra tio n (L S I), a n d now , v ery -larg e-scale in te g ra tio n (V L S I) of e le c tro n ic circuits has sp u rre d th e d e v e lo p m e n t o f p o w erfu l, sm a ller, faster, an d c h e a p e r digital c o m p u te rs an d sp e cial-p u rp o se d igital h a rd w a re . T h e se in ex p en siv e an d re lativ ely fast digital c ir cuits h av e m a d e it p o ssib le to co n stru c t highly so p h istic a te d digital system s cap ab le o f p e rfo rm in g co m p lex digital signal p ro cessin g fu n ctio n s a n d tasks, w hich are u su ally to o difficult a n d /o r to o expensive to be p e rfo rm e d by an a lo g circuitry or a n alo g signal p ro cessin g system s. H e n c e m an y of th e signal p ro cessin g task s th a t w ere c o n v en tio n ally p e rfo rm e d by an alo g m e a n s a re realized to d a y by less ex p en siv e an d o fte n m o re re lia b le digital h a rd w a re . W e do n o t w ish to im ply th a t digital signal p ro cessin g is th e p ro p e r so lu tio n fo r all signal p ro cessin g p ro b lem s. In d e e d , fo r m a n y signals w ith e x tre m e ly w ide b a n d w id th s, real-tim e p ro cessin g is a re q u ire m e n t. F o r such signals, a n a log o r, p e rh a p s, o p tical signal p ro cessin g is th e only p o ssib le so lu tio n . H o w ev er, w h ere dig ital circuits are av ailab le an d h av e sufficient sp e e d to p e rfo rm th e signal p ro cessin g , th ey a re usually p re fe ra b le . N o t only d o d igital circuits yield c h e a p e r an d m o re re lia b le system s fo r signal p ro cessin g , th e y h av e o th e r a d v an tag es as w ell. In p a rtic u la r, digital pro cessin g h a rd w a re allow s p ro g ra m m a b le o p e ra tio n s. T h ro u g h so ftw are, on e can m o re easily m o d ify th e sig n al p ro cessin g fu n ctio n s to b e p e rfo rm e d by th e h a rd w a re . T h u s dig ital h a rd w a re a n d a s so ciated so ftw are p ro v id e a g re a te r d eg re e o f flexibility in sy stem d esign. A lso , th e re is o ften a h ig h e r o rd e r of p re c isio n ach iev ab le w ith d ig ital h a rd w a re an d so ftw are c o m p a re d w ith an alo g circu its a n d an alo g signal p ro cessin g system s. F o r all th e se re a so n s, th e re h as b e e n an explosive grow th in d ig ital signal p ro cessin g th e o ry a n d a p p licatio n s o v e r th e p ast th re e decades.
Introduction
Chap. 1
In this b o o k o u r o b jectiv e is to p re se n t an in tro d u c tio n o f th e basic analysis to ols an d te c h n iq u e s fo r d igital p ro cessin g o f signals. W e b eg in by in tro d u c in g so m e o f th e n ecessa ry term in o lo g y an d by d escrib in g th e im p o rta n t o p e ra tio n s asso ciated w ith th e p ro cess of c o n v ertin g an an alo g signal to d ig ital fo rm su itab le fo r d igital p ro cessin g . A s we shall se e, digital p ro cessin g o f a n a lo g signals has som e d raw b ack s. F irst, an d fo re m o st, c o n v ersio n o f an a n a lo g signal to digital fo rm , acco m p lish ed by sa m p lin g th e signal an d q u a n tiz in g th e sa m p le s, resu lts in a d isto rtio n th a t p re v e n ts us fro m re c o n stru c tin g th e o rig in a l a n a lo g signal fro m the q u a n tiz e d sam p les. C o n tro l o f th e a m o u n t o f th is d isto rtio n is ach ie v e d by p ro p e r choice o f th e sam p lin g ra te a n d th e p recisio n in th e q u a n tiz a tio n p ro cess. S eco n d , th e re a re finite p re c isio n effects th a t m u st be c o n s id e re d in th e d igital pro cessin g o f th e q u a n tiz e d sam p les. W hile th e se im p o rta n t issues are c o n s id e re d in som e d etail in this b o o k , th e em p h asis is on th e analysis a n d d esig n o f digital signal p ro cessin g sy stem s a n d c o m p u ta tio n a l te ch n iq u es.
1.1 SIGNALS, SYSTEMS, AND SIGNAL PROCESSING

A signal is d efin ed as any physical q u a n tity th a t varies w ith tim e, sp ace, o r any o th e r in d e p e n d e n t v ariab le o r variables. M a th em atic ally , we d e sc rib e a signal as a fu n ctio n o f o n e o r m o re in d e p e n d e n t variab les. F o r e x am p le, th e fu n ctio n s
* i( r ) = 5/
(1.1.1) S2(t) = 20 r d escrib e tw o signals, o n e th a t varies lin early w ith the in d e p e n d e n t v ariab le t (tim e) an d a seco n d th a t v aries q u a d ra tic a lly w ith t. A s a n o th e r ex a m p le , co n sid e r the fu n ctio n v) = 3x + 2 x y + 1 0 y 2 (1.1.2) T his fu n ctio n d escrib es a signal o f tw o in d e p e n d e n t v a riab les x a n d y th a t could r e p re s e n t th e tw o sp a tia l c o o rd in a te s in a p lan e. T h e signals d e sc rib e d by (1.1.1) an d (1.1.2) b e lo n g to a class o f signals th a t are p recisely d efin ed by specifying th e fu n c tio n a l d e p e n d e n c e on th e in d e p e n d e n t v ariab le. H o w ev er, th e re are cases w h ere such a fu n c tio n a l re la tio n sh ip is u n k n o w n o r to o highly c o m p licated to be o f any p ractical use. F o r ex am p le, a sp e ech signal (see Fig. 1.1) c a n n o t be d e s c rib e d fu n ctio n ally by ex p ressio n s such as (1.1.1). In g e n eral, a se g m e n t o f sp e ech m ay be re p re se n te d to a high d eg re e o f accu racy as a sum of se v era l sin u so id s o f d iffe re n t am p litu d e s a n d freq u e n cies, th a t is, as N
A j ( t ) s i n [ 2 ; r f } ( r ) f + #,(/)]
(1.1.3)
i=i w h ere {/!,(/)}, {F ,(r)j, a n d {t9,(r)} a re th e se ts of (p o ssib ly tim e -v a ry in g ) a m p litu d es, freq u e n cies, an d p h a se s, resp ectiv ely , o f th e sinusoids. In fact, o n e w ay to in te rp re t th e in fo rm a tio n c o n te n t o r m essag e co n v ey ed by an y sh o rt tim e se g m e n t o f th e
Sec. 1.1
Signals, Systems, and Signal Processing
^ #
#
S
I
A
Th
... ^
ft
N D
i j ^
---------- ,|* y y y > v y y m w
'
'r r m
w m
1
Figure 1.1 Example of a speech signal.
' W W W .......... ............
sp e e c h signal is to m e a s u re the a m p litu d es, freq u e n cies, a n d p h a se s c o n ta in e d in th e sh o rt tim e se g m e n t o f the signal. A n o th e r ex am p le o f a n a tu ra l signal is an e le c tro c a rd io g ra m (E C G ). Such a signal p ro v id e s a d o c to r w ith in fo rm a tio n a b o u t th e co n d itio n o f the p a tie n t's h e a rt. S im ilarly, an e le c tro e n c e p h a lo g ra m (E E G ) signal p ro v id es in fo rm a tio n a b o u t th e activ ity o f th e b rain . S p eech , e le c tro c a rd io g ra m , a n d e le c tro e n c e p h a lo g ra m signals a re ex am p les o f in fo rm a tio n -b e a rin g signals th a t evolve as fu n ctio n s o f a single in d e p e n d e n t v ariab le, n am elv , tim e. A n ex am p le o f a signal th at is a fu n ctio n o f tw o in d e p e n d e n t v ariab les is an im age signal. T h e in d e p e n d e n t v ariab les in th is case are th e sp atial c o o rd in a te s. T h e se a re b u t a few ex am p les o f th e co u n tless n u m b e r of n a tu ra l signals e n c o u n te re d in practice. A s so c ia te d w ith n a tu ra l signals are the m ean s by w hich such signals are g e n e ra te d . F o r ex am p le, sp e ech signals are g e n e ra te d by fo rcin g air th ro u g h th e vocal co rd s. Im ag es a re o b ta in e d by ex p o sin g a p h o to g ra p h ic film to a scene o r an o b ject. T h u s signal g e n e ra tio n is usually asso ciated w ith a sy stem th a t re sp o n d s to a stim u lu s o r fo rce. In a sp e ech signal, th e system consists o f th e vocal cords a n d th e vocal tra c t, also called th e vocal cavity. T h e stim ulus in c o m b in a tio n w ith th e sy stem is called a signal source. T h u s w e have sp eech so u rces, im ag es so u rces, an d v ario u s o th e r ty p es o f signal sources. A sy stem m ay also be defin ed as a physical device th a t p e rfo rm s an o p e r a tio n on a signal. F o r e x am p le, a filter u sed to red u c e th e n o ise an d in te rfe re n c e co rru p tin g a d e s ire d in fo rm a tio n -b e a rin g signal is called a system . In this case th e filter p e rfo rm s so m e o p e ra tio n (s ) on th e signal, w hich h as th e effect o f red u cin g (filterin g ) th e n o ise a n d in te rfe re n c e from th e d e sire d in fo rm a tio n -b e a rin g signal. W h en w e pass a signal th ro u g h a system , as in filterin g , w e say th a t we h av e p ro c e sse d th e signal. In this case th e p ro cessin g of th e signal involves filtering th e n o ise an d in te rfe re n c e fro m th e d e s ire d signal. In g e n e ra l, th e system is c h a ra c te riz e d by th e ty p e o f o p e ra tio n th a t it p e rfo rm s on th e signal. F o r ex am p le, if th e o p e ra tio n is lin ear, th e system is called linear. If th e o p e ra tio n o n th e signal is n o n lin e a r, th e system is said to be n o n lin e a r, a n d so fo rth . S uch o p e ra tio n s a re u su a lly re fe rre d to as signal processing.
Introduction
Chap. 1
F o r o u r p u rp o se s, it is c o n v en ien t to b r o a d e n th e d efin itio n o f a system to include n o t o n ly physical devices, b u t also so ftw are re a liz a tio n s o f o p e ra tio n s on a signal. In d igital p ro cessin g o f signals on a digital c o m p u te r, th e o p e ra tio n s p e r fo rm e d on a signal co n sist of a n u m b e r of m a th e m a tic a l o p e ra tio n s as specified by a so ftw are p ro g ram . In this case, th e p ro g ra m r e p re s e n ts an im p le m e n ta tio n o f the system in software. T h u s we h ave a system th a t is re a liz e d on a d igital c o m p u te r by m ean s o f a se q u en ce o f m a th e m a tic a l o p e ra tio n s; th a t is, w e h av e a digital signal p ro cessin g system realized in so ftw are. F o r e x am p le, a d ig ital c o m p u te r can be p ro g ra m m e d to p e rfo rm digital filtering. A lte rn a tiv e ly , th e d igital processing o n th e signal m ay be p e rfo rm e d by digital h ard w a re (logic circu its) co nfigured to p e rfo rm th e d e sire d specified o p e ra tio n s. In such a re a liz a tio n , w e h av e a physical d ev ice th a t p e rfo rm s th e specified o p e ra tio n s. In a b r o a d e r se n se, a digital system can be im p le m e n te d as a c o m b in a tio n o f digital h a rd w a re an d so ftw are, each of w hich p e rfo rm s its ow n set of specified o p e ra tio n s. T h is b o o k d eals w ith th e p ro cessin g o f signals by digital m e a n s, e ith e r in so ft w are o r in h a rd w a re . Since m an y of the signals e n c o u n te re d in p ra c tic e are analog, w e will also co n sid er th e p ro b lem of c o n v ertin g an a n a lo g signal in to a digital sig n al fo r pro cessin g . T h u s we will be d ealin g p rim a rily w ith d ig ital system s. T he o p e ra tio n s p e rfo rm e d by such a system can u su ally be specified m ath em atically . T h e m e th o d o r set o f ru les for im p le m e n tin g th e sy stem by a p ro g ra m th a t p e r fo rm s th e c o rre sp o n d in g m a th e m a tic a l o p e ra tio n s is called an algorithm. U sually, th e re are m an y w ays o r alg o rith m s by w hich a system can be im p le m e n te d , e ith e r in so ftw are o r in h a rd w a re , to p e rfo rm th e d e sire d o p e ra tio n s a n d c o m p u tatio n s. In p ra c tic e , we h av e an in te re st in devising a lg o rith m s th a t are c o m p u ta tio n a lly efficient, fast, an d easily im p lem en ted . T h u s a m a jo r to p ic in o u r stu d y o f digi tal signal p ro cessin g is th e discussion o f efficient a lg o rith m s fo r p e rfo rm in g such o p e ra tio n s as filterin g , c o rre la tio n , an d sp e c tra l analysis.
1.1.1 Basic Elements of a Digital Signal Processing System

M o st o f th e signals e n c o u n te re d in science an d e n g in e e rin g a re a n a lo g in n a tu re. T h a t is. th e signals a re fu n ctio n s of a c o n tin u o u s v a ria b le , such as tim e o r space, an d u su ally ta k e o n v alues in a co n tin u o u s ran g e. S uch signals m ay be p ro cessed directly by a p p ro p ria te an alo g system s (such as filters o r fre q u e n c y an aly zers) or fre q u e n c y m u ltip lie rs for th e p u rp o se of ch an g in g th e ir c h a ra c te ristic s o r ex tractin g so m e d esired in fo rm a tio n . In such a case w e say th a t th e signal h as b e e n p ro cessed d irectly in its an alo g fo rm , as illu strated in Fig. 1.2. B o th th e in p u t signal a n d the o u tp u t signal a re in an a lo g form .
Analog input signal
Analog signal processor
Analog output signal Figure 1.2 A nalog signal processing.
Sec. 1.1
Signals, Systems, and Signal Processing
5
Analog output signal
Analog input
signal
Digital input signal Figure 1.3 Digital output signal
Block diagram of a digital signal processing system.
D ig ital signal p ro cessin g p ro v id e s an a lte rn a tiv e m e th o d fo r p ro cessin g th e a n a lo g signal, as illu stra te d in Fig. 1.3. T o p e rfo rm th e p ro cessin g digitally, th e re is a n e e d fo r an in te rfa c e b e tw e e n th e an a lo g signal a n d th e digital p ro cesso r. T h is in te rfa c e is called an analog-to-digital ( A / D ) converter. T h e o u tp u t of th e A /D c o n v e rte r is a d ig ital signal th a t is a p p ro p ria te as an in p u t to th e d igital p ro cesso r. T h e dig ital signal p ro c e ss o r m ay be a larg e p ro g ra m m a b le digital c o m p u te r o r a sm all m ic ro p ro c e s so r p ro g ra m m e d to p e rfo rm th e d e s ire d o p e ra tio n s on th e in p u t signal. It m ay also be a h a rd w ire d digital p ro c e ss o r co n fig u red to p e rfo rm a specified se t o f o p e ra tio n s on th e in p u t signal. P ro g ra m m a b le m ach in es p r o v id e th e flexibility to ch an g e th e signal p ro cessin g o p e ra tio n s th ro u g h a ch an g e in th e so ftw are, w h e re a s h a rd w ire d m ach in es a re difficult to reco n fig u re. C o n s e q u e n tly , p ro g ra m m a b le signal p ro c e ss o rs a re in very c o m m o n use. O n th e o th e r h an d , w h en signal p ro cessin g o p e ra tio n s are w ell d efin ed , a h a rd w ire d im p le m e n ta tio n o f th e o p e ra tio n s can be o p tim ized , re su ltin g in a c h e a p e r signal p ro c e sso r a n d , u su ally , o n e th a t ru n s fa ste r th a n its p ro g ra m m a b le c o u n te rp a rt. In a p p li c atio n s w h e re th e d ig ital o u tp u t fro m th e d igital signal p ro c e sso r is to be given to th e u se r in an alo g form , such as in sp e ech co m m u n icatio n s, w e m ust p r o vid e a n o th e r in te rfa c e fro m th e digital d o m a in to th e a n a lo g d o m ain . S uch an in te rfa c e is called a digital-to-analog ( D / A ) converter. T h u s th e signal is p r o v id ed to th e u se r in an a lo g form , as illu stra te d in th e b lo ck d iag ram o f Fig. 1.3. H o w e v e r, th e re a re o th e r p ractical a p p lic a tio n s involving signal analysis, w h ere th e d e s ire d in fo rm a tio n is co n v ey ed in digital form a n d n o D /A c o n v e rte r is re q u ire d . F o r ex am p le, in th e d igital p ro cessin g o f r a d a r signals, th e in fo rm a tio n e x tra c te d fro m th e ra d a r signal, such as th e p o sitio n o f th e aircra ft a n d its sp e ed , m ay sim ply b e p rin te d on p a p e r. T h e re is n o n e e d fo r a D /A c o n v e rte r in th is case.
1.1.2 Advantages of Digital over Analog Signal Processing

T h e re a re m an y re a so n s w hy d ig ital signal p ro cessin g o f an an alo g signal m ay be p re fe ra b le to p ro cessin g th e signal directly in th e an a lo g d o m ain , as m e n tio n e d briefly e a rlie r. F irst, a digital p ro g ra m m a b le sy stem allow s flexibility in r e c o n figuring th e digital signal p ro cessin g o p e ra tio n s sim ply by ch anging th e p ro g ra m .
Introduction
Chap. 1
R e c o n fig u ra tio n o f an an a lo g system usually im plies a re d e sig n o f th e h a rd w a re follow ed by te stin g a n d v erification to see th a t it o p e ra te s p ro p e rly . A ccu racy c o n s id e ra tio n s also p lay an im p o rta n t role in d e te rm in in g th e fo rm o f th e signal p ro cesso r. T o le ra n c e s in an alo g c ircu it c o m p o n e n ts m a k e it e x tre m e ly difficult fo r th e system d esig n er to co n tro l th e accu racy o f an an a lo g signal p r o cessing system . O n th e o th e r h an d , a digital system p ro v id e s m uch b e tte r c o n tro l o f accu racy re q u ire m e n ts . Such re q u ire m e n ts , in tu rn , re s u lt in specifying th e a c cu racy r e q u ire m e n ts in th e A /D c o n v e rte r a n d th e d igital sig n a l p ro c e sso r, in te rm s o f w ord le n g th , flo atin g -p o in t v ersu s fix ed -p o in t arith m e tic , a n d sim ilar facto rs. D ig ita l signals are easily sto re d o n m a g n e tic m ed ia (ta p e o r disk) w ith o u t d e te rio ra tio n o r loss o f signal fidelity b e y o n d th a t in tro d u c e d in th e A /D co n v ersio n . A s a c o n se q u e n c e , th e signals b e c o m e tra n s p o rta b le an d can b e p ro cessed off-line in a re m o te la b o ra to ry . T h e digital signal p ro cessin g m e th o d also allow s for th e im p le m e n ta tio n o f m o re so p h istic a te d signal p ro cessin g alg o rith m s. It is usually very difficult to p e rfo rm p recise m a th e m a tic a l o p e ra tio n s on signals in a n a lo g fo rm b u t th ese sam e o p e ra tio n s can b e ro u tin e ly im p le m e n te d on a d ig ital c o m p u te r using so ftw are. In so m e cases a d igital im p le m e n ta tio n of th e signal p ro cessin g system is c h e a p e r th a n its an a lo g c o u n te rp a rt. T h e lo w er cost m ay be d u e to th e fact th a t th e dig ital h a rd w a re is c h e a p e r, o r p e rh a p s it is a re su lt o f th e flexibility fo r m o d ifications p ro v id e d by th e digital im p le m e n ta tio n . A s a c o n se q u e n c e o f th ese ad v a n ta g e s, d igital signal p ro c e ssin g has b e e n a p p lied in p ractical sy stem s co v erin g a b ro a d ra n g e of d iscip lin es. W e cite, fo r ex am p le, th e a p p licatio n o f d igital signal p ro cessin g te c h n iq u e s in sp e ech p ro cessin g an d signal tran sm issio n o n te le p h o n e ch an n els, in im age p ro c e ssin g an d tra n sm is sio n , in seism o lo g y an d geophysics, in oil e x p lo ra tio n , in th e d e te c tio n of n u c le a r ex p lo sio n s, in th e p ro cessin g of signals receiv ed fro m o u te r sp a ce, an d in a vast v ariety o f o th e r a p p licatio n s. S om e o f th e se a p p lic a tio n s a re cited in su b s e q u e n t ch ap ters. A s a lre a d y in d icated , h o w ev er, digital im p le m e n ta tio n has its lim itatio n s. O n e p ractical lim ita tio n is th e sp e ed o f o p e ra tio n o f A /D c o n v e rte rs a n d digital signal p ro cesso rs. W e shall see th a t signals hav in g e x tre m e ly w id e b a n d w id th s re q u ire fa st-sam p lin g -rate A /D c o n v e rte rs an d fast d igital signal p ro cesso rs. H e n c e th e re a re an alo g signals w ith larg e b a n d w id th s fo r w hich a digital p ro cessin g a p p ro a c h is b ey o n d th e s ta te of th e a rt o f digital h a rd w a re .
1.2 CLASSIFICATION OF SIGNALS

T h e m e th o d s we use in p ro cessin g a signal o r in an aly zin g th e re s p o n s e o f a system to a sig n al d e p e n d h eavily on th e ch a ra c te ristic a ttr ib u te s o f th e specific signal. T h e re a re te c h n iq u e s th a t ap p ly only to specific fam ilies o f signals. C o n seq u en tly , an y in v estig atio n in signal p ro cessin g sh o u ld sta rt w ith a classification o f th e signals in v o lv ed in th e specific ap p licatio n .
Sec. 1.2
Classification of Signals
1.2.1 Multichannel and Multidimensional Signals

A s e x p lain ed in S ectio n 1.1, a signal is d escrib ed by a fu n c tio n o f o n e o r m o re in d e p e n d e n t v ariab les. T h e v alue of th e fu n ctio n (i.e., th e d e p e n d e n t v ariab le) can be a re a l-v a lu e d sc alar q u a n tity , a co m p lex -v alu ed q u a n tity , o r p e rh a p s a v ecto r. F o r e x am p le, th e signal si( r ) = A sin37rr is a re a l-v a lu e d signal. H o w e v e r, th e signal s2(f) = A e ji7Tt = A cos 37t t j'A sin 3:r r is co m p lex v alu ed . In so m e a p p lic a tio n s, signals a re g e n e ra te d by m u ltip le so u rces or m u ltip le sen so rs. Such signals, in tu rn , can be re p re s e n te d in v e c to r fo rm . F ig u re 1.4 show s th e th re e c o m p o n e n ts of a v e c to r signal th a t re p re se n ts th e g ro u n d a c c e le ra tio n d u e to an e a r th q u a k e . T h is a c c e le ra tio n is the re su lt of th re e basic ty p es of elastic w aves. T h e p rim a ry (P ) w aves an d th e se co n d a ry (S) w aves p ro p a g a te w'ithin th e b o d y o f rock a n d a re lo n g itu d in al a n d tra n sv e rsa l, resp ec tiv ely . T h e th ird ty p e o f elastic w ave is called th e su rface w ave, b e c a u se it p ro p a g a te s n e a r th e g ro u n d su rface. If $*(/). k = 1. 2. 3. d e n o te s th e electrical signal from th e th se n so r as a fu n ctio n o f tim e, th e se t of p = 3 signals can be re p re se n te d by a v e c to r S?(f )< w h ere r si (O ' S;,(r) =
Si(t)
-Sl(t) J W e re fe r to such a v e c to r o f signals as a m u ltich a n n el signal. In e le c tro c a rd io g ra p hy. for ex am p le, 3 -lead an d 12-lead e le c tro c a rd io g ra m s (E C G ) are o ften used in p ractice, w hich resu lt in 3 -ch an n el a n d 12-channel signals. L e t us n o w tu rn o u r a tte n tio n to th e in d e p e n d e n t v a ria b le (s). If the signal is a fu n ctio n o f a single in d e p e n d e n t v ariab le, th e signal is called a o ne-d im en sio n a l signal. O n th e o th e r h a n d , a signal is called M -d i m e n s i o n a l if its v alu e is a fu n ctio n of M in d e p e n d e n t v ariab les. T h e p ic tu re sh o w n in Fig. 1.5 is an ex am p le of a tw o -d im e n sio n al signal, since th e in ten sity o r b rig h tn e ss I ( x . y) a t each p o in t is a fu n ctio n of tw o in d e p e n d e n t v ariab les. O n th e o th e r h a n d , a b la c k -a n d -w h ite telev isio n p ic tu re m ay be r e p r e se n te d as I ( x . y . t ) since th e b rig h tn e ss is a fu n ctio n of tim e. H e n c e th e T V p ic tu re m ay b e tr e a te d as a th re e -d im e n s io n a l signal. In c o n tra st, a co lo r T V p ic tu re m ay b e d e sc rib e d by th re e in te n sity fu n ctio n s of th e fo rm Ir (x, y. ?), Is (x. y. t ), a n d I i , ( x . y , t ) , c o rre sp o n d in g to th e b rig h tn e ss of the th re e p rin cip al colors (red . g re e n , b lu e) as fu n ctio n s o f tim e. H e n c e th e co lo r T V p ic tu re is a th re e -c h a n n e l, th re e -d im e n s io n a l signal, w hich can b e re p re s e n te d by th e v e c to r -/,(* ,> . O ' I U , y. t) . l b(x, v ,r ) _ In this b o o k we d e a l m ainly w ith sin g le-ch an n el, o n e -d im e n sio n a l real- or co m p lex -v alu ed signals a n d w e re fe r to th e m sim ply as signals. In m a th e m a tic a l
Introduction
Up
Chap. 1
/ % East
7 jJL______
South
bouth
I ____ i
] ____ I.
1 ____ 1 ____ I
f S waves
t Surface waves
_4 P waves
I I
i_______ I ,
I____ _ J _______I _______ I----------- 1 -----------1 ----------- 1 -----------1 ----------- 1 -----------1
r1 -2
10
12
14
16
18
20
22
24
26
28
30
Time (seconds) (b) Figure 1.4 Three components of ground acceleration measured a few kilometers from the epicenter of an earthquake. (From Earthquakes, by B. A . Bold. 1988 by W. H. Freeman and Company. Reprinted with permission of the publisher.)
te rm s th ese signals are d escrib ed by a fu n ctio n o f a single in d e p e n d e n t v ariable. A lth o u g h th e in d e p e n d e n t variab le n e e d n o t be tim e, it is c o m m o n p ractice to use t as th e in d e p e n d e n t v ariab le. In m an y cases th e signal p ro c e ssin g o p e ra tio n s and a lg o rith m s d e v e lo p e d in this tex t for o n e -d im e n sio n a l, sin g le-ch an n el signals can b e e x te n d e d to m u ltic h a n n e l an d m u ltid im e n sio n a l signals.
1.2.2 Continuous-Time Versus Discrete-Time Signals

Signals can b e fu rth e r classified in to fo u r d iffe re n t c a te g o rie s d e p e n d in g on the ch a ra c te ristic s o f th e tim e (in d e p e n d e n t) v a ria b le an d th e v alu es th ey tak e. Con tin u o u s -tim e signals o r a nalog signals a re d e fin ed for ev e ry value o f tim e an d
Sec. 1.2
Figure 1.5
Example of a two-dimensional signal.
th ey ta k e on v alu es in the co n tin u o u s in terv al ( a . b ). w h e re a can be oc a n d b can be oc. M a th em atic ally , th ese signals can be d e sc rib e d by fu n ctio n s o f a c o n tin u o u s v ariab le. T h e sp eech w av efo rm in Fig. 1.1 an d th e signals x i(r) = c o s 7 it, xj{t) = e ^ 1 '1 , oc < t < oq are ex am p les o f an alo g signals. Discrete-time signals a re d efin ed o n ly at c e rta in specific v alu es o f tim e. T h e se tim e in sta n ts n eed n o t be e q u id ista n t, b u t in p ractice th ey are usually ta k e n a t e q u a lly sp a ced in terv als fo r c o m p u ta tio n a l c o n v en ien ce an d m a th e m a tic a l tra c ta b ility . T h e signal x(t) = n = 0, 1 , 2 , . . . p ro v id es an ex am p le o f a d isc re te -tim e signal. If we use th e in d ex n o f th e d isc rete-tim e in sta n ts as th e in d e p e n d e n t v ariab le, th e signal v alu e b eco m es a fu n ctio n o f an in te g e r v ariab le (i.e., a se q u e n c e of n u m b e rs). T h u s a d isc re te -tim e signal can be re p re s e n te d m a th e m a tic a lly by a se q u e n c e of real o r c o m p lex n u m b ers. T o em p h asize the d isc rete-tim e n a tu r e o f a signal, w e sh all d e n o te such a signal as x{n) in ste a d o f x ( t ) . If th e tim e in stan ts t are e q u ally sp a ced (i.e., t = n T ), th e n o ta tio n x ( n T ) is also used. F o r ex am p le, th e se q u e n c e x(n) if n > 0 o th erw ise
( 1 .2 . 1 )
is a d isc re te -tim e signal, w hich is r e p re s e n te d g rap h ically as in Fig. 1.6. In ap p licatio n s, d isc rete-tim e signals m ay arise in tw o ways: 1. B y se lectin g v alu es o f an an alo g signal a t d isc re te -tim e in stan ts. T his p ro c e ss is called s am plin g an d is discussed in m o re d etail in S ectio n 1.4. A ll m e a s u r ing in stru m e n ts th a t ta k e m e a s u re m e n ts at a re g u la r in te rv a l o f tim e p ro v id e d isc rete-tim e signals. F o r ex am p le, th e signal x ( n ) in Fig. 1.6 can be o b ta in e d
10
x{n)
Introduction
Chap. 1
I I T
Figure 1.6 Graphical representation of the discrete time signal x[n) = 0.8" for n > 0 and x(n) = 0 for n < 0.
bv sa m p lin g th e an a lo g signal x ( t ) 0 .8 ', t > 0 an d x ( t ) = 0. t < 0 once ev ery seco n d .
2. By accu m u latin g a v a ria b le o v e r a p e rio d o f tim e. F o r e x a m p le , c o u n tin g th e

n u m b e r o f cars using a given s tre e t every h o u r, o r re c o rd in g th e v alu e of gold ev ery day, resu lts in d isc re te -tim e signals. F ig u re 1.7 show s a g rap h o f the W o lfer su n sp o t n u m b e rs. E a c h sam ple o f this d isc re te -tim e signal p ro v id es th e n u m b e r o f su n s p o ts o b se rv e d d u rin g an in te rv a l o f 1 y e a r.
1.2.3 Continuous-Valued Versus Discrete-Valued Signals

T h e v alu es o f a c o n tin u o u s-tim e or d isc re te -tim e signal can be c o n tin u o u s or d is crete. If a signal ta k e s on all po ssib le values on a finite or an infinite ran g e, it
Year Figure 1.7 W olfer annual sunspot num bers (1770-1869).
Sec. 1.2
11
is said to b e c o n tin u o u s-v a lu e d signal. A lte rn a tiv e ly , if th e signal ta k e s on v alu es fro m a finite se t o f p o ssib le values, it is said to be a d isc re te -v a lu e d signal. U su ally , th e s e v alu es a re e q u id ista n t a n d h en ce can be e x p ressed as an in te g e r m u ltip le of th e d istan c e b e tw e e n tw o successive values. A d isc re te -tim e signal hav in g a set of d isc rete v alu es is called a digital signal. F ig u re 1,8 show s a d igital signal th a t ta k e s o n o n e o f fo u r p o ssib le values. In o r d e r fo r a signal to be p ro c e sse d digitally, it m u st be d isc rete in tim e a n d its v alu es m u st b e d isc re te (i.e., it m ust b e a digital sig n al). If th e signal to b e p ro c e sse d is in an a lo g fo rm , it is c o n v e rte d to a digital signal by sam pling th e an alo g signal at d isc rete in stan ts in tim e, o b ta in in g a d isc re te -tim e signal, and th e n by q ua n tizin g its v alu es to a set o f d isc re te v alu es, as d e sc rib e d la te r in th e c h a p te r. T h e p ro cess o f co n v e rtin g a c o n tin u o u s-v a lu e d signal in to a d isc re te -v a lu e d signal, called quan tizatio n, is basically an a p p ro x im a tio n p ro cess. It m ay be acco m p lish ed sim ply bv ro u n d in g o r tru n c a tio n . F o r ex am p le, if th e allo w ab le signal v alu es in th e d ig ital signal are in teg ers, say 0 th ro u g h 15, the c o n tin u o u s-v a lu e signal is q u a n tiz e d in to th ese in te g e r values. T h u s th e signal v alue 8.58 will be a p p ro x im a te d by th e v alu e 8 if th e q u a n tiz a tio n p ro cess is p e rfo rm e d by tru n c a tio n o r by 9 if th e q u a n tiz a tio n p ro c e ss is p e rfo rm e d by ro u n d in g to th e n e a re s t in teg er. A n ex p la n a tio n o f th e a n a lo g -to -d ig ita l co n v ersio n p ro cess is given la te r in th e c h a p te r.
Figure 1.8
Digital signal with four different amplitude values.
1.2.4 Deterministic Versus Random Signals

T h e m a th e m a tic a l an aly sis an d p ro cessin g o f signals re q u ire s th e availability o f a m a th e m a tic a l d e sc rip tio n fo r th e signal itself. T h is m a th e m a tic a l d e scrip tio n , o fte n re fe rre d to as th e signal m o d e l, lead s to a n o th e r im p o rta n t classification of signals. A n y signal th a t can b e u n iq u ely d esc rib e d by an explicit m a th e m a tic a l ex p ressio n , a tab le o f d a ta , o r a w ell-defined ru le is called deterministic. T his te rm is used to em p h asize th e fact th a t all p ast, p re se n t, an d fu tu re values o f th e signal are kno w n precisely , w ith o u t an y u n c e rta in ty . In m a n y p ractical ap p lic a tio n s, h o w ev er, th e re are sig n a ls th a t e ith e r c a n n o t b e d esc rib e d to an y re a so n a b le d e g re e o f accu racy by explicit m a th e m a tic a l fo r m u las, o r such a d e sc rip tio n is to o c o m p licated to b e of any p ractical use. T h e lack
12
Introduction
Chap. 1
o f such a re la tio n sh ip im plies th a t such signals ev o lv e in tim e in an u n p re d ic ta b le m a n n e r. W e re fe r to th ese signals as ra n d o m . T h e o u tp u t o f a n oise g e n e ra to r, th e seism ic signal o f Fig. 1.4, an d th e sp e ech signal in Fig. 1.1 are ex am p les of ra n d o m signals. F ig u re 1.9 show s tw o signals o b ta in e d fro m th e sa m e n o ise g e n e ra to r an d th e ir asso ciated h isto g ram s. A lth o u g h th e tw o signals do n o t re s e m b le each o th e r visually, th e ir h isto g ra m s re v e a l som e sim ilarities. T his p ro v id e s m o tiv a tio n fo r
3 > 4 1 ---------------------------------------------------------- 1 ------- * ------------------- ----- ------ ---------------------------------------
200
400
600
800
1000
1200
1400
1600
(a)
400 f ----------------- ------------------- - -------------------------- ------------------- ------------------- 1
350 r
300 250 < 200 -
-j
150 100 -
(b)
Figure 1.9 tograms.
Two random signals from the same signal generator and their his
Sec. 1.2
4
13
----- ' 200
------------ '------------------------------------------- ------------------ --------1 400 600 800 1000 1200 1400 1600
< c)
Figure 1.9
C ontinued
th e an aly sis a n d d e sc rip tio n of ra n d o m signals using statistical te c h n iq u e s in stea d o f ex p licit fo rm u las. T h e m ath e m a tic a l fra m e w o rk fo r th e th e o re tic a l analysis of ra n d o m sig n als is p ro v id e d by th e th e o ry of p ro b a b ility a n d sto c h astic processes. S o m e b asic e le m e n ts o f this a p p ro a c h , a d a p te d to th e n e e d s o f this bo o k , are p re s e n te d in A p p e n d ix A . It sh o u ld b e em p h a siz e d a t th is p o in t th a t th e classification o f a real-world sig n al as d e te rm in is tic o r ra n d o m is n o t alw ays clear. S o m etim es, b o th a p p ro a c h e s le a d to m e a n in g fu l resu lts th a t p ro v id e m o re insight in to signal b eh av io r. A t o th e r
14
Introduction
Chap. 1
tim es, th e w ro n g classification m ay le a d to e rro n e o u s resu lts, since som e m a th e m atical to o ls m ay ap p ly o n ly to d e te rm in istic signals w hile o th e rs m ay apply only to ra n d o m signals. T h is will b e c o m e c le a re r as w e ex am in e specific m a th e m a tic a l tools.
1.3 THE CONCEPT OF FREQUENCY IN CONTINUOUS-TIME AND DISCRETE-TIME SIGNALS

T h e co n cep t o f fre q u e n c y is fam iliar to stu d e n ts in e n g in e e rin g a n d th e sciences. T h is co n cep t is basic in. fo r ex am p le, th e design of a rad io re c e iv e r, a high-fidelity sy stem , o r a sp e c tra l filter fo r co lo r p h o to g ra p h y . F ro m physics w e k n o w th a t freq u e n cy is closely re la te d to a specific ty p e o f p e rio d ic m o tio n called h arm o n ic o scillatio n , w hich is d e s c rib e d by sin u so id al fu n ctio n s. T h e c o n c e p t o f freq u e n cy is d irectly re la te d to th e c o n c e p t o f tim e. A ctu ally , it has th e d im en sio n of inverse tim e. T h u s w e sh o u ld e x p ect th a t th e n a tu re of tim e (c o n tin u o u s o r d isc rete) w ould affect th e n a tu re o f th e fre q u e n c y accordingly.
1.3.1 Continuous-Time Sinusoidal Signals

A sim ple h a rm o n ic o sc illatio n is m a th e m a tic a lly d escrib ed c o n tin u o u s-tim e sin u so id al signal: x a(t) = A cos(Q t + 0). oc < t < oc by th e follow ing (1.3.1)
show n in Fig. 1.10. T h e su b sc rip t a u se d w ith x { t ) d e n o te s an an a lo g signal. T his signal is co m p letely c h a ra c te riz e d by th re e p a ra m e te rs: A is th e a m p litu d e of the sin u so id . ft is th e fre q u e n c y in ra d ia n s p e r se co n d (rad /s), a n d 6 is th e p h a se in rad ian s. In ste a d o f ft, we o fte n use th e fre q u e n c y F in cycles p e r seco n d o r h e rtz (H z), w h ere Q In F In term s o f F. (1.3.1) can be w ritte n as x a(t) = A cos(2 n F t + 6 ), oo < t < oc (1.3.3) (1.3.2)
W e will use b o th fo rm s, (1.3.1) an d (1.3.3), in re p re s e n tin g sin u so id al signals.

x j t ) = A cos(2 nFt + 8)
Figure 1.10 Example of an analog sinusoidal signal.
Sec. 1.3
Frequency Concepts in Continuous-Discrete-Tim e Signals
15
T h e an a lo g sin u so id al signal in (1.3.3) is c h a ra c te riz e d by th e follow ing p r o p erties: A L F o r ev ery fixed v alu e o f th e fre q u e n c y F, x a(r) is p erio d ic. In d e e d , it can easily b e sh o w n , using e le m e n ta ry trig o n o m e try , th a t x a(.t + Tp ) = A(r) w h e re Tp = 1 / F is th e fu n d a m e n ta l p e rio d o f the sin u so id al signal. A 2 . C o n tin u o u s -tim e sin u so id al signals w ith d istin ct (d iffe re n t) fre q u e n c ie s a re th e m se lv e s d istin ct. A 3 . In c re a sin g th e fre q u e n c y F resu lts in an in crease in th e ra te o f o sc illatio n o f th e signal, in th e sense th a t m o re p e rio d s are included in a given tim e in terv al. W e o b se rv e th a t fo r F = 0. th e value Tp oc is co n sisten t w ith the fu n d a m e n ta l re la tio n F = 1 / T r . D u e to co n tin u ity o f th e tim e v a ria b le r, w e can in crease th e fre q u e n c y F, w ith o u t lim it, w ith a c o rre sp o n d in g increase in th e ra te o f oscillatio n . T h e re la tio n sh ip s we h ave d e sc rib e d fo r sin u so id al signals carry o v er to th e class o f co m p lex e x p o n e n tia l signals xa {t) - A e JlSi,+{" (1.3.4)
T h is can easily b e seen by e x p ressin g th e se signals in te rm s o f sinusoids using th e E u le r id en tity ej4: = cos < p i j sin < p (1.3.5)
B y d efin itio n , freq u e n c y is an in h e re n tly p o sitiv e physical q u an tity . T his is o b v io u s if we in te r p re t fre q u e n c y as the n u m b e r o f cycles p e r u n it tim e in a p e rio d ic signal. H o w e v e r, in m any cases, only fo r m a th e m a tic a l co n v en ien ce , w e n e e d to in tro d u c e n e g a tiv e freq u e n cies. T o see this w e recall th a t th e sin u so id al signal (1.3.1) m ay be e x p ressed as xa (t) = A c o s (^ r + 6 ) = j eJ(Q,+f>) + ~ e - J(a+9) (1.3.6)
w hich follow s fro m (1.3.5). N o te th a t a sin u so id al signal can be o b ta in e d by ad d in g tw o e q u a l-a m p litu d e c o m p lex -co n ju g ate e x p o n e n tia l signals, so m e tim e s called p h aso rs, illu stra te d in Fig. 1.1 1. A s t i m e p ro g re sse s th e p h a s o rs ro ta te in o p p o site d ire c tio n s w ith a n g u la r fre q u e n c ie s 2 ra d ia n s p e r se co n d . Since a positive f r e q u e n c y c o rre sp o n d s to co u n te rc lo c k w ise u n ifo rm a n g u la r m o tio n , a negative f r e q u e n c y sim ply c o rre sp o n d s to clockw ise a n g u la r m o tio n . F o r m a th e m a tic a l c o n v en ien ce , w e use b o th n e g a tiv e a n d po sitiv e freq u e n cies th ro u g h o u t th is b o o k . H e n c e th e fre q u e n c y ra n g e fo r a n a lo g sinusoids is oo < F < oo.
16
Introduction
Chap. 1
Re
Figure 1.11 Representation of a cosine function by a pair of complex-conjugate exponentials (phasors).
1.3.2 Discrete-Time Sinusoidal Signals

A d isc rete-tim e sin u so id al signal m ay be e x p ressed as x ( n ) A cos (ton + 8), oo < n < oc (1.3.7)
w h ere n is an in te g e r v ariab le, called th e sam p le n u m b e r. A is th e am plitu de o f the sin u so id , co is th e fr e q u e n c y in rad ian s p e r sa m p le , and 8 is th e p h a s e in radians. If in ste a d o f a> we use the freq u e n cy v ariab le / defin ed by a; = 2 n f th e re la tio n (1.3.7) b eco m es x( n ) A cos(2 n f n + 8). oc < n < oc (1.3.9) (1.3.8)
T h e freq u e n cy / h as d im en sio n s o f cycles p e r sa m p le . In S ectio n 1.4. w here we co n sid e r th e sam p lin g o f an alo g sinusoids, w e re la te th e fre q u e n c y v ariab le / o f a d isc re te -tim e sin u so id to th e fre q u e n c y F in cycles p e r se co n d fo r the an a lo g sin u so id . F o r th e m o m e n t we co n sid er th e d isc re te -tim e sin u so id in (1.3.7) in d e p e n d e n tly of th e co n tin u o u s-tim e sinusoid given in (1.3.1). F ig u re 1.12 show s a sin u so id w ith fre q u e n c y co n /6 ra d ia n s p e r sa m p le ( f ~ ~ cycles p e r sam ple) a n d p h ase 8 n / 3 .
x(n) - A cos (urn + 8)
Figure 1.12 Example of a discrete-time sinusoidal signal (w 7t / 6 and b 7r/3).
Sec. 1.3
Frequency Concepts in Continuous-Discrete-Tim e Signals
17
In c o n tra s t to co n tin u o u s-tim e sinusoids, th e d isc re te -tim e sin u so id s are c h a r a c te riz e d by th e fo llow ing p ro p e rtie s: B l . A discrete-time sinu so id is p er iod ic o n ly i f its fr e q u e n c y f is a rational n u m b e r . B y d efin itio n , a d isc rete-tim e signal x( n ) is p erio d ic w ith p e rio d N ( N > 0) if a n d only if x ( n + N ) = x(n ) fo r all n (1.3.10)
T h e sm a llest v alu e o f N fo r w hich (1.3.10) is tru e is called th e f u n d a m e n ta l p eriod . T h e p r o o f o f th e p erio d icity p ro p e rty is sim ple. F o r a sin u so id w ith fre q u e n c y /o to b e p e rio d ic , we sh o u ld have cos[27t /o( A7 + n) + 8} co s(2 ,t/o + 6) T h is re la tio n is tru e if an d only if th e re exists an in te g e r k such th a t 2 n f ) N = 2kn o r, eq u iv alen tly . /o = 4 N (1.3.11)
A cc o rd in g to (1.3.11). a d isc rete-tim e sin u so id al signal is p e rio d ic only if its f re q u e n c y /o can be ex p re sse d as th e ra tio o f tw o in teg ers (i.e.. / () is ra tio n a l). T o d e te rm in e th e fu n d a m e n ta l p e rio d N o f a p e rio d ic sin u so id , w e e x p ress its fre q u e n c y /o as in (1.3.11) an d cancel com m on facto rs so th a t k a n d N are relativ ely p rim e . T h e n th e fu n d a m e n ta l p e rio d o f the sin u so id is e q u a l to N. O b serv e th a t a sm all ch an g e in fre q u e n c y can resu lt in a large change in th e p erio d . F o r ex am p le, n o te th a t f \ = 3 1 /6 0 im plies th at N\ = 60, w h ereas f i = 3 0 /6 0 resu lts in Nz = 2. B 2. Disc rete-tim e sinu so ids whose fre q u en cies are separa ted by an integer m ultiple o f 2 n are identical. T o p ro v e this assertio n , let us c o n sid er th e sin u so id cos(oo + 0). It easily fo llow s th a t cos[(wo + 2 n )n + B\ = cos(wo/i + 2 n n + 9) co%{a)Qn + 9) A s a re su lt, all sin u so id al se q u en ces x k(n) A cos(a>tn + 8). w h ere Wk = cl> c + 2k n , 7r < o>o < n k = 0 ,1 ,2 ,... (1.3.13) (1.3.12)
are in distinguishable (i.e., identical). O n th e o th e r h a n d , th e se q u e n c e s o f any tw o sin u so id s w ith fre q u e n c ie s in th e ra n g e - n < a> < n or ^ < f < ~ a re distinct. C o n s e q u e n tly , d isc rete-tim e sinusoidal signals w ith fre q u e n c ie s M < n o r | / | < \
18
Introduction
Chap. 1
are u n iq u e. A n y se q u e n c e re su ltin g fro m a sin u so id w ith a fre q u e n c y M > n , or | / | > j , is id en tical to a se q u e n c e o b ta in e d fro m a sin u so id al signal w ith fre q u e n c y \co\ < n . B e c a u se o f th is sim ilarity, we call th e sin u so id h av in g th e freq u e n cy M > tt an alias o f a c o rre sp o n d in g sinusoid w ith fre q u e n c y jwj < n . T h u s w e re g a rd freq u e n cies in th e ra n g e tt < a> < tt, o r 1 < / < 1 as u n iq u e a n d all freq u e n cies |o>[ > t t , o r | / | > ~, as aliases. T h e r e a d e r sh o u ld n o tice th e d iffe re n c e b e tw e e n d isc rete-tim e sin u so id s an d co n tin u o u s-tim e sinusoids, w h e re th e la tte r re su lt in d istin ct signals fo r 2 o r F in th e e n tir e ran g e oc < 2 < oc o r oc < F < oc. B 3. The highest rate o f oscillation in a discrete-time sin u s o i d is attained whe n to 7i (or cu = tt) or, eq uivalently , f \ (or f = \ ) . T o illu stra te th is p ro p e rty , let u s in v estig ate th e c h a ra c te ristic s of th e sin u so id a l signal se q u e n c e x ( n ) = cos côn w h en th e fre q u e n c y v aries fro m 0 to tt. T o sim plify th e a rg u m e n t, we ta k e values o f (l> o = 0, 7 t / 8, tt/4 , jt/2 , n c o rre sp o n d in g to / = 0, 5, w hich resu lt in p erio d ic se q u e n c e s h av in g p e rio d s N = oc, 16, 8, 4, 2. as d e p ic te d in Fig. 1.13. W e n o te th a t th e p e rio d o f th e sinusoid d e c re a se s as th e fre q u e n c y in creases. In fact, we can see th a t th e ra te o f o scillatio n in crease s as th e fre q u e n c y increases.
,, xin)
Figure 1.13
Signal x ( n ) = c o s cu^n for various values of the frequency
cdq.
Sec. 1.3
Frequency Concepts in Continuous-Discrete-Time Signals
19
T o see w h at h a p p e n s for tt < ioq < 2tt . we co n sid e r th e sinusoids w ith fre q u e n c ie s a> \ = a > (> a n d 0 J2 2 n ojq. N o te th at as co\ varies from tt to 2n . a> z v aries fro m i r to 0. it can be easily se en th at = A cos co} n A cos won X2 (n) = A cos uhn A cos(27r coo)n = A c o s (coqii) x \ (h) H e n ce ur is an alias o f w\. If we h ad u sed a sine fu n ctio n in stea d o f a cosine fu n c tio n , th e resu lt w ou ld basically be th e sam e, ex cep t fo r a 180' p h a se d ifferen ce b etw een th e sin u so id s A]() and xi ( n ) . In any case, as we increase th e re lativ e freq u e n cy coo o f a d isc re te -tim e sin u so id from tt to 27r. its ra te of o sc illatio n d e creases. F o r coo = 2 tt th e re su lt is a c o n sta n t signal, as in th e case fo r oju = 0. O b v io u sly , fo r co{) = tt (o r f = k) w e h ave th e hig h est ra te o f oscillation. A s fo r th e case o f co n tin u o u s-tim e signals, n eg ativ e fre q u e n c ie s can be in tro d u c e d as w ell for d isc rete-tim e signals. F o r this p u rp o se w e use th e id en tity
x (n ) = Acos(con + 0) = (>Jiwn+0) +
(1.3.14)
(1.3.15)
Since d isc re te -tim e sinusoidal signals w ith fre q u e n c ie s th a t are se p a ra te d by an in teg er m u ltip le o f 27r are iden tical, it follow s th a t th e fre q u e n c ie s in any in terv al co] < a> < co\ + 2 tt c o n stitu te all the existing d isc rete-tim e sinusoids o r com plex e x p o n en tials. H en ce th e fre q u e n c y ran g e fo r d isc rete-tim e sinusoids is finite with d u ra tio n 2 n . U sually, we choose th e ran g e 0 < co < 2n o r tt < co < tt ({) < f < 1. 1 < / < | ) , w hich we call th e f u n d a m e n t a l range.
1.3.3 Harmonically Related Complex Exponentials

S in u so id al signals a n d com plex e x p o n e n tia ls play a m a jo r role in th e analysis o f signals an d system s. In so m e cases we deal w ith sets of h arm on ic a lly related c o m plex e x p o n e n tia ls (o r sin u so id s). T h e se are sets of p e rio d ic co m p lex e x p o n e n tia ls w ith fu n d a m e n ta l fre q u e n c ie s th a t are m u ltip le s o f a single positive freq u e n cy . A lth o u g h we confine o u r discussion to com plex e x p o n e n tia ls, th e sam e p r o p e r ties clearly hold fo r sin u so id al signals. W e co n sid e r h a rm o n ically re la te d co m p lex e x p o n e n tia ls in b o th co n tin u o u s tim e and d isc rete tim e.
Continuous-time exponentials.
h arm o n ically re la te d e x p o n e n tia ls are sk(t) = ejkno' = e ll7TkFn'
T h e basic signals for co n tin u o u s-tim e ,
jt = 0 . l . 2 . . . .
(1.3.16)
W e n o te th a t for ea c h value o f k, sÛ) is p erio d ic w ith fu n d a m e n ta l p e rio d 1 /(kFo) = Tp/ k o r fu n d a m e n ta l freq u e n cy kFo. Since a signal th a t is p e rio d ic w ith p e rio d Tp / k is also p e rio d ic w ith p erio d k ( T p/ k ) = Tp fo r any positive in te g e r k, w e see th a t all o f th e s*(r) h av e a co m m o n p e rio d of Tp, F u rth e rm o re , acco rd in g
20
Introduction
Chap. 1
to S ectio n 1.3.1, Fo is allo w ed to ta k e any v alu e an d all m e m b e rs of th e set are d istin ct, in th e se n se th a t if k\ ^ k2, th en 5*1 (7) ^ F ro m th e basic signals in (1.3.16) we can c o n s tru c t a lin e a r c o m b in a tio n of h arm o n ically re la te d co m p lex ex p o n e n tia ls o f th e form
cc x ( ' ) = k oc = k ~ oc SC Ckelkiltit ( 1 -3 -1 7 )
w h ere ck, k = 0, 1 , 2 . . . . are a rb itra ry co m p lex co n sta n ts. T h e signal x a(t) is p erio d ic w ith fu n d a m e n ta l p erio d Tp = l / f o , a n d its r e p re s e n ta tio n in term s o f (1.3.17) is called th e F ourier series ex p an sio n fo r x a (t). T h e co m p lex -v alu ed co n sta n ts are th e F o u rie r se rie s coefficients a n d th e signal sk (r) is called th e fcth h a rm o n ic o f x (l(t).
Discrete-time exponentials. Since a d isc re te -tim e co m p le x e x p o n e n tia l is p erio d ic if its relativ e fre q u e n c y is a ra tio n a l n u m b e r, w e ch o o se f Q 1/A' an d we d efine th e sets o f h arm o n ically re la te d co m p lex e x p o n e n tia ls by
sk (n) = ej2* kf" \ k = 0. 1 . 2 , . . . (1.3.18)
In c o n tra st to th e c o n tin u o u s-tim e case, we n o te th at sk+Nln) = eJ7* n'k+N,/N = e ^ s k (n) = sk (n) T his m ean s th a t, c o n siste n t w ith (1.3.10), th e re are only N d istin ct p e rio d ic com plex e x p o n en tials in th e se t d e sc rib e d by (1.3.18). F u rth e rm o re , all m e m b e rs of the set h av e a co m m o n p e rio d o f N sam ples. C learly, w e can ch o o se a n y co n secu tiv e A' co m p lex e x p o n e n tia ls, say from k = no to k no 4- N 1 to fo rm a h arm o n ically re la te d set w ith fu n d a m e n ta l freq u e n cy /(, = 1 / N . M o st o fte n , fo r co n v en ien ce , we ch o o se th e set th a t c o rre sp o n d s to no = 0 , th a t is, th e set sk(n) = ejl n k n / s . * = 0 . 1 . 2 .........N - 1 (1.3.19)
A s in th e case o f c o n tin u o u s-tim e signals, it is o b v io u s th a t th e lin e a r co m b in atio n \-l

k= 0
N- 1
Jt= 0
x ( n ) = c * s * ( n ) = Y L ckei2nkn' N resu lts in a p e rio d ic signal w ith fu n d a m e n ta l p e rio d N . A s w e shall see later, this is th e F o u rie r series re p re s e n ta tio n for a p e rio d ic d isc re te -tim e se q u e n c e w ith F o u rie r co efficien ts {q}. T h e se q u e n c e sk (n) is called th e /tth h a rm o n ic o f x(n ). Example 1.3.1 Stored in the memory of a digital signal processor is one cycle of the sinusoidal signal
x ( n ) = sin I
. ( 2nn
+ 6
where 6 2 n q / N , where q and N are integers.
Sec. 1.4
Analog-to-Digital and Digital-to-Analog Conversion
21
(a) Determ ine how this table of values can be used to obtain values of harmonically related sinusoids having the same phase.
(b) D eterm ine how this table can be used to obtain sinusoids of the same frequency
but different phase.

Solution
(a) Let
denote the sinusoidal signal sequence ( 2yrnk xk(n ) = sin I --------- ! V N
This is a sinusoid with frequency f k = k / N. which is harmonically related to x{n). But xk{n) may be expressed as xk{n) = sin
2 tt ( k n )
= x(kn) Thus we observe that ,vt (0) = .v(0). **(1) = x ( k ) . x k (2) = x ( 2 k ) . and so on. Hence the sinusoidal sequence can be obtained from the table of values of x ( n ) by taking every k th value of * ( ) . beginning with .v(0). In this m anner we can generate the values of all harmonically related sinusoids with frequencies fk = k / N for k = 0. 1 ....... N - 1 .
(b) We can control the phase 8 of the sinusoid with frequency j\ k / N by taking
the first value of the sequence from memory location q 9 N/ 2 tt. where q is an integer. Thus the initial phase 6 controls the starling location in the table and we wrap around the table each time the index (kn) exceeds N.
1.4 ANALOG-TO-DIGITAL AND DIGITAL-TO-ANALOG CONVERSION

M o st sig n als o f p ractical in terest, such as sp e ech , bio lo g ical signals, seism ic signals, ra d a r signals, so n a r signals, and v ario u s co m m u n ic a tio n s signals such as au d io a n d v id eo signals, are an alo g . T o p ro cess an a lo g signals by digital m e a n s , it is first n ecessa ry to c o n v e rt th e m into digital form , th a t is, to c o n v e rt th e m to a se q u e n c e o f n u m b e rs h av in g finite precision. T his p ro c e d u re is called analog-to-digital ( A / D ) co n v e r sio n , an d th e c o rre sp o n d in g devices a re called A / D converters ( A D C s ) . C o n c e p tu a lly , w e view A /D co n v ersio n as a th re e -s te p p ro cess. T his p ro cess is illu stra te d in Fig. 1.14. 1. S am p lin g . T his is th e c o n v ersio n o f a c o n tin u o u s-tim e signal in to a d is c re te tim e signal o b ta in e d by tak in g sa m p le s" o f th e c o n tin u o u s-tim e signal at d isc re te -tim e in stan ts. T h u s, if x a(t) is th e in p u t to th e sa m p le r, th e o u tp u t is x a ( n T ) = x ( n ), w h ere T is called th e s a m p lin g interval. 2. Q ua ntiza tio n . T h is is th e co n v ersio n o f a d isc re te -tim e c o n tin u o u s-v a lu e d signal in to a d isc rete-tim e, d isc rete-v alu ed (d ig ital) signal. T h e v alu e o f ea c h
22 A/D converter
Introduction
Chap. 1
01011...
"7
Analog signal
Discrete-time signal
Quantized signal
Digital signal
Figure 1.14 Basic parts of an analog-to-digital (A /D ) converter. signal sam p le is re p re s e n te d by a v alu e se lected from a finite set o f p o ssi b le values. T h e d ifferen ce b e tw e e n th e u n q u a n tiz e d sa m p le x ( n ) an d the q u a n tiz e d o u tp u t x q(n) is called th e q u a n tiz a tio n erro r. 3. Coding. In th e co d in g p ro cess, each d isc rete v alu e x q{n) is re p re s e n te d by a 6 -b it b in ary se q u en ce. A lth o u g h w e m o d e l th e A /D c o n v e rte r as a sa m p le r follow ed by a q u a n tiz e r an d co d er, in p ractice th e A /D co n v ersio n is p e rfo rm e d by a single device th at ta k e s x a(t) an d p ro d u c e s a b in a ry -c o d e d n u m b e r. T h e o p e ra tio n s of sa m p lin g an d q u a n tiz a tio n can be p e rfo rm e d in e ith e r o rd e r b u t. in p ractice, sa m p lin g is alw ays p e rfo rm e d b e fo re q u a n tiz a tio n . In m an y cases o f p ractical in te re st (e.g., sp e ech p ro cessin g ) it is d esirab le to co n v ert th e p ro cessed digital signals in to an a lo g form . (O b v io u sly , we can n o t listen to th e se q u e n c e of sa m p le s re p re se n tin g a sp e ech signal o r see th e n u m b ers co rre sp o n d in g to a T V signal.) T h e p ro cess o f co n v e rtin g a digital signal in to an an alo g signal is kno w n as digital-to-analog ( D / A ) co nversion. A ll D /A c o n v e rte rs c o n n ect th e d o ts " in a digital signal by p e rfo rm in g so m e kind of in te r p o la tio n , w hose accu racy d e p e n d s on th e q u ality of th e D /A c o n v ersio n process. F ig u re 1.15 illu strates a sim ple fo rm o f D /A c o n v ersio n , called a z e ro -o rd e r hold o r a sta ircase a p p ro x im a tio n . O th e r a p p ro x im a tio n s a re p o ssib le, such as lin early co n n ectin g a p a ir o f successive sa m p le s (lin e a r in te rp o la tio n ), fittin g a q u a d ra tic th ro u g h th re e successive sa m p le s (q u a d ra tic in te rp o la tio n ), an d so on. Is th e re an o p tim u m (ideal) in te rp o la to r? F o r signals hav in g a limited f r e q u e n c y co ntent (finite b an d w id th ), th e sa m p lin g th e o re m in tro d u c e d in th e follow ing se c tio n specifies the o p tim u m form o f in te rp o la tio n . S am p lin g a n d q u a n tiz a tio n are tr e a te d in th is sectio n . In p a rtic u la r, we d e m o n s tra te th a t sa m p lin g d o e s n o t re su lt in a loss of in fo rm a tio n , n o r d o es it in tro d u c e d isto rtio n in th e signal if th e signal b a n d w id th is finite. In p rin cip le , th e an alo g signal can b e re c o n stru c te d from th e sam ples, p ro v id e d th a t th e sam p lin g ra te is sufficiently high to avoid th e p ro b le m co m m o n ly called aliasing. O n th e o th e r h an d , q u a n tiz a tio n is a n o n in v e rtib le o r irre v e rsib le p ro c e ss th a t resu lts in signal d isto rtio n . W e shall sh o w th a t th e a m o u n t o f d isto rtio n is d e p e n d e n t on
Sec. 1.4
23
Original Signal
Staircase Approximation
/ / /
V *o /
/ -/ I I o
Time Figure 1.15 Zero-ordcr hold digital-to-analog (D /A ) conversion.
67-
th e accu racy , as m e a s u re d by th e n u m b e r of bits, in th e A /D c o n v ersio n process. T h e facto rs affe ctin g the choice of th e d esired accu racy of th e A /D c o n v e rte r are cost an d sam p lin g ra te . In g en eral, th e cost in crease s w ith an in crease in accuracy a n d /o r sa m p lin g rate.
1.4.1 Sampling of Analog Signals

T h e re are m an y ways to sa m p le an an a lo g signal. W e lim it o u r discussion to p er iod ic o r u n i f o r m s a m p li n g , w hich is th e ty p e of sa m p lin g used m o st o ften in p ractice. T h is is d escrib ed by the re la tio n
x(n) = xa (nT). o c < n < c c
(1.4.1)
w h ere x ( n ) is th e d isc re te -tim e signal o b ta in e d by ta k in g sa m p le s o f th e an alo g signal x aU) ev ery T se co n d s. T his p ro c e d u re is illu stra te d in Fig. 1.16. T h e tim e in te rv a l T b etw e e n successive sa m p les is called th e sa m p li n g p er io d o r sa m p le interval an d its recip ro c al 1 / 7 Fs is called the s a m p lin g rate (sam p les p e r second) o r th e sa m p lin g f r e q u e n c y (h ertz). P e rio d ic sa m p lin g estab lish es a re la tio n sh ip b e tw e e n th e tim e v ariab les t an d n o f c o n tin u o u s-tim e an d d isc re te -tim e signals, resp ec tiv ely . In d e e d , th e s e v a ri ab les a re lin early re la te d th ro u g h th e sa m p lin g p e rio d T or, e q u iv alen tly , th ro u g h th e sa m p lin g ra te Fs l / 7 \ as (1.4.2) A s a c o n s e q u e n c e o f (1.4.2), th e re exists a re la tio n sh ip b e tw e e n th e freq u e n cy v a ria b le F (o r Q) fo r an a lo g signals a n d the freq u e n c y v a ria b le / (o r co) for d isc re te -tim e signals. T o estab lish th is re la tio n sh ip , co n sid e r an an alo g sinusoidal signal o f th e fo rm
x B{t) = A c o s ( 2 t t F t + 8 )
(1.4.3)
24
Introduction
Chap. 1
Analog signal
*(n) = xa(nT) Fs = 1IT Sampler
Discrete-time signal
Figure 1.16
Periodic sampling of an analog signal.
w hich, w h en sa m p le d p erio d ically at a ra te Fs 1 / 7 sa m p le s p e r se co n d , yields x a( n T ) == x{ n) = A cos(27 t F n T -f 9) /2 n n F \ = A cos ( + 9 \ (1-4.4)
If we c o m p are (1.4.4) w ith (1.3.9). w e n o te th a t th e fre q u e n c y v ariab les F an d / a re lin early re la te d as (1.4.5) Fs=sampling frequency F=frequency of analoa f=frequency of digital signal (1.4.6) =relative or normalized frequency
Is or, eq u iv alen tly , as co = Q T
T h e re la tio n in (1.4.5) ju stifies th e n am e relative o r n o r m a li z e d f r e q u e n c y , w hich is so m e tim e s u sed to d esc rib e th e fre q u e n c y v a ria b le / . A s (1.4.5) im p lies, w e can use / to d e te rm in e th e fre q u e n c y F in h e rtz only if th e sa m p lin g fre q u e n c y Fs is k now n. W e recall fro m S ectio n 1.3.1 th a t th e ran g e o f th e fre q u e n c y v a ria b le F o r 2 fo r co n tin u o u s-tim e sin u so id s are oc < F < oo oc < 2 < 00 ( I .4 .7 )
H o w e v e r, th e situ a tio n is d iffe re n t fo r d isc re te -tim e sinusoids. F ro m S ectio n 1.3.2 w e recall th a t 2 c / < \ 2 (1.4.8)
n < co < n
B y su b stitu tin g fro m (1.4.5) a n d (1.4.6) in to (1.4.8), w e find th a t th e freq u e n cy o f th e co n tin u o u s-tim e sin u so id w h e n sa m p le d a t a ra te Fs = 1 / 7 m u st fall in
Sec. 1.4 th e ra n a e
Analog-to-Digitai and Digital-to-Analog Conversion
25
(1.4.9) or, e q u iv alen tly .

-----------=
n F, <
. -Q < _Ti F,
(1.4.10)
T h e se re la tio n s a re su m m a riz e d in T ab le 1.1.

TA B LE 1.1 RELATIO NS AM ONG FR E Q U E N C Y VARIABLES Discrete-time sienals
Continuous-time sienals
Q = radians sec
F
u .. \
c o = 2,t f radians cycles sample sample

1A 5 IA 1
u > = nT ,f=F /F,

/ ' - /
/
\ - o c < fi < oc oc < f' < oc
= to/T.F~f- Fs
.................. - ..... ..
t t /T < Q < ,T /r -f-2,/2 5 F < /-;/-
F ro m th ese re la tio n s w e o b se rv e th at the fu n d a m e n ta l d ifferen ce b etw een c o n tin u o u s-tim e an d d isc rete-tim e signals is in th e ir ran g e of v alu es o f th e fre q u en cy v a riab les F an d / , o r Q and w. P erio d ic sa m p lin g o f a c o n tin u o u s-tim e signal im p lies a m a p p in g of th e infinite freq u e n cy ran g e fo r th e v ariab le F (o r 2) in to a finite fre q u e n c y ran g e for the v ariab le / (o r a>). Since th e highest freq u e n cy in a d isc re te -tim e signal is co tt o r / = *, it follow s th a t, w ith a sa m p lin g rate Fs, th e c o rre sp o n d in g h ig h est values o f F and 2 are
~ 2T
_J _
(1.4.11)
^max ft Fs T h e re fo re , sa m p lin g in tro d u c e s an am b ig u ity , since th e h ig h est fre q u e n c y in a co n tin u o u s-tim e signal th a t can be u n iq u ely d istin g u ish ed w h en such a signal is sa m p le d at a ra te Fs = l / T is Fm & li F J 2 , o r Qmax = n F s. T o see w h at h a p p e n s to fre q u e n c ie s a b o v e F J 2, let us co n sid er the follow ing ex am p le. Example 1.4.1 The implications of these frequency relations can be fully appreciated by considering the two analog sinusoidal signals xi (!) cos 2ji(l0)t xi(t) cos2;r(50)f (1.4.12)
26
Introduction
Chap. 1
which are sampled at a rate Fs = 40 Hz. The corresponding discrete-time signals or sequences are
(1.4.13)
However. cos5,t/2 = cos(2^n + 7rn/2) co s 7 rn / 2 . Hence = *i(n)- Thus the sinusoidal signals are identical and, consequently, indistinguishable. If we are given the sampled values generated by cos( 7r/'2 )n, there is some ambiguity as to whether these sampled values correspond to x\{i) or xz(D- Since x2(r) yields exactly the same values as when the two are sampled at F, 40 samples per second, we say that the frequency F2 50 Hz is an alias of the frequency F\ = 10 Hz at the sampling rate of 40 samples per second. It is im portant to note that F2 is not the only alias of F]. In fact at the sampling rate of 40 samples per second, the frequency F3 = 90 Hz is also an alias of F], as is the frequency F4 = 130 Hz, and so on. All of the sinusoids cos2tt(Fj -f 40k)i. k = 1. 2. 3. 4 . . . . sampled at 40 samples per second, yield identical values. Consequently, they are all aliases of F\ = 10 Hz. In g en eral, th e sam pling o f a c o n tin u o u s-tim e sin u so id al signal x a(t) = A cos(27rF{)/ + 8) w ith a sam pling rate F v = 1 / T resu lts in a d isc re te -tim e signal x ( n ) = A cos(27r/ 0 -f 6)
(1.4.15) (1 .4 .1 4 )
w h ere /o = F()/F , is th e re la tiv e fre q u e n c y o f th e sin u so id . If w e assum e th a t - F s /2 < Fo < F J 2 . th e fre q u e n c y f 0 o f x ( n ) is in th e ran g e ^ < /o < L w hich is th e freq u e n cy ran g e fo r d isc re te -tim e signals. In this case, the re la tio n sh ip b etw een Fo an d f {) is o n e -to -o n e , a n d h en ce it is p o ssib le to identify (o r re c o n stru c t) th e a n alo g signal xa (t) from th e sa m p les x ( n) . O n th e o th e r h a n d , if th e sinusoids x a {t) = A w h ere Fk = F0 + k F s . k = l. 2 . (1.4.17)
c o s (27 z F kt
+ 6)
(1.4.16)
are sam p led a t a ra te F,, it is c le a r th a t th e fre q u e n c y F* is o u tsid e th e fu n d a m e n ta l fre q u e n c y ran g e Fs / 2 < F < F J 2 . C o n se q u e n tly , th e sa m p le d signal is
= A
c o s (27zn F o / F s
+ 6 + 2nkn)
A c o s ( 2 n f o n + 6)
Sec. 1.4
27
w hich is id en tical to th e d isc re te -tim e signal in (1.4.15) o b ta in e d by sam p lin g . (1.4.14). T h u s an infinite n u m b e r of c o n tin u o u s-tim e sin u so id s is re p re se n te d by sa m p lin g th e sa m e d isc re te -tim e signal (i.e.. by th e sam e se t o f sam p le s). C o n se q u e n tly , if w e a re given th e se q u en ce an am b ig u ity exists as to w hich c o n tin u o u s-tim e signal x a(t) th e se v alu es re p re se n t. E q u iv a le n tly , we can say th a t th e fre q u e n c ie s Fk F o + k F s, oo < k < oo (k in te g e r) a re in d istin g u ish a b le fro m th e fre q u e n c y Fo a fte r sa m p lin g an d h e n c e th ey are aliases o f Fo. T he re la tio n sh ip b e tw e e n th e fre q u e n c y v a riab les o f th e c o n tin u o u s-tim e a n d d isc rete-tim e signals is illu stra te d in Fig. 1.17. A n ex a m p le o f aliasing is illu stra te d in Fig. 1.18. w h e re tw o sinusoids w ith fre q u e n c ie s F 0 = | H z an d Fj = | H z yield id en tical sa m p le s w hen a sam p lin g ra te o f Fs 1 H z is used. F ro m (1.4.17) it easily follow s th a t fo r k = - 1 , Fo = F, + Fs = ( | + 1) H z = i H z.
Figure 1.17 Relationship between the continuous-time and discrete-time fre quency variables in the case of periodic sampling.
Figure 1.18
Illustration of aliasing.
28
Introduction
Chap. 1
Since Fsf2. w hich c o rre sp o n d s to w = tt, is th e hig h est fre q u e n c y th a t can be r e p re s e n te d u n iq u ely w ith a sam p lin g ra te Fs , it is a sim ple m a tte r to d e te rm in e th e m ap p in g o f any (alias) fre q u e n c y above Fs/2 (co tt) in to th e e q u iv a le n t freq u e n cy b elo w Fs /2. W e can use F J 2 or a) t t as the p iv o ta l p o in t a n d reflect or fo ld th e alias freq u e n c y to the ran g e 0 < w < tt. Since th e p o in t o f reflectio n is Fsj 2 (co = t t ) , th e fre q u e n c y F J 2 (cu = re) is called th e f o l d i n g frequency. Example 1.4.2 Consider the analog signal Xait) = 3 cos IOOtt/ (a) Determ ine the minimum sampling rate required to avoid aliasing.
(b) Suppose that the signal is sampled at the rate Fs = 200 Hz.
What is the discrete-time signal obtained after sampling? (c) Suppose that the signal is sampled at the rate Fs ~ 75 Hz. W hat is the discretetime signal obtained after sampling?
(d) What is the frequency 0 < F < FJ2 of a sinusoid that yields samples identical to those obtained in part (c)? Solution (a) The frequency of the analog signal is F 50 Hz. Hence the minimum sampling rate required to avoid aliasing is Ff = 100 Hz. (b) If the signal is sampled at Fs = 200 Hz. the discrete-time signal is lOOtf TT x{n) = j cos n j cos n 200 2 (c) If the signal is sampled at F, = 75 Hz. the discrete-time signal is 100?r A tt x(n) 3 cos j ^ n 3 cos n
3 cos ~n 3 (d) For the sampling rate of Fs ~ 75 Hz. we have F = f F, = 7 5 / The frequency of the sinusoid in part (c) is / |. Hence F = 25 Hz Clearly, the sinusoidal signal ya(i) 3 cos I n Ft 3 cos 50 7 ti sampled at Fs 75 samples/s yields identical samples. H ence F 50 Hz is an alias of F = 25 Hz for the sampling rate Fs = 75 Hz.
Sec. 1.4
29
1.4.2 The Sampling Theorem

G iv en an y an a lo g signal, how sh o u ld w e select th e sa m p lin g p e rio d T or, eq u iv alen tly , th e sa m p lin g ra te FJ. T o an sw er this q u e stio n , w e m u st h ave som e in fo rm a tio n a b o u t th e c h aracteristics of th e signal to be sa m p le d . In p a rtic u la r, we m u st h av e so m e g e n e ra l in fo rm a tio n c o n cern in g th e fr e q u e n c y con tent o f th e sig nal. S uch in fo rm a tio n is g en erally av ailab le to us. F o r e x am p le, w e k n o w g en erally th a t th e m a jo r fre q u e n c y co m p o n e n ts o f a sp eech signal fall b elo w 3000 H z. O n th e o th e r h a n d , telev isio n signals, in g e n eral, c o n ta in im p o rta n t fre q u e n c y co m p o n e n ts u p to 5 M H z. T h e in fo rm a tio n c o n te n t of such signals is c o n ta in e d in th e a m p litu d e s, fre q u e n c ie s, an d p h ases o f the v ario u s fre q u e n c y co m p o n e n ts, b ut d e ta ile d k n o w le d g e o f th e c h aracteristics of such signals is n o t a v a ilab le to us p rio r to o b ta in in g th e signals. In fact, the p u rp o se of p ro cessin g the signals is usually to e x tra c t th is d e ta ile d in fo rm a tio n . H o w ev er, if w e k n o w th e m ax im u m freq u e n cy c o n te n t o f th e g e n e ra l class o f signals (e.g.. th e class of sp e ech signals, the class o f v id eo signals, etc.). w e can specify th e sam pling ra te n ecessa ry to co n v ert the a n alo g signals to dig ital signals. L et us su p p o se th a t any an alo g signal can be r e p re s e n te d as a sum o f sin u so id s o f d iffe re n t a m p litu d e s, freq u e n cies, a n d p h ases, th a t is.
N
(1.4.18) w h ere N d e n o te s th e n u m b e r o f freq u e n cy c o m p o n e n ts. A ll signals, such as speech an d v id eo , len d th em se lv e s to such a re p re s e n ta tio n o v er an y sh o rt tim e segm ent. T h e a m p litu d e s, freq u e n cies, a n d p h ases usually ch an g e slow ly w ith tim e from one tim e se g m en t to a n o th e r. H o w e v e r, su p p o se th a t th e fre q u e n c ie s do n o t exceed som e k n o w n fre q u e n c y , say Fmax. F o r ex am p le, F max = 3000 H z fo r th e class o f sp e ech signals a n d Fmax = 5 M H z fo r telev isio n signals. Since th e m ax im u m freq u e n cy m ay v ary slightly fro m d iffe re n t re a liz a tio n s am o n g signals of any given class (e.g., it m ay vary slightly from s p e a k e r to sp e a k e r), w e m ay wish to e n su re th a t Fmax d o e s n o t ex ceed som e p re d e te rm in e d v alue by passin g th e an a lo g signal th ro u g h a filter th a t se v ere ly a tte n u a te s freq u e n cy c o m p o n e n ts ab o v e Fmax. T hus we a re c e rta in th a t no signal in the class co n tain s fre q u e n c y c o m p o n e n ts (having significant a m p litu d e o r p o w e r) above Fmax. In p ra c tic e , such filtering is com m only u sed p rio r to sam p lin g . F ro m o u r k n o w led g e o f Fmax, w e can se lect th e a p p ro p ria te sam pling rate. W e k n o w th a t th e h ig h est freq u e n cy in an an alo g signal th a t can be u n a m b ig u ously re c o n s tru c te d w h en th e signal is sa m p le d a t a ra te F, = 1 / T is F J 7. A ny fre q u e n c y a b o v e Fsf 2 o r b elo w - F J 2 resu lts in sa m p le s th a t a re id en tical w ith a c o rre sp o n d in g fre q u e n c y in th e ra n g e F J 2 < F < Fs/2. T o avoid th e am b ig u ities re su ltin g fro m aliasin g , we m u st se lect th e sa m p lin g ra te to be sufficiently high. T h a t is, w e m u st select F J 2 to be g re a te r th an Fmax. T h u s to avoid th e p ro b le m o f aliasin g , Fs is se le c te d so th a t Fs > 2 Fmax (1.4.19)
30
Introduction
Chap. 1
w h ere Fmax is th e larg est fre q u e n c y c o m p o n e n t in the a n a lo g signal. W ith the sa m p lin g ra te se le c te d in this m a n n e r, any fre q u e n c y c o m p o n e n t, say |F ;| < Fmax, in th e an alo g signal is m a p p e d in to a d isc re te -tim e sinusoid w ith a freq u e n cy
1
2 o r, eq u iv alen tly ,
Fs ~ 2
(1.4.20)
tt < a)j = 2n f < 7 r
(1.4.21)
S ince, | / | = \ o r \co\ = n is th e h ig h est (u n iq u e ) freq u e n c y in a d isc re te -tim e signal, th e choice o f sam p lin g ra te acco rd in g to (1.4.19) avoids th e p ro b le m of aliasing. In o th e r w o rd s, th e co n d itio n Fs > 2 Fmax e n s u re s th a t all th e sin u so id al c o m p o n e n ts in th e a n a lo g signal are m a p p e d in to c o rre sp o n d in g d isc re te -tim e freq u e n cy c o m p o n e n ts w ith fre q u e n c ie s in th e fu n d a m e n ta l in terv al. T h u s all the freq u e n cy c o m p o n e n ts o f th e a n alo g signal a re re p re s e n te d in sa m p le d fo rm w ith o u t am b i guity, a n d h en ce th e an alo g signal can be re c o n stru c te d w ith o u t d isto rtio n from th e sa m p le v alu es u sing an a p p r o p ria te in te rp o la tio n (d ig ita l-to -a n a lo g c o n v e r sio n ) m e th o d . T h e ap p ro p riate o r ideal in te rp o la tio n fo rm u la is specified by the s a m p ling theorem. Sam pling T h eorem . If the h ig h est fre q u e n c y c o n ta in e d in an an alo g signal x a (t) is Fmax = B a n d th e signal is sa m p le d at a ra te F, > 2 F max = 2 B. th e n A .u(r) can b e exactly re c o v e re d fro m its sa m p le values using th e in te rp o la tio n fu n ctio n s i n 2 jr B / 8(t) = T h u s jcfl(f) m ay be e x p ressed as * . ( ) * ( ' - ) w h ere x a(n / F s ) = x a( n T ) = Jt(rc) a re th e sa m p les o f x a(t). W h en th e sa m p lin g of x a(t) is p e rfo rm e d at the m in im u m sam p lin g ra te Fs = 2 B , th e re c o n stru c tio n fo rm u la in (1.4.23) b eco m es ^ / n \ sin2nB (t n flB )
(
_ , ( }
2n B t
(1-4.23)
, ,
zb)
(1'4 2 4 )
T h e sa m p lin g r a te F ^ = 2 B = 2 Fmax is called th e N y q u is t rate. F ig u re 1.19 illus tra te s th e ideal D /A c o n v ersio n p ro cess using th e in te rp o la tio n fu n c tio n in (1.4.22). A s can b e o b se rv e d from e ith e r (1.4.23) o r (1.4.24), th e re c o n stru c tio n o f x a(t) fro m th e se q u e n c e x ( n ) is a c o m p lic a te d p ro cess, involving a w e ig h te d sum o f the in te rp o la tio n fu n ctio n g (t) an d its tim e -sh ifte d v ersio n s g ( t n T ) fo r oo < n < oo, w h ere th e w eig h tin g facto rs a re th e sa m p le s x ( n ) . B e c a u se o f th e co m p lex ity an d th e infinite n u m b e r o f sa m p les re q u ire d in (1.4.23) o r (1.4.24), th e s e re c o n stru c tio n
Sec. 1.4

sample of ,v(n
31
(/[ ^ /
{ri l )l
fll
\t -r l 1 1
Figure 1.19 Ideal D /A conversion (interpolation).
fo rm u ias are p rim a rily o f th e o re tic a l in te re st. P ra ctical in te rp o la tio n m e th o d s are given in C h a p te r 9. Example 1.4.3 Consider the analog signal xu(r) = 3cos50;rz 10sin300;n cos 100tt?
What is the Nvquist rate for this signal? Solution The frequencies present in the signal above are F = 25 Hz. F: = 150 Hz. F, = 50 Hz
Thus Fnm = 150 Hz and according to (1.4.19), F > 2 Fm ax = 300 Hz The Nvquist rale is FA = 2 Fm ;. Hence Fs = 300 Hz Discussion It should be observed that the signal component 10sin300;r/. sampled at the Nvquist raie FA - = 300, results in the samples 10 sin 7r/j. which are identically zero. In other words, we are sampling the analog sinusoid at its zero-crossing points, and hence we miss this signal component completely. This situation would not occur if the sinusoid is offset in phase by some am ount 8. In such a case we have lOsinGOOin -ffl) sampled at the Nvquist rate FA - = 300 samples per second, which yields the samples
+ -c o 10 sin(7rn+ ^) = 10(sin n n cos & stth
sin f>)
= lOsin 6 cos nn
Thus if 6 ^ 0 or tt, the samples of the sinusoid taken at the zero. However, we still cannot obtain the correct amplitude the phase 9 is unknown. A simple rem edy that avoids this situation is to sample the analog signal at a rate higher than Example 1.4.4 Consider the analog signal
Nvquist rate are not all from the samples when potentially troublesome the Nvquist rate.
*a(t) = 3cos2000irf + 5 sin6000;rr + lOcos 12.000;?;
Introduction
Chap. 1
(a) What is the Nvquist rate for this signal?
(b) Assume now that we sample this signal using a sampling rate Fs = 5000
samples/s. What is the discrete-time signal obtained after sampling? (c) What is the analog signal y(r) we can reconstruct from the samples if we use ideal interpolation?
= 2.5 kHz
2
and this is the maximum frequency that can be represented uniquely by the sampled signal. By making use of (1.4.2) we obtain
= 3 cos 2jt( j )h + 5 sin 2n- (^ ) + 10 cos 2n(^)n = 3 c o s2 ;r(|)n + 5 sin 2 7 r(l ) 4- 10cos2,t(1 4- ^)u = 3cos2jr({)n 4- 5sin27r(1) 4- 1 0 c o s 2 ^ (|) Finally, we obtain x(n) = 13 cos2^({)/i - 5sin27r( = )fl The same result can be obtained using Fig. 1.17. Indeed, since F, = 5 kHz. the folding frequency is FJ2 = 2.5 kHz. This is the maximum frequency that can be represented uniquely by the sampled signal. From (1.4.17) we have Fti = Fk kFs. Thus Fo can be obtained by subtracting from Fk an integer m ultiple of Fs such that F t/2 < F0 < F J 2. The frequency F: is less than Fsf2 and thus it is not affected by aliasing. However, the other two frequencies are above the folding frequency and they will be changed by the aliasing effect. Indeed. F '2 = Fj ~ Fs = - 2 kHz
f;
= Fj Fs = 1 kHz h and h = ^ which are in agreem ent
From (1.4.5) it follows that /] = with the result above.
Sec. 1.4
A nalog-toD igital and Digital-to-Analog Conversion
33
(c) Since only the frequency components at 1 kHz and 2 kHz are present in the sampled signal, the analog signal we can recover is
x(t)
= 13 cos 2000^r - 5sin400()-Tf
which is obviously different from the original signal xU). This distortion of the original analog signal was caused by the aliasing effect, due to the low sampling rate used. A lth o u g h aliasin g is a pitfall to be av o id ed , th e re are tw o useful p ractical a p p licatio n s b ased on th e ex p lo itatio n of the aliasing effect. T h e se a p p licatio n s are th e stro b o sc o p e an d the sam pling oscilloscope. B o th in stru m e n ts are d esigned to o p e ra te as aliasin g devices in o rd e r to re p re se n t high fre q u e n c ie s as low f re q u en cies. T o e la b o ra te , co n sid er a signal w ith h ig h -freq u en cy c o m p o n e n ts confined to a given fre q u e n c y b an d B\ < F < B2. w h ere Bz B\ = B is d efined as the b a n d w id th o f th e signal. W e assum e th a t B < < B\ < B 2. T h is co n d itio n m ean s th a t th e fre q u e n c y c o m p o n e n ts in the signal are m uch larg er th an th e b an d w id th B of th e signal. Such signals are usually called p a ssb a n d or n a rro w b a n d signals. N ow . if this signal is sa m p le d at a rate Fs > 2B. b u t F^ << B\. th e n all th e f re q u en cy c o m p o n e n ts c o n ta in e d in the signal will be aliases of fre q u e n c ie s in the ran g e 0 < F < F J 2 . C o n seq u en tly , if we o b se rv e the freq u e n cy c o n te n t of the signal in th e fu n d a m e n ta l range 0 < F < F J 2 . we k now precisely the freq u e n cy co n te n t o f th e an a lo g signal since we k now the fre q u e n c y b an d B\ < F < B2 u n d e r c o n sid e ra tio n . C o n se q u e n tly , if the signal is a n a rro w b a n d (p a ss b a n d ) signal, we can re c o n stru c t th e o riginal signal from the sam p le s, p ro v id e d th a t the signal is sa m p le d at a ra te Fs > 2 B. w h ere B is th e b an d w id th . T h is s ta te m e n t c o n stitu te s a n o th e r fo rm o f th e sam pling th e o re m , w hich we call the p a s s b a n d f o r m in o rd e r to d istin g u ish it fro m th e p rev io u s form o f the sa m p lin g th e o re m , w hich ap p lies in g en eral to all ty p es of signals. T he la tte r is so m e tim es called th e bas e ban d f or m. T h e p a s s b a n d f o r m o f th e sam pling th e o re m is d escrib ed in d e ta il in S ectio n 9.1.2.
1.4.3 Quantization of Continuous-Amplitude Signals

A s w e h av e se en , a dig ital signal is a se q u en ce of n u m b e rs (sa m p le s) in w hich each n u m b e r is re p re s e n te d by a finite n u m b e r of digits (finite p recisio n ). T h e p ro c e ss o f co n v ertin g a d isc rete-tim e c o n tin u o u s-a m p litu d e signal in to a dig ital signal by ex p ressin g each sa m p le value as a finite (in ste a d of an infinite) n u m b e r o f d igits, is called quant izati on. T he e rro r in tro d u c e d in re p re se n tin g th e c o n tin u o u s-v a lu e d signal by a finite set o f d isc rete v alu e levels is called quant i zat i on error o r quant i zat i on noise. W e d e n o te th e q u a n tiz e r o p e ra tio n o n th e sa m p le s x ( n ) as Q[x{n)] an d let x q{n) d e n o te th e se q u en ce o f q u a n tiz e d sam ples a t th e o u tp u t o f th e q u a n tiz e r. H e n ce
Xq(n)
= Q[x(n)]
34
Introduction
Chap. 1
T h e n th e q u a n tiz a tio n e r ro r is a se q u e n c e eq (n) d efin ed as th e d iffe re n c e b etw e e n th e q u a n tiz e d v alu e a n d th e actu al sa m p le value. T h u s eq (n) = x q {n) - x ( n ) W e illu strate th e q u a n tiz a tio n p ro cess w ith a n ex am p le. d isc rete-tim e signal (1.4.25) L e t us co n sid e r the
o b ta in e d by sa m p lin g th e an a lo g e x p o n e n tia l signal x a( t ) = 0 .9 ', t > 0 w ith a sam p lin g freq u e n cy f , = 1 H z (see Fig. 1.20(a)). O b se rv a tio n o f T a b le 1.2, w hich show s th e v alu es o f th e first 10 sa m p les o f x ( n ) , rev eals th a t th e d e sc rip tio n o f the sam p le v alu e x{n) re q u ire s n significant digits. It is o b v io u s th a t th is signal can n o t
(a)
Figure 1.20
Illustration of quantization.
Sec. 1.4

NU M ER IC A L ILLU S TR A TIO N O F Q U A N TIZA TIO N W ITH ONE
35
TA B LE 1.2
S IG N IF IC A N T D IG IT USING T R U N C A T IO N O R R O UN DING
x ( n) n
D iscrete-tim e signal 1 0.9 0.81 0.729 0.6561 0.59049 0.531441 0.4782969 0.43046721 0.387420489
x,(n ) (Truncation)
1.0 0.9 0.8 0.7 0.6 0.5 0.5 0.4 0.4 0.3
.voi) (Rounding)
1.0 0.9 0.8 0.7 0.7 0.6 0.5 0.5 0.4 0.4
-
(Rounding)
0.0 0.0 0.01 0.029 0.0439 0.00951 0.031441 0.0217031 0.03046721 0.012579511
0 1 2 3 4 5 6 7 8 9
be p ro cessed by u sing a c a lcu lato r o r a digital c o m p u te r since only the first few sam p les can be sto re d an d m a n ip u la te d . F o r ex am p le, m ost c alcu lato rs process n u m b e rs w ith only eig h t significant digits. H o w e v e r, let us assum e th a t w e w an t to use only o n e significant digit. To elim in ate th e excess digits, w e can e ith e r sim ply d iscard th em ( tru n ca tion ) o r dis card th e m bv ro u n d in g th e resu ltin g n u m b e r (ro un din g). T h e resu ltin g q u an tized signals x q (n) a re show n in T ab le 1.2. W e discuss only q u a n tiz a tio n by ro u n d in g , alth o u g h it is ju st as easy to tr e a t tru n c a tio n . T h e ro u n d in g p ro c e ss is g raphically illu stra te d in Fig. 1.20b. T h e values allo w ed in th e digital signal are called the quan tizatio n levels, w h ereas the d istan c e A b etw een tw o successive q u a n tiz a tio n levels is called th e q uantization step siz e o r resolution. T h e ro u n d in g q u a n tiz e r assigns each sa m p le of x ( n ) to th e n e a re s t q u a n tiz a tio n level. In c o n tra st, a q u a n tizer th a t p e rfo rm s tru n c a tio n w ould have assigned each sa m p le of jc(/z) to the q u a n tiz a tio n level b elo w it. T h e q u a n tiz a tio n e r ro r eq (n) in ro u n d in g is lim ited to th e ra n g e of A /2 to A /2 , th a t is,
- y <<?,()<f
(1A26)
In o th e r w o rd s, th e in sta n ta n e o u s q u a n tiz a tio n e r ro r c a n n o t exceed half of the q u a n tiz a tio n ste p (see T a b le 1.2). If jcmjn an d j:max r e p re s e n t th e m in im u m an d m ax im u m v alu e of x (n ) a n d L is th e n u m b e r o f q u a n tiz a tio n levels, th en A = Xmax ~ ^ L - 1 (1.4.27)
W e d efin e th e d y n a m i c range of th e signal as jrmax -*min- 1 o u r ex am p le we h av e Jtmax = 1, Jtmjn = 0, a n d L 11, w hich leads to A = 0.1. N o te th a t if the d y n am ic ran g e is fixed, in creasin g th e n u m b e r o f q u a n tiz a tio n levels, L resu lts in a d e c re a se o f th e q u a n tiz a tio n ste p size. T h u s th e q u a n tiz a tio n e r ro r d e c re a s e s and th e accu racy o f th e q u a n tiz e r in crease s. In p ra c tic e w e can re d u c e th e q u a n tiz a tio n
36
Introduction
Chap. 1
error to an insignificant am ount by choosing a sufficient num ber o f quantization levels. T h eoretically, quantization o f analog signals always resu lts in a loss o f in form ation. T his is a result o f the am biguity introduced by quantization. Indeed, quantization is an irreversible or noninvertible process (i.e., a m any-to-on e m ap ping) since all sam ples in a distance A /2 about a certain quantization level are assigned the sam e value. This am biguity m akes the exact quantitative analysis of quantization extrem ely difficult. This subject is discussed further in C hapter 9, where w e use statistical analysis.
1.4.4 Quantization of Sinusoidal Signals

Figure 1.21 illustrates the sam pling and quantization o f an an alog sinusoidal signal x a (/) = A cos using a rectangular grid. H orizontal lines w ithin the range of the quantizer indicate the allow ed levels o f quantization. V ertical lines indicate the sam pling tim es. Thus, from the original analog signal x a{t ) w e obtain a discrete tim e signal x ( n ) = x a {nT) by sam pling and a discrete-tim e, discrete-am plitude signal x q ( nT) after quantization. In practice, the staircase sign al xv (r) can be obtained by using a zero-order hold. This analysis is useful b ecau se sinusoids are used as test signals in A /D converters. If the sam pling rate Fs satisfies the sam pling theorem , quantization is the only error in the A /D conversion process. Thus w e can evalu ate the quantization error
Time Figure L21 Sampling and quantization of a sinusoidal signal.
Sec. 1.4
37
bv q u a n tiz in g th e an alo g signal x(t) in stead of th e d isc re te -tim e signal a ( ) = x a( nT) . In sp e c tio n o f Fig. 1.21 in d icates th at the signal x u {t) is alm o st lin ear b e tw e e n q u a n tiz a tio n levels (see Fig. 1.22). T he c o rre sp o n d in g q u a n tiz a tio n e rro r eq (t) x u(t) x q(t) is show n in Fig. 1.22. In Fig. 1.22. r d e n o te s th e tim e th at x a {t) stay s w ithin th e q u a n tiz a tio n levels. T he m e a n -sq u a re e r ro r p o w e r Pq is Pq = 2 t / Since eq (i) = ( A / 2 r ) t . t
e{!)cJl = 7 j we have
(1.4.28)
< t <
t.
If th e q u a n tiz e r has b bits of accuracy an d the q u a n tiz e r co v ers the e n tire range 2A . th e q u a n tiz a tio n step is A = 2 A / 2 h. H en ce
A 2/ 3
P., = T h e a v erag e p o w e r o f the signal xu(D is 1 f 1' , A2 P, = / (A cos Qi,i r d i = ~ TP Jo 2
(1.4.30)
(1.4.31)
T h e q u ality o f th e o u tp u t o f the A /D c o n v e rte r is usually m e a s u re d by th e signalto- quanti zati on noise ratio ( S Q N R ) . w hich pro v id es th e ratio o f th e signal p o w er to th e no ise po w er: S Q N R = = - 22b P " > ' i/ E x p re s se d in d ecib els (d B ), th e S Q N R is S Q N R (d B ) = 101og1(l S Q N R = 1.76 + 6.026 (1.4.32)
T his im p lies th a t th e S Q N R in crease s a p p ro x im ately 6 dB fo r ev ery bit a d d e d to th e w o rd le n g th , th a t is. fo r each d o u b lin g of th e q u a n tiz a tio n levels. A lth o u g h fo rm u la (1.4.32) was d eriv ed fo r sin u so id al signals, w e shall see in C h a p te r 9 th a t a sim ilar resu lt holds fo r every signal w hose d y n am ic ran g e sp a n s the ran g e o f th e q u a n tiz e r. T h is re la tio n sh ip is ex trem ely im p o rta n t b e c a u se it d ictates
Figure 1.22
T he quantization error eq (t) x a (t) - x q (t).
38
Introduction
Chap. 1
th e n u m b e r of bits re q u ire d by a specific a p p lic a tio n to assu re a given signal-ton o ise ratio . F o r ex am p le, m o st co m p a c t disc p lay ers use a sa m p lin g freq u e n cy o f 44.1 k H z an d 16-bit sa m p le re so lu tio n , w hich im plies a S Q N R of m o re th an 96 dB .
1.4.5 Coding of Quantized Samples

T h e co d in g p ro cess in an A /D c o n v e rte r assigns a u n iq u e b in a ry n u m b e r to each q u a n tiz a tio n level. If we h av e L levels w e n e e d at least L d iffe re n t binary' n u m b ers. W ith a w ord len g th o f b bits w e can c re a te 2b d iffe re n t b in ary n u m b e rs. H e n c e we h av e 2h > L. o r e q u iv alen tly , b > log 2 L. T h u s th e n u m b e r of bits re q u ire d in the co d e r is th e sm allest in te g e r g re a te r th a n o r eq u al to log 2 L. In o u r ex am p le it can easily be seen th a t we n eed a c o d e r w ith b = 4 bits. C o m m ercially av ailab le A /D c o n v e rte rs m ay be o b ta in e d w ith finite p recisio n of b 16 o r less. G e n e ra lly , the h ig h er th e sam p lin g sp e ed a n d th e fin er th e q u a n tiz a tio n , th e m o re ex p en siv e the d evice becom es.
1.4.6 Digital-to-Analog Conversion

T o co n v ert a d igital signal in to an an a lo g signal we can use a d ig ital-to -an alo g (D /A ) co n v erter. A s sta te d p rev io u sly , th e task o f a D /A c o n v e rte r is to in te rp o la te b e tw e e n sam ples. T h e sam p lin g th e o re m specifies th e o p tim u m in te rp o la tio n fo r a bandlim ited signal. H o w ev er, this ty p e o f in te rp o la tio n is to o c o m p lic a te d an d . h en ce im p ractical, as in d icated p rev io u sly . F ro m a p ractical v iew p o in t, the sim p lest D /A co n v e rte r is th e z e ro -o rd e r h o ld show n in Fig. 1.15. w hich sim p ly holds c o n sta n t th e v alu e o f o n e sam p le u n til th e n ex t o n e is receiv ed . A d d itio n a l im p ro v e m e n t can b e o b ta in e d by u sing lin e a r in te rp o la tio n as show n in Fig. 1.23 to c o n n e c t successive sa m p les w ith stra ig h t-lin e se g m en ts. T h e z e ro -o rd e r hold an d lin ear in te rp o la to r are an aly zed in S ectio n 9.3. B e tte r in te rp o la tio n can be ach iev ed by u sing m o re so p h isticated h ig h e r-o rd e r in te rp o la tio n tech n iq u es. In g en eral, su b o p tim u m in te rp o la tio n te c h n iq u e s re su lt in p assin g fre q u e n c ie s ab o v e th e fo ld in g freq u e n cy . S uch fre q u e n c y c o m p o n e n ts are u n d e s ira b le a n d are u su ally rem o v ed by p assin g th e o u tp u t o f th e in te rp o la to r th ro u g h a p ro p e r an alo g
Am plifude
Sec. 1.5
Summ ary and References
39
filter, w hich is called a postfilter or sm o o th in g filter. T h u s D /A c o n v ersio n usually involves a su b o p tim u m in te rp o la to r follow ed by a p ostfilter. D /A c o n v e rte rs are tre a te d in m o re d e ta il in S ectio n 9.3.
1.4.7 Analysis of Digital Signals and Systems Versus Discrete-Time Signals and Systems
W e h av e seen th a t a d igital signal is defin ed as a fu n ctio n o f an in te g e r in d e p e n d e n t v ariab le an d its v alu es are ta k e n from a finite set of po ssib le values. T he usefulness of such signals is a c o n s eq u en ce of th e possibilities o ffe re d by digital co m p u ters. C o m p u te rs o p e ra te on n u m b ers, w hich are re p re s e n te d by a strin g of 0 's an d l's . T h e len g th of th is strin g (w o r d length) is fixed an d finite an d usually is 8. 12. 16. or 32 bits. T h e effe cts o f finite w ord len g th in c o m p u ta tio n s cause co m p licatio n s in th e an aly sis of d ig ital signal p ro cessin g system s. T o avoid th ese co m p licatio n s, we neg lect th e q u a n tiz e d n a tu re of digital signals an d system s in m uch o f o u r analysis an d c o n sid e r th em as d isc rete-tim e signals an d system s. In C h a p te rs 6 . 7. and 9 we in v estig ate th e c o n se q u e n c e s o f using a finite w ord len g th . T h is is an im p o rta n t topic, since m any digital signal p ro cessin g p ro b lem s are solved w ith sm all c o m p u te rs o r m icro p ro cesso rs th at em p lo y fix ed -p o in t arith m etic. C o n se q u e n tly , o n e m ust look carefully at the p ro b le m of fin ite-p recisio n arith m e tic an d a c c o u n t for it in th e design of so ftw are an d h a rd w a re th at p e rfo rm s the d esired signal p ro cessin g tasks.
1.5 SUMMARY AND REFERENCES

In th is in tro d u c to ry c h a p te r w e have a tte m p te d to p ro v id e the m o tiv a tio n for digital signal p ro cessin g as an a lte rn a tiv e to a n a lo g signal pro cessin g . W e p re se n te d the basic e le m e n ts o f a digital signal p ro cessin g system an d d efin ed th e o p e ra tio n s n e e d e d to c o n v e rt an an alo g signal in to a digital signal re a d y fo r processing. O f p a rtic u la r im p o rta n c e is the sam pling th e o re m , w hich w as in tro d u c e d by N vquist (1928) an d la te r p o p u la riz e d in the classic p a p e r by S h a n n o n (1949). T h e sam pling th e o re m as d e sc rib e d in S ection 1.4.2 is d eriv ed in C h a p te r 4. S in u so id al signals w ere in tro d u c e d p rim a rily fo r the p u rp o se of illu stra tin g th e aliasin g p h e n o m e n o n an d fo r th e s u b s e q u e n t d e v e lo p m e n t o f th e sa m p lin g th e o re m . Q u a n tiz a tio n effects th a t are in h e re n t in the A /D co n v e rsio n of a signal w ere also in tro d u c e d in th is c h a p te r. Signal q u a n tiz a tio n is b est tre a te d in statistical term s, as d esc rib e d in C h a p te rs 6 , 7. an d 9. F in ally , th e to p ic o f signal re c o n stru c tio n , o r D /A co n v e rsio n , w as d escrib ed briefly. Signal re c o n stru c tio n b ased on sta ircase o r lin e a r in te rp o la tio n m eth o d s is tre a te d in S ectio n 9.3. T h e re a re n u m e ro u s p ractical a p p lic a tio n s of d igital signal processing. T he b o o k e d ite d by O p p e n h e im (1978) tre a ts a p p lic a tio n s to sp e ech p rocessing, im age p ro cessin g , ra d a r signal pro cessin g , so n a r signal p ro cessin g , a n d g eophysical signal p ro cessin g .
40
Introduction
Chap. 1
PROBLEMS
LI Classify the following signals according to whether they are (1) one- or multi dimensional; (2) single or multichannel, (3) continuous time or discrete time, and (4) analog or digital (in amplitude). Give a brief explanation. (a) Closing prices of utility stocks on the New York Stock Exchange. (b) A color movie. (c) Position of the steering wheel of a car in motion relative to cars reference frame. (d) Position of the steering wheel of a car in motion relative to ground reference frame. (e) Weight and height measurem ents of a child taken every month. 1.2 D eterm ine which of the following sinusoids are periodic and com pute their funda mental period. 30n \ ( 62m (a) cosO.OIjt/i (b) cs n -I (c) cos 3 (d) sin3 (e) sin 105 / ' ' V 10 1 3 Determ ine whether or not each of the following signals is periodic. In case a signal is periodic, specify its fundamental period. (a) xu(r) 3cos(5r + 7r/6) (b) = 3 cos(5n + ;r/6) (c) j:(h) = 2 e x p [j(n /6 - 7i)] (d) x(n) = cos(/8) cos(?rn/8) (e) x(n) = cos(7rn/2) sin(7rn/8) + 3cos(jrn/4 + 7t / 3) 1.4 (a) Show that the fundamental period Nr of the signals ,s>() = ei2nkr,IN. * = 0 .1 .2 ,...
is given by Np = N / C C D ( k . N ), where GCD is the greatest common divisor of k and N. (b) What is the fundam ental period of this set for N =71 (c) What is it for N = 16? 1.5 Consider the following analog sinusoidal signal: xa(t) 3 sin(1007rr) (a) Sketch the signal xa{t) for 0 < t < 30 ms. (b) The signal xa(t) is sampled with a sampling rate Fs = 300 samples/s. Determ ine the frequency of the discrete-time signal x{n) = xa( nT ), T = 1 /F. and show that it is periodic. (c) Compute the sample values in one period of .x(n). Sketch _*<n) on the same diagram with x(t). W hat is the period of the discrete-time signal in milliseconds? (d) Can you find a sampling rate Fs such that the signal x(n) reaches its peak value of 3? W hat is the minimum Fs suitable for this task? L6 A continuous-time sinusoid xa(t) with fundamental period Tp = 1/F0 is sampled at a rate F, = 1/ T to produce a discrete-time sinusoid x(n) = x(,nT). () Show that x(n) is periodic if T / T p = k / N (i.e., T/ T p is a rational num ber). (b) If x(n) is periodic, what is its fundam ental period Tp in seconds?
Chap. 1
Problem s
41
(c) Explain the statement: ,r(n) is periodic if its fundamental period Tr . in seconds,
is equal to an integer number of periods of .vu). 1.7 An analog signal contains frequencies up to 10 kHz. (a) W hat range of sampling frequencies allows exact reconstruction of this signal from its samples? (b ) Suppose that we sample this signal with a sampling frequency F, = 8 kHz. Ex amine what happens to the frequency F| = 5 kHz. (c) Repeat part (b) for a frequency F> = 9 kHz. 1.8 An analog electrocardiogram (ECG) signal contains useful frequencies up to 100 Hz. (a) W hat is the Nvquist rate for this signal? (b ) Suppose that we sample this signal at a rate of 250 samples/s. What is the highest frequency that can be represented uniquely at this sampling rate? 1.9 An analog signal a u ) = sin(480;rr) + 3sin(720:rr) is sampled 600 times per second. (a) D eterm ine the Nvquist sampling rate for x a{t). (b ) D eterm ine the folding frequency. (c) What are the frequencies, in radians, in the resulting discrete time signal t (/j )? (d) If is passed through an ideal D/A converter, what is the reconstructed signal v(n? 1.10 A digital communication link carries binary-coded words representing samples of an input signal (t) 3 cos 600.7 r 2 cos 1800jt /
The link is operated at 10.000 bits/s and each input sample is quantized into 1024 different voltage levels. (a) W hat is the sampling frequency and the folding frequency? ( b ) W hat is the Nvquist rate for the signal .*(;)? (c) What are the frequencies in the resulting discrete-time signal x(n)7 (d) W hat is the resolution A? 1.11 Consider the simple signal processing system shown in Fig. P I.11. The sampling periods of the A/D and D /A converters are T = 5 ms and T' = 1 ms. respectively. Determ ine the output vU) of the system, if the input is a:(n = 3 cos 100;rf -+ - 2 sin 250;rt (t in seconds)
The postfilter removes any frequency component above F J 2.
Figure P l . l l
1.12 (a) Derive the expression for the discrete-time signal .r(n) in Example 1.4.2 using the periodicity properties of sinusoidal functions. (b) W hat is the analog signal we can obtain from x(n) if in the reconstruction process we assume that Fs = 10 kHz?
42
Introduction
Chap. 1
1.13 The discrete-time signal x(n) = 6.35cos(jr/10)n is quantized with a resolution (a) A = 0.1 or (b) A = 0.02. How many bits are required in the A/D converter in each case? 1.14 D eterm ine the bit rate and the resolution in the sampling of a seismic signal with dynamic range of 1 volt if the sampling rate is Fs = 20 samples/s and we use an S-bit A/D converter? W hat is the maximum frequency that can be present in the resulting digital seismic signal? 1.15* Sampling o f sinusoidal signals: aliasing Consider the following continuous-time si nusoidal signal
XoO) = sin 2jr/r0f, oc < t < oo
Since xa(t) is described m athematically, its sampled version can be described by values every T seconds. The sampled signal is described by the formula
where Fs = l / T is the sampling frequency. (a) Plot the signal j:( n), 0 < n < 99 for F, = 5 kHz and Fu = 0.5, 2, 3, and 4.5 kHz. Explain the similarities and differences am ong the various plots. (b) Suppose that F0 = 2 kHz and Fs = 50 kHz. (1) Plot the signal x(n). W hat is the frequency / (l of the signal je(n)? (2) Plot the signal v(n) created by taking the even-numbered samples of x(n). Is this a sinusoidal signal? Why? If so, what is its frequency? 1.16* Quantization error in A /D conversion o f a sinuoidal signal Let xq(n) be the signal obtained by quantizing the signal x(n) = sin27r/on. The quantization error power PQ is defined by
The quality of the quantized signal can be measured by the signal-to-quantization noise ratio (SQ NR) defined by SQ N R = 10 log,0 a where Px is the power of the unquantized signal x (n). () For /o = 1/50 and N = 200, write a program to quantize the signal Jt(n), using truncation, to 64, 128, and 256 quantization levels. In each case plot the signals x (n), Xq(n), and e(n) and com pute the corresponding SQNR. (b) Repeat part (a) by using rounding instead of truncation. (c) Comment on the results obtained in parts (a) and (b). (d) Compare the experimentally m easured SQNR with the theoretical SQNR pre dicted by formula (1.4.32) and com m ent on the differences and similarities.
Discrete-Time Signals and Systems
In C h a p te r 1 w e in tro d u c e d th e re a d e r to a n u m b e r o f im p o rta n t ty p es o f signals an d d escrib ed th e sa m p lin g p ro cess bv w hich an a n a lo g signal is c o n v e rte d to a d isc rete-tim e signal. In a d d itio n , we p re s e n te d in so m e d etail th e c h a racteristics o f d isc re te -tim e sin u so id al signals. T h e sin u so id is an im p o rta n t e le m e n ta ry signal th a t se rv es as a b asic b u ild in g block in m o re co m p lex signals. H o w e v e r, th e re are o th e r e le m e n ta ry signals th a t are im p o rta n t in o u r tr e a tm e n t o f signal processing. T h ese d isc re te -tim e signals a re in tro d u c e d in this c h a p te r a n d are used as basis fu n ctio n s o r b u ild in g b locks 1o d escrib e m o re com plex signals. T h e m a jo r e m p h asis in this c h a p te r is th e c h a ra c te riz a tio n o f d isc rete-tim e sy stem s in g en era! a n d the class o f lin ear tim e -in v a ria n t (L T I) system s in p articu lar. A n u m b e r o f im p o rta n t tim e-d o m ain p ro p e rtie s o f L T I sy stem s a re d efined and d e v e lo p e d , an d an im p o rta n t fo rm u la, called th e c o n v o lu tio n fo rm u la, is d eriv ed w hich allow s us to d e te rm in e th e o u tp u t of an L T I system to any given a rb itra ry in p u t signal. In a d d itio n to th e c o n v o lu tio n fo rm u la , d iffe re n c e e q u a tio n s are in tro d u c e d as an a lte rn a tiv e m e th o d fo r describ in g th e in p u t- o u tp u t re la tio n sh ip of an L T I sy stem , an d in a d d itio n , recursive an d n o n re c u rsiv e re a liz a tio n s of LTI sy stem s are tre a te d . O u r m o tiv a tio n fo r th e em p h asis on th e stu d y o f L T I sy stem s is tw ofold. F irst, th e re is a larg e co llectio n o f m a th e m a tic a l te c h n iq u e s th a t can be ap p lied to the an aly sis o f L T I system s. S eco n d , m an y p ractical system s a re e ith e r L T I system s o r can b e a p p ro x im a te d by L T I system s. B ecau se of its im p o rta n c e in digital signal p ro cessin g a p p licatio n s an d its close re se m b la n c e to th e co n v o lu tio n form ula, we also in tro d u c e th e c o rre la tio n b e tw e e n tw o signals. T h e a u to c o rre la tio n and c ro ssc o rre la tio n o f signals a re defined an d th e ir p ro p e rtie s a re p re se n te d .
2.1 DISCRETE-TIME SIGNALS

A s w e d iscu ssed in C h a p te r 1, a d isc re te -tim e signal x{n) is a fu n ctio n o f an in d e p e n d e n t v a ria b le th a t is an in teg er. It is g rap h ically re p re s e n te d as in Fig. 2.1. It is im p o rta n t to n o te th a t a d isc re te -tim e signal is n o t defined at in sta n ts b etw een 43
44
Chap. 2
Figure 2-1
Graphical representation of a discrete-time signal.
tw o successive sam p les. A lso , it is in c o rre c t to th in k th a t .v(n) is e q u a l to z e ro if n is n o t an in teg er. S im ply, th e signal x ( n ) is n o t defin ed fo r n o n in te g e r v alu es o f n. In th e se q u el w e will assu m e th a t a d isc re te -tim e signal is d efin ed fo r every in te g e r value n fo r oo < n < oc. By tra d itio n , w e re fe r to x( n) as th e n th sa m p le o f th e signal ev en if th e signal x( n) is in h e re n tly d isc re te tim e (i.e., n ot o b ta in e d by sam p lin g an a n a lo g signal). If, in d e e d , x ( n) w as o b ta in e d fro m sa m p lin g an a n alo g signal x a( t ), th e n .i(n ) = x a( nT) , w h ere T is th e sa m p lin g p e rio d (i.e., th e tim e b etw e e n successive sam p les). B e sid es th e g rap h ical re p re s e n ta tio n of a d isc re te -tim e signal o r se q u e n c e as illu strated in Fig. 2.1. th e re a re som e a lte rn a tiv e re p re s e n ta tio n s th at are o ften m o re co n v e n ie n t to use. T h e se are: 1. F u n ctio n al re p re s e n ta tio n , such as x(n) 2. T a b u la r r e p re s e n ta tio n , such as n x ( n) -2 0 -1 0 0 0 1 1 2 4 3 1 4 0 5 0
I4,
f 1, [ 0,
fo r n = 1, 3 fo r n = 2 else w h e re
(2 . 1.1)
3. S eq u en ce re p re s e n ta tio n A n in fin ite -d u ra tio n signal o r se q u e n c e w ith th e tim e o rig in (n = 0) in d ic a te d by th e sy m b o l | is re p re s e n te d as *<n) = { . . . 0 . 0 . 1 . 4 , 1 . 0 , 0 , . . . } T A se q u e n c e j:(n ), w hich is z e ro fo r n < 0, can be re p re s e n te d as
jc()
(2.1.2)
= { 0 ,1 .4 .1 .0 .0 ....} T
(2.1.3)
T h e tim e o rig in fo r a se q u e n c e x ( n ) , w hich is z e ro fo r n < 0, is u n d e rs to o d to be th e first (le ftm o st) p o in t in th e seq u en ce.
Sec. 2.1
Discrete-Time Signals
45
A fin ite -d u ra tio n se q u e n c e can be re p re se n te d as x i n) = {3. - 1 . - 2 . 5 .0 .4 . -1 }
(2.1.4)
w h ereas a fin ite -d u ra tio n se q u e n c e th a t satisfies the c o n d itio n x(/i) ~ 0 for n < 0 can be r e p re s e n te d as jc(n) = { 0 .1 .4 . 1)
(2.1.5)
T h e signal in (2.1.4) co n sists of seven sa m p le s or p o in ts (in tim e), so it is called or id en tified as a se v e n -p o in t se q u e n c e . S im ilarly, the se q u e n c e given by (2.1.5) is a f o u r-p o in t se q u e n c e .
2.1.1 Some Elementary Discrete-Time Signals

In o u r stu d y o f d isc re te -tim e signals an d system s th e re a re a n u m b e r o f b asic signals th at a p p e a r o ften a n d play an im p o rta n t role. T h ese signals a re d efin ed below . 1. T h e unit s a m p l e se quence is d e n o te d as < 5 (n) an d is defin ed as
< 5( / i )
=
0.
for n - 0 for n ^ 0
( 2 . 1. 6 )
In w o rd s, th e u n it sa m p le se q u e n c e is a signal th at is zero e v erv w h ere, ex cep t a t n 0 w h e re its value is unity. T his signal is so m e tim e s re fe rre d to as a unit impul se. In c o n tra st to the an alo g signal 8(t). w hich is also called a u n it im p u lse an d is d efin ed to be ze ro ev ery w h ere ex cep t / = 0. an d has unit a re a , th e u n it sa m p le se q u e n c e is m uch less m ath e m a tic a lly c o m p licated . T he g rap h ical re p re s e n ta tio n o f < 5(n ) is show n in Fig. 2.2. 2. T h e unil step signal is d e n o te d a s w (n ) an d is defin ed as u(n) = 1. 0. fo r n > 0 fo r n < 0 (2.1.7)
F ig u re 2.3 illu stra te s th e u nit ste p signal. 3. T h e uni t r a m p signal is d e n o te d as u r (n) an d is d efin ed as u r (n) T h is signal is illu stra te d in Fig. 2.4. fi(n) fo r n > 0 fo r n < 0
( 2 . 1.
Figure 2.2 G raphical rep resen tatio n of the unit sample signal.
46
u(n)
Discrete-Tim e Signals and Systems
Chap. 2
12
3 4 5 6 7
Figure 2 3 G raphical rep resen tatio n of the unit step signal.
ur(n)
T
Figure 2,4 G raphical rep resen tatio n of the unit ram p signal.
4. T h e exponent i al signal is a se q u e n c e o f th e fo rm
jr(n) = a fo r all n (2.1.9)
If th e p a r a m e te r a is real, th e n jr(n) is a real signal. F ig u re 2.5 illu stra te s x ( n) fo r v ario u s v alu es o f th e p a r a m e te r a. W h en th e p a ra m e te r a is co m p lex v a lu e d , it can b e e x p re ss e d as a s rejf1 w h e re r an d 6 a re n o w th e p a ra m e te rs. H e n c e w e c a n e x p ress x ( n ) as x ( n ) = r nej0n = r n (cos On -j- y s in # n )
(2 . 1. 10 )
T iin ilU
Figure 2.5
G raphical representation of exponential signals.
Sec. 2.1
47
Since x (/j ) is now co m plex valu ed , it can be re p re s e n te d g rap h ically by p lo ttin g th e real p a rt x K(n) = r" cos6#fi as a fu n ctio n of n. an d se p a ra te ly p lo ttin g th e im ag in ary p a rt xi (n) = r 1sin 6n (2.1.12) (2.1.11)
as a fu n ctio n o f n. F ig u re 2.6 illu strates th e g ra p h s o f x R(n) a n d x / ( n ) for r 0.9 an d 6 = tt/1 0 . W e o b se rv e th a t th e signals x R(?i) a n d x / ( n ) a re a d a m p e d (decaying e x p o n e n tia l) co sin e fu n ctio n an d a d a m p e d sine fu n c tio n . T h e angle v ariab le 6 is sim ply th e freq u e n c y of th e sinusoid, p rev io u sly d e n o te d by th e (n o rm alized ) fre q u e n c y v ariab le w. C learly, if r 1. th e d a m p in g d is a p p e a rs an d x K(n). x/ ( n). an d A'(n) h av e a fixed am p litu d e, w hich is unity. A lte rn a tiv e ly , th e signal .*(/?) given by (2.1.10) can be re p re s e n te d g raphically by th e a m p litu d e fu n ctio n |.i(h )| = A{n) = r an d th e p h ase fu n ctio n _ v (/;) = < f>{n) = Bn (2.1.14) (2.1.13)
F ig u re 2.7 illu stra te s -4(/?) an d 0 (/i) fo r r 0.9 a n d B = ,t/1 0 . W e o b se rv e th at th e p h ase fu n ctio n is lin ear w ith n. H o w ev er, th e p h ase is d efin e d only o v e r the in terv al n (n) is p lo tte d o v er th e finite in terv al n < B < t t o t Q < $ < 2tt. In o th e r w o rd s, w e su b tra c t m u ltip lies o f I n fro m <p(n) b e fo re p lo ttin g . In one case. <p(n) is c o n s tra in e d to th e range n < 0 < n a n d in th e o th e r case <p(n) is c o n stra in e d to th e ran g e 0 < fi < 2n. T h e su b tra c tio n o f m u ltip le s o f 2n from <p(n) is e q u iv a le n t to in te rp re tin g th e fu n ctio n 4>(n) as 4>{n), m o d u lo 2n. T he g rap h for <p{n). m o d u lo 2 n . is show n in Fig. 2.7b.
2.1.2 Classification of Discrete-Time Signals

T h e m a th e m a tic a l m e th o d s e m p lo y ed in th e analysis of d isc re te -tim e signals and system s d e p e n d on th e c h aracteristics o f th e signals. In this se ctio n we classify d isc rete-tim e signals acco rd in g to a n u m b e r of d iffe re n t c h aracteristics.
Energy signals and power signals.

d efin ed as
OC
T h e en e rg y of a signal x( n ) is
E=
n=-oc
|jc(n)|2
(2.1.15)
W e h av e u se d th e m a g n itu d e -sq u a re d v alu es o f jr(n), so th a t o u r d efin itio n applies to c o m p le x -v a lu e d signals as w ell as re a l-v a lu e d signals. T h e e n e rg y of a signal can be fin ite o r in fin ite. If E is finite (i.e., 0 < E < oo), th e n x ( n ) is called an energy
Figure 2.6 signal.
G raph of the real and im aginary com ponents of a com plex-valued exponential
Sec. 2.1
49
-1 0
1 2
~ S
11
9
(a* Gra ph of A {n I = r" . r = 0.9
( b t G r a p h ol
n . m o d u l o 2ir p l o n e d in t h e r a n g e I t t , it I
Figure 2.7 G r a p h o f a m p l i tu d e and p h a s e function of a co m p l e x -v a l u e d e x p o n e n tial sicnal: ( a ) gr ap h of A i m = r " . 4 = 0 . 9 : ( b ) g rap h of c/>in) = I . t / I O v j . m o d u lo 2ji p lo tt e d in the ran ge i - j i . t t J .
signal. S o m etim es w e add a su b scrip t .v to an d w rite , to em p h asize th a t , is the en e rg y o f th e signal x( n). M anv signals th a t possess infinite en erg y , h av e a finite av e ra g e p o w er. T he av erag e p o w e r o f a d isc re te -tim e signal x( n) is defin ed as P = 1 lim * * 2 N -j- 1 n - -,\
(2 .1.1 6)
If we define th e signal en e rg y o f x(?i) o v er th e finite in terv al N < n < N as (2.1.17)
th e n w e can e x p ress th e signal energy as = lim /v' ,\ oc (2.1.18)
an d th e a v erag e p o w e r o f th e signal x ( n ) as 1 En P = lim n -+x 2 N + 1
(2.1.19)
50
Chap. 2
C learly , if E is finite. P = 0. O n th e o th e r h a n d , if E is infinite, th e av erag e p o w e r P m ay be e ith e r finite o r infinite. If P is finite (an d n o n z e ro ), th e signal is called a p o w e r signal. T h e follow ing e x am p le illu stra te s such a signal. Example 2.1.1 Determ ine the power and energy of the unit step sequence. The average power of the unit step signal is
A 1 -** 2 N
N + 1 lim --------------
a 9c 2
1 + l/N 1 lim ------------- --- 4- l / N 2
Consequently, the unit step sequence is a power signal. Its energy is infinite. Sim ilarly, it can b e show n th a t th e co m p lex e x p o n e n tia l se q u e n c e x(n') = A e JU Jan h as av erag e p o w e r A 2, so it is a p o w e r signal. O n th e o th e r h an d , th e unit ra m p se q u en ce is n e ith e r a p o w e r signal n o r an en e rg y signal.
Periodic signals and aperiodic signals. A s d efin e d o n S ection 1.3, a signal x( n ) is p erio d ic w ith p e rio d N ( N > 0) if a n d only if
x( n + N ) = x( n ) fo r all n (2 . 1.20 )
T h e sm allest v alu e o f N fo r w hich (2.1.20) h o ld s is called th e (fu n d a m e n ta l) p erio d . If th e re is no v alu e o f N th a t satisfies (2.1.20), th e signal is called nonper i odi c or aperiodic. W e h av e a lread y o b se rv e d th a t th e sin u so id al signal o f th e form x ( n) = A sin 2n f o n is p e rio d ic w h en /J, is a ra tio n a l n u m b e r, th a t is, if /o can be e x p re sse d as
(2 .1. 22 )
( 2 . 1 .21 )
w h ere k an d N a re integers. T h e e n erg y o f a p erio d ic signal x ( n ) o v e r a single p e rio d , say. o v er th e in terv al 0 5 n < N - 1, is fin ite if x () ta k e s on finite v alu es o v e r th e p e rio d . H o w ev er, the e n e rg y o f th e p e rio d ic signal fo r oc < n < oo is infinite. O n th e o th e r h an d , the a v e ra g e p o w e r o f th e p erio d ic signal is finite an d it is e q u a l to th e av erag e p o w er o v e r a single p e rio d . T h u s if x () is a p e rio d ic signal w ith fu n d a m e n ta l p e rio d N an d ta k e s o n fin ite v alues, its p o w e r is g iven by
(2.1.23)
n=0 C o n s e q u e n tly , p erio d ic signals are p o w e r signals.
Sec. 2.1
51
Symmetric (even) and antisymmetric (odd) signals.

n al x ( n ) is called sy m m etric (ev en ) if jr(-n ) =
j
A real v a lu ed sig
()
(2.1.24)
O n th e o th e r h a n d , a signal x ( n ) is called a n tisy m m etric (o d d ) if .v( - h ) = - x ( n ) (2.1.25)
W e n o te th a t if .v(/7) is odd, th e n x(0) = 0. E x am p le s o f signals w ith ev en an d odd sy m m etry are illu stra te d in Fig. 2.8. W e w ish to illu strate th a t any a rb itra ry signal can be e x p re sse d as th e sum of tw o signal c o m p o n e n ts , o n e o f w hich is even an d th e o th e r o d d . T h e ev en signal co m p o n e n t is fo rm e d by ad d in g x(/i) to x ( n) and div id in g by 2. th a t is. = j[.v{7!) + x ( - ) ] .v(n] (2.1.26)
< ' 4 *
ll
-4-3-2-I 0
(a)
!I j T ] ! l M ! ....................
3 4 i,
12
.r(n)
- 5 - 4 - 3 - 2 -1
.ill'
12 3 4 5 rt
< >
(b) Figure 2.8 Exam ple of even (a) and odd (b) signals.
52
Chap. 2
Clearly, x e(n ) satisfies the sym m etry con d ition (2.1.24). Sim ilarly, w e form an odd signal com p onent x(n) according to the relation x(n) = j[.x(n) - * ( - / i ) ] (2.1.27)
A gain , it is clear that x 0(n ) satisfies (2.1.25); hence it is in d eed odd. N ow , if we add the tw o signal com p onents, defined by (2.1.26) and (2.1.27), w e ob tain ;c(n), that is, *(n ) = x e(n) + x n{n) Thus any arbitrary signal can be exp ressed as in (2.1.28). (2.1.28)
2.1.3 Simple Manipulations of Discrete-Time Signals

In this section w e consider som e sim ple m odifications or m anipulations involving the in d ep en dent variable and the signal am plitude (depend en t variable).
Transformation of the independent variable (time). A signal x ( n ) may be shifted in tim e by replacing the in d ep en dent variable n by n k, w here k is an integer. If A : is a positive integer, the tim e shift results in a delay of the signal by k units o f tim e. If k is a negative integer, the tim e shift results in an advance of the signal by \k\ units in time.
Example 2.L2
A signal x ( n ) is graphically illustrated in Fig. 2.9a. Show a graphical representation of the signals x ( n 3) and x ( n -I- 2).
Solution The signal x (/i 3) is obtained by delaying ;t(n) by three units in time. The
result is illustrated in Fig. 2.9b. On the other hand, the signal x(n + 2 ) is obtained by advancing x ( n ) by two units in time. The result is illustrated in Fig. 2.9c. Note that delay corresponds to shifting a signal to the right, whereas advance implies shifting the signal to the left on the time axis. If the signal x ( n ) is stored on m agnetic tape or on a disk or, perhaps, in the m em ory o f a com puter, it is a relatively sim ple operation to m odify the base by introducing a delay or an advance. O n the other hand, if the signal is not stored but is b ein g generated by som e physical p h en om en on in real tim e, it is not p ossible to advance the signal in tim e, since such an op eration in volves signal sam ples that have not yet b een generated. W hereas it is alw ays possib le to insert a delay into signal sam ples that have already b een generated, it is physically im possible to view the future signal sam ples. C on sequ en tly, in real-tim e signal processing applications, the operation o f advancing the tim e base o f the signal is physically unrealizable. A n o th er useful m odification o f the tim e base is to replace th e in d ep en dent variable n by n. T h e result o f this op eration is a f o l d i n g or a reflection o f the signal about the tim e origin n = 0.
Sec. 2.1
xin)
4 i
53
j 1
I 4 3 2 1 0
1 2
x i n - 3)
il
T -l 0 1 2 3 4 S 6
(b)
xin +2)
- 6 - 5 - 4 -3 - 2 - 1
Figure 2.9 Graphical representation of a signal, and its delayed and advanced versions.
Example 2.1.3
Show the graphical representation of the signal x { - n ) and x i - n + 2). where x ( n ) is the signal illustrated in Fig. 2.10a.
Solution
The new signal yin) = x ( n) is shown in Fig. 2.10b. Note that y(0) = *(0). v ( l ) = x( 1). y(2) = _t( 2). and so on. Also, y (1) = jc(1 ) , v(2) = x(2), and so on.
Therefore, yin) is simply xin) reflected or folded about the lime origin n = 0. The signal yin) x ( n + 2) is simply x ( n) delayed by two units in time. The resulting signal is illustrated in Fig. 2.10c. A simple way to verify that the result in Fig. 2.10c is correct is to compute samples, such as y(0) = *(2), y (l) = jc(1 >, v(2) = ;t(0), v(1) = jr(3). and so on.
It is im p o rta n t to n o te th a t th e o p e ra tio n s o f folding an d tim e d elay in g (o r ad v an cin g ) a signal a re n o t co m m u ta tiv e . If we d e n o te th e tim e-d e la y o p e ra tio n by T D a n d th e fo ld in g o p e ra tio n by F D . we can w rite T D i[jc(n )l = jc(n - k) FD [jc(/j)] = x ( n) N ow k > 0 (2.1.29)
T D A(FD[.r(n)]l = T D *[.t(-n)] = x ( - n + k)
(2.1.30)
54
Chap. 2
y(n) = x(-n + 2 )
-2 -1
0 1 2 3 4 5 Figure 2.10 Graphical illustration of the folding and shifting operations.
in
(c)
w hereas F D {T D i[j:(n )]} = FD[j:(/i &)] = x ( n k ) (2.1.31)
N o te that becau se the signs o f n and k in x { n k) and Jt(-n + ik ) are different, the re sult is a shift o f the signals x ( n ) and x ( n) to the right by k sam p les, corresponding to a tim e delay. A third m odification o f the in d ep en dent variable in volves replacing n by fin, w here /x is an integer. W e refer to this tim e-b ase m odification as time scaling or dow nsam pling.
Example 2.L4
Show the graphical representation of the signal y(n) = x(2n), where x(n) is the signal illustrated in Fig. 2.11a.
Solution We note that the signal y(n) is obtained from x(n) by taking every other sample from jc(), starting with x(0). Thus y(0) = x(0), y{l) = x(2), y(2) = jc(4), and y ( - l ) = x(~ 2 ), y ( -2 ) = jc( 4), and so on. In other words, we have skipped
Sec. 2.1
55
v(n) = .v(2n)
-4
! -2
0 I 2 3
(b)
Figure 2.11 Graphical illustration ot down-samplinc operation.
the odd-num bered samples in *() and retained the even-num bered samples. The resulting signal is illustrated in Fig. 2.11b. If th e signal ,\ () w as o riginally o b ta in e d by sa m p lin g an a n a lo g signal x a(t), th e n jc() = Xa(nT), w h ere T is the sa m p lin g in terv al. Nowr. v(n) = x ( 2n) = x a(2Tn). H e n c e th e tim e-scalin g o p e ra tio n d escrib ed in E x a m p le 2.1.4 is e q u iv a le n t to ch an g in g th e sam p lin g ra te from 1 /T to 1/27". th a t is, to d e c re a s in g the ra te by a facto r o f 2. T h is is a d o w n s a mp l i n g o p e ra tio n .
Addition, multiplication, and scaling of sequences. A m p litu d e m o d ifi catio n s in clu d e addition, multiplication, an d scaling o f d isc re te -tim e signals. A m p l i t u d e scaling o f a signal by a c o n sta n t A is acco m p lish ed by m u ltiplying th e v alu e o f ev ery signal sa m p le by A. C o n se q u e n tly , w e o b ta in
v(n) = Ax ( n ) oc < fi < oc
T h e s u m o f tw o signals xj ( n) a n d xz i n) is a signal _v(n), w h o se v alu e at any in stan t is equal to th e sum o f th e values o f th e se tw o signals a t th a t in sta n t, th a t is. y(/t) = jq (n) + X2 <n) oc < n < oc
T h e p r o d u c t o f tw o signals is sim ilarly defin ed o n a sa m p le -to -sa m p le basis as v(n) = X](n)X 2 (n) oo < n < oo
56
Chap. 2
2.2 DISCRETE-TIME SYSTEMS

In m a n y ap p lic a tio n s o f d ig ital signal p ro cessin g w e w ish to desig n a d ev ice or an alg o rith m th a t p e rfo rm s so m e p re sc rib e d o p e ra tio n o n a d isc re te -tim e signal. S uch a d evice o r a lg o rith m is called a d isc re te -tim e system . M o re specifically, a discrete-time s y s t em is a d ev ice o r a lg o rith m th a t o p e ra te s on a d isc re te -tim e signal, called th e i nput o r excitation, acco rd in g to so m e w ell-defined ru le , to p ro d u c e a n o th e r d isc rete-tim e signal called th e out p u t o r response of th e system . In g en eral, we view a system as an o p e r a tio n o r a se t o f o p e ra tio n s p e rfo rm e d on th e in p u t sig n al x( n) to p ro d u c e th e o u tp u t signal _v(n). W e say th a t th e in p u t signal x( n) is t r ans f ormed by th e sy stem in to a signal >(), an d ex p ress th e g e n e ra l re la tio n sh ip b e tw e e n jc () a n d y ( n ) as y( n) = T[x{n) } (2 . 2. 1) w h ere th e sym bol T d e n o te s th e tra n s fo rm a tio n (also called an o p e r a to r) , o r p r o cessing p e rfo rm e d by th e sy stem on ;c(n) to p ro d u c e y(n). T h e m a th e m a tic a l re la tio n sh ip in (2.2.1) is d e p ic te d g rap h ically in Fig. 2.12. T h e re are v ario u s w ays to d escrib e th e c h a ra c te ristic s of th e system an d th e o p e ra tio n it p e rfo rm s on x( n ) to p ro d u c e y(n ). In this c h a p te r w e shall b e c o n c e rn e d w ith th e tim e -d o m a in c h a ra c te riz a tio n o f system s. W e shall begin w ith an in p u t-o u tp u t d e sc rip tio n o f th e system . T h e in p u t-o u tp u t d e sc rip tio n focuses on th e b e h a v io r at th e te rm in a ls of th e system a n d ignores th e d e ta ile d in te rn a l co n stru ctio n o r re a liz a tio n o f th e system . L a te r, in S ection 7.5. w e in tro d u c e th e sta te -sp a c e d e sc rip tio n o f a system . In this d e sc rip tio n w e d e v e lo p m a th e m a ti cal e q u a tio n s th a t n o t o nly d esc rib e th e in p u t- o u tp u t b e h a v io r o f th e sy stem b u t specify its in te rn a l b e h a v io r a n d stru c tu re .
2.2.1 Input-Output Description of Systems

T h e in p u t-o u tp u t d e sc rip tio n o f a d isc re te -tim e sy stem consists o f a m a th e m a tic a l ex p ressio n o r a ru le, w hich explicitly d efin es th e re la tio n b e tw e e n th e in p u t an d o u tp u t signals ( i n p u t - o u t p u t relationship). T h e ex act in te rn a l s tru c tu re o f th e sys tem is e ith e r u n k n o w n o r ig n o red . T h u s th e o n ly w ay to in te ra c t w ith th e sy stem is by u sin g its in p u t an d o u tp u t te rm in a ls (i.e., th e system is a ssu m ed to be a black b o x to th e u se r). T o reflec t th is p h ilo so p h y , w e u se th e g ra p h ic a l re p re se n ta -
x (n J
Input signal or excitation Figure 2.12
Discrete-time System Output signal or response
Block diagram representation of a discrete-tim e system.
Sec. 2.2
Discrete-Time Systems
57
tio n d e p ic te d in Fig. 2.12, an d th e g e n e ra l in p u t-o u tp u t re la tio n sh ip in (2.2.1) or, a lte rn a tiv e ly , th e n o ta tio n Jt(n) y(n) (2.2.2)
w hich sim ply m e a n s th a t v(n) is th e re sp o n se of the system T to th e e x c ita tio n x{n). T h e fo llo w in g e x am p les illu stra te se v era l d iffe re n t system s. Example 2.2.1 D eterm ine the response of the following sytems to the input signal x(n) = , A u, (a) (b) (c) (d) -3 < n < 3 otherwise
y(n) = x{n) v() = x in i) y(n) = x i n 4- i) y i n ) = j[A-(n + 1) + x ( n ) + x i n - D] (e) y ( n) = m a x { x( n + 1), x ( n ) . x ( n 1)1 (0 y ( n ) = Z L . x x ( k ) = x ( n ) + x ( n 1) + x{n 2) -t First, we determ ine explicitly the sample values of the input signal xin) = ( ....0 .3 ,2 .1 .0 .1 .2 , 3 ,0 ,...)
T
(2.2.3)
Solution
Next, we determ ine the output of each system using its input-output relationship. (a) In this case the output is exactly the same as the input signal. Such a system is known as the identity system. (b) This system simply delays the input by one sample. Thus its output is given by x{n) = { ...,0 ,3 .2 .1 ,0 ,1 .2 .3 ,0 ,..,) t (c) In this case the system advances the input one sample into the future. For example, the value of the output at time n = 0 is y(0) = *(1). The response of this system to the given input is x(n) = { ...,0 ,3 . 2 .1 .0 ,1 ,2 . 3 ,0 ....} t (d) The output of this system at any time is the mean value of the present, the im m ediate past, and the immediate future samples. For example, the output at time n = 0 is y(0) = + x(0) + jr(l)] = |[1 + 0 + 1] = |
R epeating this com putation for every value of n, we obtain the output signal >() = { ...0 ,1 , f , 2 , l j . l . 2 , , 1 .0 ,...) t
58
Chap. 2
(e> This system selects as its output at time n the maximum value of the three input
samples x(n - l.l, .v(n). and ,r(n + 1). Thus the response of this system to the input signal .\{n) is
v(n) = {0.3. 3. 3. 2 .1 .2 . 3, 3, 3 . 0 . . . . ) t
(f) This system is basically an accumulator that computes the running sum of all the past input values up to present time. The response of this system to the given input is v(n) = {.. ..0 .3 . 5. 6. 6, 7, 9. 1 2 .0 ....} T W e o b se rv e th a t fo r several of th e sy stem s c o n sid e re d in E x a m p le 2.2.1 the o u tp u t at tim e n no d e p e n d s n ot only on th e v alu e of the in p u t at n = n (, [i.e., jc(o)]- b u t also on th e values o f the in p u t a p p lie d to th e system b e fo re an d after n = n (). C o n sid er, fo r in stan ce, th e a c c u m u la to r in th e ex a m p le . W e see th at the o u tp u t at tim e n = ?i() d e p e n d s n ot only on th e in p u t a t tim e n = no. b u t also on x ( n ) a t tim es n = no 1. no - 2, and so on. By a sim ple a lg e b ra ic m a n ip u latio n th e in p u t-o u tp u t re la tio n o f th e a c c u m u la to r can b e w ritte n as
= y(/i - 1) + x(ti) w hich justifies th e te rm accumul at or. In d e e d , th e system c o m p u te s th e c u rre n t v alu e o f th e o u tp u t by a d d in g (a ccu m u latin g ) th e c u rre n t v alu e o f th e in p u t to th e p rev io u s o u tp u t value. T h e re are so m e in te re stin g co n clu sio n s th a t can be d raw n by tak in g a close lo o k in to this a p p a re n tly sim ple system . S u p p o se th a t we are given th e in p u t signal x(rt ) fo r n > no. a n d we wish to d e te rm in e th e o u tp u t v(/i) o f th is system fo r n > no. F o r n = no. no + 1........ (2.2.4) gives v (n n) = v(o - 1) -f x (/i0)
_v(no + l l = v ( o ) + x ( n o + 1)
an d so on. N o te th a t we h ave a p ro b le m in c o m p u tin g y ( n a), since it d e p e n d s on y(o - 1). H o w ev er, y(t io - 1) = ^

k X.
x(k)
th a t is. y(no - 1) sum m arizes* th e effect on th e system from all the in p u ts w hich h ad b e e n ap p lied to th e system b efo re tim e no- T h u s th e re sp o n s e of th e system fo r n > no to th e in p u t x(/7j th a t is a p p lie d a t tim e no is th e c o m b in e d resu lt of this in p u t an d all in p u ts th a t h ad b e e n a p p lie d p re v io u sly to th e sy stem . C o n seq u en tly . y(/i), n > no is n o t u n iq u ely d e te rm in e d by th e in p u t x ( n ) fo r n > no.
Sec. 2.2
59
T he additional inform ation required to d eterm ine y ( n) for n > no is the initial condi t i on y(no - 1). T his value sum m arizes the effect o f all p reviou s inputs to the. system . Thus the initial con d ition y(o - 1) togeth er with the input seq u en ce x ( n ) for n > no uniquely d eterm ine the output sequ en ce y(n ) for n > n0. If the accum ulator had no excitation prior to n 0, the initial con d ition is y(no 1) = 0. In such a case w e say that the system is initially relaxed. S ince y(no 1) = 0, the output seq u en ce y(n) d ep en d s only on the input seq u en ce x ( n ) for n > n0. It is custom ary to assum e that every system is relaxed at n = oo. In this case, if an input x ( n ) is applied at n = co, the corresponding output y( n) is solely and uni qu e l y determ ined by the given input.
Example 2.2.2
T he accum ulator described by (2.2.3) is excited by the sequence x(n) = nu(n). D e term ine its output under the condition that:
(a ) It is initially relaxed [i.e., v ( - l ) = 0]. <b) Initially, y ( 1 )= 1.

Solution The output of the system is defined as tl -] y(n) = ^ x(k) = x(k) +
*=-oc *=-oc
r
i= < l
= y (-l) + k=o But
x(k)
n(n -f 1)
(a) If the system is initially relaxed, v(1) = 0 and hence

n(n + l)
v (n) = -------- 2 -------" 0
(b) On the other hand, if the initial condition is y ( - l ) = 1, then n(n -I-1) n2 + n + 2 v(n) = 1 + ---- - - = ------------n > 0
2.2.2 Block Diagram Representation of Discrete-Time Systems

It is useful at this point to introduce a block diagram representation o f d iscrete tim e system s. For this purpose w e n eed to define som e basic building blocks that can b e intercon n ected to form com p lex system s.
An adder. F igure 2.13 illustrates a system (adder) that perform s the addi tion o f tw o signal seq u en ces to form another (th e sum ) seq u en ce, w hich w e d en ote
60
x|(n )
Chap. 2
y(n) = i,( n ) + x2(n)
Figure 2.13 Graphical representation of an adder.
as y (n ). N o te th a t it is n o t n ecessa ry to sto re e ith e r o n e o f th e se q u e n c e s in o rd e r to p e rfo rm th e a d d itio n . In o th e r w ords, th e a d d itio n o p e ra tio n is memor yl ess .
A constant multiplier. T his o p e ra tio n is d e p ic te d by Fig. 2.14, an d sim ply re p re s e n ts ap p ly in g a scale fa c to r on th e in p u t x ( n) . N o te th a t th is o p e ra tio n is also m em oryless.
a ------------------------ v(n) = o i( n ) Figure 2.14 Graphical representation of a constant multiplier.
A signal multiplier. F ig u re 2.15 illu stra te s th e m u ltip lic a tio n o f tw o sig nal se q u en ces to fo rm a n o th e r (th e p ro d u c t) se q u e n c e , d e n o te d in th e figure as y (n ). A s in th e p re c e d in g tw o cases, w e can view th e m u ltip lic a tio n o p e ra tio n as m em o ry less.
---- -0 ^
x2( n )
A|(n)
v(n) = jT|(n)Ai(n)
Figure 2.1S Graphical representation of a signal multiplier.
A unit delay element. T h e u n it d elay is a sp e cial sy stem th a t sim ply d elay s th e signal passing th ro u g h it by one sam p le. F ig u re 2.16 illu stra te s such a system . If th e in p u t signal is x ( n) , th e o u tp u t is x( n 1). In fact, the sa m p le x{n 1) is sto re d in m em o ry at tim e n 1 a n d it is recalled fro m m e m o ry a t tim e n to form
v ( n ) = x( n - 1) T h u s th is basic b u ild in g b lock re q u ire s m em o ry . T h e use o f th e sym bol ; _1 to d e n o te th e u n it o f d eiay will b eco m e a p p a r e n t w h e n w e discuss th e z -tra n sfo rm in C h a p te r 3.
x(n) ------------------ _____
y (n ) = jr ( n 1)
Figure 2,16 Graphical representation of the unit delay element.
A unit advance element. In c o n tra st to th e u n it d e la y , a u nit ad v an ce m o v es th e in p u t x ( n ) a h e a d by o n e sa m p le in tim e to yield x ( n + 1). F ig u re 2.17 illu stra te s th is o p e ra tio n , w ith th e o p e r a to r ; b ein g used to d e n o te th e u n it advance.
Sec. 2.2
x( n) y( n ) = x( n + I )
61
Figure 2.17 Graphical representation of the unit advance element.
W e o b se rv e th a t an y such ad v an ce is physically im possible in real tim e, since, in fact, it in v o lv es lo o k in g in to th e fu tu re o f th e signal. O n th e o th e r h an d , if we store th e signal in th e m em o ry o f th e c o m p u te r, w e can recall any sa m p le a t any tim e. In such a n o n re a l-tim e ap p lic a tio n , it is p o ssible to advance th e signal jr(?r) in tim e. Example 2.2.3 Using basic building blocks introduced above, sketch the block diagram representa tion of the discrete-time system described by the input-output relation.
v(n) = 3.v( - 1) + | x(n) + \x(n - 1)
where x(n) is the input and y(n) is the output of the system. Solution According to (2.2.5), the output v(n) is obtained by multiplying the input x(n) by 0.5, multiplying the previous input jr ( n - l) by 0.5. adding the two products, and then adding the previous output v(n 1) multiplied by j. Figure 2.18a illustrates this block diagram realization of the system. A simple rearrangem ent of (2.2.5). namely.
v ( ) =
5
.v(n -
)+
[jc(k) + x ( n -
)|
( 2 . 2 .6 )
leads to the block diagram realization shown in Fig. 2.18b. Note that if we treat "the system from the viewpoint of an input-output or an external description, we are not concerned about how the system is realized. On the other hand, if we adopt an
Black box
-i
x( n )
0.5
(a)
Black box
-i
x( n)
Figure 2.18 Block diagram realizations of the system y(n) = 0.25y(n 1) + 0.5 x{n) + 0.5j(n 1).
62
Chap. 2
internal description of the system, we know exactly how the system building blocks are configured. In terms of such a realization, we can see that a system is relaxed at time n = n o if the outputs of all the delays existing in the system are zero at n = n{ ) (i.e., all memory is filled with zeros).
2.2.3 Classification of Discrete-Time Systems

In th e analysis as w ell as in th e design o f system s, it is d e s ira b le to classify the sy stem s acco rd in g to th e g e n e ra l p ro p e rtie s th a t th e y satisfy. In fact, th e m a th e m atical te c h n iq u e s th a t w e d e v e lo p in th is an d in s u b s e q u e n t c h a p te rs fo r analyzing an d d esig n in g d isc rete-tim e system s d e p e n d h eav ily on th e g e n e ra l ch aracteristics of th e system s th a t are b ein g c o n sid e re d . F o r th is re a so n it is n ecessa ry for us to d ev elo p a n u m b e r o f p ro p e rtie s o r c a te g o rie s th a t can be u se d to d escrib e th e g e n e ra l ch aracteristics o f system s. W e stress th e p o in t th a t fo r a sy stem to po ssess a given p ro p e rty , th e p ro p e rty m u st h o ld fo r ev ery p o ssible in p u t signal to th e system . If a p ro p e rty holds for so m e in p u t signals b u t n ot fo r o th e rs, th e system d o e s n ot p o ssess th a t p ro p erty . T h u s a c o u n te re x a m p le is sufficient to p ro v e th a t a system d o e s n ot possess a p ro p e rty . H o w ev er, to p ro v e th a t th e system has so m e p ro p e rty , we m ust prove th a t th is p ro p e rty h o ld s fo r every po ssib le in p u t signal.
Static versus dynamic systems. A d isc re te -tim e sy stem is called static o r m em o ry le ss if its o u tp u t at any in sta n t n d e p e n d s at m o st o n the in p u t sam ple at th e sam e tim e, b u t n o t o n p a st o r fu tu re sa m p le s o f th e in p u t. In any o th e r case, th e system is said to b e d y n a mi c o r to h av e m em o ry . If th e o u tp u t o f a system at tim e n is co m p letely d e te rm in e d by th e in p u t sa m p le s in th e in te rv a l fro m n - N to n ( N > 0), th e system is said to h av e m e m o r y o f d u ra tio n N . U N 0. th e sy stem is static. H 0 < N < oo, th e sy stem is said to have f ini te m e m o r y , w h ereas if N = oo, th e system is said to have infinite m e m o r y . T h e system s d escrib ed by th e follow ing in p u t- o u tp u t e q u a tio n s
y(n) = a x {n) y ( n ) = nx ( n ) + b x 3(n) (2.2.7) (2.2.8)
a re b o th static o r m em o ry less. N o te th a t th e re is n o n e e d to s to re any o f th e past in p u ts o r o u tp u ts in o rd e r to c o m p u te th e p re se n t o u tp u t. O n th e o th e r h an d , th e sy stem s d escrib ed by th e follow ing in p u t-o u tp u t re la tio n s y( n) = x ( n ) + 3 x ( n 1) y(n) = J ^ x ( n - k ) k=0
X
(2.2.9) (2.2.10)
y( n) = J 2 x ( n - k ) Jt=0
(2.2.11)
are dynamic systems or systems with memory. The systems described by (2.2.9)
Sec. 2.2
63
an d (2.2.10) h av e fin ite m em o ry , w h ereas the sy stem d e sc rib e d by (2.2.11) has infinite m em o ry . W e o b se rv e th a t sta tic o r m em oryless system s are d e sc rib e d in g e n e ra l by in p u t-o u tp u t e q u a tio n s o f th e form y(n) = T [ x ( n ) , n] a n d th e y do n o t in clu d e d elay e le m e n ts (m em o ry ). (2.2.12)
Time-invariant versus time-variant systems. W e can su b d iv id e th e g en eral class o f sy stem s in to th e tw o b ro a d c a teg o ries, tim e -in v a ria n t system s and tim e -v a ria n t sy stem s. A system is called tim e -in v a ria n t if its in p u t- o u tp u t c h a ra c teristics d o n o t c h a n g e w ith tim e. T o e la b o ra te , su p p o se th a t w e h av e a system T in a re la x e d s ta te w hich, w h en ex cited by an in p u t signal x ( n) , p ro d u c e s an o u tp u t signal y(n). T h u s w e w rite
y( n) = T [ x { n) ) (2.2.13)
N o w su p p o se th a t th e sam e in p u t signal is d e la y e d by k u n its o f tim e to yield x (n - &), an d ag ain a p p lied to th e sam e system . If th e c h a ra c te ristic s of th e system d o n o t ch an g e w ith tim e, th e o u tp u t o f th e relax ed system will b e y( k). T h at is, th e o u tp u t will b e th e sa m e as th e resp o n se to x ( n) . ex cep t th a t it will be d elay ed by th e sam e k u n its in tim e th a t the in p u t w as d elay ed . T his lead s us to define a tim e -in v a ria n t o r sh ift-in v a ria n t system as follow s.
Definition.
o n ly if
A relax ed system T is time i nvariant o r shift i nvariant if and
x( n) im p lies th a t x {n k)
y(n)
y( n k)
(2.2.14)
fo r ev ery in p u t signal x (n ) a n d every tim e shift k.
T o d e te rm in e if an y given system is tim e in v a ria n t, w e n e e d to p e rfo rm the te st specified b y th e p re c e d in g definition. B asically, we ex cite th e system w ith an a rb itra ry in p u t se q u e n c e x ( n) , w hich p ro d u c e s an o u tp u t d e n o te d as y ( n) . N ext w e d elay th e in p u t se q u e n c e by sam e a m o u n t k an d re c o m p u te th e o u tp u t. In g e n eral, w e can w rite th e o u tp u t as y(, k) = T [ x ( n <:)] N o w if th is o u tp u t y{n, k) = y{n k), for all p o ssib le v alu es o f k, th e system is tim e in v a ria n t. O n th e o th e r h a n d , if th e o u tp u t y( n, k ) ^ y ( n k), ev en fo r o n e v alu e o f k, th e sy stem is tim e v arian t.
64 xin )
Chap. 2
V I7 1) = X( 7 7I- V <7 1 - ]I
D ifferentiator"
- B
x(/t) Time" multiplier
v( n ) = xl - n )
Folder"
v(n J = .u n ) cos oiii
Figure 2.19 Examples of a lime-invariant (a) and some time-variant systems (h)-(d).
Example 2.2.4 Determ ine if the systems shown in Fig. 2.19 are time invariant or time variant. Solution (a) This system is described by the inputôutput equations y(7i) = T\ xin)} = x(n\ x(n - 1) (2.2.15)
Nov, if the input is delayed by k units in time and applied to the system, it is clear from the block diagram that the output will be
y i n . k) = x i n - k) x i n k 1)
(2.2.16)
On the other hand, from (2.2.14) we note that if we delay y (n) by k units in time, we obtain yin k) = x(n k) xin k 1) (2.2.17)
Since the right-hand sides of (2.2.16) and (2.2.17) are identical, it follows that v(n. k) = yin - k). Therefore, the system is time invariant.
Sec. 2.2
65
(b) The input-output equation for this system is

y(n) = T[x(n)] = nx(n) The response of this system to x(n - Jt) is y(n, k) = nx(n - k) Now if we delay ;y(n) in (2.2.18) by k units in time, we obtain y(n - k) = (n k)x(n k)
(2 .2.20)
= nx(n k) - kx(n - k)
(2.2.18)
(2.2.19)
This system is time variant, since y(n, k) ^ y(n - k). (c) This system is described by the input-output relation >() = T[x(n)\ = x ( - n ) The response of this system to jr(n - k) is ;y(n, A ;) = T[x(n - *)] = x ( - n - k) (2.2.22) (2.2.21)
Now, if we delay the output ;y(n), as given by (2.2.21), by k units in time, the result will be y(n - k) = x ( - n + k) (2.2.23) Since y(n, k) ^ y{n - it), the system is time variant.
(d) The input-output equation for this system is

y(n) = j:(n) costDon The response of this system to x(n - k) is y(n, k) = x(n - k) cos o\)n (2.2.25) (2.2.24)
If the expression in (2.2.24) is delayed by k units and the result is compared to (2.2.25), it is evident that the system is time variant.
Linear versus nonlinear systems. The general class o f system s can also be subdivided into linear system s and nonlinear system s. A linear system is one that satisfies the su pe rp ositio n princ iple. Sim ply stated, the principle o f su p erp osi tion requires that the resp onse o f the system to a w eighted sum o f signals b e equal to the corresponding w eighted sum of the responses (outp u ts) of the system to each of the individual input signals. H en ce w e have the follow ing definition o f linearity.
Definition.
A r e la x e d T s y s te m is lin e a r if a n d only if T [ a i x i ( n ) + azx2{n)] = a\ T [ x \ ( n ) ] + a i T [ x 2{n)] (2.2.26)
for any arbitrary input seq u en ces x\ ( n) and x 2(n), and any arbitrary constants aj and 0 2 Figure 2.20 gives a pictorial illustration of the superposition principle.
66
Chap. 2
if and only if v(n) = v'(n).
T h e su p e rp o sitio n p rin cip le e m b o d ie d in th e re la tio n (2.2.26) can be s e p a r a te d in to tw o p arts. F irst, su p p o se th a t a2 = 0. T h e n (2.2.26) re d u c e s to T{a\ X\ ( n) ] = a\ T [ x \ { n ) } = a\ vi(n) w h ere
vi (fl) = T [ x x(n)}
(2.2.27)
T h e re la tio n (2.2.27) d e m o n s tra te s the mul ti pli cat i ve o r scaling p r o p e r t y of a lin ear system . T h a t is, if th e re sp o n se o f th e system to th e in p u t x i ( n ) is vi(n ), the re sp o n se to a\X](n) is sim ply a i j i(n ). T h u s any scaling of th e in p u t resu lts in an id en tical scaling o f th e c o rre sp o n d in g o u tp u t. S eco n d , su p p o se th a t ai = a2 = 1 in (2.2.26). T h e n
T [ x \ ( n ) + x 2 (n)] = T [ x \ { n ) ] + T [ x \ ( n ) }
(2.2.28) = yi ( n) + yz(n) T his re la tio n d e m o n s tra te s th e additivity pr ope r t y o f a lin e a r sy stem . T h e ad ditivity an d m u ltip lic ativ e p ro p e rtie s c o n stitu te th e su p e rp o s itio n p rin c ip le as it ap p lies to lin ear system s. T h e lin earity co n d itio n em b o d ie d in (2.2.26) can b e e x te n d e d a rb itra rily to any w eig h ted lin e a r c o m b in a tio n o f signals by in d u ctio n . In g e n e ra l, we h av e M- 1 x( n) = ^ 2 GkXk(n)
k=l
M- 1 y ( n ) = ^ akyk (n)
(2.2.29)
i= l
w h ere ^ ( n ) = T [ x k(n)} k = 1, 2, , . . , M 1 (2.2.30)
Sec. 2.2
67
W e o b se rv e fro m (2.2.27) th a t if a i 0, th e n y (n ) = 0. In o th e r w ords, a r e lax ed , lin e a r sy stem w ith z e ro in p u t p ro d u c e s a z e ro o u tp u t. If a system p ro d u c e s a n o n z e ro o u tp u t w ith a zero in p u t, th e system m ay be e ith e r n o n re la x e d o r n o n lin ear. If a re la x e d sy stem d o e s n o t satisfy th e su p e rp o s itio n p rin cip le as given by th e d efin itio n ab o v e, it is called nonlinear. Exam ple 2.2^ D eterm ine if the systems described by the following input-output equations are linear or nonlinear. (a) y(n) = ttx(n) (b) y(n) = *(n2) (e) y(n) = ex[n] (c) v(n) = v2(n)
<d) y( n ) = Ax{n) + B Solution
(a) For two input sequences jti(n) and
the corresponding outputs are (2.2.31)
Vi(n) = n.ifi(n) y2{n) = nx2(n) A iinear combination of the two input sequences results in the output Vj() = T[a\Xy (n) + oijMh)] = = ainxi (n) + a2nx2(n) () + (/i)]
(2.2.32) On the other hand, a linear combination of the two outputs in (2.2.31) results in the output a\ V ] (n) + a2y 2(n) = ainx\(n) + a2n,x2(n) (2.2.33)
Since the right-hand sides of (2.2.32) and (2.2.33) are identical, the system is iinear. (b) As in part (a), we find the response of the system to two separate input signals *i(n) and x 2(n). The result is v,(n) = X\(n2) (2.2.341 y2(rt) = X2(n2) The output of the system to a linear combination of Xi(n) and *;(?) is y3(n) = T\a\X\ () + a 2x 2(n)] = a hx,(n2) + a2x2(n2) Finally, a linear combination of the two outputs in (2.2.36) yields O ] \>i(n) + c 2V2(n) = a i JC i (n2) + < i2X2(n2) (2.2.36) (2.2.35)
By comparing (2.2.35) with (2.2.36). we conclude that the system is linear. (c) The output of the system is the square of the input. (Electronic devices that have such an input-output characteristic and are called square-law devices.) From our previous discussion it is clear that such a system is memoryless. We now illustrate that this system is nonlinear.
68
Chap. 2
The responses of the system to two separate input signals are v,() = .vf(n) ' y2(n) = x;(n) (2.2.37)
The response of the system to a linear combination of these two input signals is >'3<n ) = T[a 1*1 (n) + a2x2(n)} = [fli*i(n) + a2x2(n)]2
= cfA 'fln) -I- 2a-la 2x i ( n ) x 2(n) + a 2x 2 { n )
(2.2.38)
On the other hand, if the system is linear, it would produce a linear combination of the two outputs in (2.2.37). namely, ai_Vi(n) + fl2.V 2(n) = (rt) + a2x2(n) (2.2.39)
Since the actual output of the system, as given by (2.2.38). is not equal to (2.2.39), the system is nonlinear.
(d) Assuming that the system is excited by x\(n) and x2in) separately, we obtain
the corresponding outputs
V](n) = AX](rt) + B
(2.2.40) y2(n) = A x 2(n) + B A linear combination of X\(n) and x 2{n) produces the output
V i( ) =
T[u^X ]
(/i) + a 2x 2(n>]
= A[a,x,(/i) + a 2x 2(n)} + B
= A a \ X \ ( n ) -I- a2A x 2(n) + B
(2.2.41)
On the other hand, if the system were linear, its output to the linear com bina tion of Ji(n) and x 2 (n) would be a linear combination of vj i n) and y2(n). that is.
ai yi (n ) + a 2y 2(n) = a ] A x ] { n ) + a \ B a 2A x 2{ n ) + a 2 B
(2.2.42)
Clearly. (2.2.41) and (2.2.42) are different and hence the system fails to satisfy the linearity test. The reason that this system fails to satisfy the linearity test is not that the system is nonlinear (in fact, the system is described by a linear equation) but the presence of the constant B. Consequently, the output depends on both the input excitation and on the param eter B ^ 0. Hence, for B ^ 0. the system is not relaxed. If we set B = 0, the system is now relaxed and the linearity test is satisfied. (e) Note that the system described by the input-output equation y(n) = e,1',) (2.2.43) is relaxed. If x(n) = 0, we find that y(n) = 1. This is an indication that the system is nonlinear. This, in fact, is the conclusion reached when the linearity test, is applied.
Causal versus noncausal systems.

d isc re te -tim e system s.
W e b eg in w ith th e d efin itio n o f causal
Sec. 2.2
69
D e fin itio n , a system is said to b e causal if th e o u tp u t o f th e system at any tim e n [i.e., v(n)] d e p e n d s oniy on p re s e n t an d p ast in p u ts [i.e., x { n ), x(tt - 1), x(rt 2 ) , . . . ] , b u t d o e s n o t d e p e n d on fu tu re in p u ts [i.e., x( n + 1), x ( n + 2 ) , . . . ] . In m a th e m a tic a l te rm s, th e o u tp u t o f a cau sal sy stem satisfies an e q u a tio n o f th e form v(n) = F[x{n), x ( n - 1), x ( n - 2 ) , . . . ] w h ere /'[] is so m e a rb itra ry function. If a system d o e s n o t satisfy this d efin itio n , it is called noncausal . S uch a sy stem has an o u tp u t th a t d e p e n d s n o t oniy on p re s e n t a n d p ast in p u ts b u t also o n fu tu re in p u ts. It is a p p a re n t th a t in real-tim e signal p ro cessin g ap p licatio n s w e c a n n o t o b serv e fu tu re v alu es o f th e signal, and h e n c e a n o n cau sal system is physically u n re a l izab le (i.e., it c a n n o t b e im p le m e n te d ). O n th e o th e r h an d , if th e signal is re c o rd e d so th a t th e p ro cessin g is d o n e off-line (n o n re a l tim e ), it is p o ssible to im p lem en t a n o n cau sal sy stem , since all v alu es o f th e signal are av ailab le a t th e tim e o f p r o cessing. T h is is o fte n th e case in th e p ro cessin g o f g eophysical signals an d im ages. Example 2.2.6 D eterm ine if the systems described by the following input-output equations are causal or noncausal. (a) y(n) = x(n) - x(n - 1) (d) y(n') = x(n) + 3jr(n + 4) (g) }'(n) = x ( -n ) Solution The systems described in parts (a), (b), and (c) are clearly causal, since the output depends only on the present and past inputs. On the other hand, the systems in parts (d). (e), and (f) are clearly noncausal, since the output depends on future values of the input. The system in (g) is also noncausal, as we note by selecting, for example, n = - 1 , which yields v(1) = * 0 ) Thus the output at n = - 1 depends on the input at n = 1, which is two units of time into the future. (b) y(n) = (e) y(n) = x( n2) x(k) (c) y(n) = ax(n) (2.2.44)
(t) y(n) = x(2n)
Stable versus unstable systems. S tab ility is an im p o rta n t p ro p e rty th a t m u st b e c o n s id e re d in an y p ractical ap p lic a tio n o f a system . U n s ta b le system s u su ally ex h ib it e rra tic an d ex tre m e b e h a v io r an d cause overflow in an y p ractical im p le m e n ta tio n . H e re , w e define m a th e m a tic a lly w h at w e m e a n by a sta b le system , a n d la te r, in S ectio n 2.3.6, w e ex p lo re th e im plicatio n s o f this definition fo r lin ear, tim e -in v a ria n t system s.
Definition. A n a rb itra ry re la x e d system is said to be b o u n d e d in p u t-b o u n d e d o u tp u t (B IB O ) sta b le if a n d only if ev ery b o u n d e d in p u t p ro d u ces a b o u n d e d o u tp u t.

T h e c o n d itio n s th a t th e in p u t se q u e n c e x{n) a n d th e o u tp u t se q u e n c e y ( n) are b o u n d e d is tra n s la te d m a th e m a tic a lly to m e a n th a t th e re exist som e finite n u m b ers,
70
Chap. 2
say M x an d M v. such th at j.v(ri)! < M K < oc < M x < dc (2.2.45)
fo r all n. If. fo r so m e b o u n d e d in p u t se q u e n c e ,v(), the o u tp u t is u n b o u n d e d (in fin ite), th e sy stem is classified as u n sta b le . Exam ple 2.2.7 Consider the nonlinear system described by the input-output equation
V(/I ) = 1) ,V(/i )
As an input sequence we select the hounded signal xin ) = C&(n ) where C is a constant. We also assume that y(1) = 0. Then the output sequence is y(0) = C, y (l) = C \ y(2) = Cd............ y(n) = C : "
Clearly, the output is unbounded when ] < ICl < oc. Therefore, the system is BIBO unstable, since a bounded input sequence has resulted in an unbounded output.
2.2.4 Interconnection ot Discrete-Time Systems

D isc rete-tim e sy stem s can be in te rc o n n e c te d to form larg er sy stem s. T h e re are tw o b asic ways in w hich system s can be in te rc o n n e c te d : in c a scad e (series) o r in p arallel. T h ese in te rc o n n e c tio n s are illu strated in Fig. 2.21. N o te th at th e tw o in te rc o n n e c te d sy stem s are d ifferen t. In th e cascad e in te rc o n n e c tio n the o u tp u t of th e first system is yiO?) = 7j[;r(/j)] (2.2.46)
xin)
: T\
r
y(n) T,
7,
(a)
v | (n )
(b)
Figure 2.21 Cascade (a) and parallel (b) interconnections of systems.
Sec. 2.2
71
an d th e o u tp u t o f th e second system is v(n) = T2[\\(n)] (2.2.47)

= r 2{7 i[* (n )]}
W e o b se rv e th a t sy stem s 7"j a n d T2 can be co m b in ed o r c o n s o lid a te d in to a single o v e ra ll sy stem % = T271 C o n s e q u e n tly , w e can ex p ress th e o u tp u t o f th e co m b in ed sy stem as y( n) = Tc[x(n)] In g e n e ra l, th e o rd e r in w hich th e o p e ra tio n s T\ a n d T2 a re p e rfo rm e d is im p o rta n t. T h a t is, T27I # T,T2 fo r a rb itra ry system s. H o w ev er, if th e system s 7j a n d T2 a re iin e a r a n d tim e in v a ria n t, th e n (a) % is tim e in v arian t an d (b ) T2T\ = T \ T2, th a t is, th e o r d e r in w hich th e sy stem s p ro cess th e signal is n o t im p o rta n t. 7^71 a n d T \ T 2 yield id en tical o u tp u t se q u e n c e s. T h e p ro o f o f (a) follow s. T h e p ro o f o f (b ) is given in S e c tio n 2.3.4. T o p ro v e tim e in v a ria n c e , su p p o se th a t Tj and T2 a re tim e in v arian t; th e n x( n k ) an d Vi(n - k) ' + y( n - k) Thus x{n k) Tf y( n k ) vi(/i - k) (2.2.48)
an d th e re fo re , Tc is tim e in v arian t. In th e p a ra lle l in te rc o n n e c tio n , th e o u tp u t of th e sy stem T\ is ^ ( n ) an d the o u tp u t o f th e sy stem T2 is y2(n). H e n c e th e o u tp u t o f th e p a ra lle l in te rc o n n e c tio n is v3(n) = .V ] (n) + > > 2(n) = Ti[x{n)\ + T2[x(n)\ = (T\ + T2)[x(n)} = Tp[x(n)\ w h e re Tp = T\ + T2. In g e n e ra l, w e can u se p a rallel an d cascade in te rc o n n e c tio n o f sy stem s to c o n s tru c t la rg e r, m o re com plex system s. C o n v e rsely , w e can ta k e a la rg e r system a n d b re a k it d o w n in to sm a ller su b sy stem s fo r p u rp o se s o f an aly sis a n d im p le m e n ta tio n . W e sh all u se th e s e n o tio n s la te r, in th e design a n d im p le m e n ta tio n of d ig ital filters.
72
Chap. 2
2.3 ANALYSIS OF DISCRETE-TIME LINEAR TIME-INVARIANT SYSTEMS

In S ectio n 2.2 we classified system s in a c c o rd a n c e w ith a n u m b e r of c h aracteristic p ro p e rtie s o r categ o ries, nam ely: lin e a rity , cau sality , stab ility , a n d tim e in variance. H av in g d o n e so. we now tu rn o u r a tte n tio n to th e analysis o f th e im p o rta n t class o f lin ear, tim e -in v a ria n t (L T I) system s. In p a rtic u la r, we shall d e m o n s tra te th a t such system s are c h a ra c te riz e d in th e tim e d o m ain sim ply bv th e ir resp o n se to a u n it sam p le se q u en ce. W e shall also d e m o n s tra te th a t any a rb itra ry in p u t signal can b e d eco m p o sed an d r e p re s e n te d as a w eig h ted sum of u n it sa m p le seq u en ces. A s a c o n s eq u en ce o f th e lin e a rity a n d tim e-in v arian ce p ro p e rtie s of the system , th e resp o n se o f th e system to any a rb itra ry in p u t signal can be e x p ressed in term s of th e u n it sam p le re sp o n se of th e system . T h e g en eral form o f the ex p ressio n th a t re la te s th e u n it sam ple re sp o n se o f the system an d th e a rb itra ry in p u t signal to th e o u tp u t signal, called th e c o n v o lu tio n sum o r th e c o n v o lu tio n fo rm u la, is also d eriv ed . T h u s we are ab le to d e te rm in e the o u tp u t o f any lin e a r, tim e-in v arian t system to any a rb itra ry in p u t signal.
2.3.1 Techniques for the Analysis of Linear Systems

T h e re are tw o basic m e th o d s for an aly zin g th e b e h a v io r o r re sp o n s e of a lin ear system to a given in p u t signal. O n e m e th o d is b a se d on th e d ire c t so lu tio n o f the in p u t-o u tp u t e q u a tio n for th e system , w hich, in g e n e ra l, has th e form v(ji) - F [v ( n - 1), v (?7 2 ) ........ y(n - N) , x ( n ) . x ( n 1).......... x ( n - M)] w h ere F[-] d e n o te s so m e fu n ctio n o f th e q u a n titie s in b ra c k e ts. Specifically, fo r an L T I system , w e shall see la te r th a t th e g e n e ra l form o f th e in p u t-o u tp u t re la tio n sh ip is N M (2.3.1)
w h ere an d {b * ,.} are c o n sta n t p a ra m e te rs th a t specify th e sy stem an d a re in d e p e n d e n t o f x( n) a n d y( n) . T h e in p u t-o u tp u t re la tio n sh ip in (2.3.1) is called a d ifferen ce e q u a tio n a n d re p re se n ts o n e w ay to c h a ra c te riz e th e b e h a v io r of a d isc re te -tim e L T I system . T h e so lu tio n o f (2.3.1) is th e su b je ct o f S ection 2.4. T h e seco n d m e th o d fo r analyzing th e b e h a v io r o f a lin e a r sy stem to a given in p u t signal is first to d e c o m p o se o r reso lv e th e in p u t signal in to a sum o f e le m e n ta ry signals. T h e e le m e n ta ry signals a re se le c te d so th a t th e re sp o n se o f the system to each signal c o m p o n e n t is easily d e te rm in e d . T h e n , u sin g th e lin earity p ro p e rty o f th e sy stem , th e re sp o n se s o f th e sy stem to th e e le m e n ta ry signals are ad d e d to o b ta in th e to ta l re s p o n s e of th e sy stem to th e given in p u t signal. T h is seco n d m e th o d is th e o n e d e sc rib e d in th is se ctio n .
Sec. 2.3
Analysis o f Discrete-Time Linear Tim e-Invariant Systems
73
T o e la b o ra te , su p p o se th a t th e in p u t signal x (n ) is reso lv ed into a w eig h ted sum o f elem en tary ' signal c o m p o n e n ts {*/. ()} so th a t
w h ere th e {ct} a re th e set of a m p litu d e s (w eighting coefficients) in the d e c o m p o sitio n o f th e signal x( n) . N ow su p p o se th a t th e re sp o n se o f th e system to the e le m e n ta ry signal c o m p o n e n t x k{n) is y k(n). T hus. y*(n) = T [j.t (n)] (2.3.3)
assu m in g th a t th e sy stem is re la x e d an d th a t th e re sp o n se to ckx k(n) is cky k(n). as a c o n s e q u e n c e o f th e scaling p ro p e rty o f th e lin e a r system . F in ally , th e to ta l re sp o n se to th e in p u t x( n ) is
(2.3.4)
In (2.3.4) we u se d th e ad d itiv ity p ro p e rty o f th e lin e a r system . A lth o u g h to a larg e e x te n t, th e ch o ice o f th e e le m e n ta ry signals a p p e a rs to b e a rb itra ry , o u r se lectio n is heavily d e p e n d e n t on th e class of input signals that we w ish to co n sid er. If w e place n o re stric tio n on th e c h aracteristics of th e input signals, its re so lu tio n in to a w eig h ted sum of u n it sam ple (im p u lse) se q u en ces p ro v e s to b e m a th e m a tic a lly c o n v e n ie n t an d c o m p letely g en eral. O n th e o th e r h a n d , if w e re stric t o u r a tte n tio n to a subclass o f in p u t signals, th e re m ay be a n o th e r set o f e le m e n ta ry signals th a t is m o re c o n v e n ie n t m a th e m a tic a lly in the d e te rm in a tio n o f th e o u tp u t. F o r ex am p le, if th e in p u t signal x( n) is p erio d ic w ith p e rio d N , w e have a lre a d y o b se rv e d in S ectio n 1.3.5 th a t a m a th em atically c o n v e n ie n t set o f elem en tary 7 signals is th e set o f e x p o n en tials x k (n) = eJluin k = 0. l , . . . , i V - l (2.3.5)
w h e re th e fre q u e n c ie s {cok } a re h arm o n ic a lly re la te d , th a t is. k = 0. 1.........N - 1 (2.3.6)
T h e fre q u e n c y 2 n / N is called th e fu n d a m e n ta l fre q u e n c y , an d all h ig h er-freq u en cy c o m p o n e n ts a re m u ltip le s o f th e fu n d a m e n ta l fre q u e n c y c o m p o n e n t. T h is subclass o f in p u t sig n als is c o n s id e re d in m o re d e ta il later. F o r th e re so lu tio n o f th e in p u t signal in to a w eig h ted su m o f u n it sam ple se q u en ces, w e m u st first d e te rm in e th e re sp o n se o f th e system to a u n it sa m p le se q u e n c e a n d th e n use th e scaling a n d m u ltip lic ativ e p ro p e rtie s of th e lin ear
74
Chap. 2
sy stem to d e te rm in e th e fo rm u la fo r th e o u tp u t given any a rb itra ry input, T his d e v e lo p m e n t is d escrib ed in d etail as follow s.
2.3.2 Resolution of a Discrete-Time Signal into Impulses

S u p p o se w e h av e an a rb itra ry signal x ( n ) th a t we w ish to reso lv e in to a sum of unit sa m p le seq u en ces. T o utilize th e n o ta tio n e sta b lish e d in th e p re c e d in g se ctio n , we select th e e le m e n ta ry signals x k (n) to be x k(n) = 8{n - k) (2.3.7)
w h ere k re p re se n ts th e d elay o f th e u n it sam p le se q u e n c e . T o h a n d le an a rb itra ry signal x ( n ) th a t m ay h ave n o n z e ro v alu es o v er an infinite d u ra tio n , th e set of unit im p u lses m u st also b e infinite, to en co m p ass th e infinite n u m b e r of delays. N o w su p p o se th a t we m u ltip ly th e tw o se q u e n c e s x( n) a n d <5(n - k). Since 8{n k) is z e ro e v ery w h ere ex cep t a t n = k , w h e re its v alu e is u nity, the result o f th is m u ltip lic atio n is a n o th e r se q u e n c e th a t is z e ro e v ery w h ere e x c e p t at n k , w h ere its v alu e is x ( k) , as illu stra te d in Fig. 2,22. T h u s x( n) 8( n k) = x ( k) 8( n k) (2.3.8)
J t(n ) T T T 111 i l l i [ -2-10113 (a ) 6 (/i-J t) i , 1 I , Tt I i i 1 J C ( J c )
(b)
*(*) 6(n k )
k 0
Figure 2.22
M ultiplication of a signal x i n ) with a shifted unit sam ple sequence.
Sec. 2.3
Analysis of Discrete-Time Linear Time-Invariant Systems
75
is a se q u e n c e th a t is z e ro e v e ry w h e re ex cep t at n = k , w h e re its v a lu e is x( k ) . If we w ere to re p e a t th e m u ltip lic a tio n of x ( n ) w ith < 5(a? m ), w h ere m is a n o th e r d elay (im =6 k), th e re su lt will b e a se q u en ce th a t is z e ro e v e ry w h e re e x cep t at n = m, w h ere its v alu e is x ( m ) . H e n c e x( n ) 5 ( n m) = x ( m) 8( n m) (2.3.9)
In o th e r w o rd s, each m u ltip lic a tio n o f th e signal x( n ) by a u n it im p u lse at som e d elay k, [i.e., <5(n it)], in essen ce picks o u t th e single v alu e x ( k ) o f th e signal jc(n) at th e d e la y w h e re th e u n it im pulse is n o n z e ro . C o n s e q u e n tly , if w e re p e a t this m u ltip lic a tio n o v e r all p o ssib le delays, - o o < k < oo, a n d su m all th e p ro d u c t se q u e n c e s, th e re su lt will be a se q u e n c e e q u a l to th e se q u e n c e x ( n ) , th a t is, P C x( n) = ^ x ( k) 8( n k) k= oc W e e m p h a s iz e th a t th e rig h t-h an d side of (2.3.10) is th e infinite n u m b e r o f u n it sa m p le se q u en ces w h ere th e u n it sa m p le h as an a m p litu d e value o f x( k ) . T h u s th e rig h t-h a n d sid e o f re so lu tio n o f o r d e c o m p o s itio n o f any a rb itra ry signal jc(n) in to a sum o f sh ifted u n it sam p le seq u en ces. Exam ple 2.3.1 Consider the special case of a finite-duration sequence given as
jc()
(2.3.10) su m m a tio n of an se q u e n c e 6(n - k) (2.3.10) gives th e w e ig h te d (scaled)
= (2, 4, 0,3)
T
Resolve the sequence x(n) into a sum of weighted impulse sequences. Solution Since the sequence x(n) is nonzero for the time instants n = 1, 0. 2, we need three impulses at delays k = 1. 0, 2, Following (2.3.10) we find that x<n) = 2(5(n + 1) + 4<5(n) + 3<5(n 2)
2.3.3 Response of LTI Systems to Arbitrary Inputs: The Convolution Sum

H av in g re so lv e d an a rb itra ry in p u t signal x ( n ) in to a w eig h te d su m o f im pulses, w e a re no w re a d y to d e te rm in e th e re sp o n se of an y re la x e d lin e a r sy stem to any in p u t signal. F irs t, w e d e n o te th e re sp o n se v(n, k) o f th e sy stem to th e in p u t unit sa m p le se q u e n c e a t n = it by th e special sym bol h(n, k), oo < k < oo. T h a t is, y( n, k) = h(n. k ) = T[ S( n )] (2.3.11)
In (2.3.11) w e n o te th a t n is th e tim e index a n d k is a p a r a m e te r sh ow ing th e lo c a tio n o f th e in p u t im p u lse. If th e im pulse at th e in p u t is sc aled by an a m o u n t ct = jc(it), th e re sp o n s e of th e system is th e c o rre sp o n d in g ly sc aled o u tp u t, th a t is,
Ckh(n, k) = x(k)h(n, k)
(2.3.12)
76
Chap. 2
F in ally , if th e in p u t is th e a rb itra ry signal x(/t) th a t is e x p re ss e d as a sum of w eig h ted im p u lses, th a t is. (2.3.13) th e n th e resp o n se o f the system to x(/i) is the c o rre sp o n d in g sum o f w eig h ted o u tp u ts, th a t is, y(rt) = 7"[.r(/;)] = T ^ x( k ) S( n k)
x ( k ) T [ 5 (m - k)]
(2.3.14)
ii= oc C learly , (2.3.14) follow s from th e su p e rp o sitio n p ro p e rty of lin e a r system s, and is k n o w n as th e superposit i on s u mma t i o n . W e n o te th a t (2.3.14) is an e x p ressio n for th e resp o n se o f a lin e a r system to any a rb itra ry in p u t se q u e n c e x( n) . T his ex p ressio n is a fu n ctio n of b o th .v() and th e resp o n ses h(n. k) of the system to th e unit im pulses Sin k) fo r oc < k < oc. In d eriv in g (2.3.14) w e used th e lin earity p ro p e rty o f th e system but not its tim ein v arian ce p ro p e rty . T h u s th e ex p ressio n in (2.3.14) ap p lies to any relax ed lin ear (tim e -v a ria n t) system . If. in a d d itio n , th e system is tim e in v a ria n t, th e fo rm u la in (2.3.14) sim plifies c o n sid erab ly . In fact, if the resp o n se o f th e L T I system to th e u n it sa m p le seq u en ce <5(rc) is d e n o te d as h(n). th a t is. h(n) = T [ b( n ) \ (2.3.15)
th en by th e tim e-in v arian ce p ro p e rty , th e resp o n se o f the system to the delay ed u n it sa m p le se q u e n c e <5(n - k) is h(n k) = T [ S( n A')] C o n seq u en tly , th e fo rm u la in (2.3.14) re d u c e s to (2.3.17) k=-oc N ow we o b serv e th a t th e relax ed L T I system is co m p letely c h a ra c te riz e d by a single fu n ctio n h(n), n am ely , its resp o n se to th e u n it sam p le se q u e n c e In c o n tra st, th e g en eral c h a ra c te riz a tio n of th e o u tp u t o f a tim e -v a ria n t, lin e a r sys tem re q u ire s an in fin ite n u m b e r o f u n it sa m p le re sp o n s e fu n ctio n s, h{n, k), o n e for ea c h p o ssib le d elay . T h e fo rm u la in (2.3.17) th a t gives th e re sp o n se y( n) of th e L T I system as a fu n c tio n o f th e in p u t signal x ( n ) a n d th e u n it sa m p le (im p u lse) re sp o n se h(n) is called a convol ut i on s um. W e say th a t th e in p u t jt(n ) is c o n v o lv ed w ith th e im p u lse (2,3.16)
Sec. 2.3
Analysis of Discrete-Time Linear Tim e-Invariant Systems
77
response h(n) to yield the output y in ). W e shall now explain the procedure for com puting the resp onse y (n ). both m athem atically and graphically, given the input x ( n ) and the im pulse response h(n) o f the system . Suppose that we wish to com pute the output of the system at som e time instant, say n = n 0. A ccordin g to (2.3.17), the resp onse at n = no is given as
OC
y (n 0) =
^ 2 x ( k ) h ( n 0 - k)
Jc=-oc
(2.3.18)
O ur first observation is that the index in the sum m ation is k , and h en ce both the input signal x ( k ) and the im pulse resp onse h(no - k) are fun ction s o f k. Second, w e ob serve that the sequ en ces x ( k ) and h(nQ k) are m ultiplied togeth er to form a product seq u en ce. T h e output >(o) is sim ply the sum over all valu es o f the product sequ en ce. T he seq u en ce h ( n 0 k) is ob tain ed from h ( k ) by, first, folding h (k) about k = 0 (the tim e origin), w hich results in the seq u en ce h ( k). The folded seq u en ce is then shifted by no to yield h(no k). T o sum m arize, the process o f com p utin g the con volu tion b etw een x ( k ) and h (k) in volves the follow in g four steps. 1. Folding. F old h(k) about k = 0 to obtain h ( ~ k ) . 2. Shifting, Shift h ( k) by n 0 to the right (left) if n o is p ositive (n egative), to obtain h(no ). 3. Multiplication. M ultiply v Jk) = x ( k ) h ( n 0 - k). by h (no k) to obtain the product sequ en ce
4. S u m m a t i o n . Sum all the values o f the product seq u en ce vnt)( k ) to obtain the value o f the output at tim e n = n 0. W e note that this p rocedure results in the resp onse o f the system at a sin gle tim e instant, say n = n 0. In gen eral, we are interested in evaluating the response o f the system over all tim e instants - o o < n < oo. C onsequently, steps 2 through 4 in the sum m ary m ust be rep eated , for all p ossible tim e shifts oo < n < oo. In order to gain a b etter understanding o f the procedure for evaluating the convolution sum , w e shall dem onstrate the p rocess graphically. T he graphs will aid us in explaining the four steps in volved in the com p utation o f the convolution sum.
Example 2.3.2
The impulse response of a linear time-invariant system is /!() = [1 .2 ,1 ,-1 }
T
(2.3.19)
Determine the response of the system to the input signal x(n) = {1,2.3,1} t
(2.3.20)
Discrete-Time Signais and Systems
Chap. 2
Solution We shall com pute the convolution according to the formula (2.3.17). but we shall use graphs of the sequences to aid us in the com putation. In Fig. 2.23a we illustrate the input signal sequence x(k) and the impulse response h{k) of the system, using k as the time index in order to be consistent with (2.3.17), The first step in the com putation of the convolution sum is to fold h(k). The folded sequence h(~k) is illustrated in Fig. 2.23b. Now we can compute the output at n = 0. according to (2.3.17), which is
v(0) = *=-cx
(2.3.21)
Since the shift n = 0, we use h( k) directly without shifting it. The product sequence = x(k)h(-k) (2.3.22)
h(k) 3 * -i
x(k 4i
T 1
10
-1 0 ! j
1 2 3
h(-k)
2 .
. -2
L 'n(k 1
T t . . . -1 0 1 2
.
(b)
Shift h(\-k)
vAk)
Product
To
111
(c)
,,. L |{t) 1 k (d) 0 12
Product sequence
T .
2 *
~3 T ! j-2 -1 0 1
T 7
Br
Figure 2.23
G raphical com putation of convolution.
Sec. 2.3
79
is also shown in Fig. 2.23b. Finally, the sum of all the terms in the product sequence yields () = vott) = 4
We continue the com putation by evaluating the response of the system at n = 1. According to (2.3.17), (2.3.23) The sequence h(\ k) is simply the folded sequence h( k) shifted to the right by one unit in time. This sequence is illustrated in Fig. 2.23c. The product sequence V ] (k} = x(k)h{l k) (2.3.24)
is also illustrated in Fig. 2.23c. Finally, the sum of all the values in the product sequence yields y(l) = ui(*) = 8
In a similar m anner, we obtain y(2) by shifting h ( - k ) two units to the right, forming the product sequence ih(A) = x(k)h(2 k) and then summing all the terms in the product sequence obtaining y(2) = 8. By shifting h(k) farther to the right, multiplying the corresponding sequence, and summing over all the values of the re sulting product sequences, we obtain v(3) = 3. v(4) = - 2 , y(5) = - 1 . For u > 5, we find that v(n) = 0 because the product sequences contain all zeros. Thus we have obtained the response y(n) for n > 0. Next we wish to evaluate v(n) for n < 0. We begin with n = Then (2.3.25) Now the sequence h ( 1 k) is simply the folded sequence h ( k ) shifted one time unit to the left. The resulting sequence is illustrated in Fig. 2.23d. The corresponding product sequence is also shown in Fig. 2.23d. Finally, summing over the values of the product sequence, we obtain V (-1) = 1 From observation of the graphs of Fig. 2.23, it is clear that any further shifts of h (1 - k) to the left always results in an all-zero product sequence, and hence y(n) = 0 for n 5 2
Now we have the entire response of the system for oc < n < oc. which we summarize below as y(n) = 0 ,0,1, 4. 8, 8. 3, - 2 , - 1 , 0 . 0 . . . .) t (2.3.26)
80
Chap. 2
In E x am p le 2.3.2 w e illu stra te d th e c o m p u ta tio n o f th e co n v o lu tio n sum . using g rap h s o f th e se q u en ces to aid us in visualizing th e ste p s in volved in th e c o m p u ta tio n p ro c e d u re . B e fo re w o rk in g o u t a n o th e r ex a m p le , w e w ish to show th a t th e co n v o lu tio n o p e ra tio n is c o m m u ta tiv e in th e se n se th a t it is irre le v a n t w hich of th e tw o se q u e n c e s is fo ld ed a n d shifted. In d e e d , if w e b eg in w ith (2.3.17) a n d m ak e a ch an g e in th e v ariab le o f th e su m m a tio n , fro m k to m , by defin in g a new index m n k, th e n k = n m an d (2.3.17) b e co m es
CC
y(n) =
^2
x( n m ) h { m )
(2.3.27)
Since m is a d u m m y in dex, w e m ay sim ply re p la c e m by k so th a t y{n) = x(n-k)h(k) (2.3.28)
T h e ex p ressio n in (2.3.28) involves leav in g th e im p u lse re sp o n s e h( k) u n a lte re d , w hile th e in p u t se q u en ce is fo ld ed a n d sh ifted . A lth o u g h th e o u tp u t v(n) in (2.3.28) is id en tical to (2.3.17), th e p ro d u c t se q u e n c e s in th e tw o fo rm s o f th e co n v o lu tio n fo rm u la are not id en tical. In fact, if w e define th e tw o p ro d u c t se q u e n c e s as v(k) = x ( k ) h ( n k) w n(k) = x( n k) h(k) it can b e easily show n th at un() = w n (n k) an d th e re fo re ,
CC CC
v(n) =
^
k oc
^
oc
~ k)
since b o th se q u en ces co n tain th e sam e sa m p le v alu es in a d iffe re n t a rra n g e m e n t. Example 2.3.3 Determ ine the output y(n) of a relaxed linear tim e-invariant system with impulse response h(ri) = a"u(n), \a\ < 1 when the input is a unit step sequence, that is, x (n) = u(n ) Solution In this case both /j(n) and jc(n) are infinite-duration sequences. We use the form of the convolution formula given by (2.3.28) in which x (k) is folded. The
Sec. 2.3
Analysts of Discrete-Time Linear Tim e-Invariant Systems
81
h{k\
x(k)
TT i f
1 1 2 (b) 3 4 *
l * 1
x(- k)
I'oC A ')
-3
-2 -1
0
A(1
-1 -k) (c) >
i'i(i-) a
1
- 1 0 1 1':(!>
v( 2 k)
II
- 2 - 1 0 1 2 3 4 5
Figure 2.24
Graphical computation of convolution in Example 2.3.3.
sequences h(k), x(k). and x{k) are shown in Fig. 2.24. The product sequences vo(k). v\(k), and v2(k) corresponding to x ( k)h(k), x (l k)h(k), and x(2 - k)h(k) are illus trated in Fig. 2.24c, d. and e. respectively. Thus we obtain the outputs v(0) = 1 y(l) = 1 + a
y( 2) = 1 + a + a 1
82
Chap. 2
Clearly, for n > 0, the output is y(n) = 1 + a 4- a2 + + a 1 _ an+1 (2.3,29) = 1-a On the other hand, for n < 0, the product sequences consist of all zeros. Hence v(n) = 0 n < 0 A graph of the output y(n) is illustrated in Fig. 2.24f for the case 0 < a < 1. Note the exponential rise in the output as a function of n. Since |a| < 1, the final value of the output as n approaches infinity is v(oo) = lim v(n) = ------n-*oc ' 1 a (2.3.30)
T o su m m arize, th e co n v o lu tio n fo rm u la p ro v id e s us w ith a m ean s fo r co m p u tin g th e re sp o n s e o f a re lax ed , lin ear tim e -in v a ria n t system to an y a rb itra ry in p u t signal x( n). It ta k e s o n e o f tw o e q u iv a le n t form s, e ith e r (2.3.17) o r (2,3.28), w h ere jt(n ) is th e in p u t sig n al to th e system , h (n ) is th e im p u lse re s p o n s e of th e system , an d y (n ) is th e out p u t o f th e system in re sp o n s e to th e in p u t signal x (n ). T he e v a lu a tio n o f th e co n v o lu tio n fo rm u la involves fo u r o p e ra tio n s , n am ely: f ol di ng e ith e r th e im p u lse re sp o n s e as specified by (2.3.17) o r th e in p u t se q u e n c e as sp ec ified by (2.3.28) to yield e ith e r h ( k) o r x { k). resp ec tiv ely , shifting th e folded se q u e n c e by n u n its in tim e to yield e ith e r h{n k ) o r x { n k ). mul t i pl yi ng the tw o se q u e n c e s to yield th e p ro d u c t se q u en ce, e ith e r x { k) h{ n k) o r x ( n - k ) h ( k ) , a n d finally s u m m i n g all th e v alu es in th e p ro d u c t se q u e n c e to y ield th e o u tp u t v (n ) o f th e sy stem a t tim e n. T h e folding o p e ra tio n is d o n e only o n ce. H o w ev er, th e o th e r th re e o p e ra tio n s a re re p e a te d fo r all p o ssib le shifts oc < n < oo in o rd e r to o b ta in y (n ) fo r oo < n < oc.
2.3.4 Properties of Convolution and the Interconnection of LTI Systems

In th is se ctio n w e in v estig ate so m e im p o rta n t p ro p e rtie s o f co n v o lu tio n an d in te r p re t th e se p r o p e rtie s in te rm s o f in te rc o n n e c tin g lin ear tim e -in v a ria n t system s. W e sh o u ld stress th a t th ese p ro p e rtie s h o ld fo r e v e ry in p u t signal. It is c o n v e n ie n t to sim plify th e n o ta tio n by using an a s te ris k to d e n o te the c o n v o lu tio n o p e ra tio n . T h u s
OC
y( n) = x{n) * h(n) = ^
Jt = - O C
x ( k ) h ( n k)
(2.3.31)
In th is n o ta tio n th e se q u e n c e follow ing th e aste risk [i.e., th e im p u lse re sp o n se /i()] is fo ld e d an d sh ifted . T h e in p u t to th e sy stem is ;c(n). O n th e o th e r h a n d , we also sh o w ed th a t
OC
>>(n) = h{n) * x( n) = ^
h ( k ) x ( n - k)
(2.3.32)
k=-oc
Sec. 2.3

hin)
h(n) Figure 2.25
v(n )
83
<
>
xin)
Interpretation of the commutative property of convolution.
In th is fo rm o f th e co n v o lu tio n fo rm u la, it is the in p u t signal th a t is fo ld ed . A lte r n ativ ely . we m ay in te r p re t this fo rm o f the c o n v o lu tio n fo rm u la as re su ltin g from an in te rc h a n g e o f th e ro les o f j:(n) an d h(n). In o th e r w ords, w e m ay re g a rd x( n) as th e im p u lse re sp o n se o f th e system an d h( n) as the e x c ita tio n o r in p u t signal. F ig u re 2.25 illu stra te s th is in te rp re ta tio n . W e can view co n v o lu tio n m o re ab stractly as a m a th e m a tic a l o p e ra tio n b e tw een tw o signal se q u e n c e s, say x( n ) a n d h(n), th a t satisfies a n u m b e r o f p ro p e rtie s . T h e p ro p e rty e m b o d ie d in (2.3.31) an d (2.3.32) is called th e c o m m u ta tiv e law'.
Commutative law
jf(n) * h{n) = h(n) * x( n) (2.3.33)
V iew ed m a th e m a tic a lly , the c o n v o lu tio n o p e ra tio n also satisfies th e asso cia tive law , w hich can be sta te d as follow s.
Associative law
[-v(/i) * /?[()] * h 2(/i) x( n) * [/*!() * (2.3.34)
F ro m a physical p o in t o f view, we can in te rp re t x (n) as the in p u t signal to a lin e a r tim e -in v a ria n t system w ith im pulse re sp o n se /j|(/i). T h e o u tp u t o f this sy stem , d e n o te d as v i(n ), b eco m es the in p u t to a seco n d lin e a r tim e -in v a ria n t sy stem w ith im p u lse re sp o n se hzin). T h en th e o u tp u t is y( n) V ] (n) * h 2(n) = [j:(fi) * h\(n)] * h2(n) w hich is p recisely th e left-h an d side of (2.3.34). T h u s th e le ft-h a n d side o f (2.3.34) c o rre sp o n d s to h av in g tw o lin ear tim e -in v a ria n t system s in cascad e. N ow th e righth an d side o f (2.3.34) in d icates th a t th e in p u t x ( n ) is a p p lied to an e q u iv a le n t system h av in g an im p u lse re sp o n se , say h(n), w hich is e q u a l to the co n v o lu tio n o f th e tw o im p u lse resp o n se s. T h a t is, h(n) = h](n) * h?(n) an d y(n) = x( n ) * h(n) F u rth e rm o re , sin ce th e co n v o lu tio n o p e ra tio n satisfies th e c o m m u ta tiv e p ro p e rty , o n e can in te rc h a n g e th e o r d e r o f th e tw o system s w ith re sp o n s e s h \ ( n ) a n d hzi n) w ith o u t a lte rin g th e o v erall in p u t-o u tp u t re la tio n sh ip . F ig u re 2.26 g rap h ically il lu stra te s th e asso ciativ e p ro p e rty .
84
xin) y()

jr(n) h(n) = h\(n) * h2(n) y{/i)
Chap. 2
(a)
x(n)
h,(n)
y</i)
v(n)
h2(n)
h t(n)
(b)
Figure Z 2 6
Implications of the associative (a) and the associative and commuta tive (b) properties of convolution.
Example 2.3.4
D eterm ine the impulse response for the cascade of two iinear time-invariant systems having impulse responses h \ () = ( j W " ) and
Solution To determ ine the overall impulse response of the two systems in cascade,
we simply convolve h}{n) with h2(n). Hence
where h2(n) is folded and shifted. We define the product sequence v(k) = h\(k)h2(n - k) = ()*(*)"-* which is nonzero for k > 0 and n - k > 0 or n > k > 0. On the o th er hand, for n < 0, we have v(k) = 0 for all k, and hence h{n) = 0, n < 0 For n > k > 0. the sum of the values of the product sequence v(k) over all k yields hin) =
tscO
= Hr = (i)"(2"+l - 1 )
Sec. 2.3
85
T h e g e n e ra liz a tio n o f the associative law to m o re th a n tw o system s in cascade follow s easily fro m th e d iscu ssio n given ab o v e. T h u s if w e h av e L lin e a r tim ein v arian t sy stem s in cascade w ith im p u lse re sp o n se s h\ ( u) . h ^ i n ) ........ h L (n). th ere is an e q u iv a le n t lin e a r tim e -in v a ria n t sy stem hav in g an im p u lse re sp o n se th a t is e q u a l to th e (L l)-fo id co n v o lu tio n of th e im p u lse resp o n se s. T h a t is. h( n) = h \ { n ) * hzi n) * * h L (n) (2.3.35)
T h e c o m m u ta tiv e law im p lies th a t th e o rd e r in w hich the c o n v o lu tio n s a re p e r fo rm e d is im m a te ria l. C o n v e rsely , an y lin e a r tim e -in v a ria n t system can be d e c o m p o se d in to a cascad e in te rc o n n e c tio n o f subsystem s. A m e th o d for accom plishing th e d e c o m p o s itio n will be d escrib ed later. A th ird p ro p e rty th a t is satisfied by th e c o n v o lu tio n o p e ra tio n is the d istrib u tive law, w hich m ay be sta te d as follow s.
Distributive law
xin) * 4= .xin) * h\ {n) 4- x i n ) * hzi n) (2.3.36)
I n te rp r e te d physically, this law im plies th a t if we h ave tw o lin ear tim ein v a ria n t sy stem s w ith im p u lse re sp o n se s h \ i n ) a n d /?;() ex cited by the sam e in p u t signal .r(/;), th e sum of the tw o resp o n ses is identical to the resp o n se of an o v erall system w ith im pulse resp o n se h i n ) = )i\ in) 4- //:(/;) T h u s th e o v erall system is view ed as a p arallel co m b in a tio n of the tw o linear tim e -in v a ria n t sy stem s as illu stra te d in Fig. 2.27, T h e g e n e ra liz a tio n o f (2.3.36) to m o re th an tw o lin e a r tim e-in v arian t sys tem s in p a rallel follow's easily by m a th e m a tic a l in d u ctio n . T h u s th e in te rc o n n e c tio n of L lin ear tim e -in v a ria n t system s in p a ra lle l w ith im p u lse resp o n ses h\ i n) . h z i n ) .........h L{n) a n d ex cited by the sa m e in p u t x i n ) is e q u iv a le n t to o n e overall system w ith im p u lse re sp o n se
L
h(n) ^ hj i n) (=i
(2.3.37)
C o n v e rsely , an y lin ear tim e -in v a ria n t system can be d e c o m p o se d into a p arallel in te rc o n n e c tio n o f su b sy stem s.
Figure 2.27 Interpretation of the distributive property of convolution: two LTI systems connected in parallel can be replaced by a single system with h(n) =
(n) + k2{n).
86
Chap. 2
2.3.5 Causal Linear Time-Invariant Systems

In S ectio n 2.2.3 w e d efin ed a causal system as o n e w hose o u tp u t at tim e n d e p en d s o n ly on p re s e n t an d p ast in p u ts b u t d o es n o t d e p e n d on f u tu re in p u ts. In o th e r w o rd s, th e o u tp u t o f the. sy stem at so m e tim e in s ta n t n, say n = no, d e p e n d s only on v alu es o f jc() fo r n < n 0In th e case o f a lin e a r tim e -in v a ria n t system , cau sality can b e tra n sla te d to a c o n d itio n o n th e im p u lse resp o n se. T o d e te rm in e this re la tio n sh ip , le t us co n sid e r a lin ear tim e -in v a ria n t system having an o u tp u t a t tim e n = no given by th e c o n v o lu tio n fo rm u la O C
v ( o ) =
^2
k - o c
h ( k ) x (n o ~ k )
S u p p o se th a t w e su b d iv id e th e sum in to tw o sets o f term s, o n e se t involving p re se n t a n d p a st v alu es o f th e in p u t [i.e.. x{n) for n < n 0] a n d o n e se t involving fu tu re valu es o f th e in p u t [i.e., n > no]. T h u s we o b ta in
O C
y (n u) ^ h ( k ) x ( n o - k) +
i=0
-I
^
k oc
h( k) x( nu - k)
1) + h ( 2 ) x ( n 0 - 2 ) + ]
= [/ ?(0)x (n ()) + h( \ ) x( r tu -
+ [/?( 1 )x((] + 1) + h ( - 2 ) x ( n o + 2) + ]
W e o b se rv e th a t th e term s in th e first sum involve jr(no), x( no 1).........w hich are th e p re s e n t a n d p ast values of th e in p u t signal. O n th e o th e r h a n d , th e te rm s in th e se co n d sum in volve th e in p u t signal c o m p o n e n ts ;c(no + l) , x { n o -f 2 ) .........N ow , if th e o u tp u t a t tim e n = n 0 is to d e p e n d only on th e p re s e n t a n d p a st in p u ts, th en , clearly , th e im p u lse re sp o n se o f th e system m u st satisfy th e c o n d itio n ft(n) = 0 n < 0 (2.3.38)
Since h{n) is th e re sp o n se o f th e re lax ed lin e a r tim e -in v a ria n t sy stem to a u nit im p u lse a p p lied a t n = 0, it follow s th a t h(n) = 0 fo r n < 0 is b o th a necessary an d a sufficient c o n d itio n fo r causality. H e n c e an L T I syst em is causal i f a n d onl y i f its i mpul se respons e is z e r o f o r negative values o f n. Since fo r a cau sal sy stem , h(n) = 0 fo r n < 0, th e lim its on th e su m m a tio n of th e co n v o lu tio n fo rm u la m ay be m odified to reflect th is re stric tio n . T h u s w e h ave th e tw o e q u iv a le n t form s
OC
y ( n ) = ^ h ( k ) x ( n - k) Jt=0 n = ^ x{k) h{n k) k=~oc
(2.3.39)
(2.3.40)
A s in d icated p rev io u sly , cau sality is re q u ire d in an y re a l-tim e signal p ro c e ss ing a p p licatio n , since at a n y given tim e n w e have n o access to f u tu re v alu es o f th e
Sec. 2.3
87
in p u t signal. O n ly th e p re s e n t a n d p ast v alu es o f the in p u t signal are av ailab le in c o m p u tin g th e p re s e n t o u tp u t. It is so m e tim e s c o n v e n ie n t to call a se q u en ce th a t is z e ro fo r n < 0, a causa! s e q u e n c e , an d o n e th a t is n o n z e ro fo r n < 0 a n d n > 0. a n on c aus al sequence. T his te rm in o lo g y m e a n s th a t su c h a se q u e n c e could be th e u n it sa m p le re sp o n se of a causal o r a n o n c a u sa l system , resp ectiv ely . If th e in p u t to a causal lin e a r tim e -in v a ria n t system is a causal se q u e n c e [i.e., if jr(n) = 0 fo r n < 0]. th e lim its on th e co n v o lu tio n fo rm u la a re fu rth e r restricted . In th is case th e tw o e q u iv a le n t form s o f th e c o n v o lu tio n fo rm u la b eco m e n y ( n ) = ^ h( k ) x ( n k) (2.3.41) *=o n = -k) (2.3.42) * = < ) W e o b se rv e th a t in th is case, th e lim its on the su m m a tio n s fo r the tw o a lte rn a tiv e fo rm s are id en tical, a n d th e u p p e r lim it is grow ing w ith tim e. C learly , th e resp o n se o f a cau sal sy stem to a causal in p u t se q u e n c e is causal, since y( n) 0 fo r n < 0. Example 2.3.5 Determ ine the unit step response of the linear time-invariant system with impulse response
h{n) = a " u( n ) \a\ < 1
Solution Since the input signal is a unit step, which is a causal signal, and the system is also causal, we can use one of the special forms of the convolution formula, either (2.3.41) or (2.3.42). Since x(n) = 1 for n > 0. (2.3.41) is simpler to use Because of the simplicity of this problem, one can skip the steps involved with sketching the folded and shifted sequences. Instead, we use direct substitution of the signals sequences in (2.3.41) and obtain y(n) = y ~ v *=(I 1 - a"*1 1 -ci and y(n) = 0 for n < 0. We note that this result is identical to that obtained in Ex ample 2.3.3. In this simple case, however, we computed the convolution algebraically without resorting to the detailed procedure outlined previously.
2.3.6 Stability of Linear Time-Invariant Systems

A s in d ic a te d p rev io u sly , sta b ility is an im p o rta n t p ro p e rty th a t m u st be co n sid e re d in an y p ra c tic a l im p le m e n ta tio n o f a system . W e d efin ed an a rb itra ry relax ed system as B IB O sta b le if a n d only if its o u tp u t se q u e n c e y () is b o u n d e d fo r every b o u n d e d in p u t x ( n) .
88
Chap. 2
If x ( n ) is b o u n d e d , th e re exists a c o n s ta n t M x such th a t
l* (n )l < M x
<00
S im ilarly, if th e o u tp u t is b o u n d e d , th e re exists a c o n s ta n t M y such th a t

| ^ ( n ) | < My < OO
fo r all n. N ow , given such a b o u n d e d in p u t se q u e n c e x ( n ) to a lin e a r tim e -in v a ria n t sy stem , le t us in v e stig a te th e im p licatio n s of th e d efin itio n o f sta b ility o n th e c h a r acteristics o f th e system . T o w a rd th is en d , w e w o rk again w ith th e c o n v o lu tio n fo rm u la O O y( n) = h{k) x{n - k ) k= oc If w e ta k e th e a b so lu te value of b o th sides o f th is e q u a tio n , w e o b tain Lv()| =
Y
* = -O C
h( k ) x ( n k)
N ow , th e ab so lu te v alu e of th e sum o f term s is alw ays less th a n o r e q u a l to the sum o f th e a b so lu te v alu es o f th e term s. H e n c e
OC
\y(n)\ < Y
i= oc
IM*)II*(/1 - fc)l
If th e in p u t is b o u n d e d , th e re exists a finite n u m b e r M x such th a t |x(n)[ < M t , By su b s titu tin g th is u p p e r b o u n d fo r x( n ) in th e e q u a tio n ab o v e, w e o b ta in

OC'
|y (n )| < M X Y
k = -o c
W * )l
F ro m th is ex p ressio n w e o b se rv e th a t th e o u tp u t is b o u n d e d if th e im p u lse resp o n se o f th e sy stem satisfies th e co n d itio n

OO
Sh =
Y
k = -o c
IAWI < 00
< 2-3-43)
T h a t is, a linear time-invari ant s y s t em is stable i f its i mpu l s e respons e is absolutely s u m m a b l e . T h is c o n d itio n is n o t only sufficient b u t it is also n e c e ssa ry to e n su re th e sta b ility o f th e system . In d e e d , w e shall show th a t if 5* = 00, th e r e is a b o u n d e d in p u t fo r w hich th e o u tp u t is n o t b o u n d e d . W e c h o o se th e b o u n d e d in p u t \h*(-n)\ 0, h{n) ? 0 h(n) = 0
w h e re h*(n) is th e co m p lex c o n ju g a te o f h(n). I t is sufficient to show th a t th e re is o n e v alu e o f n fo r w h ich y( n) is u n b o u n d e d . F o r n = 0 w e h av e ,< 0 ,=

*= 00^
;
k = (x. 1 v /l
Thus, if Si, = 00, a bounded input produces an unbounded output since y(0) = 00.
Sec. 2.3
89 zero zero this, + N,
T h e c o n d itio n in (2.3.43) im plies th a t th e im p u lse re s p o n s e h(n) goes to as n a p p ro a c h e s infinity. A s a co n se q u e n c e , th e o u tp u t o f th e system g o es to as n a p p ro a c h e s infinity if th e in p u t is set to z e ro b ey o n d n > n 0. T o p ro v e su p p o se th a t |j ( h ) | < M x fo r n < no an d x ( n ) = 0 fo r n > no- T h e n , at n = no th e system o u tp u t is (no + AO = ^ h ( k ) x ( n 0 + N - k) + ^ *-=-oc h { k ) x ( n 0 + N - k)
B u t th e first su m is z e ro since * () = 0 fo r n > no- F o r th e rem ain in g p a rt, we ta k e th e a b so lu te v alu e o f th e o u tp u t, w hich is

j oc I sc
y (0 + AO| =
h( k)x(no + N )j <
|/i(/:)||jf(no + N )|
< Mx k=N
W *)!
N ow , as N a p p ro a c h e s infinity.
an d hen ce
lim ]\ {/i() + AO! = 0

T h is re su lt im p lies th a t any ex citatio n at th e in p u t to th e system , w hich is of a finite d u ra tio n , p ro d u c e s an o u tp u t th a t is tra n s ie n t in n a tu re ; th a t is, its am p litu d e d ecay s w ith tim e an d dies o u t e v en tu ally , w h en th e system is stab le. Example 2.3.6 Determ ine the range of values of the param eter a for which the linear time-invariant system with impulse response
hin) = a"u(n)
is stable. Solution First, we note that the system is causal. Consequently, the lower index on the summation in (2.3.43) begins with k = 0. Hence
Clearly, this geom etric series converges to
provided that \a\ < 1 . Otherwise, it diverges. Therefore, the system is stable if |a| < 1. Otherwise, it is unstable. In effect, h(.n) must decay exponentially toward zero as n approaches infinity for the system to be stable.
90
Chap. 2
Example 23.7
D eterm ine the range of values of a and b for which the linear time-invariant system with impulse response ... ( a", n > 0 h(-n) = [ iw. f, n < n 0 is stable.
Solution This system is noncasual. The condition on stability given by (2.3.43) yields
OC OC
1
ia i" +
y i
n= -o c
t :
n= -o c
i*()f=
fi=s(J
|fe,n
From Example 2.3.6 we have already determ ined that the first sum converges for |a| < 1. The second sum can be manipulated as follows:
= p( l + p + p 2 + )=
I - p
where p = \f\b\ must be less than unity for the geometric series to converge. Conse quently, the system is stable if both \a\ < 1 and |fc| > 1 are satisfied,
2.3.7 Systems with Finite-Duration and Infinite-Duration Impulse Response

U p to this p oin t w e have characterized a linear tim e-invariant system in term s of its im pulse resp onse h(n). It is also con ven ien t, how ever, to subdivide the class o f linear tim e-invariant system s in to tw o types, those that h ave a finite-duration im pulse resp onse (F IR ) and those that have an infinite-duration im pulse response (IIR ). Thus an F IR system has an im pulse resp onse that is zero ou tsid e o f som e finite tim e interval. W ithout loss o f generality, w e focus our atten tion on causal FIR system s, so that h (n) = 0 n < 0 and n > M
The con volu tion form ula for such a system reduces to u-1 ? ( n ) = H h ^ x(^n ~ k ) t=0 A useful interpretation o f this exp ression is ob tain ed by ob serving that th e output at any tim e n is sim ply a w eigh ted iinear com b ination o f the in p ut signal sam ples x (n ), x { n - 1 ) , , x ( n - M + 1). In other words, the system sim ply w eigh ts, by the values o f the im pulse resp onse h(k), k = 0, 1, 1, the m ost recent M signal sam ples and sum s the resulting M products. In e ffe ct, th e system acts as a w in d o w that view s on ly the m ost recent M input signal sam p les in form ing the output. It n eg lects or sim ply forgets all prior input sam p les [i.e., x ( n M ),
Sec. 2.4
Discrete-Time Systems Described by Difference Equations
91
x ( n M 1). . . .] . T h u s we say th a t an F IR svstem has a finite m e m o ry o f len g th -M sam p les. In c o n tra st, an I I R lin e a r tim e-in v arian t system has an in fin ite -d u ra tio n im pu lse resp o n se. Its o u tp u t, b ased on th e c o n v o lu tio n fo rm u la, is v(n) = ^ >2h(k) x( n - k ) w h ere cau sality h as b e e n assu m ed , a lth o u g h this a ssu m p tio n is n o t n ecessary. N ow . th e system o u tp u t is a w eig h te d [by the im p u lse re sp o n se *(/:}] lin e a r co m b in a tio n o f th e in p u t sig n al sa m p le s .r(n), x{n - 1), x( n 2 ) ........ S ince this w eig h te d sum in v olves th e p re s e n t an d all th e p ast in p u t sam ples, w e say th a t th e sy stem has an infinite m em o ry . W e in v e stig a te th e c h aracteristics of F IR an d IIR sy stem s in m o re d etail in su b s e q u e n t c h a p te rs.
2.4 DISCRETE-TIME SYSTEMS DESCRIBED BY DIFFERENCE EQUATIONS

U p to this p o in t w e h ave tre a te d linear and tim e-in v arian t sy stem s th a t are c h a r a c terized by th e ir u n it sa m p le resp o n se h(n). In tu rn . h(n) allow s us to d e te rm in e th e o u tp u t v{n) o f th e system for any given in p u t se q u e n c e jc(/i) by m e a n s o f the c o n v o lu tio n su m m a tio n . (2.4.1) In g e n e ra l, th e n , we h av e show n th a t any linear tim e -in v a ria n t system is c h a r a c terized by th e in p u t- o u tp u t rela tio n sh ip in (2.4.1). M o re o v e r, th e co n v o lu tio n su m m a tio n fo rm u la in (2.4.1) suggests a m ean s fo r th e re a liz a tio n o f th e system . In th e case o f F IR sy stem s, such a realizatio n in v o lv es a d d itio n s, m u ltip lic atio n s, a n d a finite n u m b e r o f m e m o ry locations. C o n se q u e n tly , an F IR system is read ily im p le m e n te d d irectly , as im p lied by the c o n v o lu tio n su m m atio n . If th e sy stem is IIR , h o w ev er, its practical im p le m e n ta tio n as im p lied by c o n v o lu tio n is clearly im p o ssib le, since it re q u ire s an infinite n u m b e r o f m e m o ry lo catio n s, m u ltip lic a tio n s, an d ad d itio n s. A q u e stio n th a t n a tu ra lly arises, th e n , is w h e th e r o r n o t it is possible to realize IIR sy stem s o th e r th a n in the form su g g e sted by th e c o n v o lu tio n su m m atio n . F o rtu n a te ly , th e an sw er is yes. th e re is a p ra c tic a l an d c o m p u ta tio n a lly efficient m e a n s fo r im p le m e n tin g a fam ily o f IIR sy stem s, as will b e d e m o n s tra te d in th is se ctio n . W ith in th e g en eral class o f IIR sy stem s, this fam ily of d isc re te -tim e sy stem s is m o re c o n v en ien tly d e s c rib e d by d ifferen ce eq u atio n s. T h is fam ily o r subclass of IIR sy stem s is v ery u sefu l in a v ariety o f p ractical ap p licatio n s, in clu d in g th e im p le m e n ta tio n o f d ig ita l filters, a n d the m o d elin g o f physical p h e n o m e n a a n d physical system s.
92
Chap. 2
2.4.1 Recursive and Nonrecursive Discrete-Time Systems

A s in d icated ab o v e, th e co n v o lu tio n su m m a tio n fo rm u la e x p re sse s th e o u tp u t of th e lin e a r tim e -in v a ria n t sy stem explicitly a n d only in te rm s o f th e in p u t signal. H o w e v e r, th is n e e d n o t b e th e case, as is sh o w n h e re . T h e re a re m an y system s w h ere it is e ith e r n ecessa ry o r d e sirab le to ex p re ss th e o u tp u t o f th e system n o t on ly in term s o f th e p re s e n t a n d p a st v alu es o f th e in p u t, b u t also in te rm s o f the a lre a d y av ailab le p a st o u tp u t values. T h e follow ing p ro b le m illu stra te s th is point. S u p p o se th a t w e w ish to c o m p u te th e cumul at i ve average o f a signal x ( n ) in th e in te rv a l 0 < k < n, d efin ed as 1 " v(n) = ------- = 0 . 1, . . . n + 1 ~f. (2.4.2)
A s im p lied b y (2.4.2), th e c o m p u ta tio n o f _v(n) re q u ire s th e sto ra g e o f all th e in p u t sa m p le s x (k ) fo r 0 < k < n. Since n is in creasin g , o u r m em o ry r e q u ire m e n ts grow lin early w ith tim e. O u r in tu itio n suggests, h o w ev er, th a t y( n) can be c o m p u te d m o re efficiently by utilizin g th e p re v io u s o u tp u t v alue y( n 1). In d e e d , by a sim p le alg eb raic re a rra n g e m e n t o f (2.4.2), we o b ta in it l ( + l)y (n ) = x(k) + x(n)
= ny( n - 1) + x {n) a n d h en ce y( n) = 1 -x(n) y(n - 1) + n + 1' n + 1' (2.4.3)
N ow , th e cu m u lativ e a v erag e v(n) can b e c o m p u te d recu rsiv ely b y m u ltip ly in g th e p re v io u s o u tp u t v a lu e y( n - 1) by n/ ( n 4-1 ), m u ltip ly in g th e p r e s e n t in p u t x ( n ) by 1 / (n + 1), an d a d d in g th e tw o p ro d u cts. T h u s th e c o m p u ta tio n o f y (n ) by m ean s o f (2.4.3) re q u ire s tw o m u ltip lic atio n s, o n e a d d itio n , a n d o n e m e m o ry lo catio n , as illu stra te d in Fig. 2.28. T his is an ex a m p le of a recursive s yst em. In g e n e ra l, a sy stem w hose o u tp u t v(n) a t tim e n d e p e n d s o n a n y n u m b e r o f p a s t o u tp u t values y( n - 1), y ( n - 2 ) , . . . is called a recu rsiv e system .
tin)
Figure 2.28
Realization of a recursive cum ulative averaging system.
Sec. 2.4
93
T o d e te rm in e th e c o m p u ta tio n of the recu rsiv e sy stem in (2.4.3) in m ore d etail, su p p o se th a t we begin th e p ro cess w ith n = 0 a n d p ro c e e d fo rw a rd in tim e. T h u s, acco rd in g to (2.4.3). we o b ta in y(0) jc(0)
v (l) = 5.v(0) + U ( l ) y(2) = y ( l ) + i* ( 2 ) an d so on. If o n e grow s fatig u ed w ith this c o m p u ta tio n an d w ishes to pass the p ro b le m to so m e o n e else at som e tim e, say n = no. th e only in fo rm a tio n th a t one n e e d s to p ro v id e his o r h e r su ccesso r is the p a st value y(o - 1) a n d the new' input sa m p le s jr(n), j (/j + 1 )........ T h u s the successor b eg in s w ith
an d p ro c e e d s fo rw ard in tim e until som e tim e, say n = n j. w h en he or she b e co m es fatig u ed an d passes the c o m p u ta tio n a l b u rd e n to so m e o n e else w ith the in fo rm a tio n o n th e value _v(/?i 1). an d so on. T h e p o in t wc wish to m ak e in th is discussion is th a t if o n e w ishes to co m p u te th e re sp o n se (in this case, the cu m u lativ e av e ra g e ) of the sy stem (2,4.3) to an input signal x( n ) a p p lied at n tiu. we n eed th e value y (n (, - 1) an d th e input sam ples x ( n ) for /? > /in. T h e term y(>i() 1) is called th e initial condi t i on for the system in (2.4.3) an d c o n ta in s all the in fo rm a tio n n e e d e d to d e te rm in e th e re sp o n se o f the sy stem for n > () to th e in p u t signal x (n ). in d e p e n d e n t o f w h at has o ccu rred in th e p ast. T h e fo llow ing ex am p le illu strates th e use o f a (n o n lin e a r) recu rsiv e system to co m p u te th e sq u a re ro o t of a n u m b e r. Exam ple 2.4.1 Square-Root Algorithm
Many computers and calculators compute the square root of a positive number A. using the iterative algorithm
sn ~ s_ j. T hen it easily follows that
where j_i is an initial guess (estimate) of \/~A. As the iteration converges we have = J~A. Consider now the recursive system (2.4.4)
which is realized as in Fig, 2.29. If we excite this system with a step of amplitude A [i.e.. x(n) = A u ( n )] and use as an initial condition y(1) an estimate of the response v() of the system will tend toward as n increases. Note that in contrast to the system (2.4.3), we do not need to specify exactly the initial condition. A rough estim ate is sufficient for the proper perform ance of the system. For example, if we
94
Chap. 2
4n)
-0---- -0
v(n - I) Figure Z 2 9 Realization of the square-root system.
let A = 2 and y ( - l ) = 1, we obtain j(0 ) = >-(1) = 1,4166667, v(2) = 1,4142157. Similarly, for y ( 1) = 1.5, we have v(0) = 1,416667, y (l) = 1.4142157. Compare these values with the %/2, which is approximately 1.4142136. W e h av e no w in tro d u c e d tw o sim ple recu rsiv e system s, w h ere th e o u tp u t vf/i) d e p e n d s o n th e p re v io u s o u tp u t v alu e y( n 1) a n d th e c u rre n t in p u t ;r(n). B o th system s a re causal. In g e n e ra l, w e can fo rm u la te m o re co m p lex cau sal recu rsiv e system s, in w hich th e o u tp u t y ( n ) is a fu n ctio n o f se v era l p ast o u tp u t v alu es an d p re se n t a n d p ast in p u ts. T h e system sh o u ld h av e a finite n u m b e r o f d elay s o r, eq u iv alen tly , sh o u ld re q u ire a finite n u m b e r o f sto ra g e lo catio n s to be p ractically im p le m e n te d . T h u s th e o u tp u t o f a cau sal a n d p ractically re a liz a b le recursive system can be e x p ressed in g e n e ra l as y( n) = F[y( n 1), y( n 2 ) , . . . , y( n N ) , x ( n) , x ( n 1 ) , . . . , x ( n M ) \ (2.4.5)
w h ere F [ ] d e n o te s so m e fu n ctio n of its a rg u m e n ts. T h is is a re c u rsiv e e q u a tio n specifying a p ro c e d u re fo r co m p u tin g th e system o u tp u t in term s o f p rev io u s v alu es o f th e o u tp u t a n d p re s e n t an d p ast inputs. In c o n tra st, if y{n) d e p e n d s only o n th e p re s e n t an d p a st in p u ts, th e n y (n ) = F[ x ( n ) , x ( n 1)........ x(rt M) ] (2.4.6)
S uch a sy stem is called nonrecursi ve. W e h a s te n to ad d th a t th e c a u s a l F IR system s d escrib ed in S ectio n 2.3.7 in te rm s o f th e co n v o lu tio n sum fo rm u la h av e th e fo rm o f (2.4.6). In d e e d , th e c o n v o lu tio n su m m a tio n fo r a cau sal F IR sy stem is
M
y ( n) = i=0
- *)
= h( 0 ) x ( n ) -)- h ( \ ) x ( n 1) + - + h ( M ) x ( n M ) F [j:(n ), x( n 1), . . . , x ( n M )] w h ere th e fu n ctio n F[-] is sim ply a lin e a r w eig h ted su m o f p re s e n t a n d p ast in p u ts a n d th e im p u lse re sp o n s e values h( n), 0 < n < M , c o n s titu te th e w eig h tin g c o e f ficients. C o n se q u e n tly , th e cau sal lin e a r tim e -in v a ria n t F IR sy stem s d e s c rib e d by th e c o n v o lu tio n fo rm u la in S ectio n 2.3.7, a re n o n re c u rsiv e . T h e b a sic d ifferen ces b e tw e e n n o n re c u rsiv e a n d re c u rsiv e system s a re illu stra te d in Fig. 2.30. A sim ple in sp e ctio n o f th is figure re v e a ls th a t th e fu n d a m e n ta l d iffe re n c e b e tw e e n th e s e tw o
Sec. 2.4
95
x(n)
F[Mn). xin - 1),

........ ti n - .Wl]
V(H )
K n ) ........... v(/i -
M )\
]
1 --------------------------------- 1 (b)
1 ^
x{n)
| |
y{n)
Figure 2.30 Basic form for a causal and realizable (a) nonrecursive and (b) recursive system.
system s is the fe e d b ack loop in th e recu rsiv e system , w hich feed s b ack th e o u tp u t o f the system in to th e in p u t. T h is feed b ack loop co n tain s a d e la y ele m e n t. T he p resen c e o f this d elay is crucial for the realizab ility of the system , since the ab sen ce of this d e la y w ould force the system to c o m p u te yi n) in term s o f v(n). w hich is n o t po ssib le fo r d isc re te -tim e system s. T h e p re se n c e o f the fe ed b ack loop o r, eq u iv alen tly , the recu rsiv e n a tu re of (2.4.5) c re a te s a n o th e r im p o rta n t d ifferen ce b etw een recursive an d n o n re c u rsiv e system s. F o r e x am p le, su p p o se th a t we wish to c o m p u te th e o u tp u t y(o) of a system w h en it is ex cited by an in p u t ap p lied at tim e n = 0. If th e system is recu rsiv e, to co m p u te y ( 0). we first n e e d to c o m p u te all the p re v io u s v alu es y(0). y ( l ) ........ y(o - 1)- In c o n tra st, if the system is n o n recu rsiv e. we can c o m p u te the o u tp u t y (n 0) im m e d ia te ly w ith o u t having y(no - 1), y(o 2 )........ In co n clu sio n , th e o u tp u t o f a recu rsiv e system sh o u ld be c o m p u te d in o rd e r [i.e., v(0), y ( l) , y ( 2 ) . . . w h e re a s for a n o n re c u rsiv e system , th e o u tp u t can b e c o m p u te d in any o rd e r [i.e., y(200). y (15). y{3). y(300). etc.]. T his fe a tu re is d e sira b le in so m e p ractical a p p licatio n s.
2.4.2 Linear Time-Invariant Systems Characterized by Constant-Coefficient Difference Equations

In S ectio n 2.3 w e tre a te d lin e a r tim e -in v a ria n t system s a n d c h a ra c te riz e d th em in te rm s o f th e ir im p u lse resp o n ses. In this su b sectio n w e focus o u r a tte n tio n o n a fam ily o f iin e a r tim e -in v a ria n t system s d escrib ed by an in p u t- o u tp u t r e la tio n called a d iffe re n c e e q u a tio n w ith c o n s ta n t c o e ffic ie n ts . S ystem s d e sc rib e d by c o n s tan t-co efficien t lin e a r d ifferen ce e q u a tio n s are a subclass o f the recu rsiv e a n d n o n re c u rsiv e sy stem s in tro d u c e d in th e p re ced in g su b sectio n . T o b rin g o u t th e im p o rta n t id eas, w e b eg in by tre a tin g a sim ple re c u rsiv e sy stem d e sc rib e d by a first-o rd e r d iffe re n c e e q u a tio n .
96
Chap. 2
<*>
a
Figure 131 Block diagram realization of a simple recursive system.
S u p p o se th a t w e h a v e a recu rsiv e system w ith an in p u t- o u tp u t e q u a tio n y( n) = ay( n - 1) + x ( n ) (2.4.7)
w h ere a is a c o n sta n t. F ig u re 2.31 show s a block d ia g ra m re a liz a tio n o f th e system . In co m p a rin g th is sy stem w ith th e cu m u la tiv e a v erag in g sy stem d e sc rib e d by the in p u t-o u tp u t e q u a tio n (2.4.3), w e o b se rv e th a t th e system in (2.4.7) h as a c o n sta n t co effic ie n t (in d e p e n d e n t o f tim e ), w h e re a s th e sy stem d e sc rib e d in (2.4.3) h as tim ev a ria n t coefficients. A s w e will show , (2.4.7) is an in p u t- o u tp u t e q u a tio n fo r a lin e a r tim e -in v a ria n t sy stem , w h e re a s (2.4.3) d escrib es a lin e a r tim e -v a ria n t system . N ow , su p p o se th a t w e ap p ly an in p u t signal x ( n ) to th e sy stem fo r n > 0. W e m a k e n o assu m p tio n s a b o u t th e in p u t sig n a l fo r n < 0, b u t w e d o assum e th e ex iste n ce o f th e in itial c o n d itio n v ( 1). S ince (2.4.7) d e s c rib e s th e system o u tp u t im plicitly, w e m u st solve this e q u a tio n to o b ta in an ex p licit ex p ressio n for th e sy stem o u tp u t. S u p p o se th a t w e co m p u te successive valu es o f y(n) fo r n > 0, b eg in n in g w ith y(0), T h u s y (0) = o j ( - l ) + x(0) v (l) = ay (0 ) + jc(3 ) = a 2y ( - 1) + o j:( 0 ) + *(1) y(2) = o j ( l ) + x ( 2 ) = o 3y (1) + a 2jr(0) + a ^ ( l ) + x ( 2 )
y( n) = a y( n - 1) + x( n) = e " +1y ( - l ) + a" x( 0) + fl" 1Jt(l) + + a x ( n - 1) + x( n) o r, m o re co m p actly , n > 0 (2.4.8)
T h e re sp o n se y ( n ) o f th e system as given by th e rig h t-h a n d sid e o f (2.4.8) co n sists o f tw o p arts. T h e first p a rt, w hich c o n ta in s th e te rm y ( 1), is a re su lt of th e in itial c o n d itio n y ( 1) o f th e system . T h e se co n d p a r t is th e re sp o n se o f the sy stem to th e in p u t sig n al x( n) . I f th e sy stem is initially re lax ed a t tim e n = 0, th e n its m e m o ry (i.e., the o u tp u t o f th e d e la y ) sh o u ld b e zero . H e n c e y ( 1) = 0. T h u s a rec u rsiv e system is re la x e d if it sta rts w ith z e ro in itial c o n d itio n s. B e c a u se th e m e m o ry o f th e system d escrib es, in so m e sen se, its sta te , w e say th a t th e system is a t z e ro s ta te an d its c o rre sp o n d in g o u tp u t is called th e zero-state respons e o r f o r c e d respons e , an d
Sec. 2.4
97
is d e n o te d by yzs(n). O b v iously, the z e ro -sta te re sp o n se o r fo rc e d resp o n se o f the system (2.4.7) is given by V/jittf) Y ^ a kx( n k) *=() n > 0 (2.4.9)
It is in te re stin g to n o te th at (2.4.9) is a co n v o lu tio n su m m a tio n involving the in p u t signal co n v o lv ed w ith the im p u lse re sp o n se h(n) = a':u ( n ) (2.4.10)
W e also o b se rv e th a t the sy stem d e sc rib e d by th e first-o rd e r d ifferen ce e q u atio n in (2.4.7) is cau sal. A s a resu lt, th e low er lim it on th e c o n v o lu tio n su m m atio n in (2.4.9) is k = 0. F u rth e rm o re , th e co n d itio n v (1) = 0 im plies th a t the in p u t signal can be a ssu m ed cau sal an d h en ce th e u p p e r lim it on th e co n v o lu tio n su m m atio n in (2.4.9) is n. since x( n - k) = 0 for k > n. In effect, we have o b ta in e d th e resull th a t th e re lax ed recu rsiv e system d esc rib e d by th e firs t-o rd e r d ifferen ce e q u atio n in (2.4.7), is a lin e a r tim e-in v arian t IIR svstem w ith im p u lse resp o n se given bv (2.4.10). N ow . su p p o se th at th e system d e sc rib e d by (2.4.7) is in itially n o n re la x e d [i.e.. y ( 1) ^ 0] an d th e in p u t x(/i) = 0 for all //. T h e n the o u tp u t of th e system with zero in p u t is called the zero- i nput response or nat ural r espons e an d is d e n o te d by yZj(/j). F ro m (2.4.7). w ith ,v() = 0 for oc < n < oc . we o b ta in
\Y,OM = ',+ l y ( 1)
n >
(2.4.11)
W e o b se rv e th a t a recu rsiv e system w ith n o n z e ro initial c o n d itio n is n o n relax ed in th e se n se th a t it can p ro d u c e an o u tp u t w ith o u t b ein g ex cited . N o te th at the z e ro -in p u t re sp o n se is d u e to th e m e m o ry of th e system . T o su m m a riz e , th e z e ro -in p u t re sp o n se is o b ta in e d by se ttin g the in p u t signal to z e ro , m ak in g it in d e p e n d e n t of the in p u t. It d e p e n d s only on th e n a tu re of the system a n d th e in itial co n d itio n . T h u s th e z e ro -in p u t re sp o n s e is a ch a ra c te ristic of th e sy stem itself, a n d it is also know n as th e nat ural o r f ree r esponse of th e svstem . O n th e o th e r h a n d , th e z e ro -sta te resp o n se d e p e n d s on th e n a tu re of th e system an d th e in p u t signal. Since th is o u tp u t is a re sp o n se fo rce d u p o n it by th e input signal, it is u su ally called th e f o r c e d response of th e system . In g en eral, th e total re sp o n se o f th e system can b e ex p ressed as v(n) = y Zi ( n ) + .Vzs( n). T h e sy stem d escrib ed by th e first-o rd e r d iffe re n c e e q u a tio n in (2.4.7) is the sim p lest p o ssib le recu rsiv e system in th e g en eral class o f re c u rsiv e system s d e sc rib ed by lin e a r co n stan t-co efficien t d iffe re n c e e q u a tio n s . T h e g e n eral form for such an e q u a tio n is
S M
v(/i ) = ^ ai,y{n - k) -j- ^ bkx ( n - k) i-=l k=0 or, e q u iv alen tly ,

Ar
M
(2.4,12)
' Y a ky{n - k) = ' Y ^ b kx( n - k) Jc=0 *=0
a0 m 1
(2.4.13)
98
Chap. 2
T h e in te g e r N is called th e order o f the d ifferen ce e q u a tio n o r th e o r d e r of the system . T h e n eg ativ e sign on th e rig h t-h a n d side o f (2.4.12) is in tro d u c e d as a m a tte r o f co n v en ien ce to allow us to ex p ress th e d ifferen ce e q u a tio n in (2.4.13) w ith o u t any n eg ativ e signs. E q u a tio n (2.4.12) ex p re sse s th e o u tp u t of the system at tim e rt d irectly as a w eig h ted sum o f p ast o u tp u ts \ {n - 1), y( n - 2 ).........y(/i - N ) as well as past an d p re se n t in p u t signals sam p les. W e o b se rv e th a t in o rd e r to d e te rm in e y( n) fo r n > 0, we n e e d th e in p u t x ( n ) for all n > 0, an d th e initial c o n d itio n s y ( 1), y ( 2), y ( N ) . In o th e r w o rd s, th e initial c o n d itio n s su m m a riz e all th a t we n eed to k n o w a b o u t th e p a s t h isto ry o f th e re sp o n se o f th e sy stem to c o m p u te th e p re se n t an d fu tu re o u tp u ts. T h e g e n e ra l so lu tio n o f th e /V -order c o n stan tco efficien t d ifferen ce e q u a tio n is c o n sid e re d in th e follow ing su b se c tio n . A t this p o in t we r e s ta te th e p ro p e rtie s of lin earity , tim e in v a ria n c e , an d stab ility in th e c o n te x t o f recu rsiv e system s d e sc rib e d by lin ear c o n s ta n t-c o e ffic ie n t d ifferen ce e q u a tio n s. A s we h av e o b se rv ed , a recu rsiv e system m a y b e relax ed o r n o n re la x e d , d e p e n d in g o n th e initial co n d itio n s. H e n c e th e d e fin itio n s o f th ese p ro p e rtie s m u st ta k e in to a c c o u n t th e p re se n c e o f the initial co n d itio n s. W e begin w ith th e d e fin itio n o f lin earity . A system is lin e a r if it satisfies th e follow ing th re e re q u ire m e n ts: 1. T h e to tal re sp o n se is e q u a l to th e sum o f th e z e ro -in p u t a n d z e ro -sta te r e sp o n ses [i.e.. y ( n) = y 7 .\(n) + yzs(n)]. 2. T h e p rin cip le o f su p e rp o s itio n ap p lies to th e z e ro -sta te re s p o n s e ( zero-state linear). 3. T h e p rin cip le o f su p e rp o s itio n ap p lies to the z e ro -in p u t re s p o n s e ( zero- i nput linear). A system th a t d o es n o t satisfy all three se p a ra te r e q u ire m e n ts is by d efin itio n n o n lin ear. O b v io u sly , fo r a re lax ed system , yZj(n) = 0, an d th u s re q u ire m e n t 2, w hich is th e d efin itio n o f lin e a rity given in S ection 2.2.4, is sufficient. W e illu strate th e a p p lic a tio n o f th ese re q u ire m e n ts by a sim p le ex am p le. Example 2.4.2 D eterm ine if the recursive system defined by the difference equation v(n) = av(n 1)4- x(n) is linear. Solution as By combining (2.4.9) and (2.4.11), we obtain (2.4.8), which can be expressed
y(n) = yzi(n) + y*(n) Thus the first requirem ent for linearity is satisfied.
Sec. 2.4
99 +
To check for the second requirem ent, let us assume that ,t(n) = Then (2.4.9) gives
*= {}
= ('iy ^ (n) + C2V^'(n) Hence y(/i> satisfies the principle of superposition, and thus the system is zero-state linear. Now let us assume that y ( 1) = q vj(1) 4- f 2y;( 1). From (2.4.11) we obtain 1) 4- C;V;(1)] = vi (1)4v; ( 1)
= C i v'i '(n) + r ;y ^(/i) Hence the system is zero-input linear. Since the system satisfies all three conditions for linearity, it is linear. A lth o u g h it is so m e w h at ted io u s, the p ro c e d u re used in E x am p le 2.4,2 to d e m o n s tra te lin e a rity for th e system d escrib ed by th e first-o rd e r d ifferen ce e q u a tio n , ca rrie s o v er d irectly to th e g en eral recu rsiv e system s d e sc rib e d by th e c o n stan tcoefficient d ifferen ce e q u a tio n given in (2.4.13). H en ce , a recu rsiv e system d escrib ed by th e lin ear d iffe re n c e e q u a tio n in (2.4.13) also satisfies all th re e c o n d itio n s in th e d efin itio n o f lin earity , a n d th e re fo re it is linear. T h e n ex t q u e s tio n th a t arises is w h e th e r o r n o t the cau sal lin ear system d e sc rib e d by th e lin e a r c o n stan t-co efficien t differen ce e q u a tio n in (2.4.13) is tim e in v arian t. T h is is fairly easy, w h en d e alin g w ith system s d e sc rib e d by explicit i n p u t-o u tp u t m a th e m a tic a l re la tio n sh ip s. C learly, th e system d e sc rib e d by (2.4.13) is tim e in v a ria n t b ecau se th e coefficients ak an d bk are c o n stan ts. O n th e o th e r h a n d , if o n e o r m o re o f th e s e coefficients d e p e n d s on tim e, th e system is tim e v a ria n t, since its p ro p e rtie s ch an g e as a fu n ctio n of tim e. T h u s w e co n clu d e th at the recursive sy s t em descri bed b y a linear constant-coefficient difference equat i on is linear a n d t ime invariant. T h e final issu e is th e sta b ility of th e recursive system d e sc rib e d by th e lin ear, c o n stan t-co efficien t d iffe re n c e e q u a tio n in (2.4.13). In S ection 2.3.6 w e in tro d u c e d th e c o n c e p t o f b o u n d e d in p u t-b o u n d e d o u tp u t (B IB O ) sta b ility fo r re la x e d sys tem s. F o r n o n re la x e d system s th a t m ay b e n o n lin e a r, B IB O sta b ility sh o u ld be view ed w ith so m e care. H o w e v e r, in th e case o f a lin e a r tim e -in v a ria n t recursive system d e sc rib e d by th e lin e a r co n stan t-co efficien t d iffe re n c e e q u a tio n in (2.4.13), it suffices to s ta te th a t such a system is B IB O sta b le if a n d only if fo r every b o u n d e d in p u t a n d ev ery b o u n d e d initial c o n d itio n , th e to ta l system re sp o n se is bounded.
100 Example 2.43
Chap. 2
Determ ine if the linear tim e-invariant recursive system described by the difference equation given in (2.4.7) is stable. Solution i*(n)l < Let us assume that the input signal x(n) is bounded in amplitude, that is, < oc for all n > 0. From (2.4.8) we have j v{n J | < |a',+l_y(1)| + ^ akx(n - k) , n>0
If n is finite, the bound M v is finite and the output is bounded independently of the value of a. However, as n -* oo, the bound My remains finite only if |aj < 1 because |a |B - 0 as n -* oc. Then M y = Ms j(\ - |o|). Thus the system is stable only if \a\ < 1. For the sim ple first-order system in E xam ple 2.4,3, w e w ere able to express the con d ition for B IB O stability in term s o f the system param eter a. nam ely \a\ < 1. W e should stress, how ever, that this task b ecom es m ore difficult for higher-order system s. F ortunately, as we shall see in su bsequent chapters, other sim ple and m ore efficient tech n iqu es exist for investigating the stability o f recursive system s.
2.4.3 Solution of Linear Constant-Coefficient Difference Equations

G iven a linear con stan t-coefficien t d ifferen ce eq u ation as the in p u t-o u tp u t rela tionship describing a linear tim e-invariant system , our objective in this subsection is to determ ine an explicit exp ression for the output y{n). T h e m eth od that is d ev elo p ed is term ed the direct m e t h o d . A n alternative m eth od based on the ztransform is described in Chapter 3. For reasons that will b eco m e apparent later, the z-transform approach is called the indirect m e th o d . Basically, the goal is to determ ine the output y(n ), n > 0, o f the system given a specific input x ( n ), n > 0, and a set o f initial con d ition s. T h e direct solution m ethod assum es that the total solu tion is the sum o f tw o parts: y ( n ) = y h(n ) + }'r (n) T he part y/,(n) is know n as the h o m o g e n e o u s or c o m p l e m e n ta r y solu tion , w hereas yp(n) is called the particular solution.
The homogeneous solution of a difference equation. W e begin the problem o f solving the linear con stan t-coefficien t d ifferen ce eq u ation given by
Sec. 2.4
101
(2.4.13) bv assu m in g th a t th e in p u t x (n ) = 0. T h u s w e will first o b ta in th e so lu tio n to th e h o m o g e n e o u s di ff erence equati on \ J 2 a ky ( n - k ) = 0 *=(> (2.4.14)
T h e p ro c e d u re fo r solving a lin e a r c o n s tan t-co efficien t d ifferen ce e q u a tio n d irectly is very sim ilar to th e p ro c e d u re fo r solving a iin e a r c o n stan t-co efficien t d iffe re n tia l e q u a tio n . B asically, we assum e th a t th e so lu tio n is in th e fo rm o f an e x p o n e n tia l, th a t is. yi,{n) = a" (2.4.15)
w h ere th e su b scrip t h o n \ {n) is used to d e n o te th e so lu tio n to th e h o m o g e n e o u s d ifferen ce e q u a tio n . If w e su b stitu te this assu m ed so lu tio n in (2.4.14), w e o b tain th e p o ly n o m ial e q u a tio n =0 fc=U or k" A (k* + ci)k^ ! + a2k^ * + - +
q n \ k
+ aw) = 0
(2.4.16)
T h e p o ly n o m ial in p a re n th e se s is called th e characteristic p o l y n o m i a l o f the system . In g e n e ra l, it h as N roots, w hich we d e n o te as Xi. k 2 ......... k ^ . T h e ro o ts can b e real o r co m p lex v alued. In p ra c tic e the co efficien ts a \ , a 2 ........ are usually re a l. C o m p le x -v a lu e d ro o ts o ccu r as c o m p le x -c o n ju g a te p airs. S om e o f th e N ro o ts m ay be id en tical, in w hich case w e have m u ltip Je -o rd e r ro o ts. F o r th e m o m e n t, let us assum e th a t th e ro o ts a re d istin ct, th a t is, th e re are n o m u ltip le -o rd e r ro o ts. T h e n th e m o st g e n eral so lu tio n to th e h o m o g e n e o u s d iffe re n c e e q u a tio n in (2.4.14) is yh(n) = C \ k \ + C2k 2 + + C^ k ^ , (2.4.17)
w h ere C i. C 2.........C s are w eig h tin g coefficients. T h e se co efficien ts are d e te rm in e d fro m th e in itial c o n d itio n s specified fo r th e sy stem . Since th e in p u t jr(n) = 0. (2.4.17) can b e u sed to o b ta in th e z ero- i nput response o f th e system . T h e follow ing ex am p les illu stra te th e p ro c e d u re . Example 2.4.4 D eterm ine the hom ogeneous solution of the system described by the first-order dif ference equation _v(n) + a\y(rt 1) = x(n) Solution The assumed solution obtained by setting x(n) = 0 is y*(n) = V (2.4.18)
102
D iscrete-Time Signals and System s
Chap. 2
When we substitute this solution in (2.4.18), we obtain [with x(n) = 0] X -f-1 2 ]An ^ = 0 A" *(A + ai) = 0 A= O) Therefore, the solution to the hom ogeneous difference equation is = CA" = C ( - a ,)" (2.4.19)
The zero-input response of the system can be determ ined from (2.4,18) and (2.4.19). With x(n) = 0, (2.4.18) yields v(0) =
-a, v ( - l )
On the other hand, from (2.4,19) we have v*(0) = C and hence the zero-input response of the system is Vii() = ( i)n+1 v{1) n >0 (2.4.20)
With a = ai, this result is consistent with (2.4,11) for the first-order system, which was obtained earlier by iteration of the difference equation.
Example 2AS
Determ ine the zero-input response of the system described by the homogeneous second-order difference equation y() - 3 y(n - 1) - 4y(n - 2) = 0 the solution to be the exponential yh(n ) = X " U pon substitution of this solution into (2.4.21), we obtain the characteristic equation
a"
(2.4.21)
Solution First we determ ine the solution to the homogeneous equation. We assume
- 3xn- ' - 4 r ~ 2 = o
A"-2(X2 3A 4) = 0 Therefore, the roots are X = - 1 , 4, and the general form of the solution to the homogeneous equation is }7i(n) = CjX" + C 2X2 = C1( - i r + C2(4)" The zero-input response of the system can be obtained from the homogenous solution by evaluating the constants in (2.4.22), given the initial conditions y (1) and y ( ~ 2). From the difference equation in (2.4.21) we have y(0) = 3y(1) + 4y( 2) y(l) = 3v(0) + 4_y(1) = 3 [3 y (-l)+ 4y(-2)] + 4y(-l) = 13y(1) + l 2y( 2) (2.4.22)
Sec. 2.4
103
On the other hand, from (2.4.22) we obtain v(0> = C) + C;
V ( 1 I = -C l +4C;
By equating these two sets of relations, we have C i + C ; = 3_v( 1) + 4y( 2) - C , + 4 C ; = 13 v{ 1) + 12 y( 2) The solution of these two equations is Ci = 7 v( 1 ) + ? v ( - 2 ) C; = X .V (-l) + T.v( -2 ) Therefore, the zero-input response of the system is
v,i(n) = [ 5 v( 1 >- i v(-2)](-l)''

(2.4.23) + [ x . v ( - l + t.v(-2)](4>" n >0 For example, if v{ 2) = 0 and y ( - l ) = 5. then C| = 1, C: = 16. and hence y-,i(H) = ( -1 ) " * 1 + (4)"+: n >0
T h e se e x am p les illu strate the m e th o d fo r o b ta in in g the h o m o g e n e o u s so lu tio n and th e z e ro -in p u t re sp o n se o f th e system w hen the ch a ra c te ristic e q u a tio n co n tain s d istin ct ro o ts. O n th e o th e r h an d , if th e c h a ra c te ristic e q u a tio n co n ta in s m u ltip le ro o ts, th e form o f th e so lu tio n given in (2.4.17) m ust be m odified. F o r ex am p le, if ai is a ro o t o f m u ltip licity m, th en (2.4.17) b eco m es
\ h(ii) = Ci A .'.' + C2/iX! + C v;:a',' + + Cm> 7 n'~lX'!

(2.4.24) + Cm+\)"m+\ + + C^Kl
The particular solution of the difference equation. T h e p a rtic u la r so lu tio n \ p (n) is re q u ire d to satisfy th e d ifferen ce e q u a tio n (2.4.13) fo r th e specific in p u t signal x (n ). n > 0. In o th e r w ords, y p(n) is any so lu tio n satisfying
N M
' Y ^ a ky p (n - k) ' Y ^ b kx ( n - k) A-0 *=0
an 1
(2.4.25)
T o solve (2.4.25). w e assu m e fo r yp (n), a fo rm th at d e p e n d s on th e fo rm o f th e in p u t j:(h ). T h e fo llow ing ex am p le illu strates the p ro c e d u re . Example 2.4.6 D eterm ine the particular solution of the first-order difference equation y(n) + fliy(n - 1) = x(n). | f l i | <l (2,4.26)
when the input x(n) is a unit step sequence, that is.

x(n) = u (n )
104
Discrete-Time Signals and System s
Chap. 2
Solution Since the input sequence x(n) is a constant for n > 0. the form of the solu tion that we assume is also a constant. Hence the assumed solution of the difference equation to the forcing function x(n), called the particular solution of the difference equation, is vr (n) = Ku(n) where A - is a scale factor determined so that (2.4.26) is satisfied. U pon substitution of this assumed solution into (2.4.26). we obtain
Ku i n ) + d]Ku{,n 1) = u ( n )
To determine K, we must evaluate this equation for any n > 1. where none of the terms vanish. Thus
K -t- O ]K 1
1 K = -------l+î Therefore, the particular solution to the difference equation is v (n) = -------- u(n) ' 1 +fii (2.4.27)
In th is ex am p le, th e in p u t * ( ). n > 0. is a c o n s ta n t an d th e fo rm assu m ed for th e p a rtic u la r so lu tio n is also a c o n sta n t. If x{n) is an e x p o n e n tia l, w e w ould assu m e th a t th e p a rtic u ia r so lu tio n is also an e x p o n e n tia l. If x {n) w e re a sin u so id , th e n >/,() w o uld also be a sinusoid. T h u s o u r assu m ed fo rm fo r th e p a rtic u la r so lu tio n tak es th e basic fo rm of th e signal x{n). T a b le 2.1 p ro v id e s th e g en eral fo rm o f th e p a rtic u la r so lu tio n for sev eral ty p es of excitatio n . Exam ple 2.4.7 Determ ine the particuiar solution of the difference equation v(n) = | y(n - 1) - v(w - 2) + x(n) when the forcing function x(n) = 2". n > 0 and zero elsewhere.
TABLE 2.1 GENERAL FORM OF THE PARTICULAR SOLUTION FOR SEVERAL TYPES OF INPUT SIGNALS
Input Signal, x(n) A (constant) AM" AnM
Particular Solution, vr (n) K KM" Kan + K ^ " - ' 1 + . . . + KM
AnnM
An(K i)nM+ K\ttM 1
A cos W f> n A sin a>on
Kj cos toon -I- K2 sin won
Sec. 2.4
105
Solution
The form of the particular solution is yp (n ) = K2n n >0
Upon substitution of yP(n) into the difference equation, we obtain

K2"u( n) =
-
1)
- I K 2 n- 2u(n - 2 ) + 2 " ( )
To determ ine the value of K, we can evaluate this equation for any n > 2, where none of the terms vanish. Thus we obtain 4A " = \ ( 2K) - i f f + 4 and hence K = Therefore, the particular solution is yn(n) = ^2" n > 0
W e h av e n o w d e m o n s tra te d how to d e te rm in e th e tw o c o m p o n e n ts o f the so lu tio n to a d ifferen ce e q u a tio n w ith c o n s ta n t coefficients. T h e se tw o c o m p o n e n ts are th e h o m o g e n e o u s so lu tio n and th e p a rtic u la r so lu tio n . F ro m th e s e tw o co m p o n e n ts, we c o n s tru c t the to tal so lu tio n fro m w hich w e can o b ta in th e ze ro -sta te re sp o n se .
The total solution of the difference equation. T h e lin e a rity p ro p e rty o f th e lin ear c o n stan t-co efficien t d ifferen ce e q u a tio n allow s u s to a d d th e h o m o g e n e o u s so lu tio n an d the p a rtic u la r so lu tio n in o rd e r to o b ta in th e total solution. T h u s
v() =
y h(n)
+ \ p( n)
T h e r e s u lta n t sum v(n) co n tain s th e c o n s ta n t p a r a m e te r s {C/} e m b o d ie d in th e h o m o g e n e o u s so lu tio n c o m p o n e n t >v,(n). T h e se c o n s ta n ts can be d e te rm in e d to satisfy th e in itial co n d itio n s. T h e follow ing ex a m p le illu stra te s th e p ro c e d u re , Exam ple 2.4.8 D eterm ine the total solution y(/i), n > 0, to the difference equation y(n) +a i y ( n - 1) = x(n) (2.4.28)
when x( n) is a unit step sequence [i.e., x(n) = ()] and y (1) is the initial condition. Solution From (2.4,19) of Example 2.4.4, the hom ogeneous solution is
y*(n) = C(-fli)"
and from (2.4.26) of Example 2.4.6, the particular solution is
Consequently, the total solution is y(n) = C ( - a i)" + - 1 + fli n >0 (2.4.29)
where the constant C is determ ined to satisfy the initial condition y ( - l ) .
106
Chap. 2
In particular, suppose that we wish to obtain the zero-siate response of the system described by the first-order difference equation in (2.4.28). T hen we set y( 1) = 0. To evaluate C. we evaluate (2.4.28) at n = 0 obtaining
y ( 0 ) + Oi y ( 1 ) = 1
y(0) = 1 On the other hand, (2.4.29) evaluated at n = 0 yields v(0) = C 4- i 1 + ax Consequently. 1 C + -------- = 1 1 + i
ai 1 + ai
Substitution for C into (2.4.29) yields the zero-state response of the system l-f-fli)" * ' V i*(/i) = ---- ----- 1 + i n >0
If we evaluate the param eter C in (2,4.29) under the condition that y( 1) ^ 0. the total solution will include the zero-input response as well as the zero-state response of the system. In this case (2.4.28) yields y(0) + a]V( 1) = 1
y (0) = fliy( 1) 4- 1
On the other hand. (2.4.29) yields ] 1+ U \ By equating these two relations, we obtain C + ^ = ai v(1) 4- 1
1 4 - a,
C = -g ] v( 1) +
-
1 4- a ]
Finally, if we substitute this value of C into (2.4.29). we obtain y(n) = ( a , r +1y(1) + -

= + Vzs( n)
^ 1 + a)
n >0
(2.4.30)
W e o b se rv e th a t th e system re sp o n se as given by (2.4.30) is c o n s iste n t w ith th e re sp o n se y ( n) given in (2.4.8) for th e first-o rd e r system (w ith a = - o i ) . w hich was o b ta in e d by solv in g th e d ifferen ce e q u a tio n iterativ e ly . F u rth e rm o r e , we n o te th a t th e v alu e o f th e c o n s ta n t C d e p e n d s b o th o n th e initial co n d itio n y (1) an d o n th e ex citatio n fu n ctio n . C o n se q u e n tly , th e v alue o f C in flu en ces b o th th e zeroin p u t re sp o n se a n d th e z e ro -sta te re sp o n se . O n th e o th e r h a n d , if w e w ish to
Sec. 2.4
107
obtain the zero-state resp onse only, w e sim ply solve for C under th e con d ition that y ( - l ) = 0, as d em onstrated in E xam ple 2.4.8. W e further ob serve that the particular solution to the d ifferen ce equation can b e o b tain ed from the zero-state resp onse o f the system . In d eed , if |o]| < 1, which is the con d ition for stability o f the system , as w ill be show n in S ection 2.4.4, the lim iting valu e o f >zs() as n approaches infinity, is the particular solu tion , that is, 1 ( n ) = lim >zs() = --------n-*Oo 1+ai Since this com p onent o f the system response d o es not go to zero as n approaches infinity, it is usually called the steady-state re spons e o f the system . T his response persists as lon g as the input persists. T he com p onent that d ies out as n approaches infinity is called the transient re sponse o f the system .
Example 2.4.9
D eterm ine the response y(n), n > 0, of the system described by the second-order difference equation v(n) - 3 y(n - 1) - 4v(n - 2) = x( n ) + 2x(n - 1) when the input sequence is x(n) =4"u(n) (2.4.31)
Solution We have already determ ined the solution to the homogeneous difference
equation for this system in Example 2.4.5. From (2.4.22) we have y*(n) = C , ( - l ) n + C 2(4)n (2.4.32)
The particular solution to (2.4.31) is assumed to be an exponential sequence of the same form as x(n). Normally, we could assume a solution of the form yp(n) = K(4,)au(n) However, we observe that ( n ) is already contained in the hom ogeneous solution, so that this particular solution is redundant. Instead, we select the particular solution to be linearly independent of the terms contained in the homogeneous solution. In fact, we treat this situation in the same manner as we have already treated multiple roots in the characteristic equation. Thus we assume that yp(n) = Kn{4)"u(n) Upon substitution of (2.4.33) into (2.4.31), we obtain Kn(4)"u(n) - 3 K ( n - l)(4 ),' - 1u(n - 1) - 4 K{n - 2)(4)n- 2u(n - 2) = (4)"u(n) + 2(4)"-1 m(/i 1) To determ ine K , we evaluate this equation for any n > 2, where none of the unit step term s vanish. To simplify the arithmetic, we select n = 2, from which we obtain K = | . Therefore, (2.4.33)
yP{n) |rt(4)n(/l)
(2.4.34)
108
Chap. 2
The total solution to the difference equation is obtained by adding (2.4.32) to (2.4.34). Thus
y() C ](1)" + C;(4)n + |w(4)"
n > 0
(2.4.35)
where the constants C \ and C2 are determined such that the initial conditions are satisfied. To accomplish this, we return to (2.4.31), from which we obtain v(0) = 3 y (- l) + 4y (2) + 1 y (1) = 3v(0) + 4_v(-l) + 6 = 1 3 y (- l) + 12y(-2) + 9 O n the other hand, (2.4.35) evaluated at n = 0 and n = 1 yields y(0) = Ci + C2 y (l) = -C] + 4C2 + f W e can now equate these two sets of relations to obtain C \ and C 2. In so doing, we have the response due to initial conditions y (1) and y (2) (the zero-input response), and the zero-state or forced response. Since we have already solved for the zero-input response in Exam ple 2.4.5. we can simplify the computations above by setting v (1) = y (2) = 0. Then we have C, + C; = 1 - C l + 4 C2 + f
=9
Hence C : = ^ and C 2 = Finally, we have the zero-statc response to the forcing function = (4)"(n) in the form
v B (n )=
+5< 4 )" +H 4)"
nZ0
(2.4.36)
The total response of the system, which includes the response to arbitrary initial conditions, is the sum of (2.4.23) and (2.4.36).
2.4.4 The Impulse Response of a Linear Time-Invariant Recursive System

T h e im p u lse re sp o n se o f a lin e a r tim e -in v a ria n t sy stem w as p re v io u sly defin ed as th e re sp o n se o f th e sy stem to a u n it sa m p le e x citatio n [i.e., x (n ) = <5(n)]. In th e case o f a recu rsiv e sy stem , h ( n ) is sim ply e q u a l to th e z e ro -s ta te re sp o n s e o f the system w h en th e in p u t j:(n ) = <5(n) an d th e system is initially re la x e d . F o r ex am p le, in th e sim ple first-o rd e r re c u rsiv e system g iv en in (2.4.7), th e z e ro -sta te re sp o n se g iv en in (2.4.8), is n
(2.4.37)
*=o W ith jr(n) = i ( ) is substituted into (2.4.37), we obtain n yzs(n) = Y 2 a k8(n - k)
= an n > 0
Sec. 2.4
109
H e n c e th e im p u lse re sp o n se o f th e first-o rd e r rec u rsiv e system d e sc rib e d by (2.4.7) is h(n) = a nu(n) (2.4.38)
as in d ic a te d in S ectio n 2.4.2. In th e g e n e ra l case o f an a rb itra ry , lin e a r tim e -in v a ria n t rec u rsiv e system , the z e ro -s ta te re sp o n s e ex p re sse d in te rm s o f th e co n v o lu tio n su m m a tio n is > zs(n) ^ 2 h ( k ) x ( n k)
i=0
n > 0
(2.4.39)
W h e n th e in p u t is an im p u lse [i.e., * ( ) = <5()], (2.4.39) re d u c e s to Vzs(n) = h(n) (2.4.40)
N o w , let us c o n s id e r th e p ro b le m o f d e te rm in in g th e im p u lse re sp o n se h i n ) given a lin e a r c o n s tan t-co efficien t d ifferen ce e q u a tio n d e sc rip tio n o f th e system . In term s o f o u r d isc u ssio n in th e p re c e d in g su b sectio n , we h av e e s ta b lis h e d th e fact th a t the to ta l re sp o n s e o f th e sy stem to any e x citatio n fu n ctio n co n sists o f th e sum o f tw o so lu tio n s o f th e d ifferen ce e q u atio n : th e so lu tio n to the h o m o g e n e o u s e q u atio n p lu s th e p a rtic u la r so lu tio n to th e e x c itatio n fu n ctio n . In th e case w h ere the exci ta tio n is an im p u lse, th e p a rtic u la r so lu tio n is z e ro , since x ( n ) = 0 for n > 0. th at is. yp (n) = 0 C o n s e q u e n tly , th e re sp o n se o f th e system to an im p u lse co n sists only o f th e so lu tio n to th e h o m o g e n e o u s e q u a tio n , w ith the ( Q ) p a r a m e te r s e v a lu a te d to satisfy th e in itial c o n d itio n s d ic ta te d by th e im pulse. T h e follow ing ex am p le illu strates th e p ro c e d u re fo r o b ta in in g h(n) given th e d ifferen ce e q u a tio n fo r the system . Example 2.4.10 D eterm ine the impulse response h(n) for the system described by the second-order difference equation
y(n ) 3v(n 1) 4 y ( 2) = x ( n ) + 2 x i n 1) (2.4.41)
Solution W e have already determ ined in Example 2.4.5 that the solution to the hom ogeneous difference equation for this system is ^ ( n ) = C, (-1 )" + C2(4)"
n > 0
(2.4.42)
Since the particular solution is zero when x(n) = 6(n), the impulse response of the sys tem is simply given by (2.4.42), where C] and C2 must be evaluated to satisfy (2.4.41). For n = 0 and n = 1, (2.4.41) yields v(0) = 1 y (1) = 3 y (0 )+ 2 = 5 where we have imposed the conditions y ( 1) = y ( 2) = 0. since the system must be relaxed. On the other hand, (2.4.42) evaluated at n = 0 and n = 1 yields y (0) = C, + C2
y (1) = Ci + 4C2
110
Chap. 2
By solving these two sets of equations for C] and C2, we obtain = C: = 5
Therefore, the impulse response of the system is h{n) = [i ( 1)" + f (4)-}u(n) W e m ak e th e o b se rv a tio n th a t b o th th e sim ple firs t-o rd e r recu rsiv e system an d th e se c o n d -o rd e r recu rsiv e system h av e im pulse re sp o n se s th a t a re infinite in d u ra tio n . In o th e r w o rds, b o th o f th e s e recu rsiv e system s a re IIR system s. In fact, d u e to th e recu rsiv e n a tu re of th e system , an y recu rsiv e sy stem d e sc rib e d by a lin e a r co n stan t-co efficien t d iffe re n c e e q u a tio n is an I IR system . T h e co n v erse is n o t tru e, h o w ev er. T h a t is, n o t ev ery lin e a r tim e -in v a ria n t I I R sy stem can be d e sc rib e d by a lin e a r c o n s tan t-co efficien t d ifferen ce e q u a tio n . In o th e r w ords, recu rsiv e sy stem s d e sc rib e d by lin ear co n sta n t-c o e ffic ie n t d iffe re n c e e q u a tio n s are a su bclass o f lin e a r tim e -in v a ria n t I I R system s. T h e ex ten sio n o f th e a p p ro a c h th a t w e h av e d e m o n s tra te d fo r d e te rm in ing th e im p u lse re sp o n se o f th e first- a n d s e c o n d -o rd e r system s, g en e ra liz e s in a stra ig h tfo rw a rd m a n n e r. W h e n th e system is d e sc rib e d by an A 'th -o rd e r lin e a r d ifferen ce e q u a tio n o f th e ty p e given in (2.4.13), th e so lu tio n o f th e h o m o g e n e o u s e q u a tio n is
*=i w h en th e ro o ts {a*} o f th e c h a ra c te ristic p o ly n o m ial are d istin ct. H e n c e th e im pulse re sp o n se o f th e sy stem is id en tical in fo rm , th a t is. (2.4.43) w h ere th e p a ra m e te rs {Ctl are d e te rm in e d by se ttin g th e initial c o n d itio n s v ( 1) = . . . = y ( - N ) = 0. T h is fo rm o f h{n) allow s us to easily re la te th e stab ility of a system , d escrib ed by an N th -o rd e r d iffe re n c e e q u a tio n , to th e values o f th e ro o ts o f th e ch a ra c te ristic p o ly n o m ial. In d e e d , since B IB O sta b ility re q u ire s th a t th e im p u lse re sp o n s e be ab so lu te ly su m m ab le, th e n , fo r a causal system , w e h av e
x oc N N oo oo > (* )! = Y ^ C kXn k < |C * | |A * |" =0 n =0 n=0 Ik=-l k=\ n=0
N ow if |
j < 1 fo r all k, th e n
an d h en ce
OC
IM*)[ < oo
Sec. 2.5
im plementation of Discrete-Time Systems
111
O n th e o th e r h an d , if o n e o r m o re o f th e > 1, h{n) is n o lo n g e r ab so lu te ly su m m ab le, an d c o n se q u e n tly , th e sy stem is u n sta b le . T h e re fo re , a n ecessa ry an d sufficient c o n d itio n fo r th e sta b ility o f a causal IIR system d e sc rib e d by a lin ear c o n s tan t-co efficien t d iffe re n c e e q u a tio n , is th a t al! ro o ts o f th e c h a ra c te ristic p o ly n o m ial b e less th a n u n ity in m a g n itu d e. T h e re a d e r m ay verify th a t th is c o n d itio n ca rrie s o v er to th e case w h ere th e system h as ro o ts o f m u ltip lic ity m.
2.5 IMPLEMENTATION OF DISCRETE-TIME SYSTEMS

O u r tre a tm e n t o f d isc re te -tim e sy stem s h as b e e n fo cu sed on th e tim e -d o m a in c h a r ac te riz a tio n an d analysis o f lin ear tim e -in v a ria n t system s d esc rib e d by c o n stan tco efficien t lin e a r d ifferen ce e q u a tio n s. A d d itio n a l an aly tical m e th o d s a re d e v e l o p e d in th e n ex t tw o c h a p te rs, w h ere we c h a ra c te riz e an d an aly ze L TI sy stem s in th e fre q u e n c y d o m ain . T w o o th e r im p o rta n t to p ics th a t will b e tr e a te d la te r are th e d esig n an d im p le m e n ta tio n o f th e se system s. In p ractice, system design a n d im p le m e n ta tio n a re usually tr e a te d jo in tly r a th e r th a n se p a ra te ly . O fte n , th e system design is d riv en by th e m e th o d o f im p le m e n ta tio n an d by im p le m e n ta tio n c o n stra in ts, such as cost, h a rd w a re lim itatio n s, size lim ita tio n s, a n d p o w e r re q u ire m e n ts . A t th is p o in t, we h av e n o t as y e t d e v e lo p e d th e n ecessa ry analysis an d design tools to tr e a t such com plex issues. H o w ev er, we h ave d e v e lo p e d sufficient b a c k g ro u n d to c o n sid e r so m e b a sic im p le m e n ta tio n m e th o d s for re a liz a tio n s o f L TI sy stem s d escrib ed by lin ear c o n s tan t-co efficien t d iffe re n c e e q u atio n s.
2.5.1 Structures for the Realization of Linear Time-Invariant Systems

In th is su b sectio n we d esc rib e s tru c tu re s fo r th e re a liz a tio n o f system s d escrib ed by lin e a r c o n s tan t-co efficien t d ifferen ce e q u a tio n s. A d d itio n a l stru c tu re s fo r th ese sy stem s a re in tro d u c e d in C h a p te r 7. A s a b eg in n in g , let us c o n sid e r th e first-o rd e r system y(n) = a iy (n 1) + box ( n ) + b \ x ( n - 1) (2.5.1)
w hich is re a liz e d as in Fig. 2.32a. T h is re a liz a tio n uses s e p a ra te delays (m em o ry ) fo r b o th th e in p u t a n d o u tp u t signal sa m p le s a n d it is called a direct f o r m I structure. N o te th a t th is sy stem can b e view ed as tw o lin e a r tim e -in v a ria n t system s in cascad e. T h e first is a n o n re c u rsiv e , sy stem d e sc rib e d by th e e q u a tio n
u(n) = Zoj;(n) + 6 j j ( n 1) (2.5.2)
w h e re a s th e se c o n d is a rec u rsiv e system d e sc rib e d by th e e q u a tio n y ( n ) = - a i y ( n - 1) + t>{) (2.5.3)
H o w e v e r, as we h a v e se e n in S ectio n 2.3.4, if w e in te rc h a n g e th e o r d e r o f the c a s cad ed lin e a r tim e -in v a ria n t system s, th e o v erall system re sp o n s e re m a in s th e
112
Chap. 2
sam e. T h u s if w e in te rc h a n g e the o rd e r o f the recu rsiv e an d n o n re c u rsiv e system s, we o b ta in an a lte rn a tiv e s tru c tu re fo r th e re a liz a tio n o f th e sv stem d e sc rib e d by (2.5.1). T h e re su ltin g system is show n in Fig. 2.32b. F ro m th is figure we o b tain th e tw o d ifferen ce e q u a tio n s w( n) fl] w( n 1) + jt() v(n) = t>nw(n) + b\ w{ n - 1) (2-5.4) (2.5.5)
w hich p ro v id e an a lte rn a tiv e a lg o rith m for co m p u tin g th e o u tp u t o f the system d e sc rib e d by th e single d ifferen ce e q u a tio n given in (2.5.1). In o th e r w o rd s, the tw o d ifferen ce e q u a tio n s (2.5.4) an d (2.5.5) are e q u iv a le n t to th e single differen ce e q u a tio n (2.5.1). A close o b se rv a tio n o f Fig. 2.32 re v e a ls th at th e tw o d elay e le m e n ts co n tain th e sam e in p u t w( n) an d h en c e the sam e o u tp u t w( n 1). C o n s e q u e n tly , these tw o e le m e n ts can be m e rg ed in to o n e delay, as show n in Fig. 2.32c. In c o n tra st
x(n)
-----------------~
b{ ,
r(n)
vim
(a)
(b.)
(c)
Figure 132 Steps in converting from the direct form I realization in (a) to the direct form II realization in (c).
Sec. 2.5
Implementation of Discrete-Time Systems
113
to th e d ire c t fo rm I s tru c tu re , th is new re a liz a tio n re q u ire s o n ly o n e d elay fo r th e au x iliary q u a n tity ui(n), an d h en c e it is m o re efficient in te rm s o f m em o ry re q u ire m e n ts . It is called th e direct f o r m I I structure a n d it is u sed ex ten siv ely in p ra c tic a l a p p licatio n s. T h e se stru c tu re s can read ily be g e n e ra liz e d fo r th e g e n e ra l lin e a r tim ein v a ria n t recu rsiv e sy stem d escrib ed by th e d iffe re n c e e q u a tio n
N M
y (n ) = - y ^ a ^ .y ( - k) +
*=l k=0
- k)
(2.5.6)
F ig u re 2.33 illu stra te s th e d ire c t fo rm I s tru c tu re fo r th is sy stem . T h is stru c tu re re q u ire s M + N d elay s a n d N + M + 1 m u ltip lic atio n s. I t c a n be v iew ed as the c ascad e o f a n o n re c u rsiv e system M i;() = Y bkx{n - k)
i=U
(2.5.7)
a n d a recu rsiv e system

s
y(n) = -
v(n ~ k) + v(n)
(2.5.8)
By rev ersin g th e o r d e r o f th ese tw o system s as was p rev io u sly d o n e fo r th e first-o rd e r sy stem , w e o b ta in th e d ire c t form II s tru c tu re sh o w n in Fig. 2.34 fo r
Figure 1 3 3
Direct form I structure of the system described by (2.5.6).
114

w(n) b0 N
Chap. 2
y(n )
-1 U J{ /1 - 1) J -1
w(n - 2) b2
w(n - 3)
by
Figure 2 3 4
Direct form II structure for the system described by (2.5.6).
N > M. T h is stru c tu re is the cascad e o f a recu rsiv e system

N
w( n) = ^
ai W(n - k ) -f x( n)
(2.5.9)
*=1
fo llo w ed by a n o n re c u rsiv e sy stem M v(n) = vu(n - k ) (2.5.10)
W e o b serv e th a t if N > M . this s tru c tu re re q u ire s a n u m b e r o f d elay s e q u a l to th e o rd e r N o f th e sy stem . H o w e v e r, if M > N , th e re q u ire d m e m o ry is specified by M . F ig u re 2.34 can easily by m o d ified to h a n d le th is case. T h u s th e d ire c t form II stru c tu re re q u ire s M + N + 1 m u ltip lic a tio n s a n d m ax{M , jV} d elay s. B e cau se it re q u ire s th e m in im u m n u m b e r o f d elay s fo r th e re a liz a tio n o f th e system d e sc rib e d by (2.5.6), it is so m e tim es ca lle d a canoni c f o r m .
Sec. 2.5
115
A special case o f (2.5.6) occurs if w e set th e system p a ra m e te rs ak 0. it = 1.........N. T h e n th e in p u t-o u tp u t re la tio n sh ip fo r th e sy stem re d u c e s to M v(n) = k=0 w hich is a n o n re c u rsiv e lin e a r tim e -in v a ria n t system . T his sy stem view s only the m o st re c e n t M + 1 in p u t signal sa m p les a n d , p rio r to a d d itio n , w eights each sam ple by th e a p p ro p ria te co efficien t bk from th e set {i^}. In o th e r w ords, th e system o u tp u t is b asically a weight ed m o v i n g average of th e in p u t signal. F o r this reaso n it is so m e tim e s called a m o v i n g average ( M A ) syst em. S uch a system is an F IR sy stem w ith an im p u lse re sp o n s e h (k ) e q u a l to th e coefficients bk, th at is. - k) (2.5.11)
* '* ) = { o ! '
o to rw l"
'5 1 2
If w e r e tu rn to (2.5.6) a n d se t M = 0, th e g e n eral lin e a r tim e -in v a ria n t system red u ces to a p u re ly re c u rsiv e system d escrib ed by th e d ifferen ce e q u a tio n N y ( n ) = ~ Y a O (n ~ k) + bux(n) i =1 (2.5.13)
In th is case th e sy stem o u tp u t is a w eig h ted lin ear c o m b in a tio n o f N past o u tp u ts a n d th e p re s e n t in p u t. L in e a r tim e -in v a ria n t system s d esc rib e d by a se c o n d -o rd e r d ifferen ce e q u a tio n a re an im p o rta n t subclass o f th e m o re g e n eral system s d e sc rib e d by (2.5.6) o r (2.5.10) o r (2.5.13). T h e re a so n fo r th e ir im p o rta n c e will be ex p lain ed later w h en w e discuss q u a n tiz a tio n effects. Suffice to say at this p o in t th a t se c o n d -o rd e r sy stem s a re u su ally u sed as b a sic b u ild in g b lo ck s fo r realizin g h ig h e r-o rd e r system s. T h e m o st g e n e ra l se c o n d -o rd e r sy stem is d escrib ed by th e d ifferen ce e q u a tio n y (n ) = - a[v( n - 1) - a2v(n - 2) 4- b(,x(n) (2.5.14) + b\ x ( n 1) +- b 2x ( n 2) w hich is o b ta in e d fro m (2.5.6) b y se ttin g N = 2 an d M = 2. T h e d irect form II s tru c tu re fo r realizin g th is sy stem is sh o w n in Fig. 2.35a. If we se t a\ = = 0. th e n (2.5.14) re d u c e s to y( n) = box(n) + b\x(rt 1) + b 2x ( n - 2) (2.5.15)
w hich is a sp ecial case o f th e F IR system d esc rib e d by (2.5.11). T h e stru c tu re fo r realizin g th is sy stem is sh o w n in Fig. 2.35b. F inally, if w e set b\ = bz = 0 in (2.5.14), w e o b ta in th e p u re ly recu rsiv e se c o n d -o rd e r sy stem d esc rib e d by th e d iffe re n c e e q u a tio n y( n) = a \ y ( n 1) a 2y(n - 2) 4- box(rt) (2.5.16)
w hich is a sp e cial case o f (2.5.13). T h e stru c tu re fo r realizin g th is system is show n in Fig. 2.35c.
116
Chap. 2
(a)
----- -0
(b)
x i n)
Y[n)
-^-1 H
I _a' -1
02 -1 <c)
Figure 2.35 Structures for the realization of second-order systems: (a) general second-order system; (b) FIR system; (c) purely recursive system
2.5.2 Recursive and Nonrecursive Realizations of FIR Systems

W e h a v e a lread y m a d e th e d istin ctio n b etw e e n F IR an d IIR system s, b ased on w h e th e r th e im p u lse re sp o n se h( n) o f th e system h a s a finite d u ra tio n , o r an infi n ite d u ra tio n . W e h av e also m a d e th e d istin c tio n b e tw e e n rec u rsiv e a n d n o n re c u r sive system s. B asically, a causal recu rsiv e system is d esc rib e d by an in p u t-o u tp u t e q u a tio n o f th e fo rm y(n) = F[v(î 1).........y{n N ) , x ( n ) , ------ x ( n - M)] (2.5.17)
an d fo r a lin ear tim e -in v a ria n t system specifically, by th e d iffe re n c e e q u a tio n

N U
y { n ) = - Y ak>(n ~ k ) + Y , bkX(n ~ k ) *=i *=o
(2.5.18)
Sec. 2.5
117
O n th e o th e r h a n d , cau sal n o n recu rsiv e system s d o n o t d e p e n d o n p a st values of th e o u tp u t a n d h e n c e a re d escrib ed by an in p u t-o u tp u t e q u a tio n o f th e form y (n) = F [x (n ), jc(n - 1).........x ( n Mj ] (2.5.19)
a n d fo r iin e a r tim e -in v a ria n t system s specifically, by th e d iffe re n c e e q u a tio n in (2.5.18) w ith a k : = 0 fo r k = 1, 2 , . . . , N. In th e case o f F IR system s, we h av e a lre a d y o b se rv e d th a t it is aiw ays possible to re a liz e such sy stem s n o n recu rsiv ely . In fact, w ith at . = 0, k = 1, 2 , . . . , N , in (2.5.18), w e h av e a sy stem w ith an in p u t-o u tp u t e q u a tio n v() = ] p 6 * ;t( n ~ (2.5.20)
T h is is a n o n re c u rsiv e a n d F IR system . A s in d ic a te d in (2.5.12), th e im pulse re sp o n s e o f th e sy stem is sim ply e q u a l to th e coefficients {&*). H e n c e every F IR sy stem can b e re a liz e d n o n recu rsiv ely . O n th e o th e r h a n d , an y F IR system can also b e realized recu rsiv ely. A lth o u g h the g e n eral p ro o f o f th is s ta te m e n t is given la te r, w e shall give a sim ple ex am p le to illu stra te th e p o in t. S u p p o se th a t w e h av e an F IR system o f th e form 1 m y( n) = M + If-' - k) (2.5.21)
fo r co m p u tin g th e mo v i n g average of a signal jc(h). C learly , th is sy stem is F IR w ith im p u lse re sp o n se h( n) ~ 1 M + 1 0 < n < M
F ig u re 2.36 illu stra te s th e stru c tu re of th e n o n re c u rsiv e re a liz a tio n o f th e system . N ow , su p p o se th a t w e ex p ress (2.5.21) as 1 M y (n ) = --------- x ( n 1 k) M m + ^ 1^ Jt=0 + . [x(n) ~ x ( n - 1 - M )] M + 1 1 M + V [x(n) x (n 1 Af)] (2.5.22)
y(n - 1) +
Figure 2 3 6
N onrecursive realization of an FIR moving average system.
118
Chap. 2
N ow , (2.5.22) re p re se n ts a recu rsiv e re a liz a tio n of th e F IR system . T h e stru c tu re o f th is recu rsiv e re a liz a tio n of th e m oving av erag e system is illu stra te d in Fig. 2.37. In su m m ary , w e can th in k of th e te rm s F IR a n d IIR as g e n e ra l ch aracteristics th a t distin g u ish a ty p e of lin e a r tim e -in v a ria n t system , and of th e term s recursive an d nonrecursi ve as d e sc rip tio n s of th e stru c tu re s fo r realizin g o r im p lem en tin g th e system .
vt/> - 1)
Figure 2 3 7 Recursive realization of an FIR moving averapc svstem.
2.6 CORRELATION OF DISCRETE-TIME SIGNALS

A m a th em atical o p e ra tio n th a t closely rese m b le s co n v o lu tio n is c o rre la tio n . Just as in th e case o f c o n v o lu tio n , tw o signal se q u e n c e s are in v o lv ed in c o rre la tio n . In c o n tra st to c o n v o lu tio n , h o w ev er, o u r o bjective in co m p u tin g th e c o rre la tio n b etw een th e tw o signals is to m e a su re th e d e g re e to w hich th e tw o signals are sim ilar an d th u s to e x tra c t so m e in fo rm a tio n th a t d e p e n d s to a large e x te n t on th e ap p licatio n . C o rre la tio n o f signals is o ften e n c o u n te re d in ra d a r, so n a r, digital co m m u n icatio n s, geolo g y, an d o th e r a re a s in science an d en g in eerin g . T o b e specific, le t us su p p o se th a t we have tw o signal se q u e n c e s x ( n ) an d y( n) th a t we w ish to co m p a re . In r a d a r an d active so n a r a p p lic a tio n s. x( n ) can re p re s e n t th e sa m p le d v ersio n o f th e tra n s m itte d signal and y{n) can re p re s e n t th e sam p led v ersio n o f th e receiv ed signal at th e o u tp u t o f the a n a lo g -to -d ig ita l (A /D ) c o n v erter. If a ta rg e t is p re se n t in th e space b e in g se a rc h e d by th e ra d a r o r so n a r, th e receiv ed signal y( n) consists of a d e la y e d v ersio n o f th e tra n sm itte d signal, reflec te d from th e ta rg e t, an d c o rru p te d b y a d d itiv e noise. F ig u re 2.38 d e p ic ts the ra d a r signal re c e p tio n p ro b le m . W e can re p re s e n t th e re c e iv e d signal se q u e n c e as y ( n ) = a x ( n D) + w( n ) (2.6.1)
w h ere a is som e a tte n u a tio n fa c to r re p re s e n tin g th e signal loss involved in th e ro u n d -trip tran sm issio n o f th e signal x (n ), D is th e ro u n d -trip d elay , w hich is
Sec. 2.6
Correlation of Discrete-Time Signals
119
assum ed to be an integer m ultiple o f the sam pling interval, and w (n) represents the additive n oise that is picked up by the antenna and any n oise generated by the electron ic com p onents and amplifiers contained in the front end o f the receiver. O n the other hand, if there is no target in the space searched by the radar and sonar, the received signal y(n ) consists o f noise alone. H avin g the tw o signal sequ en ces, x ( n ) , which is called the reference signal or transm itted signal, and y ( n ) , the received signal, the problem in radar and sonar d etection is to com p are y(n) and x ( n ) to determ ine if a target is present and, if so, to determ in e the tim e d elay D and com p ute the distance to the target. In practice, the signal x ( n D) is h eavily corrupted by the additive n oise to the point w here a visual inspection o f y ( n ) d oes n ot reveal the p resen ce or absence o f the desired signal reflected from the target. Correlation provides us with a m eans for extracting this im portant inform ation from y( n). D igital com m unications is an oth er area w here correlation is o ften used. In digital com m unications the inform ation to be transm itted from on e p oin t to an other is usually con verted to binary from , that is, a seq u en ce o f zeros and ones, which are then transm itted to the in tend ed receiver. T o transm it a 0 w e can trans m it the signal seq u en ce xo(n) for 0 < n < L 1, and to transm it a 1 w e can transmit the signal seq u en ce jti(n) for 0 < n < L 1, w here L is som e integer that d en otes the num ber o f sam p les in each o f the tw o sequ en ces. V ery often , x\ (n) is selected to be th e negative o f xo(n). T h e signal received by the in tend ed receiver m ay be represented as y (n ) = x,-(n) + w (n ) * = 0 ,1 0 < n < L 1 (2.6.2) w here now the uncertainty is w hether x 0(n) or *](n ) is the signal com p on en t in >(n), and w (n ) rep resents the additive n oise and other interferen ce inherent in
120
Chap. 2
any co m m u n icatio n system . A g ain , such noise has its o rig in in th e e le ctro n ic c o m p o n e n ts c o n ta in e d in th e fro n t e n d o f th e receiv er. In an y case, th e receiv er k n o w s th e p o ssib le tra n sm itte d se q u e n c e s xo(n) a n d (n) a n d is fa c e d w ith the task o f co m p a rin g th e receiv ed signal y( n) w ith b o th xo(n) a n d Jti(n) to d e te rm in e w hich o f th e tw o signals b e tte r m a tc h e s y(n). T h is c o m p a riso n p ro c e ss is p e rfo rm e d by m ean s o f th e c o rre la tio n o p e ra tio n d e sc rib e d in th e follow ing su b se c tio n .
2.6.1 Crosscorrelation and Autocorrelation Sequences

S u p p o se th a t we h av e tw o real signal se q u e n c e s x ( n ) an d y ( n ) ea c h o f w hich has finite en erg y . T h e crosscorrelation o f x ( n ) and v(n) is a se q u e n c e rxy(l), w hich is d efin ed as r.tv(l)
n ~ x
~ 0
l = 0. 1 , 2 , . . .
(2.6.3)
or, eq u iv alen tly , as

OC
riy(t) Y , X ^n + 0 y ( )
n= -oc
1 = 0, 1 , 2 , . . .
(2.6.4)
T h e in d ex I is th e (tim e ) sh ift (o r lag) p a r a m e te r an d th e su b scrip ts x y on th e c ro ss c o rre la tio n se q u e n c e rxy(l) in d icate the se q u e n c e s b ein g c o rre la te d . T h e o rd e r of th e su b scrip ts, w ith .x p re c e d in g y, in d ic a te s th e d ire c tio n in w h ich o n e se q u en ce is sh ifted , re lativ e to th e o th e r. T o e la b o ra te , in (2.6.3), th e se q u e n c e x ( n ) is left u n sh ifte d an d y (n ) is sh ifted by / u n its in tim e, to th e rig h t fo r / p o sitiv e a n d to th e left for I n eg ativ e. E q u iv a le n tly , in (2.6.4), th e se q u en ce y ( ) is left u n sh ifted a n d x{n) is sh ifted by I u n its in tim e, to th e left fo r / p o sitiv e a n d to th e rig h t fo r / n eg ativ e. B u t sh iftin g x (n ) to th e left by / u n its relativ e to y (n ) is e q u iv a le n t to sh iftin g y() to th e rig h t by / u n its re lativ e to x ( n) . H e n c e th e c o m p u ta tio n s (2.6.3) an d (2.6.4) yield id en tical c ro ssc o rre la tio n se q u en ces. If w e rev erse th e ro les o f jr() an d y (n) in (2.6.3) an d (2.6.4) a n d th e re fo re re v e rse th e o rd e r o f th e indices xy. we o b ta in th e c ro ss c o rre la tio n se q u e n c e
OC
ryx(I) = n= 0 C o r, e q u iv alen tly ,

OC
y ( n ) x ( n I)
(2.6.5)
ryx(l) = Y
n d c
y ( n + l ) x ( n)
(2.6.6)
By c o m p a rin g (2.6.3) w ith (2.6.6) o r (2.6.4) w ith (2.6.5), we c o n c lu d e th a t rxy(l) = ryx( ~l ) (2.6.7)
T h e re fo re , r y i (l) is sim ply th e fo ld ed v ersio n o f rxy(l), w h e re th e fo ld in g is d o n e w ith re sp e c t to / = 0. H e n c e , ryx(l ) p ro v id e s exactly th e sam e in fo rm a tio n as rxv(l), w ith re sp e c t to th e sim ilarity o f x ( n ) to y(n ).
Sec. 2.6
121
Exam ple 2.6.1 D eterm ine the crosscorrelation sequence rxv(l) of the sequences x(n) = { ....0 .0 .2 . 1 .3 .7 .1 .2 . 3. 0. 0 ....}
t
v(n) = { . . . . 0 . 0 . 1 . - 1 . 2 . - 2 . 4 . 1 . - 2 . 5 .0 .0 ,...]
t
Solution
Let us use the definition in (2.6.3) to compute r rv(0) = Y x(ri)v(n)
.(/). For I = 0 w e have
The product sequence u()(n) = x ( n ) y ( r ) is u(n) = { ..., 0. 0. 2. 1. 6. -1 4 . 4. 2, 6, 0, 0. . . .)

t
and hence the sum over all values of n is r,v(0) = 7 For I > 0, we simpiy shift v() to the right relaLive to a ' ( h ) hy / units, compute the product sequence v/(n) = jr(n)_v(n I), and finally, sum o v er all va lu es o f the product sequence. Thus we obtain r t ,( l) = 13, rM5) = 5. rJV(2) = -1 8 . r ,v(6) = - 3 , r vv(3) = 16. rxy (/) = 0. r,,(4 ) = - 7 I >1
For / < 0, we shift y(n) to the left relative to jr(n) by / units, compute the product sequence v/(n) = j(n ) v(n I), and sum over all values of the product sequence. Thus we obtain the values of the crosscorrelation sequence rIV(1) = 0, rJV( -5 ) = 19, riy( - 2 ) = 33. fxv(-6) = - 9 , rly{ - 3) = -1 4 . rIV( - 7) = 10, rTV ( - 4 ) = 36 rxv(l) = 0, I < - 8
Therefore, the crosscorrelation sequence of x{n) and y(n) is r,AD = (10, - 9 ,1 9 , 36, -1 4 , 33,0, 7,13, -1 8 ,1 6 . - 7 , 5, - 3 )
t
T h e sim ilarities b e tw e e n th e c o m p u ta tio n o f th e c ro ss c o rre la tio n o f tw o se q u e n c e s a n d th e c o n v o lu tio n o f tw o se q u e n c e s is a p p a re n t. In th e c o m p u ta tio n of c o n v o lu tio n , o n e o f th e se q u e n c e s is fo ld ed , th e n sh ifted , th e n m u ltip lie d by th e o th e r se q u e n c e to fo rm th e p ro d u c t se q u e n c e fo r th a t shift, a n d finally, th e values o f th e p r o d u c t se q u e n c e a re su m m ed . E x c e p t fo r th e fo ld in g o p e ra tio n , th e co m p u ta tio n o f th e c ro ss c o rre la tio n se q u e n c e involves th e sa m e o p e ra tio n s: shifting o n e o f th e se q u e n c e s , m u ltip lic atio n o f th e tw o se q u e n c e s, an d su m m in g o v er all v alu es o f th e p r o d u c t se q u e n c e . C o n se q u e n tly , if w e h av e a c o m p u te r p ro g ra m th a t p e rfo rm s c o n v o lu tio n , w e can use it to p e rfo rm c ro ss c o rre la tio n by p ro v id in g
122
Chap. 2
as in p u ts to th e p ro g ra m , th e se q u e n c e jc() an d th e fo ld ed se q u e n c e y ( n). T h e n th e c o n v o lu tio n o f x ( n ) w ith y ( n) yields th e c ro ss c o rre la tio n r rv(/). th a t is, rxv(D = x ( l ) * y ( - l ) (2.6.8)
In th e special case w h ere y( n) = x( n ) , we h av e th e aut ocorrel ati on o f *( ), w hich is d efined as th e se q u e n c e
rX x(l)=
o r, eq u iv alen tly , as
O C
x(n)j:(n - 0
ft ~ O C
(2.6.9)
O O
rxx{l) = ^ 2 , x (n +
n= oc
(2.6.10)
In d ealin g w ith fin ite -d u ra tio n se q u e n c e s, it is cu sto m a ry to ex p ress th e a u to c o rre la tio n an d c ro ss c o rre la tio n in te rm s of th e finite lim its on th e su m m a tio n . In p a rtic u la r, if x ( ) a n d v(n) a re causal se q u e n c e s o f le n g th N [i.e., .v(n) = y(n) = 0 for n < 0 an d n > N], th e c ro ssc o rre la tio n an d a u to c o rre la tio n se q u e n c e s m ay be ex p ressed as rxy(l) = an d A'-1*1-1 rxx( l ) = ^ x(n)y(n-l) (2.6.11)
Y
n=f
(2.6.12)
w h ere i = I, k = 0 fo r / > 0, a n d / = 0, k = I fo r / < 0.
2.6.2 Properties of the Autocorrelation and Crosscorrelation Sequences

T h e a u to c o rre la tio n an d c ro ssc o rre la tio n se q u e n c e s h av e a n u m b e r of im p o rta n t p ro p e rtie s th a t w e n o w p re se n t. T o d e v e lo p th e s e p ro p e rtie s , le t us assu m e th a t we h av e tw o se q u e n c e s x ( n ) a n d y (n) w ith finite en erg y from w hich w e fo rm the lin e a r c o m b in a tio n , ax(n) bv( n I)
w h ere a an d b a re a rb itra ry c o n s ta n ts a n d I is so m e tim e shift. T h e en erg y in this signal is
O C
O C
[ax(n) + by( n - I)]2 = a 2 ^
n = oc
O C
x 2( n ) + b 2 ^
/i oc
Y
fj -oc
y 2{n - I) (2 '6 -13)
. 2ab -j l ^ V""' x (in ) \ y (tn I) n +

n = oc
= a 2rxx(0) + b2r yy( 0) + l a b r xy{l)
Sec. 2.6
123
F irst, we n o te th a t rx x (0) = E x a n d /-vy(0) = v, w hich a re th e en e rg ie s o f x{n) an d y (n ), resp ectiv ely . It is o b v io u s th a t a 2rxx(0) -I- b 2r y> .(0) + 2abrxy(l) > 0 N ow , assu m in g th a t b ^ 0, w e can divide (2.6.14) by b 2 to o b ta in r (0 ) (2.6.14)
(^ )2
+ 2rxy(l) Q
+ r v_ v(0) > 0
W e view th is e q u a tio n as a q u a d ra tic w ith coefficients r XJ(0), 2 rxv(l), a n d ^ ,( 0 ) . Since th e q u a d ra tic is n o n n e g a tiv e , it follow s th a t th e d isc rim in a n t of this q u a d ra tic m u st b e n o n p o sitiv e , th a t is, 4 [ r ; v(/) - r , , ( 0 K v(0)] < 0 T h e re fo re , th e c ro ss c o rre la tio n se q u e n c e satisfies th e c o n d itio n th a t |r,v(/>| < y r ( 0 ) r , v(0) = In th e sp ecial case w h ere v(n) = x ( n ), (2.6.15) re d u c e s to \rxx( D \ < r xx(0) = E x (2.6.16) (2.6.15)
T his m ean s th a t th e a u to c o rre la tio n se q u e n c e o f a signal a tta in s its m ax im u m value at z e ro lag. T h is re su lt is c o n siste n t w ith th e n o tio n th a t a sig n a l m a tc h e s p erfe ctly w ith itself at z e ro shift. In th e case o f th e c ro ssc o rre la tio n se q u en ce, th e u p p e r b o u n d o n its v alu es is given in (2.6.15). N o te th a t if an y o n e o r b o th o f th e signals involved in th e c ro ssc o rre la tio n a re scaled , th e sh a p e o f th e c ro ss c o rre la tio n se q u e n c e d o e s n o t ch an g e, only the a m p litu d e s o f th e c ro ss c o rre la tio n se q u e n c e a re sc aled accordingly. S ince scaling is u n im p o rta n t, it is o ften d e s ira b le , in p ra c tic e , to n o rm a liz e th e a u to c o rre la tio n a n d c ro ss c o rre la tio n se q u e n c e s to th e ran g e fro m - 1 to 1. In th e case o f the a u to c o rre la tio n se q u e n c e , w e can sim ply d iv id e by ^ ( 0 ) . T h u s th e n o rm aliz ed a u to c o rre la tio n se q u e n c e is d efin ed as P. A D = (2-6 -17)
rixiO)
Sim ilarly, we d efin e th e n o rm a liz e d c ro ssc o rre la tio n se q u e n c e pXY(l) = r ' v(l) : v/ r xx(0 )rvv(0) (2.6.18)
N ow \pXI{l)\ < 1 a n d |/oX v(0! < 1, a n d h e n c e th ese se q u e n c e s a re in d e p e n d e n t of signal scaling. F in ally , as we h av e a lre a d y d e m o n s tra te d , th e c ro ss c o rre la tio n se q u e n c e sa t isfies th e p ro p e rty
r Xy ( l ) = f y x ( - 0
124
Chap. 2
W ith y(n) = x ( n) , th is re la tio n resu lts in th e follow ing im p o rta n t p ro p e rty fo r the a u to c o rre la tio n se q u en ce
r { l ) = rx s ( . - l )
(2.6.19)
H e n c e th e a u to c o rre la tio n fu n ctio n is an ev en fu n ctio n . C o n s e q u e n tly , it suffices to co m p u te rx x (l) fo r / > 0. Example 2.6.2 Compute the autocorrelation of the signal x(n) = a"u(n), 0 < a < 1 Solution Since x (n) is an infinite-duration signal, its autocorrelation also has infinite duration. We distinguish two cases. If I > 0. from Fig. 2.39 we observe that
r t J (/) =
n= l
x ( n ) x ( n - /) =
n =i
' = a
1
n~l
Since a < 1, the infinite series cot i crges and we obtain r tl (/) ~ For / < 0 we have =
n=(l
^ -V 1 a-
I> 0
x(n)x(/i /) = a -1'S~'(rr)" = ------- -ti~! I - an=(l a ' 1'.
/ < 0
r i t (!)
But when / is negative, c r 1 into the following expression: rx,(!) =
Thus the two relations for
can be combined
,,a ,t: oc < / < oc 1 a~ The sequence rxx(l) is shown in Fig. 2.42(d). We observe that r ( ~ / | = rxAD and
(2.6.20)
rlt(0) =
1 a2
Therefore, the normalized autocorrelation sequence is r (/) p s,(!) - cr|,: ~ oc < I < oc rxxW)
(2.6.21)
2.6.3 Correlation of Periodic Sequences

In S ectio n 2.6.1 w e d efin ed th e c ro ss c o rre la tio n an d a u to c o r re la tio n se q u en ces of en erg y signals. In this se ctio n we co n sid e r th e c o rre la tio n se q u e n c e s o f p o w er signals an d , in p a rtic u la r, p e rio d ic signals. L et x( n ) a n d y(rc) b e tw o p o w e r signals. T h e ir c ro ss c o rre la tio n se q u e n c e is d efin ed as 1 M
rx \ 0 )
M -oc 2 M + 1
lim
----- - Y ]
x(n)y(n~l)
( 2 .6 .2 2 )
Sec. 2.6

xi n )
125
) i'
-2-10123
(a)
in .
/ >o
xin-I)
o
(b)
x(n - I )
l< 0
(c)
r(!) =
' , a1 '1 1 - a2
- 2 - 1 0
Figure 139
1 2 (d)
/
=
Compulation of the autocorrelation o f the signal xin)
a",
0 < a < 1. If x ( n ) = y(tt), we have the definition o f the autocorrelation seq u en ce of a p ow er signal as 1 M rxx(I) = iim . . . , 1 Y ] x ( n ) x ( n - l ) (2.6.23) Af-oo 2M + 1 In particular, if x ( n ) and y ( n ) are two periodic seq u en ces, each with period the averages indicated in (2.6.22) and (2.6.23) o ver the infinite interval, are identical
126
D iscrete-Time Signals and Systems
Chap. 2
to th e av erag es o v er a single p e rio d , so th a t (2.6.22) an d (2.6.23) re d u c e to (2.6.24) an d (2.6.25) It is c lear th a t r ry(l) an d rxx(l) are p e rio d ic c o rre la tio n T h e fa c to r 1 / N can b e v iew ed as a n o rm a liz a tio n scale In som e p ractical ap p licatio n s, c o rre la tio n is u se d an o b se rv e d physical signal w hich m ay be c o rru p te d by ex am p le, c o n sid er a signal se q u e n c e y( n) o f th e form y{n) = * (n ) + w(n) se q u e n c e s w ith p e rio d N . facto r. to id en tify p erio d icitie s in ra n d o m in te rfe re n c e . F o r (2.6.26)
w h ere jc(/i) is a p erio d ic se q u e n c e o f so m e u n k n o w n p e rio d N a n d w( n) re p re se n ts an ad d itiv e ran d o m in te rfe re n c e . S u p p o se th a t w e o b se rv e M sa m p le s o f y(n ), say 0 < n < M 1, w h ere M > > N. F o r all p ractical p u rp o se s, w e can assum e th a t y( n) 0 fo r n < 0 an d n > M. N ow th e a u to c o rre la tio n se q u e n c e of y(n), using th e n o rm aliz atio n facto r o f \ / M . is (2.6.27) If we su b stitu te for y(n) fro m (2.6.26) in to (2.6.27) w e o b ta in _ _1 ry.T(/) = j M- 1 Y ] x ( n ) x i n - /)
M
j M J
H - y ^ [ .r ( n ) itj (/2 /) + w{ n) x{n /)] M
(2.6.28)
T h e first fa c to r on th e rig h t-h a n d sid e of (2.6.28) is th e a u to c o rre la tio n se q u e n c e o f x i n) . S ince x( n) is p e rio d ic , its a u to c o rre la tio n s e q u e n c e ex h ib its th e sam e p erio d icity , th u s co n ta in in g relativ ely larg e p e a k s at / = 0, N , 2 N , a n d so on. H o w ev er, as th e shift I a p p ro a c h e s M , th e p e a k s a re re d u c e d in a m p litu d e d u e to th e fact th a t w e h av e a finite d a ta re c o rd o f M sa m p le s so th a t m any o f th e p ro d u c ts ;t(n)j:(rt /) a re zero . C o n s e q u e n tly , w e sh o u ld av o id c o m p u tin g r V T(/) fo r larg e lags, say, I > M f l .
Sec. 2.6
127
T h e c ro ss c o rre la tio n s rxu,(/) an d rwx(l ) b etw e e n th e signal a n d th e a d ditiv e ra n d o m in te rfe re n c e a re e x p ected to be relativ ely sm all as a resu lt of the e x p e c ta tio n th a t ;r(/i) a n d w ( n ) will be to ta lly u n re la te d . F in ally , the last term on th e rig h t-h a n d sid e of (2.6.28) is th e a u to c o rre la tio n se q u e n c e o f th e ra n d o m se q u e n c e w( n) . T h is c o rre la tio n se q u en ce will c e rta in ly c o n tain a p e a k a t / = 0, but b ecau se o f its r a n d o m ch aracteristics, r ww(l) is ex p e c te d to d ecay rap id ly to w ard zero . C o n s e q u e n tly , o nly rxx{l) is e x p e c te d to h av e large p e a k s fo r / > 0. T his b e h a v io r allow s us to d e te c t th e p re se n c e of th e p e rio d ic signal a (/ j ) b u rie d in the in te rfe re n c e u>(n) a n d to id e n tify its p e rio d . A n e x am p le th a t illu stra te s th e use of a u to c o rre la tio n to identify a h id d e n p e rio d ic ity in an o b se rv e d physical signal is show n in Fig. 2.40. T h is figure illus tra te s th e a u to c o rre la tio n (n o rm a liz e d ) se q u e n c e fo r th e W o lfe r su n sp o t n u m b e rs fo r 0 < / < 20, w h ere an y v alu e o f / c o rre sp o n d s to o n e y ear. T h e se n u m b e rs are given in T a b le 2.2 fo r th e 100-year p e rio d 1770-1869. T h e re is c lear ev id en ce in th is fig u re th a t a p e rio d ic tre n d exists, w ith a p e rio d o f 10 to 11 years. Example 2.6.3 Suppose that a signal sequence jr(n) = sin(7r/5)/i, for 0 < n < 99 is corrupted by an additive noise sequence Kn), where the values of the additive noise are selected independently from sample to sample, from a uniform distribution over the range
TABLE 2.2 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794
YEARLY WOLFER SUNSPOT NUMBERS

101 82 66 35 31 7 20 92 154 125 85 68 38 23 10 24 83 132 131 118 90 67 60 47 41 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 21 16 6 4 7 14 34 45 43 48 42 28 10 8 2 0 1 5 12 14 35 46 41 30 24 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 16 7 4 2 8 17 36 50 62 67 71 48 28 8 13 57 122 138 103 86 63 37 24 11 15 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 40 62 9X 124 96 66 64 54 39 21
7
4 23 55 94 96 77 59 44 47 30 16 7 37 74
128
Chap. 2
Year
(a)
Figure 2.40 Identification of periodicity in the Wolfer sunspot numbers: (a) an nual Wolfer sunspot numbers; (b) autocorrelation sequence.
(A /2, A /2), where A is a param eter of the distribution. The observed sequence is y(n) = x(n) + w(n). D eterm ine the autocorrelation sequence rvt(n) and thus determine the period of the signal x(rt). Solution The assumption is that the signal sequence x(n) has some unknown period that we are attem pting to determ ine from the noise-corrupted observations {y(n)). Although x(n) is periodic with period 10, we have only a ftmte-duration sequence of
Sec. 2.6
129
length M = 100 [i.e.. 10 periods of jcm )]. The noise power level P in the sequence u'(n) is determ ined by the param eter A. We simply state that Pu = A: /12. The signal power level is P, = I. Therefore, the signal-to-noise ratio (SNR) is defined as P,
~ P ~ U ~ A : /1 2 A :
Usually, the SNR is expressed on a logarithmic scale in decibels (dB) as 101ogui ( PJPv) . Figure 2.41 illustrates a sample of a noise sequence w(n), and the observed sequence y(n) x(n) + it(n) when the SNR = 1 dB. The autocorrelation sequence
J ' t I l l Lt ttIiIt . Tllrii! ill ..tI 1. I i i h i 1 [Ijj-liii p p T T J ljw ^ IT 1 i( i| 4 * 4 i

j * 4 1 1
'
la]
(c)
Figure 2.41
Use of autocorrelation to detect the presence of a periodic signal corrupted by
noise.
130
Chap. 2
w(n)
L TT T T T . T tT J IT .T T U T tT T ._L T .T. I .T .
pp V |V v li
'
t t T .1 .
(a)
(b)
rM .
1 ?
r![It x ij|il
i o1
rl It
r!lfit
F
(c)
1 t T T * ? 1 [1
Figure 2.42 Use of autocorrelation to detect the presence of a periodic signal corrupted by noise.
r vv(/) is illustrated in Fig. 2.41c. W e observe that the periodic signal *(/?), embedded in y(n), results in a periodic autocorrelation function rZI(l) with period N 10. The effect of the additive noise is to add to the peak value at / = 0. but for I ^ 0, the correlation sequence rwu,(l) = ^ 0 a s a result of the fact that values of w(ri) were gen erated independently. Such noise is usually called white noise. The presence of this noise explains the reason for the large peak at I = 0. The smaller, nearly equal peaks at I = 10, 2 0 ,... are due the periodic characteristics of x(n). Figure 2.42 illustrates the noise sequence w(n), the noise-corrupted signal y(n), and the autocorrelation sequence r vv(/) for the same signal, within which is embedded a signal at a smaller noise level. In this case, the SNR = 5 dB. E ven with this relatively small noise level, the periodicity of the signal is not easily determ ined from observa tion of y( n ). However, it is clearly evident from observation of the autocorrelation sequence ryy(n).
2.6.4 Computation of Correlation Sequences

A s indicated on Section 2.6.1, the procedure for com puting the crosscorrelation seq u en ce b etw een x ( n ) and y ( n ) in volves shifting on e o f the seq u en ces, say x(n),
Sec. 2.6
131
to o b ta in x( n - /), m u ltip ly in g th e sh ifted se q u e n c e by v(n) to o b ta in th e p r o d u ct se q u e n c e y ( n ) x ( n - /), an d th en su m m in g all th e v alu es o f th e p ro d u c t se q u en ce to o b ta in r yx(l). T his p ro c e d u re is r e p e a te d fo r d iffe re n t values o f th e lag /. E x c e p t fo r th e fo ld in g o p e ra tio n th a t is involved in c o n v o lu tio n , th ese b a sic o p e ra tio n s fo r co m p u tin g th e c o rre la tio n se q u e n c e a re id e n tic a l to th o se in co n v o lu tio n . T h e p ro c e d u re fo r co m p u tin g th e co n v o lu tio n is directly a p p licab le to c o m p u tin g th e c o rre la tio n o f tw o se q u en ces. S pecifically, if w e fo ld y (n) to o b ta in y ( n), th e n th e c o n v o lu tio n o f x ( n ) w ith v ( n) is id en tical to th e cro ssc o rre la tio n o f x ( n ) w ith y(n). T h a t is. rxy(l) = x(rt) * y ( n ) |n=/ (2.6.29)
A s a c o n se q u e n c e , th e c o m p u ta tio n a l p ro c e d u re d escrib ed for c o n v o lu tio n can be a p p lied d irectly to th e c o m p u ta tio n o f th e c o rre la tio n se q u en ce. W e no w d e sc rib e an a lg o rith m th a t can b e easily p ro g ra m m e d to c o m p u te th e c ro ss c o rre la tio n se q u e n c e of tw o fin ite -d u ra tio n signals * ( ), 0 < n < N ~ I, an d y ( n) , 0 < n < M 1. T h e alg o rith m c o m p u te s rJV(/) fo r p o sitiv e lags. A cc o rd in g to th e re la tio n rxy(l) = r Y X(I), th e v alu es o f rXY(l) fo r n eg ativ e lags can be o b ta in e d by using th e sam e a lg o rith m fo r p o sitiv e lags, an d in te rc h a n g in g th e roles o f jt( ) an d v(n). W e o b se rv e th a t if M < N, rXY{i) can be c o m p u te d by th e re la tio n s
M - \ + l
Y ' , x ( n ) y ( n /),
0 < 1 < N M (2.6.30)
O n th e o th e r h a n d , if M > N , th e fo rm u la fo r th e c ro ss c o rre la tio n becom es rxy(l) = Y x ( n ) v ( n - 1) 0 < 1< N -1 (2.6.31)
T h e fo rm u la s in (2.6.30) a n d (2.6.31) c a n b e c o m b in ed a n d c o m p u te d by m ean s o f th e fo llo w in g sim p le alg o rith m illu stra te d in th e flow chart in Fig. 2.43. By in terch an g in g th e ro les o f j ( ) a n d y(n) an d re c o m p u tin g th e cro ssc o rre la tio n se q u e n c e , w e o b ta in th e v alu es o f rXY(l) c o rre sp o n d in g to n eg ativ e shifts I. If w e w ish to c o m p u te th e a u to c o rre la tio n se q u e n c e rxx(l), w e se t y (n ) = jr(n) an d M = N in (2.6.31). T h e co m p u ta tio n o f rxx{l) can be d o n e by m e a n s of the sa m e a lg o rith m fo r p o sitiv e shifts only.
2.6.5 Input-Output Correlation Sequences

In th is se ctio n w e d e riv e tw o i n p u t- o u tp u t re la tio n sh ip s fo r L T I system s in th e c o rre la tio n d o m a in . L e t us assu m e th a t a signal x ( n ) w ith k n o w n a u to c o r re la tio n rxx(I) is a p p lied to an L T I system w ith im p u lse re sp o n s e h( n), p ro d u c in g th e
132
Chap. 2
C j lD
Figure 2A3 Flowchart for software implementation of crosscorrelation.
Sec. 2.6
133
o u tp u t signal
v(n) = h( n) * .x (/?) = ^ h(k)x(>i k)
k= oc T h e c ro ss c o rre la tio n b e tw e e n th e o u tp u t a n d th e in p u t signal is rV J-(7) = y d ) * x ( - l ) h ( l ) * [*(/) * * ( - / ) ] or ryx(l) = /;(/) * rxx(l) (2.6.32)
w h ere w e h av e u se d (2.6.8) a n d th e p ro p e rtie s o f co n v o lu tio n . H e n ce th e c ro ssco r re la tio n b e tw e e n th e in p u t an d the o u tp u t of th e sy stem is th e co n v o lu tio n of the im p u lse re sp o n s e w ith the a u to c o rre la tio n of th e in p u t se q u e n c e . A lte rn a tiv e ly . ryx(l) m av b e view ed as th e o u tp u t of th e L T I sy stem w hen th e in p u t se q u en ce is rxx (/). T h is is illu stra te d in Fig. 2.44. If we rep lace / by - / in (2.6.32), w e o b tain r.tyU) = h ( - D * rxx(l) T h e a u to c o rre la tio n of th e o u tp u t signal can be o b ta in e d by using (2.6.8) w ith x(/i) = y(/i) an d th e p ro p e rtie s o f co n v o lu tio n . T h u s we have r lv(/) = v(/) * v (/) = [/;(/) * * (/)] * [/j ( / J * * ( /)] (2.6.33) = [/?(/) * h ( /)] * [*(/) * * ( /)] = rhh(l) * rxx(l) T h e a u to c o rre la tio n rhh(l) of th e im pulse resp o n se h(n) exists if th e system is stable. F u rth e rm o re , th e stab ility in su res th a t th e system d o e s n ot ch an g e th e ty p e (en erg y o r p o w e r) o f th e in p u t signal. By ev alu atin g (2.6.33) fo r / = 0 we o b ta in
CC
(2.6.34) w hich p ro v id e s th e en e rg y (o r p o w er) of th e o u tp u t signal in te rm s o f a u to c o r re latio n s. T h e se re la tio n sh ip s h o ld fo r b o th en erg y a n d p o w e r signals. T h e direct d e riv a tio n o f th e s e re la tio n sh ip s for en erg y an d p o w e r signals, a n d th e ir e x ten sio n s to co m p lex sig n als, are left as exercises fo r the stu d e n t.
Input
rxx<n )
LTI SYSTEM
Output rvr(n) Figure 2.44 Input-output relation for crosscorrelation ryx(n).
ft(n)
134
Chap. 2

T h e m a jo r th e m e o f th is c h a p te r is th e c h a ra c te riz a tio n o f d isc re te -tim e signals an d sy stem s in th e tim e d o m ain . O f p a rtic u la r im p o rta n c e is th e class o f lin e a r tim ein v a ria n t (L T I) sy stem s w hich a re w idely u se d in th e design a n d im p le m e n ta tio n o f d ig ital signal p ro cessin g system s. W e c h a ra c te riz e d L T I sy ste m s by th e ir u n it sa m p le re sp o n se h{n) an d d e riv e d th e co n v o lu tio n su m m a tio n , w hich is a fo rm u la fo r d e te rm in in g th e re sp o n se y( n) o f th e system c h a ra c te riz e d by h( n) to an y given in p u t se q u e n c e x(n). T h e class o f L T I system s c h a ra c te riz e d by lin e a r d ifferen ce e q u a tio n s w ith c o n s ta n t coefficients is by fa r th e m o st im p o rta n t of th e L T I sy stem s in th e th e o ry an d a p p licatio n o f d ig ital sig n a l p ro cessin g . T h e g e n e ra l s o lu tio n of a lin e a r dif fe re n c e e q u a tio n w ith c o n s ta n t coefficients w as d eriv e d in th is c h a p te r an d show n to co n sist of tw o c o m p o n en ts: th e so lu tio n o f th e h o m o g e n e o u s e q u a tio n w hich r e p re s e n ts th e n a tu ra l re sp o n se o f th e sy stem w h en th e in p u t is z e ro , an d th e p a r tic u la r so lu tio n , w hich re p re s e n ts th e re sp o n s e o f th e system to th e in p u t signal. F ro m th e d ifferen ce e q u a tio n , w e also d e m o n s tra te d h o w to d e riv e th e u n it sam ple re sp o n s e of th e L T I system . L in e a r tim e -in v a ria n t sy stem s w ere g en erally su b d iv id ed in to F IR (finited u ra tio n im p u lse re sp o n se ) a n d I I R (in fin ite -d u ra tio n im pulse re sp o n s e ) d e p e n d ing o n w h e th e r h(n) h as finite d u ra tio n o r infinite d u ra tio n , resp ec tiv ely . T he re a liz a tio n s o f such system s w ere briefly d escrib ed . F u rth e rm o re , in the re a liz a tio n o f F IR system s, we m a d e th e d istin ctio n b e tw e e n recu rsiv e a n d n o n recu rsiv e re a lizatio n s. O n th e o th e r h a n d , w e o b se rv e d th a t I I R system s can be im p le m e n te d recu rsiv ely , only. T h e re are a n u m b e r o f te x ts o n d isc re te -tim e signals a n d system s. W e m e n tio n as ex am p les th e b o o k s by M c G illem a n d C o o p e r (1984), O p p e n h e im a n d W illsky (1983), an d S ie b e rt (1986). L in e a r c o n sta n t-c o e ffic ie n t d iffe re n c e e q u a tio n s are tr e a te d in d e p th in th e b o o k s by H ild e b ra n d (1952) a n d L evy a n d L essm an (1961). T h e last to p ic in this c h a p te r, o n c o rre la tio n o f d isc re te -tim e signals, plays an im p o rta n t ro le in d ig ital signal p ro cessin g , esp ecially in a p p lic a tio n s d ealin g w ith digital c o m m u n ic a tio n s, ra d a r d e te c tio n a n d e stim a tio n , so n a r, a n d geophysics. In o u r tr e a tm e n t o f c o rre la tio n se q u e n c e s, w e a v o id e d th e use o f sta tis tic a l concepts. C o rre la tio n is sim ply d efin ed as a m a th e m a tic a l o p e r a tio n b e tw e e n tw o se q u en ces, w hich p ro d u c e s a n o th e r se q u e n c e , called e ith e r th e crosscorrelat ion s e quence w hen th e tw o se q u e n c e s a re d iffe re n t, o r th e aut ocorrel ati on sequence w h e n th e tw o se q u e n c e s are id en tical. In p ractical a p p lic a tio n s in w hich c o rre la tio n is u sed , o n e ( o r b o th ) o f th e se q u e n c e s is (a re ) c o n ta m in a te d by n o ise a n d , p e rh a p s , by o th e r fo rm s o f in te rfe r en ce. In such a case, th e noisy se q u e n c e is c alled a r a n d o m se q u e n c e a n d is c h a r a c te riz e d in sta tistical term s. T h e c o rre sp o n d in g c o rre la tio n se q u e n c e b e c o m e s a fu n ctio n o f th e sta tistical c h a ra c te ristic s of th e n o ise an d an y o th e r in te rfe re n c e . T h e statistical c h a ra c te riz a tio n o f se q u e n c e s a n d th e ir c o rre la tio n is tr e a te d in A p p e n d ix A . S u p p le m e n ta ry re a d in g o n p ro b a b ilis tic an d sta tistical c o n c e p ts deal-
Chap. 2
Problems
135
ing w ith c o rre la tio n can be fo u n d in th e b o o k s by D a v e n p o rt (1970). H e lstro m (1990). P ap o u lis (1984). an d P eeb les (1987).
PROBLEMS
2.1 A discrete-time signal x(n) is defined as
l + j, 1. 0,
3< n < 1 0 < n < 3 elsewhere
(a) Determ ine its values and sketch the signal .v(n). (b) Sketch the signals that result if we:
(1) First fold x{n) and then delay the resulting signal by four samples. (2) First delay xin) by four samples and then fold the resulting signal (c) Sketch the signal x ( n + 4 ). (d) Com pare the results in parts (b) and (c) and derive a rule for obtaining the signal ,v( n -t- k) from (e) Can you express the signal ,r(n) in term s of signals S(n) and u{n)l 2.2 A discrete-time signal ,v(n) is shown in Fig. P2.2. Sketch and label carefully each of the following signals.
.v(n)
J_
J_
L L _ ________ _
- 2 - 1 0 1 2 3 4
(a ) x(n - 2)
FigUre P2.2
(b) x(4 n) ( c)x(n + 2) (d) x(n)u(2 n) (e) x(n - 1 )8{n - 3) (F) x( n2) (g) even part of x{n) (h) odd part of x ( n ) 2 3 Show that (a ) &(n) = u(n) u(n 1) (b) u(n) =
8(k) =
2.4 Show that any signal can be decomposed into an even and an odd component. Is the decomposition unique? Illustrate your arguments using the stgnal x(n) = {2. 3, 4. 5. 6)
t
2.5 Show that the energy (power) of a real-valued energy (power) signal is equal to the sum of the energies (powers) of its even and odd components. 2.6 Consider the system
vf) = T[x( n) ] = x ( n 2)
(a) Determ ine if the system is time invariant.
136
Chap. 2
(b) To clarify the result in part (a) assume that the signal _fl, \ 0, O 5 /1 < 3 elsewhere
is applied into the system. (1) Sketch the signal *(). (2) D eterm ine and sketch the signal y ( n ) = T[x(n)}, (3) Sketch the signal y'2(n) = y(n 2). (4) D eterm ine and sketch the signal x2(n) = x(n - 2). (5) Determ ine and sketch the signal = T \ x 2(n)]. (6) Com pare the signals ^ ( n ) and y(n - 2). W hat is your conclusion? (c) R epeat part (b) for the system
y(rt) = x(n) - x(n 1)
Can you use this result to make any statem ent about the time invariance of this system? Why? (d) Repeat parts (b) and (c) for the system y(rt) = T [ x ( n )] = nx(n) 2.7 A discrete-time system can be (1) Static or dynamic (2) Linear or nonlinear (3) Time invariant or time varying (4) Causal or noncausal (5) Stable o r unstable Examine the following systems with respect to the properties above. (a) v() = cos[jc(n)] (b) (c) (d) (e) v(n) = x(k> v(n) = x(n)cos(^n) y(n) ~ x ( n + 2) y(n) = Trun[;c(n)], where Trun[jc(n)] denotes the integer part of x(n), obtained by truncation (f) y(n) = Round[jc(n)], where Round[;c(n)] denotes the integer part of Jt(n) obtained by rounding Remark: The systems in parts (e) and (f) are quantizers that perform truncation and rounding, respectively. (g) y(B) = |*(n)| (h) v(rt) = x(n)u(n) (I) y(n) = x(n) + nx{n + 1) (j) y ( n ) = x ( 2 n )
(I) y( n ) = x ( - n ) (m) y(n) = sign[j:(n)] (n) The ideal sampling system with input xaU) and output x(n) = x a(nT), oc < n < oo 2J8 Two discrete-time systems 7] and T2 are connected in cascade to form a new system T as shown in Fig. P2.8. Prove or disprove the following statements.
Chap. 2
Problems
137
xin )
yin)
T i
T - T-. T 2
T; Figure P2.8
(a) If T\ and % are linear, then T is linear (i.e.. the cascade connection of two linear systems is linear). (b) If T\ and are time invariant, then T is time invariant. (c) If T[ and 7? are causal, then T is causal. <d) If T] and T2 are linear and time invariant, the same holds for T . (e) If 7] and T2 are linear and time invariant, then interchanging their order does not change the system T. (0 As in part (e) except that 7J, T2 are now time varying. (Hint: Use an example.) (g) If 7] and T2 are nonlinear, then T is nonlinear. (h) If T< and T2 are stable, then T is stable. (i) Show by an example that the inverse of parts (c) and (h) do not hold in general. 2.9 Let T be an LTI, relaxed, and BIBO stable system with input x{n) and output y(n). Show that: (a) If x(n) is periodic with period N [i.e., jr(n) = x{n + N) for all n > 0], the output y(n) tends to a periodic signal with the same period. (b) If x(n) is bounded and tends 10 a constant, the output will also tend to a constant. (c) If x{n) is an energy signal, the output y(n) will also be an energy signal. 2.10 The following in p u t-output pairs have been observed during the operation of a umeinvariam system: x,(n) = {1.0,2} ^
t
y,(fi) = (0, 1.2}

t
x,(n) = {0.0,3} ^
t
v; (n) = (0, 1.0,2}

t
x-\(n) = {0. 0, 0. 1}
t
vj(n) = (1,2, 1}
t
Can you draw any conclusions regarding the linearity of the system. W hat is the impulse response of the system? 2.11 The following input-output pairs have been observed during the operation of a linear system: Xi(n) = {-1. 2. 1}
t
y,(/i) = (1, 2. - 1 , 0. 1}
t
x2{n) = {1, - 1 , -1}

t x 3(n) = {0, 1, 1) t
\'2in) = {-1. 1, 0, 2}
t y i ( n) = {1, 2. 1} r
Can you draw any conclusions about the time invariance of this system? 2.12 The only available information about a system consists of N input-output pairs, of signals y,(rc) = T[xj(n )], / = 1, 2........N.
138
Chap. 2
(a) What is the class of input signals for which we can determ ine the output, using
the information above, if the system is known to be linear? (b) The sam e as above, if the system is known to be tim e invariant. 2.13 Show that the necessary and sufficient condition for a relaxed LTI system to be BIBO stable is y . for some constant Mn. 2.14 Show that; (a) A relaxed linear system is causal if and only if for any input x(n) such that jc(n) = 0 for n < no => y(n) = 0 (b) A relaxed LTI system is causal if and only if h(n) = 0 for n < 0 for n < no i/i()i < Mh < oo
2.15 (a) Show that for any real or complex constant a, and any finite integer num bers M and N, we have n a M a h . if * !
1 < 3
N - M + 1, (b) Show that if |o| < 1 , then
if a = 1
y v
= - i -
l - a
2.16 (a) If y(n) = x (n) * h(n). show that v =
/ .- where ^ _w (b) Compute the convolution y(n) = x(n) * h(n) of the following signals and check the correctness of the results by using the test in (a). (1) jr(n) (1,2, 4), /)(n) = (1 ,1 ,1 ,1 ,1 } (2) x(n) = {1, 2, - 1 ) , h(n) = x (n) (3) x(n) = (0,1, - 2 , 3. - 4 ) . h(n) = {, i , 1, 1} (4) jc(n )= :{1 .2 .3.4.5J.A (n) = {l) (5) x(n) = (1, -2 ,3 } , h(n ) = (0, 0 .1 .1 ,1 ,1 ) t t (6) x(n) = { 0 ,0 ,1 ,1 ,1 ,1 ), h(n) = { 1 ,-2 . 3} t t (7) jr(u) = {0,1, 4, -31. h(n) = [1,0, - 1 , -1} t t (8) = [1,1,2], h(n) = u(n) t
(9) jt(n) = [1,1. 0,1,11, h(n) = {1, - 2 , - 3 , 4} t t (10) jc(n) = (1,2, 0,2 , l}/i(n) = Jt(n) t (11) *{n) = (i)"u(n), h(rt) = ( j ) nM(n)
2.17 Compute and plot the convolutions x(n) * h(n) and h(n) *x(n) for the pairs of signals
shown in Fig. P2.17.
Chap. 2
Problems
139
b TI f
0 12 3 j x(n) n
trln)
-
0 l 2 3 4 5 b
h\n)
Iiit.
1 2 3 x(n) i
-3-2-10 I 2 3 hin)
..III!..
3 4 5 6 J (n )
II
htnl
111
!
<di
2 3 4 5
2-1 Figure P2.17
II
2 . 1 8
Determ ine and sketch the convolution y(n) of the signals
-V(/l) =
h(n) -
0. 1, 0.
0 < Ji < 6 elsewhere 2 < n < 2 elsewhere
(a) Graphicallv
(b) Analytically
2.19 Compute the convolution y(n) of the signals

x (n) =
o'".
3 < n < 5
0. 1, 0,
elsewhere 0< n< 4 elsewhere
h(n) =
2.20 Consider the following three operations. (a) Multiply the integer numbers: 131 and 122. (b) Compute the convolution of signals: {1. 3.1) * (1,2. 2}. (c) Multiply the polynomials: 1 4- 3; + z2 and 1 4- 2z 4- 2z2. (d> Repeat part (a) for the numbers 1.31 and 12.2. (e) Comment on your results. 2.21 Com pute the convolution y(n) = x ( n ) * h(n) of the following pairs of signals. (a) x(n) = a"u(n ), h(n) = b"u{n) when a ^ b and when a ~ b
(b) x ( n) = 1. 2,
n = 2, 0, 1 n = 1
. 0,
elsewhere
h ( n ) = S(n) S (n 1) + S(n 4) + S(n 5)
140
Chap. 2
(c) x{n ) = u(n + 1) u(n 4) <5(n 5) h(n) = [u(n +2) u(n - 3)] (3 - |n|) (d ) x ( n ) = u ( n ) - u( n - 5) h(n) = u(n 2) u(n 8) 4- u(n 11) u(n 17) 2.22 Let x(n) be the input signal to a discrete-time filter with impulse response ht(n) and let y,(n) be the corresponding output. (a) Compute and sketch x(n) and \ j ( n ) in the following cases, using the same scale in all figures. x(n) = {1,4, 2. 3, 5, 3, 3. 4. 5. 7. 6. 9} h](n) = (1,1) h2(n) = {1,2.1} ]ij(n) = {i, j) *4() = {? { j)
(b) (c) (d)

(e)
Sketch x(n), yi(n), y 2(n) on one graph and *(). y3(n), y,j(n), y.s(n) on another graph W hat is the difference between yi() and \ 2(n). and between y^(n) and y^(n)? Comment on the smoothness of v2(/?) and v4(n). Which factors affect the sm ooth ness? Compare y4(n) with ysfn). What is the difference? Can you explain it? Let h(,(n) = {^, - j } . Com pute y6<n). Sketch v(n), y 2(n), and yft(n) on the same figure and comment on the results. v (n) = ny(n 1) + jr(n) n > 0
2.23 The discrete-time system
is at rest [i.e., v( 1) = 0]. Check if the system is linear time invariant and BIBO stable. 2.24 Consider the signal y(n) = a"u(n), 0 < a < 1. (a) Show that any sequence x{n) can be decomposed as
and express ck in terms of x(n). (b) Use the properties of linearity and time invariance to express the output y(n) = T[x(n)] in term s of the input x (n) and the signal g(n) = T[y(n)], where T [ ] is an LTI system. (c) Express the impulse response h(n) = T[B{n)} in terms of g(rt). 2.25 D eterm ine the zero-input response of the system described by the second-order dif ference equation x(n) - 3y(n - 1) - 4y(n - 2) = 0 2.26 Determ ine the particular solution of the difference equation y(n) = jv(ii - 1) - y(n - 2) +x ( n) when the forcing function is x(n) = 2"u(n).
Chap. 2
Problems
141
2 2 1 D eterm ine the response yin). n > 0. of the system described by the second-order difference equation yin) - 3v(n 1) 4y(/i 2) = xin) + 2x(n 1) to the input xin) = 4"w(n). 2.28 Determ ine the impulse response of the following causal system: y(n) 3y(n 1) 4v(n 2) = jr(n) + 2x(n 1) 2.29 Let xin). A'i < n < N2 and h(n), < n < M2 be two finite-duration signals. (a) Determ ine the range L\ < n < L 2 of their convolution, in term s of N\, N2, M\ and M2. (b) Determ ine the limits of the cases of partial overlap from the left, full overlap, and partial overlap from the right. For convenience, assume that h(n) has shorter duration than jc(). (c) Illustrate the validity of your results by computing the convolution of the signals -2 < n < 4 elsewhere -1 < n < 2 elsewhere 2.30 Determ ine the impulse response and the unit step response of the systems described by the difference equation (a) yin) = ().6y(;i - 1) - ().08v(n - 2) + xin) (b) _v( ) = 0.7y(;; - 1) - 0.1 yin - 2) -f 2xin) - xin - 2) 231 Consider a svstem with impulse response A,) = H r ' ( 0. - n - 4 elsewhere
Determ ine the input xin) for 0 < n < S that will generate the output sequence v(n) = 1 1 .2 .2 .5 .3 .3 .3 .2 .1 .0 ....}
t
232 Consider the interconnection of LTI systems as shown in Fig. P2.32. (a) Express the overall impulse response in terms of h \ (n), h2(n), hîn). and hîn). (b) D eterm ine h{n) when M ) = {j. 3 . 7 } h2{n) hy(n) = (n + 1 )u(n) fi4(n) = S(n 2)
Figure P 2J2
142
Discrete-Time Signals and Systems (c) Determ ine the response of the system in part (b) if
x(n) = &(n + 2) + 3S(n - 1) - 4 S(n - 3)
Chap. 2
2 3 3 Consider the system in Fig. P2.33 with h(n) = a"u(n ), 1 < a < 1. Determ ine the response y(n) of the system to the excitation *(n) = (n + 5) u(n 10)
x(n)
Figure P233
2 3 4 Com pute and sketch the step response of the system
U_ 1
2 3 5 Determ ine the range of values of the param eter a for which the linear time-invariant
system with impulse response
(Hint: The solution can be obtained easily and quickly by applying the linearity and tim e-invariance properties to the result in Exam ple 2.3.5.)
2 3 7 D eterm ine the response of the (relaxed) system characterized by the impulse response
h(n) = ( l ) u(n) to the input signal | 1, x(n) = { ^ 10, 0 < n < 10 ~ otherwise
2 3 8 D eterm ine the response of the (relaxed) system characterized by the impulse response
h(n) = ( j ) H u{n) to the input signals (a) x(n) = 2nu(n) (b) x(n) = u ( - n )
Chap. 2
Problems
143
239 T hree systems with impulse responses h\(n) 5(n) &(n 1). h2(rt) = h( n}. and = u(n), are connected in cascade. (a) W hat is the impulse response. of the overall system? (b ) Does the order of the interconnection affect the overall system? 2.40 (a) Prove and explain graphically the difference between the relations x(n )5(n no) = o) and x(n) *&(n - n0) = x(n - nu)
(b ) Show that a discrete-time system, which is described by a convolution summation,
is LTI and relaxed, (c) W hat is the impulse response of the system described by y(n) = x(n n())? 2.41 Two signals 5(n) and u(n) are related through the following difference equations j(n) + a\ j(n 1) + N) = bi)v(n)
Design the block diagram realization of: (a) The system that generates x(n) when excited by v(n). (b ) The system that generates u(n) when excited by s(n). (c) What is the impulse response of the cascade interconnection of systems in parts (a) and (b)? 2.42 Com pute the zero-state response of the system described by the difference equation y(n ) + ^ v(n 1) = x(n ) + 2x{n 2) to the input xin) = (1.2. 3. 4, 2, 1)
T
by solving the difference equation recursively.
2 .43 Determ ine the direct form II realization for each of the following LTI systems. (a) 2v(n) + y(n 1 ) - 4 v(n 3) = x(n) + 3x(n 5) (b) y(n) = Jr(n) x(n 1) + 2x(n 2) 3x(n - 4) 2 .4 4 Consider the discrete-time system shown in Fig. P2.44.
Figure P2.44
(a) Com pute the 10 first samples of its impulse response.

(b ) Find the input-output relation.
(c) Apply the input x(n) = { 1 .1 .1 ....} and com pute the first 10 samples of the output, t
144
Chap. 2
(d) Com pute the first 10 samples of the output for the input given in part (c) by using convolution. (e ) Is the system causal? Is it stable? 2 ^ 5 Consider the system described by the difference equation y(n) = ay(n - 1) + bx(n)
(a ) Determ ine b in terms of a so that
(b ) Com pute the zero-state step response s(n) of the system and choose b so that
j(oo) = 1. (c) Com pare the values of b obtained in parts (a) and (b). W hat did you notice? 2*46 A discrete-time system is realized by the structure shown in Fig. P2.46. ( a) D eterm ine the impulse response. (b) Determ ine a realization for its inverse system, that is, the system which produces x( n) as an output when y(n) is used as an input.
x in )
-o
0.8
v(n)
Figure P2.46
2 .4 7 Consider the discrete-time system shown in Fig. P2.47.
v(n)
Figure P2^f7
(a) Com pute the first six values of the impulse response of the system.
(b ) Com pute the first six values of the zero-state step response of the system.
( c ) Determ ine an analytical expression for the impulse response of the system.
1 4 8 D eterm ine and sketch the impulse response of the following systems for n 0, 1........9. (a ) Fig. P2.48(a). (b ) Fig. P 2 .4 8 (b ). ( c ) Fig. P2.48(c).
Chap. 2
Problems
145
Cc)
Figure P2.48
(d) Classify the systems above as FIR or IIR.

(e) Find an explicit expression for the impulse response of the system in part (c).
2.49 Consider the systems shown in Fig. P2.49. (a) Determ ine and sketch their impulse responses /i|(n), h2(n ), and h3(n). (b) Is it possible to choose the coefficients of these systems in such a way that h\(n) = h2(n) = h3(n) 2.50 Consider the system shown in Fig. P2.50. (a) Determ ine its impulse response h(n). (b) Show that h(n) is equal to the convolution of the following signals. h] (n) = 6(n) + 6(n - 1) M " ) = (^)"u(n)
146
Chap. 2
v (n )
yin)
2.51 Com pute the sketch the convolution y,(n) and correlation r,(n) sequences for the following pair of signals and comment on the results obtained. (a) *,{) = (1.2.4) A ,(n )= (1 ,1 .1 .1 . If
t t
(b) x2(n) = (0,1. - 2 . 3, - 4 ]

t
h2(n) = U. 1. 2 . 1 , i}
t
(c)
jt j( b )
= (1.2. 3, 4}
t
A3 (u )
= (4. 3. 2, 1)
t
(d) x4(n) = {1. 2, 3,4)

t
hA(n) = (1.2.3. 4)
t
2.52 The zero-state response of a causal LTI system to the input x ( n ) = {1,3, 3,1) is y(n) = (1,4, 6 ,4 ,1 ). Determ ine its impulse response. t
Chap. 2
Problems
147
2.53 Prove by direct substitution the equivalence of equations (2.5.9) and (2.5.10), which describe the direct form II structure, to the relation (2.5.6), which describes the direct form I structure. 2.54 D eterm ine the response y(n). n > 0 of the system described by the second-order difference equation
y ( n) 4 y(n - 1) + 4v(/i 2) = x( n) x( n 1)
when the input is x(n) = (-l)" u (n ) and the initial conditions are v(1) = y ( -2 ) = 0. 2.55 D eterm ine the impulse response h(n) for the system described by the second-order difference equation
v(ni 4v(;i 1 > + 4y(n 2) = x(r?) x(n 1)
2.56 Show that any discrete-tim e signal x(n) can be expressed as c(n) = [.v(A') x(k 1)]u(n - k)
where (/i - k ) is a unit step delayed by k units in time, that is,

u(/i k) = .
1. I 0,
n >k otherwise
2.57 Show that the output of an LTI system can be expressed in term s of its unit step response v(n) as follows. i'(n) = y " ' [,v(A :) x(k 1)]jr(fl k)
[x(AQ - x ( k - l) ] s ( /i - k)
c 2.58 Com pute the correlation sequences rIX(l) and rtv(l) for the following signal sequences.
j _ P nu - N < n < n {, + N
I 0,
f 1.
v(n) = 1
otherwise
-N <n < N
1 0,
otherwise
2.59 D eterm ine the autocorrelation sequences of the following signals. (a) x(n) = {1. 2.1.1)
t
(b) v(n) = i l . 1.2.1}

t
W hat is your conclusion? 2.60 W hat is the normalized autocorrelation sequence of the signal x(n) given by 1, x(n) = , 0, -jV < n < N otherwise
148
Chap. 2
2.61 An audio signal j(r) generated by a loudspeaker is reflected at two different walls with reflection coefficients r ] and r2. The signal *(/) recorded by a microphone close to the loudspeaker, after sampling, is x (n) = s(n) + rxs(n k\) + r2s(n k2) where kt and k2 are the delays of the two echoes. (a) Determ ine the autocorrelation rzx(I) of the signal x(n). Can we obtain ri, r2, k\, and k2 by observing r,s (1)1 (c) W hat happens if r2 = 0?
(b )
2.62* Time-delay estimation in radar Let xa(t) be the transm itted signal and yfl(r) be the received signal in a radar system, where y(r) = axa(t - id) + vu(f) and va(t ) is additive random noise. The signals xa(t) and >(/) are sampled in the receiver, according to the sampling theorem , and are processed digitally to deter mine the time delay and hence the distance of the object. The resulting discrete-time signals are jr(n) = xa(nT) y(n) = y(nT) = axu(nT - DT) + vu(nT) = ax(n D) + u(n) (a) Explain how we can measure the delay D by computing the crosscorrelation r*,.(/). (b) Let x(n) be the 13-point Barker sequence
X
(n) = (+ 1,+1,+1,+1,+1, -1, -l.+l.+l, -1,+1. -1,+1)
and u(n) be a Gaussian random sequence with zero mean and variance a 2 = 0.01. Write a program that generates the sequence v(n), 0 < n < 199 for a = 0.9 and D = 20. Plot the signals jt(), y(n), 0 < n < 199. (c) Compute and plot the crosscorrelation rT V (/), 0 < / < 59. Use the plot to estimate the value of the delay D. (d) Repeat parts (b) and (c) for a 2 = 0.1 and a 2 = 1. <e) Repeat parts (b) and (c) for the signal sequence jt(n) = j _ l , _ l , - l , + i , + i , + i . + i , - i , + 1 . - l . + l . + l , - 1 ,- 1 ,+ 1 } which is obtained from the four-stage feedback shift register shown in Fig. P2.62,
Figure P2.61 register.
Linear feedback shift
Chap. 2
Problems
149
Note that x(n) is just one period of the periodic sequence obtained from the feedback shift register. (f) Repeat parts (b) and (c) for a sequence of period N 27 1, which is obtained from a seven-stage feedback shift register. Table 2.3 gives the stages connected to the modulo-2 adder for (maximal-length) shift-register sequences of length
N =2" TABLE 2.3 SHIFT-REGISTER CONNECTIONS FOR GENERATING MAXIMAL-LENGTH SEQUENCES m Stages Connected to Modu)o-2 Adder
1
2 3 4 5 6 7 H y id li 12 13 14 15 16 17
1
1. 2 1. 3 1. 4 S. 4 L6 1. 7 1, 5, 6. 7 1.6 1. K 1. 10 i. 7. y, 12 1. H), 11. 13 1. 5. 9. 14 1. 15 1. 5. 14, 16 1. 15
2.63* Implementation o f L T I systems by the difference equation
Consider the recursive discrete-time system described
_v(n) = U\ v( n 1) a;v (n 2) 4- b^x(rt)
where a\ - 0.8, u? = 0.64. and b() = 0.866. (a) Write a program to compute and plot the impulse response h{n ) of the system for 0 < n < 49. (b) Write a program to com pute and plot the zero-state step response s(n) of the system for 0 < n < 100. (c) Define an F IR system with impulse response ^ fir (h ) given by , , .
1 h(n),
10.
0 < n < 19 elsewhere
where h(n) is the impulse response computed in part (a). W rite a program to compute and plot its step response. (d) Com pare the results obtained in parts (b) and (c) and explain their similarities and differences.
150
Chap. 2
2jS4* W rite a com puter program that computes the overall impulse response h(n) of the sys
tem shown in Fig. P2.64 for 0 < n < 99. The systems TU T2, T 3, and % are specified by Ti : hi(n) = {1 . 5 . t g.
55)
T2 : h 2(n) = {1,1,1,1,11 t Ts : + 5*(" - 1) + - 2)
T4 : y(n) = 0 .9y(n 1) 0.81y{n 2) + v(n) + u(n 1) Plot h(n) for 0 < n < 99.
Figure P2.64
The Z -Transform and Its Application to the Analysis of LTI Systems
T ra n s fo rm te c h n iq u e s a re an im p o rta n t tool in the analysis o f signals a n d lin e a r tim e -in v a ria n t (L T I) system s. In this c h a p te r w e in tro d u c e th e ^ -tran sfo rm , d ev elo p its p ro p e rtie s , and d e m o n s tra te its im p o rta n c e in th e analysis an d c h a ra c te riz a tio n o f lin ear tim e -in v a ria n t system s. T h e : -tra n sfo rm plays th e sam e role in th e analysis o f d isc re te -tim e signals an d L T I sy stem s as th e L ap lace tran sfo rm d o e s in th e analysis o f c o n tin u o u s-tim e signals a n d L T I svstem s. F o r ex am p le, w e sh a d see th a t in th e ^ -d o m ain (com plex z -p lan e) th e c o n v o lu tio n of tw o tim e-d o m ain signals is e q u iv a le n t to m u ltip lic atio n o f th e ir c o rre sp o n d in g ^ -tran sfo rm s. T h is p ro p e rty g reatly sim plifies the analysis o f th e re sp o n s e o f an L TI system to v ario u s signals. In ad d itio n , th e c-tran sfo rm p ro v id es us w ith a m ean s of ch a ra c te riz in g an L T I system , a n d its resp o n se to v ario u s signals, by its p o le - z e r o locations. W e b eg in th is c h a p te r by defining th e c-tran sfo rm . Its im p o rta n t p ro p e rtie s a re p re s e n te d in S ectio n 3.2. In S ection 3.3 the tra n sfo rm is u se d to ch aracterize signals in te rm s o f th e ir p o le - z e r o p a tte rn s. S ection 3.4 d escrib es m e th o d s fo r in v ertin g th e z-tra n sfo rm o f a signal so as to o b ta in th e tim e -d o m a in re p re s e n ta tio n o f th e signal. T h e o n e-sid ed :-tra n s fo rm is tr e a te d in S ectio n 3.5 a n d used to solve lin e a r d ifferen ce e q u a tio n s w ith n o n z e ro in itial co n d itio n s. T h e c h a p te r c o n clu d es w ith a d iscussion o n th e use of th e z -tra n sfo rm in th e analysis o f L TI system s.
3.1 THE Z-TRANSFORM

In th is sectio n w e in tro d u c e th e z -tra n sfo rm of a d isc re te -tim e signal, in v estig ate its c o n v e rg e n c e p ro p e rtie s , an d briefly discuss th e in v erse z-tran sfo rm .
151
152
The z-Transform and Its Application to the Analysis of LTI Systems
Chap. 3
3.1.1 T h e D ire c t z - T r a n s f o r m T h e ^ -tran sfo rm o f a d isc re te -tim e signal jt(fi) is d efin ed as th e p o w e r series
OC
X(z) =
n= oo
x { n ) z ~'
(3-u )
w h ere z is a co m p lex v ariab le. T h e re la tio n (3.1.1) is so m e tim e s called th e direct z - t ransform b ecau se it tra n sfo rm s th e tim e -d o m a in signal x ( n ) in to its co m p lex p la n e re p re s e n ta tio n X (z). T h e in v erse p ro c e d u re [i.e., o b ta in in g x ( n ) fro m X (z)] is called th e i nverse z- t ransf orm a n d is e x a m in e d briefly in S e ctio n 3.1.2 a n d in m o re d etail in S ectio n 3.4. F o r c o n v en ien ce , th e z -tra n sfo rm o f a signal x ( n ) is d e n o te d by X (z) - Z U ( )} w h ereas th e re la tio n sh ip b e tw e e n x ( n ) a n d X ( z ) is in d ic a te d by
jc()
(3.1.2)
< -^ * (;)
(3.1.3)
S ince th e z-tra n sfo rm is an infinite p o w e r se ries, it exists only fo r th o se v alu es of z fo r w hich th is se ries co n v erg es. T h e region o f conver gence ( R O C ) o f X (z) is th e set o f all v alu es o f z fo r w hich X ( z ) a tta in s a finite value. T h u s an y tim e w e cite a z-tra n sfo rm w e sh o u ld also in d icate its R O C . W e illu strate th e se co n c e p ts by so m e sim ple ex am p les. Example 3.1.1 D eterm ine the ^-transforms of the following finite-duration signals. (a) jf,(n) = (1 ,2 .5 .7 ,0 ,1 }
(b) jf,() = ( I . 2 . 5 . 7 . 0 . 1)
t
(c) jc3(n) = (0 ,0 ,1 ,2 , 5, 7,0.1} (d) *4(h) = (2 .4 ,5 .7 .0 ,1 )

t
(e) x j ( n ) = S( n)
(f) x$(n) = <S(n k), k > 0 (g) x j ( n ) = &(n + k) , k > 0 Solution From definition (3.1.1), we have
(a) X](z) = 1 + 2z~' + 5z~2 + 7 z '3 + z~5, ROC: entire z-plane except z = 0 (b) X 2(z ) = z2 + 2z + 5 + 7c-1 + z-3, ROC: entire z-plane except z = 0 and z = oo (c) Xj(z) = z~2 + 2z-3 + 5z-4 + 7z-5 -I- z-7, ROC: entire z-plane except z = 0 (d) X4(z) = 2z2 -I- 4z -I- 5 4- 7z_1 -I- z-3, ROC: entire z-plane except z = 0 and z = oo (e) X;(z) = l[i.e S(n) * 1], ROC: entire z-plane (f) Xb(z) = z_ t[i.e &(n k) z_t], k > 0, ROC: entire z-plane except z = 0 (g) Xy(z) = zk[i.e &(n + k) z*], k > 0, ROC: entire z-plane except z = oo
Sec. 3.1
The z-Transform
153
F ro m th is ex a m p le it is easily se en th a t th e R O C o f a f i ni t e-durat i on signal is th e e n tire ;- p la n e , e x cep t p ossibly th e p o in ts z = 0 a n d /o r z oo. T h e se p o in ts a re e x c lu d ed , b e c a u s e z t (k > 0} b eco m es u n b o u n d e d fo r z = oc an d z ~ k01 > 0) b ec o m e s u n b o u n d e d fo r z = 0. F ro m a m a th e m a tic a l p o in t of view th e z-tra n sfo rm is sim p ly an a lte rn a tiv e r e p re s e n ta tio n o f a signal. T h is is nicely illu stra te d in E x a m p le 3.1.1, w h e re w e see th a t th e co effic ie n t o f z~", in a given tra n sfo rm , is th e v a lu e o f th e signal at tim e n. In o th e r w o rd s, th e e x p o n e n t o f z co n tain s th e tim e in fo rm a tio n w e n eed to id en tify th e sa m p le s o f th e signal. In m an y cases w e can ex p ress th e sum of th e finite o r infinite se rie s fo r th e z-tra n s fo rm in a c lo sed -fo rm e x p ressio n . In such cases th e z -tra n s fo rm o ffers a co m p act a lte rn a tiv e r e p re s e n ta tio n of th e signal. Example 3.1.2 D eterm ine the z-transform of the signal Jf(n) = (5 Solution The signal jc(n) consists of an infinite num ber of nonzero values
x(n)= ( l . a u ^ . U )'1 ....

The z-transform of x(n) is the infinite power series
X( z ) = 1 + U ' + ( ^) 2z - 2 + (| ) "z"" + ---
f t = < )
nail
This is an infinite geometric series. We recall that

1 + A + ,42 + A3 - t - - - - = if | A 1 < 1
1 A
Consequently, for | 1
< 1, or equivalently* for \z\ > X(z) =

J jZ
X(z) converges to
ROC: (zl > |
We see that in this case, the z-transform provides a compact alternative representation of the signal x(n). L e t us e x p ress th e co m p lex v a ria b le z in p o la r fo rm as z = r e }6 w h ere r = |z| a n d 6 = i^z- T h e n X ( z ) can be e x p ressed as
OC
(3.1.4)
X ( z ) \ t- r ' = Y x ( n ) r ~ ne - j en n*-00
154
Chap. 3
In th e R O C o f X U ). |X ( ;) | < oc. B ut
(3.1.5) 5 \ M n ) r - ne - ' * n \ = \ x( n) r ~n\
H e n c e |X (z)| is finite if th e se q u en ce x ( n ) r ~ ' is a b s o lu te ly su m m a b le . T h e p ro b le m o f finding th e R O C fo r X ( z ) is e q u iv a le n t to d e te rm in in g the ra n g e o f v alu es o f r fo r w hich th e se q u e n c e x ( n ) r ~ " is a b so lu te ly su m m ab le. T o e la b o ra te , let us e x p ress (3.1.5) as
(3.1.6)
If X ( z ) co n v erg es in so m e region of the co m p lex p la n e , b o th su m m a tio n s in (3.1.6) m u st be finite in th a t region. If the first sum in (3.1.6) co n v erg es, th e re m ust exist v alu es o f r sm all e n o u g h such th at the p ro d u c t se q u e n c e x ( n) r" . 1 < /; < oc, is a b s o lu te ly su m m ab le. T h e re fo re , the R O C for th e first sum co n sists o f all p o in ts in a circle of so m e rad iu s r^, w here / ] < oc, as illu stra te d in Fig. 3.1a. O n the o th e r h a n d , if th e seco nd sum in (3.1.6) co n v erg es, th e re m u st exist v alu es o f r larg e en o u g h such th a t th e p ro d u c t se q u e n c e x ( n ) / r " . 0 < n < oc, is a b so lu te ly su m m ab le. H en ce th e R O C fo r the se co n d sum in (3.1.6) co n sists o f all p o in ts o u tsid e a circle o f rad iu s r > r2. as illu stra te d in Fig. 3.1b. Since th e co n v erg en ce of X (c) re q u ire s th a t b o th sum s in (3.1.6) b e finite, it follow s th a t th e R O C o f X ( z ) is g en erally specified as th e a n n u la r region in th e ;- p la n e , r: < r < r\. w hich is the co m m o n reg io n w h e re b o th su m s are finite. T his reg io n is illu stra te d in Fig. 3.1c. O n th e o th e r h a n d , if > r\, th e r e is no co m m o n reg io n o f co n v erg en ce fo r th e tw o sum s an d h en ce X ( ;) d o es n o t exist. T h e follow ing ex am p les illu strate th e se im p o rta n t co n cep ts. Example 3.1.3 Determ ine the e-transform of the signal
Solution
From the definition (3.1.1) we have
If \az 11 < 1 or equivalently, |z| > |a |, this power series converges to 1/(1 - a ; -1).
Sec. 3.1
The z-Transform
155
Im(z)
Im(;)
Im(z)
Figure 3.1 R egion of convergence for X (z) and its corresponding causal and anticausa! components.
156
Chap. 3
lm(;)
Figure 3.2 The exponential signal xi n) = txnu{n) (a), and the ROC of its transform (b).
Thus we have the z-transform pair x(n) =a" u( n ) X (;) = 1

1 - a ;-1
ROC: |z| > |cr|
(3.1.7)
The R O C is the exterior of a circle having radius |a |. Figure 3.2 shows a graph of the signal .x(n) and its corresponding ROC. Note that, in general, or need not be real. If we set or = 1 in (3.1.7), we obtain the z-transform of the unit step signal x(n) - u(n ) Example 3.1.4 Determ ine the z-transform of the signal x(n) = a"u( n 1) = Solution 0, n >0 n < -1
oc
X(z) =
1 1
ROC: |zl > 1
(3.1.6
From the definition (3.1.1) we have

-I
no c
/= i
where / = n. Using the formula A + A: + A3 + = A(1 + A + A2 + )= when | > 1f < 1 gives * ( ;) = - 1 1 a ~ lz l - a ; '1 A
1 - A
provided that |cr_1z| < 1 or, equivalently, |z| < jar j. Thus 1 (3.1.9) ROC: [z| < |a | 1-azThe R O C is now the interior of a circle having radius |a|. This is shown in Fig. 3.3. x(n) = a"u( n 1) X(Z) = -
Sec. 3.1
The z-Transform
Im(c)
157
Figure 3.3 Anticausal signal jt(h) = -crnu( n - 1) (a), and the ROC of its transform (b).
E x a m p le s 3.1.3 an d 3.1.4 illu stra te tw o very im p o rta n t issues. T h e first c o n cern s th e u n iq u e n e s s o f th e "-tran sfo rm . F ro m (3.1.7) an d (3.1.9) we see that th e cau sal signal a nu ( n ) a n d th e an ticau sal signal a " u ( n 1) have id en tical clo sed -fo rm e x p re ssio n s fo r th e ^ -tran sfo rm , th a t is,
Z ( o " w ( ) ) = Z { a nu ( n 1) ) ------ --------
1 a c ' 1
T h is im p lies th a t a clo sed -fo rm e x p ressio n fo r th e z-tra n sfo rm d o e s n o t u n iq u ely specify th e signal in th e tim e do m ain . T h e am b ig u ity can b e reso lv ed only if in a d d itio n to th e c lo sed -fo rm ex p ressio n , th e R O C is specified. In su m m ary , a discrete-time signal x ( n ) is uni quel y d et er mi ned b y its z- t rans f orm A' (;) a n d the region o f conver gence o f X( z ) . In this te x t th e te rm z -tra n s fo rm " is u se d to re fe r to b o th th e clo sed -fo rm e x p ressio n an d th e c o rre sp o n d in g R O C . E x a m p le 3.1.3 also illu stra te s th e p o in t th a t the R O C o f a causal signal is the exterior o f a circle o f s o m e radius r 2 whi l e the R O C o f an ant icausal signal is the interior o f a circle o f s o m e radi us rj. T h e fo llow ing e x am p le co n sid ers a se q u en ce th a t is n o n z e ro for 00 < n < 00. Example 3.1.5 D eterm ine the z-transform of the signal x (n) = a"u(n) + bnu( n 1) Solution From definition (3.1.1) we have X(z) =
n=0 n= oc
b"z " =
n=0
+ Y ib -'z)1
i= l
The first pow er series converges if locz-11 < 1 or |z| > |a |. The second power series converges if \b~xz\ < 1 or |z| < |6j. In determ ining the convergence of XCz), we consider two different cases.
158
Chap. 3
Case 1 |b| < |a |: In this case the two R O C above do not overlap, as shown in Fig. 3.4(a). Consequently, we cannot find values of z for which both power series converge simultaneously. Clearly, in this case, X(c) does not exist. Case 2 |f>| > |a |: In this case there is a ring in the z-plane where both power series converge simultaneously, as shown in Fig. 3.4(b). Then we obtain X(z) = 1 ccz 1 1 bz 1 b a
(3.1.10)
a + b z abz~l The R O C of X(z) ts |cr| < \z\ < \b\. T h is e x am p le show s th a t i f there is a R O C f o r an infinite durat i on two-si ded signal, it is a ring ( annul ar region) in the z-plane. F ro m E x a m p le s 3.1.1, 3.1.3, 3.1.4, an d 3.1.5. w e see th a t th e R O C of a signal d e p e n d s o n b o th its d u ra tio n (finite or in fin ite) an d o n w h e th e r it is cau sal, a n tic a u sa l, o r tw o -sid ed . T h e se facts are su m m a riz e d in T a b le 3.1. O n e special case of a tw o -sid ed signal is a signal th a t h as infinite d u ra tio n on th e rig h t sid e b u t n o t on th e left [i.e., x( n ) = 0 fo r n < (l < 0], A sec on d case is a signal th a t has infinite d u ra tio n o n th e left side b u t n o t on the
plane
Ifcl < tot

X(z) does not exisl
Im(;)
krl < IA I
ReU)
ROC for X(z) Figure 3.4 R O C fo r z-transform in E xam ple 3.1.5.
Sec. 3.1
The ^-Transform
TABLE 3.1
CH A R A C TE R IS TIC FAM ILIES O F SIGN ALS W IT H T H E IR
159
C O R R E S P O N D IN G ROC
Signal
ROC Finite-Duration Signals
Two-sided
. . TTT l i t ,
0 Causal
Jnfinite-Duration
l l T t*
Anltcausal
T ]T 1
Two-sided
I . I t.....
rig h t [i.e., x{ n ) = 0 fo r n > n\ > 0]. A th ird sp ecial case is a signal th a t has finite d u ra tio n o n b o th th e left a n d rig h t sides [i.e., x ( n ) = 0 fo r n < no < 0 a n d n > n\ > 0]. T h e se ty p e s o f signals a re so m e tim e s c alled right-sided, left sided, a n d finite-d uration two-sided, signals, resp ec tiv ely . T h e d e te rm in a tio n o f th e R O C fo r th e s e th r e e ty p es o f signals is left as an e x ercise fo r th e r e a d e r (P r o b lem 3.5). F in ally , w e n o te th a t th e z -tra n sfo rm d efin ed b y (3.1.1) is so m e tim e s re fe rre d to as th e tw o-sided o r bilateral z-transform , to d istin g u ish it fro m th e o ne-sided o r
160
Chap. 3
unilateral z- t ransf orm given by X~(z) = ^ * ( ) z ~ " (3.1.11)
T h e o n e-sid ed z -tra n sfo rm is ex am in ed in S ection 3.5. In this tex t w e use the ex p ressio n z-tra n sfo rm exclusively to m ean th e tw o -sid ed z -tra n sfo rm d efined by (3.1.1). T h e te rm 'tw o -sid e d ' will b e u se d oniy in cases w h ere w e w an t to resolve any am b ig u ities. C learly , if x ( n ) is causal [i.e., * ( 72) = 0 for n < 0], th e o n e-sid ed and tw o -sid ed z-tran sfo rm s a re eq u iv a le n t. In any o th e r case, th e y a re d ifferen t.
3.1.2 The Inverse z-Transform

O ften , we h av e th e z-tra n sfo rm X ( z ) of a signal a n d w e m ust d e te rm in e th e signal se q u en ce. T h e p ro c e d u re fo r tra n sfo rm in g from th e z-d o m ain to th e tim e dom ain is called th e inverse z- t ransform. A n in v ersio n fo rm u la for o b ta in in g x( n) from X (z) can b e d eriv ed by using th e Cau c hy integral t h e o r e m , w hich is an im p o rta n t th e o re m in th e th e o ry o f co m p lex variables. T o b egin, we h av e th e z-tra n sfo rm defin ed by (3.1.1) as
S u p p o se th a t we m u ltip ly b o th sides o f (3.1.12) by z"~' an d in te g ra te both sides o v er a closed c o n to u r w ithin the R O C o f X ( z ) w hich en clo ses th e origin. Such a c o n to u r is illu stra te d in Fig. 3.5. T h u s we have (3.1.13) w h e re C d e n o te s th e clo sed c o n to u r in th e R O C o f A'(z). ta k e n in a c o u n te rc lo c k w ise d irectio n . Since th e se rie s c o n v erg es on th is c o n to u r, w e can in te rc h a n g e th e o rd e r o f in te g ra tio n an d su m m atio n o n th e rig h t-h a n d side o f (3,1.13). T h u s
im (o
Figure 3.5 (3.1.13).
Contour C for integral in
Sec. 3.2
Properties of the z-Transform
161
(3.1.13) b eco m es x ( z ) z n- ' d z = t= 00 x ( k ) ( h z n- l~kd z * (3.1.14)
N o w w e can in v o k e th e C a u ch y in te g ra l th e o re m , w hich s ta te s th a t
ftM5)
w h ere C is an y c o n to u r th a t en clo ses th e origin. B y ap p ly in g (3.1.15), th e righth a n d side o f (3.1.14) re d u c e s to 2 n j x ( n ) a n d h en c e th e d e sire d in v e rsio n fo rm u la x ( n ) = ^ - ^ X ( z ) z n_I d z (3.1.16)
A lth o u g h th e c o n to u r in te g ra l in (3.1.16) p ro v id es th e d e s ire d in v ersio n fo r m u la fo r d e te rm in in g th e se q u e n c e * ( ) fro m th e z-tra n sfo rm , w e sh all n o t use (3.1.16) d irectly in o u r e v a lu a tio n o f in v erse z -tran sfo rm s. In o u r tre a tm e n t w e d eal w ith signals a n d sy stem s in th e z-d o m ain w hich h av e ra tio n a l z -tra n s fo rm s (i.e., ztra n sfo rm s th a t a re a ra tio o f tw o p o ly n o m ials). F o r such z -tra n sfo rm s w e d e v e lo p a s im p ler m e th o d fo r in v ersio n th a t ste m s from (3.1.16) and e m p lo y s a ta b le lo o k u p .
3.2 PROPERTIES OF THE Z-TRANSFORM

T h e z -tra n sfo rm is a v ery p o w erfu l to o l fo r th e stu d y o f d isc re te -tim e signals an d system s. T h e p o w e r o f th is tra n sfo rm is a c o n s eq u en ce o f so m e v ery im p o rta n t p ro p e rtie s th a t th e tra n sfo rm possesses. In th is sectio n w e e x am in e som e o f th e se p ro p e rtie s . In th e tre a tm e n t th a t follow s, it sh o u ld b e re m e m b e re d th a t w h e n w e c o m b in e sev eral z -tra n sfo rm s, th e R O C o f th e o v erall tra n sfo rm is, at least, th e in te rse c tio n o f th e R O C o f th e in d iv id u al tra n sfo rm s. T h is will b eco m e m o re a p p a re n t la te r, w h en w e discu ss specific ex am p les.
Linearity.
an d
If Jt,(n) Xi ( z )
x 2(n) < -U X 2(z) th e n x( n) = a \ x \ ( n ) + a 2x 2(n) X( z ) = a j Xj ( z ) -h o2X 2(z) (3.2.1)
fo r an y c o n s ta n ts a i a n d a 2. T h e p ro o f o f th is p ro p e rty follow s im m e d ia te ly from th e d efin itio n o f lin e a rity a n d is left as an ex ercise fo r th e re a d e r. T h e lin e a rity p ro p e rty can easily be g e n e ra liz e d fo r an a r b itra r y n u m b e r o f signals. B asically , it im p lies th a t th e z -tra n sfo rm o f a lin e a r c o m b in a tio n o f signals is th e sa m e lin e a r c o m b in a tio n o f th e ir z-tran sfo rm s. T h u s th e lin e a rity p ro p e rty h e lp s u s to find th e z -tra n s fo rm o f a signal by ex p ressin g th e signal as a su m o f e le m e n ta ry signals, fo r ea c h o f w hich, th e z -tra n sfo rm is a lre a d y k n o w n .
162
Chap. 3
Example 3.2.1 Determ ine the z-transform and the R O C of the signal *() = [3(2") -4 (3 ")] ( ) Solution If we define the signals jti() = 2"u(n) and x2(n) = 3"u(n) then
jc()
can be written as x(n) = 3xi(n) 4 x2(n)
According to (3.2.1), its z-transform is X(;) = 3 X ,( c ) - 4 X 2(z) From (3.1.7) we recall that a'u(n ) 1 - az ROC: |z| > |a| (3.2.2)
By setting a = 2 and or = 3 in (3.2.2). we obtain Ai(n) = 2u(n) x2(n) = 3nu(ri) X,(z) = ^ X 2(z) ~ ^ __ , ROC: |zj > 2 ROC: |zl > 3
The intersection of the ROC of X,(z) and X2(:) is |z| > 3. Thus the overall transform X (z) is ROC: |;| > 3
Example 3.2.2 Determ ine the z-transform of the signals (a) x(n) = (COSoon)u(n) (b) x(n) = (sin w^n)u(n) Solution (a) By using E uler's identity, the signal .t(n) can be expressed as x (n) = (cosa>on)u(n) = \ e Jaln"u(.n) + ê~JU Jn u(n) Thus (3.2.1) implies that
X( z ) = \ Z { e ^ u ( n ) ) + ^ { g - ^ u i n ) }
Sec. 3.2
163
If we set a = e^Ô ori = \ei< u\ = 1) in (3.2.2), we obtain eJU V "u(fi) and e - ,wnnu(n) ^ Thus X(z) = -------- ----------h I ------- - ------r 2 1 ^ f z -1 2 1 e~J^lz~I RO C : |;1 > 1
-------- T T T
---------------
R O C : |z[ > 1
1-
R O C : l; l >
A fter some simple algebraic manipulations we obtain the desired result, namely, (coscO(}f?)u(n) < - - ;---------------- 1 - 2z ~1cos ton + (b) From E ulers identity,
x(n) = (sino<o)(n) = ^-\eiun" u(n) e~,w""u()]
2j : 1 C_ ! COSW()
ROC: |-1 > 1
L r2.j)
,, ,, ,
Thus X(-) = ^ - f - ------1 ------- - ------- ------- - | 2j \ 1 r _1 1 and finally. z sm O H ) (sin coi,n)u(n) x - ^ ----- - -----------------1 - 2z_1 cos wo + z~ROC: |;| > 1 (3.2.4) ROC: |:| > 1
Time shifting.
If x( n) X (c)
th e n x{n - k ) z ~ kX{ z ) (3.2.5)
T h e R O C o f z ~kX ( z ) is th e sam e as th a t o f X ( z ) e x c e p t fo r j = 0 if k > 0 and z = oo if k < 0. T h e p r o o f o f th is p ro p e rty follow s im m e d ia te ly fro m the definition o f th e ^ -tra n sfo rm given in (3.1.1) T h e p r o p e rtie s o f lin e a rity an d tim e shifting a re th e k ey fe a tu re s th a t m ake th e z -tra n s fo rm e x tre m e ly u se fu l fo r th e analysis o f d isc re te -tim e L T I system s. Exam ple 3*23 By applying the time-shifting property, determ ine the z-transform of the signals .T 2(n,i and xi(n) in Exam ple 3.1.1 from the Z'transform of Jti(n). Solution
It can easily be seen that

x 2( n ) = * ]( + 2)
164
and
The ^-Transform and Its Application to the Analysis of LTI Systems
Chap. 3
*i(n) = j:i(n - 2) Thus from (3.2.5) we obtain X2(z) = z2X, (z) = r + 2; + 5 + 7z ' 1 + z3 and X 3(z) = r 2X,(z) = z"2 + 2z"3 + 5z-4 + 7z"s + z-7 Note that because of the multiplication by z2, the R O C of X2(z) does not include the point z = oc, even if it is contained in the R O C of Aî(z). E x am p le 3.2.3 p ro v id es a d d itio n a l insight in u n d e rs ta n d in g th e m e an in g of th e shiftin g p ro p e rty . In d e e d , if we recall th a t th e coefficient o f z ~ n is th e sam ple v alu e at tim e n, it is im m e d ia te ly se en th a t d elay in g a signal by k( k > 0) sam ples [i.e., x ( n ) x ( n A')] c o rre sp o n d s to m u ltip ly in g all te rm s o f th e z -tra n sfo rm by z~ k. T h e co efficien t o f z~" b ec o m e s th e coefficient o f ~~tn+k). Example 3.2.4 Determ ine the transform of the signal
f 1, ' " H o . 0 < n < jV - 1 e lse w h e r e , <316>
Solution We can determ ine the z-transform of this signal by using the definition (3.1.1). Indeed, * -i f N, if z = l X(z) = ^ l z ^ = l + ; - ' + - - - + z - (A f- l l = l - r * -f . , , (3.2.7) * = < ' I 1 - z-1' 1 Since .v(n) has finite duration, its R O C is the entire z-plane, except z 0. Let us also derive this transform by using the linearity and time shifting prop erties. Note that x in) can be expressed in terms of two unit step signals x(n) = u(n) u(n N) By using (3.2.1) and (3.2.5) we have X(z) = Z{u(n )} - Z{u(n - N)) = (1 - z"*')Z{u ()} However, from (3.1.8) we have Z[u(n)) = ^ ROC: jzl > 1 (3.2.8)
which, when combined with (3.2.8), leads to (3.2.7). E x a m p le 3.2,4 h e lp s to clarify a v ery im p o rta n t issue re g a rd in g th e R O C o f th e c o m b in a tio n o f se v era l z-tran sfo rm s. If th e lin ear c o m b in a tio n of several sig n als h as finite d u ra tio n , th e R O C o f its z -tra n sfo rm is exclusively d ic ta te d by the fin ite -d u ra tio n n a tu re o f th is signal, n o t by th e R O C o f th e in d iv id u al tran sfo rm s.
Scaling in the z-domain.

x{n)
If
X{z )
ROC: ri < [z| < r2
Sec. 3.2
165
th en a nx ( n ) X (c ~ z) R O C : \a\r\ < |z| < \a\r2 (3.2.9) for a n y c o n s ta n t a, re a l o r com plex. Proof. F ro m th e d efin itio n (3.1.1)
OC OC
Z{a" x( n) } =
a nx ( n ) z " =
x( n) ( a z) n
\o \r i < I;I < \a\r2 T o b e tte r u n d e rsta n d th e m ean in g a n d im plicatio n s o f th e scaling p ro p e rty , w e e x p re ss a a n d z in p o la r form as a = rae->a", z = r e i< 0, a n d w e in tro d u c e a new co m p lex v a ria b le w = a~^z- T h u s Z { x ( n ) ) = X ( z ) an d Z{a" x( n) } = X (if). It can easily b e seen th a t
T h is ch a n g e o f v a ria b le s re su lts in e ith e r sh rin k in g (if r 0 > 1) o r e x p a n d in g (if r 0 < 1) th e z -p lan e in co m b in a tio n w ith a ro ta tio n (if too # 2i-jr ) o f th e z-p lan e (see Fig. 3.6). T h is e x p lain s w hy w e h av e a ch an g e in th e R O C o f th e n ew tra n sfo rm w h e re |a | < 1. T h e case \a\ = 1, th a t is, a = e^w" is o f special in te re st b e c a u se it c o rre sp o n d s o n ly to ro ta tio n o f th e z-p lan e. Exam ple 3.2.5 Determ ine the z-transforms of the signals (a) x(n) = a"(cosu\in)u(n) (b) x(n) = tf"(sin ion)u(n) r-planc
Im(-) w-plane
(W -Cô 0 Re(z) 0
Re(w)
Figure 3.6
o _ 1 Z,
a roe^.
Mapping of the r-plane to the u -plane via the transformation ui =
166
The z-Transform and Its Application to the Analysis of LTI S ystem s Solution (a) From (3.2.3) and (3.2.9) we easily obtain 1 - a ; -1 cos wn
a (C O S a i |, ) u ( r t ) *-------* --------- ---------- ;--------------------------
Chap. 3
1 l a z r 1 co s ton -f a
(3.2.10)
(b) Similarly, (3.2.4) and (3.2.9) yield

a (sin wii/i)u(n) * --- ---- ; ---------
a r -1 sin to la I
(3.2.11)
Time reversal.
If a (//) *-> X ( z ) R O C : r\ < < r;
th en j r ( - n ) <-i-> X { z ~ ]) R O C : < |z < r2

r\
(3.2.12)
P r o o f F ro m th e d efin itio n (3.1.1), w e have Z{a (/;)) = Y

h= - x.
-x { - !i ) z~" Y ' ( h ( z ~ t )~l Ar (z~ ') /= ->;
w h ere th e ch an g e o f v ariab le / = n is m ad e. T h e R O C o f A '( ;_1) is < |c 1f < r-i o r e q u iv alen tly < ];| < r2 n
N o te th a t th e R O C fo r x( n) is the inverse o f th a t fo r x ( n). T h is m ean s th a t if co b elo n g s to th e R O C o f x( n) , th en l/" o is in th e R O C fo r x ( n). A n in tu itiv e p ro o f o f (3.2.12) is th e follow ing. W h en w e fold a signal, the co efficien t o f z ~n b e co m es th e coefficient o f z n. T h u s, fo ld in g a signal is eq u iv alen t to rep lacin g ; by in th e z-tra n sfo rm fo rm u la. In o th e r w o rd s, reflectio n in the tim e d o m ain c o rre sp o n d s to inversion in th e z-d o m ain . Exam ple 3.2.6 Determ ine the z-transform of the signal
x i n ) = u( n)
Solution
It is known from (3.1.8) that u(n) <-U i ROC:

|zl
> 1
By using (3.2.12), we easily obtain u(n) * -----ROC: |z < 1 (3.2.13)
Differentiation in the z-domain.
If
Sec. 3.2
167
then nx{n) < > - z ^ ^ y dz Proof . B y differentiating both sid es o f (3.1.1), w e have dX{z) n~ oc n= oc (3.2.14)
- z ~ l Z{nx(n)}
N o te that b oth transform s have the sam e RO C .
Example 3.2.7
D eterm ine the z-transform of the signal jr(n) = na"u(n )
Solution The signal j:(n) can be expressed as njcitn), where Xj(n) = a"u(n ). From
(3.2.2) we have that jri(rt) = aKu(n) < - X,(z) = - ----- 1 az~' Thus, by using (3.2.14), we obtain n<j"ii(n) dXji z) a : '1 X(z) = - z ^ - = ----------- r-r
az
(1 - az~' V
ROC: |z| > \a\
ROC: |z| > |a|
(3.2.15)
If we set a = 1 in (3.2.15), we find the z-transform of the unit ram p signal ntt(n)
ROC: jz| > 1
(3.2.16)
Example 3.2.8
D eterm ine the signal x(n) whose z-transform is given by X(z) = log(l -t- o z '1) |z! > |o(
Solution By taking the first derivative of X(z), we obtain

dX{z) - a z ~2 dz 1 + az~l Thus dX(z) dz = az 1 - ( - a ) z ->
>
\a]
The inverse z-transform of the term in brackets is (-a) ". The multiplication by z _1 implies a time delay by one sample (time shifting property), which results in ( - a ) " _ 1u(n 1). Finally, from the differentiation property we have nx(n ) = a ( a)n~lu(n 1)
168
The z -Transfomn and Its Application to the Analysis of LTI Systems
Chap. 3
x ( n) = ( - 1 } " +1 u(n - 1)
Convolution of two sequences.

Xi (rt) x 2(n)
th e n
If
X^z) X 2 {z)
x( n) = jfi(n) * x 2 (n)
X ( z ) = X \ ( z ) X 2 (z)
(3.2.17)
T h e R O C o f X (;) is, at least, th e in te rse c tio n of th a t fo r ATi(c) an d AS(z). Proof. T h e co n v o lu tio n o f x i(n ) a n d x 2 (n) is d efin e d as
OC
x(n) = T h e z-tra n sfo rm o f x( n ) is

oc
Y
i' = oc
x i ( k ) x 2(n - k)
oc
ac
AT{;) =
Y 2 x( n) z ~ " =
n -x
Y ,
n= - o c |_Jt = -cc
x \ ( k ) x 2(n - k)
U p o n in te rch an g in g th e o rd e r o f the su m m a tio n s a n d ap p ly in g th e tim e-sh iftin g p ro p e rty in (3.2.5). we o b tain X{z) = Y x ' {k) x 2(n k) z "
= X 2 (z) Y
^ ( k ) z ~ k = X 2 ( z ) X ](z)
Example 3.2.9
Compute the convolution x(rt) of the signals jt,(n) = { 1 .-2 ,1 ) _ f 1. 0 < n < 5
10 ,
elsewhere
Solution From (3.1.1), we have

* ,( ; ) = 1 - 2 ; ' + z ~ 2 X 2(z) = I + z ~ ' + z~2 + j ' 3 + z"4 + According to (3.2.17), we carry out the multiplication of X\(z) and X 2(z). Thus X(z) = X 1 (z)X 1 (z) = 1 - z "1 - z ~6 + z "7 Hence
x ( n) = { 1 . - 1 . 0 , 0, 0, 0 , - 1 , 1 }
t
Sec. 3.2
169
The same result can also be obtained by noting that X,U) = (1 - ; ~ 'r x 2w Then X(z) = (1 - ; - )(l - z~*) = 1 - z "6 + :" 7
1 - ; -6 = 3 p
The reader is encouraged to obtain the same result explicitly by using the convolution summation formula (tim e-domain approach). T h e co n v o lu tio n p ro p e rty is o n e o f th e m o st p o w erfu l p ro p e rtie s o f th e ztra n sfo rm b ec a u se it c o n v e rts th e co n v o lu tio n o f tw o signals (tim e d o m a in ) to m u ltip lic a tio n o f th e ir tra n sfo rm s. C o m p u ta tio n of th e c o n v o lu tio n o f tw o signals, using th e z -tra n sfo rm , re q u ire s th e follow ing steps: 1. C o m p u te th e z -tra n s fo rm s o f th e signals to be co n v o lv ed . X i(z) = Z{ x \ ( n ) \ (tim e d o m ain * -.-dom ain) X 2 (z) = Z { x 2 (n) ]
2 . M u ltip ly th e tw o z -tran sfo rm s.

X (z) = X ,(z )X 2(;) 3. F in d th e in v e rse z -tra n s fo rm o f X (z). x ( n ) = Z _ , {X(z)) (z-d o m ain tim e d o m a in ) (z-d o m ain )
T h is p ro c e d u re is, in m a n y cases, c o m p u ta tio n a lly e a s ie r th a n th e d irect e v a l u a tio n o f th e co n v o lu tio n su m m atio n .
Correlation of two sequences.

xj(n) x 2 (n) th e n
OC
If X i(z) X 2 (z)
Rx,J2 (z) = X 1 (Z)X 2(Z"1)

OC
(3.2.18)
Proof . We recall that rXix2U) = x\ ( l ) * x 2( - l )
170
Chap. 3
U sing th e co n v o lu tio n an d tim e-rev ersa l p ro p e rtie s , we easily o b ta in = Z{x] ( / ) }Z{xz ( /)} = X l ( z ) X 2 ( z ~1) T h e R O C o f R X}X2(z) is at least the in te rse c tio n o f th a t fo r X i(c ) an d X 2("- 1 )A s in th e case o f co n v o lu tio n , th e c ro ssc o rre la tio n o f tw o signals is m o re easily d o n e via p o ly n o m ial m u ltip lic atio n acco rd in g to (3.2.18) a n d th en inverse tra n sfo rm in g th e resu lt. Exam ple 3.2.10 Determ ine the autocorrelation sequence of the signal
x (n ) =
a " u {n ).
1 <
< 1
Since the autocorrelation sequence of a signal is its correlation with itself, (3.2.18) gives
S o lu tio n
Rxs(z) = Z{rxJ I ) ) = X (z)X (;~l ) From (3.2.2) we have Ale) = ---------- J - az~' and by using (3.2.15). we obtain | = -.... 1 - az Thus = - = ------------ ------------ 1 az 1 1 - az 1 a (; + ; ) + fl ROC: \a\ < |;| < |c| ROC: |z| > \a\ (causal signal)
ROC: |;| < (a|
(anticausal signal)
Since the R O C of Rxx(z) is a ring, rIX{i) is a two-sided signal, even if x(n) is causal. To obtain rsJ l ) , we observe that the ^-transform of the sequence in Exam ple 3.1.5 with b = 1 /a is simply (1 a1)Rxx(z). Hence it follows that
rXI(l) = -- a ^
\ a-
oc < I < oo
The reader is encouraged to compare this approach with the time-dom ain solution of the same problem given in Section 2.6.
Multiplication of two sequences.

x \ ( n) x 2 (n) th e n r(n) = x i ( n ) x 2 (n) X(z) =
If ATi(z) X 2 {z)
( ) v~ ^ v
(3.2.19)
w h ere C is a clo sed c o n to u r th a t e n clo ses th e o rig in a n d lies w ith in th e reg io n of co n v erg en ce c o m m o n to b o th X](i>) a n d ^ ( l / u ) .
Sec. 3.2
171
Proof . T h e z -tra n sfo rm o f J 3 (n) is
O C
X(z) = Y
ft
OC
x( n)z ~n = Y
n = oc
x ] ( n) x 2 ( n) z ~n
L e t us s u b s titu te th e in v erse tra n sfo rm * i(n ) = fo r in th e z -tra n sfo rm X (z) a n d in te rc h a n g e th e o r d e r o f su m m a tio n an d in te g ra tio n . T h u s w e o b ta in 1 X (z)= 2 / " oc v~ ^ v
T h e su m in th e b ra c k e ts is sim ply th e tra n sfo rm X 2 (:) e v a lu a te d a t z / v . T h e re fo re , X(z) = ^ - 6 x ^ x 2 2 n j Jc w hich is th e d e s ire d result. T o o b ta in th e R O C o f X (z) w e n o te th a t if X i(u ) co n v e rg e s fo r ru < |u| < r\ an d X 2 iz) c o n v erg es fo r r2! < |z| < r2ll, th e n th e R O C o f X 2 {z/v) is
H e n c e th e R O C fo r A"(z) is at least r\ir2i < |z[ < r u r2u (3.2.20)
A lth o u g h this p ro p e rty will n o t be used im m e d ia te ly , it will p ro v e u se fu l later, esp ecially in o u r tr e a tm e n t o f filter design b ased o n th e w in d o w te c h n iq u e , w h ere w e m u ltip ly th e im p u lse re sp o n s e o f an IIR system by a fin ite -d u ra tio n w in d o w w h ich se rv es to tru n c a te th e im p u lse re sp o n se o f th e I I R sy stem . F o r c o m p le x -v a lu e d se q u e n c e s .ti(rt) an d x 2 (n) w e c a n define th e p ro d u c t s e q u e n c e as x( n ) = Jti (nj xî n) . T h e n th e c o rre sp o n d in g c o m p lex co n v o lu tio n in te g ra l b eco m es x ( n ) = x i ( n ) x 2 {n) X(z) = v~ldv (3.2.21)
T h e p r o o f o f (3.2.21) is left as an ex ercise fo r th e re a d e r.
Parsevals relation.
y
If jti(n ) an d x 2 (n) a re c o m p le x -v a lu e d se q u en ces, th e n v ~ 'd v (3.2.22)
* i(n )x 2 (n) = l j ( j ^ X i ( v ) X 2
p ro v id e d th a t r ^ r y < 1 < n ur2u, w h e re ry < |z| < r \ u a n d r ^ < |z| < r2u a re th e R O C o f X ](z) a n d X 2(z). T h e p ro o f o f (3.2.22) follow s im m e d ia te ly by ev alu atin g X ( z ) in (3.2.21) at z = 1.
172
Chap. 3
The Initial Value Theorem.
If j ( ;i ) is causal [i.e.. x ( n ) = 0 fo r n < 0], th e n (3.2.23)
.r(0) = lim X ( z ) : * sc Proof. Since x{n) is causal. (3.1.1) gives X (z) y x {n )z x (0) + x (1 )z 1 + x (2 )z =o +
O b v iou sly , as z oc. z ~" 0 since n > 0 an d (3.2.23) follow s. A ll th e p ro p e rtie s o f th e z -tra n sfo rm p r e s e n te d in th is s e c tio n are su m m arized in T a b le 3.2 fo r easy re fe re n c e . T h ey are listed in th e sam e o rd e r as th ey have b een in tro d u c e d in th e tex t. T h e c o n ju g atio n p ro p e rtie s a n d P a rse v a l's relatio n are left as ex ercise s fo r th e re a d e r. W e have now' d e riv e d m o st o f th e z -tra n sfo rm s th a t are e n c o u n te re d in m any p ractical ap p licatio n s. T h e se z -tra n sfo rm pairs a re su m m a riz e d in T ab le 3.3 for easy re fe re n c e . A sim ple in sp e ctio n o f th is ta b le show s th a t th e s e z-tran sfo rm s are all rational f u n c t i o n s (i.e., ratio s o f p o ly n o m ials in z _1). A s will soon b ecom e a p p a re n t, ratio n al z -tra n sfo rm s are e n c o u n te re d n o t only as th e z-tran sfo rm s of v ario u s im p o rta n t signals b u t also in th e c h a ra c te riz a tio n of d isc re te -tim e lin ear tim e -in v a ria n t sy stem s d esc rib e d by c o n s tan t-co efficien t d iffe re n c e e q u a tio n s.
3.3 RATIONAL Z-TRANSFORMS

A s in d ic a te d in S ectio n 3.2, an im p o rta n t fam ily o f z-tra n sfo rm s a re th o se fo r w hich X ( z ) is a ra tio n a l fu n ctio n , th a t is. a ra tio of tw o p o ly n o m ials in z _l (o r z). In this sectio n w e discuss som e very im p o rta n t issues re g a rd in g th e class o f ra tio n a l z-tran sfo rm s.
3.3.1 Poles and Zeros

T h e zeros of a z -tra n sfo rm X (z) a re th e v alu es of z fo r w hich X (z) = 0. T h e pol es o f a z-tra n sfo rm are th e v alu es o f z fo r w hich X (z) = oc. If X ( z ) is a ratio n al fu n ctio n , th e n
J K ,) . D( z) ao + a i z 1 + ----- \ - aNz ~ h
= --------* k=o
(3.3.1)
If ao / 0 an d bo ^ 0, w e can avoid th e n eg ativ e p o w e rs o f z by fa cto rin g o u t the te rm s boz~M a n d a$z~N as follow s:
X(z)
N( z)
=
b 0z ~ M z M + {bx/ bo) z M~ x + + b M/ b 0 a0z N Z N + ( a \ / a a) z fJ~ l H -------- \ - aN/ a Q
D( z)
TABLE 3.2
PROPERTIES OF THE Z-TRANSFORM
Property Notation
Time Domain *(n) xt (n) ^(n) aixy(n) + a2x 2(n) x{n - k) a"x(n) x(~rj) x'(rt) Relx(n)} lm{x(n)) nx(n) x i ( n ) * x 2(n) rx,x2(l) = * i(0 * x 2 (~l) If x(n) causal Xi(n)x 2(n) = 2? i i
z-Domain X(z) Xi(z) X 2(z) <t\* l(z ) + 2X 2(z) z~kX(z) X(a~' z) X(z~') X' ( z ' ) i[X (z) + X*U*)] i[X (z )-* (; )] dX( z ) Z dz Jf|(z)X 2(z) / W z ) = Xi<z)* 2<z_l) .r(O) = lim X(z) 2~J^X,(v)X2^ J v - ' d v
R OC ROC: r2 < |z) < r\ ROC, ROC2 At least the intersection of R O Q and R O C 2 T hat of X (z), except z = 0 if k > 0 and z = oo if k < 0 \a\r2 < |z| < |a|r[ -
Linearity Time shifting Scaling in the z-domain Time reversal Conjugation Real part Imaginary part Differentiation in the z-domain Convolution Correlation Initial value theorem Multiplication Parsevals relation
ROC Includes ROC Includes ROC r2 < |z| < r. A t least, the intersection of ROCi and R O C 2 A t least, the intersection of ROC of X,(z) and -K^z 1)
r1
< |z| <
r2
A t least rj/ry < |z| < r]r2 a
5X ,( t)) ^ (l/tJ ) v - ' dv
174

TABLE 3.3 SOME COMMON Z-TRANSFORM PAIRS : -Transform, X (c) 1 1 1 - r-1 1 1 az~l 1:1 > 1 ROC A ll ;
Chap. 3
Signal. x(n) 1 2 3 4 5 6 7 8
5{n) u(n ) a"u(n) na"u(n ) au(n 1) na'u(n 1)

(cos aion)u(n) (sin a>nn)u(n)
( a" cosiH ) n ) u ( n )
\z\ > \a\ Izt > !a|

k l < la|
a:" 1 ( l - a z ^ )2
1 1 a : -1
a z~l
(1 - c : - 1)2
1 - Z '1 COSfô
|;| < N
kl > 1
1 - 2; _1 costal + z ~2
sin
1-
2 ;" 1 cos tun + z ~2 1 az~] cos a*)
\z\ > 1
10
(a'1sini*in)u{n)
1 2az~ cos wo + a 2z ~2 a ; -1 sinaiii 1-
2a ; " 1 cos a*,
a 2z ~2
Since N ( z ) and D( z ) are polynom ials in z, they can be expressed in factored form as
X (-)
b() - - M + N
D( z ) u
~ ~ Z2) (Z ~ (z - p i ) ( z - P i ) (z -
Zm )
p n
n " z*} X ( z ) = G z N~ MÂ -----------*=i
(3.3.2)
n<* ~ pjt)
w here G = 6o/ao- Thus X (r) has M finite zeros at z = z\, zi , . , z m (th e roots o f the num erator p olyn om ial), N finite p oles at z = p \ , p i .........P n (th e roots o f the d enom inator p olyn om ial), and \N M\ zeros (if N > M ) or p o le s (if N < M ) at the origin z = 0. P o les or zeros m ay also occur at z = 0 0 . A zero exists at z = oc if X ( 0 0 ) = 0 and a p o le exists at z = oc if X ( 0 0 ) = oc. If w e count the p oles and zeros at zero and infinity, w e find that X (z) has exactly the sam e num ber o f p o les as zeros. W e can represent X ( z ) graphically by a p o l e - z e r o p l o t (or pat t ern) in the com plex plane, which show s the location o f p oles by crosses ( x ) and the location o f zeros by circles (o). T he m ultiplicity of m ultipie*order p o les or zeros is indicated by a num ber clo se to the corresponding cross or circle. O b viou sly, by definition, the R O C o f a z-transform should not contain any poles.
Sec. 3.3
Rational z-Transform s
175
Exam ple 33.1 D eterm ine the pole-zero plot for the signal x(n) = a"u(n) Solution From Table 3.3 we find that X(z) = a >0
1 _____
1 a ; -1 - a
ROC; \z\ > a
Thus has one zero at n = 0 and one pole at pi = a. The pole-zero plot is shown in Fig. 3.7. Note that the pole p\ = a is not included in the R O C since the z-transform does not converge at a pole.
Re(;)
Figure 3.7
Pole-zero plot for the causal exponential signal ,v() = a"imu
Exam ple 3 3 .2 D eterm ine the pole-zero plot for the signal
1 0,
where a > 0. Solution
Of n < W - 1 elsewhere
From the definition (3.1.1) we obtain ^ X(z) = X W " ln

1}" = - r r ^ ------ r
l - t a ; ' 1) " z - a u T T t --------T 1 - az 1 = S z (z - a)
Since a > 0, the equation z M = a M has M roots at

Z t = a e ' 1* * ' " it = 0 , 1 . . . . . . M 1
The zero zo = a cancels the pole at z = a. Thus X(z) = ---------------- ^ - ---------------w ,

(Z - Z\ ) ( Z - Zl ) (Z - Z j t f - i )
which has M 1 zeros and M - 1 poles, located as shown in Fig. 3.8 for M = 8. Note th at the R O C is the entire z-plane except z = 0 because of the M - 1 poles located at the origin.
176
The z-Transfonm and Its Application to the Analysis of LTI Systems
Chap. 3
Im(c)
Red)
Figure 3.8 Pole-zero patlem for the finite-duration signal x(n ) = a", 0 < n < M l(a > 0). for M = 8. C learly , if we are given a p o le - z e r o p lo t, w e can d e te rm in e X ( ), by using (3.3.2), to w ithin a scaling fa c to r G. T his is illu stra te d in th e follow ing ex am p le. Example 3.3.3 Determine the c-transform and the signal that corresponds to the pole-zero plot of Fiji. 3.9. Solution There are two zeros (M = 2) at Z] = 0, Z2 = r cosojo and two poles (N = 2) al p] reJ , Pz = re~iw". By substitution of these relations into (3.3.2), we obtain X(z) = ( (; - Pi)(c - pi) = G(c - re}w u)(z - re~>w ")
Z(Z r C O S ttH i)
ROC:
A fter some simple algebraic manipulations, we obtain Xiz) = G-. 1 - 2 r - ~ l costou + r1 From Table 3.3 we find that x(n) = G(r" cosa>on)u(n) F ro m E x a m p le 3.3.3, w e see th a t th e p r o d u c t (; p \ ) ( z P 2 ) resu lts in a p o ly n o m ial w ith re a l coefficients, w h en p\ a n d p 2 a re c o m p lex co n ju g ates. In Im(2)
C O S a><)
ROC: 1-1 > r
Figure 3.9 P ole-zero p attern for Exam ple 3.3.3.
Sec. 3.3
Rational z - T ransforms
177
g e n e ra l, if a p o ly n o m ial h a s re a l coefficients, its ro o ts a re e ith e r real o r o ccu r in c o m p lex -co n ju g ate pairs. A s w e h av e seen , th e ; -tran sfo rm X ( z ) is a com plex fu n c tio n of th e com plex v ariab le z = R e ( z ) + j Im (r). O b v io u sly , |X (c )|, th e m a g n itu d e o f X ( z ), is a real and p o sitiv e fu n ctio n o f c. S ince : re p re se n ts a p o in t in th e co m p lex p la n e , |X ( )| is a tw o -d im e n sio n a l fu n c tio n an d d esc rib e s a su rfa c e . T h is is illu stra te d in Fig. 3.10(a) fo r th e z -tra n sfo rm
7 - 1 _ - - 2
(a)
Figure 3.10 Graph of |X (;)| for the ;;-Transform in (3.3.3). [Reproduced with permission fr om Introduction to Systems Analysis, by T. H. Glisson, 1985 by McGraw-Hill B ook Company.]
178
Chap. 3
w hich h as o n e z e ro a t z\ = 1 a n d tw o p o le s a t p \ , p 2 = 0. 9e J" /4. N o te the high p e a k s n e a r th e sin g u larities (p o les) a n d th e d e e p v alley close to th e zero. F ig u re 3.10(b) illu stra te s th e g rap h o f lX (z)j fo r z = eJO > .
3.3.2 Pole Location and Time-Domain Behavior for Causal Signals

In th is su b sectio n w e c o n sid e r th e re la tio n b e tw e e n th e z -p la n e lo catio n o f a pole p a ir a n d th e fo rm (sh a p e ) o f th e c o rre sp o n d in g signal in th e tim e d o m ain . T h e dis cu ssio n is b a s e d g en erally o n th e collectio n o f z -tra n s fo rm p a irs given in T a b le 3.3 an d th e resu lts in th e p re c e d in g su b sectio n . W e d e a l exclusively w ith real, causal signals. In p a rtic u la r, w e se e th a t th e c h a ra c te ristic b e h a v io r o f cau sal signals d e p e n d s o n w h e th e r th e p o les o f th e tra n sfo rm a re c o n ta in e d in th e reg io n |z| < 1 , o r in th e reg io n |z| > 1, o r on th e circle |z| = 1. S ince th e circle jz| = 1 h as a ra d iu s o f 1 , it is called th e unit circle. If a real signal h as a z-tran sfo rm w ith o n e p o le , th is p o le h a s to b e real. T he o n ly su ch signal is th e re a l e x p o n e n tia l x ( n ) = a nu{n) ^ X ( z ) = -- ---- r 1 az R O C : |z] > |o|
h av in g o n e z e ro a t zi = 0 an d o n e p o le a t pi a on th e re a l axis. F ig u re 3.11
x (n) x(n )
1 1
1 1 ! TTt
x(n)
x (n 0 i l l I I I r t
x(n)
'1
n i
1 -oi '
Figure 3.11 Time-domain behavior of a single-real pole causal signal as a function of the location of the pole with respect to the unit circle.
Sec. 3.3
179
illu stra te s th e b e h a v io r o f th e signal w ith re sp e c t to th e lo c a tio n o f th e p o le r e l ative to th e u n it circle. T h e signal is d ecay in g if th e p o le is inside the unit circle, fixed if th e p o le is o n th e u n it circle, an d gro w in g if th e p o le is o u t side th e u n it circle. In a d d itio n , a n eg ativ e p o le resu lts in a signal th a t a lte r n a te s in sign. O b v io u sly , causal signals w ith p o les o u tsid e th e u n it circle b e co m e u n b o u n d e d , cau se overflow in d igital system s, a n d in g e n e ra l, sh o u ld be av o id ed . A cau sal re a l signal w ith a d o u b le re a l p o le h as th e fo rm
x ( n ) = n a nu ( n )
(see T a b le 3.3) an d its b e h a v io r is illu stra te d in Fig. 3.12. N o te th a t in c o n tra s t to th e sin g le-p o le signal, a d o u b le real p o le o n th e u n it circle re su lts in an u n b o u n d e d signal. F ig u re 3.13 illu stra te s th e case of a p a ir o f c o m p le x -c o n ju g a te poles. A c c o rd ing to T a b le 3.3, th is co n fig u ratio n of p o le s resu lts in an e x p o n e n tia lly w eig h ted sin u so id al signal. T h e d istan c e r of th e p o les fro m th e o rigin d e te rm in e s th e e n v e lope o f th e sin u so id al signal an d th e ir angle w ith th e real p o sitiv e axis, its relative freq u e n cy . N o te th a t th e a m p litu d e o f th e signal is gro w in g if r > 1, c o n stan t if r 1 (sin u so id a l sig n als), a n d d ecaying if r < 1 .
x(n)
<
T T ....
xin)
. 0
T 1
Figure 3.12 Time-domain behavior of causal signals corresponding to a double (m = 2) real pole, as a function of the pole location.
180
Chap. 3
Figw e 3.13 A pair of complex-conjugate poles corresponds to causal signals with oscillatory behavior.
Finally, Fig. 3.14 show s the b eh avior o f a causal signal w ith a d ou b le pair of p o les on the unit circle. T his reinforces the corresponding results in Fig. 3.12 and illustrates that m ultiple p o les on the unit circle should b e treated with great care. T o sum m arize, causal real signals w ith sim ple real p oles or sim ple com plexconjugate pairs o f p oles, w hich are in sid e or on the unit circle are always bounded in am plitude. Furtherm ore, a signal with a p o le (or a com p lex-con ju gate pair o f p o les) near the origin decays m ore rapidly than on e a ssociated with a pole n ear (but inside) the unit circle. Thus the tim e behavior o f a signal depends strongly on the location o f its p o les relative to th e unit circle. Z eros also af fect the behavior o f a signal but n ot as strongly as p oles. F or exam ple, in the
Sec. 3.3
181
Figure 3.14 Causal signal corresponding to a double pair of complex-conjugate poles on the unit circle. case o f sin u so id al signals, th e p resen c e an d lo catio n o f ze ro s affects only th e ir p h ase. A t this p o in t, it sh o u ld be stressed th a t e v e ry th in g w e h av e said a b o u t causal signals ap p lies as w ell to causal L TI system s, since th e ir im p u lse re sp o n se is a causal signal. H en ce if a p o le of a system is o u tsid e the unit circle, th e im pulse resp o n se of th e system b eco m es u n b o u n d e d and. co n se q u e n tly , th e sy stem is u n sta b le .
3.3.3 The System Function of a Linear Time-Invariant System

In C h a p te r 2 w e d e m o n s tra te d th a t th e o u tp u t o f a (re la x e d ) lin e a r tim e -in v a ria n t sy stem to an in p u t se q u e n c e x( n) can be o b ta in e d by co m p u tin g th e co n v o lu tio n of jc(h) w ith th e u n it sa m p le resp o n se o f th e system . T h e co n v o lu tio n p ro p e rty , d eriv e d in S ectio n 3.2, allow s us to ex p ress th is re la tio n sh ip in th e z-d o m ain as Y( z ) = m z ) X ( z ) (3.3.4)
w h ere Y( z ) is th e z -tra n sfo rm o f th e o u tp u t se q u en ce v(n), X (z) is th e z-tra n sfo rm o f th e in p u t se q u e n c e x ( n ) an d H( z ) is th e z -tra n sfo rm o f th e u n it sa m p le resp o n se h{n). If w e k n o w h( n) an d x( n ) , w e can d e te rm in e th e ir c o rre sp o n d in g z-tra n sfo rm s H ( z ) a nd X ( z ) , m u ltip ly th e m to o b ta in Y(z), a n d th e r e fo r e d e te rm in e y( n) by e v a lu a tin g th e in v erse z -tra n sfo rm of K(z). A lte rn a tiv e ly , if w e k now x ( n ) an d we o b se rv e th e o u tp u t y( n) of th e system , w e can d e te rm in e th e u n it sa m p le resp o n se by first solv in g fo r H( z ) fro m th e re la tio n
a n d th e n e v a lu a tin g th e in v erse z -tra n sfo rm o f H( z ) . Since

OC
(3.3.6)
182
The z-Transform and Its Application to the Analysis of LTI System s
Chap. 3
it is c le a r th a t H ( z ) re p re s e n ts th e z -d o m ain c h a ra c te riz a tio n o f a system , w h ereas h{n) is th e c o rre sp o n d in g tim e -d o m a in c h a ra c te riz a tio n of th e system . In o th e r w o rd s, H{z ) an d h( n) are e q u iv a le n t d e s c rip tio n s o f a system in th e tw o d o m ain s. T h e tra n sfo rm H{ z ) is called th e s yst em f unct i on. T h e re la tio n in (3.3.5) is p a rtic u la rly u sefu l in o b ta in in g H{ z ) w h en th e system is d e s c rib e d b y a lin e a r c o n s tan t-co efficien t d iffe re n c e e q u a tio n o f th e fo rm M N (3.3.7) In th is case th e sy stem fu n ctio n can be d e te rm in e d d irectly fro m (3.3.7) by com p u tin g th e z -tra n s fo rm o f b o th sid es o f (3.3.7). T h u s, by a p p ly in g th e tim e-sh iftin g p ro p e rty , w e o b ta in M
y u
) = - J 2 a*Y ^ z ~k + E
b iX (z )r
M Y(z) X(z)
o r, eq u iv alen tly , > z-* (3.3.8)
T h e re fo re , a lin e a r tim e -in v a ria n t system d e s c rib e d by a c o n s ta n t-c o e ffic ie n t dif fere n ce e q u a tio n h a s a ra tio n a l system fu n ctio n . T h is is th e g e n e ra l fo rm fo r th e system fu n c tio n o f a sy ste m d e sc rib e d by a lin e a r c o n s tan t-co efficien t d ifferen ce e q u a tio n . F ro m th is g e n e ra l fo rm w e o b tain tw o im p o rta n t sp e cial fo rm s. F irst, if ajt = 0 fo r 1 < k < N , (3.3.8) re d u c e s to M H ( z ) = J 2 bkz ~k
k=0
(3.3.9)
In th is case, H ( z ) c o n ta in s M z ero s, w h o se valu es a re d e te r m in e d by the system p a ra m e te rs {*}, a n d an M th -o rd e r p o le a t th e o rigin z = 0. S ince the sy stem c o n ta in s o n ly trivial p o le s (a t z = 0) an d M n o n triv ia l z e ro s, it is called
Sec. 3.3
183
an all -zero syst em. C learly, such a system h as a fin ite -d u ra tio n im p u lse resp o n se (F IR ), a n d it is called an F IR system o r a m oving av erag e (M A ) system . O n th e o th e r h an d , if bk 0 fo r 1 < k < M, th e system fu n c tio n re d u c e s to
A'
*= 1 boz" at*.
k
(3.3.10) <io = l
In this case H ( z ) co n sists of N poles, w hose v alu es are d e te r m in e d by th e system p a ra m e te rs {a*} a n d an /V th-order zero at th e o rig in z = 0- W e usually do not m a k e re fe re n c e to th e se trivial zeros. C o n se q u e n tly , th e sy stem fu n c tio n in (3.3.10) c o n ta in s o n ly n o n triv ia l p o les an d th e c o rre sp o n d in g sy stem is called an all-pole syst em. D u e to th e p re se n c e o f poles, th e im p u lse re sp o n s e o f such a system is infinite in d u ra tio n , a n d h e n c e it is an I IR system . T h e g e n e ra l fo rm o f th e sy stem fun ctio n given by (3.3.8) c o n ta in s b o th poles a n d zero s, and h e n c e th e c o rre sp o n d in g sy stem is called a p o l e - z e r o s y s t e m , with N p o le s a n d M z ero s. P o les a n d /o r zero s at c = 0 an d z = oc a re im p lied b u t are n o t c o u n te d ex p licitly. D u e to th e p re se n c e o f p o les, a p o le - z e r o system is an IIR system . T h e fo llo w in g ex am p le illu strates th e p ro c e d u re fo r d e te rm in in g th e system fu n ctio n a n d th e u n it sa m p le resp o n se from the d iffe re n c e e q u a tio n . E xam ple 3.3.4 D eterm ine the system function and the unit sample response of the system described by the difference equation
v(n) =
1y(n -
1) + 2 x ( n )
Solution
By computing the -transform of the difference equation, we obtain Y(z) = z - > Y(z) + 2X(z)
H ence the system function is
xt:l
I - Jz -1
This system has a pole at z = \ and a zero at the origin. Using Table 3.3 we obtain the inverse transform h(n) = 2(i)"u()
This is the unit sample response of the system.
184
The z-T ransform and Its Application to the Analysis of LTI Systems
Chap. 3
W e h av e n o w d e m o n s tra te d th a t ra tio n a l z -tra n s fo rm s a re e n c o u n te re d in c o m m o n ly u se d sy stem s a n d in th e c h a ra c te riz a tio n o f lin e a r tim e -in v a ria n t sys tem s. In S ectio n 3.4 w e d esc rib e se v e ra l m e th o d s fo r d e te rm in in g th e inverse z -tra n sfo rm o f ra tio n a l fu n ctio n s.
3.4 INVERSION OF THE Z-TRANSFORM

A s w e saw in S ectio n 3.1.2, th e in v erse z -tra n s fo rm is fo rm ally given by x ( n ) = - < x ( z ) z n~ 1d z 2 njjt (3.4.1)
w h ere th e in te g ra l is a c o n to u r in te g ra l o v e r a clo sed p a th C th a t en clo ses the o rig in an d lies w ith in th e reg io n o f c o n v e rg e n c e o f ^ ( z ) . F o r sim plicity, C can be ta k e n as a circle in th e R O C o f X (z) in th e z-p lan e. T h e re a re th re e m e th o d s th a t a re o fte n u se d fo r th e e v a lu a tio n o f th e inverse z-tran sfo rm in practice: 1. D ire c t e v a lu a tio n o f (3.4.1), by c o n to u r in te g ra tio n . 2. E x p a n sio n in to a se rie s o f te rm s, in th e v a ria b le s z, an d z _1. 3. P a rtia l-fra c tio n ex p an sio n a n d ta b le lo o k u p .
3.4.1 The Inverse z-Transform by Contour Integration

In th is se ctio n w e d e m o n s tra te th e use o f th e C au ch y re sid u e th e o r e m to d e te rm in e th e in v erse z -tra n sfo rm d irectly fro m th e c o n to u r in teg ral.
Cauchy residue theorem. L e t / ( z ) b e a fu n c tio n o f th e c o m p lex v ariab le z an d C b e a clo sed p a th in th e z -p lan e. If th e d e riv a tiv e d f ( z ) / d z exists o n and inside th e c o n to u r C a n d if / ( z ) has no p o le s a t z = zo, th e n
2njjcz-zo
- L (f) J ! d z
= |^ < Z l,)' 10,
if zo is ou tsid e C
(3.4.2)
M o re g en erally , if th e (k + l ) - o r d e r d e riv a tiv e o f / ( z) exists a n d / ( z ) h a s n o p o les a t z = zo, th e n 1 d k- ' f ( z )
-!>'
0,
C
if zo is o u ts id e C
04.3)
T h e v alu es o n th e rig h t-h a n d sid e o f (3.4.2) a n d (3.4.3) a re c a lle d th e re sid u e s of th e p o le a t z = zo- T h e re su lts in (3.4.2) a n d (3.4.3) a re tw o fo rm s o f th e Cauc hy residue t heorem. W e can ap p ly (3.4.2) a n d (3.4.3) to o b ta in th e v a lu e s o f m o re g en eral c o n to u r in teg rals. T o b e specific, su p p o s e th a t th e in te g ra n d o f th e c o n to u r in te g ra l is
Sec. 3.4
Inversion of th e
2 - T ra n s fo rm
185
P(z) = f ( z ) / g ( z ) ~ w h e re f ( z ) h a s no p o les inside th e c o n to u r C an d g (z) is a p o ly n o m ial w ith d istin ct (sim p le) ro o ts c i, ^ 2 . ___-n inside C. T h e n
(3.4.4)
n
1=1
w h ere f(z) (3.4.5) A l (z) = ( z - z i ) P{ z ) = ( z - z l) J - 1x g( z) T h e v alu es (A, (-;,)} a re re sid u e s o f th e c o rre sp o n d in g p o le s at z = / = 1, 2 , . . . . n. H e n c e th e v alu e o f th e c o n to u r in te g ra l is e q u a l to th e sum o f th e resid u es o f all th e p o le s in sid e th e c o n to u r C. W e o b se rv e th a t (3.4.4) w as o b ta in e d by p e rfo rm in g a p a rtia l-fra c tio n e x p a n sion o f th e in te g ra n d an d ap p ly in g (3.4.2). W h en g(z) has m u ltip le -o rd e r ro o ts as w ell as sim p le ro o ts inside th e c o n to u r, th e p a rtia l-fra c tio n e x p a n sio n , w ith a p p ro p ria te m o d ifica tio n s, an d (3.4.3) can b e used to e v a lu a te th e resid u es at th e c o rre sp o n d in g p o les. In th e case o f th e in v erse z -tra n sfo rm , w e h ave
[resid u e of X (z )z n 1 a t z
a ll p o l e s U t) in s id e C
(3.4.6)
p ro v id e d th a t th e p o les {z,} a re sim ple. If X (z )r "_1 has no p o le s inside th e c o n to u r C fo r o n e o r m o re v alu es o f n, th e n x (n) = 0 fo r th e s e values. T h e fo llo w in g ex am p le illu stra te s th e e v a lu a tio n o f th e in v erse z-tra n sfo rm by u se o f th e C a u ch y re sid u e th e o re m . Exam ple 3.4.1 Evaluate the inverse z-transform of X(z) = ---------- r 1 az~ using the complex inversion integral. Solution We have kl > kil
where C is a circle at radius greater than |a|. We shall evaluate this integral using (3.4.2) with f ( z ) = z". We distinguish two cases.
186
The ^-Transform and Its Application to the Analysis of LTI Systems
Chap. 3
L If n > 0, f ( z ) has only zeros and hence no poles inside C. T he only pole inside C is z = a. Hence *(*) = /(zo) = a n
n > 0
2. If n < 0, / ( z ) = z" has an nth-order pole at z = 0, which is also inside C. Thus there are contributions from both poles. For n = - 1 we have x (-l) = If n = 2, we have
1 S) 1 dz = 1 2 nj j c z ( z a) z a
+ -
= 0
2)
2 n j z H z - a ) dZ
dz ( z - f l )
= 0
By continuing in the same way we can show that *(n) = 0 for n < 0. Thus x (n) = a"u(n)
3.4.2 The Inverse z-Transform by Power Series Expansion

The basic idea in this m eth od is the follow ing: G iven a z-transform X ( z ) with its corresponding R O C , w e can expand X (z) into a p ow er series o f the form
OO
X(z) =
co
cz~n
(3.4.7)
w hich con verges in the given R O C . T h en , by the u n iqu en ess o f the z-transform, x ( n ) = c for all n. W hen X ( z ) is rational, the exp an sion can b e perform ed by long division. T o illustrate this tech n iqu e, w e w ill invert som e z-transform s involving the sam e expression for X ( z ) , but different R O C . T his w ill also serve to em phasize again the im portance o f the R O C in d ealing with z-transform s. Exam ple 3A 2 D eterm ine the inverse z-transform of 1 1.5z_1 + 0.5z 2 when (a) ROC: |z| > 1 (b) ROC: |z| < 0.5 Solution (a) Since the R O C is the exterior of a circle, we expect x(n) to be a causal signal. Thus we seek a power series expansion in negative powers of z. By dividing
Sec. 3.4
Inversion of the z-Transform
187
the num erator of X{z) by its denom inator, we obtain the power series A UI = ! _ 3 _ -; +
= 1+
+ TE; "4 + " '
By com paring this relation with (3.1.1), we conclude that
Note that in each step of the long-division process, we eliminate the lowestpower term of c~*. (b) In this case the ROC is the interior of a circle. Consequently, the signal x(n) is anticausal. To obtain a power series expansion in positive powers of c. we perform the long division in the following way: 2: 2 + 6c3 + 14c4 + 30cs + 62c* + + ill 1 - 3: + 2c: 3c - 2z 2 3c - 9c: + 6c3 l z 2 - 6c3 7 r - 21 c3 + 14c4 15c3 - 14c4 15c3 - 45c4 + 30cs 31c4 - 30c5 Thus
X (c) =
1_
------:------= 2c: + 6c3 + 14c4 + 30c5 + 62cfi + 1
In this case x(n) = 0 for n > 0. By comparing this result to (3.1.1), we conclude that Jt(n) = { 62. 30. 14.6,2, 0. 0}
t
We observe that in each step of the long-division process, the lowest-power term of c is eliminated. We emphasize that in the case of anticausal sig nals we simply carry out the long division by writing down the two poly nomials in reverse order (i.e., starting with the most negative term on the left).
F ro m th is e x a m p le w e n o te th a t, in g e n eral, th e m e th o d o f long d ivision will n o t p ro v id e a n sw ers fo r x( n) w h en n is larg e b e c a u se th e lo n g division b eco m es ted io u s. A lth o u g h , th e m e th o d p ro v id es a d irect e v a lu a tio n o f x( n ) , a clo sed -fo rm so lu tio n is n o t p o ssib le , ex cep t if th e resu ltin g p a tte r n is sim p le e n o u g h to infer th e g e n e ra l te rm x ( n ) . H e n c e th is m e th o d is used only if o n e w ish e d to d e te rm in e th e v a lu e s o f th e first few sa m p le s o f th e signal.
188
Chap. 3
Example 3.4.3 Determ ine the inverse z-transform of X(z) = log(l + az_1) Solution |zj > |a|
+ x ),
Using the power series expansion for log(l ^ ( - r v r "
with
|jc| < 1,
we have
Thus
0,
n < 0
Expansion of irrational functions into power series can be obtained from tables.
3.4.3 The Inverse z-Transform by Partial-Fraction Expansion

In th e tab le lo o k u p m e th o d , w e a tte m p t to e x p ress th e fu n c tio n X (z) as a linear c o m b in a tio n X ( z ) = a j X] (;) + 02 X 2 ( 1 ) + - - + c/Kx fc(z) (3.4.8)
w h ere X] ( ; ) , . . . . X k (z ) a re ex p ressio n s w ith in v erse tra n sfo rm s x \ ( n ) , . . . , x k W av ailab le in a ta b le o f z-tra n sfo rm p airs. If such a d e c o m p o s itio n is possible, th e n x( n) , th e in v erse z -tra n sfo rm o f X ( z ) , can easily b e fo u n d using th e lin earity p ro p e rty as x ( n) = at]Xi(n) + a 2x 2 (n) H -------- \ - a Kx K (n) (3.4.9)
T h is a p p ro a c h is p a rtic u la rly useful if X ( z ) is a ra tio n a l fu n ctio n , a s in (3.3.1). W ith o u t loss o f g e n erality , w e assu m e th a t ao = 1, so th a t (3.3.1) can b e e x p ressed as _ t o D(z) _ b +b,r' + + t r
l+ fliZ - 1 H ----- + a u z ~ N
N o te th a t if a 0 ^ 1. w e can o b ta in (3.4.10) fro m (3.3.1) by div id in g b o th n u m e ra to r an d d e n o m in a to r by aoA ra tio n a l fu n ctio n o f th e form (3.4.10) is called p r o p e r if a N ^ 0 an d M < N. F ro m (3.3.2) it fo llo w s th a t th is is e q u iv a le n t to saying th a t th e n u m b e r o f finite zero s is less th a n th e n u m b e r o f fin ite p o les. A n im p ro p e r ra tio n a l fu n ctio n ( M > N ) can alw ays b e w ritte n as th e sum of a p o ly n o m ial an d a p r o p e r ra tio n a l fu n ctio n . T h is p ro c e d u re is illu stra te d by the fo llow ing ex am p le. Example 3.4.4 Express the im proper rational transform l + 3 z - ' + n z - 2 + l 2-3 + in terms of a polynomial and a proper function.
1
5
1 + 6Z
Sec. 3.4
Inversion of the z-Transform
189
Solution First, we note that we should reduce the num erator so that the term s ; -2 and c- * ' are eliminated. Thus we should carry out the long division with these two polynomials written in reverse order. We stop the division when the order of the rem ainder becomes Then we obtain ^ = 1 + 2: - i +
In g e n e ra l, an y im p ro p e r ra tio n a l fu n ctio n (M > N ) can b e e x p ressed as X ( ;) = W i = Co + C i:" 1 + ' ' ' + Cu~n Z ~im~N) + (3 A ll)
T h e in v erse z -tra n s fo rm o f the p o ly n o m ial can easily b e fo u n d by in sp ectio n . W e fo cu s o u r a tte n tio n on th e inversion o f p r o p e r ra tio n a l tra n sfo rm s, since any im p ro p e r fu n c tio n can b e tra n sfo rm e d in to a p ro p e r fu n c tio n by using (3.4.11). W e carry o u t th e d e v e lo p m e n t in tw o steps. F irst, we p e rfo rm a p a rtia l fra c tio n e x p a n sio n o f th e p r o p e r ra tio n a l fu n ctio n a n d th e n w e in v ert each o f th e term s. L e t A"(c) b e a p ro p e r ra tio n a l fu n ctio n , th a t is, * (.-) = = - " - - ' " T ,1 + +buZ^ D( z) 1 + ^ iC + * -t~ w h ere aN ^ 0 an d M < N (3.4.12)
T o sim p lify o u r discu ssion w e e lim in ate n eg ativ e p o w ers of c by m ultip ly in g b o th th e n u m e r a to r a n d d e n o m in a to r of (3,4,12) by z N. T h is resu lts in
L N I L _ y v -l I I L N -M
X{z) =
CA 4- a \ z N + ------(- <a /v
Since N > M , th e fu n ctio n ,N -2 _ i_____ l
(3.4.13,
w hich c o n ta in s only p o sitiv e p o w e rs o f
(3,4.14) ; : w + tiic w- 1 + - - - + flW is also alw ays p ro p e r. O u r ta sk in p e rfo rm in g a p a rtia l-fra c tio n ex p an sio n is to ex p ress (3.4.14) o r, e q u iv a le n tly , (3.4.12) as a sum o f sim ple fractio n s. F o r th is p u rp o se w e first fa c to r th e d e n o m in a to r p o ly n o m ial in (3.4.14) in to facto rs th a t c o n tain th e poles P i, p 2, . . . , p n o f X (z). W e d istinguish tw o cases.
Distinct poles. S u p p o se th a t th e p o le s p \ , p 2 ........ p/v a re all d iffe re n t (dis tin ct). T h e n w e se e k an ex p an sio n of th e fo rm
z z p\ z P2 + -. - + ( 3 . z Pn 4 . 1 5 )
T h e p ro b le m is to d e te rm in e th e coefficients A i , A 2 , . - . , A s - T h e re a re tw o w ays to so lv e th is p ro b le m , as illu stra te d in th e follow ing exam p le.
190
Chap. 3
Example 3.4 .5 D eterm ine the partial-fraction expansion of the proper function (3'U 6 , Solution First we elim inate the negative powers, by multiplying both num erator and denom inator by z2. Thus
.2
Xiz)= z V < + i0 n.5 2- 1 1.5z The poles of X(z) are p\ = 1 and P2 = 0.5. Consequently, the expansion of the form (3.4,15) is X(z) z C z l)(z 0.5) A\ z 1 A2 z 0.5 (3.4.17)
A very simple method to determ ine A[ and A 2 is to multiply the equation by the denom inator term (z - l)(z - 0.5). Thus we obtain z = ( z - 0 .5 M i + ( z - 1 ) A 2 (3.4.18)
Now if we set z = p\ = 1 in (3.4.18), we eliminate the term involving A2. Hence 1 = (1 -0 .5 )A , Thus we obtain the result A i = 2. Next we return to (3.4.18) and set z = p 2 = 0.5, thus eliminating the term involving Ai, so we have 0.5 = ( 0 .5 - 1)A2 and hence Ai = 1. Therefore, the result of the partial-fraction expansion is X(z) 2 z- 1 1 z - 0.5 (3.4.19)
T h e exam ple given ab ove suggests that w e can determ ine the coefficients A \, A i , . . . , Afj , by m ultiplying b oth sides o f (3.4.15) by each o f the term s (z - Pk ). k = 1 , 2 , , . . , N , and evaluating the resulting exp ression s at the corresp on d ing pole p osition s, p \ , p i .........P n T h u s w e have, in general,
(Z- Pl)X(;) = (z-wM.i+ ... + /lt+,..+ fa-P)^

z C onsequently, w ith z = Pk, Ak = Exam ple 3.4.6 Determ ine the partial-fraction expansion of 1 17- ^ 0 .5 ,z - PI z - Pn Jtth coefficient as k =
(3420)
(3.4.21)
(3.4.20) yield s the

z \z-pl
( z ~ ^ )X (; )i
1, 2. .
,3A22)
Sec. 3.4
inversion of the z-Transform
191
Solution To eliminate negative powers o f ; in (3.4.22), we multiply both num erator and denom inator by Thus X(z) z+ 1 The poles of X(z) are complex conjugates
Px = \ + J i
z2 - ; + 0.5
and
Pi
=
\ ~ j \
Since p\ ^ p 2- we seek an expansion of the form (3.4.15). Thus

* (:) ; + l
A]
Ai
(z-p i)(:-p 2)
z-P\
z-pi
To obtain A , and A2, we use the formula (3,4.21), Thus we obtain (z-pi)X(z) At = ----------( z - p 2)X(z) A- = -----------; + 1 I
- P
?+
2 \ ^ P1
3+ J5-5+./5
1 1
; + l
-
Pi
:= P 2
T h e ex p a n sio n (3,4.15) a n d the fo rm u la (3.4.21) h o ld fo r b o th real a n d c o m p lex p o les. T h e o n ly c o n s tra in t is th a t all p o les be d istin ct. W e also n o te th at A; = A*. It can b e easily seen th a t th is is a c o n s e q u e n c e o f th e fact th a t p 2 = p ' . In o th e r w o rd s, comp l ex- conj ugat e pol es result in c o mpl ex- con j ugat e coefficients in the part ial-fraction expansi on. T h is sim ple resu lt will p ro v e v ery u se fu l la te r in o u r d iscu ssio n . M u ltip le - o rd e r p o l e s . If X U) h a s a p o le o f m u ltip lic ity /, th a t is, it co n ta in s in its d e n o m in a to r th e fa c to r (z - pk )1, th e n th e e x p a n s io n (3.4.15) is n o lo n g er tru e. In th is case a d iffe re n t ex p an sio n is n e e d e d . F irst, w e in v e stig a te th e case of a d o u b le p o le (i.e., 1 = 2). Exam ple 3.4.7 D eterm ine the partial-fraction expansion of
Solution
First, we express (3.4,23) in terms of positive powers of
in the form
* ( ;) = z2 : (z + 1)(; - l ) 2 X (z) has a simple pole at p\ = - 1 and a double pole pi = p$ = 1. In such a case the appropriate partial-fraction expansion is
= ______ t ______ = + (3424)
z (z + l)(z - 1)2 z+ 1 z - l (z -1 )2 The problem is to determ ine the coefficients A 1, A2, and A3.
192
Chap. 3
We proceed as in the case of distinct poles. T o determ ine Aj, we multiply both sides of (3.4.24) by (z + 1) and evaluate the result at z = - 1 . Thus (3.4.24) becomes
------------------ = Ai + ------ - A 2 + ------- r - z A i
(z + l)X (z) z
z+ 1
z+ 1
z-1
( z - 1)2 1 4
which, when evaluated at z = - 1 , yields

A i = ------------------
(z + 1)X(z) z
Next, if we multiply both sides of (3.4.24) by (z - l ) 2, we obtain =

Z z+ 1
<3.4.25,
Now, if we evaluate (3.4.25) at z = 1, we obtain A 3. Thus (z l) 2X(z) I = 1
* Li 2 The remaining coefficient Az can be obtained by differentiating both sides of (3.4.25) with respect to z and evaluating the result at z = 1. Note that it is not necessary formally to carry out the differentiation of the right-hand side of (3.4.25), since all term s except A2 vanish when we set z = 1. Thus d A2 = dz (z l ) 2X(z)
T h e gen eralization o f the p rocedure in the exam p le ab ove to the case o f an /th-order p ole ( z p k)! is straightforward. T h e partial-fraction expansion must contain the terms Alt
Z - Pk ( z - Pk)2
Mk
(z - Pk)1
T h e coefficients {A,*} can b e evalu ated through d ifferen tiation as illustrated in E xam p le 3.4.7 for / = 2. N ow that w e have perform ed the partial-fraction exp an sion , w e are ready to take the final step in the inversion o f X (z). First, let us con sid er th e case in which X ( z ) contains distinct poles. F rom the partial-fraction exp an sion (3.4.15), it easily follow s that X ( z ) A \ -------------r + A i -------------- + -)- A n --------------1 P\Z~1 1 - PlZ 1 - P n Z~ 1
(3.4.27)
T he inverse z-transform , x ( n ) = Z ~ i { X( z ) ) , can b e ob tain ed by inverting each term in (3.4.27) and taking the corresponding linear com b ination . From T ab le 3.3 it fo llo w s that these term s can be inverted using th e form ula ( p k)nu{n), - ( P k ) nu ( - n - 1), if RO C : |*| > Ip*J (causal signals) if RO C: |z| < \P k \ (anticausal sign als)
Sec. 3.4
Inversion of the ^-Transform
193
If th e signal jc(n) is cau sal, the R O C is |r | > p max- w h e re p max = m ax{|/>i|, \p2\.........IpnII- In th is case all te rm s in (3.4.27) re su lt in causal signal c o m p o n e n ts an d th e signal x ( n ) is given by jr(/i) = (A i + A -Pz --------f" A Np n N )u(n) (3.4.29)
If all p o le s a re real. (3.4.29) is th e d e sire d e x p ressio n fo r th e signal j (). T h u s a causal sig n al, h av in g a ; -tra n s fo rm th a t co n tain s real a n d d istin ct p o les, is a linear c o m b in a tio n o f re a l e x p o n e n tia l signals. S u p p o se n o w th a t all p o les a re d istin ct b u t som e o f th e m a re co m p lex . In this case so m e o f th e te rm s in (3.4.27) re su lt in com plex e x p o n e n tia l c o m p o n en ts. H o w e v e r, if th e signal x ( n) is real, w e sh o u ld b e ab le to re d u c e th e se te rm s in to real c o m p o n e n ts. If x ( n ) is real, th e p o ly n o m ials a p p e a rin g in X (z) have real co efficients. In th is case, as w e h av e seen in S ectio n 3.3, if pj is a p o le , its com plex c o n ju g a te p j is also a p o le. A s w as d e m o n s tra te d in E x am p le 3.4.6, th e c o rre s p o n d ing coefficien ts in th e p a rtia l-fra c tio n e x p an sio n a re also co m p lex co n ju g ates. T h u s th e c o n trib u tio n o f tw o c o m p lex -co n ju g ate p o les is of th e form x k (n) = \ A k {pt )n + A H p l )"]() (3.4.30)
T h e se tw o te rm s can be c o m b in ed to form a real signal c o m p o n e n t. F irst, we ex p re ss Aj an d Pj in p o la r form (i.e., a m p litu d e an d p h a se ) as A* = \ Ak \eja' Pi: = (3.4.31) (3.4.32)
w h ere a k an d fik a re th e p h a s e c o m p o n e n ts o f A k a n d p k. S u b stitu tio n o f th ese re la tio n s in to (3.4.30) gives x k{n) = IAi or, eq u iv alen tly , x k(n) = 2|A *|r" c o s($ tn + a k)u(n) T h u s w e co n clu d e th a t Z - 1 ( - r + - ~ r ) = 2 \ A k \rR k c o s ( f tn + a k) u(n) \i - p kz ~ l i - p;z~ v (3.4.34) (3.4.33) + e - -'(A"+ l]u(n)
if th e R O C is |zj > \ pk \ = rk . F ro m (3.4.34) we o b se rv e th a t ea c h p a ir o f c o m p le x -c o n ju g a te p o le s in th e z -d o m ain resu lts in a causal sin u so id al signal c o m p o n e n t w ith an e x p o n e n tia l e n v elo p e. T h e d ista n c e rk o f th e p o le fro m th e o rig in d e te rm in e s th e e x p o n e n tia l w eig h tin g (g ro w in g if r k > 1, d ecay in g if r k < 1, c o n s ta n t if rk = 1). T h e angle of th e p o le s w ith re sp e c t to th e p o sitiv e re a l axis p ro v id e s th e fre q u e n c y o f th e sin u so id a l signal. T h e zero s, o r e q u iv alen tly th e n u m e ra to r o f th e ra tio n a l tran sfo rm , affect o n ly in d ire c tly th e a m p litu d e an d th e p h a se o f x k (n) th ro u g h A k. In th e case o f mul t i pl e p o les, e ith e r re a l o r co m p lex , th e in v erse tra n sfo rm o f te rm s o f th e fo rm A j ( z p k)n is re q u ire d . In th e case o f a d o u b le p o le the
194
The 2 -Transform and Its Application to the Analysis of LTI Systems
Chap. 3
fo llow ing tra n sfo rm p a ir (see T a b le 3.3) is q u ite useful: pz~x (1 - p z ] )2 = n p nu( n) (3.4.35)
p ro v id e d th a t th e R O C is |z| > \p\. T h e g e n e ra liz a tio n to th e case o f p o le s w ith h ig h e r m u ltip licity is left as an exercise fo r th e re a d e r. Example 3.4.8 Determ ine the inverse z-transform of X U) = 1 1 - 1 .5 ;-' +0.5z~ 2
(a) RO C III > 1 (b) RO C Izl < 0.5 (c) RO C 0.5 < |z| < 1 Solution This is the same problem that we treated in Exam ple 3.4.2. The partialfraction expansion for X(z) was determined in Example 3.4.5. The partial-fraction expansion of X(z) yields
<3-4-36>
To invert X(z) we should apply (3.4,28) for pi 1 and p 2 = 0.5. However, this requires the specification of the corresponding ROC.
(a) In case when the R O C is |z| > 1, the signal x(n) is causal and both term s in
(3.4.36) are causal terms. According to (3.4.28), we obtain x(n) = 2 (l)n(n) (0.5)"u(n) = (2 0.5 )u(n) which agrees with the result in Example 3.4.2(a). (3.4.37)
(b) When the R O C is
|z| < 0.5, the signal x(n) is anticausal. Thus both term s in (3.4.36) result in anticausal components. From (3.4.28) we obtain
x(n) = [2 + (0.5)'I]u( n 1)
(3.4.38)
(c) In this case the ROC 0.5 < |z| < 1 is a ring, which implies that the signal x(n) is two-sided. Thus one of the term s corresponds to a causal signal and the other to an anticausal signal. Obviously, the given ROC is the overlapping of the regions (z| > 0.5 and |z| < 1. Hence the pole p 2 = 0.5 provides the causal part and the pole p\ = 1 the anticausal. Thus x(n) = -2 (1 ) "u( - n - 1) - (0.5)"(n) (3.4.39)
Example 3.4.9 D eterm ine the causal signal jc(n) whose z-transform is given by
Sec, 3.4
Inversion of the z-Transtorm
195
Solution
In Exam ple 3.4.6 we have obtained the partial-fraction expansion as
where A, = Al = j - j
and
Pi = p = i Since we have a pair of complex-conjugate poles, we should use (3.4.34). The polar forms of Aj and p, are
Hence
Example 3.4.10 D eterm ine the causal signal x(n) having the ;-transiorm X(z) = Solution (1 + ; - )(] 3 + + T 41
.-1
From Example 3.4.7 we have X(Z) 41 + 2 ( 1 - ; - 1)2
By applying the inverse transform relations in (3.4.28) and (3.4.35), we obtain

1 3 1 f 1 x( n ) = - ( l)"(n) + t ( ) + - n u ( n ) = t ( - 1 )
3 w*l + - + u(n)
3.4.4 Decomposition of Rational z-Transforms

A t th is p o in t it is a p p ro p ria te to discuss som e a d d itio n a l issues c o n c e rn in g th e d e c o m p o s itio n o f ra tio n a l z-tran sfo rm s, w hich will p ro v e v ery u se fu l in th e im p le m e n ta tio n o f d isc re te -tim e system s. S u p p o se th a t w e h ave a ra tio n a l z-tran sfo rm X ( z ) e x p re ss e d as
X( z) =
(3.4.40)
196
The z-T ransform and Its Application to the Analysis of LTI Systems
Chap. 3
w h ere, fo r sim plicity, w e h av e a ssu m ed th a t ao = 1. If M > N [i.e., X ( z ) is im p ro p e r], w e c o n v e rt X (z) to a sum o f a p o ly n o m ial a n d a p r o p e r fu n c tio n M-H X ( z ) = J 2 c t z ~ k + X pA z ) k=o (3.4.41)
If th e p o le s o f X pr( z ) are d istin ct, it can b e e x p a n d e d in p a rtia l fra c tio n s as pT(z) = A \ - ------------- + A 2 ------------- r + --- + ^ jv -------------- r l-p \z~ l l ~ P 2Z~l 1 - PnZ~1 (3.4.42)
A s w e h av e a lre a d y o b se rv e d , th e re m ay b e so m e c o m p le x -c o n ju g a te p airs of p o le s in (3.4.42). S in ce we u su a lly d eal w ith real signals, w e sh o u ld av o id com plex co efficien ts in o u r d eco m p o sitio n . T h is can b e ach iev ed by g ro u p in g a n d co m b in in g te rm s co n ta in in g co m p lex -co n ju g ate p o les, in th e follow ing w ay: A A p * z ~ l + A* - A * p z ~ l 1 - pz~] 1 /5*z-1
_
1 - p z 1 - p*z ~' + PP*z ~2

bo + b -iZ ~ l
(3.4.43)
1 + a i z ~ l + 02 z ~2 w h ere
> o = 2R e(A ),
a \ = 2 R e ( p )
(3.4.44) b\ = - 2 Re (Ap*), a 2 = \ p \2 a re th e d e sire d co efficients. O b v io u sly , a n y ra tio n a l tra n sfo rm o f th e fo rm (3.4.43) w ith co efficien ts given by (3.4.44), w hich is th e case w h en a 2 4 a 2 < 0, can be in v e rte d using (3.4.34). B y c o m b in in g (3.4.41), (3.4.42), a n d (3.4.43) w e o b ta in a p a rtia l-fra c tio n e x p a n sio n fo r th e z -tra n s fo rm w ith distinct p o les th a t co n ta in s real coefficients. T h e g e n e ra l re su lt is M N Xiz) K\
1.
1 l.
= CZ-* + E 3
, / f'l -2 t l l + a u z 1+auZ 2
<3 A 4 S >
w h ere K\ + 2 AS = N . O b v io u sly , if Af = TV, th e first te rm is ju s t a c o n stan t, a n d w h e n M < N , th is te rm v an ish es. W h e n th e r e a re also m u ltip le p oles, som e a d d itio n a l h ig h e r-o rd e r te rm s sh o u ld b e in clu d ed in (3.4.45). A n a lte rn a tiv e fo rm is o b ta in e d by ex p ressin g X ( z ) as a p r o d u c t o f sim ple te rm s as in (3.4.40). H o w e v e r, th e co m p le x -c o n ju g a te p o le s a n d z e ro s sh o u ld be co m b in e d to av o id co m p lex co efficien ts in th e d e c o m p o sitio n . S u ch c o m b in atio n s re su lt in se c o n d -o rd e r ra tio n a l te rm s o f th e follow ing form : (1 - Z*Z-1 )(1 - z* kz ~ ' ) (1 - />*z- 1 ) ( l - p*kz ~ l ) w h ere b\ k = - 2 R t ( z k ) ,
b u = jz i l ,
1 + b\kZ~l + b u z ~2
1 -I- a u z -1 + a u z ~2
(3.4.46)
flu = 2 R e (p * )
(3.4.47)
02*
= \Pk\
Sec. 3.5
The One-sided 2 -Transform
197
A ssu m in g fo r sim p licity th a t M = N, w e se e th a t X (z) can be d e c o m p o s e d in the follow ing way:
1 + a kz 1
1 + a u z 1+ axz 2
<3 -4 -48>
w h ere N = K\ + 2 / ^ * W e will r e tu rn to th e s e im p o rta n t fo rm s in C h a p te rs 7 a n d 8.
3.5 THE ONE-SIDED Z-TRANSFORM

T h e tw o -sid ed z -tra n sfo rm re q u ire s th a t th e c o rre sp o n d in g signals be specified fo r th e e n tire tim e ran g e oo < n < oo. T his re q u ire m e n t p re v e n ts its u se for a v ery u se fu l fam ily o f p ra c tic a l p ro b lem s, n am ely th e e v a lu a tio n o f th e o u tp u t o f n o n re la x e d system s. A s we recall, th e s e system s a re d e sc rib e d by d ifferen ce e q u a tio n s w ith n o n z e ro initial c o n d itio n s. Since th e in p u t is a p p lie d a t a finite tim e, say n ()l b o th in p u t a n d o u tp u t signals are specified fo r n > no, b u t by no m e a n s a re z e ro fo r n < no- T h u s th e tw o -sid ed z-tran sfo rm c a n n o t b e used. In this se c tio n w e d e v e lo p th e o n e -sid e d z-tra n sfo rm w hich can be u se d to solve differen ce e q u a tio n s w ith in itial c o n d itio n s.
3.5.1 Definition and Properties

T h e one- si ded o r unilateral z -tra n sfo rm o f a signal x ( n ) is d efin ed by
CC
* + (;) n=0
W e also u se th e n o ta tio n s Z +{x(n)} a n d

*{n ) X + (z)
T h e o n e -sid e d z -tra n sfo rm d iffers fro m th e tw o -sid ed tra n sfo rm in th e low er lim it o f th e su m m a tio n , w hich is aiw ays z e ro , w h e th e r or n o t th e signal x ( n ) is zero fo r n < 0 (i.e., cau sal). D u e to th is choice o f lo w er lim it, th e o n e -sid e d z-tran sfo rm h as th e fo llow ing ch aracteristics: 1. It d o e s n o t c o n ta in in fo rm a tio n a b o u t th e signal jc(n) fo r n e g a tiv e v alu es o f tim e (i.e., fo r n < 0).
2. It is un i q u e o n ly fo r ca u sa l signals, becau se only th ese signals a re z e ro for

n < 0. 3. T h e o n e -sid e d z -tra n s fo rm A,+(z) o f x( n ) is id en tical to th e tw o -sid ed ztra n s fo rm o f th e signal x( n) u( n) . S ince x ( n ) u ( n ) is causal, th e R O C o f its tra n sfo rm , a n d h en c e th e R O C of X + {z), is alw ays th e e x te rio r o f a circle. T h u s w h en w e d e a l w ith o n e-sid ed z-tran sfo rm s, it is n o t n ecessa ry to re fe r to th e ir R O C .
198
The /-T ran sform and Its Application to the Analysis of LTI Systems
Chap. 3
Example 3.5.1 Determ ine the one-sided z-transform of the signals in Exam ple 3.1.1. Solution From the definition (3.5.1), we obtain X f ( z ) = 1 + 2 ; ' 1 + 5z~2 + 7z~3 + ;~ 5
x,(n) = {1, 2 ,5 ,7 ,0 ,1 } t x 2 (n) = {1.2, 5, 7, 0,1} t x 3(n) = (0 ,0 ,1 ,2 , 5,7, 0,1} t xA(n) = {2,4, 5, 7 ,0 ,1 ) t x 5(n) ~ S(n ) -* -1- * X+(z) = 1 x6(n) = 6(n - A), *700 = <S(k -I- k),
Jt: + (z) = 5 + 7z ~l + z ~3
X^(z ) = z ~2 + 2z"3 + 5z~4 + l z ' s + z~7
X4 + (z) = 5 + 7 ; ' 1 + z 3
k > 0 -e* k > 0
(z) = z~k Xy (z) = 0
Note that for a noncausal signal, the one-sided z-transform is not unique. Indeed, X 2 (z) = X^(z) but X2 (n) ^ x 4(n). Also for anticausal signals, (z) is always zero. A lm ost all properties we have studied for the tw o-sid ed z-transform carry o ver to the on e-sid ed z-transform with the excep tion of the shi ft i ng property.
Shifting Property
C a se 1: T im e D e la y If x(n) then k X (n - k ) X * z~*[Ar+(z) + ] [ ] j c ( - n ) z n]
n=l
X +(z )
k> 0
(3.5.2)
In case jc(n) is causal, then x{n ~ k ) z~k X +(z) (3.5.3)
Proof. F rom the definition (3.5.1) w e have Z +{x(n - k)} = z~
f S ( 0 z -'+
-k
Y t x( l ) z -
By changing the in d ex from / to n = t he result in (3.5.2) is easily obtained.
Sec. 3.5
The One-sided z-Transform
199
Example 3.5.2 Determ ine the one-sided ^-transform of the signals (a) x( n ) = a nu(n) (b) Ai(n) = xin 2) where x( n) = a " Solution (a) From (3.5.1) we easily obtain
(b) We will apply the shifting property for k = 2. Indeed, we have

Z +{ x ( n ~ 2)1 = ; " 2[X + ( ; ) + j : ( - 1 ) ; + j : ( - 2 ) - 2] = r 2^ + (:) + J T ( - l) ;- 1 + * ( -2 ) Since x( 1) = a ~ '. x ( 2) = a~2. we obtain ^ 1 - az ~l + <T2
T h e m e a n in g o f th e shifting p ro p e rty can be in tuitively ex p la in e d if w e w rite (3.5.2) as follow s: Z +[x (?i &)} = [jc (k) x ( + 1 ); ' + k> 0 + 1 ); *+* ] (3.5.4) + :.~kX +(.z) T o o b ta in x { n - k ) ( k > 0) fro m ;t(r?), w e sh o u ld shift x ( n ) by k sa m p le s to th e right. T h e n k n e w sa m p le s, x ( - k ) , x { k + 1), * ( - 1 ) , e n te r th e p o sitiv e tim e axis w ith x ( k) lo c a te d at tim e zero . T h e first te rm in (3.5.4) sta n d s fo r th e z-tran sfo rm o f th e s e sam p les. T h e o ld sa m p les o f x( n k) a re th e sa m e as th o se o f .r(n) sim ply sh ifted by k sa m p le s to th e right. T h e ir z-tra n sfo rm is o b v io u sly z _iX + (z), w hich is th e se co n d te rm in (3.5.4).
Case 2 : Time advance
If xin) X +(z)
th e n x ( n + k) z Proof . F ro m (3.5.1) we h ave

oc oc
X+( z ) - Y x ( n ) z - n k > 0
(3.5.5)
Z +[x(n + *)} =
n=0
+ k)z~n = zk Y m z ~ l
l= k
w h e re w e h a v e c h a n g e d th e in d ex o f su m m atio n fro m n t o 1 = n + k. N o w , from
200
Chap. 3
(3.5.1) w e o b ta in * + (z) = j r * ( / ) : ' = Y x { l ) z ~l + /=o i=0 i=t By co m b in in g th e last tw o re la tio n s, w e easily o b ta in (3.5.5). Exam ple 3.5.3 With x(n), as given in Example 3.5.2, determ ine the one-sided ^-transform of the signal xi (n) = x(n + 2 ) Solution obtain We will apply the shifting theorem for k = 2. From (3.5.5), with k = 2, we Z +{x(n + 2)| = -2X+(z) - x (0 );2 - *(1); But jc(0) = 1, jc(1) = a. and X + (z) = 1/(1 - a ; -1). Thus ,2 Z +{x(n + 2)) = - ----- - - z 2 - az 1 az~ T h e case o f a tim e ad v an ce can be in tu itiv e ly e x p la in e d as follow s. T o o b tain x( n- i - k) , k > 0, we sh o u ld shift jc (n ) by k sa m p le s to th e left. A s a re su lt, th e sam ples * (0 ). * ( 1 ) , ___ x ( k 1) le a v e th e p o sitiv e tim e axis. T h u s w e first re m o v e th e ir co n trib u tio n to th e X +(z ), an d th e n m u ltip ly w h at re m a in s by z k to co m p e n sa te fo r th e shifting o f th e signal by k sam p les. T h e im p o rta n c e o f th e shifting p ro p e rty lies in its a p p lic a tio n to th e so lu tio n of d ifferen ce e q u a tio n s w ith c o n s ta n t co efficien ts a n d n o n z e ro in itial co n d itio n s. T h is m a k e s th e o n e-sid ed z -tra n sfo rm a v ery useful to o l fo r th e an aly sis o f recu rsiv e lin e a r tim e -in v a ria n t d isc re te -tim e system s. A n im p o rta n t th e o re m useful in th e analysis o f signals a n d system s is th e final v alu e th e o re m . F in al V a lu e T h e o re m . If x{n) th e n lim x( n ) = lim (z - l ) X + (z) n- 00 7-l (3.5.6) X +{z)
T h e lim it in (3.5.6) exists if th e R O C o f (z - l)A '+ (z) in clu d es th e u n it circle. T h e p ro o f o f th is th e o re m is left as an ex ercise fo r th e re a d e r. T h is th e o re m is u sefu l w h en w e a re in te re s te d in th e a s y m p to tic b e h a v io r of a signal x( n) a n d w e k n o w its z -tra n sfo rm , b u t n o t th e signal itself. In such cases, esp ecially if it is co m p licated to in v e rt X + (z), w e can u se th e final v alue th e o re m to d e te rm in e th e lim it o f x{n) as n goes to infinity.
Sec. 3.5
The One-sided z-Transform
201
Example 3.5.4
The impulse response of a relaxed linear time-invariant system is k(n) = a"u(n ), | | < 1. D eterm ine the value of the step response of the system as n oo. Solution The step response of the system is y{n) = h(n) * x(n) where Jt(n) = u(n) Obviously, if we excite a causal system with a causal input the output will be causal. Since h(n), x(n), v(n) are causal signals, the one-sided and two-sided z-transforms are identical. From the convolution property (3.2.17) we know that the z-transforms of h(n) and *(n) must be multiplied to yield the z-transform of the output. Thus = . Now (z - l)y (z) Z a
1 - az 1 1 - z 1
, .
- , = 7----- tt? ------- r

(z - l ) ( z - cr)
ROC: |z| > |ar|
ROC: |z| > |a|
Since |a | < 1 the R O C of (z - l)K(z) includes the unit circle. Consequently, we can apply (3.5.6) and obtain lim v(n) = lim - = n oc' 1 a
3.5.2 Solution of Difference Equations

T he on e-sid ed z-transform is a very efficient tool for the solu tion of d ifference eq u a tio n s w ith n on zero initial conditions. It ach ieves that by reducing the dif feren ce eq u ation relating the tw o tim e-d om ain signals to an eq u ivalen t algebraic eq u ation relating their on e-sid ed z-transform s. T h is eq u ation can b e easily solved to obtain the transform o f the desired signal. T he signal in the tim e dom ain is o b ta in ed by inverting the resulting z-transform . W e will illustrate this approach with tw o exam ples. Exam ple 3-5.5 The well-known Fibonacci sequence of integer num bers is obtained by computing each term as the sum of the two previous ones. The first few terms of the sequence are 1 ,1 ,2 , 3,5. 8 ,... D eterm ine a closed-form expression for the n th term of the Fibonacci sequence. Solution Let y(n) be the nth term of the Fibonacci sequence. Clearly, y(n) satisfies the difference equation
y (n ) = y(n - 1) + y(n - 2)
(3.5.7)
202
The z -Transform and Its Application to the Analysis of LTI System s
Chap. 3
with initial conditions v(0) = v(11 + v( 2) = 1 y (l) = y(0) + V(-1) = 1 (3.5.8a) (3.5.8b)
From (3.5.8b) we have y ( 1) = 0. Then (3.5.8a) gives v( 2) = 1. Thus we have to determ ine y(n), n > 0, which satisfies (3.5.7), with initial conditions y (1) = 0 and y( 2) = 1. By taking the one-sided ^-transform of (3.5.7) and using the shifting property (3.5.2). we obtain y+(z) = [ ; - 'y +(-) + y (1) ] + [z ~2Y ^ ( z ) + y (-2 ) + . v ( - l ) ; - 1] or
K + (c) =
"
(3.5.9)
where we have used the fact that y{ 1) = 0 and v( 2) = 1. We can invert K+(;) by the partial-fraction expansion m ethod. The poles of y'+(;) are 1 -I- V5
Pi P2 =
l-V s
and the corresponding coefficients are A i = p i/V 5 and A 2 = -/> ;/V 5 . Therefore, v(n) = or, equivalently. u{n) (3.5.10)
'l + V5 / 1 + v'S \ "

2V5
1 - V5 / I - n / 5 \ "
2 ^5
u(rt)
Example 3.5.6 Determ ine the step response of the system v(n) = ay(n 1) + x(rt) when the initial condition is y (1) = 1. Solution By taking the one-sided ^-transform of both sides of (3.5.11), we obtain
K+(;) = a [ ;- 'y * (c ) + v (- D ] + X +(z)
1<
<
(3.5.11)
Upon substitution for v ( - l ) and X+(;) and solving for y + (;). we obtain the result Y+(z) = 1 a ; -1 1 (1 a ; - l )(l - ; _l) (3.5.12)
By performing a partial-fraction expansion and inverse transform ing the result, we have y(n) = a"
,u(n)
1 or"+I + ;-------- u(n) 1 a
( 3 .5 . 13)
( l - a " +2) u ( n )
1 a
Sec. 3.6
Analysis of Linear Tim e-Invariant Systems in the z-D om ain
203
3.6 ANALYSIS OF LINEAR TIME-INVARIANT SYSTEMS IN THE Z-DOMAIN

In S e ctio n 3.4.3 w e in tro d u c e d th e sy stem fu n ctio n o f a lin e a r tim e -in v a ria n t sys tem a n d re la te d it to th e u n it sa m p le re sp o n se a n d to th e d iffe re n c e e q u a tio n d e sc rip tio n o f sy stem s. In th is sectio n w e d escrib e th e use o f th e sy stem fu n c tio n in th e d e te rm in a tio n o f th e re sp o n s e of th e system to so m e ex c ita tio n signal. F u rth e rm o re , w e e x te n d th is m e th o d o f a nalysis to n o n re la x e d sy stem s. O u r a tte n tio n is fo c u se d o n th e im p o rta n t class o f p o le - z e r o sy stem s r e p re s e n te d by lin e a r c o n s tan t-co efficien t d iffe re n c e e q u a tio n s w ith a rb itra ry in itial co n d itio n s. W e also c o n s id e r th e to p ic o f sta b ility o f lin e a r tim e -in v a ria n t sy stem s an d d esc rib e a test fo r d e te rm in in g th e sta b ility o f a sy stem b a s e d o n th e co efficien ts o f th e d e n o m in a to r p o ly n o m ial in th e system fu n c tio n . F in ally , w e p ro v id e a d e ta ile d an aly sis o f se c o n d -o rd e r system s, w hich fo rm th e b asic b u ild in g b lo ck s in th e re a liz a tio n o f h ig h e r-o rd e r system s.
3.6.1 Response of Systems with Rational System Functions

L e t us c o n s id e r a p o le - z e r o system d e s c rib e d by th e g e n e ra l lin e a r c o n s ta n tc o efficien t d iffe re n c e e q u a tio n in (3.3.7) a n d th e c o rre sp o n d in g system fu n ctio n in (3.3.8). W e r e p re s e n t H ( z ) as a ra tio o f tw o p o ly n o m ials B ( z ) / A ( z ) , w h ere B (z) is th e n u m e r a to r p o ly n o m ia l th a t c o n ta in s th e z e ro s o f H( z ) , a n d A ( z ) is the d e n o m in a to r p o ly n o m ia l th a t d e te rm in e s th e p o le s o f H ( z ) . F u rth e rm o r e , let us a ssu m e th a t th e in p u t signal x ( n ) h as a ra tio n a l z -tra n sfo rm X (z) o f th e fo rm X(z) = Q( z) (3.6.1)
T h is a s su m p tio n is n o t o v erly restrictiv e, since, as in d ic a te d p rev io u sly , m o st signals o f p ra c tic a l in te re s t h av e ra tio n a l z -tran sfo rm s. If th e sy stem is in itially re lax ed , th a t is, th e in itia l c o n d itio n s fo r th e d iffe re n c e e q u a tio n are z e ro , y ( 1) = y ( 2) = = y ( N ) = 0, th e z -tra n s fo rm o f th e o u tp u t o f th e sy stem h a s th e fo rm Y(z) = H ( z ) X ( z ) = (3.6.2)
Mz)Q(z)
N o w su p p o s e th a t th e sy stem c o n ta in s sim ple p o le s p \ , p j ......... p s a n d th e ztra n sfo rm o f th e in p u t signal co n ta in s p o le s < 71, qt, , q u w h e re p t ^ qm fo r all it = 1, 2 a n d m = 1, 2 , . . . , L. In a d d itio n , w e a ssu m e th a t th e z e ro s o f th e n u m e r a to r p o ly n o m ia ls B( z ) a n d N ( z ) d o n o t co in cid e w ith th e p o les {p t } an d {<7i}, so th a t th e re is n o p o le - z e r o c a n c e lla tio n . T h e n a p a rtia l-fra c tio n ex p an sio n o f K(z) yield s
t o 1 - PkZ 1
*-* 1 - qkZ 1
204
Chap. 3
T h e in v erse tra n sfo rm o f ^ (z ) y ield s th e o u tp u t signal fro m th e sy stem in th e form

N L
y (n ) = Y A t ( P k ) n u ( n ) + 5 2 2*0?*)'1n (n) *=i *-i
(3.6.4)
W e o b se rv e th a t th e o u tp u t se q u e n c e y( n) can b e su b d iv id e d in to tw o p a rts. T he first p a r t is a fu n ctio n o f th e p o les {pt} o f th e system an d is called th e natural response o f th e sy stem . T h e in flu e n ce o f th e in p u t signal o n th is p a r t o f th e re sp o n s e is th ro u g h th e scale facto rs {A*}. T h e se c o n d p a r t o f th e re sp o n se is a fu n ctio n o f th e p o le s {qk} o f th e in p u t signal an d is called th e f o r c e d response of th e sy stem . T h e in flu e n ce o f th e sy stem o n th is re sp o n s e is e x e rte d th ro u g h th e scale facto rs { Qk } W e sh o u ld e m p h asize th a t th e scale fa c to rs {A*} an d { Q k } a re fu n ctio n s o f b o th se ts o f p o le s {pk } a n d { ^ ) . F o r e x am p le, if AXz) = 0 so th a t th e in p u t is zero , th e n K(z) = 0, a n d c o n s e q u e n tly , th e o u tp u t is z e ro . C lea rly , th e n , th e n a tu r a l re sp o n se o f th e sy stem is zero . T h is im p lies th a t th e n a tu ra l re sp o n se of th e sy stem is d iffe re n t fro m th e z e ro -in p u t re sp o n se . W h en X (z) a n d H ( z ) h av e o n e o r m o re p o les in c o m m o n o r w h e n X (z) a n d /o r H{ z ) c o n ta in m u ltip le -o rd e r p o les, th e n K(z) will h av e m u ltip le -o rd e r poles. C o n se q u e n tly , th e p a rtia l-fra c tio n ex p a n sio n o f Y( z ) will c o n ta in fa c to rs o f th e form 1/(1 - p / z ~ l )t , k = 1, 2 , . . . , m, w h ere m is th e p o le o rd e r. T h e in v ersio n o f th ese facto rs will p ro d u c e te rm s o f th e fo rm n k~ xp* in th e o u tp u t y( n ) o f th e system , as in d ic a te d in S ectio n 3.4.2.
3.6.2 Response of Pole-Zero Systems with Nonzero Initial Conditions

S u p p o se th a t th e sig n al x ( n ) is a p p lie d to th e p o le - z e r o sy stem at n = 0. T hus th e sign al x ( n ) is a ssu m ed to be cau sal. T h e effe cts o f all p re v io u s in p u t signals to th e system a re reflec te d in th e in itial c o n d itio n s y ( 1), y ( ~ 2 ) .........y ( N ) . Since th e in p u t x ( n ) is ca u sa l a n d since w e a re in te re s te d in d e te rm in in g th e o u tp u t y ( n ) fo r n > 0, w e can u se th e o n e -sid e d z-tra n s fo rm , w hich allow s us to d e a l w ith th e in itial co n d itio n s. T h u s th e o n e -sid e d z -tra n s fo rm o f (3.4.7) b e c o m e s akz + f c * z - * X + (z) (3.6.5)
f t)
Y +{z) + Y j y ( ' - n ) z ''
S ince x ( n) is cau sal, we can se t X + (z) = X (z). In an y case (3.6.5) m ay b e e x p ressed as
52 ^
Y +(z) =
1+5 2 < :
= H(z)X(z) + No(z) A (z)
*=0 s
-X(z)akz
(3.6.6)
Sec. 3.6
205
w h e re
N k
No(z) = - 5 Z a *z_,t k= 1 n=l
(3.6.7)
F ro m (3.6.6) it is a p p a r e n t th a t th e o u tp u t o f th e sy stem w ith n o n z e ro initial c o n d itio n s can b e su b d iv id e d in to tw o p a rts. T h e first is th e z e ro -sta te re sp o n se of th e sy stem , d e fin ed in th e z -d o m ain as Ya (z) = H ( z ) X ( z ) (3.6.8)
T h e se c o n d c o m p o n e n t c o rre sp o n d s to th e o u tp u t re su ltin g fro m th e n o n z e ro initial co n d itio n s. T h is o u tp u t is th e z e ro -in p u t re sp o n se o f th e system , w hich is defined in th e z -d o m ain as
(z) = TTT A(z)
a 6 9 )
H e n c e th e to ta l re sp o n se is th e su m o f th ese tw o o u tp u t c o m p o n e n ts , w hich can b e e x p ressed in th e tim e d o m a in by d e te rm in in g th e in v erse z -tra n s fo rm s o f K zs(;) a n d Y A z ) s e p a ra te ly , a n d th e n ad d in g th e resu lts. T h u s y ( n) = >zs() + >'zi(n) (3.6.10)
S ince th e d e n o m in a to r of Fz |( z ) , is A(z), its p o le s are p i , P 2 ........ Ps- C o n se q u e n tly , th e z e ro -in p u t re sp o n se h a s th e form >'zi(n) =
52
L
(3.6.11)
T h is can b e a d d e d to (3.6.4) a n d th e te rm s involving th e p o les {/?*} can be co m b in ed to yield th e to ta l re sp o n s e in th e form y( n) = w h e re , by d efin itio n , A'k = + Dk (3.6.13)
52 A k ' ( Pk) nu( n) + 51 Q k i q k T u ( n )

k= 1 k=\
(3.6.12)
T h is d e v e lo p m e n t in d icates clearly th a t th e effect o f th e initial co n d itio n s is to a lte r th e n a tu ra l re sp o n s e o f th e system th ro u g h m o d ifica tio n o f th e scale f a c to rs {Ak}. T h e re a re n o n ew p o le s in tro d u c e d by th e n o n z e ro in itial co n d itio n s. F u rth e rm o r e , th e r e is no effect on th e fo rce d re sp o n s e o f th e system . T h ese im p o rta n t p o in ts a re re in fo rc e d in th e follow ing e x am p le. Exam ple 3.6.1 D eterm ine the unit step response of the system described by the difference equation y{n) = 0.9y(n 1) - 0.81y(n - 2) + x( n ) under the following initial conditions: (a) y ( - l ) = > (-2 ) = 0
(b) y ( - 1) = y ( - 2 ) = 1
206
Chap. 3
Solution
The system function is

H {z) =
1 1 - 0 . 9 : - 1 + 0 .8 1 ;- 2
This system has two complex-conjugate poles at

p l = Q.9e^ P 2 = G. 9e- jn/i
The z-transform of the unit step sequence is 1 X(z) = Therefore, 1

(1 - 0 . 9 e j ^ z - l K l - 0 . 9 e - j '1z -'l ) ( l - z ~ ' )
1-z -1
0.542 - /0 .0 4 9
1 - 0 .9 e ^ f h ~ l
0.542 + y0.049 + :---- r +
1.099
1 - z~l
1 - 0 .9e->nPz - 1
and hence the zero-state response is

y^(n) =
l.099
1.088(0.9)" cos
(~ n
- 5.2C )J
u(n)
(a) Since the initial conditions are zero in this case, we conclude that v(n) = >'is(fl). (b) For the initial conditions v ( - l ) = v (-2 ) = 1, the additional com ponent in the z-transform is I'ziW Mi(z) A(z)
0.09 - 0.S ir " 1 1 - 0 . 9 ; - ' + 0.81r2
0.026 + j0.4936 i 0.026 - jO.4936
1-
0 H e i ' f i z - 1 + 1 - 0 . 9 e - '* P z ~ '
Consequently, the zero-input response is yzi(n) = 0.988(0.9)'' cos ^ ~ n + 87^J u(n) In this case the total response has the z-transform
y (z )
= J ^ w + y^ u ) 1.099 0.568 + y'0.445 0.568 - /0.445 + :---- r + 1 - z "1 1 - Q.9eJ*Vz- 1 1 - 0.9e' ff/3; -1
The inverse transform yields the total response in the form y(n) = 1.099u(n) + 1.44(0.9)" cos ( ^ n + 3 8 ^ u (n)
3.6.3 Transient and Steady-State Responses

A s w e h av e seen fro m o u r p re v io u s d iscussion, th e re sp o n se o f a sy ste m to a given in p u t can b e se p a ra te d in to tw o c o m p o n e n ts , th e n a tu r a l re sp o n s e a n d th e fo rce d
Sec. 3.6
Analysis of Linear Tim e-Invariant Systems in the z-Dom ain
207
re sp o n se . T h e n a tu ra l re sp o n s e o f a cau sal system h a s th e fo rm N >nr(n) =
52 A k i p k ) nu{ n)
*=1
(3.6.14)
w h e re {pk), k = 1, 2, N a re th e p o le s o f th e sy stem a n d {A*} a re sc ale fac to rs th a t d e p e n d o n th e in itial c o n d itio n s an d on th e c h a ra c te ristic s o f th e in p u t se q u en ce. If |/>*| < 1 fo r all k, th e n , y nT(n) d ecay s to z e ro as n a p p ro a c h e s infinity. In such a case w e re fe r to th e n a tu ra l re sp o n se of th e system as th e t ransient response. T h e ra te a t w hich >'nr(n) d ecay s to w a rd z e ro d e p e n d s on th e m a g n itu d e o f th e p o le p o sitio n s. I f all th e p o le s h av e sm all m a g n itu d e s, th e d ec a y is v ery ra p id . O n the o th e r h a n d , if o n e o r m o re p o le s a re lo c a te d n e a r th e u n it circle, th e c o rre sp o n d in g te rm s in >nr(n) w ill d e c a y slow ly to w a rd z e ro a n d th e tra n s ie n t will p ersist fo r a relativ ely lo n g tim e. T h e fo rc e d re sp o n s e o f th e system h as th e fo rm i Vfr() = 5 2 <2*(<?*)"() k=\ (3.6.15)
w h e re {^t), k = 1, 2 , . . . , L a re th e p o le s in th e forcing fu n c tio n a n d { Qk } a re scale fa c to rs th a t d e p e n d o n th e in p u t se q u e n c e an d o n th e c h a ra c te ristic s o f th e system . If all th e p o le s o f th e in p u t signal fall in sid e th e u n it circle, ^ ( n ) will d ecay to w a rd z e ro as n a p p ro a c h e s infinity, ju st as in th e case o f th e n a tu ra l resp o n se. T h is sh o u ld n o t b e su rp risin g since th e in p u t signal is also a tra n s ie n t signal. O n th e o th e r h a n d , w h e n th e cau sal in p u t signal is a sin u so id , th e p o le s fall o n th e unit circle a n d c o n s e q u e n tly , th e fo rce d re sp o n se is also a sin u so id th a t p ersists fo r all n > 0. In th is case, th e fo rce d re sp o n se is called th e steady-state respons e o f th e system . T h u s, fo r th e sy stem to sustain a ste a d y -s ta te o u tp u t fo r n > 0, th e in p u t sig n al m u st p e rsist fo r all n > 0. T h e fo llo w in g ex a m p le illu stra te s th e p re se n c e o f th e s te a d y -s ta te resp o n se. Exam ple 3*6.2 D eterm ine the transient and steady-state responses of the system characterized by the difference equation >{n) = 0.5;y(n - 1) + jc(n) when the input signal is x ( n ) = 10cos(jrn/4)u(n). The system is initially at rest (i.e., it is relaxed). Solution The system function for this system is
and therefore the system has a pole at z = 0.5. The z-transform of the input signal is (from Table 3.3)
10(1 ( l/y / 2 ) z ~ 1) X (z ) ---------------f ---------------1 - y / 2 _1 + Z~1
208
The z-Transform and Its Application to the Analysis of LTI S ystem s Consequently.
K (o = H (:)X (z )
Chap. 3
10(1 - ( l / v '2 ) ; - 1) (1 0.5~- )(1 - e - w 4; - 1)!! 6.3

1 - 0.5;- 1
i) 6.7&ej2K1
6.78e~J2 & 1
1-
, "T
1 - e-J*iA z~x
The natural or transient response is ynr(n) = 6.3(0.5)"u() and the forced or steady-state response is
Vfr (n) =
+6. 1&eJ2*J e - inn!A]u(},)
= 13.56cos (^~^n 2 8 . 7 w() Thus we see that the steady-state response persists for all n > 0. just as the input signal persists for all n > 0.
3.6.4 Causality and Stability

A s d efin ed p rev io u sly , a causal lin e a r tim e -in v a ria n t system is o n e w hose unit sa m p le resp o n se h (n ) satisfies the co n d itio n
h (n ) = 0 n < 0
W e h av e also show n th at the R O C o f th e ;;-tra n sfo rm o f a cau sal se q u e n c e is the e x te rio r o f a circle. C o n se q u e n tly , a lin e a r t im e -in v a r ia n t sy ste m is c a u s a l i f an d
o n ly i f the R O C o f the system f u n c t io n is the e x te rio r o f a c ir c le o f ra d iu s r < 00, in c lu d in g the p o in t z
do.
T h e stab ility o f a lin ear tim e -in v a ria n t system can also be e x p re ss e d in term s o f th e ch a ra c te ristic s o f th e system fu n ctio n . A s w e recall fro m o u r prev io u s d iscu ssio n , a n ecessa ry an d sufficient co n d itio n fo r a lin e a r tim e -in v a ria n t system to b e B IB O sta b le is n=- x In tu rn , this c o n d itio n im plies th a tt H ( z ) m u st c o n ta in th e u n it circle w ith in its R O C . musi In d e e d , since
OC
52
H (z) =
h (n)z
it follow s th a t
OC OC
n oc
n= oc
W h en e v a lu a te d o n th e u n it circle (i.e., |z| = 1),

OC
\ h (z )\ <
52
Sec. 3.6
Analysis of Linear Tim e-Invariant Systems in the 7 -Domain
209
H en ce, if the system is B IB O stable, the unit circle is con tained in the R O C o f H(z)- T h e con verse is also true. T h erefore, a linear tim e-in va ria n t sy stem is B IB O stable i f a n d o n ly i f th e R O C o f the sy stem fu n c tio n includes the u n it circle. W e should stress, how ever, that the con d ition s for causality and stability are d ifferent and that o n e d oes not im ply the other. F or exam p le, a causal system m ay b e stable or unstable, just as a noncausal system m ay b e stable or unstable. Sim ilarly, an unstable system m ay be eith er causal or n oncausal, just as a stable system m ay be causal or noncausal. For a causal system , h ow ever, the con d ition on stability can be narrowed to so m e exten t. In d eed , a causal system is characterized by a system function H ( z ) having as a R O C the exterior o f som e circle o f radius r. For a stable system , the R O C m ust include the unit circle. C on sequ en tly, a causal and sta ble system m ust have a system function that con verges for |z| > r < 1. Since the R O C cannot contain any p oles o f H ( z ) , it follow s that a causal linear tim ein va ria n t sy stem is B I B O stable i f a n d o n ly i f all the p o le s o f H ( z ) are inside the u n it circle. Example 3.63 A linear tim e-invariant system is characterized by the system function 3 - 4 z-' H(z) ~ 1 - 3.5z-> + 1.5: ' 2
1
"
l - ' i z - 1 + 1 - 3 z- 1
Specify the R O C of H(z) and determ ine h(n) for the following conditions: (a) The system is stable.
( b ) The system is causal.
(c) The system is anticausal. Solution T he system has poles at z = 5 and z = 3.
(a) Since the system is stable, its R O C must include the unit circle and hence it is \ < \z\ < 3. Consequently, h(n) is noncausal and is given as h(n) = (i)"n() - 2(3)-it ( - n - 1) (b) Since the system is causal, its R O C is jz| > 3. In this case /i(n) = ( i r ( n)+ 2 (3 )"u (n ) This system is unstable. (c) If the system is anticausal, its R O C is |z| < 0.5. Hence h{n) = - { ( \ y + 2 ( 3 T ) u ( - n - l )
In this case the system is unstable.
210
The z - T ransform and Its Application to the Analysis of LTI Systems
Chap. 3
3.6.5 Pole-Zero Cancellations

W h en a z -tra n sfo rm h as a p o le th a t is a t th e sam e lo catio n as a z e ro , th e pole is ca n c e le d by th e z e ro an d , c o n s e q u e n tly , th e te rm c o n ta in in g th a t p o le in the in v erse z-tra n sfo rm v an ish es. S uch p o le - z e r o c a n ce llatio n s a re v ery im p o rta n t in th e an alysis o f p o le - z e r o system s. P o le -z e ro c a n ce llatio n s can o c c u r e ith e r in th e sy stem fu n c tio n itself o r in th e p ro d u c t o f th e sy stem fu n ctio n w ith th e z -tra n sfo rm o f th e in p u t signal. In the first case w e say th a t th e o r d e r o f th e sy stem is re d u c e d by o n e . In th e la tte r case w e say th a t th e p o le o f th e system is su p p re s se d by th e z e ro in th e in p u t signal, o r vice v ersa. T h u s, by p ro p e rly se lectin g th e p o sitio n o f th e ze ro s o f th e in p u t signal, it is p o ssib le to su p p re ss o n e o r m o re sy stem m o d es (p o le fa c to rs) in th e re sp o n se o f th e sy stem . S im ilarly, by p r o p e r se le c tio n o f th e z e ro s o f th e system fu n ctio n , it is p o ssib le to su p p re s s o n e o r m o re m o d e s o f th e in p u t signal fro m the re sp o n se o f th e system . W h e n th e z e ro is lo c a te d v ery n e a r th e p o le b u t n o t ex actly a t th e sa m e loca tio n , th e te rm in th e re sp o n se h as a v ery sm all a m p litu d e . F o r e x a m p le , n o n ex act p o ie - z e r o c an ce llatio n s can o ccu r in p ra c tic e as a re su lt o f in su ffician t n u m erical p recisio n u sed in re p re se n tin g th e co efficien ts o f th e system . C o n s e q u e n tly , one sh o u ld n o t a tte m p t to stab ilize an in h e re n tly u n sta b le system by p lacing a z e ro in th e in p u t signal at th e lo catio n o f th e pole. Example 3.6.4 Determ ine the unit sample response of the system characterized by the difference equation v(n) = 2.5 v(n 1) >(n 2) + jr(n) 5jr(n 1) + 6x(n 2) Solution The system function is
1 - 5*-1 + 6z ' 2 (1 - i z - ') ( l - 2 z - This system has poles at p\ = 2 and p x = ~, Consequently, at first glance it appears that the unit sample response is Y(z) = H( z) X( z) = 1 - 5 z ~ ' + 6z2 (1 - j z _ ,)(l - 2z_l
By evaluating the constants at z = j and z = 2, we find that A = * B =0
The fact that 5 = 0 indicates that there exists a zero at z = 2 which cancels the pole at z = 2. In fact, the zeros occur at z = 2 and z = 3. Consequently, H{z)
Sec. 3.6
Analysis of Linear Tim e-Invariant Systems in the z-Dom ain
211
reduces to H(z) = 1 - 3 : -1 1- k -1 2 .5 ;-1
= 1and therefore
h(rt) = S{n) -
2 .5 ( 5)"
~ ^
The reduced*order system obtained by canceling the common pole and zero is char acterized by the difference equation y(n) = iy(n - 1) + x(n) - 3x{n - 1) Although the original system is also BIBO stable due to the pole-zero cancellation, in a practical im plementation of this second-order system, we may encounter an instability due to imperfect cancellation of the pole and the zero. Example 3.6.5 D eterm ine the response of the system v(n) = jjy(>i - 1) - y(n - 2) + x(n) to the input signal x{n) = S(n) ^&(n 1). Solution The system function is 1 WU) =
(1 -
(1 - 1 c-1) The ^transform of
This system has two poles, one at z = ^ and the other at : = the input signal is
X {z ) = 1 - j i ' 1
In this case the input signal contains a zero at ; = i which cancels the pole at ; = Consequently, Y(z) = HU) X( z ) m i r p
and hence the response of the system is y(n) = ( 5)n(n) Clearly, the m ode ( |)" is suppressed from the output as a result of the pole-zero cancellation.
3.6.6 Multiple-Order Poles and Stability

A s w e h a v e o b se rv e d , a n ecessa ry an d sufficient c o n d itio n fo r a causal lin e a r tim ein v a ria n t sy stem to b e B IB O sta b le is th a t all its p o le s lie in sid e th e u n it circle. T h e in p u t sig n al is b o u n d e d if its z -tra n s fo rm c o n ta in s p o le s {qk}, k = 1 , 2 , L,
212
Chap. 3
w hich satisfy th e c o n d itio n \qk \ < 1 fo r all k. W e n o te th a t th e fo rc e d re sp o n se of th e system , given in (3.6.15). is also b o u n d e d , ev en w h e n th e in p u t signal contains o n e o r m o re d istin ct p o les on th e u n it circle. In view o f th e fact th a t a b o u n d e d in p u t signal m ay h av e p o le s on th e unit circle, it m ight a p p e a r th a t a sta b le system m ay also h av e p o le s on th e u n it circle. T his is n o t th e case, h o w ev er, since such a sy stem p ro d u c e s an u n b o u n d e d resp o n se w h en ex cited by an in p u t signal th a t also has a p o le a t th e sa m e p o sitio n o n the u n it circle. T h e follow ing ex a m p le illu stra te s th is p o in t. Example 3.6.6 D eterm ine the step response of the causal system described by the difference equation v(n) = y{n - 1) + x{n) Solution The system function for the system is
We note that the system contains a pole on the unit circle at c = 1. The ^-transform of the input signal x(n) = u{n) is
which also contains a pole at ; = 1. Hence the output signal has the transform
K(=) = H {z)X {z)
1 (1 - ; - 1)2 which contains a double pole at z = 1. The inverse z-transform of Y(z) is >-()= (n + l)n(B) which is a ramp sequence. Thus y(n) is unbounded, even when the input is bounded. Consequently, the system is unstable. E x am p le 3.6.6 d e m o n s tra te s clearly th a t B IB O stab ility r e q u ire s th a t th e sys tem p o le s b e strictly in side th e u n it circle. If th e sy stem p o les a re all inside th e unit circle an d th e e x c ita tio n se q u e n c e x ( n ) c o n ta in s o n e o r m o re p o le s th a t coincide w ith th e p o les o f th e sy stem , th e o u tp u t Y(z ) w ill c o n ta in m u ltip le -o rd e r poles. As in d ic a te d p rev io u sly , such m u ltip le -o rd e r p o les re su lt in an o u tp u t se q u e n c e th at co n tain s term s o f th e form A kn b( p k)nu(n) w h ere 0 < b < m 1 an d m is th e o r d e r o f th e p o le. If |p*| < 1, th e s e te rm s decay to z e ro as n a p p ro a c h e s infinity b e c a u s e th e e x p o n e n tia l f a c to r ( pk) n d o m in ates th e te rm n b. C o n se q u e n tly , n o b o u n d e d in p u t signal can p r o d u c e an u n b o u n d e d o u tp u t signal if th e sy stem p o les a re all in sid e th e u n it circle.
Sec. 3.6
Analysis of Linear Tim e-Invariant Systems in the z - D o m a in
213
F in ally , w e sh o u ld s ta te th a t th e o n ly u sefu l system s w h ich c o n ta in p o les on th e u n it circle a re th e d ig ital o sc illato rs discussed in C h a p te r 4. W e call such sy stem s mar gi nal l y stable.
3.6.7 The Schur-Cohn Stability Test

W e h av e sta te d p rev io u sly th a t th e sta b ility o f a sy stem is d e te rm in e d by th e p o sitio n o f th e p o les. T h e p o les o f th e system are th e ro o ts o f th e d e n o m in a to r p o ly n o m ia l o f H ( z ), n am ely , A(z) = 1 + a \ z 1 + Q -2Z 2 + + QpfZ ^ (3.6.16)
W h en th e sy stem is cau sal all th e ro o ts o f A (z) m u st lie inside th e u n it circle fo r th e sy stem to b e sta b le . T h e re a re se v e ra l c o m p u ta tio n a l p ro c e d u re s th a t aid u s in d e te rm in in g if any o f th e ro o ts o f A (z) lie o u tsid e th e u n it circle. T h e se p ro c e d u re s a re called stability criteria. B e io w w e d esc rib e th e S c h u r-C o h n test p ro c e d u re fo r th e sta b ility o f a sy stem c h a ra c te riz e d by th e system fu n c tio n H( z ) = B ( z ) / A { z). B e fo re w e d e sc rib e th e S c h u r-C o h n te st w e n e e d to e sta b lish so m e useful n o ta tio n . W e d e n o te a p o ly n o m ia l o f d e g re e m by Am(z) = a m(* )z - A i=0 a m(0) = l (3.6.17)
T h e reci procal o r reverse p o l y n o m i a l Bm(z) o f d e g ree m is d e fin e d as Bm(z) = z-') (3.6.18) UmV"
k*= 0
W e o b se rv e th a t th e coefficients o f B m( z ) are th e sa m e a s th o se o f Am(z), b u t in re v e rse o rd e r. In th e S c h u r-C o h n sta b ility te s t, to d e te rm in e if th e p o ly n o m ia l A (z) has all its r o o ts in sid e th e u n it circle, w e c o m p u te a set o f coefficients, called reflection coefficients , K \ , K i .........K n fro m th e p o ly n o m ials A m(z). F irst, w e set A N (z) = A( z ) and A T jV = a # ( N ) T h e n w e c o m p u te th e lo w e r-d e g re e p o ly n o m ials Am(z), m = N , N 1, N 2 , . . . , 1, a c c o rd in g to th e rec u rsiv e e q u a tio n A ^ ^m (z) K mB m(z) A m- i ( z ) = --------1 -------w h e re th e c o effic ie n ts K m a re d e fin ed as ^ c (3.6.20) (3.6.19)
Km = am{m)
(3.6.21)
214
Chap, 3
T h e S c h u r-C o h n sta b ility test sta te s th a t the p o l y n o m i a l A (;) gi en by (3.6.16) has all its roots inside the unii circle i f a n d onl y i f the coefficients K m satisfy the condit i on \Km \ < 1 f o r all m 1, 2 ........ N. W e shall n o t p ro v id e a p r o o f o f th e S c h u r-C o h n te s t at th is p o in t. The th e o re tic a l ju stificatio n fo r th is te s t is given in C h a p te r 11. W e illu stra te th e com p u ta tio n a l p ro c e d u re w ith th e follow ing exam ple. Example 3.6.7 Determ ine if the system having the system function
is stable. Solution We begin with A ;(;), which is defined as

A 2 (z ) =
1 -
lz~] -
k " 2
Hence
Now #>(:) = and /\2(r) 1 - K;
Therefore. Ki = - I Since |A Ti | > 1 il follows that the system is unstable. This fact is easily estab lished in this example, since the denom inator is easily factored to yield the two poles at pi = 2 and p 2 However, for higher-degree polynomials, the Schur-Cohn test provides a simpler test for stability than direct factoring of //(- ) . T h e S c h u r-C o h n sta b ility te s t can b e easily p ro g ra m m e d in a d igital c o m p u ter an d it is very efficien t in te rm s o f a rith m e tic o p e ra tio n s. S pecifically, it req u ires o n ly N 2 m u ltip lic a tio n s to d e te rm in e th e co efficien ts {Km}, m = 1 , 2 .........N. The recu rsiv e e q u a tio n in (3.6.20) can b e e x p ressed in te rm s o f th e p o ly n o m ial coef ficients by e x p a n d in g th e p o ly n o m ia ls in b o th sides o f (3.6.20) a n d e q u a tin g the co efficien ts c o rre sp o n d in g to e q u a l p o w ers. In d e e d , it is easily e s ta b lis h e d that (3.6.20) is e q u iv a le n t to th e follow ing alg o rith m : Set a N (k) = ak Kfj = a f j ( N) T h e n , fo r m = N , N 1 , . . . , 1, c o m p u te K m = a m(rn) <jm_ i(0 ) = l
it
= 1 ,2 .........N
(3.6.22)
(3.6.23)
Sec. 3.6
215
an d
it = 1, 2 , . . . , m 1
(3.6.24)
w h ere bm(k) = am(m k) k = 0,1,...,m (3.6.25)
T h is recu rsiv e a lg o rith m fo r th e c o m p u ta tio n o f th e c o effic ie n ts {A T m} finds a p p lic a tio n in v a rio u s signal p ro cessin g p ro b le m s, especially in sp e e c h signal p r o cessing.
3.6.8 Stability of Second-Order Systems

In th is se c tio n w e p ro v id e a d e ta ile d an aly sis o f a sy stem h av in g tw o p o les. A s w e sh all se e in C h a p te r 7, tw o -p o le system s fo rm th e b asic b u ild in g b lo ck s fo r th e r e a liz a tio n o f h ig h e r-o rd e r system s. L e t us c o n s id e r a cau sal tw o -p o le sy stem d e s c rib e d by th e s e c o n d -o rd e r dif fe re n c e e q u a tio n y (n ) = - a ^ y ( n - I) - a 2y ( n - 2) + b 0x ( n ) T h e sy stem fu n ctio n is (3.6.26)
X (z)
1 4- a i r 1 + a 2z ~ :
(3.6.27)
boz2 z 2 + a \z + 2 T h is sy stem h a s tw o z ero s at th e origin a n d p o les a t
(3.6.28) T h e sy stem is B IB O sta b le if th e p o le s lie in sid e th e u n it circle, th a t is, if |P i| < 1 a n d \ p2\ < 1. T h e se c o n d itio n s can b e re la te d to th e v alu es o f th e co effic ie n ts a\ a n d a 2. In p a rtic u la r, th e ro o ts o f a q u a d ra tic e q u a tio n satisfy th e re la tio n s fil = (P5 + pi ) (3.6.29) (3.6.30) F ro m (3.6.29) a n d (3.6.30) w e easily o b ta in th e c o n d itio n s th a t a\ a n d a 2 m u st satisfy fo r sta b ility . F irst, aj m u st satisfy th e c o n d itio n \ai\ = \ p\ pi \ - \p\Wpi\ < 1 T h e c o n d itio n fo r a\ can b e e x p re sse d as (3.6.31)
1 1 < 1 + 2
(3.6.32)
216
Chap. 3
T h e c o n d itio n s in (3.6.31) a n d (3.6.32) can also be d e riv e d fro m th e S c h u rC o h n stab ility te st. F ro m th e recu rsiv e e q u a tio n s in (3.6.22) th r o u g h (3.6.25), we find th a t (3.6.33) an d K 2 = 02 T h e sy stem is sta b le if a n d o n iy if [AT]| < 1 an d lAÎ < 1. C o n s e q u e n tly , 1 < 2 < 1 o r eq u iv a le n tly \a% \ < 1, w hich a g rees w ith (3.6.31). A lso, (3.6.34)
or, eq u iv alen tly , ai < 1 ~b 02 d\ > 1 a 2 w hich a re in a g re e m e n t w ith (3.6.32). T h e re fo re , a tw o -p o le sy stem is sta b le if and only if th e co efficien ts a\ a n d a ; satisfy th e c o n d itio n s in (3.6.31) a n d (3.6.32). T h e stab ility c o n d itio n s given in (3.6.31) an d (3.6.32), d efin e a reg io n in the co efficien t p lan e (o i. ai), w hich is in th e fo rm o f a tria n g le , as sh o w n in Fig. 3.15. T h e sy stem is sta b le if an d only if th e p o in t (o j, q t) lies inside th e tria n g le , w hich w e call th e stability triangle. T h e ch a ra c te ristic s o f th e tw o -p o le system d e p e n d o n th e lo catio n o f the p o les o r. eq u iv alen tly , on th e lo catio n o f th e p o in t (e i, 02) in th e sta b ility triangle. T h e p o le s o f th e sy stem m ay be re a l o r co m p lex c o n ju g a te , d e p e n d in g on the v alu e o f th e d isc rim in a n t A = a* Aa2- T h e p a ra b o la a 2 = a \ f 4 splits th e stability
Figmre 3 .15 Region of stability (stability triangle) in the ( a i, a?) coefficient plane for a second-order system.
Sec. 3.6
Analysis of Linear Tim e-Invariant Systems in the ^-D om ain
217
tria n g le in to tw o reg io n s, as illu stra te d in Fig. 3.15. T h e re g io n b elo w th e p a ra b o la (,a\ > 4 d 2 ) c o rre sp o n d s to re a l an d d istin ct poles. T h e p o in ts o n th e p a ra b o la ( af = 4* 22) re su lt in re a l a n d e q u a l (d o u b le ) poles. F inally, th e p o in ts ab o v e th e p a ra b o la c o rre sp o n d to co m p lex -co n ju g ate poles. A d d itio n a l in sig h t in to th e b e h a v io r o f th e system can be o b ta in e d fro m th e u n it sa m p le re sp o n s e s fo r th e s e th re e cases.
Real and distinct poles (af = 4a2)- Since p 1 , system fu n c tio n can b e ex p re sse d in th e fo rm
A\
1 - Piz 1
pi are re a l a n d p\ ^ p2 . th e
(3.6.35)
A2 1 - P 2Z' 1 -boP2
P 1 - P2
w h ere koPi
Pi - P2 Ai ~
(3.6.36)
C o n se q u e n tly , th e u n it sa m p le re sp o n se is b ' ( P T ' - P ? l )u(n) (3.6.37)
Pi P 2
T h e re fo re , th e u n it sa m p le re sp o n se is th e d ifferen ce of tw o d e c ay in g e x p o n e n tia l se q u en ces. F ig u re 3.16 illu stra te s a ty p ical g rap h for h( n) w h en th e p o les are distinct.
Real and equal poles (af = 4a2).

system fu n ctio n is
In th is case p\
P 2 p = a \ / 2. T he
( 1 - p z - 1)2 an d h e n c e th e u n it sa m p le re sp o n se o f th e system is h{n) = b0(n + 1 ) p nu(n)

h(n)
(3.6.38)
(3.6.39)
Figure 3.16 Plot of h(n) given by (3.6.37) with p\ = 0.5, pz = 0.75; h(n) =
~ P2)](P+1 - P 2 +I)u(n).
218
Chap. 3
h(n)
Figure 3.17
Plot of h(rr) given by (3.6.39) with p =
ti(n) = (n + 1
W e o b serv e th a t h( n) is th e p ro d u c t o f a ra m p se q u e n c e a n d a real decaying e x p o n e n tia l se q u en ce. T h e g ra p h o f h( n) is show n in Fig. 3.17.
Complex-conjugate poles (af < 4a2). Since the p o le s a re co m p lex c o n ju g a te , th e system fu n ctio n can be fa c to re d and ex p re sse d as
A 1 p : -1 A 1 - re^'z' 1 + A 1 p*z~l + A' 1 - r e ~ )tU i' z ~x (3.6.40)
w h ere p = r e Jai an d 0 < coq < tt. N o te th a t w hen th e p o le s are co m p lex co n ju g ates, th e p a ra m e te rs a\ a n d 02 a re re la te d to r an d a>o acco rd in g to a\ = 2 r cos coo (3.6.41) az = r T h e c o n stan t A in th e p a rtia l-fra c tio n ex p an sio n o f H( z ) is easily sh o w n to be A = bo p p p* b 0eJ < I* j 2 sin ti>o C o n seq u en tly , th e u n it sa m p le re sp o n se o f a system w ith c o m p ie x -c o n ju g a te poles is h{n) = sin coo born sin coo boreJW " r ( e i 0 e~JW 0) (3.6.42)
2j
sin(rt -t- l)cuou(n)
-(n) (3.6.43)
In this case h ( n ) h a s an o sc illato ry b e h a v io r w ith an e x p o n e n tia lly decaying e n v elo p e w hen r < 1. T h e an g le wo o f th e p o les d e te rm in e s th e fre q u e n c y of o scillatio n an d th e d istan c e r o f th e p o le s from th e origin d e te rm in e s th e ra te of
Sec. 3.7
ft(n)
Summ ary and References
219
Figure 3.18
h(n) = [fcor"/(sin
Plot of h(n) given by (3.6.43) with bo = 1, a*> = n/4, r = 0.9; sin[(n + l)twn)u(n).
decay . W h e n r is clo se to u n ity , th e d ecay is slow . W h en r is close to th e origin, th e d ecay is fast. A typical g ra p h o f h(n) is illu stra te d in Fig. 3.18.

T h e z -tra n s fo rm p lay s th e sa m e ro le in d isc re te -tim e signals a n d sy stem s as th e L a p la c e tra n sfo rm d o e s in c o n tin u o u s-tim e signals a n d system s. In th is c h a p te r we d e riv e d th e im p o rta n t p ro p e rtie s o f th e z-tra n sfo rm , w hich a re e x tre m e ly useful in th e an aly sis o f d isc re te -tim e system s. O f p a rtic u la r im p o rta n c e is th e co n v o lu tio n p ro p e rty , w hich tra n sfo rm s th e c o n v o lu tio n o f tw o se q u e n c e s in to a p ro d u c t o f th e ir z-tran sfo rm s. In th e c o n te x t o f L T I system s, th e co n v o lu tio n p ro p e rty re su lts in th e p ro d u c t o f th e z -tra n s fo rm X ( z ) o f th e in p u t signal w ith th e sy stem fu n c tio n H ( z ) , w h e re th e la tte r is th e z -tra n s fo rm o f th e u n it sa m p le re sp o n s e o f th e sy stem . T his re la tio n sh ip allo w s u s to d e te rm in e th e o u tp u t o f an L T I sy stem in re s p o n s e to an in p u t w ith tra n s fo rm X ( z ) b y c o m p u tin g th e p ro d u c t Y ( z ) = H ( z ) X ( z ) a n d th e n d e te rm in in g th e in v e rse z -tra n sfo rm of Y ( z ) to o b ta in th e o u tp u t se q u e n c e y(n). W e o b se rv e d th a t m an y signals o f p ra c tic a l in te re s t h a v e r a tio n a l z-tran sfo rm s. M o re o v e r, L T I sy stem s c h a ra c te riz e d b y c o n s tan t-co efficien t lin e a r d ifferen ce
220
Chap. 3
e q u a tio n s , also p o ssess r a tio n a l system fu n ctio n s. C o n s e q u e n tly , in determ in in g th e in v erse z -tran sfo rm , w e n a tu ra lly e m p h a siz e d th e in v ersio n o f ra tio n a l trans fo rm s. F o r such tra n sfo rm s, th e p a rtia l-fra c tio n ex p an sio n m e th o d is relatively easy to apply, in c o n ju n ctio n w ith th e R O C , to d e te rm in e th e c o rre sp o n d in g se q u e n c e in th e tim e d o m ain . T h e o n e-sid ed z -tra n sfo rm w as in tro d u c e d to solve for th e resp o n se o f cau sal system s ex cited by causal in p u t signals w ith n o n z e ro initial co n d itio n s. F inally, w e c o n s id e re d th e c h a ra c te riz a tio n of L T I sy stem s in th e z-transform d o m a in . In p a rtic u la r, w e re la te d th e p o le - z e r o lo catio n s of a sy stem to its timed o m a in c h a racteristics an d re sta te d th e re q u ire m e n ts fo r sta b ility a n d cau sality of L T I sy stem s in te rm s o f th e p o le lo catio n s. W e d e m o n s tra te d th a t a causal system h as a system fu n ctio n H( z ) w ith a R O C |z| > w h e re 0 < r\ < oc. In a stable an d cau sal system , th e p o les o f H( z ) lie inside th e u n it circle. O n th e o th e r hand, if th e system is n o n cau sal, th e c o n d itio n for stab ility re q u ire s th a t th e u n it circle be c o n ta in e d in th e R O C o f H( z ) . H en ce a n o n cau sal sta b le L T I sy stem has a system fu n c tio n w ith p o les b o th in sid e an d o u tsid e th e u n it circle w ith an a n n u la r R O C th a t in clu d es th e u n it circle. T h e S c h u r-C o h n test fo r th e sta b ility o f a causal LTI sy stem w as d escrib ed and th e stab ility o f se c o n d -o rd e r system w as co n sid e re d in so m e d etail. A n ex cellen t c o m p reh en siv e tre a tm e n t o f the z -tra n s fo rm a n d its application to th e analysis o f L TI system s is given in the tex t by Ju ry (1964). T h e S churC o h n test for sta b ility is tr e a te d in se v era l texts. O u r p r e s e n ta tio n w as given in th e c o n tex t o f reflec tio n co efficien ts w hich are used in lin e a r p re d ic tiv e cod in g of sp e ech signals. T h e te x t by M a rk e l an d G ra y (1976) is a go o d re fe re n c e for the S c h u r-C o h n te st a n d its a p p licatio n to sp e ech signal processing.
PROBLEMS
3.1 Determ ine the z-transform of the following signals, (a) x(n) = {3. 0. 0. 0, 0, 6, 1. 4} (} )\ > 5 0. n <4 3.2 D eterm ine the z-transforms of the following signals and sketch the corresponding pole-zero patterns. (a) x ( n ) = (1 + n ) u ( n ) (b) x ( n ) = (a" + a ' ) u( n) , a real (b) x(n) =
(c) W) (e ) (I) x(n) x(n) x(n) x (n) = = = = ( 1 )n2 -n u (r) ( n a " sinaon)w(rc) (na" CQSwon) u( n) Ar " c o s (w^n + <j>)u(n). 0 < r < 1
(g) *(n) = j( n : i ~ 1) (h) jr(rt) = ( i ) n[(n) - u(n - 10)]
Chap. 3
Problems
221
n> _0
33 Determine the z-transforms and sketch the ROC of the following signals.
(a) x,(n)
(b) x 2(n) =
f(i)\ 1 ( f) " " .
"< n >0 n < 0
( i ) B- 2 \ 0, (c) *j() = x 1(n + 4 ) <d) x4(n) = j t , ( - n ) 3.4 D eterm ine (a) x(n) = (b) x(n) = (c) x(n) = (d) x(n) = <e) x(n) = (f) jr(n) =
the z-transform of the following signals. n(l)"w(n) n2u(n) nanu( n 1) (-1 )" (cos ~n) u(n) ( l)"u(n) ( 1 ,0 .- 1 ,0 . 1 ,- 1 , . . . }
t
3.5 D eterm ine the regions of convergence of right-sided, left-sided, and finite-duration two-sided sequences. 3.6 Express the z-transform of y(n) = Y x(k) k= * oc in term s of X (-). [Him: Find the difference y(n) - y(n - 1).] 3.7 Com pute the convolution of the following signals by m eans of the z-transform. *i(n)
= f ()".
1 ()"".
n >0
n<0
*2 (n) = ( j ) nH(n) 3.8 Use the convolution property to: (a) Express the z-transform of y(n) = 5 2 x tt-oc in terms of X(z).
(b) Determine the z-transform of x(n) = (n + l)u(n). [Hint. Show first that x ( n ) =
u(n) * n(n).] 3.9 The z-transform X(z) of a real signal x(n) includes a pair of complex-conjugate zeros and a pair of complex-conjugate poles. What happens to these pairs if we multiply x(n) by eJW nl (Hint. Use the scaling theorem in the z-domain.) 3.10 Apply the final value theorem to determine 1, 10,
jt( oo )
for the signal
if n is even otherwise
3.11 Using long division, determine the inverse z-transform of 1 + 2Z-1 1 - 2z 1 + z 2 if (a) jr(n) is causal and (b) x(n) is anticausal.
222
Chap. 3
3.12 Determ ine the causal signal
having the z-transform
*(z) =
1 ( l - Z z - ' H l - z - 1)2
3.13 Let ;t(n) be a sequence with z-transform X(z)- Determ ine, in terms of X(z), the z-transforms of the following signals.
(a) j : i ( n )=
if n odd 0, (b) x2(n) = x(2n) 3.14 D eterm ine the causal signal x (n) if its z-transform X(z) is given by: l+ 3 z - (a) 1 + 3 z- + 2 z-2 1 (b) 1 - z - ' + ^ z -2 I I * I I (c) (d) z -6 + z^7 I I 11 + 2z-2 1+ :-2
1 2 1 + 6z + z " 2 1.5z1 4 (1 - 2 ; - 1 + 2 z '2)(l - O.Sz"1) 1 - 1.5--1 + C.5z^: 1 + 2z + z-2
i,n
even
(e) X(z) = (0 X(z) = * * I I (g)
1 + 4z_1 + 4z-2 (h) X(z) is specified by a pole-zero p atten 1- k -1 (i) I I (j) I I

1 -az1
3.15 Determ ine all possible signals x(n) associated with the z-transform
X U ) = {1 - 2 z - ') ( 3 - z - ])
3.16 Determ ine the convolution of the following pairs of signals by means of the ztransform.
> < > <
I I
Figure P3.14
Chap. 3
Problems
223
(a) x\(ft) =
(b)
Xi(n)
= u(n),
- 1), x 2(n) = [1 + jc 2(n) = 5(n) + (i)"u (n ) * 2(71) = c-Osnnu(n)

x 2(n) = 2 "u(n 1)
(c) x\ (n) = ( i ) nM(n),
(d)
ari(n) = nu(n),
3.17 Prove the final value theorem for the one-sided z-transform. 3.18 If X(z) is the z-transform of x(n), show that: (a) Z { x m (n)} = X ' ( z ' ) (b) Z{Re[jr(n}]} = j[X (z) + X*(z*)] (c) Z{Im[jr()]l = |[X (z) - **(=*)] (d) If * * ( )= { * (? ) 1 0, then
X t U) = X (z k)
if " A integer otherwise
(e) = X(ze~Jw ) 3.19 By first differentiating X(z) and then using appropriate properties of the z-transform. determ ine x(n) for the following transforms.
(a) X(z) = l o g ( l 2 z),
\z\<{
(b) X(z) = log(l - z-1), |z! > 5 3.20 (a) Draw the pole-zero pattern for the signal jcj(n) = (r" sin a>on)u(fi) 0 < r < 1 (b) Com pute the z-transform A^tz), which corresponds to the pole-zero pattern in part (a). (c) Com pare X]{z) with X 2(z). Are they indentical? If not. indicate a m ethod lo derive Xi(z) from the pole-zero pattern. 3.21 Show that the roots of a polynomial with real coefficients are real or form complexconjugate pairs. The inverse is not true, in general.
3.22 Prove the convolution and correlation properties of the z-transform using only its
definition.
3.23 D eterm ine the signal x(n) with z-transform

X (z) = e: + e l/z
|z|^0
3.24 D eterm ine, in closed form, the causal signals x(n) whose z-transforms are given by:
(a) X(z) = T T T 5 ^ T
o3
(b) X(Z) = 1 - 0.5z~! + 0.6z 2 Partially check your results by computing *(0), x (l), *(2), and m ethod. 1
jt( oo)
by an alternative
3.25 D eterm ine all possible signals that can have the following z-transforms.
224
The z - T ransform and Its Application to the Analysis of LTI Systems
Chap. 3
3.26 Determ ine the signal x(n) with z-transform XU) = if X(z) converges on the unit circle.
3 .2 7 Prove the complex convolution relation given by (3.2.22). 3.28 Prove the conjugation properties and Parse vals relation for the z-transform given in
1-
+ z- 2
Table 3.2. 3.29 In Example 3.4.1 we solved for .r(n), n < 0, by perform ing contour integrations for each value of n. In general, this procedure proves to be tedious. It can be avoided by making a transform ation in the contour integral from z-plane to the uj = 1/z plane. Thus a circle of radius R in the z-plane is mapped into a circle of radius 1/ R in the wplane. As a consequence, a pole inside the unit circle in the z-plane is mapped into a pole outside the unit circle in the m-plane. By making the change of variable w = 1/z in the contour integral, determ ine the sequence x ( n ) for n < 0 in Example 3.4.1,
3.30 Let *(n), 0 <
n < N 1 be a finite-duration sequence, which is also real-valued and even. Show that the zeros of the polynomial X(z) occur in mirror-image pairs about the unit circle. That is. if z = rej/> is a zero of X(z), then z = (1 / r ) e J" is also a zero.
3.31 Compute the convolution of the following pair of signals in the time domain and by
using the one-sided z-transform. (a) .v,(n) = {1. 1. 1. 1. 1). x 2(n) = (1. 1. 1)
t t
(b) Xi (n) = ( j)"u(h). (c) .ii(n) = (1.2, 3.4}.

t
x 2(n) = (i)"u(n) x 2(n ) =

Jr2 ( )
(4,
3, 2. 1}
t
(d) *,() = {1.1. 1.1.1}.

t
= {1.1,1}
t
Did you obtain the same results by both methods? Explain.

3.32 D eterm ine the one-sided z-transform of the constant signal x ( n ) = 1. oo < n < 00. 3 .3 3 Prove that the Fibonacci sequence can be thought of as the impulse response of the system described by the difference equation ,y(/i) = v(n - 1) + y ( n - 2) 4- x(n). Then
determ ine h(n) using z-transform techniques.

3 .3 4 Use the one-sided z-transform to determ ine y(n), n > 0 in the following cases.
(a) y(n) + \ y( n - 1) - \ y( n - 2) = 0; (c) v(n) = \ y ( n - 1) + x ( n )

x ( n ) = ()"u(rt). y(-l) = 1
y ( - l ) = y (-2 ) = 1 y ( - l ) = 1. v( 2) = 0
(b) y(n) - 1.5y(n - 1) + 0.5y(n - 2) = 0;
(d) _v(r) = j v(n - 2) + x(n)

x ( n ) = u{n)
> (1) = 0;
y( 2) = 1
3.3 5 Show that the following systems are equivalent. (a) y ( n ) = 0 .2 y (n - 1) + x { n ) - 0.3.r(n - 1) + 0.02jt(n - 2)
(b) _y{) = x ( n ) - 0 .1 x (n - 1)
Chap. 3
Problems
225
3.36 Consider the sequence xin) = ci"uin). 1 < a < 1. D eterm ine at least two sequences that are not equal to xin) but have the same autocorrelation. 3.37 Com pute the unit step response of the system with impulse response f3\ n < 0 n >0
3.38 Com pute the zero-state response for the following pairs of systems and input signals. (a) hin) = ,v(h) = <j)"^cos w r j ;<('?)
(b) h(n) = ( | Yu Oi ). xi n) = { \ ) u{n) + ( \ r " u ( - n - 1) (c) yin) = 0.1 yin 1) + 0.2y(>i - 2) + xin) + xin - 1)
,v ( / i) = ( ^ V m (i i )
(d) yin) = |.v() -
1)
x{n) = lO^Cos n'ju(n) (e) yin) = yin 2 ) + lOjr(n) ,v(/7) = lO^cos n^jnin)
(f) h{n) = i ^ Y ii(n). x i n ) i/ln) it in 7)
(g) hin) = (|V'if(H). xin) = ( 1)", 3C < n < x (h) hin) = U) ' i 3.39 Consider the system
H(Z) =
= in + 1)(j)"h<h)
1 - 2 r ' + 2:~: ----------;---- ----------- : -----------------( 1 - 0 . 5 ; - ) ( l - 0. 2: - ' )
ROC: 0.5 < |d < l
(a) Sketch the pole-zero pattern. Is the system stable? (b) D eterm ine the impulse response of the system. 3.40 Com pute the response of the system ytn) = 0.7v(n - 1) 0.12y(n - 2) + x(n 1) + xin 2) to the input v(/?) = nuin), Is the system stable? 3.41 D eterm ine the impulse response and the step response of the following causal systems. Plot the pole-zero patterns and determ ine which of the systems are stable.
(a ) y i n ) = j v( - 1) - jr v( n - 2) + x ( n )
(b) y(n) = v in 1) 0.5y(n - 2) + x(n) + xin I )

; ~ ' f l +
(c> = (d) v(n) = 0.6y(n 1) 0.08v(n 2) -f- x(n) (e) v() = 0.7y(n 1) 0.1y(/3 2) + 2xin) x(n 2) 3.42 Let xin) be a causal sequence with ;-transform X(z) whose pole-zero plot is shown in Fig. P3.42. Sketch the pole-zero plots and the R O C of the following sequences; (a) Xt(n) = x ( ~ n -)- 2) (b) x 2(n) = eim/iu,x(n)
226

Im(z)
Chap. 3
Re(;)
Figure P3.42 3.43 W e want to design a causal discrete-time LTI system with the property that if the
input is
x( n ) = ( l ) nu ( n) - 1)
then the output is V(lt) = (|)"tt(n) (a) D eterm ine the impulse response h(n) and the system function H(z) of a system that satisfies the foregoing conditions. (b) Find the difference equation that characterizes this system. (c) D eterm ine a realization of the system that requires the minimum possible amount of memory. (d) D eterm ine if the system is stable. 3 .4 4 D eterm ine the stability region for the causal system 1 + o ir -1 + a2z~2 by computing its poles and restricting them to be inside the unit circle.
3.45 Consider the svstem
H{z) =
_ L J?i-r-2 1 _ l --1 + I Determine: (a) The impulse response (b) The zero-state step response (c) The step response if y(1) = 1 and y ( -2 ) = 2 3 .4 6 D eterm ine the system function, impulse response, and zero-state step response of the system shown in Fig P3.46.
3 .4 7 Consider the causal system
H{z) =
y in ) = - a i y i n - 1) + b0x(n) + b^xin - 1)
Determine: (a) The impulse response
Chap. 3
Problems
227 y(n)
X i.f l)
a
Figure P3.46
(b) The zero-state step response

(c) The step response if y ( 1) = A / 0 (d) The response to the input
x (n ) ~ CO S a> u f!
<
<
oo
3.48 Determ ine the zero-state response of the system

y(n) = i v(h - 1) + 4*(n) + 3x(n 1)
to the input jt(n) = em 'nu{n) W hat is the steady-state response of the system? 3.49 Consider the causal system defined by the pole-zero pattern shown in Fig. P3.49. (a) Determ ine the system function and the impulse response of the system given that W U )U i = 1. (b) Is the system stable? (c) Sketch a possible im plementation of the system and determ ine the corresponding difference equations. Im(;)
3.50 An FIR LTI system has an impulse response h(n), which is real valued, even, and has finite duration of 2/V + 1. Show that if ;i = rejaJa is a zero of the system, then d = ( l / r ) e j a *> is also a zero. 3.51 Consider an LTI discrete-time system whose pole-zero pattern is shown in Fig. P3.51. (a) D eterm ine the R O C of the system function H(z) if the system is known to be stable.
228
Chap. 3
Im(;)
-0 .5
Re<z)
Figure P3.51
(b ) It is possible for the given pole-zero plot to correspond to a causal and stable
system? If so, what is the appropriate ROC? (c) How many possible systems can be associated with this pole-zero pattern? 3.52 Let x(n) be a causal sequence. (a) What conclusion can you draw about the value of its z-transform A(c) at ; = oo? (b ) Use the result in part (a) to check which of the following transform s cannot be associated with a causal sequence.
3.53 A causal pole-zero system is BIBO stable if its poles are inside the unit circle. Con sider now a pole-zero system that is BIBO stable and has its poles inside the unit circle. Is the system always causal? [Hint: Consider the systems h\(n) = anu(n ) and ti2 (n) = anu{n + 3), |a| < 1.] 3 3 4 Let *(/i) be an anticausal signal [i.e., x (n) = 0 for n > 0]. Form ulate and prove an initial value theorem for anticausal signals. 3.55 The step response of an LTI system is J(w) = + 2)
(a) Find the system function H(z) and sketch the pole-zero plot.
(b ) D eterm ine the impulse response ft(n).
(c) Check if the system is causal and stable. 3 3 6 Use contour integration to determ ine the sequence jc(n) whose z-transform is given
Chap. 3
Problems
229
(c) X'U) = --------<d) A'C)
; a 1 a: 1-
3.57 Let ,v(h) be a sequence with c-transform

X d = -- --------- (1 - t;r)(l - a ; " 1) R O C : a < |;| < 1ja
with 0 < a < 1. Determ ine ,v() by using contour integration
3.58 The ^-transform of a sequence .v(n) is given by X(;> =
(: - i ) ( ; - 2)5< - + ?): (r + 3)
Furthermore it is known that X (:) converges for |;| = 1. (a ) Determ ine the R O C of X(c). (b) Determ ine xin ) at n = -18. {Hint: Use contour integration.)
34
Frequency Analysis of Signals and Systems
T h e F o u rie r tra n sfo rm is o n e o f se v e ra l m a th e m a tic a l to o ls th a t is useful in the an alysis an d d esign o f L T I system s. A n o th e r is th e F o u rie r se ries. T h ese signal re p re se n ta tio n s basically involve th e d e c o m p o sitio n o f th e sig n als in te rm s o f sinu so id al (o r co m p lex e x p o n e n tia l) c o m p o n e n ts. W ith such a d e c o m p o s itio n , a signal is said to be re p re se n te d in th e f r e q u e n c y domai n . A s w e shall d e m o n s tra te , m ost signals o f p ractical in te re st can be d eco m p o sed in to a sum o f sin u so id al signal c o m p o n e n ts. F o r th e class of p e rio d ic signals, such a d eco m p o sitio n is called a Fouri er series. F o r th e class o f finite e n e rg y signals, the d eco m p o sitio n is called th e Fouri er t ransf orm. T h e se d e c o m p o s itio n s are ex trem ely im p o rta n t in th e an aly sis o f L T I system s becau se th e re sp o n se o f a n L T I system to a sin u so id al in p u t signal is a sin u so id o f th e sam e freq u e n c y b u t o f d iffe re n t am pli tu d e a n d p h ase. F u rth e rm o re , th e lin e a rity p ro p e rty o f th e L T I sy stem im plies th a t a lin ear sum o f sin u so id al c o m p o n e n ts at th e in p u t p ro d u c e s a sim ilar lin e a r sum o f sin u so id al c o m p o n e n ts a t th e o u tp u t, w hich d iffer only in th e a m p litu d e s and p h ases fro m th e in p u t sin u so id s. T h is c h a ra c te ristic b e h a v io r o f L T I sy stem s ren d e rs th e sin u so id al d e c o m p o sitio n o f signals v ery im p o rta n t. A lth o u g h m an y o th er d e c o m p o sitio n s o f signals a re p o ssib le, only th e class o f sin u so id al (o r co m p lex ex p o n e n tia l) signals p o ssess th is d e sira b le p ro p e rty in passin g th ro u g h an L T I system . W e begin o u r stu d y o f fre q u e n c y analysis o f signals w ith th e re p re se n ta tio n o f co n tin u o u s-tim e p e rio d ic a n d a p e rio d ic signals by m e a n s of th e F o u rie r series an d th e F o u rie r tra n sfo rm , resp ec tiv ely . T h is is follow ed by a p a ra lle l tre a tm e n t o f d isc rete-tim e p erio d ic a n d a p e rio d ic signals. T h e p r o p e rtie s o f th e F o u rie r tra n sfo rm a re d e sc rib e d in d e ta il an d a n u m b e r of tim e -fre q u e n c y d u alities are p re se n te d .
4.1 FREQUENCY ANALYSIS OF CONTINUOUS-TIME SIGNALS

It is w ell k n o w n th a t a prism can be u sed to b re a k u p w h ite light (su n lig h t) in to the co lo rs o f th e ra in b o w (see Fig. 4.1a). In a p a p e r su b m itte d in 1672 to th e R oyal Society, Isaac N e w to n u se d th e te rm spe ct r um to d escrib e th e c on t i n u o u s b an d s 230
Sec. 4.1
Frequency Analysis of Continuous-Time Signals
231
Glass prism
Figure 4.1
(a) Analysis and
( b ) sy n th e sis o f th e w h ite light (s u n lig h t) u sin g glass p rism s.
o f co lo rs p ro d u c e d by this a p p a ra tu s. T o u n d e rsta n d this p h e n o m e n o n , N ew to n p laced a n o th e r p rism u pside-dow rn w ith resp ec t to th e first, an d sh o w ed th a t th e co lo rs b le n d e d back into w hite light, as in Fig. 4.1b, By in se rtin g a slit b etw een the tw o p rism s a n d blo cking o n e o r m o re colors from h ittin g the se co n d prism , he sh o w ed th a t th e rem ix ed light is no lo n g er w hite. H e n c e th e light passing th ro u g h th e first p rism is sim ply an aly zed into its c o m p o n e n t co lo rs w ith o u t any o th e r ch an g e. H o w e v e r, o n ly if we mix again all o f th ese c o lo rs d o we o b ta in the o rig in a l w hite light.
L a te r, Jo s e p h F r a u n h o f e r ( 1 7 8 7 - 1 8 2 6 ) . in m a k in g m e a s u r e m e n ts o f lig h t e m itt e d b y th e su n a n d s ta r s , d is c o v e r e d th a t th e s p e c tr u m o f th e o b s e r v e d lig h t c o n s is ts o f d is tin c t c o l o r lin e s . A fe w y e a r s la te r ( m i d - ] 8 0 0 s ) G u s ta v K i r c h h o f f a n d R o b e r t B u n s e n fo u n d th a t e a c h c h e m ic a l e le m e n t, w h e n h e a te d to in c a n d e s c e n c e , r a d ia te d its o w n d is tin c t c o lo r o f lig h t. A s a c o n s e q u e n c e , e a c h c h e m ic a l e le m e n t c a n b e id e n tifie d b y its o w n lin e sp e ctru m . F r o m p h y s ic s w e k n o w th a t e a c h c o lo r c o r r e s p o n d s to a s p e c ific f r e q u e n c y o f th e v is ib le s p e c tr u m . H e n c e th e a n a ly sis o f lig h t in to c o lo r s is a c tu a lly a f o r m o f
fr e q u e n c y a n a ly s is .
F r e q u e n c y a n a ly s is o f a s ig n a l in v o lv e s th e r e s o lu tio n o f th e s ig n a l in to its fr e q u e n c y ( s in u s o id a l) c o m p o n e n ts . to o ls th a t w e w ill d e v e lo p : In s te a d o f lig h t, o u r s ig n a l w a v e fo r m s a re The b a s ic a lly fu n c tio n s o f tim e . T h e r o le o f th e p ris m is p la y e d b y th e F o u r i e r a n a ly sis th e F o u r ie r s e r ie s a n d th e F o u r i e r t r a n s f o r m . r e c o m b in a tio n o f th e s in u s o id a l c o m p o n e n ts to r e c o n s tr u c t th e o r ig in a l s ig n a l is b a s ic a lly a F o u r i e r s y n th e s is p r o b le m . T h e p r o b le m o f s ig n a l a n a ly s is is b a s ic a lly th e s a m e fo r th e c a s e o f a s ig n a l w a v e fo rm a n d f o r th e c a s e o f th e lig h t f r o m h e a te d c h e m ic a l c o m p o s itio n s . J u s t a s in th e c a s e o f c h e m ic a l c o m p o s itio n s , d if f e r e n t s ig n a l w a v e fo r m s h a v e d iffe r e n t s p e c tr a . T h u s th e s p e c tr u m p r o v id e s a n id e n tity
232
Chap. 4
o r a s ig n a tu r e f o r th e s ig n a l in th e s e n s e t h a t n o o t h e r s ig n a l h a s th e s a m e sp e c tru m . A s w e w ill s e e , th is a ttr ib u te is r e la te d to th e m a th e m a tic a l t r e a t m e n t o f fre q u e n c y d o m a in te c h n iq u e s . I f w e d e c o m p o s e a w a v e fo r m in to s in u s o id a l c o m p o n e n ts , in m u c h th e sa m e w ay t h a t a p ris m s e p a r a te s w h ite lig h t in to d iffe r e n t c o lo r s , th e su m o f th e s e s in u s o id a l c o m p o n e n ts re s u lts in th e o r ig in a l w a v e fo r m . O n th e o t h e r h a n d , if any o f th e s e c o m p o n e n ts is m is s in g , th e r e s u lt is a d if f e r e n t s ig n a l. I n o u r tr e a t m e n t o f fr e q u e n c y a n a ly s is , w e w ill d e v e lo p th e p r o p e r m a th e m a tic a l to o ls ( p r is m s ) f o r th e d e c o m p o s itio n o f s ig n a ls ( l i g h t ) in to s in u so id a l f r e q u e n c y c o m p o n e n ts ( c o l o r s ) . F u r t h e r m o r e , th e t o o ls ( in v e r s e p r is m s " ) f o r sy n th e s is o f a g iv e n s ig n a l f r o m its f r e q u e n c y c o m p o n e n ts w ill a ls o b e d e v e lo p e d . T h e b a s ic m o t iv a tio n f o r d e v e lo p in g th e fr e q u e n c y a n a ly s is to o ls is to p ro v id e a m a th e m a tic a l a n d p ic to r ia l r e p r e s e n t a t io n f o r th e f r e q u e n c y c o m p o n e n ts th a t a re c o n ta in e d in a n y g iv e n s ig n a l. A s in p h y s ic s , th e te r m s pe ct r um is u s e d w h e n r e f e r rin g t o th e fr e q u e n c y c o n t e n t o f a s ig n a l. T h e p r o c e s s o f o b ta in in g th e s p e c tru m o f a g iv e n sig n a l u sin g th e b a s ic m a th e m a tic a l t o o ls d e s c r ib e d in th is c h a p te r is k n o w n a s f re q u e n c y o r spectral analysis. is c a lle d spect rum estimation. In c o n tr a s t, th e p r o c e s s o f d e te r m in in g In a p r a c tic a l th e s p e c tr u m o f a s ig n a l in p r a c tic e , b a s e d o n a c tu a l m e a s u r e m e n ts o f th e sig n a l, T h is d is tin c tio n is v e r y im p o r ta n t. p r o b le m th e s ig n a l to b e a n a ly z e d d o e s n o t le n d it s e lf to a n e x a c t m a th e m a tic a l d e s c r ip tio n . T h e s ig n a l is u s u a lly s o m e i n f o r m a tio n - b e a r in g s ig n a l fr o m w h ich we a r e a tte m p tin g t o e x t r a c t th e r e le v a n t in f o r m a tio n . I f th e in f o r m a tio n th a t w e wish to e x t r a c t c a n b e o b ta in e d e i t h e r d ir e c tly o r in d ir e c tly fr o m th e s p e c tr a l c o n te n t o f th e s ig n a l, w e c a n p e r fo r m spe ct rum est imation o n th e in f o r m a t io n - b e a r in g sig n a l, a n d th u s o b ta in a n e s tim a te o f th e s ig n a l s p e c tr u m . In f a c t, w e c a n v iew s p e c tra l e s tim a tio n as a ty p e o f s p e c tr a l a n a ly s is p e r fo r m e d o n s ig n a ls o b t a i n e d f r o m p h y si ca l s o u r c e s (e .g ., s p e e c h , E E G , E C G , e t c .) . T h e in s tr u m e n ts o r s o f tw a r e p r o g ra m s u se d t o o b ta in s p e c tr a l e s tim a te s o f s u c h s ig n a ls a r e k n o w n a s sp e c t r u m analyzers. H e r e , w e w ill d e a l w ith s p e c tr a l a n a ly s is . H o w e v e r , in C h a p t e r 12 w e sh all t r e a t th e s u b je c t o f p o w e r s p e c tr u m e s tim a tio n .
4.1.1 The Fourier Series for Continuous-Time Periodic Signals

In th is s e c tio n w e p r e s e n t th e f r e q u e n c y a n a ly s is to o ls fo r c o n tin u o u s - tim e p e r io d ic s ig n a ls . e x p o n e n tia ls . T h e b a s ic m a th e m a tic a l r e p r e s e n ta tio n o f p e r io d ic s ig n a ls is th e F o u r i e r s e r ie s , w h ic h is a lin e a r w e ig h te d su m o f h a r m o n ic a lly r e la te d s in u s o id s o r c o m p le x e x p o n e n tia ls . J e a n B a p t is t e J o s e p h F o u r i e r ( 1 7 6 8 - 1 8 3 0 ) , a F r e n c h m a th e m a tic ia n , u se d s u c h t r ig o n o m e t r ic s e r ie s e x p a n s io n s in d e s c r ib in g th e p h e n o m e n o n o f h ea t c o n d u c tio n a n d te m p e r a tu r e d is tr ib u tio n th r o u g h b o d ie s . A lth o u g h h is w o r k was m o t iv a te d b y th e p r o b le m o f h e a t c o n d u c tio n , th e m a th e m a tic a l te c h n iq u e s th a t E x a m p le s o f p e r io d ic s ig n a ls e n c o u n te r e d in p r a c t i c e a r e s q u a re w a v e s , r e c ta n g u la r w a v e s , tr ia n g u la r w a v e s , a n d o f c o u r s e , s in u s o id s a n d c o m p le x
Sec. 4.1
Frequency Analysis of C ontinuous-Tim e Signals
233
h e d e v e lo p e d d u rin g th e e a r ly p a r . o f th e n in e t e e n t h c e n tu r y n o w fin d a p p lic a tio n in a v a r ie ty o f p r o b le m s e n c o r r .r \ is s in g m a n y d if f e r e n t f ie ld s , in c lu d in g o p tic s , v ib r a tio n s in m e c h a n ic a l s y s te m s , s y s t e m th e o r y , a n d e le c t r o m a g n e t ic s . F r o m C h a p t e r 1 w e r e c a ll t h .i : a lin e a r c o m b in a t io n o f h a r m o n ic a lly r e la te d c o m p le x e x p o n e n tia ls o f th e fo r m
x {r) =
cke j 2 * kF
1/Fo.
( 4 .1 .1 )
i = -3C
is a p e r io d ic s ig n a l w ith f u n d a m e n t a l p e r io d Tp = th e e x p o n e n t ia l s ig n a ls H e n c e w e c a n th in k o f
{ e i i x k p k = Q i i 2 l
a s th e b a s ic b u ild in g b lo c k s f r o m w h ic h w e c a n c o n s t r u c t p e r io d ic s ig n a ls o f s p e c ify W e ca n
v a r io u s ty p e s b y p r o p e r c h o ic e o f t h e fu n d a m e n ta l f r e q u e n c y a n d th e c o e f f ic ie n ts ( q ). F o d e te r m in e s th e f u n d a m e n ta l p e r i o d o f x ( t ) a n d th e c o e f f ic ie n t s { } th e s h a p e o f th e w a v e fo rm . S u p p o s e th a t w e a r e g iv e n a p e r io d i c s ig n a l x { i) w ith p e r io d Tp . r e p r e s e n t th e p e r io d ic sig n a l by t h e s e r ie s ( 4 .1 .1 ) , c a lle d a F o u r ie r series, w 'here th e f u n d a m e n ta l fr e q u e n c y Fo is s e l e c t e d to b e th e r e c ip r o c a l o f th e g iv e n p e r io d
Tp . T o d e te r m in e th e e x p r e s s io n t o r th e c o e f f i c i e n t s ( q ) , w e firs t m u ltip ly b o th
sid e s o f ( 4 .1 .1 ) b y th e c o m p le x e x p o n e n t i a l
F ltl!
w h e r e I is a n in t e g e r an d th e n in t e g r a t e b o th s id e s o f th e r e s u ltin g e q u a tio n o v e r a s in g le p e r io d , s a y fr o m 0 to T r , o r m o r e g e n e r a lly , fr o m fo to r0 + T p, w h e r e i o is a n a r b itr a r y b u t m a th e m a tic a lly c o n v e n i e n t s ta r t in g v a lu e . T h u s w e o b ta in fh>+Tr r'o+Tr I oc \
J'
x ( t ) e - j2*IF,''dt = J'
g- j t oi Kt /
cke+J2nkF"' J di
(4.1.2)
T o e v a lu a te th e in te g r a l o n th e r ig h t- h a n d s id e o f ( 4 .1 .2 ) , w e in te r c h a n g e th e o r d e r o f th e s u m m a tio n a n d in te g r a tio n a n d c o m b in e th e tw o e x p o n e n tia ls . H e n c e

sc
rl
c*
,, + r .
OC
p j27rF < M -hl
- |'n + r ,
i= oc
ei2* F {k-'"dt =
k= oc
Ck
J l n F o i k - /)_
( 4 .1 .3 )
F o r k ^ I, th e r ig h t-h a n d s id e o f ( 4 .1 .3 ) e v a lu a te d a t th e lo w e r a n d u p p e r lim its , Z o a n d t0 + Tp , r e s p e c tiv e ly , y ie ld s z e r o . O n th e o t h e r h a n d , if k = /, w e h a v e
f Jtn
Consequently, (4.1.2) reduces to
' o+Tp
di i
f J /ft
x ( t ) e - j2 * ,Fo'd t = c ,T p
234
Frequency Analysis of Signals and System s
Chap. 4
a n d t h e r e f o r e th e e x p r e s s io n f o r th e F o u r i e r c o e f fic ie n ts in t e r m s o f th e g iv e n p e r io d ic s ig n a l b e c o m e s
I c, =
Tp
fto+TP x ( l ) e - jlw,F"'dt
S in c e fo is a r b itr a r y , th is in te g r a l c a n b e e v a lu a te d o v e r a n y i n te r v a l o f le n g th Tp, th a t is, o v e r a n y in te r v a l e q u a l t o th e p e r io d o f th e s ig n a l jr ( r ) . C o n s e q u e n tly , th e in te g r a l fo r th e F o u r i e r s e r ie s c o e f f ic ie n t s w ill b e w r itte n as
c, = f
Tp J t
x ( t ) e ~ j2nlF,>'d t
( 4 .1 .4 )
A n im p o r ta n t is s u e th a t a r is e s in th e r e p r e s e n ta tio n o f th e p e r io d ic s ig n a l
x ( t ) b y th e F o u r ie r s e r ie s is w h e t h e r o r n o t th e s e r ie s c o n v e r g e s t o * ( f ) f o r e v e ry
v a lu e o f r, th a t is, if th e s ig n a l x (t ) a n d its F o u r ie r s e r ie s r e p r e s e n t a t io n
OC
c ke j2 * kFtt'
( 4 .1 .5 )
A=-oc
a r e e q u a l a t e v e r y v a lu e o f t. d is c o n tin u o u s . T h e s o -c a lle d D ir ic h le t c o n d it io n s g u a r a n te e th a t th e s e r ie s ( 4 .1 .5 ) w ill b e e q u a l to x ( t ), e x c e p t a t th e v a lu e s o f i f o r w h ich jc (r ) is A t th e s e v a lu e s o f r, ( 4 .1 .5 ) c o n v e r g e s t o th e m id p o in t (a v e r a g e v a lu e ) o f th e d is c o n tin u ity . T h e D i r ic h l e t c o n d itio n s a r e : 1 . T h e s ig n a l .r ( f ) h a s a fin ite n u m b e r o f d is c o n tin u itie s in a n y p e r io d . 2 . T h e sig n a l x ( t ) c o n ta in s a fin ite n u m b e r o f m a x im a an d m in im a d u rin g an y p e rio d . 3 . T h e s ig n a l x (t ) is a b s o lu te ly in te g r a b le in a n y p e r io d , t h a t is.
f
JT r
\x (t )\d t < oo
( 4 .1 .6 )
A lt p e r io d ic s ig n a ls o f p r a c tic a l i n te r e s t s a tis fy th e s e c o n d itio n s . T h e w e a k e r c o n d itio n , t h a t th e s ig n a l h a s fin ite e n e r g y in o n e p e r io d ,
j
J tp
\x ( t ) \2d i < o c
( 4 .1 .7 )
g u a r a n te e s th a t th e e n e r g y in th e d if f e r e n c e s ig n a l
OC
e (t) = x { t ) k= oc
Ckej2kF"'
is z e r o , a lth o u g h * ( ; ) a n d its F o u r i e r s e r ie s m a y n o t b e e q u a l f o r a ll v a lu e s o f t. N o te th a t ( 4 .1 .6 ) im p lie s ( 4 .1 .7 ) , b u t n o t v ic e v e r s a . A l s o , b o t h ( 4 .1 .7 ) a n d th e D ir i c h le t c o n d itio n s a r e s u f fic ie n t b u t n o t n e c e s s a r y c o n d itio n s ( i .e ., th e r e a r e s ig n a ls t h a t h a v e a F o u r ie r s e r ie s r e p r e s e n t a t i o n b u t d o n o t s a tis fy t h e s e c o n d itio n s ) . In s u m m a r y , i f x ( t ) is p e r io d ic a n d s a tis fie s th e D i r ic h l e t c o n d itio n s , it ca n b e r e p r e s e n te d in a F o u r i e r s e r ie s a s in ( 4 .1 .1 ) , w h e r e th e c o e f f i c i e n t s a r e s p e c ifie d b y ( 4 .1 .4 ) . T h e s e r e la t io n s a r e s u m m a r iz e d b e lo w .
Sec. 4.1

FR E Q U E N C Y A N A LY S IS O F C O N T IN U O U S -TIM E PE R IO D IC S IG N A LS
235
Synthesis equation
-vU) = ^
cke,2~kl'"'
(4.1.8)
Analysis equation
ct =
j 1r Jrp
(4.1.9)
In g e n e r a l, th e F o u r i e r c o e ff ic ie n ts c k a r e c o m p le x v a lu e d . e a s ily sh o w n t h a t if th e p e r io d ic s ig n a l is r e a l, c k a n d A s a r e s u lt, if
M o r e o v e r , it is
a r e c o m p le x c o n ju g a te s .
ct = \ck\e}b
th e n
C-U ~ ~
C o n s e q u e n tly , th e F o u r i e r s e r ie s m ay a ls o b e r e p r e s e n te d in t h e fo r m
rv
a (M = (-(, +
k i I c o s 0 -7 1 k
F()t
6k )
( 4 .1 .1 0 )
n-i
w h e r e t,, is r e a l v a lu e d w h e n x ( i) is r e a l. F in a lly , w e s h o u ld in d ic a te th a t y e t a n o th e r fo r m f o r th e F o u r ie r s e r ie s ca n b e o b ta in e d by e x p a n d in g th e c o s in e fu n c tio n in ( 4 .1 .1 0 ) as c o s ( 2 jt A / v + 0k) c o s 2 n k F o i c o s 0t s i n l n k F ^ r s in fy C o n s e q u e n tly , w e c a n re w r ite ( 4 .1 .1 0 ) in th e fo rm
3t
ji ( r ) ao + Y 2 ( a ic c o $ 2 n k F o t bk s i n 2 -n k F o i) ( 4 .1 .1 1 )
1=1
w h ere < 3() = Co fl* = 2|C|- [ c o s 0k
bk = 2|q| sin # *
T h e e x p r e s s io n s in ( 4 .1 .8 ) , ( 4 ,1 .1 0 ) , a n d ( 4 .1 ,1 1 ) c o n s t itu te t h r e e e q u iv a le n t fo r m s f o r th e F o u r i e r s e r ie s r e p r e s e n ta tio n o f a re a l p e r io d ic s ig n a l.
4.1.2 Power Density Spectrum of Periodic Signals

A p e r io d ic s ig n a l h a s in fin ite e n e r g y a n d a fin ite a v e r a g e p o w e r , w h ic h is g iv e n a s
Px = ~
f
P 'T .
\x ( t ) \2d t
( 4 .1 .1 2 )
236
Chap. 4
I f w e ta k e th e c o m p le x c o n ju g a te o f ( 4 .1 .8 ) an d s u b s t itu te f o r * * ( / ) in ( 4 .1 .1 2 ) . we o b ta in
OC
( 4 .1 .1 3 )
DC
E k~ 3 C
T h e r e f o r e , w e h a v e e s ta b lis h e d th e r e la tio n ( 4 .1 .1 4 ) w h ich is c a lle d P a r s e v a l's re la tio n fo r p o w e r s ig n a ls . T o illu s tr a te th e p h y s ic a l m e a n in g o f ( 4 .1 .1 4 ) , s u p p o s e th a t .v(r) c o n s is ts o f a s in g le c o m p le x e x p o n e n tia l
x (t) = c , e j27!tFn'
In th is c a s e , all th e F o u r ie r s e r ie s c o e f f ic ie n ts e x c e p t c* a r e z e r o . th e a v e r a g e p o w e r in th e sig n a l is C o n s e q u e n tly ,
It is o b v io u s th a t | q | : r e p r e s e n ts th e p o w e r in th e Ath h a r m o n ic c o m p o n e n t o f th e sig n a l. H e n c e th e to ta l a v e r a g e p o w e r in th e p e r io d ic sig n a l is s im p ly th e su m o f th e a v e r a g e p o w e r s in a ll th e h a r m o n ic s . I f w e p lo t th e |q|2 as a fu n c t io n o f th e f r e q u e n c ie s kF o, k = 0 , 1 , 2 ..........th e d ia g r a m th a t w e o b ta in sh o w s h o w th e p o w e r o f th e p e r io d ic s ig n a l is d is tr ib u te d a m o n g th e v a r io u s f r e q u e n c y c o m p o n e n ts . T h is d ia g r a m , w h ic h is illu s tr a te d in F ig . 4 .2 , is c a lle d th e p o w e r d e n sity sp e ctru m * o f th e p e r io d ic s ig n a l x ( t ). S in c e th e

Pow er density spectrum lct P
- 4 F 0 - 3 F 0 2Fn F0 Figure 4.2
Fn
2F0 3F0 4F0
Frequency. F
Pow er density spectrum of a continuous-tim e periodic signal.
This function is also called the power spectral density or. simply, the power spectrum.
Sec. 4.1
237
p o w e r in a p e r io d ic s ig n a l e x is ts o n ly at d is c r e te v a lu e s o f f r e q u e n c ie s ( i .e .. F 0.
F o . 2 F q . . . . ) . th e s ig n a l is s a id to h a v e a lin e sp e ctru m . T h e s p a c in g b e tw e e n
tw o c o n s e c u tiv e s p e c tr a l lin e s is e q u a l to th e r e c ip r o c a l o f th e fu n d a m e n ta l p e rio d
Tp . w h e r e a s th e s h a p e o f th e s p e c tr u m ( i.e .. th e p o w e r d is tr ib u tio n o f th e s ig n a l),

d e p e n d s o n th e tim e -d o m a in c h a r a c t e r is t ic s o f th e sig n a l. A s in d ic a te d in th e p r e c e d in g s e c t i o n , th e F o u r ie r s e r ie s c o e f f ic ie n ts ( q ) a re c o m p le x v a lu e d , th a t is. th e y c a n b e r e p r e s e n te d as
C k = \ck\eJhL
w h ere
6k = 4-Q
In s te a d o f p lo ttin g th e p o w e r d e n sity s p e c tr u m , w e c a n p lo t th e m a g n itu d e v o lta g e s p e c tr u m {|ot|} a n d th e p h a s e s p e c tr u m {(?*} as a fu n c tio n o f f r e q u e n c y . C le a r ly , th e p o w e r s p e c tr a l d e n s ity in th e p e r io d ic s ig n a l is s im p ly th e s q u a r e o f th e m a g n itu d e s p e c tr u m . T h e p h a s e in fo r m a tio n is to ta lly d e s tr o y e d ( o r d o e s n o t a p p e a r ) in th e p o w e r s p e c tr a l d e n s ity . I f th e p e r io d ic sig n a l is re a l v a lu e d , th e F o u r ie r s e r ie s c o e f f ic ie n ts { c * } sa tis fy th e c o n d itio n
c -k =
C o n s e q u e n tly . Ki|: = |q|: . H e n c e th e p o w e r s p e c tru m is a s y m m e tr ic fu n c tio n o f f r e q u e n c y . T h i s c o n d itio n a ls o m e a n s th a t th e m a g n itu d e s p e c tr u m is s y m m e tr ic ( e v e n f u n c t io n ) a b o u t th e o rig in an d th e p h a s e s p e c tru m is a n o d d fu n c tio n . As a c o n s e q u e n c e o f th e s y m m e tr y , it is s u ffic ie n t to s p e c ify th e s p e c tr u m o f a re a l p e r io d ic sig n a l f o r p o s itiv e f r e q u e n c ie s o n ly . F u r th e r m o r e , th e to ta l a v e r a g e p o w e r c a n b e e x p r e s s e d as
Px
+ 2
1q |
( 4 .1 .1 5 )
( 4 .1 .1 6 ) w h ich fo llo w s d ir e c t ly fro m th e r e la tio n s h ip s g iv e n in S e c t io n 4 .1 .1 a m o n g { aa }, {bn }. a n d ( q ) c o e f fic ie n ts in th e F o u r ie r s e r ie s e x p r e s s io n s .
Example 4 .1.1
Determ ine the Fourier series and the power density spectrum of the rectangular pulse train sienal illustrated in Fie. 4.3.
x(t)
-T
Figure 4 J Continuous-time periodic tram of rectangular pulses.
238
Chap. 4
Solution The signal is periodic with fundam ental period Tp and. clearly, satisfies the Dirichlet conditions. Consequently, we can represent the signal in the Fourier series given by (4.1.8) with the Fourier coefficients specified by (4.1.9). Since x(t) is an even signal [i.e.. x(t) = * ( - ;) ] , it is convenient to select the integration interval from ~ T p/2 to Tp/2. Thus (4.1.9) evaluated for k = 0 yields (4.1.17) The term c(I represents the average value (dc com ponent) of the signal .r (r). For k ^ 0 we have Tp l ~j2TTkFn\ _ jri A
tt FkTp
ej*yf-i,r
_
)2
(4.1.18)
A t sin TrkF^r
Tp
n k F\\T
k = 1 . 2 . ...
It is interesting to note that the right-hand side of (4.1.18) has the form (sin 4> )/4> , where takes on discrete values since F(, and r are fixed and the index k varies. However, if we plot (sin $)/</> with 0 as a continuous param eter over the range oc < < t> < oc, we obtain the graph shown in Fig. 4.4. We observe that this function decays to zero as 0 < x . has a maximum value of unity at xkFu t and scaled in amplitude by A t / T p. Since the periodic function x{t) is even, the Fourier coefficients c* are real. Consequently, the phase spectrum is either zero, when c* is positive, or t t when c k is negative. Instead of plotting the m agnitude and phase spectra separately, we may sim ply plot |c*} on a single graph, showing both the positive and negative values ck on the graph. This is commonly done in practice when the Fourier coefficients {c*} are real. Figure 4.5 illustrates the Fourier coefficients of the rectangular pulse train when Tp is fixed and the pulse width t is allowed to vary. In this case Tp = 0.25 second, so that F() = \ j T p = 4 Hz and t = 0.057),, t = 0.17},, and r = 0.27),. We observe that the effect of decreasing r while keeping Tp fixed is to spread out the signal power over the frequency range. The spacing between adjacent spectral lines is F{) = 4 Hz, independent of the value of the pulse width r.
sin 
The function (sin <p)/<p.
Sec. 4.1
239
ck
t = 0.1 Tp
,mTrnit
J lU jii "
vfffilllll I l k ,
j ct
JJiLLUTT
u -itrnrm-i.
F
T
= 0.05 Tr
..................rrnTfniniiin|fnijnnfiiifiii........... ...
0
and the pulse width r varies.
On the other hand, it is also instructive to fix t and vary the period Tp when Tp > r. Figure 4 .6 illustrates this condition when T,, ~ 5r. Tr = lOr. and Tp = 2 0 t . In this case, the spacing between adjacent spectral lines decreases as Tp increases. In the limit as Tr oc, the Fourier coefficients q approach zero due to the factor of Tp in the denom inator of (4 .1 .1 8 ). This behavior is consistent with the faci that as Tp < oc and r remains fixed, the resulting signal is no longer a power signal. Instead,
C l ...... ...................IMn 1' ..,.((111111 lllllllli,.. 0

'M illin '"
T p = 20r
Figure 4.6 Fourier coefficient of a rectangular pulse train with fixed pulse width t and varying period Tp .
240
Chap.
it becomes an energy signal and its average power is zero. The spectra of finite energy signals are described in the next section. We also note that if k / 0 and sm(7ikFnx) = 0. then ct = 0. The harmonics with zero power occur at frequencies kF0 such that n{kFG )r = m n , m = 1, 2 ,. o r at JtF0 = m jx. For example, if F() = 4 Hz and t = Q.2TP, it follows that the spectral components at 20 Hz, 40 H z , . .. have zero power. These frequencies correspond to the Fourier coefficients k = 5, 10, 15........O n the other hand, if r = 0.1Tp, the spectral components with zero power are k = 10, 20, 3 0 ........ The power density spectrum for the rectangular pulse train is
4.1.3 The Fourier Transform for Continuous-Time Aperiodic Signals

In Section 4.1.1 w e d evelop ed the Fourier series to represent a periodic signal as a linear com bination o f harm onically related com p lex exp on en tials. A s a con seq u en ce o f the periodicity, w e saw that these signals possess line spectra with equidistant lines. T h e line spacing is equal to the fundam ental frequency, which in turn is the inverse o f the fundam ental period of the signal. W e can view the fundam ental period as providing the num ber o f lin es per unit o f frequency (line d en sity), as illustrated in Fig. 4.6. W ith this interpretation in m ind, it is apparent that if w e allow the period to increase w ithout limit, the line spacing tends toward zero. In the limit, when the period b eco m es infinite, the signal b eco m es aperiodic and its spectrum becom es continuous. This argum ent suggests that the spectrum of an ap eriod ic signal will b e the en v elo p e o f the line spectrum in the corresponding p eriod ic signal obtained by repeating the aperiodic signal with som e period Tp. L et us consider an aperiodic signal x ( t ) with finite duration as show n in Fig. 4.7a. From this aperiodic signal, w e can create a periodic signal * , , ( 0 with p e riod Tp, as shown in Fig. 4.7b. Clearly, x r (t) = x ( t ) in the limit as Tp oo, that is, x{t) = lim x p(t)
This interpretation im plies that w e should be able to obtain the spectrum of * (/) from the spectrum o f x p(i) sim ply by taking the limit as Tp -* oo. W e begin with the Fourier series representation o f x p(t). xp{t) = where ckej l ^ ' 1 F0 = y (4.1.20)
(4.1.21)
Sec. 4.1
Frequency Analysis o1 Continuous-Time Signals
241
,r(n
- T
-TJ2
TJ2
TJ2
Figure 4.7
(a) Aperiodic signal ,v(/)
(h)
and (b) periodic signal xr U) constructed bv repeating x(t) with a period T r.
S in c e x p (t) x (/ ) f o r ~ T r f 2 < t < Tp/ 2 , ( 4 .1 .2 1 ) c a n b e e x p r e s s e d as
1 'Tr /2 = / x ( t ) e ~ )27' kF'', d t Tp J-Trfl
( 4 .1 .2 2 )
It is a ls o tr u e th a t .* (r ) = 0 fo r | r| > Tpj2 . C o n s e q u e n tly , th e lim its o n th e in te g r a l in ( 4 .1 .2 2 ) c a n b e r e p la c e d b y o c an d o c . H e n c e
ct =
x { t ) e ~ j2 * kF" d t
( 4 .1 .2 3 )
Tp J-oc
L e t us n o w d e fin e a fu n c tio n X ( F ) , c a lle d th e F o u r ie r tra n sfo rm o f * { / ) , as
X (F) =
A' ( F )
x (t)e ~ lln F , dt
( 4 .1 .2 4 )
J-oc
is a fu n c t io n o f th e c o n tin u o u s v a r ia b le F . I t d o e s n o t d e p e n d o n Tp o r
F q.
H o w e v e r , i f w e c o m p a r e ( 4 .1 .2 3 ) a n d ( 4 .1 .2 4 ) , it is c l e a r th a t th e F o u r ie r
c o e f f ic ie n t s ct c a n b e e x p r e s s e d in te r m s o f X ( F ) as
c* = ^ X ( k F o )
1p
o r e q u iv a le n tly .
Tpc k = X ( k F 0) = X
( 4 A .2 5 )
T h u s th e F o u r i e r c o e ff ic ie n ts a r e s a m p le s o f X ( F ) ta k e n a t m u ltip le s o f f o a n d s c a le d b y F 0 (m u ltip lie d b y \ / T p). y ie ld s S u b s titu tio n f o r c t f r o m ( 4 .1 .2 5 ) in to ( 4 .1 .2 0 )
V O
= ^r 'jh x ( ^ r ) er M
(4 .1 .2 6 )
242
Chap, 4
W e w ish to ta k e th e lim it o f ( 4 .1 .2 6 ) a s T r a p p r o a c h e s in fin ity . A F = 1/7},. W ith th is s u b s t itu tio n , ( 4 .1 .2 6 ) b e c o m e s
F i r s t , w e d e fin e
xpU ) =
Y2
k=-x
X (k & F )e
JlxkAFi
( 4 .1 .2 7 )
I t is c l e a r th a t in th e lim it as Tp a p p r o a c h e s in fin ity , x p {t) r e d u c e s to jc(/ ). A ls o , A F b e c o m e s th e d iff e r e n tia l d F a n d k A F b e c o m e s th e c o n tin u o u s f r e q u e n c y v a ria b le
F.
In tu r n , th e s u m m a tio n in ( 4 .1 .2 7 ) b e c o m e s a n in te g r a l o v e r th e fr e q u e n c y
v a r ia b le F . T h u s
( 4 .1 .2 8 )
T h is in te g r a l r e la tio n s h ip y ie ld s x ( t ) w h e n X ( F )
is k n o w n , a n d it is c a lle d th e
in erse F o u r ie r tra n sfo rm .

T h is c o n c lu d e s o u r h e u r is tic d e r iv a tio n o f th e F o u r i e r tr a n s f o r m p a ir g iv e n b y ( 4 .1 .2 4 ) an d ( 4 .1 .2 8 ) f o r an a p e r io d ic s ig n a l x ( t ). A lth o u g h th e d e r iv a tio n is n o t m a th e m a tic a lly rig o r o u s , it le d to th e d e s ir e d F o u r i e r t r a n s f o r m re la tio n s h ip s w ith r e la tiv e ly s im p le in tu itiv e a r g u m e n ts . In s u m m a r y , th e f r e q u e n c y a n a ly s is o f c o n tin u o u s -tim e a p e r io d ic s ig n a ls in v o lv e s th e fo llo w in g F o u r i e r tr a n s f o r m p a ir.
FR EQ U EN C Y ANALYSIS O F C O N T IN U O U S -T IM E A P ER IO DIC SIGN ALS
Synthesis equation inverse transform Analysis equation direct transform
(4 .1 .2 9 )
( 4 .1 .3 0 )
I t is a p p a r e n t th a t th e e s s e n tia l d if f e r e n c e b e tw e e n th e F o u r i e r s e r ie s a n d th e F o u r i e r tr a n s fo r m is th a t th e s p e c tr u m in th e l a t t e r c a s e is c o n tin u o u s a n d h e n c e th e s y n th e s is o f an a p e r io d ic s ig n a l fr o m its s p e c tr u m is a c c o m p lis h e d b y m e a n s o f in te g r a tio n in s te a d o f s u m m a tio n . F in a lly , w e w ish to in d ic a te th a t th e F o u r ie r t r a n s f o r m p a ir in ( 4 .1 .2 9 ) and ( 4 .1 .3 0 ) c a n b e e x p r e s s e d in te r m s o f th e r a d ia n fr e q u e n c y v a r ia b le Q = S in c e d F = d S l / l n . ( 4 .1 .2 9 ) a n d ( 4 .1 .3 0 ) b e c o m e ( 4 .1 .3 1 )
2nF.
( 4 .1 .3 2 ) T h e s e t o f c o n d itio n s th a t g u a r a n te e th e e x is t e n c e o f th e F o u r i e r tr a n s f o r m is th e
Sec. 4.1
243
D ir ic h le t c o n d it io n s , w h ich m a y b e e x p r e s s e d as:
1 . T h e s ig n a l v (r) h as a fin ite n u m b e r o f fin ite d is c o n tin u itie s . 2 . T h e s ig n a l x ( t ) h a s a fin ite n u m b e r o f m a x im a a n d m in im a . 3 . T h e s ig n a l x ( t ) is a b s o lu te ly in te g r a b le , th a t is.
i:
g iv e n in ( 4 .1 .3 0 ) . In d e e d .
\x (t )\d t < o c
( 4 .1 .3 3 )
T h e th ir d c o n d itio n fo llo w s e a s ily fr o m th e d e fin itio n o f th e F o u r i e r tr a n s f o r m ,
\ X(F)\
jj
x ( t ) e - J2 F ,dt
< j
\x ( t )\d t
H e n c e ] X ( F ) ! < o c if ( 4 .1 .3 3 ) is s a tis fie d . A w e a k e r c o n d itio n f o r th e e x is te n c e o f th e F o u r i e r t r a n s f o r m is th a t x {t) h a s fin ite e n e r e v ; th a t is. |.v(/)|^r < o c ( 4 .1 .3 4 )
N o te th a t if a s ig n a l x ( i) is a b s o lu te ly in te g r a b le . it w ill a ls o h a v e fin ite e n e rg y . T h a t is. if
J C M
th e n
|.v(/)!^r < o c
rx
|.v(/)i^/ < o c ( 4 .1 .3 5 )
H o w e v e r , th e c o n v e r s e is n o t tr u e .
T h a t is. a s ig n a l m a y h a v e fin ite e n e r g y b u t
m a y n o t b e a b s o lu te ly in te g r a b le . F o r e x a m p le , th e s ig n a l sin 2 n t x {t) = -
7
T t
( 4 .1 .3 6 ) T h is s ig n a l h a s th e F o u r i e r
is s q u a r e in te g r a b le b u t is n o t a b s o lu te ly in te g r a b le . tr a n s fo r m * ( F ) = { o ;
f1
\FI < 1
(4 -1 .3 7 )
S in c e th is s ig n a l v io la te s ( 4 .1 .3 3 ) , it is a p p a r e n t th a t th e D i r i c h l e t c o n d itio n s a re s u ffic ie n t b u t n o t n e c e s s a r y f o r th e e x is te n c e o f th e F o u r i e r tr a n s f o r m . In a n y c a s e , n e a r ly all fin ite e n e r g y s ig n a ls h a v e a F o u r ie r tr a n s f o r m , s o t h a t w e n e e d n o t w o rry a b o u t th e p a th o lo g ic a l s ig n a ls , w h ich a r e s e ld o m e n c o u n t e r e d in p r a c tic e .
4.1.4 Energy Density Spectrum of Aperiodic Signals

L e t x (t ) b e a n y fin ite e n e r g y s ig n a l w ith F o u r i e r t r a n s fo r m X ( F ) . Its e n e r g y is
244
Chap. 4
w hich, in turn, may be expressed in term s o f X ( F ) as follows:

Ex f
J OC
x ( t ) x '( i) d t
1 oo
X '( F ) d F \
'OC
r /*oc / x { t ) e ~ j2 n F 'd t
\ X ( F ) \ 2d F
T h erefore, w e con clu d e that (4.1.38) This is P a r s e v a ls re la tio n for aperiodic, finite energy signals and expresses the principle o f conservation of energy in the tim e and frequency dom ains. T he spectrum X ( F ) o f a signal is in general, com p lex valued. C onsequently, it is usually expressed in polar forms as
X ( F ) = | X ( F ) | ^ W(f)
where |X ( F ) | is the m agnitude spectrum and (F ) is the phase spectrum . (F ) - i U ( F ) O n the other hand, the quantity
SXX( F ) = \X ( F ) | 2
(4.1.39)
which is the integrand in (4.1.38), represents the distribution of en ergy in the signal as a function o f frequency. H en ce SXX( F ) is called the e n e rg y d e n sit y s p e ctru m of x ( t ). T he integral o f S XX( F ) over all freq u en cies gives the total en ergy in the signal. V iew ed in another way, the energy in the signal x ( t ) over a band o f frequencies F \ < F < F [ + A F is
From (4.1.39) w e observe that S XX( F ) d o es not con tain any p h ase inform ation [i.e., SXX( F ) is purely real and n on n egative]. Since the phase spectrum of ;t(r) is not contained in SXX( F ) , it is im possible to reconstruct the signal given S XX( F ) . Finally, as in the case of Fourier series, it is easily show n that if the signal x (t ) is real, then
\X ( -F ) \ = X { -F ) \X ( F ) \
(4.1.40) (4.1.41)
= -* X (F )
Sec. 4.1
245
By com bining (4.1.40) and (4.1.39), w e obtain

SXX( ~ F ) = S xA F )
(4.1.42)
In other words, the energy density spectrum o f a real signal has even sym m etry. Exam ple 4.1.2 D eterm ine the Fourier transform and the energy density spectrum of a rectangular pulse signal defined as
and illustrated in Fig. 4.8(a). Solution Clearly, this signal is aperiodic and satisfies the Dirichlet conditions. Hence its Fourier transform exists. By applying (4.1.30), we find that (4.1.44) We observe that X( F) is real and hence it can be depicted graphically using only one diagram, as shown in Fig. 4.8(b). Obviously, X( F) has the shape of the (sin0)/<? function shown in Fig. 4.4. Hence the spectrum of the rectangular pulse is the en velope of the line spectrum (Fourier coefficients) of the periodic signal obtained by
0 (a)
X<F)
~ >
Ar
F (b)
Figure 4.8 (a) R ectangular pulse and (b) its F ourier transform .
246
Chap. 4
periodically repeating the pulse with period Tp as in Fig. 4.3. In other words, the Fourier coefficients ck in the corresponding periodic signal xp{t) are simply samples of X( F) at frequencies kF0 = k/ Tp. Specifically,
From (4.1.44) we note that the zero crossings of X( F ) occur at multiples of 1 /r. Furtherm ore, the width of the main lobe, which contains most of the signal en ergy, is equal to 2/z. As the pulse duration t decreases (increases), the main lobe becomes broader (narrow er) and m ore energy is moved to the higher (lower) frequencies, as illustrated in Fig. 4.9. Thus as the signal pulse is expanded (com pressed) in time, its transform is compressed (expanded) in frequency. This be havior, between the time function and its spectrum, is a type of uncertainty principle that appears in different forms in various branches of science and engi neering. Finally, the energy density spectrum of the rectangular pulse is
, /sm 7 rj S ( F ) = ( A t )1 ( " ~ ' ^ T ) } ( ttF-
(4.1.46)
*(r)
X 0 I
x (r) A
X(F)
- L
X(l)
Figure 4.9
F o u rier transform of a re d a n g u la r pulse for various w idth values.
Sec. 4.2
Frequency Analysis of Discrete-Time Signals
247
4.2 FREQUENCY ANALYSIS OF DISCRETE-TIME SIGNALS

In Section 4.1 we d ev elop ed the Fourier series representation for continuous-tim e periodic (pow er) signals and the Fourier transform for finite energy aperiodic signals. In this section we repeat the d evelop m en t for the class o f discrete-tim e signals. A s we have ob served from the discussion o f Section 4.1, the Fourier series representation o f a con tinu ou s-tim e periodic signal can consist o f an infinite num ber o f frequency com p onents, where the frequency spacing b etw een tw o successive harm onically related freq u en cies is 1 / T p, and w here Tp is the fundam ental period. Since the frequency range for continuous-tim e signals exten d s from oo to oc, it is p ossib le to have signals that contain an infinite num ber o f frequency com p o nents. In contrast, the frequency range for discrete-tim e signals is unique over the interval ( - r t . T i ) or (0 .2 ^ ). A discrete-tim e signal o f fundam ental period N can consist o f frequency co m p on en ts separated by 2 n / N radians or / = 1/jV cycles. C onsequently, the Fourier series representation o f the discrete-tim e periodic signal will contain at m ost N frequency com ponents. This is the basic d ifference b etw een the Fourier series representations for continuous-tim e and discrete-tim e periodic signals.
4.2.1 The Fourier Series for Discrete-Time Periodic Signals

Suppose that w e are given a periodic sequ en ce a (/ i ) with period N , that is, x i n ) = x(n + N ) for all n. T he Fourier series representation for x(h) consists o f N har m onically related exp on en tial functions k = 0 A ........ N 1 and is expressed as A'-l X () = Y ^ c kejlKkn/N *=o (4.2.1)
where the {<:*) are the coefficients in the series representation. T o derive the expression for the Fourier coefficients, w e use the follow ing formula: y '1 n=0 = ( I 0, 1 * = 0. N , 2 N, . . . otherw ise 2 2
N o te the sim ilarity o f (4.2.2) with the con tinu ou s-tim e counterpart in (4.1.3). T he p roof o f (4.2.2) follow s im m ediately from the application of the geom etric sum m ation form ula jv-i f a = l
248
Chap. 4
T h e e x p r e s s io n f o r th e F o u r i e r c o e f f ic ie n ts c* c a n b e o b t a i n e d b y m u ltip ly in g b o th s id e s o f ( 4 .2 .1 ) b y th e e x p o n e n tia l e ~ j2nin//v a n d s u m m in g th e p r o d u c t fro m = 0 t o n = yV 1. T h u s A'-l N 1 A- ]
^ 2 x ( n ) e ~ il!rln /N
c ke J2n<k~ l)n/N
(4 .2 .4 )
n=(J
w e o b ta in y ^ (fj2.TU-/)n/A' _ | N, I 0,
n=0 t=0
I f w e p e r fo r m th e s u m m a tio n o v e r n first, in th e r ig h t-h a n d s id e o f (4 .2 .4 ) ,
k - I = 0, N , 2 N , . . . o th e r w is e
(4 2 5)
w h e r e w e h a v e m a d e u se o f ( 4 .2 .2 ) . r e d u c e s to N c / a n d h e n c e j A'-l q = -
T h e r e f o r e , th e rig h t-h a n d s id e o f ( 4 .2 .4 )
J ' x ( n ) e ~ J2*" /tl

/T = (J
1 = 0. 1 .......... N - 1
(4 .2 .6 )
T h u s w e h a v e th e d e s ir e d e x p r e s s io n fo r th e F o u r i e r c o e f f ic ie n t s in te r m s o f th e s ig n a l x ( ) . T h e r e la tio n s h ip s ( 4 .2 .1 ) a n d ( 4 .2 .6 ) fo r th e f r e q u e n c y a n a ly s is o f d is c r e te * tim e s ig n a ls a r e s u m m a r iz e d b e lo w .
FR E Q U E N C Y ANALYSIS O F D IS C R E TE -TIM E P E R IO D IC SIGN ALS
Synthesis equation
A'-l X(/! ) = ^ C k(j2jTt"!r' *=0
(4.2.7)
iW l
-1 % II
Analysis equation
(4.2.8)
E q u a t i o n ( 4 .2 .7 ) is o ft e n c a lle d th e d isc re te -tim e F o u r ie r s e rie s ( D T P S ) . T h e F o u r i e r c o e ff ic ie n ts {c* }. k = 0 . 1 .......... N 1 p r o v id e th e d e s c r ip tio n o f jc ( ) in th e f r e q u e n c y d o m a in , in th e s e n s e th a t c k r e p r e s e n t s th e a m p litu d e a n d p h a s e a s s o c ia t e d w ith th e fr e q u e n c y c o m p o n e n t
sk(n) = e ^ kn^ ' = e jW tB

w h e r e to* = 2 n k / N . W e re c a ll f r o m S e c tio n 1 .3 .3 th a t th e f u n c t io n s s k (n ) a r e p e r i o d i c w ith p e r io d
N . H e n c e sk(n) = sk {n + N ) . In v iew o f th is p e r io d ic ity , it fo llo w s t h a t th e F o u r ie r

c o e f f ic ie n ts c k, w h e n v ie w e d b e y o n d th e r a n g e k ~ have 0 , 1 , ____A ' - l , a ls o s a tis fy a p e r io d ic ity c o n d itio n . In d e e d , fro m ( 4 .2 .8 ) , w h ich h o ld s f o r e v e r y v a lu e o f k , w e
Ck+ N = ^ J ^ x ( n ) e - ^ (k+N)n/N = ~ J 2 x { n ) e - ^ kn/N = c t
( 4 .2 .9 )
n=0
n=0
Sec. 4.2
249
T h e r e f o r e , th e F o u r i e r s e r ie s c o e f f ic ie n ts { q } fo r m a p e r io d ic s e q u e n c e w h e n e x te n d e d o u ts id e o f t h e r a n g e k = 0 , 1 ..........jV 1. H e n c e
Ck+N = < ^k T hu s the s p e ctru m o f a signa l x(n ), w h ich is per iod ic with p e r i o d N , is a p er io d ic s e qu en ce with p e r io d
N . C o n s e q u e n tly , a n y N c o n s e c u tiv e s a m p le s o f th e s ig n a l o r its s p e c tr u m p r o v id e
a c o m p le t e d e s c r ip tio n o f th e s ig n a l in th e tim e o r f r e q u e n c y d o m a in s . A lth o u g h th e F o u r i e r c o e f fic ie n ts fo r m a p e r io d ic s e q u e n c e , w e w ill fo c u s o u r a tte n tio n o n th e s in g le p e r io d w ith ra n g e k = 0 , 1 .......... N 1. T h i s is c o n v e n ie n t, s in c e in th e f r e q u e n c y d o m a in , th is a m o u n ts to c o v e r in g th e f u n d a m e n ta l ra n g e 0 < a>t = th a t is , { c t } is a p e r io d ic s e q u e n c e w ith f u n d a m e n ta l p e r io d N.
2 n k /N
< 2 n , fo r 0 < k < N 1.
In c o n t r a s t , th e f r e q u e n c y ra n g e
< j t , c o r r e s p o n d s to N / 2 < k < N / 2 , w h ic h c r e a t e s an in c o n v e n ie n c e w h e n N is o d d . C le a r ly , i f w e u se a s a m p lin g f r e q u e n c y F s , th e r a n g e 0 < k < N 1 c o r r e s p o n d s to th e f r e q u e n c y ra n g e 0 < F < F , .
it < a)* = 2 7 i k / N
Example 4.2.1 D eterm ine the spectra of the signals (a) jr() = cos -Ji nn (b) x(n) = cos nn/ 3 (c) x( n ) is periodic with period N = 4 and x(n) = {1, 1.0.0}
T
Solution (a) For m = J i n , we have / = 1j - Jl . Since f , is not a rational number, the signal is not periodic. Consequently, this signal cannot be expanded in a Fourier series. N evertheless, the signal does possess a spectrum. Its spectral content consists of the single frequency component at id = wo = -Jin. (b) In this case / (l = | and hence x{n) is periodic with fundam ental period N = 6. From (4.2.8) we have 5 It = 0 . 1 ........5 However, x(n) can be expressed as x(n) = cos " r = + \ e~i2*nib 6 * which is already in the form of the exponential Fourier series in (4.2.7). In comparing the two exponential terms in x{n) with (4.2.7), it is apparent that ci = j. The second exponential in x(n) corresponds to the term Jt = 1 in (4.2.7), However, this term can also be written as
- j 2 j r n / 6 _ ^ j2jr (5 n> /6
which means that c_i = c$. But this is consistent with (4.2.9), and our previous observation that the Fourier series coefficients form a periodic sequence of
250
Chap. 4
period N . Consequently, we conclude that C (, = c, = ^ (c) From (4.2.8). we have 1

Ck
= C\ = f 4 = 0 i
A-= 0 . 1 ,2 ,3
or Ct = l ( l + e - i ^ ) For k = 0, 1, 2, 3 we obtain
C 'd = S f'l j ( l j ) f'2 = 0 Cl = j { l + j }
A-= 0 . 1.2.3
The m agnitude and phase spectra are K 'd I

4-Co = 0
ki
4 f| = 2tr- = undefined 4_o =
4 4 Figure 4.10 illustrates the spectral content of the signals in (b) and (c).
4.2.2 Power Density Spectrum of Periodic Signals

T h e a v e r a g e p o w e r o f a d is c r e te - tim e p e r io d ic sig n a l w ith p e r io d N w as d efin ed in ( 2 .1 .2 3 ) as
P, = -
M )|
( 4 .2 .1 0 )
W e s h a ll n ow d e riv e a n e x p r e s s io n fo r P x in te r m s o f th e F o u r i e r c o e f f ic ie n t {c<J. I f w e u se th e r e la tio n ( 4 .2 .7 ) in ( 4 .2 .1 0 ) , w e h a v e
j v -i
Pi = T
n = ()
x(rt)x*(n)
A'-l
N
N o w . w e c a n in te r c h a n g e t h e o r d e r o f th e tw o s u m m a tio n s a n d m a k e u s e o f ( 4 .2 .8 ) , o b ta in in g
p * = Y , ct
*=<j IV 1 A'-l
, A-i - y x ( n ) e - j2kn/N
-i A I
( 4 . 2 . 11 )
Sec. 4.2
Frequency Analysis of Discrete-Tim e Signals
251
(a)
Zc t
Jr
"4
-3
- 2 - 1 0 1
5 2 3 4
... t
7 1
4 Figure 4.10 Spectra of the periodic signals discussed in Example 4.2.1 (b) and (c).
(c)
which is the desired exp ression for the average p ow er in the p eriod ic signal. In other w ords, the average pow er in the signal is th e sum o f th e pow ers o f the individual frequency com p onents. W e view (4.2.11) as a P arsevals relation for d iscrete-tim e period ic signals. T h e seq u en ce | a |: for k = 0, 1 , N - 1 is the distribution o f p ow er as a function o f freq u en cy and is called the p o w e r density spectrum o f the periodic signal. If w e are interested in the energy o f th e se q u en ce Jt(n) o v er a single period, (4.2.11) im plies that
/VI N -l
(4.2.12) n=0 k () which is consistent with our p reviou s results for con tin u ou s-tim e p eriodic signals. If the signal x ( n ) is real [i.e., x (n) = jr(n)], then, p roceed in g as in S ec tion 4.2.1, w e can easily sh ow that ct = c . k
(4.2.13)
252
Chap. 4
or equivalently, |c_*| = - 4 c_t =

|c * f
(even sym m etry) (odd sym m etry)
(4.2.14) (4.2.15)
T h ese sym m etry properties for the m agnitude and phase spectra o f a periodic sig nal, in conjunction with the periodicity property, have very im portant im plications on the frequency range o f discrete-tim e signals. In d eed , by com bining (4.2.9) with (4.2.14) and (4.2.15), w e obtain
Iq I =
(4.2.16)
and
4-c* =
(4.2.17)
M ore specifically, w e have

II = ~ = O |c il \c N p \ k (/V -l)/2 * k jV /2 |, k ( A f + l)/2 l< 4 .c 0 = - % - c' n 0
4 - c' l
% -C N -l
(4.2.18)
A -c N[2 = o
if N is even
= ~ 4 - C ( /V + l)/2
4-C (N -1)/2
if N is odd
Thus, for a real signal, the spectrum c*, k = 0, 1 ,. . . , N [ 2 for N even, or k = 0. 1........ ( N l ) / 2 for N odd, com p letely specifies the signal in the frequency dom ain. Clearly, this is consistent with the fact that the highest relative frequency that can be represented by a discrete-tim e signal is eq u al to n . In d eed , if 0 < a> k= 2 n k / N < jt, then 0 < k < N f 2. By m aking use o f these sym m etry p roperties o f the Fourier series coefficients o f a real signal, the F ourier series in (4.2.7) can also be exp ressed in the alternative forms x ( n ) = co + 2 lc*l cos (4.2.19)
= ao + ^
( ^ k cos kn - b k sin kn 'j
(4.2.20)
w here a0 = c0. < * k = 2 [ct|co s0 * . bk = 2|c*|sin 0*, and L = N [2 if N is even and L ( N ~ l ) / 2 if N is odd. F inally, we n o te that as in the case o f continuous-tim e signals, the power density spectrum |ct |2 d oes n ot contain any phase inform ation. Furtherm ore, the spectrum is discrete and periodic with a fundam ental period eq u al to that o f the signal itself.
Example 4.2.2 Periodic Square-Wave Signal

Determ ine the Fourier series coefficients and the power density spectrum of the periodic signal shown in Fig. 4.11.
Sec. 4.2
Frequ ency A nalysis of Discrete-Time Signals
253
A'
Figure 4.11 Discrete-time periodic square-wave signal.
Solution By applying the analysis equation (4.2.8) to the signal shown in Fig. 4.11. we obtain = 1 A j 1 L~l = -L Y Ae- P^n/ A J i/ k = 0. 1. N - 1
which is a geom etric sum mation. Now we can use (4.2.3) to simplify the summation above. T hus we obtain AL TT A 1 - e- j2*kL/s I N 1' k= 0 k = 1.2. N - 1
The last expression can be simplified further if we note that

1 _ e -j2xkL;S -jzk L tN I.ikl/K _
1 - <-
e -ink;'\
( rrl. \ _ ,-jxkl.\
-jxHt.-u.'K sin (xkL. /N )

s in (7 r /A /
T herefore. AL
A
k = 0. +A'. 2 N , , s i ni n kL/ N) sin(7zk/N) (4.2.21) otherwise
I A'
The pow er density spectrum of this periodic signal is

/ AL_
k = 0, +N. 2 A'. (4.2.22) otherwise
k*l" =
V N A V / sinnkL/N
Ar / V s i n n k / N
Figure 4.12 illustrates the plots of jct |2 for L = 5 and 7, A' = 40 and 60. and A = 1.
4.2.3 The Fourier Transform of Discrete-Time Aperiodic Signals

Just as in the case o f con tinu ou s-tim e aperiodic energy signals, the frequency anal ysis o f d iscrete-tim e aperiodic finite-energy signals in volves a Fourier transform of the tim e-d om ain signal. C on sequ en tly, th e d evelop m en t in this section parallels to a large ex ten t, that given in Section 4.1.3.
254
Chap. 4
L = 5, N = 40
_xlJ
-2 0
-1 0
I 1II I 1. . I I I I
0
10
20
L = 7, N = 60
L J il
-3 0 - 20 -1 0 0 10 20 30
= 5. N =
-3 0
- 20
10
10
20
30
Figure 4.12 Plot of the pow er density spectrum given by (4.2.22).
T he Fourier transform o f a finite-energy discrete-tim e signal x ( n ) is defined as

OC
X (co) =
FI OC
x ( n ) e ~ Jwr
(4.2.23)
Physically, X(co) represents the frequency con ten t o f the signal x ( n ) . In other words. X{(o) is a d ecom p osition o f x ( n ) into its frequency com p onents. W e observe tw o basic d ifferen ces b etw een the Fourier transform of a discrete tim e finite-energy signal and the F ourier transform o f a finite-energy analog signal. First, for continuous-tim e signals, the Fourier transform , and h en ce the spectrum o f the signal, have a frequency range o f {0 0 , 0 0 ). In contrast, the frequency range for a discrete-tim e signal is unique over the frequency interval o f (n , n) or, equivalently, (0. 2 tt). T h is property is reflected in the Fourier transform o f the
Sec. 4.2
255
signal. In d eed . X(co) is period ic with period 2 n . that is. X(a> + 2 7 T k )=

22
cc oc
x { n ) e 'i{w^ 7 ' k)n
j{u> ^-27ik)u
J 2 x ( n ) e - ' ame - JZ7,kn

/l = OC OC
(4.2.24)
2 2
x ( n ) e - , Mn
= X(w)
H en ce X ( t o ) is p eriod ic with period 2 t t . B u t this property is just a con seq u en ce of the fact that the frequency range for any discrete-tim e signal is lim ited to (n , n ) or (0 , 27r). and any frequency outside this interval is eq u ivalen t to a frequency within the interval. T h e secon d basic d ifference is also a con seq u en ce o f the discrete-tim e nature o f the signal. Since the signal is discrete in tim e, the Fourier transform o f the signal in v o lv es a sum m ation o f term s instead o f an integral, as in the case of continuous-tim e signals. Since X (w ) is a periodic function o f the frequency variable a), it has a Fourier series exp an sion , provided that the con d ition s for the existen ce o f the Fourier series, described previously, are satisfied. In fact, from th e definition o f the Fourier transform X (co) o f the sequ en ce x(n), given by (4.2.23), w e ob serve that X (a > ) has the form o f a Fourier series. T he Fourier coefficients in this series expansion are the valu es o f the seq u en ce x ( n ) . T o dem onstrate this point, let us evaluate the seq u en ce x ( n ) from X(co). First, we m ultiply both sides (4.2.23) by ej< um and integrate over the interval ( tt, t t ) . Thus w e have (4.2.25) The integral on the right-hand side o f (4.2.25) can be evaluated if w e can inter change the order o f sum m ation and integration. This interchange can be m ade if the series
N
X N (a>) = 2 2 x ( n ) e - llon
n=-N
con verges uniform ly to X(o)) as A1 -> oc. U niform con vergen ce m ean s that, for every w, Xh((d) * X(ca), as /V -* oo. T he con vergen ce o f the Fourier transform is discussed in m ore detail in the follow in g section. For the m om en t, let us as sum e that the series con verges uniform ly, so that w e can interchange the order of sum m ation and integration in (4.2.25). Then
256
Chap.
C onsequently,
rj = OC
x{n)
f J 7 1
n'da) =
2:Tx(m). 0.
(4.2.26)
By com bining (4.2.25) and (4.2.26). w e obtain the desired result that
x (n )
.7T J jt
fn
X ( o j) e J
do)
(4.2.27)
If w e com pare the integral in (4.2.27) with (4,1.9), we n o te that this is just the expression for the Fourier series coefficient for a function that is periodic with period 2 n . The only difference b etw een (4.1.9) and (4.2.27) is the sign on the exp on en t in the integrand, which is a con sequ en ce o f our definition o f the Fourier transform as given by (4.2.23). T h erefore, the Fourier transform of the sequence x ( n ) , defined by (4.2.23), has the form o f a Fourier series expansion. In summary, the Fourier tra nsform p a ir f o r discrete-time signals is as follows.
FR EQ U EN C Y ANALYSIS OF DtS C R E TE -TIM E AP ER IO D IC S IG N A LS
Synthesis equation inverse transform Analysis equation direct transform
x(
O C
X {io )= 2 2
X < w ) i lu ' ' ' d w
(4 .2 .2 8 )
x ( n ) e ~ , w '1
(4.2.29)
4.2.4 Convergence of the Fourier Transform

In the derivation o f the inverse transform given by (4.2.28), w e assum ed that the series
Xjv(cw) =
22 xW e~
(4.2.30)
converges uniform ly to X(a>), given in the integral o f (4.2.28), as N -* oc. By uniform convergence w e m ean that for each
lim { s u p X (oj) ~ X s (>>)'} = 0 (4.2.31) N >* < * > U niform convergence is guaranteed if ;c(/j) is ab solu tely sum m able. Indeed, if D C 2 2 |j c ) i < oc (4.2.32) n=-oc then
|X M l =
22
x (n )e '
<
2 2
<
00
H en ce (4.2.32) is a sufficient condition for the existen ce o f the d iscrete-tim e Fourier transform. W e n o te that this is the discrete-tim e counterpart of th e third Dirich-
Sec. 4.2
257
let condition for the Fourier transform o f con tinu ou s-tim e signals. T he first two conditions d o n ot apply due to the discrete-tim e nature o f [*()}. S om e se q u en ces are not absolutely sum m able, but they are square sum m able. That is, they have finite energy
OC
Ex =
^
n = DC
|jc(n ) |2 < oo
(4.2.33)
which is a w eak er condition than (4.2.32). W e w ould like to d efine the Fourier transform o f finite-energy sequ en ces, but w e m ust relax the con d ition o f uniform con vergen ce. For such seq u en ces w e can im pose a m ean-square con vergen ce co n dition: lim f |X M - X N ((ti)\2dto = 0 A f<* J-71 (4.2.34)
Thus the energy in the error X(a>) - X/v(oj) tends toward zero, but the error |X (w ) - Xw(a))j d o es not necessarily tend to zero. In this way w e can include finite-energy signals in the class o f signals for which the Fourier transform exists. Let us con sid er an exam p le from the class o f finite-energy signals. Suppose that X M = ( 1
[ U.
H ^
< ja>| < 7 T
|< 7 r
(4-2.35)
T he reader should rem em ber that X(a>) is periodic with p eriod 2 n . H en ce (4.2.35) represents only on e period o f X(co). The inverse transform o f X { cj) results in the sequence i jr() =
1
r I
X { w ) e ja>nda) sin<u,-n Jin
-~
In
eJ
dm = ------------
nÔ
For n = 0, w e have *( 0 ) H en ce n (j)c sin a)cn n cocn n = 0 (4.2.36) n

0
x (n ) =
This transform pair is illustrated in Fig. 4.13. S om etim es, the seq u en ce {*()) in (4.2.36) is exp ressed as sin cocn x ( n ) = --------- oo < n < oo (4.2.37)
258
Chap. 4
X(w)
1
-JT UJt 0 (i), 7T
(b) Figure 4.13 Fourier transform pair in (4.2.35) and (4.2.36j.
with the understanding that at n = 0, x ( n ) = w j n . W e should em p hasize, however, that (sina)(7i)/jrfi is not a con tinu ou s function, and hence L H osp ital's rule cannot be used to determ ine jr(0 ). N ow iet us consider the determ ination o f the Fourier transform o f the se q uence given by (4.2.37), T he seq u en ce {jr(n)j is not absolutely sum m able. H ence the infinite series (4.2.38)
does not converge uniform ly for all w. H ow ever, the seq u en ce {jc()} has a finite energy E x = o)c/ tc as will be show n in Section 4.3. H en ce the sum in (4.2.38) is guaranteed to converge to the X (a>) given by (4.2.35) in the m ean-square sense. T o elaborate on this point, let us con sid er the finite sum
X N (a>)=
(4.2.39)
Figure 4.14 show s the function X N{w) for several values of N . W e n ote that there is a significant oscillatory oversh oot at co = coc, in d ep en dent o f the value o f N . As
Sec. 4.2
Frequency Analysis of Discrele-Ttme Signals
259
x , s(w)
^50<)
A "7o(ul)
Figure 4.14 Illustration of convergence of the F ourier transform and the G ibbs phenom enon at the point of discontinuity
N increases, the oscillation s b ecom e m ore rapid, but the size o f the ripple rem ains the sam e. O ne can sh ow that as N -* oo, the oscillations con verge to the point o f the discontinuity at to a > t . but their am plitude d oes not go to zero. H ow ever, (4.2.34) is satisfied, and therefore converges to X ( w ) in the m ean-square sense. T he oscillatory behavior o f the approxim ation Xîc o) to the function X(a>) at a point o f discontinuity o f is called the G ib b s p h e n o m e n o n . A sim ilar effect is ob served in the truncation o f the F ourier series o f a con tinu ou s-tim e periodic signal, given by the syn th esis eq u ation (4.1.8). For exam ple, the truncation o f the Fourier series for the period ic square-w ave signal in E xam ple 4.1.1, gives rise to the sam e oscillatory behavior in the finite-sum approxim ation o f x ( t ) . T he G ibbs ph en om en on will be en cou n tered again in the design o f practical, discrete-tim e FIR system s considered in C hapter 8 .
260
Chap. 4
4.2.5 Energy Density Spectrum of Aperiodic Signals

R ecall that the energy o f a discrete-tim e signal x ( n ) is defined as
OC
Ex = 2 2 l*(n )l2 n= oc
we have
(4.2.40)
Let us n o w e x p r e s s th e e n e r g y E x in te r m s o f th e s p e c tr a l c h a r a c t e r i s t i c X (w). F irst 00 22 x(n)x*(n)=

n ------- -v
Ex =
I" J rJT 2 2 x(n) X* n -v _ J JT
If we interchange the order o f integration and sum m ation in the equation above, we obtain dw
|X (co)\~du> T h erefore, the energy relation b etw een x ( n ) and X(a>) is

E,=
oc \ 2 2 \x^)\2 = n=-oc
fn
\X (c o )l2dco
(4.2.41)
This is Parseval's relation for discrete-tim e aperiodic signals with finite energy. T he spectrum X(a>) is, in general, a com p lex-valu ed function of frequency. It may be expressed as
X(u>) = \X ( c u ) \e J* M
(4.2.42)
where Q(co) - ^ X ( c o ) is the phase spectrum and \X (a > )\ is the m agnitude spectrum . A s in the case o f continuous-tim e signals, the quantity S(o>) = | X M |2 (4.2.43)
represents the distribution of energy as a function o f frequency, and it is called the ener gy density sp ectrum o f x ( n ) . Cleariy, Sxx(a>) d oes not contain any phase inform ation. Suppose now that the signal x ( n ) is real. T hen it easily follow s that X*(o>) = X ( - oj) (4.2.44)
or equivalently, |X ( ti>)| = |X(tt>)| (even symmetry) (4.2.45)
Sec. 4.2
261
(4.2.46)
(4.2.47) From these sym m etry properties we con clu d e that the frequency range of real d iscrete-tim e sign als can be lim ited further to the range 0 < co < n (i.e., on e-half o f the p erio d ). Indeed, if we know X ( a > ) in the range 0 < w < n , we can d eterm in e it for the range - n < co < 0 using the sym m etry p roperties given above. A s w e have already ob served , sim ilar results hold for discrete-tim e periodic signals. T h erefo re, the frequency-dom ain description o f a real discrete-tim e signal is co m p letely sp ecified by its spectrum in the frequency range 0 < co < n . U su a lly , w e w ork with the fundam ental interval 0 < a > < 7r o r 0 < F < F J 2, exp ressed in H ertz. W e sketch m ore than half a period only w hen required by the specific application. Example 4.2.3 D eterm ine and sketch the energy density spectrum 5,,r (w) of the signal = a"u(/i) 1 < a < 1
Solution Since \a\ < 1. the sequence x(n) is absolutely summable, as can be verified by applying the geometric summation formula.
H ence the Fourier transform of x(n) exists and is obtained by applying (4.2.29). Thus
Since \ae~J,Jl = |a| < 1. use of the geom etric summation formula again yields
The energy density spectrum is given by Sl x(w) = |X(w)|2 = X(a>)X'(w) = or, equivalently, as 1 1 2a cos a >+ a Note that S (-o> ) = Ss i (w) in accordance with (4.2.47). Figure 4.15 shows the signal *() and its corresponding spectrum for a = 0.5 and a = -0 .5 . Note that for a = -0 .5 the signal has more rapid variations and as a result its spectrum has stronger high frequencies. (1 ae~Ju,)( 1 aeJW )
262
Chap. 4
x(n) - (-0.5)"u(!
Figure 4.15 spectra.
(a) Sequence .\in) =
and mii I = i
(b) ih eir energy density
Figure 4.16 pulse.
D iscrete-tim e rectangular
Example 4.2.4 Determ ine the Fourier transform and the energy density spectrum of the sequence xin) = which is illustrated in Fig. 4.16. Solution Before computing the Fourier transform, we observe that 0 < n <L otherwise (4.2.48)
Hence xin) is absolutely summable and its Fourier transform exists. Furthermore, we note that Jt(n) is a finite-energy signal with = \A\: L, The Fourier transform of this signal is
Sec. 4.2

1 - e~uL
263
= Ae - ' ,u/2'a - 1 1
s m( w / 2 )
(4.2.49)
For w = 0 the transform in (4.2.49) yields X(0) = AL, which is easily established by setting w = 0 in the defining equation for X(a>), or by using L H ospitals rule in (4.2.49) to resolve the indeterm inate form when w = 0. The m agnitude and phase spectra of ;t(n) are
f \A\L,
|X<)I = and
$X (a>) = $ A - - ( L - l ) + $
w= 0 otherwise sin(ctiZ./2) ' sin(w/2) (4-2.50)
sin(L /2) sin(aj/2) w 2
(4.2.51)
where we should rem em ber that the phase of a real quantity is zero if the quantity is positive and n if it is negative. The spectra |X(o>)| and ^.X(w) are shown in Fig. 4.17 for the case A 1 and L = 5. The energy density spectrum is simply the square of the expression given in (4.2.50). T here is an interesting relationship that exists b etw een the F ourier transform o f the constant am plitude pulse in E xam p le 4.2.4 and the period ic rectangular
lX(w)l
Figure 4.17 Magnitude and phase of Fourier transform of the discrete-time rectangular pulse in Fig. 4.16.
264
Chap. 4
w ave considered in E xam ple 4.2.2. If w e evaluate the F ourier transform as given in (4.2.49) at a set o f equally spaced (harm onically related) freq u en cies
w e obtain (4.2.52) If w e com pare this result with the expression for the F ourier series coefficients given in (4.2.21) for the period ic rectangular w ave, w e find that k = 0. 1 , . . . , N 1 (4.2.53)
T o elab orate, w e have established that the Fourier transform o f the rectangular p ulse, which is identical with a single p eriod o f the period ic rectangular pulse train, evaluated at the freq u en cies co = I n k / N , k = 0, 1 , ___ N 1, which are identical to the harm onically related frequency com p onents u sed in the Fourier series representation o f the periodic signal, is sim ply a m ultiple o f the Fourier coefficients f a ) at the corresponding frequencies. T he relationship given in (4.2.53) for the F ourier transform o f the rectangular pulse evaluated at co 2 n k / N , k = 0. 1........ A ' - l , and the F ourier coefficients o f the corresponding periodic signal, is not only true for these tw o signals but, in fact, holds in general. This relationship is d evelop ed further in Chapter 5.
4.2.6 Relationship of the Fourier Transform to the z-Transform

T h e z-transform o f a seq u en ce * () is defined as
OC
RO C: r2 < \z\ < r\ w here ri < |;) < rj is the region o f con vergen ce o f X (z). com plex variable z in polar form as
(4.2.54) L et us express the
(4.2.55) where r = |z| and co = 4 ;. T h en , w ithin the region o f con vergen ce of X (z), we can substitute z = r e jai into (4.2.54), T his yields (4.2.56) From the relationship in (4.2.56) w e n ote that X ( z ) can be interpreted as the Fourier transform o f the signal seq u en ce x ( n ) r ~ . T he w eigh tin g factor r ~ n is growing with n if r < 1 and decaying if r > 1 . A ltern atively, if X ( z ) con verges for
Sec. 4.2
265
1-1 = 1 , then
X (:)U , = X H =
2 2 x { , ! ) e - 1,0,1
(4.2.57)
T h erefore, the Fourier transform can be view ed as the z-transform o f the sequ en ce evaluated on the unit circle. If X ( z ) d oes not converge in the region |z| 1 [i.e.. if the unit circle is n o t contained in the region o f con vergen ce o f X (z)], the Fourier transform X ( i o ) d o es not exist. W e should n o te that the existen ce of the z-transform requires that the s e quence {jr ( )r "} be ab solu tely sum m able for som e value o f r. that is.
O C
22
n~-oc
| < oc
(4.2.58)
H en ce if (4.2.58) con verges only for values o f r > ro > 1. the z-transform exists, but the Fourier transform d oes not exist. This is the case, for exam p le, for causal seq u en ces o f the form jr(rc) = a " u(n ), where jo > 1 . T here are seq u en ces, how ever, that do not satisfy the requirem ent in (4.2.58). for exam p le, the seq u en ce x ( n ) ---------Tin sin avrc
~ oo < n < oc
(4.2,5^)
This seq u en ce d o es not have a .--transform. Since it has a finite energy, its Fourier transform con verges in the m ean-square sense to the d iscon tin uou s function X {(jo ). defined as M <*v
[ (J, < w ( < \ a > \< 7 1
(4 2 6 0 )
In con clu sion , the existen ce of the z-transform requires that (4,2.58) be sat isfied for so m e region in the z-plane. If this region contains the unit circle, the Fourier transform X(cu) exists. H ow ever, the existen ce o f the Fourier transform, which is defined for finite energy signals, d oes not n ecessarily ensure the existence o f the z-transform .
4.2.7 The Cepstrum

Let us con sid er a seq u en ce {jc(/j)} having a z-transform X (z). W e assum e that (jr(n)) is a stable seq u en ce so that X ( z ) con verges on the unit circle. The c o m p l e x ce pstru m o f the seq u en ce {jc()} is defined as the seq u en ce (cr(n)}, which is the inverse z-transform o f Cjj(z), w here Cjciz) = In X (z) (4.2.61)
T h e com p lex cepstrum exists if C^tz) con verges in the annular region n < \z\ < n , w here 0 < r\ < 1 and r2 > 1. W ithin this region o f con vergen ce, C t (z) can be represen ted by the Laurent series Cx(z) = In X (z) = c^ z ~n (4.2.62)
266
Chap. 4
where
cAn) =
f In X ( z ) z n' 1d z Jc
(4.2.63)
C is a closed contour about th e origin and lies w ithin the region o f convergence. Clearly, if C, (z) can be rep resen ted as in (4.2.62), the com plex cepstrum sequence {cr(rt)} is stable. F urtherm ore, if the com p lex cepstrum exists, Cx (z) converges on the unit circle and hence w e have
CC
CAat) = lnX(cu) =
2 2 c A n ) e ~ JU ,n
n-OC
(4.2.64)
where (cv(/i)} is the sequ en ce ob tain ed from the inverse F ourier transform of In X(to). that is, 1 r c r(n) = / ln X (a >)eJwnd w 2tt J _ n If w e express X(co) in term s of its m agnitude and p h ase, say X(a>) = \X (co)\eJf>iw) th en in X(co) = In |X(ct))| + jd(a>) (4.2.67) (4.2.66) (4.2.65)
By substituting (4.2.67) into (4.2.65), w e obtain the com plex cepstrum in the form
1 fn cx (n) I [In | X ( a >)| + j6(cL>)]eJU ,ndco 2jt J _ n
(4,2.68)
W e can separate the inverse F ourier transform in (4.2.68) into the inverse Fourier transforms o f In |X (w )| and 9(a>)\ cm(n) = 2 ce(n) = - ^
ln\X(a>)\eJl^dcL>
(4.2.69)
6(co)eJa> ndco
(4.2.70)
In som e applications, such as sp eech signal processing, only the com p onent c(n) is com puted. In such a case the phase o f X (a>) is ignored. T h erefore, the sequence {*()} cannot b e recovered from {cm(n)j. That is, the transform ation from (jr(n)} to {cm(n)} is not invertible. In speech signal p rocessing, the (real) cepstrum has b een used to separate and thus to estim ate the spectral con ten t o f the sp eech from the pitch frequency of the speech. T he com plex cepstrum is used in practice to separate signals that are con volved . T h e process o f separating tw o con volved signals is called d e c o n volution and the use o f the com p lex cepstrum to perform the separation is called h o m o m o r p h i c d econ vo lutio n. T his topic is discussed in Section 4.6.
Sec. 4.2
Frequency Analysts of Discrete-Time Signals
267
4.2.8 The Fourier Transform of Signals with Poles on the Unit Circle
A s was show n in Section 4.2.6, the Fourier transform o f a seq u en ce jc(h) can be determ ined by evaluating its z-transform X (z) on the unit circle, provid ed that the unit circle lies within the region o f con vergen ce o f X (z). O therw ise, the Fourier transform d o es not exist. T h ere are som e aperiodic sequ en ces that are neither ab solu tely sum m able nor square sum m able. H en ce their Fourier transform s do not exist. O n e such seq u en ce is the unit step seq u en ce, which has the z-transform
A n oth er such seq u en ce is the causal sinusoidal signal seq u en ce x ( n ) = (coscoon) u(n). T his seq u en ce has the z-transform
y , , _ 1 - Z ~ COS (at) 1 - 2 z _i
cosojo
+ z -2
N o te that both o f these seq u en ces have p oles on the unit circle. For seq u en ces such as these tw o exam ples, it is som etim es useful to extend the Fourier transform representation. T his can be accom plished, in a m ath em ati cally rigorous w ay, by allow ing the Fourier transform to contain im pulses at certain frequencies corresponding to the location o f the p oles o f X ( z ) that lie on the unit circle. T he im pu lses are functions o f the con tinu ou s frequency variable co and have infinite am plitude, zero width, and unit area. A n im pulse can be view ed as the lim iting form o f a rectangular pulse of height 1 /a and width a, in the limit as a -+ 0. Thus, by allow ing im pulses in the spectrum of a signal, it is p ossible to exten d the Fourier transform representation to som e signal seq u en ces that are neither absolutely sum m able nor square sum m able. T h e follow in g exam ple illustrates the exten sion of the F ourier transform rep resentation for three sequ en ces. Exam ple 4.2.5 D eterm ine the Fourier transform of the following signals. (a) x i(n) = u(n) (b) jr2(n) = (-l)" (n )
( c )j c 3(n) = (coswon)u(n)
by evaluating their z-transforms on the unit circle. Solution (a) From Table 4.3 we find that XiU) = r - t - j = - ^ r ROC: |c| > 1 1 - z '1 z -1 Xi(z) has a pole, p\ = 1, on the unit circle, but convenges for |z| > 1.
268
Chap. 4
If we evaluate A'l (:i on the unit circle, except at : = 1. we obtain

X \ ( w ) = ------ : ------------ = - : --------- e to = 2 r r k k = 0 . J . . ..
2i/ sin(w/2>
2sm(j/2)
At a) = 0 and multiples of 2, t, A |( w ) contains impulses of area . t. Hence the presence of a pole at ; = 1 (i.e.. at w = 0) creates a problem only when we want to compute A'i(w) at w = 0 .because |A](tu)i * oc as to > 0. For any other value of to. X \ (to) is finite (i.e.. well behaved). Although, at first glance one might expect the signal to have zero-frequency components at all frequencies except at as = 0. this is not the case. This happens because the signal .\|(n) is not a constant for all oc < < oc. Instead, it is turned on at n = 0. This abrupt jump creates all frequency com ponents existing in the range 0 < t o < tt. Generally, all signals which start at a finite time have nonzero-frequency com ponents everywhere in the frequency axis from zero up to the folding frequency.
(b ) From Table 3.3 we find that the .'-transform of a''uin) with a = 1 reduces to 1
XA:)
---------- r =
1+ : !
;+ l
R O C :
|;i >
which has a pole at : = - 1 = c-n . The Fourier transform evaluated at frequen cies other than a > = tt and multiples of 2tt is A'lfw)
= - ------- a> * 2 , t (k r |) k = 0.
2 cos (to/2)
1.
. ..
In this case the impulses occurs at w = Hence the magnitude is

I X-> ( t o ) | = -------- --
t t
+ 2rr k.
2[ cos(to/2)
t!)
2 tZ k +
TT
^ = 0 . 1 . . . .
and the phase is
X; (to) =
if cos < 0 Note that due to the presence of the pole at a = 1 (i.e.. at frequency w = tt), the m agnitude of the Fourier transform becomes infinite. Now \X(w)\ * oc as w > n. We observe that (])nu(n) = (cos Trn)u(n), which is the fastest possible oscillating signal in discrete time. (c) From the discussion above, it follows that A?(w) is infinite at the frequency com ponent w = too. Indeed, from Table 3.3. we find that
- - -----------) = (coStonfi)tt(n) < A-i(;) = --
1 C_1 CO S to,) 1 - 2: ~ cos wo +
ROC: | |> 1
The Fourier transform is

X 3 (to) - --------- : ------------------------------(1 -
1 - e ' 1" cos cuo
to ^ ito o 4- 27tk
k = 0. 1 . . . .
Sec. 4.2
269
The magnitude of X3 M is given by

|1 e~,)| becomes infinite. For all other frequencies, the Fourier transform is well behaved.
4.2.9 The Sampling Theorem Revisited

T o p rocess a con tinu ou s-tim e signal using digital signal processing techniques, it is necessary to convert the signal into a seq u en ce o f num bers. A s w as discussed in Section 1.4, this is usually d on e by sam pling the analog signal, say x a(t), periodically every T secon ds to produce a discrete-tim e signal x ( n ) given by x ( n ) = x a( n T ) 00 < n < 00 (4.2.71)
T h e relationship (4.2.71) describes the sam pling process in the tim e dom ain. A s discussed in C hapter 1, the sam pling frequency Fs = l / T m ust be selected large enough such that the sam pling d oes not cause any loss o f spectral inform ation (no aliasing). In d eed , if the spectrum of the analog signal can be recovered from the spectrum o f the discrete- tim e signal, there is no loss of inform ation. C onsequently, w e investigate the sam pling p rocess by finding the relationship b etw een the spectra o f signals x a(t) and x ( n ). If x a (t) is an aperiodic signal with finite energy, its (voltage) spectrum is given by the F ourier transform relation (4.2.72) w hereas the signal x a (i ) can be recovered from its spectrum by the inverse Fourier transform (4.2.73) N o te that u tilization o f all frequency com p on en ts in the infinite frequency range 00 < F < 00 is necessary to recover the signal x(t) if the signal x(t) is not bandlim ited. T h e spectrum o f a discrete-tim e signal x ( n ) , ob tain ed by sam pling x a (t), is given by the Fourier transform relation O O (4.2.74) or, equivalently,
O O
X (f) = 2 2 x ( n ) e - ^ n
dc
(4.2.75)
270
Chap. 4
The sequ en ce x(n) can be recovered from its spectrum X(a>) or X ( / ) by the inverse transform
-*{) -
2 tt J . x
X(co)eJU ,n dco
(4.2.76)
= /
J-\/2
X { f ) e j2nJnd f
In order to determ ine the relationship b etw een the spectra of the discrete time signal and the analog signal, w e n ote that period ic sam pling im poses a rela tionship betw een the in d ep en dent variables t and n in the signals x a {t) and x(n), respectively. That is, r = nT = F, (4.2.77)
This relationship in the tim e dom ain im plies a corresponding relationship betw een the frequency variables F and / in Xa(F) and X( f ) . respectively. Indeed, substitution o f (4.2.77) into (4.2.73) yields
x(n) = xa( " T ) =
Xtl(F)ej2,TnF/F'd F
(4.2.78)
If we com pare (4.2.76) with (4.2.78), we conclude that r 1/2 1/2 / \a X ( f ) e j2nf " d f = J -oc f oc X u ( F ) e )2nnFlFd F (4.2.79)
From the d evelop m en t in C hapter 1 w e know that periodic sam pling im poses a relationship b etw een the frequency variables F and / of the corresponding analog and discrete-tim e signals, respectively. That is, / = (4.2.80)
With the aid o f (4.2.80), w e can m ake a sim p le change in variable in (4.2,79), and obtain the result y
X ^ - 0 ej2nF/F' d F =
X a ( F ) e j2,,nF/F- d F
(4.2.81)
W e now turn our atten tion to the integral on the right-hand side o f (4.2.81)T he integration range of this integral can be divided into an infinite num ber of intervals o f width F,. Thus the integral over the infinite range can be expressed as a sum o f integrals, that is,
oc
3C
[{k+XflsF,
X a ( F ) e J2,TnF/F' d F =
x
/
Jt= -o c J ( k - \ f l ) F s
X a ( F ) e J2jrnF/Fd F
(4.2.82)
Sec. 4.2
271
W e o b serve that X a {F) in the frequency interval (k I ) F t to (k + ~)FS is identical to X a( F k F s) in the interval Fs/ 2 to Fs/2. C onsequently. oc V
Jk=-oc
p{k+\l2)F,
x.
*F \ :2
X a( F ) e i27TnF/F' d F =
Y2
X a ( F - k F , ) e p^ nF!f d F
J^ \fl\Fs
k=-?c
'FJ2 J2
J-FJ2
X a( F - k F < )
' ~"r ! d F (4.2.83)
where w e have used the p eriodicity of the exp on en tial, nam ely.
e j 2 n n { F + k F s )/Ft _ ^jTnnF/F,
Com paring (4.2.83). (4.2.82), and (4.2.81), w e conclude that (4.2.84)
X ( t ) = F' T . X A F - k F . ) ' *t k=--XL or, equivalently.

OC
X(f) = F
X a [ ( f - k ) F s]
(4.2.85)
This is the desired relationship betw een the spectrum X ( F / F , ) or X ( f ) o f the discrete-tim e signal and the spectrum X a( F) o f the analog signal. The righl-hand side o f (4.2.84) or (4.2.85) consists o f a periodic repetition of the sealed spectrum Fs X a( F) with period F,. This periodicity is necessary because the spectrum X ( f ) or X ( F / F ,) o f the discrete-tim e signal is periodic with period f p = 1 or Fp = Fs . For exam p le, su p p ose that the spectrum o f a band-lim ited analog signal is as sh ow n in Fig. 4.18(a). T h e spectrum is zero for [FI > B. N ow . if the sam pling frequency Fs is selected to be greater than 2 5 . the spectrum X ( F / F S) of the discrete-tim e signal will appear as show n in Fig. 4.18(b). Thus, if the sam pling frequency is selected such that Fv > 2 B. where 2 B is the N yquist rate, then = FsX a(F ) |f | < F J 2 (4.2.86)
In this case there is no aliasing and therefore, the spectrum o f the discrete-tim e signal is identical (w ithin the scale factor F,) to the spectrum o f the analog signal, within the fundam ental frequency range |F | < Fsf 2 or [ f \ < O n the other hand, if the sam pling frequency Fs is selected such that Fs < 2 B, the periodic continuation o f X a( F ) results in spectral overlap, as illustrated in Fig. 4.18(c) and (d). T h u s the spectrum X { F / F S) o f the discrete-tim e signal contains aliased frequency com p onents o f the analog signal spectrum X a(F). The end result is that the aliasing which occurs prevents us from recovering the original signal x (r) from the sam ples. G iven the discrete-tim e signal x(n) with the spectrum X ( F / F S), as illustrated in Fig. 4.18(b ), w ith no aliasing, it is n ow p ossible to reconstruct the original analog
Figure 4.18 Sampling of an analog bandlimited signal and aliasing of spectral components.
Sec. 4.2
273
signal from the sam ples j ( n ) . Since in the absence of aliasing
*.<-(*()
lo ,
r -
(4.2.87)
\ F \ > F 5/ 2
and by the Fourier transform relationship (4.2.75).
f )-
/ n= -o c
(4.2.88)
the inverse Fourier transform o f X a ( F ) is

xAD =
/
f F' p -
X a ( F ) e s~*F ' d F
(4.2.89)
Let us assum e that F s = 2 B . W ith the substitution o f (4.2.87) in to (4.2.89), we have

Xa(t)
F, J - F J
x { n ) e ~ j2 * Fn/F
dF
(4.2.90) s\n(n/T )(t - nT) (ir/T W -n T ) where x (n ) = x a( n T ) and w here T = \ / F s 1 /2 B is the sam pling interval. This is the reconstruction form ula given by (1.4.24) in our discussion o f the sam pling theorem . T h e reconstruction form ula in (4.2.90) in volves the function s in (jr /7 ); i KO =
(n/T)t
s in 2 ^ B f
2n B t
(4.2.91)
appropriately shifted by n T , n = 0 , 1 , 2 ........ and m ultiplied or w eigh ted by the corresponding sam ples x (n T ) o f the signal. W e call (4.2.90) an in terp ola tion form ula for reconstructing x a U) from its sam ples, and g (f). given in (4.2.91), is the interpolation function. W e note that at t = k T , the in terp olation function g(t n T ) is zero excep t at k = n. C onsequently. xa (t) evalu ated at t = k T is simply the sam ple x 0{k T ). A t all other tim es the w eigh ted sum o f the tim e shifted versions o f the interpolation function com b ine to yield exactly x a (t). T his com b ination is illustrated in Fig. 4.19. T he form ula in (4.2.90) for reconstructing the analog signal xa (t) from its sam ples is called the ideal interpolation fo rm u la . It form s the basis for the sa m p lin g theorem, which can b e stated as follow s. Sam pling T h eo rem . A bandlim ited con tinu ou s-tim e signal, with highest fre quency (bandw idth) B H ertz, can be uniquely recovered from its sam ples provided that the sam pling rate Fs > 2 B sam ples per second.
274
Chap, 4
A ccordin g to the sam pling theorem and the reconstruction form ula in (4.2,90), the recovery o f x a(t) from its sam ples ;c(), requires an infinite num ber o f sam ples. H ow ever, in practice w e use a finite num ber o f sam ples o f the signal and deal with finite-duration signals. A s a con seq u en ce, w e are con cern ed only with reconstructing a finite-duration signal from a finite num ber o f sam ples. W hen aliasing occurs due to to o low a sam pling rate, the effect can be de scribed by a m ultiple folding o f the frequency axis o f the frequency variable F for the analog signal. Figure 4.20(a) show s the spectrum X a(F ) o f an analog signal. A ccordin g to (4.2.84), sam pling of the signal with a sam pling frequency Fs results in a periodic repetition o f X a( F ) with period Fs . If Fs < 2 B , the shifted replicas of X g {F) overlap. T he overlap that occurs within the fundam ental frequency range F.J2 < F < Fs/2, is illustrated in Fig. 4.20(b ). T h e corresponding spectrum of the discrete-tim e signal w ithin the fundam ental frequency range, is obtained by adding all the shifted portions within the range | / | < j , to yield the spectrum shown in Fig. 4.20(c). A careful inspection of Fig. 4.20(a) and (b) reveals that the aliased spectrum in Fig. 4.20(c) can be obtained by folding the original spectrum lik e an accordian with pleats at every odd m ultiple o f Fs/2. C on sequ en tly, the frequency F J 2 is called the f o l d i n g fre q u e n c y , as indicated in C hapter 1. Clearly, then, periodic sam pling autom atically forces a folding o f the frequency axis o f an analog signal at odd m ultiples o f Fs/2, and this results in the relationship F = f Fs b etw een the frequencies for con tinu ou s-tim e signals and discrete-tim e signals. D u e to the fold ing o f the frequency axis, the relationship F f F, is not truly linear, but piecew ise linear, to accom m odate for the aliasing effect. T his relationship is illustrated in Fig. 4.21. If the analog signal is bandlim ited to B < Fs/2, the relationship betw een / and F is linear and o n e-to-on e. In other words, there is no aliasing. In practice, prefiltering with an antialiasing filter is usually em p loyed prior to sam pling. This ensures that frequency com p onents o f the signal above F > B are sufficiently attenuated so that, if aliased, they cause n egligib le distortion on th e desired signal. T h e relationships am ong the tim e-d om ain and freq u en cy-d om ain functions x a (t), x ( n ) , X ( F ), and X ( f ) are sum m arized in Fig, 4.22. T h e relationships for
Sec. 4.2

XJFi
275
0
(a)
o T
(c)
Figure 4.20 Illustration of aliasing around the folding frequency.
Figure 4.21
R elationship betw een frequency variables F and / .
276

XJF)= Fourier transform pair Xa(t) xa(t) = j ~ Xa(F )ei -^' dF
a:0(F)
Chap. 4
\. Xa(F)
Reconstruction:
= tfrt)
sin 7r(r - nT)!T
F, IFI < - j
7r(t - nT)!T
Sampling-, x(n) - x a(nT)
x(n )
X (/) Fourier transform pair x(n) = J
T
Figure 4.22
X(f)eill,f"df
Time-domain and frequency-domain relationships for sampled sig
nals. recovering the con tinu ou s-tim e functions, x 0{t) and X a{F ), from th e discrete-tim e quantities x ( n ) and X ( f ) , assum e that the analog signal is bandlim ited and that it is sam pled at the N yquist rate (or faster). T h e follow in g exam p les serve to illustrate the problem o f the aliasing of frequency com ponents.
Example 4.2.6 Aliasing i n Sinusoidal Signals

The continuous-time signal
xa(r) ~ cos 2nF^t
LgjZxFai
has a discrete spectrum with spectral lines at F = F U > as shown in Fig. 4.23(a). The process of sam pling this signal with a sam pling frequency Fs introduces replicas of the spectrum about multiples of Fs. This is illustrated in Fig. 4.23(b) fo r Fs/2 < F0 < F,To reconstruct the continuous-time signal, we should select the frequency com ponents inside the fundam ental frequency range \F\ < Fs f l . The resulting spectrum
Sec. 4.2
277
Spectrum
_| -Fo 0
U)
Spectrum
F0
' it
-F t
Fu - Fs 0 F,. - F0
(b) Spcctrum
F, -<FS- / (,) f-,~F0 ^ 2 (c)

Spectrum
2T\
- F q - Fs Fs F, - F0 0 F0 - F, F,
(d) Spectrum
F 0 + F,
2 -y
F ,
0
(e)
f3
Figure 4.23
A liasing of sinusoidal signals.
278
Chap. 4
is shown in Fig. 4.23(c). The reconstructed signal is

* ( /) = c o s 2 j t ( F , F (t)i
Now. if Fv is selected such that Fs < F(l < 3F 5/2, the spectrum of the sampled signal is shown in Fig. 4.23(d). The reconstructed signal, shown in Fig. 4.23(e). is xu(t) = cos2jr(F ( 1 - Fs)i In both cases, aliasing has occurred, so that the frequency of the reconstructed signal is an aliased version of the frequency of the original signal.
E x a m . e 4.2.7
S am p lin g a N o n b a n d lim ited Signal
Consider the continuous-time signal *(/) = e - A I,i whose spectrum is given hv X J F ) = ----- -------A- 4 - (2;r F Y Determ ine the spectrum of the sampled signal ,r(n) = xu{nT). Solution If we sample x(t) with a sampling frequency F, = 1/7", we have jr(n) xuinT) e~A1'"' = (e~A1 oc < n < oc A>0
The spectrum of x(n) can be found easily if we use a direct com putation of the Fourier transform. We find that F\ F, J 1 - e~1AT 1 - 2e~A1 cos I n FT + e~2AT T = 1 Fs
Clearly, since c o s 2 x F T = c o s 2 t t ( F/ Fs ) is periodic with period Fs, so is X ( F / F S), Since X a(F) is not bandlim ited. aliasing cannot be avoided. The spectrum of the reconstructed signal i(r) is i - i . Xa(F) =
[o .
\F\<!
I F I > J
( t.
Figure 4.24(a) shows the original signal xa(t) and its spectrum X a( F ) for A = 1The sampled signal x(n) and its spectrum X ( F / F S) are shown in Fig. 4.24(b) for = 1 Hz. The aliasing distortion is clearly noticeable in the frequency domain. The reconstructed signal xa(r) is shown in Fig. 4.24(c). The distortion due to aliasing can be reduced significantly by increasing the sampling rate. For example, Fig. 4.24(d) illustrates the reconstructed signal corresponding to a sampling rate Fs = 20 Hz. It is interesting to note that in every case xa( nT) = xa(nT), but x{t) / xa(t) at other values of time.
Sec. 4.2
279 14
A - + (2 ttF) 2
.*,,(/) = e~A > 1 1 .A = I
x (n :
(al
1.0
-2
-1
1 t
[
.
<b)
(c }
(d)
Figure 4.24 (a) A nalog signal xa(i ) and its spectrum Xa(Fy. (b) xin i = x a ( ti T) and the spectrum of *(n) for A = 1 and F, = 1 Hz; (c) reconstructed signal x(i) for F, - 1 Hz: (dj reconstructed signal i aU) for Fs = 20 Hz.
4.2.10 Frequency-Domain Classification of Signals: The Concept of Bandwidth

Just as w e have classified signals according to their tim e-dom ain characteristics, it is also desirable to classify signals according to their frequency-dom ain character istics. It is com m on practice to classify signals in rather broad terms according to their frequency content. In particular, if a p ow er signal (or energy signal) has its p ow er density sp ec trum (or its energy density spectrum ) concentrated about zero frequency, such a signal is called a low -freq ue ncy signal. Figure 4.25(a) illustrates the spectral characteristics o f such a signal. O n the other hand, if the signal p ow er density
280 Xa{F)

X{a> )
Chap. 4
fa) Xa(F) X(w)
(b) XJF) X(oj)
(c) Figure 4.25 (a) Low-frequency, (b) high-frequency, and (c) medium-frequency signals. spectrum (or the energy density spectrum ) is concentrated at high frequencies, the signal is called a h igh -fr eq uency signal. Such a signal spectrum is illustrated in Fig. 4.25(b). A signal having a p ow er density spectrum (or an energy density spectrum ) concentrated som ew h ere in the broad frequency range b etw een low fre q uencies and high frequencies is called a m e d i u m - f r e q u e n c y signal or a bandpass signal. Figure 4.25(c) illustrates such a signal spectrum . In addition to this relatively broad frequency-dom ain classification o f signals, it is often desirable to express q uantitatively the range o f freq u en cies over which the p ow er or energy density spectrum is concentrated. This q u an titative m easure is called the b a n d w i d t h o f a signal. For exam p le, su p p ose that a con tinu ou s tim e signal has 95% o f its p ow er (or energy) density spectrum con cen trated in the frequency range F\ < F < F 2. T hen the 95% bandw idth o f the signal is F2 F 1 . In a sim ilar m anner, w e may define the 75% or 90% or 99% bandw idth o f the signal.
Sec. 4.2
281
In the case o f a bandpass signal, the term n a r r o w b a n d is used to describe the signal if its bandw idth Fz - F\ is m uch sm aller (say, by a factor o f 10 or m ore) than th e m edian frequency {Fi + F \) f2. O therw ise, the signal is called wideband. W e shall say that a signal is b a n d lim ite d if its spectrum is zero ou tsid e the freq u en cy range > B. For exam ple, a con tinu ou s-tim e finite-energy signal x (t) is bandlim ited if its Fourier transform X ( F ) = 0 for m > B. A d iscrete-tim e finite-en ergy signal x ( n ) is said to be (periodically) b a n d lim ited if jX(ti>)| = 0 for a>o < M < x
Sim ilarly, a p eriod ic co n tinu ou s-tim e signal x p (r) is p eriodically bandlim ited if its F ourier coefficien ts c* = 0 for || > M , w here M is som e p ositive integer. A p erio d ic discrete-tim e signal w ith fundam ental period N is periodically bandlim ited if the Fourier coefficients ck = 0 for kG < |j < N . Figure 4.26 illustrates the four types o f bandlim ited signals. B y exploiting the duality b etw een the frequency dom ain and the tim e dom ain, we can provide sim ilar m ean s for characterizing signals in the tim e dom ain. In particular, a signal x ( t ) will be called tim e- lim ited if x(r) =
0 |f|
>
If the signal is p eriod ic with period Tn, it will be called periodic ally time- lim ited if
x p it) = 0 r < \t\ < T p f l
If w e have a d iscrete-tim e signal x( n ) o f finite duration, that is, .r(n) = 0 In I > N
it is also called tim e-lim ited. W hen the signal is period ic with fundam ental period it is said to be p eriod ically tim e-lim ited if x(n} = 0 no < )n\ < N
Figure 4 J 6
Some exam ples of bandlim ited signals.
282
Chap. 4
W e state, w ithout proof, that no sign al can be time-lim ited a n d ban dlimited sim ultaneously. Furtherm ore, a reciprocal relationship exists b etw een the time duration and the frequency duration o f a signal. T o elaborate, if w e have a shortduration rectangular pulse in the tim e dom ain, its spectrum has a width that is inversely proportional to the duration o f the tim e- dom ain pulse. T he narrower the pulse b eco m es in the tim e dom ain, the larger the bandwidth of the signal b ecom es. C onsequently, the product o f the tim e duration and the bandwidth of a signal cannot be m ade arbitrarily sm all. A short-duration signal has a large bandwidth and a sm all bandwidth signal has a lon g duration. T hus, for any signal, the tim e-b an dw id th product is fixed and cannot b e m ade arbitrarily small. Finally, w e n o te that w e have discussed frequency analysis m eth od s for peri odic and aperiodic signals with finite energy. H o w ev er, there is a family o f deter m inistic aperiodic signals with finite pow er. T h ese signals consist o f a linear super p osition o f com plex exp on en tials with nonharm onically related frequencies, that is, M x{n) = Y l A ke ^ n k= 1 w here &>i, a s , . . . , a > m are nonharm onically related. T h ese signals have discrete spectra but the distances am ong the lin es are nonharm onically related. Signals with discrete nonharm onic spectra are som etim es called quasi-periodic.
4.2.11 The Frequency Ranges of Some Natural Signals

T he frequency analysis tools that w e have d ev elo p ed in this chapter are usually applied to a variety o f signals that are en cou n tered in practice (e.g., seism ic, b io lo g ical, and electrom agn etic signals). In gen eral, the frequency analysis is perform ed for the purpose o f extracting inform ation from the ob served signal. For exam ple, in the case o f biological signals, such as an E C G signal, the analytical tools are used to extract inform ation relevant for d iagn ostic purposes. In th e case o f seism ic signals, w e may b e interested in detectin g th e p resen ce o f a nu clear exp losion or in d eterm ining the characteristics and location o f an earthquake. A n electrom agnetic signal, such as a radar signal reflected from an airplane, con tains inform ation on the p osition o f the plane and its radial velocity. T h ese param eters can be estim ated from observation o f the received radar signal. In processing any signal for the p urpose o f m easuring param eters or ex tracting other types o f inform ation, o n e m ust know approxim ately the range o f freq u en cies contained by the signal. For referen ce, T ables 4.1, 4.2, and 4.3 give approxim ate lim its in the frequency d om ain for b iological, seism ic, and electro m agnetic signals.
4.2.12 Physical and Mathematical Dualities

In the p revious section s o f the chapter w e have introduced several m eth od s for the frequency analysis o f signals. Several m eth od s w ere n ecessary to accom m odate the
Sec. 4.2

TABLE 4.1
SIGN ALS FR EQ U E N C Y R A NG ES OF SO M E BIO LOG ICAL
283
Type of Signal E lectroretinogram 8 Electronystagtnogram h P neum ogram ' Electrocardiogram (EC G ) E lectroencephalogram (E E G ) Electrom yogram d Sphygm om anogram e Speech "A A CA dA CA
F requency R ange (Hz) 0-20 0-20
0-40 0-100 0-100

10-200 0-200 100-^000
graphic recording of retina characteristics. graphic recording of involuntary m ovem ent of the eyes. graphic recording of respiratory activity. graphic recording of m uscular action, such as m uscular contraction. recording of blood pressure.
FR E Q U E N C Y R A NG ES OF SO M E S E IS M IC SIGNALS
TABLE 4.2
Type o f Signal W ind noise Seismic exploration signals E arthquake and nuclear explosion signals Seismic noise
F requency R ange (Hz) 100-KXX) 10-11X1 (1.01-1(1 0.1-1
TABLE 4.3
F R E Q U E N C Y RANGES O F E L E C T R O M A G N ETIC SIGN ALS
Type of Signal R adio broadcast S hortw ave radio signals R adar, satellite com m unications. space com m unications. com m on-carrier microwave Infrared Visible light U ltraviolet G am m a rays and x-rays
W avelength (m )
10M 02 io-- i o - :
F requency R ange (Hz) 3 x lO4^ x JO6 3 x 1(^-3 x 10U 1
3.9
1-10 u rM o ^ 6 10_7-8.1 10~7 10 -7-]O - 8 u r M o - 10
3 x 3x 3.7 x 3 x 3 x
10^-3 x 10 11 3x 10w-7.7 1015-3 x 1017 3 x
1010 I0 i4 x 1014 lO16 101 K
differen t types o f signals. T o sum m arize, the follow in g frequency analysis tools have b een introduced: 1. T he F ourier series for con tinu ou s-tim e p eriod ic signals. 2. T he F ourier transform for con tinu ou s-tim e ap eriod ic signals. 3. T he F ourier series for discrete-tim e periodic signals.
4. T he F ourier transform for discrete-tim e aperiodic signals.
284
Chap. 4
Figure 4.27 sum m arizes the analysis and synthesis form ulas for these types of signals. A s w e have already indicated several tim es, there are two tim e-dom ain char acteristics that determ ine the type of signal spectrum w e obtain. T h ese are whether the tim e variable is con tinu ou s or discrete, and w h eth er the signal is periodic or aperiodic. Let us briefly sum m arize the results o f the previous sections.
Continuous-time signals have aperiodic spectra. A close inspection of the F ourier series and Fourier transform analysis form ulas for continuous-tim e signals d o es not reveal any kind of periodicity in the spectral dom ain. This lack of p eriodicity is a con sequ en ce of the fact that the com p lex ex p on en tial exp(y2jr Ft) is a function o f the continuous variable t, and hence it is not p eriod ic in F . Thus the frequency range o f continuous-tim e signals exten d s from F = 0 to F = 0 0 . Discrete-time signals have periodic spectra. In d eed , both the Fourier series and the Fourier transform for discrete-tim e signals are p eriod ic with period co = 2n . A s a result o f this periodicity, the frequency range o f d iscrete-tim e signals is finite and exten d s from co = 1r to w tt radians, where w = n corresponds to the highest p ossible rate o f oscillation. Periodic signals have discrete spectra. A s we have ob served , periodic signals are described by m eans o f Fourier series. T h e Fourier series coefficients provide the lines" that constitute the discrete spectrum . T h e line spacing A F or A f is equal to the inverse o f the period Tp or N , resp ectiveiy, in the time dom ain. That is. A F = 1/7,, for continuous-tim e period ic signals and A f = 1 / N for discrete-tim e signals. Aperiodic finite energy signals have continuous spectra. This prop erty is a direct con seq u en ce o f the fact that both X ( F ) and X ( w ) are functions o f exp { j 2 n F t ) and exp {jeon), respectively, which are con tinu ou s functions o f the variables F and co. The continuity in frequency is necessary to break the harmony and thus create aperiodic signals. In sum m ary, w e can conclude that periodic ity with p e r i o d a in o n e domain automatically im plies discretization with s p a c i n g " o f 1 jot in the o th e r do m ain , and vice versa. If w e k eep in m ind that p eriod in the frequency d om ain m eans the fre quency range, spacing'' in the tim e dom ain is the sam pling p eriod T , line spacing in the frequency dom ain is A F , then a = Tp im plies that 1 /a = \ j T p = A F , a = N im plies that A f = \ / N , and a = Fs im plies that T = 1 /F S. T h ese tim e-frequency dualities are apparent from observation o f Fig. 4.27. W e stress, how ever, that the illustrations u sed in this figure do n ot correspond to any actual transform pairs. Thus any com parison am ong them sh ou ld be a v o i d e d . A careful in sp ection o f Fig. 4 .2 7 also reveals som e m athem atical s y m m e tr ie s and dualities am ong the several frequency analysis relationships. In particular,
Figure 4.27
Summary of analysis and synthesis form ulas.
286
Chap. 4
we observe that there are dualities b etw een the follow in g analysis and synthesis equations: 1. T he analysis and synthesis eq u ation s o f the con tinu ou s-tim e Fourier trans form. 2. T he analysis and synthesis eq u ation s of the discrete-tim e F ourier series. 3. T he analysis eq u ation o f the con tinu ou s-tim e F ourier series and the synthesis equation o f the discrete-tim e Fourier transform.
4. T he analysis equation o f the discrete-tim e F ourier transform and the synthesis

equation o f the continuous-tim e Fourier series. N o te that all dual relations differ on ly in the sign of the exp on en t of the corresponding com p lex exp on en tial. It is interesting to note that this change in sign can be thought o f either as a folding of the signal or a folding o f the spectrum , since
e - j 2 n Fl _ e j 2n( - FU _
If w e turn our atten tion now to the spectral density o f signals, w e recall that we have used the term ener gy density sp e ctru m for characterizing finite-energy aperiodic signals and the term p o w e r density sp e ctru m for period ic signals. This term inology is con sistent with the fact that periodic signals are p ow er signals and aperiodic signals w ith finite energy are energy signals.
4.3 PROPERTIES OF THE FOURIER TRANSFORM FOR DISCRETE-TIME SIGNALS

The Fourier transform for aperiodic finite-energy discrete-tim e signals described in the preceding section p ossesses a num ber of properties that are very useful in reducing the com p lexity o f frequency analysis problem s in m any practical appli cations. In this section w e d evelop the im portant p roperties o f the Fourier trans form. Similar p roperties hold for the Fourier transform of ap eriod ic finite-energy continuous-tim e signals. For co n v en ien ce, w e adopt the notation
OC
X(a>) = F {x{)} =
/J CC
x ( n ) e ~ Jcun
(4.3.1)
for the direct transform (analysis eq u ation ) and x ( n ) = F ~ 1{X(tv)] = [ X{u>)ejmndoo (4.3.2)
2 * J 2n
for the inverse transform (syn thesis eq u ation ). W e also refer to x ( n ) and X(o)) as a Fourier tra nsfo rm p a i r and d en ote this relationship with the n otation
Sec. 4.3
Properties of the Fourier Transform for Discrete-Time Signals
287
R ecall that X (a>) is periodic with period 2 n . C on sequ en tly, any interval o f length 2 n is sufficient for the specification o f the spectrum . U sually, we plot the spectrum in the fundam ental interval [ jr. tt J. W e em p hasize that all the spectral inform ation contained in the fundam ental interval is necessary for the co m p lete d escription or characterization o f the signal. For this reason, the range o f integration in (4.3.2) is always 2jt, in d ep en dent o f the specific characteristics o f the signal w ithin the fundam ental interval.
4.3,1 Symmetry Properties of the Fourier Transform

W hen a signal satisfies som e sym m etry p roperties in the tim e dom ain, these prop erties im pose so m e sym m etry conditions on its Fourier transform . E xp loitation of any sym m etry characteristics leads to sim pler form ulas for both the direct and inverse F ourier transform. A discussion o f various sym m etry properties and the im plications o f these properties in the frequency dom ain is given here. Suppose that both the signal x(n) and its transform X(co) are com plex-valued functions. T hen they can be expressed in rectangular form as jr(n) = x K(n) + jX f(n ) X'(co) = X k {(o ) + j X [ ( co) (4.3.4) (4.3.5)
By substituting (4.3.4) and e~>w cosu> - / s in co into (4.3.1) and separating the real and im aginary parts, we obtain
OC
Xk(lo) =
22
ft = OC
costu/i -f- xj{n) sin con]
(4.3.6)
S C X/(a>) = 2 2 [*jfOO sin a>n.*/() cos&inj n=-oc
(4.3.7)
In a sim ilar m anner, by substituting (4.3.5) and eJm = cos co + j s m u > into (4.3.2), w e obtain x R(n) = - f [X/?(a>) cosam X/(co) sin con]dco 2 n J2n (4.3.8)
X/(n)
f [X ft (co) sincon + X/ ( w) cos con]dco
(4.3.9)
"T J2k N o w , let us in vestigate som e special cases.
Real signals. If x(n) is real, then x/t(n) = x(n) and xi(n) = 0. (4.3.6) and (4.3.7) reduce to
OC
H en ce
X R(co) 2 2 x ( n ) cos con
(4.3.10)
288
Chap. 4
and X i { w) = * ( )s in o jrt (4.3.11)
Since c o s (con) = cos ton and sin (con) = sin cun, it follow s from (4.3.10) and (4.3.11) that
X R( - w ) = X R(co) X, ( -co) = -X /(o > )
(ev en ) (o d d )
(4.3.12) (4.3.13)
If w e com bine (4.3.12) and (4.3.13) into a single eq u ation , w e have
X*(co) = X ( - w )
(4.3.14)
In this case we say that the spectrum o f a real signal has H e r m itia n sy m m etry. W ith the aid o f Fig. 4.28, w e observe that the m agnitude and phase spectra for real signals are X(to)\ =
J X\{co)
X
r
+ Xj(co)
(4.3.15) (4.3.16)
4 _X|oj| = tan
( co)
A s a con sequ en ce o f (4.3.12) and (4.3.13), the m agnitude and p h ase spectra also p ossess the sym m etry properties |,Y(co)| = \ X { co)\ liX (-co) = -& X (to ) (ev en ) (od d ) (4.3.17) (4.3.18)
In the case o f the inverse transform o f a real-valued signal [i.e., Jt(n) = jcj?(n)], (4.3.8) im plies that (n) =
[X/f(a>) cosa>n X /(a>)sinam ]dw
(4.3.19)
Jin
R(co) Since both products X R (to) cos ton and X /( oj) sin con are even fun ction s o f co, we have 1 r = I [X^(cu) cos con X/(co) sin con]dco
x Ja
(4.3.20)
Imaginary axis
functions.
Sec. 4.3
289
Real and even signals. If .*<) is real an d ev e n [i.e., x ( ~ n ) = * ( )], th en jf(/i)cosw ?i is ev en a n d x ( n )s in w n is od d . H e n c e , fro m (4.3.10). (4.3.11). an d (4.3.20) w e o b ta in
Xff(a>) = x (0 ) + 2 X i ( w) = 0 jr(n) = j f * X R(o>) cos cun da> x(n)coswn (e v e n ) (4.3.21) (4.3.22) (4.3.23)
T h u s real a n d ev en signals po ssess real-v a lu e d sp e c tra , w hich, in ad d itio n , a re even fu n ctio n s o f th e freq u e n cy v ariab le a> .
Real and odd signals. If x(/i) is real an d o d d [i.e., x ( n) = - x ( ) ] , th e n x ( n )c o s w n is o d d an d .vODsinwn is even. C o n se q u e n tly . (4.3.10). (4.3.11) and (4.3.20) im ply th a t
A'k(oi) = 0 (o d d ) ( 4.3.24) (4.3.25)
(4.3.26) T h u s re a l-v a lu e d o d d signals possess p u rely im ag in arv -v alu ed sp e ctral c h a ra c te ris tics. w hich, in a d d itio n , a re o d d f unct i ons o f th e freq u e n cy v ariab le co.
Purely imaginary signals. In th is case x K(n) = 0 an d x ( n ) - j x f (n). T h u s (4.3.6). (4.3.7), a n d (4.3.9) re d u c e to

(o d d ) (4.3.27)
(ev en )
(4.3.28)
(4.3.29) If x / ( n ) is o d d [i.e., x / ( n) = x/ (n)], th e n (o d d ) X/(a>) = 0 (4.3.30) (4.3.31) (4.3.32)
290
Chap. 4
S im ilarly, if xj ( n) is ev en [i.e.. x / ( n) .v/()]. we h ave X *M 0 (ev en ) (4.3.33) (4.3.34) X X [ ( oj) = x /(0 ) + 2 Y ^ x i ( n ) cos con
n= 1
1 r x ; (n) = I
XJ o
X 1 (co) cos cun d w
(4.3.35)
A n a rb itra ry , possibly co m p lex -v alu ed signal x (n ) can be d e c o m p o s e d as x ( n ) = Xfi(n) + j x i ( n ) = x R e (n) + x ^ ( n) + j [ x ] ( n ) + *"( )] (4.3.36) = Ar (/i) + Jr(rt) w h ere, by d efin itio n , x f (n) = x K f ( n) - h j Xf ( n ) = j[at() + jr* (- n )] x(n) = x'x(n) + j x/ ( r i ) - |[ jr( n ) - x * ( - n ) ] T h e su p e rscrip ts e a n d o d e n o te th e even an d o d d signal c o m p o n e n ts , respectively. W e n o te th a t x e(n) x e(n) a n d x r,(n) = - a :(). F ro m (4.3.36) a n d the F o u rie r tra n sfo rm p ro p e rtie s e s tab lish ed ab o v e, w e o b ta in th e follow ing relatio n sh ip s: x{n) = [*() + j x l (n)\ + [.i*(/0+y'A'/(n)] ] xito) = [x( R(co) + j x f u o ) 1 + ' + j x ) \ w) ] (4.3.37)
T h e se sy m m e try p ro p e rtie s of th e F o u rie r tra n sfo rm are su m m a riz e d in T a ble 4.4 an d in Fig. 4.29. T h ey a re o ften used to sim plify F o u rie r tra n sfo rm calcu latio n s in p ractice.
Example 4.3.1
Determ ine and sketch X R(w), X , ( w ), lîoOi. and ^X(co) for the Fourier transform X(a>) = 1
1 a e ~ ,w
- 1< a < 1
(4.3.38)
Solution By multiplying both the num erator and denom inator of (4 .3 .3 8 ) by the complex conjugate of the denom inator, we obtain
1 ae!w 1 a co s co j a sin co X(w) - ----------------------------- = ------------------ ------ (1 a e ~ }U,) ( 1 - a e JU >) 1 2a c o s o j + a "
This expression can be subdivided into real and imaginary parts. Thus we obtain
X r (w )
1
1
2
a cos co
a cos o j + a 1
a sin w X,(u>) = - 1 2a cos co + a 2

Substitution of the last two equations into (4.3.15) and (4.3.16) yields the mag nitude and phase spectra as
|X(<u)i -
v l 2a cos co
1 - ==
a2
(4.3.39)
Sec. 4.3
Properties of the Fourier Transform for Discrete-Time Signals TABLE 4.4

SYMMETRY PROPERTIES OF THE DISCRETE-TIME FOURIER TRANSFORM
291
Sequence
xin)
x (n )
DTFT
A ' (w) Xi-w) X (a>)
j 'l - n ) J(n)
jx /[n )
A>(n) =
x(n) =
" IX R\.u
^)]
A'/jlo)) =: ^[X(t^) -V*( aj)j

j X, { t o)
Real Sienals
4 [ j r ( n } + a ' ( n ) ]
- , v ( n)]
A n y real signal x (n )
X (w ) = X ' ( - u i ) X r (w ) = X r { id) Xfiuj)

IX ( oj)| = ] X ( <y)|
\ t.(n)
x.An)
i (a (> > )
i|.r(n) -
,r( n )]
r l n)]
= -iX (-w ) An(to)

(real and even)
(r e a l an d e v e n )
]X
(imaginary and odd)
(r e a l a n d o d d )
F igure 4.29
Sum m ary of symmetry properties for the F o u rier transform .
292
Chap, 4
and
XrX(w) = - tan 1 1 a cos w
(4.3.40)
Figures 4.30 and 4.31 show the graphical representation of these spectra foT a = 0.8. The reader can easily verify that as expected, all symmetry properties for the spectra of real signals apply to this case. Example 4.3.2 D eterm ine the Fourier transform of the signal
| A.
M < n < M
1 0,
elsewhere
Teal
(4.3.41) and even signal. From (4.3.21)
Solution CleaTly, x { n) = x{n). Thus *() is a we obtain
X(a>) = X K(a> ) = A I 1 4- 2 ^ c o s
X g((i> )
(a)
(b)
Figure 4 J 0 G raph of X r ( o> ) and X / ( w ) for the transform in Exam ple 4,3.1.
Sec. 4.3

IXlcoM
293
(a)
^X{w)
Figure 4.31
Magnitude and phase spectra of the transform in Example 4.3.1.
If we use the identity given in Problem 4.13, we obtain the simpler form X(u>) = A sinfAf + l)oj . - -J sin(tu/2)
Since X (a>) is real, the magnitude and phase spectra are given by sin(Af +
IX M I = A
I |
sin(a)/2)
(4,3.42)
and 0,
jt,
if X{w) > 0 if X(a> )< 0
(4.3.43)
Figure 4,32 shows the graphs for X(w).
294

X(n)
Chap. 4
- MOM
X(u>)
IXMI
Z X(w)
4.3.2 Fourier Transform Theorems and Properties

In th is se ctio n w e in tro d u c e se v era l F o u rie r tra n sfo rm th e o re m s a n d illu stra te th eir u se in p ra c tic e b y ex am p les.
Linearity.
If
Sec. 4,3
Properties of the Fourier Transform for Discrete-Time Signais
295
an d X2 (n) * X ;(a/) th en (4.3.44) Sim ply sta te d , th e F o u rie r tra n sfo rm a tio n , view ed as an o p e ra tio n on a signal jt(rt), is a iin e a r tra n sfo rm a tio n . T h u s th e F o u rie r tra n sfo rm o f a lin e a r c o m b in a tio n o f tw o o r m o re signals is eq u al to th e sam e lin e a r co m b in a tio n o f th e F o u rie r tra n sfo rm s o f th e in d iv id u al signals. T h is p ro p e rty is easily p ro v e d by using (4.3.1). T h e lin earity p ro p e rty m a k e s th e F o u rie r tra n sfo rm su ita b le fo r th e stu d y o f lin ear system s. Example 4.3.3 Determ ine the Fourier transform of the signal (4.3.45)
F
Beginning with the definition of the Fourier transform in (4.3.1), we have Xi(a>) =
2 2 M'Oe- -""" = 22a 'e i=D

1 ae~)w
The summation is a geometric series that converges to
provided that
\ae ""I = Icij |e ""| = \a] < 1
which is a condition that is satisfied in this problem. Similarly, the Fourier transform of X j i n ) is
296
Chap, 4
By combining these two transforms, we obtain the Fourier transform of x(n) in the form X ((i> ) = l _ a:
1 2a co so j + a 2
(4.3.46)
Figure 4.33 illustrates *(n) and X(w) for the case in which a = 0.8.
Time shifting.
th e n
If x (n ) x ( n - k) X (a)) e - ja,kX ( w ) (4.3.47)
T h e p ro o f o f th is p ro p e rty follow s im m e d ia te ly from (he F o u rie r tra n sfo rm of x ( n k) by m ak in g a change in th e su m m a tio n index. T h u s F[x{n - k ) } = X(a>)e~J< uk = \X(aj )\ej[^ Xuu]- < ,,k]
X(a>)
Figure 4.33
Sequence x I n ) and its F ourier transform in Exam ple 4.3.3 with
Sec. 4.3
297
T h is re la tio n m ean s th a t if a signal is sh ifted in th e tim e d o m a in by k sam p les, its m a g n itu d e sp e c tru m re m a in s u n c h a n g e d . H o w e v e r, th e p h ase sp e c tru m is ch an g ed by an a m o u n t -cok. T his re su lt can easily b e e x p la in e d if w e recall th a t th e fre q u e n c y c o n te n t o f a signal d e p e n d s only on its sh a p e. F ro m a m a th e m a tic a l p o in t o f view , w e can say th a t shifting by k in th e tim e d o m a in , is e q u iv a le n t to m u ltip ly in g th e sp e c tru m by e~i'ok in th e fre q u e n c y d o m ain . T im e r e v e r s a f . If j:() th e n x(-n) X(-a> ) (4.3.48) X(co)
T h is p ro p e rty can be e s tab lish ed by p e rfo rm in g th e F o u rie r tra n sfo rm a tio n o f jt ( n) an d m ak in g a sim ple ch an g e in th e su m m a tio n in d ex . T h u s O C F (x (-n )j =
22 x ( l '> e>W i = f= O C
If x ( n ) is real, th e n from (4.3.17) an d (4.3.18) w e o b ta in F ( * ( - n ) | = X{-a>) = = \ X( w) \ e ~ / ^ XiM' T h is m e a n s th a t if a signal is folded a b o u t th e origin in tim e, its m a g n itu d e sp e ctru m re m a in s u n c h a n g e d , a n d th e p h a se sp e c tru m u n d e rg o e s a c h a n g e in sign (p h ase re v e rsa l). C o n v o lu tio n t h e o r e m . If F x\ (n) 4 X ](w) an d x 2 (n) th e n x (n ) = Xi (h) * X2 (n) X(co) = X] (a))X 2 (& >) (4.3.49) X 2 {oj)
T o p ro v e (4.3.49), w e recall th e co n v o lu tio n fo rm u la O C x ( n ) = ^ i(n ) * x 2 (n) = Y x \ ( k ) x 2 (n - k)

k= oc
B y m u ltip ly in g b o th sides o f this e q u a tio n by th e e x p o n e n tia l e x p ( jeon) a n d su m m in g o v e r all n, w e o b ta in O C X(a>)= ^ x ( n ) e ~ Jwn = O C ^ " O C ^ x \ { k ) x 2(n k) e~]a> n
298
Chap. 4
A fte r in te rc h a n g in g th e o rd e r o f th e su m m a tio n s a n d m ak in g a sim ple ch an g e in th e su m m atio n in d ex , th e rig h t-h a n d side o f this e q u a tio n re d u c e s to th e p ro d u ct Xi(a>)X 2 (io). T h u s (4.3.49) is estab lish ed . T h e co n v o lu tio n th e o re m is o n e o f th e m o st p o w erfu l to o ls in lin e a r system s analysis. T h a t is, if w e co nvolve tw o signals in th e tim e d o m a in , th e n this is e q u iv a le n t to m u ltip ly in g th e ir sp e c tra in th e fre q u e n c y d o m ain . In la te r c h ap ters we will see th a t th e co n v o lu tio n th e o re m p ro v id es an im p o rta n t c o m p u tatio n al to o l fo r m an y d ig ital signal p ro cessin g ap p licatio n s. Example 4.3.4 By use of (4.3.49). determ ine the convolution of the sequences A'l(/1) = A:(h ) = {1. 1. 1)
t
Solution By using (4.3.21). we obtain X i ((o) = X 2(w ) = 1 + 2 cos w Then X (co) = A't (cijjASfa)) = (1 + 2coso>) = 3 + 4 cos w -i- 2 cos ho = 3 4- 2 (c"" + ( "") + ic' 2"' + Hence the convolution of with is
.v (ii) = ( 1 2 3 2 1 )
Figure 4.34 illustrates the foregoing relationships.
The correlation theorem.
If x i(n ) X \(a>)
an d A ;( ) th en ri,x2(m ) S ,; (w) = X \ ( c d) X 2 (oj) (4.3.50) Xzito))
T h e p ro o f o f (4.3.50) is sim ilar to th e p ro o f o f (4.3.49). In th is case, w e have X

r^xAn) =
^
x ] ( k ) x 2( k - n )
k=~-'X . B y m u ltip ly in g b o th sides o f this e q u a tio n by th e e x p o n e n tia l ex p (jeon) and su m m in g o v er all n , w e o b ta in

OC X oc
Sxll2(w) =
x i ( k )x 2(k - n )
e ~ JmR
X2M
Figure 4 3 4
Graphical representation of the convolution property.
F in ally , w e in te rc h a n g e th e o rd e r o f th e su m m a tio n s an d m a k e a ch an g e in th e su m m a tio n in d ex . T h u s w e find th a t th e rig h t-h a n d sid e o f th e e q u a tio n above re d u c e s to X \ ( u ) ) X 2 ( a>). T h e fu n ctio n SXlX2(a)) is called th e cross-energy density s pe c t r um o f th e sig n a ls x\ (n) an d x 2 (n).
The W iener-Khintchine theorem.

rxz(l)
L e t jc(n) be a re a l signal. T h e n s(a) (4-3.51)
T h a t is, th e e n e rg y sp e c tra l d en sity o f an en erg y signal is th e F o u rie r tra n sfo rm o f its a u to c o rre la tio n se q u e n c e . T h is is a sp ecial case o f (4.3.50). T h is is a v ery im p o rta n t resu lt. It m ean s th a t th e a u to c o rre la tio n se q u e n c e o f a sig n al a n d its en e rg y sp e c tra l d e n sity c o n tain th e sam e in fo rm a tio n a b o u t th e signal. S ince, n e ith e r o f th e s e c o n ta in s any p h ase in fo rm a tio n , it is im p o ssib le to u n iq u e ly re c o n stru c t th e signal fro m th e a u to c o rre la tio n fu n ctio n o r th e en erg y d e n sity sp e ctru m .
Example 4JJ
D eterm ine the energy density spectrum of the signal x(n) = tfûfn) - 1<a < 1
300 Solution signal is
Frequency Analysts of Signals and Systems
Chap. 4
From Example 2.6.2 we found that the autocorrelation function for this
= --------- oc < / < oo
f a il)
1 a*
By using the result in (4.3.46) for the Fourier transform of a 1 '1 , derived in Exam ple 4.3.3. we have
1 0
= ----- ----- - 1 2a cos w + a1 1 2a cos w + a 2
Thus, according to the W iener-K hintchine theorem , Sxx(w) =
Frequency shifting.
th en
If x( n) < F > X( u ) )
e i< onnx ( n )
X(cu-m )
(4.3.52)
T h is p ro p e rty is easily p ro v e d by d ire c t s u b s titu tio n in to th e analysis e q u a tio n (4.3.1). A cco rd in g to this p ro p e rty , m u ltip lic a tio n o f a se q u e n c e x( n) by e iu> )n is eq u iv a le n t to a fre q u e n c y tra n sla tio n o f th e sp e c tru m X(co) by co< > . T h is freq u e n cy tra n sla tio n is illu stra te d in Fig. 4.35. S ince th e s p e c tru m X(co) is p e rio d ic, th e shift a\) ap p lies to th e sp e c tru m o f th e signal in every p erio d .
The modulation theorem.
If x( n) < F > X (a>)
X(u>)
_2
0
(a)
1 2
X( uj-
cjq)
- 2w
- 2* + wo
(b)
2i
2 + cjo
Figure 4 3 5 form.
Illustration of the frequency-shifting property of the Fourier trans
Sec. 4.3
Properties o f the Fourier Transform for Discrete-Time Signals
301
th e n
x ( n ) cos coqt i
+ cuo) + X(w cô)]
(4.3.53)
T o p ro v e th e m o d u la tio n th e o re m , w e first ex p ress th e signal coswo'J as cos won = + e - J^'")
U p o n m u ltip ly in g x (n ) by th e se tw o e x p o n e n tia ls a n d using th e freq u e n cy -sh iftin g p ro p e rty d e sc rib e d in th e p re c e d in g sectio n , w e o b ta in th e d e sire d re su lt in (4.3.53). A lth o u g h th e p ro p e rty given in (4.3.52) can also be v iew ed as (com plex) m o d u la tio n , in p ractice w e p re fe r to use (4.3.53) b ecau se th e signal j (h ) c o s a ^ ? is real. C learly , in th is case th e sy m m etry p ro p e rtie s (4.3.12) and (4.3.13) are p re se rv e d . T h e m o d u la tio n th e o re m is illu stra te d in Fig. 4.36, w hich c o n ta in s a plot of th e sp e c tra o f th e signals jr(/;). vi(n) = ;c (n )c o s 0 .5 ;rn an d y ifn ) = .* ( )cos7n;.
2
(a)
(b)
2
(c)
Figure 4.36
G raphical representation of the m odulation theorem .
302
Chap. 4
Parsevals theorem.
If jri(n)
an d X2 (n) th en
oc 1 / ' ,7
X 2 (a>)
xi(n)jL*(n) = - /
2jr J - *
X\((i>)X*
(4.3.54)
T o p ro v e this th e o re m , w e use (4.3.1) to elim in a te X\(<o) on th e rig h t-h an d side o f (4.3.54). T h u s we h ave Xî^dto x
=
h i
T ,
X](n)e'
i
f
r
( w ) e ~ Jujnda> =
*' 2n-
^
22
n ~ oc
22
n= -3 c
In th e special case w h ere x 2 (n) = X](/i) = x ( n) , P a rse v a ls re la tio n (4.3.54) red u ces to x y
k ( )| 2 = r
2T T J 2,
/f
\X (a > )\2dw
(4.3.55)
W e o b se rv e th a t th e le ft-h a n d side o f (4.3.55) is sim ply th e e n e rg y E x o f th e signal x(fj). It is also e q u a l to th e a u to c o rre la tio n o f * ( ), rxx(l), e v a lu a te d a t / = 0. T h e in te g ra n d in th e rig h t-h a n d side o f (4.3.55) is e q u a l to th e en erg y density sp e c tru m , so th e in te g ra l o v e r th e in terv al it < a> < it y ields th e to ta l signal en erg y . T h e re fo re , w e co n clu d e th a t E x = r xA ) = y i \x ( n )\2 = 2jt r l cn Ss x (w)da> (4.3.56) J2 n\X(a))\ 2d w = 2it
Multiplication of two sequences (Windowing theorem).

x\ ( n) < Xi(cu) an d x 2 (n) < X 2 (co)
If
then
f i x j i n ) = x \ { n ) x 2 {n) * X j ( w) =
r / X] ( k ) X 2 (a) - k ) d k J_n
(4.3.57)
Sec. 4.3
Properties of the Fourier Transform f o r Discrete-Time Signals
303
T h e in te g ra l on th e rig h t-h a n d side of (4.3.57) re p re se n ts th e co n v o lu tio n of the F o u rie r tra n sfo rm s Xi(cl>) a n d X 2(w). T h is re la tio n is th e d u a l o f th e tim e-d o m ain c o n v o lu tio n . In o th e r w ords, th e m u ltip lic atio n o f tw o tim e -d o m a in se q u en ces is e q u iv a le n t to th e co n v o lu tio n of th e ir F o u rie r tran sfo rm s. O n th e o th e r h a n d , the co n v o lu tio n o f tw o tim e -d o m a in se q u e n c e s is e q u iv a le n t to th e m u ltip lic a tio n o f th e ir F o u rie r tran sfo rm s. T o p ro v e (4.3.57) we b egin w ith th e F o u rie r tra n sfo rm o f (n) = jci(rt)x 2(n) an d u se th e fo rm u la fo r th e in v erse tra n sfo rm , nam ely,
T h u s, w e h av e
CC CC
X3(oj) =
x 3(n)e Jwn =
2 2 x ' ( n )x 2 (n )e
X\(k)
k)d k
T h e co n v o lu tio n in teg ral in (4.3.57) is kn o w n as th e per i od i c convol ut i o n o f Xi(o>) a n d X 2 (to) b e c a u se it is th e co n v o lu tio n o f tw o p e rio d ic fu n ctio n s h av in g th e sam e p e rio d . W e n o te th a t th e lim its o f in te g ra tio n e x te n d o v e r a single p erio d . F u rth e rm o re , w e n o te th a t d u e to th e p erio d icity o f th e F o u rie r tra n sfo rm fo r d isc re te -tim e signals, th e re is no p e rfe c t" d u ality b e tw e e n th e tim e an d freq u e n c y d o m a in s w ith re sp e c t to th e c o n v o lu tio n o p e ra tio n , as in th e case o f c o n tin u o u s tim e signals. In d e e d , co n v o lu tio n in th e tim e d o m ain (a p e rio d ic su m m a tio n ) is e q u iv a le n t to m u ltip lic a tio n o f co n tin u o u s p e rio d ic F o u rie r tra n sfo rm s. H o w ev er, m u ltip lic a tio n o f a p e rio d ic se q u e n c e s is e q u iv a le n t to p e rio d ic co n v o lu tio n o f th e ir F o u rie r tra n sfo rm s. T h e F o u rie r tra n sfo rm p a ir in (4.3.57) will p ro v e u se fu l in o u r tr e a tm e n t of F IR filter d esign b a se d on th e w indow te c h n iq u e .
Differentiation in the frequency domain.

x( n) X(a>)
F
If
then
f
nx(n)
. ii X( ( o) da)
(4.3.58)
304
Chap. 4
T o p ro v e this p ro p e rty , we use th e d efin itio n o f the F o u rie r tra n sfo rm in (4.3.1) an d d iffe re n tia te th e series te rm by te rm w ith resp ec t to w. T h u s we o b ta in d X ( u> ) doj d du>
7 2 n x ( n ) e Ju
N ow w e m u ltip ly b o th sides of th e e q u a tio n bv j to o b ta in th e d esired resu lt in (4.3.58). T h e p ro p e rtie s d eriv ed in this section are su m m a riz e d in T a b le 4.5, which serv es as a co n v e n ie n t refe ren ce . T ab le 4.6 illu stra te s som e useful F o u rie r tran s form p a irs th a t will be e n c o u n te re d in la te r c h a p te rs.
TABLE 4.5
S IG N ALS
PR O P E R TIE S OF TH E FO U R IE R T R A N S F O R M FOR D IS C R E TE -TIM E
Property Notation
Time Domain ,v() A](ii) x2(n) a\X] (n) + x(rt ~ k) x(~n) x \ { n ) * x 2 (n) r llt,(/) = ^, (/) * j z( - / )
Frequency Domain A '(co) X](w) A'iUu) a\X](a)) + 02 X 2 (01) e_ ""l X() X ( a> ) Xi(u>)X;(oj) 5,,.,, iw) = X, ( w)X2(-to) - A j (u l > )X , (iu) [if X2 (h ) is real] -S .t.r (w) X( w o^j) \ X ( w + tut,) + i X ( w - wn) 1 f I X |(/.)X (w X)dk 2jr d X (u) du>
Linearity Time shifting Time reversal Convolution C orrelation
W iener-Khm tchine theorem Frequency shifting M odulation Multiplication D ifferentiation in the frequency domain Conjugation
r (/) eja,0"x(n) x(n) cosa>nr
J-n
nx(n) x*{n)
X ( - w) X\ ( w) X*(io)doj
Sec. 4.4 TABLE 4.6

SIGNALS
Frequency-Domain Characteristics of Linear Tim e-Invariant Systems
305
SO M E U S EFU L F O U R IE R TR A N S FO R M PAIRS FO R D IS C R E TE -TIM E A P ER IO D IC
4.4 FREQUENCY-DOMAIN CHARACTERISTICS OF LINEAR TIME-INVARIANT SYSTEMS

In th is se ctio n w e d e v e lo p th e c h a ra c te riz a tio n o f lin e a r tim e -in v a ria n t system s in th e fre q u e n c y d o m a in . T h e basic e x c ita tio n signals in th is d e v e lo p m e n t a re th e co m p lex e x p o n e n tia ls a n d sin u so id al fu n ctio n s. T h e c h a ra c te ristic s o f th e system a re d e s c rib e d by a fu n ctio n o f th e fre q u e n c y v a ria b le co called th e fre q u e n c y r e sp o n se, w hich is th e F o u rie r tra n sfo rm o f th e im p u lse re sp o n s e h( n) o f th e system . T h e fre q u e n c y re sp o n s e fu n ctio n c o m p le te ly c h a ra c te riz e s a lin e a r tim ein v a ria n t sy stem in th e fre q u e n c y d o m a in . T h is allow s us to d e te rm in e th e
306
Chap. 4
ste a d y -s ta te re sp o n se o f th e system to an y a rb itra ry w eig h ted lin e a r co m b in atio n o f sin u so id s o r co m p lex e x p o n en tials. Since p e rio d ic se q u e n c e s, in p a rtic u la r, lend th em se lv e s to a F o u rie r series d e c o m p o sitio n as a w eig h te d sum o f h a rm o n ically re lated co m p lex e x p o n e n tia ls, it b eco m es a sim ple m a tte r to d e te rm in e th e response o f a lin e a r tim e -in v a ria n t system to this class o f signals. T his m e th o d o lo g y is also a p p lied to a p e rio d ic signals since such signals can b e view ed as a su p e rp o s itio n of infin itesim al size co m p lex ex p o n en tials.
4.4.1 Response to Complex Exponential and Sinusoidal Signals: The Frequency Response Function
In C h a p te r 2, it w as d e m o n s tra te d th a t th e re sp o n se o f any re la x e d lin e a r tim ein v arian t sy stem to an a rb itra ry in p u t signal jr(n), is given by th e c o n v o lu tio n sum fo rm u la
X
yin)-
ft(k)x(n-k)
(4.4.1)
O C
In th is in p u t- o u tp u t re la tio n sh ip , th e sy stem is c h a ra c te riz e d in th e tim e dom ain by its u n it sam p le re sp o n s e {h ( n ). - o c < n < oo}. T o d ev e lo p a fre q u e n c y -d o m a in c h a ra c te riz a tio n o f th e sy stem , let us excite th e sy stem w ith th e co m p lex e x p o n e n tia l jr(n) = A e Jam 00 < n < cc (4.4.2)
w h ere A is th e a m p litu d e and a> is an y a rb itra ry fre q u e n c y c o n fin ed to th e freq u en cy in terv al [ - t t . j t ] . By su b stitu tin g (4.4.2) in to (4.4.1), w e o b ta in th e re sp o n se v(n) =
22
X
h ( k ) [ A e Jb,in- k}] (4.4.3)
k~ ~ X OC = A
2 2 h ( k ) e ~ Juk
W e o b se rv e th a t th e te rm in b ra c k e ts in (4.4.3) is a fu n ctio n o f th e freq u en cy v ariab le a> . In fact, th is te rm is th e F o u rie r tra n sfo rm o f th e u n it sa m p le resp o n se h{k) o f th e system . H e n c e w e d e n o te th is fu n ctio n as
0 0
H(w)=
22
h ( k ) e ~ ja> k
(4.4.4)
C learly , th e fu n ctio n H(co) exists if th e sy stem is B IB O sta b le , th a t is, if O C

y
n =-oc |/i(rc)| <
00
W ith th e d efin itio n in (4.4.4), th e re sp o n se o f th e system to th e com plex e x p o n e n tia l giv en in (4.4.2) is y(n) = AH(oo)eJwn
(4.4.5)
Sec. 4.4
307
W e n o te th a t th e re sp o n se is also in th e fo rm o f a co m p lex e x p o n e n tia l w ith the sam e fre q u e n c y as th e in p u t, b u t a lte re d by th e m u ltip lic ativ e fa c to r H ( m ). A s a re su lt o f th is c h a ra c te ristic b e h a v io r, th e e x p o n e n tia l signal in (4.4.2) is called an ei genf unct i on o f th e system . In o th e r w o rd s, an eig e n fu n c tio n o f a system is an in p u t signal th a t p ro d u c e s an o u tp u t th a t differs fro m th e in p u t by a c o n s ta n t m u ltip lic ativ e facto r. T h e m u ltip lic ativ e fa c to r is called an ei genval ue o f th e system . In th is case, a c o m p lex e x p o n e n tia l signal o f th e fo rm (4.4.2) is an e ig en fu n ctio n of a lin e a r tim e -in v a ria n t system , a n d H(u>) e v a lu a te d at th e fre q u e n c y o f the in p u t signal is th e c o rre sp o n d in g eig en v alu e. Example 4.4.1 D eterm ine the output sequence of the system with impulse response k{n) = (iY u(/i) when the input is the complex exponential sequence xin) = Ac^"' oo < n < oc (4.4.6)
Solution First we evaluate the Fourier transform of the impulse response hin), and then we use (4.4.5) to determ ine v(n). From Example 4.2.3 we recall that
(4.4.7 >
At ui = 7t/2, (4.4.7) yields
and therefore the output is
(4.4.8) oc < n < oc
T h is ex a m p le clearly illu stra te s th a t th e only effe ct o f th e system on th e in p u t signal is to scale th e a m p litu d e by 2 /\/5 a n d shift th e p h ase by - 2 6 .6 . T h u s th e o u tp u t is also a co m p lex e x p o n e n tia l o f fre q u e n c y n / 2 , a m p litu d e 2 A / v /5. an d p h a se - 2 6 .6 C. If w e a lte r th e fre q u e n c y of th e in p u t signal, th e effe ct o f th e system on th e in p u t also c h an g es a n d h en c e th e o u tp u t ch an g es. In p a rtic u la r, if th e in p u t se q u e n c e is a c o m p lex e x p o n e n tia l o f fre q u e n c y tt. th a t is, x(n) = A e ^ n then, at co = jr. 1 2 3 oo < n < oc (4.4.9)
308
Chap. 4
an d th e o u tp u t o f th e sy stem is v(n) = 5 A e jnn 00 < n < 00 (4.4.10)
W e n o te th a t H ( n ) is p u re ly real [i.e., th e p h a se a s so ciated w ith H{u>) is zero at co t t ] . H e n c e , th e in p u t is scaled in a m p litu d e by th e fa c to r H ( n ) = =, b u t the p h ase shift is zero . In g e n e r a l H(co) is a co m p lex -v alu ed fu n ctio n o f th e fre q u e n c y variable w. H e n c e it can b e e x p ressed in p o la r form as H(to) = \H(a>)\eJH{w) w h ere |//(a> )| is th e m ag n itu d e o f H(a>) and
@(w) ^ H (c o)
(4.4.11)
w hich is th e p h a se shift im p a rte d on th e in p u t signal by the system at th e fre quen cy CO. Since H( w) is th e F o u rie r tra n sfo rm o f i/i()}, it follow s th a t H(to) is a peri od ic fu n ctio n w ith p e rio d 2 n . F u rth e rm o re , w e can view (4.4.4) as th e ex ponential F o u rie r series ex p an sio n for H(co ), w ith h( k) as th e F o u rie r se rie s coefficients. C on se q u e n tly , th e u n it im p u lse h( k) is re la te d to H(co) th ro u g h th e in te g ra l expression h( k) =
1
H( t o) elwkdco
(4.4.12)
F o r a lin ear tim e -in v a ria n t system w ith a re a l-v a lu e d im p u lse resp o n se, the m ag n itu d e an d p h a s e fu n ctio n s possess sy m m etry p ro p e rtie s w h ich are d eveloped as follow s. F ro m th e d efin itio n o f H(co). w e have H(co) = h( k)e'
ju / k
52
k D c
h(k) coscok j 5 2
k=oc
h (jt)s in w
(4 4 13)
= H R{io) -f j H 1 (to)
= y j H l i t o ) + H f ( c o ) e J tan- 1["/<")/"*<-)]
w h ere H K(to) an d Hi(co) d e n o te th e real a n d im ag in a ary c o m p o n e n ts o f H(a>), d e fined as

DC
Hn(co) =
52
h ( k ) c os c ok (4.4.14)
H/(co) = 5 2 h( k) sin cok

k = -x
It is c lear from (4.4.12) th a t th e m a g n itu d e an d p h a s e o f H(co), ex p re sse d in term s o f H r (co) an d Hi(co), are \H(co)\ = J H 2r(co) + Hf(co)
(4.4.15)
(w ) = tan --------H R(to)
Sec. 4.4
309
W e n o te th a t H R (a>) = H R( - t o ) a n d Hi {a>) = / / / ( oj). so th a t H r ( c o ) is an ev en fu n c tio n o f oj a n d Hi (co) is an o d d fu n ctio n o f w. A s a c o n s e q u e n c e , it follow s th a t is an ev en fu n ctio n of w an d 0(o>) is an o d d fu n c tio n o f a> . H ence, if w e k n o w j / f ( a > ) J a n d 0 ( a > ) fo r 0 < w < n . w e also k n o w th e s e fu n c tio n s for jr < w < 0. Example 4.4.2 Moving Average Filter
D eterm ine the m agnitude and phase of H(w) for the three-point moving average (MA) system y ( n ) = ^fjc(n + 1) + x( n ) + x( n - 1)] and plot these two functions for 0 < a > < n. Solution Since h{n) = [ i i i } t it follows that H(co) = Hence IW M I
B ( w ) = ^ + 1 + = 1 ( 1 +
2cos<w) (4.4.16)
i 11 + 2 cos a) ' 0. 0
<
a> < <
2?r/3
w <
jt
2 jt / 3
Figure 4.37 illustrates the graphs of the m agnitude and phase of H(w). As indicated previously, |W(w)| is an even function of frequency and C-)(a>) is an odd function of
3tt 4 Figure 4.37 Magnitude and phase
responses for the M A system in

Example 4.4.2.
310
Chap. 4
frequency. It is apparent from the frequency response characteristic H(u>) that this moving average filter smooths the input data, as we would expect from the inputoutput equation. T h e sy m m etry p ro p e rtie s satisfied by th e m a g n itu d e a n d p h a se fu n ctio n s of H(u> ), an d th e fact th a t a sinusoid can be e x p re sse d as a su m o r d ifferen ce of tw o co m p lex -co n ju g ate e x p o n e n tia l fu n ctio n s, im ply th at th e re sp o n s e o f a linear tim e -in v a ria n t sy stem to a sin u so id is sim ilar in fo rm to th e re sp o n se w hen the in p u t is a co m p lex e x p o n e n tia l. In d e e d , if th e in p u t is
Jfi (n) Ae-' w'
th e o u tp u t is vi() = A \ H( o j ) \ e j<rnw' eilM ' O n th e o th e r h a n d , if th e in p u t is

j::( )
th e re sp o n se o f th e system is
y : ( / j) =
A\H{~
A \ H (aj)\e~j H ""' e ~J'l> "
w h ere, in th e last ex p ressio n , w e h av e m a d e use of the sy m m etry p ro p e rtie s \H(to)\ = \ H ( cd)\ an d (~)(a;) = (-)(co). N ow , by ap p ly in g the su p e rp o sitio n p ro p e rty o f th e lin e a r tim e -in v a ria n t system , w e find th at th e re sp o n se o f th e sys tem to th e in p u t i[jt] (n) + _ V 2 (/;)] = A cos con is }(>0 = j[ v , (n) + \ 2(/r)] (4.4.17) y i n ) = A \ H( c o) \ cosjw/? + (-)(o>)] Sim ilarly, if th e in p u t is x i n ) = r[jf |(n ) jt-*(/t) 1 = A sin ton J2 th e resp o n se o f th e system is v W = i [ v l( ,, ) v(h) = A \ H { w ) \ sin[cu + 0 (tn )] I t is a p p a r e n t fro m th is discussion th a t H{co ), o r e q u iv a le n tly . |//(c u )| and 0(cw), co m p le te ly c h a ra c te riz e the effect o f th e system on a sin u so id a l in p u t signal o f any a rb itra ry freq u e n cy . In d e e d , w e n o te th a t \H(a>)\ d e te rm in e s th e am plifi catio n ( | / / M | > 1) o r a tte n u a tio n (\H(co)\ < 1) im p a rte d by th e system o n the in p u t sin u so id . T h e p h a se @(co) d e te rm in e s th e a m o u n t o f p h a s e shift im p a rte d (4 4 1 8 )
Sec. 4.4
311
by th e sy stem o n th e in p u t sinusoid. C o n se q u e n tly , by k n o w in g H{a j), w e are ab le to d e te rm in e th e re sp o n s e o f the system to an y sin u so id a l in p u t signal. Since H ( w ) specifies th e re sp o n s e o f th e sy stem in th e fre q u e n c y d o m a in , it is called th e f r e q u e n c y response o f th e system . C o rre sp o n d in g ly , \H{to) \ is called th e ma g ni t ude respons e a n d 0(o>) is c a lle d th e phas e response o f th e system . I f th e in p u t to th e sy stem co n sists o f m o re th a n o n e sin u so id , th e s u p e rp o sitio n p ro p e rty o f th e lin e a r system can b e u se d to d e te rm in e th e re sp o n se . T h e fo llo w in g e x a m p le s illu stra te th e use o f th e su p e rp o s itio n p ro p e rty . Exam ple 4.43 D eterm ine the response of the system in Example 4.4.1 to the input signal ;c(n) = 10 5 sin n + 2 0 co s nn
oc < n < oc
Solution
The frequency response of the system is given in (4.4.7) as H(w) = ------* 1-
The first term in the input signal is a fixed signal com ponent corresponding to w = 0. Thus H( 0) = The second term in response of the system is =2 At this frequency the frequency
has a frequency jr/2.
Finally, the third term in x ( n ) has a frequency w =

H { tt) = |
tt,
At this frequency
Hence the response of the system to *(n) is 10 . / jt \ 40 y (n ) = 2 0 j= . sm ^ 2 ^ 2 6 .6 J + co sttw

o c < n < oc
Example 4.4.4
A linear tim e-invariant system is described by the following difference equation: y(n) = ay(n 1) + bx(n) 0 < a < 1
(a) D eterm ine the magnitude and phase of the frequency response H(w) of the
system,
(b ) Choose the param eter b so that the maximum value of \H{u>)\ is unity, and
sketch fH(<d)\ and 4 . H( w) for a = 0,9.
312
Chap. 4
(c) D eterm ine the output of the system to the input signal
x (n ) = 5
+ 12 sin
2 0 cos
(^ z n
+ ^
Solution
The impulse response of the system is
h(n) = ba"u(n)
Since |a| < 1. the system is BIBO stable and hence H(co) exists, (a) The frequency response is oc
H (w) = 22, h ^ e ~iam
b
1 ae~JW
Since
1 ae~, = (1 a cos co) 4- j a sin co it follows that [1 ae J '| = ^ /(l a cos co)2 4- (a sin co)2
= y/\ + a2 2a cos co
and 4 (1 - ae J) = tan T herefore, \H(co)\ = \b\ V l 4- a2 2a cos 1u ^H(co) = &(co) = 4-b - tan 1 a sin co 1 a cos ( osinct) 1 a cos co
(b) Since the param eter a is positive, the denom inator of \H(w)\ attains a minimum at w = 0. Therefore, |//(a>)| attains its maximum value at a> = 0. A t this frequency we have 1 *1 |ff(0)| = r U - = ' l 1 a which implies that b = (1 a). We choose b = 1 a, so that l o
|W M I -
J\ + a 2 2a cos , asinoj 1 a cos co
and &(w) = tan"
The frequency response plots for |tf(co)| and 0(o>) are illustrated in Fig. 4.38. W e observe that this system attenuates high frequency signals.
Sec. 4.4
313
Magnitude and phase responses for the system in Example 4.4.4 with a =
Figure 4.38
(c) The input signal consists of components of frequencies to = 0. tt/2, and n. For co = 0. |//(0)| = 1 and 0 (0) = 0. For to = jr/2, 1 a (-)l = _________ V2 / 1 y/\ -f a 2 For co = tt, 1 -a 0.1 \H(n)\ = = = 0.053 l+ o 1.9 0(7r) = 0 Therefore, the output of the system is y(n ) = 5|ff(0)| + 1 2 | / / ( | ) | s i i i | j + e ( ! ) ]
2 0 |/ / ( jt ) |
0.1
______
= 0.074
-s/1.81
-- - tan-1 a = - 4 2
cos
^ 7 rn + + 0 ( 7 r ) j
= 5 + 0.888sin
42^ 1.06 cos ^jrn + ^
oc < n < oc
314
Frequency Analysis ol Signals and Systems
Chap. 4
In th e m o st g e n e ra l case, if th e in p u t to th e system co n sists o f an arbitrary lin ear co m b in a tio n o f sinusoids o f th e form

L
x ( n ) = 5 2 A, cos(co,n 4 4> i)
oc < < oc
1=1
w h ere (A, | an d {<,} a re th e am p litu d e s a n d p h ases o f th e c o rre sp o n d in g sinusoidal c o m p o n e n ts, th e n th e re sp o n s e of th e sy stem is sim ply
L
v{n) = 5 2 Aj \ H{ a ),)| cos[oj, 4- ,)]

1= 1
(4.4.19)
w h ere \H(u>i)\ an d (oj,) are th e m a g n itu d e an d p h a se , re sp e c tiv e ly , im p a rte d by th e system to th e in d iv id u al fre q u e n c y c o m p o n e n ts o f th e in p u t signal. It is clear th a t d e p e n d in g on th e fre q u e n c y re sp o n se H ( uj) o f th e system , input sinu soid s o f d iffe re n t fre q u e n c ie s will be affe c te d d ifferen tly by th e system . F o r ex am p le, so m e sin u so id s m ay be co m p letely su p p re sse d by th e sy stem if H ( w ) = 0 at th e fre q u e n c ie s o f th e s e sinusoids. O th e r sin u so id s m ay receiv e n o a tte n u a tio n (or p e rh a p s, som e am p lifica tio n ) by th e system . In effect, w e can view th e lin ear timein v arian t system fu n c tio n in g as a filter to sin u so id s o f d iffe re n t fre q u e n c ie s, passing som e o f th e fre q u e n c y c o m p o n e n ts to th e o u tp u t a n d su p p re s sin g o r preventing o th e r freq u e n cy c o m p o n e n ts fro m re ach in g th e o u tp u t. In fact, as discussed in C h a p te r 8, th e basic dig ital filter design p ro b le m involves d e te rm in in g th e p a ra m e ters o f a lin e a r tim e -in v a ria n t system to achieve a d e sire d fre q u e n c y re sp o n s e H (co).
4.4.2 Steady-State and Transient Response to Sinusoidal Input Signals

In th e discu ssio n in th e p re ced in g sectio n , w e d e te rm in e d th e re s p o n s e o f a linear tim e -in v a ria n t sy stem to e x p o n e n tia l a n d sin u so id al in p u t sig n als a p p lied to the system at n = oc. W e usually call such signals e te rn a l e x p o n e n tia ls o r etern al sin u so id s, b ecau se th e y w ere a p p lied at n = - o c . In such a case, th e re sp o n se that w e o b se rv e a t th e o u tp u t o f th e system is th e ste a d y -s ta te re s p o n s e . T h e re is no tra n sie n t re sp o n se in th is case. O n th e o th e r h a n d , if th e e x p o n e n tia l o r sin u so id al signal is a p p lie d a t som e finite tim e in sta n t, say at n = 0, th e re sp o n se of th e sy stem co n sists of tw o term s, the tra n sie n t re sp o n se a n d the ste a d y -sta te re sp o n se . T o d e m o n s tra te this b ehavior, let us co n sid er, as an e x am p le, th e sy stem d e sc rib e d by th e firs t-o rd e r d ifferen ce e q u a tio n y( n) = av( n 1 )4- j:(/i) (4.4.20) T his sy stem w as c o n s id e re d in S ectio n 2.4.2. Its re sp o n se to an y in p u t x ( n) applied at n = 0 is given by (2.4.8) as y( n) = a n+ly ( - l ) + J 2 a kx ( n - k) n> 0 (4.4.21)
*=0
w h ere y ( 1) is th e in itial co n d itio n .
Sec. 4.4
315
N o w , let us assu m e th a t th e in p u t to th e system is th e c o m p le x e x p o n e n tia l x ( n ) - A e Jan n> 0 (4.4.22)
a p p lied a t n = 0. W h e n w e s u b s titu te (4.4.22) in to (4.4.21), w e o b ta in n v(n) = a " +1y ( 1) +
= a
v ( 1) + A (4.4.23) n > 0
A
k=0 1 _ gn+le ~juiUi +\) = o "+1v ( - l ) + A ----- -------------------- e jmn 1 -ae-J

A/jn^
= an+' v ( - l ) ---------- : --- eJua + ------ - e ia*

1 a e ~ JW 1 ~ a e JW
n>0
W e recall th a t th e system in (4.4.20) is B IB O sta b le if |a | < 1 . In th is case th e tw o te rm s in v o lv in g ^ " +1 in (4.4.23) d ecay to w a rd z e ro as n a p p ro a c h e s infinity. C o n s e q u e n tly , w e a re left w ith th e ste a d y -s ta te re sp o n se Vss(n) =
=
A i hm v() = ------------ e1 l-ae-J* AH(a))eJwn
(4.4.24)
T h e first tw o te rm s in (4.4.23) c o n stitu te th e tra n s ie n t re s p o n s e o f th e svstem , th a t is, Aan + ' I Vlr(n) = a n+1 v ( - l ) ------- ;----------- : ----- e JU ,n n > 0 (4.4.25) 1 - a e ~ JW w hich d e c a y to w a rd z e ro as n a p p ro a c h e s infinity. T h e first te r m in th e tra n s ie n t re sp o n se is th e z e ro -in p u t re sp o n se o f th e system an d th e se c o n d te rm is th e tra n s ie n t p ro d u c e d by th e e x p o n e n tia l in p u t signal. In g e n e ra l, all lin e a r tim e -in v a ria n t B IB O system s b e h a v e in a sim ilar fashion w h e n e x c ite d by a co m p lex e x p o n e n tia l, o r by a sin u so id a t n = 0 o r at so m e o th e r finite tim e in sta n t. T h a t is, th e tra n s ie n t re sp o n s e d ecay s to w a rd z e ro as n oo, leav in g o n ly th e s te a d y -s ta te re sp o n s e th a t w e d e te rm in e d in th e p re c e d in g section. In m an y p ractical a p p lic a tio n s, th e tra n s ie n t re sp o n se of th e sy ste m is u n im p o rta n t, a n d th e r e fo r e it is u su ally ig n o re d in d e a lin g w ith th e re sp o n s e o f th e system to sin u so id al in p u ts.
4.4.3 Steady-State Response to Periodic Input Signals

S u p p o se th a t th e in p u t to a sta b le lin e a r tim e -in v a ria n t system is a p e rio d ic signal x ( n ) w ith fu n d a m e n ta l p e rio d N. S ince such a signal exists fro m co < n < oc, th e to ta l re sp o n s e o f th e sy stem a t an y tim e in s ta n t n, is sim p ly e q u a l to the s te a d y -s ta te resp o n se.
316
Chap. 4
T o d e te rm in e th e resp o n se v(n) o f th e system , we m a k e use o f th e Fourier se ries re p re s e n ta tio n of th e p e rio d ic signal, w hich is
A- l
jr(n) = 5 2 c ^ j27rkn/N k=0
k = 0, 1........ N - 1
(4.4.26)
w h e re th e {c*} are th e F o u rie r se ries coefficients. N ow th e re sp o n s e of th e system to th e co m p lex e x p o n e n tia l signal Xk (n) is v* ( n ) = t \ H w h ere H / Ink \ ( J = H ( w ) L = 2, i7A k = 0.
1 -
ckej2 7 rk n /N
k^j
k=
0 , 1 ............ N
fc = 0. 1.........A ' - l
(4.4.27)
........ A ' - l
B y u sin g th e su p e rp o s itio n prin cip le fo r lin e a r system s, w e o b ta in th e resp o n se of th e system to th e p erio d ic signal x( n) in (4.4.26) as A 1 v() = 5 2 / ?t k \ gjlxknlf* oc < 7 ; < OC (4.4.28)
T his resu lt im p lies th a t th e re sp o n se of th e system to th e p e rio d ic in p u t signal x ( n ) is also p e rio d ic w ith th e sam e p e rio d N . T h e F o u rie r se rie s coefficients for y( n) a re dk = ckH ( ^ ^ j * = 0 , 1 .........A ' - l (4.4.29)
H e n c e , th e lin e a r system can ch an g e th e sh a p e o f th e p e rio d ic input signal by scaling th e a m p litu d e an d shifting th e p h ase o f th e F o u rie r se rie s c o m p o n en ts, but it d o e s n o t affe ct th e p e rio d of the p e rio d ic in p u t signal.
4.4.4 Response to Aperiodic Input Signals

T h e c o n v o lu tio n th e o re m , given in (4.3.49). p ro v id e s th e d e s ire d freq u e n cy -d o m ain re la tio n sh ip fo r d e te rm in in g the o u tp u t of an L T I system to an ap e rio d ic finiteen erg y signal. If {jt(n)} d e n o te s th e in p u t se q u e n c e , (v(n)} d e n o te s th e o u tp u t se q u e n c e , an d (/?()} d e n o te s th e u n it sa m p le re sp o n se o f th e system , th e n from th e co n v o lu tio n th e o re m , w e have Y(w) = H(a>)X(w) (4.4.30)
w h e re Y(to), X (o j), an d H(co) are th e c o rre sp o n d in g F o u rie r tra n sfo rm s o f {v()J. (*()}, an d {h( n ) ), resp ectiv ely . F ro m th is re la tio n sh ip w e o b se rv e th a t th e sp e c tru m o f th e o u tp u t signal is e q u a l to th e sp e c tru m o f th e in p u t signal m ultiplied by th e fre q u e n c y resp o n se o f th e system .
Sec. 4.4
317
If w e e x p re ss K(cu), an d X(o>) in p o la r fo rm , th e m a g n itu d e a n d p h ase o f th e o u tp u t signal can b e ex p re sse d as |y M ! = I / / M I I X M 1 4-Y (o>) + 4 - H ( co) (4.4.31) (4.4.32)
w h ere |//(a> )| a n d a re th e m a g n itu d e an d p h ase re sp o n s e s o f th e system . B y its v ery n a tu r e , a fin ite-en erg y a p e rio d ic signal c o n ta in s a c o n tin u u m of fre q u e n c y c o m p o n e n ts . T h e lin e a r tim e -in v a ria n t system , th ro u g h its fre q u e n c y re sp o n s e fu n c tio n , a tte n u a te s so m e fre q u e n c y c o m p o n e n ts of th e in p u t signal an d am p lifies o th e r fre q u e n c y co m p o n e n ts. T h u s th e system acts as a filter to th e in p u t signal. O b s e rv a tio n o f th e g ra p h o f \H(to)\ show s w hich fre q u e n c y c o m p o n e n ts a re am p lified a n d w h ich a re a tte n u a te d . O n th e o th e r h a n d , th e angle o f H(a>) d e te rm in e s th e p h a s e sh ift im p a rte d in th e c o n tin u u m o f fre q u e n c y c o m p o n e n ts of th e in p u t signal as a fu n c tio n o f freq u e n cy . If th e in p u t signal s p e c tru m is ch an g ed by th e sy stem in an u n d e s ira b le way, we say th a t th e system h a s cau sed mag ni t ude a n d p h a s e distortion. W e also o b se rv e th a t the o ut put o f a linear ti me-i nvariant sy s t em c annot c o n tain f r e q u e n c y c o m p o n e n t s that are n o t cont ai ned in the input signal. It tak es e ith e r a lin e a r tim e* v aria n t sy stem o r a n o n lin e a r system to c re a te fre q u e n c y c o m p o n e n ts th a t a re n o t n ecessa rily c o n ta in e d in th e in p u t signal. F ig u re 4.39 illu stra te s th e tim e-d o m ain an d fre q u e n c y -d o m a in relatio n sh ip s th a t can b e u se d in th e an aly sis o f B IB O -s ta b le L T I system s. W e o b se rv e th a t in tim e -d o m a in an aly sis, w e d eal w ith th e co n v o lu tio n o f th e in p u t signal w ith th e im p u lse re sp o n s e o f th e system to o b ta in th e o u tp u t s e q u e n c e o f th e system . O n th e o th e r h a n d , in fre q u e n c y -d o m a in analysis, w e deal w ith th e in p u t signal sp e c tru m X(a)) a n d th e fre q u e n c y re sp o n s e H(co) o f th e sy stem , w hich a re re la te d th ro u g h m u ltip lic a tio n , to yield th e sp e c tru m o f th e signal a t th e o u tp u t o f th e system . W e c a n u se th e re la tio n in (4.4.30) to d e te rm in e th e sp e c tru m Y( u > ) o f th e o u tp u t signal. T h e n th e o u tp u t se q u e n c e {v()} can b e d e te rm in e d fro m th e in v erse F o u rie r tra n sfo rm (4.4.33) H o w e v e r, th is m e th o d is se ld o m used. In ste a d , th e z -tra n s fo rm in tro d u c e d in C h a p te r 3 is a s im p le r m e th o d fo r solving th e p ro b le m o f d e te rm in in g th e o u tp u t se q u e n c e {y(n)}.
X(n)
X( w)
Input
Linear time-invariant system hin), H(a>)
Output
v(n) = h(n)+x(n) KM = ma>)X(a>)
F,gre 4 3 9
Tlm e' and
frequency-domain inpul-output relationships in LTI systems.
318
Chap. 4
L e t us r e tu rn to th e b asic in p u t-o u tp u t re la tio n in (4,4.30) a n d c o m p u te the s q u a re d m a g n itu d e o f b o th sides. T h u s w e o b ta in \Y(co)\2 = \H(oj)\ 2 \X(co )\2 (4.4.34) Syy(u}) = \ H{ co)\2S xx {0> ) w h ere Sxx(oo) a n d S vv(cu) a re th e en erg y d en sity sp e c tra o f th e in p u t a n d o u tp u t signals, resp ec tiv ely . By in te g ra tin g (4.4.34) o v e r th e fre q u e n c y ran g e ( - n , n ) , we o b ta in th e e n e rg y o f th e o u tp u t signal as i r (cu)da> (4.4.35) Example 4.4.5 A linear time-invariant system is characterized by its impulse response
h( n ) = (y)" u (n )
r
27r
J-y,
H ( co ) |2 S j x ( co) d co
Determ ine the spectrum and the energy density spectrum of the output signal when the system is excited by the signal *() = ( j )"() Solution The frequency response function of the system O C
H(w) =
1 1 - \e-^ Similarly, the input sequence U(n)) has a Fourier transform X(w) = 1
1 - e ~ JU I
Hence the spectrum of the signal at the output of the system is

Y (co) = H( w) X( c o )
1 (1 1-
The corresponding energy density spectrum is

= \Y(<o)\2 = \H(o>)\2\X(o>)\2
1 ( j - coso))(j| - 1 cosa>)
Sec. 4.4
319
4.4.5 Relationships Between the System Function and the Frequency Response Function
F ro m th e d iscu ssio n in S ectio n 4.2.6 w e k now th a t if th e system fu n ctio n H( z ) co n v erg es on th e u n it circle, w e can o b ta in th e freq u e n c y re sp o n s e o f th e system by e v a lu a tin g H ( z ) o n th e u n it circle. T h u s (4.4.36) In th e case w h e re H ( z ) is a ra tio n a l fu n ctio n of th e fo rm H{ z ) = B ( z ) / A ( z ) . w e have
(4.4.37)
M (4.4.38) ]~~[U - pk? /< ") w h e re th e (fli) a n d {>*} a re real, b u t {ct } and {pk} m ay be c o m p lcx -v alu cd . It is so m e tim e s d e s ira b le to ex p ress th e m a g n itu d e s q u a re d o f H(a>) in term s o f H( z ) . F irst, w e n o te th a t |f f( a O r = H{ w) H*( u >) F o r th e ra tio n a l sy stem fu n ctio n given by (4.4.38). we have M (4.4.39)
ft
It fo llo w s th a t H*( w) is o b ta in e d by e v a lu a tin g H*( 1/ z*) on th e unit circle, w h ere fo r a ra tio n a l sy stem fu n ctio n . M (4.4.40) *=i H o w e v e r, w h e n {*()} is re a l or, e q u iv alen tly , th e coefficients {a*} and {b k} a re real, c o m p lex -v alu ed p o le s a n d z e ro s o ccu r in c o m p le x -c o n ju g a te pairs. In this
320
Chap. 4
case. H*{ 1/;* ) = H ( z ~ l ). C o n se q u e n tly , H'(a>) = | H(co )\2 = H{oj)H*{oj) = H ( w ) H ( - o > ) =
an d (4.4.41)
A cc o rd in g to th e c o rre la tio n th e o re m fo r th e z -tra n sfo rm (see T a b le 3.2), the fu n ctio n H ( z ) H ( z ~ l ) is th e z -tra n sfo rm o f th e a u to c o rre la tio n se q u e n c e {rhh( m )} o f th e u n it sa m p le re sp o n se (/i(n)J. T h e n it follow s fro m th e W ie n e r-K h in tc h in e th e o re m th a t \ H { w )\2 is th e F o u rie r tra n sfo rm o f Sim ilarly, if H ( z ) = B ( z ) / A ( z ) , th e tra n sfo rm s D( z ) = B tz J B tz - 1 ) an d C( z) = /4 (;)A (c ') are th e z-tra n sfo rm s o f th e a u to c o rre la tio n s e q u e n c e s {c/} an d \di\, w h ere
\rh h (m )}.
N - |J i
Q = t=o
M - \t\
akak+i
N < I < N
(4.4.42)
d, = 2 2 bkbk+i ~ M < / <M (4.4.43) Jk=0 Since th e system p a ra m e te rs (at) a n d {*} a re re a l valu ed , it follow s th a t c, c_/ an d di = d-i . B y using th is sy m m etry p ro p e rty , \ H( u >)\2 m ay be ex p ressed as do + 2 2^2 dk cos kco \H(co )\2 = ----------^ -------------(4.4.44) ri co 4- 2 ^ ' c'jt cos k.u) i =1 F inally, w e n o te th a t c o s k w can be e x p re ss e d as a p o ly n o m ia l fu n ctio n of cosoj. T h a t is, t cos kco = ^ / U c o s a , ) " (4.4.45) m=0 w h ere [j3m] a re th e coefficients in th e ex p an sio n . C o n s e q u e n tly , th e n u m e ra to r an d d e n o m in a to r o f \H(u ))\2 can b e v iew ed as p o ly n o m ial fu n c tio n s o f coso;. T he fo llow ing e x am p le illu stra te s th e fo reg o in g re la tio n sh ip s. Example 4.4.6 Determ ine for the system y(n) = 0.1 v(n 1) + 0.2 y(n 2 ) + x (n) + x(n 1) Solution The system function is
14
:~ l
H(Z) ~ 1 4 0 . 1 - - 0 . 2 : - 2 and its R O C is |z| > 0.5. Hence H{ o d ) exists. Now

1 + z
1+ z
1 + 0 .1 Z '1 - 0 .2 z - 2
1 4- O.lz - 0.2z2
2 4- z + z 1 1.05 + 0.08(z 4- z_1) - 0.2(z~2 + z-2)
Sec. 4.4
321
By evaluating H{z)H(z~' ) on the unit circle, we obtain 2 -j- 2 cos w \H(u>)~ = --------------------------------------1.0^ i- 0.16 cos w 0.4 cos 2w However, cos2w = 2cos: w 1. Consequently. | H(u>)\2 may be expressed as 2(1 cos u> ) 1.45 4 0.16cos w - 0 .8 cos: W e n o te th a t given H ( z ), it is stra ig h tfo rw a rd to d e te rm in e H ( z ~ l ) an d th e n |//{ w ) |2. H o w e v e r, th e inverse p ro b lem o f d e te rm in in g H( z ) given 1/ / (cu)]2 or th e c o rre sp o n d in g im p u lse resp o n se {/?())- is n o t stra ig h tfo rw a rd . Since |//(a > )l: d o es n o t c o n tain th e p h ase in fo rm atio n in H(a>), it is n o t p o ssible to u n iq u ely d e te rm in e H (:). T o e la b o ra te on th e p oint, let us assum e th a t th e N p o les an d M ze ro s of H{z) are [pk] an d {;.*). respectively. T h e c o rre sp o n d in g p o les and zero s o f H ( z ~ ]) are (1 j p^ j an d ll/c*. }, respectively. G iv en \ H( u >)\2 o r, eq u iv a le n tly . H ( z ) H ( z ~ l ). w e can d e te rm in e d iffe re n t system fu n ctio n s H( z ) by assigning to H( z ) . a pole pt o r its recip ro c al 1f p k . an d a z e ro zt o r its recip ro c al 1 fzk- F o r ex am p le, if A" = 2 a n d M = 1. th e p o les an d z ero s of H l z ) H ( : _ l ) are [ p j. p z , 1/ p \ , l / p z ) an d (~ i, 1 /c i }. If p i a n d pz are real, th e possible p o les for H(z.) a re {pi. 1. U /P i 1 /P :K i / 'i 1/ / :}. a n d [/52. 1//>]} an d th e possible ze ro s are {~i) o r {1 / - 1 }, T h e re fo re , th e re are eight p o ssib le ch o ices o f sy stem functions, all o f w hich re su lt in th e sam e j//(a> )|: . E ven if we restrict th e p o les of H( z ) to be inside th e unit circle, th e re a re still tw o d iffe re n t ch o ices for H( z ) . d e p e n d in g on w h e th e r w e pick th e zero j;i} o r { l/:.i). T h e re fo re , we c a n n o t d e te rm in e H{z ) u n iq u ely given only th e m a g n itu d e resp o n se
\H(w)\.
4.4.6 Computation of the Frequency Response Function

In e v a lu a tin g th e m a g n itu d e re sp o n se an d th e p h a s e re sp o n se as fu n c tio n s o f f r e q u en cy , it is c o n v e n ie n t to ex p ress H ( oj) in te rm s o f its p o le s an d zeros. H e n c e we w rite H(co) in fa c to re d form as M
n a - z t e - i wk )
H(co) =
-------------------f ] ( 1 - p ke->*) k=1
(4.4.46)
or, e q u iv alen tly , as M H M = b0eJU ^ - M ) ---- --------------]"~[(eJ - p k) (4.4.47)
322
Chap. 4
L e t us ex p re ss th e c o m p lex -v alu ed facto rs in (4.4.47) in p o la r form as ej a - Z k = Vk (co)ejB*M

an d
(4.4.48)
e J0J - p k = Uk (a>)e] * k(m ) w h ere V*(w) s \eJW - z*|, an d Uk(co) = \eJ - p k I, = %.(eJW - p k) &k (co) = 4. (eJW - zk)
(4.4.49)
(4.4.50)
(4.4.51)
T h e m a g n itu d e o f H ( w ) is eq u al to th e p ro d u c t o f m a g n itu d e s of all te rm s in (4.4.47). T h u s, u sin g (4.4.48) th ro u g h (4.4.51), w e o b ta in \H(co)\ =
|fr0l------
Vi() VM(co) ------- V\ (co)U2 ((o) Un(co)
(4.4.52)
since th e m a g n itu d e o f eJM(N~M) is 1. T h e p h ase o f H(to) is th e sum o f th e p h a s e s o f th e n u m e r a to r facto rs, m i n us th e p h ases o f th e d e n o m in a to r facto rs. T h u s, by co m b in in g (4.4.48) th ro u g h (4.4.51), we h av e
2 ^H
(co)
4-bo
co(N -
) +
0 i (co)
0T(dti)
&m(oj) (4.4.53)
[<t>i(a>) + 2(^) + + $>(o>)] T h e p h ase o f th e gain te rm is z e ro o r jt, d e p e n d in g on w h e th e r bo is po sitiv e or n eg ativ e. C learly , if w e k n o w th e z ero s an d th e p o les o f th e sy ste m fu n ctio n H(z), we can ev a lu a te th e fre q u e n c y re sp o n se fro m (4.4.52) an d (4.4.53). T h e re is a g e o m e tric in te r p re ta tio n of th e q u a n titie s a p p e a rin g in (4.4.52) a n d (4.4.53). L e t us c o n sid e r a p o le p k a n d a z e ro z* lo c a te d a t p o in ts A a n d B o f th e z-p lan e, as sh o w n in Fig. 4 .40(a). A ssu m e th a t w e w ish to c o m p u te H{io) a t a specific v alu e o f fre q u e n c y co. T h e given v alu e o f co d e te rm in e s th e an g le of ejw w ith th e p o sitiv e real axis. T h e tip o f th e v e c to r e specifies a p o in t L o n the u n it circle. T h e e v a lu a tio n o f th e F o u rie r tra n sfo rm fo r th e g iv en v alue of co is e q u iv a le n t to e v a lu a tin g th e z -tra n sfo rm at th e p o in t L o f th e co m p lex p la n e . L et us d raw th e v ecto rs A L a n d B L from th e p o le an d z e ro lo c a tio n s to th e p o in t L, at w hich w e w ish to c o m p u te th e F o u rie r tra n sfo rm . F ro m Fig. 4 .4 0 (a) it follow s th at CL = CA + A L an d CL = CB + BL H o w ev er, C L = ej< u, C A = p k a n d C B = zk. T h u s
AL = ej* - p k
an d
(4.4.54)
BL = e JW - z k
(4.4.55)
Sec. 4.4
323
ImU)
(a) im(r)
Figure 4.40 Geom etric interpretation

of the contribution of a pole and a zero to the Fourier transform (1) magnitude: the factor V * / V k, (2) phase; the factor
* - < t> * B y c o m b in in g th e s e re la tio n s w ith (4.4.48) a n d (4.4.49), w e o b ta in
A L = eJ) is th e len g th o f AL, th a t is, th e d istan c e o f th e p o le p k fro m th e p o in t L c o rre sp o n d in g to e }w, w h e re a s Vk(cu) is th e d ista n c e o f th e z e ro z k fro m th e sam e p o in t L. T h e p h a se s <!>*(& >) a n d Q*(a>) a re th e an g les o f th e v e c to rs A L an d BL
324 Im(c)
Chap. 4
Pt = e>^
Zt =
mz)
Figure 4.41 A zero on the unit circle causes |f/((u)| - 0 and w = 4-Zk- In contrast, a pole on the unit circle results in \H{w)\ = 0 0 at w = 4-Pt-
w ith th e p o sitiv e re a l axis, resp ec tiv ely . T h e s e g e o m e tric in te r p re ta tio n s a re show n in Fig. 4.40(b). G e o m e tric in te rp re ta tio n s are very useful in u n d e rs ta n d in g how th e location o f p o le s an d ze ro s affects th e m ag n itu d e an d p h a s e o f th e F o u rie r tran sfo rm . S u p p o se th a t a zero , say z*. an d a pole, say p k, a re o n th e u n it circle as sh o w n in Fig. 4.41. W e n o te th a t at w = 4 z* , Vk (co) an d c o n s e q u e n tly | H(co)\ b eco m e zero. S im ilarly, at to = 4 -pk th e len g th Uk(w) b eco m es z e ro a n d h e n c e \H(co)\ b ecom es infinite. C learly , th e e v a lu a tio n o f p h a s e in th e s e case s h a s n o m ean in g . F ro m th is discu ssion we can easily see th a t th e p re se n c e o f a ze ro close to th e u n it circle cau ses th e m a g n itu d e o f th e fre q u e n c y re sp o n s e , a t freq u e n cies th a t c o rre sp o n d to p o in ts o f th e u n it circle close to th a t p o in t, to b e sm all. In c o n tra st, th e p re se n c e o f a p o le close to th e u n it circle c au ses th e m a g n itu d e of th e fre q u e n c y re sp o n s e to b e larg e a t fre q u e n c ie s close to th a t p o in t. T h u s poles h av e th e o p p o site effe ct o f zero s. A lso , p lacin g a z e ro close to a p o le cancels th e effe ct o f th e p o le, an d vice v ersa. T h is can b e also se en fro m (4.4.47), since if Zk = Pk, th e te rm s eJW - z* an d ejb> pk cancel. O b v io u sly , th e p re se n c e of b o th p o les an d z e ro s in a tra n sfo rm resu lts in a g r e a te r v a rie ty o f sh a p e s for \H(to)\ a n d ^ H( c o ) . T his o b se rv a tio n is very im p o rta n t in th e desig n o f digital filters. W e co n clu d e o u r discussion w ith th e follow ing ex a m p le illu stra tin g th ese co n cep ts. Example 4.4.7 Evaluate the frequency response of the system described by the system function
Sec. 4.4
325
S o lu tio n
Clearly. H(z) has a zero at ; = 0 and a pole at p = 0.8. frequency response of the system is H (w) = el- - 0.8 1 \e-lw - 0.8| Vl-64 1.6cose sin oj cos oj - 0.8
Hence the
The m agnitude response is | H (to) j = and the phase response is B(co) = ui tan
The magnitude and phase responses are illustrated in Fig. 4.42. Note that the peak of the m agnitude response occurs at a s = 0. the point on the unit circle closest to the pole located at 0.8. If th e m a g n itu d e re sp o n se in (4.4.52) is ex p ressed in d ecib els, M \ \H(co)\lllt = 2 0 lo g ,(l |/j()| + 2 0 5 2 lg.jci v/a M - 2 o 5 2 logui A =1 i-l (4.4.58)
T h u s th e m a g n itu d e resp o n se is ex p ressed as a sum o f th e m ag n itu d e facto rs in | H i w )!.
4.4.7 Input-Output Correlation Functions and Spectra

In S ectio n 2.6.5 w e d e v e lo p e d several c o rre la tio n re la tio n sh ip s b etw een the input an d th e o u tp u t se q u e n c e s o f an L TI system . S pecifically, w e d eriv ed th e follow ing e q u a tio n s: r vv(m) = rhh(m) * r xx(m) ryx( m ) = h( m) * rxx(m) (4.4.59) (4.4.60)
w h ere rxx(m) is th e a u to c o rre la tio n se q u e n c e of th e in p u t signal {-*(/?)), ryv(m) is th e a u to c o rre la tio n se q u e n c e o f th e o u tp u t {.y(n)K rhh(m ) is th e a u to c o rre la tio n se q u e n c e o f th e im p u lse re sp o n se {Ai(/t)}. a n d ryx(m) is th e c ro ss c o rre la tio n se q u e n c e b e tw e e n th e o u tp u t an d th e in p u t signals. Since (4,4.59) an d (4.4.60) involve th e co n v o lu tio n o p e ra tio n , th e z-tra n sfo rm o f th e s e e q u a tio n s yields Svvf;) = 5 m ( : ) 5 ( c ) (4.4.61) = H ( z ) H ( z ~ 1 )Sxx(z) SV Jr(z) = H ( z ) S ( z ) If we su b s titu te z = e JW in (4,4.62), we o b tain
S y jc M =
(4.4.62)
H(o s )Sxx(o j)
(4.4.63)
= H(o>)\X(o>)\2
326
Chap. 4
- T
_ t 2
I 2
Figure 4.42 Magnitude and phase of system with H(z) = 1/(1 - 0 .8 ; - 1).
w h ere Syx(u>) is th e cro ss-en e rg y d en sity sp e c tru m o f [y(n)} a n d (jc(/7)}. Sim ilarly, ev alu atin g Svv(z) o n th e u n it circle yields th e en erg y d e n sity s p e c tru m o f th e o u tp u t signal as S v v M = | / / ( w ) |2S M w h ere Sxx(a>) is th e en erg y d en sity sp e c tru m o f th e in p u t signal. S ince ryy(m) an d Syy(co) a re a F o u rie r tra n sfo rm p air, it fo llo w s th at ryy(m) = ~
j
(4.4.64)
Syy(co)ejumdco
(4.4.65)
Sec. 4.4
327
T h e to ta l e n erg y in th e o u tp u t signal is sim ply Ey --- f Syy{CO)dlO = T,.v(0 > (4.4.66)
~jT i r = I \ H(to)\' S xx(cv)dcv J-*
T h e re su lt in (4.4.66) m ay be used to easily p rove th a t v > 0. F in ally , we n o te th a t if the in p u t signal has a flat sp e c tru m [i.e.. 5 v., (w) = , = c o n s ta n t fo r n < co < t t ] , (4.4.63) red u ces to 5 , v(oj) = H { w ) E x w h ere , is th e c o n s ta n t v alu e o f the sp e ctru m . H e n c e H ( w ) = 5vv(w) (4.4.67)
E \
(4.4.68)
o r . e q u iv a le n tly .
h(n) r yxim)
(4.4.69)
T h e re la tio n in (4,4.69) im plies th at Inn) can be d e te rm in e d by exciting th e input to th e system by a sp e ctrally flat signal ja ()), and c ro ssc o rre la tin g th e in p u t w ith th e o u tp u t o f th e system . T his m e th o d is usef ul in m e a su rin g th e im pulse resp o n se o f an u n k n o w n svstem .
4.4.8 Correlation Functions and Power Spectra for Random Input Signals
T his d e v e lo p m e n t p a ra lle ls th e d e riv a tio n s in S ection 4.4.7, w ith th e e x cep tio n th a t we no w d e a l w ith sta tistical m o m e n ts o f th e in p u t an d o u tp u t signals o f an L TI system . T h e v ario u s sta tistical p a ra m e te rs are in tro d u c e d in A p p e n d ix A . L et us c o n s id e r a d isc re te -tim e lin e a r tim e -in v a ria n t sy stem w ith u n it sam ple re sp o n se (/z(n)} an d fre q u e n c y re sp o n se H ( f ) . F o r this d e v e lo p m e n t w e assum e th a t {/?()} is real. L et _*() be a sam ple fu n ctio n of a sta tio n a ry ra n d o m process X ( n ) th a t ex cites th e system an d let y(/;) d e n o te th e re sp o n se o f th e system to xi n) . F ro m th e c o n v o lu tio n su m m atio n th a t re la te s th e o u tp u t to th e in p u t w e have
OC
v(n) =
5 2 h ( k ) x ( n k)
Jt = - oc
(4.4.70)
S ince jc(n) is a ra n d o m in p u t signal, th e o u tp u t is also a ra n d o m se q u en ce. In o th e r w o rd s, fo r each sa m p le se q u e n c e x ( n ) o f th e p ro cess X ( n ), th e r e is a c o rre sp o n d in g sa m p le se q u e n c e v(n) o f th e o u tp u t ra n d o m p ro c e ss Y(n) . W e w ish to relate th e sta tistical c h a ra c te ristic s o f th e o u tp u t ra n d o m p ro cess Y( n) to th e statistical c h a ra c te riz a tio n o f th e in p u t p ro c e ss a n d th e c h aracteristics o f th e system .
328
Chap. 4
T h e ex p e c te d v alu e o f th e o u tp u t y(n) is
OC
m v s= [y (n )] = E[
oc
22
k ~ rx.
h(k)x{rt fc)]
A=ôc
cc
22
A (* ) 0 (n - * ) ]
(4.4.71)
my = mx F ro m th e F o u rie r tra n sfo rm re la tio n sh ip

CC
22
k oc
h( k)
H{u>)= we have
22
k = - oc
* (*)*"-'*
OC
(4.4.72)
H ( 0) =
22 h (*)
A DC
(4.4.73)
w hich is th e dc gain o f th e system . T h e re la tio n sh ip in (4.4.73) allow s us to express th e m e a n v alu e in (4.4.71) as m y = m xH{ 0) (4.4.74) T h e a u to c o rre la tio n se q u en ce fo r th e o u tp u t ra n d o m p ro c e ss is yyy(m) [ y * ( n ) y ( + m ) ] -- E
2 2 h( k)x*(n ~ k) Y 2 _k O Q
OC OC
+ m - j) (4.4.75)
22
OC
V .
cc
h ( k ) h ( j ) E [ x * { n - k)x(rt -f- m - _/)]
= C C _/ = OC
22 Y2 h(k'> hU)Yxx(k - j + m)
k= co cc
T h is is th e g en eral fo rm fo r th e a u to c o rre la tio n o f th e o u tp u t in term s of the a u to c o rre la tio n o f th e in p u t a n d th e im p u lse re sp o n s e o f th e system . A sp ecial fo rm o f (4.4.75) is o b ta in e d w h en th e in p u t ra n d o m p ro c e ss is w hite, th a t is, w h en m , = 0 an d Yxx(m) = o jS(m) (4.4.76) w h ere o x = ^ j(O ) is th e in p u t signal p o w er. T h e n (4.4.75) re d u c e s to
OC
Yyy(m)
= <72
22
t= -o c
h{k) h{k + m)
(4.4.77)
U n d e r th is co n d itio n th e o u tp u t p ro cess h a s th e a v e ra g e p o w e r MO) = ^ h ^ n) = n=-e (4-4.78) J - 1/2
w h ere w e h av e a p p lie d P a rse v a ls th e o re m .
Sec. 4.4
Frequency-Domain Characteristics of Linear Tim e-invariant Systems
329
T h e re la tio n sh ip in (4.4.75) can b e tra n sfo rm e d in to th e fre q u e n c y d o m ain by d e te rm in in g th e p o w e r d e n sity sp e c tru m of yvv(m ). W e h ave r vvM =
22
E E
/vv
E E
h ( k ) h ( l ) Y, A k - I + m) (4.4.79)
k ^ o o i = dc
h{k)h(l )
Yxx(k I + m) e
OC
k = CC I = DC
= r r, ( / )
2 2 h{k) eJwk
^k= - c c
h { l ) e - ]wl
= tw (tu )|-r^ .v(w) T h is is th e d e sire d re la tio n sh ip fo r th e p o w e r d en sity sp e c tru m o f th e o u tp u t p r o cess, in te rm s o f th e p o w e r d en sity sp e c tru m o f th e in p u t p ro cess an d th e freq u e n cy re sp o n s e o f th e system . T h e e q u iv a le n t e x p ressio n fo r co n tin u o u s-tim e system s w ith ra n d o m in p u ts is r vv( F ) = \ H{ F ) \ 2 r xA F ) (4.4.80)
w h ere th e p o w e r d e n sity s p e c tra T w ^ ) and T ( F ) are th e F o u rie r tra n sfo rm s o f th e a u to c o rre la tio n fu n c tio n s y v v ( T ) a n d yJ t ( r ) , resp ec tiv ely , a n d w h e re H { F ) is th e fre q u e n c y re sp o n s e of th e system , w hich is re la te d to th e im p u lse re sp o n se by th e F o u rie r tra n sfo rm , th a t is. H(F)
- J-C f
h( t ) e
- j 2n Fi
dt
(4.4.81)
A s a final ex ercise , w e d e te rm in e th e c ro ssc o rre la tio n o f th e o u tp u t v(n) w ith th e in p u t signal jcOi). If w e m u ltip ly b o th sides o f (4.4.70) by x*(n m) a n d ta k e th e e x p e c te d v a lu e , w e o b ta in [ v (n )-**( wz)] = E y
= OC
h( k)x*{n m ) x ( n k)
y VJ(m) =
2 2 h( k) E[ x * ( n m ) x ( n k)]
= CC OC
(4.4.82)
22
h ( k ) Y x x ( m - k)
Since (4.4.82) h as th e fo rm o f a c o n v o lu tio n , th e fre q u e n c y -d o m a in e q u iv a le n t e x p re ssio n is
r(w ) = ff(<u)r(<u)
In the special case where x ( n ) is white noise, (4.4.83) reduces to
r y, M = ax 2 H{a>)
(4.4.83)
(4.4.84)
330
Chap. 4
w h ere a 2 is th e in p u t n o ise p o w er. T h is re su lt m e a n s th a t an u n k n o w n sy stem w ith fre q u e n c y resp o n se H ( w ) can be id en tified by exciting th e in p u t w ith w h ite noise, c ro ssco rrelatin g th e in p u t se q u e n c e w ith th e o u tp u t se q u e n c e to o b ta in y V jr(m ), and finally, c o m p u tin g th e F o u rie r tra n sfo rm o f yyx(m). T h e re su lt o f th e s e c o m p u ta tio n s is p ro p o rtio n a l to H{co).
4.5 LINEAR TIME-INVARIANT SYSTEMS AS FREQUENCY-SELECTIVE FILTERS

T h e te rm filter is co m m o n ly used to d escrib e a d ev ice th a t d isc rim in a te s, a c co rd ing to so m e a ttr ib u te o f th e o b jects a p p lied at its in p u t, w h at passes th ro u g h it. F o r ex am p le, an a ir filter allow s a ir to p ass th ro u g h it b u t p re v e n ts d u st p a r ticles th a t are p r e s e n t in th e a ir from passing th ro u g h . A n oil filter p erfo rm s a sim ilar fu n ctio n , w ith th e e x cep tio n th a t oil is th e su b sta n c e allo w ed to pass th ro u g h th e filter, w h ile p a rtic le s o f dirt are c o llected a t th e in p u t to th e filter a n d p re v e n te d fro m p assin g th ro u g h . In p h o to g ra p h y , an u ltra v io le t filter is of ten u sed to p re v e n t u ltra v io le t light, w hich is p re s e n t in su n lig h t a n d w hich is not a p a rt o f visible light, from passin g th ro u g h an d affe ctin g th e chem icals on the film. A s w e h av e o b se rv e d in th e p re c e d in g se ctio n , a lin e a r tim e -in v a ria n t system also p e rfo rm s a ty p e o f d isc rim in atio n o r filtering am o n g th e v ario u s freq u e n cy c o m p o n e n ts a t its in p u t. T h e n a tu re o f th is filtering actio n is d e te rm in e d by the fre q u e n c y re sp o n se c h a ra c te ristic s H{co), w hich in tu r n d e p e n d s o n th e ch o ice of th e sy stem p a ra m e te rs (e.g., th e coefficients (at ) a n d [bk ] in th e d iffe re n c e e q u a tio n ch a ra c te riz a tio n o f th e sy stem ). T h u s, by p r o p e r selectio n o f th e coefficients, we can d esig n freq u e n cy -selectiv e filters th a t p ass signals w ith fre q u e n c y co m p o n e n ts in so m e b an d s w h ile th e y a tte n u a te signals co n ta in in g fre q u e n c y c o m p o n e n ts in o th e r fre q u e n c y b an d s. In g e n eral, a lin e a r tim e -in v a ria n t system m odifies th e in p u t signal sp e c tru m X((o) a cco rd in g to its fre q u e n c y re sp o n s e H(co) to yield an o u tp u t signal w ith sp e c tru m Y(to) = H(o>)X(cl>). In a sense, H(co) acts as a wei ghti ng f u n c tion o r a spectral s h a pi ng f un c t i o n to th e d iffe re n t fre q u e n c y c o m p o n e n ts in th e in p u t signal. W h e n v iew ed in th is c o n te x t, any lin e a r tim e -in v a ria n t system can b e c o n sid e re d to b e a fre q u e n c y -sh a p in g filter, ev en th o u g h it m ay n o t n ecessa r ily c o m p letely b lo ck a n y o r all fre q u e n c y c o m p o n e n ts. C o n s e q u e n tly , th e term s lin e a r tim e -in v a ria n t sy stem a n d filte r a re sy n o n y m o u s a n d a re o fte n used in terch an g eab ly . W e use th e te r m filter to d escrib e a lin e a r tim e -in v a ria n t system u se d to p e rfo rm sp e c tra l sh a p in g o r freq u e n cy -selectiv e filtering. F ilte rin g is u se d in dig ital sig n al p ro cessin g in a v a rie ty o f w ays. F o r e x am p le, re m o v a l o f u n d e sira b le n o ise fro m d e s ire d signals, sp e c tra l sh a p in g such as e q u a liz a tio n o f c o m m u n icatio n c h a n n e ls, signal d e te c tio n in r a d a r, so n a r, a n d c o m m u n ic a tio n s, a n d fo r p e rfo rm in g sp e c tra l an aly sis o f signals, a n d so on.
Sec. 4.5
Linear Tim e-Invariant Systems as Frequency-Selective Filters
331
4.5.1 Ideal Filter Characteristics

F ilters a re usu ally classified acco rd in g to th e ir fre q u e n c y -d o m a in ch aracteristics as low pass, h ig h p ass. b a n d p a ss, and b a n d s to p o r b a n d -e lim in a tio n filters. T he id eal m a g n itu d e re sp o n s e ch a ra c te ristic s o f th ese ty p e s o f filters are illu strated in Fig. 4.43. A s sh o w n , th ese ideal filters h ave a c o n sta n t-g a in (usually ta k e n as u n itv -g a in ) p a ssb a n d c h a ra c te ristic an d z e ro gain in th e ir sto p b a n d .
Lowpass
IHu)\
I
Hiphpa*
\HUo)I
Bandpass
-n
a>0
ai]
cu, iii(|Wi
t
Bandstop
~o>
All-pass
Figure 4.43 Magnitude responses for some ideal frequency-selective discrete-time filters.
332
Chap. 4
A n o th e r ch a ra c te ristic o f an id eal filter is a lin e a r p h ase re sp o n s e . T o dem on stra te this p o in t, let us assum e th a t a signal se q u e n c e (*(/?)} w ith fre q u e n c y com p o n e n ts co n fin ed to th e fre q u e n c y ran g e coj < w < un is p a s se d th ro u g h a filter w ith freq u e n cy resp o n se H(co) = J [ w h ere C an d JWnv'
(J,
< co<c^ o th erw ise
(4_5
a re c o n stan ts. T h e signal at th e o u tp u t o f th e filte r h as a spectrum Y( co) = X(co)H(co) - CX(co)e }tL "" u> \ < co < co2 (4.5.2)
By app ly in g th e scaling a n d tim e-sh iftin g p ro p e rtie s o f the F o u rie r tran sfo rm , we o b ta in th e tim e -d o m a in o u tp u t v(ri) = Cx ( n - h 0) (4.5.3)
C o n se q u e n tly , th e filter o u tp u t is sim ply a d e la y e d an d a m p litu d e -s c a le d v ersio n of th e in p u t signal. A p u re d elay is usually to le ra b le a n d is n o t c o n s id e re d a distortion o f th e signal. N e ith e r is a m p litu d e scaling. T h e re fo re , ideal filters have a linear p h ase c h a ra c te ristic w ithin th e ir p a ssb a n d , th a t is. 0(a>) = cori(i (4.5.4)
T h e d eriv ativ e o f th e p h a se w ith re sp e c t to fre q u e n c y has th e u n its o f delay. H e n c e we can defin e th e signal d elay as a fu n ctio n o f fre q u e n c y as
d&(a>)
( o j) = --------- ----------
aco
. (4.5.5)
Zg(co) is usually called th e e nvel ope del ay o r th e g r o u p del ay o f th e filter. W e in te r p re t zg(co) as th e tim e d elay th a t a signal c o m p o n e n t o f fre q u e n c y co u n d erg o es as it p asses fro m th e in p u t to th e o u tp u t o f the system . N o te th a t w h en 0(co) is lin e a r as in (4.5.4), r s (a)) no = c o n s ta n t. In th is case all fre q u e n c y co m p o n en ts o f th e in p u t signal u n d e rg o th e sam e tim e delay. In co n clu sio n , id eal filters h av e a c o n s ta n t m a g n itu d e c h a ra c te ristic an d a lin ear p h ase c h a ra c te ristic w ith in th e ir p assb an d . In all cases, such filters a re not p h ysically re a liz a b le b u t se rv e as a m a th e m a tic a l id e a liz a tio n o f p ra c tic a l filters. F o r ex am p le, th e id eal low pass filter h a s an im p u lse re sp o n se
sin coc7Tn hip(n) -----------nn
00 < /7 < 00
(4.5.6)
W e n o te th a t th is filter is n o t causal an d it is n o t a b so lu te ly su m m a b le an d th e re fo re it is also u n sta b le . C o n se q u e n tly , this id eal filter is physically u n re a liz a b le . N ev e rth e le ss, its fre q u e n c y re sp o n se c h a racteristics can be a p p ro x im a te d v ery closely by p ractical, ph y sically re a liz a b le filters, as will be d e m o n s tra te d in C h a p te r 8. In th e fo llo w in g d iscu ssio n , w e tr e a t th e design o f so m e sim p le d igital filters by th e p la c e m e n t o f p o les a n d zero s in th e z-p lan e. W e h a v e a lre a d y d escrib ed h o w th e lo catio n o f p o les a n d zero s affects th e fre q u e n c y re s p o n s e characteristics
Sec. 4.5
333
o f th e system . In p a rtic u la r, in S ection 4.4.6 w e p re s e n te d a g ra p h ic a l m e th o d for c o m p u tin g th e fre q u e n c y re sp o n se c h a racteristics fro m th e p o le - z e r o p lo t. T his sam e a p p ro a c h can be u sed to design a n u m b e r o f sim ple b u t im p o rta n t digital filters w ith d e s ira b le freq u e n c y re sp o n se ch aracteristics. T h e basic p rin cip le u n d e rly in g th e p o le - z e r o p la c e m e n t m e th o d is to lo cate p o les n e a r p o in ts o f th e u n it circle c o rre sp o n d in g to fre q u e n c ie s to be em p h asized , a n d to p lace ze ro s n e a r th e fre q u e n c ie s to be d e e m p h a siz e d . F u rth e rm o re , the fo llo w in g c o n s tra in ts m ust be im posed: 1. A ll p o le s sh o u ld be p laced inside th e unit circle in o r d e r fo r the filter to be stab le. H o w e v e r, ze ro s can be p laced an y w h ere in th e z-p lan e. 2. A ll co m p lex zero s an d p o les m ust o ccu r in c o m p le x -c o n ju g a te p airs in o rd e r fo r th e filter coefficients to be real. F ro m o u r p re v io u s discussion w e recall th a t fo r a given p o le - z e r o p a tte rn , th e sy stem fu n ctio n H i : ) can be ex p ressed as
(4.5.7)
w h ere bi} is a jiain c o n sta n t selected to n o rm aliz e th e freq u e n cy re sp o n se at som e sp ecified freq u e n cy . T h a t is, bo is se le c te d such th at | H (c l > o) | = 1 (4.5.8)
w h ere is a fre q u e n c y in th e p a ssb a n d o f the filter. U sually, N is se le c te d to eq u al o r ex cee d M , so th a t th e filter h as m o re n o n triv ial p o le s th a n zeros. In th e n ex t se ctio n , we illu strate th e m e th o d o f p o le - z e r o p la c e m e n t in the d esig n o f so m e sim p le low pass. highpass. an d b an d p a ss filters, d igital re so n a to rs, a n d co m b filters. T h e d esign p ro c e d u re is facilita ted w h en c a rrie d o u t in te ra c tiv e ly on a d ig ital c o m p u te r w ith a g raphics term in a l.
4.5.2 Lowpass, Highpass, and Bandpass Filters

In th e d esig n o f lo w pass digital filters, th e p oles sh o u ld be p la c e d n e a r th e unit circle at p o in ts c o rre sp o n d in g to low fre q u e n c ie s (n e a r cm = 0) and ze ro s sh ould b e p la c e d n e a r o r on th e u n it circle at p o in ts c o rre sp o n d in g to high freq u e n cies ( n e a r co = t t ) . T h e o p p o site h o ld s tru e fo r h ighpass filters. F ig u re 4.44 illu strates th e p o le - z e r o p la c e m e n t o f th re e low pass a n d th re e h ig h p ass filters. T h e m a g n itu d e an d p h a se re sp o n se s fo r th e sin g le-p o /e filter w ith sy stem fu n ctio n
334
Chap. 4
Highpass
Figure 4.44
Pole-zero patterns for several lowpass and highpass fillers.
are illu stra te d in Fig. 4.45 fo r a = 0.9. T h e gain G w as se le c te d a s 1 a, so th at th e filter h as u n ity gain a t co = 0. T h e gain of this filter a t h ig h fre q u e n c ie s is relativ ely sm all. T h e a d d itio n o f a ze ro a t z = 1 f u rth e r a tte n u a te s th e re s p o n s e o f th e filter a t high freq u e n cies. T h is lead s to a filter w ith a sy stem fu n ctio n
H 2 (z)
= 1- ~ ~ ~ ~ l~~' . 2 1 az~
(4.5.10)
an d a freq u e n cy re sp o n se c h a ra c te rstic th a t is also illu stra te d in Fig. 4.45. In this case th e m ag n itu d e o f H 2 (co) g o es to z e ro a t co = n. S im ilarly, w e can o b ta in sim p le h ig h p ass filters by reflec tin g (fo ld in g ) the p o le -z e ro lo catio n s o f th e low pass filters a b o u t th e im ag in ary axis in th e z-plane. T h u s w e o b tain th e sy stem fu n ctio n H i(z)= l ~ a \ 2 1 -f a z ~ [ (4.5.11)
w hich has th e freq u e n cy re sp o n s e c h a ra c te ristic s illu stra te d in Fig. 4.46 fo r a = 0.9.
Example 4 . 5 . 1
A two-pole lowpass filter has the system function b0
(1 pz
W(z) =
-')2
Sec. 4.5
Linear Tim e-invariant Systems as Frequency-Selective Filters
335
Figure 4.45 Magnitude and phase response of (1) a singie-pole filter and (2) a one-pole, one-zero filter; Wi(;) = (1 - a ) /( 1 a ; ' 1), H j(z) = [(1 a )/2 ][(l + r ~ * )/(l - a ; - 1 )] and a = 0.9.
D eterm ine the values of h{ >and p such that the frequency response H(w) satisfies the conditions H(0) = 1 and I / K \ I2 1
r W I =2
(1 - p )2
Solution
A t o) = 0 we have H( 0) = = 1
Hence
bo = (1 - p Y
336
Chap. 4
2 0 log|0]HM|
Figure 4.46 Magnitude and phase response of a simple highpass fitter; H(C) = [(1 - a )/2 ][(l - z - ' ) / ( l + a z - 1)]
with a = 0.9. At w = tt/4. (1 ) = (1 ~ P? v4 / (1 pe~i*/4)2 (1 ~ P )2 (1 - p cos(;r/4) + j p sin (tt/4))2

(1 - P ) 2
(1 - pj s / 2 + jp/ ^/ 2)2 Hence (1 - p t 1 2
[(1 - p / - j2 ) 2 + p 212]2
Sec. 4.5
Unear Tim e-Invariant Systems as Frequency-Selective Filters
337
or, equivalently. V ^d - p y = 1 + p1 - y'lp The value of p = 0.32 satisfies this equation. Consequently, the system function for the desired filter is 0.46 H( z) (1 - 0.32c 1)= T h e sam e p rin cip le s can be a p p lied fo r th e desig n of b a n d p a s s filters. B asi cally, th e b a n d p a s s filter sh o u ld c o n tain o n e o r m o re p airs o f co m p lex -co n ju g ate p o les n e a r th e u n it circle, in the vicinity of th e freq u e n cy b a n d th a t c o n stitu te s the p assb a n d o f th e filter. T h e follow ing ex am p le serves to illu strate th e basic ideas.
E x a m p le 4.5.2
Design a two-pole bandpass filler that has the center of its passband at w = n72. zero in its frequency response characteristic at w = 0 and at = n . and its magnitude response is 1 /V 2 at w = 4w /9 .
S o lu tio n
Clearly, the filter must have poles at Pl : = r i b
and zeros at
1 and : = 1. Consequently, the system function is /z o = a = c(: - 1)<: + 1) (: - j r)(: + j r) - 1
The gain factor is determ ined by evaluating the frequency response H(w) of the filler at u > = jr/2. Thus we have
" ( ) - cr b - >
G
1 - r2
The value of r is determ ined by evaluating H( w ) at w = 4tt/9. Thus we have 4tt \

9 /|
(1 - r1)2
4 1 4-
2 - 2 c o s ( 8 7 t/9 ) r4 + 2r2 c o s ( 8 jt/9 )
1
2
or. equivalently,
1.94(1
- r 1)1 =
1 .8 8 r2 + r 4
The value of r2 = desired filter is
0.7
satisfies this equation. Therefore, the system function for the
"<:l = a,5IT5&
Its frequency response is illustrated in Fig.
4.47.
338
Chap. 4
_ T
x ~2
r 2
Figure 4.47 Magnitude and phase response of a simple bandpass filter in Example 4.5.2; H(z) = 0.15[(1 - z ~ 2) / a + 0 . 7 ; - 2)].
It sh o u ld b e e m p h a siz e d th a t th e m ain p u rp o se of th e fo re g o in g m eth o d o lo g y fo r d esig n in g sim p le digital filters b y p o le - z e r o p la c e m e n t is to p ro v id e insight in to th e effect th a t p o les a n d z e ro s have on th e fre q u e n c y re sp o n s e characteristic o f system s. T h e m e th o d o lo g y is n o t in te n d e d as a goo d m e th o d fo r designing d igital filters w ith w ell-specified p a s sb a n d an d sto p b a n d ch aracteristics. System atic m e th o d s fo r th e d esig n of so p h is tic a te d d ig ital filters fo r p ra c tic a l ap p lic a tio n s are d iscussed in C h a p te r 8.
A simple lowpass-to-highpass filter transformation. S u p p o se th a t we h av e d esig n ed a p r o to ty p e low pass filter w ith im p u lse re sp o n se h\p(n). B y us-
Sec. 4.5
339
ing th e fre q u e n c y tra n sla tio n p ro p e rty o f th e F o u rie r tra n sfo rm , it is p o ssib le to co n v ert th e p ro to ty p e filter to e ith e r a b a n d p a ss or a h ig h p ass filter. F re q u en cy tra n sfo rm a tio n s fo r c o n v ertin g a p ro to ty p e low pass filter in to a filter of a n o th e r ty p e are d e sc rib e d in d etail in S ection 8.3. In th is sectio n w e p re se n t a sim plefre q u e n c y tra n sfo rm a tio n fo r co n v e rtin g a low pass filter in to a h ig h p ass filter, and vice v ersa. If /i|p(n) d e n o te s th e im p u lse resp o n se of a low pass filter w ith freq u e n cy re sp o n se H\r (co). a h ighpass filter can be o b ta in e d by tra n sla tin g H]P(a>) by t t rad ian s (i.e., rep lacin g co by co n ) . T h u s
^ h p ( ^ ) H\p(co TT )
(4.5.12)
w h e re H hP(w) is th e fre q u e n c y re sp o n se o f th e h ig h p ass filter. Since a freq u e n cy tra n sla tio n o f t t ra d ia n s is e q u iv a le n t to m u ltip lic atio n o f th e im pulse resp o n se /iip(;i) by e 177", th e im p u lse resp o n se o f th e highpass filter is /ihp(n) (p 7't )''/7|p (/7) = ( 1 )"/ ii p( )
(4.5.13)
T h e re fo re , th e im p u lse resp o n se of th e h ig h p ass filter is sim ply o b ta in e d from the im pulse re sp o n se o f the low pass filter by ch an g in g th e signs o f th e o d d -n u m b e re d sa m p le s in inp(n). C o n v ersely . = ( - 1 )"/ihp() If th e lo w pass filter is d e sc rib e d by th e d ifferen ce e q u a tio n
(4.5.15) (4.5.14)
its fre q u e n c y re sp o n s e is M k 0
(4.5.16)
N ow , if w e re p la c e co by co n , in (4.5.16). th en M
tfhpM =
-------------1 + ( - 1 )kake ~ iak
(4.5.17)
w hich c o rre sp o n d s to th e d iffe re n c e e q u a tio n

( 4 . 5 . 18 )
340 Example 4.5.3
Chap. 4
Convert the lowpass filter described by the difference equation v(n) = 0.9v(n 1) + 0 .1 x 0 ) into a highpass filter. Solution The difference equation for the highpass filter, according to (4.5.18), is y(n) = -0.9v(n - 1) + 0.1x(n) and its frequency response is
- 1+ Q9e_ju
The reader may verify that H^((o) is indeed highpass.
0.1
4.5.3 Digital Resonators

A digital resonat or is a sp ecial tw o-pole b a n d p a ss filter w ith th e p a ir o f com plexc o n ju g ate p o les lo c a te d n e a r th e u n it circle as show n in Fig. 4 .4 8 (a). T h e m ag n itu d e of th e fre q u e n c y re sp o n se of th e filter is show n in Fig. 4 .48(b). T h e n am e re so n a to r re fe rs to th e fact th a t th e filter has a large m a g n itu d e re sp o n se (i.e., it re so n a te s ) in th e v icinity o f th e p o le lo catio n . T h e a n g u lar p o sitio n of th e p o le d e te rm in e s th e re so n a n t freq u e n c y o f th e filter. D igital re so n a to rs are useful in m a n y applications, in clu d in g sim p le b a n d p a ss filterin g an d sp e ech g e n e ra tio n . In th e d esig n o f a digital r e s o n a to r w ith a re so n a n t p e a k at o r n e a r to o>o, w e se lect th e c o m p lex -co n ju g ate p o les at Pi .2 = r e ja* 0 < r < 1 In a d d itio n , we can select u p to tw o zeros. A lth o u g h th e re a re m an y possible choices, tw o cases a re o f special in terest. O n e choice is to lo cate th e z ero s a t the origin. T h e o th e r ch o ice is to locate a z e ro a t z = 1 an d a z e ro a t z = 1. T his choice co m p letely e lim in a te s th e re sp o n s e o f th e filter a t fre q u e n c ie s co = 0 and co = n , an d it is u se fu l in m an y p ractical a p p licatio n s. T h e sy stem fu n ctio n o f th e digital r e s o n a to r w ith ze ro s a t th e origin is H ( z ) = ----------- : ------ r ~ -----------: -------r (1 r e ^ z )(1 re -' "z- 1 ) (4.5.19)
H ( z ) = ------- ----------(4.5.20) 1 (2 r coscwo);-1 + r 2z ~2 S ince \H(co)\ h as its p e a k at o r n e a r co = coo, w e select th e gain bo so th a t |W(too)I = 1- F ro m (4.5.19) w e o b ta in
b0
H{(Oo) =
---------: ------- ----- ^-----------: ------- :---(1 - r e j<*,e -J*>)0. - r e - w e - w ) (1 r ) ( l r e - -'2'00)
(4 5 21)
_______________
an d h e n c e
\H(coo)\ = ------;---I - === = 1 (1 r)V 1 + r 2 2 r cos2a)o
bo
Sec. 4.5
341
2
(b)
tt
_ n
2
0 (c)
r
2
T t
Figure 4.48 (a) Pole-zero pattern and (b) the corresponding magnitude and phase response of a digital resonator with (1) r 0.8 and (2) r = 0.95.
342
Chap. 4
T h u s th e d esired n o rm a liz a tio n fa c to r is i?o = (1 r ) > /l + r 2 2r cos2a>o T h e freq u e n cy re sp o n se o f th e r e s o n a to r in (4.5.19) can b e e x p re ss e d as bo \H(a>)\ = ----------------Ux(to)U2 ((o) (w ) = 2 to < J> i (tfj) < t> 2 (co) w h ere U\ (w) a n d 1/ 2 (01) are th e m a g n itu d e s of th e v e c to rs fro m p\ an d p 2 to the p o in t w in th e u n it circle a n d <t>i(<w) an d < J> 2 (d) a re th e c o rre sp o n d in g angles of th e se tw o v ecto rs. T h e m a g n itu d e s Ui(to) a n d U2 (co) m ay be e x p re ss e d as Ui(co) = J \ + r 2 2 r cos(a>o to) (4.5.24)
U 2 (<o) = y / 1 + r 2 2 r c o s (w q + co)
(4.5.22)
(4.5.23)
F o r any v alu e o f r , U\(to) ta k e s its m in im u m v alu e (1 r ) a t to = too- T he p ro d u c t U\((o)U2 ({o) re a c h e s a m in im u m v alu e at th e fre q u e n c y air = co s-1 coswo^j (4.5.25)
w hich defin es p recisely th e r e s o n a n t fre q u e n c y o f th e filter. W e o b se rv e th a t w hen r is very close to u n ity , tor a>o, w hich is th e a n g u la r p o sitio n o f th e p o le. W e also o b se rv e th a t as r a p p ro a c h e s unity, th e re so n a n c e p e a k b ec o m e s s h a rp e r becau se U\(to) ch an g es m o re rap id ly in relativ e size in th e vicinity o f ojo- A q u a n tita tiv e m e a su re o f th e sh a rp n e s s o f th e re so n a n c e is p ro v id e d by th e 3-dB b a n d w id th A w o f th e filter. F o r v alu es o f r close to u nity. Aw % 2(1 - r ) (4.5.26)
F ig u re 4.48 illu stra te s th e m a g n itu d e a n d p h a s e o f d igital re s o n a to rs w ith c o q = n / 3 , r = 0.8 an d too ~ *73, r = 0.95. W e n o te th a t th e p h a se resp o n se u n d erg o es its g re a te st ra te o f c h a n g e n e a r th e re s o n a n t freq u e n cy . If th e zero s o f th e digital r e s o n a to r a re p laced a t z = 1 an d z = 1, th e re s o n a to r has th e sy stem fu n ctio n H(z) = G
(1 - z - ' X l + z " 1)
(1 re>aK ,z ~ l )( 1 r e - -'a* 2 -1 ) 1 -z -2 1 (2r coscdo)z-1 + r 2z ~2
(4.5.27)
= G
an d a fre q u e n c y re sp o n s e c h a ra c te ristic H{ w) = b [ \ _ (*-->][! _ r e - n o +*>)J (4.5.28)
W e o b se rv e th a t th e z e ro s a t z = 1 affe ct b o th th e m a g n itu d e a n d p h a se resp o n se o f th e re so n a to r. F o r ex a m p le , th e m a g n itu d e re sp o n s e is

\H(to)\=b0
U\(to)U2(to)
( 4 -5 2 9 )
Sec. 4.5
343
Figure 4.49 Magnitude and phase response of digital resonator with zeros at 10 = 0 and - t and ( I ) r = O.N and (2) r = 0.95.
to=
w h ere N { oj) is d efin ed as N ( w) 7 2 (1 - cos2a>) D u e to th e p re se n c e o f th e zero fa c to r, th e re s o n a n t fre q u e n c y is a lte re d from th at given by th e ex p ressio n in (4.5.25). T h e b a n d w id th of th e filter is also a lte re d . A lth o u g h ex act v alu es for th e s e tw o p a ra m e te rs are ra th e r te d io u s to d eriv e, we can easily c o m p u te th e fre q u e n c y re sp o n se in (4.5.28) an d c o m p a re th e re su lt with th e p re v io u s case in w hich th e zeros a re lo c a te d at th e origin. F ig u re 4.49 illu strates th e m a g n itu d e an d p h a s e c h a ra c te ristic s fo r o> q n/3. r = 0.8 an d a> o tt/ 3, r = 0.95. W e o b se rv e th a t th is filter h a s a slightly sm a ller b a n d w id th th a n th e re s o n a to r, w hich h as zero s a t th e origin. In a d d itio n , th e re a p p e a rs to be a v ery sm all shift in th e re so n a n t fre q u e n c y d u e to th e p re se n c e of th e zero s.
4.5.4 Notch Filters

A n o tc h filter is a filter th a t c o n ta in s o n e o r m o re d e e p n o tc h e s o r, ideally, p e rfe c t nulls in its fre q u e n c y re sp o n se ch a ra c te ristic . F ig u re 4.50 illu stra te s th e freq u e n cy re sp o n se c h a ra c te ristic o f a n o tc h filter w ith nulls a t fre q u e n c ie s cl>o an d . N otch filters are useful in m an y a p p lic a tio n s w h e re specific fre q u e n c y c o m p o n e n ts m ust be elim in a te d . F o r ex am p le, in stru m e n ta tio n a n d re c o rd in g sy stem s re q u ire th a t th e p o w er-lin e fre q u e n c y o f 60 H z a n d its h a rm o n ic s b e e lim in ated .
344
Chap. 4
Figure 4,50 Frequency response characteristic of a notch filter.
T o c re a te a null in th e fre q u e n c y re sp o n se o f a filter at a fre q u e n c y ojo, we sim ply in tro d u c e a p a ir of c o m p lex -co n ju g ate ze ro s on th e u n it circle at an angle too. T h a t is, S,.2 = f j T h u s th e system fu n ctio n fo r an F IR n o tch filter is sim ply H( z ) = boQ - eJlo" z ~] Ml (4.5.30) bo (1 2 C O S tr> ()c ' z ~) A s an illu stratio n . Fig. 4.51 show s th e m a g n itu d e re sp o n se fo r a n o tc h filter having a null a t a> = tt/4 . T h e p ro b lem w ith th e F IR notch filter is th a t th e n o tch has a relativ ely large b an d w id th , w hich m ean s th a t o th e r fre q u e n c y c o m p o n e n ts a ro u n d th e d esired null are se v ere ly a tte n u a te d . T o re d u c e th e b an d w id th o f th e null, w e can re so rt to a m o re so p h isticated , lo n g er F IR filter d esig n ed acco rd in g to c rite ria describ ed in C h a p te r 8. A lte rn a tiv e ly , w e could, in an ad h o c m a n n e r, a tte m p t to im prove on th e freq u e n cy re sp o n se ch a ra c te ristic s by in tro d u c in g p oies in th e system func tion. S u p p o se th a t w e p lace a p a ir o f c o m p lex -co n ju g ate p o les at
Ph2 = r e i m
T h e effect o f th e p o les is to in tro d u c e a re so n a n c e in th e vicinity o f th e null and th u s to red u ce th e b a n d w id th o f th e n o tch . T h e sy stem fu n ctio n fo r th e resulting filter is H
(c) =
1 - 2 cos wo; 1 + z~ bo1 - 2 r cosojoZ-1 + r 2z ~2
(4.5.31)
T h e m ag n itu d e re sp o n se \H(u>)\ o f th e filter in (4.5.31) is p lo tte d in Fig. 4.52 for coo = t t /4 , r = 0.85, an d fo r ti>o = n / 4 , r = 0.95. W h en c o m p a re d w ith the freq u e n cy resp o n se o f th e F IR filter in Fig. 4.51, w e n o te th a t th e effect o f the p o les is to red u ce th e b an d w id th o f th e no tch . In a d d itio n to red u cin g th e b a n d w id th o f th e n o tc h , th e in tro d u c tio n o f a p o le in th e vicinity o f th e null m ay re su lt in a sm all rip p le in th e p a s sb a n d o f th e filter d u e to th e re so n a n c e c re a te d by th e pole. T h e effe ct o f th e rip p le can be re d u c e d by in tro d u c in g a d d itio n a l p o le s a n d /o r ze ro s in th e sy ste m fu n c tio n o f the n o tch filter. T h e m a jo r p ro b le m w ith th is a p p ro a c h is th a t it is b a sic a lly an ad hoc, tria l-a n d -e rro r m e th o d .
Sec. 4.5
345
Figure 4.51 Frequency response characteristics of a notch filter with a notch at co = n / 4 or / = 1/8; H(z) = G[1 - 2cosooz-1 + z-2 ]-
4.5.5 Comb Filters

In its sim p lest fo rm , a c o m b filter can be v iew ed as a n o tc h filter in w h ich th e n u lls o c c u r p e rio d ic a lly ac ro ss th e fre q u e n c y b a n d , h e n ce th e an a lo g y to an o rd i n a ry c o m b th a t h as p erio d ically sp a ced te e th . C o m b filters find ap p lic a tio n s in a w ide ra n g e o f p ra c tic a l sy stem s such as in th e re je c tio n o f p o w e r-lin e h arm o n ics, in th e se p a ra tio n o f so la r a n d lu n a r c o m p o n e n ts fro m io n o sp h e ric m e a s u re m e n ts o f e le c tro n c o n c e n tra tio n , a n d in th e su p p re ssio n o f c lu tte r fro m fixed o b je c ts in m o v in g -ta rg e t-in d ic a to r (M T I) ra d a rs.
346
Chap. 4
Figure 4.52
Frequency response characteristics of two notch filters with poles at ( I ) r = 0,85 and 12) r (1.95:
H(z) = />n|(l - - cos to u r'1 2r cosgju:-1 + r2z ~~)].
r " ")/(1 -
T o illu strate a sim ple form of a co m b filler, c o n sid er a m oving a v erag e (F IR ) filter d escrib ed by th e d ifferen ce e q u a tio n v() = T h e system fu n ctio n o f this F IR filter is Hiz xin - k ) (4.5.32
+ * Jt= 0
1 M + 1 an d its freq u e n cy re sp o n se is S]na)( j d ) H(a>) = [ l - c - * ^ 11] ( 1 - - - )
(4.5.33)
(4.5.34) M + 1 sin(w /2)
F ro m (4.5.33) we o b se rv e th a t th e filter has ze ro s on th e u n it circle at k = 1 , 2 . 3 .........M (4.5.35)
N o te th a t th e p o le a t ; = 1 is actu ally can ce le d by th e zero at ; = 1, so th a t in effect th e F IR filter d o e s n ot c o n tain p o le s o u tsid e z = 0. A p lo t o f th e m ag n itu d e c h a ra c te ristic o f (4.5.34) clearly illu stra te s th e ex isten ce o f th e p erio d ically sp a c e d zero s in fre q u e n c y at cot = 2 n k / ( M + 1) for = 1 , 2 , ----- M. F ig u re 4,53 sh o w s \ H( w) \ fo r M = 10.
Sec. 4.5
347
Figure 4.53 Magnitude response characteristic for the comb filter given by (5,4.32) with M = 10.
In m o re g e n eral te rm s, w e can c re a te a com b filter by ta k in g an F I R filter w ith sy stem fu n ctio n M H( z ) = 2 2 h ( k ) z ~ k (4.5.36) k=0 an d rep la c in g z by z L, w h ere L is a po sitiv e in teg er. T h u s th e new F IR filter has a sy stem fu n ctio n
H l (z ) = Y , h ( k ) z
(4.5.37)
If th e fre q u e n c y re sp o n se of th e original F IR filter is //(w ), th e freq u e n c y resp o n se o f th e F IR in (4.5.37) is M H l (u> ) = 2 2 h ( k ) e ~ jkLa (4.5.38) k=I) = H(Lco) C o n s e q u e n tly , th e fre q u e n c y re sp o n se c h a ra c te ristic H l {oj) is sim ply an L -o rd e r re p e titio n o f H(co) in th e ra n g e 0 < co < 2 n . F ig u re 4.54 illu stra te s th e re la tio n sh ip b e tw e e n H l ( c o) a n d H (w) fo r L 5. N o w , su p p o se th a t th e o rig in al F I R filter w ith system fu n ctio n H ( z ) h as a sp e c tra l n ull (i-e -, a z e ro ), at so m e fre q u e n c y coo. T h e n th e filter w ith system fu n ctio n H L(z) h a s p erio d ically sp a ced nulls a t av = coo + I n k j L , k = 0, 1 , 2 , . . . , L 1. A s an illu stra tio n , F ig. 4.55 show s an F IR co m b filter w ith Af = 3 an d L = 3. T h is F IR filter can b e v iew ed as an F IR filter o f le n g th 10, b u t only fo u r o f th e 10 filter coefficients are n o n zero . L e t us no w r e tu rn to th e m oving a v erag e filter w ith system fu n ctio n given by (4.5.33). S u p p o se th a t we re p la c e z by z L. T h e n th e re su ltin g co m b filter h as th e sy stem fu n ctio n 1 1_ 1) Hl(z) = T 7~~r / (4-5.39) M + 1 l-z~ L
and a frequency response

H l ( co) =
1 M + 1
sin[a>L(M + l ) /2 ] sin(coL/2)
(4.5.40)
348

H (u > )
Chap. 4
(a)
Hl (co)
(b) Figure 4.54 Comb fiJter with frequency response W/Joi) obtained from H(a>).
Figure 4.55
Realization of an FIR comb filter having M = 3 and L ~ 3.
T his filter has zero s on th e u n it circle at

Zk = e j 2* k / L l M+ h
(4.5.41)
fo r alt in te g e r v alu es o f k e x cep t k 0, L, 2 L .........M L . F ig u re 4.56 illu strates \ H l {(d )\ fo r L = 5 an d M = 10. T h e co m b filter d esc rib e d by (4.5.39) finds a p p lic a tio n in th e s e p a ra tio n of so la r a n d lu n a r sp e c tra l c o m p o n e n ts in io n o sp h e ric m e a s u re m e n ts o f e le c tro n c o n c e n tra tio n as d e sc rib e d in th e p a p e r by B e rn h a rd t e t al. (1976). T h e so la r p e rio d is Ts = 24 h o u rs an d resu lts in a so la r c o m p o n e n t o f o n e cycle p e r d ay a n d its h arm o n ics. T h e lu n a r p e rio d is Tl = 24.84 h o u rs a n d p ro v id e s sp e c tra l lin es at 0.96618 cycle p e r d ay an d its h arm o n ics. F ig u re 4.57a show s a p lo t o f th e p o w e r d en sity sp e c tru m o f th e u n filtered io n o sp h e ric m e a s u re m e n ts o f th e e le c tro n con-
Sec. 4.5
Linear Time-Invariant Systems as Frequency-Selective Filters
349
Figure 4.56 Magnitude response characteristic for a comb filter given by (4.5.40). with L 3 anti M = II).
c e n tra tio n . N o te th a t th e w eak lu n a r sp e ctral c o m p o n e n ts are alm o st h id d en by th e stro n g so la r sp e c tra l co m p o n en ts. T h e tw o se ts o f sp e c tra l c o m p o n e n ts can be se p a ra te d by th e use o f com b filters. If w e w ish to o b ta in th e so lar co m p o n e n ts, we can use a co m b filter w ith a n a rro w p a s sb a n d a t m u ltip le s of o n e cycle p e r day. T h is can be ach iev ed by selectin g L such th a t Fs/ L = 1 cycle p e r day. w h ere Fs is th e c o rre sp o n d in g sa m p lin g freq u e n cv . T h e resu lt is a filter th a t has p e a k s in its freq u e n cy resp o n se at m u ltip le s o f o n e cycle p e r day. By se lectin g M = 58. the filter will h ave nulls at m u ltip le s o f ( F J L ) I ( M + 1) = 1/59 cycle p e r day. T h e se nulls a re very close to th e lu n a r c o m p o n e n ts a n d result in good rejectio n . F ig u re 4.57(b) illu strates
Freq ue ncy (cycles/day)

<c)
Figure 4.57 (a) Spectrum of unfiltered electron content data; (b) spectrum of out put of solar filter; (c) spectrum of output of lunar filter. [From paper by Bernhardt et al. (1976). Reprinted with permission of the American Geophysical Union.]
350
Chap. 4
the p ow er spectral density of the output o f the com b filter that isolates the solar com ponents. A com b filter that rejects the solar com p onents and passes the lunar com ponents can be d esigned in a sim ilar m anner. Figure 4.57(c) illustrates the pow er spectral density at the output o f such a lunar filter.
4.5.6 All-Pass Filters

An all-pass filter is defined as a system that has a constant m agnitude resp onse for all frequencies, that is. |ff{oj)l = l 0<o><7r (4.5.42)
The sim plest exam ple o f an all-pass filter is a pure delay system with system func tion H(z) = z~k This system passes all signals w ithout m odification except for a d elay of k sam ples. This is a trivial all-pass system that has a linear phase response characteristic. A m ore interesting all-pass filter is described by the system function n(z) = a x + a /v '-i" ^ 1 + + a \ Z 1 + 0\Z
,. .- N + k
;
n
' + + Qn Z
(4.5.43)
----- IT
a = 1
where all the filter coefficients \ak ) are real. If w e define the polyn om ial /\(~) as A (;) = Y ^ a kz k k=U then (4.5.43) can be expressed as
H(z) = z ~ n M
A(z)
an 1
(4.5.44)
Since
\H(a>)\2 = = 1 the system given by (4.5.44) is an all-pass system . Furtherm ore, if zo is a pole o f H( z ) . then 1/zu is a zero o f H ( z ) (i.e., the p oles and zeros are reciprocals of o n e another). Figure 4.58 illustrates typical p o le -z e r o patterns for a single-pole, single-zero filter and a tw o-p ole, tw o-zero filter. A plot o f the p h ase characteristics o f these filters is shown in Fig. 4.59 for a = 0.6 and r = 0.9, too = tt/4. The m ost genera! form for the system function o f an all-pass system with real coefficients, expressed in factored form in term s of p oles and zeros, is Nr H, f ^ - T T - _____ - 1J i n Pk l \ a - A z - x i - P i z ~ x) (4 5 45) (4 5 '
where there are N R real p o les and zeros and N c com p lex-con ju gate pairs o f poles and zeros. For causal and stable system s w e require that - 1 < a* < 1 and |/S*| < 1.
Sec. 4.5
351
(a)
Figure 4.58 Pole-zero patterns of (a) a first-order and (b) a second-order all-pass filter.
10 0
2 0 tog I H M I 0(u)
-10
-20
-3 0 -4 0
Figure 4.59 Frequency response characteristics of an all-pass filter with system functions (1) H(z) = (0.6 + z - ') / ( l + 0 . 6 Z '1), (2) H(z) = (r2 - 2tcoswqz~1 + Z ~ 2 ) /
(1 2rcoscuoz-i + r2z~2), r = 0.9, (jo n/A.
352
Chap. 4
E xp ression s for the phase response and group delay o f all-pass system s can easily be obtained using the m ethod described in Section 4.4.6. For a single p o lesingle zero all-pass system w e have Manioc) H en ce C -)ap(w) = cd 2 tan'
1

Ju r sin(a) 6}
r COS(w - 6 )
and di~).Ap(ct>) ciaj 1 - r2 1 + i - 2r cosiw 0) (4.5.46)
W e n ote that for a causal and stable system , r < 1 and hence rt, (w) > 0. Since the group delay o f a higher-order p o le-zero system consists o f a sum o f positive terms as in (4.5.46), the group delay will always be positive. A ll-pass filters find application as phase equalizers. W hen placed in cascade with a system that has an undesired phase response, a phase eq u alizer is designed to com pensate for the poor phase characteristics of the system and therefore to p roduce an overall linear-phase response.
4.5.7 Digital Sinusoidal Oscillators

A digital si nusoi dal oscillator can be view ed as a lim iting form o f a tw o-p ole res onator for which the com plex-conjugate p oles He on the unit circle. From our previous discussion o f secon d-ord er system s, we recall that a system with system function H ( z ) = -----------------------1 + a j z " 1 + ai z and param eters a i = 2rcosa>o and a2 = r 2 (4.5.48) (4.5.47)
has com plex-conjugate poles at p = r e m \ and a unit sam ple response h(n) = ~ sin(n + l)fDow(n) sin cun (4.5.49)
If the p oles are placed on the unit circle (r = 1) and bo is set to A sincuo, then h(n) = A s i n ( n + l)u>ou(n) (4.5.50)
Thus the im pulse response o f the second-order system with com plex-conjugate poles on the unit circle is a sinusoid and the system is called a digital sinusoidal oscillator or a digital si nusoi dal generator. A digital sinusoidal gen erator is a basic com ponent o f a digital frequency synthesizer.
Sec. 4.5
353
T h e block diagram representation o f the system function given by (4.5.47) is illustrated in Fig. 4.60. The corresponding difference equation for this system is v ( n ) = a i y0? 1) y( n 2) + boS(n) (4.5.51)
w here the param eters are a\ = 2cosa>o and bo = Asincwo, and the initial con d i tions are y ( 1) = v (2) = 0. By iterating the differen ce eq u ation in (4.5.51), we obtain y (0) = ^ sin w o y ( l ) = 2cos<woy(0) - 2 A sin too cos tt>o = Asin2tL>o y(2) 2costL>oy(l) - y(0) = 2 A cos ao sin 2wo A sin - A {4 cos2 coq 1) sin
wq lo q
3 A sin too ~ 4 sin 3 coq = A sin Ixoq and so forth. W e n ote that the application of the im pulse at n = 0 serves the purpose o f beginning the sinusoidal oscillation. T hereafter, th e oscillation is selfsustaining b ecau se the system has no dam ping (i.e., r = 1). It is in terestin g to note that the sinusoidal oscillation ob tain ed from the sys tem in (4.5.51) can also be ob tain ed by setting the input to zero and setting the initial co n d ition s to y ( - l ) = 0, v (2) = -A sin a> o. Thus the zero-input response to the secon d-ord er system described by the h om ogen eou s differen ce equation y{n) = - a i y ( n - 1) - y( n - 2) (4.5.52)
with initial con d ition s y ( - l ) 0 and >(2) = A sin two, is exactly the sam e as the resp onse o f (4.5.51) to an im pulse excitation. In fact, the d ifferen ce equation in (4.5.52) can b e ob tain ed directly from the trigonom etric identity sin a + sin fi = 2 sin - cos -
a +8
a - B
(4.5.53)
w here, by definition, or = (n + !)&*>, 0 = (n l)^ o , and y( n) = s i n( n + l)a>o-
354
Chap. 4
In som e practical applications involving m odulation of tw o sinusoidal carrier signals in phase quadrature, there is a need to generate the sinusoids A sin ojqn and A cosaion. T hese signals can be gen erated from the so-called coupl ed- f orm oscillator, which can be ob tain ed from the trigonom etric form ulas co s(a + ft) = cos a cos ft sin a sin ft sin (a + ft) = sin a cos ft + cos a sin ft where, by definition, a =
ft = u>o, and
yr (n) = COsnajoK(rc) y , (n ) = sinna>t)U(fj) Thus we obtain the tw o cou p led d ifference equations
)'(() = (COS a>o)} < ( - 1) - (s in a jo )y , ( - 1)
(4.5.54) (4.5.55)
(4.5.56) (4.5.57)
y.,() = (sin o>o)yt-(n - 1) + (cosw o)yr(n - 1) which can also be expressed in matrix form as
_vt(n) _
yAn)
cos co{) sin too sin t?o cos lo{)
yt.(n - 1) yAn ~
(4.5.58)
T h e structure for the realization of the coupled-form oscillator is illustrated in Fig. 4.61. We note that this is a tw o-output system which is not driven by any input, but which requires the initial con d ition s y<( ]) = /Icosaio and y.f ( 1) = -A sin a> o in order to begin its self-sustaining oscillations. Finally, it is interesting to note that (4.5.58) corresponds to vector rotation in the tw o-dim ensional coordinate system with coordinates yc(n) and y.T (n). A s a con sequ en ce, the coupled-form oscillator can also be im plem en ted by use o f the so-called C O R D IC algorithm [see the b ook by K ung et al. (1985)].
Figure 4.61 Realization of the coupled-form oscillator.
Sec. 4.6
inverse Systems and Deconvolution
355
4.6 INVERSE SYSTEMS AND DECONVOLUTION

A s we have seen , a linear tim e-invariant system takes an input signal v(/j) and produces an output signal y (n ), which is the con volu tion of * ( ) with the unit sam ple response h{n) of the system . In m any practical applications w e are given an output signal from a system w hose characteristics are unknow n and we are asked to determ ine the input signal. For exam p le, in the transm ission of digital inform ation at high data rates over telep h on e channels, it is w ell known that the channel distorts the signal and causes intersym bol interference am ong the data sym bols. T h e intersym bol interference m ay cause errors when w e attem pt to re cover the data. In such a case the problem is to design a corrective system which, w hen cascaded with the channel, produces an output that, in som e sense, corrects for the distortion caused by the channel, and thus yields a replica of the desired transm itted signal. In digital com m unications such a corrective system is called an equalizer. In the general con text of linear system s theory, how ever, w e call the corrective system an inverse s y s t e m , because the corrective system has a fre quency resp onse which is basically the reciprocal o f the frequency response o f the system that caused the distortion. F urtherm ore, since the distortivc system yields an output y i n ) that is the con volu tion of the input x ( n ) wiih the im pulse response h(n). the inverse system operation that takes y(/i) and produces a ( / i ) is called deconvol ut i on. If the characteristics o f the distorlive system are unknow n, it is often nec essary, when p ossib le, to excite the system with a known signal, observe the output, com pare it with the input, and in som e m anner, determ in e the charac teristics o f the system . For exam ple, in the digital com m unication problem just described, w here the frequency response of the channel is unknow n, the m ea surem ent o f the channel frequency response can be accom p lish ed by transm itting a set o f equal am plitude sinusoids, at different freq u en cies with a specified set o f phases, within the frequency band o f the channel. The channel will atten uate and phase shift each o f the sinusoids. By com paring the received signal with the transm itted signal, the receiver obtains a m easu rem en t o f the channel frequency resp onse which can be used to design the inverse system . The pro cess o f determ ining the characteristics of the unknow n system , either h( n) or H(co ), by a set o f m easurem ents perform ed on the system is called syst em identi fication. T h e term decon volu tion " is often used in seism ic signal p rocessing, and m ore generally, in geophysics to describe the op eration of separating the input signal from the characteristics o f the system which w e wish to m easure. T h e decon volu tion operation is actually in tend ed to identify the characteristics o f the system , which in this case, is the earth, and can also be view ed as a system id en tification problem . T he inverse system , in this case, has a frequency response that is the reciprocal o f the input signal spectrum that has b een used to excite the system .
356
Chap. 4
4.6.1 Invertibility of Linear Time-Invariant Systems

A system is said to be invertible if there is a o n e-to -o n e corresp on d en ce betw een its input and output signals. This definition im plies that if w e know the output sequ en ce y(n), oc < n < oc, of an invertible system T , w e can un iqu eiv determ ine its input A(n), oc < n < oo. T he inverse syst em with input v() and output x{n) is d en oted by T ~ l . Clearly, the cascade con n ection o f a system and its inverse is equivalent to the identity system , since w( n) = 7 1[>'()] = T ~ x [T[x{n)]) - x(n ) (4.6.1)
as illustrated in Fig. 4.62. For exam ple, the system s defined by the input-output relations y{n) = ax ( n ) and y( n) = x ( n 5) are invertible, w h ereas the input-output relations y( n) = x 2(n) and y{n) = 0 represent noninvertible system s. A s indicated above, inverse system s are im portant in m any practical appli cations. including geop h ysics and digital com m unications. Let us begin by con sidering the problem of determ ining the inverse o f a given system . W e limit our discussion to the class o f linear tim e-invariant discrete-tim e system s. N ow . suppose that the linear tim e-invariant system T has an im pulse response h(n) and let h t (n) d en ote the im pulse response o f the inverse system T _ l . Then (4.6.1) is equivalent to the con volu tion equation U(n) = h / ( n) * h( n) * x ( n ) = x( n) But (4.6.2) im plies that h(n) * h/ ( n) = S(n) (4.6.3) (4.6.2)
The convolution equation in (4.6.3) can be used to solve for h/ ( n) for a given h(n). H ow ever, the solution o f (4.6.3) in the tim e dom ain is usually difficult. A sim pler approach is to transform (4.6.3) into the z-dom ain and so lv e for T _1. Thus in the z-transform dom ain, (4.6.3) becom es H{z )H, (z) = 1 and therefore the system function for the inverse system is
*,(:)= ^
H( z) =
( 4 . 6 . 4 )
(4.6.5)
If H ( z ) has a rational system function
/1(c)
Identity system
v( n ) T Direct system T -i Inverse system
n '(n ) = x(n)
Figure 4.62 System T in cascade with its inverse T _1.
Sec. 4.6 then
Inverse Systems and Deconvolution
357
(4.6.6) Thus the zeros o f H ( z ) b ecom e the p oles of the inverse system , and vice versa. Furtherm ore, if H{ z ) is an F IR system , then Hi ( z) is an all-p ole system , or if H{z ) is an all-p ole system , then H s {z) is an F IR system . Example 4.6.1 Determine the inverse of the system with impulse response h(n) = UyuOO Solution The system function corresponding to h(n ) is H(z) 1-
This system is both causal and stable. Since H(z) is an all-pole system, its inverse is FIR and is given by the system function
h ,{ z)
= i - h-1
Hence its impulse response is = &(n) 1)
Example 4.6.2 Determine the inverse of the system with impulse response
h(n) = <5(n) j<5(h 1)
Solution
This is an FIR system and its system function is H(z) = 1 - k ~ ' ROC: |z| > 0
The inverse system has the system function
Thus H/ ( z ) has a zero at the origin and a pole at z = j- In this case there are two possible regions of convergence and hence two possible inverse systems, as illustrated in Fig. 4.63. If we take the ROC of Hi(z) as |j| > j, the inverse transform yields
which is the impulse response of a causal and stable system. On the other hand, if the ROC is assumed to be |j| < | , the inverse system has an impulse response
In this case the inverse system is anticausal and unstable.
358
Chap. 4
(b)
Figure 4.63 Two possible regions of convergence for H (z) = z/(z - Y>-
W e observe that (4.6.3) cannot be solved uniquely by using (4.6.6) unless we specify the region of con vergen ce for the system function o f the inverse system. In som e practical applications the im pulse response h(n) d o es not possess a z-transform that can be expressed in closed form. A s an alternative w e may solve (4.6.3) directly using a digital com puter. Since (4.6.3) d oes not. in general, possess a unique solution, we assum e that the system and its inverse are causal. Then (4.6.3) sim plifies to the equation ^ h(k)hi(rt k) S(/i) (4.6.7)
By assum ption, h/ ( n) = 0 for n < 0. For n = 0 we obtain

h,(0 ) = l / h ( 0 )
(4.6.8)
T he values o f hi {n) for n > 1 can be ob tain ed recursively from the equation
^ h(k)h/(n-k) ,
= ~
mo)
n21
( A -6
This recursive relation can easily be program m ed on a digital com puter.
Sec. 4.6
359
T here are two problem s associated with (4.6.9). First, the m ethod d oes not work if h(0) = 0. H ow ever, this problem can easily be rem ed ied by introducing an appropriate delay in the right-hand side o f (4.6.7), that is, by replacing <50) by S(n m) , where m = 1 if *(0) = 0 and /i( l) ^ 0, and so on. Second, the recursion in (4.6.9) gives rise to rou n d -off errors which grow with n and, as a result, the num erical accuracy o f h(n) d eteriorates for large n. Example 4.6.3 D eterm ine the causal inverse of the FIR system with impulse response
h(n) = S(n) aS(n 1)
Solution and
Since /i(0) = 1. h(1 ) = a, and h(n) = 0 for n > a , we have
MO) = 1 /MO) = 1
h/(n) = a M n _ 1 ) Consequently, n 1
M l) = a .
M 2) = a 2 ..........
h,(n) =a"
which corresponds to a causal IIR system as expected,
4.6.2 Minimum-Phase, Maximum-Phase, and Mixed-Phase Systems

T h e invertibility o f a linear tim e-invariant system is intim ately related to the char acteristics o f the phase spectral function o f the system . T o illustrate this point, let us consider tw o FIR system s, characterized by the system functions H\ {z) = 1 + j ; -1 = z_1(z + \ ) Hj i z ) = z ! ( |z + 1) (4.6.10) (4,6.11)
The system in (4.6.10) has a zero at z = and an im pulse response fc(0) = 1, /i(l) = 1/2. T h e system in (4.6.11) has a zero at z = 2 and an im pulse response h(0) = 1/2, /i(1) = 1, which is the reverse of the system in (4.6.10). This is due to the reciprocal relationship b etw een the zeros o f H\ ( z) and In the frequency dom ain, the tw o system s are characterized by their fre quency response functions, which can be expressed as |ffi(<y)| = |H 2(o))! = y | + cos to and i((w) = to + tan
,
(4.6.12)
sin w -j------------| + cos O J sin to ------------2 + cos to
(4.6.13)
2 (cl>) = to + tan
_i
(4.6.14)
The m agnitude characteristics for the tw o system s are identical becau se the zeros o f Hi ( z ) and Hi ( z ) are reciprocals.
360 fr|I co)
Chap. 4
94a.)
i
(b)
Figure 4.64 Phase response charactcnslics for the systems in (4.6.10) and (4.6.11).
The graphs o f 0 ] (co) and (~)2 (cv) are illustrated in Fig. 4.64. W e observe that the phase characteristic 0i(cd ) for the first system begin s at zero phase at the fre quency w = 0 and term inates at zero phase at the frequency cv 7t. H en ce the net phase change, 0 i( ;r ) - 0i(O ) is zero. On the other hand, the p h ase characteristic for the system with the zero outside the unit circle undergoes a n et phase change 0 ;(;r ) 2(0) = n radians. A s a con seq u en ce o f these different p h ase character istics, w e call the first system a m i n i mu m - p h a s e s y s t e m and the secon d system is called a m a x i mu m - p h a s e system. Th ese definitions are easily exten d ed to an F IR system o f arbitrary length. T o be specific, an F IR system o f length M + 1 has M zeros. Its frequency response can be expressed as H(co) = bQ(\ - z \ e ~ iw)( 1 - z 2e ~ n ( ! - z Me ~ J) (4.6.15)
where {;,} d en ote the zeros and bo is an arbitrary constant. W h en all the zeros are inside the unit circle, each term in the product o f (4.6.15), corresponding to a real-valued zero, will undergo a net phase change o f zero b etw een a> = 0 and (* ) = n . A lso , each pair of com p lex- conjugate factors in H(a>) will undergo a net phase change o f zero. T herefore, ^ H ( n ) - ^H(O) = 0 (4.6.16)
and h en ce the system is called a m inim um -phase system . O n the o th er hand, when all the zeros are outside the unit circle, a real-valued zero will con trib ute a net
Sec. 4.6
361
phase change o f tt radians as the frequency varies from w = 0 to co it. and each pair o f com p lex-con ju gate zeros will contribute a net phase change of 2 tt radians over the sam e range o f ai. T herefore.
iL H ( jt )
- ^ H ( O ) = M tt
(4.6.17}
which is the largest possible phase change for an FIR system with M zeros. H ence the svstem is called m axim um phase. It follow s from the discussion above that 4- Hmax{7T) > 4 tf m,n(jr) (4.6.18)
If the FIR system with M zeros has som e o f its zeros inside the unit circlc and the rem aining zeros ou tsid e the unit circle, it is called a mi x e d - p h a s e system or a n o n m i n i m u m - p h a s e syst em. Since the derivative o f the phase characteristic o f the system is a m easure of the tim e d elay that signal frequency com p onents undergo in passing through the system , a m inim um -phase characteristic im plies a m inim um delay function, while a m axim um -phase characteristic im plies that the delay characteristic is also m axim um . N ow suppose that we have an FIR system with real coefficients. .Then the m agnitude square value of its frequency response is |t f ( w ) |: = )|r^ ,(4.6.19)
This relationship im plies that if we replace a zero o f the system by its inverse 1 /zk- the m agnitude characteristic o f the system d oes not change. Thus if we re flect a zero zt that is inside the unit circle into a zero I/:* ou tsid e the unit circle, we se e that the m agnitude characteristic o f the frequency response is invariant to such a change. It is apparent from this discussion that if \ H ( c o ) \ 2 is the m agnitude square frequency resp onse o f an F IR system having M zeros, there are 2 M possible con figurations for the M zeros, som e of which are inside the unit circle and the re m aining are ou tsid e the unit circle. Clearly, one configuration has all the zeros inside the unit circle, which corresponds to the m inim um -phase system . A sec ond configuration has all the zeros outside the unit circle, which corresponds to the m axim um -phase system . T h e rem aining 2 W 2 configurations correspond to m ixed-phase system s. H ow ever, not all 2 M - 2 m ixed-phase configurations n ec essarily correspond to F IR system s with real-valued coefficients. Specifically, any pair o f com p lex-con ju gate zeros result in on ly two possib le configurations, w hereas a pair o f real-valued zeros yield four possib le configurations. Example 4.6.4 D eterm ine the zeros for the following FIR systems and indicate whether the system is minimum phase, maximum phase, or mixed phase. tf,(z) = 6 + z_1 - z~2
H 2{ z ) = 1 - r 1 - 6 ;" 2
362
Frequency Analysis of Signals and Systems H: Az) = 1 - ^c"1 - $z~z

HAZ) =
Chap. 4
1+
f r ' -
Solution
By factoring the system functions we find the zeros for the four systems H]{z) * Ci.: = * { * minimum phase
H 2 (z )
are
- ci,: = - 2 , 3 maximum phase
H:,(z) * ci.: = - J. 3 * mixed phase Hi\z) * Ci.: = 2, t * mixed phase Since the zeros of the fouT systems are reciprocals of one another, it follows that all four systems have identical m agnitude frequency response characteristics but different phase characteristics. The m inim um -phase property o f F I R system s carries over to I I R system s that have rational system functions. Specifically, an I I R system with system function B(z) H( z ) =
A (c )
(4.6.20)
is called m i n i m u m p h a s e if all its poles and zeros are inside the unit circle. For a stable and causal system [all roots of A (c) fall inside the unit circle] the system is called m a x i m u m pha s e if all the zeros are outside the unit circle, and m i x e d phase if som e, but not all. o f the zeros are ou tsid e the unit circle. This discussion brings us to an im portant point that should be em phasized. That is. a stable p o le-zero system that is m inim um phase has a stab le inverse which is also minim um phase. The inverse system has the system function = d(z) (4 -6 -21)
H en ce the m inim um -phase property of H ( z ) ensures the stability o f the inverse system H ~ l (z) and the stability o f H( z ) im plies the m inim um -phase property of H ~ \ z ) . M ixed-phase system s and m axim um -phase system s result in unstable in verse system s.
D ecom p osition
of
n on m in im u m -p h ase
p o le -z e r o
sy stem s.
Any
nonm inim um -phase p o le -z e r o system can be expressed as

(z) = /W z)tfa p (z) (4.6.22)
w here is a m inim um -phase system and / / ap(z) is an all-pass system . We dem onstrate the validity o f this assertion for the class o f causal and stable systems w-ith a rational system function H ( z ) = B ( z ) / A ( z ) . In general, if B(z) has one or m ore roots ou tsid e the unit circle, w e factor B(z) in to the product B i( z ) # 2 (z)> w here B i(z) has all its roots inside the unit circle and B 2(z) has all its roots outside
Sec. 4.6
363
the unit circle. T h en m inim um -phase system
has all its roots inside the unit circle. W e define the B i i - J B j i z - 1) Mz)
and the all-pass system Hw (z) = Biiz) B 2(z - 1)
Thus H ( z ) Fiminiz)Hap(z) N ote that / / ap(z) is a stable, all-pass, m axim um -phase system .
Group delay of nonminimum-phase system. B ased on the d ecom p osi tion o f a nonm inim um -phase system given by (4.6.22), w e can express the group delay o f H( z ) as
Tg(o>) = rin(co) + âpM (4,6.23)
Since r r (u>) > 0 for 0 < to < n , it follow s that 1 ^( 0;) > rn(o;), 0 < a>< tt. From (4.6.23) w e co n clu d e that am ong all p o le -z e r o system s having the sam e m agnitude response, the m inim um -phase system has the sm allest group delay.
Partial energy o f nonminimum-phase system.

causal system with im pulse response h(n) is defined as Cn) = W * ) I 2 t=u
T he partial energy o f a
(4.6.24)
It can be show n that am ong all system s having the sam e m agnitude response and the sam e total en ergy ( 0 0 ), the m inim um -phase system has the largest partial energy [i.e., E min(n) > E( n) , where Emj(n) is the partial energy o f the m inim um phase system ].
4.6.3 System Identification and Deconvolution

Suppose that w e excite an unknow n linear tim e-invariant system with an input se quen ce x ( n ) and w e observe the output seq u en ce y( n). From the output sequ en ce w e wish to determ ine the im pulse resp onse o f the unknow n system . This is a prob lem in s y s t em identification, which can be solved by deconvol ut i on. Thus w e have y(n) = h( n) * *(n) ^ = k= oo A n analytical solution o f the d econ volu tion problem can b e obtained by w orking with the z-transform o f (4.6.25). In the z-transform dom ain w e have Y( z ) = H( z ) X( z ) h( k ) x ( n k) (4-6.25)
364 an d h en ce
Chap, 4
K(;) HrJ = - * (;)
(4.6.26)
X u ) and are th e ^ -tra n sfo rm s of the av ailab le in p u t signal x(/;) a n d the o b se rv e d o u tp u t signal yi n) , resp ectiv ely . T his a p p ro a c h is a p p r o p ria te only w hen th e re are clo sed -fo rm ex p ressio n s fo r X ( z ) an d F (c). Example 4.6.5 A causal system produces the output sequence f 1. yin) = | . I 0. when excited by the input sequence (1 . x(n) = i
H i'
n=0 n= 1 otherwise n= 0
I 0,
otherwise
Determ ine its impulse response and its input-outpul equation. Solution The system function is easily determ ined by taking the .--transforms of ,v(n) and yin). Thus we have
H( z ) = ----- =
no
1 + 77,:"
-------: ----- 7
1 - 17i1-r (1 )(1 -
+ i7>"~
Since the system is causal, its ROC is |;| > i. The system is also stable since its poles lie inside the unit circle. The input-output difference equation for the system is y i n ) = -jj;y(/i - 1 ) - ^ y i n - 2 ) + ,r() + y^.v(n - 1 ) Its impulse response is determined by performing a partial-fraction expansion of H(z) and inverse transforming the result. This com putation yields hin) = [4(i)" - 3({ )n]u(n) W e observe that (4.6.26) d eterm in es the unknow n system uniquely if it is known that the system is causal. H ow ever, the exam ple above is artificial, since the system response {>()} is very likely to be infinite in duration. C onsequently, this approach is usually im practical. A s an alternative, we can deal directly with the tim e-dom ain expression given by (4.6.25). If the system is causal, w e have n y( n) = ^ h{ k) x( n k) n> 0 *=o
Sec. 4.6 and h ence
365
n-l
y { n ) y ~ ^ h ( k ) x ( n k ) n > 1
(4.6.27)
Hn) =
x(0)
This recursive solu tion requires that jr(0) ^ 0. H ow ever, we n ote again that w hen (/i(n)) has infinite duration, this approach m ay not be practical unless w e truncate the recursive solu tion at sam e stage [i.e., truncate {/z()}]. A n o th er m eth od for identifying an unknow n system is b ased on a crosscor relation technique. R ecall that the in p ut-ou tp u t crosscorrelation function derived in Section 2.6.5 is given as
oc
r y x ( r 77) = ^ h ( k ) r ( X ( m k ) = h ( n ) * r x ! ( m )
(4.6.28)
t=0 where r y x ( m ) is the crosscorrelation seq u en ce o f the input {x(^)J to the system with the output {>()} o f the system , and r x x { m ) is the autocorrelation sequ en ce o f the input signal. In the frequency dom ain, the corresponding relationship is
S vx((d)
J-!(io)S.fX (co)
=
-
H ( c o ) \X (u >)\2
H en ce Svi(a>)
H ( w ) = ---------- =
5V j(w )
1 r
SX i ( w )
\ X ( a >)\2
(4.6.29)
T h ese relation s suggest that the im pulse response (/?(k)1 or the frequency re sp on se o f an unknow n system can be determ ined (m easu red ) by crosscorrelating the input seq u en ce {*()} with the output seq u en ce (y(n)}, and then solvin g the d econ volu tion problem in (4.6.28) by m eans o f the recursive eq u ation in (4.6,27). A ltern atively, w e cou ld sim ply com pute the Fourier transform o f (4.6.2 8) and d e term ine the freq u en cy response given by (4.6.29). Furtherm ore, if w e select the input seq u en ce (jc(n)} such that its autocorrelation seq u en ce { ^ ( n ) } , is a unit sam ple seq u en ce, or eq u ivalen tly, that its spectrum is flat (con stan t) over the passband o f H(a>), the valu es o f the im pulse resp onse {/?()} are sim ply equal to the values o f the crosscorrelation seq u en ce {rVJ()}. In general, the crosscorrelation m eth od described above is an effective and practical m eth od for system identification. A n oth er practical approach based on least-squares optim ization is described in C hapter 8.
4.6.4 Homomorphic Deconvolution

T h e com p lex cepstrum , introduced in Section 4.2.7, is a useful to o l for perform ing d econ volu tion in so m e applications such as seism ic signal processing. T o describe this m eth od , let us su p p ose that {>>()} is the output seq u en ce o f a linear timeinvariant system w hich is excited by the input seq u en ce (x(n )f. Then Y( z) = X{ z ) H( z )
(4.6.30)
366
Chap. 4
w here H( z ) is the system function. T h e logarithm o f Y ( z ) is C A z ) = In Y (c) = in X (c) + In H( z ) = C ,(c) + C>,U) C onsequently, the com plex cepstrum o f the output sequence (y(n)} is expressed as the sum o f the cepstrum o f |x(n )} and {/i(n)J, that is, c y( n ) = cx (n) + ch(n) (4.6.32) (4.6.31)
Thus w e observe that con volu tion of the tw o seq u en ces in the tim e dom ain corre sponds to the sum m ation o f the cepstrum seq u en ces in the cepstral dom ain. The system for perform ing these transform ations is called a h o m o r m o r p h i c sy s t em and is illustrated in Fig. 4.65. In som e applications, such as seism ic signal processing and speech signal processing, the characteristics of the cepstral sequ en ces (c,(/j)} and {c>(n)} are suf ficiently different so that they can be separated in the cepstral dom ain. Specifically, suppose that {c* (/?)} has its main com p on en ts (m ain energy) in th e vicinity o f small values o f n, w hereas |c r(n)} has its com p on en ts concentrated at large values of n. W e m ay say that |c(n)} is lowpass" and {cx()l is highpass. W e can then sepa rate {cy,(n)} from {c,(r;)) using appropriate low p ass and h igh pass w indow s, as illustrated in Fig. 4.66. Thus ch(n) = c v(rt)wir (rr) and cx (n) = c_v(n )Whp(n) (4.6.34) (4.6.33)
z-Transform
logarithm
c j.)
;-iransform
Figure 4.65 Homomorphic system for obtaining the cepstrum (cv(n)} of the se quence (y(n)l-
F igure 4.66 S ep aratin g th e two cepstral com ponents by low pass" and highpass windows.
Sec. 4.7 w here
Summary and References
367
U ) p ( J l)
1. 0. 0,
< N, otherw ise ll < N\ \n\ >
(4.6.35)
U hpO ) = 1. 1
(4.6.36)
O nce w e have separated the cepstrum sequ en ces ( o ,( ) } and (c,-(n)} by w indow ing, the seq u en ces {x(n)} and {/i(n)( are obtained b y p a ssin g (o ,( )| and (c.v(n)) through the inverse hom om orphic system , shown in Fig. 4.67. In practice, a digital com puter w ould be used to com p ute the cepstrum o f the seq u en ce {v()}. to perform the w indow ing functions, and to im plem ent the inverse h om om orphic system shown in Fig. 4.67. In place o f the --transform and inverse z-transform . we w ould substitute a special form of the Fourier transform and its inverse. T his special form , called the discrete Fourier transform, is described in C hapter 5. C ,<M )
.--Transform C,(.v) Com plex exponential W(c) XO | .v(n) Inverse "-rnmt l^nT> ( ir i /iijn
Figure 4.67 Inverse homomorphic system for recovering the sequences {.kii )| mid |/j()) from the corresponding cepstru.

T he Fourier series and the Fourier transform are the m athem atical tools lor an alyzing the characteristics o f signals in the frequency dom ain, T he Fourier series is appropriate for representing a periodic signal as a w eigh ted sum of harm oni cally related sinusoidal com p onents, w here the w eighting coefficients represent the strengths o f each o f the harm onics, and the m agnitude squared of each w eighting coefficient represents the pow er of the corresponding harm onic. A s we have in dicated, the Fourier series is on e o f m any possible orthogonal series expansions for a p eriodic signal. Its im portance stem s from the characteristic behavior o f LTI system s, as we shall see in Chapter 5. T h e Fourier transform is appropriate for representing the spectral charac teristics o f aperiodic signals with finite energy. T he im portant properties of the F ourier transform were also presented in this chapter. There are m any excellen t texts on Fourier series and Fourier transforms. F or reference, w e include the texts by B racew ell (1978), D avis (1963), Dvm and M cK ean (1972). and P apoulis (1962). In this chapter w e also considered the frequency-dom ain characteristics o f LTI system s. W e show ed that an LTI system is characterized in the frequency d om ain by its frequency response function H ( ai), w hich is the Fourier transform
368
Chap. 4
o f the im pulse response o f the system . W e also observed that the frequency response function d eterm ines the effect o f the system on any input signal. In fact, by transform ing the input signal into the frequency dom ain, w e ob served that it is a sim ple matter to determ ine the effect of the system on the signal and to determ ine the system output. W hen view ed in the frequency dom ain, an LTI system performs spectral shaping or spectral filtering on the input signal. The design o f som e sim ple IIR filters was also considered in this chapter from the view point o f p o le-zero placem ent. B y m eans o f this m eth od , we w ere able to design sim ple digital resonators, notch filters, com b filters, ail-pass filters, and digital sinusoidal generators. T h e design o f m ore com p lex IIR filters is treated in detail in Chapter 8. which also includes several references. D igital sinusoidal gen erators find use in frequency synthesis applications. A com p reh en sive treatm ent of frequency synthesis tech n iqu es is given in the text edited by G orski-P op iel (1975). Finally, w e characterized LTI system s as either m inim um -phase, maximumphase, or m ixed-phase, d ep en d ing on the p osition o f their p o les and zeros in the frequency dom ain. U sing these basic characteristics of LTI system s, w e considered practical problem s in inverse filtering, d econ volu tion , and system identification. W e concluded with the description o f a d econ volu tion m eth od based on cepstral analysis o f the output signal from a linear system . A vast am ount o f technical literature exists on the topics o f inverse filter ing. d econ volu tion , and system identification. In the context o f com m unications, svstem identification, and inverse filtering as they relate to channel equalization are treated in the book by Proakis (1995). D econ volu tion tech n iqu es are widely used in seism ic signal processing. For reference, w e suggest the papers by W ood and Treitel (1975), P eacock and T reitel (1969), and the b ook s by R obinson and Treitel (1978, 1980). H om om orphic d econ volu tion and its ap p lication s to speech processing is treated in the book by O p p en heim and Schafer (1989).
PROBLEMS
4.1 Consider the full-wave rectified sinusoid in Fig. P4.1. (a ) Determine its spectrum X a(F). (b) Compute the power of the signal.
Xa( ! )
Figure P4.1
Chap. 4
Problems
369
(c) Plot the power spectral density. (d) Check the validity of Parseval's relation for this signal. 4.2 Com pute and sketch the m agnitude and phase spectra tor the following signals (a > 0). Ae~a' , i > 0 la) V - , ' , = '0 . ,< 0 (b) xu(t) = Ae~'r 4.3 Consider the signal
1 l r |h . |r | < t
' 0.
elsewhere
(a) Determ ine and sketch its magnitude and phase spectra. |X (F)| and 2^ X a(F), respectively. ( b) Create a periodic signal x,.(t) with fundam ental period Tr > 2r. so that ,v(/) = .v,,(n for |f| < T;,/2. What are the Fourier coefficients ci for the signal x;,u)? (c) Using the results in p an s (a) and (b). show that a = (1/X;,)X (k/Tr ).
4.4 Consider the following periodic signal:
xin) = { ..., I. (I. 1.2. 3.2. 1.0. I. . . ,|

t
(a) Sketch the signal vt/i) and its magnitude and phase spectra. ( b) Using the results in pari (a), verily Parseval's relation by computing the power in the time and frequency domains. 4.5 Consider the signal rr/i nn 1 3nn xin ) = 2 -f- 2 cos ------Hcos h c o s ----4 2 2 4 (a) Determ ine and sketch its power density spectrum.
( b) Evaluate the power of the signal.
4.6 Determ ine and sketch the magnitude and phase spectra of the following periodic signals. n(n - 2) (a) ,v<(t)=4sin 3 In . In (b) ,v(n) = cos n + sin n 3 ^ 2n . 2n (c) x(n) = cos n sin n
(d) x( n)
= {..., - 2 . - 1 . 0 . 1 . 2 . - 2 . - 1 . 0 . 1.2.... )
t
( e) x i n ) = ( . . . . - 1 . 2 . 1. 2. - 1 . 0 . - 1 . 2 . 1 . 2 . . . . J (f) x (n) = ( . . . . 0 , 0 . 1 . 1 . 0 . 0 . 0 . 1. 1 . 0 . 0 , . . . )
t (g) x i n) = 1 . oc < n < oc ( h) x(n) = ( - 1 ) " . - o c < n < oc 4.7 D eterm ine the periodic signals * 0 ), with fundam ental period N = 8. if their Fourier coefficients are given by: kn 3kn
370 kn
(b) c * = { Slny -
Chap. 4
Q < k <6
0,
k= 7
t
(c) {c*} = { . . . . O . i . i . l . 2 . 1 . i , i . O . 4.8 Two DT signals. j*() and s{(n ), are said to be orthogonal over an interval [N\, A^] if Y2st ( n) s ?( n) = j 0 *' If At = 1. the signals are called orthonorm al, (a) Prove the relation N ' ej2*kn/* = ' N, 0. k = 0, jV, 2N, otherwise k =t k / I
(b) Illustrate the validity of the relation in part (a) by plotting for every value of
k ~ 1 ,2 ........ 6. the signals st (n) = eJI2,,/('}k" , n = 0, 1 ,___ 5. [Aro/e: For a given k, n the signal can be represented as a vector in the complex plane.] (c) Show that the harmonically related signals i*(n) = ejan,K'kB are orthogonal over any interval of length N. 4.9 Compute the Fourier transform of the following signals. (a) x(n) = u(n) u(n 6 ) (b) x(n) = 2"u{-n) (c) x(n) = (j)"u(n + 4) (d) x(n) = (a" sin a>on)u(n) |a | < 1 (e) x(n) = laTsiniuiin jar| < 1 (f) x(n) = 2 -U )n . 0,
t
|| 5 4 elsewhere
(g) xin) = { - 2 . - 1 . 0 . 1 . 2 ) ^ . U (2 A / + l - | n | ) . (h) x(n) = | 0.
. . |n| > M
\ n\ <M
Sketch the magnitude and phase spectra for parts fa), (f), and (g). 4.10 D eterm ine the signals having the following Fourier transforms. 0, 0 < \a> | < ) = ,
I.
W() < \a>\ < 7T
(b) X (u>) = cos2 a >
(n\ Yt \ - \ O* - &i/2 < M < O *) + to /2 W 1 0, elsewhere (d) The signal shown in Fig. P4.10. 4 .11 Consider the signal x{n) = {1 , 0,- 1 ,2 .3 }
t
Chap. 4
Problems Xitu)
371
3tt
jt
7r
3n
6 jt
7tt
tt
Figure P4.10
with Fourier transform Xtw) = X's (ct>) + j ( X t (a>)). D eterm ine and sketch the signal y(n) with Fourier transform Y(w) = X,(o>) + X R( ^) ei2,il 4.12 Determ ine the signal
jc(h )
if its Fourier transform is as given in Fig. P4.12.
8k 10
0
(a)
8rr
971
7 T
10
To
X(ai)
(b) X(u> )
(c )
Figure P4.12
372
Chap. 4
4.13 In Example 4.3.3. the Fourier transform of the signal

x (n) =
1, 0.
-M < n < M otherwise
was shown to be M X (a> ) = 1 + 2 2 2 cos am f!=1 Show that the Fourier transform of X|(n) = and x 2(n) = are, respectively. ] _ Xi(a>) = X 2{to) = Thus prove that
X (to) - X 1 (tt>) + X 2 (tt)}
1. 0. 1. 0,
0< n <M otherwise M < n < 1 otherwise I
sin(M + \) w sintw / 2 ) and therefore. coswn = 4.14 Consider the signal x(n) = ( - 1 ,2 , - 3 ,2 , -1}
t
sin(M + \ ) w sin(a/2 j
with Fourier transform X((u). Compute the following quantities, without explicitly computing X(a>): (a) X(0) (b) AX(co) (c) f * n X(w) dw (d) X (jt) (e) f *JX( co ) \ 2 dco 4.15 The center of gravity of a signal x(w) is defined as T , nx(n)
yx(n)
7I = OC
and provides a measure of the time delay of the signal.
Chap. 4
Problems Xui
373
Figure P4.15
(a) Express c in terms of X{<o). (b) Compute c for the signal x(n) whose Fourier transform is shown in Fig. P4.15. 4.16 Consider the Fourier transform pair a"u(n) -----------1 at
|ti) < 1
Use the differentiation in frequency theorem and induction to show that < + /-!> ! i 1 _v(/i) = --------------- a u(n} - * \ Uo) ----------------/i!(f-l)! (1 - f 4.17 Let vUil he an arbitrary signal, not necessarily real-valued, with Fourier transform Express the Fourier transforms of the following signals in terms of X(a>). (a) _v i)
(h) x ' ( - n )
(f) yin) = vf/1 ) - ,v(/; - 1 ) (d) v(fi I = ' y xik i
(L ) V|H)=.V(2) _ I x(n/2). n even (0 v(n) = L ,. ( 0. n odd 4.18 Determ ine and sketch the Fourier transforms Xiiw), X 2(co). and Xi(a>) of the following signals. (a) -V!(fi) = (1. 1. 1. 1.1) (b) x2(n) = ( 1 . 0. 1 . 0. 1 , 0. 1 . 0, 1 ) (c) x: j n ) = ( 1 . 0 . 0. 1 . 0 . 0 . 1 . 0 . 0 , 1 , 0 . 0 . 1 )
t
(d) Is there any relation between Xi(w). X:(g;). and X3(w)? What is its physical meaning? (e) Show that if **() = then
, ( r ). 0.
if n j k integer otherwise
Xt(a>) = X (ka> )
4.19 Let x(n) be a signal with Fourier transform as shown in Fig. P4.19. D eterm ine and sketch the Fourier transform s of the following signals.
374
Chap. 4
Figure P4.19 (b ) A;(n) = x(n) sin(jrn/2) (d) a4(/7) = x(n) cos nn
(a ) jri(fl) = jr(n) cos(;rn/4) (c) xy(n) = x( n)cos(nn/2)
Note that these signal sequences are obtained by amplitude modulation of a carrier co s wrn or sinw,,/! by the sequence x(n). 4.20 Consider an aperiodic signal *(n) with Fourier transform X(u>). Show that the Fourier series coefficients CA ' of the periodic signal
are eiven bv
c;
4.21 Prove that
I . N
k N
k = 0. 1........A '- l
X v (w ) ^ '
' n=~ N
sin w,n nn
may be expressed as
X v (o j )
1 = ^
.L ,
r sinf( 2 A ^ + 1 ){w 8f lj \ db - 0)/2\
4.22 A signal jfn ) has the following Fourier transform:

X (oj) = ---------------
1 - ae~JW
Determine the Fourier transforms of the following signals: (b) e*nf2x(n + 2 ) (b) x { 2n) (d) x(n) cos( 0.37rn) (c) jt() * x(n 1 ) (0 x(n) * x ( n) 4.23 From a discrete-time signal x ( n ) with Fourier transform X(cv), shown in Fig. P4.23, determine and sketch the Fourier transform of the following signals: , 1 x(n). n even ( a )v i( n ) = I 0, n odd (b) yiiri) - x(2n) [ x( nj 2 ). n even (c) y,() = ,, I 0, n odd Note that vj (n) = x(n)s(n), where s(n) = {... 0, 1. 0, 1. 0. 1. 0. 1. ...} t
(a) x(2n + \ )
Chap. 4
Problems Xlw)
375
ai
4
4
Figure P4.23
4.24 The following inpul-output pairs have been observed during the operation of various svstems:
V(IJ) = _V' " ' (C) .t(H) C --- * (d) .v(u) = e n u(ii) yin) j r ' " 7* (e) ,v(/i = ,v(/i + A ri ) vin) = yin 4- N 21 A', A;. A't. N2 prime Determ ine their frequency response if each of the above systems is LTI. 4.25 (a) D eterm ine and sketch the Fourier transform W/ffw) of the rectangular sequence [I. 1 0. (b) Consider the triangular sequence ( ) < / i < M/2 Mj2 < 2 < M otherwise Determ ine and sketch the Fourier transform Wr (a)) of u'70 )) by expressing it as the convolution of a rectangular sequence with itself. (c) Consider the sequence , in) = ^ (l -t- cos ^f-) u'R{n) Determ ine and skctch Wricu) by using 4.26 Consider an LTI system with impulse response h i n ) = u(n). (a) Determ ine and skctch the magnitude and phase response \H(co)\ and respectively. (b) D eterm ine and sketch the magnitude and phase spectra for the input and output signals for the following inputs: (I < n < M otherwise
a s ,
(2) jr(n) = ( . . . . 1.0.0. 1. 1. 1.0. 1. 1. 1,0, 1. ...}
r
4.27 D eterm ine and sketch the magnitude and phase response of the following systems: (a) y(n) = ^[.v(/i) + xin - 1 )] (b) yfn) = 4 [v(/i) - xin - 1 )] (c) v(n) = + 1) - x(n - 1)]
376 (d) y(n) = (f) y(n) = (g) (h) (i) (j) (k) (I) (m) (n)
Frequency Analysis of Signals and Systems + 1 ) + x(n - 1 )]

jc (/t - 2 ) ]
Chap. 4
(e) v (n )= i[jr(rt) +
- x(n - 2 )]
v(n) = j[jr(n) + x(n - 1 ) + x{n - 2 )] y(n) = x(n) x(n 8) y(n) = 2x(n 1 ) x(n 2 ) v(n) = + x(n - 1) + x(n - 2) + x( - 3)] y() y(n) y(n) y(n) = + 3x(n - 1) + 3x(n - 2) + x(n - 3)] = x(n - 4) = x(n + 4) = j[x(n) - 2x(n - 1 ) + xin - 2 )]
4.28 An FIR filter is described by the difference equation y(n) = x(n) + xin - 10 ) (a) Compute and sketch its magnitude and phase response. (b) Determ ine its response to the inputs (1) x(n) = cos n + 3sm \ ^- n + ^ (2 ) jr(n) = 10 + 5 cos ( ^ - n + y
TT . / TT IT '
oc < n < oc O O< < oc
4.29 D eterm ine the transient and steadv-statc responses of the FIR filler shown in Fig. P4.29 to the input signal A(n) = 10ejr"',~uin). Let b = 2 and v ( - l ) = v (-2 ) = v( 3) = v( 4) = 0.
Figure P4.29 4.30 Consider the FIR filter v(n) = x ( n ) + x ( n - 4) (a) Compute and sketch its m agnitude and phase response. (b) Compute its response to the input x (n) = cos n + cos n 2 4 oc < n < oc
(c) Explain the results obtained in part (b) in terms of the m agnitude and phase responses obtained in part (a).
4.31 Determ ine the steady-state and transient responses of the system
y(n) = j [*() x ( n - 2)]
Chap. 4
Problems
377
to the input signal x(n) = 5 + 3 cos + 60^ oo < n < oc
4.32 From our discussions it is apparent that an LTI system cannot produce frequencies at its output that are different from those applied in its input. Thus, if a system creates new" frequencies, it must be nonlinear and/or time varying. D eterm ine the frequency content of the outputs of the following systems to the input signal 4 (a) v (/i) = .v(2 n) (b) y(n) = x 2(n) <c) y(n) = (cos nn)x(n) 4.33 D eterm ine and sketch the m agnitude and phase response of the systems shown in Fig. P4.33(a) through (c).
(a)
(b) l
8
(c) Figure P4J3 4.34 Determ ine the magnitude and phase response of the m ultipath channel y( n ) = x (n) + xfn M) A t what frequencies does H(co) = 0? 4.35 Consider the filter v() = 0.9v(n - 1) + bx(n ) (a) D eterm ine b so that |H (0)| = 1. (b) D eterm ine the frequency at which j/ / (cu)| = l/%/2.
378
Frequency Analysis of Signals and Systems (c) Is this filter lowpass, bandpass, or highpass? (d) Repeat parts (b) and (c) for the filter v(n) = 0.9y(n - 1) + O.l.r(n).
Chap. 4
4.36* Harmonic distortion in digital sinusoidal generators An ideal sinusoidal generator produces the signal x(n) = cos 2nf)n oc < n < cc which is periodic with fundam ental period N if /<i = ki,/N and ky, N are relatively prime numbers. The spectrum of such a pure" sinusoid consist of two lines at k = ko and k = N ka (we limit ourselves in the fundam ental interval 0 < k < N - 1). In practice, the approximations made in com puting the samples of a sinusoid of relative frequency f t result in a certain am ount of power falling into other frequencies. This spurious power results in distortion, which is referred to as harmonic distortion. Harmonic distortion is usually m easured in terms of the total harmonic distortion (TH D), which is defined as the ratio THD = (a) Show that iQn t2 THD = = 1 - 2 P, where spurious harmonic power total power
*-j
(b) By using the Taylor approximation
n = (I
CO S 0 = 1 -------- --------- 2! 4! 6! compute one period of jr(n) for / 0 = 1/96, 1/32, 1/256 by increasing the number of terms in the Taylor expansion from 2 to 8. (c) Compute the THD and plot the power density spectrum for each sinusoid in part (b) as well as for the sinusoids obtained using the com puter cosine function. Comment on the results.
4.37* Measurement o f the total harmonic distortion in quantized sinusoids
Let x(n) be a
periodic sinusoidal signal with frequency fo = k / N , that is, x(n) = sin 2 jr/on (a) Write a computer program that quantizes the signal x(n) into b bits or equivalently into L = 2h levels by using rounding. The resulting signal is denoted by xq(n). (b) For f a = 1/50 compute the THD of the quantized signals xq(n) obtained by using b = 4, 6, 8, and 16 bits. (c) Repeat part (b) for / = 1/100. (d) Comment on the results obtained in parts (b) and (c). 438* Consider the discrete-time system y(n) = ay(n 1 ) + (1 a)x(n) where a 0.9 and y( 1) = 0. n > 0
Chap. 4
Problems
379
(a) Compute and sketch the output y, (n) of the system to the input signals
.v ,(n ) = s in 2 jT /,n 0 ft < 100
where / , = ./? = (b) Compute and sketch the magnitude and phase response of the system and use these results to explain the response of the system to the signals given in part (a). 4.39* Consider an LTI system with impulse response h{n) = ( t )1"1 (a) Determ ine and sketch the magnitude and phase response Hi m) and respectively. (b) D eterm ine and sketch the magnitude and phase spectra for the input and output signals for the following inputs: ( 1 ) .v(n ) = cos
3,t ii
. rc < n < cc T
(2) .xin) = (....-1, 1. - 1 . 1 . - 1 . 1 . - 1 . 1 . - I . 1, - 1 . 1...-I
4.40* Time-domain sampling
Consider the continuous-time signal
(a) (b) (c) (d) (e)
Compute analytically the spectrum X(F) of a U ) Compute analytically the spectrum of the signal a (/?> = x{nT). T = 1 /F ,. Plot the magnitude spectrum |X (F)| for Ft, = K) Hz. Plot the magnitude spectrum (X(F)I for F, = 10. 20, 40. and 100 Hz. Explain the results obtained in part (d) in terms ol the aliasing effect.
4.41 Consider (he digital filter shown in Fig. P4.41. (a) Determ ine the input-output relation and the impulse response h(n). (b) Determ ine and sketch the magnitude ]W(w)| and the phase response 2^ H ( a of the filter and find which frequencies are completely blocked by the filter. (c) When w,, = t / 2 , determ ine the output yin) to the input .v(n) = 3 cos + 30 ^ cc < n < oc
-*
1 ---- ( +
y(n>
u - - 2 cos aif, 4.42 Consider the FIR filter
Figure P4.41
y(n) = xin) - xin - 4 ) (a) Compute and sketch its m agnitude and phase response. (b) Com pute its response to the input n n x ( n ) = cos n + cos n c c < n < oc 2 4
380
Chap. 4
(c) Explain the results obtained in part (b) in terms of the answer given in part (a). 4.43 Determ ine the steady-state response of the system y(n) = i[.v(n) - x(n - 2 )] to the input signal oc < n < oc 4.44 Recall from Problem 4.32 that an LTI system cannot produce frequencies at its output that are different from those applied in its input. Thus if a system creates new frequencies, it must be nonlinear and/or time varying. Indicate w hether the following systems are nonlinear and/or time varying and determ ine the output spectra when the input spectrum is
(a) _v(n) = j:(2nt (b) y ( ) = x : (n) (c) v(n) = (cos nn)xin) 4.45 Consider an LTI svstem with impulse response
(a) Determ ine its system function H(z). (b) Is it possible to implement this system using a finite number of adders, multipliers, and unit delays? If yes. how? (c) Provide a rough sketch of \ H(w )| using the pole-zero plot. (d) Determine the response of the system to the input x(n) = ( 4.46 An FIR filter is described by the difference equation yin) = x(n) xin 10 ) (a) Compute and sketch its m agnitude and phase response. (b) Determ ine its response to the inputs oc < n < oc oc < n < oc4.47 The frequency response of an ideal bandpass filter is given by
Chap. 4
Problems
381
(a) D eterm ine its impulse response (b) Show that this impulse response can be expressed as the product of cos(n 7t / 4 ) and the impulse response of a lowpass filter. 4.48 Consider the system described by the difference equation v(n) = jy(n I ) + .r(n) 4- |.r(n - 1 > (a) D eterm ine its impulse response. (b) D eterm ine its frequency response: (1) From the impulse response (2) From the difference equation (c) D eterm ine its response to the input x {n > = cos y r +
I TT .T \
- oc < n < oc
4.49 Sketch roughly the m agnitude |A"u<>)! of the Fourier transforms corresponding to the pole-zero patterns given in Fig. P4.49.
Figure P4.49 4.50 Design an F IR filter that completely blocks the frequency a*, = rr/4 and then compute its output if the input is
x(n) = ^sin
u(n)
for n = 0, 1 , 2, 3, 4. Does the filter fulfill your expectations? Explain.
382
Chap. 4
4.51 A digital filter is characterized by the following properties: (1) It is highpass and has one pole and one zero. (2) The pole is at a distance r = 0.9 from the origin of the ; -plane. (3) Constant signals do not pass through the system. (a) Plot the pole-zero pattern of the filtei and determ ine its system function H(z). (b) Compute the m agnitude response \H(cu}\ and the phase response ^ H( t n) of the filter, (c) Normalize the frequency response H( cd) so that = 1. (d) Determ ine the input-output relation (difference equation) of the filter in the time domain. (e) Compute the output of the system if the input is
ex: < n < oc
(You can use either algebraic or geom etncal arguments.) 4.52 A causal first-order digital filter is described by the system function
(a) Sketch the direct form I and direct form II realizations of this filter and find the corresponding difference equations. (b) For a = 0.5 and b 0.6, sketch the pole-zero pattern, is the system stable? Why? (c) For a = 0.5 and b = 0.5, determ ine bu. so that the maximum value of \H(w)\ is equal to 1 . (d) Sketch the m agnitude response \H{co)\ and the phase response 2t//(o>) of the filter obtained in part (c). (e) In a specific application it is known that a 0.8. Does the resulting filter amplify high frequencies or low frequencies in the input? Choose the value of b so as to improve the characteristics of this filter (i.e., m ake it a better lowpass or a better highpass filter). 4.53 Derive the expression for the resonant frequency of a two-pole filter with poles at Pi = r e 6 and p 2 = p*x. given by (4.5.25). 4.54 Determ ine and sketch the magnitude and phase responses of the Hanning filter char acterized by the (moving average) difference equation y(n) ~ jx (n ) + \ x( n - 1 ) + j*( - 2 ) 4.55 A causal LTI system excited by the input x(rt') = (j)"(n) + u( n - 1) produces an output v(n) with r-transform
_ 3 -l
(a) D eterm ine the system function H(z) and its ROC. (b) Determ ine the output y(n) of the system. (Hint: Pole cancellation increases the original ROC.)
Chap. 4
Problems
383
4.56 Determ ine the coefficients of a linear-phase FIR filter v(n) = b^x(n) + btx(n 1) + thx(n 2) such that: (a) It rejects completely a frequency com ponent at a* = 2jt/3. (b) Its frequency response is normalized so that H{0) ~ 1. (c) Compute and sketch the magnitude and phase response of the filter to check if it satisfies the requirements. 4.57 D eterm ine the frequency response H(w) of the following moving average filters. v(") = 2F
t t
> - * ) J x(n - k) + ~r~rzx(n - M) AM k=- M + i Y
1 1 (b) y(ff) = r-x(n + Af) + AM 2M
Which filter provides better smoothing? Why? 4.58 The convolution *(r) of two continuous-time signals *((/) and x2(t), from which at least one is nonperiodic, is defined by x(t) = X] (/) * xj(t) j x\ {X)x2(t k)dk
(a) Show that X(F) = X\ ( F) X2(F), where X, (F) and X 2(F) are the spectra of *1 (r) and a:(/), respectively. Com pute .r(r) if jrt(r) = = ( I' < X p' 10. elsewhere (c) Determ ine the spectrum of x(t) using the results in part (a).
4.59 Com pute the magnitude and phase response of a filter with system function H(z) = l + c If the sampling frequency is F, = 1 kHz, determ ine the frequencies of the analog sinusoids that cannot pass through the filter. 4.60 A second-order system has a double pole at p li2 = 0.5 and two zeros at Z\.2 = es3*,A Using geom etric arguments, choose the gain G of the filter so that |tf(0 )| = 1. 4.61 In this problem we consider the effect of a single zero on the frequency response of a system. Let ; = re9 be a zero inside the unit circle (r < 1). Then H,(w) = 1 re]0e~iu > = 1 r cos(a) B) + j r sin(ti) (?) (a) Show that the m agnitude response is \HZ((D )\ = [ 1 - 2 r cosito - 6 ) + r 2f n
or, equivalently,
201og10 |//:(cd)| = 101og10[l 2r cos{ct> 6) + r2]
384
Frequency Analysis of Signals and Systems (b) Show that the phase response is given as 0 : (ct>) = tan -i rsin(w - 8) ------------------1 - r cos(ct) -
Chap. 4
(c) Show that the group delay is given as r - rcos(w - 8) 1 + r1 2r COS(oj (d) Plot the magnitude ()]dB. the phase ( oj) and the group delay and 0 = 0, n j 2, and n. for r = 0.7
4.62 In this problem we consider the effect of a single pole on the frequency response of a system. Hence, we let '<> = Show that [Wr (a>)ljB = ~ |W-(fi))|dD
4 // (u > ) = - Z - H . ( c o )
' =1
where H:(< x > ) and
are defined in Problem 4.61.
4.63 In this problem we consider the effect of complex-conjugate pair of poles and zeros on the frequency response of a system. Let H.{u>) = (1 - reJ*'e~J'")(l - re~il'e~-") (a) Show that the magnitude response in decibels is
| W - M | d B 1 0 1 o g 1()[ l + 2 r c o s ( o ) # ) ]
+ 101og,[l + r 2 - 2r cos(a) + 6)] (b) Show that the phase response is given as : (co) = tan ' rsinUo 8) l-r c o s (tu -# ) _ Asin(w + fi) l- r c o s ( o > + 0)
(c) Show that the group delay is given as r2 r cos(o> 8) r2 r c o s ( c j + 8) = r ------ + 1 + r 2 - 2r cos(to 9 ) 1 4- r2 - 2r cos(a> + (d) If Hp(w) = l / H : (a> ), show that
&p(w) - - .(oj) Tg(a>) = (e) Plot p(ci>) and (w) for r = 0.9, and 8 0, jt/2.
Chap. 4
Problems
385
4.64 D e te rm in e th e 3-dB b a n d w id th o f the filte rs (0 < a < 1)
Which is a better lowpass filter? 4.65 Design a digital oscillator with adjustable phase, that is. a digital filter which produces the signal
y(n) = cos(<W|, + ft)u(n)
4.66 This problem provides another derivation of the structure for the coupled-form os cillator by considering the system
y i n ) = av(ri 1 ) 4 xi / i )
for a e' ". Let xi/i) be real. Then yi n} is complex. Thus

yin) =
+ ./'/()
(a) Determ ine the equations describing a system with one input x i n) and the two outputs y t i U t ) and y f Ui ) . (b) Determ ine a block diagram realization (c) Show that if ,v() = bin), then yîn) = (cosw()'i ii/tn I y/ {ti ) = (sin mull )u(n) (d) Compute ynin), \/in). n = 0. 1....... 9 for & j(, = tt/6. Compare these with the true values of the sine and cosine, 4.67 Consider a filter with system function H(Z) = h\ --------: ------j --- ;-------------- r( 1 reJ U I"z )(1 re~,w"z ) (a) Sketch the pole-zero pattern. <b) Using geom etric arguments, show that for r ^ 1, the system is a notch filter and provide a rough sketch of its m agnitude response if ttx> = 60 . (c) For (i> n = 60 , choose bo so that the maximum value of \Hia>)\ is 1. (d) Draw a direct form II realization of the system (e) D eterm ine the approximate 3-dB bandwidth of the system. 4.68* Design an FIR digital filter that will reject a very strong 60-Hz sinusoidal interference contaminating a 200-Hz useful sinusoidal signal. Determ ine the gain of the filter so that the useful signal does not change amplitude. The filter works at a sampling frequency F, = 500 samples/s. Compute the output of the filter if the input is a 60-Hz sinusoid or a 200-Hz sinusoid with unit amplitude. How does the perform ance of the filter compare with your requirem ents? 4.69 Determ ine the gain bo for the digital resonator described by (4.5.28) so that I//(n) I = 1.
(1 1 -
386
Chap. 4
4.70 D em onstrate that the difference equation given in (4.5.52) can be obtained by apply ing the trigonometric identity a + [i a cos or + cos p = 2 cos - cos - where a (n-t-Dwu, ~ (n l)a>o, and v(n) = coswo n. Thus show that the sinusoidal signal x(n) A co s co^n can be generated from (4.5.52) by use of the initial conditions v(1) = A cos an) and y ( 2) = j4cos2a>o. 4.71 Use the trigonometric identity in (4.5.53) with a = na> o and (i = (n 2)a^t to derive the difference equation for generating the sinusoidal signal y(n) = A sin no*,. Determine the corresponding initial conditions. 4.72 Using the --transform pairs 8 and 9 in Table 3.3. determine the difference equations for the digital oscillators that have impulse responses h(n) = A cosna>it(n) and h(n) = A sin ncDf>u(n), respectively. 4.73 Determ ine the structure for the coupled-form oscillator by combining the structure for the digital oscillators obtained in Problem 4.72. 4.74 Convert the highpass filter with system function H(z) = a < 1
az
into a notch filter that rejects the frequency a*, = tt/ 4 and its harmonics. (a) Determ ine the difference equation. (b) Sketch the pole-zero pattern. (c) Sketch the magnitude response for both filters. 4.75 Choose L and M for a lunar niter that must have narrow passbands at (k AF) cycles/dav. where k = i. 2, 3 ,... and A F = 0.067726. 4.76 (a) Show that the systems corresponding to the pole-zero patterns of Fig. 4.58 are all-pass. (b) What is the number of delays and multipliers required for the efficient implemen tation of a second-order all-pass system? 4.77 A digital notch filter is required to remove an undesirable 60-Hz hum associated with a power supply in an ECG recording application. The sampling frequency used is Fs = 500 samples/s. (a) Design a second-order FIR notch filter and (b) a secondorder pole-zero notch filter for this purpose. In both cases choose the gain by so that \H(w)\ = 1 for w = 0. 4.78 Determ ine the coefficients {/?()} of a highpass linear phase FIR filter of length M = 4 which has an antisymmetric unit sample response h{n) = h(M - I - n) and a frequency response that satisfies the condition
4.79 In an attem pt to design a four-pole bandpass digital filter with desired magnitude response
0,
elsewhere
Chap. 4
Problems
387
we select the four poles at
/>,, = 0 .8 e-MT" and four zeros at
(a) Determ ine the value of the gain so that
HDb
(b) Determ ine the system function H(:).
(c) Determ ine the magnitude of the frequency response H(a>) for 0 < a> < tt and compare it with the desired response \Hd(w)[. 4.80 A discrete*time system with input and output v ( n ) is described in the frequency domain by the relation
V(oj) v
dXiuj) ~ 1 X (as) -t-

das
(a) Compute the response of the svstem to the input x i n ) =

(b ) C heck if the system is LTI and stable.
4.81 Consider an ideal lowpass filter with impulse response It in) and frequency response Hiw) = |
I 1. K"! 5
C IJ, <
a),
|< D | < TT
I 0,
What is the frequency response of the iiher defined by X<n) =

h , n ev e n
f).
n odd
4.82 Consider the system shown in Fig. P4.S2. Determ ine its impulse response and its frequency response if the system H(a>) is: (a) Lowpass with cutoff frequency w,. (b) Highpass with cutoff frequency o j , .
xi n ) Hiw)
I'
4.83 Frequency inverters have been used for many years for speech scrambling. Indeed, a voice signal x(n) becomes unintelligible if we invert its spectrum as shown in Fig. P4.S3. (a) Determ ine how' frequency inversion can be perform ed in the time domain. (b) Design an unscrambler. (Hint: The required operations are very simple and can easily be done in real time.)
388 X(a> )
Chap. 4
-x
0 (b)
Figure P4.83
(a) Original spectrum; (b) frequency-inverted spectrum.
4.84 A lowpass filter is described by the difference equation v(n) = 0.9v(/i - 1) + 0.1x(n) (a) By performing a frequency translation of n f l . transform the filter into a bandpass filter. (b) What is the impulse response of the bandpass filter? (c) What is the major problem with the frequency translation m ethod for transform ing a prototype lowpass filter into a bandpass filter? 4.85 Consider a system with a real-valued impulse response h(n) and frequency response H(w) = \H (a>)\em -' The quantity
D = 2 2 2^ ( n )
n = ~ oc
provides a m easure of the effective duration of h(n). (a) Express D in terms of H(w). (b) Show that D is minimized for 0(a>) = 0. 4.86 Consider the lowpass filter
v(n) = a y { n 1) + fejr(n) 0 < a < 1
(a) (b) (c) <d)
D eterm ine b so that |//{0)| = 1. D eterm ine the 3-dB bandwidth an, for the normalized filter in part (a). How does the choice of the param eter a affect uyf! Repeat parts (a) through (c) for the highpass filter obtained by choosing - 1 <
a < 0.
4.87 Sketch the magnitude and phase response of the m ultipath channel
y (n ) = x ( n ) + a jr(n M ) fo r a < < 1.
a > 0
Chap. 4
Problems
389
4.88 Determ ine the system functions and the pole-zero locations for the systems shown in Fig. P4.88(a) through (c). and indicate w hether or not the systems are stable. l
lc) Figure P4.88
4.89 D eterm ine and sketch the impulse response and the magnitude and phase responses of the FIR filter shown in Fig. P4.89 for b = 1 and b = -1 .
4.90 Consider the system v(n) = x(n) - 0.95.r(fl - 6) (a) (b) (c) (d) Sketch its pole-zero pattern. Sketch its magnitude response using the pole-zero plot. D eterm ine the system function of its causal inverse system. Sketch the magnitude response of the inverse system using the pole-zero plot.
4.91 Determ ine the impulse response and the difference equation for all possible systems specified by the system functions
390
Chap. 4
(a)
H( Z)
1 _ rZ _'_
,_2
0< < 1
0 " W = ! _ eL z-4
4.92 D eterm ine the impulse response of a causal LTI system which produces the response v(n) = {1. - 1 .3 . - 1 ,6 ) t when excited by the input signal x(n) = (1,1,2} t 4.93 The system y(n) = i_v(n - 1) + x(n) is excited with the input x(n) = (j)"(n) Determ ine the sequences aj t (/), rhh(l), r,y(i). and rvy(l). 4.94 Determ ine if the following FIR systems are minimum phase. (a) h(n) = {10,9. - 7 , - 8 , 0 , 5 . 3} r (b) h(n) = (5,4, - 3 . - 4 , 0 ,2 ,1 ) t 4.95 Can you determ ine the coefficients of the all-pole svstem
*=i if you know its order N and the values /?(0), /i(l)........ h ( L - l ) of its impulse response? How? W hat happens if you do not know N1 4.96 Consider a system with impulse response h(n) = baS(n) + bi$(n D) + ^<5(n 2D)
(a ) Explain why the system generates echoes spaced D samples apart.
(b) Determ ine the m agnitude and phase response of the system. (c) Show that for \b0 + b2\ < < |>]|, the locations of maxima and minima of \H(a>) are at k w = n k = 0. 1, 2. . . . D (d) Plot |//(w )| and for b$ 0.1, b\ = 1, and b2 = 0.05 and discuss the results. 4.97 Consider the pole-zero system B(z) 1 + bz * v ' W(z) = A(z) = T r=Y 1T + -----az 1 l h^ z (a) Determine >i(0), >i(l), h(2), and h( 3) in terms of a and b. (b) Let rhH (l) be the autocorrelation sequence of h(n). D eterm ine rkh{0), /-^(l), rw,(2), and r/,h(3) in terms of a and b.
Chap. 4
Problems
391
4.98 Let be a real-valued minimum-phase sequence. Modify xin I to obtain another real-valued minimum-phase sequence y(/i) such that y(0) = x(0) and y(n) = |jr(n)|. 4.99 The frequency response of a stable LTI system is known to be real and even. Is the inverse system stable? 4.100 Let h(n) be a real filter with nonzero linear or nonlinear phase response. Show that the following operations are equivalent to filtering the signal x(n) with a zero-phase filter. (a) g(n) Ji(n) * ,v(n)
/ (h ) = h (/ t ] * g t - i i )
yon = f( ~n)
(b) g(n) h(n) * / (r?) = h{n) * x ( n)
y(nl = ain) + / ( - i i )
( Hint: D eterm ine the frequency response of the composite system _v(/i) = //[.v()].) 4.101 Cheek the validity of the following statements: (a) The convolution of two minimum-phase sequences is always mimmum-phase se quence. (h) The sum of two minimum-phase sequences is always minimum phase. 4.102 Determ ine the minimum-phase syslem whose squared magnitude response is given by: (a) i H (</)}'- = - cos w ^
(b )
I H (u>)',: =
(i
a - ) - 2 a cos w
4.103 Consider an FIR syslem with the following system function: H(z) = (1 - 0 . 8 r '7V ' K l - l..v " T,V 1HI - L S ^ ' V 1) Which is the
(a) Determ ine all systems that have the same magnitude response. minimum-phase system? (b) Deierm ine the impulse response of al) systems in part (a). (c) Plot the partial energy
i=t> for every system and use il lo identify the minimum- and maximum-phase systems. 4.104 The causal system //(.-) = ,v
392
Frequency Analysis of Signals and Systems (a) Show that by properly choosing A we can obtain a new stable system. (b) What is the difference equation describing the new system?
Chap, 4
4.105* Given a signal x(n), we can create echoes and reverberations by delaying and scaling the signal as follows > () = 22 Skx (n ~ kD) where D is positive integer and gk > g i i > 0. (a) Explain why the comb filter H(z) =
1-az-
can be used as a reverberator (i.e.. as a device to produce artificial reverberations). (Hint: Determ ine and sketch its impulse response.) (b) The all-pass comb filter H(z) = -a 1 - az-1
is used in practice to build digital reverberators by cascading three to five such filters and properly choosing the param eters a and D. Com pute and plot the impulse response of two such reverberators each obtained by cascading three sections with the following parameters. UNIT 1
Seclion 1 2 3 I) 50 40 32 a 0.7 0.665 0.63175 Section 1 2 3
UNIT 2
D 50 17 6 a 0.7 0.77 0.847
(c) The difference between echo and reverberation is that with pure echo there are clear repetitions of the signal, but with reverberations, there are not. How is this reflected in the shape of the impulse response of the reverberator? Which unit in part (b) is a better reverberator? (d) If the delays D\, D2, Dj in a certain unit are prime numbers, the impulse response of the unit is more dense." Explain why. (e) Plot the phase response of units 1 and 2 and comm ent on them. (f) Plot h(n) for D |, D2, and being nonprime. W hat do you notice? More details about this application can be found in a paper by J. A. M oorer, Signal Processing Aspects of Com puter Music: A Survey," Proc, IEEE, vol. 65, No. 8, Aug. 1977, pp. 1108-1137. 4.106* By trial-and-error design a third-order lowpass filter with cutoff frequency at wc = Jt/9 radians/sample interval. Start your search with (a) zi = Z2 = Z3 = 0, pi = r, p2J = re,tu' . r = 0.8 (b) r = 0.9, zi = Z2 = Z3 = - 1
Chap. 4
Problems
393
4.107* A speech signal with bandwidth B = 10 kHz is sampled at F2 = 20 kHz. Suppose that the signal is corrupted by four sinusoids with frequencies F, = 10, 000 Hz. F2 = 8889 Hz, F3 = 7778 Hz F4 = 6667 Hz
(a) Design a FIR filter that eliminates these frequency components. (b) Choose the gain of the filter so that |H (0)| = 1 and then plot the log m agnitude response and the phase response of the filter. (c) Does the filter fulfill your objectives? Do you recom m end the use of this filter in a practical application? 4.108* Com pute and sketch the frequency response of a digital resonator with co = t t / 6 and r = 0.6, 0.9, 0.99. In each case, compute the bandwidth and the resonance frequency from the graph, and check if they are in agreem ent with the theoretical results. 4.109* The system function of a communication channel is given by H(z) = (1 - 0 . 9 f - 'u4'T;;-1)(l - 1.5^'afor; - 1)(l - 1.5<- '" tez - 1)
D eterm ine the system function Ht.(;) of a causal and stable compensating system so that the cascade interconnection of the two systems has a flat magnitude response. Sketch the pole-zero plots and the m agnitude and phase responses of all systems in volved into the analysis process. [Hint: Use the decomposition H(z) = W ap(;) Wn,in(;)-]
The Discrete Fourier Transform: Its Properties and Applications
F requency analysis o f discrete-tim e signals is usually and most con ven ien tly per form ed on a digital signal processor, which may be a general-purpose digital com puter or specially d esigned digital hardware. T o perform frequency analysis on a d iscrete-tim e signal w e convert the tim e-dom ain sequ en ce to an equivalent frequencv-dom ain representation. We know that such a representation is given by the Fourier transform X(cu) of the seq u en ce |a(>;)). H ow ever, A'u<j) is a contin uous function o f frequency and therefore, it is not a com p utationally convenient representation o f the sequence {.v(/i)). In this section we consider the representation o f a sequ en ce by sam ples o f its spectrum X ( uj). Such a frequency-dom ain representation lead s to the discrete Fourier transform (D F T ), which is a pow erful com putational tool for perform ing frequency analysis o f discrete-tim e signals.
5.1 FREQUENCY DOMAIN SAMPLING: THE DISCRETE FOURIER TRANSFORM

B efore w e introduce the D F T , w e consider the sam pling of the F ourier transform of an aperiodic discrete-tim e sequ en ce. Thus, w e establish the relationship betw een the sam pled Fourier transform and the D FT.
5.1.1 Frequency-Domain Sampling and Reconstruction of Discrete-Time Signals

W e recall that aperiodic finite-energy signals have con tinu ou s spectra. Let us consider such an aperiodic discrete-tim e signal x ( n ) with Fourier transform
(5.1.1)
394
Sec. 5.1
Frequency Domain Sampling: The Discrete Fourier Transform
395
S uppose that w e sam ple X(a>) periodically in frequency at a spacing o f Sco radians b etw een successive sam ples. Since X(a>) is period ic with period 2 tt, only sam ples in the fundam ental frequency range are necessary. For con ven ien ce, w e take N equidistant sam ples in the interval 0 < w < 2tt with spacing Sco = 2 ttf N , as shown in Fig. 5.1. First, w e consider the selection of N , the num ber o f sam ples in the frequency dom ain. If w e evaluate (5.1.1) at u> = 2 n k / N , w e obtain (5.1.2) T h e sum m ation in (5.1.2) can be subdivided in to an infinite n um ber of sum m ations, where each sum con tains N term s. Thus -i
N-1
oc
IN + N - 1
-)2nkn/\
}=-~yL f}=i A
If we change the index in the inner sum m ation from n to n I N and interchange the order o f the sum m ation, we obtain the result (5.1.3) for k = 0, 1, 2 ......... N 1. T h e signal
OC
(5.1.4) ob tain ed by the periodic repetition o f x{n) every N sam ples, is clearly periodic with fundam ental period N. C onsequently, it can be expanded in a Fourier
0 F igure 5.1
kSa)
tt
Swh
Frequency-dom ain sam pling of the F ourier transform .
396 series as
Chap. 5
A '-l
x p (n) = Y ^ c ke,2nkn/N
k=Q
n = 0 , 1 .........N - 1
(5.1.5)
with Fourier coefficients

l A '-l
ck = - - ' x p ( n ) e - j2,rkn/N N n=0
k = 0 , 1 ........ N 1
(5.1.6)
U p on com paring (5.1.3) with (5.1.6), we conclude that 1 f 2* \ ck = X l k ) N \ N J T herefore, 1 ^ ~ ^ / ^7T \ x.,(n) = ' y * ( * ) e llnkntN \ N / n = 0, 1........ A ' - l (5.1.8) k = 0 . 1.........A ' - l (5.1.7) J
The relationship in (5.1.8) provides the reconstruction of the periodic signal x r (u) from the sam ples of the spectrum X ( oj). H ow ever, it d o es not imply that we can recover X(u>) or x{n) from the sam ples. T o accom plish this, we need to consider the relationship betw een x p (n) and j:(). Since x p(n) is the periodic exten sion of x{n) as given by (5.1.4). it is clear that x (/ j) can be recovered from x p (n) if there is no aliasing in the tim e domain, that is, if x( n) is tim e-lim ited to less than the period N of x p(n). This situation is illustrated in Fig. 5.2, w here without loss of generality, we consider a finite-duration rtn)
H it t ,....
L
N > L
xp(n)
ITI
i t t t t
IITI
0
t t t
.
L
JTTTTttt
N
N< L
IT- I TNt t t O I TN T t
TTTt 1-
Figure 5.2 Aperiodic sequence x(n) of length L and its periodic extension for N > L (no aliasing) and N < L (aliasing).
Sec. 5.1
397
sequ en ce x( n) , which is n on zero in the interval 0 < n < L 1. W e ob serve that when N > L. x(ti) = Xp(n) 0 < n < N 1
so that x ( n ) can be recovered from x r (n) w ithout am biguity. On the other hand, if N < L, it is not possib le to recover from its periodic exten sion due to timed o m a i n aliasing. Thus, w e conclude that the spectrum of an aperiodic discrete-tim e signal with finite duration L . can be exactly recovered from its sam ples at frequen cies tot: = 2j rk/ N. if N > L. The procedure is to com pute x p (n). n = 0, 1.........N - 1 from (5.1.8); then 0 < n < N elsew h ere and finally, X(io) can be com p uted from (5.1.1). A s in the case o f con tinu ou s-tim c signals, it is p ossible to express the spectrum X ( w ) directly in term s o f its sam ples X Q n k / N ) , k = 0. 1........ N 1. T o derive such an interpolation formula for X{co), we assum e that N > L and begin with (5.1.8). Since x( n) = x,,(n) for 0 < /; < A' - 1,
k ]
()<//< N - I
(5.1.1(1)
If w e use (5.1.1) and substitute for v(/;), we obtain

A'-I
X(a>) =
n = (l
A '-l
(5.1.11)
2>
The inner sum m ation term in the brackets o f (5.1.11) represents the basic interpolation function shifted by 2 ttk / N in frequency. Indeed, if w e define v/tuA -' N 1 - e-J sin(o>Ar/2 ) N sin(w /2) jiuiN (5.1.12)
then (5.1.11) can be expressed as
398
Chap. 5
The interpolation function P(co) is not the familiar ( si n 8 ) / 6 but instead, it is a periodic counterpart of it, and it is due to the periodic nature of A(co). The phase shift in (5.1.12) reflects the fact that the signal j:(n) is a causal, finite-duration seq u en ce o f length N. The function sln( a>N/ 2)/ ( N sin(a>/2)) is plotted in Fig. 5.3 for N = 5. We observe that the function P(to) has the property k = * = 1 .2 .........A ' - l
(5.1.14)
C onsequently, the interpolation form ula in (5.1,13) gives exactly the sam ple val ues X ( 2 j r k / N ) for oj = 2 i r k / N . A t all other freq u en cies, the form ula provides a properly w eighted linear com bination o f the original spectral sam ples. T h e follow ing exam ple illustrates the frequency-dom ain sam pling of a discrete-tim e signal and the tim e-dom ain aliasing that results. Example 5.1.1 Consider the sisinal xin ) = a"uin) 0 < a < 1
The spectrum of this signal is sampled at frequencies a > t = l i r k / S . k 0. 1....... A '- l. Determ ine the reconstructed spectra for a = 0.8 when A' = 5 and N = 50.
Solution
The Fourier transform of the sequence x(n) is
Suppose thai we sample X(w) at N equidistant frequencies w *. = 2 x k / N , k = 0, A' - 1. Thus we obtain the spectral samples
Xiw)
1.0
JV= 5
Figure S.3
Plot o f the function
[sin(ct>W/2)]/[jty sin(tu/2)].
Sec. 5.1
399
The periodic sequence xp(n), corresponding to the frequency samples X ( 2 n k / N ) , k = 0, 1 , . . . . N 1, can be obtained from either (5.1.4) or (5.1.8). Hence
oc
0 jc(n - IN) = ^ a~IN
x p(n) = ^
0 <n < N - 1 where the factor 1/(1 - a N) represents the effect of aliasing. Since 0 < a < 1, the aliasing error tends toward zero as N -+ oo. For a = 0.8, the sequence x(n) and its spectrum X{w) are shown in Fig. 5,4a and b, respectively. The aliased sequences xp(n) for N = 5 and N = 50 and the corresponding spectral samples are shown in Fig. 5.4c and d, respectively. We note that the aliasing effects are negligible for N = 50. If we define the aliased finite-duration sequence x(n) as xp{n), 0, then its Fourier transform is X(w) = y ^ i ( n ) e 1 1 a N 1 1 ae~,w 0 < n < N 1 otherwise
Note that although X(ou) ^ X(a>), the sample values at a>t = I n k f N are identical. That is, 1 1- a N 1-a N l-ae-'2 1"
5.1.2 The Discrete Fourier Transform (DFT)

The d ev elo p m en t in the preceding section is concerned with the frequency-dom ain sam pling o f an aperiodic finite-energy sequ en ce j:(n). In general, the equally spaced freq u en cy sam ples X (2 n k / N ) , k = 0 , 1 ___ _ N 1, do n ot uniquely represent the original seq u en ce x ( n ) when x(n) has infinite duration. Instead, the frequency sam ples X ( 2 n k / N ) , k = 0, 1........ N - I, correspond to a period ic seq u en ce x p(n) o f period N , w here x p (n ) is an aliased version o f *{), as indicated by the relation in (5.1.4), that is, (5.1.15) W hen the seq u en ce x ( n ) has a finite duration o f length L < N , then x p{n) is sim ply a periodic repetition o f x ( n ) , w h ere x p (n) over a single period is
400

IAWH
Chap. 5
1.0?
Tfrrmnx. (a)
(b)
x(n) 1.0*
x \ f k)
1 W
50
(1
50
fdi Figure 5.4 (a) Plot of sequence xin) = (0 .X)"h (h ): (b) its Fourier transform (magnitude
only): (c) effect of aliasing with A' = 5: (d) reduced effect of aliasing with A' = 50.
given as x r (n) = x( n) .
0.
0 < n < L 1 L < n < N 1
(5.1.16)
C onsequently, the frequency sam ples X ( 2 i r k / N ) , k = 0. 1.........N 1, uniquely represent the finite-duration seq u en ce *{/;) Since x ( n ) = x p (n) over a single pe riod (padded by N L zeros), the original finite-duration seq u en ce x ( n ) can be obtained from the frequency sam ples \ X ( 2 n k / N \ by m eans o f the formula (5.1.8)It is im portant to n ote that zero p a d d i n g d oes not provide any additional inform ation about the spectrum X(a>) o f the seq u en ce \x(n)}. T he L equidis-
Sec. 5.1
401
tant sam ples o f X(u>) are sufficient to reconstruct X(co) using the reconstruction form ula (5.1.13). H o w ever, padding the seq u en ce {jc( )} with N L zeros and com puting an A'-point D F T results in a better display" of the Fourier transform X(o>). In sum m ary, a finite-duration seq u en ce x( n ) o f length L [i.e., x(n ) = 0 for n < 0 and n > L\ has a Fourier transform J L 1 X(u>) = ^ 2 x ( n ) e ~ Jwn n=0 0 < a> < 2jt (5.1.17)
where the upper and low er indices in the sum m ation reflect the fact that x ( n ) = 0 outside the range 0 < n < L 1. W hen w e sam ple X{a>) at equally spaced freq u en cies u> k = 2 n k / N . k ~ 0, 1, 2 ........ N 1, where N > L. the resultant sam ples are X(k) = X = J 2 x ( n ) e ^ j27,k"/N N_ ) N 7 "= X(k) = y
n=<l
(5.1.18)
x ( n ) e - p^ h,IN
k = 0, 1, 2, , . . , N - 1
where for co n v en ien ce, the upper index in the sum has been increased from L 1 to /V - 1 since x ( n ) = 0 for n > L. T he relation in (5.1.18) is a form ula for transform ing a seq u en ce {jc()} of length L < N in to a seq u en ce o f frequency sam ples ( X()) o f length N . Since the frequency sam ples are ob tain ed by evaluating the Fourier transform X (a>) at a set o f N (eq ually spaced) discrete frequencies, the relation in (5.1.18) is called the discrete Fouri er t rans f or m (D F T ) o f jc(). In turn, the relation given by (5.1.10), which allow s us to recover the seq u en ce jr(n) from the frequency sam ples 1 A'-l x ( n ) = - X f c ) e ,23tknlN n = 0 . 1 .........N ~ 1 (5.1.19)
is called the i nverse D F T (ID F T ). Clearly, when x ( n ) has length L < N , the Ap point ID F T yield s x (n ) = 0 for L < n < N 1. T o sum m arize, the form ulas for the D F T and ID F T are
DFT
A'-l X ( k ) = J ^ x ( n ) e - j2nkn/N ^ 7 0 k = 0 , 1, 2, (5.1.18)
IDFT
, A'-l x( n) = Y * X ( k ) e JZnkn/N N *=o n = 0 , 1 , 2 .........N - 1 (5.1.19)
402
Chap. 5
Example 5.1.2 A finite-duration sequence of length L is given as x(n) I 1 0, 0 < n <L -l otherwise
Determine the /V-point DFT of this sequence for N > L. Solution The Fourier transform of this sequence is -1
L-l
The magnitude and phase of X (w) are illustrated in Fig. 5.5 for L = 10. The Appoint DFT of x (n) is simply X(w) evaluated at the set of N equally spaced frequencies wt = 2it k / N , k = 0, 1........N 1. Hence p-jkrkLiN
sm(*kL/N)
, - j x k l L - 1 I/A/
sm(i t k/ N)
IX(w)l
1"
2
0(a>) T T
7 C
2n
W | H 0
ID
Figure SS Magnitude and phase characteristics of the Fourier transform for signal in Exam ple 5.1.2.
Sec. 5.1
Frequency Domain Sampling: The Discrete Fourier Transform If N is selected such that N = L, then the D FT becomes L, * (* ) = 10. k =0 * = 1 .2 ........L - l
403
Thus there is only one nonzero value in the DFT. This is apparent from obser vation of X(w), since X(a>) = 0 at the frequencies an = I n k f L , k ^ 0. The reader should verify that x(n) can be recovered from X(k) by perform ing an Z.-point IDFT. Although the L-point D FT is sufficient to uniquely represent the sequence x{ n ) in the frequency domain, it is apparent that it does not provide sufficient detail to yield a good picture of the spectral characteristics of x(n). If we wish to have better picture, we must evaluate (interpolate) X(a>) at more closely spaced frequencies, say wt, = 2n k / N , where N > L. In effect, we can view this com putation as expanding the size of the sequence from L points to N points by appending A '- L zeros to the sequence x ( n ). that is, zero padding. Then the A*-point DFT provides finer interpolation than the L-point DFT. Figure 5.6 provides a plot of the Appoint DFT, in magnitude and phase, for L = 10, N = 50, and N = 100. Now the spectral characteristics of the sequence are more clearly evident, as one will conclude by comparing these spectra with the continuous spectrum XUo).
5.1.3 The DFT as a Linear Transformation

T he form ulas for the D F T and ID F T given by (5.1.18) and (5.1.19) m ay be ex pressed as
N -1
X ( k ) = J ^ j r O i) ^ " F T (I ] JV-1 x( n) = J ] x a ) ^ in ^ k=o w here, by definition,
A - = 0 , 1 , . . . , A/ 1
(5.1.20)
n = 0 , 1 .........N - 1
(5.1.21)
W N = e ~ ^ IN
(5.1.22)
which is an N th root o f unity. W e n o te that the com putation o f each point o f th e D F T can be accom plished by N com p lex m ultiplications and ( N ~ 1) com plex additions. H en ce the W-point D F T valu es can b e com p uted in a total o f N 2 com p lex m ultiplications and N ( N 1) com p lex additions. It is instructive to view the D F T and ID F T as linear transform ations on seq u en ces {jc()} and {X(fc)}, respectively. Let us d efine an Af-point vector %N o f th e signal se q u en ce x( n) , n 0, 1 , . . . , N 1, an N -p oin t vector X N o f frequency
404
Chap. 5
Figure 5.6 Magnitude and phase of an N -poini DFT in Example 6.4.2; (a) L = N = 50; (b) L = 10, N = 100.
sam ples, and an N x N m atrix
as 1 , X,v
- *(0) X* =
'1 1 3 * 1 !
=
1
x(l)
x(0) X ( l)
-I
.x(N-l).
1 w*
- X ( N - 1 ).
1. ..
K w*
(5.1.23)
N
y y 2 ( N 1)
.1
<
_1
W%N~ 0
(* -!)< A '-l)
Sec. 5.1
405
W ith th ese definitions, the W-point D F T m ay b e exp ressed in matrix form as X* = (5.1.24)
w h ere is the m atrix o f the linear transform ation. W e ob serve that Wjv is a sym m etric matrix. If w e assum e that the inverse o f W * exists, then (5.1.24) can be inverted by prem ultiplying both sid es by W ^1. T hus w e ob tain x/v = (5.1.25)
B u t this is just an exp ression for the ID F T . In fact, th e ID F T as given by (5.1.21), can b e exp ressed in m atrix form as ** = N (5.1.26)
406
Chap. 5
where d en o tes the com p lex conjugate o f the m atrix W A. (5.1.26) with (5.1.25) leads us to con clu d e that W *1 = -W * w w hich, in turn, im plies that W W ;, = N l N
C om parison of
(5.1.27)
(5.1.28)
w here I * is an N x N identity matrix. T h erefore, the m atrix in the trans form ation is an orthogonal (unitary) m atrix. F urtherm ore, its inverse exists and is given as W * n / N . O f course, the existen ce o f the inverse o f W,v w as established p reviously from our derivation o f the ID FT . Example 5.L3 Compute the D FT of the four-point sequence x(n) = (0 1 2 3)
Solution The first step is to determ ine the matrix W4. By exploiting the periodicity property of W 4 and the symmetry property = -v v the matrix W 4 may he expressed as
W< =
~w " w "
Wl -W -1 1 1 .1 1
1 - 1 1
w "
w4 4 w;
w w1. w9
1 j -1 -j
< 1 1
'1
1
1 w4
1 M'i W?
I" W 4 -
1 J
W ; w *
-j
-1 j
T 6
Then
X 4 = W 4X 4=
-2
-2 + 2;
L-2-2JJ The IDFT of X 4 may be determ ined by conjugating the elements in W 4 to obtain WJ and then applying the formula (5.1.26). T h e D F T and ID F T are com putational to o ls that play a very im portant role in m any digital signal processing applications, such as frequency analysis (spectrum analysis) o f signals, p ow er spectrum estim ation , and lineaT filtering. T h e im por tance o f the D F T and ID F T in such practical ap p lication s is du e to a large extent on the ex isten ce o f com putationally efficient algorithm s, know n co llectiv ely as fast
Sec. 5.1
407
F ourier transform (F F T ) algorithm s, for com puting the D F T and ID F T . T his class o f algorithm s is d escrib ed in Chapter 6.
5.1.4 Relationship of the DFT to Other Transforms

In this d iscu ssion w e h ave indicated that the D F T is an im portant com putational to o l for perform ing frequency analysis o f signals on digital signal p rocessors. In v iew o f the other frequency analysis to o ls and transform s that w e have d e v e l o p ed , it is im portant to establish the relationships b etw een the D F T to th ese other transform s.
Relationship to the Fourier series coefficients of a periodic sequence.

A p eriod ic se q u en ce F ourier series o f th e form with fundam ental period N can b e represented in a
iV~l x p (n) ^ kei2l,nk,N t=o 1 A'-l Ct = ~ J 2 XP(/1 )r~j2"k/N ^ n=(l
oo < n < oo
(5.1.29)
w h ere the F ourier series coefficients are given by the expression

Jt = 0 , 1 .........A T- 1
(5.1.30)
If w e com p are (5.1.29) and (5.1.30) with (5.1.18) and (5.1.19), w e ob serve that the form ula for the F ourier series coefficients has the form o f a D F T . In fact, if we d efine a se q u en ce x(rt) = x p(n), 0 < n < N - 1 , the D F T o f this se q u en ce is sim ply X(k) = Nc k (5.1.31)
F urtherm ore, (5.1.29) has the form o f an ID F T . T hus th e N -p oin t D F T provid es the exact line spectrum o f a p eriod ic seq u en ce with fundam ental period N.
Relationship to the Fourier transform of an aperiodic sequence. W e have already sh ow n that if -t(n) is an aperiodic finite en ergy seq u en ce w ith Fourier transform X(a>), w h ich is sam pled at N eq u ally spaced freq u en cies = 2n k / N , k = 0 , 1 , . . . , N 1, the spectral com p onents O O
X(k) = =
n oo
x ( n ) e - j2* nk,N
k = 0,1,..., N - 1
(5.1.32)
are th e D F T coefficien ts o f the period ic seq u en ce o f period N , given by
O O
J;p (n) = ^ x(n-lN ) (5.1.33)
/* O O
Thus x p (n) is d eterm in ed by aliasing {jc(n)J o ver th e interval 0 < n < N - 1. T h e finite-duration seq u en ce
408
Chap. 5
bears no resem blance to the original seq u en ce {x(n )), u nless ;c(n) is o f finite dura tion and length L < N, in which case x( n) = x{n) 0 < n < N 1 (5.1.35)
O nly in this case will the ID F T o f {X(jfc)} yield the original seq u en ce {*()}.
Relationship to the z*transform.

the ^-transform
Let us consider a seq u en ce x( n) having
(5.1.36) with a R O C that includes the unit circle. If X ( z ) is sam pled at the N equally spaced points on the unit circle z t = e i2*k/ N , 0, 1, 2 , . . . , N 1, w e obtain X( k ) = X(z)U=, it = k = 0, 1 , . . . , N - 1 (5.1.37) x ( n ) e ~ j2nnk/N
T he expression in (5.1.37) is identical to the Fourier transform X(io) evalu ated at the N equally spaced freq u en cies < ot = 2 n k / N , k = 0, 1 , ___ N ~ 1, which is the topic treated in Section 5.1.1. If the seq u en ce x( n) has a finite duration o f length N or less, the seq u en ce can be recovered from its /V-point D F T . H en ce its z-transform is un iqu ely determ ined by its N -p oin t D F T . C onsequently, X ( z ) can be expressed as a function o f the D F T fX (k)} as follow s N~ 1 * (Z ) = ^ * ( ) 2 n n=0 /V-l X(z) =
n=0 L
*=0
(5.1.38)
-N N - 1
X ( Z) =
X( k) ej2*k/Nz - i
W hen evaluated on the unit circle, (5.1.38) yield s the F ourier transform o f the finite-duration seq u en ce in term s o f its D F T , in th e form
k=0 1
T his expression for th e Fourier transform is a p olyn om ial (L agrange) interpolation form ula for X ( w ) exp ressed in term s o f the valu es {X (Jt)) o f th e polyn om ial at a set o f equally spaced d iscrete freq u en cies < ok = 2 n k / N , k = 0, 1 ......... N - 1. With
Sec. 5.2
Properties of the DFT
409
so m e algebraic m anipulations, it is p ossib le to reduce (5.1.39) to the interpolation form ula given p reviously in (5.1.13).
Relationship to the Fourier series coefficients of a continuous-time signal. S uppose that xa (t) is a con tinu ou s-tim e periodic signal with fundam ental
p eriod Tp = 1 /F 0. T he signal can be expressed in a Fourier series
OC
xaU )=
t = 3C
C teJ2,T kF "
(5-1-40)
w here { q ) are the Fourier coefficients. If we sam ple xc,(t) at a uniform rate Fs = N / T p = 1 / T , w e obtain the discrete-tim e sequ en ce x ( n ) = x a( nT) = y C l;ej2,rkF",T = ckej2nt,l/N k= '\, k= cc
N - 1 J 2 x k n /N
- E
l~~CX.
It is clear that (5.1.41) is in the form o f an ID F T form ula, where rv X{ k) = N j 2 Ck-iN = N c t /=--v and V Q = y t'k-iN /=-ac T hus the {q } seq u en ce is an aliased version o f the seq u en ce (cA}.
(5.1.42)
(5.1.43)
5.2 PROPERTIES OF THE DFT

In S ection 5.1.2 w e introduced the D F T as a set o f N sam p les {X(Jt)} of the F ourier transform X(a>) for a finite-duration seq u en ce |jr(n)} o f length L < N. T h e sam pling o f X (to) occurs at the N equally spaced freq u en cies cd* = 2 n k / N , k 0, 1, 2 ......... N 1. W e dem onstrated that the N sam p les (X(A)} uniquely represent the seq u en ce (;c(n)} in the frequency dom ain. R ecall that the D F T and inverse D F T (ID F T ) for an //-p o in t seq u en ce {*()} are given as N- 1 DFT: X( k ) = J ^ x ^ n)WN n=0
1 N- 1
* - 0 , 1 .........A ' - l
(5.2.1)
ID FT: jc(n) = X ( k ) W ~ kn i=0
n = 0 , 1 .........N - 1
(5.2.2)
where Wn is defined as
WN = e~ j2n,N
(5.2.3)
410
Chap. 5
In this section w e present the im portant p rop erties o f the D F T . In view o f the relationships estab lish ed in Section 5.1.4 b etw een the D F T and Fourier series, and Fourier transform s and --transform s o f d iscrete-tim e signals, w e exp ect the p roperties o f the D F T to resem ble the properties o f these o th er transform s and series. H ow ever, som e im portant d ifferen ces exist, o n e o f w hich is the circular con v o lu tio n property derived in the follow in g section . A good understanding of these properties is extrem ely helpful in the application o f the D F T to practical p roblem s. T h e notation used b elow to d en o te the N -point D F T pair x ( n ) and AT(Jt) is
*() S
x(k)
5.2.1 Periodicity, Linearity, and Symmetry Properties Periodicity.

If x(n) and X (k ) are an Af-point D F T pair, then x(n + N) = x( n) X( k + N) = X( k ) for all n for all k (5.2.4) (5.2.5)
T h ese periodicities in .r(n) and A' (A:) follow im m ed iately from form ulas (5.2.1) and (5.2.2) for the D F T and ID FT , respectively. W e previously illustrated the periodicity property in the seq u en ce x (n) for a given D FT. H ow ever, we had not p reviously view ed the D F T X( k ) as a periodic seq u en ce. In so m e applications it is ad van tageou s to do this.
Linearity.
If * ,( )
N DFT
X l (k)
and x 2(n) K N X 2(k)
then for any real-valued or com p lex-valu ed con stan ts a\ and a2,
a\ X] ( n) + a 2x 2(n) < + a {X\ ( k) + a2X 2(k)
DFT
(5.2.6)
This property follow s im m ediately from the definition o f the D F T given by (5.2.1).
Circular Symmetries of a Sequence. A s w e h ave seen , the Appoint D FT o f a finite duration seq u en ce, x(n) o f length L < N is eq u ivalen t to the W-point D F T o f a periodic seq u en ce xp (n ), o f period N, w hich is ob tain ed by periodically exten d ing jc(n), that is,
OO
xp (n) =
x( n - I N) /= O O
(5.2.7)
Sec. 52.
411
N o w su p p ose that w e shift the periodic seq u en ce x p (n) by k units to the right. Thus w e obtain a n oth er period ic sequ en ce
OC
x'p (n) = x p (n - k) = T h e finite-duration seq u en ce 1 0,
^ x( n - k - IN) /=-OC
(5.2.8)
otherw ise
( 5 .2 .9 )
(5.2.10)
is related to the original seq u en ce x ( n ) by a circular shift. T h is relationship is illustrated in Fig. 5.7 for N = 4. In gen eral, th e circular shift of the seq u en ce can b e rep resen ted as the index m o d u lo N . T h u s w e can w rite x'(n) = x(n k, m odu lo N ) s x((n ~ k) ) N For exam p le, if k = 2 and N = 4, w e have x'(n) = * (( - 2))4 which im plies that jc'(0) = x ( ( - 2 ) ) 4 = x ( 2 ) x' { l ) =
jc (
( - 1))4 = ;c ( 3 )
* '( 2 ) = jt<(0))4 = j (0)

x
'( 3 ) =
j t ( ( 1))4
j t (1 )
H en ce x'(n) is sim ply x (n) shifted circularly by tw o units in tim e, w here the cou n terclock w ise direction has b een arbitrarily selected as the p ositive direction. Thus w e con clu d e that a circular shift o f an -point seq u en ce is eq u ivalen t to a linear shift o f its p eriod ic ex ten sion , and vice versa. T h e in h eren t p eriod icity resulting from the arrangem ent o f the Af-point se q u en ce o n th e circum ference o f a circle dictates a differen t definition o f even and od d sym m etry, and tim e reversal o f a sequ en ce. A n Af-point seq u en ce is called circularly even if it is sym m etric ab ou t the p oin t zero on th e circle. T h is im plies that x ( N n ) = x ( n ) 1 < n < N 1 (5.2.11)
A n W -point seq u en ce is called circularly o d d if it is antisym m etric ab ou t the point zero o n the circle. T h is im plies that x ( N n) = x( n) 1 < n < N 1 (5.2.12)
T h e tim e reversal o f an Af-point seq u en ce is attained by reversing its sam ples about the p o in t ze r o on the circle. T h u s th e seq u en ce jc((n ) ) # is sim ply given as * ((- ))* = x ( N - n ) 0 < n < N - l (5.2.13)
T h is tim e reversal is eq u ivalen t to plottin g j:(/i ) in a clock w ise direction on a circle.
412
Chap. 5
j(n) 2t '
4< 3
<a)
xM 3
4'
41
T2J
-4 -3 - 2 - I
. . t
0 I 2 3 (b)
r l l '
4 5 6 7
Jr,,(n - 2) 4< 3 <
4'
4'
t I -6 -5
-4
-3
, - t " l 1 m i I
-2 -I 0 i 2 3 4
(O
4 3
TI
*(0)
*'(2)
<e)
Figure 5.7 Circular shift o f a sequence.
Sec. 5.2
413
A n eq u ivalen t definition o f even and od d seq u en ces for th e associated peri odic seq u en ce x p (n) is given as follow s even: odd: x(n) = = x p( N - n) (5.2.14) x p (n) = - x p( n) - - x p ( N - n) If the periodic seq u en ce is com p lex-valu ed , w e have conjugate even: conjugate odd: x (n) = x U N n) P x p (n) = x * ( N n ) (5.2.15)
T h ese relationships suggest that w e d ecom p ose the seq u en ce x p (n) as xp (n) = xP'(n) + x ^ i n ) w here xPAn) = \ [ x p (n) + x p( N - n)] , P x p{n) = \ [ x P(n) - x * ( N - n ) ] (5.2.17) (5.2.16)
Symmetry properties of the DFT. T h e sym m etry properties for the D F T can be o b tain ed by applying the m eth od ology p reviously used for the Fourier transform . Let us assum e that the N -p oin t seq u en ce jc(n) and its D F T are both com plex valued. T h en the seq u en ces can be exp ressed as
jr(n) = XR(n) + j x / ( n ) X( k ) = Xg ( k ) + j X , ( k ) 0 < n < N 1 0 < k < N - 1 (5.2.18) (5.2.19)
B y substituting (5.2.18) in to the expression for the D F T given by (5.2.1), w e obtain * * (* ) = j*J?(") cos + x , ( n ) sin
(5.2.20)
X,(k) = E
[**(" )sin ~ J p ' -*/() cos

J
(5-2.21)
n=0 L
Sim ilarly, by substituting (5.2.19) into the expression for the ID F T given by (5.2.2), w e obtain x K(n) = i [* * ( * ) cos *=0 L f . - X ,( fc ) s in ^ ^ j J 2n kn " I cos ^ J (5.2.22)
1 * /(n ) =
. 2nkn sin ^ h
(5.2.23)
Real-valued sequences.
from (5.2.1) that
If the seq u en ce jc(n) is real, it follow s directly
X( N - k ) = X m (k) = X(-Jfc)
(5.2.24)
414
Chap. 5
C onsequently, \ X ( N ~ jfc)| = |AT(A:)| and I X ( N - k) = l X ( k ) . Furtherm ore, xi (n) = 0 and therefore ;r(rj) can b e d eterm in ed from (5.2.22). which is another form for the ID F T .
Real and even sequences.
If x (n) is real and even , that is, 0 < n < N 1
x(n ) x ( N n)
then (5.2.21) yield s X / ( k ) = 0. H en ce the D F T reduces to Cy 2nkn X ( k ) = y x ( n ) c o s ------0 < k < N \ (5.2.25)
^
which is itself real-valued and even . reduces to
N
Furtherm ore, since X / ( k ) = 0, the ID FT
jc () ^ X (k ) c o s -------*=o
0 < n < N 1
(5.2.26)
Real and odd sequences.
If x ( n ) is real and odd, that is, 0 < n < N ~ 1
x ( n ) = x ( N n) then (5.2.20) yields X R(k) = 0. H en ce X( k ) j Y x ( n ) sin t z
o < k < N 1 N
(5.2.27)
which is purely im aginary and odd. Since X K(k) = 0, the ID F T reduces to * ( ) = j ~ y X (k ) sin 0 < n < N - 1 (5.2.28)
N
Purely imaginary sequences.
(5.2.20) and (5.2.21) reduce to
N
In this case, x( n) = j x f (n). C onsequently,
**< *) -
E
n=0
JC,(n)sin
N
(5.2.29)
x >(k) -
* / ( " ) 0 0 8 77n=0
(5.2.30)
W e ob serve that Xn( k) is odd and X } (k) is even. If x / ( n ) is odd, then X / ( k ) = 0 and h en ce X(Jt) is purely real. O n th e other hand, if x r (n ) is ev en , then X * (Jt) = 0 and h en ce X(i fc) is purely im aginary.
Sec. 5.2

TABLE 5.1 SYMMETRY PROPERTIES OF THE DFT
415
yV-Point Sequence x(n). 0< n < N - 1 x(n) x*(n) x*( N - n ) xx(n) JXj (n) xcf(n) = i[jr(n) + x*(N - )] xcJ n ) = |[T (n) - x*( N - /j)] Real Signals A ny real signal jr{n)
/V-Point DFT X(k) X ( N k) X' ( k) X t, *)] X,-,<*) = HX( k ) - X ' ( N - *)] X K(k) j X/ ( k ) X(k) = X*( N - k ) X K(k) = X W - k ) X, (k) = - X , i N - k ) \X(k)\ = \ X( N - k ) \ I X( k) = - I X { N - k) X K{k) jX/tt)
x, An) = j[* (n ) + x ( N - n)] x.(n) = 5 (a-() - x( N - )]
T h e sym m etry properties given ab ove may be sum m arized as follow s: Mu) - xjf(n) +
x' kOi )
+ jxj {n) + jx'/in) (5-2-3D
I \ x(k) = x H (k) + x ; (k) + j x ; t (k) + j x l < ,(k)
AH the sym m etry properties o f the D F T can easily be d ed u ced from (5.2.31). For exam p le, the D F T o f the seq u en ce x pr(n) = j ^ f n ) + x * ( N - n)] is * * ( * ) = X'K(k) + XK(k) T h e sym m etry p rop erties of the D F T are sum m arized in T able 5.1. E x p loitation o f th e se properties for the efficient com putation o f the D F T o f special se q u en ces is con sid ered in so m e o f the problem s at the end o f the chapter.
5.2.2 M ultiplication o f Two DFTs and Circular C on volu tion

S u p pose that w e have tw o finite-duration seq u en ces o f length N, Jti(n) and T h eir resp ective Appoint D F T s are JV-1 X, (k ) = Y L jc, (n)e~i2* HklN iV-1 X 2(k) = Y x 2( n ) e - j2*nk/N n=0 k = 0,1,..., N 1 (5.2.32)
k = 0,1,..., N - 1
(5.2.33)
416
Chap. 5
If w e m ultiply the tw o D F T s togeth er, the result is a D F T , say o f a se quen ce jf3(n) o f length N. L et us determ in e the relationship b etw een x3(n) and the sequ en ces X] ( n) and jr2(n). W e have Xi ( k) = X t ( k ) X2(k) T h e ID F T o f {X3(Jt)} is 1 N-1 x 3 (m) = ~ X j ( k ) e i2* km/N *=o k = 0 , 1 .........N - 1 (5.2.34)
A'-l
(5.2.35)
Suppose that we substitute for X\ ( k ) and X 2(k) in (5.2.35) using the D F T s given in (5.2.32) and (5.2.33). T hus w e obtain
j27rkn/N /=0
-j2nkl/N
j2nkm/N (5.2.36)
k=0 Af-l
T he inner sum in the brackets in (5.2.36) has the form -i *- w here a is defined as
Q _ e j2n(m-n-l)/N
f N.
a = 1 (5.2.37) 0751
y > A= l l - a "
W e observe that a = 1 w hen m n I is a m ultiple o f N. O n the other hand, a N = 1 for any value o f a = 0. C on sequ en tly, (5.2.37) reduces to /V -1
t a
ks=0
, I N, 1 0, 1
I = m - n + p N = ((m - n ) ) N, otherw ise
p an in teger
2 381
If w e substitute the result in (5.2.38) in to (5.2.36), w e obtain the desired expression for ^ ( m ) in the form xi ( m) = Y x \ ( n) x 2((m - n ) ) N m = 0,1,..., N - 1 (5.2.39)
T h e expression in (5.2.39) has th e form o f a con volu tion sum . H ow ever, it is not the ordinary linear con volu tion that w as introduced in C hapter 2, which relates the output seq u en ce y( n) o f a linear system to th e input seq u en ce x( n) and the im pulse response h(n). Instead, the con volu tion sum in (5.2.39) in volves the index
Sec. 5.2
417
((m fl))w and is called circular co nvolution. Thus w e con clu d e that m ultiplication o f the D F T s o f tw o seq u en ces is eq u ivalen t to the circular con volu tion o f the tw o seq u en ces in the tim e dom ain. T h e fo llow in g exam p le illustrates the op eration s in volved in circular con vo lution.
Example Si.1
Perform the circular convolution of the following two sequences: *!() = { 2,1,2( 1} t x 2(n) = {1,2,3,4} t
Solution Each sequence consists of four nonzero points. For the purposes of illus
trating the operations involved in circular convolution, it is desirable to graph each sequence as points on a circle. Thus the sequences x\{n) and x 2(n) are graphed as illustrated in Fig. 5.8(a). We note that the sequences are graphed in a counterclock wise direction on a circle. This establishes the reference direction in rotating one of the sequences relative to the other. Now, xi(m) is obtained by circularly convolving jci<) with J 2() as specified by (5.2.39). Beginning with m = 0 we have
jrj(O ) = y ^ j r i ( n ) j r 2( ( - n ) ) y
n= U jc2(( )>4 is simply the sequence x 2(n) folded and graphed on a circle as illustrated in Fig. 5.8(b). In other words, the folded sequence is simply xz(n) graphed in a clockwise direction. The product sequence is obtained by multiplying jci(h) with * j( ( - n ) ) 4, point by point. This sequence is also illustrated in Fig. 5.8(b). Finally, we sum the values in the product sequence to obtain
*3(0) = 14
For m = 1 we have
3
*3(1) = ^ x l ( n ) x 2( ( l - 4
*0 It is easily verified that *2((1 n ))4 is simply the sequence * 2((~ n ))4 rotated coun terclockwise by one unit in time as illustrated in Fig. 5.8(c). This rotated sequence multiplies x\ (n) to yield the product sequence, also illustrated in Fig. 5.8(c). Finally, we sum the values in the product sequence to obtain *3(1). Thus * 3d) = 16 F or m = 2 we have
3
*3(2) = E
Xl (n )*2((2 - n ))
n0 Now *2 ((2 n ))4 is the folded sequence in Fig. 5.8(b) rotated two units of time in the counterclockwise direction. The resultant sequence is illustrated in Fig. 5.8(d)
*|0) = 1
jr2( l) = 2
jri(2> - 2
* ,( 0) = 2
* 2(2) = 3
*2( 0 ) =1
(a)
jc2 ( 2 )
=3
* 2( 0 ) = 1
jc 2(I) = 2 Folded sequence
Product sequence (b)
x2(0) = 1
jr2(3 )= 4
x2(l) = 2
jc 2(2> = 3 Folded sequence rotated by one unit in time
(c)
x 2( l )
=2
*2(0) = 1
*2(2)= 3
*2(3) = 4 Folded sequence rotated by two units in time (d) *j(2) = 3
*2(3) = 4
*j(0)=l
Folded sequence rotated by three units in time Figure 5.8 Product sequence
(e)
Circular oonvolutioa o f tw o sequences.
Sec. 5.2
419
along with the product sequence x l (n)xi((2 - n))A. By summing the four term s in the product sequence, we obtain
* j ( 2 ) = 14
For m = 3 we have
3
X 30 ) = y^xi(n)x;((3 - n ) ) 4
i,= Q The folded sequence X2(( n))4 is now rotated by three units in time to yield jc2(<3 n))4 and the resultant sequence is multiplied by *j(n) to yield the product sequence as illustrated in Fig. 5.8(e). The sum of the values in the product sequence is
x 3( 3) = 16
We observe that if the com putation above is continued beyond m = 3, we simply repeat the sequence of four values obtained above. Therefore, the circular convolution of the two sequences x \ ( n ) and x2(n) yields the sequence xi(n) = {14,16,14,16) t From this exam p le, w e observe that circular con volu tion involves basically the sam e four step s as the ordinary linear co n vo lu tio n introduced in C hapter 2: fo l d i n g (tim e reversing) on e seq u en ce, shifting the fold ed seq u en ce, m u ltip ly in g the tw o seq u en ces to obtain a product seq u en ce, and finally, s u m m i n g the valu es o f the product seq u en ce. T h e basic d ifference b etw een th ese tw o types o f con volu tion is that, in circular con v olu tion , the foldin g and shifting (rotatin g) op eration s are p erform ed in a circular fashion by com puting the index o f o n e o f the seq u en ces m o d u lo N . In linear co n volu tion , there is n o m odu lo N op eration . T h e reader can easily sh ow from our previous d evelop m en t that eith er on e o f the tw o seq u en ces m ay b e fold ed and rotated w ithou t changing the result o f the circular con volu tion . Thus N- 1 x$(m) = Y ' X 2 ( n )x i ( ( m n ) ) s m = 0 , 1 , ... t N 1 (5.2.40) n=0 T h e fo llo w in g exam p le serves to illustrate the com p utation o f x 3 (n) by m eans o f the D F T and ID F T . Exam ple SJJ1 By m eans of the D FT and IDFT, determ ine the sequence xj(n) corresponding to the circular convolution of the sequences x\ (n) and X2 (n) given in Exam ple 5.2.1. Solution xi(n) is First we com pute the DFTs of xi(n) and * 2(0 ). T he four-point D FT of
3
Xi<*) = Y Xl ft-0
k = 0 ,1 ,2 ,3
420 Thus
Chap. 5
*1(0) = 6 The D FT o f Jt2(n) is
JCid) = 0
Jlf,(2) = 2
A'1(3) = 0
X 2(.k) = Y fl0
k = 0 ,1 ,2 ,3
= 1 + 2 e - iltkri + 3e~>*k + 4e~J^ kri
Thus X2(0) = 10 X 2a ) = - 2 + j 2 X2( 2 ) = - 2 X 2Q) = - 2 - j 2
When we multiply the two DFTs, we obtain the product X3(*) = X t (k)X 2(k) or, equivalently, X3(0) = 60 * 3(1) = 0 X3( 2 ) = - 4 Xj (3) = 0
Now, the ID FT of X 3(k) is jr3(n) = Y X3(Jt)cj7,rn*/4 *41 = j(6 0 - 4eJn) Thus j 3(0> = 14
jc 3(1)
n = 0, 1, 2, 3
= 16
x3(2) = 14
;c3(3) = 16
which is the result obtained in Exam ple 5.2.1 from circular convolution. W e conclude this section by form ally stating this im portant property o f the DFT.
Circular co n v o lu tio n .
If * i(n )
DFT
x m
and
then *1 (H) (g> * 2 (n ) ^

DFT
(*)X 2(Jt)
(5.2.41)
w here x\ (n) (N) x2(n) d en o te s th e circular con volu tion o f the se q u en ce x i(n ) and x%(n).
Sec. 5.2
Properties of the DFT 42)

*6)
421
*(6)
jt (2)
Figure 5.9 Tim e reversal of a sequence.
5.2.3 Additional DFT Properties Time reversal of a sequence.

If
DFT
th e n x((-n))v = x ( N - n ) X ( ( - k ) ) N = X ( N - k) (5.2.42)
H e n c e re v e rsin g th e /V-point se q u e n c e in tim e is e q u iv a le n t to rev ersin g th e D F T values. T im e rev ersal of a s e q u e n c e x(rt) is illu stra te d in Fig. 5.9. Proof . F ro m th e d efin itio n of th e D F T in (5.2.1) w e h a v e JV-I D F T { * W - )} = x ( N ~ n ) e ~ j2nkn/N n=0 I f w e c h a n g e th e in d ex fro m n to m = N - n, th e n N -l D F T [ x ( N - n ) } = x ( m ) < T j7,r*(* - m)/,v =o
A '-l
= Y , x ( m ) e > 2*kmIN m=0 = Y x ( m ) e - i2*m(N- k)/N = X ( N k) m=0 W e n o te th a t X ( N - k) = X { ( - k ) ) N, 0 < k < N - l .
Circular time shift of a sequence.
If
422 th e n
Chap. 5
* n - l))N
X ( k ) e ~ j2* k!/N
(5.2.43)
Proof. From the definition o f the D F T w e have

N- 1
D F T { * ((# i- / ) ) * } = * ( ( n - / ) ) e - ' ' 2ir*"/N n=0 = > ( ( - l ) ) Ne ~ * * k/N n=0 + Y x ( n - l ) e - j kn/N n=l B ut x( (n - 1)) h = x ( N - l + n). C onsequently, x ( ( n - l ) ) Ne~j2*kn/N = ] T n 0 rr= s() = F urtherm ore, y x { n - l ) e - J2,Tkn/N = /=/ T herefore, N~ 1 DFT{jc(( - / ) ) } = ^ x{m)e~-'2xk(m+,)/N
m=0
~ 1 + n ) e ' j2jrkn/N
x ( m ) e ' ^ k^ m=N-l
x ( m ) e - j2kim+IWN
= X ( k ) e ' ^ u/N
Circular freq u en cy sh ift.
If x( n)
DFT
X( k)
then x ( n ) e j2*In/N 5 - X ( ( k - / * (5.2.44)
H en ce, the m ultiplication o f the seq u en ce jc(/i) w ith the com p lex exp on en tial se quence eP***1* is eq u ivalen t to th e circular shift o f the D F T by I u n its in frequency. This is th e dual to the circular tim e-shifting p roperty and its p r o o f is sim ilar to the latter.
Sec. 5.2
Properties of the DFT If X(k)
423
C o m p lex -co n ju g a te p roperties.

jr(n) K th e n x\n)
DFT
* * ( ( - * ) ) * = * * ( * - k)
(5.2.45)
T h e p r o o f o f th is p r o p e rty is left as an ex ercise fo r th e re a d e r. T h e ID F T o f X m (k) is Y T h e re fo re , x * ( ( - n ) ) N = x+( N - b) A / X'(k) (5.2.46) X * ( k ) e J2nkn/N = x ( J k ) e ' 2**(A '- " )' A '
*=o
Circular correlation.
V(n), if
In g e n e ra l, fo r c o m p le x -v a lu e d se q u e n c e s x ( n ) and
DF!'
x(n) an d v() th e n FJV(/)

n
/v
X(A-)
y<*)
R ( k ) = X ( k ) Y r (k)
(5.2.47)
w h ere r ,,.(/) is th e (u n n o rm a liz e d ) c irc u la r c ro ssc o rre la tio n se q u e n c e , d efin e d as

N- 1
^ v (0 = ^ * ( n ) / ( ( n - l))N F r o o / W e c a n w rite f*v(/) as th e circ u la r co n v o lu tio n o f x ( n ) w ith y*(n), th a t is,
T h e n , w ith th e a id o f th e p r o p e rtie s in (5.2.41) a n d (5.2.46), th e W -point D F T o f rxy(l) is R i y (k) = X (k)F *(k) In th e sp e c ia l case w h e re y (n ) = x ( n ) , w e h a v e th e c o rre sp o n d in g e x p ressio n fo r th e c irc u la r a u to c o r re la tio n o f x ( n ) ,
424
Chap. 5
Multiplication of two sequences.

jr,(n ) and x 2(n) then x A n ) x 2(n) K
If * ,( * )
X 2(k)
^ X ,( J t ) ( N ) X 2(*)
(5.2.49)
This property is the dual o f (5.2.41). Its p ro o f follow s sim ply by interchanging the roles o f tim e and frequency in the exp ression for the circular con volu tion of tw o sequ en ces.
Parsevals theorem.
eral, if
For com p lex-valu ed seq u en ces * (n ) and _y(n), in gen
and
then
/V 1
A N- 1
/V 1
Jt(n)y*(n) = - X ( J t ) r (/:) N *=0 n=0
(5-2 50)
Proof. T h e property follow s im m ed iately from the circular correlation prop erty in (5.2.47). W e have
N -1
^ x ( n ) y * ( n ) = r JV(0) and
k~0 H en ce (5.2.50) fo llo w s by evalu atin g the ID F T at I = 0. T he exp ression in (5.2.50) is th e gen eral form o f P arsevals theorem . In th e sp ecial case w h ere ;y(n) = x ( n ) , (5.2.50) red u ces to
/11=0
\x{n)\2 = i
JtseO
|X ( * ) | 2
(5.2.51)
Sec. 5.3
Linear Filtering Methods Based on the DFT

PROPERTIES OF THE DFT
425
TABLE 5.2
Property Notation Periodicity Linearity Time reversal Circular time shift Circular frequency shift Complex conjugate Circular convolution Circular correlation Multiplication of two sequences Parsevals theorem
Time Dom ain x(n), yOi) x (n) =s x(rt + iV) a]Xi(n) + a2x 2(n) x ( N n) *((" - D) n x(n)ei2jr,n/N x "(n) xi (n )@ jr2(n) x (n ) y * (-n ) jri(n)x2(n) /V 1 n=(!
Frequency Domain *<*), Y(k) X( k) = X( k + N) 0\ Xi ( k) + a2X 2(k) X(N - k ) X(k)e~J2*kl/N X ((k-l))N X*(N k) xm xiik) X(k) Y' ( k)
*3= 0
w hich ex p resses the energy in the finite-duration seq u en ce x ( n) in term s o f the frequency co m p o n en ts {X()l T h e properties o f the D F T given ab ove are sum m arized in T able 5.2.
5.3 LINEAR FILTERING METHODS BASED ON THE DFT

Since the D F T provides a discrete frequency rep resen tation o f a finite-duration seq u en ce in the frequency dom ain, it is in terestin g to exp lore its use as a com p utational to o l for linear system analysis and, especially, for linear filtering. W e have already estab lish ed that a system with freq u en cy resp on se H { w ) y w hen e x cited with an input signal that has a spectrum X(a>), p o ssesses an output spectrum Y(a>) = X ( oj) H ( w ). T he output seq u en ce y ( n) is d eterm in ed from its spectrum via the inverse F ourier transform . C om putationally, the p rob lem with this frequencydom ain approach is that X(a>), H(a>), and Y(a>) are fun ction s o f the continuous variable o>. A s a con seq u en ce, the com p utations can n ot be d o n e on a d igital com puter, since th e com puter can on ly store and perform com p utations on quantities at discrete frequencies. O n the other hand, th e D F T d o es len d itself to com p utation on a digital com puter. In th e discussion that follow s, w e d escrib e h ow th e D F T can b e used to perform lin ear filtering in the frequency dom ain. In particular, w e p resent a com p utational p rocedure that serves as an alternative to tim e-d om ain co n v o lution. In fact, the frequency-dom ain approach b ased o n the D F T , is com pu tationally m ore efficien t than tim e-dom ain con volu tion d u e to the existen ce o f efficient algorithm s for com p utin g the D F T . T h ese algorithm s, w hich are d e scribed in C hapter 6, are collectively called fast F ourier transform (FFT ) algo rithms.
426
Chap. 5
5.3.1 Use of the DFT in Linear Filtering

In the p receding section it w as d em onstrated that the product o f tw o D F T s is equivalent to th e circular con volu tion o f the corresponding tim e-d om ain sequ en ces. U nfortunately, circular con volu tion is o f no use to us if our ob jectiv e is to deter m ine th e output o f a linear filter to a given input seq u en ce. In this case w e seek a frequency-dom ain m eth od ology eq u ivalen t to lin ear con volu tion . S u ppose that w e have a finite-duration seq u en ce x( n) o f length L which excites an F IR filter o f length Af. W ithout loss o f generality, let jc(n) = 0, h(n) = n < 0 and n > L n <
0,
and n > M
where h(n) is the im pulse resp onse o f the F IR filter. The output seq u en ce y ( n ) o f the FIR filter can be exp ressed in the tim e dom ain as the con volu tion o f x( n) and h(n), that is y(n)
M- 1
h(k) x( n k )
(5.3.1)
Since h(n) and jc(n) are finite-duration seq u en ces, their con volu tion is also finite in duration. In fact, the duration o f y( n) is L + M 1. T h e frequency-dom ain equivalent to (5.3.1) is Y(co) = X( w) H( ( o ) (5.3.2)
If the seq u en ce y (n ) is to be represented u n iqu ely in the freq u en cy dom ain by sam ples o f its spectrum Y (to) at a set o f d iscrete freq u en cies, th e n um ber o f distinct sam ples m ust equal or exceed L + M - 1. T h erefore, a D F T o f size N > L + M I, is required to represent [y(n)\ in the frequency dom ain. N ow if Y(k) = Y{a>)\m 2 ,tkfN k = 0 , 1 ,. . . , N 1 it = 0 ,1 ....... N - 1
= X M H M U h t /N
then Y(k) = X ( k ) H ( k )
J fc = 0 , l , . . . , t f - 1
(5.3.3)
w h ere {X(Jfc)} and {f/(Jt)} are the N -p oin t D F T s o f th e corresp on d ing sequences x( n) and h (n), resp ectively. Since the seq u en ces x ( n) and h ( n ) have a duration less than N , w e sim ply pad th ese seq u en ces w ith zeros to in crease their length to N. T his increase in th e size o f the seq u en ces d o e s n ot alter their spectra X(o>) and H(a>), which are con tin u ou s spectra, sin ce the seq u en ces are ap eriod ic. H ow ever, by sam pling their spectra at N equally sp aced p oin ts in freq u en cy (com p u ting the JV-point D F T s), w e have increased the num ber o f sam p les that represent these seq u en ces in the frequency dom ain b eyon d the m inim um n u m b er (L or M, re sp ectively).
Sec. 5.3
427
Since the N = L + M 1-point D F T o f the output seq u en ce y ( n ) is sufficient to represent y ( n ) in the frequency dom ain, it follow s that the m ultiplication o f the N -point D F T s X( k ) and H{ k) , according to (5.3.3), follow ed by the com putation o f the Appoint ID F T , m ust yield the seq u en ce {,v(n)J. In turn, this im plies that the Appoint circular con volu tion o f x( n) with h(n) m ust b e eq u ivalen t to th e linear con v o lu tio n o f x ( n ) with h(n). In other words, by increasing th e length o f the seq u en ces x( n) and h(n) to N points (by appending zeros), and then circularly con volvin g the resulting seq u en ces, w e obtain the sam e result as w ou ld have b een ob tain ed with linear con volu tion . Thus with zero padding, the D F T can b e used to perform linear filtering. T h e follow in g exam p le illustrates the m eth od ology in the u se o f the D F T in linear filtering. Example 5.3.1 By m eans of the D FT and IDFT, determine the response of the FIR filter with impulse response
h(n) = 11.2.3} t
to the input sequence
x(n) = |1. 2, 2.1) t Solution The input sequence has length L = 4 and the impulse response has length M = 3. Linear convolution of these two sequences produces a sequence of length N = 6. Consequently, the size of the DFTs must be at least six. For simplicity we com pute eight-point DFTs. We should also m ention that the efficient com putation of the DFT via the fast Fourier transform (FFT) algorithm is usually perform ed for a length N that is a power of 2. Hence the eight-point D FT of jr(n) is
7
Jkn/8
= 1 + 2e~W This com putation yields
+ le->*k2 -I- e~**k,A
k = 0 .1 ........7
X (4) = 0
X to -i+ J
XC7
428
The Discrete Fourier Transform: Its Properties and Applications The eight-point DFT of h(n) is
7
Chap. 5
H(k) = = 1 + 2e~J*k/* + 3> e- jnkr2 Hence H(0) = 6, H (l) = l + V 2 - > (3 + V 5 ) , H( 4) = 2 H( 6) = ~2 + >2 H( 2 ) = - 2 - j 2
HO) = 1 - ^ 2 + > (3 - V 2 ) , H( 5) = 1 - 7 2 - y (3 - J 2 ) , W(7) = 1 + V2 + j ( 3 + J 2 )
The product of these two DFTs yields Y{k), which is Y (0) = 36, y(4) = 0, y (l) = - 1 4 .0 7 -> 1 7 .4 8 V(5) = 0.07 y 0.515
7
Y (2) = >4 Y( 6) = ~ j 4
X(3) = 0.07 + ,0.515 Y(7) = -14.07 + >17.48
Finally, the eight-point ID FT is v(n) = Y y(k)ej2nk"/H n = 0, 1........7
* = 0
This computation yields the Tesult >() = (1 ,4 ,9 ,1 1 .8 ,3 ,0 ,0 }
t
We observe that the first six values of y(rt ) constitute the set of desired output values. The last two values are zero because we used an eight-point D FT and IDFT, when, in fact, the minimum number of points required is six. A lthough the m ultiplication o f tw o D F T s corresp on d s to circular convolution in the tim e dom ain, w e have ob served that padding the seq u en ces x ( n ) and h(n) with a sufficient num ber o f zeros forces the circular con volu tion to yield the sam e output sequ en ce as linear con volu tion . In the case o f the F IR filtering problem in E xam p le 5.3.1, it is a sim ple m atter to d em onstrate that the six-point circular convolution o f the sequ en ces h( n) = {1, 2 , 3 , 0 , 0, 0} t x ( n ) = {1 ,2 , 2 , 1 , 0 , 0} t results in the output sequ en ce y ( n) = {1, 4, 9 , 1 1 , 8 , 3} t w hich is the sam e seq u en ce ob tain ed from linear con volu tion . (5.3.6) (5.3.4)
(5.3.5)
Sec. 5.3
429
I t is im p o rta n t f o r u s to u n d e rs ta n d th e aliasing th a t re su lts in th e tim e d o m a in w h e n th e size o f th e D F T s is sm a lle r th a n L + M I. T h e fo llo w in g e x am p le focuses o n th e aliasin g p ro b le m . Exam ple 5.3.2 D eterm ine the sequence v(n) that results from the use of four point DFTs in Exam ple 5.3.1. Solution The four-point D F T of h(n) is .-jink*!* fi n \ t * H(k) = ^ A ( )jr e ' ff-0 H(k) = \ + 2 e - ink/1 + 3 e - ikn Hence H(0) = 6, H( \ ) = - 2 - j 2 , H( 2) = 2, H ( 3 ) = - 2 + j2 k = 0 ,1 ,2 , 3
The four-point DFT of x (} is X(k) = \ + 2c ~jnk/2 + 2e~,7,k + 3e ~J%7,t/2 Hencc *(()) = 6, X (l) = - l - j , X(2) = 0, Jf(3) = 1 + y k = 0, 1, 2, 3
The product of these two four-point DFTs is K(0) = 36, f ( l ) = ;4 , Y( 2) = 0. K(3) = ~ j 4
The four-point ID FT yields y(n) = \ Y ^ kêJ2xk,A *=o = 0 , 1 ,2 ,3
= 1(36 + j 4e Jn"^ - j 4e J1*na) Therefore, v(n) = {9,7, 9,11} t The reader can verify that the four-point circular convolution of h(n) with x(n) yields the same sequence y(n). I f w e c o m p a re th e re su lt y (n ), o b ta in e d fro m fo u r-p o in t D F T s w ith th e s e q u e n c e y (n ) o b ta in e d fro m th e use o f e ig h t-p o in t (o r six -p o in t) D F T s, th e tim ed o m a in aliasin g effe cts d e riv e d in S ectio n 5.2.2 a re clearly e v id e n t. In p a rtic u la r, >(4) is a liased in to y (0) to yield y (0) = y(( + y ( 4) = 9 S im ilarly, _y(5) is a liased in to _y(l) to yield
?(1) = ? (!) + ?(5) = 7
430
Chap. 5
A ll other aliasing has n o effect since y(n ) = 0 for n > 6 . C on sequ en tly, w e have y ( 2) = y ( 2) = 9 y(3) = y(3) = 1 1 T h erefore, on ly the first tw o p oints o f y ( n ) are corrupted by th e effect o f aliasing [i.e., y(0) 7^ y(0) and v( l ) ^ y (l)]. T his observation has im portan t ram ifications in the discussion o f the follow in g section , in which w e treat th e filtering o f long sequ en ces.
5.3.2 Filtering of Long Data Sequences

In practical applications involving linear filtering o f signals, the input sequ en ce x ( n ) is often a very long sequ en ce. T his is especially true in so m e real-tim e signal processing applications concerned with signal m onitorin g and analysis. Since linear filtering perform ed via the D F T in volves op eration s on a block o f data, w hich by necessity m ust be lim ited in size due to lim ited m em ory o f a digital com puter, a long input signal seq u en ce m ust be se g m en ted to fixed-size blocks prior to processing. Since the filtering is linear, su ccessive blocks can be p rocessed on e at a tim e via the D F T and the output blocks are fitted togeth er to form the overall output signal sequ en ce. W e now d escribe tw o m eth od s for linear F IR filtering a lon g seq u en ce on a block-by-block basis using the D F T . T h e input seq u en ce is segm en ted in to blocks and each block is p rocessed via the D F T and ID F T to prod u ce a block o f output data. T h e output blocks are fitted togeth er to form an overall output sequ en ce which is identical to the seq u en ce ob tain ed if the lon g block had b een processed via tim e-dom ain con volu tion . T h e tw o m eth od s are called the overlap-save m e t h o d and the o verlap-a dd m eth o d . For b oth m eth od s w e assum e that the F IR filter has duration M . The input data seq u en ce is segm en ted in to blocks o f L points, w h ere, by assum ption, L M w ithout lo ss o f generality.
Overlap-save method. In this m eth od the size o f the in p ut data blocks is N L 4- M 1 and th e size o f the D F T s and ID F T are o f len gth N . E ach data block consists o f the last M - 1 data points o f the p reviou s data b lock follow ed by L new data p oin ts to form a data seq u en ce o f len gth N = L 4- M 1. A n N -point D F T is com p uted for each data block. T h e im pulse resp onse o f the F IR filter is increased in length by appending L - l zero s and an Appoint D F T o f the sequ en ce is com p uted on ce and stored. T h e m ultiplication o f the tw o Af-point D F T s {//(Jt)} and {Xm(Jt)} for the m th b lock o f data yields
Ym(k) = H ( k ) X m(k) T hen the Appoint ID F T yield s the result L ( n ) = { ^ (0 )y ffl(l) ym(M - 1)ym{M ) ym(N - 1)} k = 0 , 1 ......... N - 1 (5.3.7)
(5.3.8)
Sec. 5.3
431
Sin ce the data record is o f length N, the first Af 1 p oin ts o f >m(n) are corrupted by aliasing and m ust be discarded. T h e last L p oints o f y( n) are exactly the sam e as the result from linear con volu tion and, as a con seq u en ce, >(n) = y( n) , n = M, M + 1 .........N - 1 (5.3.9)
T o avoid lo ss o f data du e to aliasing, the last Af 1 p oin ts o f each data record are saved and th ese p oin ts b ecom e the first Af 1 data p oin ts o f the subsequent record, as indicated above. T o begin the processing, the first Af 1 p oints o f the first record are set to zero. Thus the blocks o f data seq u en ces are jci(n) = (0, 0 ........ 0, x(0), x ( l ) .......... x ( L - 1)}
M 1 points
(5.3.10)
x2(n) = {x ( L - M + 1).........x ( L 1 ) , x ( L ) , . . . , x ( 2 L - l) j
M - 1 data points from i|(n) L new data points
(5.3.11)
x3(rt) = {x ( 2 L - M + l ) .........x ( 2 L 1), x ( 2 L ) .......... x ( 3 L - l } )

M - 1 data points from I new data points
(5.3.12)
and so forth. T h e resulting data sequ en ces from the ID F T are given by (5.3.8), w here the first M 1 points are discarded due to aliasing and the rem aining L p oin ts con stitute the d esired result from linear con volu tion . T h is segm en tation o f th e input data and the fitting o f the output data blocks togeth er to form the output seq u en ce are graphically illustrated in Fig. 5.10.
Overlap-add method. In this m eth od the size o f the input data block is L p oin ts and the size o f th e D F T s and ID F T is N ~ L + Af - 1. T o each data block w e append Af 1 zeros and com p ute the N -point D F T . T h u s the data b lock s may be rep resen ted as
jti(n) = {jc(0),
jc(1 )......... jc(
- 1 ) , 0 , 0 _____0}
M -l
(5.3.13)
zeros
X2(n) = [ x( L) , x ( L + 1), . . , X(2L - 1), 0, 0 , . . . , 0}

M-1 zeros
(5.3.14)
jt3(n) = { x ( 2 L ) .........x( 3 L - 1), 0 , 0 , . . . , 0)

M -\
(5.3.15)
zeros
and so on . T h e tw o Appoint D F T s are m ultiplied togeth er to form Ym(k) = H ( k ) X m(k) * = 0 , 1 ......... N - 1 (5.3.16)
T h e ID F T y ield s data blocks o f length N that are free o f aliasing sin ce the size o f th e D F T s and ID F T isW = Z, + Af 1 and the seq u en ces are increased to N -p oin ts b y ap p en d in g zero s to each block.
432
Chap. 5
Output signal
points
/ Discard Pints
Figure 5.10 L in ear F IR filtering by the overlap-save m ethod.
Since each data block is term inated with M 1 zeros, the last M 1 points from each output block m ust be overlap p ed and added to the first M - 1 p oin ts of the succeeding block . H en ce this m eth od is called th e overlap-add m ethod. This overlapping and adding yields the output seq u en ce
y ( ) = {y i(0 ), ^ i ( l ) , . . . , y \ ( L - l ) , y i ( L ) + y2(0 ), ^ ( L + 1) +
(5.3.17) > 2( 1 ) ......... y \ ( N - 1) + y i ( M 1), y 2 ( M ) , . . . } T he segm en tation o f the input data in to b lock s and th e fitting o f th e output data blocks to form the output seq u en ce are graphically illustrated in F ig. 5.11. A t this poin t, it m ay appear to the reader that th e use o f th e D F T in linear F IR filtering is n ot o n ly an indirect m eth od o f com puting the o u tp u t o f an FIR filter, b ut it m ay a lso be m ore exp en sive com p utationally since th e input data must first b e converted to the frequency d om ain via th e D I T , m ultip lied by th e D F T o f th e F IR filter, and finally, converted back to th e tim e d om ain via the ID FT. O n th e contrary, h ow ever, by using the fast F ourier transform algorithm , as will b e show n in C hapter 6, th e D F T s and ID F T require few er com p utations to com pu te th e output seq u en ce than th e direct realization o f the F IR filter in the time
Sec. 5.4
Input data
Frequency Analysis of Signals Using the DFT
433
\
MA
Output data
M-1 point.*;/ PZJ add V/y together M-l points^ add together Figure 5.11 Linear FIR filtering by the overlap-add method.
dom ain. This com putational efficiency is the basic advantage o f using the D F T to com p ute the output o f an F IR filter.
5.4 FREQUENCY ANALYSIS OF SIGNALS USING THE DFT

T o com p ute the spectrum o f either a con tinu ou s-tim e or d iscrete-tim e signal, the valu es o f the signal for all tim e are required. H ow ever, in practice, w e observe signals for on ly a finite duration. C on sequ en tly, the spectrum o f a signal can on ly b e approxim ated from a finite data record. In this section w e exam ine the im plications o f a finite data record in frequency analysis using the D F T . If the signal to b e analyzed is an analog signal, w e w ou ld first pass it through an an tialiasing filter and th en sam ple it at a rate F, > 2 5 , w h ere B is the band width o f the filtered signal. T h u s th e highest frequency that is contained in the sam pled signal is Fs f l . F inally, for practical purposes, w e lim it the duration o f the signal to th e tim e interval To L T , w h ere L is the num ber o f sam ples and T
434
Chap. 5
is the sam ple interval. A s w e shall ob serve in the follow in g d iscu ssion , the finite observation interval for the signal places a limit on the freq u en cy resolution; that is, it lim its our ability to distinguish tw o frequency com p on en ts that are separated by less than 1 /To = 1/Z-7" in frequency. L et {*()} d en o te the seq u en ce to b e analyzed. L im iting th e duration o f the seq u en ce to L sam ples, in the interval 0 < n < L 1, is eq u ivalen t to m ultiplying {jc(/i)} by a rectangular w indow w ( n ) o f length L. That is, x (n) = x ( n ) w ( n ) w h ere u;w = {J ; , v,, 0 < n < L - 1 otherw ise (5.4.2) (5.4.1)
N ow su p p ose that the sequ en ce x ( n ) consists o f a single sin u soid , that is, x( n ) = coscoon (5.4.3)
Then the Fourier transform o f the finite-duration seq u en ce x ( n ) can be expressed X(u) = $[W '(cu-<u0) + W 'to + wo)] (5.4.4)
where W(a>) is the Fourier transform o f the w indow seq u en ce, w h ich is (for the rectangular w indow ) W{0}) = sin(df/ 2 ) D/2 (5.4.5)
T o com pute X(a>) w e use the D F T . B y padding the seq u en ce x ( n ) w ith N L zeros, we ca n c o m p u te the N -point D F T o f the truncated (L p oin ts) seq u en ce {Jt(n)J. T he m agnitude spectrum |A W I = |X(t*) | for o* = 2 n k / N , k = 0, l , . . . , A f , is illustrated in Fig. 5.12 for L = 25 and N = 2048. W e n ote that the w indow ed spectrum X((o) is n ot localized to a single frequency, but instead it is spread out over the w h ole frequency range. Thus the p ow er o f the origin al signal sequence {jr(n)} that was concentrated at a single frequency has b een spread by the window into the entire freq u en cy range. W e say that the p ow er has lea k ed o u t in to the entire frequency range. C onsequently, this p h en om en on , which is a characteristic o f w indow ing the signal, is called leakage.
Frequency
Figure 5.12 Magnitude spectrum for I = 25 and n = 2048, illustrating the occurrence of Leakage.
Sec. 5.4
435
W in d o w in g n o t o n ly d isto rts th e sp e ctral e stim a te d u e to th e le a k a g e effects, it also re d u c e s s p e c tra l re so lu tio n . T o illu stra te th is p ro b le m , let us c o n s id e r a signal s e q u e n c e co n sistin g o f tw o fre q u e n c y co m p o n e n ts, x ( n ) = cos co\ n + c o s a ^ n (5.4.6)
W h e n th is se q u e n c e is tr u n c a te d to L sam ples in th e ra n g e 0 < n < L 1, th e w in d o w e d s p e c tru m is X ( w ) \ [ W (to u>i) + W (a> a> 2 ) + W((o + ati) + W(a> + a>2)] (5.4.7)
T h e s p e c tru m W (to) o f th e re c ta n g u la r w indow se q u e n c e h as its first z e ro crossing a t co = 2 n / L . N o w if |<yi a ^ l < 2 n / L , th e tw o w in d o w fu n c tio n s W(co a>\) a n d V/(o) an) o v e rla p a n d , as a co n se q u e n c e , th e tw o sp e c tra l lin es in x ( n ) a re n o t d istin g u ish a b le . O n ly if (a>\ a>i) > 2 n / L will w e se e tw o s e p a ra te lo b es in th e sp e c tru m X(a>). T h u s o u r a b ility to reso lv e sp e c tra l lines o f d iffe re n t fre q u e n c ie s is lim ite d b y th e w in d o w m ain lo b e w idth. F ig u re 5.13 illu s tra te s th e m a g n itu d e sp e c tru m |X (a)|, c o m p u te d via th e D F T , fo r th e se q u e n c e x ( n ) = costuo n + cos + cosa^n (5.4.8)
Magnitude
Frequency (a)
Frequency <b)
(c)
Figure 5.13 Magnitude spectrum for the signal given by (5.4.8), as observed through a
rectangular window.
436
Chap. 5
w here at o = 0.2 n , = 0. 22n, and coi = 0.6jt. T h e w indow len gth s selected are L = 25, 50, and 100. N o te that o>o and are not resolvab le for L = 25 and 50, but th ey are resolvab le for L = 100. T o reduce leak age, w e can select a data w indow w( n) that h as low er sid elob es in the frequency d om ain com pared with the rectangular w in d ow . H ow ever, as we d escribe in m ore detail in Chapter 8, a reduction o f the sid elo b es in a w indow W( w) is obtained at the ex p en se o f an increase in the width o f th e m ain lo b e o f W(a>) and h en ce a loss in resolution. T o illustrate this point, let us consider the H anning w indow , w hich is specified as w( n ) = ^(1 - cos jTTj-n), 0, 0 < n < i. 1 otherw ise (5.4.9)
Figure 5.14 show s |A"(<y)| for the w indow o f (5.4.9). Its sid elo b es are significantly sm aller than th o se o f the rectangular w in d ow , but its m ain lob e is approxim ately tw ice as w ide. Figure 5.15 sh ow s the spectrum o f the signal in (5.4.8), after it is w indow ed by the H anning w indow , for L = 50, 75, and 100. T h e reduction o f the sid elo b es and th e d ecrease in the resolu tion , com pared with the rectangular w indow , is clearly evident. For a general signal seq u en ce ljr(n)}, the frequency-dom ain relationship be tw een the w indow ed seq u en ce i ( n ) and the original seq u en ce x ( n ) is given by the con volu tion form ula (5.4.10) T h e D F T o f the w in d ow ed seq u en ce x( n) is the sam pled version o f the spectrum X(a>). T hus w e have X( k ) = X ( ) U W (5.4.11) k = 0,1,..., N - 1 Just as in the case o f th e sinusoidal seq u en ce , if the spectrum o f th e w indow is relatively narrow in w idth com pared to the spectrum X (to) o f th e signal, the win dow function has o n ly a sm all (sm ooth in g) effect on the spectrum X (w). O n the other hand, if the w in d ow function has a w ide spectrum com pared to the w idth of
6
25
0
Frequency
Figure 5.14 Magnitude spectrum of tbe Hanning window.
Sec. 5.4
437
Magnitude
L = 100
; 6
-r
- T
2
0
Frequency
(O Figure 5.15 Magnitude spectrum of the signal in (5.4.8) as observed through a Hanning window. X(a>), as w ould b e the case w hen the num ber o f sam ples L is sm all, the w indow spectrum m asks the signal spectrum and, con sequ en tly, the D F T o f the data re flects th e spectral characteristics o f the w in d ow function. O f course, this situation should b e avoided. Example 5.4.1 The exponential signal *..(0
H r
t >o t <0
is sampled at the rate F, = 20 samples per second, and a block of 100 samples is used to estimate its spectrum. Determine the spectral characteristics of the signal x(t) by computing the DFT of the finite-duration sequence. Compare the spectrum of the truncated discrete-time signal to the spectrum of the analog signal. Solution The spectrum of the analog signal is Xa(F) = 1 1 + ;2 jtF
The exponential analog signal sampled at the rate of 20 samples per second yields
438
The Discrete Fourier Transform: Its Properties and Applications the sequence x(n) = <r"r = c - '/ 20, n > 0
Chap. 5
Now, let 0 < n < 99 (0.95)", 0, otherwise The A/-point D F T of die L = 100 point sequence is
99
X(k) = Y
i<-n '> e~i2*k/N
* = 0 ,1 ........A ' - l
To obtain sufficient detail in the spectrum we choose N = 200. This is equivalent to padding the sequence x ( n ) with 100 zeros. The graph of the analog signal xa(t) and its magnitude spectrum |Xa(F )| are illustrated in Fig. 5.16(a) and (b), respectively. The truncated sequence xin) and its N = 200 point D FT (m agnitude) are illustrated in Fig. 5.16(c) and (d), respectively.
1.0
0.8 0.6 0.4 0.2
(a)
_
-5 0
-4 0
-3 0
-2 0 -1 0
JL
0
i
10 20 30
40
50
(b)
Figure 5.1ti Effect o f windowing (truncating) the sampled version o f the analog signal in Example 5.4.1.
Sec. 5.4
439
(d)
(*) Figure 5.16 Continued In this case the D FT {X (Jt)} bears a close resem blance to the spectrum of the analog signal. The effect of the window function is relatively small. O n the other hand, suppose that a window function of length L = 20 is selected. T hen the truncated sequence x(n) is now given as .() _ [ ( - 95)"1 0, 0 < n < 19 otherwise
440
Chap. 5
Its N = 200 point DFT is illustrated in Fig. 5.16(e). Now the effect of the wider spectral window function is dearly evident. First, the main peak is very wide as a result of the wide spectral window. Second, the sinusoidal envelope variations in the spectrum away from the main peak are due to the large sidelobes of the rectangular window spectrum. Consequently, the DFT is no longer a good approximation of the analog signal spectrum.

T h e m a jo r focus o f th is c h a p te r w as o n th e d isc re te F o u rie r tra n s fo rm , its p ro p e rtie s an d its ap p licatio n s. W e d e v e lo p e d th e D F T by sa m p lin g th e sp e c tru m X (&>) o f th e se q u en ce x(n'). F re q u e n c y -d o m a in sa m p lin g o f th e sp e c tru m o f a d is c re te -tim e signal is p a r ticu larly im p o rta n t in th e p ro cessin g o f d ig ital signals. O f p a r tic u la r significance is th e D F T , w hich w as show n to u n iq u e ly re p re s e n t a fin ite -d u ra tio n se q u e n c e in th e fre q u e n c y d o m a in . T h e ex iste n ce o f c o m p u ta tio n a lly efficien t a lg o rith m s fo r th e D F T , w hich a re d e sc rib e d in C h a p te r 6, m a k e it p o ssib le to d ig itally p ro cess sig n als in th e fre q u e n c y d o m a in m u ch fa ste r th a n in th e tim e d o m a in . T h e p ro cessin g m e th o d s in w hich th e D F T is esp ecially su ita b le in clu d e lin e a r filtering as d e sc rib e d in th is c h a p te r a n d c o rre la tio n , an d sp e c tru m an aly sis, w h ich a re tre a te d in C h a p te rs 6 a n d 12. A p a rtic u la rly lucid an d co n cise tre a tm e n t o f th e D F T and its ap p licatio n to fre q u e n c y an aly sis is given in th e b o o k by B rig h a m (1988).
PROBLEMS
5.1 The first five points of the eight-point DFT of a real-valued sequence are (0.25, 0.125 - j 0.3018, 0, 0.125 - y0.0518, 0}. Determine the remaining three points.
5.2 Compute the eight-point circular convolution for the following sequences. (a) *,() = (1,1,1,1,0,0,0,0}
. 3jt x2(n) = sin -n
8
_ 0 < n < 7
(b) *,() = (i)"
0 <n <7 3n . _ *2(n) = cos n 0< n < 7 O (c) Compute the DFT of the two circular convolution sequences using the DFTs of *i(n) and X2 (n).
S 3 Let X (Jt), 0 < We define
Jt <
N 1, be the Appoint DFT of the sequence *(n), 0 < n < N 1. 0 < k < k c, N - k c < k < N - 1 kc < k < N ke
y /ia _ f {) 1 0,
and we compute the inverse //-point DFT of X(k), 0 < k < N - 1. What is the effect of this process on the sequence x ()? Explain.
Chap. 5
Problems
441
5.4 F or the sequences jci(n) = cos ^ - n N jc2() = sin N n 0 <n <N - 1
determ ine the N -point: (a) Circular convolution jc^n) (N) x 2(n) (b) Circular correlation of *i(n) and x 2(n) (c) Circular autocorrelation of xi(n) (d) Circular autocorrelation of x 2(n) 5.5 Com pute the quantity N-l y ^ x ](n)x 2(n) =0 for the following pairs of sequences. (a) JC](n) = x 2 (n) = cos ~ n N 0 < n < N 1
(b) Xi (n) = cos n
x 2(n) = sin n 0 < n < N 1 N N (c) *](n) = 6(n) + 5(n - 8) *2() = ( ) - u(n - N)
5.6 D eterm ine the A/-point D FT of the Blackman window if(n) = 0.42 0.5 cos - -I- 0.08c o s ------ A r 1 A '- l 0 < n < N 1
5.7 If X (Jt) is the D FT of the sequence *(n), determ ine the Af-point D FTs of the sequences 2nkn xc(n) x(n) c o s ------N and . 2nkn x,(n) = x ( n ) sin A r in terms of X (Jt). 5.8 Determine the circular convolution of the sequences *i(n) = {1,2,3,1} t x2(n) = {4,3,2,2} t using the time-domain formula in (5.2.39). 5.9 Use the four-point DFT and IDFT to determine the sequence X3(n) = *i(n)()*2(*) where xi(n) and x 2(n) are the sequence given in Problem 5.8. 5.10 Compute the energy of the N -point sequence 0 < n < N 1 . . . 0< n < N - 1
442
Chap. 5
5.11 Given the eight-point D FT of the sequence *() = compute the D FT of the sequences: 1, n= 0 (a) x\(n) = 0, 1< n < 4 1. 5 <n < 7 0, 0 <n < 1 (b) jr2(n) = 1 , 2 <n < 5 0, 6< n < 7 5.12 Consider a finite-duration sequence jc(n) = {0 ,1 .2 ,3 .4 ) t (a) Sketch the sequence .t(n) wilh six-poini DFT (*) = W j XOfc) k = 0 .1 ........6 1. 0, 0 <n < 3 4 <n < 7
(b) D eterm ine the sequence y(n) with six-point D FT K(jt) = Re |X(A)|. (c) Determ ine the sequence u(w) with six-point D FT V(k) = Im |X(A-)|. 5.13 Let x p(n) be a periodic sequence with fundamental period N. Consider the following DFTs: x p(n) * * *,(*)
N
DFT
x ()
D FT
3^
X ,(k)
(a) W hat is the relationship between Xi(^) and XiOt)? (b) Verify the result in part (a) using the sequence xp(n) = { 1. 2 ,1 ,2 ,1 ,2 ,1 .2 - ] t 5.14 Consider the sequences X] (n) = {0,1, 2,3,4} t x 2(n) = {0 ,1 ,0 ,0 .0 ] t s(n) = {1. 0, 0 .0 ,0 )
t
and their 5-point DFTs. (a) Determine a sequence y( n ) so that Y(k) = A-! fJt). (b) I s there a sequence x3(n) such that S(k) = X^(k)X^(k)'! 5.15 Consider a causal LTI system with system function
The output y(n) of the system i s known for 0 < n < 63. Assuming that H(z) i s available, can you develop a 64-point DFT method to recover the sequence x{n), 0 < n < 63? Can you recover a l l values of x(n) in t h i s interval? 5.16* The impulse response of an LTI system i s given by h(n) = S(n) - ji(n - / t o ) . To determine the impulse response g(n) of the inverse system, an engineer computes the Af-point D FT H(k), N = 4ko, of h(n) and then defines g(n) as the inverse DFT of
Chap. 5
Problems
443
G (k ) = 1j H ( k ), k = 0, 1, 2 , . . . , A' - l . Determine g(n) and the convolution h(n)*g(n), and comment on whether the system with impulse response g(n) is the inverse of the system with impulse response h(n). 5.17* Determine the eight-point DFT of the signal *() = [ 1 , 1 , 1 . 1 . 1 , 1 , 0 , 0 } and sketch its magnitude and phase. 5.18 A linear time-invariant system with frequency response H(u>) is excited with the periodic input O O jc(n) = ^ S(n kN) *=-oo Suppose that we compute the Af-point DFT Y(k) of the samples >(), 0 < it < N - 1 of the output sequence. How is ^(Jt) related to //(w)? 5.19 D F T o f real sequences with special symmetries (a) Using the symmetry properties of Section 5.2 (especially the decomposition prop erties), explain how we can compute the DFT of two real symmetric (even) and two real antisymmetric (odd) sequences simultaneously using an W-point DFT only. (b) Suppose now that we are given four real sequences *,(), i = 1, 2, 3, 4, that are all symmetric [i.e., X j ( n ) = x j ( N n ), 0 < n < N 1], Show that the sequences
j,-(n) = X j ( n + 1) X j ( n 1)
are antisymmetric [i.e., s, (n) = s;(N - n) and 5, (0 ) = 0 ]. (c) Form a sequence *(/i) using Jti(n), J C 2 (n), ^(n), and ^(n) and show how to compute the DFT Xj(k) of x,(n), i = 1, 2, 3, 4 from the N-point DFT X(k) of x(n). (d) Are there any frequency samples of X, (k) that cannot be recovered from X(k)? Explain. 5.20 D F T o f real sequences with o dd harmonics only Let x(n) be an A'-point real sequence with Af-point DFT X(k) (N even). In addition, x(n) satisfies the following symmetry property: / N \ x ( n + y j = -x (n ) N n = 0 , 1 ........y - 1
that is, the upper half of the sequence is the negative of the lower half. (a) Show that X (k) = 0 k even
that is, the sequence has a spectrum with odd harmonics.
(b) Show that the values of this odd-harmonic spectrum can be computed by evaluat
ing the A /2-point DFT of a complex modulated version of the original sequence x(n). 5.21 Let x(t) be an analog signal with bandwidth B = 3 kHz. We wish to use a N = 2"point DFT to compute the spectrum of the signal with a resolution less than or equal to 50 Hz. Determine (a) the minimum sampling rate, (b) the minimum number of required samples, and (c) the minimum length of the analog signal record.
444
Chap. 5
5.22 Consider the periodic sequence

x p(n) = COS n
2x
oo < n < oo
with frequency /o = ^ and fundamental period N = 10. Determine the 10-point DFT of the sequence x ( n ) = x p(n), 0 < n < Af - 1.
5.23 Compute the W-point DFTs of the signals

(a) >:(") = ^()
(b) x ( n) = S(n no)
(c) Jt(n) = a "
0 < fiq < N
0 < n < N 1
(d) x(n)
1,
, Q
0 < n < N/ 2 - l ( N even) N p.< n< N
(e) x(n) = ea*IN )k< > 0<n<N - 1 2x (I) x(n) = cos kon 0 < n < N 1 N (g) j:(/i) = sin -j^kort 0< n < N - 1
i I n even ) ^ 10, odd 0 < n < Af 1 5.24 Consider the finite-duration signal x(n) = {1, 2, 3, 1}
(a) Compute its four-point DFT by solving explicitly the 4-by-4 system of linear
equations defined by the inverse DFT formula. (b) Check the answer in part (a) by computing the four-point DFT, using its defini tion.
5.25 (a) Determine the Fourier transform X (tu) of the signal

x(n) =
{ 1 ,2 ,3 ,2 ,1 ,0 1
(b) Compute the 6-point DFT V (k) of the signal

i(n) = { 3 ,2 ,1 ,0 ,1 ,2 }
(c) Is there any relation between X(w) and V(Jt)? Explain.
5.26 Prove the identity

Y & ( n + l N ) = Y eJQ*/,'*K JO C km C (Hint: Find the DFT of the periodic signal in the left-hand side.) 5 J 7 Computation o f the even and odd harmonics using the DFT sequence with an Appoint DFT X(k) ( N even) (a) Consider the time-aliased sequence
x ( n + I M ),
Let x(n) be an Appoint
0 < n < M 1
fv-00 0, elsewhere What is the relationship between the Af-point DFT K(Jt) of y(/i) and the Fourier transform X(w) of x(n)?
Chap. 5
Problems
445
(b) Let J x (n ) + x ( n + ~ l ' I 0, and y(n) D FT K(Jt) N/l
(-*?)
0 < n < N 1
elsewhere
Show that AT(Jt) = Y(k/2), k = 2, 4 , . . . , N - 2. (c) Use the results in parts (a) and (b) to develop a procedure that computes the odd harmonics of X(k) using an jV/2-point DFT. 528"' Frequency-domain sampling Consider the following discrete-time signal C ( n ) = h '" ' } 0, where a = 0.95 and L = 10 (a) Com pute and plot the signal x(n). (b) Show that X (w )= x(n)e = x(0) + 2 ^ ^ x(n) cos wn
|n| > L
Plot X{i) by computing it at w = jri/1 0 0 , k = 0, 1........100. (c) Com pute N \N )
for N = 30. (d) D eterm ine and plot the signal x(n) = *=o W hat is the relation between the signals x(n) and x ( n)l Explain. (e) Com pute and plot the signal ii(n ) = X x(n - IN), - L < n < L for N = 30. Com pare the signals x(rt) and X](n). (f) R epeat parts (c) to (e) for N = 15.
5.29* Frequency-domain sampling The signal x(n) = a 1 "1 , - 1 < a < 1 has a Fourier
transform X(w) = 1 1 2a cos < o + a2
(a) Plot X(w) for 0 < w < 2jt, a = 0.8. Reconstruct and plot X(w) from its samples X Qj r k / N) , 0 < k < N - 1 for: (b) N = 20 (c) N = 100 (d) Com pare the spectra obtained in parts (b) and (c) with the original spectrum X(a>) and explain the differences. (e) Illustrate the time-domain aliasing when N = 20.
446
Chap. 5
5 J0 * Frequency analysis o f amplitude-modulated discrete-time signal
T he discrete-time
(a) (b) (c) (d) (e)
Sketch the signals jr(n), xc(/i), and *m(n), 0 < n < 255. Com pute and sketch the 128-point D FT of the signal 0 < n < 127. Compute and sketch the 128-point D FT of the signal xtm(n), 0 < n < 99. Com pute and sketch the 256-point D FT of the signal jcani{n), 0 < n < 179. Explain the results obtained in parts (b) through (d), by deriving the spectrum of the am plitude-m odulated signal and comparing it with the experim ental results.
5.31* The sawtooth waveform in Fig. P5.31 can be expressed in the form of a Fourier series as
)
(a) D eterm ine the Fourier series coefficients c*. (b) Use an N -point subroutine to generate samples of this signal in the time domain using the first six term s of the expansion for N 64 and N = 128. Plot the signal x(f) and the samples generated, and com m ent on the results.
Figure P531 5 3 2 Recall that the Fourier transform of x (r) = eJ0* is X ( j i 2) = 2jtS(2 - i2o) and the Fourier transform of 0 < t < To otherwise is
e-jOT<i/l
(a) D eterm ine the Fourier transform Y ( j n ) of y(r) = p(t)eja' and roughly sketch |K0 '2)| versus 2.
Chap. 5
Problems
447
(b) Now consider the exponential sequence

jr(n) = where <uo is some arbitrary frequency in the range 0 < ojo < tt radians. Give the most general condition that a> o must satisfy in order for x(n) to be periodic with period P (P is a positive integer). (c) Let y(n) be the finite-duration sequence
v (n ) =
x ( n ) w N (n)
e iw " u i s ( n )
where w^( n) is a finite-duration rectangular sequence of length N and where x(n) is not necessarily periodic. D eterm ine Y(a)) and roughly sketch \Y(a>)\ for 0 < to < 2n. W hat effect does N have in | y ( o j ) | ? Briefly com ment on the similarities and differences between jY(a>)\ and ]y(j'S2)|. (d> Suppose that *(n) = e> {1' ! p)n and y(n) = w N(n)x(n) where N = IP, I a positive integer. Determ ine and sketch the N-point D FT of y(n). R elate your answer to the characteristics of |K{w)|. Is the frequency sampling for the D FT in part (d) adequate for obtaining a rough approxim ation of |K(w)| directly from the magnitude of the D FT sequence |K(/t)|? If not. explain briefly how the sampling can be increased so that it will be possible to obtain a rough sketch of |K(o)| from an appropriate sequence |y (Jt) | . p a positive integer
Efficient Computation of the DFT: Fast Fourier Transform Algorithms
A s w e h av e o b se rv e d in th e p reced in g c h a p te r, th e D isc re te F o u rie r T ra n sfo rm (D F T ) play s an im p o rta n t ro le in m an y a p p lic a tio n s o f digital signal processing, in clu d in g lin e a r filterin g , c o rre la tio n an aly sis, a n d sp e c tru m analysis. A m ajor re a so n fo r its im p o rta n c e is th e ex isten ce o f efficient a lg o rith m s fo r c o m p u tin g the DFT. T h e m ain to p ic o f this c h a p te r is th e d e sc rip tio n o f c o m p u ta tio n a lly efficient a lg o rith m s fo r e v a lu a tin g th e D F T . T w o d iffe re n t a p p ro a c h e s a re d e sc rib e d . O n e is a d iv id e -a n d -c o n q u e r a p p ro a c h in w hich a D F T o f size N , w h e re jV is a c o m p o site n u m b e r, is re d u c e d to th e c o m p u ta tio n o f sm a lle r D F T s fro m w hich th e larg er D F T is co m p u te d . In p a rtic u la r, w e p r e s e n t im p o rta n t c o m p u ta tio n a l alg o rith m s, called fast F o u rie r tra n sfo rm (F F T ) a lg o rith m s, fo r c o m p u tin g th e D F T w h en the size N is a p o w e r o f 2 a n d w h e n it is a p o w e r o f 4. T h e se co n d a p p ro a c h is b a s e d o n th e fo rm u la tio n o f th e D F T as a lin ear filterin g o p e ra tio n o n th e d a ta . T h is a p p ro a c h le a d s to tw o a lg o rith m s, th e G o ertzel alg o rith m an d th e ch irp -z tra n sfo rm a lg o rith m fo r co m p u tin g th e D F T via linear filterin g o f th e d a ta se q u e n c e .
6.1 EFFICIENT COMPUTATION OF THE DFT: FFT ALGORITHMS

In th is sectio n w e p re s e n t se v e ra l m e th o d s fo r c o m p u tin g th e D F T efficiently. In view o f th e im p o rta n c e o f th e D F T in v a rio u s d ig ital sig n a l p ro cessin g ap p licatio n s, such as lin e a r filtering, c o rre la tio n an aly sis, a n d s p e c tru m analysis, its efficien t c o m p u ta tio n is a to p ic th a t h a s re c e iv e d c o n s id e ra b le a tte n tio n by m any m a th e m a tic ia n s, e n g in e e rs, a n d a p p lie d scien tists. B asically , th e c o m p u ta tio n a l p ro b le m fo r th e D F T is t o c o m p u te th e se q u en ce {X(*)} o f N c o m p lex -v alu ed n u m b e rs g iv en a n o th e r se q u e n c e o f d a ta (x(n)} of 448
Sec. 6.1
Efficient Computation of the DFT: FFT Algorithms
449
le n g th N , a c c o rd in g to th e fo rm u la
N- 1
* (* ) = V w h ere
Jt(n) W N kn
0 < k < N -l
(6.1.1)
WN = e - }7* ,N In g e n e ra l, th e d a ta se q u e n c e x ( n ) is also a s su m e d to b e c o m p lex v alu ed . S im ilarly , th e I D F T b ec o m e s 1 n -i jr(n ) = ^ X ( J k ) W ^ " * ^ *=o 0 < n < N I
(6.1.2)
(6.1.3)
Since th e D F T a n d ID F T involve b asically th e sa m e ty p e o f c o m p u ta tio n s, o u r d iscu ssio n o f efficien t c o m p u ta tio n a l a lg o rith m s fo r th e D F T a p p lie s as w ell to th e efficien t c o m p u ta tio n o f th e ID F T . W e o b se rv e th a t fo r each v alu e o f k , d ire c t c o m p u ta tio n o f X ( k ) involves N co m p lex m u ltip lic a tio n s ( 4 N real m u ltip lic a tio n s) a n d N 1 co m p lex a d d itio n s (4 JV -2 re a l a d d itio n s). C o n se q u e n tly , to c o m p u te all N v alu es o f th e D F T re q u ire s jV2 c o m p lex m u ltip lic atio n s a n d N 2 N c o m p lex ad d itio n s. D ire c t c o m p u ta tio n o f th e D F T is b asically in efficien t p rim a rily b e c a u se it d o e s n o t e x p lo it th e sy m m etry a n d p e rio d ic ity p ro p e rtie s o f th e p h a s e fa c to r WV In p a rtic u la r, th e s e tw o p r o p e rtie s are: S y m m etry p ro p e rty : P e rio d ic ity p ro p e rty : Wk N+N/2 = W N L W#+N = W N k (6.1.4) (6.1.5)
T h e c o m p u ta tio n a lly efficient a lg o rith m s d e s c rib e d in th is se c tio n , k n o w n collec tiv ely as fa st F o u rie r tra n sfo rm (F F T ) a lg o rith m s, e x p lo it th e s e tw o b asic p ro p e rtie s o f th e p h a s e facto r.
6.1.1 Direct Computation of the DFT

F o r a co m p le x -v a lu e d se q u e n c e x ( n ) o f N p o in ts, th e D F T m a y b e ex p re sse d as X R{k) = Y |* * ( n ) c o s ^ ~ p - + x / ( n ) s i n ^ ^ - j (6.1.6)
Xi(k) = - Y
|x ,f ( n ) s in
- x , ( n ) co s
(6.1.7)
T h e d ire c t c o m p u ta tio n o f (6.1.6) a n d (6.1.7) re q u ire s : L 2 N 2 e v a lu a tio n s o f trig o n o m e tric fu n ctio n s. 2. 4 N 2 re a l m u ltip lic atio n s.
450
Chap. 6
3. AN ( N 1) real additions.
4. A num ber o f indexing and addressing op eration s.

T hese operations are typical o f D F T com putational algorithm s. T he operations in item s 2 and 3 result in the D F T values X K(k) and X i ( k ) . T h e indexing and addressing op eration s are necessary to fetch the data x(n'), 0 < < Ar 1, and the phase factors and to store th e results. T h e variety o f D F T algorithm s optim ize each o f these com p utational p rocesses in a different way.
6.1.2 Divide-and-Conquer Approach to Computation of the DFT

T h e d evelop m en t o f com p utationally efficient algorithm s for the D F T is m ade pos sible if w e adopt a divide-and-conquer approach. T his approach is based on the d ecom p osition o f an Af-point D F T in to su ccessively sm aller D F T s. T his basic ap proach leads to a fam ily o f com putationally efficient algorithm s k n ow n collectively as FFT algorithm s. T o illustrate the basic n otion s, let us consider the com p utation o f an Appoint D F T , where N can be factored as a product o f tw o integers, that is, N = LM (6.1.8)
T h e assum ption that N is not a prim e num ber is n ot restrictive, sin ce w e can pad any sequ en ce with zeros to ensure a factorization o f the form ( 6 . 1 .8 ). N o w the seq u en ce j ( n ) , 0 < n < N 1, can be stored in either a one dim ensional array in d exed by n or as a tw o-d im en sion al array in d exed by I and m, w here 0 < / < L 1 and 0 < m < A / - l a s illustrated in Fig. 6.1. N o te that / is the row index and m is the colum n index. T hus, the se q u en ce x (n ) can b e stored in a rectangular array in a variety o f w ays, each o f which d ep en d s on the mapping o f index n to the in d exes (/, m). For exam p le, su p p ose that w e select the m apping n = Ml + m (6.1.9)
T his leads to an arrangem ent in which the first row consists o f th e first M elem ents o f x ( n ) , the secon d row consists o f the n ext M elem en ts o f x ( n ) , and so on, as illustrated in Fig. 6.2 (a ). O n the other hand, the m apping n l + mL (6.1.10)
stores the first L elem en ts o f x ( n ) in the first colu m n, the n ext L elem en ts in the second colum n, and so on , as illustrated in Fig. 6.2(b ). A sim ilar arrangem ent can be u sed to store th e com p u ted D F T valu es. In particular, the m apping is from the in d ex it to a pair o f in d ices (p , q), where 0 < p < L 1 and 0 < q < M 1. If w e se lect the m apping
k Mp
(6 .1 .1 1 )
Sec. 6.1

n ---------0 1
x(\) M2)
451
N
...
- 1
JcCW -1)
(a) column index row index
K
0 1 2
0 *(0,0)
* 1 .0 )
1 x(0,1)
* 1 ,1 )
Af-1
*(2,0)
*2,1)
L- 1
(b)
Figure 6.1 N -l.
T w o dim ensional data array for storing the sequence
x(n).
0 < n
the D F T is stored on a row -w ise basis, w here the first row contains the first M elem en ts o f the D F T X ( k ) , the second row contains the next set of M elem en ts, and so on . O n the other hand, th e m apping k = qL + p
(6 . 1. 12)
results in a colu m n-w ise storage o f X (Jt), w here the first L elem en ts are stored in the first colu m n, th e secon d se t o f L elem en ts are stored in the secon d colum n, and so on. N o w su p p ose that x ( n) is m apped in to the rectangular array x ( l , m ) and X( k ) is m app ed in to a corresp on d ing rectangular array X ( p , q). T h en the D F T can be exp ressed as a d o u b le sum o ver th e elem en ts o f the rectangular array m ultiplied b y th e corresp on d ing p h ase factors. T o b e specific, let us ad op t a colum n-w ise m apping fo r x ( n ) g iven by (6.1.10) and th e row -w ise m apping for the D F T given by (6.1.11). T hen X ( p , q) = Y Y x{1' m ) W < p+qHmL+t) rrtacO 1=0 But
p+ q){m L+ i)
(6.1.13)
^M L rnp ^m L q
(6.1.14)
= W[l
H o w ev er, W%mp = 1, W%*L = W $ L = W**, and W * pl = W?'
452

Row-wise
n
Chap.
Ml + m
M- 1 0
1
*0)
MM)
M D
*2)
M M + 2)
M M -
1)
MM
+ 1)
1)
M2M
- 1) 1)
M 2M )
M 2M +
M 2M
+ 2)
M 'iM -
L - 1 x( (L- IW) M(L - 1 )Af + 1) x L -l)M + 2) (a)

Column-wise
x{L M -
1)
M- 1
0 1 2
MO ) ML) M 2L) x((M
1) L ) I )L + 1)
jt(1)
M2)
M L+
1)
x(2L
+ 1)
x((M -
M L + 2)
M 2 L + 2)
M (M -l)L + 2 )
L -l
x(L
- I)
M U -
1)
M IL
- 1) (b)
MLM -
1)
Figure <L 2 Two arrangements for the data arrays. W ith th ese sim p lification s, (6.1.13) can be exp ressed as L- l X( p, q) = Y /*0 (6.1.15) ^|R0
The expression in (6.1.15) in volves the com p utation o f D F T s o f length M and length L. T o elab orate, let us subdivide th e com p utation into th ree steps: L First, w e com p ute th e M -point D F T s M-l F(l,q) = x ( / ,m ) W ^ \ 0 < q < M - 1
( 6 .1 .1 6 )
for each of the rows I = 0 ,1 ........ L - l .
Sec. 6.1
453
2 . S econ d , w e com p ute a new rectangular array G ( l , q ) defined as (6.1.17) 3. F inally, w e com p ute the L -point D F T s L-l X ( p , q ) = J 2 G ^l ' ^ WL for each colum n q = 0 , 1 , . . . , M 1, o f the array G (l, q). O n the surface it m ay appear that the com putational procedure outlined above is m ore com p lex than the direct com p utation o f the D F T . H ow ever, let us evalu ate the com putational com p lexity o f (6.1.15). T h e first step in volves the com p utation o f L D F T s, each o f M points. H en ce this step requires L M 2 com plex m ultiplications and L M { M 1) com p lex additions. T h e secon d step requires L M com p lex m ultiplications. Finally, the third step in the com p utation requires M L 2 com p lex m ultiplications and M L ( L 1) com p lex additions. T h erefore, the com p utational com plexity is C om plex m ultiplications: C om plex additions: N ( M + L + 1) (6.1.19) N ( M + L 2) w here N = M L . T hus the num ber o f m ultiplications has b een reduced from N 2 to N ( M + L + 1 ) and the num ber o f additions has b een reduced from N ( N 1) to N ( M + L 2). For exam p le, suppose that N = 1000 and w e select L = 2 and M = 500. T h en , instead o f having to perform 106 com p lex m ultiplications via direct com pu tation o f the D F T , this approach leads to 503,000 com p lex m ultiplications. This rep resents a reduction by approxim ately a factor o f 2. T h e num ber o f additions is also red u ced by ab ou t a factor o f 2. W h en N is a highly com p osite num ber, that is, N can be factored in to a product o f prim e num bers o f th e form N = r\ r 2 - r v (6.1.18)
(6.1.20)
then the d ecom p osition ab ove can b e rep eated (v - 1 ) m ore tim es. T his procedure results in sm aller D F T s, w hich, in turn, lead s to a m ore efficient com putational algorithm . In effect, th e first segm en tation o f th e seq u en ce x ( n ) in to a rectangular array o f M colu m ns w ith L elem en ts in each colu m n resu lted in D F T s o f sizes L and M . Further d eco m p o sitio n o f th e data in effect in volves th e segm en tation o f each row (or colu m n ) into sm aller rectangular arrays w hich result in sm aller D F T s. This p roced u re term inates w h en N is factored in to its prim e factors.
Example 6.1.1
To illustrate this computational procedure, let us consider the computation of an N = 15 point DFT. Since N = 5 x 3 = 15, we select L = 5 and M = 3. In other
454
Efficient Computation of the DFT: Fast Fourier Transform Algorithms words, we store the 15-point sequence *(n) column-wise as follows: Row Row Row Row Row 1: 2: 3: 4: 5: *(0, 0) *(1,0) *(2,0) *(3,0) * (4,0) = *(0) = * (1 ) = * (2 ) = * (3 ) = * (4 ) *(0.1) *(1, 1) x ( 2 , 1) *(3,1) *(4,1) = * (5 ) = jc(6) = x(7) = * (8 ) = * (9 ) *(0, 2) = *(10) *(1,2) = *(11) x(2,2)=x(12) *(3,2) = .r{13) *(4,2) = *(14)
Chap. 6
Now. we com pute the three-point DFTs for each of the five rows. This leads to the following 5 x 3 array: F(0, 0) F a , o) F( 2. 0) F( 3, 0) F(4. 0) F (O .l) F O .l) F( 2,1) FO. 1) F (4.1) F( 0.2) F (1.2) F( 2.2) FO- 2) F( 4.2)
The next step is to multiply each of the terms F(l, q) by the phase factors = M/jj. 0 < / < 4 and 0 < q < 2. This computation results in the 5 x 3 array: Column 1 G (0.0) G (1,0) G(2, 0) G (3.0) G (4 .0) Column 2 C(0. 1) C ( l.l) C(2. 1) G (3 ,1) G (4 .1) Column 3 C (0.2) G (l. 2) C ( 2 .2) G (3 .2) G (4.2)
The final step is to com pute the five-point DFTs for each of the three columns. This com putation yields the desired values of the D FT in the form X (0.0) X (1,0) X (2.0) X (3,0) X (4,0) = = = = = X(0) X(3) X (6) X (9) X (12) X (0 ,1) = X (1.1) = X (2 .1) = X (3,1) = X (4,1) = X (l) X (4) X(7) X(10) X (13) X (0,2) X (1,2) X (2,2) X (3 ,2) X (4 ,2) = = = = = X(2) X(5) X(8> X ( ll) X(14)
Figure 6.3 illustrates the steps in the computation. It is interesting to view the segm ented data sequence and the resulting D FT in term s of one-dim ensional arrays. When the input sequence x(n) and the output DFT X(jfc) in the two-dimensional arrays are read across from row 1 through row 5, we obtain the following sequences: IN PU T A R R A Y *(0) x{5) *(10) *(1) *(6) *(11) x(2) *(7) *(12) *(3) *(8) *(13) x(4) *{9) *(14) O U T PU T A R R A Y X(0) X (l) X(2) X(3) X(4) X(5) X(6) X(7) X(8) X(9) X(10) X ( ll) X(12) X(13) X(14) W e observe that the input data sequence is shuffled from the norm al order in the com putation of the D FT. On the other hand, the output sequence occurs in norm al order. In this case the rearrangem ent of the input data array is due to the
Sec. 6.1
455
Figure 63 DFTs.
Computation of N = 15-point DFT by means of 3-point and 5-point
segm entation of the one-dim ensional array into a rectangular array and the order in which the D FTs are com puted. This shuffling of either the input data sequence or the output D FT sequence is a characteristic of most FFT algorithms. T o sum m arize, the algorithm that w e have introduced in volves the follow in g com putations:
Algorithm 1
1. S tore the signal colu m n-w ise. 2. C om pute the Af-point D F T o f each row. 3. M ultiply the resulting array by the p h ase factors 4. C om p ute the L -point D F T o f each colum n 5. R ea d the resulting array row -w ise. A n additional algorithm with a sim ilar com putational structure can b e o b tained if th e input signal is stored row -w ise and the resulting transform ation is. colu m n-w ise. In this case w e select as n = Ml + m
k
(6.1.21)
= qL + p
This ch o ice o f in d ices lead s to the form ula for the D F T in the form
X {p ,q )
= * (/,
msO 1*0 M- l = E > c urmp
WN
(6. 1.22)
Thus w e obtain a secon d algorithm .
456
Efficient Computation of the DFT: Fast Fourier Transform Algorithms A lg o rith m 2
Chap. 6
1. S to re th e sig n al row -w ise. 2. C o m p u te th e L -p o in t D F T a t each co lu m n . 3. M u ltip ly th e re su ltin g a rra y by th e fa c to rs W%m. 4. C o m p u te th e A f-point D F T o f e a c h row . 5. R e a d th e re su ltin g a rra y colum n-w ise. T h e tw o a lg o rith m s given a b o v e h a v e th e sa m e c o m p lex ity . H o w e v e r, they d iffer in th e a rra n g e m e n t o f th e c o m p u ta tio n s. In th e follow ing se c tio n s w e exploit th e d iv id e -a n d -c o n q u e r a p p ro a c h to d e riv e fast a lg o rith m s w h e n th e size o f the D F T is re stric te d to b e a p o w e r o f 2 o r a p o w e r o f 4.
6.1.3 Radix-2 FFT Algorithms

In th e p re c e d in g se ctio n w e d e sc rib e d fo u r a lg o rith m s fo r efficien t c o m p u ta tio n of th e D F T b ased o n th e d iv id e -a n d -c o n q u e r a p p ro a c h . S uch an a p p ro a c h is applica b le w h en th e n u m b e r N o f d a ta p o in ts is n o t a p rim e . In p a rtic u la r, th e a p p ro ach is v ery efficien t w h en N is highly c o m p o s ite , th a t is, w h en N can b e fa c to re d as N = r\r2ry r v, w h e re th e {r,} are prim e. O f p a rtic u la r im p o rta n c e as th e case in w hich r i = r2 = r v = r , so th at N = r ' \ In such a case th e D F T s a re o f size r , so th a t th e c o m p u ta tio n o f the N -p o in t D F T h a s a re g u la r p a tte rn . T h e n u m b e r r is called th e rad ix o f th e F F T alg o rith m . In th is se c tio n w e d escrib e rad ix -2 a lg o rith m s, w hich a re by far th e m ost w idely u sed F F T alg o rith m s. R ad ix -4 a lg o rith m s a re d e sc rib e d in th e follow ing sectio n . L e t us c o n s id e r th e c o m p u ta tio n o f th e N 2 V p o in t D F T by th e dividea n d -c o n q u e r a p p ro a c h specified by (6.1.16) th ro u g h (6.1.18). W e select M = N / 2 a n d L = 2. T h is se lectio n re su lts in a sp lit o f th e N -p o in t d a ta se q u e n c e in to two // /2 - p o in t d a ta se q u e n c e s f \ ( n ) a n d f 2(ri), c o rre sp o n d in g to th e ev en -n u m b ered a n d o d d -n u m b e re d sa m p le s o f x ( n ), resp ec tiv ely , th a t is, / i ( n ) = x ( 2 n) f i ( n ) = x (2n + 1), N n = 0 ,1 ,..., ~ 1 (6.1.23)
T h u s f i ( n ) a n d f j ( n ) a re o b ta in e d by d e c im a tin g x ( n) b y a f a c to r o f 2, a n d hence th e re su ltin g F F T a lg o rith m is called a d e c im a tio n -in -tim e a lg o rith m . N o w th e Af-point D F T c a n b e e x p re ss e d in te rm s o f th e D F T s o f th e deci m a te d se q u e n c e s as follow s: A'-l X(k) = Y x ^ n*0 * = 0 ,1 ,..., A f-1
wn
Sec. 6.1
Efficient Computation of the DFT: FFT Algorithms =

n even <W/2)-l
457 (6.1.24)
x (n )H # + 5 3
n odd (Af/2)-l
= B ut
J ! x ( 2 m ) W ] f k 4- 5 3 *(2m + l)W * (2'n+1> m=0 m=0
= W ^/ 2- W ith this su b stitu tio n , (6.1.24) c a n b e e x p re ss e d as

(N/Zi-i (N/2) l
x(k)=
5 3
M m )w *ra + K
= fi(Jfc) + W * f 2(*)
it = 0 , 1 , . . . , W - l
w h e re F i(it) a n d F2 (k) a re th e N /2 -p o in t D F T s o f th e se q u e n c e s f \ ( m ) an d f 2 (m), resp ec tiv ely . S ince F\ ( k) a n d F 2 (k) a re p e rio d ic, w ith p e rio d N f l , w e h a v e F](A -f N / 2) = F i(* ) a n d /^(jfc + N f l ) = / i t * ) - 1 a d d itio n , th e fa c to r W ^ Nfl = Wfa. H e n c e (6.1.25) can b e ex p re sse d as X ( k ) = f i(J t) + W* F2(fc) * = 0 , 1 .........~ 1 (6.1.26)
= F , ( * ) - < F 2(/:)
* = 0 , 1 ....... y - 1
(6.1.27)
W e o b se rv e th a t th e d ire c t c o m p u ta tio n o f F\ ( k) re q u ire s ( N /2 )2 com plex m u ltip lic a tio n s. T h e sa m e a p p lie s to th e c o m p u ta tio n o f F2 (k). F u rth e rm o r e , th e re a re N f l a d d itio n a l co m p lex m u ltip lic a tio n s re q u ire d to c o m p u te W k N F 2 (k). H en ce th e c o m p u ta tio n o f X ( k ) re q u ire s 2 ( N f l )1 4- N f l = N 2/ 2 + N f l c o m p lex m u ltip li c a tio n s. T h is first s te p resu lts in a re d u c tio n o f th e n u m b e r o f m u ltip lic a tio n s from N 2 to N 1 f l + N f l , w h ich is a b o u t a fa c to r o f 2 fo r N large. T o b e c o n s iste n t w ith o u r p rev io u s n o ta tio n , w e m ay define </!(*) = F l(* ) G 2 (k) = W N l F2 (k) T h e n th e D F T X (it) m ay b e ex p re sse d as X ( k ) = G x(k) + G 2 (k) X(k + j ) = G x( k ) - G 2 (k) it = 0 , 1 ____ y - 1 (6.1.28) * = 0 , 1 .........y - 1 * = 0 ,l,...,y - l * = 0 , l , . . . , y 1
T h is c o m p u ta tio n is illu stra te d in Fig. 6.4. H a v in g p e rfo rm e d th e d e c im a tio n -in -tim e o n ce, w e c a n r e p e a t th e p ro cess fo r e a c h o f th e se q u e n c e s f \ ( n ) a n d f 2 (n). T h u s f \ ( n ) w o u ld re s u lt in th e tw o
458

JrW -2)
Chap. 6
* 0 ) x{2) x(4)
Figure 6.4 F irst step in the decim ation-in-tim e algorithm . /V /4-point se q u e n c e s v n (n ) = / ] ( 2n) v\ 2 (n) = f \ Q n + 1) an d f 2 (n) w o u ld yield V2\(n) = f 2 (2 n) V22(n) = /2(2n + 1) N n = 0, 1......... - 1 4 N n = 0 , 1......... - 1 N = 0 , 1 ................... 1 4 n = 0, 1.........j - 1
(6.1.29)
(6.1.30)
B y co m p u tin g jV /4-point D F T s , w e w o u ld o b ta in th e ///2 - p o in t D F T s Fi(Jfc) and F2 (k) fro m th e re la tio n s FiOfc) = V(Jfc) + W k Nf2 Vn (k) k = 0 ,1 , 1 (6.1.31)
F x (* + t ) = V l1{k ) - KpynW
F 2 (k) = V21(*) + W N k / 2 V22(k)
k=
1..... 7 - 1
- 1 (6.1.32)
k = 0 ,1 ,..., J N k = 0, . , , , j - l
F2 ( * + j )
= V2i (*) -
where the (Vi; (jt)} are the ///4 -p o in t D F T s o f th e seq u en ces {u,;(n)}.
Sec. 6.1
459
TABLE 6.1
COMPARISON OF COMPUTATIONAL COMPLEXITY FOR THE DIRECT COMPUTATION OF THE DFT VERSUS THE FFT ALGORITHM Number of Points, N 4 8 16 32 64 128 256 512 1,024 Complex Multiplications in Direct Computation, N2 16 64 256 1,024 4,096 16,384 65,536 262.144 1,048,576 Complex Multiplications in FFT Algorithm, (JV/2) log2 N 4 12 32 80 192 448 1,024 2,304 5,120 Speed Improvement Factor 4.0 5.3 8.0 12.8 21.3 36.6 64.0 113.8 204.8
W e o b se rv e th a t th e c o m p u ta tio n o f {V(J(*)} re q u ire s 4{W /4)2 m u ltip lic a tio n s a n d h e n c e th e c o m p u ta tio n o f F\ ( k) a n d F 2(Jt) can b e a c c o m p lish ed w ith N 2/ 4 + N f l c o m p lex m u ltip lic a tio n s. A n a d d itio n a l N f l co m p lex m u ltip lic a tio n s a re re q u ire d to c o m p u te X ( k ) fro m F i(it) a n d Fi{k), C o n se q u e n tly , th e to ta l n u m b e r o f m u ltip lic a tio n s is re d u c e d a p p ro x im a te ly by a fa c to r o f 2 ag ain to N 2/ 4 + N. T h e d e c im a tio n o f th e d a ta se q u e n c e can b e re p e a te d ag ain a n d ag ain until th e re su ltin g se q u e n c e s a re re d u c e d to o n e -p o in t se q u en ces. F o r N = 2 V, this d e c im a tio n can b e p e rfo rm e d v = log2 N tim es. T h u s th e to ta l n u m b e r o f co m p lex m u ltip lic a tio n s is re d u c e d to { N f l ) log2 N . T h e n u m b e r o f co m p lex a d d itio n s is N log2 N . T a b le 6.1 p re s e n ts a c o m p a riso n o f th e n u m b e r o f co m p lex m u ltip lic a tio n s in th e F F T a n d in th e d ire c t c o m p u ta tio n o f th e D F T . F o r illu stra tiv e p u rp o se s , Fig. 6.5 d e p ic ts th e c o m p u ta tio n o f an N = 8 p o in t D F T . W e o b se rv e th a t th e c o m p u ta tio n is p e rfo rm e d in th r e e stag es, b eg in n in g w ith th e c o m p u ta tio n s o f fo u r tw o -p o in t D F T s, th e n tw o fo u r-p o in t D F T s , an d
Figure <L5 Three stages in the computation o f an N = 8-point DFT.
460

Stage Stage 2 Stage 3
Chap. 6
X (0 )
X(l)
X(2)
X(3)
X(4)
X{5)
*(6)
X p )
finally, o n e eigh t-p oin t D F T . T he com bination o f the sm aller D F T s to form the larger D F T is illustrated in Fig. 6. 6 for N = 8 . O bserve that the basic com putation perform ed at every stage, as illustrated in Fig. 6 .6 , is to take tw o com p lex num bers, say th e pair (a, b), m ultiply b by W N r, and then add and subtract the product from a to form tw o new com p lex numbers (A, B ). This basic com putation, which is show n in Fig. 6.7, is called a butterfly b ecau se the flow graph resem bles a butterfly. In general, each butterfly involves on e com p lex m ultiplication and tw o com plex additions. F or N = 2 V , there are N f l butterflies per stage o f th e com putation p rocess and log 2 N stages. T h erefore, as previously indicated th e total num ber of com plex m ultiplications is ( N f l ) log 2 N and com p lex additions is Arlog 2 N . O nce a butterfly operation is perform ed on a pair o f com p lex num bers (a, b) to p roduce ( A , B ) , there is no n eed to 'sa v e the input pair ( a , b ) . H en ce w e can >A = a + W^b
F igure 6.7 Basic butterfly com putation in th e decim ation-in-tim e F FT algorithm .
B=a-Wt/b
Sec. 6.1
461
store th e result (A , B ) in the sam e location s as ( a , b ) . C on sequ en tly, w e require a fixed am ount o f storage, n am ely, 2 N storage registers, in order to store the results ( N com p lex num bers) o f the com p u tation s at each stage. Since th e sam e 2 N storage loca tio n s are used throughout the com p utation o f the JV-point D F T , w e say that the c o m p u ta tio n s are d o n e in place. A secon d im portant observation is con cern ed with the ord er o f the input data seq u en ce after it is d ecim ated (v - 1 ) tim es. For exam p le, if w e consider the case w h ere N = 8 , w e know that th e first d ecim ation yield s the seq u en ce jc(0), x ( 2 ) , x (4 ), * ( 6 ), * (1 ), Jt(3), jr(5), jc(7), and the secon d d ecim ation results in the seq u en ce jc(0), x (4 ), x (2 ), x ( 6 ), jt(1), x (5 ), jc(3), jc(7). T h is sh u fflin g o f the input data seq u en ce has a w ell-d efin ed order as can b e ascertained from observing Fig. 6 .8 , w hich illustrates the d ecim ation o f th e eigh t-p oin t seq u en ce. B y expressing the in d ex n, in the seq u en ce x ( n ) , in binary form , w e n o te that th e order o f the d ecim ated d ata seq u en ce is easily ob tain ed by reading the binary representation o f th e index n in reverse order. T hus the data p oin t jt(3) = *(011) is placed in position m = 110 or m = 6 in the d ecim ated array. Thus w e say that the data x ( n ) after d ecim ation is stored in bit-reversed order. W ith th e input data seq u en ce stored in bit-reversed order and the butterfly com p utations perform ed in p lace, the resulting D F T seq u en ce X ( k ) is ob tain ed in natural order (i.e., k = 0 , 1 , . . . , N 1). O n the oth er hand, w e should indi ca te that it is p ossib le to arrange the F F T algorithm such that the input is left in natural order and the resulting output D F T will occur in bit-reversed order. Furtherm ore, w e can im pose the restriction that b oth the input data x ( n ) and the output D F T X ( k ) be in natural order, and d erive an FFT algorithm in which the com p utations are not d on e in place. H en ce such an algorithm requires additional storage. A n o th e r im portant radix-2 FFT algorithm , called th e decim ation -in-freq u en cy algorithm , is o b tain ed by using the divide-and-conquer approach described in S ec tion 6.1.2 w ith th e ch oice o f M = 2 and L = N f l . T his ch oice o f param eters im plies a colu m n-w ise storage o f the input data seq u en ce. T o d erive the algo rithm, w e begin by splitting the D F T form ula in to tw o sum m ations, o n e o f which in v o lv es the sum over the first N t 2 data p oin ts and th e secon d sum in volves the last N I 2 data points. Thus w e obtain
W 2 > -1
N- 1
X (k) =
x( n)W *? + Y
x(n)W % (6.1.33)
Since WN/2 = (1)*, the exp ression (6.1.33) can b e rew ritten as
(6.1.34)
462

Data decimation 1 Memory address (decimal) (binary) Memory
Chap. 6
000
(ninino) (0 0 0) (0 0 1) (0 1 0) (0 1 1) (1 0 0) (1 0 !) (1 1 0) (1 1 1)
- * - - -*> -*> 4
(/To"2 () (0 0 0 ) (1 0 0 ) (0 0 1) (1 0 1 ) (0 10) (1 1 0 ) (0 1 1) (1 1 1 ) (b)
- - - * - -
(nnn in ;) (0 0 0 ) (1 0 0) (0 10) (t 10) (0 0 1) (1 0 1) (0 1 1) (1 1 ))
Figure 6Jt
Shuffling of the data and bit reversal.
N ow , let us split (d e c im a te ) X ( k ) in to th e ev en - a n d o d d - n u m b e re d sam p les. Thus w e o b ta in

(W /2)-l r
x(n) + x an d
(A 72)-l
,
v tr k n
K
/
N/2
k = Q, 1 , . . . , y - 1
(6.1.35)
JV \"1
X{2k + \) =
"
N * = 0 , 1 ......y - 1 (6.1.36)
w h ere w e h av e u se d th e fact th a t Wf, = 'Wsp.-
Sec. 6.1
463
If w e d e fin e th e N /2 -p o in t se q u e n c e s gi ( n) a n d gz(n) as g i( n ) = * ( ) + * (6.1.37) g 2 (n) = |* ( n ) - x th e n (N/2)-l x ( 2 k) = Y

n=0
+ y^)
n = 0 , 1 , 2 .........J
- 1
s m K )2 (6.1.38) & W WN/2
(AT/2)1
X (2* + l ) =
n=0
T h e c o m p u ta tio n o f th e s e q u e n c e s g i(n ) a n d g 2 (n) a c co rd in g to (6.1.37) a n d th e su b s e q u e n t u se o f th e s e se q u e n c e s to c o m p u te th e N /2 -p o in t D F T s a re d e p ic te d in Fig. 6.9. W e o b se rv e th a t th e b asic c o m p u ta tio n in th is figure in v o lv es th e b u tte rfly o p e r a tio n illu stra te d in Fig. 6.10. T h is c o m p u ta tio n a l p ro c e d u re can b e re p e a te d th ro u g h d e c im a tio n o f th e N /2 -p o in t D F T s , X ( 2 k ) a n d X ( 2 k + 1). T h e e n tire p ro cess in v o lv es v = log2 N
464

A = a + b
Chap. 6
wi,
B = ( a b)W f/
-
Figure 6.10 Basic butterfly com putation in the decim ation-in-frequency F FT algorithm .
stages o f decim ation , w here each stage in volves N t l butterflies o f the type shown in Fig. 6.10. C onsequently, the com putation o f the Af-point D F T via the decim ationin-frequency FFT algorithm , requires ( N / 2) iog 2 N com p lex m ultiplications and N lo g 2 N com p lex additions, just as in the d ecim ation -in-tim e algorithm . For il lustrative purposes, the eight-point d ecim ation-in-frequency algorithm is given in Fig. 6.11. W e observe from Fig. 6.11, that the input data x ( n ) occurs in natural order, but the output D F T occurs in bit-reversed order. W e also n ote that the com puta tions are perform ed in place. H ow ever, it is p ossib le to reconfigure the decim ationin-frequency algorithm so that the input seq u en ce occurs in bit-reversed order w hile the output D F T occurs in norm al order. F urtherm ore, if w e abandon the requirem ent that the com putations b e d on e in place, it is also p ossib le to have both the input data and the output D F T in norm al order.
Figure 6.11
N=
8-point decimation-in-frequency FFT algorithmn.
Sec. 6.1
465
6.1.4 Radix-4 FFT Algorithms

W h e n th e n u m b e r o f d a ta p o in ts N in th e D F T is a p o w e r o f 4 (i.e., N = 4 l ), w e can , o f c o u rse, alw ays use a radix-2 a lg o rith m fo r th e c o m p u ta tio n . H o w e v e r, fo r th is case, it is m o re efficien t c o m p u ta tio n a lly to em p lo y a rad ix -4 F F T alg o rith m . L e t u s b eg in by d escrib in g a rad ix -4 d e c im a tio n -in -tim e F F T a lg o rith m , w hich is o b ta in e d by se lectin g L = 4 an d M = N / 4 in th e d iv id e -a n d -c o n q u e r a p p ro a c h d e s c rib e d in S ectio n 6.1.2. F o r this ch o ice o f L a n d M , w e h av e /, p 0 ,1 , 2, 3: m, q = 0, 1.........N J4 - 1; n = 4m + /; a n d k = ( N / 4) p + q. T h u s w e sp lit o r d e c im a te th e W -point in p u t se q u e n c e in to f o u r su b s e q u e n c e s, x ( 4 n ), jc(4 n + 1), x( 4n + 2), x ( 4 n -f 3), n = 0, 1.........N / 4 1By a p p ly in g (6.1.15) w e o b ta in
3
* ( /> . ) =
[w ^F il'q ^W ?
0 ,1 .2 .3
(6.1.39)
w h ere F ( I . q ) is giv en by (6.1.16), th a t is.

(iV/4 |!
F(l.q)= an d
mq x ( l - n i ) W N/A
I = 0 .1 , 2. 3. = N .........4 - 1 (6.1.40)
x(l . m ) = x ( 4 m -j- /) (N X(p.q) = X / p + q
(6.1.41) (6.1.42)
T h u s, th e fo u r ///4 -p o in t D F T s o b ta in e d fro m (6.1.40) a re c o m b in e d a cco rd in g to (6.1.39) to yield th e W -point D F T . T h e ex p re ssio n in (6.1.39) fo r co m b in in g th e ///4 -p o in t D F T s d efin es a rad ix -4 d e c im a tio n -in -tim e b u tterfly , w hich can be e x p re ss e d in m atrix fo rm as ~X( Q , q ) X(\,q) X (2,q) -X(3,q)J
- 1 1 1 1 1 - 1 1 - 1 Li j -i - j
1 ' j -1 j
W F(0,<7) WF(Lq) W % F(2 ,q) w l qF { X q ) (6.1.43)
T h e ra d ix -4 b u tte rfly is d e p ic te d in Fig. 6 .1 2 (a) a n d in a m o re co m p a c t fo rm in Fig. 6 .1 2 (b ). N o te th a t since W = 1, e a c h b u tte rfly in v o lv es th r e e co m p lex m u ltip lic a tio n s , a n d 12 c o m p lex a d d itio n s. T h is d e c im a tio n -in -tim e p ro c e d u re can b e re p e a te d recu rsiv ely v tim es. H e n c e th e re su ltin g F F T alg o rith m co n sists o f v sta g es, w h e re e a c h sta g e c o n ta in s A74 b u tte rflie s . C o n s e q u e n tly , th e c o m p u ta tio n a l b u r d e n fo r th e a lg o rith m is 3 v N / 4 = (3jV /8) lo g ; N c o m p lex m u ltip lic a tio n s a n d O N f l ) log2 N co m p lex a d d itio n s. W e n o te th a t th e n u m b e r o f m u ltip lic a tio n s is re d u c e d by 2 5 % , b u t th e n u m b e r o f a d d itio n s h a s in c re a se d b y 50% fro m N log2 N to O N f l ) log2 N .
466
Chap. 6
Figure 6.12 Basic butterfly computation in o radix-4 FFT algorithm.
It is in te re stin g to n o te , h o w ev er, th a t by p e rfo rm in g th e a d d itio n s in tw o step s, it is p o ssib le to re d u c e th e n u m b e r o f a d d itio n s p e r b u tte rfly fro m 12 to 8. T h is can b e acco m p lish ed by e x p ressin g th e m atrix o f th e lin e a r tra n s fo rm a tio n in (6.1.43) as a p ro d u c t o f tw o m atric es as follow s: m ? n X ( l,9 ) X ( 2,q) -1 0 1 .0 0 1 0 1 1 0 -1 0 0 -j 0 j 1 1 0 .0 0 0 1 1 1 -1 0 0 0 0 1 -1 w jjm g ) WqF(\,q) W % F ( 2 .q) lW *F(3,q). (6.1.44)
N ow ea c h m atrix m u ltip lic a tio n involves fo u r a d d itio n s fo r a to ta l o f e ig h t ad d i tio n s. T h u s th e to ta l n u m b e r o f co m p lex a d d itio n s is re d u c e d to N log2 N , w hich is id en tical to th e ra d ix -2 F F T a lg o rith m . T h e c o m p u ta tio n a l sa v in g s re su lts from th e 25% re d u c tio n in th e n u m b e r o f co m p lex m u ltip lic atio n s. A n illu stra tio n o f a rad ix -4 d e c im a tio n -in -tim e F F T a lg o rith m is sh o w n in Fig. 6.13 fo r N = 16. N o te th a t in th is a lg o rith m , th e in p u t se q u e n c e is in norm al o r d e r w h ile th e o u tp u t D F T is shuffled. In th e rad ix -4 F F T a lg o rith m , w here th e d e c im a tio n is b y a f a c to r o f 4, th e o r d e r o f th e d e c im a te d se q u e n c e can be d e te rm in e d b y re v e rsin g th e o r d e r o f th e n u m b e r th a t re p re s e n ts th e in d ex n in a q u a te rn a ry n u m b e r sy stem (i.e., th e n u m b e r sy stem b a s e d o n th e digits 0, 1, 2, 3). A rad ix -4 d e c im a tio n -in -fre q u e n c y F F T a lg o rith m can b e o b ta in e d by se lect ing L = N / 4 , M = 4; /, p = 0, 1.........N / 4 - 1; m, q = 0, 1, 2, 3; n = {N / 4 ) m + /; a n d k = 4 p + q. W ith th is ch o ice o f p a ra m e te rs , th e g e n e ra l e q u a tio n g iven by
467
Figure 6.13 Sixteen-point radix-4 decimation-in-time algorithm with input in nor mal order and output in digit-reversed order. (6.1.15) can be exp ressed as
(A y 4 )- l
X(p,q) = w here
1=0
C ( l , q) W' NfA l
Ip
(6.1.45)
q = 0 , 1 , 2, 3 G (l,q) =
w hF (l,
q) i - 0 . 1 ......... - 1 4
(6.1.46)
and q F(l,q) = Y x ( l , m ) W ?
=0,1,2,3
(6.1.47)
N / = 0 , 1 , 2 , 3 .........- - 1 4
W e n o te that X ( p , q ) = X ( 4 p + q ), q = 0, 1, 2, 3. C on sequ en tly, the Af-point D F T is d ecim a ted into four N /4 -p o in t D F T s and h en ce w e have a decim ationin -frequency F F T algorithm . T h e com putations in (6.1.46) and (6.1.47) define
468
Chap. 6
Figure 6.14 Sixteen-point, radix-4 decimation-in-frequency algorithm with input in normal order and output in digit-reversed order. the basic radix-4 butterfly for the d ecim ation-in-frequency algorithm . N o te that the m ultiplications by the factors W% occur after the com b ination o f the data p oin ts x (/, m), just as in the case o f th e radix-2 decim ation -in-freq u en cy algo rithm. A 16-point radix-4 decim ation -in-freq u en cy F F T algorithm is show n in Fig. 6.14. Its input is in norm al order and its output is in digit-reversed order. It has exactly the sam e com putational com p lexity as the d ecim ation -in-tim e radix4 F F T algorithm . For illustrative purposes, let us rederive the radix-4 decim ation-in-frequency algorithm by breaking the jV-point D F T form ula in to four sm aller D F T s. We have
N- 1
X( k) = T x ( n ) W N kn n=0 = JV/4-1 N/Z-1 3N/4-1 fif-l x{n)W k N n+ x{n)W%' + * (* )< " + x(n)W N kn n=0 n=N/4 n=Nfi n=JN/4
Sec. 6.1

A / 74/ * -1 J A/ A /4 t- iI
469
,< ,< + <

h= (I
=o
E*("
/ '
A/ N
t
N / A \ / 'l \ ! \
A'/4 1
\i \
< V/2 E
* ( + y ) <
+ < A'/4 E
lNk/4 w\N
A (" + t )
(6.1.48)
F ro m th e d efin itio n o f th e tw id d le facto rs, w e h av e (jf (6.1.49)
A fte r su b s titu tio n o f (6.1.49) in to (6.1.48). we o b ta in

N/ 4-1
xa)=
x() + (
(6.1.50) N + { -1 f x [ n + - J + ( ;) W\
T h e re la tio n in (6.1.50) is n o t an N /4 -p o in t D F T b e c a u se th e tw id d le facto r d e p e n d s o n N a n d n o t on N / 4 . T o c o n v ert it in to an A '/4-point D F T , wc su b d iv id e th e D F T se q u e n c e in to four /V /4-point su b se q u e n c e s, X ( 4 k ) . X ( 4 k + !), X (4 + 2), an d X { 4 k + 3), k 0, 1........ N / 4 1. T h u s we o b ta in th e rad ix -4 d ecim atio n -in fre q u e n c y D F T as x ( n ) + .v (6.1.51)
+? )
+a(', + i ) +j:(" + t
X (4k + l ) =
0 urkn w"w N r r N /4
*( ) ~
ix
( +
"j}
IV " w kn
(6.1.52)
N w yv/4
X ( 4 k + 2) =
T ; 1r ( n \ E * ( ) * ( + j J
(6.1.53)
- K
r X ( 4 k + 3) = E
' +t )
n
W 2" w kn w Nf4
/ x (n '>+ J x ( " +
\ J
kn S/4
(6.1.54)
~ x (n ~ h j ) - Jx (b+ t
) ] w^
470
Chap. 6
w h ere w e h av e u se d th e p ro p e rty W^ kn = W^"4. N o te th a t th e in p u t to ea c h N/ 4p o in t D F T is a lin e a r co m b in a tio n o f fo u r signal sa m p le s sc aled by a tw id d le factor. T h is p ro c e d u re is r e p e a te d v tim es, w h ere v = log,, N.
6.1.5 Split-Radix FFT Algorithms

A n in sp e ctio n o f th e radix-2 d e c im a tio n -in -fre q u e n c y flo w g rap h show n in Fig. 6.11 in d icates th a t th e e v e n -n u m b e re d p o in ts o f th e D F T can b e c o m p u te d in d e p e n d en tly o f th e o d d -n u m b e re d p o in ts. T h is suggests th e p o ssib ility o f using d ifferen t c o m p u ta tio n a l m e th o d s fo r in d e p e n d e n t p a rts o f th e a lg o rith m w ith th e ob jectiv e o f re d u c in g th e n u m b e r o f c o m p u ta tio n s. T h e sp lit-ra d ix F F T (S R F F T ) alg o rith m s ex p lo it th is id ea by u sing b o th a radix-2 a n d a radix-4 d e c o m p o s itio n in th e sam e F I T alg o rith m . W e illu stra te th is a p p ro a c h w ith a d e c im a tio n -in -fre q u e n c y S R F F T alg o rith m d u e to D u h a m e l (1986). F irst, w e recall th a t in th e rad ix -2 d ec im a tio n -in -fre q u e n c y F F T alg o rith m , th e e v e n -n u m b e re d sa m p le s o f th e /V -point D F T a re given as
N o te th a t th ese D F T p o in ts can b e o b ta in e d fro m an N /2 -p o in t D F T w ith o u t any a d d itio n a l m u ltip lic atio n s. C o n se q u e n tly , a radix-2 suffices fo r th is c o m p u ta tio n . T h e o d d -n u m b e re d sa m p le s {X{2k + 1)) o f th e D F T re q u ire th e p re m u ltip li catio n o f th e in p u t se q u e n c e w ith th e tw id d le fa c to rs W N n . F o r th e s e sa m p les a rad ix -4 d eco m p o sitio n p ro d u c e s so m e c o m p u ta tio n a l efficiency b e c a u se th e fourp o in t D F T h as th e larg est m u ltip lic a tio n -fre e b u tterfly . I n d e e d , it can b e show n th a t usin g a rad ix g r e a te r th a n 4, d o e s n o t re su lt in a significant re d u c tio n in com p u ta tio n a l co m p lex ity . I f we u se a rad ix -4 d e c im a tio n -in -fre q u e n c y F F T a lg o rith m fo r th e oddn u m b e re d sa m p le s o f th e /V -point D F T , w e o b ta in th e fo llo w in g N /4 -p o in t D FTs:
N/4-1
(6.1.56) - j [ x ( n + N / 4 ) - x( n + 3 N / 4 )]}
A74-1
(6.1.57) + j [ x ( n + N / 4 ) - x{ n + 3 N / 4 ) ] } W ^ T h u s th e N -p o in t D F T is d e c o m p o se d in to o n e N /2 -p o in t D F T w ith o u t ad d itio n al tw id d le facto rs a n d tw o N /4 -p o in t D F T s w ith tw id d le facto rs. T h e /V-p o in t D F T is o b ta in e d by su ccessiv e u se o f th e s e d e c o m p o s itio n s u p to th e la st stag e. T hus w e o b ta in a d e c im a tio n -in -fre q u e n c y S R F F T alg o rith m . F ig u re 6.15 sh o w s th e flow g ra p h fo r a n in -p la ce 3 2 -p o in t d ecim atio n in -freq u en cy S R F F T a lg o rith m . A t stag e A of th e c o m p u ta tio n fo r N = 32, the
Sec. 6.1
471
A
w ith perm ission from the IE E E .
Figure 6,15 L ength 32 split-radix F F T algorithm s from p ap er by D uham el (1986); rep rin ted
to p 16 p o in ts c o n s titu te th e se q u e n c e go() = x ( n ) + x ( n + N / 2) 0 < n < 15 (6.1.58)
T h is is th e s e q u e n c e re q u ire d fo r th e c o m p u ta tio n o f X ( 2 k ) . T h e n e x t 8 p o in ts c o n s titu te th e se q u e n c e
gi(n) = x(n) - x(n + N/ 2)
0<n<7
(6.1.59)
472
Chap. 6
T h e b o tto m e ig h t p o in ts c o n s titu te th e se q u e n c e j g 2 (71). w h ere
82 (n) = x ( n + N / 4 ) x ( n + 3 N / 4 )
0 < < 7
(6.1.60)
T h e se q u e n c e s gj ( n) a n d gi ( n) a re u sed in th e c o m p u ta tio n o f X ( 4 k 4 - 1) and A'( 4 * + 3). T h u s, a t stag e A w e h ave c o m p le te d th e first d e c im a tio n fo r th e radix-2 c o m p o n e n t o f th e alg o rith m . A t sta g e B , th e b o tto m eig h t p o in ts c o n stitu te the c o m p u ta tio n o f [# i(n ) + 7 ( )]^32 >0 < /i < 7, w hich is u sed to c o m p u te X (4A -f 3), 0 < k < 7. T h e n ex t eig h t p o in ts fro m th e b o tto m c o n s titu te th e c o m p u ta tio n of [tfi(n) j g 2(h)] VVjj, 0 < n < 7, w hich is u se d to c o m p u te X ( 4 k 4 -1 ), 0 < k < 7. T h u s a t stag e B , w e h av e c o m p le te d th e first d e c im a tio n fo r th e ra d ix -4 alg o rith m , w hich resu lts in tw o 8 -p o in t se q u e n c e s. H e n c e th e b asic b u tte rfly c o m p u ta tio n fo r th e S R F F T a lg o rith m h a s th e L -s h a p e d form illu stra te d in Fig. 6.16. N ow w e r e p e a t th e ste p s in th e c o m p u ta tio n a b o v e. B e g in n in g w ith th e to p 16 p o in ts a t stag e A , w e r e p e a t th e d e c o m p o s itio n f o r th e 1 6 -p o in t D F T . In o th e r w o rd s, w e d e c o m p o se th e c o m p u ta tio n in to an e ig h t-p o in t, rad ix -2 D F T an d tw o fo u r-p o in t, rad ix -4 D F T s. T h u s a t sta g e B , th e to p eig h t p o in ts c o n s titu te the se q u e n c e (w ith N = 16)
g'o(*) = 8o(n) + go(n 4- N / 2)
<n<7
(6.1.61)
an d th e n ex t eig h t p o in ts c o n s titu te th e tw o fo u r-p o in t se q u e n c e s g[(n) a n d jg'2(n), w h ere g[ (n) = go(n) ~ go(n + N f l ) 0 < n < 3 (6.1.62) 82 () = 8o(n + N / 4 ) - g0(n + 3 N / 4 ) 0 < n < 3 T h e b o tto m 16 p o in ts o f sta g e B a re in th e fo rm o f tw o e ig h t-p o in t D F T s. H en ce ea c h e ig h t-p o in t D F T is d e c o m p o s e d in to a fo u r-p o in t, rad ix -2 D F T a n d a fourp o in t, rad ix -4 D F T . In th e final stag e, th e c o m p u ta tio n s in v o lv e th e co m b in atio n o f tw o -p o in t se q u en ces. T a b le 6.2 p r e s e n ts a c o m p a riso n o f th e n u m b e r o f nont ri vi al re a l m u ltip li ca tio n s a n d a d d itio n s re q u ire d to p e rfo rm a n jY -point D F T w ith co m p lex -v alu ed
Sec. 6 .1

TABLE 6.2 NUMBER OF NONTRIVIAL REAL MULTIPLICATIONS AND ADDITIONS TO COMPUTE AN N-POINT COMPLEX DFT
Real M ultiplications Radix
N
473
Real A dditions Split Radix 20 68 196 516 1.284 3.076 7,172 Radix 2 152 408 1.032 2.504 5,896 13.566 30.728 Radix 4 148 976 5,488 12,420 28.336 972 R adix 8 Split Radix 148 388 964 2308 5.380 12.292 27,652
R adix 4 20 208 1.392
Radix 8
16 32 64 128 256 512 1,024
24 88 264 712 1,800 4.360 10.248
204
3.204 7,856
Source: E xtracted from D uham el (1986).
d a ta , using a rad ix -2 , ra d ix -4, radix-8, a n d a sp lit-ra d ix F F T . N o te th a t th e S R F F T alg o rith m re q u ire s th e lo w est n u m b e r o f m u ltip lic a tio n a n d a d d itio n s. F o r this re a so n , it is p re fe ra b le in m an y p ra c tic a l a p p licatio n s. A n o th e r ty p e o f S R F F T a lg o rith m has b e e n d e v e lo p e d by P rice (1990). Its re la tio n to D u h a m e ls a lg o rith m d e sc rib e d p rev io u sly can b e seen by n o tin g th a t th e rad ix -4 D F T te rm s X ( 4 k 4- 1) an d X ( 4 k + 3) involve th e N /4 -p o in t D F T s o f th e se q u e n c e s [g i(n ) a n d [ # i( '0 + ,/ 2( )]W $ \ resp ec tiv ely . In effect, th e se q u e n c e s g i(/i) a n d g 2 (n) are m u ltip lie d by th e fa c to r (v e c to r) (1, j ) = (1, H ^ ) an d by WJJ fo r th e c o m p u ta tio n o f X ( 4 k + 1), w hile th e c o m p u ta tio n o f X (4k + 3) in v o lv es th e fa c to r (1 , j ) = (1, W{2*) an d W j f . In s te a d , o n e can r e a rra n g e th e c o m p u ta tio n so th a t th e fa c to r fo r X ( 4 k + 3) is ( j , 1) = (W 8, 1). A s a resu lt o f th is p h a s e ro ta tio n , th e tw id d le fa c to rs in th e c o m p u ta tio n o f X (4k -f 3) b eco m e ex actly th e sam e as th o se fo r X ( 4 k + 1), e x cep t th a t th e y o c c u r in m irr o r im age o rd e r. F o r e x am p le, at sta g e B o f Fig. 6.15, th e tw id d le fa c to rs W21, W 18, . . . , W3 a re re p la c e d by (V1, W 2, . . . , W 1, resp ec tiv ely . T h is m irro r-im a g e sy m m etry occurs a t ev ery s u b s e q u e n t sta g e o f th e alg o rith m . A s a c o n s e q u e n c e , th e n u m b e r o f tw id d le facto rs th a t m u st b e c o m p u te d a n d s to re d is r e d u c e d by a fa c to r o f 2 in c o m p a riso n to D u h a m e ls a lg o rith m . T h e re su ltin g a lg o rith m is called th e m irr o r F F T (M F F T ) alg o rith m . A n a d d itio n a l facto r-o f-2 savings in sto ra g e o f tw id d le fa c to rs can be o b ta in e d b y in tro d u c in g a 90 p h a s e o ffset a t th e m id p o in t o f ea c h tw id d le a rra y , w hich can b e re m o v e d if n ecessa ry a t th e o u tp u t o f th e S R F F T c o m p u ta tio n . T h e in co r p o r a tio n o f th is im p ro v e m e n t in to th e S R F F T ( o r th e M F F T ) re su lts in a n o th e r a lg o rith m , also d u e to P ric e (1990), c a lle d th e p h a s e F F T (P F F T ) alg o rith m .
6.1.6 Implementation of FFT Algorithms

N o w th a t w e h av e d e s c rib e d th e b asic rad ix -2 a n d rad ix -4 F F T a lg o rith m s, let us c o n s id e r so m e o f th e im p le m e n ta tio n issues. O u r r e m a rk s ap p ly d ire c tly to
474
Chap. 6
rad ix -2 a lg o rith m s, a lth o u g h sim ilar c o m m e n ts m ay b e m a d e a b o u t rad ix -4 a n d h ig h er-rad ix alg o rith m s. B asically, th e rad ix -2 F F T a lg o rith m co n sists o f ta k in g tw o d a ta p o in ts a t a tim e fro m m e m o ry , p e rfo rm in g th e b u tte rfly c o m p u ta tio n s a n d r e tu rn in g th e r e su ltin g n u m b e rs to m em o ry . T h is p ro c e d u re is r e p e a te d m an y tim e s ( ( N log2 N) [ 2 tim es) in th e c o m p u ta tio n o f an JV-point D F T . T h e b u tterfly c o m p u ta tio n s re q u ire th e tw id d le fa c to rs {W^} a t v a rio u s stages in e ith e r n a tu ra l o r b it-re v e rse d o rd e r. In an efficien t im p le m e n ta tio n o f th e algo rith m , th e p h ase fa c to rs a re c o m p u te d o n ce a n d s to re d in a ta b le , e ith e r in n o rm a l o r d e r o r in b it-re v e rse d o rd e r, d e p e n d in g o n th e specific im p le m e n ta tio n o f th e alg o rith m . M e m o ry re q u ire m e n t is a n o th e r fa c to r th a t m u st b e c o n s id e re d . If th e c o m p u ta tio n s a re p e rfo rm e d in p lace, th e n u m b e r o f m e m o ry lo c a tio n s re q u ire d is 2 N since th e n u m b e rs a re com plex. H o w e v e r, w e can in ste a d d o u b le th e m e m o ry to 4N , th u s sim p lifying th e in d ex in g a n d c o n tro l o p e r a tio n s in th e F F T a lg o rith m s. In th is case w e sim p ly a lte rn a te in th e use o f th e tw o se ts o f m e m o ry lo c a tio n s from o n e sta g e o f th e F F T a lg o rith m to th e o th e r. D o u b lin g o f th e m e m o ry also allow s us to h av e b o th th e in p u t se q u e n c e a n d th e o u tp u t se q u e n c e in n o rm a l o rd e r. T h e re are a n u m b e r o f o th e r im p le m e n ta tio n issues re g a rd in g ind ex in g , bit rev ersal, an d th e d e g re e o f p arallelism in th e c o m p u ta tio n s. T o a larg e ex ten t, th e se issues a re a fu n ctio n o f th e specific a lg o rith m a n d th e ty p e o f im p le m e n ta tio n , n am ely , a h a rd w a re o r so ftw are im p le m e n ta tio n . In im p le m e n ta tio n s b ased o n a fix ed -p o in t a rith m e tic , o r flo atin g -p o in t a rith m e tic o n sm a ll m a ch in es, th e re is also th e issue o f ro u n d -o ff e rro rs in th e c o m p u ta tio n . T h is to p ic is co n sid e re d in S ectio n 6.4. A lth o u g h th e F F T a lg o rith m s d e sc rib e d p re v io u sly w e re p re s e n te d in th e c o n te x t o f co m p u tin g th e D F T efficiently, th e y can also b e u s e d to c o m p u te th e ID F T , w hich is
j A '- l
* (") = 7 T ] C * ( *=0
(6' ll63)
T h e o n ly d iffe re n c e b e tw e e n th e tw o tra n sfo rm s is th e n o rm a liz a tio n fa c to r l / N a n d th e sign o f th e p h a se fa c to r WN. C o n s e q u e n tly , a n F F T a lg o rith m fo r co m p u tin g th e D F T , c a n b e c o n v e rte d to an F F T a lg o rith m fo r c o m p u tin g th e ID F T by ch an g in g th e sign o n all th e p h a se fa c to rs a n d d iv id in g th e final o u tp u t o f th e alg o rith m by N. In fact, if w e ta k e th e d e c im a tio n -in -tim e a lg o rith m t h a t w e d e sc rib e d in S ectio n 6.1.3, re v e rse th e d ire c tio n o f th e flow g ra p h , c h a n g e th e sign o n th e p h ase facto rs, in te rc h a n g e th e o u tp u t a n d in p u t, a n d finally, d iv id e th e o u tp u t by N , w e o b ta in a d e c im a tio n -in -fre q u e n c y F F T a lg o rith m fo r c o m p u tin g th e ID F T . O n th e o th e r h a n d , if w e b eg in w ith th e d e c im a tio n -in -fre q u e n c y F F T a lg o rith m d escrib ed in S ectio n 6.1.3 a n d r e p e a t th e ch an g es d e s c rib e d a b o v e , w e d b ta in a d ecim atio n in -tim e F F T a lg o rith m fo r c o m p u tin g th e ID F T . T h u s it is a sim p le m a tte r to devise F F T a lg o rith m s fo r co m p u tin g th e ID F T .
Sec. 6.2
Applications of FFT Algorithms
475
F in ally , w e n o te th a t th e em p h asis in o u r discussion o f F F T a lg o rith m s w as o n rad ix -2 , rad ix -4 , a n d sp lit-ra d ix a lg o rith m s. T h e se are by fa r th e m o st w idely u se d in p ra c tic e . W h e n th e n u m b e r o f d a ta p o in ts is n o t a p o w e r o f 2 o r 4. it is a sim p le m a tte r to p a d th e se q u e n c e x ( n ) w ith zero s such th a t /V = 2 1 o r N = 4 '. T h e m e a s u re o f co m p lex ity fo r F F T alg o rith m s th a t w e h av e e m p h a siz e d is th e r e q u ire d n u m b e r o f a rith m e tic o p e ra tio n s (m u ltip lic a tio n s an d a d d itio n s). A lth o u g h this is a v ery im p o rta n t b e n c h m a rk fo r c o m p u ta tio n a l co m p lex ity , th e re a re o th e r issues to b e c o n s id e re d in p ractical im p le m e n ta tio n o f F F T a lg o rith m s. T h e se in clu d e th e a rc h ite c tu re o f th e p ro cesso r, th e av ailab le in stru c tio n set, th e d a ta s tru c tu re s fo r s to rin g tw id d le facto rs, an d o th e r c o n sid e ra tio n s. F o r g e n e ra l-p u rp o s e c o m p u te rs, w h e re th e cost o f th e n u m e ric a l o p e ra tio n s d o m in a te , rad ix -2 , rad ix-4, an d sp lit-ra d ix F F T a lg o rith m s a re go o d c a n d id a te s. H o w e v e r, in th e case o f sp e c ia l-p u rp o se digital signal p ro c e ss o rs , fe a tu rin g sin g le cycle m u ltip ly -a n d -a c c u m u la te o p e ra tio n , b it-re v e rse d ad d re ssin g , a n d a high d e g ree o f in stru c tio n p a ra lle lism , th e stru c tu ra l re g u la rity o f th e a lg o rith m is eq u ally im p o rta n t as a rith m e tic co m p lex ity . H e n c e fo r D S P p ro c e sso rs, rad ix -2 o r radix4 d e c im a tio n -in -fre q u e n c y F F T a lg o rith m s are p re fe ra b le in te rm s o f sp e e d an d a ccu racy . T h e irre g u la r stru c tu re o f th e S R F F T m ay r e n d e r it less su ita b le fo r im p le m e n ta tio n o n d ig ital signal pro cesso rs. S tru c tu ra l re g u la rity is also im p o rta n t in th e im p le m e n ta tio n o f F F T a lg o rith m s on v e c to r p ro c e sso rs, m u ltip ro c e sso rs, a n d in V L SI. I n te rp ro c e s s o r co m m u n icatio n is an im p o rta n t c o n s id e ra tio n in such im p le m e n ta tio n s o n p a rallel pro cesso rs. In co n clu sio n , we h av e p re s e n te d several im p o rta n t c o n s id e ra tio n s in th e im p le m e n ta tio n o f F F T a lg o rith m s. A d v an ce s in digital signal p ro cessin g te c h n o l ogy, in h a rd w a re a n d so ftw are, will c o n tin u e to influence th e ch o ice a m o n g F F T a lg o rith m s fo r v a rio u s p ractical ap p licatio n s.
6.2 APPLICATIONS OF FFT ALGORITHMS

T h e F F T a lg o rith m s d e sc rib e d in th e p re c e d in g se ctio n find a p p lic a tio n in a v ariety o f a re a s , in clu d in g lin e a r filtering, c o rre la tio n , a n d s p e c tru m analysis. B asically, th e F F T a lg o rith m is u sed as a n efficient m e a n s to c o m p u te th e D F T a n d th e ID F T . In th is se c tio n w e c o n s id e r th e u se o f th e F F T a lg o rith m in lin e a r filterin g a n d in th e c o m p u ta tio n o f th e c ro ssco rrelatio n o f tw o se q u e n c e s. T h e use o f th e F F T in s p e c tru m an aly sis is c o n sid e re d in C h a p te r 12. In a d d itio n w e illu strate h o w to e n h a n c e th e efficiency o f th e F F T a lg o rith m by fo rm in g co m p le x -v a lu e d se q u e n c e s fro m re a l-v a lu e d se q u e n c e s p rio r to th e c o m p u ta tio n o f th e D F T .
6.2.1 Efficient Computation of the DFT of Two Real Sequences

T h e F F T a lg o rith m is d esig n e d to p e rfo rm co m p lex m u ltip lic a tio n s a n d a d d itio n s, ev en th o u g h th e in p u t d a ta m ay b e re a l v alued. T h e b asic re a s o n fo r th is s itu a tio n is
476
Chap. 6
th a t th e p h a se fa c to rs a re co m p lex a n d h e n c e , a fte r th e first sta g e o f th e alg o rith m , all v a riab les a re b asically co m p lex -v alu ed . In view o f th e fact th a t th e a lg o rith m can h a n d le c o m p le x -v a lu e d in p u t se q u en ces, w e c a n ex p lo it th is cap ab ility in th e c o m p u ta tio n o f th e D F T o f tw o re a l-v a lu e d se q u e n c e s. S u p p o se th a t * i(n ) a n d x 2(n) are tw o re a l-v a lu e d se q u e n c e s o f len g th N , and let x( n ) b e a co m p le x -v a lu e d se q u e n c e d efin ed as x ( n ) = X] (n) + j x 2 (n) 0 < n < N 1 (6.2.1)
T h e D F T o p e ra tio n is lin e a r a n d h en ce th e D F T o f x ( n ) can be ex p re sse d as X ( k ) = X ](k) + j X 2(k) (6.2.2)
T h e se q u e n c e s ^ i( ) an d J 2 OO can be ex p re sse d in te rm s o f x ( n ) as follow s: , 4 ac(H) + J:*(n) * i( ) = -------- 2-------x(n)-x*(n) *:() = ------- tt.-------H e n c e th e D F T s o f jri(n ) an d x 2(n) are * ,( * ) = ]\ D F T [ x { n ) } + D F T [ x \ n ) } ) (6.2.5) (6.2.6) (6.2.3) (6.2.4)
X 2(k) = j - \ D F T [ x ( n )] - DF T [ x * ( n ) ] ) R ecall th a t th e D F T o f x*( n) is X * ( N k). T h e re fo re , X] (k) = i[X (* r) + X * ( N - jt)] X 2(k) = -^ [X (* > - X * ( N - *)] ;2
(6.2.7) (6.2.8)
T h u s, by p e rfo rm in g a single D F T o n th e co m p le x -v a lu e d se q u e n c e x ( n ), we h av e o b ta in e d th e D F T o f th e tw o re a l se q u e n c e s w ith only a sm all a m o u n t of a d d itio n a l c o m p u ta tio n th a t is involved in co m p u tin g Xi (Jt) a n d X 2 (k) fro m X(k) by u se o f (6.2.7) a n d (6.2.8).
6.2.2 Efficient Computation of the DFT of a 2/V-Point Real Sequence

S u p p o se th a t g( n) is a re a l-v a lu e d se q u e n c e o f 2 N p o in ts. W e n o w d e m o n s tra te h o w to o b ta in th e 2 N -p o in t D F T o f g( n) fro m c o m p u ta tio n o f o n e A ppoint D F T involving c o m p le x -v a lu e d d a ta . F irst, w e define
* i(n ) = g(2 n)
(6.2.9)
*2(n) = g ( 2 n + 1)
Sec. 6.2
Applications of FFT Algorithms
477
T h u s w e h a v e su b d iv id e d th e 2 N -p o in t re a l se q u e n c e in to tw o W -point real se q u e n c e s. N o w w e can ap p ly th e m e th o d d escrib ed in th e p re c e d in g sectio n . L et jc(n) b e th e A7-p o in t c o m p lex -v alu ed se q u e n c e

A-(n) = * i ( n ) + j x i i n ) (6 .2 .10)
F ro m th e re su lts o f th e p re c e d in g se ctio n , w e h av e x m = ^ [* (* ) + * * ( * - * ) ] j X 2(k) = [ X( k) - X * ( N - k)] F inally, w e m u st ex p re ss th e 2/V -point D F T in te rm s o f th e tw o /V -point D F T s, Xi(A) a n d X 2(k). T o acco m p lish this, w e p ro c e e d as in th e d e c im a tio n -in -tim e F F T a lg o rith m , n am ely ,
N -1 N-1
(6.2.11)
C( k ) = s < 2 h ) H $ * + J 2 s ( 2 n + ^ W7 N ^ k
n=tl N- l n=0 N-1
=() C o n s e q u e n tly , G( k ) = X t (k) + W i N X 2(k) G( k + N ) = X i ( k ) - W%N X 2(k)
n=()
k = 0 . 1 ..........N - 1
( 6 . 2 . 12 )
k = Q . \ ..........N - l
T h u s w e h av e c o m p u te d th e D F T o f a 2/V -point real se q u e n c e from o n e jV-point D F T an d so m e a d d itio n a l c o m p u ta tio n as in d icated by (6.2.11) an d (6.2.12).
6.2.3 Use of the FFT Algorithm in Linear Filtering and Correlation

A n im p o rta n t ap p lic a tio n o f th e F F T a lg o rith m is in F IR lin e a r filterin g o f lo n g d a ta se q u e n c e s. In C h a p te r 5 w e d e sc rib e d tw o m e th o d s, th e o v e rla p -a d d an d th e o v e rla p -sa v e m e th o d s fo r filterin g a lo n g d a ta se q u e n c e w ith an F I R filter, b a s e d o n th e u se o f th e D F T . In th is se ctio n w e c o n sid e r th e u se o f th e s e tw o m e th o d s in c o n ju n c tio n w ith th e F F T a lg o rith m fo r co m p u tin g th e D F T an d th e ID F T . L e t h( n), 0 < n < M - 1 , b e th e u n it sa m p le re sp o n s e o f th e F IR filter an d let x ( n ) d e n o te th e in p u t d a ta se q u e n c e . T h e block size o f th e F F T alg o rith m is N , w h e re N = L + M 1 an d L is th e n u m b e r o f new d a ta sa m p le s b e in g p ro cessed by th e filter. W e a ssu m e th a t fo r a n y given v alu e o f Af, th e n u m b e r L o f d a ta sa m p le s is se le c te d so th a t N is a p o w e r o f 2. F o r p u rp o se s o f th is discussion, w e c o n s id e r o n ly rad ix -2 F F T alg o rith m s. T h e /V -point D F T o f h(n), w hich is p a d d e d b y L 1 z e ro s, is d e n o te d as H( k ) . T h is c o m p u ta tio n is p e rfo rm e d o n c e via th e F F T an d th e re su ltin g N co m p lex n u m b e rs a r e sto re d . T o be specific w e a ssu m e th a t th e d e c im a tio n -in -fre q u e n c y
478
Chap. 6
F F T a lg o rith m is u se d to c o m p u te H( k ) . T h is y ields H ( k ) in b it-re v e rse d o rd e r, w hich is th e w ay it is s to re d in m em ory. In th e o v e rlap -sav e m e th o d , th e first M 1 d a ta p o in ts o f e a c h d a ta b lo ck are th e last M 1 d a ta p o in ts o f th e p rev io u s d a ta b lo ck . E a c h d a ta b lo c k c o n ta in s L new d a ta p o in ts, su ch th a t N = L + M 1. T h e N -p o in t D F T o f ea c h d a ta block is p e rfo rm e d by th e F F T alg o rith m . If th e d e c im a tio n -in -fre q u e n c y alg o rith m is e m p lo y ed , th e in p u t d a ta b lo ck re q u ire s n o shuffling a n d th e v a lu e s o f th e D F T o ccu r in b it-re v e rse d o rd e r. S ince th is is ex actly th e o r d e r o f H ( k ) . w e can m ultiply th e D F T o f th e d a ta , say Xm(fc), w ith //(Jt) a n d th u s th e re su lt Ym(k) = H ( k ) X m(k) is also in b it-re v e rse d o rd e r. T h e in v erse D F T (ID F T ) can b e c o m p u te d by use o f an F F T alg o rith m th a t ta k e s th e in p u t in b it-re v e rse d o r d e r a n d p ro d u c e s an o u tp u t in n o rm al o rd er. T h u s th e r e is n o n e e d to shuffle any b lo ck o f d a ta e ith e r in c o m p u tin g th e D F T o r th e ID F T . If th e o v e rla p -a d d m e th o d is used to p e rfo rm th e lin e a r filterin g , th e co m p u ta tio n a l m e th o d u sin g th e F F T a lg o rith m is basically th e sa m e. T h e only differen ce is th a t th e N -p o in t d a ta b lo ck s consist o f L new d a ta p o in ts a n d M 1 a d d itio n a l zero s. A fte r th e I D F T is c o m p u te d fo r ea c h d a ta b lo ck , th e W -point filtered blocks a re o v e rla p p e d as in d ic a te d in S ectio n 5.3.2, a n d th e M - 1 o v e rla p p in g d a ta p o in ts b e tw e e n successive o u tp u t re c o rd s a re a d d e d to g e th e r. L et u s assess th e c o m p u ta tio n a l co m p lex ity o f th e F F T m e th o d fo r lin e a r fil terin g . F o r th is p u rp o se , th e o n e -tim e c o m p u ta tio n o f H ( k ) is in sig n ifican t an d can b e ig n o red . E ach F F T re q u ire s ( N / 2) log2 N co m p lex m u ltip lic a tio n s an d N Iog2 N a d d itio n s. Since th e F F T is p e rfo rm e d tw ice, o n ce fo r th e D F T a n d o n ce fo r th e ID F T , th e c o m p u ta tio n a l b u rd e n is N log2 N co m p lex m u ltip lic a tio n s an d 2 N log2 N a d d itio n s. T h e re a re also N co m p lex m u ltip lic a tio n s a n d N 1 a d d itio n s re q u ire d to c o m p u te ym(Jfc). T h e re fo re , w e h av e ( N \ o g 2 2 N ) / L co m p lex m u ltip lic a tio n s p er o u tp u t d a ta p o in t a n d a p p ro x im a te ly ( 2 N \ o g 2 2 N ) / L a d d itio n s p e r o u tp u t d ata p o in t. T h e o v e rla p -a d d m e th o d re q u ire s an in c re m e n ta l in c re a se o f ( M \ ) / L in th e n u m b e r o f ad d itio n s. B y w ay o f c o m p a riso n , a d ire c t fo rm re a liz a tio n o f th e F I R filter involves M real m u ltip lic atio n s p e r o u tp u t p o in t if th e filter is n o t lin e a r p h a s e , a n d M / 2 if it is lin e a r p h ase (sy m m etric ). A lso , th e n u m b e r o f a d d itio n s is M - 1 p e r o u tp u t p o in t (see Sec. 8.2). I t is in te re stin g to co m p a re th e efficiency o f th e F F T a lg o rith m w ith th e direct fo rm re a liz a tio n o f th e F IR filter. L e t us focus o n th e n u m b e r o f m ultip lic atio n s, w h ich a re m o re tim e co n su m in g th a n a d d itio n s. S u p p o se th a t M = 128 = 27 an d N = 2 V. T h e n th e n u m b e r o f co m p lex m u ltip lic a tio n s p e r o u tp u t p o in t fo r an F F T size o f N = 2 V is
Sec. 6.3
A Linear Filtering Approach to Computation of the DFT TABLE 6.3

COMPUTATIONAL COMPLEXITY
479
Size of FFT i) log2 N 9 10 11 12 14
f(v) Number of Complex Multiplications per Output Point 13.3 12.6 12.8 13.4 15.1
T h e v alu es o f c( v) fo r d iffe re n t v alu es o f i> are given in T a b le 6.3. W e o b se rv e th a t th e re is an o p tim u m v a lu e o f i< w h ich m in im iz es c(u ). F o r th e F IR filter of size M = 128, th e o p tim u m o ccu rs at d = 10. W e sh o u ld e m p h asize th a t c ( f ) r e p re s e n ts th e n u m b e r o f co m p lex m u ltip lic a tio n s fo r th e F F T -b a se d m e th o d . T h e n u m b e r o f re a l m u ltip lic a tio n s is fo u r tim es th is n u m b e r. H o w e v e r, ev en if th e F IR filter has lin e a r p h a s e (see Sec. 8.2), th e n u m b e r o f c o m p u ta tio n s p e r o u tp u t p o in t is still less w ith th e F F T -b a se d m eth o d . F u rth e rm o r e , th e efficiency o f th e F F T m e th o d can be im p ro v e d by c o m p u tin g th e D F T o f tw o successive d a ta b lo ck s sim u lta n e o u sly , ac c o rd in g to th e m eth o d ju st d e sc rib e d . C o n s e q u e n tly , th e F F T -b a se d m e th o d is in d e e d su p e rio r from a c o m p u ta tio n a l p o in t o f view w h en th e filter len g th is re lativ ely large. T h e c o m p u ta tio n o f th e cross c o rre la tio n b e tw e e n tw o se q u e n c e s by m e a n s o f th e F F T a lg o rith m is sim ilar to th e lin e a r F IR filtering p ro b le m ju st d esc rib e d . In p ractical a p p lic a tio n s involving c ro ssc o rre la tio n , a t least o n e o f th e se q u e n c e s has finite d u ra tio n an d is a k in to th e im p u lse re sp o n s e o f th e F IR filter. T h e seco n d s e q u e n c e m ay be a lo n g se q u e n c e w hich c o n ta in s th e d e s ire d se q u e n c e c o rru p te d b y a d d itiv e n o ise. H e n c e th e se co n d se q u e n c e is a k in to th e in p u t to th e F I R filter. B y tim e rev e rsin g th e first se q u e n c e a n d co m p u tin g its D F T , w e h av e r e d u c e d th e cro ss c o rre la tio n to an e q u iv a le n t co n v o lu tio n p ro b le m (i.e.. a lin e a r F I R filtering p ro b le m ). T h e re fo re , th e m e th o d o lo g y w e d e v e lo p e d fo r lin e a r F IR filterin g by u se o f th e F F T a p p lie s directly.
6.3 A LINEAR FILTERING APPROACH TO COMPUTATION OF THE DFT

T h e F F T alg o rith m ta k e s N p o in ts o f in p u t d a ta a n d p ro d u c e s an o u tp u t se q u e n c e o f N p o in ts c o rre sp o n d in g to th e D F T o f th e in p u t d a ta . A s w e h a v e show n, th e rad ix -2 F F T a lg o rith m p e rfo rm s th e c o m p u ta tio n of th e D F T in ( N f l ) log2 N m u ltip lic a tio n s a n d N log2 N a d d itio n s fo r a n N -p o in t se q u e n c e . T h e re a re so m e a p p lic a tio n s w h e re o n ly a se le c te d n u m b e r o f valu es o f th e D F T a re d e s ire d , b u t th e e n tire D F T is n o t re q u ire d . In such a case, th e F F T a lg o rith m m ay n o lo n g e r be m o r e efficien t th a n a d ire c t c o m p u ta tio n o f th e d e s ire d v alu es o f th e D F T . In fact, w h e n th e d e s ire d n u m b e r o f valu es o f
480
Chap. 6
th e D F T is less th a n log2 N , a d ire c t c o m p u ta tio n o f th e d e s ire d v alu es is m o re efficient. T h e d irect c o m p u ta tio n o f th e D F T can b e fo rm u la te d as a lin e a r filtering o p e ra tio n o n th e in p u t d a ta se q u en ce. A s w e will d e m o n s tra te , th e lin e a r filter ta k e s th e fo rm o f a p a ra lle l b a n k o f re so n a to rs w h e re ea c h r e s o n a to r se lects o n e o f th e fre q u e n c ie s a> k = 2 n k / N , k = 0, 1 , . . . , N 1, c o rre sp o n d in g to th e N fre q u e n c ie s in th e D F T . T h e re a re o th e r a p p lic a tio n s in w hich w e re q u ire th e e v a lu a tio n o f th e ztra n sfo rm o f a fin ite -d u ra tio n se q u e n c e a t p o in ts o th e r th a n th e u n it circle. If th e se t o f d e sire d p o in ts in th e z-p lan e po ssesses so m e re g u la rity , it is possible to also ex p ress th e c o m p u ta tio n o f th e z -tra n s fo rm a s a lin e a r filte rin g o p e ra tio n . In th is c o n n e c tio n , w e in tro d u c e a n o th e r alg o rith m , called th e c h irp -z tra n sfo rm alg o rith m , w hich is su ita b le fo r e v a lu a tin g th e z -tra n s fo rm o f a se t o f d a ta o n a v ariety o f c o n to u rs in th e z-p lan e. T h is alg o rith m is also fo rm u la te d as a lin ear filtering o f a set o f in p u t d a ta . A s a co n se q u e n c e , th e F F T a lg o rith m can b e used to c o m p u te th e ch irp -z tra n sfo rm a n d th u s to e v a lu a te th e z -tra n s fo rm at various c o n to u rs in th e z -p la n e , in clu d in g th e u n it circle.
6.3.1 The Goertzel Algorithm

T h e G o e rtz e l a lg o rith m ex p lo its th e p erio d icity o f th e p h ase fa c to rs {W} an d allow s us to ex p re ss th e c o m p u ta tio n o f th e D F T as a lin e a r filterin g o p e ra tio n . Since W # kN = 1, w e can m u ltip ly th e D F T by th is fa c to r. T h u s (6.3.1) W e n o te th a t (6.3.1) is in th e fo rm of a c o n v o lu tio n . se q u e n c e yk(n) as In d e e d , if w e d efin e the
(6.3.2)
m=0
th e n it is c le a r th a t ( n ) is th e co n v o lu tio n o f th e fin ite -d u ra tio n in p u t se q u en ce x( n ) o f len g th N w ith a filter th a t h as an im pulse re sp o n s e h k(n) = W ~ knu ( n ) (6.3.3)
T h e o u tp u t o f th is filter a t n = N y ields th e v alu e o f th e D F T a t th e freq u e n cy an = h r k / N . T h a t is, X ( k ) = >*(n)|n=JV as can b e verified b y c o m p a rin g (6.3.1) w ith (6.3.2). T h e filter w ith im p u lse re s p o n s e h k (n) h a s th e sy stem fu n c tio n (6.3.4)
(6.3.5)
Sec. 6.3
A Linear Filtering Approach to Computation of the DFT
481
T h is filter h as a p o le o n th e u n it circle a t th e fre q u e n c y cd* = 2 n k / N . T h u s, the e n tire D F T can b e c o m p u te d by passin g th e block o f in p u t d a ta in to a p a ra l lel b a n k o f N sin g le-p o le filters (re s o n a to rs), w h ere each filter h as a p o le at the c o rre sp o n d in g fre q u e n c y o f th e D F T . I n s te a d o f p e rfo rm in g th e c o m p u ta tio n o f th e D F T as in (6.3.2), via co n v o lu tio n , w e can use th e d iffe re n c e e q u a tio n c o rre sp o n d in g to th e filter given by (6.3.5) to c o m p u te y k(ir) recu rsiv ely . T h u s we h av e y t (n) = W ^ kyt ( n - 1) + x i n ) V i-(-l) = 0 (6.3.6)
T h e d e sire d o u tp u t is X ( k ) = y k( N) , fo r k = 0, 1 , . . . , N 1. T o p e rfo rm this c o m p u ta tio n , w e can c o m p u te o n ce a n d sto re th e p h a s e facto rs W # k. T h e co m p lex m u ltip lic a tio n s an d a d d itio n s in h e re n t in (6.3.6) can be av o id ed by co m b in in g th e p airs o f re so n a to rs p o ssessin g c o m p le x -c o n ju g a te p oles. T his lead s to tw o -p o le filters w ith system fu n c tio n s o f th e form ] _ iy* HkL ) ~ 1 - 2 c o s ( 2 t i k / N ) : ~ l + C ' 2 (6'3 '7)
T h e d irect form II re a liz a tio n o f th e system illu stra te d in Fig. 6.17 is d e sc rib e d by th e d iffe re n c e e q u a tio n
v k(n) = 2 cos
2:rk N
v*.(zi 1) - vk(n - 2) + x( i t )
(6.3.8) (6.3.9)
Vi(h) = vk in) - W N k vk (n - 1)
w ith in itial c o n d itio n s iv - ( - l) = vk{ - 2 ) = 0. T h e recu rsiv e re la tio n in (6.3.8) is ite ra te d for n = 0, 1.........N , b u t th e e q u a tio n in (6.3.9) is c o m p u te d o n ly o n ce a t tim e n = N. E ach ite ra tio n re q u ire s o n e real m u ltip lic a tio n a n d tw o a d d itio n s. C o n se q u e n tly , fo r a re a l in p u t se q u e n c e x ( n) . th is a lg o rith m re q u ire s N + 1 re a l m u ltip lic a tio n s to yield n o t o n ly X ( k ) b ut also, d u e to sy m m etry , th e v a lu e o f X ( N k). T h e G o e rtz e l alg o rith m is p a rtic u la rly a ttra c tiv e w h en th e D F T is to b e c o m p u te d at a re lativ ely sm all n u m b e r M o f values, w h e re M < Iog2 N . O th erw ise, th e F F T a lg o rith m is a m o re efficient m e th o d .
Figure 6.17 Direct form It realization of two-pole resonator for computing the DFT.
482
Chap. 6
6.3.2 The Chirp-z Transform Algorithm

T he D F T o f an W -point data seq u en ce x(n ) has b een view ed as the z-transform o f x i n ) evaluated at N equally spaced p oin ts on the unit circle. It has also been view ed as N equally spaced sam ples o f the Fourier transform o f th e data sequ en ce x (n ). In this section w e consider the evaluation o f X ( z ) on other contours in the z-plane, including th e unit circle. S u ppose that w e wish to com p ute the values o f the z-transform o f jc(n) at a set o f p oints {z*}. T hen,
A '- l
X i z k) = J 2 x ( n ) z r
n=0
* = 0 , 1 .........L - 1
(6.3.10)
For exam ple, if the contour is a circle o f radius r and the z* are N equally spaced points, then Zk = r e j 2*"/" 2=1 X ( z k) = J 2 i x M r ~n}e n=0 k = n/N
0, 1,2 ..... N - 1
(6.3.11) k = 0 , 1 , 2 .........N - 1
In this case the FFT algorithm can be applied on the m odified seq u en ce x { n ) r ~ n. M ore generally, suppose that the p oin ts z* in the z-plane fall on an arc which begins at som e point
Zo
= r0eJlk
and spirals either in toward the origin or out away from the origin such that the points are defined as zk = rQ e je(Roei *)i k = 0 ,1 ,..., L - 1 (6.3.12)
N o te that if R0 < 1, the points fall on a con tour that spirals tow ard th e origin and if R0 > 1, the contour spirals away from the origin. If Ro 1, the con tou r is a circular arc o f radius ro. If r0 = 1 and Ro = l , the con tour is an arc o f th e unit circle. The latter contour w ould allow us to com p ute the frequency con ten t o f the sequence x ( n ) at a dense set o f L freq u en cies in the range covered by the arc w ithout having to com pute a large D F T , that is, a D F T o f the seq u en ce x ( n ) pad d ed with many zeros to obtain the desired resolution in frequency. Finally, if r0 = Ro = 1, = 0, 0o = 2 n / N , and L = N , the contour is the entire unit circle and the frequencies are those o f the D F T . T h e various contours are illustrated in Fig. 6.18. W hen points {z*J in (6.3.12) are substituted in to the exp ression for the ztransform, w e obtain * ( z t ) = X ! -* ( > z r i n=0 n=
N-1
(6.3.13)
= j > ( n ) ( r 0e j * ) ~ V
Sec. 6.3
A Linear Filtering Approach to Computation of the DFT lm(r) ImU)
483
n=0
lm (;l
Im(;)
Figure 6.18 Some examples of contours on which we may evaluate the ztransform. w here, by definition. V = R veJ^ W e can exp ress (6.3.13) in the form o f a con volu tion , by n oting that nk = j[n 2 + k 2 (k n) 2] Substitution o f (6.3.15) into (6.3.13) yield s
N- 1
(6.3.14)
(6.3.15)
X(C*) = V -
l ' Z /2
J 2 [ x ( n ) ( r 0eJlk' ) - nV - n2f2] V (k- n)2fZ
(6.3.16)
Let us define a n ew seq u en ce g ( n ) as g (n) = x ( n )( r ^ e j<hr n V - n^ T h en (6.3.16) can b e exp ressed as

X ( z k) = V - k2/2y g ( n ) V {k-',)1/2
(6.3.17)
(6.3.18)
484
Chap. 6
T he sum m ation in (6.3.18) can be interpreted as the con volu tion o f the sequ en ce g (n) with the im pulse resp onse h (n) o f a filter, where h(n) = V n2/2 C onsequently, (6.3.18) m ay b e expressed as X(zt) = V -*Vy(k) (6.3.20) (6.3.19)
w here y(Jt) is the ou tp u t o f the filter s 1 y ( k ) = Y ' g ( n ) h ( k n) n=0 k = 0. 1.........L 1 (6.3.21)
W e observe that b oth h (n) and g(n) are com p lex-valu ed seq u en ces. T he sequ en ce h (n) with R 0 = 1 has the form o f a com plex exp on en tial with argum ent (on = n 2< (> o /2 = (n 0 o /2 )n. T h e quantity rep resents the frequency o f the com plex exp on en tial signal, which increases linearly with tim e. Such signals are used in radar system s and are called chi rp signals. H en ce th e z-transform evaluated as in (6.3.18) is called the chi rp-z t ransf orm. T h e linear con volu tion in (6.3.21) is m ost efficien tly d on e by use o f the FFT algorithm . T he seq u en ce g( n) is o f length N . H ow ever, h{n) has infinite du ration. Fortunately, only a portion h{n) is required to co m p u te the L values o f X (z). Since w e will com p ute the con volu tion in (6.3.1) via the F FT, let us consider the circular con volu tion o f the W-point seq u en ce g{n) with an M -p oint section of /i(n), w here M > N . In such a case, w e k n ow that the first N 1 p oin ts contain aliasing and that the rem aining M N + 1 p oints are identical to the result that would b e obtained from a linear con volu tion o f h( n) with g(n). In view o f this, we should select a D F T o f size M - L + N - 1 which would yield L valid p oin ts and N - 1 points corrupted by aliasing. T he section o f h(n) that is n eed ed for this com putation corresp on d s to the values o f h{ri) for ( N - 1) < n < (L 1), which is o f length M = L + N 1, as observed from (6.3.21). Let us define the seq u en ce h \{n ) o f length M as /ii(n ) = h(n N -f 1) n 0 , 1 .........M 1 (6.3.22)
and com p ute its Af-poin t D F T via the FFT algorithm to obtain H \ ( k ) . F rom x (n ) w e com p ute g ( n ) as specified by (6.3.17), pad g(n ) w ith L 1 zeros, and com pute its Af-point D F T to yield G(Jfc). T h e ID F T o f th e product y i(* ) = G ( k ) H \( k ) yields the Af-point seq u en ce > i(n ), n = 0, 1 , . . . , Af 1. T h e first N 1 p oints of y i( ) are corrupted by aliasing and are discarded. T h e desired valu es are yi(n) f o r N 1 < n < M 1, w hich correspond to the range 0 < n < L l i n (6.3.21),
Sec. 6.3 that is,
A Linear Filtering Approach to Computation of the DFT
485
y(n) = y t ( n + N 1)
n = 0, 1.........L 1
(6.3.23)
A ltern atively, w e can define a seq u en ce ft 2 (n) as h 2(n) = h (n), h ( n ~ N - L + l),
< n < L 1 L < ti < M 1
(6.3.24)
The A f-point D F T o f h 2{n) yields H2(k), which w hen m ultiplied by G( k ) yields Y2(k) = G( k ) Hz ( k ) . T he ID F T o f Y2(k) yield s the seq u en ce y2( n ) for 0 < n < A f - 1 . N o w the desired valu es o f > 2 (") are in the range 0 < n < L 1, that is, y ( n ) = y 2(n) n = 0, 1 , . . . , L 1 (6.3.25)
Finally, the com p lex valu es X(Zi) are com p uted by dividing y( k) by h ( k ), k = 0, 1.........L 1, as specified by (6.3.20). In gen eral, the com p utational com p lexity o f the chirp-z transform algorithm described ab ove is o f the order of Af log 2 M com plex m ultiplications, where M = N + L ~ 1. T h is num ber should be com pared with the product, N L, the num ber o f com p utations required by direct evaluation o f the z-transform . Clearly, if L is sm all, direct com p utation is m ore efficient. H ow ever, if L is large, then the chirp-z transform algorithm is m ore efficient. T h e chirp-z transform m eth od has b een im plem ented in hardware to com pute the D F T o f signals. For the com putation of the D FT, w e select ro = /?(i = 1, 6\j = 0, < /> o = 2 n / N , and L = N. In this case
y-ir/2 _
e -jjin -/N
nn2 . Tin2 = c o s --------- j s i n ------N N T he chirp filter with im pulse response h( n) = V nlfl t2 . nn2 = c o s -------i n ---- (- j/ s sin N N = h r(n) + jh , { n )
< 6 '3 '26 >
(6.3.27)
has b een im p lem en ted as a pair o f F IR filters w ith coefficients h r (n) and A,(n), resp ectively. B o th su rface acous tic w av e (SAW ) d evices and charge co u p l e d d e vices (C C D ) h ave b een u sed in practice for the F IR filters. T h e cosine and sine seq u en ces given in (6.3.26) n eed ed for the prem ultiplications and postm ultiplica tion s are usually stored in a read-only m em ory (R O M ). Furtherm ore, w e n ote that if o n ly the m agnitu d e o f the D F T is desired, the postm ultiplications are u n n eces sary. In this case, |X (z*)l = \y(k)\ k = 0 ,1 ,...,
n
-1
(6.3.28)
as illustrated in Fig. 6.19. T hus the linear F IR filtering approach using th e chirp-z transform has b een im p lem en ted for the com putation o f the D F T .
486
Chap. 6
Chirp Fillers
Figure 6.19 Block diagram illustrating the implementation of the chirp-z transform for com puting the DFT (magnitude only).
6.4 QUANTIZATION EFFECTS IN THE COMPUTATION OF THE DFT*

A s w e have ob served in our p reviou s discussions, the D F T plays an im portant role in m any digital signal p rocessing applications, including F IR filtering, the com pu tation o f the correlation betw een signals, and spectral analysis. For this reason it is im portant for us to kn ow th e effect o f quantization errors in its com puta tion. In particular, w e shall consider the effect o f rou n d -off errors due to the m ultiplications perform ed in the D F T with fixed-point arithm etic. T h e m odel that w e shall adopt for characterizing rou n d -off errors in m ulti plication is the additive w hite n oise m o d el that w e use in the statistical analysis o f Tound-off errors in IIR and F IR filters (see Fig. 7.34). A lth ou gh the statistical
*It is recommended that the reader review Section 7.5 prior to reading this section.
Sec. 6.4
Quantization Effects in the Computation of the DFT
487
analysis is perform ed for rounding, the analysis can be easily m odified to apply to truncation in tw o's-com p lem en t arithm etic (see Sec. 7.5.3). O f particular interest is the analysis o f rou n d -off errors in the com putation o f the D F T via the FFT algorithm . H ow ever, w e shall first establish a benchm ark by determ ining the round-off errors in the direct com p utation o f the D F T .
6.4.1 Quantization Errors in the Direct Computation of the DFT

G iven a finite-duration seq u en ce (jt(n)], 0 < n < N 1, the D F T o f {jc(h)1 is defined as
A/-1
* (* ) = Y l x ( n ) w "' j,=0
= 0 , 1 ........ N - 1
(6.4.1)
w here IV yv = c ~ )2r,/N. W e assum e that in general, {*()] is a com p lex-valu ed se quence. W e also assum e that the real and im aginary com p on en ts o f {a (h)I and {VV^"] are represented by b bits. C onsequently, the com putation o f the product requires four real m ultiplications. Each real m ultiplication is rounded from 2b bits to b bits, and hence there are four quantization errors for each com p lex-valu ed m ultiplication. In the direct com putation o f the D F T , there are N com p lex-valu ed m ultiplica tions for each point in the D FT. T herefore, the total num ber o f real m ultiplications in the com putation o f a single point in the D F T is 4 N. C on sequ en tly, there are 4 N quantization errors. Let us evaluate the variance o f the quantization errors in a fixed-point com putation o f the D F T . First, w e m ake the follow in g assum ptions about the statistical properties o f the quantization errors. 1. T h e quantization errors due to rounding are uniform ly distributed random variables in the range (A /2 , A /2 ) where A = 2~ b. 2. T h e 4 N quantization errors are m utually uncorrelated.
3. T h e 4 N quantization errors are uncorrelated with the seq u en ce |jc {}}.

Since each o f the quantization errors has a variance
A 2
7~2b <6A2>
" ' = 1 2 = 1 2 the variance o f the quantization errors from the 4 N m ultiplications is <r2 = 4 N o ]
(6A3)
3
H en ce the variance o f the quantization error is p roportional to the size o f D FT. N o te that w hen Af is a p ow er o f 2 (i.e., N = 2 1 ), the variance can be expressed
488
Chap. 6
2 2(h 1 * /2 a ] = -------------
(6.4.4)
This expression im plies that every fourfold increase in the size N o f the D F T requires an additional bit in com putational precision to offset the additional quan tization errors. T o prevent overflow , the input seq u en ce to the D F T requires scaling. Clearly, an upper bound on | X (A :) | is
A '- l
[* ()! < Y
n=0
|*(n )|
1
(6.4.5) requires that (6.4.6)
If the dynam ic range in addition is ( - 1 , 1 ) , then |X (/:)| <

A '-l
Y
n=0
|jr(/i)| < 1
If U (/i)| is initially scaled such that |a (/j)| < 1 for all n, then each point in the seq u en ce can be divided by N to ensure that (6.4.6) is satisfied. T h e scaling im plied by (6.4.6) is extrem ely severe. For exam p le, su p p ose that the signal seq u en ce {*()} is white and. after scaling, each valu e |.r(n)l o f the seq u en ce is uniform ly distributed in the range (-1 /7 V , I / N ) . T h en the variance of the signal sequ en ce is 2 = ---------< 2 / * > 2 = 1 o,
12
3N1
tzA-n (6.4.7)
>
and the variance o f the output D F T coefficients |Jf(/t)l is al = N a2 1 ~ 3 ~N Thus the signal-to-noise p ow er ratio is (6.4.9) W e observe that the scaling is responsible for reducing th e S N R by N and the com bination o f scaling and quantization errors result in a total reduction that is proportional to N 2. H en ce scaling the input seq u en ce (j(n )} to satisfy (6.4.6) im poses a severe p en alty on the signal-to-noise ratio in the D F T . Exam ple 6.4.1 Use (6.4.9) to determ ine the num ber of bits required to com pute the D FT of a 1024point sequence with a SNR of 30 dB. Solution (6.4.8)
The size of the sequence is N = 210. Hence the SNR is
Sec. 6.4
489
For an SNR o f 30 dB, we have 3(2* - 20) = 30 b = 15 bits N ote that the 15 bits is the precision for both multiplication and addition. Instead o f scaling the input sequ en ce {Jt(n)}, suppose w e sim ply require that |x(n)l < 1. T h en w e m ust provide a sufficiently large dynam ic range for addition such that |* ( * ) l < N . In such a case, the variance o f the seq u en ce {|jc(n)|) is a 2 = 5 , and h en ce th e variance o f |X (* )| is (6.4.10) C on sequ en tly, the S N R is (6.4.11) If w e repeat the com putation in E xam ple 6.4,1, w e find that the num ber o f bits required to a ch ieve a S N R o f 30 dB is b = 5 bits. H ow ever, w e n eed an additional 1 0 bits for the accum ulator (th e adder) to accom m odate the increase in the dynam ic range for addition. A lthou gh w e did not ach ieve any reduction in the dynam ic range for addition, we have m anaged to reduce the p recision in m ultiplication from 15 bits to 5 bits, which is highly significant.
6.4.2 Quantization Errors in FFT Algorithms

A s w e have sh ow n , the F F T algorithm s require significantly few er m ultiplications than the direct com p utation o f the D F T . In view o f this w e m ight con clu d e that the com p utation o f the D F T via an FFT algorithm w ill result in sm aller quantization errors. U n fortu n ately, that is n ot the case, as w e will dem onstrate. L et us con sid er the use o f fixed-point arithm etic in th e com putation o f a radix-2 F F T algorithm . T o be specific, w e select the radix-2, decim ation-in-tim e algorithm illustrated in Fig. 6.20 for the case N = S. T h e results on quantiza tion errors that w e ob tain for this radix-2 FFT algorithm are typical o f th e results o b ta in ed w ith o th er radix - 2 and higher radix algorithm s. W e o b serv e that each butterfly com putation in volves o n e com plex-valued m ultiplication or, eq u ivalen tly, four real m ultiplications. W e ignore the fact that so m e butterflies con tain a trivial m ultiplication by 1 . If w e consider th e but terflies that affect the com p utation o f any on e valu e o f the D F T , w e find that, in gen eral, there are N /2 in the first stage of the FFT , N / 4 in the secon d stage, N / 8 in the third state, and so on , until the last stage, w here there is on ly on e. C on sequ en tly, th e num ber o f butterflies per output point is
2 " - '+ 2 " - 2 + --- + 2 + l = 2v [ l + ( ! ) + + ( j ) ']

= 2 [ l - ( j n = W -l
490

Stage 1 Stage 2 Stage 3
Chap. 6
For exam ple, the butterflies that affect the com p utation o f A"(3) in the eight-point FFT algorithm o f Fig. 6.20 are illustrated in Fig. 6.21. T h e quantization errors introduced in each butterfly propagate to the output. N o te that the quantization errors introduced in the first stage p ropagate through (v - 1 ) stages, th ose introduced in the second stage propagate through (v - 2 ) stages, and so on. A s these quantization errors propagate through a num ber of su bsequent stages, th ey are phase shifted (ph ase rotated) by th e phase factors W^n. T h ese phase rotations do not change the statistical p rop erties o f the quan tization errors and, in particular, the variance o f each q uantization error remains invariant. If w e assum e that the quantization errors in each butterfly are uncorrelated with the errors in other butterflies, then there are 4(W - 1 ) errors that affect the output o f each point o f the FFT. C on sequ en tly, th e variance o f the total quanti zation error at the output is
A
f f| = 4 (A r- ! )
(6.4.13)
Sec. 6.4
491
Figure 6.21 w here A = 2 h. H en ce
Butterflies that affect the computation o f X (3).
a2= j
2"
(6.4.14)
T his is exactly the sam e result that w e ob tain ed for the direct com p utation o f the DFT. T h e result in (6.4.14) should n ot b e surprising. In fact, the FFT algorithm d o es not reduce the num ber o f m ultiplications required to com p ute a single point o f the D F T . It d o es, h ow ever, exp loit the p eriod icities in W^n and thus reduces the num ber o f m ultiplications in the com p utation o f the entire block o f N points in the D F T . A s in the case o f the direct com p utation o f the D F T , w e m ust scale the input seq u en ce to prevent overflow . R ecall that if |jc(n) | < \ / N , 0 < n < N 1, then |X (* )| < 1 for 0 < k < N 1. T hus overflow is avoided. W ith this scaling, the relation s in (6.4.7), (6.4.8), and (6.4.9), ob tain ed previously for the direct com p utation o f the D F T , apply to the F F T algorithm as w ell. C onsequently, the sam e S N R is o b tain ed for the FFT. Since the F F T algorithm consists o f a seq u en ce o f stages, w h ere each stage con tains butterflies that in volve pairs o f points, it is p ossib le to d evise a differ en t scaling strategy that is n ot as severe as dividing each input p oin t by N . This
492
Chap. 6
alternative scaling strategy is m otivated by the observation that the in term edi ate values [Xn(/r)| in the n = 1, 2,..., u stages o f th e F F T algorithm satisfy the conditions (see P roblem 6.35) m ax[|X n+1 ( * ) U X n+1(/)|] > m ax[|X n( * ) U X B( 0 |] (6.4.15) m ax [|X n+1 ( * ) |,|X B+1(/)|] <
2
m ax[jX (Jt)|,|X n(/)|]
In view o f th ese relations, w e can distribute the total scalin g o f 1 / N in to each o f the stages o f the F F T algorithm . In particular, if |jr(n)| < 1, w e apply a scale factor o f 5 in the first stage so that |jr(n)| < T h en the output o f each subsequent stage in the FFT algorithm is scaled by | , so that after v stages w e have achieved an overall scale factor o f ( j ) 1 ' = 1 //V. Thus overflow in the com p utation o f the D F T is avoided. This scaling procedure d o es not affect the signal level at the output o f the FFT algorithm , but it significantly reduces the variance o f the quantization errors at the output. Specifically, each factor o f ^ reduces the variance o f a quantization error term by a factor o f Thus the 4 ( N / 2 ) quantization errors introduced in the first stage are reduced in variance by (^ V - 1 , the 4 ( N / 4 ) quantization errors introduced in the second stage are reduced in variance by ( j ) 1 - 2 . and so on. C on sequ en tly, the total variance o f the quantization errors at the output o f the FFT algorithm is
w here the factor (^ )tJ is negligible. W e now o b serve that (6.4.16) is n o longer proportional t o N . O n th e other hand, the signal has th e variance a \ = 1 /3 N , as given in (6.4.8). H e n c e the S N R is
f l = _ L .2 ^ _
2N 2 2b v \ (6.4.17)
Thus, by distributing th e scaling o f l / N uniform ly throughout th e FFT algorithm , w e have achieved an S N R that is inversely proportional to N in stead o f N 2. Example 6.4.2 Determine the number of bits required to compute an FFT of 1024 points with an SNR of 30 dB when the scaling is distributed as described above.
Sec. 6.5
493
Solution
The size of the FFT is N = 210. Hence the SNR according to (6.4.17) is 101gio 22h~v~l = 30 3(2b - 11) = 30 b = bits)
This can be com pared with the 15 bits required if all the scaling is perform ed in the first stage of the FFT algorithm.

T h e fo cu s o f this chapter w as on the efficien t com putation o f the D F T . W e d em on strated that by taking advantage o f the sym m etry and p eriodicity p roperties o f the ex p on en tial factors W#", w e can reduce the num ber o f com p lex m ultiplications n eed ed to com p ute the D F T from N 2 to N log 2 N w hen Af is a p ow er o f 2. A s w e indicated, any seq u en ce can be augm ented with zeros, such that N 2''. For d ecad es, FFT -type algorithm s were o f interest to m athem aticians w ho w ere con cern ed w ith com p utin g values o f F ourier series by hand. H ow ever, it w as not until C o o ley and T u k ey (1965) published their w ell-k now n paper that the im pact and significance o f the efficient com putation o f the D F T was recognized. Since then the C o o le y -T u k e y FFT algorithm and its various form s, for exam ple, the algorithm s o f S in gleton (1967, 1969), have had a trem en dou s influence on the use o f the D F T in con v olu tion , correlation, and spectrum analysis. For a historical p erspective on the F FT algorithm , the reader is referred to the paper by C ooley et al. (1967). T h e split-radix FFT (SR F F T ) algorithm d escribed in Section 9.3.5 is due to D u h a m el and H ollm an n (1 9 8 4 ,1 9 8 6 ). T he m irror F F T (M FF T ) and p h ase F F T (PF FT ) algorithm s w ere described to the authors by R. Price. T he exp loitation o f sym m etry p rop erties in the data to reduce the com putation tim e are described in a paper by Sw arztrauber (1986). O ver the years, a num ber o f tutorial papers have b een published on FFT algorithm s. W e cite the early papers by Brigham and M orrow (1967), Cochran et al. (1967), B ergland (1969), and C ooley et al. (1967, 1969). T h e reco g n itio n that the D F T can b e arranged and com p uted as a linear con volu tion is also highly significant. G o ertzel (1968) indicated that the D F T can b e com p uted via linear filtering, although the com p utational savings o f this approach is rath er m odest, as w e have observed. M ore significant is the work o f B lu estein (1 9 7 0 ), w ho d em onstrated that the com p utation o f the D F T can be form u lated as a chirp linear filtering operation. T his w ork led to the d evelop m en t o f th e chirp-z transform algorithm by R ab in er et al. (1969). In addition to the F F T algorithm s describ ed in this chapter, there are other efficien t algorithm s for com p utin g the D F T , som e o f w hich further reduce the
494
Chap. 6
num ber o f m ultiplications, but usually require m ore additions. O f particular im portance is an algorithm due to R ader and B renner (1976), the class o f prim e factor algorithm s, such as the G o o d algorithm (1971), and the W inograd algorithm (1976, 1978). For a d escription o f these and related algorithm s, the reader m ay refer to the text by Blahut (1985).
PROBLEMS
6.1 Show that each of the numbers eja*/NH o < /t < W - 1
corresponds to an Wth root of unity. Plot these numbers as phasors in the complex plane and illustrate, by m eans of this figure, the orthogonality property
-ja ir/N M n _ |
n=< )
N, I 0, I
if k ~ 1 if k ^ l 1
6.2 (a) Show that the phase factors can be com puted recursively by W$ = (b) Perform this com putation once using single-precision floating-point arithmetic and once using oniy four significant digits. Note the deterioration due to the accumulation of round-off errors in the later case. (c) Show how the results in part (b) can be improved by resetting the result to the correct value - j . each time gl = N/4. 6 3 Let x(n) be a real-valued N -point (N = 2' ) sequence. Develop a method to compute an N -point D FT X (k), which contains only the odd harmonics [i.e., X'(k) = 0 if Jt is even] by using only a real A,/2-spoint DFT. 6.4 A designer has available a num ber of eight-point FFT chips. Show explicitly how he should interconnect three such chips in order to com pute a 24-point DFT. 6.5 The ^-transform of the sequence x(n) = u(n) - u(n - 7) is sampled at five points on the unit circle as follows x(k) = X(z) 1 - = eJ'2jr*/5 Jt* 0,1 ,2 ,3 ,4
Determ ine the inverse D FT x'(n) of X (Jt). Com pare it with *(/t) and explain the results. 6.6 Consider a finite-duration sequence x(n), 0 < n < 7, with z-transform X(z). We wish to com pute X (:) at the following set of values: zk = 0.8ejf|(2)r*/*,+( ' /8)] 0 < J fc < 7
( b )
(a) Sketch the points {z*} in the complex plane. D eterm ine a sequence s ( n ) such that its D F T provides the desired samples of
*(z).
Chap. 6
Problems
495
6.7 Derive the radix-2 decimation-in-time FFT algorithm given by (6.1.26) and (6.1.27)
as a special case of the more general algorithmic procedure given by (6.1.16) through (6.1.18). 6.8 Com pute the eight-point D FT of the sequence X^ 1. I 0, 0 < n < 7 otherwise
by using the decimation-in-frequency FFT algorithm described in the text. 6.9 Derive the signal flow graph for the N = 16 point, radix-4 decimation-in-time FFT algorithm in which the input sequence is in norm al order and the computations are done in place. 6.10 D erive the signal flow graph for the N = 16 point, radix-4 decimation-in-frequency FFT algorithm in which the input sequence is in digit-reversed order and the output D FT is in normal order. 6.11 Com pute the eight-point DFT of the sequence
x{n) = U . i
1,0. 0,0.0
12 2 2 2
using the in-placc radix-2 dccimation-in-time and radix-2 decimation-in-frequency al gorithms. Follow exactly the corresponding signal flow graphs and keep track of all the interm ediate quantities by putting them on the diagrams, 6.12 Com pute the 16-point D FT of the sequence x(n) = cos 0 < n < 15
using the radix-4 decimation-in-time algorithm. 6.13 Consider the eight-point decimation-in-tim e (D IT) flow graph in Fig. 6.6. (a) What is the gain of the signal p ath that goes from x(7) to X(2)7 (b) How many paths lead from the input to a given output sample? Is this true for every output sample? (c ) Com pute X (3) using the operations dictated by this flow graph. 6.14 Draw the flow graph for the decimation-in-frequency (D IF) SRFFT algorithm for N = 16. W hat is the num ber of nontrivial multiplications? 6.15 D erive the algorithm and draw the N = 8 flow graph for the D IT SRFFT algorithm. Com pare your flow graph with the D IF radix-2 FFT flow graph shown in Fig. 6.11. 6.16 Show that the product of two complex numbers (a + j b ) and ( C + j d ) can be perform ed with three real multiplications and five additions using the algorithm
x R xi
=
=
(a (a
b)d b)d
+ (c + (c +
d)a d)b
where
X = x R
jx ,
= (a +
jb )(c
jd )
6.17 Explain how the D FT can be used to com pute N equispaced samples of the ztransform , of an iV-point sequence, on a circle of radius r.
496
Chap. 6
6.18 A real-valued A/-point sequence Jt(n) is called D FT bandlim ited if its D FT X(k) = 0 for ko < k < N An. We insert (L 1)N zeros in the middle of A'(Jt) to obtain the following L N -point DFT X(k), X'(k) = { 0, X(k + N - LN) . Show that Lx'(Ln) = x(n) where x\n)
LN
0 <i < A o 1 Jt0 < J k < L N - kt, L N - J t + 1 < Jt < L N 1
0 < n < N 1
X (k)
Explain the meaning of this type of processing bv working out an example with N = 4, L = 1. and A ( J t ) = { 1 , 0 . 0 . 1 ) . 6.19 Let X(k) be the A'-point DFT of the sequence .v(n). 0 < n < A - 1. What is the A-point DFT of the sequence s(n) = X(n). 0 < n < N - 1? 6.20 Let X(k) be the A-point DFT of the sequence \(n ), (I < < N 1. We define a 2 N -point sequence _v() as V{;I) 1-v^V --vcn n odd Express the 2 Appoint DFT of y(n) in terms of X(k). 6.21 (a) Determ ine the ^-transform W (-) of the Hanning window w () = (1 - cos fl. (b) Determ ine a formula to compute the JV-point DFT X(k) of ihe signal .r.(n) = w(n)x(n)< 0 < n < N 1. from the JV-point DFT A'(A ) of the signal .r(n). 6.22 Create a DFT coefficient table that uses only N/4 memory locations to store the first quadrant of the sine sequence (assume N even). 6.23 D eterm ine the com putational burden of the algorithm given by (6.2.12) and compare it with the com putational burden required in the 2 N -point DFT of g(n). Assume that the FFT algorithm is a radix-2 algorithm. 6.24 Consider an IIR system described by the difference equation
S' M
v(n) = - ^
*= 1
ai-v< n~
+ y , b k x ( n - k)
Describe a procedure that com putes the frequency response H A ' - l using the FFT algorithm ( N = 2).
Jt = 0, 1........
6.25 Develop a radix-3 decimation-in-time FFT algorithm for N = 3 and draw the corre sponding flow graph for N = 9. W hat is the num ber of required com plex multiplica tions? Can the operations be perform ed in place? 6.26 Repeat Problem 6.25 for the D IF case. 6.27 FFT input and output pruning In many applications we wish to com pute only a few points M of the Appoint D FT of a finite-duration sequence of length L (i.e., M N and I < < N).
Chap. 6
Problems
497
(a) Draw the flow graph of the radix-2 D IF FFT algorithm for N = 16 and eliminate [i.e., prune] all signal paths that originate from zero inputs assuming that only jc(0) and jc(1) are nonzero. (b) R epeat part (a) for the radix-2 D IT algorithm. (c) Which algorithm is better if we wish to com pute all points of the D FT? What happens if we want to compute only the points X (0), X (l), X (2), and X (3)? Establish a rule to choose between D IT and D IF pruning depending on the values of M and L. (d) Give an estim ate of saving in computations in term s of M, L, and N. 6.28 Parallel computation o f the D F T Suppose that we wish to com pute an N = 2P2V point D FT using 2P digital signal processors (DSPs). For simplicity we assume that p = v = 2. In this case each DSP carries out all the com putations that are necessary to com pute 2V D FT points. (a) Using the radix-2 D IF flow graph, show that to avoid data shuffling, the entire sequence x(n) should be loaded to the memory of each DSP. (b) Identify and redraw the portion of the flow graph that is executed by the DSP that com putes the D FT samples X(2), X(10), X(6), and X(14). (c) Show that, if we use M = 2'' DSPs, the com putation speed-up S is given by S= M log-, N log2 N - log, M + 2 (M - 1)
6.29 Develop an inverse radix-2 D IT FFT algorithm starting with the definition. Draw the flow graph for com pulation and com parc with the corresponding flow graph for the direct FFT. Can the IFFT flow graph be obtained from the one for the direct FFT? 6.30 R epeat Problem 6.29 for the D IF case. 6.31 Show that an FFT on data with H erm itian symmetry can be derived by reversing the flow graph of an FFT for real data. 6.32 D eterm ine the syslem function H(z) and the difference equation for the system that uses the G oertzel algorithm to compute ihe DFT value X ( N - k). 6.33 (a) Suppose that x(n) is a finite-duration sequence of N = 1024 points. It is desired to evaluate the z-transform X (;) of the sequence at the points
Zk
eiQ * * *
k = 0 . 100,200 ......1000
by using the most efficient m ethod or algorithm possible. Describe an algorithm for perform ing this com putation efficiently. Explain how you arrived at your answer by giving the various options or algorithms that can be used. (b) R epeat part (a) if X (z) is to be evaluated at zt = 2(0.9) V '1(2,r/5000*+r/21 k = 0 ,1 ,2 ........999
634 R epeat the analysis for the variance of the quantization error, carried out in Sec
tion 6.4.2, for the decimation-in-frequency radix-2 FFT algorithm.
635 The basic butterfly in the radix-2 decimation-in-time FFT algorithm is

*+,<*) = Xn( * )+ W X(l)
498
Efficient Computation of ttie DFT: Fast Fourier Transform Algorithms
Chap. 6
(a) If we require that !X(Jt)| < j and |Jf(/)| < 5 , show that
|Re[X +,(*)]| < 1,
| I m [ X + 1 (*)]l < 1,
|/te[X n+1(/)]| < 1

| / m [ A B + 1 (/)]| < 1
Thus overflow does not occur. (b) Prove that max[|X,,+i(A)|. |X+i(/)|] > m ax[|Xn(*)l.l*,,(/)|] max[|X+1(*)UX,,+1(/)|] < 2 m ax [|X (* )|.|* n (/)l] 636* Computation o f the D F T Use an FFT subroutine to compute the following DFTs and plot the m agnitudes |X(Jt)| of the DFTs. (a) The 64-point D FT of the sequence , , [1 , x(n) = I 10, n = 0 ,1 ........15 otherwise (Ni ~ 16)
(b) The 64-point D FT of the sequence 1, 10, n = 0 ,1 ........... 7 otherwise (N\ = 8)
(c) The 128-point DFT of the sequence in part (a). (d) The 64-point D FT of the sequence | 10f >'*/*>", x (n) { J 0. = 0 ,1 ........63 otherwise (Ni = 64)
Answer the following questions. (1) What is the frequency interval between successive samples for the plots in parts (a), (b). (c), and (d)? (2) What is the value of the spectrum at zero frequency (dc value) obtained from the plots in parts (a), (b), (c), (d)? From the formula X(k) = Y ^ x (n ) e ' compute the theoretical values for the dc value and check these with the com puter results. (3) In plots (a), (b), and (c), what is the frequency interval betw een successive nulls in the spectrum ? W hat is the relationship between N 1 of the sequence x( n ) and the frequency interval between successive nulls? (4) Explain the difference between the plots obtained from parts (a) and (c). 637* Identification o f pole positions in a system difference equation Consider the system described by the
y(n) = r2y(n 2) + x(n)
(a) Let r = 0.9 and x(n) = &(n). G enerate the output sequence y(n) for 0 < n < 127.
Compute the N = 128 point D FT [ I'M ) and plot {|y(Jt)t).
Chap. 6
Problems
499
(b) Com pute the N = 128 point DFT of the sequence ui(n) = (0.92)- " v(n) where y(/t) is the sequence generated in part (a). Plot the D FT values | W (t)|. W hat can you conclude from the plots in parts (a) and (b)? (c) Let r = 0.5 and repeat part (a). (d) R epeat part (b) for the sequence ui(n) = (0.55)~nv(n) where y( n ) is the sequence generated in part (c). W hat can you conclude from the plots in parts (c) and (d)? (e) Now let the sequence generated in part (c) be corrupted by a sequence of mea surem ent" noise which is Gaussian with zero mean and variance a 2 = 0.1. Repeat parts (c) and (d) for the noise-corrupted signal.
The focus o f this chapter is on the realization o f linear tim e-invariant discrete tim e system s in eith er softw are or hardware. A s w e noted in C h ap ter 2, there are various configurations or structures for the realization o f any F IR and IIR discrete tim e system . In C hapter 2 w e described the sim plest o f these structures, nam ely, the direct-form realizations. H ow ever, there are other m ore practical structures that offer som e distinct advantages, especially w hen q uantization effects are taken into consideration. O f particular im portance are the cascade, parallel, and lattice structures, which exhibit robustness in finite-w ord-length im plem en tation s. A lso described in this chapter is the frequency-sam pling realization for an F IR system , which often has the advantage o f being com putationally efficient w hen com pared with alternative FIR realizations. O ther im portant filter structures are ob tain ed by em ploying a state-sp ace form ulation for linear tim e-invariant system s. A n analysis o f system s characterized by th e state-variable form is p resen ted in both the tim e and frequency dom ains. In addition to describing the various structures for the realization o f discrete tim e system s, w e a lso treat problem s associated with q uantization effects in the im plem en tation o f digital filters using finite-precision arithm etic. T his treatm ent includes the effects on the filter frequency response characteristics resulting from coefficient quantization and the round-off noise effects inherent in the digital im p lem entation o f d iscrete-tim e system s.
7.1 STRUCTURES FOR THE REALIZATION OF DISCRETE-TIME SYSTEMS

Let us consider the im portant class o f linear tim e-invariant d iscrete-tim e system s characterized by the general linear constant-coefficient d ifferen ce eq u ation
N
(7.1.1)
500
Sec. 7.1
Structures for the Realization of Discrete-Time Systems
501
A s w e have sh ow n by m eans o f the z-transform , such a class o f linear tim e-invariant d iscrete-tim e system s are also characterized by the rational system function M H(z) = (7.1.2)
which is a ratio o f tw o polyn om ials in z - 1 . From the latter characterization, we obtain the zero s and p o les o f the system function, which d ep en d on the ch oice o f the system p aram eters {>*} and {a*} and which determ ine the frequency response characteristics o f the system . O ur fo cu s in this chapter is on the various m eth od s o f im plem en ting (7.1.1) or (7.1.2) in eith er hardware, or in softw are on a program m able digital com puter. W e shall sh ow that (7.1.1) or (7.1.2) can be im plem en ted in a variety o f ways d ep en d ing on the form in which th ese tw o characterizations are arranged. In g en eral, w e can view (7.1.1) as a com putational p rocedure (an algorithm ) for determ in in g the output seq u en ce v(/t) o f the system from the input sequ en ce x ( n ) . H o w ev er, in various ways, the com putations in (7.1.1) can be arranged into eq u ivalen t sets o f d ifferen ce equations. Each set o f eq u ation s defines a com pu tational procedure or an algorithm for im plem enting the system . From each set o f eq u ation s w e can construct a block diagram consisting o f an interconnection o f d elay elem en ts, m ultipliers, and adders. In Section 2.5 w e referred to such a block diagram as a realization o f the system or, equivalently, as a structure for realizing the system . If the system is to be im plem en ted in softw are, the block diagram or, eq u iv alently, the set o f eq u ation s that are obtained by rearranging (7.1.1), can be co n verted in to a program that runs on a digital com puter. A ltern atively, the structure in block diagram form im plies a hardware configuration for im plem enting the system . Perhaps, the o n e issue that m ay not be clear to the reader at this point is w hy w e are con sid erin g any rearrangem ents o f (7.1.1) or (7.1.2). W hy not just im plem en t (7.1.1) or (7.1.2) directly w ithout any rearrangem ent? If either (7.1.1) or (7.1.2) is rearranged in som e m anner, what are the benefits gained in the corresponding im plem en tation ? T h ese are the im portant q u estion s which are answ ered in this chapter. A t this p oin t in our d ev elop m en t, w e sim ply state that the m ajor factors that influ en ce our ch o ice o f a specific realization are com putational com p lexity, m em ory requirem ents, and finite-w ord-length effects in the com putations. C o m p u t a t i o n a l c o m p le x ity refers to the num ber o f arithm etic op eration s (m u l tiplications, d ivisions, and additions) required to com p ute an output valu e y (n ) for the system . In the past, th ese w ere the on ly item s used to m easure com putational com p lexity. H o w ev er, w ith recent d evelop m en ts in the design and fabrication o f rather sop h isticated program m able digital signal p rocessin g chips, oth er factors,
502
Chap. 7
such as the num ber o f tim es a fetch from m em ory is p erform ed or the num ber of tim es a com parison betw een tw o num bers is perform ed per output sam ple, have b ecom e im portant in assessing the com putational com p lexity o f a given realization o f a system . M e m o r y requ irem ents refers to the num ber o f memory locations required to store the system param eters, past inputs, past outputs, and any interm ediate com puted values. F inite- wor d-length effects or finite-precision effects refer to the quantization effects that are inherent in any digital im plem en tation o f the system , either in hardware or in softw are. T h e param eters o f the system m ust necessarily be repre sented with finite precision. T h e com p utations that are p erform ed in the process o f com puting an output from the system must be rounded- o ff or truncated to fit within the lim ited precision constraints o f the com p uter or the hardware used in the im plem en tation . W hether the com p utations are perform ed in fixed-point or floating-point arithm etic is an oth er consideration. A ll these p rob lem s are usually called finite-w ord-length effects and are extrem ely im portant in influencing our ch oice o f a system realization. W e shall see that different structures o f a system, which are equivalent for infinite precision, exhibit different behavior when finiteprecision arithm etic is used in the im plem entation. T h erefore, it is very important in practice to select a realization that is not very sensitive to finite-w ord-length effects. A lthou gh these three factors are the major o n e s in influencing our ch oice o f the realization o f a system o f the type described by either (7.1.1) or (7.1.2), other factors, such as w hether the structure or the realization len d s itself to parallel processing, or w h eth er the com p utations can b e pip elined, m ay play a role in our selection o f the specific im plem entation. T h ese additional factors are usually im portant in the realization o f m ore com p lex digital signal processin g algorithm s. In our discussion o f alternative realizations, w e con cen trate on the three m ajor factors just outlined. O ccasionally, w e will include som e additional factors that m ay be im portant in som e im plem entations.
7.2 STRUCTURES FOR FIR SYSTEMS

In general, an F IR system is described by the differen ce equation
Af-1
y(n) = *=0 or, equivalently, by the system function
bkx (n ~ k)
(7.2.1)
A f-1
H{z)=*Yib kZ ~ k
*=o
{122)
Furthermore, the unit sample response of the FIR system is identical to the coef-
Sec. 7.2
Structures for FIR Systems
503
ficients {>*}, that is, ,, , \ b, 0 < n < Af 1 ,7 , , ,
'K,) = | o .
otherwise
<7-2 '3>
T h e length o f the F IR filter is selected as M to conform w ith the established n otation in the technical literature. W e shall present several m eth od s for im plem en ting an F IR system , b egin ning w ith the sim p lest structure, called the direct form. A secon d structure is the cascade-form realization. T h e third structure that w e shall d escribe is the frequency-sam pling realization. Finally, w e present a lattice realization o f an FIR system . In this discussion w e follow the con ven tion often used in the technical literature, which is to use for the param eters o f an F IR system . In addition to the four realizations indicated ab ove, an F IR system can be realized by m eans o f the D F T , as described in S ection 6.2. From on e point o f view, the D F T can be considered as a com putational procedure rather than a structure for an F IR system . H ow ever, when the com putational procedure is im plem ented in hardw are, there is a corresponding structure for the F IR system . In practice, hardware im plem en tation s o f the D F T are based on the use o f the fast Fourier transform (FFT ) algorithm s described in C hapter 6 .
7.2.1 Direct-Form Structure

T he direct-form realization follow s im m ed iately from the nonrecursive difference eq u ation given by (7.2.1) or, equivalently, by the con volu tion sum m ation M- 1 y ( n ) = ^ h ( k ) x ( n - k) t=o (7.2.4)
T h e structure is illustrated in Fig. 7.1. W e o b serve that this structure requires Af 1 m em ory locations for stor ing th e Af 1 previous inputs, and has a com p lexity o f Af m ultiplications and M 1 additions p er output point. Since the output con sists o f a w eigh ted linear com b ination o f Af 1 past values o f the input and the w eigh ted current value o f th e input, th e structure in Fig. 7.1, resem b les a tapped d elay line or a transversal
Figwre 7.1
Direct-form realization of FIR system.
504
Chap. 7
Figure 7.2 Direct-form realization of linear-phase FIR system (Af odd). system . C on sequ en tly, the direct-form realization is often called a transversal or tapped-delay-line filter. W hen the F IR system has linear phase, as described in S ection 8.2, the unit sam ple response o f the system satisfies either the sym m etry o r asym m etry condition
h(n) = h ( M - 1 - n)
(7.2.5)
For such a system the num ber o f m ultiplications is reduced from M to M f l for Af even and to (Af l ) / 2 for M odd. For exam p le, the structure that takes advantage o f this sym m etry is illustrated in Fig. 7.2 for the case in which M is odd.
7.2.2 Cascade-Form Structures

T he cascade realization follow s naturally from the system fun ction given by (7.2.2). It is a sim ple m atter to factor H ( z ) into secon d-ord er F IR system s so that
K
fl(z) = [] fl* ( z )
*=i where H k (z) = bk0 + bk]Z~l + bk2z~ 2 * = 1 ,2 ......... K
0-2.6)
(7.2.7)
and K is the in teger part o f ( M + l ) /2 . T h e filter param eter bo m ay be equally distributed am ong the K filter section s, such that bo = biobw b Ka or it m ay be assigned to a sin gle filter section . T h e zeros o f H ( z ) are grou p ed in pairs to pro d uce the secon d-ord er F IR system s o f the form (7.2.7). It is alw ays desirable to form pairs o f com p lex-con ju gate roots so that th e coefficien ts \b ki} in (7.2.7) are real valued. O n th e other hand, real-valued roots can b e paired in any arbitrary m anner. T h e cascade-form realization alon g with th e basic secon d -ord er section are sh ow n in Fig, 7.3.
Sec. 7.2 jr(n)=jr,(n)
Structures for FIR Systems H](z)

>'[(") = H2(z) *2< ) jr,(n)
(a)
505
>'2<n) =
Vjr(n) = ,Y (n )
Figure 7.3 Cascade realization of an FIR system.
In the case o f linear-phase F IR filters, the sym m etry in h (n) im plies that the zeros o f H ( z ) a lso exhibit a form o f sym m etry. In particular, if Zk and z* k are a pair o f co m p lex-con ju gate zeros then 1 f z t and 1 / z \ are also a pair o f com plex-conjugate zero s (see Sec. 8.2). C on sequ en tly, w e gain som e sim plification by form ing fourthorder section s o f the FIR system as follow s Hk(z) = c ,( l - z / t z - K l - c ^ ' K l - z ~ l / z k )(\ - z - ' / z V , (7.2.8)
= Q o + Q - l t " 1 + Ck2Z~ " + Q t Z - + Z ~ A
w here the coefficien ts {ct i } and (c^ f are functions o f zt- Thus, by com bining the tw o pairs o f p o les to form a fourth-order filter section , w e have reduced the n um ber o f m ultiplications from six to three (i.e., by a factor o f 50% ). Figure 7.4 illustrates the b asic fourth-order F IR filter structure.
Figure TA Fourth-order section in a cascade realization o f an FIR system.
506
Chap. 7
7.2.3 Frequency-Sampling Structures*

T he frequency-sam pling realization is an alternative structure for an F IR filter in which the param eters that characterize the filter are the valu es o f the desired frequency resp onse instead o f the im pulse response h(n). T o d erive the frequencysam pling structure, w e specify the desired frequency response at a set o f equally spaced frequencies, nam ely 2 tt cot = (.k + a ) M M - 1 A : = 0 , 1 , . . . , - 2 M A : = 0, 1 , . . . , -------1 M od d M even
a =
or j
and so lv e for the unit sam ple response h( n) from these equally spaced frequency specifications. Thus we can write the frequency response as M- 1 h ( n ) e ~ jum
n 0
and the values o f H(a>) at freq u en cies a>t = ( 2 n / M ) { k + a ) are sim ply H (k + a ) = H + cr)^ (7.2.9) = h (n )e -j2*lk+aWM k _ 0< 1____ A/ - 1
T he set o f values { //( -(-a )} are called the frequency sam ples o f H(a)). In the case w here a = 0, j//(Jt)} corresponds to th e M -point D F T o f (A(n)}. It is a sim ple m atter to invert (7.2.9) and express h(n) in term s o f the fre quency sam ples. T he result is i M- 1 h(n) = J ' H ( k + a ) e j2*(i+a)n/M n = 0, 1 , . . . , M - 1 (7.2.10)
W hen a = 0, (7.2.10) is sim ply the ID F T o f {//(& )}. N ow if w e use (7.2.10) to substitute for h (n) in the z-transform H ( z ) , w e have
u- \
H( z ) = > ( " ) ; "
n=0 M- 1
-L
filters.
i u-1 Y \ H ( k + a ) e j2* il'H' )"/" M *=0
(7.2.11)
^The reader may also refer to Section 8.2.3 for additional discussion o f frequency-sampling FIR
Sec. 7.2
507
B y interchanging the order o f the tw o sum m ations in (7.2.11) and perform ing the sum m ation over the index n w e obtain
H( z ) = ^ H ( k + a)
I M-1 ^ êj2n(i+a)/M - \ y (7.2.12)
} - Z- Mej2na y l
M 2. 1
H ( k + a)
_ e j2nik+a)/M z - 1
Thus the system function H ( z ) is characterized by the set o f frequency sam ples (W(Jt-t-cr)j instead o f {h ()). W e v iew this F IR filter realization as a cascade o f tw o filters [i.e., H { z ) = / / i ( z ) / / 2 (s)]- O ne is an all-zero filter, or a com b filter, with system function H x(z) = (1 - r wf y2l ) M Its zeros are located at equally spaced points on the unit circle at Zt = T he second filter with system function
Af 1
(7.2.13)
jt = 0 , 1 .........A f - 1
H l i z ) 2 ^ -J _ Jt=0
j2n(t+a)/M *
(7.2.1 )
consists o f a parallel bank o f sin gle-p ole filters with resonant frequencies Pl = e}2* ik+ayM k = 0 , l .........M 1
N o te that the p o le locations are identical to the zero locations and that both occur at a)k = 2 k ( k + a ) / M , which are the freq u en cies at which the desired fre quency resp onse is specified. T he gains o f the parallel bank o f resonant filters are sim ply the com p lex-valu ed param eters \ H ( k + a )}. T h is cascade realization is illustrated in Fig. 7.5. W hen the desired frequency resp onse characteristic o f the F IR filter is nar row band, m ost o f the gain param eters \ H ( k + a )) are zero. C on sequ en tly, the corresponding resonant filters can be elim inated and on ly the filters with nonzero gains n eed be retained. T h e n et result is a filter that requires few er com p uta tion s (m ultiplications and additions) than the corresponding direct-form realiza tion. Thus w e ob tain a m ore efficient realization. T h e frequency-sam pling filter structure can be sim plified further by exploiting the s y m m e t r y in H ( k + a ) , nam ely, H ( k ) = H * ( M k) for a = 0 and H (* + i ) = H* ( M - k - | ) for a = \
T h ese relations are easily deduced from (7.2.9). A s a result o f this sym m etry, a pair o f sin g le-p o le filters can b e com b ined to form a sin gle tw o -p o le filter with
508
Chap. 7
real-valued param eters. Thus for a = 0 the system function Hz (z) reduces to H ( 0) , fiiiz) = t _ l + H i 0) A ik ) + B(lc)z - 1 ----- 7 2 c o s { 2 n k / M ) z ~ ' + z~2 , lM ^ ~ ] A{k) + B { k ) z ~ x w M odd
^ t=1
, H iM /2)
H^ > = r r p r + T T P - +
2cos(2;rfc/Af)z -1 + z ~ 2
M even
(7.2.15)
Sec. 7.2
509
w h ere, by definition, A( k ) = H ( k ) + H ( M - k ) B( k) = H { k ) e ~ i2nklM + H ( M Sim ilar exp ression s can b e ob tain ed for a = Example 7.2.1 Sketch the block diagram for the direct-form realization and the frequency-sampling realization of the M = 32, a = 0, linear-phase (symmetric) F IR filter which has frequency samples 1, * = 0 ,1 ,2
(7.2.16)
(f)-
5 0,
4= 3
* = 4 . 5 ........15
Compare the computational complexity of these two structures. Solution Since the filter is symmetric, we exploit this symmetry and thus reduce the num ber of multiplications per output point by a factor of 2, from 32 to 16 in the direct-form realization. The number of additions per output point is 31. The block diagram of the direct realization is illustrated in Fig. 7.6. We use the form in (7.2.13) and (7.2.15) for the frequency-sampling realization and drop all terms that have zero-gain coefficients |H(k)}. T he nonzero coefficients are H(k) and the corresponding pairs are H( M - k), for k = 0 ,1 , 2, 3. The block diagram of the resulting realization is shown in Fig. 7.7. Since H( 0) = 1, the single pole filter requires no multiplication. The three double-pole filter sections require three multiplications each for a total of nine multiplications. The total num ber of additions is 13. Therefore, the frequency-sampling realization of this FIR filter is computationally more efficient than the direct-form realization.
Figure 7.6
Direct-form realization of Af = 3 2 FIR filter.
510
Chap. 7
Figare 7.7
Frequency-sampling realization for the FIR filter in Exam ple 7.2.1.
Sec. 7.2
511
7.2.4 Lattice Structure

In this sectio n w e introduce an oth er F IR filter structure, called the lattice filter or lattice realization. Lattice filters are used exten sively in digital sp eech processing and in the im plem en tation o f adaptive filters. L et us b egin the d evelop m en t by considering a seq u en ce o f F IR filters with system functions Hm(z) = A m(z) m = 0 . 1 , 2 ......... M - 1 (7.2.17)
w here, by definition, A m(z) is the polyn om ial m A m(z) = 1 + a m(k)z~ k *=i
m > 1
(7.2.18)
and /lo(^) = ! T he unit sam ple resp onse o f the m th filter is /j,(0) = 1 and h m{k) = a m(k), k = 1, 2 , . . . , m. T he subscript m on the p olyn om ial Am(c) d en otes the d egree o f the polynom ial. For m athem atical con ven ien ce, w e define a, (0) = 1. If (*(n)} is the input seq u en ce to the filter A,{z) and (y(n )( is the output seq u en ce, w e have m v() = -*() + ^ o r m(A:)A-(?f k) (7.2.19) *=l T w o direct-form structures o f the F IR filter are illustrated in Fig. 7.8.
Figure 1A
Direct-form realization of the FIR prediction filter.
512
Chap. 7
In C hapter 11, w e show that the F IR structures sh ow n in Fig. 7.8 are inti m ately related w ith the topic o f linear prediction, w here
m
x ( n ) = ~ ^ 2 a m( k )x (n - k) *=i
(7.2.20)
is the o n e-step forward predicted value o f x (n), based on m past inputs, and y ( n ) = x (n) x (n ), given by (7.2.19), rep resents the p red iction error sequ en ce. In this context, the top filter structure in Fig. 7.8 is called a p re d ictio n er ror filter. N o w suppose that w e have a filter of order m = 1. T h e ou tp u t o f such a filter is >-() = x ( n ) - f a i( l)j r ( n - 1) (7.2.21) This output can also be ob tain ed from a first-order or sin gle-stage lattice filter, illustrated in Fig. 7.9, by exciting both o f the inputs by x ( n ) and selectin g the output from the top branch. Thus the output is exactly (7.2.21), if w e select /li = ori (1). T he param eter K i in the lattice is called a reflection coefficient and it is identical to the reflection coefficient introduced in the S ch u r-C ohn stability test described in Section 3.6.7. N ext, let us consider an FIR filter for w hich m = 2. In this case the output from a direct-form structure is y (n ) = x ( n ) + ff;>(l)jr(/i 1
) + ct2(2)x (n -
(7.2.22)
By cascading tw o lattice stages as show n in Fig. 7.10, it is p ossib le to obtain the sam e output as (7.2.22). Indeed, the output from the first stage is
f i (fl) = -*(n) + K]X{n - 1)

(7.2.23) i(n ) = K {x ( n ) + x ( n T h e output from the secon d stage is flin ) = f\(n ) + 1 1
) (7.2.24)
g 2(n) = K 2f i ( n ) + g i ( n - 1)
/o(") = &>(") =*()
/i() =/o(") + *io{n - 1) = *00 + X , x ( n
1)
j,{n) = tf,/oOO + SoO1 - 1) = fCix(n) + x(n - 1) Figirc 7.9 Single-stage lattice filter.
Sec. 7.2
513
Figure 7.10 Two-stage lattice filter. If w e focus our atten tion on f 2{n) and substitute for f \ ( n ) and g\ ( n 1) from (7.2.23) in to (7.2.24), w e obtain f 2(n) = jr(n) + K \ x ( n - 1) + - 1) + x ( n - 2)] (7.2.25) = x ( n ) + ^ i ( l + K 2) x( n - 1) + K 2x {n 2) N o w (7.2.25) is identical to the output of the direct-form F IR filter as given by (7.2.22), if w e eq u ate the coefficients, that is, a 2 (2) = K 2 or, eq u ivalen tly, K 2 = a 2(2) Kx = 1 + a 2(2) (7.2.27) a 2( l) = ATi(l + K 2) (7.2.26)
Thus the reflection coefficients K\ and K 2 o f the lattice can be ob tain ed from the coefficien ts {am()} o f th e direct-form realization. B y con tinu in g this process, o n e can easily d em onstrate by induction, the eq u iv a len ce b etw een an m th-order direct-form FIR filter and an w -ord er or m stage lattice filter. T h e lattice filter is gen erally d escribed by the follow in g set o f order-recursive equations: /o (n ) = gain) = x ( n ) U i n ) = / ffl_ j( ) + K mgm- X{n - 1) g m(n) = ^ / B- i W + i . - i ( i i - l ) m = 1 ,2 .........M - 1 m = 1,2,..., M 1 (7.2.28) (7.2.29) (7.2.30)
T h en the ou tp u t o f th e (A fl)-sta g e filter corresponds to the output o f an (A f1)order F IR filter, that is, y (n ) = Figure 7.11 illustrates an (Af - l)-sta g e lattice filter in b lock diagram form along w ith a typical stage that show s the com p utations sp ecified by (7.2.29) and (7.2.30). A s a co n seq u en ce o f the eq u ivalen ce b etw een an F IR filter and a lattice filter, th e ou tp u t f m(n) o f an m -stage lattice filter can b e exp ressed as
m
f m( n) = < xm( k) x( n - k) *=o
m(0) = 1
(7.2.31)
Sin ce (7.2.31) is a con v olu tion sum , it follow s that th e z-transform relationship is
Fm(z) = Am(z)X(z)
514
Chap. 7
(a)
Figure 7.11 (Af l)-stage lattice filter. or, equivalently,

F (?) F
(
A m(z) =
7}
X(z)
Fo(z)
(7.2.32)
T h e other output com p on en t from the lattice, nam ely, g m(n), can also be exp ressed in the form o f a con volu tion sum as in (7.2.31), by using another set o f coefficients, say {^m(Jt)}. T hat this in fact is the case, b ecom es apparent from ob servation o f (7.2.23) and (7.2.24). From (7.2.23) w e n ote that the filter coeffi cients for the lattice filter that produces / i ( n ) are {1 , Al} = { 1 , ofi( 1 )} w hile the coefficients for the filter with output g \ ( n ) are (AT], 1} = | a i ( l ) , 1}. W e n ote that these tw o sets o f coefficients are in reverse order. If w e con sid er the tw o-stage lattice filter, with th e output given by (7.2.24), w e find that g 2 (n) can be expressed in the form g 2(n) = K 2f \ ( n ) + i(n - 1) = K 2[x in) + K \ x { n 1)] + K \ x { n - 1) + x{ n 2) = K 2x{ n ) + K \ ( l + K 2)x{n - 1) -I- x i n - 2) = cr2 ( 2 )jc(n) + *2 ( 1 ) * (n 1
) + xin -
C onsequently, the filter coefficients are {a 2 (2), a 2 ( l ) , 1}, w h ereas the coefficients for the filter that produces the output f i i n ) are {1, a 2( \ ) , a 2 (2)}. H ere, again, the tw o sets o f filter coefficients are in reverse order. F rom this d ev elo p m en t it follow s that the ou tp u t gm{n) from an m -stage lattice filter can b e exp ressed by the con volu tion sum o f the form
m
gmin) =
A *(*M - k)
(7.2.33)
Sec. 7.2
515
w h ere the filter coefficients {& ,(*)} are associated with a filter that produces f m(n) = y ( n ) but op erates in reverse order. C onsequently, f}m (k) = a m (m k) * = 0 , 1 .........m (7.2.34)
with p m(m) = 1 . In the co n tex t o f linear prediction, su p p ose that the data x(n), x{n - 1), . . . , x ( n m + 1) is u sed to linearly predict th e signal valu e x ( n m ) by u se o f a linear filter w ith coefficients {fim(k)}. Thus the predicted value is m-l (7.2.35) Since the data are run in reverse order through the predictor, the prediction per form ed in (7.2.35) is called b a c k w a r d predictio n. In contrast, the F IR filter with system function Am(z) is called a f o r w a r d predictor. In the z-transform dom ain, (7.2.33) b ecom es G m(z) = B m( z ) X ( z ) or, eq u ivalen tly, (7.2.37) w here Bm(z ) represents the system function o f the F IR filter with coefficients {& ,(*)}, that is, (7.2.38)
4=0
(7.2.36)
Since fim(k) = a m(m *), (7.2.38) m ay b e exp ressed as m
f f l
(7.2.39)
m
T h e relationship in (7.2.39) im plies that the zeros o f the F IR filter w ith system function B m(z) are sim ply the reciprocals o f the zeros o f A ( z ) . H en ce B m(z) is called the reciprocal or reverse p olyn om ial o f A m(z). N o w that w e have estab lish ed th ese interesting relationships b etw een the direct-form F IR filter and th e lattice structure, let us return t o th e recursive lattice eq u a tio n s in (7.2.28) through (7.2.30) and transfer them to th e z-dom ain. Thus
516 w e have
Chap. 7
F0(z) = G 0(z) = X ( z ) Fm(z) = Fm_ i(z ) + J:mz - 1 Gm_ i(z ) G m(z) = + m = 1, 2 , . . . , M - 1 m =

1
(7.2.40) (7.2.41) (7.2.42)
, 2 , . . . , M 1
If w e divide each eq u ation by X (z), w e obtain the desired results in the form A0(z) = B0(z) = 1 A m{z) (7.2.43) K z ~ } Bm- i ( z ) m = 1 , 2 ......... M -
A m- X(z)
(7.2.44) (7.2.45)
Bm(z) = K n A n - i W + z - ' B ^ i z )
m = l,2 ,...,M -l
T hus a lattice stage is described in the z-dom ain by the m atrix equation (7.2.46) B efo re concluding this discussion, it is desirable to d ev elo p the relationships for converting the lattice param eters that is, the reflection coefficients, to the direct-form filter coefficients (a m(Jt)), and vice versa.
Conversion of lattice coefficients to direct-form filter coefficients. The direct-form F IR filter coefficients {am(Jk)} can b e obtained from the lattice coeffi cien ts {AT;} by using the follow ing relations:
A 0 (z) = B 0(z) = 1 A m(z) = A m- \ ( z ) + K mz ~ xBm~\(z) Bm(z) = Z- mA m(z~ l ) m = (7.2.47)
1,2- 1
-M -
(7.2.48) (7.2.49)
m = 1 ,2 .........M -
T h e solu tion is o b tain ed recursively, beginning with m = 1 . Thus w e obtain a seq u en ce o f (M 1) F IR filters, o n e for each valu e o f m. T h e procedure is best illustrated by m eans o f an exam ple.
Example 12J2
Given a three-stage lattice filter with coefficients A T ] = j, K 2 = the FIR filter coefficients for the direct-form structure. we have Ai(z) = A<z) + tfiZ ^ B oW = 1 + K n - 1 = 1 + Jz "1 Hence the coefficients of an FIR filter corresponding to the single-stage lattice are ori(0) = 1 , ari(l) = Ki = j. Since Bm(z) is the reverse polynomial of A(z), we have
5
, Ky =
, determine
Solution We solve the problem recursively, beginning with (7.2.48) for m = 1. Thus
Sec. 7.2
Structures for FIR Systems Next we add the second stage to the lattice. For m = 2 , (7.2.48) yields AjU) = ^ i(i) + KiZ~* B\(z)
1 4- J - -1 4- 1 ,- 2 1 + Rc T 2'-
517
Hence the FIR filter param eters corresponding to the two-stage lattice are *2(0) = 1, 2(1) = | , *2(2) = Also,
*2(z) = J + i = +z-2
Finally, the addition of the third stage to the lattice results in the polynomial M( z ) = * z (;)+ * :3 z _ ,ft(z ) = 1+ z _1 + %z~2 + | z -3
Consequently, the desired direct-form FIR filter is characterized by the coefficients *3(0) = 1 a 3(l) = g of3(2) = 2 o3{3) = i
A s this exam ple illustrates, the lattice structure with param eters K \ , K i ......... K m , corresp on d s to a class o f m direct-form F IR filters with system functions A\ (z), y4i ( z) , . . . , A,(z). It is interesting to note that a characterization o f this class o f m FIR filters in direct form requires m( m + l ) / 2 filter coefficients. In contrast, the lattice-form characterization requires on ly the m reflection coefficients {A',}. T he reason that the lattice provides a m ore com pact representation for the class o f m F IR filters is sim ply du e to the fact that the addition o f stages to the lattice d o es n ot alter the param eters o f the previous stages. O n the other hand, the addition o f the mth stage to a lattice with ( m - 1 ) stages results in a F IR filter with system function A m(z) that has coefficients totally d ifferent from the coefficients o f the low er-order F IR filter with system function Am_ i(z ). A form ula for determ ining the filter coefficients [am(Jt)} recursively can be easily d erived from p olyn om ial relationships in (7.2.47) through (7.2.49). From the relationship in (7.2.48) w e have Am(z) = Am_ i(z ) + K mz ~ l B m- i ( z ) m _i m-t ^ a m ( k ) z ~ k = ^ a m-] (k)z~ k + K m ^ a m- \ ( m - 1 - k ) z (i+1)
k=0 k=0
(7.2.50)
*=0
B y eq u atin g the coefficients o f equal pow ers o f z - 1 and recalling that a m(0) = 1 for m = 1, 2 ......... M - 1, w e obtain the desired recursive eq u ation for the F IR filter coefficients in the form *m(0) = 1 a m(m) = K m a m(k) = ctm- \ ( k ) + - *) (7.2.51) (7.2.52)
= *m i (k) + a m(rn)am-i(rn - k)
(7-2. 53)
518
Chap. 7
W e n o te that (7.2.51) through (7.2.53) are sim ply the L ev in so n -D u rb in recursive equations given in Chapter 11.
Conversion of direct-form FIR filter coefficients to lattice coefficients.

S u p pose that w e are given the F IR coefficien ts for the direct-form realization or, equivalently, the polyn om ial A m(z), and w e w ish to determ in e the corresponding lattice filter param eters { ,} . For the m -stage lattice w e im m ed iately obtain the param eter K m = a m (m). T o obtain K m- \ w e n eed the p olyn om ials Am_ i(z ) since, in general, K m is ob tain ed from the p olyn om ial A m (z) for m = M 1, Af 2 , . . . , 1. C onsequently, w e need to com p ute the p olyn om ials A m(z) starting from m = Af 1 and stepping d o w n su ccessively t o m = 1 . T h e d esired recursive relation for the polyn om ials is easily determ ined from (7.2.44) and (7.2.45). W e have A m(z) = A m- \ ( z ) + K mz ~ x B m- \ { z ) = A m- \ ( z ) + K m[Bm(z) - ^ mAm_ i( ;) ] If w e solve for Am_i (z), w e obtain A m- i ( z ) = -*"*"<*>
m = M - h M - 2 .........1
(7.2.54)
which is just the step-dow n recursion used in the S ch u r-C ohn stability test d e scribed in S ection 3.6.7. T hus w e com p ute all low er-d egree p olyn om ials A m(z) beginning with A M- \ ( z ) and obtain the desired lattice coefficien ts from the rela tion K m = a m(m). W e ob serve that th e procedure works as lon g as \ Km \ 1 for m = 1, 2 .........A f - 1 . Example 7.23 Determ ine the lattice coefficients corresponding to the FIR filter with system function H(z) = A3(z) = 1 -ISolution First we note that # 3 (z) = -I- f z 2 -I- z-J
= a 3(3) = | . Furtherm ore,

5
+ jU- 1 +
z~2 + z~y
The step-down relationship in (7.2.54) with m = 3 yields A2(z) = A3(z) - K i B ^ z ) 1- K
1 ,-2 = l + fz -^ z
Hence K2 a 2(2) = 5 and B i (z) = recursion in (7.2.51), we obtain , , ,
+ | z _1 + z_1- By repeating the step-down
A i b ) ~ K2B2(z )
Sec. 7.3
Structures for HR Systems
519
From the step -d ow n recursive eq u ation in (7.2.54), it is relatively easy to obtain a form ula for recursively com puting K m, begin n ing with m = M 1 and stepping dow n to m = 1. For m = M 1, M 2 , . . . , 1 w e h ave K m = a m( m ) Q fm l (0) = 1 (7.2.55)
a m(k) - K mfim(k) J __ g 2 a m(k) otm ( r n ) a m ( m
- k)
1 -al(m )
1 5 k < m 1
(7.2.56)
w hich is again the recursion w e introduced in the S ch u r-C oh n stability test. A s indicated ab ove, the recursive eq u ation in (7.2.56) breaks dow n if any lattice param eters | Armj = 1. If this occurs, it is ind icative o f the fact that the polyn om ial A m- \ ( z ) has a root on the unit circle. Such a root can b e factored out from Am_](z) and the iterative p rocess in (7.2.56) is carried out for th e reducedorder system .
7.3 STRUCTURES FOR IIR SYSTEMS

In this section w e consider different IIR system s structures d escribed by the dif ference eq u ation in (7.1.1) or, eq u ivalen tly, by the system function in (7.1.2). Just as in the case o f F IR system s, there are several types o f structures or realizations, including direct-form structures, cascade-form structures, lattice structures, and lattice-ladder structures. In addition, IIR system s lend th em selv es to a parallelform realization. W e begin by describing tw o direct-form realizations.
7.3.1 Direct-Form Structures

T h e rational system function as given by (7.1.2) that ch aracterizes an IIR system can be v iew ed as tw o system s in cascade, that is,
H ( z ) = H i (z ) H 2( z )
(7.3.1)
w h ere Hi (z) con sists o f th e zeros o f H ( z ) , and H z (z ) con sists o f the p o les o f H {z), M H l (z ) = Y ! f bkZ~k *=o and
H 2 ( z ) = ------- -- ---------
(7-3.2)
(7.3.3)
1+ X > z"* *=i In S ection 2.5.1 w e describe tw o different direct-form realizations, character ized by w h eth er H \{z) p reced es H j i z ) , or vice versa. Since H \ ( z ) is an F IR system , its direct-form realization w as illustrated in Fig. 7.1. B y attaching th e all-p ole
520
Chap. 7
All-zero system
All-pole system
Figure 7.12 Direct form I realization. system in cascade with H\ ( z) , w e obtain the direct form I realization d epicted in Fig. 7.12. T his realization requires M + N + 1 m ultiplications, M + N additions, and M + N + 1 m em ory locations. If th e all-p ole filter f y i z ) is placed b efore th e all-zero filter H i(z), a m ore com pact structure is ob tain ed as illustrated in S ection 2.5.1. R ecall that the differ en ce eq u ation for th e all-pole filter is
N
w(n) = - ^
ai(W{n - k) + jr(n)
(7.3.4)
Since w ( n ) is the input to the all-zero system , its ou tp u t is M yW = *=o W e n o te that both (7.3.4) and (7.3.5) in volve d elayed versions o f the sequ en ce (u>(n)}- C on sequ en tly, on ly a single d elay line or a sin gle set o f m em ory locations is required for storing the past values o f {u>(n)}. T h e resulting structure that im plem en ts (7.3.4) and (7.3.5) is called a direct form II realization and is depicted in Fig. 7.13. T his structure requires M + N + 1 m ultiplications, M + N additions, bk w (n - k) (7.3.5)
Sec. 7.3
Structures for IIR Systems
521
Figure 7.13
D irect form II realization
(N
Af).
and the m axim um o f {M, N } m em ory locations. Since the direct form II realization m inim izes the num ber o f m em ory locations, it is said to b e canonic. H ow ever, w e should indicate that other IIR structures also p ossess this property, so that this term in ology is perhaps unjustified. T h e structures in Figs. 7.12 and 7.13 are both called direct form realiza tions b ecau se th ey are ob tain ed directly from the system function H ( z ) w ithout any rearrangem ent o f H ( z ) . U n fortu n ately, both structures are extrem ely sen si tive to param eter quantization, in gen eral, and are not recom m en d ed in practical applications. T his to p ic is discussed in detail in Section 7.6, w h ere w e dem onstrate that w h en N is large, a sm all change in a filter coefficient d u e to param eter quan tization, results in a large change in the location o f th e p o les and zeros o f the system .
7.3.2 Signal Flow Graphs and Transposed Structures

A signal flow graph provides an alternative,N but eq u ivalen t, graphical rep resen tation to a b lock diagram structure that w e have b een using to illustrate various system realization s. T h e basic elem en ts o f a flow graph are branches and n od es. A signal flow graph is basically a set o f directed branches that con n ect at n od es. B y d efinition , th e signal ou t o f a branch is equal to th e branch gain (system func tio n ) tim es the signal in to the branch. Furtherm ore, the signal at a n od e o f a flow graph is eq u al to the sum o f the signals from all branches con n ectin g to the node. T o illustrate th e se basic n otion s, let us con sid er the tw o -p o le and tw o-zero IIR system d ep icted in b lock diagram form in Fig. 7.14a. T h e system block
522
Chap. 7
(a)
Source node Sink node
5
(b)
Fignre 7.14 Second-order filter structure (a) and its signal flow graph (b). diagram can be converted to the signal flow graph show n in F ig. 7.14b. W e note that the flow graph contains five n od es lab eled 1 through 5. T w o o f the nodes ( 1 ,3 ) are sum m ing n od es (i.e., they contain ad d ers), w hile the other three nodes represent branching points. Branch transm ittances are ind icated for the branches in the flow graph. N o te that a d elay is indicated by the branch transm ittance z- 1 . W h en the branch transm ittance is unity, it is left u n lab eled . T h e input to the system originates at a so ur ce n o d e and the ou tp u t signal is extracted at a sink node. W e observe that the signal flow graph con tains th e sam e b asic inform ation as the block diagram realization o f the system . T h e on ly ap p aren t d ifference is that b o th branch p oints and adders in the b lock diagram are rep resen ted by nodes in th e signal flow graph. T h e subject o f linear signal flow graphs is an im portant o n e in the treatm ent o f netw ork s and m any interesting results are available. O n e b asic n otion involves the transform ation o f o n e flow graph in to a n oth er w ithout changing the basic in p u t-ou tp u t relationship. Specifically, o n e tech n iq u e that is usefu l in deriving new system structures for F IR and IIR system s stem s from th e transposition or flo w - g ra p h reversal theorem. T his th eorem sim ply states that if w e reverse the
Sec. 7.3
523
directions o f all brancb transm ittances and interchange the input and output in the flow graph, the system function rem ains unchanged. T h e resulting structure is called a transposed s tructure or a tran sp osed f o r m . F or exam p le, the transposition o f the signal flow graph in Fig. 7.14b is illus trated in Fig. 7.15a. T h e corresponding b lock diagram realization o f the transposed form is d ep icted in Fig. 7.15b. It is interesting to n ote that the transposition o f the original flow graph resulted in branching n od es b ecom in g ad d er n od es, and vice versa. In Section 7.5 w e p rovid e a p ro o f o f the transposition theorem by using state-sp ace techniques. L et us apply the transposition theorem to the direct form II structure. First, w e reverse all the signal flow d irections in Fig. 7.13. Second, w e change n od es into adders and adders into n od es, and Anally, w e interchange the input and the output. T h ese op eration s result in the transposed direct form II structure show n in Fig. 7.16. T his structure can b e redrawn as in Fig. 7.17, which show s the input on the left and the output on the right.
~ ai
s^\
(b)
Figure 7.15 Signal flow graph of transposed structure (a) and its realization (b).
524
Chap. 7
Figure 7.16 Transposed direct form II structure. T h e transposed direct form II realization that w e have ob tain ed can be d e scribed by the set o f d ifference equations
y ( n ) = wr f n - l ) + b o x ( n )
(7.3.6)
Wt(n) = wt+i(n - 1) w N(n ) = bNx(n) - a Ny(n)
+ b kx(n)
Jt = 1 ,2 ......... N - 1
(7.3.7) (7.3.8)
W ithout loss o f generality, w e have assum ed that Af = W in w riting equ ation s. It is also clear from observation o f Fig. 7.17 that this set o f d ifferen ce eq u ation s is equ ivalen t to the sin gle differen ce equation
y ( n ) = ~ ^ 2 a ky (n - k ) + ^ b kx ( n - k )
(7.3.9)
Sec. 7.3
525
Figure 7.17 Transposed direct form II structure.
Finally, w e o b serv e that th e transposed direct form II structure requires the sam e n um ber o f m ultip lication s, additions, and m em ory locations as the original direct form II structure. A lth o u g h our discussion o f transposed structures has b een concerned with the general form o f an IIR system , it is interesting to n o te that an F IR system , ob tain ed from (7.3.9) by setting the a* = 0, k = 1, 2 , . . . , N , also has a transposed direct form as illustrated in Fig. 7.18. T his structure is sim ply ob tain ed from Fig. 7.17 by settin g a* = 0, k = 1, 2 , . . . , N . T his transposed form realization m ay
Figure 7.18
Transposed FIR structure.
526
Chap. 7
be described by the set o f differen ce eq u ation s w M(n) = b t f x ( n ) u>*(/i) = u>*+i(n 1) + bkx(n) y ( n ) = w i ( n - 1) +/>&*() k = M 1, M ~ 2, . . . , 1 (7.3.10) (7.3.11) (7.3.12)
In sum m ary, T ab le 7.1 illustrates the direct-form structures and the corresponding d ifference eq u ation s for a basic tw o-p ole and tw o-zero IIR system with system function x b0 + b i z ~ l + b 2z ~ 2 H ( z ) = --- -------- ------------ 3 -
1 + fli z 1 +azz 1
(7.3.13)
This is th e basic building block in the cascade realization o f h igh-order IIR system s, as described in the follow in g section. O f the three direct-form structures given in T ab le 7.1, th e direct form II structures are p referable due to the sm aller number o f m em ory locations required in their im plem entation. Finally, w e n o te that in the z-dom ain, the set o f d ifferen ce eq u ation s describ ing a linear signal flow graph con stitute a linear set o f equations. A n y rearrange m ent o f such a set o f eq u ation s is equ ivalen t to a rearrangem ent o f the signal flow graph to obtain a n ew structure, and vice versa.
7.3.3 Cascade-Form Structures

Let us con sid er a high-order IIR system with system function given by (7.1.2). W ithout loss o f generality w e assum e that N > M . T h e system can be factored into a cascade o f secon d-ord er subsystem s, such that H (z) can b e exp ressed as
K
*=i w here K is the in teger part o f (N + l ) /2 . Hk (z) has th e general form rr , , bto + b u z ' 1 + bk2Z ' 2 Hk (z) = --------------------- 3 5 1 + fljtiz 1 + a k2z 2 (7.3.15)
A s in the case o f F IR system s b ased on a cascad e-form realization, the param eter bo can be distributed equally am ong the K filter se ctio n s so that bo = > 1 0 ^ 2 0 bxoT he coefficients {a*j} and {*>*, }in the secon d -ord er su b system s are real. This im plies that in form ing th e secon d-ord er su b system s or quadratic factors in (7.3.15), w e should group to g eth er a pair o f com p lex-con ju gate p o les and w e should group togeth er a pair o f com p lex-con ju gate zeros. H ow ever, th e pairing o f tw o com plexconjugate p o les with a pair o f com p lex-con ju gate zeros or real-valu ed zeros to form a subsystem o f the type given by (7.3.15), can b e d o n e arbitrarily. Furtherm ore, any tw o real-valued zeros can b e paired to g eth er to form a quadratic factor and, likew ise, any tw o real-valued p o les can b e paired togeth er to form a quadratic factor. C on sequ en tly, th e quadratic factor in th e num erator o f (7.3.15) m ay consist
TABLE 7.1
SOME SECOND-ORDER MODULES FOR DISCRETE-TIME SYSTEMS Structure Implementation Equations System Function
xin)
v(n) = fc().t(n) +fc]X(n - I) -I- i>2.t(rt - 2) - oi v(n - 1) <t2\(n 2) x<n)
H( z) =
bft + bjZ 1 +biz 2 1 + a u -1 +aiz~2
u.( n )
~ d i w ( n I ) a 2 w ( n 2)
+ x(n)
H( Z) =
,v(n) = bow(n) -+ - fcj utr( 1) + b2U)(n 2) x<n)
bp + bjZ 1 +bjz 1 + a u _1 +aiz~2
v(n) = bu.x(n) + wt (n - 1) = ht x( n) ai_v(n)

-t- w 2 ( n 1)
H( Z)
bp -t-fri? 1 +bjz 2 1 + c u - * +aiz~2
w 2 (n)
b ix(n) -
a 2y ( n )
527
528
Chap. 7
o f either a pair o f real roots or a pair o f com p lex-con ju gate roots. T h e sam e statem ent applies to the den om in ator o f (7.3.15). If N > M , so m e o f the secon d-ord er subsystem s h ave num erator coefficients that are zero, that is, eith er bk2 = 0 or bk\ = 0 or b oth bk2 = bk\ = 0 for som e k. Fur therm ore, if N is odd, o n e o f the subsystem s, say Hk (z), m ust h ave a k2 = 0, so that the subsystem is o f first order. T o p reserve the m odularity in th e im plem en tation o f H ( z ) , it is o ften preferable to use the basic secon d-ord er su b system s in the cas cade structure and h ave som e zero-valu ed coefficients in so m e o f the subsystem s. E ach o f the secon d-ord er subsystem s with system function o f the form (7.3.15) can be realized in eith er direct form I, or direct form II, or tran sp osed direct form II. Since there are m any w ays to pair the p oles and zeros o f H (z ) into a cascade o f secon d-ord er section s, and several w ays to order the resulting subsystem s, it is p ossib le to obtain a variety o f cascade realizations. A lth ou gh all cascade realiza tions are equ ivalen t for infinite precision arithm etic, the various realizations may differ significantly w hen im plem en ted with finite-precision arithm etic. T he general form o f th e cascade structure is illustrated in Fig. 7.19. If we use the direct form II structure for each of the subsystem s, the com putational algorithm for realizing the IIR system with system function H ( z ) is described by the follow in g set o f equations. >>o(n) = x ( n ) w k(n) = - a ki w k(n - 1) - a k2w k (n - 2) + y*_i(n) y k(n) = bk0w k(n) + bt i w k(n - 1) -I- bk2w t (n - 2) y (n ) = ytcin) k = 1 ,2 ........... K k = 1 ,2 .......... K (7.3.16) (7.3.17) (7.3.18) (7.3.19)
(a)
**0 S ~ \ > '* ( * ) = X t + 1 ( n )
-*l
0 -
^ -0
(b)
Ftg*rc 7.19 Cascade structure of second-order systems and a realization of each second-order section.
Sec. 7.3
529
Thus this set o f eq u ation s p rovides a com p lete description o f the cascade structure based on direct form II section s.
7.3.4 Parallel-Form Structures

A parallet-form realization o f an IIR system can be ob tain ed by p erform ing a partial-fraction expansion o f H( z ) . W ithout loss o f generality, w e again assum e that N > M and that the p o le s are distinct. T hen, by perform ing a partial-fraction expansion o f H( z ) , w e obtain the result N Ai H ( z ) = C + V ------- r t t 1 - P*z~' (7.3.20)
w here {p*} are the p oles, {A t \are the coefficients (residues) in the partial-fraction exp an sion , and the constant C is defined as C = b s / a s i . T h e structure im plied by (7.3.20) is sh ow n in Fig. 7.20. It consists o f a parallel bank o f sin gle-p ole filters. In gen eral, som e o f the p oles o f H ( z ) may be com plex valued. In such a case, the corresp on d ing coefficients A t are also com plex valued. T o avoid m ultiplica tions by com p lex num bers, w e can com b ine pairs o f com p lex-con ju gate p oles to form tw o -p o le subsystem s. In addition, w e can com b ine, in an arbitrary m anner,
Figure 7.20
Parallel structure of IIR system.
530
Chap. 7
Figure 7*21
Structure of second-order section in a parallel IIR system realization.
pairs o f real-valued p o les to form tw o-p ole subsystem s. Each o f th ese subsystem s has the form Hk{z) =
1
+ a kiz ] + a k2r
(7.3.21)
w here the coefficients {bk,} and (at;] are real-valued system param eters. T he over all function can n ow be expressed as
H( z) = C + J 2 h* < * >
(7.3.22)
w here K is the in teger part o f ( N + 1)/2. W hen N is odd, on e o f the H k (z) is really a sin gle-p ole system (i.e., bk\ = a k 2 = 0 ). T h e individual second-order section s which are th e basic b uilding blocks for H ( z ) can be im plem en ted in eith er o f the direct form s or in a transposed direct form. T h e direct form II structure is illustrated in Fig. 7.21. W ith this structure as a basic building block, the parallel-form realization o f the F IR system is described by the follow in g set o f equations
wk(n) = aki w k(n - 1) - ak2wk(n - 2) + x(n)
* = 1,2,..., k = 1 ,2 ......... K
(7.3.23) (7.3.24) (7.3.25)
yt(n) = bk0wk(n ) + bktwt (n - 1)

K y(n) = Cx ( n ) + ^ y t ( n ) Example 7.3.1
Determine the cascade and parallel realizations for the system described by the system function
10(1 -
H(z) =
1)(1 l z - 1 ) ( l +
z-1)
(1 - |z _1)(l - 5Z-1)[1 - (j + ; j ) z -1][l - (5 -
Sec. 7.3
531
Solution The cascade realization is easily obtained from this form. One possible pairing of poles and zeros is 1 - lz~* = i _ 2 .- i + 2, z- 2
1 8' T 32
l + l z - '- z " 2 H2< c) = i ; 1 - z ~ l + \rz~ ~2 and hence H(z) = 10//,(z)Jfe(z) The cascade realization is depicted in Fig. 7.22a. To obtain the parallel-form realization, H(z) must be expanded in partial frac tions. Thus we have Ai Ai " (2 ) = +
1-fz-1 1-iz-1 i - ( i + yi)z-i i - ( l - y i );-i
where j4t , A 2, A 3, and A j are to be determ ined. A fter some arithmetic we find that A { = 2.93, A, = -17.68, A3 = 12.25 - yl4.57, A\ = 12.25 + >14.57
upon recom bining pairs of poles, we obtain % 14.75 12.90z_l 24.50 + 26.82;"' 1 ;----5 i I- 1 _ -- I -L J.--2 H (z) ------1 _ Z--1 _1 _ 2 .7 -1 x*- + 324 * 2^ The parallel-form realization is illustrated in Fig. 7.22b.
7.3.5 Lattice and Lattice-Ladder Structures for IIR Systems

In Section 7.2.4 w e d ev elop ed a lattice filter structure that is eq u ivalen t to an F IR system . In this sectio n w e exten d the d evelop m en t to IIR system s. L et us b egin with an all-p ole system with system function H (z) = ------- ^ ----------- = A t a n (z ) 1+ ^cis(k)z k=l (7. 3. 26)
T h e direct form realization o f this system is illustrated in Fig. 7.23. T h e difference eq u ation for this IIR system is
N
y( n ) = ~ ^ a N (k) y( n - k) + x { n) *=i
(7.3.27)
It is in terestin g to n ote that if w e interchange the roles o f input and output [i.e., interchange x ( n ) w ith y(n ) in (7.3.27)], w e obtain
x(n) = - ^ T a N(k)x(n - k) 4- y(ri)
532
Chap. 7
(a)
(b) Figure 7.22 Cascade and parallel realizations for the system in Example 7.3.1. or, equivalently,
N
y ( n ) = x ( n ) 4- ^ a w(fc)*(n - k) *=i
(7.3.28)
W e n o te that the eq u ation in (7.3.28) describes an F IR system having the system function H ( z ) = A N (z), w h ile the system describ ed by th e differen ce equa tion in (7.3.27) rep resen ts an IIR system with system function H ( z ) = lM w(z)*
Sec. 7.3
533
*()
yi.n)
Figure 7.23 Direct-form realization of an all-pole system. O n e system can b e o b tain ed from the other sim ply by interchanging th e roles o f th e input and output. B ased on this o b servation , w e shall use the all-zero (F IR ) lattice describ ed in S ection 7.2.4 to ob tain a lattice structure for an all-p ole IIR system by interchanging the ro les o f the input and output. First, w e take the all-zero lattice filter illustrated in Fig. 7.11 and th en redefine the input as x ( n ) = f N(n) and the ou tp u t as v(n) = fo(n) (7.3.30) (7.3.29)
T h ese are exactly the o p p osite o f the definitions for the all-zero lattice filter. T h ese d efinitions d ictate that the qu an tities { / m(n)l be com p uted in d escen din g ord er [i.e., / v ( n ) , / v _ i ( n ) , . . . ) . T h is com p utation can be accom plished by rearranging the recursive eq u ation in (7.2.29) and thus solvin g for / m_\ (n) in term s o f f m(n ), that is,
/ m - i ( n ) = f m( n) -
K mgm- i ( n - 1 )
m =
N, N - 1 ,..., 1
T h e eq u a tio n (7.2.30) for gm(n) rem ains unchanged. T h e result o f th ese ch an ges is th e set o f equations f s ( n ) = x (n) f m - \ ( n ) = f m(n) - K mgm. ] ( n - 1) gm(n) = K mf m- i ( n ) + g m- \ ( n - 1) y ( n ) = /o (n ) = go(n) w hich corresp on d to the structure sh ow n in Fig. 7.24. Input m = N , N - 1,. m = N , N - 1, (7.3.31) (7.3.32) (7.3.33) (7.3.34)
Figure 7.24
Lattice structure for an all-pole
IIR system.
534
Chap. 7
T o d em onstrate that the set o f eq u ation s (7.3.31) through (7.3.34) represent an all-pole IIR system , let us con sid er th e case w h ere N = 1. T h e eq u ation s reduce to * (n ) = /i( n )
fain) = h i n ) - K\goin - 1)
g i(n ) = ATi/o(n) + go(n - 1) (7.3.35)
?() = foin)
= xin ) - K \yin - 1 ) Furtherm ore, the eq u ation for g i(/i) can be exp ressed as i(n ) = ATi^n) + y in - 1) (7.3.36)
W e observe that (7.3.35) represents a first-order all-p ole IIR system w hile (7.3.36) represents a first-order F IR system . T h e p o le is a result o f the feed b ack introduced by the solution o f th e ( / m(n)} in descen din g order. T his feed b ack is dep icted in Fig. 7.25a.
Forward
(a)
Forward
Reverse
(b)
Figure 125
Single-pole and two-pole lattice system.
Sec. 7.3
535
N ex t, let us con sid er the case N = 2, which corresponds to the structure in Fig. 7.25b. T h e eq u ation s corresponding to this structure are f 2(n) = x (n) f i ( n ) = f 2(n) - K 2g i(n - 1) g2(n) = K 2f i ( n ) + g i(n 1
) (7.3.37)
/o (n ) = / i ( n ) - K igo in - 1 ) gi (n) = K ifo (n ) + g o ( n - l ) y (n ) = /o (n ) = go() A fter so m e sim ple substitutions and m anipulations w e obtain y (n ) = + K 2)y (n 1) - K 2y (n - 2) + x ( n ) (7.3.38) (7.3.39)
g2(n) = K 2v(n ) + K XQ + K 2)y (n - 1) + v(n - 2)
C learly, the differen ce equation in (7.3.38) represents a tw o-p ole IIR system , and the relation in (7.3.39) is the in p u t-ou tp u t equation for a tw o-zero F IR system . N o te that the coefficients for the F IR system are identical to th ose in the IIR system excep t that they occu r in reverse order. In general, th ese con clu sion s hold for any N . Indeed, with the definition of Am( i) given in (7.2.32). the system function for the all-p ole IIR system is , , Y (z) Fo(z) 1 Ha (z) = -------= ----------= ---------X (z ) Fm(z) Am( z) Sim ilarly, the system function o f the all-zero (F IR ) system is H b(z) = = B m(z) = z~ mA m( z ~ l ) (7.3.41) 7 , ,, (7.3.40)
Y (z )
G 0(Z)
w h ere w e used the p reviously established relationships in (7.2.36) through (7.2.42). Thus the coefficients in the F IR system H b(z) are identical to the coefficients in Am(z), excep t that they occur in reverse order. It is interesting to n o te that the all-p ole lattice structure has an all-zero path with input go(n) and output g x ( n ), w hich is identical to its counterpart all-zero path in th e all-zero lattice structure. T h e polynom ial Bm (z), w hich represents the system function o f the all-zero path com m on to b oth lattice structures, is usually called the b a c k w a rd sy stem fu n c tio n , b ecau se it provid es a backward path in the a ll-p ole lattice structure. From this discussion th e reader should ob serve that the all-zero and all-pole lattice structures are characterized by th e sam e set o f lattice param eters, nam ely, K \, K 2, . . K n . T h e tw o lattice structures differ on ly in the in tercon n ection s o f their signal flow graphs. C on sequ en tly, the algorithm s for con vertin g b etw een the system param eters (a m(fc)} in th e direct form realization o f an F IR system , and the p aram eters o f its lattice counterpart apply as w ell to the all-p ole structure.
536
implementation of Discrete-Time Systems
Chap. 7
W e recall that the roots o f the polynom ial A N (z) lie in sid e th e unit circle if and o n ly if the lattice param eters | ^ m| < 1 for all m = 1, 2 .........N . T h erefore, the all-p o le lattice structure is a stable system if and on ly if its param eters \K m\ < 1 for all m. In practical applications the all-pole lattice structure has b een used to m odel the hum an vocal tract and a stratified earth. In such cases the lattice param eters, {K m) have the physical significance o f b ein g identical to reflection coefficients in the physical m edium . T his is the reason that the lattice param eters are o ften called reflection coefficients. In such applications, a stable m od el o f the m edium requires that the reflection coefficients, obtained by perform ing m easu rem en ts o n output signals from the m edium , b e less than unity. T h e all-p ole lattice provides the basic b uilding block for lattice-type structures that im plem en t IIR system s that contain both p oles and zeros. T o d evelop the appropriate structure, let us consider an IIR system with system function M
H (z ) = -------------------------=
A , 1 + J 2 a N(k)z ~ L k= 1
* n (z )
(1 3 A 2 )
w here the notation for the num erator polynom ial has b een changed to avoid con fusion with our previous develop m en t. W ithout loss o f generality, w e assum e that N > M. In the direct form II structure, the system in (7.3.42) is described by the difference equations
N
w (n ) = ^ a tf(fc )u > (n k) + x (n ) *=i M y (n ) = c M(fc)w(n - k) *=o
(7.3.43)
(7.3.44)
N o te that (7.3.43) is the in p ut-ou tp u t o f an all-p ole IIR system and that (7.3.44) is the in p u t-o u tp u t o f an all-zero system . Furtherm ore, w e ob serve that the output of the all-zero system is sim ply a linear com bination o f delayed ou tp u ts from the all p o le system . This is easily seen by observing th e direct form II structure redrawn as in Fig. 7.26. Since zeros result from form ing a linear com bination o f previous outputs we can carry o ver this observation to construct a p o le -z e r o IIR system using the all p ole lattice structure as the basic building block. W e have already observed that gm(n ) is a linear com bination o f present and past outputs. In fact, the system Hh(z) = Y (z) = Bm(z)
Sec. 7.3
537
^ - 0 - ------0 -------0
w (n )
0>
w(n - Af + 1) -----t - M ------------- ------
w(n - 1 )
w(n - 2)
w ( n - M)
e 0 )
c*<2 )
cu ( M -
1)
c^M) yirt)
- 0 - -------0 ------- 0 ^
Figure 7.26 D irect form II realization of IIR system.
is an all-zero system . T h erefore, any linear com bination o f {gm(n)} is also an all-zero system . Thus w e begin with an all-p ole lattice structure with param eters K m, 1 < m < N , and w e add a la dder part by taking as the output a w eigh ted linear com bination o f ( * (n )}. T h e result is a p o le-zero IIR system w hich has the latticela d d e r structure show n in Fig. 7.27 for M = N . Its output is M (7.3.45) m srO w here (um) are the param eters that determ ine the zeros o f the system . T h e system
Figure 7.27
Lattice-ladder structure for the realization o f a p ole-zero system.
538
Chap. 7
function corresponding to (7.3.45) is

H(z) = Y (z ) X (z) (7.3.46) G m(z) X{z)
-Z>
rr/ ^ H(z) =
Since X ( z ) = F N(z) and fb (z) = G q( z ), (7.3.46) can be written as G m{z) Fq( z ) > u m 7 ^ G 0 (z) FN (z) Bm{z)
a n (z )
(7.3.47)
M y , vmBm{z) A \(z) If w e com pare (7.3.41) with (7.3.47), w e con clu d e that M C M(z) = ' / vmBm(z) m=0 (7.3.48)
T his is the desired relationship that can be used to determ in e the w eighting coef ficients {um). Thus, w e have dem onstrated that the coefficients o f the num erator polyn om ial C*f(z) determ in e the ladder param eters {um}, w h ereas the coefficients in the d en om in ator polyn om ial A ^(z) determ ine the lattice p aram eters {Km\. G iven the polyn om ials C (z ) and A n ( z ), w here N > M , th e param eters of the all-pole lattice are d eterm ined first, as described p reviou sly, by the conver sion algorithm given in Section 7.2.4, which con verts the direct form coefficients in to lattice param eters. B y m ean s o f the step -d ow n recursive relations given by (7.2.54), w e obtain the lattice param eters {Km} and th e p olyn om ials Bm(z), m = 1, 2 ,..., N. T h e ladder p aram eters are d eterm in ed from (7.3.48), w hich can be expressed as T O 1 C m( z ) = vk B k(z) + vmB m(z) (7.3.49) or, equivalently, as C m(z) = Cm_ ,(z ) + vmB m( z ) (7.3.50)
Thus Cm(z) can b e com p uted recursively from the reverse p olyn om ials Bm(z), m = 1 , 2 , . . . , M . Since 1 for all m, th e param eters vm, m = 0, 1 , . . . , M can be determ ined by first noting that vm = cm(m) m = 0,1,..., M
(7.3.51)
Sec. 7.4
State-Space System Analysis and Structures
539
T h en , by rew riting (7.3.50) as Cm- \ (z) = Cm (z ) - iv, Bm(z) (7.3.52)
and running this recursive relation backward in m (i.e., m = M , M 1 , . , , , 2 ), w e o b tain cm(m ) and therefore the ladder param eters according to (7.3.51). T h e lattice-ladder filter structures that w e h ave p resen ted require the m in im um am ount o f m em ory but not the m inim um num ber o f m ultiplications. A l thou gh lattice structures w ith only on e m ultiplier per lattice stage exist, the tw o m ultiplier-per-stage lattice that w e have described, is by far th e m ost w idely used in practical applications. In con clu sion , th e m odularity, the built-in stability charac teristics em b o d ied in the coefficients {A Tm}, and its robustness to finite-w ord-length effe cts m ake th e lattice structure very attractive in m any practical applications, including sp eech processing system s, adaptive filtering, and geophysical signal pro cessing.
7.4 STATE-SPACE SYSTEM ANALYSIS AND STRUCTURES

U p to this point our treatm ent o f linear tim e-invariant system s has b een lim ited to an in p u t-o u tp u t or external description of the characteristics o f the system . In other w ords, the system was characterized by m athem atical eq u ation s that relate the input signal to the output signal. In this section w e introduce the basic concepts in the state-sp ace description o f linear tim e-invariant causal system s. A lth ou gh the state-space or in tern a l description of the system still in volves a relationship betw een the input and output signals, it also in volves an additional set o f variables, called state variables. Furtherm ore, the m athem atical eq u ation s describing the system , its input, and its output are usually divided into tw o parts: 1. A set o f m athem atical eq u ation s relating th e state variables to the input signal. 2. A seco n d set o f m athem atical eq u ation s relating the state variables and the current input to the output signal. T h e state variables provide inform ation ab ou t all th e internal signals in the system . A s a result, the state-sp ace description provides a m ore d etailed descrip tion o f the system than the in p u t-ou tp u t description. A lth o u g h our treatm ent o f state-sp ace analysis is confined prim arily to sin gle in p u t-sin g le output linear tim einvariant causal system s, the state-sp ace tech n iqu es can also b e applied to non linear system s, tim e-variant system s, and m ultip le in p u t-m u ltip le ou tp u t system s. In fact, it is in the characterization and analysis o f m ultip le in p u t-m u ltip le output system s that th e p ow er and im portance o f state-sp ace m eth od s are clearly evident. B o th in p u t-o u tp u t and state-variable descriptions o f a system are useful in practice. T h e description w e use d ep en d s on the p rob lem , the available inform a tion, and the q u estion s to b e answ ered. In our p resen tation , the em phasis is on
540
Chap. 7
the use o f state-sp ace tech n iqu es in system analysis, and in the d evelop m en t o f state-sp ace structures for th e realization o f discrete-tim e system s.
7.4.1 S ta te -S p a ce D e sc rip tio n s o f S y s te m s C haracterized by D ifference E qu ations

A s w e have already o b served , the determ ination o f the output o f a system requires that w e know the input signal and the set o f initial con d ition s at th e tim e the input is applied. If a system is n ot relaxed initially, say at tim e no, th en k n ow led ge o f the input signal x ( n ) for n > n Q is not sufficient to un iqu ely d eterm in e the output y(n) for n > no- T h e initial con d ition s o f the system at n = no m ust also b e know n and taken in to account. T his set o f initial con d ition s is called th e state o f the system at n = no- H en ce we defin e the state o f a sy ste m a t tim e no a s the a m o u n t o f in fo rm a tio n that m u s t be p r o v id e d at tim e n0, w hich, together w ith th e in p u t signal x ( n ) f o r n > no, u n iq u e ly d eterm in e the o u tp u t o f th e sy stem f o r a ll n > noFrom this definition w e infer that the con cep t o f state lead s to a d ecom p o sition o f a system in to tw o parts, a part that contains m em ory, and a m em oryless com p onent. T he inform ation stored in the m em ory co m p on en t con stitu tes the set o f initial conditions and is called the state o f the system . T h e current ou tp u t o f the system then b eco m es a function o f the current value o f the input and the current state. T hus, to d eterm in e the output o f the system at a given tim e, w e n eed the current value o f the state and the current input. Since the current value o f the input is available, w e only n eed to provide a m echanism for upd atin g the state o f the system recursively. C on sequ en tly, the state o f the system at tim e no + 1 should d epend on the state o f the system at tim e n 0 and the value o f th e input signal x (n) at n = noT h e follow in g exam p le illustrates the approach in form ulating a state-space description o f a system . L et us con sid er a linear tim e-invariant causal system described by the d ifferen ce equation
3 3
(7.4.1) T h e direct form II realization for the system is show n in Fig. 7.28. A s state variables, w e u se the con ten ts o f the system m em ory registers, cou n t ing them from the b o ttom , as show n in Fig. 7.28. W e recall that th e output o f a delay elem en t rep resen ts the present value stored in th e register and the input represents the next v alu e to b e stored in th e m em ory. C on seq u en tly, with the aid o f Fig. 7.28, w e can w rite t>i(n + 1 ) = V 2 (n) v i (n +
1
) = U 3 ()
(7.4.2)
v$(n + 1 ) = - a 3U i(n ) - 0 2 2 (n ) - o m 3( n ) + x(n)
Sec. 7.4
541
<*>
o - l
v2 ()
6
U,()
" 3 3
Figure 7.28 Direct form II realization of system described by the difference equa
tion in (7.5.1).
It is in terestin g to n o te that the state-variable form ulation for the third-order system o f (7.4.1) in v o lv es three first-order difference eq u ation s given by (7.4.2). In general, an n th-order system can be described by n first-order differen ce equations. T h e ou tp u t eq u ation , w hich exp resses y ( n ) in term s o f the state variables and the presen t input v alu e x (n ), can also be ob tain ed by referring to Fig. 7.28. W e have y (n ) = i>oVi(n + 1 ) + />3 Ui(n) + b2v2 (n) + 6 it>3 (/i) W e can elim in a te v$(n + 1 ) by using the last equation in (7.4.2). Thus w e obtain the desired ou tp u t eq u ation y (n ) = (f >3 - boa2) v i (n ) + ( 2 boa2) v2(n) + (b\ - i>o<ai)u3 (n) + b0x ( n ) If w e put (7.4.2) and (7.4.3) in to m atrix form w e have
V[(n + 1 )0 0 .-0 3
(7.4.3)
1
0
t> ] (n ) -
-0 + 0
V2 (n + 1) = _ v 3 (n + 1 ) . and
-a 2
1 -a \.
V2 (n)
-i> 3 (n )-
x(n)
(7.4.4)
.1 .
y ( n ) = [(b3 - b0a 3) ( ^ - M
) (i>i ~ M l ) ]
~ v i(n ) v2(n ) + box(n)

-V 3 ( * ) -
(7.4.5)
T h e eq u a tio n s (7.4.4) and (7.4.5) provide a com p lete description o f the sys tem . F urtherm ore, th e variables v i(n ), V2 (n), and 113( 0 ), w hich sum m arize all the n ecessary past inform ation, are the state variables o f the system . W e also observe that as in d icated previou sly, eq u ation s (7.4.4) and (7.4.5) split the system in to tw o co m p o n en t parts, a dynam ic (m em ory) subsystem and a static (m em oryless) sub system . W e say that this se t o f eq u ation s provides a state-space description o f the system .
542
Chap. 7
B y generalizing the previous exam p le, it can easily be se en that the A'th-order system described by A f ,( ) = - a ty(n ~ *=i + J t=0 N bkX^n ~ k) (7.4.6)
can be expressed as a linear tim e-invariant state-space realization by the relations
State equation
v(n -j- 1 ) = Fv(n) + q x(n ) (7.4.7)
Output equation
>() = gf v ( n ) + d x ( n ) (7.4.8)
w here the elem en ts o f F, q, g, and d are con stan ts (i.e., they d o n ot change as a function o f the tim e index n ), given by F =
0 0 0 1 0 0 1 0 0 1
0 0
"0 "
0
q=
0
.
(7.4.9)
.-a s
~a -1
-a2
~a\
b N boaN b fj - 1 bo a n-\
g=
b\ boa\ A n y discrete-tim e system w h ose input x (n ), ou tp u t y( n ) , and state v{n), for all n > no, are related by the state-sp ace eq u ation s ab ove, w h ere F, q, g, and d are arbitrary but fixed quantities, will be called linear and tim e invariant. If at least o n e o f the quantities in F, q, g, or d d ep en d s on tim e, the system b ecom es time variant. W e will refer to (7.4.7) through (7.4.8) as the linear tim e-in va rian t state-space m o d el, which can b e represented by the sim ple vector-m atrix block diagram in Fig. 7.29. In this figure the d ou b le lin es represent vector qu an tities and the blocks represent the v ector or m atrix coefficients. Example 7.4.1 Determine the state-space equations for the transposed direct form II structure shown in Fig. 7.30. Solution The validity of this structure can be seen if we rewrite (7.4.1) as y(n) =
- k) - aky(n - * )] + box(n)
Sec. 7.4
543
Figure 7.29 General state-space description of a linear time-invariant system. *()
Figure 7JO State-space realization for the system described by (7.4.1). Due to the linearity and time invariance of the system, instead of first delaying the signals x(n) and y(n) and then computing the terms bkx(n - k) aky(n k) as in Fig. 7.28, we first compute the terms bkx(n) at y(n) and then delay them. If we use the state variables indicated in Fig. 7.30, we obtain 'v i(n + 1 )" Vi (n + 1 ) = . v3(n + 1 ) . 0
1 .0 0 0 1
'M n ) bi boaj " -a 2 = * 2 - M 2 x(n) i> 2 (n) -b\ bod\ . - a i . .t>B(n).

03
(7.4.10)
Di(n) y(n) = [0
0 1 ]
l> 2 (fl)
-f box(n)
(7.4.11)
V3(n) J T h e state-sp ace description specified by (7.4.4) and (7.4.5) is know n as a ty p e 1 state-sp ace realization, w h ereas the o n e described by (7.4.10) and (7.4.11) is ca lled a typ e 2 state-sp ace realization.
7.4 .2 S o lu tio n o f th e S ta te -S p a ce E q u ation s

T h ere are several m eth od s for solvin g the state-sp ace eq u ation s. H ere w e discuss a recursive solu tion w hich m akes u se o f th e fact that the state-sp ace eq u ation s are a se t o f linear first-order differen ce equations.
544
Implementation of Discrete-Time Systems For the jV-dim ensional state-sp ace m odel v( + 1) = F v(/i) + qjr(n) y(n) = g'v(n) + d x(n )
Chap. 7
(7.4.12) (7.4.13)
and given the initial co n d ition v ( 0), we have for n > n0, v(n 0 + 1) = Fv(n0) + q*(n) v(no + 2 ) = F v(n 0 + 1 ) + qjr(n0 + 1 ) = F 2 v(n0) + F q j( o ) + q-^(o + 1) w here F 2 represents the matrix product FF and Fq is the product o f the m atrix F and the vector q. If w e continue as in the on e-dim en sion al case, w e obtain, for n > n0, (7.4.14) The matrix F is defined as the N x N identity matrix, having unity on the main diagonal and zeros elsew h ere. T he m atrix F'- -' is often d en o te d as 4>(/ - j ) , that is, * (/ - j) = F ~ J (7.4.15)
for any p ositive in tegers i > j . T his m atrix is called the state transition matrix of the system . T h e output o f the system is obtained by substituting (7.4.14) in to (7.4.13). T he result o f this substitution is y(n) = g'F" nov(n0) +
*=0
n- 1
g'F" 1 kq x ( k ) + d x { n ) (7.4.16) 2 1 - &)q*(*) + d x ( n )
- floM no) + ^
From this general result, w e can determ ine the output for tw o special cases. First, the zero-input resp onse o f the system is yzM ) = g'F" nv(n0) = g '$ ( n - floM no) O n the other hand, the zero-state response is n 1 yzs(n) = g ' $ { n - 1 - fc)q*(/:) + d x ( n ) (7.4.18) (7.4.17)
Clearly, the A'-dim ensional state-sp ace system is zero-in p u t linear, zero-state linear, and since y ( n ) = y Zi(n) + y zs(n ), it is linear. F urtherm ore, sin ce any system described by a linear con stan t-coefficien t d ifferen ce eq u ation can be put in the state-space form, it is linear, in agreem en t with the results ob tain ed in S ection 2.4.
Sec. 7.4
545
7.4.3 Relationships Between Input-Output and State-Space Descriptions

From our p reviou s discussion w e h ave seen that there is n o unique ch oice for the state variables o f a causal system . Furtherm ore, d ifferent ch o ices for th e state vecto r lead to differen t structures for the realization o f the sam e system . H en ce, in g en eral, the in p u t-o u tp u t relationship d oes not un iqu ely d escrib e the internal structure o f th e system . T o illu strate these assertions, let us consider an N -dim en sion al system with the sta te-sp a ce rep resen tation v(n + 1) = Fv(rt) 4- qjf(n) y(n ) = g'v(n) -1- d x ( n ) L et P b e any N x N matrix w h ose inverse m atrix state v ecto r v(n) as (7.4.19) (7.4.20)
P-1 exists. W e d efine a new

(7.4.21) (7.4.22)
(7.4.23)
(7.4.24) N o w , w e d efine a n ew system param eter matrix
F and the vectors q and g as
F = PFP-1
(7.4.25)
W ith th ese d efinition s, th e state eq u ation s can b e exp ressed in term s o f th e new system qu an tities as
v(rt 4 1) Fv(n) 4 - q *(n )

y ( n ) = fv(n) -j- d x ( n )
(7.4.26) (7.4.27)
If w e com p are (7.4.19) and (7.4.20) with (7.4.26) and (7.4.27), w e ob serve that by a sim p le linear transform ation o f the state variables, w e have gen erated a n ew set o f sta te eq u ation s and an output eq u ation , in w hich th e input x ( n ) and the ou tp u t y ( n ) are unchanged. Since there is an infinite n um ber o f ch oices o f the transform ation m atrix P, there is also an infinite num ber o f state-sp ace eq u ation s
546
Chap. 7
and structures for a system . S om e o f these structures are different, w hile som e others are very sim ilar, differing o n ly by scale factors. A ssociated with any state-sp ace realization o f a system is the con cep t o f a m i n i m a l realization. A state-sp ace realization is said to b e m i n i m a l if the d im ension o f the state sp a ce (th e num ber o f state variab les) is th e sm allest o f all p ossible realizations. Since each state variable rep resen ts a quantity that m ust be stored and updated at every tim e instant n , it fo llo w s that a m inim al realization is o n e that requires the sm allest num ber o f d elays (storage registers). W e recall that the direct form II realization requires the sm allest num ber o f storages registers, and con sequ en tly, a state-sp ace realization based on the con ten ts o f the d elay elem en ts results in a m inim al realization. Sim ilarly, an F IR system realized as a direct form structure leads to a m inim al state-sp ace realization if th e valu es o f the storage registers are defined as the state variables. O n the other hand, the direct form I realization o f an IIR system d oes n ot lead to a m inim al realization. N o w , let us determ in e the im pulse resp on se o f the system from the statespace realization. T he im pulse resp onse provid es o n e o f the links betw een the in p u t-o u tp u t and state-space description o f system s. B y definition the im pulse resp onse h ( n ) o f a system is the zero-state re sponse o f the system to the excitation x ( n ) = 6 (n). H e n c e it can be obtained from equation (7.4.16) if w e set no = 0 (th e tim e w e apply th e input), v(0) = 0, and x ( n ) S(n). T hus the im pulse resp onse o f the system describ ed by (7.4.19) and (7.4.20) is given by h (n) g, Fn~ 1 q(n 1) + d 8(n) (7.4.28) = g '$ ( n l)qw (n 1 ) + d 8 ( n ) G iven a state-sp ace description, it is straightforward to d eterm in e the im pulse re sp on se from (7.4.28). H ow ever, the inverse is n ot easy since there is an infinite number o f state-sp ace realizations for the sam e in p u t-o u tp u t description.
The transpose system. T h e transpose o f a m atrix F is ob tain ed by inter changing its colum ns and rows, and it is d en oted by F . F or exam p le, r /n
/21 /12
/2 2
f\N '
r /11 F* =
/12
/21 /2 2
f2N
fm '
/V 2
-/v 1
fs2
fsN
-flN
flN
f NN -
N o w define the tra nspose s y s t e m (7 .4 .1 9 )-(7 .4.20) as v'tn + 1) = F v * (rt) + g x ( n ) y '( n ) = q V (n ) - \-d x ( n ) A ccording to (7.4.28), the im pulse resp on se o f this sy stem is given as t i ( n ) = qr(F')"-1 g(n - 1) + d S ( n ) (7.4.31) (7.4.29) (7.4.30)
Sec. 7.4
547
From m atrix algebra w e know that (F )" 1 = ( F -1 )'- H en ce h '( n ) ~ q ' ( F _ 1 ), gw(n - 1 ) + d5(n) W e claim that h'(n) = h(n). In d eed , the term q ' ( F -1 )'g is a scalar. H en ce it is eq u al to its transpose. C onsequently, [ q ' ( F - 1 )'g]' = g'(F y - ' q S ince this is true, it fo llow s that (7.4.31) is identical to (7.4.28) and, therefore, h '(n) = h (n). T hus a single in p ut-sin g le o u t p u t sy stem a n d its trans po se have i d e n tical im p u lse respons es a n d h en ce the s a m e i n p u t - o u t p u t relationship. T o support this claim further, w e n ote that the type 1 and type 2 state-sp ace realizations, d escribed by (7.4.3), (7.4.4), (7.4.10), and (7.4.11) are transpose structures, which stem from the sam e in p ut-ou tp u t relationship (7.4.1). W e have introduced the transpose structure b ecau se it provides an easy m eth o d for generating a new structure. H ow ever, som etim es this new structure m ay either differ trivially or be identical to the original on e.
The diagonal system. A closed -form solu tion o f the state-space equations is easily ob tain ed w hen the system m atrix F is diagonal. H en ce, by finding a m atrix P so that F = P F P - 1 is diagonal, the solu tion of the state eq u ation s is sim plified considerably. T h e d iagonalization o f the m atrix F can be accom plished by first d eterm ining the eigen valu es and eigen vectors o f the matrix. A num ber A is an eigenvalue o f F and a n on zero vector u is the associated eigen vecto r if
Fu = Au T o determ ine the eigen valu es o f F, w e n ote that (F - Xl)u = 0 (7.4.33) (7.4.32)
T his eq u ation has a (nontrivial) non zero solu tion u if the m atrix F - XI is singular [i.e., if (F Al) is n oninvertible], which is the case if the determ inant o f (F /.I) is zero, that is, if d et (F - XI) = 0 (7.4.34)
This determ inant in (7.4.34) yield s the characteristic p o l y n o m i a l o f the m atrix F. For a n N x N m atrix F, the characteristic polyn om ial o f F is degree N and h en ce it has jV roots, say X,, i = 1, 2 .........N . T h e roots m ay b e distinct or som e roots m ay b e repeated. In any case, for each root A,, w e can d eterm ine a vector u (, called the eigen vector corresponding to the eigen valu e X,, from the equation FU; = X(U; T h ese eigen vectors are orthogonal, that is, uju, = 0, for i ^ j . If w e form a m atrix U w h ose colum ns consist o f the eigen vectors {u, }, that is,
f t
U = Ul - 1 1
t
U2
t
U jV i J
548
Chap. 7
then the m atrix F = U -1 F U is diagonal. Thus w e have solved for the m atrix that diagonalizes F. T h e follow in g exam ple illustrates the procedure o f d iagon alizin g F. Example 7.4.2 The Fibonacci sequence, which is the sequence {1,1,2, 3, 5 ,8 .1 3 ....} , can be gener ated as the impulse response of the system that satisfies the state-space equations
v ( + 1) = j j v < n ) + j ^ J .x ( n )
y(rt) = [ 1
1 ] v ( n ) + x(n)
Determine the impulse response {/()} of the system. Solution Now we wish to determine an equivalent system v(n + 1) = Fv(n) + qx(n) y(n) = g'v(n) + dx(n) such that the matrix F is diagonal. From (7.4.25) we recall that the two systems are equivalent if F = PFP' q = P? g '= g 'P " 1
Given F. the problem is to determine a matrix P such that F = PFP - 1 is a diagonal matrix. First, we compute the determinant in (7.4.34). We have d e t(F -A l) = d e t [ j or x, = 1 + V5 2
X 1 = 0
X 2 = - 1 -
1 - Vs
To find the eigenvector U] corresponding to A.], we have
[?
Similarly, we obtain
r '=[i]
We observe that ur ,u 2 = 1 + = 0 (i.e., the eigenvectors are orthogonal). Now matrix U, whose columns are the eigenvectors of F, is
Then the matrix U -1FU is diagonal. Indeed, it easily follows that Xj

0 1
Sec. 7.4
549
and since the transformation matrix is P = U -1, we have

p
_ !_ _r ^ -M = _ A .2 A .i L A -i 1 J
Thus the diagonal matrix F has the form
M o':]
where the diagonal elements are the eigenvalues of the characteristic polynomial. Furthermore, we obtain _1_ q = Pq = and r = vl L 'V 5 J
'3 + V s 3 - V 5
2 2
The impulse response of this equivalent diagonal system is fi(n) = g'Fqu(n - 1) +d&(n )
(^K^y
m m
u(n - 1 ) + 5(n) which is the general formula for the Fibonacci sequence. An alternative expression can be found by noting that the Fibonacci sequence can be considered as the zero-input response of the system described by the difference equation y(n) = y(n - 1 ) + y(n - 2 ) + x (n) with initial conditions j>< 1) = 1, y ( 2) = 1. From the type 1 state-space realization, we note that U](0) = y { - 2 ) = - 1 and t^(0) = ;y (-l) = 1. Hence r * ( n Lt> 2 (0 ) J p r ^ n Lv2 < 0 )J zi 5 r - 3 + V5 n 2
3 + VS
and the zero-input response is y n(") = r ^ ^ ( 0 ) () This is the more familiar form for the Fibonacci sequence, where the first term of the sequence is zero, that is, the sequences is {0 , 1 , 1 , 2 , 3 , 5 , 8 , . . .)
550
Chap. 7
This exam ple illustrates the m eth od for diagonalizing the matrix F. The diagonal system y ield s a set o f N d ecou p led , first-order linear d ifferen ce equations that are easily solved to yield the state and the ou tp u t o f the system . It is im portant to n ote that the eigen valu es o f the matrix F are identical to the roots o f the characteristic polynom ial, which are ob tain ed from th e h om ogen eou s d ifferen ce eq u ation that characterizes the system . F or exam p le, the system that g en erates the F ibonacci seq u en ce is characterized by the h o m o g e n e o u s difference equation y(n)
-
y(n
I)
y{n - 2) = 0
(7.4.35)
R ecall that the solu tion is ob tain ed by assum ing that the h o m o g e n e o u s solution has the form yh(n ) = kn Substitution o f this solu tion into (7.4.35) yield s the characteristic polynom ial A2 - / - 1 = 0 B ut this is exactly the sam e characteristic polynom ial ob tain ed from the determ i nant o f (F - XI). Since the state-variable realization o f the system is not u n iq u e, the matrix F is also not unique. H ow ever, the eigen valu es o f the system are unique, that is, they are invariant to any nonsingular linear transform ation o f F. C onsequently, the characteristic polynom ial o f F can be determ in ed eith er from evaluating the determ inant o f ( F - A l ) or from the d ifferen ce eq u ation characterizing the system . In conclusion, th e state-sp ace description provides an alternative character ization o f the system that is eq u ivalen t to the in p u t-ou tp u t description. O n e ad vantage o f the state-variable form ulation is that it provid es us with the additional inform ation concerning the internal (state) variables o f the system , inform ation that is not easily ob tain ed from the in p u t-ou tp u t description. Furtherm ore, the state-variable form ulation o f a linear tim e-invariant system allow s us to represent the system by a set o f (usually cou p led ) first-order differen ce eq u ation s. T he d e coupling o f the eq u ation s can be ach ieved by m eans o f a linear transform ation that can b e ob tain ed by solvin g for the eigen valu es and eigen vectors o f the system . The d ecou p led eq u ation s are then relatively sim ple to solve. M ore im portant, how ever, the state-space form ulation provides a pow erful, yet straightforward m eth od for d ealing with system s that have m ultiple inputs and m ultiple ou tp u ts (M IM O ). A l though w e have n ot con sid ered such system s in our study, it is in the treatm ent of M IM O system s w h ere the true p ow er and the b eau ty o f the sp ace-sp ace form ula tion can be fully appreciated.
7.4.4 State-Space Analysis in the z-Domain

T h e state-sp ace analysis in th e previous section s has b een perform ed in the time d om ain. H o w ev er, as w e have ob served previously, the analysis o f linear timeinvariant discrete-tim e system s can also b e carried ou t in the z-transform
Sec. 7.4
551
d om ain , often w ith greater ease. In this section w e treat the state-sp ace rep resen tation o f linear tim e-invariant d iscrete-tim e system s in the z-transform d o main. L et us con sid er the state-sp ace eq u ation v(rt + 1) = Fv(n) + <pr(n) If w e define the v ecto r V (z) as * \(z ) 1 V2(z) V (z) = (7.4.36)
(7.4.37)
LVV(z).
then (7.4.36) can b e ex p ressed in m atrix form as zV (z) = F V (z) + q * ( z ) T h e tw o term s in volving can b e used to so lv e for (7.4.38)
V(z) can be V(z). Thus
collected togeth er and the resulting equation
(zi - F)V(z) = q*(z) V(z) = ( z I - F r V ( z )
(7.4.39)
T h e inverse z-transform o f (7.4.39) yield s the solu tion for the state equations. N ex t, w e turn our atten tion to the output eq u ation , which is given as j (n ) = g*v (n ) + d x ( n ) T h e z-transform o f (7.4.40) is y ( z ) = g ' V ( z ) + d X (z) (7.4.41) (7.4.40)
B y using the solu tion in (7.4.39) w e can elim in ate th e state vector V (z) in (7.4.41). Thus w e obtain y (z ) = [g, ( z I - F ) ~ 1q + d ]X (z) w hich is the z-transform o f the zero-state resp onse o f th e system . fun ction is easily ob tain ed from (7.4.42) as H (z ) = y = g ' ( z I - F ) - 1q + <f (7.4.42) T h e system
(7.4.43)
T h e state eq u ation given by (7.4.39), th e ou tp u t eq u ation given by (7.4.42) and the system function given by (7.4.43) all h ave in com m on the factor (z i F )- 1 . This is a fun d am en tal quantity that is related to the z-transform o f the state transition m atrix o f th e system . T h e relationship is easily estab lish ed by com puting the
552
Chap. 7
z-transform o f the im pulse resp on se h (n), which is g iv en by (7.4.28). T hus w e have

00
H (z) = X > ( )z ~ " n=0

oc
= - 1) + d i ( n ) \ z ~ n
(7.4.44)
T h e term in p a ren theses in (7.4.44) can b e w ritten as

00
= z - H l + F z 1 + F 2 z - 2 + ---) = z -H l-F z " 1)" 1 = ( z I - F ) ' 1
(7.4.45)
If w e substitute the result in (7.4.45) into (7.4.44), w e obtain the expression for H ( z ) as given in (7.4.43). Since the state transition m atrix is given by (/i) = F the z-transform o f (n) is
00
(7.4.46)
V z~" = I + F z - 1 + F 2 z - 2 + F V 3 + = ( I - F z - 1) - 1 = z ( z I - F ) - 1
T h e relation in (7.4.47) p rovides a sim p le m eth od for determ ining the state transition m atrix by m eans o f z-transform s. W e recall that (z I-F T 1 = (7.4.48)
d e t(z l - F)
w here a d j(A ) d en o te s the a djo int m a trix o f A and d et ( A ) d en o te s the determ inant o f the matrix A . Substitution o f (7.4.48) in to (7.4.43) yield s the result
(7 A 4 9 >
C on sequ en tly, the den om in ator D ( z) o f th e system fun ction H ( z ) , w hich contains the p o les o f the system is sim ply D ( z ) = d e t(z l - F) (7.4.50)
But the d e t(z l F) is just the characteristic p olyn om ial o f F. Its roots, w hich are the p o les o f system , are th e eigen valu es o f th e m atrix F.
Sec. 7.4
553
Example 7A 3 Determine the system function H(z), the impulse response h(n), and the state tran sition matrix 3>(n) of the system that generates the Fibonacci sequence. This system is described by the state-space equation v(n + l) = j^ y(n) = [ 1 Solution
j j v(n) +
1 ] v(n) + jr(n)
x(n)
First, we determine H(z) and h(n) by computing (zl - F)_t. We have
Hence
..[71 !][?] +'

1
z2 - ; - 1
l- z - '-z - 2
By inverting / / ( ), we obtain h(n) in the form h(n) = y/5

(H)
We note that the poles of H(z) are />] = (] + \/5 )/2 and p 2 = (1 - \/5)/2. Since | Pi | > 1, the system that generates the Fibonacci sequence is unstable. The state transition matrix # (n ) has the z-transform z(zl - F) " 1 =
2 1 [ ;2 ~ Z z2 - z - 1 L z z J
The four elements of ^(n ) are obtained by computing the inverse transform of the four elements of z(zI - F)_1. Thus we obtain
L<fci(n) <h2(n)l
where
<hi(n) = ~7 =
H T -(4 T >
()
We note that the impulse response h{n) can also be computed from (7.4.28) by using the state transition matrix.
554
Chap. 7
This analysis m ethod applies specifically to the com p utation o f the zero-state response o f the system . T his is the con seq u en ce o f the fart that w e have used the tw o-sided z-transform . If w e wish to determ in e the total resp on se o f the system , beginning at a nonzero state, say v(/i0), w e m ust use the on e-sid ed z-transform . Thus, for a given initial state v(n0) and a given input * ( ) for n > no, w e can d eterm in e th e state vector v(n) for n > no and the output y ( n ) for n > no, by m ean s o f the one-sided z-transform. In this d ev elo p m en t w e assum e that no = 0, w ithou t loss o f generality. T hen, given ;r(n) for n > 0 , and a causal system , d escrib ed by th e state eq u ation s in (7.4.36), the on e-sid ed z-transform o f the state eq u ation s is zV + (z) - zv(0) = F V + (z) + q X (z) or, equivalently, V + (z) = z (z l - F r V O ) + (z l - F ) - 1 q X (z) N o te that X+ (z) = X (z), since x ( n ) is assum ed to b e causal. Similarly, the z-transform o f the output eq u ation given by (7.4.40) is Y +(z) = i V + (z) + d X ( z ) If w e substitute for V + (z) from (7.4.51) in to (7.4.52), w e obtain the result Y +(z) = z g ' ( z I - F r 1 v(0) + [ g ' ( z I - F r 1q + rf]X (z) (7.4.53) (7.4.52) (7.4.51)
O f the terms on the right-hand side o f (7.4.53), the first rep resen ts the zero-input response o f the system due to the initial con d ition s, w hile the secon d represents the zero-state resp onse o f the system that w e ob tain ed previously. C onsequently, (7.4.53) constitutes the total resp onse o f th e system , which can b e exp ressed in the tim e dom ain by inverting (7.4.53). T h e result o f this inversion y ield s the form for y(n) given p reviously by (7.4.16).
7.4.5 Additional State-Space Structures

In Section 7.4.2 w e described h ow state-sp ace eq u ation s can b e ob tain ed from a given structure and, con versely, h ow to obtain a realization o f the system given the state equations. In this section w e revisit the parallel-form and cascade-form realizations described previously and con sid er th ese structures in the con text o f a state-space form ulation. T h e parallel-form state-sp ace structure is ob tain ed by exp an d ing the system function H ( z ) in to a partial-fraction exp an sion , d ev elo p in g the state-sp ace form u lation for each term in the expansion and the corresponding structure, and finally, connecting all the structures in parallel. W e illustrate the p roced u re under the assum ption that the p o les are distinct and N M . T he system fun ction H ( z ) can be exp ressed as N D H (z) = C + Y (7.4.54)
Sec. 7.4
555
N o te that this is a d ifferent exp an sion from that given in (7.3.20). T h e output o f the system is Y (z ) = H ( z ) X ( z ) = C X ( z ) + k=1 w here, by definition, X(Z) it = 1 ,2 ......... N Z - Pt In the tim e dom ain, the eq u ation s in (7.4.56) b ecom e Yk (z) = y t (n + 1) = p kyk(n) + * ( ) W e define the state variables as vt(n ) = yt (n) k = 1 ,2 ,..., N * = 1 , 2 ......... N 0 0 v(n) +
_
Bk Yk(z)
(7.4.55)
(7.4.56)
k = 1 , 2 , , N
(7.4.57) (7.4.58) (7.4.59)
T h en the d ifferen ce eq u ation s in (7.4.57) b ecom e Vk(n + 1 ) = PkVkin) + x ( n ) "Pi 0 v(n + 1) = 0 B2 Pn o pj T h e state eq u ation s in (7.4.59) can b e exp ressed in m atrix form as "1" 1 jc(rt) .1 . (7.4.61) (7.4.60)
and the output equation is y (n ) = [ B] B n ] v(n ) + C x ( n ) This parallel-form realization is called the n o r m a l f o r m representation, b e cause the m atrix F is d iagonal, and h en ce the state variables are uncoupled. A n alternative structure is ob tain ed by pairing com p lex-con ju gate p oles and any tw o real-valued p o les to form secon d-ord er section s, which can be realized by using eith er type 1 or type 2 state-sp ace structures. T h e ca scade-fo rm state-sp ace structure can b e ob tain ed by factoring H ( z ) in to a product o f first-order and secon d-ord er section s, as described in S ection 7.2.2, and then im plem en ting each section by using eith er type 1 or type 2 state-sp ace structures. L et us con sid er the state-sp ace representation o f a single secon d-ord er section in volving a pair o f com p lex-con ju gate p oles. T h e system function is H (z) = bo + b i z ~ l + f a z ~ 2
1 + O iZ - 1 + 0 2 Z ' 2
boz2 + b i z + b 2 z 2 + a i z + 02 A A* + z - p Z -p * The output of this system can be expressed as = bo +
(7.4.62)
t (z) + , ----------AX& h . -------- y(z) = b0x

z - p Z -p
(7.4.63)
556 W e define the quantity
Chap. 7
S (z) =
z -p
(7.4.64)
This relationship can be exp ressed in the tim e dom ain as s{n + 1) = p s ( n ) + A x { n ) Since s(n ), p , and A are com p lex valued, w e define .y(n) as s(n ) = (/i) + j v 2(n) (7.4.66) (7.4.65)
P = i + ;' 2 A q \+ jq2
U p o n substitution o f these relations into (7.4.65) and separating its real and im aginary parts, w e obtain i> i ( + ] ) = cfiui (n) ct2V2( n ) + q \ x { n ) (7.4.67)
U2( + 1) = <*2i>i(n) + a ] U 2 ( n ) + q2x ( n )
W e ch oose vi(n) and v2(n) as the state variables and thus ob tain the coupled pair o f state equations w hich can be expressed in matrix form as v(n + 1 ) = ai a2
a2
ai
v() +
L</2 j
j:( n)
(7.4.68)
T he output eq u ation can be exp ressed as y ( n ) = b0x ( n ) + s ( n ) + j*(n ) (7.4.69)
U p on substitution for s(n ) in (7.4.69), w e obtain the desired result for the output in the form y ( n ) = [2 0 ] v ( ) + b0x ( n ) (7.4.70)
A realization for the secon d-ord er section is show n in Fig. 7.31. It is simply called the co u p le d - fo rm state-sp ace realization. T his structure, w hich is used as the building block in the im plem en tation o f cascade-form realizations for higher-order IIR system s, exh ib its low sensitivity to finite-w ord-length effects.
7.5 REPRESENTATION OF NUMBERS

U p to this point w e h ave con sid ered the im plem en tation o f d iscrete-tim e systems without being co n cern ed ab ou t the finite-w ord-length effects that are inherent in any digital realization, w h eth er it b e in hardware or in softw are. In fact, w e have analyzed system s that are m od eled as linear w h en , in fact, digital realizations of such system s are inherently nonlinear. In this and th e fo llow in g tw o section s, w e con sid er th e variou s form s of quantization effects that arise in digital signal processing. A lth o u g h w e describe
Sec. 7.5
Representation of Numbers
i> o
557
i
Figure 7 3 1 system. Coupled-form state-space realization of a two-pole, two-zero IIR
floating-point arithm etic operations briefly, our m ajor concern is with fixed-point realizations o f digital filters. In this sectio n w e con sid er the representation o f n um bers for digital com pu tations. T h e m ain characteristic o f digital arithm etic is the lim ited (usually fixed) nu m b er o f digits used to represent num bers. T his constraint leads to finite n u m erical p recision in com putations, w hich leads to rou n d -off errors and nonlinear effects in th e p erform ance o f digital filters. W e n ow provide a brief introduction to digital arithm etic.
7.5.1 Flxed-Point Representation of Numbers

T h e rep resen tation o f n um bers in a fixed-point form at is a generalization o f the fam iliar d ecim al rep resen tation o f a num ber as a string o f digits w ith a decim al p oin t. In this n otation , th e digits to th e left o f th e d ecim al point represent the in teger part o f th e num ber, and the digits to the right o f the decim al p oin t represent the fractional part o f th e num ber. T hus a real num ber X can be represented as X = {b~A ........ b ^ M M ..........b B)r = t
i rzA
>,r~
0 < ,< ( r - l)
<7'5 1 )
w h ere bt rep resen ts the digit, r is th e radix or base, A is th e num ber o f integer
sse
Chap. 7
digits, and B is the num ber o f fractional digits. A s an exam p le, th e decim al num ber (123.45)io and the binary num ber ( 1 0 1 .0 1 ) 2 r e p r e se n t th e fo llo w in g sums: (123.45)io =
1
x 10 2 + 2 x 10 1 + 3 x 10 + 4 x 10_I + 5 x 10 ~ 2
( 1 0 1 .0 1 ) 2 = 1 x 2 2 + 0 x 2 1 + 1 x 2 + 0 x 2 _ 1 + 1 x 2 2 L et us focus our atten tion on the binary representation sin ce it is the m ost im portant for digital signal processing. In this case r = 2 and the digits {,} are called binary digits or bits and take the valu es {0 ,1 }. T h e binary digit b - A is called the m ost significant bit (M S B ) o f th e num ber, and the binary d igit b B is called the least significant bit (L S B ). T h e binary p o in t b etw een the digits bo and b\ d oes not exist physically in the com puter. Sim ply, the logic circuits o f the com puter are design ed so that the com p utations result in num bers that correspond to the assum ed location o f this point. B y using an n-bit integer form at (A = n 1, B = 0), w e an represent unsigned integers with m agnitude in the range 0 to 2" - 1. U sually, w e use the fraction format (A = 0, B = n - 1), with a binary p oin t b etw een bo and b \, that perm its num bers in the range from 0 to 1 - 2~n. N o te that any in teger or m ixed num ber can be represented in a fraction form at by factoring o u t the term r A in (7.5.1). In the sequ el w e focu s our atten tion on the binary fraction form at because m ixed num bers are difficult to m ultiply and the num ber o f b its representing an integer cannot be reduced by truncation or rounding. T here are three w ays to represent n egative num bers. T h is leads to three form ats for the representation o f signed binary fractions. T h e form at for positive fractions is the sam e in all three representations, nam ely,
B
X = 0.bib2 bB = Y l b <- 2 ', X > 0 ;i
(7.5.2)
N o te that the M SB bo is set to zero to represent th e positive sign. C onsider now the negative fraction
B
X = - O M h bB = - J 2 b i - 2 ~ T his num ber can b e represented using o n e o f th e follow in g th ree form ats.
(7 -5 -3>
S ign -m agn itu d e form at.

the n egative sign,
In this form at, th e M SB is set to 1 to represent for X < 0 (7.5.4)
.Xsm l.b \b 2 - b B
O n e's-com p lem en t form at.

resen ted as
In this form at th e negative num bers are rep X < 0 (7.5.5)
X iC = l b i h - b B
w here fc; = 1 - i>, is th e o n e s com p lem en t o f bt . T h u s if X is a p ositive number, the corresponding negative num ber is d eterm in ed by com p lem en tin g (changing l s
Sec. 7.5
559
to 0 s and 0 s to l s) all th e bits. A n alternative definition for X ic can b e obtained by notin g that
B
X 1C = 1 x 2 + ( 1 - bi) 2 -' = 2 - 2 ~ b \X\

r=l
(7.5.6)
Twos-complement format. In this form at a negative num ber is repre sen ted by form ing the tw o s com p lem en t o f the corresponding p ositive num ber. In o th er w ords, the negative num ber is ob tain ed by subtracting the p ositive num ber from 2.0. M ore sim ply, the tw o s com p lem en t is form ed by com p lem en tin g the p ositive num ber and adding on e L SB . Thus
X 2c = l . b i b 2 - - b B + 0 0 - - 0 1 X < 0 (7.5.7)
w here + represents m o d u lo -2 addition that ignores any carry gen erated from the sign bit. For exam p le, the num ber 2 is sim ply ob tain ed by com p lem en tin g 0011 ( | ) to obtain 1100 and then adding 0001. T his yield s 1101, which rep resents j in tw o s com p lem en t. From (7.5.6) and (7.5.7) is can easily be seen that X 2Q = X ]C + 2 ~ H = 2 - \X\ (7.5.8)
T o d em onstrate that (7.5.7) truly represents a n egative num ber, w e use the identity
B
l = J 2 2 ~i + 2 ' B i=i T h e negative num ber X in (7.5.3) can b e exp ressed as

B
(7 -5 -9 )
X 2C = - J 2 b i - 2 "' + 1 - 1
1= 1
B
= - 1 + J ] ( l - bi)2~' + 2~B i=i = - l + ^ b i -2 -1 + 2~b

i= i
w hich is exactly th e tw o s-com p lem ent rep resen tation o f (7.5.7). In sum m ary, the value o f a binary string b o h b B d ep en d s on the form at used. For p ositive num bers, bo = 0, and the num ber is given by (7.5.2). For negative num bers, w e use th ese corresponding form ulas for the three form ats. Example 7.5.1 Express the fraction g and | in sign-magnitude, twos-complement, and onescomplement format.
560 Solution

X = | is rep resen ted as 2~' + 2 - 2 + 2~3, s o that X = 0 .1 1 1 .
Chap. 7
In sign-
m agnitude form at, X = | is rep resen ted as 1.111. In o n e 's co m p lem e n t, we have
X,c = 1.0 0 0
In tw os co m p lem en t, th e result is
X 2C = 1.0 0 0 + 0.00 1 = 1 .0 0 1 T he basic arithm etic op eration s o f addition and m ultiplication d epend on the form at used. For o n e s-com p lem ent and tw o s-com p lem ent form ats, addition is carried out by adding the num bers bit by bit. T h e form ats differ only in the way in which a carry bit affects the M SB. F or exam p le, f = g. In tw o s com p lem en t, we have 0 1 0 0 1101 = 0 0 0 1 where indicates m o d u lo-2 addition. N o te that the carry bit, if p resent in the M SB. is dropped. O n the o th er hand, in o n e s- com p lem en t arithm etic, the carry in the M SB, if present, is carried around to the LSB. T h u s the com p utation g | = g b ecom es
0100 1100 = 0000 0001 = 0001

A ddition in the sign-m agnitude form at is m ore com p lex and can in volve sign checks, com plem enting, and the generation o f a carry. O n the other hand, di rect m ultiplication o f tw o sign- m agnitude num bers is relatively straightforward, w hereas a special algorithm is usually em p loyed for o n e s co m p lem en t and tw os com plem ent m ultiplication. M ost fixed-point digital signal processors use tw o s-com p lem en t arithm etic. H en ce, the range for ( B + l ) bit num bers is from 1 to 12 - f i . T h ese num bers can be view ed in a w h eel form at as show n in Fig. 7.32 for B = 2. T w o s-com p lem ent arithm etic is basically arithm etic m odu lo- 2 B +1 [i.e., any num ber that falls ou tsid e
0
0.0
-2
0.5
-4 (a)
1.0
(b)
Figure 7.32 Counting wheel for 3-bit twos-complement numbers (a) integers and (b) functions.
Sec. 7.5
561
the range (overflow or underflow ) is reduced to this range by subtracting an appro* priate m ultiple o f 2 B+1]. T h is type o f arithm etic can be view ed as cou n tin g using the w h eel o f Fig. 7.32. A very im portant property o f tw o s- com p lem en t addition is that if the final sum o f a string o f num bers X \ , X i , . . . , X N is w ithin the range, it w ill be com p uted correctly, even if individual partial sum s result in overflow s. T h is and other characteristics o f tw o s-com p lem ent arithm etic are con sid ered in Problem 7.44. In g en eral, th e m ultiplication o f tw o fixed-point num bers each o f b bits in length results in a product o f 2b bits in length. In fixed-point arithm etic, the product is eith er truncated or rounded back to b bits. A s a result w e have a truncation or rou n d -off error in the b least significant bits. T he characterization o f such errors is treated b elow .
7.5.2 Binary Floating-Point Representation of Numbers

A fixed-point representation o f num bers allow s us to cover a range o f num bers, say* *max *min with a resolution
. -*max -Xmin
w h ere m = 2* is the num ber o f levels and b is the num ber o f bits. A basic character* istic o f the fixed-point represen tation is that the resolution is fixed. Furtherm ore, A increases in direct p roportion to an increase in the dynam ic range. A floating-point representation can be em p loyed as a m ean s for covering a larger dynam ic range. T h e binary floating-point representation com m on ly used in practice, con sists o f a m antissa M , which is the fractional part o f the num ber and falls in the range | < M < 1, m ultiplied by the exp on en tial factor 2 E, w here the exp on en t E is either a p ositive or n egative integer. H en ce a num ber X is represented as X = M -2E T h e m antissa requires a sign bit for representing p ositive and negative num bers, and the ex p o n en t requires an additional sign bit. Since th e m antissa is a signed fraction, w e can use any o f the four fixed-point rep resen tation s just described. For exam p le, the num ber X i = 5 is represented by the follow in g m antissa and exponent: M i = 0 .10 10 0 0 = 0 11 w hile the num ber X 2 = | is represented by th e follow in g m antissa and exp on en t M2 0.110000
E 2 = 101
where the leftmost bit in the exponent represents the sign bit.
562
Chap. 7
If the tw o num bers are to b e m ultiplied, the m antissas are m ultiplied and the exp on en ts are added. Thus the product o f th e se tw o num bers is X i X 2 = M \ M 2 2 El+E2
=
( 0 .011110) - 2010 ( 0 . 111100) - 20 0 1
O n the other hand, the addition o f the tw o floating-point n um bers requires that the exp on en ts be equal. T his can b e accom p lish ed by sh iftin g th e m antissa o f the sm aller num ber to the right and com p en sating by increasing the corresponding exp on en t. Thus the num ber X 2 can b e exp ressed as M 2 = 0 .0 0 0 0 1 1 E 2 = O il W ith E 2 = E], w e can add the tw o num bers X \ and X 2. T h e result is X i + X 2 = (0.101011) -2 011 It should be ob served that the shifting operation required to equalize the exp on en t o f X 2 with that for X 1 results in loss o f precision, in general. In this exam p le the six-bit m antissa w as sufficiently lon g to a ccom m od ate a shift o f four bits to the right for M 2 w ithout dropping any o f the on es. H ow ever, a shift o f five bits w ould have caused the loss o f a single bit and a shift o f six bits to the right w ould have resulted in a m antissa o f M 2 = 0.000000, u n less w e round upward after shifting so that M 2 = 0.000001. O verflow occurs in the m ultiplication o f tw o floatin g-poin t num bers w hen the sum o f the exp on en ts exceed s the dynam ic range o f the fixed -p oin t representation o f the exp on en t. In com paring a fixed-point represen tation w ith a floatin g-poin t representa tion, each w ith the sam e num ber o f total bits, it is apparent that the floating p oin t representation allow s us to cover a larger dynam ic range by varying the resolution across the range. T h e resolu tion d ecreases w ith an increase in the size o f successive num bers. In other words, the distance b etw een tw o successive floating-point num bers increases as the num bers increase in size. It is this vari able resolution that results in a larger dynam ic range. A ltern atively, if w e wish to co v er the sam e dynam ic range with b oth fixed-point and floating-point rep resentations, th e floating-point representation provid es finer resolu tion for sm all num bers but coarser resolution for th e larger num bers. In contrast, the fixedp oin t representation provides a uniform resolu tion through ou t the range o f num bers. For exam p le, if w e have a com p uter with a w ord size o f 32 bits, it is p ossible to represent 2 32 num bers. If w e wish to represent the p o sitiv e integers beginning w ith zero, th e largest p ossib le in teger that can b e accom m od ated is 2 32 1 = 4,294,967,295
Sec. 7.5
563
T h e distance b etw een su ccessive num bers (th e resolu tion ) is 1. A ltern atively, w e can d esign ate the leftm ost bit as the sign bit and u se the rem aining 31 bits for th e m agnitu d e. In such a case a fixed-point rep resen tation allow s us to cover the range (2 31 - 1) = 2,147,483,647 to (2 31 - 1) = 2,147,483,647
again with a resolu tion o f 1 . O n the other hand, su p p ose that w e increase the resolu tion by allocating 10 bits for a fractional part, 21 bits for the in teger part, and 1 bit for the sign. T hen this representation allow s us to cover the d ynam ic range or, eq u ivalen tly, -2 ,0 9 7 ,1 5 1 .9 9 9 to 2,097,151.999 ( 2 31 1 ) 2 ~ 10 = ( 2 21 2 -t0 ) to ( 2 31 1 ) 2 - 1 0 = 2 21 2 ~ 10
In this case, the resolution is 2 ~ 10. Thus, th e dynam ic range has b een decreased b y a factor o f approxim ately 1 0 0 0 (actually 2 10), w hile th e resolution has b een increased by the sam e factor. For com parison, su p p ose that the 32-bit w ord is u sed to represent floating poin t num bers. In particular, let the m antissa be rep resen ted by 23 bits plus a sign bit and let th e exp on en t b e represented by 7 bits plus a sign bit. N ow , the sm allest num ber in m agnitude will have the representation, sign 0,
10 0
23 bits - 0
sign
1
7 bits
1111111
= i x 2 127 ^ 0 . 3 x 1 0 ' 38
A t the o th er extrem e, the largest num ber that can be represented w ith this floating p oin t representation is sign 0 23 bits 111 -- 1 sign 0 7 bits 1111111 = ( l - 2 23) x 2 m 1.7 x 1038
Thus, w e h ave ach ieved a dynam ic range o f ap p roxim ately 1076, but with varying resolu tion . In particuiar, w e have fine resolu tion for sm all num bers and coarse resolu tion for larger num bers. T h e representation o f zero p o ses so m e special p rob lem s. In general, on ly th e m antissa has to be zero, but not the exp on en t. T h e ch oice o f M and , the representation o f zero, th e handling o f overflow s, and other related issu es h ave resu lted in various floating-point rep resen tation s o n different digital co m puters. In an effort to define a com m on floating-point form at, the Institute o f E lectrical and E lectron ic E n gin eers (IE E E ) in trodu ced th e IE E E 754 standard, w hich is w id ely used in practice. For a 32-bit m achine, th e IE E E 754 standard sin gle-p recision , floating-point num ber is rep resen ted as X = ( 1)* 2 - l 27 (Af), w here
8 9
31
564
Chap. 7
T his num ber has th e follow in g interpretations: If E = 255 and M ^ 0, then X is n ot a num ber If E 255 and M 0, then X ( - 1 ) 5 oo If 0 < E < 255, then X = ( - 1 ) 5 2 E~ l21( l . M ) If E - 0 and M ^ 0, then X = ( - 1 ) 5 2"I26 (0.Af) If E = 0 and M = 0, then X = ( - 1 ) 5 0 w here 0 . M is a fraction and l . M is a m ixed num ber with on e in teger bit and 23 fractional bits. F or exam p le, th e num ber
0 S 1 0 0 0 0 0 10 E 10 1 0 M 00
has th e value X =
1 x 2 13 0 -12 7 x 1 .1 0 1 0 ...0 = 2 3 x y
13.
T h e m agni
tude range o f the 32-bit IE E E 754 floating-point num bers is from 2 ~ m x 2 -2 3 to (2 2 23) x 2 127 (i.e., from 1,18 x 10 ~ 38 to 3.40 x 1038). C om p utations with num bers ou tsid e this range result in eith er underflow or overflow .
7.5.3 Errors Resulting from Rounding and Truncation

In perform ing com p utations such as m ultiplications with eith er fixed-point or floating-point arithm etic, w e are usually faced with the p rob lem o f quantizing a num ber via truncation or rounding, from a given level o f p recision to a lev el o f low er precision. T h e effect o f rounding and truncation is to in trodu ce an error w h ose value d ep en d s on th e num ber o f bits in th e original num ber relative to the num ber o f bits after quantization. T h e characteristics o f th e errors introduced through eith er truncation or rounding d ep en d on th e particular form o f num ber representation. T o be specific, let us con sid er a fixed-point representation in w hich a num ber x is quantized from b u bits to b bits. Thus the num ber bu x = 0 .1 0 1 1 0 1 consisting o f bu bits prior to quantization is rep resen ted as b
x = o.ioi i
after quantization, w h ere b < bu. F or exam p le, if x rep resen ts th e sam p le o f an analog signal, then bu m ay b e taken as infinite. In any ca se if th e quantizer truncates th e value o f x , th e truncation error is defined as Et = Q , ( x ) - x
(7.5.10)
Sec. 7.5
565
First, w e con sid er th e range o f valu es o f the error for sign-m agnitude and tw o s-com p lem ent representation. In both o f th ese representations, the positive num bers have identical rep resen tation s. For p ositive num bers, truncation results in a num ber that is sm aller than the unquantized num ber. C on sequ en tly, the trunca tion error resulting from a reduction o f the num ber o f significant bits from b to b is - Q T b - 2 ~ b) < E, < 0 (7.5.11)
w h ere the largest error arises from discarding bu b bits, all o f which are ones. In the case o f n egative fixed-point num bers based on the sign-m agnitude representation, the truncation error is p ositive, since truncation basically reduces the m agnitude o f the num bers. C on sequ en tly, for negative num bers, w e have 0 < E, < ( X b - 2 ~ b-) (7.5.12)
In the tw o s-com p lem en t rep resen tation , the negative o f a num ber is obtained by subtracting the corresp on d ing p ositive num ber from 2. A s a con seq u en ce, the effect o f truncation on a negative num ber is to increase the m agnitude o f the negative num ber. C on sequ en tly, x > Q t (x) and hence (2~h 2~b) < E, < 0 (7.5.13)
H en ce w e con clu d e that the truncation er ror f o r the sig n-m ag nitu de representation is sy m m e t r i c a b o u t ze ro a n d falls in the range - ( 2 ~ h - 2~b) < , < ( 2~h - 2 ~b") (7.5.14)
O n the other hand, f o r tw o 's-co m p lem e n t representation, the trunc ation error is a lw a ys neg ative a n d falls in the range (2~h 2 ~ b ) < E, < 0 (7.5.15)
N ex t, let us con sid er the quantization errors due to rounding o f a num ber. A num ber x , rep resen ted by bu bits b efore quantization and b bits after quantization, incurs a q uantization error r = Q r(x ) - X (7.5.16)
B asically, rounding in v o lves on ly the m agnitude o f the num ber and, con sequ en tly, the rou n d -off error is in d ep en d en t o f the type o f fixed-point representation. T he m axim um error that can be introduced through rounding is ( 2~ b 2~b,) /2 and this can be eith er p ositive or n egative, d ep en d ing on the valu e o f x . T h erefore, the r o u n d - o f f er ro r is s y m m e t r i c a b o u t z e r o a n d falls in the ra nge - \ { 2 ~ b - 2~ h ) < E r < \ { 2 ~ b - 2~b' ) (7.5.17)
T h ese relation sh ip s are sum m arized in Fig. 7.33 w hen x is a continuous signal am plitude (b u = oo). In a floating-point rep resen tation , the m antissa is either rounded or truncated. D u e to the n onuniform resolu tion , the corresponding error in a floating-point rep resen tation is p rop ortional to the num ber b ein g quantized. A n appropriate
566

QM)
Chap. 7
E, - QM) - *
2'
(a )
,= Q,M - x 2-b < , < 0

(b)
, = (2,M - x - 2 -* 5 , < 2 -
(c )
Figure 733 Quantization errors in rounding and truncation: (a) rounding; (b) truncation in twos complement; (c) truncation in sign-magnitude. represen tation for th e quantized value is Q ( x ) = x + ex w here e is called the relative error. N o w Q ( x ) - x = ex (7.5.19) (7.5.18)
Sec. 7.5
567
In the case o f truncation based on tw o s-com p lem en t rep resen tation o f the m antissa, w e have - 2 E2~b < e,x < 0 for p ositiv e num bers. Since 2 _ 1 < x < 2 , it follow s that - 2~ m < e, < 0 x > 0 (7.5.21) (7.5.20)
O n the other hand, for a negative num ber in tw o s-com p lem en t representation, the error is
0 < e,x < 2E2~b
and h en ce 0 < e, < 2~ b+l x <0 (7.5.22)
In the case where the m antissa is rounded, the resulting error is sym m etric relative to zero and has a m axim um valu e o f 2 ~ hf l . C on sequ en tly, the round-off error b eco m es - 2 e 2~hf l < erx < 2 E 2~ hf l (7.5.23)
A gain , since x falls in the range 2 _ 1 < x < 2 , w e divide through by 2 E~ ] so that - 2~h < er < 2~b (7.5.24)
In arithm etic com putations in volving q uantization via truncation and round ing, it is con ven ien t to adopt a statistical approach to the characterization o f such errors. T h e quantizer can be m od eled as introducing an additive n oise to the unquantized value x . T hus w e can w rite Q(x )
= X +
w h ere e = E r for rounding and e = E, for truncation. T h is m od el is illustrated in Fig. 7.34. S ince x can be any num ber that falls w ithin any o f the levels o f the quan tizer, the quantization error is usually m odeled as a random variable that falls
I Quantizer CM
(a)
{*)---(b)
Figure 734 Additive noise model for the nonlinear quantization process: (a) actual system; (b) model for quantization.
568
Chap. 7
within the lim its specified. T his random variable is assum ed to be uniform ly distributed w ithin th e ranges sp ecified for th e fixed-point rep resen tation s. Fur therm ore, in practice, bu > > b, so that w e can n eglect th e factor o f 2~ b- in the form ulas given b elow . U n d er th ese con d ition s, the probability density func tions for the rou n d -off and truncation errors in the tw o fixed-point representations are illustrated in Fig. 7.35. W e n ote that in the case o f truncation o f the tw o scom plem ent representation o f the num ber, the average value o f the error has a bias o f 2~ bf2, w h ereas in all other cases just illustrated, th e error has an average value o f zero. W e shall use this statistical characterization o f the q uantization errors in our treatm ent o f such errors in digital filtering and in the com p utation o f the D F T for fixed-point im plem entation. p(r)
A=
2~h
A
2
A
2
Er
(a)
P(E,)
2A
A = 2-*
-A
E,
(b) P<|)
A A = 2~b
0
(c)
Figure 7J5 Statistical characterization of quantization errors: (a) round-off error, (b) truncation error for sign-magnitude; (c) truncation error for twos complement.
Sec. 7.6
Quantization of Filter Coefficients
569
7.6 QUANTIZATION OF FILTER COEFFICIENTS

In th e realization o f F IR and IIR filters in hardware or in softw are on a generalpu rp ose com puter, th e accuracy w ith w hich filter coefficients can b e specified is lim ited by th e w ord length o f the com puter or the length o f the register provided to store the co efficients. Since the coefficients used in im plem en ting a given filter are n ot exact, the p o les and zeros o f the system function w ill, in general, be d ifferent from th e d esired p o les and zeros. C onsequently, w e obtain a filter having a frequency resp on se that is d ifferent from the frequency resp onse o f the filter with unquantized coefficients. In S ection 7.6.1, w e d em onstrate that the sensitivity o f th e filter frequency resp onse characteristics to quantization o f the filter coefficien ts is m inim ized by realizing a filter having a large num ber o f p o les and zeros as an interconnection o f secon d -ord er filter section s. T his lead s us to the parallel-form and cascadeform realizations in which the b asic building blocks are secon d-ord er filter section s.
7.6.1 Analysis of Sensitivity to Quantization of Filter Coefficients

T o illustrate the effect o f quantization o f the filter coefficients in a direct-form realization o f an IIR filter, let us consider a general IIR filter with system function M I > * -
Jt = 0
(7.6.1)
T h e direct-form realization o f the IIR filter w ith quantized coefficients has the system function
*3=0
(7.6.2)
w h ere th e quantized coefficients {jt}and {a*) can b e related to the unquantized coefficien ts {*} and [ak ] by th e relations a t ~ ok + A a* b t = bt + A b t jfc = 1 ,2 (7.6.3) k = 0, 1 ,,.., M
and {Aa*} and {Ai*} represent the quantization errors.
570
Implementation of Discrete-Time Systems T h e d en om in ator o f H ( z ) m ay be exp ressed in the form
Chap. 7
d (z )
= i + Y l akZ * = n a
~ PkZ ^
(7.6.4)
w h ere H { z ) as
are th e p o les o f H ( z ) . Sim ilarly, w e can express th e d en om in ator o f
N (7.6.5)
w here ~pk = p k + A p k, k = 1 , 2 , . . . , N , and A p t is the error or perturbation resulting from the quantization o f the filter coefficients. W e shall now relate th e perturbation A p t to the q u an tization errors in the fa tl. T he perturbation error Ap, can b e expressed as
w here dp,/dan, the partial derivative o f with respect to ak, rep resen ts the incre m ental change in the p ole p, d u e to a change in the coefficien t a k. Thus the total error Ap, is exp ressed as a sum o f the increm ental errors d u e to ch an ges in each o f the coefficients {ak }. T h e partial d erivatives d p t / d a k, k = 1 , 2 , . . . , N , can be ob tain ed by differen tiating D (z) with respect to each o f the {a*}. First w e have (7.6.7)
T hen dPi _ { d D ( z ) / d a k)z=Pi dak T he num erator o f (7.6.8) is O D ( z ) / d z ) z=Pl (7.6.8)
(7.6.9)
T h e den om in ator o f (7.6.8) is
Sec. 7.6
571
(7.6.10)
7 * n (n -p i)
* /=i
T h erefore, (7.6.8) can be expressed as (7.6.11)
1=1
Substitution o f the result in (7.6.11) into (7.6.6) yield s the total perturbation error A p, in the form
N N-k
(7.6.12)
/=] Mt T his exp ression provides a m easure o f the sensitivity o f the /th p ole to changes in the coefficients {a*). A n analogous result can b e ob tain ed for the sensitivity o f the zero s to errors in the param eters ( it ) . T h e term s (p,- pi) in th e denom inator o f (7.6.12) represent vectors in the z -plane from the p o les {/?/} to the p ole p,. If th e p oles are tightly clustered as th ey are in a narrowband filter, as illustrated in Fig. 7.36, the lengths |p, - p t \ are sm all for th e p o les in the vicinity o f p, . T h ese sm all len gth s will contribute to large errors and h en ce a large perturbation error Ap, results. T h e error Ap, can b e m inim ized by m axim izing the lengths |p,- p/|. T his can be accom p lish ed by realizing the high-order filter with eith er sin gle-p ole or d o u b le p o le filter section s. In general, how ever, sin gle-p ole (and sin gle-zero) filter section s have com p lex-valu ed p o les and require com p lex-valu ed arithm etic op eration s for their realization. T h is problem can b e avoided by com b inin g com plex-valued p o les (and zero s) to form secon d-ord er filter section s. Since th e com p lex-valu ed p oles are usually sufficiently far apart, the perturbation errors {A p,} are m inim ized. A s a co n seq u en ce, the resulting filter with quantized coefficien ts m ore closely ap proxim ates th e freq u en cy resp onse characteristics o f th e filter w ith unquantized coefficients. It is interesting to n ote that ev en in the case o f a tw o -p o le filter section , the structure used to realize the filter section plays an im portant role in the errors
572 Im(z)
Chap. 7
Figure 7 3 6
Pole positions for a bandpass HR filter.
caused by coefficient quantization. T o be specific, let us con sid er a tw o-p ole filter with system function H{z) =
1 1 (2 /- c o s 0 )z - 1 + r 2z~ 2
(7.6.13)
This filter has p o les at z = r e j<i. W hen realized as shown in Fig. 7.37, it has two coefficients, a\ 2 r cos 6 and a 2 = r 2. W ith infinite precision it is p ossib le to achieve an infinite num ber o f p ole p osition s. Clearly, with finite precision (i.e., q uantized coefficients oi and a2), th e p ossib le p ole p osition s are also finite. In fact, w hen b bits are used to represent the m agnitudes o f a\ and a2, there are at m ost (2h l ) 2 p ossib le p osition s for the p o les in each quandrant, exclu d in g the case a\ = 0 and a2 = 0 . F or exam ple, su p p ose that b = 4. T hen there are 15 p ossib le n on zero values for a\. T here are a lso 15 p ossib le valu es for r 2. W e illustrate th e se p ossib le valu es in Fig. 7.38 for the first quandrant o f the z-plane only. T here are 169 p ossib le pole Mn)
<*>
2rcos 6
-y(n)
Figure 7.37 IIR filter.
Realization of a two-pole
Sec. 7.6
573
Figure 7.38 Possible pole positions for two-pole IIR filter realization in Fig. 7.37. p o sition s in this case. T h e nonuniform ity in their p osition s is d u e to th e fact that w e are quantizing r 2, w h ereas the p ole p osition s lie on a circular arc o f radius r. O f particular significance is the sparse set o f p oles for valu es o f 6 near zero and, due to sym m etry, near 6 = n . T h is situation w ould b e highly unfavorable for low pass fil ters and highpass filters w hich norm ally h ave p o les clustered near 6 = 0 and 8 = jr. A n alternative realization o f the tw o*pole filter is th e coupled-form realiza tion illustrated in Fig. 7.39. T h e tw o cou p led eq u ation s are > i( ) = x i n ) + r c o s # >j( 1 ) - r s i n # y (n 1 ) (7.6.14) y (n ) = r sin # y \( n 1 ) + r c o s 0 y (n 1 ) B y transform ing these tw o eq u ation s into the z-d om ain , it is a sim p le m atter to sh ow that y (z) X (z ) ~ . ( r s in 0 ) z - 1 1 - ( 2 r c o s 6 )z 1
+ r2 z~ 2
574
. 0

0
Chap. 7
>i(n)
rcos 9
V|{n - 1 )
rsin 6 t sin 6
V|(n - 1)
Figure 7.39 Coupled-form realization o f a two-pole IIR filter.
In the cou p led form w e ob serve that there are also tw o coefficients, a i r s in # and a 2 = r c o s 9. Since they are both linear in r, th e possib le p o le positions are now equally spaced points on a rectangular grid, as sh ow n in Fig. 7.40. A s a co n seq u en ce, the p ole p osition s are now uniform ly distributed inside the unit circle, which is a m ore d esirable situ ation than the p reviou s realization, especially for low pass filters. (T h ere are 198 possib le p o le p osition s in this case.) H ow ever, the price that w e pay for this uniform distribution o f p ole p osition s is an increase in com putations. T h e cou p led -form realization requires four m ultiplications per ou tp u t point, w h ereas the realization in Fig. 7.37 requires o n ly tw o m ultiplications p er output point. It is interesting to com pare the coupled-form realization o f Fig. 7.39 with the cou p led (or norm al) form state-sp ace structure o f Fig. 7.31. T h e p oles o f the state-space structure are directly related to its coefficients, sin ce on and a 2 are the real and im aginary parts o f the roots. Since aj = r co s and a 2 r sin , it is clear that quantizing cri and a 2 results in a rectangular grid o f possib le p o le p osition s, as sh ow n in Fig. 7.40. Since there are various w ays in which o n e can realize a secon d-ord er filter section , there are o b v iou sly m any p ossib ilities for d ifferen t p o le locations with quantized coefficients. Ideally, w e should select a structure that provides us with a d en se set o f p o in ts in th e region s w here th e p o les b e. U n fortu n ately, how ever, there is n o sim ple and system atic m eth od for determ ining th e filter realization that yield s this d esired result. G iven that a higher-order IIR filter should be im p lem en ted as a com bination o f secon d-ord er section s, w e still m ust d ecid e w h eth er to em p lo y a parallel config uration or a cascade configuration. In other w ords, w e m ust d ecid e b etw een the realization
(7.6.16)
Sec. 7.6
575
and the realization

h
(z ) = T
1 CM + C 1Z' 1 + fltiZ"1 + a*2Z-2
(7.6.17)
If the IIR filter has zeros on the unit circle, as is gen erally the case with ellip tic and C h eb ysh ev type II filters, each secon d-ord er sectio n in th e cascade configu ration o f (7.6.16) contains a pair o f com p lex-con ju gate zeros. T h e coefficients {&*} directly d eterm in e the location o f th ese zeros. If th e [by] are quan tized , the sen sitivity o f th e system resp onse to th e q uantization errors is easily and directly con trolled by allocating a sufficiently large num ber o f b its to th e representation o f th e {bki }. In fact, w e can easily evalu ate th e perturbation effe ct resulting from quan tizin g the coefficients to so m e sp ecified p recision. T h u s w e h ave direct con trol o f both th e p o les and th e zeros that result from th e quantization p rocess. O n th e other hand, the p arallel realization o f H ( z ) provid es direct con trol o f th e p o le s o f the system only. T h e num erator co efficien ts {chi} and {c*i} d o not
576
Chap. 7
specify the location o f the zeros directly. In fact, the {c*o) and fc*i} are ob tain ed by perform ing a partial-fraction exp an sion o f H ( z ). H en ce th ey d o not directly influence the location o f th e zeros, but on ly indirectly through a com b in ation o f all th e factors o f H ( z ) . A s a co n seq u en ce, it is m ore difficult to d eterm in e the effect o f quantization errors in the coefficients {c*, }, on the location o f the zeros o f the system . It is apparent that quantization o f the param eters {c*, J is lik ely to produce a significant perturbation o f the zero p osition s and usually, it is sufficiently large in fixed-point im plem en tation s to m ove the zeros o ff the unit circle. T his is a highly undesirable situation, w hich can b e easily rem edied by use o f a floating-point representation. In any case the cascade form is m ore robust in th e p resen ce o f co efficient quantization and should b e the preferred ch oice in practical applications, especially w here a fixed-point representation is em p loyed . Example 7.6.1 Determine the effect of parameter quantization on the frequency response of the 7-order elliptic filter given in Table 8.11 when it is realized as a cascade of secondorder sections. Solution The coefficients for the elliptic filter given in Table 8.11 are specified for the cascade form to six significant digits. We quantized these coefficients to four and then three significant digits (by rounding) and plotted the magnitude (in decibels) and the phase of the frequency response. The results are shown in Fig. 7.41 along the frequency response of the filter with unquantized (six significant digits) coefficients. We observe that there is an insignificant degradation due to coefficient quantization for the cascade realization. Example 7.(L2 Repeat the computation of the frequency response for the elliptic filter considered in Example 7.6.1 when it is realized in the parallel form with second-order sections. Solution The system function for the 7-order elliptic filter given in Table 8.11 is
0 .2 7 8 1 3 0 4 + 0 .0 0 5 4 3 7 3 1 0 8 Z -1 (2) " 1 - 0 .7 9 0 1 0 3 Z '1 - 0 .3 8 6 7 8 0 5 + 0 .3 3 2 2 2 2 9 ;- ' + 1 - 1 .5 1 7 2 2 3 j-1 + 0 .7 1 4 0 8 8 z ~ 2 0.1 2 7 7 0 3 6 - 0 .1 5 5 8 6 9 6 :- 1 + 1 - 1 .4 2 1 7 7 3 * -1 + 0 .8 6 1 8 9 5 z 2 - 0 .0 1 5 8 2 4 1 8 6 + 0 .3 8 3 7 7 3 5 6 z " ]
+ 1 - 1.387447*-1 +0.962242z-i The frequency response of this filter with coefficients quantized to four digits is shown in Fig. 7.42a. When this result is compared with the frequency response in Fig. 7.41, we observe that the zeros in the parallel realization have been perturbed sufficiently so that the nulls in the magnitude response are now at - 8 0 , - 8 5 , and 92 dB. The phase response has also been perturbed by a small amount.
Sec. 7.6
577
Phase (degree)
Gain (db)
Figure 7.41 Effect of coefficient quantization of the magnitude and phase response of an N 7 elliptic filter realized in cascade form. When the coefficients are quantized to three significant digits, the frequency response characteristic deteriorates significantly, in both magnitude and phase, as illustrated in Fig. 7.42b. It is apparent from the magnitude response that the zeros are no longer on the unit circle as a result of the quantization of the coefficients. This result clearly illustrates the sensitivity of the zeros to quantization of the coefficients in the parallel form. When compared with the results of Example 7.6.1, it is also apparent that the cascade form is definitely more robust to parameter quantization than the parallel form.
578
Chap. 7
Phase (degree)
Gain (db)
(a) Quantization to 4 digits
Figure 7.42 Effect of coefficient quantization of the magnitude and phase response of an N = 7 elliptic filter realized in cascade form: (a) quantization to four digits; (b) quantization to three digits.
7.6.2 Quantization of Coefficients In FIR Filters

A s ind icated in the preceding section , the sensitivity analysis p erform ed o n the p o le s o f a system a lso ap p lies directly to th e zeros o f th e IIR filters. C on sequ en tly, an exp ression analogou s to (7.6.12) can b e ob tain ed for th e zeros o f an F IR filter. In effect, w e should gen erally realize F IR filters w ith a large n u m b er o f zeros as
Sec. 7.6
579
Phase (degree)
Gain (db)
(b) Quantization to 3 digits
Figure 7.42 Continued
a cascad e o f secon d-ord er and first-order filter section s to m inim ize the sensitivity to coefficient quantization. O f particular interest in practice is the realization o f linear p h ase F IR filters. T h e direct-form realizations show n in Figs. 7.1 and 7.2 m aintain th e linear-phase property ev en w h en th e coefficients are quantized. T his follo w s easily from the o b servation that the system function o f a linear-phase F IR filter satisfies th e property H( z ) =
580
Chap. 7
in d ep en dent o f w h eth er the coefficien ts are quantized or unq uan tized (see S ec tion 8.2). C on sequ en tly, co efficien t quantization d o es not affect th e p h ase charac teristic o f the F IR filter, but affects o n ly the m agnitude. A s a result, coefficient quantization effects are n ot as severe on a linear-phase F IR filter, since the on ly effect is in the m agnitude. Example 7.63 Determine the effect of parameter quantization on the frequency response of an Af = 32 linear-phase FIR bandpass filter. The filter is realized in the direct form. Solution The frequency response of a linear-phase FIR bandpass filter with unquantized coefficients is illustrated in Fig. 7.43a. When the coefficients are quantized to four significant digits, the effect on the frequency response is insignificant. However, when the coefficients are quantized to three significant digits, the sidelobes increased by several decibels, as illustrated in Fig. 7.43b. This result indicates that we should use a minimum of 10 bits to represent the coefficients of this FIR filter and, preferably, 12 to 14 bits, if possible. From this exam ple w e learn that a m inim um o f 10 bits is required to represent the coefficients in a direct-form realization o f an F IR filter o f m od erate length. A s the filter length increases, the num ber o f bits per coefficient m ust be increased to m aintain the sam e error in the freq u en cy resp onse characteristic o f the filter. For exam ple, suppose that each filter coefficient is rounded to {b + 1) bits. T hen the m axim um error in a coefficien t value is bounded as _ 2 -<fr+t> < e h(n) < 2 ~ (b+l) Since the quantized valu es m ay b e rep resen ted as h(n) = h (n) + eh(n), th e error in the frequency resp onse is = y ^ f ek {n)e' Since eh(n) is zero m ean, it follow s that E m (to) is also zero m ean . A ssu m in g that the coefficient error seq u en ce e h(), 0 < n < M - 1, is uncorrelated, the variance o f the error E m {co) in the frequency resp on se is just the sum o f the variances o f the M terms. Thus w e have 2 - 2 (i>+i) 2~2(b+2> cr| = : M = ----------- M 12 3 H ere w e note that the variance o f the error in H(a>) increases linearly with M . H en ce the standard d eviation o f the error in H { a>) is
2 ~(H-2 ) a = \ /3
C onsequently, for every factor o f 4 in crease in M , th e precision in the filter coeffi cients m ust b e increased by o n e ad d ition al bit to m aintain th e standard deviation fixed. This result, taken togeth er w ith th e results o f E xam p le 7.6.3, im plies that
Sec. 7.6
581
Gain (dB)
Relative frequency (a) No quantization
G ain (dB)
(b) Quantization to 3 digits Figure 7.43 Effect of coefficient quantization of the magnitude of an M = 32 linear-phase FIR filter realized in direct form: (a) no quantization; (b) quantization to three digits. th e freq u en cy error rem ains tolerab le for filter lengths up to 256, provided that filter coefficients are rep resen ted by 12 to 13 bits. If th e w ord length o f the d igi tal signal processor is less than 12 bits or if th e filter len gth excee d s 256, the filter sh ou ld b e im plem en ted as a cascad e o f sm aller length filters to reduce the precision requirem ents.
582
Implementation of Discrete-Time Systems In a cascade realization o f the form K H ( z ) = G ]" [//* ( 0 i=l
Chap. 7
(7.6.18)
w here the secon d-ord er section s are given as H k(z) = 1 + bklz ~ l + b k2z~ 2 the coefficients o f com p lex-valu ed zeros are exp ressed as b (7 A 1 9 ) = 2rk cos 6k and
bk2 = r\. Q uantization o f bki and bk2 result in zero location s as sh ow n in Fig. 7.38,
ex cep t that the grid exten d s to p oin ts ou tsid e th e unit circle. A problem m ay arise, in this case, in m aintaining the lin ear-ph ase property, b ecau se th e quantized pair o f zeros at z = ( l / r k)ej6k m ay n ot b e th e m irror im age o f the quantized zero s at z = rke j6k. T his problem can be avoid ed by rearranging the factors corresponding to the m irror-im age zero. That is, w e can write the m irror-im age factor as f l co s0 * z _ 1 + -~rz- 2 ) = -^-(r? 2 rk cos 0 *7 + z ~ 2) (7.6.20)
rk
r\
T h e factors {1 / r 2} can be com bined with the overall gain factor G, or they can be distributed in each o f the secon d-ord er filters. T h e factor in (7.6.20) contains exactly the sam e param eters as the factor ( 1 2rk c o s 0 *z - 1 + r^z- 2 ), and co n se quently, the zeros now occur in m irror-im age pairs even w hen th e param eters are quantized. In this brief treatm ent w e have given the reader an introduction to the problem s o f coefficient q uantization in IIR and F IR filters. W e have dem onstated that a high-order filter should b e reduced to a cascad e (for F IR or IIR filters) or a parallel (for IIR filters) realization to m inim ize th e effects o f quan tization errors in the coefficients. T his is esp ecially im portant in fixed-point re alizations in w hich the coefficients are represented by a relatively sm all num ber o f bits.
7.7 ROUND-OFF EFFECTS IN DIGITAL FILTERS

In Section 7.5 w e characterized the q uantization errors that occu r in arithm etic op eration s p erform ed in a digital filter. T h e p resen ce o f o n e or m ore quantiz ers in the realization o f a d igital filter results in a n on linear d evice w ith char acteristics that m ay be significantly different from the id eal linear filter. For exam p le, a recursive digital filter m ay exh ib it undesirable oscillation s in its out put, as show n in the follow in g section , even in the ab sen ce o f an input signal. A s a result o f the finite-precision arithm etic op eration s perform ed in the digital filter, so m e registers m ay overflow if th e input signal le v e l b ecom es large.
Sec. 7.7
Round-Off Effects in Digital Filters
583
O verflow rep resents another form o f undesirable n on linear distortion on the d e sired signal at the output o f the filter. C on sequ en tly, sp ecial care m ust be exercised to scale th e input signal properly, eith er to p revent overflow com p letely or, at least, to m inim ize its rate o f occurrence. T h e non linear effects due to finite-precision arithm etic m ake it extrem ely difficult to precisely analyze the p erform ance o f a digital filter. T o perform an analysis o f q uantization effects, w e adopt a statistical characterization o f qu an ti zation errors w hich, in effect, results in a linear m od el for th e filter. Thus w e are able to quantify the effects o f quantization errors in th e im plem en tation o f d igi tal filters. O ur treatm ent is lim ited to fixed-point realization s where quantization effects are very im portant.
7.7.1 Limit-Cycle Oscillations in Recursive Systems

In the realization o f a digital filter, either in digital hardware or in software on a digital com p uter, the quantization inherent in the finite- precision arithm etic op era tio n s render the system nonlinear. In recursive system s, the nonlinearities due to the finite-precision arithm etic op eration s o ften cau se periodic oscillation s to occur in the output, even w hen the input seq u en ce is zero or som e n on zero con stan t value. Such oscillations in recursive system s are called limit cycles and are directly attributable to round-off errors in m ultiplication and overflow errors in addition. T o illustrate the characteristics o f a lim it-cycle oscillation , let us con sid er a sin gle-p ole system described by the linear d ifferen ce eq u ation y(n) = ay( n - 1) + x ( n ) (7.7.1)
w h ere th e p o le is at z = a. T h e ideal system is realized as show n in Fig. 7.44. O n the other hand, the actual system , which is described by the non linear differen ce eq u ation v( n ) = Q[ av( n - 1)] + x ( n ) (7.7.2)
is realized as show n in Fig. 7.45. S u p pose that the actual system in Fig. 7.45 is im p lem en ted w ith fixed-point arithm etic based on four bits for the m agnitude plus a sign bit. T he quantization that takes p lace after m ultiplication is assum ed to round the resulting product upward.
Figure 7.44 system.
Idea] single-pole recursive
584
Implementation of Discrete-Time Systems W n)
Chap. 7
Figure 7.45 Actual nonlinear system.
In T able 7.2 w e list the resp onse o f the actual system for four different locations o f the p o le z a, and an input x(n ) = /9S(n), w here fi = 15/16, which has the binary representation 0.1111. Ideally, the resp onse o f th e system should decay toward zero exp on en tially [i.e., y( n ) = a" * 0 as n * oo]. In the actual system , how ever, the resp onse v(n) reaches a steady-state p eriod ic output sequ en ce with a period that d ep en d s on the value o f the p ole. W hen the p o le is p ositive, the oscillations occur with a period N p = 1, so that the output reach es a con stan t value of ^ for o = j and g for a = O n th e other hand, w hen the p o le is n egative, the output sequ en ce o scillates b etw een p ositive and negative valu es ( ^ for a = | for a = He n c e the p eriod is N p = 2. T h ese limit cycles occur as a result o f the quantization effe cts in m ultipli cations. W hen the input seq u en ce x ( n ) to the filter b ecom es zero, the output o f the filter then, after a num ber o f iterations, enters into the lim it cycle. T h e ou t put rem ains in the lim it cycle until an oth er input o f sufficient siz e is applied that drives the system ou t o f the limit cycle. Sim ilarly, zero-input lim it cycles occur from nonzero initial con d ition s with the input x ( n ) = 0 . TTie am plitudes o f the output during a lim it cycle are confined to a range o f values that is called the d ead b a n d o f the filter. and
TABLE 7.2 LIMIT CYCLES FOR LOWPASS SINGLE-POLE FILTER
0 .1 0 0 0
a = ~
a =
1 .1 0 0 0 1
2
a =
_
2 0 .1 1 1 1 1 .1 0 0 0 0 .0 1 0 0 1 .0 0 1 0 0 .0 0 0 1 1 .0 0 0 1 0 .0 0 0 1 1 .0 0 0 1 0 .0 0 0 1
0 .1 0 1 1 0 .1 0 0 0 0 .0 1 1 0 0 .0 1 0 1 0 .0 1 0 0 0 .0 0 1 1 0 .0 0 1 0 0 .0 0 1 0 0 .0 0 1 0
0 .1 1 0 0 3 4
a =
1 .1 1 0 0
3
0 1
0 .1 1 1 1 0 .1 0 0 0 0 .0 1 0 0 0 .0 0 1 0 0 .0 0 0 1 0 .0 0 0 1 0 .0 0 0 1 0 .0 0 0 1 0 .0 0 0 1
\ 16 /
(f t) (f t) (f t) (f t) (f t) (f t) (f t)
(if) ( - ft)
(S )
(f t)
0 .1 0 1 1 1 .1 0 0 0 0 .0 1 1 0 1 .0 1 0 1 0 .0 1 0 0 1 .0 0 1 1 0 .0 0 1 0 1 .0 0 1 0 0 .0 0 1 0
(B )
( - ft) (f t) (- ft) (f t) (-ft) (f t) (- ft)
m
(f t)
(ft)
(f t) (f t) (f t) (f t) (f t)
3 4 5
6
( - ft)
(-ft) (f t) (- ft)
7
8
(ft)
(ft )
(ft)
(ft)
Sec. 7.7
Round-Off Effects in Digital Fitters
585
It is interesting to n ote that w h en the resp onse o f the sin gle-p ole filter is in the lim it cycle, the actual n onlinear system op erates as an equivalent linear system w ith a p o le at z = 1 w h en the p o le is p ositive and z = 1 w h en the p o le is negative. That is, (7.7.3) Since the qu an tized product av(n 1) is ob tain ed by rounding, it fo llo w s that the quantization error is b ou n ded as \Q r[av(n - 1)] - a v ( n - 1)| < \ 2~b (7.7.4)
w h ere b is th e num ber o f bits (exclu sive o f sign) used in th e representation o f the p o le a and u(n). C on sequ en tly, (7.7.4) and (7.7.3) leads to |v(n - 1 )| - \ctv{n - 1 )| < \ 2~b and h en ce
1
2~b
N " ~ DI < f 7 1 - \a\
(7.7.5)
T he exp ression in (7.7.5) d efines the dead band for a sin gle-p ole filter. F or ex a m p le, w hen b = 4 and ja| = 1, w e have a dead band with a range o f am plitudes ^ ) . W hen b = 4 and \a\ = 2, th e dead band increases to (g, | ) . T h e lim it-cycle behavior in a tw o-p ole filter is m uch m ore com plex and a larger variety o f o scillation s can occur. In this case the ideal tw o-p ole system is described by the linear differen ce eq u ation , ( >>(n) = ai)'(n - 1) + a2y ( n - 2) + jc(n) (7.7.6)
w h ereas the actual system is described by the non linear d ifferen ce eq u ation v(n) = Qr [a iv( n - 1)] + Q r f a v i n - 2)] + x ( n ) (7.7.7)
W hen th e filter coefficients satisfy the con d ition a \ < 4o2, the p oles o f the system occur at z = r e ie w h ere a 2 = r 2 and ai = 2 r c o s 0 . A s in the case o f th e sin gle-p ole filter, w hen the system is in a zero-input or zero-state lim it cycle, Qr[a2v(n - 2)] = - v ( n - 2) (7.7.8)
In other w ords, th e system b eh aves as an oscillator with com p lex-con ju gate p o les on the unit circle (i.e., a2 r 2 = 1). R ou n d ing the product a v ( n 2) im plies that | Q r[a2v(n - 2)] - a2v{n - 2 ) \ < \ - 2~ b U p o n substitution o f (7.7.8) into (7.7.9), w e obtain the result |u(n - 2 )| - |a2v( n - 2 )| < \ 2~b (7.7.9)
586
Chap. 7
or equivalently,
(7.7.10) T h e expression in (7.7.10) d efines the dead band o f th e tw o-p ole filter with com plexcon gu gate p oles. W e ob serve that the dead-band lim its d ep en d on ly on |ci2 1 - T he param eter a\ = 2r cos 8 d eterm in es th e frequency o f oscillation. A n oth er p o ssib le lim it-cycle m od e with zero input, w hich occurs as a result o f rounding the m ultiplications, corresponds to an eq u ivalen t secon d -ord er system with p o les at z = 1 . In this case it w as show n by Jackson (1969) that the tw o-p ole filter exh ib its oscillation s with an am plitude that falls in the d ead band b ounded by 2 - * / ( l - | il - 0 2 )It is interesting to n ote that these lim it cycles result from rounding the prod uct o f the filter coefficients with the previous outputs, t>( 1 ) and v(n 2 ). Instead o f rounding, w e m ay ch o o se to truncate the products to b bits. W ith trun cation, w e can elim in ate m any, although not all, o f the lim it cycles as show n by Claasen et al. (1973). H ow ever, recall that truncation results in a biased error unless the sign-m agnitude representation is u sed, in which case the truncation er ror is sym m etric about zero. In general, this bias is undesirable in digital filter im plem entation. In a parallel realization o f a high-order IIR system , each secon d-ord er filter section exhibits its ow n lim it-cycle behavior, with n o interaction am ong the secondorder filter section s. C on sequ en tly, the output is the sum o f the zero-input limit cycles from the individual section s. In the case o f a cascade realization for a highorder IIR system , the lim it cycles are m uch m ore difficult to an alyze. In particular, w hen the first filter section exhibits a zero-input limit cycle, the output lim it cycle is filtered by the su cceed in g section s. If the frequency o f the lim it cycle falls near a resonance frequency in a su cceed in g filter section , th e am plitu d e o f the seq u en ce is en h an ced by the reson ance characteristic. In general, w e m ust be careful to avoid such situations. In addition to lim it cycles caused by rounding the result o f m ultiplications, there are limit cycles caused by overflow s in addition. A n o v erflow in addition o f tw o or m ore binary num bers occurs w hen the sum e x cee d s the w ord size available in the digital im plem en tation o f the system . F or exam p le, let us con sider the secon d-ord er filter section illustrated in Fig. 7.46, in w hich th e addi tion is perform ed in tw os-com p lem ent arithm etic. Thus w e can w rite the output y i n ) as y i n ) = g [ a i y in - 1 ) + a 2y i n - 2 ) + -c(/i)] (7.7.11)
w h ere the function g [ ] represents th e tw os-com p lem ent ad d ition . It is easily verified that th e function g(u ) versus v is described by the graph in Fig. 7.47. R ecall that th e range o f valu es o f the param eters (0 1 , 0 2 ) for a stab le filter is g iven by the stability triangle in Fig. 3.15. H ow ever, th e se con d ition s are no longer sufficient to p reven t overflow oscillation w ith tw os-com p lem en t arithm etic.
Sec. 7.7
Round-Off Effects in Digital Filters x(n)
587 y(n)
<>
Q>
02
g\v\
Figure 7.46 Two-pole filter realization.
Figure 7.47 Characteristic functional relationship for two's complement addition of two or more numbers. In fact, it can easily be show n that a necessary and sufficient con d ition for ensuring that n o zero-input overflow limit cycles occur is |oi | + \ai | < 1 (7.7.12)
w hich is extrem ely restrictive and h en ce an u n reason ab le constraint to im pose on an y secon d -ord er section. A n effectiv e rem edy for curing the prob lem o f overflow oscillation s is to m o d ify the a d d er characteristic, as illustrated in Fig. 7.48, so that it perform s sat uration arithm etic. T h u s w h en an overflow (or underflow ) is sen sed , th e output o f the ad d er w ill be the full-scale valu e o f 1 . T h e distortion caused by this nonlinearity in the adder is usually sm all provid ed that saturation occurs in fre q u en tly. T h e u se o f such a nonlinearity d o es n ot p reclu d e the n eed for scal ing o f the signals and th e system param eters, as d escrib ed in the follow in g se c tion.
586 g(v)
Chap. 7
Figure 7.48 Characteristic functional

relationship for addition with clipping at
1.
Saturation arithm etic as just d escribed elim in ates lim it cycles d u e to overflow , on the o n e hand, but o n the other hand, it cau ses undesirable signal distortion due to the nonlinearity o f the clipper. In order to limit the am ount o f n on linear distortion, it is im portant to scale the input signal and the unit sam ple resp onse, b etw een the input and any internal sum m ing n od e in the system , such that overflow b ecom es a rare event. For fixed-point arithm etic, let us first con sid er the extrem e con d ition that overflow is not perm itted at any nod e o f th e system . Let yk (n ) d en o te the response o f the system at the Jtth n od e w hen the input seq u en ce is x{n) and the unit sam ple response b etw een the n od e and the input is h t i n ) . T hen !>*() I = h k( m ) x ( n - m )
-o o oo
S u ppose that jc(n) is upper b ou n ded by A x . T hen O C !>*()! < A z \hk (m)\ for all n (7.7.13)
N o w , if the dynam ic range o f th e com p uter is lim ited to (1 ,1 ) , th e con d ition !>*(") I < 1 can be satisfied by requiring that the input x ( n ) b e scaled such that
1
(7.7.14)
IM m)l
for all p ossib le n o d es in th e system . T h e con d ition in (7.7.14) is b oth necessary and sufficient to p revent overflow .
Sec. 7.7
589
T h e con d ition in (7.7.14) is overly con servative, h ow ever, to the point w here the input signal m ay be scaled to o m uch. In such a case, m uch o f the precision used to represent x ( n ) is lost. T his is especially true for narrowband sequ en ces, such as sinusoids, w here the scaling im plied by (7.7.14) is extrem ely severe. For narrowband signals w e can use the frequency resp on se characteristics o f the system in determ ining the appropriate scaling. Since |//(a>)| rep resents the gain o f the system at frequency co, a less severe and reason ably ad eq u ate scaling is to require that m ax | (ct>) | (7-7 -15)
0 <o><;r
w h ere Hk(w) is the Fourier transform o f {/t*(n)). In the case o f an F IR filter, the con d ition in (7.7.14) reduces to A, < ^ m=0 which is now a sum over the M n on zero term s o f the filter unit sam ple response. A n oth er approach to scaling is to scale the input so that
OO OO
------l*t(m )|
(7.7.16)
\yt(n )\2 < C 2 \ x(n)\2 = C 2E x n= = -00 OC = 00 F rom P arsevals theorem ;m w e have f]

n ôo
(7.7.17)
|>-t( ) | 2 =
[
2 * J -*
\H(co)X(co)\2da> (7.7.18)
^ /I
B y com b inin g (7.7.17) with (7.7.18), w e obtain
fn
00
T \hk(n )\2 tô o
( 1 /2 * ) / \H(a>)\2da) J ~n
If w e com pare the d ifferent scaling factors given ab ov e, w e find that
n = -o o
I 1 /2
00
I M n )|2
< m ax|flt(o> )| <

n-oo
IM )I
(7-7-20)
C learly, (7.7.14) is th e m ost pessim istic constraint. In the fo llo w in g section w e ob serve the ram ifications o f this scaling on the o u tp u t sign al-to-n oise (pow er) ratio (S N R ) from a first-order and a second-order filter section.
590
Chap. 7
7.7.3 Statistical Characterization of Quantization Effects in Fixed-Point Realizations of Digital Filters

It is apparent from our treatm ent in the previou s section that an analysis o f quan tization errors in digital filtering, based on d eterm inistic m o d e ls o f quantization effects, is not a very fruitful approach. T h e basic problem is that the nonlinear effects in quantizing th e products o f tw o num bers and in clipping the sum o f tw o num bers to prevent overflow are n ot easily m od eled in large system s that contain many m ultipliers and m any sum m ing nodes. T o obtain m ore general results on the q uantization effects in digital filters, w e shall m od el the quantization errors in m ultiplication as an ad d itive n oise sequ en ce e ( n ), just as w e did in characterizing th e q uantization errors in A /D con version o f an analog signal. For addition, w e con sid er the effect o f scaling the input signal to prevent overflow . L et us begin our treatm ent w ith the characterization o f th e rou n d -off n oise in a sin gle-p ole filter w hich is im plem en ted in fixed-point arithm etic and is described by the nonlinear d ifferen ce eq u ation v(n) = Qr[av( n 1)] + x ( n ) (7.7.21)
The effect o f rounding the product a v ( n 1) is m od eled as a n oise seq u en ce e(n) added to the actual product a v(n 1 ), that is, Q r[av(n 1)] = a v ( n 1) + e ( n ) (7.7.22)
W ith this m odel for the q uantization error, the system under consideration is described by the linear difference eq uation v( n) = a v ( n 1) + x ( n ) + e(n) (7.7.23)
T he corresponding system is illustrated in b lock diagram form in Fig. 7.49. It is apparent from (7.7.23) that the output seq u en ce v( n ) o f the filter can be separated in to tw o com p on en ts. O n e is th e resp onse o f th e system to the input seq u en ce x( n ) . T h e secon d is the resp onse o f the system to the additive quantization n o ise e( n). In fact, w e can express the output se q u en ce u(n) as a sum o f th ese tw o com p onents, that is, v ( n ) = y( n ) + q ( n ) (7.7.24)
Figure 7.49 Additive noise model for the quantization error in a single-pole filter.
Sec. 7.7
591
w h ere y ( n ) rep resen ts the response o f the system to * (n ), and q ( n ) rep resen ts the resp onse o f th e system to the quantization error e(n). U p o n substitution from (7.7.24) fo r v(n) in to (7.7.23), w e obtain y ( n ) + q(rt) = a y(n - 1 ) + a q (n - 1) + je(n) + e(n) (7.7.25)
T o sim p lify the analysis, w e m ake the follow in g assu m p tion s ab ou t the error seq u en ce e(n). 1. For any n, th e error sequ en ce {e(n)\ is uniform ly distributed o v er the range ( _ i . 2 ~b, i - 2~b). T his im plies that the m ean value o f e(n) is zero and its variance is 2 - 2b e = - f i 0 -1.26 ) 2 . T h e error {()) is a stationary w hite n oise seq u en ce. In o th er words, the error e( n) and the error e(m ) are uncorrelated for n ^ m. 3. T h e error seq u en ce {e(n)} is uncorrelated with the signal se q u en ce {*()}. T h e last assum ption allow s us to separate the differen ce eq u ation in (7.7.25) into tw o uncou p led differen ce equations, nam ely, y (n ) = a y(n - 1) + x (n) q (n) = aq(tt - 1) + e ( n ) (7.7.27) (7.7.28)
T h e differen ce equation in (7.7.27) represents th e in p u t-ou tp u t relation for the desired system and the d ifferen ce equation in (7.7.28) rep resen ts the relation for th e q uantization error at the output o f the system . T o co m p lete the analysis, w e m ake use o f tw o im portant relation sh ip s d evel op ed in A p p en d ix A . T h e first is the relationship for the m ean valu e o f the output q ( n ) o f a linear shift-invariant filter with im pulse resp onse h{h) w h en excited by a random seq u en ce e( n) h aving a m ean value m e. T h e resu lt is
00
mq = me
fl= 00
h(rt)
(7.7.29)
or, eq u ivalen tly, m q ~ m eH { 0) (7.7.30)
w h ere H { 0 ) is the v alu e o f the frequency resp onse H(co) o f th e filter evalu ated at (o = 0. T h e seco n d im portant relationship is the exp ression for th e au tocorrelation seq u en ce o f th e output q ( n ) o f the filter with im pulse resp on se h ( n ) w h en the input random se q u en ce e{ri) has an autocorrelation y ( n ) . T h is result is = E h ( k ) h ( l ) y te(k l + n) (7.7.31)
i oo/= o o
592
Chap. 7
In th e im portant special case w h ere the random seq u en ce is w h ite (spectrally flat), the autocorrelation yee(n) is a unit sam p le seq u en ce scaled by the variance a 2, that is, Yee(n) = cr2S (n) (7.7.32)
U p o n substituting (7.7.32) in to (7.7.31), w e ob tain th e desired result for the au to correlation seq u en ce at the ou tp u t o f a filter excited by w hite n o ise, nam ely,
00
y(n) =
Y , h ik> h ^k + n)
fc= 00
(7 7 33)
T he variance a 2 o f the output n oise is sim ply ob tain ed by evaluating y ^ i n ) at n = 0. Thus

OO
q 2 = < Y , h2^
* = s -O C
<7 -7-34>
and w ith the aid o f P arsevals th eorem , w e have the alternative expression a] = ^ / \H(a>)\2du> (7.7.35)
In the case o f the sin gie-p ole filter under consideration, the unit sam ple resp onse is h (n) = a nu ( n ) (7.7.36)
Since the quantization error d u e to rounding has zero m ean, th e m ean value o f the error at the ou tp u t o f the filter is m q = 0. T h e variance o f the error at the output o f the filter is
00
*=
,2 1 a2
(7.7.37)
W e observe that the n oise p ow er a 2 at the output o f the filter is enhanced relative to the input n oise p ow er a 2 by the factor 1 / ( 1 a 2). T his factor increases as the p o le is m oved closer to th e unit circle. T o obtain a clearer picture o f the effect o f the quantization error, w e should also co n sid er the effect o f scaling th e input. L et us assum e that th e input seq u en ce (.t(n)} is a w hite n o ise sequ en ce (w ideband signal), w h ose am plitude has been scaled according to (7.7.14) to p revent overflow s in addition. T h en A , < 1 - |a| If w e assum e that x i n ) is uniform ly distributed in th e range ( - A * , A x), then, according to (7.7.31) and (7.7.34), the signal p ow er at the ou tp u t o f the filter
Sec. 7.7 is
Round-Off Effects in Digital Fitters
593
O O *
1 - a2
(7.7.38)
w h ere a 2 = (1 - |a |) 2/3 is the variance o f the input signal. T h e ratio o f the signal p ow er a 2 to the quantization error p ow er a 2, w hich is called the signal-to-noise ratio (S N R ), is sim ply
= d
<*} = ( 1 - | s |)2 2 2(6+l) T his exp ression for th e output S N R clearly illustrates the severe penalty paid as a con seq u en ce o f the scaling o f the input, expecially w h en the p ole is near the unit circle. B y com parison, if the input is not scaled and the adder has a sufficient num ber o f bits to avoid overflow , then th e signal am plitude m ay b e confined to the range ( 1 , 1 ). In this case, cr2 = which is in d ep en d en t o f the p ole position. T hen
2
(7.7.39)
% = 2 1(b+l)
(7.7.40)
T h e d ifferen ce b etw een th e S N R s in (7.7.40) and (7.7.39) clearly dem onstrates the need to use m ore bits in addition than in m ultiplication. T h e num ber o f additional bits d ep en d s on the p osition o f the p ole and sh ou ld b e increased as the p ole is m oved closer to the unit circle. N ex t, let us consider a tw o-p ole filter w ith infinite p recision which is described by the linear d ifferen ce equation
3>(n) = a i y ( n - 1) + a 2y ( n - 2) + .x(n)
(7.7.41)
w here a\ = 2 r co s 6 and a2 = r 2. W h en the tw o products are rounded, w e have a system which is d escribed by th e non linear differen ce eq u ation u(n) = Q r[a\v{n - 1)] + Q r[a2v(n - 2)] + x (n ) (7.7.42)
This system is illustrated in block diagram form in Fig. 7.50. N o w there are tw o m ultiplications, and h en ce tw o quantization errors are produced for ea ch output. C on sequ en tly, w e sh ou ld introduce tw o n o ise seq u en ces e\ (n) and e2(n), w hich correspond to the quantizer outputs Q r[aiv(n - 1)] = a i v ( n - 1) + *i(n ) (7.7.43) Q r[a2v(n - 2 )] = a2v(n ~ 2 ) + e2(n )
594
Chap. 7
Figure 7.50 Tw o-pole digital filter with rounding quantizers.
A block diagram for the corresponding m od el is show n in Fig. 7.51. N o te that the error sequ en ces e\ (n) and e 2 (n) can b e m oved directly to the input o f the filter. A s in the case o f the first-order filter, the output o f the secon d -ord er filter can be separated in to tw o com p onents, the desired signal com p on en t and the quantization error com p onent. T he form er is described by the differen ce equation y(n) = a t y ( n - 1) + a 2 y(n 2) + x ( n ) w hile the latter satisfies the d ifferen ce eq u ation q ( n ) = a \ q ( n - l ) + a 2 q (n - 2) + e i ( n ) + e 2 (n) (7.7.45) (7.7.44)
It is reasonable to assum e that the tw o seq u en ces e\( n ) and e 2 (n) are uncorrelated. N o w the secon d-ord er filter has a unit sam ple resp onse h(n) = sin(/t + sin 6 H en ce ^
X j
1
) 9 u (n)
(7.7.46)
,
~
+ r2
1 _ r 2 r 4 + J __ 2 r 2 c o s 26
(7.7.47)
j(n)
O'
<l(n)
u(n)
O'0e2(n)
Fijjnre 7.51 A dditive noise model for the quantization errors in a two-pole filter realization.
Sec. 7.7
595
B y applying (7.7.34), w e ob tain the variance o f the quantization errors at the output o f the filter in th e form
k2
(7.7.48) In the case o f the signal com p on en t, if w e scale the input as in (7.7.14) to avoid overflow , the p o w er in the output signal is al = o2 l ' Y J h 2 {n) where the p ow er in the input signal
jc()
(7.7.49)
is given by the variance

1
(7.7.50)
X > ( )l C on sequ en tly, th e S N R at the output o f the tw o-p ole filter is
_\,
0
a o:
X
(^+1 ) (7.7.51]
cr,:
> ( n ) l
A lth ou gh it is difficult to determ in e the exact value o f the den om in ator term in (7.7.51), it is easy to obtain an upper and a low er bound. In particular, |/7 (n)| is upper b ounded as \H n )\ < - J r n sm 6 so that
OO 4 OC -I
n>
(7.7.52)
A; n=o
|/i(n)j < - Y V = --------1
- s in 0
1 - r ) s in e (1
(7.7.53)
T h e low er bound m ay be ob tain ed by notin g that I //M l = B ut

1
y^h(n)e
jam
< X
> ( ) 1
H(a>) =
(1
- r e j s e-j>)( 1 - r e ~ j ee ~ im)
A t a> = $, w hich is th e reson ant freq u en cy o f the filter, w e ob tain the largest value o f \HU d )\. H en ce
1
(7.7.54)
(1 r) V 1 + r 2 2r cos 26
596
Chap. 7
T h erefore, the S N R is b ounded from a b ove and b elo w according to the relation
2 2 2
>+i)(i _
r ) 2 s in 2
<
2< 6+1)( l - r ) 2( 1 + r 2 -
r c o s 2 9)
( 7 . 7 .5 5 )
F or exam ple, w hen 9 = n [ 2, the exp ression in (7.7.55) red u ces to

2 22
+i>( i _
r ) 2
< f l < 22< i,+1>(l - r )2( l + r ) 2
(7.7.56)
T h e dom inant term in this bound is (1 r ) 2 which acts to reduce the S N R dram atically as the p o les m ove tow ard the unit circle. H en ce the effect o f scaling in the secon d-ord er filter is m ore severe than in the sin gle-p ole filter. N o te that if d = 1 - r is the distance o f the p ole from the unit circle, the S N R in (7.7.56) is reduced by d 2, w hereas in the sin gle-p ole filter the reduction is proportional to d. T h ese results serve to reinforce the earlier statem ent regarding the use of m ore bits in addition than in m ultiplication as a m echanism for avoiding the severe penalty due to scaling. T he analysis o f the quantization effects in a secon d-ord er filter can be ap plied directly to higher-order filters b ased on a parallel realization. In this case each secon d-ord er filter section is in d ep en d en t o f all the other section s, and there fore the total quantization n oise p ow er at the output o f the p arallel bank is sim ply the linear sum o f the quantization n oise p ow ers o f each o f th e individual sections. On the other hand, the cascade realization is m ore difficult to analyze. For the cascade interconnection, the n oise gen erated in any secon d-ord er filter section is filtered by the succeeding section s. A s a co n seq u en ce, there is the issue o f how to pair togeth er real-valued p oles to form secon d -ord er section s and h ow to arrange the resulting secon d-ord er filters to m inim ize the total n oise p ow er at the output o f the high-order filter. T h is gen eral top ic was investigated by Jackson (1970a, b), w h o sh ow ed that p o les close to the unit circle sh ou ld be paired w ith nearby zeros to reduce the gain o f each secon d-ord er section . In ordering the secon d-ord er sectio n s in cascade, a reasonable strategy is to place the section s in the order o f decreasin g m axim um frequency gain. In this case the n oise p o w er gen erated in the early high-gain section is not b o o sted significantly by the latter sections. T h e follow in g exam p le illustrates th e p oin t that proper ordering o f section s in a cascade realization is im portant in con trollin g th e rou n d -off n o ise at th e output o f th e overall filter.
Example 7.7.1
Determine the variance of the round-off noise at the output of the two cascade realizations of the filter with system function
H{z) = Hi (z)H2(z)
where
1
Sec. 7.7
597
m z) =
i- iz -
SolntioD Let h(n), hj(n), and h 2 (n) represent the unit sample responses correspond
ing to the system functions H(z), M ) = ( jV ^ n ) and H2 (z), respectively. It follows that M ) = (;)*<()
h(n) = [2 ( 1 )" - (i)"](n) The two cascade realizations are shown in Fig. 7.52. In the first cascade realization, the variance of the output is O O 00 V ( n) + 5 2 ^ (n) B = e O nxO
In the second cascade realization, the variance of the output noise is O O O O *i<) 5 2 h 2 (n) + ^
xM
>
vO O
-O
e\(")
K *)
e 2(n)
(a) Cascade realization I
( b ) Cascade r e a l i z a t i o nI I Figwe 7.52 Two cascade r e a l i z a t i o n si n Example 7 . 8 . 1 :( a ) cascade r e a l i z a t i o nI ; (b ) cascade r e a l i z a t i o nI I .
598
Now
Chap. 7
1
X > i< " ) = R bO >
2
l-l
1
1
3
16
(") =
- 15
4 4 1
Ys V ( n ) = i _ i r - T + i _ j. = 1-83 i i_ i
Therefore,
= 2.90er?
a2 2 = 3. 16c t ; and the ratio of noise variances i s

a\
= 1.09
Consequently, the noise power i n the second cascade realization i s 9% larger than the f i r s t realization.

From the treatm ent in this chapter w e have seen that there are various realizations o f discrete-tim e system s. FIR system s can be realized in a direct form, a cascade form , a frequency sam pling form, and a lattice form. IIR system s can also be realized in a direct form , a cascade form , a lattice or a lattice-lad d er form, and in a parallel form. For any given system described by a linear con stan t-coefficien t differen ce equation, these realizations are eq u ivalen t in that they represent the sam e system and produce the sam e output for any given input, provided that the internal com putations are perform ed w ith infinite precision. H ow ever, th e various structures are n o t equivalent w hen they are realized with finite- p recision arithm etic. T h e state-sp ace form ulation provides an internal d escrip tion o f a system and, as a con seq u en ce, w e o b tain ed additional system realizations, called state-space re alizations. T h ese realizations represent additional p ossib le structures that provide good alternative candidate realizations for th e system . T hree im portant factors are p resen ted for ch oosin g am on g the various FIR and IIR system realizations. T h ese factors are com p utational com p lexity, m em ory requirem ents, and finite-w ord-length effects. D ep en d in g o n eith er th e tim ed om ain or the frequency-dom ain characteristics o f a system , so m e structures m ay require less com p utation and/or less m em ory than others. H e n c e our selection m ust co n sid er th e se tw o im portant factors. M uch research has b een d on e o v er th e p ast tw o d ecad es o n state-sp ace rep resentation and realization o f system s. For referen ce, w e cite the b ook s by Chen
Sec. 7.8
599
(1970), D e R u sso et al. (1965), Z ad eh and D eso er (1963), and G u p ta (1966). T h e use o f state-sp ace filter structures in the realization o f IIR system s has b een p ro p o sed by M ullis and R ob erts (1976a,b), and further d ev elo p ed by H w an g (1977), Jackson et al. (1979), Jackson (1979), M ills et al. (1981), and B om ar (1985). In deriving the transposed structures in S ection 7.3, w e introduced several con cep ts and op eration s on signal flow graphs. Signal flow graphs are treated in d ep th in th e b ook s by M ason and Zim m erm an (1960) and C how and Cassignol (1962). A n o th e r im portant structure for IIR system s, a w a v e digital filter, has b een in vestigated by F ettw eis (1971) and further d ev elo p ed by S ed lm eyer and F ettw eis (1973). A treatm ent o f this filter structure can also b e found in the b ook by A n to n io u (1979). F in ite-w ord -len gth effects are an im portant factor in th e im plem en tation o f digital signal processing system s. In this chapter w e described the effects o f a finite w ord length in digital filtering. In particular, w e con sid ered the follow in g problem s dealin g with finite-w ord length effects: 1. P aram eter quantization in digital filters 2. R o u n d -o ff n oise in m ultiplication 3. O verflow in addition 4. Lim it cycles T h ese four effects are internal to the filter and influence th e m eth od by w hich the system w ill be im plem en ted . In particular, w e d em onstrated that high-order system s, especially IIR system s, should be realized by using secon d-ord er section s as building blocks. W e advocated the use o f the direct form II realization, eith er th e co n v en tio n a l or th e transposed form. E ffects o f rou n d -off errors in fixed-point im plem en tation s o f F IR and IIR filter structures have b een in vestigated by m any researchers. W e cite the papers by G old and R ad er (1966), R ader and G old (1967b), Jackson (1970a,b), Liu (1971), C han and R ab in er (1973a,b,c), and O p p en heim and W ein stein (1972). A s an alternative to the use o f direct form II secon d-ord er filters as building b lock s for high-order filters, w e can use secon d-ord er state-variable form s. Such state-variable form s can b e op tim ized with respect to th e state transition m atrix to m inim ize rou n d -off errors. T he op tim ization lead s to m inim u m -rou n d -off-n oise secon d -ord er state-variable filters that are highly robust for im plem en ting both narrow band and w ideband filters. F or a treatm ent o f m inim um -round-off-noise secon d -ord er state-sp ace real izations, th e reader can refer to the papers o f M ullis and R ob erts (1976a,b), H w an g (1977), Jackson et al. (1979), M ills et al. (1981), B om ar (1985), and the b o o k by R ob erts and M ullis (1987). L im it-cycle o scillation s occur in IIR filters as a result o f quantization effects in fixed-point m ultiplication and rounding. Investigation o f lim it cycles in d igital filtering and their characteristic behavior is treated in the p apers by Parker and
600
Chap. 7
H ess (1971), B rubaker and G ow d y (1972), Sandberg and K aiser (1972), and Jack son (1969, 1979). T he latter paper deals with lim it cycles in state-sp ace structures. M eth od s have also b een d evised to elim in ate lim it cycles cau sed by round-off er rors. For exam ple, the papers by B arnes and Fam (1977), F am and B arnes (1979), C hang (1981), B utterw eck et al. (1984), and A u e r (1987) discuss this problem . O verflow oscillation s have b een treated in the pap er by E bert et al. (1969). T he effects o f param eter quantization has b een treated in a num ber o f papers. W e cite for reference the work o f R ad er and G old (1967b), K n ow les and O lcayto (1968), A v en h a u s and S chuessler (1970), H errm ann and Sch u essler (1970b), Chan and R abiner (1973c), and Jackson (1976). Finally, w e m en tion that the lattice and lattice-ladder filter structures are know n to b e robust in fixed-point im plem en tation s. For a treatm ent o f th ese types o f filters, the reader is referred to the p apers o f G ray and M arkel (1973), M akhoui (1978), and M orf et al. (1977) and to the b ook by M arkel and Gray (1976).
PROBLEMS
7.1 Determine a direct form realization for the following linear phase f i l t e r s . (a) h(n) = {1,2,3,4,3,2,1)
t
(b) h(n) = { 1,2 ,3,3,2 ,1}

t
12. Consider an FIR f i l t e r with system function H(z) = 1 + 2.88Z"1 + 3.4048Z"2 + 1.74z"3 + 0.4z~4 Sketch the direct form and l a t t i c e realizations of the f i l t e rand determine in detail the corresponding input-output equations. Is the system minimum phase? 73 Determine the system function and the impulse response of the system shown i n Fig. P7.3.
3
Figwe P7J
1A Determine the system function and the impulse response of the system shown in Fig. P7.4. 7.5 Determine the transposed structure of the systems in Fig. P7.4 and verify that both the original and the transposed system have the same system function.
Chap. 7
Problems
601
Figure P7.4
7.6 Determine a , , a2, and f i and c0 in terms of b\ and b2 so that the two systems i n F ig . P7.6 are equivalent. 7.7 Consider the f i l t e r shown i nF ig. P7.7. (a) Determine i t s system function. (b) Sketch the pole-zero plot and check for s ta bi li ty i f
(3) = ^2 = 1, (2) b{) = > 2 = 1,
b\ = 2,
b] = 2,
! = 1.5, fl2 = -0 .9 flj = 1,^2 = 2
( c ) Determine the response to x(n) = cos(;rn/3) i fbo = 1 , b\ = > 2= 0 , a, = 1 , and 02 = 0 .9 9 . 7.8 Consider an LTI system, i n i t i a l l ya tr e s t , described by the difference equation
y(n) = \y(n - 2) + x(n)
(a ) Determine the impulse response, h(n), of the system. (b) What i s the response of the system to the input signal
*() = [(j)" + (-f)" ] (n )
( c ) Determine the direct form I I ,parallel-form, and cascade-form realizations for t h i s system. <d) Sketch roughly the magnitude response |//(w)| of t h i s system. 7.9 Obtain the direct form I , direct form I I , cascade, and parallel structures for the fol lowing systems.
(a) y(n) = |y (n - 1) - g>>(n - 2) + x(n) + ijc(n - 1)
(b) y(n) = - O .ly O i - 1 ) + 0.72y(n - 2) + 0.7z(n) - 0.252*(n - 2) (c) y(n) = -0.1;y(n -1)4- 0.2y( - 2) + 3x(n) + 3.6x(n - 1) + 0 . 6 j t( n- 2 )
(d ) H( z ) =
2(1 z-1)(l + -J2z~x + z~2) (1 + O-Sz-Od - 0.9t-1 + 0.81z-2)
602
Chap. 7
Figure P7.6 *()
(e) y(n) = \y(n - 1)+ iy(n - 2)+ x(n) + x{n - 1) (f) y(n) = y(n - l ) - \ y ( n - 2 ) + x ( n ) - x ( n - 1) + x(n - 2) Which of the systems above are stable? 7.10 Show that the systems in Fig. P7.10 are equivalent. 7.11 Determine a l l the FIR f i l t e r s which are specified by the l a t t i c e parameters Ki = K2 = 0 . 6 , K3 = -0.7, and KA= i . 7.12 Determine the set of difference equations for describing a realization of an IIR s y s tem based on the use of the transposed direct form II structure for the second-order subsystems.
Chap. 7
Problems
603
Figure P7.10
7.13* Write a program that implements a parallel-form realization based on transposed
direct form II second-order modules.
7.14* Write a program that implements a cascade-form realization based on regular direct
form II second-order modules.
7.15 Determine the parameters {

described by the system function
of the lattice filter corresponding to the FIR filter
ff(z) = A 2(z ) = 1 + 2 :7 1 + z~2
7.16 (a) Determine the zeros and sketch the zero pattern for the FIR lattice filter with
parameters
Ki = \,
( b )
K2 = - l
= -1 .
*3 = 1
The same as in part (a) but with
(c) You should have found that all the zeros lie exactly on the unit d rd e. Can this
( d )
result be generalized? How? Sketch the phase response of the filters in parts (a) and (b). What did you notice? Can this result be generalized? How?
7.17 Consider an FIR lattice filter with coefficients K i = 0.65, K 2 = -0 .3 4 , and K 3 = 0.8.
(a) Find its impulse response by tracing a unit impulse input through the lattice structure. ( b ) Draw the equivalent direct-form structure.
604
Chap. 7
7.18 Consider a causal IIR system with system function 1 + 2z-1 + 3z ~2 + 2z ~ 3 1 + 0.9z-1 - 0.8z-2 + 0.5z-3
(a) Determine the equivalent lattice-ladder structure. (b) Check if the system is stable.
7.19 Determine the input-output relationship, the system function, and plot the pole-zero pattern for the discrete-time system shown in Fig. P7.19. rcos $
v(n)
rcos 8
Figure P7.19
7.20 Determine the coupled-form state-space realization for the digital resonator
1
H(z) =
1
- (2 r cos o)z - 1 + r 2 z ~2
7.21 (a) Determine the impulse response of an FIR lattice filter with parameters Aj = 0.6, Ki = 0.3, = 0.5, and KA = 0.9. (b) Sketch the direct form and lattice all-zero and all-pole filters specified by the ^-parameters given in part (a). 7.22 (a) Sketch the lattice realization for the resonator H(z) =
1 1
- (2 r cos a> o )z-:i + r 2 z ~2
(b) What happens if r = 1? 7.23 Sketch the lattice-ladder structure for the system 1 - 0.8z _ 1 + 0.15z-2 1 + O.lz"1 - 0.72z~2 7.24 Determine a state-space model and the corresponding realization for the following FIR system:
Chap. 7
Problems
605
7.25 Determine the state-space model for the system described by
y(n)= y(n - 1)+ O.ll^n - 2)+ x(n) and sketch the type 1 and type 2 state-space realizations. 1J26 Determine the type 1 and type 2 state-space realizations for the Fibonacci system and i t s diagonal form. 7.27 By means of the z-transform, determine the impulse response of the system described by the state-space parameters ro 0.111
i J' H
ro.11-1
i J-
roi
=LiJ-
. .
<i=1
F = Li
7.28 Determine the characteristic polynomial of the coupled-form state-space structure described by (7.4.68) and solve for the roots. 729 Determine the transpose structure for the coupled-form state-space structure shown in Fig. 7. 31 . 730 Consider a pole-zero system with system function
(1
(1
- O.Se^'V'Xl - 0.5 e - J*/4 z~l ) - 0 .8 e J ^ z ~ l ) a - 0 . 8
(a) Sketch the regular and transpose direct form I I realizations of the system. (b) Determine and sketch the type 1 and type 2 state-space realizations. ( c ) Determine the impulse response of the system by inverting H(z) and by using state-space techniques. (d) Determine the coupled-form state-space realization. (e) Repeat parts (a) through (d) for the system obtained by changing the angle of the poles from j t/ 3 to n/4. 731 (a) Determine a parallel and a cascade realization of the system
(1
- z - ^ l - O -^ ^ Z -O d - O .& r-^ Z -1)
(b) Determine the type 1 and type 2 state-space descriptions of the system in part (a).
7 3 2 Show how to use a lattice structure to implement the following all-pass filter
0.5 + 0.2z _1 - 0.6z - 2 + z - 3 1 - 0.6z- + 0.2z - 2 + 0.5z - 3 Is the system stable?

7 3 3 Consider a system described by the following state-space equations:
T<" + 1 ) = [-0 8 1 y(n) = [-1 .8 1
! ] ' <") + [?]*<"> l]v (n ) + * (n )
(a) Determine the characteristic polynomial and the eigenvalues of the system. (b) Determine the state transition matrix $ (n ) for n > 0. (c) Determine the system function and the impulse response of the system. ( d ) Compute the step response of the system if v(0) = [0 1 ]'. (e) Sketch a state-space re alizatio n for the system.
606
Chap. 7
7 3 4 Repeat Problem 7.33 if the system is described by the state-space equations v(n + 1 ) = y(n) =
[1 0
^ J v(n) +
] v(n)
j x{n)
7 3 5 Repeat Problem 7.33 for the system described by the state-space equations
, ( " + 1 ) - [ c u
y (n )
] * (") + [ o ] - ' (")
= [1
1 ] v (n ) + x ( n )
7 3 6 Consider the system

y ( n ) = 0 .9 y (n -
1) - 0 .0 8 y (n 2 ) + jc(r) + x ( r t - 1)
(a) Determine the type 1 and type 2 state-space realizations of the system. (b) Determine the parallel and cascade state-space realizations of the system. (c) Determine the impulse response of the system by at least two different methods,
7 3 7 Consider the causal system
y(n) = J:Kn - 1) - gyfn - 2)+ x(n) +
- 1)
(a) Determine its system function. (b) Determine the type 1 state-space model. (c) Determine the state transition matrix &(n) = F , for any n, using z-transform techniques. (d) Determine the system function using the formula
tf<2) = g'UI - F)_1q + < *

Compare the answer with that in part (a). (e) Compute the characteristic polynomial det(zl F) and check if the system is stable.
7 3 8 Determine the impulse response of the system
- n - .-[? ]
using the z-transform approach.
7 3 9 A discrete-time system is described by the following state-space model:
v(/i + 1) = Fv(n) + qx(n)
y{n) = g'v(n) +dx(n )

where
F = [-0 ji i] -[!] -[?] -
(a) Sketch the corresponding state-space structure. (b) Calculate the impulse response for n = 0, 1 , . . . , 5 and for n == 17 by using the state-space approach (c) Find the difference equation description of the system. (d) Repeat part (b) by using the difference equation. (e) Sketch the direct form II implementation o f the system.
Chap. 7
Problems
607
7.40 Determine the state-space parameters F, q, g, and d for:

(a) the all-zero lattice structure
(b) the all-pole lattice structure

7.41 The generic floating-point format for a DSP microprocessor is the following:
| sign bit
exponent
The value of the number X is given by Ol.M x 2 10.M x 2 e if 5 = 0 if 5 = 1
0 if is the most negative twos-complement value Determine the range of positive and negative numbers for the following two formats:
I
E E
15
12
11 S
10 M 22 M
0 short format 0 single-precision format
()
31
24
23 S
(b)
7.42 Consider the IIR recursive filter shown in Fig. P7.42 and let h F(n), h K(n), and h(n)
denote the impulse responses of the FIR section, the recursive section, and the overall filter, respectively, (a) Find all the causal and stable recursive second-order sections with integer coef ficients (a x, a2) and determine and sketch their impulse responses and frequency responses. These filters do not require complicated multiplications or quantiza tion after multiplications.
Figure P7.42
608
Chap. 7
(b) Show that three of the sections obtained in part (a) can be obtained by intercon nection of other sections. (c ) Find a difference equation that describes the impulse response h{n) of the f i l t e r and determine the conditions for the overall f i l t e r to be FIR. (d) Rederive the results in parts (a) to (c ) using z-domain considerations. 7.43 This problem il lu st ra te s the development of d ig ital f i l t e r structures using Horner s rule for polynomial evaluation. To this end consider the polynomial p(x) = atpXp + ap- 1*'-1 I --- h aix + ao which computes p(x) with the minimum cost of p multiplications and p additions. (a) Draw the structures corresponding to the factorizations
Hr(z) = b0(l + b i z - l (l + b2 z - l a + b i z ^ ) ) ) H(z) = bv(z ~ 3 + (] z ~ 2 + (b2 Z~ 2 +!%)))
and determine the system function, number of delay elements, and arithmetic operations for each structure (b) Draw the Horner structure for the following linear-phase system:
3
H( z ) = Z ~ l
*o +
7.44 Let x\ and xi be (b + l) -bit binary numbers with magnitude l es s than 1 . To compute the sum of x i and x2 using two s-complement representation we treat them as (b + 1) bi t unsigned numbers, we perform addition modulo-2 and ignore any carry after the sign b i t . (a) Show that i f the sum of two numbers with the same sign has the opposite s i g n , t h i s corresponds to overflow. (b) Show that when we compute the sum of several numbers using two s-complement representation, the result will be correct, even i fthere are overflows, i fthe correct sum i sl e ss than 1 in magnitude. Il lu strate this argument by constructing a simple example with three numbers. 7.45 Consider the system described by the difference equation y(n) = ay(n - 1) - ax(n) I -x(n 1) (a) Show that i ti sa l l p a s s . (b) Obtain the direct form I I realization of the system ( c )I f you quantize the coefficients of the system in part ( b ) ,i si ts t i l la l l - pass? (d) Obtain a realization by rewriting the difference equation as y(n) = a[y{n - 1)- x(n)] + x(n - 1) (e) I f you quantize the coefficients of the system in part ( d) ,i si ts t i l l all-pass? 7.46 Consider the system
y(n) = \y ( n - 1) + *()
(a) Compute i t s response to the input x(n) = (j)"n(n) assuming infinite-predsion arithmetic.
Chap. 7
Problems
609
(b) Compute the response of the system y(n), 0 < n < 5 to the same input, assuming finite-precision sign-and-magnitude fractional arithmetic with five bits (i.e., the sign bit plus four fractional bits). The quantization is performed by truncation. (c) Compare the results obtained in parts (a) and (b).
7 .4 7 The input to the system
y(n) = 0.999y(n - 1) + x(n)

is quantized to b = 8 bits. What is the power produced by the quantization noise at the output of the filter?
7 .4 8 Consider the system
y(n) = 0.875y(n 1) 0.125;y(n 2) -I- x(n) (a) Compute its poles and design the cascade realization of the system.
(b) Quantize the coefficients of the system using truncation, maintaining a sign bit plus three other bits. Determine the poles of the resulting system. (c) Repeat part (b) for the same precision using rounding. (d) Compare the poles obtained in parts (b) and (c) with those in part (a). Which realization is better? Sketch the frequency responses of the systems in parts (a), (b), and (c). 7 .4 9 Consider the system
(a) Draw all possible realizations of the system. (b) Suppose that we implement the filter with fixed-point sign-and- magnitude frac
tional arithmetic using (b + 1) bits (one bit is used for the sign). Each resulting product is rounded into b bits. Determine the variance of the round-off noise created by the multipliers at the output of each one of the realizations in part (a).
7 .5 0 The first-order filter shown in Fig. P7.50 is implemented in four-bit (including sign)
fixed-point twos-complement fractional arithmetic. Products are rounded to four-bit representation. Using the input x(n) = 0.105(n), determine:
(a) The first five outputs if a = 0.5. D oes the filter go into a limit cycle? (b) The first five outputs if a = 0.75. Does the filter go into a limit cycle?
7 .5 1 The digital system shown in Fig. P7.51 uses a six-bit (including sign) fixed-point twos-
complement A /D converter with rounding, and the filter H(z) is implemented us ing eight-bit (including sign) fixed-point twos-complement fractional arithmetic with rounding. The input x(t) is a zero-mean uniformly distributed random process hav ing autocorrelation yZJ(t) = 3(r). Assume that the A /D converter can handle input values up to 1 . 0 without overflow.
610
Chap. 7
Figure P7.51
(a) What value of attenuation should be applied prior to the A /D converter to assure
that it does not overflow? (b) With the attenuation above, what is the signal-to-quantization noise ratio (SNR) at the A /D converter output? (c) The six-bit A /D samples can be left-justified, right-justified, or centered in the eight-bit word used as the input to the digital filter. What is the correct strategy to use for maximum SNR at the filter output without overflow? (d) What is the SNR at the output of the filter due to all quantization noise sources?
7 .5 2 Shown in Fig. P7.52 is the coupled-form implementation of a two-pole filter with
poles at x = reje. There are four real multiplications per output point. Let c, (n), i = 1, 2, 3, 4 represent the round-off noise in a fixed-point implementation of the filter. Assume that the noise sources are zero-mean mutually uncorretated stationary white noise sequences. For each n the probability density function p(e) is uniform in the range - A f l < e < A /2 , where A 2~h. (a) Write the two coupled difference equations for y(/i) and t>(n), including the noise sources and the input sequence x(rt). (b) From these two difference equations, show that the filter system functions H\(z) and H2 (z) between the input noise terms ej(n) + e2 (n) and e3 (n) + et (n) and the output y(n) are:
Hiiz)
1
rsin<?z-1 ~lrCO& 6 z~v + r 2 z ~2

1 1
m t) We know that
r cos 6 z~l r c o s# z - 1 + r 2 z ~2
H(z) = - --------~ 2 j => Hri) = 1 - 2 TCOS0 Z - 1 + r 2 z ~2 Determine hi(n) and h%(n).
sin(n + l)0u(/i) sin# (*)>
(c ) Determine a closed-form expression for the variance of the total noise from
i = 1, 2, 3, 4 at the output of the filter.
Chap. 7
Problems
611
7.53 Determine the variance of the round-off noise at the output of the two cascade real
izations of the filter shown in Fig. P7.53, with system function H{z) = where " =
H 2(z ) =
7.54 Quantization effects in direct-form FIR filters
Consider a direct-form realization of an FIR filter of length Af. Suppose that the multiplication of each coefficient with the corresponding signal sample is performed in fixed-point arithmetic with b bits and each product is rounded to b bits. Determine the variance of the quantization noise at the output of the filter by using a statistical characterization of the round-off noise as in Section 7.7.3.
7.55* Consider the system specified by the system function

*(*>
A(z) (1 - Q-8ej," V 1)(l L 1 (1 - h - W + 3Z-1 )
la
1r (1 + i z - H i - sz~l) 1 JL 2( l - 0 . ^ - > ) ( l - 0 . 8 e- > ^ - 1)J
(a) Choose Gi and G 2 so that the gain of each second-order section at to = 0 is equal to 1.
612 x (n)
Chap. 7
Xn)
k>-f
c,(n) *2 (n)
(a) Cascade realization I
(b) Cascade realization n
Figure P7.53
(b) Sketch the direct form 1 , direct form 2, and cascade realizations of the system.
(c) Write a program that implements the direct form 1 and direct form 2, and compute the first 1 0 0 samples of the impulse response and the step response of the system. (d) Plot the results in part (c) to illustrate the proper functioning of the programs. 736* Consider the system given in Problem 7.55 with Gi = G 2 = 1. (a) Determine a lattice realization for the system H(z) = B(z)
(b) Determine a lattice realization for the system

1
A{z) (c) Determine a lattice-ladder realization for the system H(z) = B(z)/A(z).
(d) Write a program for the implementation of the lattice-ladder structure in part (c).
(e) Determine and sketch the first 100 samples of the impulse responses of the sys tems in parts (a) through (c) by working with the lattice structures. (f) Compute and sketch the first 100 samples of the convolution of impulse responses in parts (a) and (b). What did you find? Explain your results.
Chap. 7
Problems
613
7.57* Consider the system given i n Problem 7. 55 . (a) Determine the parallel-form structure and write a program for i t simplementation. (b) Sketch a parallel structure using second-order coupled-form state-space sections. ( c ) Write a program for the implementation of the structure in part ( b ) . (d) Verify the programs in parts (a) and (c ) by computing and sketching the impulse response of the system.
Design of Digital Filters
W ith the background that w e have d evelop ed in the p reced in g chapters, w e are n ow in a position to treat the subject o f digital filter d esign . W e shall describe several m eth od s for designing F IR and IIR digital filters. In the design o f freq u en cy-selective filters, the desired filter characteristics are specified in th e frequency dom ain in term s o f the desired m agnitude and phase resp onse o f the filter. In the filter design p rocess, w e d eterm in e th e coefficients of a causal F IR or IIR filter that closely approxim ates the desired frequency response specifications. T h e issue o f which type o f filter to design, F IR or IIR , depends on the nature o f the p roblem and on the specifications o f th e desired frequency response. In practice, F IR filters are em p loyed in filtering p rob lem s w here there is a requirem ent for a linear-phase characteristic w ithin the passband o f the filter. If there is n o requirem ent for a linear-phase characteristic, eith er an IIR or an F IR filter m ay b e em p loyed . H ow ever, as a gen eral rule, an IIR filter has lower sid elo b es in the stopband than an F IR filter having th e sam e n u m b er o f parameters. F or this reason, if som e p h ase distortion is either tolerab le or unim portant, an IIR filter is preferable, prim arily because its im plem en tation in v o lv es few er parameters, requires less m em ory and has low er com putational com p lexity. In conjunction w ith our discussion o f d igital filter d esign , w e describe fre quency transform ations in b oth the analog and digital d om ain s for transforming a low pass p rototyp e filter in to either an oth er low p ass, bandpass, bandstop, or highpass filter. T oday, F IR and IIR digital filter design is greatly facilitated by the availability o f num erous com puter softw are program s. In describing th e various digital filter design m eth od s in this chapter, our primary objective is to give the reader the background n ecessary to select the filter that b est m atches th e application and satisfies the design requirem ents.
8.1 GENERAL CONSIDERATIONS

In Section 4 .5 , w e d escribed the characteristics o f ideal filters and d e m o n s t r a t e d that such filters are n ot causal and therefore, are n ot physically realizable. In t 5
614
Sec. B.1
General Considerations
615
section , the issu e o f causality and its im plications is con sid ered in m ore detail. F ollo w in g this discussion, w e presen t the frequency resp on se characteristics o f causal F IR and IIR digital filters.
8.1.1 Causality and Its Implications

Let us co n sid er the issue o f causality in m ore d etail by exam ining the im pulse resp o n se h ( n ) o f an ideal low pass filter with frequency resp onse characteristic H(a>) = T h e im pulse resp onse o f this filter is C t)c h(n) = n coc sin a> cn n 0
C 0c < O ) < 7T
( 8 . 1. 1)
(8.1.2) nÔ
A p lot o f h{n) for < oc = n / 4 is illustrated in Fig. 8.1. It is clear that the ideal low p ass filter is noncausal and h en ce it cannot be realized in practice. O ne p ossib le solu tion is to introduce a large delay no in h (n) and arbitrarily to set h in ) = 0 for n < no- H ow ever, the resulting system no longer has an id eal frequency resp onse characteristic. In d eed , if w e set h ( n ) 0 for n < n0, the F ourier series exp an sion o f H(to) results in the G ibbs p h en om en on , as will be described in Section 8.2. A lth o u g h this discussion is lim ited to the realization o f a low pass filter, our con clu sion s hold, in gen eral, for all th e other ideal filter characteristics. In brief, non e o f the id eal filter characteristics previously illustrated in Fig. 4.43 are causal, h en ce all are physically unrealizable. A q u estion that naturally arises at this p oin t is the follow ing: W hat are the necessary and sufficient con d ition s that a frequency resp onse characteristic H(co)
h (n )
Figure 8.1
Unit sample response of an ideal lowpass filter.
616
Design of Digital Fitters
Chap. 8
m ust satisfy in order for th e resulting filter to b e causal? T h e answ er to this question is given by the P a ley -W ien er theorem , w hich can b e stated as follows:
Paley-Wiener Theorem. If h(n) has finite energy and h( n ) = 0 for n < 0 , then [for a reference, se e W ien er and P aley (1934)]
f | I n \H(to)\\dco < oo J ir (8.1.3)
Conversely, if |//(a>)I is square integrable and if th e integral in (8.1.3) is finite, then w e can associate w ith \H{a>)\ a phase resp onse 0 (o>), so that the resulting filter w ith frequency response H{a>) = \H(a>)\em w ] is causal. O ne im portant con clu sion that w e draw from the P a le y -W ien er theorem is that the m agnitude function |//(&>) I can b e zero at som e freq u en cies, but it cannot be zero over any finite band o f freq u en cies, since the integral th en b ecom es infinite. C onsequently, any ideal filter is noncausal. A pparently, causality im poses so m e tight constraints on a linear tim einvariant system . In addition to the P a ley -W ien er con d ition , causality also im plies a strong relationship b etw een H r {cd) and H r (a>), the real and im aginary com p o nents o f the frequency resp onse H(o>). T o illustrate this d ep en d en ce, w e d ecom p o se h (n) into an even and an odd seq u en ce, that is, h(n) = h e(n) + h a(n) w here h e(n) = \ [ h ( n ) + h ( - n ) ] and h 0 (n) = \ [ h ( n ) - h { - n )] ( 8 . 1 .6 ) (8.1.5) (8.1.4)
N ow , if h ( n ) is causal, it is p ossib le to recover h(n) from its ev en part h e{n) for < n < oo or from its odd co m p on en t h 0 (n) for 1 < n < oo. Indeed, it can b e easily seen that h{n) = 2 h r(n)u (n) h e{0)S{n) n> 0 (8.1.7)
and h ( n ) = 2 h 0 (n )u (n ) + h (0)S (n) n > 1 (8.1.8)
Since h 0 {n) = 0 for n = 0, w e cannot recover ft(0) from h 0 (n) and h en ce w e also m ust know /i(0). In any case, it is apparent that h 0 (n) = h e(n) for n > 1, so there is a strong relationship b etw een h 0 (n) and h t (n). If h ( n ) is ab solu tely sum m able (i.e., B IB O stab le), th e frequency response H{co) exists, and
H(o>) = HR(a>) + j H , (co)
(8.1.9)
Sec. 8.1
617
In addition, if h (n) is real valued and causal, the sym m etry p rop erties o f the F ourier transform im ply that h e(n) h 0 (n) H R(a>)
( 8 . 1. 10)
Hi (to)
Since h(n) is co m p letely specified by h e(n), it follow s that H(to) is com p letely d eterm in ed if w e k n ow H R(to). A ltern atively, H(to) is com p letely determ in ed from H j(a > ) and fc(0). In short, H R(to) and Hi(to) are in terd ep en d en t and cannot b e specified in d ep en d en tly if the system is causal. E q u ivalen tly, the m agnitude and phase resp onses o f a causal filter are in terd ep en dent and h en ce cannot b e specified ind ep en dently. G iven H g(w) for a corresponding real, even , and ab solu tely sum m able s e q u en ce h e(rt), w e can determ in e H(to). T h e follow in g exam p le illustrates the p ro cedure. Example 8 .L 1 Consider a stable LTI system with real and even impulse response h(n). Determine H(w) if 1 a cos to HR(w) S B ----- ----------- - J < * j< 1 1 2 a cos w + a 1 Solution The first step is to determine he(n). This can be done by noting that
H r (w ) = H r ( z )
where H ( )=
RZ
1 ~ Q(z + _ z - a(z 2 + l ) / 2 1 - a(z + z_1) + a 2 (z - a ) ( l ~ a z )
The ROC has to be restricted by the poles at pi = a and pj = \ / a and should include the unit circle. Hence the ROC is (oj < |z| < l/|o |. Consequently, ht (n) is a twosided sequence, with the pole at z = a contributing to the causal part and p 2 = 1 /a contributing to the anticausal part. By using a partial-fraction expansion, we obtain
ht(n) =
k(n)
ifl1 1+
\& (h)
(8.1.11)
By substituting (8.1.11) into (8.1.7), we obtain h(n) as a"()
Finally, we obtain the Fourier transform of h(n) as

1 1
ae~lto
T h e relationship b etw een the real and im aginary com p on en ts o f th e Fourier transform o f an ab solu tely sum m able, causal, and real seq u en ce can be easily established from (8.1.7). T h e F ourier transform relationship for (8.1.7) is
H(a>) = H r (co) + j H ] ( t o ) = I H R ( k ) U ( t o - k ) d k ~ h e ( 0)
(8.1.12)
618
Chap. 8
w h ere U(u>) is th e Fourier transform o f the unit step seq u en ce u (n). A lth ou gh the unit step seq u en ce is n ot ab solu tely sum m able, it has a F ourier transform (see S ection 4.2.8). U(u)) = n&(a>)+
1 1
- e-J*
(8.1.13)
7TS(a))
r/ x + 1 1 " yr cot
J2
n < a> < n

~
B y substituting (8.1.13) in to (8.1.12) and carrying ou t the in tegration , w e obtain the relation b etw een H R(to) and Hj(a>) as H,(to) =
H (k) cot
(8.1.14)
T hus HiUo) is uniquely determ in ed from H r (uj) through this integral relationship. T h e integral is called a discrete H ilb ert tra nsform . It is left as an exercise to the reader to establish the relationship for H R(a>) in term s o f the discrete H ilbert transform o f T o sum m arize, causality has very im portant im plications in the design o f freq u en cy-selective filters. T h ese are: (a) th e frequency resp on se H(a>) cannot b e zero, excep t at a finite set o f p oin ts in frequency; (b) th e m agnitude }tf(ti>)| cannot be con stan t in any finite range o f freq u en cies and the transition from pass band to stopband cannot b e infinitely sharp [this is a co n seq u en ce o f the G ibbs p h en o m en o n , w hich results from the truncation o f h (n) to a ch iev e causality]; and (c) the real and im aginary parts o f H(a>) are in terd ep en dent and are related by the discrete H ilb ert transform . A s a con seq u en ce, the m agnitu d e \H(a>)\ and phase 0 (<y) o f H(a>) cannot be ch osen arbitrarily. N o w that w e k n ow th e restrictions that causality im p oses on the frequency resp onse characteristic and th e fact that ideal filters are not ach ievab le in practice, w e lim it our atten tion to the class o f linear tim e-invariant system s specified by the differen ce eq u ation
N U
3,(n ) = ~ H ak y(n - *) +
*= 1 *= 0
bkX<Jl ~ k )
w hich are causal and physically realizable. A s w e h ave d em onstrated , such system s have a frequency response
H(co) =
------------a ke -jwk
Jc=l
(8.1.15)
T h e basic digital filter d esign prob lem is to approxim ate any o f th e ideal frequency resp onse characteristics w ith a system that has the freq u en cy resp on se (8.1.15), by properly selectin g the coefficients {ak} and {bt}. T h e approxim ation problem is
Sec. 8.1
619
treated in d etail in Sections 8.2 and 8.3, w h ere w e discuss tech n iqu es for digital filter design.
8.1.2 Characteristics of Practical Frequency-Selective Filters

A s w e ob served from our discussion o f the preced in g section , ideal filters are n oncausal and h en ce physically unrealizable for real-tim e signal processing appli cations. Causality im plies that the frequency resp onse characteristic H ( a>) o f the filter cannot be zero, except at a finite set o f p oin ts in the frequency range. In addition, H ( a>) cannot have an infinitely sharp cu toff from passband to stopband, that is, H(oj) cannot drop from unity to zero abruptly. A lth o u g h the frequency resp onse characteristics p ossessed by ideal filters m ay b e desirable, they are not absolutely necessary in m ost practical applications. If w e relax th ese con d ition s, it is p ossib le to realize causal filters that approxim ate the ideal filters as closely as w e desire. In particular, it is n ot necessary to insist that the m agnitude \H{u>)\ b e constant in the entire passband o f th e filter. A sm all am ount o f ripple in the passband, as illustrated in Fig. 8.2, is usually tolerab le. Similarly, it is not necessary for the filter response (</>)|to be zero in the stopband. A sm all, n on zero value or a sm all am ount of ripple in the stopband is also tolerable. T he transition o f the frequency response from passband to stopband defines the transition b a n d or transition region o f the filter, as illustrated in Fig. 8.2. T he b an d -ed ge frequency wp defines the ed ge o f the passband, w hile the frequency ojs d en o tes the beginning o f the stopband. Thus the width o f the transition band is aij o)p . T h e w idth o f the passband is usually called the b a n d w i d t h o f the filter. For exam p le, if the filter is low pass with a passband ed ge frequency w p, its bandwidth is wp .
Figure &2
Magnitude characteristics of physically realizable filters.
620
Chap. 8
If there is ripple in th e passband o f the filter, its v alu e is d en o te d as 8 lt and the m agnitude |//(a ))| varies b etw een the lim its 1 i i . T h e ripple in the stopband o f the filter is d en o ted as S2. T o accom m odate a large dynam ic range in the graph o f the frequency re sp o n se o f any filter, it is com m on practice to use a logarithm ic scale for th e m ag nitude | H(a)) |. C on sequ en tly, the ripple in the passband is 2 01og 10 5] d ecib els, and that in the stopband is 2 0 1 o g 10 & 2In any filter design p roblem w e can sp ecify (1) the m axim um tolerab le pass band ripple, (2) the m axim um tolerab le stopband ripple, (3) the passband edge frequency a> p, and (4) th e stopband ed ge freq u en cy cos. B ased on th e se speci fications, w e can select the param eters {a*} and {>*} in the freq u en cy response characteristic, given by (8.1.15), w hich best ap proxim ates th e d esired specification. T he d egree to which H(co) approxim ates the sp ecification s d ep en d s in part on the criterion used in th e selection o f the filter coefficien ts {a*} and {&*) as w ell as on th e num bers ( M , N ) o f coefficients. In the follow in g section w e presen t a m ethod for designing linear-phase F IR filters.
8.2 DESIGN OF FIR FILTERS

In this section w e describe several m eth od s for design in g F IR filters. O ur treatm ent is fo cu sed o n the im portant class o f linear-phase F IR filters.
8.2.1 Symmetric and Antisymmetric FIR Filters

A n F IR filter o f length M with input x ( n ) and output y ( n ) is d escribed by the differen ce equation y ( n ) s= box(n) + b \ x ( n 1
) -I-------- H bM- \ x ( n - M +
)
(8.2.1)
w here {*} is the set o f filter coefficients. A ltern atively, w e can exp ress the output seq u en ce as the con volu tion o f the unit sam p le resp onse h (n) o f th e system with the input signal. T hus w e have y(n) = h (k)x(n-k)
(8 .2 .2 )
w here th e low er and upper lim its on the con volu tion sum reflect the causality and finite-duration characteristics o f the filter. Clearly, (8.2.1) and (8.2.2) are identical in form and h ence it fo llow s that i>* = h ( k ), k = 0 , 1 , . . . , M ~ 1 . T h e filter can a lso be characterized by its system function M1
(8.2.3)
Sec. 8.2
Design of FIR Filters
621
w hich w e view as a p olyn om ial o f d egree Af 1 in the variable z _1. T h e roots o f this p oly n o m ia l con stitu te the zeros o f the filter. A n F IR filter has linear p h ase if its unit sam ple resp onse satisfies the con d i tion h{n) = h { M - 1 - n) n = 0,1,..., M - 1 (8.2.4)
W hen th e sym m etry and antisym m etry conditions in (8.2.4) are incorporated into (8.2.3), w e have H ( z ) = /i(0) + M D z "
1
+ h ( 2 ) z ' z + --- + h ( M - 2 ) z ~(M~ 2) + h ( M - 1) M od d
(M / 2 ) - l
-m
Y ,
h ( n )[ ziM- l~ 2k)/2 z - ^ - 2k}/2]
M even
(8.2.5)
N o w , if w e sub stitu te z 1 for z in (8.2.3) and m ultiply both sides o f the resulting eq u ation by z_tw_1), w e obtain z- (w - i ) H ( z - i ) _ H { Z ) (82>6) T h is result im plies that the roots o f the polyn om ial H ( z ) are identical to the roots o f the p olyn om ial H { z - 1 ). C on sequ en tly, the roots o f H ( z ) m ust occur in reciprocal pairs. In other w ords, if zj is a root or a zero o f H ( z ), then \ / z \ is also a root. Furtherm ore, if the unit sam ple resp onse h(n) o f the filter is real, com p lex-valu ed roots m ust occur in com p lex-con ju gate pairs. H en ce, if zi is a com p lex-valu ed root, z* is a lso a root. A s a con sequ en ce o f (8.2.6), H ( z ) also has a zero at 1 /z j . Figure 8.3 illustrates the sym m etry that exists in the location of th e zeros o f a linear-phase F IR filter. T he frequency resp onse characteristics o f linear-phase F IR filters are o b tained by evaluating (8.2.5) on th e unit circle. This substitution yield s the exp res sion for H ( w ) .
Figure &3 Symmetry o f zero locations for a linear-phase FIR filter.
622
Chap. 8
W hen h i n ) = h ( M - 1 n), H{a>) can be expressed as H(to) = H r (co)ejm iM ~ 1)/2
(8.2.7)
w here H r(a>) is a real function o f co and can be expressed as
fu / __ i \
(M - 3 ) f2
__ i
^2 n= 0
^ ( n ) c o s a ) ^ -------- n j
M odd
( 8 .2 .8)
(Af/2)1
Hr (co) = 2
L7 ( M 1 \ ^ h (n) cosco ( -------- n j =o \ '
M even
(8.2.9)
T he phase characteristic o f the filter for both M odd and M ev en is

-to
i f t f r fcu) > 0
@(<o) =
to + 71 * i i H r ( a > ) < 0
(8.2.10)
W hen h(n) = h ( M 1
- n)
the unit sam ple resp onse is antisym m etric. For M odd, the cen ter p oin t o f the antisym m etric h(n) is n = ( M - l ) / 2 . C onsequently,
H ow ever, if M is ev en , each term in h( n ) has a m atching term o f o p p o site sign. It is straightforward to sh ow that the frequency response o f an F IR filter with an antisym m etric unit sam ple resp onse can be exp ressed as H(to) = Hr (co)e*-a> {M - ' )rz+1' f t where
( -3 )/2 / M - 1 \
(8.2.11)
Hr (co) 2
^ o ^ =o
A( n ) s i n a > ( -------- n l \ 2 / A ( n ) s i n o ; f -------- n J \ 2 /
M od d
(8.2.12)
H r(io) = 2
M ev en
(8.2.13)
The phase characteristic o f the filter for b oth Af od d and Af ev e n is n

2
( M - 1\ "_ < w ( 2 ) ~ )
'&Hr (a>)> H H r (co)
(o>) = T ~ w( ~
2
(8.2.14)
< 0
T h ese general freq u en cy resp onse form ulas can b e used to design linearphase F IR filters w ith sym m etric and antisym m etric unit sam ple resp onses. W e
Sec. 8.2
623
n o te that, for a sym m etric h ( n ), the num ber o f filter coefficien ts that specify th e frequency resp onse is (Af + l ) / 2 w hen M is odd or M / 2 w h en M is even . O n the other hand, if the unit sam ple resp onse is antisym m etric,
so that there are (Af l ) / 2 filter coefficients w hen M is odd and M / 2 coefficients w h en M is ev en to be specified. T he ch o ice o f a sym m etric or antisym m etric unit sam p le resp on se dep en d s on the application. A s w e shall se e later, a sym m etric unit sam ple resp onse is suitable for so m e applications, w h ile an antisym m etric unit sam p le resp onse is m ore su itab le for other applications. For exam p le, if h(n) = h ( M - 1 - ri) and M is odd, (8.2.12) im plies that Hr (0) = 0 and Hr ( n ) = 0. C on sequ en tly, (8.2.12) is n ot suitable as eith er a low pass filter or a highpass filter. Sim ilarly, the antisym m etric unit sam ple resp onse w ith M even also results in H r (0) = 0, as can b e easily verified from (8.2.13). C onsequently, w e w ould not use the antisym m etric con d ition in the design o f a low pass linear-phase F IR filter. O n the other hand, the sym m etry con d ition h(n) = h ( M 1 - n) yields a linear-phase F IR filter w ith a nonzero resp onse at to = 0 , if desired, that is, / Af 1 \ w -W tfr(0) = / i ( - 1 + 2 h{n), M odd (8.2.15)
2 /
n=0
Hr ( 0) = 2
J n= 0
h ( n ),
M even
(8.2.16)
In sum m ary, the problem o f F IR filter design is sim ply to d eterm in e the Af coefficients A( n) , n = 0 , l , . . . , A f 1 , from a specification o f the desired frequency resp onse Hj(to) o f th e F IR filter. T he im portant param eters in the specification o f Hd (co) are given in Fig. 8.2. In the follow in g subsections w e describe design m eth od s based on specifica tion o f Hjico).
8.2.2 Design of Linear-Phase FIR Filters Using Windows

In this m eth od w e b egin with the desired frequency resp onse specification Hj(to) and determ in e the corresponding unit sam ple resp onse h j ( n ) . In d eed , hd {n) is related to Ha(a>) by the F ourier transform relation
ff*(o>) = f > ( " ) * '"

ItsO
(8.2.17)
(8.2.18)
where hd {n) = / 2j t
1
f*
Hd {co)ejamd(o
T hus, given Hj((o), w e can d eterm in e th e unit sam p le resp on se h j ( n ) by evaluating the integral in (8 .2 .1 8 ).
624
Chap. 8
In general, the unit sam ple resp onse h d {n) ob tain ed from (8.2.17) is infinite in duration and m ust be truncated at som e p oin t, say at n = M 1 , to yield an F IR filter o f length Af. Truncation o f h d (n) to a length M 1 is equivalent to m ultiplying h d (n) by a rectangular w in d ow , defined as
" " - { i
othervrise " ^
< 8 '2 1 9 )
Thus the unit sam ple response o f the F IR filter b ecom es h{n) = h d (n )w in ) _ | h d in), {0, n =
0
, 1 ......... M
- 1
(8 .2 .2 0 )
otherw ise
It is instructive to con sid er the effect o f the w indow fun ction on the d e sired frequency response Hd ito). R ecall that m ultiplication o f th e w indow function w i n ) with h d (n) is equivalent to con volu tion o f H d (w) with Wi<o), w here W(<o) is the frequency-dom ain representation (Fourier transform ) o f th e w indow function, that is,
A f-1
W i w ) = Y i w ( n ) e - jwn n= 0
(8.2.21)
T hus the con volu tion o f H d i<o) with W i w ) yields the freq u en cy resp onse of the (truncated) F IR filter. That is,
1 f* H i w ) = / H d( v ) W ( w - v ) d v 2n J.
(8.2.22)
T h e Fourier transform o f the rectangular w indow is

Af-1
W (w ) = Y e ~ jan n* 0 =
1
(8.2.23) = sin(a>M/2 ) sin(a>/2 )
e~ iw
T his w indow function has a m agnitude response \W(<o)\ = and a p iecew ise linear phase
co
|sm (<u/ 2 )|
- 7t < w < 7t
(8.2.24)
^ ~ 2~ )
w h en sin(uAf/2) > 0 (8.2.25) w h en sm (tuAf/2) < 0
( oj) =
}M 1\ & I - I + n ,
T h e m agnitude resp onse o f th e w in d ow function is illustrated in F ig. 8.4 for M = 31 and 61. T he w idth o f the m ain lob e [width is m easured to th e first zero o f W(tu)]
Sec. 8.2
625
Figure 8.4 Frequency response for rectangular window of lengths (a) M = 31, (b) M = 61.
is 4j t / M . H en ce, as M increases, the m ain lob e b ecom es narrower. H ow ever, the sid elo b es o f |W(a>)| are relatively high and rem ain u n affected by an increase in M . In fact, even though the w idth o f each sid elob e d ecreases with an increase in M , the height o f each sid elo b e increases with an increase in M in such a m anner that the area under each sid elob e rem ains invariant to changes in M . T h is character istic b eh avior is not evid en t from observation o f Fig. 8.4 b ecau se W ( oj) has b een norm alized by M such that the norm alized peak valu es o f the sid elob es rem ain invariant to an increase in M . T h e characteristics o f the rectangular w in d ow play a significant role in d eter m ining the resulting frequency resp onse o f the F IR filter obtained by truncating hd (n) to length M . Specifically, th e con volu tion o f Hd {u)) with W(a)) has the effect o f sm ooth in g A s M is increased, W(a) b ecom es narrower, and the sm o o th ing provided by W(cv) is reduced. O n the other hand, th e large sid elob es o f W(a>) result in so m e undesirable ringing effects in th e F IR filter frequency resp onse H(co), and also in relatively larger sid elob es in H(a>). T h ese undesirable effects are b est alleviated by th e use o f w indow s that d o not con tain abrupt discon tin u ities in their tim e-d om ain characteristics, and h ave correspondingly low sid elob es in their frequency-dom ain characteristics. T ab le 8.1 lists several w in d ow functions that p ossess desirable frequency re sp on se characteristics. Figure 8.5 illustrates th e tim e-d om ain characteristics o f th e w indow s. T h e frequency resp onse characteristics o f the H anning, H am m ing, and B lackm an w in d ow s are illustrated in Figs. 8 . 6 through 8 .8 . A ll o f th ese w indow fun ction s h ave significantly low er sid elob es com pared w ith the rectangular w in dow . H o w ev er, for the sam e value o f M , the width o f the m ain lob e is also w ider for th e se w in d ow s com pared to the rectangular w indow . C on sequ en tly, these w in d ow fun ction s provide m ore sm ooth in g through the con volu tion operation in the frequency dom ain, and as a result, th e transition region in th e F IR filter resp onse is w ider. T o reduce the w idth o f this transition region, w e can sim ply increase the length o f th e w indow w hich results in a larger filter. T ab le 8.2 sum m arizes th ese im portant frequency-dom ain featu res o f th e various w in d ow functions. T he w in d ow tech n iqu e is b est described in term s o f a specific exam ple. Sup p o se that w e w ant to design a sym m etric low p ass linear-phase F IR filter having a
626 TABLE 8.1
Design of Digital Filters WINDOW FUNCTIONS FOR FIR FILTER DESIGN

Time-domain sequence,
h (n ),
Chap. 8
Name of window
0<
n < M
M -l
Bartlett (triangular) Blackman Hamming Hanning 0.42 0.5 cos
MM -l '
+ 0.08 cos ^
M -l
0.54 0.46 cos
2n n M -l
- 11 c o s ------- r 2 \ M-lJ
1 /.
2n n \
Kaiser
* [.(!? )]
Lanczos
L >0
, 1
1. ------- c
(8
.2 6 )
o t h e r w is e
A d elay o f ( M l ) / 2 units is incorporated in to Hd (w) in anticipation o f forcing the filter to be o f length M . T he corresp on d ing unit sam ple resp on se, ob tain ed by evaluating the integral in (8.2.18), is h d (n) = ^ - f e ^ - ^ d c o 2n J_U c
- _1
/ M - l\ n *
2
(*2.27,
2 )
Clearly, h d (n) is noncausal and infinite in duration.
Sec. 8.2
Design of FIR Fitters
627
l
0.8
------------------------------- 7--------/ / ^ \ \ Rectangular y Hamming
0.6
0.4
f t * Z Hanning
f t
// / // *.
Blackman
0.2
0
\ \\ ' *\ \ \\ \ W V <v.
M- 1
Figure 8J5
Shapes of several window
Magnitude (dB )
Figure 8.6 Frequency responses of Hanning window for (a) M 31 and (b) M = 61.
Magnitude (dB)
Figure 8.7 Frequency responses for Hamming window for (a) M = 31 an< (b) M = 61.
628
Chap. 8
Figure & 8 Frequency responses for Blackman window for (a) M = 31 and (b) M = 61. TABLE 8.2 IMPORTANT FREQUENCY-DOMAIN CHARACTERISTICS OF SOME WINDOW FUNCTIONS Approximate transition width of main lobe Ait/M &7T/M %ixjM &JI/M YI tz/ M
Type of window Rectangular Bartlett Hanning Hamming Blackman
Peak sidelobe (dB) -1 3 -2 7 -3 2 -4 3 -5 8
If w e m ultiply h j ( h ) by the rectangular w indow seq u en ce in (8.2.19), w e obtain an F IR filter o f length M having the unit sam ple resp onse sin a), h (n) =
0
< n < M 1
(8.2.28)
If M is selected to b e odd, the value o f h ( n ) at n = ( M Y)/2 is . ( ) , = (8.2.29)
T h e m agnitude o f the frequency resp onse H(o)) o f this filter is illustrated in Fig. 8.9 for M = 61 and M = 101. W e o b serve that relatively large oscilla tion s or ripples occur near the band ed ge o f th e filter. T h e oscillation s increase in freq u en cy as M increases, but they d o n ot dim inish in am plitude. A s indicated pre viously, these large oscillation s are the direct result o f the large sid elo b es existing in th e frequency characteristic W (a>) o f th e rectangular w in d ow . A s this window fun ction is co n v o lv ed with th e desired freq u en cy resp onse characteristic the oscillation s occur as the large con stan t area sid elob es o f W ( cd) m ove across th e discontinuity that exists in Since (8.2.17) is basically a F ourier series represen tation o f the m ultiplication o f h j ( n ) with a rectangular w in d ow is id en tical to truncating the Fourier series rep resen tation o f th e desired filter char-
Sec. 8.2
Design of FIR Fitters
629
(a)
0.8 n a 0.6 oc 2 0.4 0.2 o O1
M = 101
0.2
(b)
0.3
0.4
Normalized frequency
Figure 8.9 Lowpass filter designed with a rectangular window (a) M = 61 and (b) M = 101.
acteristic H j (co). T he truncation o f th e F ourier series is know n to introduce ripples in the frequency resp onse characteristic H(co) du e to th e nonuniform con vergen ce o f the F ourier series at a discontinuity. T h e oscillatory b eh avior near the band ed g e o f the filter is called the G i b b s p h e n o m e n o n . T o a lleviate the p resence o f large oscillation s in b oth the passband and th e stopband, w e should use a w in d ow function that contains a taper and decays to ward zero gradually, instead o f abruptly, as it occurs in a rectangular w indow . F igures 8.10 through 8.13, illustrate the frequency resp onse o f the resulting filter w h en so m e o f the w indow fun ction s listed in T ab le 8.1 are used to taper h d (n). A s illustrated in Figs. 8.10 through 8.13, the w in d ow functions d o indeed elim inate the ringing effects at the band ed ge and d o result in low er sid elo b es at the exp en se o f an increase in the w idth o f the transition band o f the filter.
Fignv 8.10 Lowpass FIR filter designed with rectangular window (M = 61).
630
Chap. 8
Figure 8.11 Lowpass FIR filter designed with Hamming window (Af = 61).
Figure 8-12 Lowpass FIR filter designed with Blackman window (Af = 61).
Figure 8.13 Lowpass FIR filter designed with a = 4 Kaiser wind* (Af = 61).
8.2.3 Design of Linear-Phase FIR Filters by the Frequency-Sampling Method

In the frequency sam pling m eth od for F IR filter design, w e specify the desired frequency resp onse Hj(co) at a set o f equally sp aced freq u en cies, nam ely Vk = tj-(* + ) * =
0
,1,
M -l
M
M od d (8.2.30)
and solve for the unit sample response h ( n ) of the FIR filter from these equally
Sec. 8.2
631
spaced frequency specifications. T o reduce sid elob es, it is desirable to optim ize the frequency specification in the transition band o f the filter. T his optim ization can be accom plished num erically on a digital com p uter by m eans o f linear program m ing tech n iqu es as show n by R abiner et al. (1970). In this section w e exploit a basic sym m etry property o f th e sam pled frequency resp onse function to sim plify the com putations. L et us begin with the desired frequency resp onse o f the F IR filter, w hich is [for sim plicity, w e drop the subscript in Hd (a>)], MH(a>) = =o Su p pose that w e specify the frequency resp on se o f the filter at the freq u en cies given by (8.2.30). T hen from (8.2.31) w e obtain H (k + a) = H ^ - i k + a y j M -1 Y h ( n ) e ~ i2n{k+a)n,M n= 0
1
h ( n )e ~ iwn
(8.2.31)
H (k + a) s
k = 0 , 1 , . . . . Af 1
(8.2.32)
It is a sim ple m atter to invert (8.2.32) and express h{n) in term s o f H ( k + a ) . If w e m ultiply both sid es of (8.2.32) by the exp on en tial, z x p ( j 2 x k m / M ) , m 0, 1 . . . . , M 1, and sum over k = 0, 1 , Af ~ 1, the right-hand side o f (8.2.32) reduces to A f h ( m ) e x p ( j 2 n a m / A f ) . Thus w e obtain i M -1 = 77 Y L H ( k + ct)ej2,T(k+a)nlM *=o n = 0,1,...,A/ 1 (8.2.33)
T h e relationship in (8.2.33) allow s us to com p ute the valu es o f th e unit sam p le resp onse h (n) from the specification o f the frequency sam ples H ( k + a ) , k = 0 , 1 . . . . , M 1. N o te that w hen a = 0, (8.2.32) reduces to the discrete F ourier transform (D F T ) o f the seq u en ce [h(n)} and (8.2.33) red u ces to the inverse D F T (ID F T ). Since (fc(/i)} is real, w e can easily show that the freq u en cy sam ples { H ( k + a ) } satisfy the sym m etry condition H (k + a) = H*(Af k a) (8.2.34)
T h is sym m etry co n d ition , along with the sym m etry con d ition s for \h (n )), can b e u sed to reduce the frequency specifications from M p oin ts to ( M + l ) / 2 p oin ts for M od d and M f l p oin ts for M even . T hus the linear eq u ation s for determ ining {h(n)} from [ H ( k + a )} are considerably sim plified. In particular, if (8.2.11) is sam pled at the freq u en cies = 2n ( k + a ) / M , k = 0, 1 , . . . , Af 1, w e obtain / 2 jt H ( k + a ) = Hr l (k + a ) J
(8.2.35)
632
Chap. 8
w here j3 = 0 w h en is sym m etric and p = 1 w hen {h(n)} is antisym m etric. A sim plication occurs by defining a set o f real frequency sam ples { G (k + m)} G ( k + a ) = ( ~ l ) kH r + k = 0 , 1 .........M - l (8.2.36)
W e u se (8.2.36) in (8.2.35) to elim in ate Hr (a>k). T h u s w e ob tain H (k + a) = G ik + (8.2.37)
N o w the sym m etry con d ition for H ( k + a ) given in (8.2.34) tran slates into a corre sponding sym m etry con d ition for G (k + a ), which can be ex p lo ited by substituting into (8.2.33), to sim plify th e expressions for the F IR filter im pu lse resp onse {h(n)) for the four cases ar = 0, a = j , = 0, and f$ = \ . T h e results are sum m arized in T able 8.3. T h e d etailed derivations are left as exercises for the reader. A lth o u g h the frequency sam pling m eth od provides us w ith another m ean s for designing linear-phase F IR filters, its major advantage lies in th e efficien t frequency sam pling structure, w hich is obtained w hen m ost o f the freq u en cy sam ples are zero, as dem onstrated in S ection 7.2.3. T h e follow in g exam p les illustrate the design o f linear-phase F IR filters based on the frequency sam pling m ethod. T h e optim um valu es for th e sam ples in the transition band are ob tain ed from the tables in A p p en d ix C w h ich are taken from the paper by R abiner et al. (1970). Example 8 .2 . 1 Determine the coefficients of a linear-phase FIR filter of length M 15 which has a symmetric unit sample response and a frequency response that satisfies the conditions 1, 0.4, 0. * = 0 ,1 , 2, 3 k=4 k = 5 ,6 , 7
Solution Since h(n) is symmetric and the frequencies are selected to correspond to the case a = 0, we use the corresponding formula in Table 8.3 to evaluate h(n). In this case G(k) = ( ~ l ) kHr The result of this computation is *(0) = A(14) A(l) = A(13) h (2) = A(12) A(3) = A (ll) h( 4) = *(10) h( 5) = A(9) h{6) = A(8) A(7) = -0.014112893 = -0.001945309 = 0.04000004 = 0.01223454 -0.09138802 = -0.01808986 = 0.3133176 = 0.52 ' 2n k \ ) u j k = 0 , 1 ........7
Sec. 8.2
633
TABLE 8.3 UNIT SAMPLE RESPONSE: h(n) = h{M ~ 1 - n)
Symmetric H(k) = G(k)e>*k/M G( Jt ) C l)*r

a =
0
k = 0,1...., M - 1 <?(*) = ~G(M - *) G(*) cos + i )
= *
G(0) + 2
Jk = l M -l
M odd A/ even
t f (* + i) = G (k + A) e>*/2e>0*+D/2*f
f-> ,v
|(+ ^ = EC ( * S in ( *
G ( * + i ) = ( - l) * f f r
G (i + i) = G (M - k - i)
2
. ,
jt
+ 5) (" + 5)
Antisymmetric
H(k) = G(k)eJ,r/2eJ,lk/M G{k) = (-1)kHr ( j p j

=
0
k = 0,1,..., M - 1 G(k) = G(M - k) Af odd ( + j) Af even
h{n) = -^7 V G(fc)sin^^ (n + ;) Af ' Af v

*() =
{ l)"+1G(A//2) - 2
G(&)sin
(* + i) = G (* + i) e 0k+l)fUI G{k + \) = ( - l ) kHr \ ~ (* + l)j G (k + i) = - G (Af - k - \) ;G(M/2) = 0 for Af odd

* (n )=
j; 12 G(k+ 5) 005
Af - 3
( * + j ) (w + 2)
V =
Af odd Af even
634
Chap. 8
3ir 4
Figure 8.14 Frequency response of linear-phase FIR filter in Example 8.2.1.
The frequency response characteristic of th is f i l t e ri s shown in F ig. 8. 14 . We should emphasize that Hr{w) i s exactly equal to the values given by the specifications above at a)k = 2 j r ) t / 1 5 .
Example HJ- 2
Determine the coefficients of a linear-phase FIR f i l t e r of length M = 32 which has a symmetric unit sample response and a frequency response that s a t i s f i e s the condition k ~ 0 , 1,2,3,4, 5
k=
6
*= 7 , 8,.... 15 where T\ = 0.3789795 for a = 0 , and 7 \ = 0.3570496 for a = 5 . These values of T\ were obtained from the tables of optimum transition parameters given in Appendix C. The appropriate equations for t h i s computation are given in Table 8. 3 for a = 0 and a = These computations yield the unit sample responses shown i n Table 8 . 4 . The corresponding frequency response characteristics are illustrated in Figs. 8.15 and 8.16, respectively. Note that the bandwidth of the f i l t e r for a = 5 i s wider than that for a = 0.
Solution T h e optim ization o f th e frequency sam p les in th e tran sition region o f the frequency resp onse can b e explained by evalu atin g th e system fun ction H ( z ) , given by (7.2.12), on th e unit circle and using th e relation sh ip in (8.2.37) to express H ( u > )
Sec. 8.2
635
TABLE 8.4
M = 32 ALPHA 0 . T1 * 0 . 3 7 8 9 7 9 5 E + 0 0 h ( 0) h ( 1) h ( 2) h ( 3) h ( 4) h ( 5) h ( 6) h ( 7) h ( 8) h ( 9) h(10) h (ll) h (1 2) h(13) h (1 4) h (15) = -0 .7 1 4 1 9 7 8 E -0 2 = - 0 . 3 0 7 0 8 0 I E - 02 0 . 5B 91327E-02 = 0 . 1349923E-01 = 0 . 8087033E-02 * -0 .1 1 0 7 2 5 8 E -0 1 - 0 . 2420687E-01 - 0 . 9446550E-02 = 0 . 2544464E-01 = 0 . 3985050E-01 0 . 2753036E-02 = - 0 . 5913959E-01 = - 0 . 6841660E-01 = 0 . 3 175741E-01 = 0.2080981E+00 = 0 . 3471138E+00 M * 32 ALPHA 0 . 5 T1 = 0 . 3 5 7 0 4 9 6 E + 0 0 h ( 0) h ( 1) h ( 2) h ( 3) h ( 4) h ( 5) h ( 6) h ( 7) h ( 8) h ( 9) h(10) h (1 1 ) h (1 2) h(13) h (14) h(15) = - 0 . 40B9120E-02 = - 0 . 9973779E-02 = - 0 . 7379891E-02 = 0 . 5949799E-02 = 0 . 1727056E-01 = 0 . 7878412E-02 = - 0 . 1798590E-01 = -0 .2 6 7 0 5 8 4 E -0 1 = 0 . 3778549E-02 = 0 . 4191022E-01 = 0.283 9 3 4 4 E -0 1 = - 0 . 4 1 6 3 144E-01 = - 0 . 8254962E-01 = 0 . 2802212E-02 = 0 . 2013655E+00 = 0 . 3717532E+00
Figure 8.15 Frequency response oflinear-phase FIR f i l t e ri nExample 8 . 2 . 2 (M = 32 and a = 0 ) .
Magnitude (d B )
636
Chap. 8
Figure 8.16 32 and a =
5).
Frequency response o f linear-phase FIR filter in Example 8.2.2 (M =
in term s o f G{k + a ) . T hus for th e sym m etric filter w e obtain sin a)M MM w here [-G (M -k), G{k + or) _ | C (M _ k _ j ^ a = 0 asci (8.2.39)
1
G (k + a )
l)/2
(8.2.38)
Similarly, for the antisym m etric linear-phase F IR filter w e obtain sin H(co) =
.
(w M M
[
,
u > I X
2
G ( k + a) ~ M (k +
a)\
a = 0 (8.2.41)
where r(lr, f G ( M k) ,
W ith these ex p ression s for the frequency resp onse H(co) given in term s o f the desired frequency sam p les {G (*+<*)}, w e can easily explain th e m eth o d for selectin g the param eters {G(k + a )} in th e transition band w hich result in m inim isin g the peak sid elo b e in th e stopband. In brief, th e valu es o f G ( k + a ) in the passband are
Sec. 8.2
Design of FiR Filters
637
set to (1)* and th ose in th e stopband are set to zero. F or any ch oice o f G (k +ot) in the transition band, the value o f H ( w ) is com puted at a d en se set o f frequencies (e.g., at co = 2 n n / K , n = 0, 1 , . . . , K 1, w h ere, for exam p le, K = 10 M ) . T he valu e o f the m axim um sid elo b e is d eterm ined, and the valu es o f the param eters [G (k + a)} in th e transition band are changed in a d irection o f steep est descent, w hich, in effect, red u ces th e m axim um sid elob e. T h e com p utation o f H(a>) is now rep eated w ith the n ew ch oice o f {G(k + a )}. T h e m axim um sid elob e o f H ( a>) is again determ in ed and the values o f the param eters {G (ik+a)} in the transition band are adjusted in a direction o f steep est d escen t that, in turn, reduces the sidelobe. T his interactive process is perform ed until it con verges to the optim um choice o f the param eters {G (jfc + a )} in the transition band. T h ere is a p otential p roblem in the frequency-sam pling realization o f the FIR linear-phase filter. T h e frequency sam pling realization o f the F IR filter introduces p o les and zeros at equally spaced p oints on the unit circle. In the ideal situation, the zeros can cel the p o les and, con sequ en tly, the actual zeros o f H ( z ) are determ ined by the selectio n o f the frequency sam ples { / / ( i k - f a ) } . In a practical im plem entation o f the frequency-sam pling realization, how ever, q uantization effects preclude a perfect can cellation o f the p oles and zeros. In fact, the location o f p oles on the unit circle provide no dam ping o f th e rou n d -off n oise that is introduced in the com putations. A s a result, such n oise tends to increase w ith tim e and, ultim ately, m ay destroy the norm al operation o f the filter. T o m itigate this problem , w e can m ove b oth the p o le s and zeros from the unit circle to a circle just inside the unit circle, say at radius r = 1 e, w here e is a very sm all num ber. T hus the system function o f the linear-phase F IR filter b eco m es (8.2.42) T h e corresponding tw o -p ole filter realization given in Section 7.2.3 can be m odified accordingly. T h e dam ping provided by selecting r < 1 ensures that rou n d off noise will be b ou n ded and thus instability is avoided.
8.2.4 Design of Optimum Equiripple Linear-Phase FIR Filters

T h e w indow m eth od and th e frequency-sam pling m eth od are relatively sim ple tech n iqu es for d esigning linear-phase F IR filters. H ow ever, th ey also p ossess som e m inor disadvantages, described in Section 8.2.6, which m ay render them undesir able for so m e applications. A major probiem is the lack o f p recise control o f the critical freq u en cies such as cop and ws . T h e filter d esign m eth od described in this section is form ulated as a C hebyshev approxim ation problem . It is view ed as an optim um design criterion in the sen se that the w eigh ted approxim ation error b etw een th e desired frequency re sp on se and the actual frequency resp onse is spread even ly across the passband
638
Chap. 8
and even ly across th e stopband o f the filter m inim izing the m axim um error. T he resulting filter designs have ripples in b oth the passband and th e stopband. T o describe the design procedure, let us con sid er the design o f a lowpass filter with passband ed ge frequency a> p and stop b an d ed ge frequency oos. From the general specifications given in Fig. 8.2, in th e passband, the filter frequency resp onse satisfies the condition 1 - i < Hr (w) < 1 + Si M < o)p (8.2.43)
Similarly, in the stopband, the filter frequency resp on se is sp ecified to fall betw een the lim its 5 2 > that is, 8 2 < Hr ((D) < S2 M > O J, (8.2.44)
Thus Si represents the ripple in the passband and &2 rep resen ts th e attenuation or ripple in the stopband. T h e rem aining filter param eter is M , th e filter length or the num ber o f filter coefficients. L et us focus on the four different cases that result in a linear-phase F IR filter. T h ese cases w ere treated in Section 8.2.2 and are sum m arized b elow .
Case 1: Symmetric unit sample response h(n) = h (M 1 n) and M Odd.

In this case, the real-valued frequency resp onse characteristic Hr (io) is H r {co) = h i - X h(ri)
COS
c u f -------- n )
(8.2.45)
If w e let k = ( M l ) / 2 n and define a new set o f filter p aram eters (a(Jk)} as
a (k) =
*V (^ )7
(8.2.46)
( ^ - 0 then (8.2.45) reduces to the com pact form Hr (a>) (M - 1 ) /2 ^ a (k) cos (ok
kse0
(8.2.47)
Case 2: Symmetric unit sample response /i(n ) = h (M 1 n ) and M Even.

In this case, H r(co) is exp ressed as H r{(d ) = 2 ^
n= o
/ i ( ) c o s g j | -------- n l
\ 2 /
(8.2.48)
A g a in , w e change the sum m ation in d ex from n t o k = M / 2 n and define a new set o f filter param eters {>(*)} as
(M
Sec. 8.2
639
W ith th ese substitutions (8.2.48) b ecom es
up.
H r(w) = X
/ 1\ b(k) cos to [ k - - j
(8.2.50)
In carrying o u t the optim ization, it is co n ven ien t to rearrange (8.2.50) further in to the form H r (o>) = cos (A//2 ) 1 _ H k)co sa > k o (8.2.51)
w h ere the coefficients {(*)] are linearly related to the coefficients {i>(fc)}. In fact, it can b e show n that the relationship is m = \ h i) k = 1,2,3,..., y -2 (8.2.52)
b(k) = 2b(k) - b(k - 1)
C ase 3: A n tisym m etric unit sam p le resp onse h ( n ) = h ( M 1 n ) and M O dd. T he real-valued frequency resp onse characteristic H r (co) for this case is (M 'i) /2 H r (to) = 2 n=o / M _ 1 \ A( ) s i n a > f -------- n l \ 2 /
(8.2.53)
If w e ch an ge th e sum m ation in (8.2.53) from n to k = ( M I ) / 2 n and define a n ew set o f filter param eters {c(*)} as c(jk) = 2h then (8.2.53) b eco m es
(* f-l)/2
* = L 2 .........(Af l ) / 2
(8.2.54)
H r (co )=
5Z *=i
c(Jfc) sin (ok
(8.2.55)
A s in th e p reviou s case, it is con ven ien t to rearrange (8.2.55) in to the form

(W-3)/2
Hr (u>) = sin a> ^
c(k) cos tok
(8.2.56)
w h ere th e coefficients {c(Jfc)} are linearly related to the p aram eters {c(Jfc)}. T h is d e sired relationship can b e derived from (8.2.55) and (8.2.56) and is sim ply
640
given as
Chap. 8
' M - l
m - -m
c(k 1) c(k 4- 1) = 2c(k) 5(0) + \ c ( 2 ) = c( l ) 2 < k < M -5
(8.2.57)
Case 4: Antisymmetric unit sample response h(ri) - h ( M 1 n) and

M Even. In this case, the real-valued frequency response characteristic Hr (co) is
H r (a>) = 2
(Mf2 ) 1 / M _ 1 \ X /i(n)sincti ( - - n \ =o \ 2 /
(8,2.58)
A change in the sum m ation index from n to k = M f l n com b ined with a definition o f a n ew set o f filter coefficients {^(*)), related to {/i(n)} according to d (k)~ 2 h (^-k^j results in th e expression
M/2
* = l,2,...,y
(8.2.59)
j \
H r ((o) = X ^ ( i) s in & > ( k - 1 ' '
(8.2.60)
A s in the previous tw o cases, w e find it con ven ien t to rearrange (8.2.60) in to the form
(j)
ft
m m
Hr (a>) = s i n
*=o
d(k )c osa >k
(8.2.61)
w here the n ew filter p aram eters {d(fc)} are related to {<*()} as follow s:
H
d(k 1
- ' H
) - d(k) =
2 d(k)
M < k <
(8.2.62)
d(0 ) - \ d ( 1 ) = d( 1 ) T h e exp ression s for H r (a>) in these four cases are sum m arized in T ab le 8.5. W e n o te that the rearrangem ents that w e m ade in ca ses 2, 3, an d 4 h ave allow ed
Sec. 8.2
641
TABLE 8.5 REAL-VALUED FREQUENCY RESPONSE

FUNCTIONS FOR LINEAR-PHASE FIR FILTERS
Filter type h(n) ~ h(M M odd (case 1 ) h(n) = h(M Af even (case 2 )
1
Q(<o)
P(co) y ' a(k)coscok k=0 (M /2 J -1 b(k) cos cok M

(M-31/2
- n)
1
- n)
C O C O S
2
h(n) = ~ h ( M - I - n) Af odd (case 3) h(n) = h{M 1 n) M even (case 4)
sin to
Y km 0 (*/2 )-l
cfc) cos cok
. co sin
2
d(k) cos cok km O
us to express H r (co) as H r (co) = Q(co)P(co) where case

COS
1 2
(8.2.63)
Q(<o) =
case
sin co . C O sin
case 3 case 4
(8.2.64)
and P(co) has th e com m on form P(co) = 7 > ( f c ) c o s ^ (8.2.65)
with {(*)} representing the param eters o f the filter, w hich are linearly related to th e unit sam ple resp onse h (n) o f th e F IR filter. T h e upper lim it L in the sum is L = ( M - l ) / 2 for Case 1, L = (Af 3 )/2 for C ase 3, and L = A f/2 1 for C ase 2 and C ase 4. In addition to the com m on fram ew ork given ab ove for the representation o f H r (eo), w e also define th e real-valued desired frequency resp onse H j r ((o) and the w eigh ting function W(co) on the approxim ation error. T h e real-valued d e sired frequency resp onse Hdr (co) is sim ply defined to b e unity in the passband and zero in the stopband. For exam p le, Fig. 8.17 illustrates several d ifferen t typ es o f characteristics for Hdr(co). T he w eigh ting function on the approxim ation er ror allow s u s to ch o o se the relative size o f the errors in the different frequency b ands (i.e., in th e passband and in the stopband). In particular, it is con ven ien t to norm alize W(co) to unity in the stopband and set W(a>) = B2 /& 1 in the passband,
642
HdJ&)
Chap. 8
(a)
(b)
H dr{w )
(o]
(c)
o>2
tlj 4lj (t) 4 <d)
tu
Figure 8 .17 D esired frequency response characteristics for different types of filters.
that is, W(a>) I S2 /S 1 , co in the passband & in the stopband
(8.2.66)
Then w e sim ply select W(a>) in the passband to reflect our em p h asis o n the relative size o f the ripple in the stopband to the ripple in the passband. W ith th e specification o f Hdr(oi) and W O ), w e can n ow d efin e the w eighted approxim ation error as E(a>) = W ( o ) [ H dr(co) - Hr (a>)\ = W (w )[ H dr((o) - Q(co)P(w)]
Sec. 8.2
643
= W(co)Q(co)
Hdr(co) . Q(C0 )
P(co)
(8.2.67)
For m athem atical con ven ien ce, w e define a m odified w eigh ting function W (co) and a m odified desired frequency resp onse Hdr(co) as W(a>) = W(co)Q(co) Hdr(co) = Hdr(co) ( 8 .2 .68)
Q(co)
T h en the w eigh ted approxim ation error m ay b e exp ressed as E(co) = W(co)[Hdr(co) ~ P(o>)\ (8.2.69)
for all four d ifferent types o f linear-phase F IR filters. G iven the error function E(co), the C h eb ysh ev approxim ation problem is basically to determ in e the filter param eters (of(Jt)} that m inim ize the m aximum ab solu te value o f E(co) over the frequency bands in which the approxim ation is to be p erform ed . In m athem atical term s, w e seek the solu tion to the problem
L
m ax \W(co)[Hdr(co) y ^ g ( * ) c o s cok]\ . WfS *=o (8.2.70) w here 5 rep resents the set (disjoint union) o f frequency bands over which the op tim ization is to be perform ed. B asically, the set 5 consists o f the passbands and stopbands o f the desired filter. T h e solu tion to this problem is due to Parks and M cC lellan (1972a), w ho applied a th eorem in the theory of C h eb ysh ev approxim ation. It is called the al ternation theo rem , which w e state w ithout proof.
over (a ( i) J
min
m ax [ ( oj)|
toeS
min
over jo<(t)}
Alternation Theorem: Let 5 b e a com pact subset o f the interval [0, n ecessary and sufficient con d ition for
i P(co) = y ^ g ( f c )c o s cok k= 0
tt).
to b e the unique, best w eigh ted C hebyshev approxim ation to Hdr(co) in S, is that the error function E (co) exhibit at least L + 2 extrem al freq u en cies in S. That is, there m ust exist at least L + 2 freq u en cies {co,} in 5 such that co\ < c& i < < coL+2, E(coj) = E(a>i+j), and jfO oi)! = m a x \E(cv)\
axS
i = 1,2,..., L + 2
W e n o te that the error function E(co) alternates in sign b etw een tw o su cces sive extrem al frequencies. H en ce the theorem is called the alternation theorem . T o elab orate on the alternation theorem , let us con sid er the design o f a low p ass filter with passband 0 < co < cop and stopband cos < co < t t . Since the
644
Chap. 8
desired frequency resp onse Hdr(&) and the w eigh ting function W (w ) are p iecew ise constant, w e have dE{w) d ~ = ( W O ) [Hdr(w) - H r {w) } dw aw _ d H r (w) = Q dw C onsequently, the freq u en cies {oj,} corresponding to the p eak s o f E (w ) also cor respond to peaks at w hich H r (w) m eets the error toleran ce. Since H r(w) is a trigonom etric polyn om ial o f d egree L , for C ase 1, for exam p le,
L
H r (w) Y _ a ( k ) co s wk
k=0
l
]P a r (* )
i=0
L
X ^ni(C O SW )n
= ][V (Jfc)(cosa> )k *=o
(8.2.71)
it follow s that Hr (w) can have at m ost L 1 local m axim a and m inim a in the op en interval 0 < w < n . In addition, w = 0 and w = n are usually extrem a o f H r (w) and, also, o f E{w). T h erefore, H r {w) has at m ost L - l - l extrem al frequencies. Further m ore. the ban d -ed ge freq u en cies w p and ws are also extrem a o f E (w ), since |(a>)| is m axim um at w = wp and w = ws . A s a con seq u en ce, there are at m ost L + 3 ex trem al frequencies in ( w) for the unique, best approxim ation o f the ideal low pass filter. O n the other hand, the alternation theorem states that there are at least L + 2 extrem al freq u en cies in E (w ). T hus the error function for th e low pass filter design has either L + 3 or L + 2 extrem a. In gen eral, filter designs that contain m ore than L + 2 alternations or ripples are called extra ripple filters. W h en the filter design contains the m axim um num ber o f alternations, it is called a m a x i m a l ripple filter. T he alternation th eorem guarantees a unique solu tion for the C hebyshev optim ization problem in (8.2.70). A t the desired extrem al freq u en cies {w}, w e have the set o f equations W (w n)[Hdr(wn) - P ( w n)] = ( - i y \ 5 = 0 , 1 ......... L + 1 (8.2.72)
w here & represents the m axim um valu e o f th e error function E (w ). In fact, if we select W (w ) as indicated by (8.2.66), it follow s that S = S2. The set o f linear eq u ation s in (8.2.72) can b e rearranged as ( - 1 )"$ P((On) + V - = Hdr(wn) W(o>) or, equivalently, in the form a(k)cosw k + to ^ = H dr((On) W (v*) = 0,1,..., L + 1 (8.2.73) n = 0 , 1 .........L + 1
Sec. 8.2
645
I f w e treat the {()} and S as the param eters to be determ in ed , (8.2.73) can be exp ressed in matrix form as
1
co s C O o co s cd\
co s 2 c d o co s 2 o> ]
cos Lwo W(coo) cos Lcoi

- 1
a (0 ) a(l) W(<ui) (-l)i a(L) S
Hdr (< * > 0) Hdr(0 h )
c o s w t+ i
c o s 2 a>z,+i
cosZ.gj+ i
_ H dr (&1 +1 ) _
(8.2.74) Initially, w e k n ow neither the set o f extrem al freq u en cies {io} nor the p a ram eters {(*()} and S. T o solve for the param eters, w e use an iterative algorithm , called th e R e m e z e xch an ge a lgorithm [see R abiner et al. (1975)], in which w e begin by guessing at the set o f extrem al freq u en cies, d eterm in e P (w ) and < 5, and then co m p u te the error function E (w ). From (w ) w e d eterm in e an oth er set o f L + 2 extrem al freq u en cies and repeat the process iteratively until it con verges to the optim al set o f extrem al freq u en cies. A lth ou gh the m atrix eq u ation in (8.2.74) can b e used in the iterative procedure, matrix inversion is tim e consum ing and inefficient. A m ore efficient procedure, su ggested in the paper by R abiner et al. (1975), is to co m p u te < 5 analytically, according to the form ula S= yoHdr((oo) + y\ Hdr (<oi) H --------1- yL+\Hdr O l + i)
tf>
(8.2.75)
Y\
W (w i)
( - r V u i W { a>t + i)
W(o)o) w here
co s (O k cos (on nykk
(8.2.76)
T h e expression for S in (8.2.75) fo llo w s im m ed iately from th e m atrix equation in (8.2.74). Thus with an initial guess at th e L + 2 extrem al freq u en cies, w e com pute 5. N o w since P(a>) is a trigom etric p olyn om ial o f the form L P(co) = Y a ( k ) x k *o x = cosa)
and since w e k n ow that th e polynom ial at the p oin ts x = co s ( o , n ~ 0 , 1 , . . . , L + 1, has th e corresponding values
P(a>n) - Hjr((On) ~ -- ---W(con)
( 1)"5
n = 0 , 1.....L + 1
(8.2.77)
646
Chap. 8
w e can use the Lagrange interpolation form ula for P(a>). T h u s P(to) can be ex pressed as [see H am m ing (1962)]
L
> ( o * )[& /(* JC *)]
(8.2.78)
w here P(a>n) is given by (8.2.77), x = co sw , x k = cosw *, and (8.2.79)
H aving the solu tion for P(co), w e can n ow com p ute the error function E(co) from (o>) = W{u>)[Hdr{a>) - P(co)] (8.2.80)
on a den se set o f frequency points. U sually, a num ber o f p oints equal to 16Af, w here M is the length o f the filter, suffices. If |(&))| > 8 for so m e freq u en cies on the den se set, then a new set o f frequencies corresponding to the L + 2 largest peaks o f |(<y)| are selected and the com putational procedure b egin n ing with (8.2.75) is repeated. Since the new set o f L + 2 extrem al freq u en cies are selected to correspond to the peaks o f the error function |(a>)|, the algorithm forces 5 to increase in each iteration until it con verges to the upper b ou n d and h en ce to the optim um solu tion for the C h eb ysh ev approxim ation problem . In other words, w hen [(a>)| < 5 for all frequencies on the d en se set, the optim al solu tion has b een found in term s o f the polynom ial H(co). A flowchart o f the algorithm is sh ow n in Fig. 8.18 and is d u e to R em ez (1957). O nce the optim al solu tion has b een ob tain ed in term s o f P(co), the unit sam ple resp onse h (n) can b e com p uted directly, w ithout h aving to com p ute the p aram eters (ar(fc)}. In effect, w e have d eterm ined H r(co) = Q ( oj) P ( co) w hich can be evalu ated at co = 2 n k / M , k = 0, 1 , . . . , (Af l ) /2 , for Af odd, or A f/2 for M ev en . T hen, d ep en d ing on the type o f filter b ein g d esign ed , h (n) can b e determ in ed from th e form ulas given in T able 8.3. A com puter program written by Parks and M cC lellan (1972b) is available for d esigning iinear phase F IR filters b ased on the C h eb ysh ev approxim ation criterion and im plem en ted with the R em ez exch an ge algorithm . T h is program can be used to design low pass, highpass or bandpass filters, differentiators, and H ilbert trans form ers. T he latter tw o types o f filters are described in the follow in g section s. A num ber o f softw are packages for designing equiripple lin ear-ph ase F IR filters are n ow available. T h e P ark s-M cC lellan program requires a num ber o f in p u t param eters which determ in e th e filter characteristics. In particular, the follow in g param eters must
Sec. 8.2
647
Input filter parameters
Initial guess of M + 2 extremal freq.
Fifnre 8.18
Flowchart o f R em ez algorithm.
b e specified:
L IN E 1
N F 1 LT: JT Y P E :
T h e filter length, d en oted ab ove as M . T yp e o f filter JT Y P E = 1 results in a m ultiple passband/stopband filter.
648
Chap. 8
JT Y P E = 2 results in a differentiator. JT Y P E = 3 results in a H ilb ert transform er. N BA NDS: L G R ID : L IN E 2 EDGE: T he freq u en cy bands specified by low er and upper cu toff fre qu en cies, up to a m axim um o f 1 0 bands (an array o f size 2 0 , m axim um ). T he freq u en cies are given in term s o f the variable / = a ) / 2 n , w h ere / = 0.5 corresponds to the folding frequency. The num ber o f frequency bands from 2 (for a low pass filter) to a m axim um o f 1 0 (for a m ultiple-band filter). T he grid density for interpolating the error function E{u>). T he d efault valu e is 16 if left unspecified.
L IN E 3 FX: L IN E 4 W TX; A n array o f m axim um size 10 that specifies the w eight function in each band. A n array o f m axim um size 10 that sp ecifies the desired fre q uency resp onse Hdr (& >) in each band.
The follow in g exam p les d em onstrate the use o f this program to design a low pass and a bandpass filter. Example 8.23 Design a lowpass filter of length M = 61 with a passband edge frequency f p = 0.1 and a stopband edge frequency f s = 0.15. Solution The lowpass filter is a two-band filter with passband edge frequencies (0,0.1) and stopband edge frequencies (0.15,0.5). The desired response is (1, 0) and the weight function is arbitrarily selected as ( 1 , 1 ). 61, 1, 2 0.0, 0.1,0.15,0.5 1. 0. 0.0 1. 0, 1.0 The result of this design is illustrated in Table 8 .6 , which gives the filter coefficients. The frequency response is shown in Fig. 8.19. The resulting filter has a stopband attenuation of - 5 6 dB and a passband ripple of 0.0135 dB. If w e increase the length o f the filter to M = 101 w h ile m aintaining all the other param eters given ab ove the sam e, the resulting filter h as the frequency re sp o n se characteristic sh ow n in Fig. 8.20. N o w , th e stopband atten u ation is 85 dB and the passband ripple is reduced to 0.00046 dB . W e should indicate that it is p ossib le to in crease the atten u ation in the stop band by k eep in g th e filter len gth fixed, say at Af = 61, and decreasin g the weighting function W(co) = S2 / S 1 in th e passband. W ith Af = 61 and a w eigh ting function
TABLE 8.6 PARAMETERS FOR LOWPASS FILTER DESIGN IN

EXAMPLE 8.2.3 F I N I T E I M P U L S E R E S P O N S E (FIR) LINEAR PHASE DIGITAL FILTER DESIGN REMEZ EXCHANGE ALGORITHM FILTER LENGTH = 6 1 I M P U L S E R E S P O N S E *** ** H ( 1) = - 0 . 1 2 1 0 9 3 5 1 E - 0 2 = H ( 61) H ( 2) = - 0 . 6 7 2 7 0 6 8 7 E - 0 3 = H ( 60) H ( 3) = 0 . 9 8 0 9 0 2 4 0 E - 0 4 = H ( 59) H( 4) = 0 . 1 3 5 3 6 6 6 4 E - 0 2 = H ( 58) 0. 22969784E-02 H ( 5) H ( 57) H ( 6) = 0 . 1 9 9 6 3 4 9 5 E - 0 2 = H ( 56) 0 . 9 7 0 2 6 0 9 5 E - 0 4 = H( 55) H ( 7) = H ( 8) = - 0 . 2 6 4 6 6 6 9 5 E - 0 2 = H ( 54) H ( 9) = - 0 . 4 5 1 3 3 1 0 3 E - 0 2 = H ( 53) H (10) = - 0 . 3 7 7 0 4 9 4 4 E - 0 2 = H ( 52) H (11) = 0 . 1 3 0 7 9 6 5 5 E - 0 4 = H ( 51) HS12) = 0 . 5 1 7 913 5 6 E - 0 2 = H( 50) H (13) = 0 . 8 4 8 8 3 4 7 8 E - 0 2 = H ( 49) H (14) = 0 . 6 9 5 3 2 1 1 0 E - 0 2 = H ( 48) H (15) = 0 . 7 1 0 3 7 0 5 9 E - 0 4 = H ( 47) H (16) = - 0 . 9 0 4 07 8 9 7 E - 0 2 = H ( 46) H (17) = - 0 . 1 4 7 2 3 0 4 7 E - 0 1 = H ( 45) H (18) = - 0 . 1 1 9 5 8 9 4 5 E - 0 1 = H ( 44) H (19) = - 0 . 2 9 7 9 9 2 1 4 E - 0 4 = H ( 43) H (20) = 0 . 1 5 7 1 3 4 2 2 E - 0 1 = H ( 42) H (21) = 0 . 2 5 6 5 7 1 5 1 E - 0 1 = H ( 41) H (22) = 0 . 2 1 0 5 7 3 7 3 E - 01 = H { 40) H (23) = 0. 6 8 6 3 7 7 6 8 E - 0 4 = H ( 39) H (24) = -0. 2 8 9 0 2 0 5 4 E - 0 1 = H ( 38) H (25) = - 0 . 4 9 1 1 8 5 4 1 E - 0 1 = H ( 37) H (26) - 0 . 4 2 7 1 3 9 7 0 E - 01 = H ( 36) H (27) = - 0 . 5 0 1 1 4 3 0 4 E- 0 4 = H ( 35) H (28) = 0 . 7 3 5 7 4 2 1 5 E - 0 1 = H ( 34) H (29) = 0 . 1 5 7 8 2 0 4 0 E + 0 0 = H( 33) H (30) = 0 . 2 2 4 6 5 5 1 2 E + 0 0 = H ( 32) H (31) = 0 . 25007001E+'p0 = H ( 31) BAND 1 BAND 2 LOWER BAND EDGE 0.0000000 0.1500000 UPPER BAND EDGE 0.1000000 0.5000000 DESIRED VALUE 1.0000000 0.0000000 WEIGHTING 1.0000000 1.0000000 DEVIATION 0.0015537 0.0015537 D E V I A T I O N IN D B 0.0134854 -56.1724014 EXTREMAL FREQUENCIES--MAXIMA OF THE ERROR CURVE 0.0000000 0.0252016 0.0423387 0.0584677 0.0735887 0.0866935 0.0957661 0.1000000 0.1500000 0.1540323 0.1631048 0.1762097 0.1903225 0.2054435 0.2215725 0.2377015 0.2538306 0.2699596 0.2860886 0.3022176 0.3183466 0.3354837 0.3516127 0.3677417 0.3848788 0.4010078 0.4171368 0.4342739 0.4504029 0.4665320 0.4836690 0.5000000
649
650
Design of Digital Fillers
Chap. 8
Magnitude (dB)
Figure 8.19
Frequency response o f M = 61 FIR filter in Example 8.2.3.
Figure
Frequency response o f M 101 FIR filter in Exam ple 8.2.3.
(0 .1 ,1 ), w e obtain a filter that has a stopband attenuation o f 65 dB and a pass band ripple o f 0.049 dB.
Example 8*2,4 Design a bandpass f i l t e r of length M = 32 with passband edge frequencies f pi = 0.2 and fpi = 0.35 and stopband edge frequencies of f M \ = 0.1 and f ,2 = 0.425.
Sec. 8.2
651
Solution This passband filter is a three-band filter with a stopband range of (0,0.1), a passband range of (0.2,0.35), and a second stopband range of (0.425,0.5). The weight ing function is selected as (1 0 .0 , 1 .0 , 1 0 .0 ), or as ( 1 .0 , 0 .1 , 1 .0 ), and the desired response in the three bands is (0.0,1.0,0.0). Thus the input parameters to the program are
32, 1, 3 0.0,0.1,0.2,0.35,0.425,0.5 0 . 0, 1. 0, 0.0 10.0, 1. 0, 10.0 The results of this design are shown in Table 8.7, which gives the filter coefficients. We note that the ripple in the stopbands Bz is 10 times smaller than the ripple in
TABLE 8.7 PARAMETERS FOR BANDPASS FILTER IN EXAMPLE 8.2.4 F I N I T E I M P U L S E R E S P O N S E (FIR) LINEAR PHASE DIGITAL FILTER DESIGN REMEZ EXCHANGE ALGORITHM BANDPASS FILTER FILTER LENGTH = 3 2 ***** I M P U L S E R E S P O N S E ***** H ( 1) = -0 .5 7 5 3 4 0 2 6 E - 0 2 = H ( 32) H ( 2) = 0 .9 9 0 2 6 6 9 1 E - 0 3 = H ( 31) H< 3) = 0 . 7 5 7 3 3 4 7 1 E - 0 2 = H ( 30) -0 .6 5 1 4 1 2 0 4 E - 0 2 = H ( 29) H ( 4) H ( 5) = 0 .1 3 9 6 0 5 0 9 E - 0 1 = H ( 28) H ( 6) = 0 .2 2 9 5 1 6 4 4 E - 0 2 = H ( 27) H ( 7) = -0 .1 9 9 9 4 0 4 1 E - 0 1 = H( 26) H ( 8) = 0,. 7 1 3 6 9 6 5 6 E - 0 2 = H ( 25) H< 9} = - 0 . .3 9 6 5 7 3 7 3 E - 0 1 H( 24) H (10) e 0 . ,1 1 2 6 0 0 6 6 E - 0 1 = H ( 23) H( l l > = 0 . .6 6 2 3 3 6 3 5 E - 0 1 = H ( 22) H (12) * -0.,1 0 4 9 7 2 0 2 E - 0 1 = H ( 21) H H H H (13) (14) (15) (16) = 0 . ,8 5 1 3 6 1 6 0 E - 0 1 = H( = -0..1 2 0 2 4 9 8 8 E + 0 0 = H ( = - 0 .,2 9 6 7 8 5 8 0 E + 0 0 * = H( = 0 . 3 0 4 1 0 9 1 3 E + 0 0 = H( BAND 1 BAND 2 20) 19) 18) 17)
LOWER BAND EDGE 0.0000000 0.2000000 0.4250000 UPPER BAND EDGE 0.1000000 0.3500000 0.5000000 DESIRED VALUE 0.0000000 1.0000000 0.0000000 WEIGHTING 10.0000000 1.0000000 10.0000000 DEVIATION 0.0015131 0.0151312 0.0015131 DEVIATION IN DB -56.4025536 0.1304428 -56.4025536 EXTREMAL FREQUENCIES MAXIMA OF THE ERROR CURVE 0.0000000 0.0273438 0.0527344 0.0761719 0.0937500 0.1000000 0.2000000 0.2195313 0.2527344 0.2839844 0.3132813 0.3386719 0.3500000 0.4250000 0.4328125 0.4503906 0.4796875
652
Chap. 8
Magnitude (d B )
Figure JL21
Frequency response of M = 32 FIR filter in Example 8.2.4.
the passband due to the f ac t that errors in the stopband were given a weight of 10 compared to the passband weight of unity. The frequency response of the bandpass f i l t e ri s illustrated i nF ig. 8 .2 1 .
T h ese exam p les serve to illustrate the relative ease w ith which optim al low pass, highpass, bandstop, bandpass, and m ore general m ultiband linear-phase FIR filters can be designed based on the C h eb ysh ev approxim ation criterion im ple m ented by m eans o f the R em ez exch an ge algorithm . In the n ext tw o sections w e consider the design o f differentiators and H ilbert transform ers.
8.2.5 Design of FIR Differentiators

D ifferen tiators are used in m any analog and digital system s to take the derivative o f a signal. A n ideal differentiator has a frequency resp onse that is linearly pro portional to frequency. Sim ilarly, an ideal digital differentiator is defined as one that has the frequency resp onse
Hd(ao) = j t o
7r < u) < n
(8.2.81)
T h e unit sam ple response corresponding to Hj(co) is M ) = =

2
/ f
Hd((*>)eiwnda> j(jDeJconda> oo < n < oo, n ^ 0
2n
cosnn
(8.2.82)
Sec. 8.2
653
W e ob serve that the ideal differentiator has an antisym m etric unit sam ple resp onse [i.e., h d (n) = h d( n)]. H en ce, h d (0) = 0. In this section w e consider the design o f lin ear-ph ase F IR differentiators b ased o n the C h eb ysh ev approxim ation criterion. In view o f the fact that the ideal differentiator has an antisym m etric unit sam p le resp onse, w e shall con fine our atten tion to F IR designs in which h ( n ) = h ( M 1 n). H en ce w e con sid er the filter types classified in the p receding section , as C ase 3 and C ase 4. W e recall that in C ase 3, w h ere Af is od d , the real-valu ed frequency resp onse o f the F IR filter H r(co) has the characteristic that H r (0) = 0. A zero resp onse at zero freq u en cy is just the con d ition that the d ifferen tiator should satisfy, and w e se e from T ab le 8.5 that b oth filter types satisfy this con d ition . H ow ever, if a fullband differentiator is desired, this is im possib le to ach ieve with an F IR filter having an od d num ber o f coefficients, since H r ( n ) = 0 for Af odd. In practice, h ow ever, full-band differentiators are rarely required. In m ost cases o f practical interest, the desired freq u en cy resp onse character istic n eed o n ly be linear over the lim ited frequency range 0 < co < 2 n f p , w h ere f p is called the bandw idth o f the differentiator. In the frequency range 2 n f p < co < n , the desired resp onse m ay be either left unconstrained or constrained to b e zero. In the design o f F IR differentiators b ased on the C h eb ysh ev approxim ation criterion, the w eigh ting function W(co) is sp ecified in th e program as (8.2.83) W (co) = 0 < co < 2 n f p co in order that the relative ripple in the passband be a constant. T hus the ab solu te error b etw een the desired resp onse co and th e approxim ation Hr(co) increases as co varies from 0 to 2 n f p. H ow ever, the w eigh ting fun ction in (8.2.83) ensures that the relative error
8
m ax
O s a><2nfp
{ W (co )[(o - Hr (a>)]}

H r (CO)
m ax
1 -
(8.2.84)
is fixed within the passband o f the differentiator.
Example 8^5
Use the Remez algorithm to design a linear-phase FIR differentiator of length hi = 60. The passband edge frequency is 0.1 and the stopband edge frequency is 0.15.
Solution The input parameters to the program are

60, 0.0, 1.0, 1.0,
2 , 0.1, 0.0 1.0 2
0.15,
0.5
The results of this design including the filter coefficients are shown in Table 8 .8 . The frequency response characteristic is illustrated in Fig. 8.22. Also shown in the same figure is the approximation error over the passband 0 < / < 0 . 1 of the filter.
TABLE 8.8
PARAMETERS FOR FIR DIFFERENTIATOR IN EXAMPLE 8.2.5
F I N I T E I M P U L S E R E S P O N S E (FIR) LINEAR-PHASE DIGITAL FILTER DESIGN REMEZ EXCHANGE ALGORITHM DIFFERENTIATOR FILTER LENGTH = 6 0 IMPULSE RESPONSE ' H ( 1) = - 0 . 1 2 4 7 8 0 7 5 E - 0 2 = -H{ 60) - 0 . 1 5 7 1 3 5 6 0 E - 02 = -H( 59) H ( 2) H ( 3) = 0 . 3 6 8 4 6 7 3 7 E - 0 2 = -H ( 58) 0 . 1 9 2 9 8 0 2 0 E - 0 2 = - H ( 57) H ( 4) = H( 5) = 0 . 1 4 2 6 4 1 4 I E - 02 = - H ( 56) H ( 6) = - 0 . 1 7 6 1 5 2 7 7 E - 0 2 = -H( 55) H ( 7) = - 0 . 4 3 1 1 0 5 7 3 E ~ 02 = -H( 54) - 0 . 4 6 9 5 3 4 0 5 E - 0 2 = -H( 53) H ( 8) H ( 9) = - 0 . 1 4 1 0 5 2 4 4 E - 02 * -H( 52) 0 . 4 1 6 9 4 2 2 2 E - 0 2 = -H( 51) H (10) H (11) = 0 . 8 5 7 3 6 2 1 5 E - 0 2 - H ( 50) H (12) = 0 . 7 9 8 1 3 0 3 1 E - 0 2 = -H( 49) 0.11833385E-02 = H(13) H ( 48) H (14) = - 0 . 8 7 3 9 6 0 6 5 E - 0 2 = -H( 47) H ( 46) H (15) = - 0 . 1 5 4 0 1 8 4 7 E - 0 1 = H (16) K - 0 . 1 2 8 7 8 4 4 5 E - 0 1 = -H ( 45) H (17) = - 0 . 1 8 8 2 6 8 7 2 E - 0 3 = ~H( 44) H (18) = 0 . 1 6 6 2 0 5 0 6 E - 0 1 = -H ( 43) s 0 . 2 6 7 4 1 5 2 3 E - 0 1 = -H( 42) H (19) s H (20) = 0.2 0 8 9 2 0 1 8 E - 0 1 = -H ( 41) H(21) = - 0 . 1 8 5 8 4 0 9 5 E - 02 = -H( 40) H (22) = - 0 . 3 1 1 0 9 9 0 9 E - 0 1 = -H ( 39) H (23) = - 0 . 4 8 8 2 2 1 7 6 E - 0 1 = -H( 38) H (24) = - 0 . 3 8 6 7 3 4 5 3 E - 01 = " H ( 37) H (2 5) = 0 . 3 6 7 6 0 1 2 2 E - 0 2 = -H( 36) H (26) = 0 . 6 5 4 6 2 4 7 8 E - 0 1 = -H ( 35) H (27) = 0 . 1 2 0 6 6 3 1 7 E + 0 0 = -H( 34) H (28) = 0 . 1 4 1 8 2 1 3 4 E + 0 0 = -H( 33) H (29) = 0 . 1 1 4 0 3 7 5 7 E + 0 0 * -H( 32) H ( 3 0 > = 0 . 4 3 6 2 0 0 8 0 E - 0 1 = - H ( 31) BAND 1 BAND 2 LOWER BAND EDGE 0.0000000 0.1500000 UPPER BAND EDGE 0.1000000 0.5000000 DESIRED SLOPE 10.0000000 0.0000000 WEIGHTING 1.0000000 1.0000000 DEVIATION 0.0073580 0.0073580 E X T R E M A L F R E Q U E N C I E S --- M A X I M A O F T H E E R R O R C U R V E 0.0010417 0.0156250 0.0312500 0.0468750 0.0614583 0.0750000 0.0875000 0.0968750 0.1000000 0.1500000 0.1552083 0.1666667 0.1822916 0.1979166 0.2156249 0.2322916 0.2489582 0.2666668 0.2843754 0.3020839 0.3187508 0.3364594 0.3541680 0.3718765 0.3906268 0.4083354 0.4260439 0.4447942 0.4625027 0.4812530 0.5000000
654
Sec. 8.2
655
Normalized Error X .001195
Error X .000736
Magnitude
Normalized frequency Fifure Frequency response and approximation error for M = 60 FIR differentiator of Example 8.2.5.
The im portant p aram eters in a differentiator are its len gth Af, its bandwidth {band-edge frequency) f p, and the p eak relative error S o f th e approxim ation. T he in terrelationship am ong th ese three param eters can be easily displayed param et rically. In particular, th e value o f 20 log 10 S versus f p w ith Af as a param eter is sh ow n in Fig. 8.23 for Af even and in Fig. 8.24 for Af odd. T h ese results, due to R abiner and Schafter (1974a), are useful in the selection o f th e filter length, given specifications on the inband ripple and the cu toff frequency f p. A com parison o f th e graphs in Figs. 8.23 and 8.24 reveals that even -len gth differentiators result in a significantly sm aller approxim ation error S than com pa rable od d -len gth d ifferentiators. D esig n s based o n Af odd are particularly p oor if
656
Design of Digital niters
Chap. 8
Passband cutoff frequency {fp) Figure &23 Curves of 20 log 105 versus f p for M = 4, 8 , 16, 32, and 64. [From paper by Rabiner and Schafer (1974a). Reprinted with permission o f AT&T.]
the bandw idth ex ceed s f p = 0.45. T h e problem is b asically th e zero in the fre q u en cy resp onse at co = n ( f = 1 /2 ). W h en f p < 0.45, g o o d d esign s are obtained for M odd, but com parable-length d ifferen tiators with M even are always better in the sense that the approxim ation error is sm aller. In view o f the ob viou s advantage o f even -len gth over od d -len gth differentia tors, a conclusion m ight b e that even -len gth differentiators are alw ays preferable in practical system s. T his is certainly true for m any applications. H ow ever, w e should n o te that th e signal d elay introduced by any linear-phase F IR filter is ( M 1)/2, w hich is n ot an integer w hen M is even . In m any practical applications, this is unim portant. In so m e applications w here it is d esirable to h ave an integer-valued d elay in the signal at th e output o f th e differentiator, w e m ust se lect M to be odd. T h ese num erical results are b ased on d esign s resulting from the C hebyshev approxim ation criterion. W e w ish to indicate it is also p o ssib le and relatively
Sec. 8.2
657
Passband cutoff frequency (fp) Figure 8-24 Curves of 201ogto S versus Fp for M 5, 9, 17, 33 and 65. [From paper by Rabiner and Schafer (1974a). Reprinted with permission o f AT&T.]
easy to design linear phase F IR differentiators based o n the frequency sam pling m eth od . F or exam p le, Fig. 8.25 illustrates the frequency resp onse characteristics o f a w ideband ( f p = 0.5) differentiator, o f length M 30. T h e graph o f the ab solu te valu e o f th e approxim ation error as a function o f freq u en cy is also show n in this figure.
8.2.6 Design of Hilbert Transformers

A n ideal H ilb ert transform er is an all-pass filter that im parts a 90 phase shift on th e signal at its input. H en ce th e frequency resp on se o f th e ideal H ilbert transform er is specified as
(8.285)
658
Chap. 8
Error (dB)
Magnitude
Figure fL25 Frequency response and approximation error for M = 30 FIR differentiator designed by frequency sampling method.
H ilbert transform ers are frequently used in com m u n ication system s and signal processing, as, for exam ple, in the gen eration o f sin gle-sid eb an d m odulated signals, radar signal processin g, and sp eech signal p rocessing. T h e unit sam ple resp onse o f an idea! H ilbert transform er is
hd(n) = J
Hd(u>)eja,ndco
Sec. 8.2
659 2 sin2(jrn/2)
n 0,
(8.2.86)
A s ex p ected , hd(n) is infinite in duration and n oncausal. W e n ote that hd(n) is antisym m etric [i.e., hd (n) = k d( n)]. In view o f this characteristic, w e focus our atten tion on th e design o f linear-phase F IR H ilb ert transform ers w ith antisym m et ric unit sam ple resp onse [i.e., h(n) h ( M 1 n)]. W e also ob serve that our ch oice o f an antisym m etric unit sam ple resp onse is con sisten t with having a purely im aginary frequency resp onse characteristic Hd (co). W e recall o n ce again that w h en h{n) is antisym m etric, the real-valued fre quency resp onse characteristic H,(co) is zero at co = 0 for b oth M od d and ev en and at co = n w h en M is odd. C learly, then, it is im p ossib le to design an all-pass digital H ilb ert transform er. Fortunately, in practical signal processing applications, an all-pass H ilbert transform er is unnecessary. Its bandw idth n eed on ly cover the bandwidth o f the signal to b e p h ase shifted. C on sequ en tly, w e specify the desired real-valued frequency resp onse o f a H ilbert transform filter as Hdr(co) =
2 :z
f < co <
2n f u
(8.2.87)
w here f t and / are the low er and upper cu toff freq u en cies, respectively. It is interesting to n ote that the ideal H ilbert transform er with unit sam ple resp onse hd (n) as given in (8.2.86) is zero for n even . T his property is retained by the F IR H ilbert transform er under som e sym m etry con d ition s. In particular, let us consider the C ase 3 filter type for which (W-l)/2
(8.2.88)
and su p p ose that ft = 0.5 f u. This ensures a sym m etric passband about the m idpoint frequency / = 0.25. If w e have this sym m etry in the frequency resp onse, Hr (co) = Hr ( n co) and h en ce (8.2.88) yields (*
IM-D/2 J ' c(k) sin<w/k c o s 7 T /k fc-i (M-D /2 7 ! c ( k) ( 1)*+1 sini>* or equivalently, (M- 1)/2
(8.2.89)
Clearly, c(k) m ust b e eq u al to zero for jfc 0, 2, 4,
660
Chap. 8
N o w , th e relationship b etw een (c()} and the unit sam p le resp on se {/?() j is, from (8.2.54),
or, equivalently, (8.2.90) If c(k) is zero for k = 0, 2, 4 , . . . , then (8.2.90) yield s
0,
h{k) =
0 ,2 , 4 , . . . . for k 1, 3 , 5 , . . . , for ^
ev en (8.2.91)
0,
^ od d
U n fortu n ately, (8.2.91) h old s only for M odd. It d o es n ot h old for M even . This m ean s that for com parable values o f M , the case M odd is preferab le since the com p utational com p lexity (num ber o f m ultiplications and ad d ition s per output p oin t) is roughly o n e half o f that for M even. W hen th e design o f th e H ilbert transform er is perform ed b y the C hebyshev approxim ation criterion using the R em ez algorithm , w e se lect th e filter coefficients to m inim ize the peak approxim ation error S = = max m ax [Hdr(ai) Hr (co) ] (8.2.92)
2jrfi<w<2x / , 2jtft <a><2xfm
[1 - Hr ((t>)]
Thus the w eighting function is set to unity and th e op tim ization is perform ed over the single frequency band (i.e., the passband o f th e filter). Example 8.2.6 Design a Hilbert transformer with parameters M = 31, // = 0.05, and / , = 0.45. Solution We observe that the frequency response is symmetric, since f u = 0.5 - fiThe parameters for executing the Remez algorithm are 31. 0.05,
1.0
3, 0.45
1.0 The result of this design is the unit sample response coefficients and the peak ap proximation error & = 0.0026803 or 51.4 dB given in Table 8.9. We observe that, indeed, every other value of h(n) is essentially zero (these values are of the order of 10~7). The frequency response of the Hilbert transformer is shown in Fig. 826.
Sec. 8.2
661
TABLE 8.9 PARAMETERS FOR FIR HILBERT TRANSFORM FILTER IN

EXAMPLE 8.2.6
F I N I T E I M P U L S E R E S P O N S E (FIR) L INEAR PHASE DIGITAL FILTER DESIGN REMEZ EXCHANGE ALGORITHM HILBERT TRANSFORMER FILTER LENGTH = 3 1 * *** ** IMPULSE RESPONSE * 0 .4 1 9 5 7 5 1 6 E - 0 2 = -H< 31) 0 .6 4 3 1 0 2 5 7 E - 0 7 = ~H( 30) 0 . 9 2 8 2 2 4 4 4 E - 0 2 = -H ( 29) 0 .5 2 6 9 3 9 2 7 E - 0 7 = -H< 28) 0,.1 8 8 3 5 9 8 8 E - 0 1 = -H ( 27) 0 .8 2 3 0 8 2 8 3 E - 0 7 = -H( 26) = 0..3 4 4 0 1 1 9 0 E - 0 1 = -H ( 25) = 0..9 3 3 2 8 7 9 4 E - 0 7 = -H( 24) 0.,5 9 5 5 1 7 3 8 E - 0 1 = -H( 23) = = 0..5 0 8 2 1 1 7 1 E - 0 7 = -H( 22) H (11) = 0. 1 0 3 0 3 7 8 2 E + 0 0 = ~H( 21) H (12) = 0. 1 7 6 1 2 1 3 8 E - 0 7 = -H( 20) H (13) = 0. 1 9 6 8 3 1 6 7 E + 0 0 = -H< 19) H (14) = -0. 2 3 9 7 7 6 0 6 E - 0 7 = -H ( 18) H {15) = 0. 6 3 1 3 5 3 7 4 E + 0 0 = -H ( 17) H (16) = 0. 0 BAND 1 LOWER BAND EDGE 0.0500000 UPPER BAND EDGE 0.4500000 DESIRED VALUE 1.0000000 WEIGHTING 1.0000000 DEVIATION 0.0026803 E X T R E M A L F R E Q U E N C I E S --- M A X I M A O F T H E E R R O R C U R V E = = = = = 0.0500000 0.1583333 0.3124998 0.0562500 0.1874999 0.3416664 0.0750000 0.2187499 0.3708331 0.1000000 0.2499998 0.3999997 0.1291666 0.2812498 0.4249997 H { 1) H ( 2) H ( 3) H ( 4) H ( 5) H ( 6) H { 7) H ( 8) H( 9) H {10) =
R abiner and Schafer (1974b) h ave investigated the characteristics o f H ilbert transform er designs for b oth M odd and M ev en . If the filter design is restricted to a sym m etric frequency response, then there are basically three param eters o f interest, M, 8 , and f\. Figure 8.27 is a p lot o f 201og10< 5 versus f i with M as a param eter. W e ob serve that for com parable valu es o f M, th ere is no perform ance advantage o f using M od d over M even , and vice versa. H ow ever, the com puta tional com p lexity in im plem enting a filter for M odd is less by a factor o f 2 over M even as previou sly indicated. T herefore, M odd is p referab le in practice. For design purposes, the graphs in Fig. 8.27 suggest that, as a rule o f thum b,
Mfi % -0 .6 1 log10 8
(8.2.93)
662
Chap. 8
Magnitude (d B )
Figure 8.26
Frequency response of FIR Hilbert transform filter in Example
8.6.6.
H en ce this form ula can be used to estim ate the size o f on e o f th e three basic filter param eters w h en the other tw o param eters are specified. In concluding this section , w e wish to sh ow that H ilb ert transform ers can also be designed by the w indow m eth od and th e freq u en cy sam pling m ethod. F or exam p le, Fig. 8.28 illustrates the frequency resp onse o f an M = 31 H ilbert transform er design ed using the frequency sam pling m ethod. T h e corresponding values o f the unit sam ple resp onse are given in T ab le 8.10. A com parison o f these filter param eters with th ose given in T able 8.9 indicates so m e sm all differences. In particular, it appears that the C h eb ysh ev approxim ation criterion gives signifi cantly sm aller valu es for the filter coefficients that should b e zero. In general, the C h eb ysh ev approxim ation criterion results in b etter filter designs.
8.2.7 C om parison of D esign M ethods for L inear-P hase FIR Filters

H istorically, the design m eth od based on the use o f w in d ow s to truncate the im pulse resp onse hd(n) and obtaining the d esired spectral shaping, was the first m eth od p rop osed for designing linear-phase F IR filters. T h e frequency-sam pling m eth od and the C h ebyshev approxim ation m eth od w ere d ev elo p ed in the 1970s and have since b eco m e very popular in the design o f practical linear-phase FIR filters. T he major disadvantage o f the w in d ow d esign m eth od is the lack o f precise control o f th e critical frequencies, such as cop and tos , in th e design o f a lowpass F IR filter. T h e valu es o f w p and cos, in general, d ep en d on the typ e o f w indow and the filter length M.
Sec. 8.2
663
-10
-20
-3 0
I
I
o s o
"
-4 0
-50
1 - 60
-7 0
-8 0
-9 0
-100
-1 1 0
-1 2 0
0.02
0.04
0.06
0.08
0.10
Transition width ( A f ) Figure JL27 Curves o f 201og 10S versus A f for M 3, 4, 7, 8 , 15, 16, 31, 32, 63, 64. [From paper by Rabiner and SchafeT (1974b). Reprinted with permission of AT&T.]
T he frequency sam pling m eth od provides an im provem en t o ver the w in dow design m eth od , sin ce Hr (ui) is specified at the freq u en cies cok 2 n k / M or cok = n( 2k + 1 )/M and th e transition band is a m ultiple o f 2 n / M . This filter d e sign m eth od is particularly attractive w hen the F IR filter is realized eith er in the frequency dom ain by m eans o f th e D F T or in any o f the frequency sam pling re alizations. T h e attractive feature o f th ese realizations is that Hr (cok) is either zero or unity at all freq u en cies, excep t in the transition band. T h e C h eb ysh ev approxim ation m ethod provides total control o f the filter specifications, and, as a co n seq u en ce, it is usually p referable o v er the other tw o m eth od s. For a low pass filter, the specifications are given in term s o f the param eters < D p, (ds, $i, 52, and Af. W e can specify th e param eters cdp, u>s, Af and 8 , and o p tim ize th e filters relative to 82- B y spreading th e approxim ation error over the
664
Chap. 8
Magnitude ( d B )
Figure 8-28 Frequency response of M 31 FIR Hilbert transform filter designed by the frequency sampling method.
passband and the stopband o f the filter, this m eth od results in an optim al filter d e sign, in the sense that for a given set o f specifications just d escrib ed , the m axim um sid elob e level is m inim ized. T h e C hebyshev design p rocedure based on the R em ez exch an ge algorithm requires that w e sp ecify the length o f the filter, the critical freq u en cies iop and ws, and the ratio S2/Si- H o w ever, it is m ore natural in filter design to specify iop , cos , 5j, and S2 and to d eterm in e th e filter length that satisfies the sp ecification s. A lth ou gh there is n o sim ple form ula to determ in e the filter length from th e se specifications, a num ber o f approxim ations h ave b een prop osed for estim ating M from cop , ojs , Si, and S2. A particularly sim p le form ula attributed to K aiser for approxim ating M is A = - 2 0 1 o g ,,( ^ ) - 1 3 +1 14.6 A f where A f is the transition band, defined as A f = (cos - a)p) f l u . This form ula has b een given in the paper by R abiner et al. (1975). A m ore accurate form ula proposed by H errm ann et al. (1973) is M = w here, by definition, >00( 51, S2) = [0.005309(loglo < 5 ; )2 + 0 .0 7 1 14(log105i) - 0.4761 J(Iog10 S2) - [0.00266(log10 Sx)2 + 0.5941 lo g 10 5i + 0.4278] (8.2.96) >oo(Si, S2) - f ( S u S2 ) ( A f ) 2 * 2n }> + 1 A/ (8.2.95)
/(S i, 52) = 11.012 + 0.51244(log10 5i - log10 S2)
(8.2.97)
Sec. 8 2
665
TABLE 8.10 PARAMETERS A FOR M = 31 HILBERT TRANSFORM FILTER DESIGNED BY THE FREQUENCYSAMPLING METHOD
LINEAR PHASE FIR HILBERT TRANSFORM FREQUENCY SAMPLING METHOD F I L T E R L E N G T H =31 LOWER CUTOFF FREQUENCY UPPER CUTOFF FREQUENCY IMPULSE RESPONSE: H ( 0) H ( 1) H ( 2) H ( 3) H ( 4) H ( 5) H< 6) H ( 7) H ( 8) H ( 9) H (10) H (11) H(12) H (13) H(14) = -0.1342662E-03 = 0.2133148E-02 0.4848863E-02 = 0.2286159E-02 0.1423532E-01 = 0.1517075E-02 = 0.3001805E-01 = 0.5263533E-03 = 0 .5 5 7 4 7 2 1 E - 0 1 = -0.2281570E-03 = 0.1001032E+00
(RELATIVE) (RELATIVE)
= 0.5000000E-01 = 0.4500000E+00
-0. 5 3 3 8 3 2 6 E - 0 3 = 0.1949848E+00 = -0.3994641E-03 = 0.6307253E+00 H {15) = - 0 . 9 3 3 5 9 5 6 E - 0 6 H (16) s - 0 . 6 3 0 7 2 4 5 E + 0 0 H ( 1 7 ) = 0.3996222E-03 H (18) = - 0 . 1 9 4 9 8 5 3 E + 0 0 H (19) = 0 . 5 3 4 1 3 0 7 E - 0 3 H (20) = - 0 . 1 0 0 1 0 3 5 E + 0 0 H (21) = 0 . 2 2 8 5 3 3 8 E - 0 3 H (22) H(23) H(24) H(25) H (26) H (27) = = = = = -0.5574735E-01 -0.5263340E-03 -0.3001794E-01 -0.1517240E-02
-0.1423557E-01 = -0.2285915E-02 H (28) = - 0 . 4 8 4 8 2 1 5 E - 0 2 H (29) = - 0 . 2 1 3 3 8 0 0 E - 0 2 H (30) = 0 . 1 3 4 4 1 6 2 E - 0 3
666
Chap. 8
T h ese form ulas are extrem ely useful in obtaining a g o o d estim ate o f the filter length required to ach ieve the given specifications A / , 61, and 82 . T he estim ate is used to carry out the design and if the resulting 8 ex ceed s th e specified 82 , the length can b e increased until w e obtain a sid elo b e level that m eets the specifica tions.
8.3 DESIGN OF IIR FILTERS FROM ANALOG FILTERS

Just as in the design o f F IR filters, there are several m eth od s that can be used to design digital filters having an infinite-duration unit sam ple resp onse. T he tech niques described in this section are all based on converting an analog filter into a digital filter. A n a lo g filter design is a m ature and w ell d ev elo p ed field, so it is n ot surprising that w e begin the design o f a digital filter in the analog dom ain and then convert the design into the digital dom ain. A n analog filter can be described by its system function. M (8.3.1)
0
w here {a*} and {/S*} are the filter coefficients, or by its im pulse resp onse, which is related to Ha (s ) by the L aplace transform (8.3.2) A ltern atively, the analog filter having the rational system fun ction H ( s ) given in (8.3.1), can b e described by the linear con stan t-coefficien t d ifferen tial equation (8.3.3) w here x ( t ) d en o tes the input signal and ;y(f) d en o tes the ou tp u t o f the filter. E ach o f th ese three eq u ivalen t characterizations o f an an alog filter leads to alternative m eth od s for converting the filter in to the digital dom ain, as will be described in S ection s 8.3.1 through 8.3.4. W e recall that an analog linear tim einvariant system with system function H ( s ) is stable if all its p o le s lie in the left half o f the j-p la n e. C on sequ en tly, if the con version tech n iqu e is to be effective, it should p ossess th e follow in g d esirable properties: L T h e j Q axis in the s-p lan e should m ap in to the unit circle in the z-plane. T hus there w ill b e a direct relationship b etw een the tw o frequency variables in the tw o dom ains.
Sec. 8.3
Design of IIR Filers From Analog Fitters
667
2. T h e left-h alf plane (L H P ) o f the 5-plane should m ap into the inside o f the unit circle in the z-plane. T hus a stable analog filter w ill b e con verted to a stable digital filter. W e m en tio n ed in the p receding section that physically realizable and stable IIR filters cannot have linear phase. R ecall that a linear-phase filter m ust have a system function that satisfies the con d ition H ( z ) = z ~ n H ( z ~ 1) (8.3.4)
w here z ~ N represents a d elay o f N units o f tim e. B u t if this w ere the case, the filter w ou ld have a m irror-im age p o le ou tsid e the unit circle for every p o le inside th e unit circle. H en ce the filter w ou ld be unstable. C on sequ en tly, a causal and stable IIR filter cannot have linear phase. If the restriction o n physical realizability is rem oved , it is p ossib le to obtain a linear-phase IIR filter, at least in principle. T his approach in volves perform ing a tim e reversal o f the input signal x ( n ), passing x ( n) through a digital filter H( z ) , tim e-reversing the output o f H( z ) , and finally, passing th e result through H( z ) again. T his signal processing is com putationally cu m bersom e and appears to offer no advantages over linear-phase F IR filters. C on sequ en tly, w hen an application requires a linear-phase filter, it should be an F IR filter. In the design o f IIR filters, w e shall specify the desired filter characteristics for the m agnitude response only. T his d oes not m ean that w e con sid er the phase resp onse unim portant. Since the m agnitude and phase characteristics are related, as indicated in Section 8.1, w e specify the desired m agnitude characteristics and accept the phase response that is ob tain ed from the design m eth od ology.
8.3.1 IIR Filter Design by Approximation of Derivatives

O n e o f the sim plest m eth od s for converting an analog filter in to a digital filter is to approxim ate the differential eq u ation in (8.3.3) by an eq u ivalen t d ifferen ce eq u a tion. T his approach is often used to solve a linear con stan t-coefficien t differential eq u ation num erically on a digital com puter. For the derivative d y { t ) / d t at tim e t n T , w e substitute the b a c k w a r d di f ference ^ ( n T ) y ( n T l ) ] / 7 \ Thus dy ( t ) dt (g.3.5) w h ere T represents the sam pling interval and y(n) = y { n T ) . T h e analog differ en tiator w ith output d y ( t ) / d t has th e system function H ( s ) = s, w hile the digi tal system that produces the output [y(n) >-(n 1 ) ] /T has the system function H ( z ) = (1 z ~ y) / T . C on sequ en tly, as show n in Fig. 8.29, th e frequency-dom ain y ( n T ) - y i n T - T)
668 yit)
H(s) = s
(a)
Chap. 8
dyjt) dt
y(n)
H(Z)
i-z
(b)
y(n) - v(n -l) FigHre fL29 Substitution of the backward difference for the derivative implies the mapping s (1 z~l )/T.
equ ivalen t for the relationship in (8.3.5) is
1 - z "1
(8.3.6)
T h e second derivative d 2 y ( t ) / d t 2 is replaced by the secon d differen ce, which is derived as follow s: d 2y{t) dt2
i= nT
d dt
'dy(t)
t=nT
\ y ( n T) - y ( n T - T ) ] / T - [ y ( n T - T) - y { n T - 2 T ) ] / T
y ( n ) - 2 y ( n - 1) + y( n - 2)
r2
(8.3.7)
In the frequency dom ain, (8.3.7) is equ ivalen t to

2
2 z ] +
Z~2
( \
- I n z _1
2
(8.3.8)
T2
~ \
It easily fo llow s from th e discussion that th e substitution for th e Jfcth derivative of y ( t ) results in the eq u ivalen t frequency-dom ain relationship k s~ = t = I (8.3.9)
C onsequently, th e system function for the digital IIR filter ob tain ed as a result of the approxim ation o f the d erivatives by finite d ifferen ces is H(z) = (8.3.10)
w here Ha(s ) is th e system function o f th e analog filter characterized by th e differ ential equation given in (8.3.3). L et us investigate the im plications o f the m apping from th e s-p lan e to the z-plane as given by (8.3.6) or, equivalently,
1 2 1
1-sT If w e substitute s = jQ, in (8.2.11), w e find that
(8.3.11)
Sec. 8.3
Design of IIR Filters From Analog Filters
669
(8.3.12)
1 . QT 1+ &T2+ J \ + &T2
A s Q varies from oo to oo, the corresponding locus o f p oin ts in th e z-p lane is a circle o f radius \ and with cen ter at z as illustrated in Fig. 8.30. It is easily d em onstrated that the m apping in (8.3.11) takes p oin ts in the L H P o f the 5-plane in to corresponding p oints inside this circle in the z-plane and p oints in the R H P o f the 5-plane are m apped in to p oints ou tsid e this circle. C on sequ en tly, this m apping has the desirable property that a stable analog filter is transform ed into a stable digital filter. H ow ever, the p ossib le location o f the p oles o f the d igital filter are confined to relatively sm all freq u en cies and, as a co n se q u en ce, the m apping is restricted to the design o f low p ass filters and bandpass filters having relatively sm all resonant frequencies. It is n ot p ossib le, for exam p le, to transform a highpass analog filter into a corresponding highpass digital filter. In an attem pt to overcom e the lim itations in the m apping given ab ove, m ore com p lex substitutions for the derivatives have b een p rop osed . In particular, an Lth-order d ifferen ce o f the form
y(nT + kT) y(nT kT)
_ _
(8.3.13)
has b een p rop osed , w h ere {a*} are a set o f param eters that can b e selected to op tim ize the approxim ation. T h e resulting m apping b etw een the 5-plan e and the z-plane is n ow (8.3.14)
Unit circle
j-plane
Figure 8 3 0
The mapping s = (1 z l ) / T takes LHP in the J-plane into points inside the circle o f radius j and center r = j in the 2 -plane.
670
W h en z = w e have
L
Chap. 8
j Y ak sin w * *=i w hich is purely im aginary. T hus
(8.3.15)
2 x ^ 1 = ^ a* sin<uJt * k=l
(8.3.16)
is the resulting m apping b etw een the tw o frequency variables. B y proper choice o f the coefficients {a*} it is p ossib le to m ap the j l - axis in to th e unit circle. Fur therm ore, p oin ts in the L H P in 5 - can b e m apped into points in sid e the unit circle in z. D esp ite achieving the tw o d esirable characteristics w ith th e m apping of (8.3.16), the prob lem o f selectin g the set o f coefficients (a*} rem ains. In general, this is a difficult problem . Since sim pler tech n iqu es exist for converting analog filters in to IIR digital filters, w e shall not em p hasize the u se o f the Lth-order d ifferen ce as a sub stitu te for the derivative.
Example 83.1
Convert the analog bandpass filter with system function H{S) = (*'+0.1 )2 + 9 into a digital IIR filter by use of the backward difference for the derivative.
Solution Substitution for s from (8.3.6) into H(s) yields

H(z) =
(i^ + 0 .l) +9
r 2/(i + o . 2 r + 9.oir2) 2(1 + 0.17) , 1

1 + 0 . 2 7 + 9.01 f 2 l+ 0 .2 7 + 9 .0 ir 2
The system function H(z) has the form of a resonator provided that T is selected small enough (e.g., T < 0 .1 ), in order for the poles to be near the unit circle. Note that the condition af < 4ai is satisfied, so that the poles are complex valued. For example, if T = 0.1 , the poles are located at
P i,2 = 0.91 jO.27
= 0.949e y i6 ' 5
W e n o te that th e range o f reson ant freq u en cies is lim ited to lo w freq u en cies, due to the characteristics o f the m apping. T h e reader is en cou raged to p lo t the frequency resp o o se H(co) o f th e digital filter for d ifferen t valu es o f T and com p are th e results w ith th e frequency resp onse o f the an alog filter.
Sec. 8.3
671
Example 8-3.2 Convert the analog bandpass f i l t e r in Example 8 . 3 . 1 into a di gi ta l IIR f i l t e r by use of the mapping
s = j(z~ z-')
Solution By substituting for j in H ( j ) , we obtain H{z) *
1
+9
z2T
~ z4 + 0.27 V + (2 + 9.01T2)z2 - 02Tz + 1 W e observe that t h i s mapping has introduced two additional poles in the con version from Ha(s) to H(z). As a consequence, the dig it al f i l t e ri s significantly more complex than the analog f i l t e r . This i sa major drawback to the mapping given above.
8.3.2 IIR Filter Design by Impulse Invariance
In the im pulse invariance m ethod, our objective is to d esign an IIR filter having a unit sam ple response h(n) that is the sam pled version o f th e im pulse response o f the analog filter. That is, h(n) = h( n T) n = 0,1,2,... (8.3.17)
w h ere T is th e sam pling interval. T o exam in e the im plications o f (8.3.17), w e refer back to S ection 4.2.9. R ecall that w hen a con tinu ou s tim e signal xa (t) with spectrum X a ( F) is sam pled at a rate Fs \ / T sam ples per second, the spectrum o f th e sam pled signal is the period ic rep etition o f the scaled spectrum FsX a( F) w ith period Fs . Specifically, the relationship is O O X ( f ) = Fs X a[ ( f r k)F, ] (8.3.18)
**-00
w h ere / = F / F s is the norm alized frequency. A liasin g occurs if the sam pling rate Fj is less than tw ice th e highest frequency con tained in X a (F). E xp ressed in the con text o f sam pling th e im pulse resp onse o f an analog filter with frequency response Ha (F), the digital filter w ith unit sam ple response h(n) s ha{ nT) has the frequency resp onse
OO
W ) = F,
Ha[ ( f - k)Fs]
(8.3.19)
** 0 0
or, eq u ivalen tly,
0 0
H(a>) = Fs Y H a [ (to - 2 n k ) F s] (8.3.20)
672
H J C iT )
Chap. 8
Figure 8J 1 Frequency response H (& ) of the analog f i l t e r and frequency re sponse of the corresponding d i g i t a lf i l t e rwith a l i a s i n g .
or H&T) = J ; Ha ( n (8.3.21)
Figure 8.31 depicts the frequency resp onse o f a low pass an alog filter and the frequency response o f the corresponding digital filter. It is clear that the digital filter w ith frequency resp onse H(<d ) has the fre quency response characteristics o f th e corresponding analog filter if the sam pling interval T is selected sufficiently sm all to com p letely avoid or at least m inim ize the effects o f aliasing. It is also clear that the im pulse invariance m eth od is in appropriate for d esigning highpass filters due the to spectrum aliasing that results from the sam pling process. T o investigate the m apping o f p oin ts b etw een the z-p lane and the j-p la n e im plied by the sam pling process, w e rely on a generalization o f ( 8.3 .2 1 ) which relates the z-transform o f h(n) to the L aplace transform o f ha (t). This relation ship is fl(z )i * = j Y H ( * ^ 3 -22)
Sec. 8.3
673
w here
O O H(z) = X > ( n ) z - n 73=0 O O * Y h M e ' sTn <8 -3-23)
N o te that w hen s = j Q , (8.3.22) reduces to (8.3.21), where th e factor o f j in Ha (Q) is suppressed in our notation. L et us con sid er the m apping o f p oin ts from the s-p lan e to the z-p lane im plied by the relation z - esT (8.3.24)
If w e sub stitu te j = a + j Q. and express the com p lex variable z in polar form as e = r e jai, (8.3.24) b eco m es r e j = eaTejnT Clearly, w e m ust have r = eaT (8.3.25)
co = Q.T
C on sequ en tly, a < 0 im plies that 0 < r < 1 and a > 0 im plies that r > 1. W hen a = 0, w e have r ~ 1. T h erefore, the L H P in s is m apped inside the unit circle in z and the R H P in s is m apped ou tsid e the unit circle in z. A lso , the j Q- ax\ s is m apped into th e unit circle in z as indicated above. H o w ever, the m apping o f the y'fi-axis in to the unit circle is n ot on e-to -o n e. Since co is unique over th e range (tt, tt), the m apping co = Q.T im p lies that the interval n / T < 2 < n / T m aps into the corresponding values o f tt < co < n. Fur therm ore, the frequency interval n / T < Cl < 3 n / T also m aps in to the interval n < co < 7i and, in general, so d o es the interval (2k - V ) n / T < 2 < (2k -(- \ ) t t / T , w hen k is an integer. T hus the m apping from th e analog frequency 2 to the fre q uency variable co in the digital dom ain is m any-to-on e, w hich sim ply reflects the effects o f aliasing due to sam pling. Figure 8.32 illustrates the m apping from the j-p la n e to the z-p lane for the relation in (8.3.24). T o exp lore further the effect o f the im pulse invariance design m eth od on th e characteristics o f the resulting filter, let us express the system function o f the analog filter in partial-fraction form. On the assum ption that the p o les o f the analog filter are distinct, w e can write
N
Ha(s) = Y
(8.3.26)
w here {/>*} are the p o les o f the analog filter and {c*} are the coefficients in the partial-fraction exp an sion . C onsequently,
N
M O = Y CkePkt k- i
(8.3.27)
674
Chap. 8
Figure 8_32 The mapping of z e*T maps strips of width 2 n / T (for cr < 0) in the 5 -plane into points in the unit circle in the z-plane.
If w e sam ple h a(t) p eriodically at t ~ nT, w e have

h(n) = h A n T )
(8.3.28) N ow , with the substitution o f (8.3.28), the system fun ction o f the resulting digital IIR filter b eco m es H( z ) = ] T / i ( n ) z - n w *0
oo / N
= E n=0 \*=1
N
(8.3.29) i= l =o T he inner sum in (8.3.29) con verges b ecau se p* < 0 an d yields (8.3.30) T herefore, the system function o f the digital filter is
H (z) = Y - ------- f
f t
1 - e Tz - 1
(8.3.31)
W e observe that the digital filter has p o les at

Zk = e
_
ffPkT
k = 1 ,2 ,..., A T
(8.3.32)
Sec. 8.3
Design of HR Fitters From Analog Filters
675
A lth o u g h the p o les are m apped from the .y-plane to th e z-p lane by the relationship in (8.3.32), w e should em phasize that the zeros in th e tw o dom ains do not satisfy the sam e relationship. T herefore, the im pulse invariance m eth od d o es not correspond to the sim p le m apping o f points given by (8.3.24). T h e d ev elo p m en t that resulted in H ( z ) given by (8.3.31) was based on a filter having distinct p oles. It can b e generalized to in clu d e m ultiple-order p oles. For brevity, h ow ever, w e shall not attem pt to gen eralize (8.3.31).
Example 833 Convert the analog f i l t e r with system function
into a d i g it al IIR f i l t e r by means of the impulse invariance method. Solution W e note that the analog f i l t e rhas a zero at s = -0.1 and a pair of complexconjugate poles a t pk = - 0.1 j 3 as illustrated in Fig. 8 . 3 3 . W e do not have to determine the impulse response ha{t) in order to design the d i g i ta l IIR f i l t e r based on the method of impulse invariance. Instead, we directly determine H{z), as given by (8 . 2 . 3 1 ) , from the partial-fraction expansion of Ha(s). Thus we have
l
2
Then
i
Figure 8 3 3 Pole-zero locations for analog filter in Example 83.3.
676
Chap. 8
Since the two poles are complex conjugates, we can combine them to form a single two-pole f i l t e rwith system function
H{z) =
1 (e 017 cos3T)z~ 1 ( 2e~0 lT cos 3T)z~l + e~2Tz~
The magnitude of the frequency response characteristic of t h i sf i l t e ri s plotted in Fi g. 8.34 for T = 0. 1 and T = 0 . 5 . For purpose of comparison, we have also plotted the magnitude of the frequency response of the analog f i l t e r in Fig. 8 .35, W e note that aliasing i s significantly more prevalent when T 0.5 than when T = 0 . 1 . Also, note the s h i f t of the resonant frequency as T changes.
Figure 8.35 Frequency response of analog filler in Exam ple 8.3.3.
T he preceding exam ple illustrates the im portance o f selectin g a sm all value for T to m inim ize the effect o f aliasing. D u e to the presen ce o f aliasing, the im pulse invariance m ethod is appropriate for the design o f low p ass and bandpass filters only.
8.3.3 IIR Filter Design by the Bilinear Transformation

T h e IIR filter design tech n iqu es describ ed in the p receding tw o section s have a severe lim itation in that they are appropriate only for low pass filters and a lim ited class o f bandpass filters. In this section w e describe a m apping from th e J-plane to th e z-plane, called the bilinear transform ation, that overcom es the lim itation o f th e other tw o design
Sec. 8.3
677
m eth od s d escribed previously. T h e bilinear transform ation is a conform al m apping that transform s the yft-axis in to the unit circle in the z-p lane only on ce, thus avoiding aliasing o f frequency com p onents. Furtherm ore, all points in the L H P o f s are m app ed inside th e unit circle in th e z-p lane and all p oin ts in the R H P o f s are m apped in to corresponding p oin ts ou tsid e the unit circle in the z-plane. T he bilinear transform ation can b e linked to the trapezoidal form ula for num erical integration. For exam ple, let us con sid er an analog linear filter with system function H( s ) = s + a T his system is also characterized by the differential equation dy(t) ------h ay( t ) bx( t ) (8.3.34) at Instead o f substituting a finite d ifferen ce for th e d erivative, suppose that w e in tegrate the derivative and approxim ate the integral by th e trapezoidal form ula. Thus y O ) = / y ' < j ) d r + y ( t 0) = J/o /> ' (8.3.35) (8.3.33)
w here y'(t) d en o tes the derivative o f y( t ). T he approxim ation o f the integral in (8.3.35) by the trapezoidal form ula at r ~ nT and t 0 = nT T yields y ( n T ) = j [ y \ n T ) + y' ( nT - T)] + y { n T - T) N o w the differential eq u ation in (8.3.34) evalu ated at t = n T yields y ' ( n T ) - a y ( n T ) + bx ( nT ) (8.3.37) (8.3.36)
W e u se (8.3.37) to substitute for th e derivative in (8.3.36) and thus ob tain a dif feren ce eq u ation for the equivalent discrete-tim e system . W ith y(n) = y(n T ) and x( n) = x( n T ) , w e obtain the result ^1 +
y
'J y ( n) ~
~ f ) y(~ n
~ f l x (n ) + x (n ~
(8.3.38)
T h e z-transform o f this differen ce equation is
( i + ~ ) y(z) - ( ! ~
) z ~l y ( z ) = T
(1 +
Z 1)X (Z)
C on sequ en tly, th e system function o f th e equivalent digital filter is = {Z) X( z ) ( bT pL)(\ + z~') 1 + a T / 2 ~ (1 a T pL)z~l
or, equivalently,
H(z) =
------
(8.3.39)
678
Clearly, the m apping from the
j - p lane
Chap. 8
to th e z-p lane is
5= H r r )
(8'3'40>
T his is called the bilinear transformation. A lth ou gh our derivation o f the bilinear transform ation w as p erform ed for a first-order d ifferential eq u ation , it holds, in general, for an N th -ord er differential equation. T o investigate the characteristics o f the bilinear transform ation, let z = r e i(u
s = a + ji2
T h en (8.3.40) can b e exp ressed as = 2 z ~ Tz +

1
2 r e Ja> - 1
T re>w + 1 r2 1 2r sin co = - ( + jT \ 1 + r 2 + 2r co s co J 1 + r 2 + 2r cos co, C onsequently,
2 r2- T 1 + r2 + 2 r cos <

_ A = 2 2 r sin co ------------- ---------- --
(8.3.41)
( 8 .3 .4 2 )
T l+ r
2r
cos co
First, w e n o te that if r < 1, then a < 0, and if r > 1, th en a > 0. C on se quently, the L H P in s m aps into the in sid e o f the unit circle in the z-plane and the R H P in s m aps in to the ou tsid e o f the unit circle. W hen r = 1, then a = 0 and
Q =
sin to
T 1 + c o s co (8.3.43) K
= - tan 2 T or, equivalently, oT co=2tan~1
(8.3.44)
T h e relationship in (8.3.44) b etw een the frequency variables in the tw o dom ains is illustrated in Fig. 8.36. W e ob serve that the entire range in 2 is m apped only o n ce in to the range n < a> < n . H ow ever, the m apping is high ly n onlinear. W e o b serve a frequency com p ression or f r equency warping, as it is usually called , due to the n onlinearity o f the arctangent function. It is also interesting to n o te that th e b ilinear transform ation m aps th e point s = oo into th e p oin t z 1. C on sequ en tly, th e sin gle-p ole low p ass filter in
Sec. 8.3
Design of IIR Fitters From Analog Filters
679
Figure 836
Mapping between the frequency variables to and Q resulting from the bilinear transformation.
(8.3.33), which has a zero at s = oo, results in a digital fi l t e r that has a zero at z = 1. Example 83.4 Convert the analog f i l t e r with system function
into a di gi ta l IIR f i l t e r by means of the bilinear transformation. The d i g it al f i l t e ri s to have a resonant frequency of wr = j t / 2. Solution F i r s t , we note that the analog f i l t e r has a resonant frequency =4 . This frequency i s to be mapped into cu r= j t /2 by selecting the value of the parameter T. From the relationship i n( 8 . 3 . 43 ), we must select T = \ i n order to have cor - it f l . Thus the desired mapping i s
The resulting d i g it al f i l t e r has the system function

0.128 + 0.006Z 1 - 0 .1 2 2 r 2 (Z) 1 + 0 .0 0 0 6 z - 5 + 0 .9 7 5 z -2
W e note that the coefficient of the z~l term in the denominator of H(z) i s extremely small and can be approximated by zero. Thus we have the system function
0.128 + 0 .0 0 6 z-1 - 0 .1 2 2 z -2
;---------T h is filter h as p o le s at p i .2 = 0.987e,r/2
680
and ze ro s at zi
Chap. 8
.2 =
- 1 , 0.95
Therefore, we have succeeded in designing a two-pole filter that resonates near u) = nfl . In this exam ple the param eter T w as selected to m ap th e resonant frequency o f the analog filter in to the desired reson ant frequency o f the digital filter. U sually, the design o f the digital filter begins w ith specifications in the digital dom ain, which in volve the frequency variable < d . T h ese sp ecification s in frequency are converted to the analog dom ain by m eans o f the relation in (8.3.43). T h e analog filter is then d esign ed that m eets these sp ecification s and con verted to a digital filter by m eans o f the bilinear transform ation in (8.3.40). In this procedure, the param eter T is transparent and m ay b e set to any arbitrary value (e.g., T = 1). T he follow ing exam p le illustrates this point. Example 8.3.5 Design a single-pole lowpass digital filter with a 3-dB bandwidth of 0.2jt, using the bilinear transformation applied to the analog filter His) = where is the 3-dB bandwidth of the analog filter.
Solution The digital filter is specified to have its -3-d B gain at wc = 0.2n. In the frequency domain of the analog filter we 0 .2 n corresponds to
fZ c = tan0.l7r
_
065
7
Thus the analog filter has the system function 0.65/7

j + 0 .6 5 / 7
This represents our filter design in the analog domain. Now, we apply the bilinear transformation given by (8.3.40) to convert the analog filter into the desired digital filter. Thus we obtain II(z) ( ) ~ 1 - 0 .5 0 9 Z - 1
where the parameter 7 has been divided o u t The frequency response of the digital filter is 0.245(1 + e~ju)
]
1 - 0.5 0 9 c-.'* '
At w = 0, H( 0) = 1, and at a> = 0.2?r, we have |//(0.2jr)| = 0.707, which is the desired response.
Sec. 8.3
681
8.3.4 The Matched-z Transformation

A n o th er m eth od for converting an analog filter into an eq u ivalen t digital filter is to m ap the p o les and zeros o f H ( s ) directly into p o les and zeros in the z-plane. Sup p o se that the system function of the analog filter is exp ressed in the factored form M Y [ ( S ~ z k) H ( s ) = *=!-----------r > *=i
pa
(8.3.45)
w here {z*} are the zeros and {p*} are the p o les of the filter . T hen the system function for the digital filter is M Y [ ( l - e ^ Tz ~ x) (8.3.46) f ] a - ^ V ' ) *=i w here T is the sam pling interval. Thus each factor o f the form (s a) in H( s ) is m apped into the factor (1 eaTz ~ l ). This m apping is called the matched- z transformation. W e ob serve that the p oles ob tain ed from the m atched-z transform ation are identical to the p oles ob tain ed with the im pulse invariance m eth od . H ow ever, the tw o techniques result in d ifferent zero positions. T o preserve the frequency resp onse characteristic o f the analog filter, the sam pling interval in th e m atched-z transform ation m ust b e properly selected to yield the p o le and zero locations at the eq u ivalen t position in the z-plane. Thus aliasing m ust be avoid ed by selectin g T sufficiently sm all.
8.3.5 Characteristics of Commonly Used Analog Filters

A s w e have seen from our discussion ab ove, IIR digital filters can easily b e o b tained by b eginning with an analog filter and then using a m apping to transform the s -plane into the z-plane. Thus the design o f a digital filter is reduced to designing an appropriate analog filter and then p erform ing the con version from H( s ) to H( z ) , in such a w ay so as to preserve as m uch as p ossib le, the desired characteristics o f the analog filter. A n a lo g filter design is a w ell-d evelop ed field and m any b ook s have b een written on the subject. In this section w e briefly describe the im portant character istics o f com m on ly used analog filters and introduce the relevan t filter param eters. O ur discussion is lim ited to low pass filters. S u b seq u en tly, w e describe several frequency transform ations that convert a low p ass p rototyp e filter in to either a bandpass, highpass, or band-elim ination filter.
682
Chap. 8
Butterworth filters. L ow pass B utterw orth filters are all-p ole filters char acterized by the m agnitude-squared freq u en cy response
1H (S2) 12 = l + ( 2 / n c)2* =
i + cH n/np)M
(8'3 47)
where N is the order o f the filter, 2C is its 3-dB freq u en cy (usually called the cu toff freq u en cy), Qp is the passband ed ge frequency, and 1/(1 + e 2) is th e banded ge valu e o f |H (2)|2. Since H ( s ) H ( s) evaluated at j = j l is sim ply equal to |//( 2)|2, it fo llo w s that
H(sW-s) - t
= ( - l ) 1//v = e j & + * / * and h ence sk - n ce j 7 r/2e j(2k+l)7r/2N
t f W
< 8 3'48)
T he p o les o f H ( s ) H ( s) occur on a circle o f radius 2C at equally spaced points. From (8.3.48) w e find that k = 0,1,..., N - 1
* = 0 , 1 ......... N 1
(8.3.49)
For exam ple, Fig. 8.37 illustrates the p ole p osition s for an N = 4 and N 5 Butterw orth filters. T he frequency response characteristics o f the class o f B utterw orth filters are show n in Fig. 8.38 for several valu es o f N. W e n ote that (2)|2 is m on oton ic in both the passband and stopband. T h e ord er o f the filter required to m eet an attenuation 82 at a specified frequency 2, is easily d eterm in ed from (8.3.47). Thus at 2 = 2* w e have
*
i + e H n I / a p ) 2N
-2
61
and hence
_ tog [ ( l / g ) - 1 ] _ . W O 21og(G,/W log(S J,/n,)
w here, by definition, < 5 2 = l / V l + 8 2. T hus th e B utterw orth filter is com p letely characterized by the param eters N, 82, e, and the ratio Qs / l p. Example 8.3.6 Determine the order and the poles of a lowpass Butterworth filter that has a -3-d B bandwidth of 500 Hz and an attenuation of 40 dB at 1000 Hz. Solution The critical frequencies are the 3-dB frequency Qc and the stopband frequency 2 ,, which are = IOOOjt 2 , =
2000jt
Sec. 8.3
Design of HR Filters From Analog Filters
683
Figure 8.37
Pole positions for Butterworth filters.
For an attenuation of 40 dB, S2 = 0.01. Hence from (8.3.50) we obtain log,o ( 10* -
1)
21ogio 2
= 6.64
To meet the desired specifications, we select N = 7. The pole positions are st = \0007rei{nfl+cu+l,*/u] 0,1.2.....
Chebyshev filters. T h ere are tw o types o f C h eb ysh ev filters. T yp e I C h eb ysh ev filters are all-pole filters that exhibit equiripple b eh avior in th e passband and a m o n o to n ic characteristic in the stopband. O n the other hand, the fam ily o f type II C h eb ysh ev filters contains both p o le s and zeros and exh ib its a
684
l#(2)P
Chap. 8
m on o to n ic b eh avior in the passband and an eq u irip p le b eh avior in the stopband. T h e zeros o f this class o f filters lie on the im aginary axis in th e s -plane. T he m agnitude squared o f the frequency resp onse characteristic o f a type I C h ebyshev filter is given as |t f ( f i )|2 = ------- J-----------i + 2r 2 ( n / n , ) (8.3.51)
w here e is a param eter o f the filter related to the ripple in th e passband and Tn (x ) is the W th-order C hebyshev polyn om ial defined as
t _
\ cos(Af c o s- 1 * ),
|x |
<1
, ox
c o s h w c o d .- 1* ),
|X|>1
( 8 J '52)
Sec. 8.3
685
T he C h eb ysh ev p olyn om ials can be generated by the recursive equation Tn+ i (.x ) = 2 xTN(x) - Tn i ( x ) A T= 1,2,... (8.3.53)
w h ere 7 o U ) = 1 and T\(x) = x. From (8.3.53) w e obtain T^C*) = 2*2 1- T-six) 4 r 3 3 x , and so on. Som e o f the p roperties o f these polyn om ials are as follow s:
1 . | 7 VU) | < 1 for all 1*1 < 1. 2 . 7V(1) = 1 for all N.

3. A ll the roots o f th e p olyn om ial T?j(x) occur in th e interval 1 < * < 1. T h e filter param eter is related to the ripple in th e passband, as illustrated in Fig. 8.39, for N odd and N even. For N odd, TN( 0) = 0 and h en ce | / / ( 0 )|2 = 1. O n th e other hand, for N even, TN(0) = 1 and h en ce | / / ( 0 )|2 = 1/(1 -f e 2). A t the band ed ge frequency SI = Qp, w e have T,y(l) = 1 , so that
or, equivalently.
w h ere < $ i is the value o f the passband ripple. The p o les o f a type I C hebyshev filter lie on an ellip se in the j-p la n e with m ajor axis (8.3.55) iffrti)!*
N odd Figure 8-39
N even Type I Chebyshev filter characteristic.
686
and m inor axis -1
20
Chap. 8
(8.3.56)
w here
is related to e according to the equation vT
+72 + 1
1/ N
(8.3.57)
T he p ole locations are m ost easily d eterm in ed for a filter o f ord er N by first locating the p o les for an equ ivalen t iVth-order B utterw orth filter that lie on circles o f radius rj or radius r2, as illustrated in Fig. 8.40. If w e d en o te the angular p osition s o f the p o les o f the B utterw orth filter as (8.3.58) 2N then the p osition s o f the p oles for the C hebyshev filter lie on the ellip se at the coordinates (xk, yt ) , k = 0, 1 , . . . , N 1 , w here xk = r 2 cos <pk, yk = r i sin < pk. k = Q ,l,...,N -l (8.3.59) k = 0,1, . . . , N - 1 (2k +
1)7T
Figure 8.40
Determination o f the pole locations for a Chebyshev filter.
Sec. 8.3
687
A type II C hebyshev filter contains zeros as w ell as p oles. T h e m agnitude squared o f its frequency resp onse is given as \ H (2)|2 = (8.3.60)
w here Ty( x) is, again, the N th -ord er C hebyshev p olyn om ial and 2* is the stopband frequency as illustrated in Fig. 8.41. T h e zeros are located on the im aginary axis at th e points Sk = j sin f a k = 0,1,..., N - 1 (8.3.61)
T h e p o les are located at the points (vk, w k), w here vk = Slsxk * =
0, 1 ,
-1
(8.3.62)
yf*l + Vk
wk Qyk y/xk + y k 2 where j**} and {yk} are defined in (8.3.59) with 0 now related to the ripple in the stopband through the equation * = 0,1,..., A f-1 (8.3.63)
1/jV
1
+ J l - S 2' (8.3.64)
From this description, w e o b serve that the C hebyshev filters are characterized by the param eters N, e, S2, and the ratio Q .s / Sl p . For a given set o f specifications IWDI2 |tf(n)P
n , n,
N odd Figure 8.41 N even Type II Chebyshev filters.
688
Chap. 8
on (, S2, and Q .s/Q.p, we can detennine the order of the fi lter from the equation log N =
l(v
fl-Sl + y /l-S ld + ^ j/e S !

(8.3.65)
log (ns/ Q p) + y/ ( n s / n p)2 - 1 cosh (5/e)

co sh -1 (2, / ^ ) w here, by definition, S2 =
1 /V l + < 5 2.
Example 83.7 Determine the order and the poles of a type I lowpass Chebyshev f i l t e r that has a 1-dB ripple in the passband, a cutoff frequency 2 P = IOOOjt, a stopband frequency of 2000jt, and an attenuation of 40 dB or more for 2> S 2 s. Solution F i r s t , we determine the order of the f i l t e r . W e have
101og1 0 ( l + e2) = 1
1 + 2 = 1.259
2 = 0.259
e = 0.5088 Also, 20 log10< 5 2 = -40

82
0.01
Hence from (8.3.65) we obtain N = ^Sio 196.54 log10( 2 +V3)
= 4.0 Thus a type I Chebyshev f i l t e r having four poles meets the specifications. The pole positions are determined from the relations in (8.3.55) through ( 8 . 3 . 5 9 ) . F i r s t , we compute / ? ,n, and r2. Hence P = 1.429
ri = 1.062P
r 2 = 0.3650, The angles { 0* } are < P t= n
2+
(2k + 1) 7 T
8
k = 0 ,1,2 ,3
Therefore, the poles are located a t

*1 + j y 1 = ~0.13972P j0.979np
+ jyi = 0.3372,, _/'0.4056Qp
Sec. 8.3
689
T h e filter specifications in E xam p le 8.3.7 are very sim ilar to the specifications g iven in E xam p le 8.3.6, w hich in volved the design o f a B utterw orth filter. In that case th e num ber o f p o les required to m eet th e sp ecification s was seven . O n the other hand, the C hebyshev filter required on ly four. T h is result is typical o f such com parisons. In general, the C hebyshev filter m eets the specifications with a few er num ber o f p o les than the corresponding B utterw orth filter. A ltern atively, if w e com p are a B utterw orth filter to a C h eb ysh ev filter having th e sam e num ber o f p o les and the sam e passband and stopband specifications, the C hebyshev filter will h a v e a sm aller transition bandwidth. For a tabulation o f the characteristics o f C h eb ysh ev filters and their p o le-zero locations, the in terested reader is referred to th e h an d b ook o f Z verev (1967).
Elliptic filters. E llip tic (or C auer) filters exhibit eq uiripple b eh avior in b oth the passband and the stopband, as illustrated in Fig. 8.42 for N odd and N even. T h is class o f filters contains both p o les and zeros and is characterized by the m agnitude-squared frequency response
|H (^ )|2 = 1
u.
277~; 0 / o ,
(8-3 -66)
w h ere UN{:r) is the Jacobian elliptic function o f order N , w hich has b een tabulated by Z v erev (1967), and e is a param eter related to the passband ripple. T h e zeros lie on the ;'fi-axis. W e recall from our discussion o f F IR filters that the m ost efficient designs occur w hen w e spread the approxim ation error equally over the passband and the stop b an d . E llip tic filters accom plish this ob jective and, as a con seq u en ce, are the m ost efficien t from the view p oin t o f yielding the sm allest-ord er filter for a given IW (Q)|2
1 + e2
\S
N odd Figure 8.42 Magnitude-squared frequency characteristics o f elliptic filters.
690
Chap. 8
set o f specifications. E quivalently, w e can say that for a given order and a given set o f specifications, an elliptic filter has the sm allest transition bandwidth. T he filter order required to ach ieve a given set o f specifications in passband ripple Si, stopband ripple 62, and transition ratio ftp /ft* is given as K ( n pM ) K U i - f i / s * ) ) N = --------------- v L K W K ( y i - ( 8 , / n , )2j w h ere K ( x ) is the co m p lete elliptic integral o f the first kind, defined as
K (X ) =
(8.3.67)
r a
- ........... .
v 1 jc 2 s u v
d$
(8.3.68)
6
'0
and 62 = 1 / V l + 62. V alu es o f this integral have b een tabu lated in a num ber o f texts [e.g., th e b ook s by Jahnke and E m d e (1945) and D w igh t (1957)]. T h e passband ripple is 10 lo g 10( l + e2). W e shall n ot attem pt to describe ellip tic functions in any d etail b ecau se such a discussion w ould take us to o far afield. Suffice to say that com p u ter program s are available for designing ellip tic filters from the frequency specifications indicated above. In v iew o f the optim ality o f ellip tic filters, the reader m ay qu estion the reason for considering the class o f Butterw orth or th e class o f C h eb ysh ev filters in practical applications. O n e im portant reason that th ese other types o f filters m ight be prefer able in som e applications is that they p ossess b etter phase resp on se characteristics. T h e phase resp onse o f ellip tic filters is m ore non linear in the passband than a com parable B utterw orth filter or a C h eb ysh ev filter, especially near the band edge.
Bessel filters. B e sse l filters are a class o f all-p ole filters that are charac terized by the system function
H( s ) =
B n (s )
( 8. 3. 69)
w h ere B n (s ) is the N th -ord er B e sse l polynom ial. T h ese p olyn om ials can b e ex p ressed in the form
N
B * (j) = a * j *
0
(8-3.7)
w h ere the coefficients {a*} are given as
&N - t y
,
* - ( U .... N <8371)
' 2 - > * ! ( * - * ) !
A ltern atively, the B essel p olyn om ials m ay b e gen erated recursively from the rela tion B n (s) = (2 N - 1)B*_j(i) + s 2B n . 2(s) with Bo(s) = 1 and fli(s ) = s + 1 as initial conditions. (8.3.72)
Sec. 8.3
Magnitude
691
Phase
Figure 8.43 Magnitude and phase responses o f Bessel and Butterworth filters of order N = 4.
A n im portant characteristic o f B essel filters is the linear-phase response o ver the passband o f the filter. For exam p le, Fig. 8.43 sh ow s a com parison o f th e m agnitude and phase resp onses o f a B e sse l filter and B utterw orth filter o f order N = 4. W e n o te that the B essel filter has a larger transition bandwidth, but its phase is linear w ithin the passband. H ow ever, w e should em p hasize that the
692
Chap. 8
linear-phase chacteristics o f the analog filter are d estroyed in th e p rocess o f con verting the filter into the digital dom ain by m eans o f the transform ations decribed previously.
8.3.6 Some Examples of Digital Filter Designs Based on the Bilinear Transformation
In this section w e presen t several exam p les o f digital filter d esign s ob tain ed from analog filters by applying the bilinear transform ation to con vert H( s ) to H(z). T h ese filters design s are p erform ed with the aid o f o n e o f several softw are packages now available for use on a p ersonal com puter. A low pass filter is design ed to m eet specifications o f a m axim um ripple o f - dB in the passband, 60-dB attenuation in th e stopband, a passband ed ge fre quency o f (i)p = 0.25;r, and a stopband ed ge frequency o f cos = 0.30tt. A B utterw orth filter o f order N = 37 is required to satisfy the specifications. Its frequency response characteristics are illustrated in Fig. 8.44. If a C heby shev filter is used, a filter o f order N = 13 satisfies the sp ecification s. T he fre quency resp onse characteristics for a type I and type II C h eb ysh ev filters are show n in Figs. 8.45 and 8.46, respectively. T he type I filter has a passband rip p le o f 0.31 dB. Finally, an elliptic filter o f order N 7 is d esign ed which also satisfied the specifications. For illustrative purposes, w e sh ow in T able 8.11, the num erical valu es for the filter param eters and the resulting frequency specifica tions are show n in Fig. 8.47. T he follow in g n otation is used for th e param eters in the function H(z): , r r W ' > + W
' 1)z_1 + k(i, 2) z ~2
/C Q w (8.3.73)
A lth o u g h w e have described only low pass analog filters in th e p reced in g section, it is a sim ple m atter to convert a low pass analog filter into a ban d pass, bandstop, or highpass analog filter by a frequency transform ation, as is d escrib ed in Section 8.4. T he bilinear transform ation is then applied to convert the an alog filter into an equivalent digital filter. A s in the case o f the low p ass filters describ ed ab ove, the entire design can b e carried ou t on a com puter.
8.4 FREQUENCY TRANSFORMATIONS

T he treatm ent in th e p receding section is focu sed prim arily on the d esign o f lowpass IIR filters. If w e wish to design a highpass or a bandpass or a bandstop filter, it is a sim ple m atter to take a low pass prototyp e filter (B u tterw orth , Chebyshev, elliptic, B e sse l) and perform a frequency transform ation. O n e possib ility is to perform the frequency transform ation in th e analog dom ain and then to convert the analog filter into a corresp on d in g digital filter by a m apping o f the 5-plane in to the z-plane. A n alternative approach is first to
Sec. 8.4
Frequency Transformations
693
Figure 8.44
Frequency response characteristics of a 37-order Butterworth filter.
convert the analog lowpass f i l t e r into a lowpass digital f i lt er and then to transform the lowpass digital f i l t e r into the desired digital f i l t e r by a digital transformation. In general, these two approaches yield different results, except for the bilinear transformation, in which case the resulting f i l t e r designs are identical. These two approaches are described below.
8.4.1 Frequency Transformations in the Analog Domain
First, we consider frequency transformations in the analog domain. Suppose that we have a lowpass f i l t e r with passband edge frequency Qp and we wish to convert
694
Chap. 8
Figure 8.45
Frequency response characteristics o f a 13-order type I Chebyshev filter.
it to another low pass filter with passband ed ge frequency 2^,. T h e transform ation that accom plishes this is j y s (low pass to low pass) (8.4.1)
Thus w e obtain a low pass filter w ith system function Hi ( s ) = Hp [(Qp/ t t p)s], w here Hp(s) is th e system function o f the p rototyp e filter w ith passband ed ge frequency Qp.
Sec. 8.4
695
Figure M
Frequency response characteristics o f a 13-order type II Chebyshev filter.
If w e w ish to convert a low pass filter in to a highpass filter with passband ed g e frequency Q !p, the desired transform ation is s -------(low pass to highpass) (8.4.2) s T h e system function o f th e highpass filter is Hk(s) Hp (Q.pQ!p/s). T h e transform ation for converting a low pass analog filter with passband ed ge frequency S2P in to a band filter, having a lo w er band ed ge frequency Qj and an upper band ed g e frequency 2, can b e accom plished by first converting the low pass
696
Chap. 8
TABLE 8.11
FILTER COEFFICIENTS FOR A 7-ORDER ELLIPTIC FILTER

I N F I N I T E I M P U L S E R E S P O N S E (IIR) ELLIPTIC LOWPASS FILTER
UNQUANTIZED COEFFICIENTS FILTER ORDER = 7 SAMPLING FREQUENCY = 2.000 KILOHERTZ I. I 2 A(I, 1) A(I, 2) B (I , 0} B (I , 1) B (I , 2)
-.790103 .000000 .104948 . 104948 .000000 .714 0 8 8 .102450 -.007817 -1.517223 .102232 3 -1.421773 .861895 .420 1 0 0 -.399842 .419864 4 .962252 -1.387447 -.826743 .714841 . 714929 *** C H A R A C T E R I S T I C S O F D E S I G N E D F I L T E R *** BAND 1 BAND 2 LOWER BAND EDGE .30000 .00000 1.00000 UPPER BAND EDGE .25000 NOMINAL GAIN 1.00000 .00000 NOMINAL RIPPLE .00100 .05600 MAXIMUM RIPPLE .00071 . 04910 R I P P L E I N DB .41634 -63.00399
filter in to another low pass filter having a band ed ge frequency Sl'p = perform ing the transform ation s S s(lu S 2/) (low pass to bandpass)
1 and then
(8.4.3)
E quivalently, w e can accom plish the sam e result in a single step by m ean s o f the transform ation s 2 + 2 / S2 s y Qp ------ j( 2u 2i) where 2/ = low er band ed ge frequency S2U = upper band ed ge frequency Thus w e obtain ( s 2 + 2i 2B \ (low pass to bandpass) (8.4.4)
Finally, if w e w ish to convert a low pass analog filter w ith band ed g e frequency 2p in to a bandstop filter, th e transform ation is sim ply th e in verse o f (8.4.3) with the additional factor Slp serving to norm alize for th e band ed ge freq u en cy o f the low pass filter. T h u s th e transform ation is
j lpS2 - ^ -
(lowpass to bandstop)
(8.4.5)
Sec. 8.4
697
Figure 8.47
Frequency response characteristics o f a 7-order elliptic filter.
which leads to
T h e m appings in (8.4.1), (8.4.2), (8.4.3), and (8.4.5) are sum m arized in T able 8.12. T h e m appings in (8.4.4) and (8.4.5) are non linear and m ay appear to distort the frequency response characteristics o f the low p ass filter. H ow ever, the effects o f the nonlinearity on the frequency resp onse are m inor, prim arily af fecting the frequency scale but preserving the am plitude resp onse characteristics o f the filter. Thus an equiripple low p ass filter is transform ed in to an equiripple b andpass or b andstop or highpass filter.
698
Chap. 8
TABLE 8.12
FREQUENCY TRANSFORMATIONS FOR ANALOG FILTERS (PROTOTYPE LOWPASS FILTER HAS BAND EDGE FREQUENCY ftp) Band edge frequencies o f new filter
Type of transformation Lowpass
Transformation np s * ^ ^ s _ s2 + Sit S --- iin ------------- p s ( n u - n () s --- ---------- p s2 + n u n . ftp
Highpass Bandpass Bandstop
ft/, ftp Q t , Clu
Example 8.4.1 Transform the single-pole lowpass Butterworth f i l t e r with system function
H^ s) = - T 7 T
into a bandpass f i l t e r with upper and lower band edge frequencies 2 and t i v e l y . Solution The desired transformation i s given by ( 8 . 3 . 4 ) . Thus we have
H (s) =
respec
1
-f-
i(n - a,) +
(K K/)j s 2 + (2 S2 j)s +
The resulting f i l t e r has a zero at s = 0 and poles at
8.4.2 Frequency Transformations in the Digital Domain

A s in the analog dom ain, frequency transform ations can be perform ed on a digital low pass filter to convert it to either a bandpass, bandstop, or highpass filter. T h e transform ation in volves replacing th e variable z -1 by a rational function g (z -1 ), w hich must satisfy the follow in g properties. L T he m apping z _1 g (z -1 ) m ust m ap p oin ts inside th e unit circle in the z-plane in to itself. 2. The unit circle m ust also b e m apped into itself.
Sec. 8.4
699
C ondition (2) im plies that for r = 1, e~Ja g ( e - J,) = g(u})
It is clear that w e m ust have |g(ti>)| = 1 for all all-pass. H en ce it is o f the form
Z
oj .
That is, the m apping m ust b e
-ak
S ( z - ) = n f r akz-
(8.4.6)
w h ere \ak \ < 1 to ensure that a stable filter is transform ed into another stable filter (i.e., to satisfy co n d ition 1 ). From the general form in (8.4.6), w e obtain the desired set o f digital trans form ations for converting a prototyp e digital low pass filter in to either a bandpass, a bandstop, a highpass, or an oth er low pass digital filter. T h ese transform ations are tabulated in T able 8.13.
TABLE 8.13 FREQUENCY TRANSFORMATION FOR DIGITAL FILTERS (PROTOTYPE LOWPASS FILTER HAS BAND EDGE FREQUENCY a>P)
Type of transformation
Transformation
Parameters (o'p band edge frequency of new filter
Lowpass
1 - az'
sin[(^p - w'p ) p \
sin[(a>p + w'p)p.]
w'p = band edge frequency new filter
Highpass
1 + az~x
cos[(top + to'p ) /2] cos[(a>p - a>p)/2] wi = lower band edge frequency toy = upper band edge frequency
Bandpass
z 1 aiz + a ; a2z~2 - a\z~l + 1
2aK/ ( K + 1) 0l = a2 = ( K - 1 )/(K + 1)
COs[ (tD, + Ci>;)/2]

c o s K c D j, c d / ) / 2 ]
K as COt
0> 1
tan -r-
o> i = lower band edge frequency co = upper band edge frequency
Bandstop
a{ = - 2a/(K + 1) 02 = {I - K) / ( 1 + K)
ai z- 1 + 1
cos[( i> u + <u;)/2]
c o s [ ( a ) a M a>i CL
700 Example 8.42
Chap. 8
Convert the single-pole lowpass Butterworth filter with system function H(z)
0 .2 4 5 ( 1 + Z ' 1) 1 - 0 .5 0 9 Z -1
into a bandpass filter with upper and lower cutoff frequencies a> u and a>i, respectively. The lowpass filter has 3-dB bandwidth cop = 0.2jt (see Example 8.3.5).
Solution The desired transformation is

Z
_i
---- ----------z------------;----- -
z~2 - a i z ~ l + a 2 a 2z ~2 - ai z ' 1 + 1
where a\ and a 2 are defined in Table 8.13. Substitution into H(z) yields
0.245
H(z) =
z~l ~ aiz~' + a2 a2z ~2 - fliZ-1 + 1
1 + o.509
\ a 2z~2 -
+1/
^ ________0.245(1 - a2)( 1 - r 2) ________ (1 + 0.509o2) - 1.509fliz1 + (a2 + 0.509)z~2

Note that the resulting filter has zeros at z = 1 and a pair of poles that depend on the choice of o> u and a> t. For example, suppose that o> u = 3jt/5 and a> i = 2tt/5. Since wn = 0 . 2 ? r , we find that K = 1 ,a2 = 0 , and ay = 0 . Then
() l+0.509z~2 This filter has poles at z = y'0.713 and hence resonates at < d = np..
Since a frequency transform ation can be perform ed eith er in the analog d o m ain or in the digital dom ain, the filter design er has a ch o ice as to w hich ap proach to take. H o w ever, som e caution m ust b e exercised d ep en d in g on the types o f filters b ein g d esign ed . In particuiar, w e know that the im pulse invari an ce m ethod and th e m apping o f d erivatives are inappropriate to use in designing highpass and m any bandpass filters, due to the aliasing p rob lem . C on sequ en tly, o n e w ould n ot em p loy an analog frequency transform ation follow ed by con ver sio n o f the result into the digital d om ain by use o f these tw o m appings. Instead, it is m uch b etter to perform the m apping from an analog low pass filter in to a digital low pass filter by eith er o f th ese m appings, and th en to perform the fre q uency transform ation in the digital dom ain. Thus the p rob lem o f aliasing is avoided. In the case o f the b ilinear transform ation, w here aliasing is not a problem , it d o es n ot m atter w h eth er the frequency transform ation is p erform ed in the analog dom ain or in th e digital dom ain. In fact, in this case on ly, th e tw o approaches result in identical d igital filters.
-245(1 ~ Z~2)
Sec. 8.5
Design of Digital Filters Based on Least-Squares Method
701
8.5 DESIGN OF DIGITAL FILTERS BASED ON LEAST-SQUARES METHOD

E xcep t for th e im pulse invariance m eth od , the design tech n iqu es for IIR filters d e scribed in S ection 8.3 in volved the conversion o f an analog filter in to a digital filter by so m e m apping from the ,y-plane to the z-plane. A s an alternative, on e can design digital IIR filters directly in the z-dom ain w ithou t referen ce to th e analog dom ain. W e n ow describe several m eth od s for designing digital filters directly. In th e first three tech n iqu es, the P ade approxim ation m eth o d and least-squares design m eth od s, the specifications are given in the tim e d om ain and th e design is carried out in th e tim e dom ain. The final section describes a least-squares technique in w hich the design is carried out in the frequency dom ain.
8.5.1 Pade Approximation Method

Suppose that the desired im pulse resp onse hj { n) is sp ecified for n > 0. T he filter to b e d esign ed has the system function M
Y b k Zk
H(z) =
(8.5.1) where h(k) is its unit sam ple response. T he filter has L = M + N + 1 param eters, nam ely, th e coefficients [ak} and {&*}, w hich can b e selected to m inim ize so m e error criterion. T h e least-sq u ares error criterion is often u sed in optim ization problem s o f this type. Su p pose that w e m inim ize the sum o f the squared errors u (8.5.2) with respect to th e filter p aram eters {a*} and {>*}, w h ere U is som e p reselected upper lim it in the sum m ation. In g en eral, h(n) is a non linear function o f the filter param eters and h en ce th e m inim ization o f in v o lves th e solu tion o f a set o f n on linear equ ation s. H ow ever, if w e select th e upper lim it as U = L 1, it is p ossib le to m atch h(n) p erfectly to the desired resp onse hd (n ) for 0 < n < Af + N. T h is can b e achieved in th e follow in g m anner. T h e differen ce eq u ation for th e d esired filter is y (n ) = - a i y ( n -
1 ) - a 2y( n - 2) ---------- a Ny{n - N )

(8.5.3)
+ box(n) + b\ x( n - 1) H -------- 1 - b Mx( n - M)
702
Chap. 8
Suppose that the input to the filter is a unit sam ple [i.e., x( n) S(n)]. T hen the resp onse o f the filter is y(/i) = h(n) and h en ce (8.5.3) b eco m es h(n) = a\ h{n 1) ai hi n 2) -------a s h { n N) --------1 - b MS(n Af) -)- boS(n) -I- b\S(n 1) H Since S(n Jt) = 0 excep t for n = k, (8.5.4) red u ces to h(n) = a \ h ( n ~ 1) aj h(n 2 ) ----------a \ h ( n N ) + bn For n > Af, (8.5.4) b eco m es h(n) a\ h(n 1 ) ai h(n 2 ) a Nh(n - N) (8.5.6) 0 < n < M (8.5.5) (8.5.4)
T he set o f linear eq u ation s in (8.5.5) and (8.5.6) can b e used to solve for the filter param eters {a*} and {b*}. W e set h(n) = hd {n) for 0 < n < Af -I- N, and use the linear eq u ation s in (8.5.6) to solve for the filter param eters {a *}. T hen w e use values for the {a*} in (8.5.5) and solve for the param eters {>*}. Thus we obtain a perfect m atch b etw een h(n) and the desired resp onse hj ( n) for the first L values o f th e im pulse response. T his design tech n iqu e is u su ally called the Pade approxi mat i on procedure. T h e d egree to which this design tech n iqu e produces accep tab le filter designs d ep en d s in part on the num ber o f filter coefficients se lecte d . Since the design m eth od m atches /i</(n) o n ly up to the num ber o f filter param eters, the m ore com p lex the filter, the b etter th e approxim ation to h j ( n ) for 0 < n < M + N . H ow ever, this is also the m ajor lim itation with the P ade approxim ation m eth od , nam ely, the resulting filter m ust contain a large num ber o f p o les and zeros. For this reason, the P ade approxim ation m ethod has found lim ited use in filter designs for practical applications.
Example 8.5.1
Suppose that the desired unit sample response is
hj(n) = 2(i)"w(n)
Determine the parameters of the filter with system function
using the Pad6 approximation technique.
Solution In this simple example, H(z) can provide a perfect match to Hjiz), by
selecting bo = 2, f>i = 0, and we indeed obtain the same result. Let us apply the Pad6 approximation to see if
Sec. 8.5
703
With the substitution for hd(ri), we obtain a i = j . To solve for b^ and b\, we use the form (8 .5 . 5 ) with h(n) = hj(n). Thus
h<t(n) = ~ M - 1) + b08(n) + b xS(n - 1)
For n = 0 t h i s equation yields b0 = 2 . For n = 1 we obtain the result b\ = 0 . Thus H(z) HAz).
This exam ple illustrates that the P ade approxim ation results in a perfect m atch to Hd(z) w hen the desired system function is rational and w e have prior k n ow led ge o f the num ber o f p oles and zeros in the system . In general, how ever, this is n ot the case in practice, since hd (n) is d eterm in ed from som e desired fre quency resp onse specifications a>). In such a case th e P ade approxim ation m ay n ot result in a good filter design. T o illustrate a p oten tial p roblem and suggest a solu tion , let us consider the follow in g exam ples.
Example 8.5.2 A fourth-order Butterworth f i l t e r has the system function _ 4.8334 x 10-3( ; + l)4 A:} ~ (z2 - 1.3205- + 0.6326)(z2 - 1.0482z + 0.2959) The unit sample response corresponding to /^(z) i s illustrated i n Fig. 8 . 4 8 . Use the Pade approximation method to approximate / / r f ( ~ ) . Solution We observe that the desired f i l t e r has M = 4 zeros and j V = 4 poles. I ti s instructive to determine the coefficients i n the Pade approximation when the number of zeros and/or poles are not identical to the desired number of f i l t e r parameters.
704
Magnitude response
Chap. 8
4 Figure 8.49
Filter designs based on Pade approximation (Example 8.5.2).
In Fig. 8.49 we plot the frequency response of the f i l t e r obtained by the Padd approximation method. We have considered four cases: M = 3 ,N = 5 ;M = 3 ,N = 4 ; M= 4 ,N = 4 ;M= 4 ,N = 5 . W e observe that when M = 3 , the resulting frequency response i s a relatively poor approximation to the desired response. However, an increase in the number of poles from A ' = 4 to N = 5 appears to compensate in part for the lack of the one zero. When M i s increased from three to four, we obtain a perfect match with the desired Butterworth f i l t e r not only for N = 4 but for N = 5 , and, in f a c t , for larger values of N.
Example 8.5.3
A three-pole and three-zero type I I lowpass Chebyshev d i g it al f i l t e r has the system function " 0.3060(1 + z )(0.2652 - 0.09z' + 0.2652z~2) ( 1 - OJSSOz-^d - 1.1318z-> + 0.5387z-2)
I t s unit sample response i s illustrated in Fig. 8. 50 . Use the Padf approximation method to approximate Hj(z).
Solution By following the same procedure as in Example 8 . 5 . 2 , we determined the Pad6 approximation of H < /( z) based on the selection of M = 2, N = 3 ;M = 2 ,N = 4 ; M= 3 ,N = 3 ; M = 3, N = 4 . The frequency responses of the resulting designs are illustrated in Fig. 8.5 1.
Sec. 8.5
705
Figure fL50 Impulse response h j i n) of type II Chebyshev digital filter given in Example 8.5.3.
Magnitude
706
Chap. 8
As in Example 8 . 5 . 2 , we note that when we underestimate the number of zeros we obtain a relatively poor design, as evidenced by the two cases in which M = 2 . On the other hand, i fM = 3 , we obtain a perfect match for N = 3 and N = 4 .
T h ese tw o exam ples suggest that an effectiv e approach in using the Pade approxim ation is to try different values o f M and N until the frequency resp onses o f the resulting filters con verge to the d esired frequency resp onse w ithin som e sm all, acceptable approxim ation error. H ow ever, in practice, this approach appears to be cum bersom e.
8.5.2 Least-Squares Design Methods

A gain , let us assum e that hd {n) is specified for n > 0. W e begin with the sim ple case in w hich th e digital filter to b e d esign ed contains only p oles, that is, H(z) = (8.5.7)
N ow , con sid er the cascade con n ection o f the desired filter Hd (z) w ith the reciprocal, all-zero filter 1 / / / (z), as illustrated in Fig. 8.52. N ow suppose that the cascade configuration in Fig. 8.52 is excited by the unit sam ple seq u en ce S(n). Thus the input to the inverse system 1/ H { z ) is hd (n) and the output is v(n). Ideally, y d (n ) = S(n). T he actual output is (8.5.8) T h e condition that >v(0) = _y(0) 1 is satisfied by selectin g bo h d {0). For n > 0, y( n) represents the error b etw een th e desired output yd {n) 0 and the actual output. H en ce the param eters {a*} are selecte d to m inim ize the sum of M )
H d ( z)
yin)
/~ns
S(n)
H (z)
T
error
Minimize the sum of squared errors
Figure 8J?2
Least-squares inverse filter design m ethod.
Sec. 8.5
707
squares o f th e error sequence, = X > 2()

n=l
Z. n=l
;v hd ( n ) + Y akhdin - k) k=\ h2 A 0)
(8.5.9)
B y differen tiatin g with respect to the param eters {a*}, it is easily estab lish ed that w e ob tain the set o f linear equations o f the form
N
J 2 w hh( k, l ) = - r hh( l , 0 )
k-\
/ =
1 , 2, . . . , W
(8.5.10)
w h ere, by definition,
O O
rkh(k,l) = rt=l
0 0
= m * ) M + * - n = rhhik - l ) n=0 (8.5.11)
T h e solu tion o f (8.5.10) yields th e desired param eters for th e inverse system 1 / H ( z ) . Thus w e obtain the coefficients o f the all-p ole filter. In a practical design problem , the desired im pu lse resp onse hd (n) is specified for a finite set o f points, say 0 < n < L, w h ere L > > N. In such a case, the correlation sequ en ce rdd(k) can be com p uted from the finite seq u en ce hd (n) as
L \kl\
?hh (k I)
hd(n)hd(n + k - I)
0 < k - l < N
(8.5.12)
and these valu es can b e used to solve the set o f linear eq u ation s in (8.5.10). T he least-squares m eth od can also b e used in a p o le -z e r o approxim ation for Hd (z). If the filter H( z ) that approxim ates Hd (z) has b oth p oles and zeros, its resp onse to the unit im pulse <5(n) is
S M
^ ) = - J f l i A(fi t ) + ^ M ( - i t ) * -i *=o or, eq u ivalen tly, h(n) - Y a*h(n - V + b * **i F or n > M , (8.5.13) red u ces to
N
n > 0
(8.5.13)
0 < n < M
(8.5.14)
h(n ) = ^ a * h ( n k)
n> M
(8.5.15)
708
Chap. 8
Clearly, if H j i z ) is a p o le-zero filter, its resp onse to Bin) w ou ld satisfy the sam e eq u ation s (8.5.13) through (8.5.15). In gen eral, h ow ever, it d o es not. N everth eless, w e can use the desired response h j ( n) for n > M to construct an estim ate o f hd(n), according to (8.5.15). That is, (8.5.16)
hd(n) -
- *)
T hen w e can select the filter param eters {a*} to m inim ize the sum o f squared errors b etw een the desired response hd(n) and the estim ate hd (n) for n > M. Thus we have O O n=M+1 ih d^ ~ hd(n )]2
(8.5.17) The m inim ization o f \, with respect to the p o le param eters {a*}, leads to the set o f linear equations Y a irhh(k,l) = ~ r hh( k , 0 ) k =
1 , 2, . . . , N
(8.5.18)
where rhh( k, l ) is n ow defined as r>, h( k, l )= Y hd (n - k) hd (n - I) n~M+1 (8.5.19)
Thus these linear eq u ation s yield the filter p aram eters {a*}. N o te that these eq u a tions reduce to the all-pole filter approxim ation w h en M is set to zero. T h e param eters {bk} that determ in e the zeros o f the filter can be obtained sim ply from (8.5.14), w here h(n) = hd (rt), by sub stitu tion o f th e values [ak] ob tained by solving (8.5.18). Thus bn = hd(n) + ^ akhd (n k)
0<n<M
(8.5.20)
T h erefore, the param eters {a*} that d eterm in e the p o les are obtained by the m ethod o f least squares w hile the p aram eters {bk} that d eterm in e the zeros are ob tain ed by the the P ade approxim ation m eth od . T he foregoin g approach for determ ining the p o les and zeros o f H ( z ) is so m etim es called P r o n y s m e t h o d . T h e least-squares m eth od provid es go o d estim ates for th e p o le param eters {ak}. H ow ever, P ron ys m eth od m ay n ot b e as effective in estim ating the pa ram eters {>*}, prim arily b ecau se the com p utation in (8.5.20) is not based on the least-squares m ethod.
Sec. 8.5
Design of Digital Fitters Based on Least-Squares Method

(n) all-pole filter
W ,( z )
709
all-zero filter
H 2(z)
M")
Figure 8^ 3
Least-squares method for determining the poles and zeros of a filter.
A n alternative m eth od in which both sets o f param eters {fl*) and {*} are d e term ined by application o f the least-squares m eth od has b een prop osed by Shanks (1967). In Shanks m eth od , the param eters {a*) are com p uted on the basis o f the least-squares criterion, according to (8.5.18), as indicated ab ove. This yields the estim ates {a*}, which allow us to syn th esize the all-pole filter. Hi (z) = ------- rr--------(8-5.21)
i+
akz k
T h e resp onse o f this filter to the im pulse <5(n) is u(rc) = ^ fl; v(n k) + S(n) n > 0 (8.5.22)
If the seq u en ce (u(n)} is used to excite an all-zero filter with system function
H 2( z ) =
(8.5.23)
as illustrated in Fig. 8.53, its resp onse is hd(n) = ^ & * u ( n - k) N o w w e can define an error seq u en ce e ( n ) as e{n) = hd {n) - hd{n) = hd (n) - Y b*v (n ~ k) (8.5.25) (8.5.24)
and, con sequ en tly, the param eters {&*} can also be determ in ed by m eans o f the least-squares criterion, nam ely, from the m inim ization of -l 2 M ) - J ^ M ( ~k) (8.5.26)
Thus w e obtain a set o f linear eq u ation s for the param eters {bk}, in the form M Y bt rvv(k,l) = r * (0
/ = 0 ,1 ........ Af
(8.5.27)
710
w here, by definition, O O
Chap. 8
r vv(k, I) = Y v (n ~ k ^ n - *> r *=0 oo rhvik) = hd (n)v(n - Jt) n=0
(8.5.28)
(8.5.29)
Example 8.5.4 Approximate the fourth-order Butterworth f i l t e r given in Example 8.5.2 by means of an all-pole f i l t e r using the least-squares inverse design method. Solution From the desired impulse response hd(n), which i s illustrated in F ig. 8 . 4 8 , we computed the autocorrelation sequence rkh(kj) = rhh(k - / ) and solved to set of linear equations in (8.5.10) to obtain the f i l t e r coefficients. The results of t h i s computation are given in Table 8.14 for N = 3 ,4 ,5 , 10, and 15. In Table 8.15 we l i s t the poles of the f i l t e r designs for N = 3, 4, and 5 along with the actual poles of the fourth-order Butterworth f i l t e r . W e note that the poles obtained from the designs are fa r from the actual poles of the desired f i l t e r .
TABLE 8.14
{a k}
ESTIMATES OF FILTER COEFFICIENTS IN LEAST-SQUARES INVERSE FILTER DESIGN METHOD
i V= 02 = a-s = N = a{ = a2 = a3 = a4 = JV = f l != a2 = a2 = aj, = a5 =
0) =
3 0.254295E + 01 0.241800E + 01 0.853829E + 00 4 0.319047 + 01 0.425176 + 01 0.278234 + 01 0.758375E + 00 5 0.368733E + 01 0.607422 + 01 0.556726E + 01 -0.284813 + 01 0.654996E + 00
N = 10 a i= 5.008451 a2 - -12.660761 = 21.557365 04 = -27.804110 as = 28.683949 6 = -24.058558 a7 = 16.156847 f l gs s -8.247148 ag = 2.854789 f l i o -0.502956
N 15 = 2.993620 -1.143053 a2 = = -12.132861 39.663433 a4 ~ ai = -75.749001 109.247757 06 = a7 = -129.513794 a8 = 131.026794 a9 ~ -114.905266 87.449211 flio = an = -57.031906 30.915134 < 312 = a 13 = -13.124536 3.879295 014 = -0.597313 015 =
ai
Sec. 8.5
711
TABLE 8.15 ESTIMATES OF POLE

POSITIONS IN LEAST-SQUARES INVERSE FILTER DESIGN METHOD (EXAMPLE 8.5.4)
Number of poles N =3 N = 4 N = 5
Pole positions 0.9305 0.8062 yO.5172 0.8918 j'0.2601 0.7037 yO.6194 0.914 0.8321 y0.4307 0.5544 j'0.7134 0.6603 y0.4435 0.5241 yo.1457
N =4 Butterworth filter
The frequency responses of the f i l t e r designs are plotted in Fig. 8 .54. W e note that when N i ssmall, the approximation to the desired f i l t e ri spoor. As N i sincreased to N = 10 and N = 15, the approximation improves s i g n i f i c a l l y . However, even for N = 15, there are large ripples in the passband of the f i l t e r response. I ti s apparent that t h is method, which i sbased on an all-pole approximation, does not provide good approximations to f i l t e r s that contain zeros.
Example 8.5.5
Approximate the type I I Chebyshev lowpass f i l t e r given i n Example 8 .5.3 by means of the three least-squares methods described above.
Solution The results of the f i l t e r designs obtained by means of the least-squares inverse method, Prony s method and Shanks method, are illustrated i nF ig. 8.5 5. The f i l t e r parameters obtained from these design methods are l is te d in Table 8.1 6. The frequency response characteristics in Fig. 8.55 il l u s t r a t e that the leastsquares inverse (all-pole) design method yields poor designs when the f i l t e r contains zeros. On the other hand, both Prony s method and Shanksmethod yield very good designs when the number of poles and zeros equals or exceeds the number of poles and zeros i n the actual f i l t e r . Thus the inclusion of zeros in the approximation has a significant effect in the resulting f i l t e r design.
8.5.3 FIR Least-Squares Inverse (Wiener) Filters

In th e preced in g sectio n w e described the use o f the least-squares error criterion in the d esign o f p o le -z e r o filters. In this section w e u se a sim ilar approach to d eterm in e a least-squares F IR inverse filter to a d esired filter. T h e inverse to a linear tim e-invariant system w ith im pulse resp onse h(n) and system fun ction H(z) is defined as the system w h ose im pu lse resp onse hr(n) and
712
Chap. 8
Magnitude (dB )
(a)
Magnitude (dB)
Figure &54 method.
Magnitude responses for filter designs based on the least-squares inverse filter
system function H / ( z ), satisfy the resp ective equ ation s. h ( n ) * h , ( n ) = 5(o) H ( z ) H , (z) = 1 (8.5.30) (8.5.31)
In gen eral, H, { z) is IIR , unless H { z ) is an all-p ole system , in w hich case H, { z) is HR.
Magnitude (dB)
Magnitude (dB)
Magnitude (dB)
F ifu n &55 Filter designs based on least-squares methods (Example 8-5.5): (a) least-squares design; (b) Pronys method; (c) Shanks m ethod.
714
Chap. 8
TABLE 8.16 POLE-ZERO LOCATIONS FOR FILTER DESIGNS IN EXAMPLE 8.5.5

Chebyshev Filter: Zeros: -1 .0.1738311 ;0.9847755 Poles: 0.3880,0.5659 y'0.467394 Filter Order N = 3 Poles in Least-Squares Inverse
0.8522 0.6544 j0.6224 N = 4 0.7959 y0.3248 ____________________ ___________________0.4726 j'0.7142___________________ Filter Order N = 3 M = 2 N = 4 M = 2 N - 3 M = 3 N = 4 M = 3 Pronys Method Shanks M ethod ________________ 1__________________________________________________ Poles Zeros Poles Zeros 0.5332 0.6659 j 0.4322 0.7092 -0 .2 9 1 9 0.6793 y0.4863 0.3881 0.5659 J0.4671 -0 .00014 0.388 0.5661 y0.4672 0.5348 0.6646 j0.4306 0.7116 -0.2921 0.6783 _/0.4855 0.3881 0.5659 _/0.4671 -0.00014 0.388 0.566 ;0.4671
-0 .1497 j0.4925
-0 .2 4 3 7 y0.5918
-0 .1 9 8 2 ;0.37
0.306 y0.4482
-1
0.1736 /0.9847
-1
0.1738 y'0.9848
-1
0.1738 ;0.9848
-1
0.1738 y0.9848
In m any practical applications, it is d esirable to restrict th e inverse filter to b e F IR . O bviously, o n e sim ple m eth od is to truncate In so doing, w e incur a total squared approxim ation error equal to
CO
,=
Y * / ( ") fi Af+1
(8.5.32)
where Af + 1 is the length o f the truncated filter and E, rep resen ts the energy in the tail o f the im pulse resp onse /i/(n ). A ltern atively, w e can use the least-squares error criterion to op tim ize the Af + 1 coefficients o f the F IR filter. First, let d( n) d en o te the desi red out put sequence o f the F IR filter o f length Af -I-1 and let h(n) be th e input sequence. T hen, if y( n) is the output seq u en ce o f the filter, as illustrated in Fig. 8.56, the error seq u en ce b etw een the desired output and the actual ou tp u t is M e ( n ) = d( n) ^ * / i ( n k) k=0
(8.5.33)
where the {*} are the FIR filter coefficients.
Sec. 8.5

d(n) = S(n)
715
m -
FIR filter
v(")
(M ~ ~ r~
e(n)
Minimize the sum of squared errors
Figure 8.56
Least-squares FIR inverse filter.
T h e sum o f squares o f the error seq u en ce is

OC
2 d{n) - ^ b k h i r i - k)
(8.5.34)
n=0 l
i -0
W hen is m inim ized with respect to the filter coefficients, w e obtain the set o f linear eq u ation s M Y , b k r hh{ k - l ) = rdh{l) *=o I = 0, 1.........M (8.5.35)
w here rhh(l) is the au tocorrelation o f h(n), defined as
0 0
rhhd) = Y ^ h { n ) h { n - I) (8.5.36)
=0
and rdh{n) is the crosscorrelation b etw een the desired output d(n) and the input seq u en ce h(rt), defined as O O rdh{l) = Y , d ( n ) h { n - l ) (8.5.37) =o T h e optim um , in the least-squares sen se, F IR filter that satisfies the linear eq u ation s in (8.5.35) is called the Wiener filter, after th e fam ou s m athem atician N orbert W ien er, w h o introduced optim um least-squares filtering m eth od s in en g i n eerin g [see b o o k by W iener (1949)]. If the optim um least-squares F IR filter is to b e an approxim ate inverse filter, the desired resp onse is d( n) = S(n) T h e crosscorrelation b etw een d(n) and h(n) red u ces to ^ ( / ) = { J (0) (8.5.38)
10
/ = . otherw ise
(8.5.39)
716
Chap. 8
T h erefore, the coefficients o f the least-squares F IR filter are ob tain ed from the solution o f the linear eq u ation s in (8.5.35), w hich can b e exp ressed in m atrix form as
~ r hh (0 )
^ a(I)
. r hh{ M )
rhhi 1 ) rhh(0)
rhh (2 ) rhh(l)
rhh(M)
r hfl( M - l )
" bo ~ b\
'H O )!
(8.5.40)
r hh{ M ~ l )
rhhi 0 )
-bu -
W e observe that the matrix is n ot only sym m etric but it also has the special p roperty that all the elem en ts along any d iagon al are equal. Such a matrix is called a T oeplitz matrix and lends itself to efficient inversion by m eans o f an algorithm due to L evin son (1947) and D urbin (1959), which requires a num ber o f com putations p roportional to M 2 instead o f the usual M 3. T h e L evin son -D u rb in algorithm is described in Chapter 11. T h e m inim um value o f the least-squares error ob tain ed w ith th e optim um F IR filter is d ( n ) - ^ b k h i n - k) o
d{n)
= L
M ^ ^ ( n ) - ^ 2 b krdh(k)
(8.5.41)
In the case w here the F IR filter is the least-squares inverse filter, d ( n ) = S(n) and rdh(n) h(0)8(n). T herefore,
Erin = 1 - h( 0)bo
(8.5.42)
Example 8.5.6 Determine the least-squares FIR inverse filter of length 2 to the system with impulse response h(n) 1, a, . 0,
n=0 n = 1 otherwise
where |a| < 1. Compare the least-squares solution with the approximate inverse obtained by truncating hj(n). Solution Since the system has a system function H(z) = 1 a z ~ \ the exact inverse is IIR and is given by H,{z) = or, equivalently,
hi(n) = a"u(n)
1 - az"
Sec. 8.5
717
If this is truncated after n terms, the residual energy in the tail is
, =
a 2^ ! + a 2 + a 4 H------- )
From (8.5.40) the least-squares FIR filter of length 2 satisfies the equations
which have the solution 1 + a 2 4- a 4
For purposes of comparison, the truncated inverse filter of length 2 has the coefficients bo = 1 , b\ = a. The least-squares error is ^ which compares with ~ 1 +
for the truncated approximate inverse. Clearly, , > , so that the least-squares FIR inverse filter is superior. In this exam p le, th e im pulse resp onse h( n ) o f the system w as m inim um phase. In such a case w e selected the d esired resp onse to b e J (0) = 1 and d (n) = 0, n > 1. O n the other hand, if the system is nonm inim um phase, a d elay should b e inserted in th e desired resp onse in order to obtain a g o o d filter design. T he value o f the ap propriate d elay d ep en d s on the characteristics o f h(n). In an y case w e can com pute the least-squares error filter for d ifferen t d elays and se lect the filter that produces the sm allest error. T h e follow in g exam p le illustrates the effect o f the delay.
Example 8-5.7
Determine the least-squares FIR inverse of length 2 to the system with impulse response a, h(n) =
1, 0,
n= 0 n= 1 otherwise
where |a| < 1.
718 Solution
Chap. 8
T h is is a m axim um -phase system . If we select d(n) = [ 1 0 ] w e obtain the sam e solu tion as in E x am p le 8.5.6, with a m inim um least-squ ares erro r
E m = 1 - h( 0 )bo
= 1 + a
l+ a 2
1 + a2 + a4
If 0 < a < 1 , then > 1, which rep resen ts a p o o r inverse filter. If 1 < a < 0, then mjn < 1. In p articu lar, fo r a we obtain m in = 1.57. F o r a = - 5 . m = 0 .8 1 , which is still a very large value fo r th e squared erro r. N ow suppose that the desired response is specified as d(n) = & ( n 1). T h e n th e set o f eq u atio ns fo r the filter coefficien ts, o btain ed from (8 .5 .3 5 ), are th e solution to the eq uations fl + a2 L a
-a
-I p o ] _ [ M D ] _ r 1 1 L / i(0)J L -o fJ
l + a 2 J Lf c 1J
T h e solu tion o f th ese eq uations is
a 2 +
O '3
1 + a 2 + or4 T h e least-sq u ares erro r, given by (8.5.41), is
mm = 1 - b(,rdh(0) - birdh(l) = 1 - M ( 1 ) - M ( 0) ^m in 1
1 ^ . " t" 1 + a2 + a4 I + a2 + a4 1 - a4 1 + a2 + a4 In p articu lar, suppose th at a = j . T h en m in = 0.29. C o n sequ en tly, the desired response d(n) = S(n - 1) results in a significantly b e tte r in v erse filter. F u rth e r im provem en t is possib le by increasing th e length o f th e inverse filter.
In gen eral, w hen the desired resp onse is specified to con tain a delay D , then the crosscorrelation rd/,(l), defined in (8.5.37), b ecom es
r dh(l) = h ( D - I) l =
T h e set o f linear eq u ation s for the coefficients o f the least-squares F IR inverse filter given by (8.5.35) reduce to M ] T V a / . ( * - 0 = h ( D - 1) *=o 1 = 0 , 1 .........M (8.5.43)
T h en the exp ression for the corresponding least-squares error, given in general by
Sec. 8.5
719
(8.5.41), b eco m es u ftr in s l - J ^ b M D - k ) *=o (8.5.44)
L east-squares F IR in verse filters are often used in m any practical applications for d eco n v o lu tio n , including com m unications and seism ic signal p rocessing.
8.5.4 Design of HR Filters in the Frequency Domain

T h e IIR filter design m eth od s described in S ection s 8.5.1 through 8.5.3 are carried out in the tim e dom ain. T h ere are also direct design tech n iqu es for IIR filters that can b e p erform ed in the frequency dom ain. In this section w e d escribe a filter param eter optim ization tech n iqu e carried ou t in the frequency d om ain that is rep resen tative o f frequency-dom ain design m ethods. T h e design is m ost easily carried out with the system function for the IIR filter exp ressed in the cascade form as (8.5.45)
1+
+ak 2Z~2
w here the filter gain G and the filter coefficients {a*i], {**2 ], {#ti}, {Pki} are to b e determ in ed . T h e frequency response o f the filter can be exp ressed as H{to) = GA(w)e>B(a,) w here A((D) - ]""[ (8.5.46)
1 + Pk\Z~l + f$k2Z~2
1 + a * U - 1 + a * 2 Z 2
(8.5.47)
and (a>) is the phase response. In stead o f d ealin g with the phase o f the filter, it is m ore co n ven ien t to deal w ith the en v elo p e delay as a function o f frequency, w hich is </(*) or, eq u ivalen tly,
Zg (u>) = Tg ( z ) I z > -
(8.5.48)
r d Q (z )~ [ L dz
dz
J WJ db)
(8.5.49)
It can b e show n that r?(z) can b e exp ressed as zg (z) = R e
K r Atiz + 2/5*2
a*iz + 2a*2
(8.5.50)
i l[ L
w here R e ( u ) d en o tes the real part o f the com p lex-valu ed quantity u. N o w su p p ose that the d esired m agnitude and d elay characteristics A(a>) and Tg(co) are specified at arbitrarily ch osen discrete freq u en cies coi, in
720
Chap. 8
the range 0 < M < n . T hen the error in m agnitude at the frequency cok is GA(cok) A d (cok) w here A d (cok) is the desired m agnitude resp onse at cok. Sim i larly, the error in delay at a> k can b e defined as zg(a>k) zd (a>k), w here zd (cok) is the desired d elay response. H ow ever, the ch oice o f zd (a> k) is com plicated by the difficulty in assigning a nom inal delay to th e filter. H en ce, w e are led to define the error in d elay as zg(cok) rg(wo) zd (a> k), w h ere rg(wo) is the filter d elay at som e nom inal center frequency in the passband o f the filter and zd (wk) is the desired delay response o f the filter relative to r?(o^). B y defining th e error in delay in this m anner, w e are w illing to accept a filter having w h atever nom inal delay zg (a>o) results from the op tim ization procedure. A s a perform ance index for determ ining th e filter param eters, o n e can ch oose any arbitrary function o f the errors in m agnitude and d elay. T o be specific, let us select the total w eigh ted least-squares error o ver all freq u en cies &i, a> 2, . . . , < oL, that is,
L
( p ,
G)
( 1
A )
wn[GA(co) - A d (con )
] 2
n 1
v[zg(a)) 7 1 =1
(o>o) - zd (con)]2
(8.5.51)
w here p d en o tes the 4 K -dim ensional vector o f filter coefficients {a*i}, {<**2 }, (An), and {/ta h and A , {u>}, and {t>} are w eigh ting factors se lecte d by the designer. Thus the em p hasis on the errors affecting the design m ay b e placed entirely on the m agnitude ( A = 0), or on the delay ( A = 1) or, perhaps, eq u ally w eigh ted betw een m agnitude and d elay ( A = 1/2). Similarly, th e w eigh ting factors in frequency {u;} and {u} determ in e the relative em phasis on the errors as a fun ction o f frequency. T h e squared-error function E (p , G ) is a non linear fun ction o f (AK + 1) pa ram eters. T h e gain G that m inim izes is easily d eterm in ed and given by the relation
L
Y ^ u M v n) A d (con) G = ^ L--------------------^ i o nA 2 (a))

n=l
(8.5.52)
T h e optim um gain C can b e substituted in (8.5.51) to yield L (p , G) = (1 *) v)n[GA(a>) - A d (co)]2 n* 1

L
+ k
n 1
v K
> ~ zs ( t o n ) ] 2
(8.5.53)
D u e to th e nonlinear nature o f (p, G) , its m inim ization o v er the rem aining 4 K param eters is perform ed by an iterative n um erical op tim ization m eth od such
Sec. 8.5
721
as the F letch er and P o w ell m ethod (1963). O n e begins the iterative process by assum ing an initial set o f param eter values, say p (0). W ith the initial values su b sti tuted in (8.5.51), we obtain the least-squares error ( p (0), G). If w e also evaluate the partial derivatives d / d a k\, d / d a k2, 3/3/J*i, and b / b 0 k2 at the initial value p (0), w e can u se this first derivative inform ation to change the initial values o f the p aram eters in a direction that lead s toward the m inim um o f the function (p . G) and thus to a new set o f param eters p (1). R ep etitio n o f the ab ove step s results in an iterative algorithm which is d e scribed m athem atically by the recursive eq u ation p (m+l) _ p (m) A (m)Q()g (m) m _ 0, 1, 2, . . .
w here A (m ) is a scalar representing the step size o f the iteration, Q (m) is a (4K x 4 K ) m atrix, w hich is an estim ate o f the H essian , and g (m) is a (4 K x 1) vector consisting o f the four AT-dimensional vectors o f gradient com p on en ts o f (i.e., d / d a \ n , d / d a k l, d / d p kU d / d p k2), evaluated at a kl = or" \ a k2 = a ^ \ /3k] = fik2 = and Th*s iterative p rocess is term inated w hen the gradient com p on en ts
are nearly zero and the value o f the function (p , G) d o es not change appreciably from o n e iteration to another. T he stability constraint is easily incorporated into the com puter program through the param eter vector p. W hen la ^ l > 1 for any k = 1 , . . . , K , the param eter a k2 is forced back inside the unit circle and the iterative p rocess continued. A sim ilar p rocess can be used to force zeros in sid e the unit circle if a m inim um -phase filter is desired. T h e m ajor difficulty with any iterative p roced u re that searches for the param eter valu es that m inim ize a nonlinear function is that the process m ay con verge to a local m inim um instead o f a global m inim um . Our on ly recourse around this p rob lem is to start th e iterative process w ith d ifferent values for the param eters and o b serve the end result. Example 8.5.8 Let us design a lowpass filter using the Fletcher-Powell optimization procedure just described. The filter is to have a bandwidth of 0.3 n and a rejection band commencing at 0.45jr. The delay distortion can be ignored by selecting the weighting factor X = 0. Solution We have selected a two-stage (K = 2) or four-pole and four-zero filter which we believe is adequate to meet the transition band and rejection requirements. The magnitude response is specified at 19 equally spaced frequencies, which is con sidered a sufficiently dense set of points to realize a good design. Finally, a set of uniform weights is selected. This filter has the response shown in Fig. 8.57. It has a remarkable resemblance to the response of the elliptic lowpass filter shown in Fig. 8.58, which was designed to have the same passband ripple and transition region as the computer-generated filter. A small but noticeable difference between the elliptic filter and the computer generated filter is the somewhat flatter delay response of the latter relative to the former.
722
Chap. 8
Frequency (a)
(b) Figure (L57 Filter designed by Fletcher-Powell optimization method (Exam ple 8.5.8). Example 8.5.9 Design an IIR filter with magnitude characteristics smm, Arf(a>) =
0,
<\(0 \ < 7 t
and a constant envelope delay in the passband
Sec. 8.5
723
Frequency Figure &58 Amplitude and delay response for elliptic filter. Solution The desired filter is called a modified duobinary filter and finds application in high-speed digital communications modems. The frequency response was specified at the frequencies illustrated in Fig. 8.59. The envelope delay was left unspecified in the stopband and selected to be flat in the passband. Equal weighting coefficients {ujn} and {t>} were selected. A weighting factor of A = 1/2 was selected. A two-stage (four-pole, four-zero) filter is designed to meet the foregoing spec ifications. The result of the design is illustrated in Fig. 8.60. We note that the magnitude characteristic is reasonably well matched to sin a> in the passband, but the stopband attenuation peaks at about - 2 5 dB, which is rather large. The envelope delay characteristic is relatively flat in the passband.
724
Chap. 8
n 2
Figure 8^9 Frequency response of an ideal modified duobinary filter. A four-stage (eight-pole, eight-zero) filter having the same frequency response specifications was also designed. This design produced better results, especially in the stopband where the attenuation peaked at 36 dB. The envelope delay was also considerably flatter.

W e have described in so m e detail the m ost im portant tech n iq u es for designing FIR and IIR digital filters based on eith er frequency-dom ain specifications expressed in term s o f a d esired frequency resp onse Hd (to), or in term s o f the desired im pulse response h d (n). A s a general rule, F IR filters are used in applications w h ere there is a need for a linear-phase filter. T his requirem ent occurs in m any applications, especially in telecom m u n ication s, w h ere there is a requirem ent to separate (dem u ltiplex) signals such as data that have b een frequency-division m ultip lexed , w ithou t distorting th ese signals in the process of dem ultiplexing. O f the several m eth od s described for designing F IR filters, the frequency sam pling design m eth od and the optim um C hebyshev approxim ation m ethod yield the b est designs. IIR filters are generally used in applications w here so m e phase distortion is tolerab le. O f the class o f IIR filters, ellip tic filters are th e m ost efficient to im plem en t in th e sen se that for a given set o f specifications, an elliptic filter has a low er order or few er coefficients than any other IIR filter type. W hen com pared w ith F IR filters, elliptic filters are also considerably m ore efficient. In view o f this, on e m ight con sid er the use o f an elliptic filter to ob tain th e desired frequency selectivity, fo llo w ed then by an all-pass phase equ alizer th a t com p en sates for the p h ase distortion in th e ellip tic filter. H ow ever, attem pts to accom plish this have resulted in filters w ith a num ber o f coefficients in the cascad e com bination that
Sec. 8.6
725
Frequency (kHz)
4.000 3.500 3.000 | 2.500 'I S' 2 000 2 1.500 u
tt
1.000
0.500
0.000
0.00
0.48
0.96 . 1.44 Frequency (kHz)
1.92
2.40
Figare &60 Frequency response of filter in Example 8.5.9. Designed by the Fletcher-Powell optimization method.
eq u aled or ex ceed ed th e num ber o f coefficients in an eq u ivalen t linear-phase F IR filter. C on sequ en tly, no reduction in com p lexity is ach ievab le in using phaseeq u alized ellip tic filters. In addition to th e filter d esign m eth od s based on th e transform ation o f an alog filters in to th e digital dom ain, w e also p resen ted several m eth od s in w hich th e d esign is d o n e directly in the discrete-tim e dom ain. T h e least-squares m ethod is particularly appropriate for designing IIR filters. T h e least-squares m ethod is also u sed for th e design o f F IR W iener filters.
726
Chap. 8
Such a rich literature n ow exists on the design o f digital filters that it is not p ossible to cite all the im portant referen ces. W e shall cite on iy a few . S om e of the early w ork on digital filter design w as d on e by K aiser (1963, 1966), Steiglitz (1965), G old en and K aiser (1964), R ad er and G old (1967a), Shanks (1967), H elm s (1968), G ibbs (1969, 1970), and G old and R ader (1969). T h e design o f analog filters is treated in the classic b ook s by Storer (1957), G u illem in (1957), W einberg (1962), and D an iels (1974). T h e frequency sam pling m ethod for filter design was first p rop osed by G old and Jordan (1968, 1969), and op tim ized by R abiner et al. (1970). A d d ition al results w ere published by H errm ann (1970), H errm ann and S chuessler (1970a), and H ofstetter et al. (1971). T h e C h eb ysh ev (m inim ax) approxim ation m eth od for d esigning linear-phase F IR filters w as p rop osed by Parks and M cC lellan (1972a,b) and discussed further by R ab in er et al. (1975). T h e design o f ellip tic digital filters is treated in the b ook by G old and R ad er (1969) and in the paper b y Gray and M arkel (1976). T he latter includes a com puter program for designing digital ellip tic filters. T h e use o f frequency transform ations in the digital d om ain w as prop osed by C onstantinides (1 9 6 7 ,1 9 6 8 , 1970). T h ese transform ations are appropriate on ly for IIR filters. T he reader should n ote that w hen these transform ations are applied to a low pass F IR filter, the resulting filter is IIR. D irect design techniques for digital filters have b een con sid ered in a num ber o f papers, including Shanks (1967), Burras and Parks (1970), S teiglitz (1970), D eczk y (1972), B rophy and Salazar (1973), and B andler and Bardakjian (1973).
PROBLEMS
8.1 Design an FIR linear phase, digital filter approximating the ideal frequency response
1,
for \a)\ <
HAco) =
0,
for < \w\ < n
(a) Determine the coefficients of a 25-tap filter based on the window method with a rectangular window. (b) Determine and plot the magnitude and phase response of the filter. (c) Repeat parts (a) and (b) using the Hamming window. (d) Repeat parts (a) and (b) using a Bartlett window. 8.2 Repeat Problem 8.1 for a bandstop filter having the ideal response for M < 6 t n i . n for | < M < j for < \w\ < jt 8 J Redesign the filter of Problem 8.1 using the Hanning and Blackman windows. 8.4 Redesign the filter of Problem 8.2 using the Hanning and Blackman windows.
HAo>) =
Chap. 8
Problems
727
8 ^ D eterm ine the unit sample response {h(n)} of a linear-phase F IR filter of length M = 4 for which the frequency response at u> = 0 and w = n/ 2 is specified as HA 0) = 1 H,
--I
8.6 D eterm ine the coefficients {A(n)} of a linear-phase FIR filter of length M = 15 which has a symmetric unit sample response and a frequency response that satisfies the condition /2 7 T * \ fl, * = 0 ,1 .2 ,3 r V 15 j ~ I 0, * = 4, 5, 6, 7 8.7 R epeat the filter design problem in Problem 8.6 with the frequency response specifi cations 1 * = 0 .1 ,2 ,3 0.4 *= 4 , i f 0 * = 5 .6 ,7 8.8 The ideal analog differentiator is described by dxa(r) y. ( 0 = where *(/) is the input and ya(t) the output signal. (a) D eterm ine its frequency response by exciting the system with the input jc (/) =
e jlxFt
(b ) Sketch the magnitude and phase response of an ideal analog differentiator band-
limited to B hertz. (c) The ideal digital differentiator is defined as H(w) sc ja> |w| < x
Justify this definition by comparing the frequency response \H(w )|, 4 . H(<d) with that in part (b). (d) By computing the frequency response H(o>), show that the discrete-time system y(n) = x ( n ) - x ( n - 1) is a good approximation of a differentiator at low frequencies. (e) Com pute the response of the system to the input x(n) = A cos(a>on + 6) 8.9 Use the window m ethod with a Hamming window to design a 21-tap differentiator as shown in Fig. P8.9. Com pute and plot the magnitude and phase response of the resulting filter.
W,f()l
Figure P8.9
728
Chap. 8
8.10 Use the matched-z transform ation to convert the analog filter with system function 5 + 0 .1 H ( s ) ~ (s + 0.1 )2 + 9 into a digital IIR filter. Select T = 0.1 and com pare the location of the zeros in H (z) with the locations of the zeros obtained by applying the impulse invariance m ethod in the conversion of H(s). 8.11 Convert the analog bandpass filter designed in Example 8.4.1 into a digital filter by means of the bilinear transformation. Thereby derive the digital filter characteristic obtained in Example 8.4.2 by the alternative approach and verify that the bilinear transform ation applied to the analog filter results in the same digital bandpass fil ter. 8.12 A n ideal analog integrator is described by the system function Ha(s) = 1fs. A digital integrator with system function //(z) can obtained by use of the bilinear transform a tion. T hat is, 7 1 + z -1 W(z) 2 1 - - i = (a) W rite the difference equation for the digital integrator relating the input x ( n ) to the output v(n). (b) Roughly sketch the magnitude and phase (ft) of the analog integra tor. (c) It is easily verified that the frequency response of the digital integrator is
H
(a>) =
- J-
. T cos(<u/2) ; = 2 sin(a>/2 )
- J
.T u) COt 2 2
Roughly sketch |W(w)| and 6(w). (d) Com pare the magnitude and phase characteristics obtained in parts (b) and (c). How well does the digital integrator m atch the m agnitude and phase character istics of the analog integrator? (e) The digital integrator has a pole at z = 1. If you im plem ent this filter on a digital com puter, what restrictions might you place on the input signal sequence x ( n ) to avoid com putational difficulties? 8.13 A z-plane pole-zero plot for a certain digital filter is shown in Fig. P8.13. The filter has unity gain at dc. (a) D eterm ine the system function in the form H( z) (1 + aiz~1)(l + b\Z 1 + biz (1 +CiZ_)(l +<fiz - 1 + d2z 3
giving numerical values for the parameters A, ai, b\, bj, c\, d\, and d2. (b) Draw block diagrams showing numerical values for path gains in the following forms: (1) Direct form II (canonic form) (2) Cascade form (make each section canonic, with real coefficients) 8.14 Consider the pole-zero plot shown in Fig. P8.14. (a) D oes it represent an FIR filter? (b) Is it a linear-phase system?
Chap. 8
Problems
729
Figure P8.13
Figure P8.14 (c) Give a direct form realization that exploits all symmetries to minimize the number of multiplications. Show all path gains. 8.15* A digital low-pass filter is required to meet the following specifications: Passband ripple: < 1 dB Passband edge: 4 kHz Stopband attenuation: > 40 dB Stopband edge: 6 kHz Sample rate: 24 kHz The filter is to be designed by performing a bilinear transformation on an analog system function. Determine what order Butterworth, Chebyshev, and elliptic analog designs must be used to meet the specifications in the digital implementation.
730
Chap. 8
8.16* An IIR digital low-pass filter is required to meet the following specifications: Passband ripple (or peak-to-peak ripple): < 0.5 dB Passband edge: 1.2 kHz Stopband attenuation: > 40 dB Stopband edge: 2.0 kHz Sample rate: 8.0 kHz Use the design formulas in the book to determine the required filter order for (a) A digital Butterworth filter (b) A digital Chebyshev filter (c) A digital elliptic filter 8.17* Determine the system function H(z) of the lowest-order Chebyshev digital filter that meets the following specifications: (a) 1-dB ripple in the passband 0 < \co\ < 0 .3 jt. (b) At least 60 dB attentuation in the stopband 0.35jt < |w] < jr. Use the bilinear transformation. 8.18* Determine the system function H(z) of the lowest-order Chebyshev digital filter that meets the following specifications: (a) |-d B ripple in the passband 0 < |a>) < 0.24 jt. (b) At least 50-dB attenuation in the stopband 0 .3 5 jt < |u| < jr. Use the bilinear transformation. 8.19* An analog signal x (r) consists of the sum of two components jri(r) and jr2 (r). The spectral characteristics of x(t) are shown in the sketch in Fig. P8.19. The signal x{t) is bandlimited to 40 kHz and it is sampled at a rate of 100 kHz to yield the sequence *(n). It is desired to suppress the signal x 2(f) by passing the sequence *(n) through a digital lowpass filter. The allowable amplitude distortion on |X i(/) | is 2% (Si = 0.02) over the range 0 < |F | < 15 kHz. Above 20 kHz, the filter must have an attenuation of at least 40 dB (52 = 0.01). (a) Use the Remez exchange algorithm to design the minim um -order linear-phase FIR filter that meets the specifications above. From the plot of the magni tude characteristic of the filter frequency response, give the actual specifications achieved by the filter. (b) Compare the order M obtained in part (a) with the approximate formulas given in equations (8.2.94) and (8.2.95). (c) For the order M obtained in part (a), design an FIR digital lowpass filter using the window technique and the Hamming window. Compare the frequency response characteristics of this design with those obtained in part (a).
\X(F)\
Frequency in kilohertz
Figure P8.14
Chap. 8
Problems
731
(d) Design the minimum -order elliptic filter that meets the given amplitude specifica tions. Compare the frequency response of the elliptic filter with that of the FIR filter in part (a). (e) Compare the complexity of implementing the FIR filter in part (a) versus the elliptic filter obtained in part (d). Assume that the FIR filter is implemented in the direct form and the elliptic filter is implemented as a cascade of two-pole filters. Use storage requirements and the number of multiplications per output point in the comparison of complexity.
8.20 The impulse response of an analog filter is shown in Fig. P8.20.
Figure P&20 (a) Let h(n) = h(nT), where T = 1, be the impulse response of a discrete-time filter. Determine the system function H(z) and the frequency response H{w) for this FIR filter. (b) Sketch (roughly) \H{w) and compare this frequency response characteristic with (c) The FIR filter with unit sample response h{n) given above is to be approximated by a second-order IIR filter of the form a2z ' Use the least-squares inverse design procedure to determine the values of the coefficients &o, ai, and a2.
1 - ayz~
G(z) =
bgz
8.21 In this problem you will be comparing some of the characteristics of analog and digital implementations of the single-pole low-pass analog system HAs) = J < M O = e~t" s -(- Of. (a) What is the gain at dc? At what radian frequency is the analog frequency re sponse 3 dB down from its dc value? At what frequency is the analog frequency reponse zero? At what time has the analog impulse response decayed to 1/e of its initial value? (b) Give the digital system function H(z) for the impulse-invariant design for this filter. What is the gain at dc? Give an expression for the 3-dB radian frequency. A t what (real-valued) frequency is the response zero? How many samples are there in the unit sample time-domain response before it has decayed to 1/e o f its initial value? (c) Prewarp the parameter a and perform the bilinear transformation to obtain the digital system function H(z) from the analog design. What is the gain at dc? At what (real-valued) frequency is the response zero? Give an expression for the 3-dB radian frequency. How many samples in the unit sample time-domain response before it has decayed to 1 /e of its initial value?
732
Chap. 8
JL22 We wish to design a FTR bandpass filter having a duration M = 201. Hj(co) represents the ideal characteristic of the noncausal bandpass filter as shown in Fig. P8.22.
------------ ------------------- --------------- -----------jt
O.Sjt
0.4 k
0.4 k
0.5 k
jt
Figure P&22 (a) Determine the unit sample (impulse) response hd(n) corresponding to Hj(a>),
(b) Explain how you would use the Hamming window

/ 2* \ u>(n) = 0.54 + 0.46 cos j ------ - ) \M -1 J M -l M -l ------ - < n < - 2 ~ ~ 2
to design a FIR bandpass filter having an impulse response h(n) for 0 < n < 200. (c) Suppose that you were to design the FIR filter with M = 201 by using the fre quency sampling technique in which the DFT coefficients H(k) are specified instead of h(n). Give the values of H(k) for 0 < k < 200 corresponding to Hd{eJU) and indicate how the frequency response of the actual filter will differ from the ideal. Would the actual filter represent a good design? Explain your answer. 6 2 3 We wish to design a digital bandpass filter from a second-order analog lowpass But terworth filter prototype using the bilinear transformation. The specifications on the digital filter are shown in Fig. P8.23(a). The cutoff frequencies (measured at the half power points) for the digital filter should lie at w = 5jt/12 and co ln/12. The analog protoype is given by H(s) ----------- l - -----s2 + v 2 j + 1 with the half-power point at 2 = 1 .
(a) Determine the system function for the digital bandpass filter. (b) Using the same specs on the digital filter as in part (a), determine which of the
analog bandpass prototype filters shown in Fig. P.8.23(b) could be transform ed directly using the bilinear transformation to give the proper digital filter. Only the plot of the magnitude squared of the frequency is given. 6 2 4 Figure P8.24 shows a digital filter designed using the frequency sampling method. (a) Sketch a z-plane pole-zero plot for this filter. (b) Is the filter lowpass, highpass, or bandpass? (c) Determine the magnitude response \H(a>) at the frequencies wk ~ n k / 6 for k = 0, 1, 2, 3, 4, 5, 6 .
Chap. 8
Problems
733
m a>)\2
TT 6
t\ J V
TT
7r
m m 2
Figure P&23
734
Chap. 8
(d) Use the results of part (c) to sketch the magnitude response for 0 < a> < n and confirm your answer to part (b). 8.25 An analog signal of the form xa(t) = a(t) cos 2000^/ is bandlimited to the range 900 < F < 1100 Hz. It is used as an input to the system shown in Fig. P8.25. (a) Determine and sketch the spectra for the signals x(n) and w(n). (b) Use a Hamming window of length M = 31 to design a lowpass linear phase FIR filter H(w) that passes (c(n)}. (c) Determine the sampling rate of the A /D converter that would allow us to elim inate the frequency conversion in Fig. P8.25. 8.26 System identification Consider an unknown LTI system and an FIR system model as shown in Fig. P8.26. Both systems are excited by the same input sequence (x(n)}. The problem is to determine the coefficients (A(n), 0 < n < M - 1} of the FIR model
Chap. 8
Problems
735
Rx
= ~ = 2500
cos (0.8
nn)
Figure P8.25
Figure P8.26 of the system to minimize the average squared error between the outputs of the two systems. (a) Use the least-squares criterion to determ ine the equations for the optimum F IR filter coefficients. (b) R epeat part (a) if the output of the unknown system is corrupted by an additive white noise {u>()} sequence with variance cr2. 8.27 D eterm ine the least-squares F IR inverse of length 3 to the system with impulse re sponse
2,
h(n) =
1. 0.
n= 0 n= 1 otherwise
Also, determ ine the minimum squared error min. 8.28 D eterm ine the least-squares FIR inverse filter of length 3 for the system with im pulse response h(n) given in Example 8.5.6, when a = | and the desired response is specified as d(n) = 8(n 2). Also compute the minimum least-squares error.
736
Chap. 8
8.29* A linear time-invariant system has an input sequence x(n) and an output sequence y(n). The user has access only to the system output y(n). In addition, the following information is available. (a ) The input signal is periodic with a given fundamental period N and has a flat spectral envelope, that is, N- 1 jt(n) = ^ 2 ct e Ji2xmk" all n k =0 where c * k 1 for all k. (b) The system H(z) is all-pole, that is, H(z) = ------- -------i+ X > -* *=4 but the order p and the coefficients (a*, 1 < k < p) are unknown. Is it possible to determine the order p and the numerical values of the coefficients {ak, 1 < k < p\ by taking measurements on the output y(n)7 If yes, explain how. Is this possible for every value of p i (c) Repeat Problem 8 . 3 1 for a system with system function M Y , b*: ~k
H (z) =
-----------
k=\ (d) FIR system modeling Consider an unknown FIR system with impulse response h ( n ) , 0 < n < 11, given by
h(0) h(l) h ( 2)
= = = /i( 1 1 ) /i(10) A (9) = = = = = = 0.309828 0.416901 x x 1 0 ~ 1 0 ~ x x l O -1 1 0 " 1
-0.577081 - 0 . 8 52502 0.147157 x 0.449188 x
h(3) = h ( 8 ) h( 4 ) = h( 7 ) h(5)
=
10 10
h(6)
A potential user has access to the input and output of the system but does not have any information about its impulse response other than that it is FIR. In an effort to determine the impulse response of the system, the user excites it with a zero mean, random sequence x(n) uniformly distributed in the range [ 0 . 5 , 0 . 5 ] , and records the signal x(n) and the corresponding output y(n) for
0 <
<
199.
(1) By using the available information that the unknown system is FIR, the user employs the method of least-squares to obtain an FIR model h(n), 0 < n < M 1. Set up the system of linear equations, specifying the param eters / i ( 0 ) , A ( l),. . . , h(M 1 ) . Specify formulas we should use to determine the necessary autocorrelation and crosscorrelation values.
Chap. 8
Problems
737
(2) Since the order of the system is unknown, the user decides to try models of different orders and check the corresponding total squared error. Clearly, this error will be zero (or very close to it if the order of the model be comes equal to the order of the system). Com pute the F IR models hM{n), 0 < n < M - 1 for M = 8, 9, 10, 11, 12, 13, 14 as well as the corre sponding total squared errors E M, M = 8, 9........14. What do you ob serve? (3) D eterm ine and plot the frequency response of the system and the models for M = 1 1 ,1 2 , 13. Comment on the results. (4) Suppose now that the output of the system is corrupted by additive noise, so instead of the signal v(n), 0 < n < 199, we have available the sig nal v(n) = y(n) + O.Oliy(n) where w(n) is a Gaussian random sequence with zero mean and variance ff2 = 1 . R epeat part (b) by using u(n) instead of _v(n) and comment on the results. The quality of the model can be also determined by the quantity O C
J2h 2 (n ) n=0
Sampling and Reconstruction of Signals
In C hapters 1 and 4 w e treated the sam pling o f con tin u ou s-tim e signals and d em on strated that if the signals are bandlim ited, it is p ossib le to reconstruct the original signal from the sam ples, provided that the sam pling rate is at least tw ice the highest frequency contained in th e signal. W e also briefly describ ed th e sub seq u en t o p er ations o f quantization and coding that are n ecessary to con vert an analog signal to a digital signal appropriate for digital processing. In this chapter w e con sid er tim e-d om ain sam pling, analog-to-digital (A /D ) con version (quantization and cod in g), and d igital-to-an alog ( D /A ) conversion (sig nal reconstruction) in greater depth. First, w e con sid er th e sam pling o f the sp e cial class o f signals that are characterized as bandpass signals. T h en w e treat an alog-to-digital converters and their characteristics. O f particular interest is the use o f oversam pling and sigm a-delta m odu lation in th e design o f high p re cision A /D converters. T h e final top ic o f th e chapter is d igital-to-an alog con ver sio n or, sim ply, the reconstruction o f th e con tinu ou s-tim e sign al from its sam pled values.
9.1 SAMPLING OF BANDPASS SIGNALS

O ur m ain focus in this section is th e sam pling o f bandpass signals. W e begin by describing th e tim e and frequency dom ain rep resen tation s o f bandpass signals.
9.1.1 Representation of Bandpass Signals

S u p pose that a real-valued signal x ( t ) has a freq u en cy co n ten t concentrated in a narrow band o f freq u en cies in the vicinity o f a frequency Fc, as show n in Fig. 9.1. O ur ob jective is to d ev elop a m athem atical rep resen tation o f such signals. First, w e construct a signal that contains o n ly the p ositive freq u en cies in x ( t ) . Such a signal can b e exp ressed as X+(F) = 2 V ( F ) X ( F ) 738 (9.1.1)
Sec. 9.1
Sampling of Bandpass Signals

|X(F)|
739
-Fr
Figure 9.1 Spectrum of a bandpass signal.
w here X ( F ) is the Fourier transform o f x ( t ) and V ( F ) is the unit step function. T he equivalent tim e-dom ain expression for (9.1.1) is x +(t)
OC
X + ( F ) e i2* f ' d F -O C
(9.1.2) = f - 1 [2 V ( F ) ] * F , [ ^ ( ^ ) ] T h e signal A+(r) is called the analytic signal or the pre -en v elo p e o f jr(r). W e n ote that F ~ l [X ( F ) ] = x ( t ) and / r [2 V < /)] = <$(r) + ^ 7 T/ H en ce, -M O = SU) + *x(t)
TT t
(9.1.3)
x(t) + j * x ( t)
TT 1
(9.1.4)
W e define x { t ) as x ( t ) = x ( t ) nt =
i r oc ^ H J t
(9.1.5) ^
T h e signal x ( t ) can be view ed as the output o f the filter with im pulse response hit) = , nt
00 < t < oo
(9.1.6)
w hen excited by the input signal * ( /). Such a filter is called a Hilber t trans form er . T he frequency resp onse of this filter is sim ply O / O h ( t ) e ~ j2,lF,d t
*00
-n I Jf i.e .oc t
dt
( F > 0)
(F = 0)
(F < 0)
(9.1.7)
We observe that |tf(F )| = 1 and that the phase response (F) = - I n for F > 0
740
Chap. 9
and @ (F) = jjr for F < 0. T h erefore, this filter is basically a 90 p h ase shifter for all frequencies in the input signal, and it is akin to the discrete-tim e H ilbert transform filter described in Section 8.2.6. T he analytic signal * + (/) is a bandpass signal. W e can obtain an equivalent low pass representation by perform ing a frequency translation o f X + ( F ) . Thus, w e define X ,(F ) as X , ( F ) = X + ( F + Fc) T h e equivalent tim e-dom ain relation is * /(/) = x + ( l )e ~ ]2 Fr' = [*(f) + ; x ( r ) K j2* fr' or, equivalently, * (/) + j x ( t ) = x , ( t ) e j2nF'' (9.1.10) (9.1.9) (9.1.8)
In general, the signal x/(t) is com p lex-valu ed (see P roblem 9.3), and can be expressed as
jc ,(0
= M / ) + jf 'M 0
(9.1.11)
If w e substitute for x/(t) in (9.1.10) and eq u ate real and im aginary parts on each side, we obtain the relations x(r) = u< (t) c o s 2 n F ct - u s (t) s i n 2 j t F ct x ( t ) = ,.(/) sin 2?r Fct + u s {t) cos 27rF,t (9.1.12) (9.1.13)
T he expression (9.1.12) is the desired form for the representation o f a band pass signal. T h e low -frequency signal com p on en ts uc (t) and u s (t) can be view ed as am plitude m odulations im pressed on the carrier com p on en ts cos 2 n F ct and s i n 2 n F ct, respectively. Since th ese carrier com p onents are in phase quadrature, u c(t) and u s (t) are called the q ua dra ture c o m p o n e n t s o f the bandpass signal x ( t ) . A n o th er representation o f the signal in (9.1.12) is x ( t ) = R e{[Wf(r) + y.t ( r ) y 2jr^ } = R e M f W 2^ ' ] (9.1.14)
w here R e d en o tes the real part o f the com p lex-valu ed quantity in the brackets follow ing. T h e low pass signal x/(r) is usually called the c o m p l e x envelo pe o f the real signal j(? ), and is basically the eq uivalent l ow p a ss signal. Finally, a third p ossib le representation o f a bandpass signal is ob tain ed by expressing * ,(/) as x,U ) = a ( t ) e Jfnn where a(t ) = yj u l ( t ) + u](t ) (9.1.16) (9.1.15)
Sec. 9.1
741
Then .r(r) = R e[x,(r)e-/2irF''] = R e [ a ( f ) ^ 23rF,+^>J] = a (r )c o s[ 2 jr /v / + 0 ( 0 ] (9.1.18)
T h e signal a ( t ) is called the en velop e o f jc(r), and 6 ( t) is called the p h a s e o f jc(r). T h erefore, (9.1.12), (9.1.14), and (9.1.18) are eq u ivalen t rep resen tation s o f b and pass signals. T h e Fourier transform o f x ( t ) is
f
U se o f the identity
[R e[ xi(t)ej27fFr' ] } e - i2,tF,d t
(9.1.19)
R e ( f) = i ( * + n in (9.1.19) yield s the result X (F ) = l

J
(9.1.20)
f X [x,(t)eJ2 *F ' ' + x ; ) e - J2* F' ' ] e - J2* F'd t

-oc (9.1.21)
= \ [ X t ( F Fc) + X j i F F(.)]
w here X i ( F ) is the Fourier transform o f x/(t). T his is the basic relationship b etw een the spectrum o f the real bandpass signal x ( t ) and the spectrum o f the equivalent low pass signal xi(t). It is apparent from (9.1.21) that the spectrum o f the bandpass signal x ( t ) can b e ob tain ed from the spectrum o f the com p lex signal x t (t) by a frequency translation. T o be m ore precise, su p p ose that the spectrum o f the signal *;(/) is as sh ow n in Fig. 9.2(a). T hen the spectrum X ( F ) for p ositive freq u en cies is sim ply X f ( F ) translated in frequency to the right by Fc and scaled in am plitude by 5 . T he spectrum X ( F ) for negative freq u en cies is ob tain ed by first foldin g X / ( F ) about F = 0 to obtain X i ( - F ), conjugating X t ( ~ F ) to ob tain X * ( - F ) , translating X * ( - F ) in frequency to the left by /> , and scaling the result by T h e folding and conjugation o f X i ( F ) for the negative-frequency co m p on en t o f the spectrum result in a m agnitude spectrum |X (F )| that is even and a phase spectrum 2 ^X (F ) that is od d as show n in Fig. 9.2(b). T h ese sym m etry p roperties m ust hold since the signal x ( t ) is real valued. H ow ever, they do not apply to the spectrum o f the equ ivalen t com plex signal X i(t). T h e d ev elo p m en t ab ove im plies that a n y b a nd p ass signal x ( t ) can b e re pre se nted b y an eq u ivalent lo w pa ss signal x/(t). In general, the equivalent low pass signal x /( t) is com p lex valued, w hereas the bandpass signal x ( t ) is real. T h e latter can b e ob tain ed from the form er through the tim e-d om ain relation in (9.1.14) or through the frequency-dom ain relation in (9.1.21).
742
Sampling and Reconstruction of Signals |*,(F)|
Chap. 9
-f2
0 &i(F)
F,
(P
(a)
|X(F)|
-F-F,
-F-F+F,
F - F -, - Fr F + F ,
K X (n
Figure 9.2 (a) Spectrum of the lowpass signal and (b) the corresponding spectrum for the bandpass signal.
9.1.2 Sampling of Bandpass Signals

W e have already dem onstrated that a con tin u ou s-tim e signal with highest fre q uency B can be uniquely represented by sam p les taken at the m inim um rate (N yq uist rate) o f 2 B sam ples per secon d . H ow ever, if the signal is a bandpass signal w ith frequency com p onents in the band B\ < F < B i, as show n in Fig. 9.3, a blind ap plication o f the sam pling theorem w ould have us sam p lin g the signal at a rate o f 2 B 2 sam ples per secon d. If that w ere the case and B 2 w as an ex trem ely high freq u en cy, it would certainly be ad van tageou s to perform a freq u en cy shift o f the bandpass signal by
Sec. 9.1

X(F)
743
-B, an am ount Fc =
B-,
Figure 9.3 Bandpass signal with frequency components in the range

B\ < F 5 B2.
B x + B2
(9.1.22)
and sam pling the equivalent low pass signal. Such a frequency shift can be achieved by m ultiplying the bandpass signal as given in (9.1.12) by the quadrature carriers co s 2 jt F c[ and sin 2 n F ct and low pass filtering the products to elim inate the signal co m p o n en ts at 2 Fc. Clearly, the m ultiplication and the su b seq u en t filtering are first perform ed in the analog dom ain and then the ou tp u ts o f the filters are sam pled. T h e resulting eq u ivalen t low pass signal has a bandw idth B /2 , w h ere B = B2 B\. T h erefore, it can be represented uniquely by sam ples taken at the rate o f B sam p les per secon d for each o f the quadrature com p onents. Thus the sam pling can be p erform ed on each o f the low pass filter outputs at the rate o f B sam p les per secon d, as indicated in Fig. 9.4. T herefore, the resulting rate is 2 B sam ples per second. In v iew o f the fact that frequency con version to low pass allow s u s to reduce the sam pling rate to I B sam ples per secon d, it should be p ossib le to sam ple the bandpass signal at a com parable rate. In fact, it is. S u p p ose that the upper frequency Fc + B p , is a m ultiple o f th e bandwidth B (i.e., Fc + B / 2 = k B ) , w here k is a p ositive integer. If w e sam ple * ( /) at the rate
Figure 9.4 Sampling of a bandpass signal by first converting to an equivalent low-pass signal.
744
Chap. 9
2 B = 1 / r sam ples per secon d , w e have x ( n T ) = u c( n T ) o o s 2 n Fcn T us ( n T ) s i n 2 j i Fcn T n n ( 2 k - 1) . nn{2k - 1 ) = u c( n T ) c o s ------- ------------ us { n T ) s i n -------- -------(9.1.23)
w here the last step is ob tain ed by substituting Fc ~ k B B [2 and T = 1 /2 B. For n ev en , say n = 2m , (9.1.23) red u ces to x ( 2 m T ) = x ( m T \ ) = u c(m T i) cos 7rm(2k 1) = (l ) mu c(mT ]) w here T\ = 2 T = 1 /B . For n odd, say n = 2m 1, (9.1.23) red u ces to x(2 m T - T ) = x (m Ti - ^ = u s (' m T j - ^ ( - l ) m + t+ 1 (9.1.25) (9.1.24)
T h erefore, the even -n um b ered sam ples o f x{ t), which occur at the rate o f B sam p les per secon d, produce sam p les o f the low p ass signal com p on en t u c(t). T he odd-num bered sam ples o f x ( t ) , which also occu r at the rate o f B sam ples per secon d , produce sam ples o f the low pass signal com p onent K ,(f). N o w , the sam p les {uc{mT\)} and the sam ples (w,(m 7i 7 \/2 )) can b e used to reconstruct the equivalent low pass signals. Thus, according to the sam pling theorem for low p ass signals with T\ = 1 / B , uA t) = f.\
M , ) = J
^ , .,s in ( 7 r /r i) (/ - m 7 j ) 2 ^ u ^ m T 0 , / T Ul------- = rr( x / T \ ) ( t m T \) V'

L r
,n 1 ^ (9.1.26)
t
n _
^ ^ sin (?r/7i)(r mT\ + 7 \/2 )

2
) M w = * T r + T r m ( M 2 7 )
Furtherm ore, th e relation s in (9.1.24) and (9.1.25) allow us to express u c(t) and 3 (i) directly in term s o f sam p les o f x ( t ) . N ow , sin ce x ( t ) is exp ressed as x ( t ) = u c(t) c o s 2 n F ct u s (t) s m 2 n Fct (9.1.28)
substitution from (9.1.27), (9.1.26), (9.1.25), and (9.1.24) in to (9.1.28) yield s 'O = \, XQa ^ s in (jr /2 r )(r -2 m 7 )
) ( n / 2 T W - 2 r , T ) - C S 2 * F '
(9.1.29) B ut ( l ) m c o s 2 ti Fct = c o s 2 j r Fc{t 2 m T )
and
(l ) m+* sin27r Fct =
cos2nrFc(r 2m T + T )
Sec. 9.1
745
W ith th ese substitutions, (9.1.29) reduces to x (i)= ^ T sin (jr/27')(/ - m T ) ------ - c o s l n F c{ t - m T ) x(m T) [n/2T )(t - m T) (9.1.30)
w here T = 1 /2 B. T his is the desired reconstruction form ula for the bandpass signal x ( i ) , with sam p les taken at the rate o f I B sam ples per secon d , for the special case in w hich the upper band frequency Fc + B f 2 is a m ultiple o f the signal ban d w idth B. In the general case, where only the condition Fc > B f 2 is assum ed to hold, let us define the in teger part o f the ratio Fc + B / 2 to B as Fc + B / 2 B (9.1.31)
W hile holding the upper cu toff frequency Fc + B / 2 constant, w e increase the bandwidth from B to B' such that Fc + B / 2 D = r (9.1.32)
Furtherm ore, it is con ven ien t to define a new center frequency for the increased bandwidth signal as + (9.1.33)
Clearly, the increased signal bandw idth B' includes the original signal spectrum of bandwidth B. N o w the upper cu to ff frequency Fc + B / 2 is a m ultip le o f B', C onsequently, the signal reconstruction form ula in (9.1.30) h old s with Fc replaced by F'c and T replaced by T \ w here T ' = 1 /2 B \ that is, ^ * > - ^ /2 T '){t m T ') r ) l d r w - r ) . c m 2 7 , F {' ~ m T )
This p roves that x{ t) can b e represented by sam ples taken at the uniform rate 1 / T 2 B r ' / r , w here r' is the ratio ,
Fc + B / 2
j -
Fc
7 +
1
-
( 9 .1 .3 5 )
and r = \r' J. W e o b serve that w hen the upper cu toff frequency Fc + B / 2 is n ot an integer m ultiple o f the bandwidth B , the sam pling rate for the bandpass signal m ust be increased by the factor r ' f r . H ow ever, n ote that as Fc/ B increases, the ratio r ' f r ten d s tow ard unity. C on sequ en tly, the p ercent increase in sam pling rate tends to zero. T he derivation given ab ove also illustrates the fact that the low pass signal co m p o n en ts u c(t) and u s(t) can b e exp ressed in term s o f sam p les o f the bandpass
746
Chap. 9
signal. In d eed , from (9.1.24), (9.1.25), (9.1.26), and (9.1.27), w e obtain the result
mo and , > = V
j l <-ir,
- r ) Sjn(" / ? r ') ( , ~ 2 " r + n (ff /2 r ')( f - 2 n T ' + V ) ( 9 .U 7 )
w here r = Lr'J. In con clu sion , w e have dem onstrated that a bandpass signal can b e repre sen ted u n iqu ely by sam ples taken at a rate 2 B < Fs < 4 B w here B is the bandw idth o f the signal. T h e low er lim it ap p lies w h en the upper frequency F t. + B / 2 is a m ultiple o f B, T he upper lim it on Fs is ob tain ed under w orst-case con d ition s w hen r = 1 and r' % 2 .
9.1.3 Discrete-Time Processing of Continuous-Time Signals

A s indicated in our introductory rem arks in C hapter 1, th ere are num erous ap p lications w here it is advantageous to p rocess con tin u ou s-tim e (analog) signals on a digital signal processor. Figure 9.5 illustrates the general configuration of the system for d igital processing o f an analog signal. In design in g the processing to be p erform ed, w e m ust first select the bandw idth o f the signal to be processed since the signal bandw idth determ ines the m inim um sam pling rate. For exam ple, a speech signal, w hich is to be transm itted digitally, can con tain frequency com p o nen ts ab ove 3000 H z, but for p u rp oses o f sp eech intelligibility and sp eak er id en tification, the preservation o f frequency com p on en ts b elow 3000 H z is sufficient. C on sequ en tly, it w ould be inefficient from a processin g view p oin t to preserve the h igher-frequency com p on en ts and w asteful o f channel bandw idth to transm it the extra bits n eed ed to represent these h igher-frequency co m p o n en ts in the speech signal. O nce th e desired frequency band is selected w e can sp ecify the sam pling rate and the characteristics o f the prefilter, w hich is also called an antialiasing filter.
Antialiasing filte r T he antialiasing filter is an an alog filter which has a tw ofold purpose. First, it ensures that the bandw idth o f the signal to b e sam pled is lim ited to the desired frequency range. Thus any freq u en cy com p on en ts o f th e signal ab o v e th e foldin g frequency Fs f l are sufficiently atten u ated so that the
Figure 9.5
Configuration of system for digital processing of an analog signal.
Sec. 9.1
747
am ount o f signal distortion due to aliasing is negligible. F or exam ple, the sp eech signal to be transm itted digitally over a telep h o n e channel w ould be filtered by a low p ass filter having a passband exten d ing to 3000 H z, a transition band o f approxim ately 400 to 500 H z, and a stopband ab ove 3400 to 3500 H z. T he speech signal m ay be sam pled at 8000 H z and h en ce the fold in g frequency w ould be 4000 H z. Thus aliasing would be n egligible. A n o th e r reason for using an antialiasing filter is to lim it the additive n oise spectrum and other interference, w hich often corrupts the desired signal. U su ally, ad d itive n oise is w ideband and exceed s the bandw idth o f the desired signal. B y prefiltering w e reduce the additive noise p ow er to that which falls w ithin the bandwidth o f the desired signal and w e reject the ou t-of-b an d noise. Ideally, w e w ould like to em p loy a filter with steep cu toff frequency resp onse characteristics and with no delay distortion w ithin the passband. Practically, h ow ever, w e are constrained to em p loy filters that have a finite-w idth transition region, are relatively sim ple to im plem ent, and introduce som e tolerab le am ount o f delay distortion. V ery stringent filter specifications, such as a narrow transition region, result in very com plex filters. In practice, w e may c h o o se to sam p le the signal well ab ove the Nyquist rate and thus relax the design sp ecification s on the antialiasing filter. O n ce w e have specified the prefilter requirem ents and have selected the d e sired sam pling rale, w e can p roceed with the design o f the digital signal processing op eration s to be perform ed on the discrete-tim e signal. T h e selection o f the sam pling rate Fs = 1 / 7 , where T is the sam pling interval, not only determ ines the highest frequency (Fsf 2) that is preserved in the analog signal, but also serves as a scale factor that influences the design specifications for d igital filters and any other discrete-tim e system s through w hich the signal is processed . For exam ple, su p p ose that w e have an analog signal to be differentiated that has a bandw idth o f 3000 H z. A lth ou gh differentiation can be perform ed directly on the analog signal, w e ch oose to do it digitally in discrete tim e. H en ce w e sam ple th e signal at the range Fs = 8000 H z and d esign a digital d ifferentiator as described in Sec. 8.2.4. In this case, the sam pling rate Fs = 8000 H z estab lish es the foldin g frequency o f 4000 H z, which corresponds to the frequency w = n in the discrete-tim e signal. H en ce the signal bandwidth o f 3000 H z corresponds to the frequency wc = 0.75tt. C on sequ en tly, the discrete-tim e differentiator for p rocessing the signal would be designed to have a passband o f 0 < \w\ < 0.75jt. A s an oth er exam ple o f digital p rocessing, the sp eech signal that is b andlim ited to 3000 H z and sam pled at 8000 H z m ay be separated in to tw o or m ore frequency subbands by digital filtering, and each subband o f sp eech is digitally e n cod ed with different precision, as is d on e in subband cod in g (see Section 10.9.5 for m ore d etails). T h e frequency resp onse characteristics o f the digital filters for sep arating th e 0- to 3000-H z signal in to subbands are sp ecified relative to the folding frequency o f 4000 H z, which corresponds to the frequency co = n for the d iscrete tim e signal. Thus w e m ay process any con tinu ou s-tim e signal in the discrete-tim e dom ain by perform ing equivalent op eration s in discrete tim e.
746
Chap. 9
T he o n e im plicit assum ption that w e h ave m ade in this discussion on the eq u ivalen ce o f continuous-tim e and d iscrete-tim e signal p rocessin g is that the quan tization error in analog-to-digital con version and rou n d -off errors in digital signal processing are negligible. T h ese issu es are further discussed in this chapter. H o w ever, w e should em p hasize that analog signal processin g op eration s cannot be d on e very p recisely either, since electron ic com p on en ts in analog system s have toleran ces and th ey introduce n oise during their operation. In general, a d igi tal system designer has b etter con trol o f toleran ces in a digital signal processing system than an analog system d esign er w h o is designing an eq u ivalen t analog system .
9.2 ANALOG-TO-DIGITAL CONVERSION

T h e discussion in Section 9.1 focu sed on the con version o f con tinu ou s-tim e signals to discrete-tim e signals using an id eal sam pler and ideal in terp olation . In this section w e deal with the d evices for p erform ing th ese con version s from analog to digital. R ecall that the p rocess o f con vertin g a con tin u ou s-tim e (analog) signal to a digital seq u en ce that can b e p rocessed by a digital system requires that w e quantize the sam pled values to a finite num ber o f levels and represent each level by a num ber o f bits. The electron ic d evice that perform s this conversion from an analog signal to a digital seq u en ce is called an an alog-to-digital ( A / D ) converter (A D C ). On the other hand, a digital-to-analog ( D / A ) co nverter (D A C ) takes a digital seq u en ce and produces at its ou tp u t a voltage or current proportional to the size o f th e digital word applied to its input. D /A con version is treated in Section 9.3. Figure 9.6(a) show s a block diagram o f the basic elem en ts o f an A /D co n verter. In this section w e con sid er the perform ance requirem ents for these e l em ents. A lth o u g h we focu s m ainly on ideal system characteristics, w e shall also m en tion som e k ey im perfection s en cou n tered in practical d evices and indicate how th ey affect the perform ance o f the con verter. W e con cen trate on th ose aspects that are m ore relevant to signal processing applications. T h e practical aspects o f A /D converters and related circuitry can b e foun d in the m anufacturers specifications and data sheets.
9.2.1 Sample-and-Hold
In practice, the sam pling o f an analog signal is p erform ed b y a sam ple-and-hold (S /H ) circuit. T h e sam pled signal is th en q uantized and con verted to digital form. U sually, the S/H is integrated in to the A /D converter. T he S/H is a digitally con trolled analog circuit that tracks the analog input signal during the sam ple m od e, and then h old s it fixed during th e hold m ode to the instantaneous value o f the signal at the tim e th e system is sw itched from the
Sec. 9.2
Analog-to-Digital Conversion
749
(a)
Tracking
Figure 9.6 (a) Block diagram of basic elements of an A/D converter; (b) timedomain response of an ideal S/H circuit.
sam ple m o d e to the hold m ode. Figure 9.6(b) show s the tim e-dom ain resp onse o f an ideal S/H circuit (i.e., a S/H that responds instantaneously and accurately). T he goal o f the S/H is to continuously sam ple the input signal and then to hold that value constant as lon g as it takes for the A /D converter to ob tain its digital representation. T h e use o f an S/H allow s the A /D con verter to op erate m ore slow ly com pared to the tim e actually used to acquire the sam ple. In the ab sen ce o f a S/H , the input signal must n ot change by m ore than on e-h alf o f the quantization step during the conversion, which m ay be an im practical constraint. C on sequ en tly, the S/H is crucial in h igh-resolution (12 bits per sam ple or higher) digital conversion o f signals that have large bandw idths (i.e., they change very rapidly). A n ideal S/H introduces n o distortion in the conversion p rocess and is ac curately m o d eled as an ideal sam pler. H ow ever, tim e-related degradations such as errors in the p eriodicity o f the sam pling p rocess ( jitter), n onlinear variations in the duration o f the sam pling aperture, and changes in the voltage held during con version ( d ro o p ) d o occur in practical devices. T h e A /D converter begins the conversion after it receives a convert co m m and. T h e tim e required to com p lete the con version should b e less than the duration o f the hold m o d e o f the S/H . F urtherm ore, the sam pling p eriod T should b e larger than the duration o f the sam ple m od e and the hold m ode. In th e follow in g section s w e assum e that the S/H introduces n egligib le errors and w e fo cu s on the digital conversion o f the analog sam ples.
750
Chap. 9
9.2.2 Quantization and Coding

T h e basic task o f the A /D converter is to con vert a con tin u ou s range o f input am plitudes in to a discrete set o f digital cod e w ords. T his con version in volves the processes o f q ua ntiza tio n and coding. Q uantization is a n on linear and noninvertible p rocess that m aps a given am plitude x( n ) = x ( n T ) at tim e t = n T into an am plitude x k, taken from a finite set o f values. T h e p roced u re is illustrated in Fig. 9.7(a), w h ere the signal am plitude range is divided in to L intervals h = {** < x( n ) < x * + i} k = 1 ,2 ,..., L (9.2.1)
by the L + 1 decision levels x \ , x L, . . . , x L+\- T h e p ossib le ou tputs o f the quantizer (i.e., the quantization levels) are d en oted as x\ , x 2, . . . , x i . T h e op eration o f the quantizer is defined by the relation *<?() = G [*(n )j = *k if * (n ) Ik (9.2.2)
In m ost digital signal processing operations the m apping in (9.2.2) is in d ep en dent o f n (i.e., th e quantization is m em oryless and is sim ply d e n o te d as x q = Q[x]). Furtherm ore, in signal processing w e o ften use u n i f o r m or linea r q u an tizer s defined by 4+i - x k = A k = l,2 ,...,L -l (9.2.3) Xk+i - x k = A for finite x k, x t+] w here A is the q u a n tize r step size. U n iform quantization is usually a requirem ent if the resulting digital signal is to b e p rocessed by a digital system . H ow ever, in transm ission and storage applications o f signals such as sp ee ch , n onlinear and tim e-variant quantizers are frequently used. If a zero is assigned a quantization level, the quantizer is o f the m id tre a d type. If zero is assigned a decision level, the quantizer is called a m id rise type. Quantization levels Decision levels
<
\
*3 *3
\
*4 H x5
r~
xk xk xk*\
Instantaneous amplitude ------- (a)

x >~ xl x2 *2 x3 x5
*4 *4 -* 5
A O
x5
x6
X1
X1
xi
-* 8 -*9= 0 0
3A I
! - 4A
- 3A
- 2A
2A
>
Instantaneous amplitude-------
! --------------------------- Range of quantizer------------------------------ -I (b)

Figure 9.7 Quantization process and an example o f a midtread quantizer.
Sec. 9.2
751
Figure 9 .7 (b ) illustrates a m idtread quantizer with L = 8 levels. In theory, the extrem e d ecisio n lev els are taken as *i = oo and x l +i = oo, to cover the total d y n a m i c ra n g e o f the input signal. H ow ever, practical A /D converters can handle on ly a finite range. H en ce w e define the range R o f th e quantizer by assum ing that /] = I I = A. For exam ple, the range o f the quantizer show n in Fig. 9.7(b) is equal to 8 A , In practice, the term full-scale range (F S R ) is used to describe the range o f an A /D converter for bipolar signals (i.e., signals with b oth p ositive and negative am plitu d es). T h e term f u l l scale (F S ) is used for unipolar signals. It can b e easily seen that the q uantization error eq (n) is always in the range - A / 2 to A /2 : A A - < eq{n) < (9.2.4) In other w ords, the instantaneous quantization error cannot exceed half o f the q uantization step. If the dynam ic range o f the signal, defined as x m ax jcmjn, is larger than th e range o f the quantizer, the sam ples that exceed the quantizer range are clip p ed , resulting in a large (greater than A /2 ) quantization error. T h e op era tio n o f the quantizer is b etter described by the quantization char acteristic fun ction , illustrated in Fig. 9.8 for a m idtread quantizer with eight
Figure 9.8
Example of a midtread quantizer.
752
Chap. 9
quantization levels. This characteristic is preferred in practice over the midriser b ecau se it provides an output that is in sen sitive to infinitesim al changes o f the input signal about zero. N o te that the input am plitu d es o f a m idtread quantizer are rounded to the nearest quantization levels. T h e c od ing process in an A /D con verter assigns a unique binary num ber to each quantization level. If w e have L levels, w e n eed at least L different binary num bers. W ith a w ord length o f b + 1 bits w e can represent 2 b+l distinct binary num bers. H en ce w e should have 2b+1 > L or, eq u ivalen tly, b + 1 > log 2 L. T hen the step size or the re solution o f the A /D converter is given by
where R is the range o f the quantizer. There are various binary cod in g schem es, each with its advantages and dis advantages. Table 9.1 illustrates som e existin g sch em es for 3-bit binary coding. T h ese num ber representation schem es w ere d escribed in detail in Section 7.5. The tw o s-com p lem ent representation is used in m ost digital signal pro cessors. Thus it is con ven ien t to u se the sam e system to represent digital sig nals becau se w e can operate on them directly w ithou t any extra format conver-
TABLE 9.1
C O M M O N LY US ED BIPOLAR C O D E S Decimal Fraction
Number +7 +6 +5 +4 +3 +2 +1 0 0 -1 -2 -3 -4 -5 -6 -7 -8
Positive Reference +i +i +5 +i +1 +1 0+ 01 B
2 s
4
Negative Reference 7 8 h K 5 * 4 5 3 K 2 S 1 8 00+ +s +1
Sign + Magnitude 0111 0 110 0 10 1 0 1 00 00 1 1 0010 000 1 0 00 0 1000 100 1 10 10 10 11 1100 110 1 1110 1111
Two's Complement 0111 0110 0 10 1 0 10 0 0 0 11 0 0 10 000 1 0000 (0 0 0 0) 1111 1110 110 1 1100 10 11 10 10 10 0 1 (1 0 0 0)
Offset Binary 1111 1110 110 1 1100 10 11 10 10 100 1 1000 ( 1 0 0 0) 0111 0 110 0 10 1 0 10 0 00 1 1 00 1 0 000 1 (0 0 0 0)
Ones Complement 0111 0 110 0 10 1 0 1 00 00 1 1 00 1 0 000 1 0000 1111 1110 110 1 1100 10 11 10 10 10 0 1 1000
3
5
~S
+i!
+! +! +1 +i
~s 6 ~s 7 ~g 8 S
Sec. 9.2
753 P b has the
sion. In gen eral, a (b + 1) -bit binary fraction o f the form value - A > - 2 + f t 2 -1 + 02 2~2 H --------h f o 2~ b
if w e use th e tw o s-com p lem ent representation. N o te that 0 q is the m ost signifi cant bit (M S B ) and 0 b is the least significant bit (L S B ). A lth ou gh the binary cod e used to represent the quantization levels is im portant for the design o f the A /D converter and the su bsequent n um erical com putations, it d o e s n ot h ave any effect in the p erform ance o f the quantization process. Thus in our sub seq u en t discus sions w e ign ore the p rocess o f cod in g when w e analyze the perform ance o f A /D converters. Figure 9.9(a) sh ow s the characteristic o f an ideal 3-bit A /D converter. The only degradation introduced by an ideal con verter is the q uantization error, which can b e reduced by increasing the num ber o f bits. This error, that d om in ates the p erform ance o f practical A /D converters, is analyzed in the next section. Practical A /D con verters differ from id eal con verters in several w ays. V ariou s d egradations are usually en cou n tered in practice. A num ber o f these perform ance degradations are illustrated in Fig. 9 .9 (b )-(e ). W e n ote that practical A /D con vert ers m ay have offset error (the first transition m ay not occur at exactly + |L S B ) , scale-factor (or gain) error (the d ifferen ce b etw een the values at which the first transition and the last transition occur is n ot equal to FS 2L S B ), and linearity error (th e differen ces b etw een transition valu es are n ot all eq u al or uniform ly changing). If the differential linearity error is large en ou gh , it is possib le for on e or m ore cod e words to b e m issed. P erform ance data on com m ercially available A /D con verters are specified in the m anufacturers data sh eets.
9.2.3 Analysis of Quantization Errors

T o d eterm in e the effects o f quantization on the p erform ance o f an A /D converter, w e adopt a statistical approach. T h e d ep en d en ce o f the q uantization error on the characteristics o f the input signal and the n onlinear nature o f the quantizer m ake a determ in istic analysis intractable, excep t in very sim ple cases. In the statistical approach, w e assum e that the q uantization error is random in nature. W e m odel this error as n oise that is added to the original (unquan tized ) signal. If the input analog signal is within the range o f the quantizer, the q uantization error eq(n) is b ou n ded in m agnitu d e [i.e., \eq {n)\ < A /2 ], and the resulting error is called gra nular noise. W hen the input falls ou tsid e the range o f the quantizer (clipping), eq (n) b eco m es unb oun d ed and results in o verloa d noise. T his type o f n o ise can result in severe signed distortion w h en it occurs. O ur on ly rem ed y is to scale the in p ut signal so that its dynam ic range falls within the range o f the quantizer. T h e follow in g analysis is b ased on the assum ption that there is n o overload n oise. T h e m athem atical m od el for th e quantization error ev (n) is sh ow n in Fig. 9.10. T o carry ou t th e analysis, w e m ake th e follow in g assum ptions about the statistical
754
Output digital number (code and fractional value)
Chap. 9
Normalized analog input (a)
(b)
(c)
2
(d)
(e)
Figure 9.9 Characteristics of ideal and practical A/D converters.
properties o f eq( n ): L T h e error eq(n) is uniform ly distributed over the range A /2 < eq{n) < A /2. 2. T h e error seq u en ce is a stationary w h ite n o ise seq u en ce. In other w ords, the error eq (n ) and the error eq (m) for m are uncorrelated.
Sec. 9.2
Quantizer
755
x(n)
GM>]
(a) Actual system
xAn)
x(n)-
&
X(fn) = x (n) + e^n)
eq{n)
(b) Mathematical model
Figure 9.10 M athem atical m odel of

quantization noise.
3. T h e error se q u en ce {eQ(n)} is uncorrelated with th e signal seq u en ce x ( n ). 4. T h e signal se q u en ce x ( n ) is zero m ean and stationary. T h ese assu m p tion s do not hold, in general. H ow ever, they do hold w hen the quan tization step size is sm all and the signal seq u en ce x ( n ) traverses several quan tization levels b etw een tw o successive sam ples. U n d er these assum ptions, the effect o f the additive n o ise eq(n) on the d esired signal can b e quantified by evaluating the signal-to-quantization n oise (pow er) ratio (S Q N R ), w hich can be exp ressed on a logarithm ic scale (in decib els or d B ) as S Q N R = 101og 10 M I (9.2.6)
w here Px = a j = E [ x 2{n)} is the signal p ow er and P = a ] E[e^(n)] is the p ow er o f the qu an tization n oise. If th e quantization error is uniform ly distributed in the range (A /2 , A /2 ) as sh ow n in Fig. 9.11, th e m ean valu e o f th e error is zero and th e variance (the quantization n o ise p ow er) is
/
Pie) 1
A
A/2
f A/2
A2
. , / m d e = z L A/2
(9.2.7)
Figure 9.11 Probability density function for the quantization error.
756
Chap. 9
By com bining (9.2.5) w ith (9.2.7) and substituting the result in to (9.2.6), the exp ression for the S Q N R b ecom es S Q N R = 10 log = 20 log
P a,
(9.2.8) = 6.02* + 16.81 - 20 log ox dB
T he last term in (9.2.8) d ep en d s o n the range R o f the A /D converter and the statistics o f the input signal. For exam p le, if w e assum e that jf(n) is Gaussian distributed and the range o f the quantizer exten d s from 3ax to 3 ax (i.e., R = 6 ax ), then less than 3 o u t o f every 1000 input signal am plitudes w ou ld result in an overload on the average. F or R = 6ax , (9.2.8) b eco m es S Q N R = 6.02* + 1.25 dB T h e form ula in (9.2.8) is frequently used to specify the p recision n eed ed in an A /D converter. It sim ply m ean s that each additional bit in the quantizer increases the signal-to-quantization n oise ratio by 6 dB. (It is interesting to n ote that the sam e result w as derived in Section 1.4 for a sinusoidal signal using a determ inistic approach.) H o w ev er, w e should bear in m ind the con d ition s under which this result has b een derived. D u e to lim itations in the fabrication o f A /D converters, their perform ance falls short o f the theoretical value given by (9.2.8). A s a result, the effective number o f bits m ay b e som ew h at less than the num ber o f bits in the A /D converter. For instance, a 16-bit converter m ay have on ly an effective 14 bits o f accuracy.
9.2.4 Oversampling A/D Converters

T he basic idea in oversam pling A /D converters is to increase the sam pling rate o f the signal to th e point w here a low -resolu tion quantizer suffices. B y oversam pling. w e can reduce the dynam ic range o f the signal valu es b etw een successive sam ples and thus reduce the resolu tion requirem ents on the quantizer. A s w e have ob served in the p receding section , the variance o f the q uantization error in A /D conversion is cr,2 = A 2 /1 2 , w here A = R f l b+l. Since the dynam ic range o f the signal, which is proportional to its standard deviation o x , should m atch the range R o f the quantizer, it follow s that A is p roportional to ax . H en ce for a given num ber o f bits, the p ow er o f the quantization n oise is proportional to the variance o f the signal to b e quantized. C on sequ en tly, for a given fixed S Q N R , a reduction in th e variance o f the signal to be q uantized allow s us to red u ce the num ber o f bits in the quantizer. T h e basic id ea for reducing the dynam ic range leads us to con sid er differential quantization. T o illustrate this p oin t, let us evalu ate th e variance o f the difference b etw een tw o su ccessive signal sam ples. Thus w e have d ( n ) x (n ) jc (n 1) (9.2.9)
Sec. 9.2
Analog-toDigital Conversion
757
T h e variance o f d ( n ) is u l = E [ d 2(n)] = {[x (n ) - x ( n - l ) ] 2} = E [ x \ n ) } - 2 E [ x ( n ) x ( n - 1)] + E [ x 2(n - 1)] = 2<7j[1 y* A 1
(9.2.10)
)]
w here y xx( l ) is the value o f the autocorrelation seq u en ce yxx(m) o f x ( n ) evalu ated at m = 1. If Yx.r (l) > 0.5, w e ob serve that a J < cr^. U n d er this condition, it is better to quantize the d ifferen ce d{n) and to recover x( n ) from the quantized valu es {d9(n)}. T o obtain a high correlation b etw een successive sam ples o f the signal, w e require that the sam pling rate be significantly higher than the N yquist rate. A n ev en b etter approach is to quantize the d ifference d ( n ) = x ( n ) a x ( n 1) (9.2.11)
w here a is a param eter selected to m inim ize the variance in d{n). This leads to the result (see Problem 9.7) that the optim um ch oice o f a is
a = Yxx^ = Yxx^
Yxx (0)
and ad 2 = a2 x [\ - a 2] (9.2.12)
In this case, ctJ < a 2, since 0 < a < 1. T he quantity a x ( n 1 ) is called a first-order predictor o f jc(). Figure 9.12 sh ow s a m ore general differential predictiv e signal quantizer sys tem . This system is u sed in sp eech en cod in g and transm ission over telep h on e channels and is know n as differential p u lse co d e m odulation (D P C M ). T h e goal o f the predictor is to provide an estim ate x ( n ) o f x ( n ) from a linear com bination o f past values o f x ( n ) , so as to red u ce the dynam ic range o f the difference signal d ( n ) = x { n ) x ( n ). T hus a predictor o f order p has the form p x ( n ) = ^ 2 akx (n - k ) (9.2.13)
Figure 9.12 Encoder and decoder for differential predictive signal quantizer system.
758
Chap. 9
T he use o f the feed b ack loop around the quantizer as sh ow n in Fig. 9.12 is n ec essary to avoid the accum ulation o f quantization errors at the d ecoder. In this configuration, the error e ( n ) = d ( n ) dq (n) is e(n) = d ( n ) - dq(n) = x(n ) - x (n) - dq (n) = Jf(n) - x q (n) Thus the error in the reconstructed quantized signal x q (n) is equal to the quan tization error for the sam ple d (n). T h e d ecod er for D P C M that reconstructs the signal from the q uantized values is also show n in Fig. 9.12. T h e sim p lest form o f differential predictive quantization is called delta m o d ulation (D M ). In D M , the quantizer is a sim ple 1-bit (tw o -lev el) quantizer and the predictor is a first-order predictor, as shown in Fig. 9.13(a). B asically, D M pro vid es a staircase approxim ation o f the input signal. A t every sam pling instant, the sign o f the d ifferen ce b etw een the input sam ple x ( n ) and its m ost recent staircase approxim ation x ( n ) = a x q (n 1 ) is determ ined, and then the staircase signal is updated by a step A in the direction o f the d ifference. From Fig. 9.13(a) w e ob serve that x q (n) = a x g(n - 1) + dq {n) (9.2.14)
which is the d iscrete-tim e equivalent of an analog integrator. If a = 1, w e have an ideal accum ulator (integrator) w hereas the ch oice a < 1 results in a leaky integrator. Figure 9.13(c) show s an analog m od el that illustrates the basic prin ciple for the practical im plem en tation o f a D M system . T h e analog low pass filter is necessary for the rejection o f out-of-band com p on en ts in the frequency range b etw een B and Fs f l , since Fs > > B due to oversam pling. T he crosshatched areas in Fig. 9.13(b) illustrate tw o typ es o f quantization error in D M , slo p e-o v erload distortion and granular n oise. Since the m axim um slop e A ( T in x ( n ) is lim ited by the step size, slop e-overload distortion can be avoided if m ax \ d x ( t ) / d t \ < A / T . T h e granular n oise occurs w h en the D M tracks a relatively flat (slow ly changing) input signal. W e n ote that increasing A reduces overload distortion but increases th e granular n oise, and vice versa. O ne way to reduce these tw o types o f distortion is to use an integrator in front o f the D M , as sh ow n in Fig. 9.14(a). This has tw o effects. First, it em phasizes the low freq u en cies o f x (t) and increases the correlation o f th e signal in to the D M input. S econd, it sim plifies the D M d ecod er b ecau se the d ifferen tiator (inverse system ) required at th e d ecod er is can celed by the D M integrator. H en ce the d ecod er is sim ply a low pass filter, as show n in Fig. 9.14(a). Furtherm ore, the two integrators at the en co d er can b e replaced by a single in tegrator placed before the com parator, as show n in Fig. 9.14(b). T h is system is know n as sigma-delta m o d u l a t i o n (S D M ). S D M is an ideal candidate for A /D con version . Such a con verter takes advantage o f the high sam pling rate and spreads the quantization n oise across the band up to Fs f2 . Since Fs > > B, the n oise in the signal-free band B < F < F , f l can b e rem oved by appropriate digital filtering. T o illustrate this principle, let
Sec. 9.2
759
Coder (a)
Decoder
(b) T T
(c)
Figure 9.13 Delta modulation system and two types of quantization errors. us con sid er the discrete-tim e m od el o f S D M , sh ow n in Fig. 9.15, where w e have assum ed that the com parator ( 1 -bit quantizer) is m od eled by an additive w hite n o ise source w ith variance a f = A 2 /12. T h e integrator is m od eled b y the d iscrete tim e system with system function (9.2.15)
760
Chap. 9
Figure 9.14
Sigm a-detta m odulation system.
H{z)
ein)
T h e z-transform o f the sequ en ce {dq (n)} is

1
1 + H {z )
1 + H (z)
E (z ) (9.2.16)
= H A z)X (z) + H(z)E(z) w h ere H s {z) and H{z) are the signal and n oise system fun ction s, respectively. A go o d S D M system has a flat frequency resp onse Hs {co) in th e signal frequency
Sec. 9.2
Analog-to-Digrtal Conversion
761
band 0 < F < B . O n the other hand, H(z) should have high attenuation in the frequency band 0 < F < B and low attenuation in the band B < F < Fs/2. For th e first-order S D M system with th e integrator specified by (9.2.15), w e have Hs (z) = z ~ 1 H n(z) = l - z ~ 1 (9.2.17)
T hus H s (z) d o es n o t distort the signal. T h e p erform ance o f the S D M system is therefore d eterm in ed by the n oise system function H n(z), w hich has a m agnitude frequency response J (/^) I = 2 sin (9.2.18) Fs
as show n in Fig. 9.16. T h e in-band quantization n oise variance is given as (9.2.19) w here Se( F ) = a ] j F s is the p ow er spectral density o f the quantization n oise. From this relationship w e n ote that doubling Fs (increasing the sam pling rate by a factor o f 2), w hile k eep in g B fixed, reduces the p ow er o f the quantization n oise by 3 dB. T his result is true for any quantizer. H ow ever, additional reduction m ay be possib le by properly ch oosin g the filter H ( z ) . For the first-order S D M , it can be sh ow n (see P roblem 9.10) that for Fs > > 2 B, the in-band q uantization n oise p ow er is (9.2.20) N o te that doubling th e sam pling frequency reduces the n oise p ow er by 9 dB o f which 3 dB is due to the reduction in St ( F ) and 6 dB is due to the filter charac teristic H (F). A n additional 6 -dB reduction can be ach ieved by using a d ou b le integrator (see P rob lem 9.11). In sum m ary, th e n oise p ow er can be reduced by increasing the sam pling rate to spread the quantization n oise ppw er over a larger frequency band ( - F J 2, Fs/ 2), and then shaping th e n oise p ow er spectral density by m eans o f an
F, ~T
Figure 9.16
F,
2 Frequency (magnitude) response o f noise system function.
762
Analog section

Digital section
Chap. 9
SDM - to - PCM Converter
Figure 9.17
Basic elem ents o f an oversam pling A /D converter.
appropriate filter. Thus, S D M provides a 1-bit quantized signal at a sam pling fre q uency Fs = 2I B , w here the oversam pling (interp olation) factor I d eterm in es the SN R o f the S D M quantizer. N ext, we explain how to convert this signal in to a b-bit q uantized signal at the N yquist rate. First, w e recall that the S D M d ecod er is an an alog low p ass filter with a cu toff frequency B. T h e output o f this filter is an approxim ation to the input signal x(r). G iven the 1-bit signal dq {n) at sam pling frequency Fs , w e can obtain a signal x q {n) at a low er sam pling frequency, say the N yquist rate o f 2 B or som ew hat faster, by resam pling the output o f the low pass filter at the 2 B rate. T o avoid aliasing, w e first filter out the out-of-band (fl, F J 2) n o ise by processing the w ideband signal. T h e signal is then passed through the low pass filter and resam pled (dow nsam pled ) at the low er rate. T he dow nsam pling process is called decimation and is treated in great detail in Chapter 10. For exam p le, if the interpolation factor is I = 256, the A /D con verter output can be obtained by averaging su ccessive non-overlapping blocks o f 128 bits. This averaging w ould result in a digital signal with a range o f valu es from zero to 256{b as 8 bits) at the N yquist rate. T he averaging p rocess also provid es the required antialiasing filtering. Figure 9.17 illustrates the basic elem en ts o f an oversam p lin g A /D converter. O versam pling A /D con verters for voice-band (3-k H z) signals are currently fab ricated as integrated circuits. T ypically, they op erate at a 2 -M H z sam pling rate, dow nsam ple to 8 kH z, and provide 16-bit accuracy.
Sec. 9.3
Digital-to-Analog Conversion
763
9-3 DIGITAL-TO-ANALOG CONVERSION

In Section 4.2.9 w e dem onstrated that a bandlim ited low p ass analog signal, which has b een sam p led at the N yquist rate (or faster), can be reconstructed from its sam ples w ithou t distortion. T h e ideal reconstruction form ula or ideal interpolation form ula derived in Section 4.2.9 is sin(:rr/T)(r n T ) (n/T )(t - n T ) (9.3.1)
w here the sam pling interval T l / F s = 1 /2 B , Fs is th e sam pling frequency and B is the bandwidth o f the analog signal. W e have view ed the reconstruction o f the signal x { t ) from its sam ples as an in terp olation prob lem and have described the function s in ^ r /T ) g ( t) = n t/T (9.3.2)
as the ideal interpolation function. T he interpolation form ula for x(r), given by (9.3.1), is basically a linear superposition o f tim e-sh ifted versions o f g (t). with each g (t n T ) w eigh ted by the corresponding signal sam ple x ( n T ) . A ltern atively, w e can view the reconstruction o f th e signal from its sam ples as a linear filtering p rocess in which a discrete-tim e seq u en ce o f short pulses (ideally im pulses) with am plitudes equal to the signal sam ples, excites an analog filter, as illustrated in Fig. 9.18. T h e analog filter corresponding to the ideal interpolator has a frequency response T, H (F) =
0, 1
IFf <
Fs
2T 1
(9.3.3)
|F | >
2T
H ( F ) is sim ply the F ourier transform of the in terp olation function g{t). In other w ords, H { F ) is the frequency resp onse o f an analog reconstruction filter w h ose
Ideal analog lowpass filter
H (F)
Input signal
F.
Reconstructed signal oc sin

x(nT )~ 0 - riT)
Y .
x(nT)& (t ~ nT)
x (t)=
Y l
--------------
n=-oc
(! nT )
Figure 9.18 Signal reconstruction viewed as a filtering process.
764

H (F )
Chap. 9
L
'
Fi
2
(a) hU)
(b) Figure 9.19 Frequency response (a) and the impulse response (b) of an ideal low-pass filter. im pulse response is h(t) = g(t). A s show n in Fig. 9.19, the id eal reconstruction filter is an ideal low pass filter and its im pulse resp on se exten d s for all tim e. H ence the filter is noncausal and physically nonrealizable. A lth ou gh th e interpolation filter with im pulse resp onse given by (9.3.1) can be approxim ated closely with som e delay, the resulting function is still im practical for m ost ap p lication s where D /A con version is required. In this sectio n w e p resent som e practical, alb eit nonideat. in terp olation tech niques and interpret them as linear filters. A lth ou gh m any sophisticated poly nom ial interpolation tech n iqu es can b e d evised and analyzed, our discussion is lim ited to constant and linear interpolation. Q uadratic and h igh er polyn om ial in-
Sec. 9.3
765
terp olation is often used in num erical analysis, but is it less likely to be used in digital signal processing.
9.3.1 Sample and Hold

In practice, D /A conversion is usually perform ed by com bining a D /A converter w ith a sam ple-and-hold (S/H ) and follow ed by a low pass (sm ooth in g) filter, as show n in Fig. 9.20. T h e D /A converter accepts at its input, electrical signals that correspond to a binary w ord, and p rod u ces an output voltage or current that is proportional to the valu e o f the binary w ord. Ideally, its in p u t-ou tp u t characteristic is as show n in Fig. 9.21(a) for a 3-bit bipolar signal. T he line connecting the d ots is a straight line through the origin. In practical D /A converters, the line connecting the d ots m ay d eviate from the ideal. Som e o f the typical d eviation s from ideal are offset errors, gain errors, and nonlinearities in the in p u t-ou tp u t characteristic. T h ese types o f errors are illustrated in Fig. 9.21(b). A n im portant param eter o f a D /A converter is its settling tim e , which is defined as the tim e required for the output o f the D /A converter to reach and rem ain w ithin a given fraction (usually, i^ L S B ) o f the final value, after applica tion o f the input cod e word. O ften , the ap plication o f the input cod e word results in a high-am plitude transient, called a glitch . This is especially the case w hen tw o co n secu tive cod e words to the A /D differ by several bits. T h e usual w ay to rem edy this problem is to use a S/H circuit d esigned to serve as a d eglitcher. H en ce the basic task o f the S/H is to hold th e output o f the D /A converter c o n stant at the p revious output value until the n ew sam ple at the output o f the D /A reach es steady state, then it sam ples and h olds the new valu e in the next sam pling interval. Thus the S/H approxim ates the analog signal by a series o f rectangular p u lses w h o se height is equal to the corresponding value o f the signal pulse. F ig ure 9.22(a) illustrates the approxim ation o f the analog signal x ( t ) by a S/H . A s sh ow n , the approxim ation, d en oted as x ( t ), is basically a staircase function which ta k es the signal sam ple from the D /A con verter and h old s it for T secon ds. W hen the n ext sam ple arrives, it jum ps to the n ext value and h old s it for T secon ds, and so on. W hen view ed as a linear filter, as show n in Fig. 9.22(b), the S/H has an im pulse response h ( t) = 1,
0,
0 < t< T otherw ise
(9.3.4)
Digital input
signal
Digitalto-analog convener
Sample
and hold
Lowpass smoothing filter
Analog output signal
Figure 9.20
Basic operations in converting a digital signal into an analog signal.
766
Chap. 9
Analog output voltage 3A Ideal D/A
2A
/
_l_______1 _______L .
100 101 110 111 000
001
010
011
Input code words
/
/
/
-A
-2 A
-4A
O ffset : e rro r
//
sy
i i i
Z x "
Scale factor or gain error
JL,
r /1 i i i
L _i _u-i i i i
Nonmonoionicity
Figure 9.21 (a) Ideal D/A converter characteristic and (b) typical deviations from ideal performance in practical D/A converters.
Sec. 9.3
767
Sampled signal -
S /H
- x(t )
(b)
*(/)
(c)
Figure 9.22 (a) Approximation of an analog signal by a staircase; (b) linear filtering interpretation; (c) impulse response of the S/H.
T his is illustrated in Fig. 9.22(c). T h e corresponding freq u en cy resp onse is
O O h{t)e->2F'd t
00
[ T e - j2* F,d t Jo
,-jxFT
(9.3.5)
r ( sin n F T \ \ 7TF T )
T h e m agnitude and p h ase o f H ( F ) are p lotted in Figs. 9.23. For com parison, the freq u en cy resp o n se o f th e id eal interpolator is su p erim p osed o n th e m agnitude characteristics. It is apparent that the S/H d o es not p ossess a sharp cu toff frequency re sp on se characteristic. T his is du e to a large exten t to the sharp transitions o f its im pulse resp onse h ( t). A s a con seq u en ce, the S /H p asses undesirable aliased frequency co m p o n en ts (freq u en cies ab ove Fsf 2) to its output. T o rem edy this problem , it is com m on practice to filter i( r ) by passing it through a low pass filter
768
Sampling and Reconstruction of Signals l(F)|
Chap. 9
9(F)
which highly atten u ates frequency com p onents ab ove F J 2. In effect, the low pass filter follow ing the S/H sm ooth s the signal x ( t ) by rem oving th e sharp discontinu ities.
9.3.2 First-Order Hold

A first-order hold approxim ates x ( t ) by straight-line segm en ts w hich have a slop e that is d eterm ined by the current sam ple x { n T ) and the previou s sam ple x ( n T T). A n illustration o f this signal reconstruction tech n iqu es is given in Fig. 9.24. T h e m athem atical relationship b etw een the input sam p les and the output w aveform is s(r)
= x(nT) +
T)
X ^ n T ~ T - (t
raf)
nr<K(n
l)r
(9.3.6)
Sec. 9.3
769
W hen view ed as a linear filter, the im pulse resp onse o f the first-order hold is
1 + T'
0 <t <T T < t < IT otherw ise (9.3.7)
h(t) =
1 0,
t T
T his im pulse response is dep icted in Fig. 9.25(a). T h e Fourier transform o f h ( t) y ield s the frequency response, which can be exp ressed in the form H ( F ) = T ( \ + A n F 2T 2) X f l ( ^ ^ ^ j w here the phase 0 (F ) is (F ) = n F T + tan - 1 2 n F T (9.3.9) jm F ) e} (9.3.8)
T h ese frequency resp onse characteristics are graphically illustrated in Fig. 9.25(b) and (c). Since this reconstruction tech n iqu e also suffers from distortion due to passage o f frequency com p o n en ts ab ove Fs/ 2, as can be ob served from Fig. 9.25(b), it is fo llo w ed by a low pass filter that significantly atten u ates freq u en cies ab ove the fold in g frequency Fs f l . T h e p eak s in H { F ) within the band |F | < Fsf l m ay be undesirable in som e applications. In such a case it is possib le to m odify th e im pulse response by reducing the slop e by so m e factor 0 < 1. T his results in th e im pulse response h (t) illustrated in Fig. 9.26(a). T he corresponding frequency resp onse is given by
-r[i-
0 + 0(1 + j l n F T )
sin n F T , - j xFT I sin jrF T nFT * n FT
(9.3.10)
T h e m agnitude |/ / ( F ) | is illustrated in Fig. 9.26(b) for 0 = 0.5, 0 = 0.3, and 0 = 0.1. W e n o te that the peak in H ( F ) is relatively sm all for 0 = 0.3 and
k(l)
\H {F )\
(b)
9(F)
Figure 9.25 Impulse response (a) and frequency response characteristics (b) and (c) for a first-order hold. 770
Sec. 9.3
h(t)
771
MF)!
2T
IT
(b)
Figure 9.26 Impulse response (a) and frequency (magnitude) response (b) for a modified first-order hold. d o es not exist w h en 0 = 0.1. T hus this m odified first-order hold exhibits b etter frequency resp onse characteristics in the frequency range |F | < Fs/ 2.
9.3.3 Linear Interpolation with Delay

T h e first-order hold perform s signal reconstruction by com p utin g th e slo p e o f the straight lin e b ased on th e current sam p le x ( n T ) and th e past sam p le x ( n T T ) o f
772
Chap. 9
the signal. In effect, this tech n iqu e linearly extrapolates or attem p ts to linearly p r e dict the next sam ple o f the signal based on the sam ples x ( n T ) and x { n T - T). A s a co n sequ en ce, the estim ated signal w aveform i ( r ) contains jum ps at the sam ple points. The jum ps in x ( t ) can be avoided by providing a o n e-sa m p le delay in the re construction process. Then successive sam ple p oin ts can b e con n ected by straightlin e segm ents. Thus the resulting interpolated signal x ( t ) can b e exp ressed as x(t) = x{nT - T ) + x ( n T ) x ( n T 7") (t n T ) n T < t < (n + 1 )T (9.3.11)
W e ob serve that at t n T , x { n T ) = x ( n T T ) and at t = n T + T , x ( n T + T ) = x ( n T ) . T herefore, x ( t ) has an inherent delay o f T secon d s in interpolating the actual signal jc(r). Figure 9.27 illustrates this linear in terp olation technique. V iew ed as a linear filter, the linear interpolator with a 7 -se c o n d d elay has an im pulse response 0 < t < T i t IT. (9.3.12) hit) = 2 -t/T . T < t < 2T otherw ise 10 . T he corresponding frequency response is
+ ( 2 r
dt (9.3.13)
-j2x Fi
T h e im pulse response and frequency resp onse characteristics o f this interpolation filter are illustrated in Fig. 9.28. W e observe that the m agnitude characteristic falls o ff rapidly and contains sm all sid elo b es b eyon d the sam p lin g frequency Fs. Furtherm ore, its p h ase characteristic is linear due to the d elay T. B y follow ing this interpolator with a low pass filter that has a sharp cu toff b eyon d the frequency Fs/2, the high-frequency com p on en ts in x ( t ) can b e further reduced.
Figure 9.27
Linear interpolation of x{t) with a 7-second delay.
Sec. 9.3
A (r)
773
(a)
IW )I
_3 T
__2 T
_ J _ __ 1_ 0 T IT (b) 9(F)
J_ IT
J_ T
2_
2 T
(c)
Figure 9.28 Impulse response (a) and frequency response characteristics (b) and (c) for the linear interpolator with delay.
774
Chap. 9
Digital section
Analog section
Figure 93.9 Elements of an oversampling D/A converter. This con clu d es our discussion o f signal reconstruction b ased on sim ple inter p olation techniques. T he tech n iqu es that w e have described are easily incorporated into the design o f practical D /A converters for the reconstruction o f analog signals from digital signals. W e shall con sid er interpolation again in C hapter 10 in the context o f changing the sam pling rate in a digital signal processin g system .
9.3.4 Oversampling D/A Converters

T he elem en ts o f an oversam pling D /A converter are shown in Fig. 9.29. A s we ob serve, it is subdivided in to a digital front end follow ed by an an alog section. The digital section con sists o f an interpolator w h ose function is to increase th e sam pling rate by som e factor / , and th en is follow ed by a SD M . T he in terp olator sim ply increases the digital sam pling rate by inserting I 1 zeros b etw een successive low rate sam ples. T he resulting signal is then processed by a digital filter with cu toff frequency Fc = B / F s in order to reject the im ages (replicas) o f the input signal spectrum . T his higher rate signal is fed to the S D M , w hich creates a n o ise shaped 1-bit sam ple. E ach 1-bit sam ple is fed to the 1-bit D /A . w hich provides the analog interface to th e an tialiasing and sm ooth in g filters. T h e output analog filters have a passband o f 0 < F < B hertz and serve to sm ooth the signal and to rem ove the quantization n oise in the frequency band B < F < Fs/2. In effect, the oversam pling D /A converter uses S D M with the roles of th e analog and digital section s reversed com pared to the A /D converter. In practice, oversam pling D /A (and A /D ) converters have m any advantages over the m ore co n v en tion al D /A (and A /D ) converters. First, the high sam pling rate and the su b seq u en t digital filtering m inim ize or rem ove the n eed for com plex and exp en sive analog an tialiasing filters. Furtherm ore, any an alog n o ise introduced during the con version p h ase is filtered out. A lso , there is no n eed for S/H circuits. O versam pling S D M A /D and D /A con verters are very robust with resp ect to vari ations in the analog-circuit param eters, are inherently linear, and have low cost.

T h e m ajor focus o f this chapter w as on the sam pling and reconstruction o f sig nals. In particular, w e treated the sam pling o f con tinu ou s-tim e signals and the subsequent op eration o f A /D con version . T h ese are necessary op eration s in the
Chap. 9
Problems
775
digital p rocessin g o f a n alog signals, either on a gen eral-p u rp ose com puter or on a cu stom -d esign ed digital signal processor. T he related issue o f D /A conversion was also treated. In addition to the con ven tional A /D and D /A con version techniques, w e also described a n oth er type o f A /D and D /A con version , based on the principle o f oversam p lin g and a type o f w aveform en cod in g called sigm a-delta m odulation. S igm a-d elta con version tech n ology is especially suitable for audio band signals due to their relatively sm all bandwidth (less than 20 k H z) and in som e applications, th e requirem ents for high fidelity. T h e sam pling th eorem was introduced by N yquist (1928) and later p opular ized in the classic paper by Shannon (1949). D /A and A /D con version techniques are treated in a b ook by Sheingold (1986). O versam pling A /D and D /A con ver sion has b een treated in the technical literature. Specifically, w e cite the work o f C andy (1986), C andy et al. (1981) and Gray (1990).
PROBLEMS
9.1 Consider the sampling of the bandpass signal whose spectrum is illustrated in Fig. P9.1. Determine the minimum sampling rate Fx to avoid aliasing.
X(F)
9 2 Consider the sampling of the bandpass signal whose spectrum is illustrated in Fig. P9.2. Determine the minimum sampling rate Fs to avoid aliasing. X(F)
9.3 Prove that xi(t) is generally a complex-valued signal and give the condition under which it is real. Assume that x(t) is a real-valued bandpass signal. 9.4 Consider the two systems shown in Fig. P9.4. (a) Sketch the spectra of the various signals if xa(t ) has the Fourier transform shown in Fig. 9.4(b) and F, = 2 B. How are >](f) and ^ (z) related to xa(t)f ( b ) Determine >i(r) and ^(z) if xa(t) = cos2nFot, Fo = 20 Hz, and F, = 50 Hz or Fs = 30 Hz.
776
Chap. 9
F,
(a)
F,
XJF)
O
(b)
Figure P9.4
9.5 A contin uou s-tim e signal xa(t) with bandw idth B and its ech o xa(t - r) arrive sim ul taneously at a T V receiver. T h e received an aiog signal
sa(t) = xa(r) + a x a( t - T )
|| < 1
is processed by the system shown in F ig, P 9.5. Is it possible to specify F , and H (z ) so that ya(t) = xa(t) [i.e., rem ov e th e g h ost xa(i - r ) from th e received signal]?
F,
Figure P9J
9 j6 A bandlim ited co n tin u ou s-tim e signal x(t) is sam pled at a sam plin g frequen cy F s > 2 B . D ete rm in e th e en ergy E d o f th e resulting d iscrete -tim e signal x(n) as a function o f th e energy o f th e an alog signal, E a, and th e sam pling period T = 1 / F S.
9 .7 L e t x(n) b e a zero-m ean statio n ary p rocess with v arian ce and a u to co rrela tio n y j l ) .
Chap. 9
Problems
777
a j = a 2[l + a2 - 2apx(l)] where px(l) = j'x(l)/?'* (0) is the normalized autocorrelation sequence, (b) Show that <rj attains its minimum value
for a = Kv(l)/Xr() = ArC1) (c) U nder what conditions is a j < a ;? (d) R epeat steps (a) to (c) for the second-order prediction error d(n) = x(n) a^x(rt 1) aixln 2)
9 .8 Consider a DM coder wilh input ,v(n) = A cos(2 n n F/ F, ). W hat is the condition for
avoiding slope overload? Illustrate this condition graphically. 9.9 Let x(t) be a bandlimited signal with fixed bandwidth B and variance rr*. (a) Show that the signal-lo-quantization noise ratio. SQNR = 101og,n(fl7/<v). in creases by 3 dB each time we double the sampling frequency F,. Assume that the quantization noise model discussed in Section 9.2.3 is valid. (b ) If we wish to increase the SQNR of a quantizer by doubling its sampling fre quency, what is the most efficient way to do it? Should we choose a linear muUibil A/D converter or an oversampling one? 9.10 Consider the first-order SDM model shown in Fig. 9.15. (a) Show that the quantization noise power in the signal band {- B . B ) is given by
(b ) Using a two-term Taylor series expansion of the sine function and assuming that
Fs > > B, show that
9.11 Consider the second-order SDM model shown in Fig. P 9 .ll. (a) D eterm ine the signal and noise system functions H,(z) and H(z ), respectively. (b ) Plot the magnitude response for the noise system function and compare it with the one for the first-order SDM, Can you explain the 6-dB difference from these curves? (c) Show that the in-band quantization noise power a 2 is given approximately by
which implies a 15-dB increase for every doubling of the sampling frequency.
778
Chap. 9
e(n)
Figure P 9 .ll
9.12 Figure P9.12 illustrates the basic idea for a lookup table based sinusoidal signal gen erator. The samples of one period of the signal
x ( n ) = cos ( n
'2n
n = 0 ,1 ........N - l
are stored in memory. A digital sinusoidal signal is generated by stepping through the table and wrapping around at the end when the angle exceeds 2x. This can be done by using modulo-N addressing (i.e.. using a circular buffer). Samples of x ( n ) are feeding the ideal D /A converter every T seconds. (a) Show that by changing F, we can adjust the frequency F{, of the resulting analog sinusoid. (b) Suppose now that Fs = l / T is fixed. How many distinct analog sinusoids can be generated using the given lookup table? Explain.
xa(t) = cos 2 ttF0t
Figure P9.12
9.13 Suppose that we represent an analog bandpass filter by the frequency response H( F) = C( F - Fc) + C*{F - Fc) where C ( f ) is the frequency response of an equivalent lowpass filter, as shown in Fig. P9.13.
Chap. 9
Problems
779
(a) Show that the impulse response c(r) of the equivalent lowpass filter is related to the impulse response h(t) of the bandpass filter as follows: h(t) = 2Re[c(t)ej2jrF{'] (b) Suppose that the bandpass system with frequency response H ( F ) is excited by a bandpass signal of the form x(t) = Re[u(f)e-'2rF,''3 where u(r) is the equivalent lowpass signal. Show that the filter output may be expressed as y(r) = Re[r(/)e-'2)rfV'] where
(Hint: Use the frequency domain to prove this result.)

C(F)
-8
Figure P9.13
9.14* Consider the sinusoidal signal generator in Fig. P9.14, where both the stored sinusoidal data
and the sampling frequency Fs = 1/ T are fixed. An engineer wishing to produce a sinusoid with period 2N suggests that we use either zero-order or first-order (linear) interpolation to double the num ber of samples per period in the original sinusoid as illustrated in Fig. P9.14(a). (a) Determ ine the signal sequences v(n) generated using zero-order interpolation and linear interpolation and then com pute the total harm onic distortion (TH D ) in each case for N = 32, 64. 128. (b) R epeat part (a) assuming that all sample values are quantized to 8 bits. (c) Show that the interpolated signal sequences y(n) can be obtained by the system shown in Fig. P9.14(b). The first module inserts one zero sample between suc cessive samples of x ( n ). D eterm ine the system H(z) and sketch its magnitude response for the zero-order interpolation and for the linear interpolation cases. Can you explain the difference in perform ance in term s of the frequency response functions?
780
Interpolated values
Chap. 9
Interpolated values
(a>
Figure P9.13
(a)
Insert zeros
jc/n)
U(r\ n \i)
Rpire P9.13 (b)
(d) Determine and sketch the spectra of the resulting sinusoids in each case both analytically [using the results in part (c)] and evaluating the DFT of the resulting signals. (e) Sketch the spectra of x, (n) and j(n), if x(n) has the spectrum shown in Fig. P9.14(c) for both zero-order and linear interpolation. Can you suggest a better choice for
Chap. 9
Problems
781
XM
0
3
(c)
n
3 Figure P9.13 (c)
9.15 Let xa(t) be a time-limited signal: that is, xa(t) = 0 for |/| > t , with Fourier transform X ( F ) . The function X ( F ) is sampled with sampling interval &F = 1/T<. (a) Show that the function C C
can be expressed as a Fourier series with coefficients 1

ct = y X J k & F )
(b ) Show that X ( F) can be recovered from the samples X a(kSF ), - o c < k < oc if
T, >
2t.
(c) Show lhal if Tx < 2r. there is time-domain aliasing that prevents exact recon
struction of X U (F). (d) Show that if 7, > 2 r, perfect reconstruction of X a(F) from the samples X(k5F) is possible using the inlerpolation formula O O
XJF) = ^ 2 X a (k&F)
sin { ( j i / S F ) ( . F ~ k S F ) ]
(n /8 F )(F - k S F )
Multirate Digital Signal Processing
10
In m an y p ractical a p p lic a tio n s o f d igital signal pro cessin g , o n e is faced w ith th e p ro b le m o f c h an g in g th e sa m p lin g ra te o f a signal, e ith e r in c re a sin g it o r d e creasin g it by so m e a m o u n t. F o r e x a m p le , in te le c o m m u n ic a tio n sy stem s th a t tra n sm it an d receiv e d ifferen t ty p e s o f signals (e.g., te le ty p e , facsim ile, sp e e c h , v id eo , e tc .), th e re is a re q u ire m e n t to p ro cess th e v ario u s signals at d iffe re n t ra te s c o m m e n s u ra te w ith th e c o rre sp o n d in g b a n d w id th s o f th e signals. T h e p ro cess o f c o n v e rtin g a signal fro m a given ra te to a d iffe re n t ra te is called samp l i n g rate conversi on. In tu rn , sy stem s th a t em p lo y m u ltip le sa m p lin g ra te s in th e p ro cessin g o f d ig ital signals a re called multirale digital si gnal proc es s i ng systems. S am p lin g ra te c o n v ersio n o f a d igital signal can b e a c c o m p lish ed in o n e of tw o g en eral m e th o d s. O n e m e th o d is to p ass th e digital signal th ro u g h a D /A co n v e rte r, filter it if n ecessa ry , a n d th e n to re sa m p le th e re su ltin g a n a lo g signal at th e d esired ra te (i.e., to p ass th e a n a lo g signal th ro u g h an A /D c o n v e rte r). T h e se co n d m e th o d is to p e rfo rm th e sa m p lin g ra te c o n v ersio n e n tire ly in th e d igital d o m ain . O n e a p p a re n t a d v a n ta g e o f th e first m e th o d is th a t th e new sa m p lin g ra te can b e a rb itra rily se le c te d a n d n e e d n o t h av e a n y special re la tio n s h ip to th e old sa m p lin g ra te . A m a jo r d isa d v a n ta g e , h o w ev er, is th e signal d is to rtio n , in tro d u c e d by th e D /A c o n v e rte r in th e signal re c o n stru c tio n , a n d by th e q u a n tiz a tio n effects in th e A /D co n v e rsio n . S am p lin g ra te c o n v ersio n p e rfo rm e d in th e d igital d o m ain av o id s this m a jo r d isa d v a n ta g e . In th is c h a p te r w e d e sc rib e sa m p lin g r a te co n v ersio n a n d m u ltira te signal p ro cessin g in th e d ig ital d o m a in . F irs t w e d esc rib e sa m p lin g r a te c o n v ersio n by a ra tio n a l fa c to r a n d p re s e n t se v e ra l m e th o d s fo r im p le m e n tin g th e ra te c o n v e rte r, in clu d in g sin g le-stag e a n d m u ltista g e im p le m e n ta tio n s. T h e n , w e d e sc rib e a m e th o d fo r sa m p lin g ra te co n v e rsio n by an a rb itra ry fa c to r a n d discuss its im p le m e n ta tio n . F in ally , w e p re s e n t se v era l a p p lic a tio n s o f sa m p lin g r a te c o n v e rsio n in m u ltira te signal p ro cessin g sy stem s, w hich in c lu d e th e im p le m e n ta tio n o f n a rro w b a n d fil ters, d igital filter b a n k s, a n d q u a d r a tu re m irro r filters. W e a lso discuss th e use o f
782
Sec. 10.1
Introduction
783
q u a d r a tu re m irro r filters in su b b a n d coding, tra n sm u ltip le x e rs, a n d finally o v ersam p lin g A /D an d D /A co n v e rte rs.
10.1 INTRODUCTION
T h e p ro c e ss o f sa m p lin g ra te co n v ersio n in th e d ig ital d o m a in can b e v iew ed as a lin e a r filterin g o p e ra tio n , as illu stra te d in Fig. 10.1(a). T h e in p u t signal jr(n) is c h a ra c te riz e d b y th e sa m p lin g ra te Fx = \ / T x a n d th e o u tp u t signal y ( m ) is c h a ra c te riz e d by th e sa m p lin g ra te F v = 1/7V, w h e re Tx a n d Ty are th e c o r re s p o n d in g sam p lin g in terv als. In th e m ain p a r t o f o u r tr e a tm e n t, th e ra tio Fy/ F x is co n s tra in e d to b e ra tio n a l,
=L Fx D
w h ere D a n d / a re relativ ely p rim e in teg ers. W e shall show th a t th e lin e a r filter is c h a ra c te riz e d by a tim e -v a ria n t im pulse re sp o n se , d e n o te d as h ( n , m ) . H e n c e th e in p u t x ( n ) an d th e o u tp u t v( m) a re re la te d by th e c o n v o lu tio n su m m a tio n fo r tim e -v a ria n t system s. T h e sam p lin g ra te co n v ersio n p ro cess can also b e u n d e rsto o d fro m th e p o in t o f view o f d igital re sam p lin g o f th e sam e an a lo g signal. L et x ( i ) b e th e a n a log signal th a t is sa m p le d at th e first ra te Fx to g e n e ra te x ( n) . T h e goal o f ra te co n v ersio n is to o b ta in a n o th e r se q u e n c e y(m ) d ire c tly from jr(n). w hich is eq u a l to th e sa m p le d v alu es o f x ( t ) at a se co n d ra te Fy. A s is d e p ic te d in Fig. 10.1(b), y(m ) is a tim e-sh ifted version o f * (n ). S uch a tim e shift can be x(n) Tx (a) Linear filter k(n, m) y{i") Fx mT ky
(b)
Figure 10.1 Sampling rate conversion viewed as a linear filtering process.
7B4
Chap. 10
r e a liz e d by u sin g a lin e a r filter th a t h as a flat m a g n itu d e re sp o n s e a n d a lin e a r p h a s e re sp o n s e (i.e., it h a s a fre q u e n c y re sp o n s e o f e~>u > r, w h e re r, is th e tim e d e la y g e n e ra te d by th e filter). If th e tw o sa m p lin g r a te s a r e n o t e q u a l, th e r e q u ire d a m o u n t o f tim e sh iftin g will vary fro m sa m p le to sa m p le , a s show n in Fig. 10.1(b). T h u s th e ra te c o n v e rte r can b e im p le m e n te d u sin g a se t o f lin ear filters th a t h av e th e sa m e flat m a g n itu d e re sp o n s e b u t g e n e r a te d iffe re n t tim e d elays. B e fo re co n sid e rin g th e g e n e ra l case o f sa m p lin g r a te co n v e rsio n , w e shall c o n s id e r tw o sp ecial cases. O n e is th e case o f sa m p lin g ra te r e d u c tio n by an in te g e r fa c to r D, a n d th e se co n d is th e case o f a sa m p lin g ra te in c re a se b y an in te g e r facto r I . T h e p ro c e ss o f re d u c in g th e sa m p lin g ra te b y a fa c to r D (d o w n sa m p lin g by D) is called deci mati on. T h e p ro c e ss o f in c re a sin g th e sa m p lin g r a te by an in te g e r fa c to r I (u p sam p lin g b y I ) is ca lle d interpolation.
10.2 DECIMATION BY A FACTOR D

L e t us assu m e th a t th e signal x ( n ) w ith sp e c tru m X(co) is t o be d o w n sam p led by a n in te g e r fa c to r D. T h e sp e c tru m X(a>) is assu m ed to b e n o n z e ro in th e freq u e n cy in te rv a l 0 < )co\ < n o r, e q u iv a le n tly , \F\ < Fx /2. W e k n o w th a t if w e re d u c e th e sam p lin g ra te sim p ly by se lectin g ev ery D th v alu e o f x ( n ) , th e resu ltin g sig n al will b e an alia se d v ersio n o f jr(n), w ith a fo ld in g f re q u e n c y o f FX/ 2 D . T o av o id aliasing, w e m u st first re d u c e th e b a n d w id th o f x ( n ) to F max = FX/ 2 D or, eq u iv alen tly , to a>max = n j D . T h e n we m ay d o w n sam p le by D a n d th u s avoid aliasing. "Hie d e c im a tio n p ro c e ss is illu stra te d in Fig. 10.2. T h e in p u t s e q u e n c e x{n) is p a s se d th ro u g h a lo w p ass filter, c h a ra c te riz e d by th e im p u lse re s p o n s e h ( n ) an d a fre q u e n c y re sp o n se H D(a>), w hich id eally satisfies th e c o n d itio n H d (o>) = ( J 10, M * * ! D o th erw ise (10.2.1)
T h u s th e filter e lim in a te s th e sp e c tru m o f X (a>) in th e ra n g e n / D < w < n . O f c o u rse, th e im p licatio n is th a t o n ly th e fre q u e n c y c o m p o n e n ts o f x i n ) in th e ran g e M < tc/ D a re o f in te re st in f u rth e r p ro cessin g o f th e signal. T h e o u tp u t o f th e filter is a s e q u e n c e v(n) given as
OC
v(n) = Y ^ ^ ) x ( n - k) t=o X(fl)

1 \ i
(10.2.2)
h(n)
i<n)
1 !
Downsampler W
yim)
Fx
Figure 103.
D ecimation by a factor D.
Sec. 10.2
Decimation by a Factor
785
w hich is th e n d o w n sa m p le d by th e fa c to r D to p ro d u c e y ( m) . T h u s y(m) = v ( m D ) (10.2.3) = h ( k ) x { m D k)
A lth o u g h th e filtering o p e ra tio n o n x ( n ) is iin e a r an d tim e in v a ria n t, th e d o w n sam p lin g o p e r a tio n in c o m b in a tio n w ith th e filterin g resu lts in a tim e -v a ria n t sy stem . T h is is easily verified. G iv en th e fact th a t x (n) p ro d u c e s y ( m) , w e n o te th a t x ( n no) d o e s n o t im ply y( n no) u n less no is a m u ltip le o f D. C o n seq u en tly , th e o v e ra ll lin e a r o p e ra tio n (lin e a r filtering fo llo w ed by d o w n sam p lin g ) on x ( n ) is n o t tim e in v arian t. T h e fre q u e n c y -d o m a in c h a racteristics o f th e o u tp u t se q u e n c e y ( m ) can be o b ta in e d b y re la tin g th e sp e c tru m o f y (m ) to th e s p e c tru m o f th e in p u t se q u e n c e x ( n ) . F irs t, it is c o n v e n ie n t to define a se q u e n c e v(n) as n = 0, D , 2 D , . . . o th erw ise (10.2.4)
C learly , v(n) can b e view ed as a se q u e n c e o b ta in e d by m u ltiplying v(n) w ith a p e rio d ic tr a in o f im p u lses p (n ), w ith p e rio d D, as illu stra te d in Fig. 10.3. T h e d isc re te F o u rie r se ries re p re s e n ta tio n o f p( n) is (10.2.5) H ence v(n) = v ( n ) p ( n ) an d y( m) = v (mD) = v ( m D ) p ( m D ) v( mD) (10.2.7)
( 10.2 .6 )
n
-6 -5 -4 -3 - 2 (-1 0 1 2 3 4 5
Pin)
n
-6
Figure 1 0 3
-3
Multiplication o f v(n) with a periodic impulse train pin) with period
D = 3.
786
Chap. 10
N o w th e z -tra n sfo rm o f th e o u tp u t se q u e n c e y ( m ) is y(z) = O O

m= oc
oo J2
ms=oc
v( r nD) z ~m
(10.2.8)
Y(z) =
00 ]T
m= oc
v ( m ) z - m/D
w h ere th e last ste p follow s fro m th e fact th a t v ( m) = 0, e x c e p t a t m u ltip le s o f D. B y m ak in g use o f th e re la tio n s in (10.2.5) a n d (10.2.6) in (10.2.8), w e o b ta in Y( z ) =
m= cc j D 1 oc
__
u ~ l
ejlxmk/D | -m/D
<t=0
j2nk/D 1/D^-m (10.2.9)
*=0 m=-oo
"* -0 = 7J H D ( e - i2* t /Dz l/D) X ( e - W > z l / D) *=0 w h ere th e last ste p fo llow s fro m th e fact th a t V (z) = H q { z ) X ( z ). B y e v a lu a tin g y ( 2) in th e u n it circle, w e o b ta in th e s p e c tru m o f th e o u tp u t signal y{m) . S ince th e ra te o f y ( m ) is Fy = 1/TV, th e fre q u e n c y v a ria b le , w hich w e d e n o te as w v, is in ra d ia n s a n d is re la tiv e to th e sa m p lin g r a te F v, ti)y = Fy S ince th e sa m p lin g ra te s a re r e la te d by th e ex p ressio n Fy = ^ it fo llow s th a t th e fre q u e n c y v a ria b le s wy a n d 2tr F cox = - = 2 n F T x Fx a re re la te d by a>v = D ojx (10.2.13) (10.2.12) (10.2.11) = 2 n F T s. (10.2.10)
T h u s, as e x p ected , th e fre q u e n c y ra n g e 0 < \ojx \ < n / D is stre tc h e d in to th e c o rre sp o n d in g fre q u e n c y ra n g e 0 < \a> y l < n by th e d o w n sam p lin g p ro cess.
Sec. 10.3
Interpolation by a Factor I
787
W e c o n clu d e th a t th e sp e c tru m Y (wv), w hich is o b ta in e d by e v a lu a tin g (10.2.9) o n th e u n it circle, can b e e x p ressed as (10.2.14) W ith a p ro p e rly d esig n e d fitter H D(a)), th e aliasin g is e lim in a te d a n d , c o n seq u en tly , all b u t th e first te rm in (10.2.14) vanish. H e n c e
(10.2.15)
fo r 0 < |w v| < tt. T h e s p e c tra fo r th e se q u en ces x i n ) , v{n), a n d y ( m) a re illu stra te d in Fig. 10.4.
10.3 INTERPOLATION BY A FACTOR I

A n in crease in th e sam p lin g rate by an in te g e r fa c to r o f I can be acco m p lish ed b y in te rp o la tin g I 1 new sam ples b etw e e n successive v alu es o f th e signal. T h e in te rp o la tio n p ro cess can b e acco m p lish ed in a v ariety o f w ays. W e sh all d escrib e a p ro cess th a t p re se rv e s th e sp e c tra l sh a p e o f th e signal se q u e n c e x ( n ) . L et u(m ) d e n o te a se q u e n c e w ith a ra te F v = I Fx , w hich is o b ta in e d fro m x i n ) by a d d in g / - 1 z e ro s b e tw e e n successive v alu es o f x{n) . T h u s (10.3.1) a n d its sa m p lin g ra te is id en tical to th e ra te o f y{m) . tra n sfo rm V( z ) = Y ,
m = oc oc
T h is se q u e n c e has a z-
v^ z ~ m
(10.3.2) = Xiz') T h e c o rre sp o n d in g sp e c tru m of v ( m) is o b ta in e d by e v a lu a tin g (10.3.2) o n th e u nit circle. T h u s Vicoy) = X { w y I) (10.3.3)
w h e re <uv d e n o te s th e fre q u e n c y v a ria b le re lativ e to th e n ew sa m p lin g ra te F , (i.e., o)y = 2 n F / F y ) . N o w th e re la tio n sh ip b e tw e e n sa m p lin g ra te s is F y = 1 Fx an d h e n c e , th e fre q u e n c y v a ria b le s (oI a n d toy a re re la te d a c c o rd in g to th e fo rm u la
768
Chap. 10
l* K )l
I Via,)I
Figure 10.4 Spectra of signals in the decimation of x( n ) by a factor D.
T h e s p e c tra X(a>x) a n d V(a>y) a re illu stra te d in Fig. 10.5. W e o b se rv e th a t the sa m p lin g ra te in c re a se , o b ta in e d by th e a d d itio n o f / 1 z e ro sa m p le s b etw e e n successiv e v alu es o f x ( n) , re su lts in a sig n a l w h o se sp e c tru m V(<wv) is an /-fo ld p e rio d ic re p e titio n o f th e in p u t signal sp e c tru m X S ince o n ly th e fre q u e n c y c o m p o n e n ts o f x (n) in th e ra n g e 0 < coy < n j l a re u n iq u e , th e im ag es o f X ( oj) ab o v e coy = n f l sh o u ld b e r e je c te d b y passing th e se q u e n c e v ( m ) th ro u g h a low pass filter w ith fre q u e n c y r e s p o n s e Hi(a>y) th a t
Sec. 10.3
Interpolation by a Factor /
789
l* K )l
Figure 10.5 Spectra of .run and run where V(a>v) - X(wvl).
id eally has th e ch a ra c te ristic Hl((i}y) = C. 0. 0 < |a>v| < n / I o th erw ise (10.3.5)
w h e re C is a scale fa c to r re q u ire d to p ro p e rly n o rm a liz e th e o u tp u t se q u e n c e v(/7?). C o n s e q u e n tly , th e o u tp u t sp e c tru m is (10.3.6) T h e scale fa c to r C is se le c te d so th a t th e o u tp u t y ( m ) = jc ( m/ 1 ) fo r m = 0, / , + 2 1 .........F o r m a th e m a tic a l co n v e n ie n c e , w e select th e p o in t m = 0. T h u s
y( 0 )=
2~
/
(10.3.7) X {(I)vI)d(Dx
.J r ~ _ n
c i r v(0) = / I 2n C = -x (0 )
S in ce a> y = tox / I , (10.3.7) can be e x p re ss e d as X ( w x) dwx (10.3.8)
T h e re fo re , C = / is th e d esired n o rm a liz a tio n facto r. F in ally , w e in d ic a te th a t th e o u tp u t se q u e n c e y ( m ) c a n b e ex p re sse d as a c o n v o lu tio n o f th e s e q u e n c e v(n) w ith th e u n it sa m p le re sp o n s e h( n) o f th e low pass
790
Chap. 10
filter. T h u s
OC
y( m) =
Y 2 h( m k) v( k)
k= oc
(10.3.9)
S in ce v(k) = 0 e x cep t at m u ltip le s o f / , w h e re v ( k l ) = x ( k) , (10.3.9) b ec o m e s y (m )= O O Y2 h{m -kl)x(k) (10.3.10)
10.4 SAMPLING RATE CONVERSION BY A RATIONAL FACTOR U D

H a v in g discussed th e sp ecial cases o f d e c im a tio n (d o w n sa m p lin g by a fa c to r D ) a n d in te rp o la tio n (u p sa m p lin g by a fa c to r / ) , w e no w c o n s id e r th e g e n e ra l case o f sam p lin g ra te co n v e rsio n by a ra tio n a l fa c to r I / D . B asically, w e can ach iev e th is sa m p lin g ra te c o n v ersio n by first p e rfo rm in g in te rp o la tio n by th e fa c to r I a n d th e n d e c im atin g th e o u tp u t o f th e in te r p o la to r by th e fa c to r D. In o th e r w o rd s, a sa m p lin g ra te co n v ersio n by th e ra tio n a l fa c to r I ( D is ac c o m p lish e d by cascad in g an in te rp o la to r w ith a d e c im a to r, as illu stra te d in Fig. 10.6. W e em p h asize th a t th e im p o rta n c e o f p e rfo rm in g th e in te rp o la tio n first an d th e d e c im atio n se co n d , is to p re se rv e th e d e sire d sp e c tra l c h a ra c te ristic s o f x(n). F u rth e rm o re , w ith th e cascad e co n fig u ra tio n illu stra te d in Fig. 10.6, th e tw o filters w ith im p u lse re sp o n se {/(/)} a n d {hd (l)\ a re o p e ra te d at th e sa m e r a te , n am ely I F X an d h e n c e can b e c o m b in e d in to a single low pass filter w ith im p u lse re sp o n s e h(l) as illu stra te d in Fig. 10.7. T h e fre q u e n c y re sp o n s e H( t ov) of th e co m b in e d filter m u st in c o rp o ra te th e filtering o p e ra tio n s fo r b o th in te rp o la tio n an d d ecim atio n , a n d h e n c e it sh o u ld id eally p o ssess th e fre q u e n c y re sp o n s e c h a ra c te ristic 0 < \&v\ < m m ( n / D , n / 1 ) o th e rw ise w h e re cut. = 2 x F / F v = 2 n F / I Fx = a)x/ / . (10.4.1)
Rate = / F t
Rate = Fx = Fv D
Figure 10.6
M ethod for sampling rate conversion by a factor f/D .
Sec. 10.4
Sampling Rate Conversion by a Rational Factor
IID
791
x(n) Rate = F,
Upsampler T/
u(t)
Lowpass filter h(D
w([)
Down sampler 4D
v(m)
Rate =
Rate = IFX = Fv Figure 10.7 Method for sampling rate conversion by a factor 1/D.
In th e tim e d o m a in , th e o u tp u t o f th e u p sa m p le r is th e se q u e n c e ( ///) , u (0 = ( j ( [ 0. / = 0, / , 2 / , . . . , o th erw ise (10.4.2)
a n d th e o u tp u t o f th e lin e a r tim e -in v a ria n t filter is O O w( l ) = ^ h(! k) v( k) k=-oc oc

=
(10.4.3)
h(l k l ) x ( k )
F inally, th e o u tp u t o f th e sa m p lin g ra te c o n v e rte r is th e se q u e n c e (y(m )}, w hich is o b ta in e d b y d o w n sa m p lin g th e se q u e n c e {>(/)} by a fa c to r o f D. T h u s v(m) = w{mD) = ~ 2_j h ( m D k l ) x ( k ) k oc (10.4.4)
It is illu m in a tin g to ex p ress (10.4.4) in a d iffe re n t fo rm by m a k in g a ch an g e in v ariab le. L e t k =
mD
(10.4.5)
w h e re th e n o ta tio n [ r j d e n o te s th e la rg e st in te g e r c o n ta in e d in r. W ith th is ch an g e in v ariab le, (10.4.4) b ec o m e s y( m) = W e n o te th a t mD mD I mD m o d u lo I Y

j
oc
h ( m D ----V - ^ -
V- ^ _
~ 1 /
(10.4.6)
= ( mD) ! C o n s e q u e n tly , (10.4.6) can b e e x p re sse d as
0 0
y(m )Y
j
/
h(~ n I + (m D ) i ) x (
VL
L)
j J
\
~ n)
/
(10.4.7)
n=- oc
792
Chap. 10
It is a p p a r e n t fro m th is fo rm th a t th e o u tp u t y ( m ) is o b ta in e d by p assin g th e in p u t se q u e n c e x ( n ) th ro u g h a tim e -v a ria n t filter w ith im p u lse re sp o n s e g(n,m) h(n l + (mD)i) oo< m ,n< oo (10.4.8)
w h e re h(k) is th e im p u lse re sp o n s e o f th e tim e -in v a ria n t lo w p ass filter o p e ra tin g at th e sam p lin g r a te 1FX. W e fu rth e r o b se rv e , th a t fo r any in te g e r k, g ( n , m -I- k l ) = h ( n l -I- ( m D -I- k D I ) , ) = h(nl + (mD),) = g(n,m) H e n c e g(n, m) is p e rio d ic in th e v a ria b le m w ith p e rio d I. T h e fre q u e n c y -d o m a in re la tio n sh ip s can b e o b ta in e d by c o m b in in g th e resu lts o f th e in te rp o la tio n a n d d e c im a tio n p ro cesses. T h u s th e s p e c tru m a t th e o u tp u t of th e lin e a r filter w ith im p u lse re sp o n s e h(l) is V (a) v) = H( ai v)X(a>vl ) I X ( w vI ) , 0, 0 < IojJ < m in (7 r /0 , n / I ) o th e rw ise (10.4.10) (10.4.9)
T h e sp e ctru m o f th e o u tp u t se q u e n c e y (m ), o b ta in e d by d e c im a tin g th e se q u e n c e v(n) by a fa c to r o f D , is
a o .4 .1 1 )
w h e re w v = D(ov. S in ce th e lin e a r filter p re v e n ts aliasin g as im p lie d by (10.4.10), th e sp e c tru m o f th e o u tp u t se q u e n c e given by (10.4.11) re d u c e s to
o th e rw ise
10.5 FILTER DESIGN AND IMPLEMENTATION FOR SAMPLING-RATE CONVERSION

A s in d ic a te d in th e discu ssion a b o v e , sa m p lin g ra te c o n v e rsio n b y a fa c to r I } D can b e a c h iev ed by first in creasin g th e sa m p lin g r a te b y 1, ac c o m p lish e d by in se rtin g 7 - 1 z e ro s b e tw e e n su ccessive v alu es o f th e in p u t signal x( n ) , fo llo w e d by lin e a r filterin g o f th e re su ltin g se q u e n c e to elim in a te th e u n w a n te d im a g e s o f X (co), an d finally, by d o w n sam p lin g th e filte re d signal by th e fa c to r D. I n th is se c tio n we c o n s id e r th e d esig n a n d im p le m e n ta tio n o f th e lin e a r filter.
Sec. 10.5
Filter Design and Implementation for Sampling-Rate Conversion
793
10.5.1 Direct-Form FIR Filter Structures

In p rin c ip le , th e sim p lest re a liz a tio n o f th e filter is th e d ire c t-fo rm F IR s tru c tu re w ith sy stem fu n ctio n
M -l
H{z ) = Y b ( k ] z ~k t =o
(10-5 -1)
w h e re {/j {/:)} is th e u n it sam ple re sp o n se o f th e F IR filter. T h e low pass filter c an b e d e sig n e d to h av e lin e a r p h a s e , a specified p a s sb a n d rip p le a n d sto p b a n d a tte n u a tio n . A n y o f th e sta n d a rd , w ell k n o w n F IR filter design te c h n iq u e s (e.g., w in d o w m e th o d , freq u e n cy sa m p lin g m e th o d ) can be u se d to c a rry o u t this design. T h u s w e w ill h av e th e filter p a ra m e te rs {/j ()}, w hich allo w us to im p le m e n t th e F IR filter d ire c tly as sh ow n in Fig. 10.8. A lth o u g h th e d ire c t-fo rm F IR filter re a liz a tio n illu stra te d in Fig. 10.8 is sim p le, it is also v ery in efficient. T h e inefficiency resu lts fro m th e fact th a t th e u p sa m p lin g p ro cess in tro d u c e s / 1 ze ro s b e tw e e n successive p o in ts o f th e in p u t sig n al. If I is larg e, m o st o f th e signal c o m p o n e n ts in th e F IR filter are zero. C o n se q u e n tly , m o st o f th e m u ltip lic atio n s an d a d d itio n s re su lt in zero s. F u rth e rm o re , th e d o w n sam p lin g p ro cess at th e o u tp u t o f th e filter im p lies th a t only on e o u t o f
Figure 10.8 factor I/D.
Direct-form realization of FIR filter in sampling rate conversion by
794
Mullirate Digital Signal Processing
Chap. 10
ev ery D o u tp u t sa m p le s is re q u ire d a t th e o u tp u t o f th e filter. C o n s e q u e n tly , only o n e o u t o f ev ery D p o ssib le v a lu e s a t th e o u tp u t o f th e filter sh o u ld be co m p u te d . T o d ev elo p a m o re efficien t filter s tru c tu re , let us b eg in w ith a d e c im a to r th a t re d u c e s th e sa m p lin g ra te by an in te g e r fa c to r D. F ro m o u r p re v io u s discussion, th e d e c im a to r is o b ta in e d by p assin g th e in p u t se q u e n c e x ( n ) th r o u g h an F IR filter a n d th e n d o w n sam p lin g th e filter o u tp u t by a fa c to r >, as illu stra te d in Fig. 10.9a. In th is co n fig u ratio n , th e filter is o p e ra tin g a t th e high sa m p lin g ra te Fx, w hile o n ly o n e o u t o f ev ery D o u tp u t sa m p le s is ac tu a lly n e e d e d . T h e logical so lu tio n to th is inefficiency p ro b le m is to e m b e d th e d o w n sam p lin g o p e ra tio n w ith in th e filter, as illu stra te d in th e filter re a liz a tio n given in Fig. 10.9b. In this filter stru c tu re , all th e m u ltip lic a tio n s an d a d d itio n s a re p e rfo rm e d a t th e lo w er sam p lin g ra te Fx / D. T h u s w e h av e ach ie v e d th e d e s ire d efficiency. A d d itio n a l re d u c tio n in c o m p u ta tio n can b e ach ie v e d by e x p lo itin g th e sy m m etry c h a ra c te ristic s of F ig u re 10.10 illu stra te s an efficient re a liz a tio n o f th e d e c im a to r in w hich th e F IR filter h as lin ear p h a s e , a n d h e n c e {/i{&)} is sym m etric. N ex t, let us co n sid e r th e efficien t im p le m e n ta tio n o f an in te rp o la to r, w hich is realized by first in se rtin g I l z e ro s b e tw e e n sa m p le s o f x ( n ) a n d th e n filtering th e re su ltin g se q u en ce. T h e d ire c t-fo rm re a liz a tio n is illu stra te d in Fig. 10.11. T h e m a jo r p ro b lem w ith th is s tru c tu re is th a t th e filter c o m p u ta tio n s a re p e rfo rm e d at th e hig h sam p lin g ra te 1FX. T h e d e s ire d sim p lificatio n is a c h ie v e d by first using th e tra n sp o s e d fo rm o f th e F IR filter, as illu stra te d in Fig. 10.12a, a n d th e n e m b ed d in g th e u p sa m p le r w ithin th e filter, as show n in Fig. 10.12b. T h u s, all th e filter m u ltip li c a tio n s a re p e rfo rm e d at th e low ra te F x, w hile th e u p sa m p lin g p ro c e ss in tro d u c e s 1 1 ze ro s in ea c h o f th e filte r b ra n c h e s o f th e stru c tu re show n in Fig. 10.12b. T h e r e a d e r can easily verify th a t th e tw o filteT s tru c tu re s in Fig. 10.12 a re e q u iv a le n t. It is in te re stin g to n o te th a t th e s tru c tu re o f th e in te rp o la to r, sh o w n in Fig. 10.12b, can b e obtained by tra n sp o s in g th e stru c tu re o f th e d e c im a to r sh o w n in Fig. 10.9. W e o b se rv e th a t th e tra n s p o s e o f a d e c im a to r is a n in te rp o la to r, an d vice v ersa. T h ese re la tio n sh ip s a re illu stra te d in Fig. 10.13, w h e re (b ) is obtained by tra n sp o sin g (a ) a n d (d) is obtained by tra n sp o s in g (c). C o n s e q u e n tly , a d e c i m a to r is th e d u al o f an in te rp o la to r, a n d vice v ersa. F ro m th e s e re la tio n sh ip s, it fo llow s th a t th e re is an in te r p o la to r w h o se s tru c tu re is th e d u a l o f th e decimatoT sh o w n in Fig, 10.10, w hich ex p lo its th e sy m m etry in h(n).
10.5.2 Polyphase Filter Structures

T h e c o m p u ta tio n a l efficiency o f th e filter s tru c tu re show n in Fig. 10.12 can also b e ach iev ed by re d u c in g th e larg e F IR filter o f le n g th M in to a se t o f sm a ller filters o f len g th K = M / 1 , w h ere M is se le c te d to b e a m u ltip le o f I . T o d e m o n s tr a te th is p o in t, le t u s c o n s id e r th e in te r p o la to r given in F ig. 10.11. S ince th e u p sa m p lin g p ro cess in se rts 1 1 z e ro s b e tw e e n successive v a lu e s o f x( n) , o n ly K o u t o f th e M in p u t v alu es sto re d in th e F IR filter a t an y o n e tim e a re n o n z e ro . A t o n e tim e in sta n t, th e s e n o n z e ro v a lu e s co in cid e a n d a re m u ltip lie d by th e fil te r co efficien ts h{0), h ( I ) , h ( 2 l ) , . . . , h ( M I ) . In th e fo llo w in g tim e in sta n t, th e
Sec. 10.5
795
(a)
Figure 10.9
Decimation by a factor D.
Chap. 10
Sec. 10.5 J T (n )
Rlter Design and Implementation for Sampling-Rate Conversion
797
h(M - 2)
MM-
1)
Figure 10.11 Direct-form realization of FIR filter in interpolation by a factor /,
n o n z e ro v a lu e s o f th e in p u t se q u e n c e co in cid e a n d a re m u ltip lie d by th e filter co efficien ts /?(!). h ( I + 1), h ( 2 I + 1)___ h ( M - 7 + 1), a n d so o n . T h is o b se rv a tio n le a d s us to d efin e a se t o f sm a ller filters, called p o ly p h a s e filters, w ith u n it sa m p le re sp o n se s p t ( n ) = h( k + n ! ) k = 0 ,1 ,..., 7 1 (10.5.2) n = 0, ~ 1 w h e re K = A //7 is an in teg er. F ro m th is d iscussion it follow s th a t th e se t o f 7 p o ly p h a se filters can b e a rra n g e d as a p a ra lle l re a liz a tio n , a n d th e o u tp u t o f e a c h filter can be se le c te d b y a c o m m u ta to r as illu stra te d in Fig. 10.14. T h e ro ta tio n o f th e c o m m u ta to r is in th e co u n te rc lo c k w ise d ire c tio n b e g in n in g w ith th e p o in t a t m ~ 0. T h u s, th e p o ly p h a s e filters p e rfo rm th e c o m p u ta tio n s a t th e low sa m p lin g ra te Fx , a n d th e r a te c o n v e rsio n resu lts fro m th e fact th a t 7 o u tp u t sa m p le s a re g e n e ra te d , o n e fro m each o f th e filters, fo r each in p u t sa m p le . T h e d e c o m p o s itio n o f {/r{)} in to th e se t o f 7 su b filters w ith im p u lse re sp o n se Pk(n), k = 0, 1 , . . . , 7 1, is c o n siste n t w ith o u r p re v io u s o b s e rv a tio n th a t th e in p u t sig n a l w as b e in g filte re d by a p erio d ically tim e -v a ria n t lin e a r filter w ith im p u lse re sp o n se g ( n , m ) = h ( n l + ( mD) [ ) (10.5.3) w h e re D = 1 in th e c a se o f th e in te rp o la to r. W e n o te d p rev io u sly th a t g( n, m) v a rie s p e rio d ic a lly w ith p e rio d 7. C o n se q u e n tly , a d iffe re n t se t o f coefficients a re u se d to g e n e ra te th e se t o f 7 o u tp u t sa m p le s y ( m ) , m = 0, 1 , . . . , I - 1. A d d itio n a l in sig h t can b e g ain e d a b o u t th e c h a ra c te ristic s o f th e se t o f p o ly p h a s e su b filte rs b y n o tin g th a t pk(n) is o b ta in e d fro m h( n ) by d e c im a tio n w ith a f a c to r 7. C o n s e q u e n tly , if th e o rig in a l filter fre q u e n c y re s p o n s e H {co) is flat o v e r th e ra n g e 0 < M < c o/ I , e a c h o f th e p o ly p h a se su b filters p o ssesses a re lativ ely flat
(b)
Figure 10.12 Efficient realization of an interpolator.
Sec. 10.5
799
(a)
< b )
(O (d ) F ig u re 10.13 D u ality relatio n sh ip s obtain edth ro u g htran sp o sitio n .
Pi- i< ") R a te=F x

Figure 10.14
R a te - Fr - IFX R a te=F, In terp o latio nb yu se of p o ly p h ase filters.
re sp o n s e o v e r th e ra n g e 0 < |a>| < n (i.e., th e p o ly p h a s e su b filters a re b asically all-p ass filters a n d d iffer p rim a rily in th e ir p h a se c h a ra c te ristic s). T h is ex p lain s th e re a so n fo r th e te rm p o ly p h a s e in d e sc rib in g .th e s e filters. T h e p o ly p h a se filter can also b e view ed as a se t o f I su b filters c o n n e c te d to a co m m o n d e la y line. Id eally, th e fcth su b filter will g e n e ra te a fo rw a rd tim e sh ift o f (.k / I ) T x, fo r k = 0, 1. 2 ........./ 1, re la tiv e to th e z e ro th subfilter. T h e re fo re , if th e z e ro th filter g e n e ra te s z e ro d elay , th e fre q u e n c y re sp o n s e o f th e Jtth su b filter is
pk((o) = e)wk/I
A tim e sh ift o f an in te g e r n u m b e r o f in p u t sa m p lin g in te rv a ls (e.g., ITX) can b e g e n e ra te d by sh iftin g th e in p u t d a ta in th e d e la y line by I sa m p le s an d using th e sa m e su b filters. B y c o m b in in g th e s e tw o m e th o d s, w e c a n g e n e ra te an o u tp u t th a t is sh ifted fo rw a rd b y an a m o u n t (/ + i / I ) T x re la tiv e to th e p re v io u s o u tp u t. B y tra n sp o s in g th e in te rp o la to r stru c tu re in Fig. 10.14, w e o b ta in a c o m m u ta to r stru c tu re fo r a d e c im a to r b a s e d o n th e p a ra lle l b a n k o f p o ly p h a se filters, as illu stra te d in Fig. 10.15. T h e u n it sa m p le re sp o n s e s o f th e p o ly p h ase filters a re
800
Chap. 10
Figure 10.15
Decimation by use of polyphase filters.
n o w d efin ed as Pk{n) = h( k + n D ) k = 0, 1____ O - l (10.5.4) n = 0 , 1 .........K - l w h e re K = M / D is an in te g e r w h en M is se le c te d to be a m u ltip le o f D. T h e c o m m u ta to r ro ta te s in a c o u n terclo ck w ise d ire c tio n sta rtin g w ith th e filter po(n) at
0.
A lth o u g h th e tw o c o m m u ta to r s tru c tu re s fo r th e in te r p o la to r a n d th e d e c i m a to r ju st d escrib ed r o ta te in a c o u n te rc lo c k w ise d ire c tio n , it is also p o ssib le to d e riv e an e q u iv a le n t p a ir o f c o m m u ta to r s tru c tu re s hav in g a clockw ise ro ta tio n . In th is a lte rn a tiv e fo rm u la tio n , th e se ts o f p o ly p h a se filters a re d efin ed to h ave im p u lse resp o n ses
P k ( n ) = h ( n l k)
k = 0 ,1 ,...,/- 1 * = 0 , 1 ......... D - 1
(10.5.5) (10.5.6)
Pk( n ) = h { n D - k )
f o r th e in te rp o la to r a n d d e c im a to r, resp ectiv ely .
10.5.3 Time-Variant Filter Structures

H av in g d esc rib e d th e filter im p le m e n ta tio n fo r a d e c im a to r a n d an in te rp o la to r, let us no w c o n sid e r th e g e n e ra l p ro b le m o f sa m p lin g ra te c o n v e rsio n by th e fa c to r I / D . In th e g e n e ra l case o f sa m p lin g r a te c o n v e rsio n by a fa c to r I / D , th e filtering can b e acco m p lish ed by m e a n s o f th e lin e a r tim e -v a ria n t filte r d e s c rib e d by th e re sp o n se fu n ctio n g( n, m) = h ( n l (m D )/) (10.5.7) w h ere h(n) is th e im p u lse re sp o n s e o f th e low -pass F IR filter, w hich id eally , has th e fre q u e n c y re sp o n s e specified by (10.4.1). F o r c o n v e n ie n c e w e se lect th e length o f th e F I R filter {/i(n)} to a m u ltip le o f I (i.e., M = K I ) . A s a c o n s e q u e n c e , th e s e t o f coefficien ts {#(, m)} fo r e a c h m = 0 , 1, 2 , 1, c o n ta in s K elem en ts.
Sec. 10.5
801
S in ce g( n, m ) is also p e rio d ic w ith p e rio d I , a s d e m o n s tr a te d in (10.4.9), it follow s th a t th e o u tp u t y { m) can b e e x p re ss e d as
y(m ) = X I#
n 0
- [_ y j l ) x
\L J
-n )
/
(10.5.8)
C o n c e p tu a lly , w e can th in k o f p e rfo rm in g th e c o m p u ta tio n s specified by (10.5.8) b y p ro cessin g blocks o f d a ta o f le n g th A T b y a se t o f K filter coefficients g ( n , m j_ m / l \ I ) , n = 0 , 1 , . . . , K 1. T h e re a re I such se ts o f coefficients, o n e se t fo r e a c h b lo ck o f I o u tp u t p o in ts o f y{m) . F o r e a c h b lock o f 1 o u tp u t p o in ts, th e r e is a c o rre sp o n d in g b lo ck o f D in p u t p o in ts o f x ( n ) th a t e n te r in th e c o m p u ta tio n . T h e b lo ck p ro cessin g a lg o rith m fo r co m p u tin g (10.5.8) can b e visualized as illu stra te d in Fig. 10.16. A b lo ck o f D in p u t sa m p le s is b u ffe re d a n d sh ifted in to a se co n d b u ffe r o f len g th K , o n e sa m p le a t a tim e. T h e sh iftin g fro m th e in p u t b u ffe r to th e se co n d b u ffe r o ccu rs a t a ra te o f o n e sa m p le ea c h tim e th e q u a n tity \ m D / I J in c re a se s by o n e. F o r ea c h o u tp u t sa m p le y( f ) , th e sa m p le s fro m th e se co n d b u ffe r a re m u ltip lie d by th e c o rre sp o n d in g se t o f filter coefficients g( n, I) fo r n = 0, 1 , ___ K 1, an d th e K p ro d u c ts a re a c c u m u la te d to give y (/), fo r 1 = 0,
x(n) Coefficient storage
Figure 10.16 cessing.
Efficient implementation of sampling-rate conversion by block pro
802
Chap. 10
1 , 1. T h u s th is c o m p u ta tio n p ro d u c e s / o u tp u ts. It is th e n re p e a te d fo r a new se t o f D in p u t sam p les, a n d so on. A n a lte rn a tiv e m e th o d fo r co m p u tin g th e o u tp u t o f th e sa m p le ra te c o n v e rte r, sp ecified by (10.5.8), is b y m e a n s o f an F IR filter stru c tu re w ith p e rio d ic a lly varying filter coefficients. Such a stru c tu re is illu stra te d in Fig. 10.17. T h e in p u t sam ples x{n) a re p a sse d in to a sh ift re g iste r th a t o p e ra te s a t th e sa m p lin g ra te Fx an d is of le n g th K = M / 1 , w h e re M is th e le n g th o f th e tim e -in v a ria n t F I R filter specified by th e fre q u e n c y re sp o n s e g iv en by (10.4.1). E a c h sta g e o f th e re g is te r is c o n n e c te d to a h o ld -a n d -sa m p le dev ice th a t se rv es to co u p le th e in p u t sa m p le ra te Fx to th e o u t p u t sa m p le r a te Fy = ( I / D ) F X. T h e sa m p le at th e in p u t to ea c h h o ld -a n d -sa m p le d ev ice is h eld u n til th e n e x t in p u t sa m p le arriv es a n d th e n is d isc a rd e d . T h e o u tp u t sa m p les o f th e h o ld -a n d -sa m p le d ev ice a re ta k e n a t tim es m D / l , m = 0, 1, 2 ........ W h e n b o th th e in p u t a n d o u tp u t sa m p lin g tim e s co in cid e (i.e.. w h en m D / I is an in te g e r), th e in p u t to th e h o ld -a n d -sa m p le is ch an g e d first a n d th e n th e o u tp u t sa m p le s th e n ew in p u t. T h e K o u tp u ts fro m th e K h o ld -a n d -s a m p le devices a re m u ltip lie d by th e p e rio d ic a lly tim e-v a ry in g co effic ie n ts g( n, m - \ m / I \ l ) , fo r n = 0, 1 , . . . , K - 1, an d th e re su ltin g p ro d u c ts a re su m m e d to yield y(m ). T h e c o m p u ta tio n s a t th e o u tp u t o f th e h o ld -a n d -sa m p le d ev ices are r e p e a te d a t th e o u tp u t sa m p lin g ra te o f F v = ( I / D ) F X. F inally, ra te co n v ersio n by a ra tio n a l fa c to r 1 / D can also be p e rfo rm e d by use o f a p o ly p h ase filter h av in g 1 subfilters. If w e assu m e th a t th e m th sam p le n(0.l).l = 0. I......./ - I
Figure 10.17
Efficient realization of sampling-rate conversion by a factor //>.
Sec. 10.5
Fitter Design and Implementation for Sampling-Rate Conversion
803
y ( m ) is com p uted by taking the output o f the imth subfilter with input data x( n ) , x ( n - 1 ) , . . . , x ( n K + 1 ) , in the delay line, the n ext sam p le y ( m + 1) is taken from the (i + i)st subfilter after shifting l m+j new sam ples in the d elay lin es w here i m+i = (i + D )mod / and l m +1 is the in teger part o f (im + D ) / I . T h e in teger im+i should be saved to b e used in determ ining the subfilter from w hich the n ext sam p le is taken. L et us now dem onstrate the filter design procedure, first in the design o f a d ecim ator, secon d in th e design o f an interpolator, and finally, in th e design o f a rational sam ple-rate converter.
Example 10.5.1
Design a decim ator that downsamples an input signal x(n) by a factor D = 2. Use the Rem ez algorithm to determ ine the coefficients of the F IR filter that has a 0.1-dB ripple in the passband and is down by at least 30 dB in the stopband. Also determ ine the polyphase filter structure in a decim ator realization that employs polyphase filters.
Solution A filter of length M = 30 achieves the design specifications given above. The impulse response of the FIR filter is given in Table 10.1 and the frequency response is illustrated in Fig. 10.18. Note that the cutoff frequency is wc = jr/2. The polyphase filters obtained from h{n) have impulse responses
p*(n) = h(2n + k) k = 0,1; n = 0 ,1 , ,.. ,1 4
N ote that po(n) = h(2n ) and p\{tt) = h(2n + 1). Hence one filter consists of the evennum bered samples o f h(n) and the other filter consists of the odd-num bered samples of h(n).
Example 10.5.2
Design an interpolator that increases the input sampling rate by a factor of I = 5. Use the R em ez algorithm to determ ine the coefficients of the F IR filter that has 0.1-dB ripple in the passband and is down by at least 30 dB in the stopband. Also, determ ine the polyphase filter structure in an interpolator realization based on polyphase filters.
Solution A filter o f length M = 30 achieves the design specifications given above. The frequency response of the F IR filter is illustrated in Fig. 10.19 and its coefficients are given in Table 10.2. The cutoff frequency-is wc = n/5. The polyphase filters obtained from h(n) have impulse responses
pk{n) = h(Sn + jt) Consequently, each filter has length 6. * = 0 ,1 ,2 ,3 ,4
Example 10.53
Design a sam ple-rate converter that increases the sampling rate by a factor 2.5. Use the Rem ez algorithm to determ ine the coefficients of the F IR filter that has 0.1-dB ripple in the passband and is down by at least 30 dB in the stopband. Specify the sets of time-varying coefficients g( n, m) used in the realization of the sampling-rate converter according to the structure in Fig. 10.17.
Solution The F IR filter that m eets the specifications of this problem is exactly the same as the filter designed in Example 10.5.2. Its bandwidth is tt/5.
804
Muttirate Digital Signal Processing TABLE 10.1 COEFFICIENTS OF LINEAR-PHASE FIR FILTER IN EXAMPLE 10.5.1
FINITE IMPULSE RESPONSE (FIR)
Chap. 10
LINEAR-PHASE DIGITAL FILTER DESIGN REMEZ EXCHANGE ALGORITHM FILTER LENGTH =30
H { 1) H ( 2) H ( 3) H( 4) H( H( 5) 6) H ( 7) 8)
IMPULSE RESPONSE 0 .6 0 2 5 6 1 6 5 E - 0 2 = H ( 30) = -0 .1 2 8 1 7 1 4 3 E - 0 1 = H ( 29) -0..2 8 5 8 2 0 6 6 E - 0 2 = HI 28) = 0..1 3 6 6 3 3 4 6 E - 0 1 H ( 27) = -0,,4 6 6 8 8 9 6 1 E - 0 2 = H ( 26) = -0,.1 9 7 0 4 4 1 5 E - 0 1 = H ( 25) = 0,.1 5 9 8 4 6 2 3 E - 0 1 = H ( 24) = = 0 .21384886E-01 = -0..3 4 9 7 9 4 4 0 E - 0 1 = -0 . 1 5 6 1 5 5 2 2 E - 0 1 = 0..6 4 0 0 6 1 1 3 E - 0 1 = - 0 .7 3 4 5 1 7 7 2 E - 0 2 = -0. 1 1 8 7 3 1 8 5 E + 0 0 = 0. 9 8 0 4 7 8 4 5 E - 0 1 0. 4 9 2 2 5 0 6 8 E + 0 0 = BAND 1 = H ( = H( = H ( = H( = H ( H( = H{ = H( 23) 22) 21) 20) 19) 18) 17) 16) BAND 2 0.. 3 1 0 0 0 0 0 0.. 5 0 0 0 0 0 0 0.. 0 0 0 0 0 0 0 1.. 0 0 0 0 0 0 0 0.. 0 2 1 4 3 0 2 33. 3 7 9 4 7 4 6
9) H (10) H (11) H (12 ) H(13) H (14) H (15)
H( H(
LOWER BAND EDGE UPPER BAND EDGE DESIRED VALUE WEIGHTING DEVIATION DEVIA T I O N IN DB
0.0000000 0.2500000 1.0000000 2.0000000 0.0107151 0.0925753
EXTREMAL FREQUENCIES-MAXIMA OF THE ERROR CURVE 0.0000000 0.0416667 0.0791667 0.1166666 0.1854166 0.2145832 0.2395832 0.2500000 0.3225000 0.4829165 0.3495833 0.3808333 0.4141666
0.1520833 0.3100000 0.4474999
The coefficients of the filter are given by (10.4.8) as g( n, m) ~ h(rtl + (mD)/) = h ^ nil l + mD By substituting 1 = 5 and D = = 2, we obtain g(n, x,m) m) = h {^5 5n + 2 m - 5
m DA
I
y\)
2m \
By evaluating g(n, m) for n = 0, 1 , . . . . 5 and m = 0 , 1 , . . . , 4 we obtain the following
Sec. 10.5
805
Figure 10.18 Magnitude response o f linear-phase FIR filter of length M = 30 in Example 10.5.1.
Figure 10.19 Magnitude response o f linear-phase FIR filter of length M = 30 in Example 10.5.2.
coefficients for the time-variant filter: *(0,m ) = {*(0) * (l,m ) = {*(5) g ( 2 , m ) = {*(10) g ( 3 , m ) = {*(15) g ( 4 , m ) = {*(20) g ( 5 , m ) = {*(25) *(2) *(7) *(12) *(17) *(22) *(27) h{ 4) *( 9) *(14) *(19) *(24) *(29) *(1) *( 6) *(11) *(16) *(21) *(26) *(3)} A(8)J *(13)} *(18)} *(23)} *(28)}
A polyphase filter im plementation would employ five subfilters, each of length six. To decim ate the output of the polyphase filters by a factor of D = 2 simply means
806
TABLE 10.2
Multirate Digital Signal Processing COEFFICIENTS OF LINEAR-PHASE FIR FILTER IN EXAMPLE 10.5.2
F I N I T E I M P U L S E R E S P O N S E (FIR) LINEAR-PHASE DIGITAL FILTER DESIGN REMEZ EXCHANGE ALGORITHM FILTER LENGTH = 3 0 IMPULSE RESPONSE 0 .6 3 9 8 7 2 1 6 E - 0 2 = H ( = = -0..1 4 7 6 1 3 0 4 E - 0 1 = H ( = -0..1 0 8 8 6 5 7 7 E - 0 2 = H ( = -0..2 8 7 1 4 9 5 7 E - 0 2 = H { 0 .1 0 4 8 6 4 3 0 E - 0 1 = H< 0 .2 1 4 7 7 1 4 2 E - 0 1 = H ( = 0,.1 9 4 7 9 3 6 2 E - 0 1 = H( = -0,.3 1 0 6 7 4 3 1 E - 0 3 = -0 , .3 0 0 5 3 0 3 3 E - 0 1 = - 0 .4 9 8 7 7 0 2 9 E - 0 1 = - 0 .3 7 3 7 1 2 8 5 E - 0 1 = 0. 1 8 4 8 2 8 9 6 E - 0 1 = 0. 1 0 7 4 7 1 4 1 E + 0 0 0..1 9 9 5 1 0 9 8 E + 0 0 0 .2 5 7 9 4 8 2 8 E + 0 0 = BAND 1 0.0000000 0.1000000 1.0000000 3.0000000 0.0097524 0.0842978 = H( = H ( = H ( = H { = H( = H { = H (
Chap. 10
H ( 1) H ( 2) H< 3) H ( 4) H ( 5) H( 6) H ( 7) H< 8) H( 9) H (10) H {11) H (12) H (13) H (14) H(15)
30! 29) 28) 27) 26) 25) 24) 23) 22) 21) 20) 19) 18) 17)
= H ( 16) BAND 2 0 . 1600000 0. 5 0 0 0 0 0 0 0., 0 0 0 0 0 0 0 1 .0000000 0,.0 2 9 2 5 7 2 30. 6 7 5 3 3 4 9
LOWER BAND EDGE UPPER BAND EDGE DESIRED VALUE WEIGHTING DEVIATION DEVIATION : IN D B
EXTREMAL FREQUENCIES-MAXIMA OF THE ERROR CURVE 0.0000000 0.1600000 0.3058332 0 .4829164 0.0333333 0.1745833 0.3412498 0.0645834 0.2016666 0.3766665 0.0895833 0.2370833 0.4120831 0.1000000 0.2704166 0.4474997
that we take every other output from the polyphase filters. Thus the first output v(0) is taken from pQ (n), the second output y (l) is taken from P2(n), the third from p*(n), the fourth from pi(n), the fifth from p-i(n). and so on.
10.6 MULTISTAGE IMPLEMENTATION OF SAMPLING-RATE CONVERSION

In practical applications o f sam pling-rate con version w e often en cou n ter decim a tion factors and in terp olation factors that are m uch larger than unity. For exam ple, su p p ose that w e are given the task o f altering th e sam pling rate by the factor
Sec. 10.6
Multistage Implementation of Sampling-Rate Conversion
807
I / D = 130/63. A lth o u g h , in th e o ry , th is ra te a lte ra tio n can b e a c h iev ed exactly, th e im p le m e n ta tio n w o u ld re q u ire a b a n k o f 130 p o ly p h a se filters a n d m ay b e c o m p u ta tio n a lly in efficien t. In th is section w e c o n s id e r m e th o d s f o r p e rfo rm in g sa m p lin g -ra te c o n v e rsio n fo r e ith e r D 1 a n d /o r 7 > 1 in m u ltip le stages. F irst, let us c o n s id e r in te rp o la tio n by a fa c to r 7 1 a n d le t u s a ssu m e th a t 7 can b e fa c to re d in to a p r o d u c t o f p o sitiv e in te g e rs as
'-fl*
( 10 . 6 . 1)
T h e n , in te rp o la tio n by a fa c to r 7 c a n b e a c c o m p lish e d b y cascad in g L sta g es o f in te rp o la tio n a n d filterin g , as sh o w n in Fig. 10.20. N o te th a t th e filter in ea c h o f th e in te rp o la to rs e lim in a te s th e im ag es in tro d u c e d by th e u p sa m p lin g p ro cess in th e c o rre sp o n d in g in te rp o la to r. In a sim ilar m a n n e r, d e c im a tio n by a fa c to r D, w h e re D m ay b e fa c to re d in to a p ro d u c t o f p o sitiv e in te g e rs as D = \\D .
( 10.6 .2)
can be im p le m e n te d as a cascad e o f J stag es o f filtering an d d e c im atio n as illu s tr a te d in Fig. 10.21. T h u s th e sa m p lin g ra te a t th e o u tp u t o f th e ith sta g e is F, = Fj-i D, i = 1 ,2 ........ J (10.6.3)
w h ere th e in p u t ra te fo r th e se q u e n c e {*()] is F0 = Fx . T o e n s u re th a t n o aliasing o ccu rs in th e o v erall d e c im a tio n p ro cess, w e can d esig n each filter stag e to avoid aliasin g w ithin th e fre q u e n c y b a n d o f in te re st. T o
*(n)
F, Stage 1 Figure 10.20 Stage 2 h h Fi Stage L
Multistage implementation of interpolation by a factor /,
Stage 1
Stage 2
Fx
Staged
D\Dt
Figure 10.21 Multistage implementation of decimation by a factor D.
808
Chap. 10
e la b o ra te , let us d efin e th e d e s ire d p a s sb a n d an d th e tra n sitio n b a n d in th e overall d e c im a to r as P assb an d : 0 < F < (10.6.4) T ra n s itio n b a n d : Fpc 2; F < Fsc w h e re Fx < FX/ 2 D . T h e n , aliasing in th e b a n d 0 < F < Fx is a v o id e d by selectin g th e fre q u e n c y b a n d s o f ea c h filter sta g e a s follow s: P a ssb a n d : 0 < F < F ^ T ra n s itio n b an d : Fpc < F < Fj - Fx F , S to p b a n d : Fj - Fx < F < ~ F o r ex am p le, in th e first filter sta g e we h av e F\ = Fx / D \ . a n d th e filter is d esig n ed to h av e th e fo llow ing fre q u e n c y bands: P assb an d : 0 < F < F ^ T ra n s itio n b an d : F ^ < F < Fy - Fsz fT S to p b a n d : F\ - Fx < F < - y A f te r d e c im atio n b y D\ , th e re is aliasing from th e signal c o m p o n e n ts th a t fall in th e filter tra n sitio n b a n d , b u t th e aliasin g o ccu rs at fre q u e n c ie s ab o v e Fsc. T h u s th e re is n o aliasin g in the fre q u e n c y b a n d 0 < F < Fx . B y d esig n in g th e filters in th e su b s e q u e n t sta g es to satisfy th e sp ecificatio n s given in (10.6.5). w e e n su re th at n o aliasin g o ccu rs in th e p rim a ry fre q u e n c y b a n d 0 < F < Fsc. Example 10.6.1 Consider an audio-band signal with a nominal bandwidth of 4 kH z that has been sampled at a rate of 8 kHz. Suppose that we wish to isolate the frequency components below 80 Hz with a filter that has a passband 0 < F < 75 and a transition band 15 < F < 80. H ence F ^ = 75 Hz and FK = 80. The signal in the band 0 < F < 80 may be decim ated by the factor D = F J 2 F X = 50. We also specify that the filter have a passband ripple 8, = 102 and a stopband ripple of < 5 2 = 10-4. The length o f the linear phase F IR filter required to satisfy these specifications can be estim ated from one of the well known formulas given in the literature. R e call that a particularly simple formula for approxim ating the length M , attributed to Kaiser, is ^ = - ' 0 lo ^ - 1 3 14.6 A / + 1 (1 M 7 ) (10 6 6) ^^
where A/ is the norm alized (by the sampling rate) width of the transition region [i.e., A f = ( Fx Fpc)/Fs]. A m ore accurate formula proposed by H errm ann et al. (1973) is ^ l M A/ + 1 (10.6.8)
Sec. 10.6
Multistage Implementation of Sampling-Rate Conversion
809
where Dî Si , ^2) and /( 5 i, 52) are defined as A (S i.fc) = [0.005309(logini i ) 2 + G.07114(log]0Si) -0.4761] log1052 -[0.00266(logi0 5i)2 + 0.5941 log10Si + 0.4278] f ( S 1, i 2) = 11.012 + O.51244[log105, - log102] (10.6.9) (10.6.10)
Now a single FIR filter followed by a decim ator would require (using the Kaiser form ula) a filter of (approximate) length 14.6(5/8000) 1 3 + ^ ^ 2 As an alternative, let us consider a two-stage decimation process with D\ = 25 and Di = 2. In the first stage we have the specifications F\ = 320 Hz and Passband: 0 < F < 75 Transition band: 15 < F < 240 J61 A^ 8000 fin = y
^21 = 2
Note that we have reduced the passband ripple < 5 [ by a factor of 2, so that the total passband ripple in the cascade of the two filters does not exceed Si. On the other hand, the stopband ripple is maintained at S2 in both stages. Now the Kaiser formula yields an estim ate of A/, as
M l=
~ 101ogn > ^ 2 1 - 1 3 + J % 167 14.6A/
For the second stage, we have F2 = F\fL 160 and the specifications Passband: 0 < F < 75 Transition band: 75 < F < 80 A f = 320 < 5 i2 ~
y
&n = h
Hence the estim ate of the length M2 of the second filter is
M2 ~ 220
Therefore, the total length of the two F IR filters is approxim ately M\ + M2 = 387. This represents a reduction in the filter length by a factor of more than 13. The reader is encouraged to repeat the com putation above with D\ = 10 and >2 = 5. It is apparent from the com p utations in E xam p le 10.6.1 that the reduction in th e filter length results from increasing the factor A f , w hich appears in the d en om in ator in (10.6.7) and (10.6.8). By decim ating in m ultiple stages, w e are
810
Chap. 10
ab le to in c re a se th e w id th o f th e tra n sitio n reg io n th ro u g h a re d u c tio n in th e sa m p lin g rate. In th e case o f a m u ltista g e in te rp o la to r, th e sa m p lin g r a te a t th e o u tp u t of th e / t h stag e is //_! = ItFi i = J , J - l .........1 a n d th e o u tp u t ra te is Fo = I F j w h e n th e in p u t sa m p lin g ra te is Fj . T h e c o r re sp o n d in g freq u e n cy b a n d sp e cificatio n s a re P a ssb a n d : 0 < F < Fp T ra n s itio n b an d : Fp < F < F; Fsc T h e fo llow ing e x am p le illu stra te s th e a d v a n ta g e s o f m u ltista g e in te rp o la tio n . Example 10.6.2 Let us reverse the filtering problem described in Exam ple 10.6.1 by beginning with a signal having a passband 0 < F < 75 and a transition band of 75 < F < 80. We wish to interpolate by a factor of 50. By selecting I\ = 2 and l2 = 25. we have basically a transposed form of the decim ation problem considered in Exam ple 10.6.1. Thus we can simply transpose the two-stage decim ator to achieve the two-stage interpolator with /, = 2 , i2 = 25. Mi 220. and M2 % 167.
10.7 SAMPLING-RATE CONVERSION OF BANDPASS SIGNALS

A b an d p a ss signal is a signal w ith fre q u e n c y c o n te n t c o n c e n tra te d in a n a rro w b an d o f fre q u e n c ie s ab o v e z e ro freq u e n cy . T h e c e n te r fre q u e n c y Fc o f th e signal is g en erally m uch la rg e r th a n th e b a n d w id th B (i.e., Fc B). B a n d p a ss signals arise fre q u e n tly in p ra c tic e , m o st n o ta b ly in c o m m u n ic a tio n s, w h e re in fo rm a tio n b e a rin g signals su c h as sp e ech an d v id eo a re tra n s la te d in fre q u e n c y a n d th e n tra n sm itte d o v e r such c h a n n e ls as w ire lines, m icro w av e ra d io , a n d sa tellites. In th is se ctio n w e c o n s id e r th e d e c im a tio n a n d in te rp o la tio n o f b a n d p a ss signals. W e b eg in b y n o tin g th a t an y b a n d p a s s signal has an e q u iv a le n t low pass re p re se n ta tio n , o b ta in e d by a sim ple fre q u e n c y tra n sla tio n o f th e b a n d p a s s signal. F o r ex am p le, th e b a n d p a s s signal w ith sp e c tru m X ( F ) sh o w n in Fig. 10.22a can b e tra n s la te d to lo w p ass by m e a n s o f a fre q u e n c y tra n sla tio n o f Fc, w h e re Fc is an a p p ro p ria te ch o ice o f fre q u e n c y (u su ally , th e c e n te r fre q u e n c y ) w ith in th e b a n d w id th o c c u p ied b y th e b a n d p a s s signal. T h u s w e o b ta in th e e q u iv a le n t low pass signal as illu stra te d in Fig. 10.22b. F ro m S ectio n 9.1 w e re c a ll th a t an a n a lo g b a n d p a s s signal c a n b e re p re se n te d as x ( t ) = ,4(/) cos[27rFcr + ?(/)] = A( t ) cosf?(r) c o s 2 jrF cr - 4 (/) s in 0 ( /)s in 2 7 rF c/ (10.7.1) = u c(t) cos 2 n F ct us (t) s i n 2 n F ct
= R & [x i( t)e J7* Fcl]
Sec. 10.7
Sampling-Rate Conversion of Bandpass Signals
811
|W)|
Bandpass signal (a)
Equivalent lowpass signal

<b)
Figure 10.22
Bandpass signal and its equivalent lowpass representation.
w h ere, by d efin itio n , u c( t ) = y 4 (r)c o s# (0 u s (t) ~ j4 (O sin 0 (O x t (t) = u c(t) + j u s (t) (10.7.2) (10.7.3) (10.7.4)
A( t ) is called th e a mp l i t u d e o r envel ope o f th e signal, O(t') is th e p h a se , a n d u c(t) a n d u s( t ) a re called th e quadrat ur e c o mp o n e n t s o f th e signal. P h y sically , th e tra n s la tio n o f x {t ) to low pass in volves m u ltip ly in g (m ixing) x (r) by th e q u a d r a tu re c a rrie rs cos 2 tz Fct a n d s i n 2 ^ F c; a n d th e n low pass filtering th e tw o p ro d u c ts to e lim in a te th e fre q u e n c y c o m p o n e n ts g e n e ra te d a ro u n d th e f re q u e n c y 2FC (th e d o u b le freq u e n c y te rm s). T h u s all th e in fo rm a tio n c o n te n t c o n ta in e d in th e b a n d p a s s signal is p re se rv e d in th e low pass signal, a n d h en ce th e la tte r is e q u iv a le n t to th e fo rm er. T h is fact is o b v io u s fro m th e sp e c tra l r e p re s e n ta tio n o f th e b a n d p a s s signal, w hich can b e w ritte n as X ( F ) = \ [ X , { F - Fc) + X ; { - F - Fc)] (10.7.5)
w h e re X i ( f ) is th e F o u rie r tra n sfo rm o f th e e q u iv a le n t low pass sig n al xi{t) a n d X { F ) is th e F o u rie r tra n s fo rm o f ;c(f).
812
Chap. 10
It w as sh o w n in S ectio n 9.1 th a t a b a n d p a s s signal o f b a n d w id th B can be u n iq u e ly re p re s e n te d by sa m p le s ta k e n at a ra te o f 2 B sa m p le s p e r se co n d , p r o v id ed th a t th e u p p e r b a n d (h ig h e st) fre q u e n c y is a m u ltip le o f th e signal b a n d w id th B. O n th e o th e r h a n d , if th e u p p e r b a n d fre q u e n c y is n o t a m u ltip le of B, th e sa m p lin g r a te m u st b e in c re a se d by a sm all a m o u n t to a v o id aliasing. In an y case, th e sa m p lin g ra te fo r th e b a n d p a s s sig n a l is b o u n d e d fro m a b o v e a n d b elo w as 2 B < Fs < 4 B (10.7.6)
T h e re p re s e n ta tio n o f d isc re te -tim e b a n d p a ss signals is b asically th e sa m e as th a t fo r an alo g sig n als given by (10.7.1) w ith th e su b s titu tio n o f t = n T , w h e re T is th e sa m p lin g in terv al.
10.7.1 Decimation and Interpolation by Frequency Conversion

T h e m a th e m a tic a l eq u iv a le n c e b e tw e e n th e b a n d p a s s signal x ( t ) a n d its e q u iv a le n t lo w p ass re p re s e n ta tio n xi( t ) p ro v id e s o n e m e th o d fo r a lte rin g th e sa m p lin g rate o f th e signal. S pecifically, w e can ta k e th e b a n d p a s s signal w h ich h as b e e n sa m p led a t ra te Fx , c o n v e rt it to low pass th ro u g h th e fre q u e n c y c o n v e rsio n p ro cess illu stra te d in Fig. 10.23, a n d p e rfo rm th e sa m p lin g -ra te c o n v e rsio n o n th e low pass sig n al u sing th e m e th o d s d e sc rib e d prev io u sly . T h e low pass filters fo r o b ta in in g th e tw o q u a d r a tu re c o m p o n e n ts can b e d esig n e d to h av e lin e a r p h a s e w ithin the
Figure 10.23
Conversion of a bandpass signal to lowpass.
Sec. 10.7
Sampling-Rate Conversion of Bandpass Signals
813
b a n d w id th o f th e sig n al an d to a p p ro x im a te th e id eal fre q u e n c y re sp o n s e c h a ra c te ristic
*<>-(
< m 77 )
w h e re coB is th e b a n d w id th o f th e d isc re te -tim e b a n d p a s s sig n al (a>B < n ) . I f d e c im a tio n is to b e p e rfo rm e d by a n in te g e r fa c to r D, th e an tialiasin g filter p re c e d in g th e d e c im a to r can be c o m b in e d w ith th e low pass filter u se d fo r fre q u e n c y c o n v e rsio n in to a single filter th a t a p p ro x im a te s th e id eal freq u e n c y r e sp o n s e H D(a>) = h 0, M <a >p / D o th e rw ise (10.7.8)
w h e re coD is an y d e s ire d freq u e n cy in th e ra n g e 0 < coD < n . F o r e x am p le, w e m ay se le c t < d d = m b / 2 if w e a re in te re ste d only in th e fre q u e n c y ran g e 0 < co < coB/ 2 D o f th e o rig in a l signal. If in te rp o la tio n is to be p e rfo rm e d by an in te g e r fa c to r I o n th e fre q u e n c y tra n s la te d sig n al, th e filter used to re je c t th e im ag es in th e sp e c tru m sh o u ld b e d e sig n e d to a p p ro x im a te th e low pass filter c h a ra c te ristic
"'<->-{
(10-7-9>
W e n o te th a t in th e case o f in te rp o la tio n , th e low pass filte r n o rm ally u se d to re je c t th e d o u b le -fre q u e n c y c o m p o n e n ts is re d u n d a n t an d m ay b e o m itte d . Its fu n ctio n is e sse n tia lly se rv e d by th e im age re je c tio n filter H[(co). F in ally , w e in d ic a te th a t sa m p lin g -ra te co n v e rsio n by an y ra tio n a l fa c to r //> can b e acco m p lish ed o n th e b a n d p a s s signal as illu stra te d in Fig. 10.24. A g a in , th e lo w p ass filter fo r re je c tin g th e d o u b le -fre q u e n c y c o m p o n e n ts g e n e ra te d in th e fre q u e n c y -c o n v e rsio n p ro cess can b e o m itte d . Its fu n c tio n is sim ply se rv e d by th e im a g e -re je c tio n /a n tia lia s in g filter follow ing th e in te r p o la to r , w hich is d esig n ed to a p p ro x im a te th e id eal fre q u e n c y re sp o n s e c h a ra c te ristic : = (10.7.10)
[ 0,
o th erw ise
O n c e th e sa m p lin g ra te o f th e q u a d r a tu re signal c o m p o n e n ts h a s b e e n a lte re d by e ith e r d e c im a tio n o r in te rp o la tio n o r b o th , a b a n d p a s s signal c a n b e r e g e n e r a te d by a m p litu d e m o d u la tin g th e q u a d r a tu re c a rrie rs cos<wcn a n d sin cocn by th e c o rre s p o n d in g signal c o m p o n e n ts a n d th e n a d d in g th e tw o signals. T h e c e n te r
Figure 10.24
Sampling rate conversion of a bandpass signal.
814
Chap. 10
fre q u e n c y
is an y d e s ira b le fre q u e n c y in th e ran g e mi n(u>B/ 2D, (oB/ 2 l ) < coc < n (10.7.11)
10.7.2 Modulation-Free Method for Decimation and Interpolation

B y re stric tin g th e fre q u e n c y ra n g e fo r th e signal w hose fre q u e n c y is to be a lte re d , it is p o ssib le to av o id th e c a rrie r m o d u la tio n p ro c e ss a n d to ach iev e freq u e n cy tra n sla tio n d irectly . In th is case w e e x p lo it th e freq u e n cy tra n sla tio n p ro p e rty in h e re n t in th e p ro c e ss o f d e c im a tio n an d in te rp o la tio n . T o b e specific, let us c o n s id e r th e d e c im a tio n o f th e sa m p le d b an d p a ss signal w h o se sp e ctru m is sh o w n in Fig. 10.25. N o te th a t th e signal sp e c tru m is confined to th e freq u e n c y ran g e mir (m + l ) 7 r , _ ----- < co < :--------- (10.7.12) D ~ ~ D w h ere m is a p o sitiv e in te g e r. A b a n d p a s s filter w o u ld n o rm a lly b e u sed to elim in ate signal fre q u e n c y c o m p o n e n ts o u tsid e th e d e s ire d fre q u e n c y ran g e. T h e n d irect d e cim atio n o f th e b a n d p a s s signal by th e fa c to r D re su lts in th e sp e c tru m sh o w n in Fig. 10.26a, fo r m o d d , a n d Fig. 10.26b fo r m even. In th e case w h ere m is o d d . th e re is an in v ersio n o f th e sp e c tru m o f th e signal. T h is in v ersio n can b e u n d o n e by m u ltip ly in g ea c h sa m p le o f th e d e c im a te d signal by ( - 1 ) " , n = 0. 1, N o te th a t v io la tio n o f th e b a n d w id th c o n s tra in t given by (10.7.12) re su lts in signat aliasing, M o d u la tio n -fre e in te rp o la tio n o f a b a n d p a s s signal by an in te g e r fa c to r I can b e acco m p lish ed in a sim ilar m a n n e r. T h e p ro c e ss o f u p sa m p lin g by in se rtin g zero s b e tw e e n sa m p les o f x( n ) p ro d u c e s I im ag es in th e b a n d 0 < w < i t . T h e d e sire d im ag e can b e se le c te d by b a n d p a s s filtering. N o te th a t th e p ro c e ss o f in te r p o la tio n also p ro v id es u s w ith th e o p p o rtu n ity to ach iev e fre q u e n c y tra n sla tio n o f th e sp e ctru m . F inally, m o d u la tio n -fre e sa m p lin g r a te co n v e rsio n fo r a b a n d p a s s signal by a ra tio n a l fa c to r I / D can b e acco m p lish ed b y cascad in g a d e c im a to r w ith an in te r p o la to r in a m a n n e r th a t d e p e n d s on th e ch o ice o f th e p a r a m e te r s D a n d I. A b a n d p a s s filter p re c e d in g th e sa m p lin g c o n v e rte r is u su ally re q u ire d to isolate th e signal freq u e n c y b an d o f in te re st. N o te th a t th is a p p ro a c h p ro v id e s us w ith a m o d u la tio n -fre e m e th o d fo r ach iev in g fre q u e n c y tra n sla tio n o f a signal by selecting D = 1.
(m + i) n
mTi
mn
{m + 1)n
D Figure 10.25
D Spectrum o f a bandpass signal.
Sec. 10.8
Sampling-Rate Conversion by an Arbitrary Factor
815
m odd
(>
m even
(b>
Figure 10.26
Spectrum of decimated bandpass signal.
10.8 SAMPLING-RATE CONVERSION BY AN ARBITRARY FACTOR

In the previou s sectio n s o f this chapter, w e have show n h ow to perform sam pling rate con version exactly by a rational num ber I / D . In som e applications, it is either inefficient or, so m etim es im possib le to use such an exact rate con version schem e. W e first con sid er the fo llow in g tw o cases.
Case 1. W e n eed to perform rate con version by the rational num ber I / D , w h ere / is a large in teger (e.g., l / D = 1023/511). A lth ou gh w e can ach ieve exact rate con version by this num ber, w e w ould n eed a p olyp hase filter with 1023 subfilters. Such an exact im plem en tation is ob viou sly inefficient in m em ory usage b ecau se w e n eed to store a large num ber o f filter coefficients. Case 2. In so m e applications, th e exact con version rate is n ot know n w hen w e design the rate converter, or th e rate is con tinu ou sly changing during the co n version p rocess. For exam p le, w e m ay en cou n ter the situation w here th e input and ou tp u t sam p les are con trolled by tw o in d ep en dent clocks. E ven though it is still p ossib le to d efine a n om in al con version rate that is a rational num ber, the actual
816
Chap. 10
ra te w o u ld b e slig h tly d iffe re n t, d e p e n d in g on th e fre q u e n c y d iffe re n c e b etw een th e tw o clocks. O b v io u sly , it is n o t p o ssible to design an e x a c t ra te c o n v e rte r in th is case. T o im p le m e n t sa m p lin g r a te c o n v ersio n fo r a p p lic a tio n s sim ilar to th e se cases, we re s o rt to n o n e x a c t ra te c o n v ersio n sc h em es. U n a v o id a b ly , a n o n ex act sc h em e will in tro d u c e so m e d isto rtio n in th e c o n v e rte d o u tp u t signal. (It sh ould be n o te d th a t d isto rtio n exists ev en in an ex a c t ra tio n a l r a te c o n v e rte r because th e p o ly p h a se filter is n e v e r id eal.) S uch a c o n v e rte r will b e a d e q u a te , as long as th e to ta l d isto rtio n d o e s n o t ex cee d th e specification r e q u ire d in th e a p p li catio n . D e p e n d in g on th e ap p lic a tio n re q u ire m e n ts a n d im p le m e n ta tio n c o n stra in ts, w e can use first-o rd e r, se c o n d -o rd e r, o r h ig h e r-o rd e r a p p ro x im a tio n s . W e shall d e sc rib e first-o rd e r a n d se c o n d -o rd e r a p p ro x im a tio n m e th o d s a n d p ro v id e an analysis o f th e resu ltin g tim in g e rro rs.
10.8.1 First-Order Approximation

L e t u s d e n o te th e a rb itra ry c o n v e rsio n ra te by r a n d su p p o se th a t th e in p u t to th e ra te c o n v e rte r is th e se q u e n c e (jr(n)}. W e n e e d to g e n e ra te a se q u e n c e o f o u tp u t sa m p le s se p a ra te d in tim e by Tx f r , w h e re Tx is th e sa m p le in te rv a l for {x(n)i. By co n stru ctin g a p o ly p h a se filter w ith a larg e n u m b e r o f su b filters as ju s t d esc rib e d , we can a p p ro x im a te such a se q u e n c e w ith a n o n u n ifo rm ly sp a c e d se q u e n c e . W ith o u t loss o f g e n e ra lity , w e can e x p ress 1 j r as
w h ere k a n d I a re p o sitiv e in te g e rs a n d 0 is a n u m b e r in th e ra n g e 0 < &< y C o n se q u e n tly , 1j r is b o u n d e d fro m ab o v e an d b e lo w as

_
<
<
-------------
jfc + 1
I r I I c o rre sp o n d s to th e in te rp o la tio n fa c to r, w hich w ill b e d e te r m in e d to satisfy th e specificatio n on th e a m o u n t o f to le ra b le d isto rtio n in tro d u c e d b y ra te c o n v ersio n . / is also eq u a l to th e n u m b e r o f p o ly p h a se filters. F o r e x am p le, su p p o se th a t r = 2.2 a n d th a t w e h a v e d e te r m in e d , as we will d e m o n s tra te , th a t 1 = 6 p o ly p h a se filters a re r e q u ire d to m e e t th e d isto rtio n specificatio n . T h e n k ^ 2 1 3 k + 1 / 6 r 6 I so th a t k = 2. T h e tim e sp a cin g b e tw e e n sa m p le s o f th e in te r p o la te d se q u e n c e is Tx / I . H o w e v e r, th e d e s ire d c o n v ersio n ra te r = 2.2 fo r / = 6 c o rre sp o n d s to a d e c im a tio n fa c to r o f 2.727, w h ich falls b e tw e e n k = 2 a n d k = 3. In th e first-o rd e r a p p ro x im a tio n , w e ach iev e th e d e s ire d d e c im a tio n r a te b y se le c tin g th e o u tp u t
Sec. 10.8
817
Figure 10.27
Sample rate conversion by use of first-order approximation.
sa m p le fro m th e p o ly p h a se filter closest in tim e to th e d e s ire d sa m p lin g tim e. T h is is illu stra te d in Fig. 10.27 fo r 1 = 6. In g e n e ra l, to p e rfo rm ra te c o n v ersio n by a fa c to r r, w e em p lo y a p o ly p h ase filter to p e rfo rm in te rp o la tio n an d th e re fo re to in c re a se th e fre q u e n c y o f th e o rig in al s e q u e n c e o f a fa c to r o f I. T h e tim e sp acin g b e tw e e n th e sa m p le s o f th e in te r p o la te d se q u e n c e is e q u a l to Tx/ J . If th e id eal sa m p lin g tim e o f th e m th s a m p le , y ( m) , o f th e d e sire d o u tp u t se q u e n c e is b e tw e e n th e sa m p lin g tim e s o f tw o sa m p le s o f th e in te rp o la te d se q u en ce, w e se lect th e sa m p le clo se r to y ( m ) as its ap p ro x im a tio n . L e t u s a ssu m e th a t th e m th se le c te d sa m p le is g e n e ra te d b y th e (im)th su b filter u sin g th e in p u t sa m p le s x( n) , x ( n 1 ) , . . . , x ( n K + 1) in th e d e la y line. T h e n o rm a liz e d sa m p lin g tim e e rro r (i.e., th e tim e d iffe re n c e b e tw e e n th e se le c te d sa m p lin g tim e a n d th e d e s ire d sam p lin g tim e n o rm a liz e d b y Tx) is d e n o te d by tm. T h e sign o f tm is p o sitiv e if th e d e s ire d sa m p lin g tim e le a d s th e se le c te d sam pling tim e , a n d n e g a tiv e o th e rw ise . It is easy to show th a t \tm | < 0 .5 / / . T h e n o rm a liz e d tim e a d v a n c e fro m th e m th o u tp u t y ( m ) to th e (m + l ) s t o u tp u t y ( m + 1) is e q u a l t o ( 1 / r ) + tm. T o c o m p u te th e n e x t o u tp u t, w e first d e te rm in e a n u m b e r clo sest to im/ J + + km/ I th a t is o f th e fo rm /m_] + im+ i / I , w h e re b o th /m+1 a n d i m+1 a re in te g e rs a n d im+i < I. T h e n , th e (m + l ) s t o u tp u t y(m + 1) is c o m p u te d using th e (jm+i) th su b filte r a f te r sh iftin g th e signal in th e d e la y line by lm+i in p u t sam p les. T h e n o rm a liz e d tim in g e r r o r fo r th e (m + l ) th sa m p le is tm+1 = ( i m/ I + l / r + t m) ~ (l m+] + im+i / I ) . I t is sa v ed fo r th e c o m p u ta tio n o f th e n e x t o u tp u t sam ple.
818
Chap. 10
B y in creasin g th e n u m b e r o f su b filters u se d , w e can a rb itra rily in c re a se th e co n v ersio n accu racy . H o w e v e r, w e also re q u ire m o re m e m o ry to sto re th e larg e n u m b e r of filter coefficients. H e n c e it is d e s ira b le to use as few su b filters as possible w hile k e e p in g th e d isto rtio n in th e c o n v e rte d signal below th e sp ecificatio n . T he d isto rtio n in tro d u c e d d u e to th e sa m p lin g -tim e a p p ro x im a tio n is m o st c o n v en ien tly e v a lu a te d in th e fre q u e n c y d o m a in . S u p p o se th a t th e in p u t d a ta se q u e n c e {x( n)\ h a s a flat s p e c tru m fro m - t o x to cox , w h e re u> x < w ith a m a g n itu d e A. Its to ta l p o w e r can b e c o m p u te d using P a rse v a ls th e o re m , n am ely ,
Ssi r j \X{(o)\2d ( o = 5 2?r ; _ tU
(10.8.1)
F ro m th is d iscussion given, w e k n o w th a t fo r each o u tp u t v(m ), th e tim e d ifferen ce b e tw e e n th e d e s ire d filter a n d th e filter actu ally u se d is tm, w h e re | r j < 0 .5 / / . H e n c e th e fre q u e n c y re sp o n se o f th e s e filters c a n b e w ritte n as eJa> T and resp ectiv ely . W h e n / is la rg e , wtm is sm all. By ig n o rin g h ig h -o rd e r e rro rs, we can w rite th e d ifferen ce b e tw e e n th e fre q u e n c y re sp o n s e s as e jur _ eja>U-l) _ 1 _ e - j w,)
( 10 .8 .2 )
= e i0J1 (1 cos cot,,, + j sin colm) % j e JU 'Tcot,,, B y using th e b o u n d | irn\ < 0 .5 / / , w e o b ta in an u p p e r b o u n d fo r th e to ta l e rro r p o w e r as Pe = - L ' \ X ( c o ) e ^ T ~ X(co)eJU,lT~ ^ \ 2d w ^ [ ' \X(<o)jeJâ>tm\2dco J o jx 2-7T J (u ,
-^La 2 {t) 2 d w = ^
( 1 0 '8 '3 )
T h is b o u n d show s th a t th e e r r o r p o w e r is in v ersely p ro p o rtio n a l to th e s q u a re o f th e n u m b e r o f su b filters I. T h e re fo re , th e e r r o r m a g n itu d e is in v e rse ly p ro p o rtio n a l to I. H e n c e w e call th e a p p ro x im a tio n o f th e ra te c o n v ersio n m e th o d d e sc rib e d a b o v e a first-o rd e r a p p ro x im a tio n . B y u sin g (10.8.3) a n d (10.8.1), th e r a tio o f th e sig n a l-to -d isto rtio n d u e to a sa m p lin g -tim e e r r o r fo r th e firs t-o rd e r a p p ro x im a tio n , d e n o te d as S D rR l , is lo w e r b o u n d e d as S D ,R 1 = P pe M l2 > ~ (4 (10.8.4)
It can be seen fro m (10.8.4) th a t th e sig n a l-to -d isto rtio n r a tio is p ro p o rtio n a l to th e sq u a re o f th e n u m b e r o f su b filters. Exam ple 10.8.1 Suppose that the input signal has a flat spectrum betw een - 0 .8 jt and 0.8 tt. Determ ine the num ber of subfilters to achieve a signal-to-distortion ratio of 50 dB.
Sec. 10.8
819
Solution T o achieve an SDrR > 105, we set S D iR l = 12/2/wJ equal to 105. Thus we find that fw 1 s w ,,/ ~ ~ ~ 230 subfilters
10.8.2 Second-Order Approximation (Linear Interpolation)

T h e d isa d v a n ta g e o f th e first-o rd e r a p p ro x im a tio n m e th o d is th e larg e n u m b e r o f su b filte rs n e e d e d to ach iev e a specified d isto rtio n re q u ire m e n t. In th e follow ing d iscu ssio n , w e d esc rib e a m e th o d th a t u se s lin e a r in te rp o la tio n to ach iev e th e sa m e p e rfo rm a n c e w ith a r e d u c e d n u m b e r o f subfilters. T h e im p le m e n ta tio n o f th e lin e a r in te rp o la tio n m e th o d is v ery sim ilar to th e firs t-o rd e r a p p ro x im a tio n d iscu ssed ab o v e. In ste a d o f u sin g th e sa m p le fro m th e in te rp o la tin g filter clo sest to th e d e s ire d c o n v ersio n o u tp u t as th e a p p ro x im a tio n , w e c o m p u te tw o a d ja c e n t sa m p le s w ith th e d e s ire d sa m p lin g tim e falling b e tw e e n th e ir sa m p lin g tim es, as is illu stra te d in Fig. 10.28. T h e n o rm a liz e d tim e spacing
y(+1M i a m=Ilm
Figure 10.28 Sample rale conversion by use o f linear interpolation.
820
Chap. 10
b e tw e e n th e se tw o sa m p le s is 1 j l . A ssu m in g th a t th e sa m p lin g tim e o f th e first sa m p le lags th e d e s ire d sa m p lin g tim e by j m, th e sa m p lin g tim e o f th e se c o n d sa m p le is th e n lead in g th e d e s ire d sa m p lin g tim e by ( 1 / / ) tm. If w e d e n o te th ese tw o sa m p le s by vi (m) a n d y i ( m ) a n d use lin e a r in te rp o la tio n , w e can c o m p u te the a p p ro x im a tio n to th e d e s ire d o u tp u t as v(m ) = (1 - a m)vi(m ) + a n,y2(m) (10.8.5)
w h e re a m = l t m. N o te th a t 0 < a, < 1 . T h e im p le m e n ta tio n o f lin e a r in te rp o la tio n is sim ilar to th a t fo r th e firsto r d e r a p p ro x im a tio n . N o rm a lly , b o th y i(m ) a n d yz i m) a re c o m p u te d using th e /th an d (/ + l) th su b filters, re sp e c tiv e ly , w ith th e sa m e se t o f in p u t d a ta sa m p le s in th e d elay line. T h e o n ly e x c e p tio n is in th e b o u n d a ry case, w h e re / = / - 1. In th is case w e u se th e ( / l ) t h su b filte r to c o m p u te >'i(/w), b u t th e se co n d sam ple yi ( m) is c o m p u te d u sin g th e z e ro th su b filter a fte r new in p u t d a ta a re sh ifted into th e d e la y line. T o an aly ze th e e r r o r in tro d u c e d by th e se c o n d -o rd e r a p p ro x im a tio n , we first w rite th e fre q u e n c y re sp o n s e s o f th e d esired filter an d th e tw o su b filters u se d to c o m p u te vi(m ) a n d V 2(w ), as e-/W r, ancj ja>u-t+'\/i)^ resp ec tiv ely . B e cau se lin e a r in te rp o la tio n is a iin e a r o p e ra tio n , w e can also use lin e a r in te rp o la tio n to co m p u te th e fre q u e n c y re sp o n se o f th e filter th a t g e n e ra te s y(w?) as
= e Jtur[(l a m)e~JW '"' + ameJ'i~! "+l / n ]

( 10 . 8 .6 )
- ejmT0 - a,)(cos coin, - ;s in o tf m) -h e ^ Ta [ c o s a j ( l / / - tm) + j sin c o (l/7 - r , ) ] B y ig n o rin g h ig h -o rd e r e rro rs, w e can w rite th e d ifferen ce b e tw e e n th e d esired fre q u e n c y resp o n se s an d th e o n e given by (10.8.6) as ei m - (1 - a m)e Jtu<t ," ) \ - (i cim) c o s u)tm am costL>(l// - tm)] (10.8.7)
+ j [ ( l - a m) sin (D tm - a m s in a > ( l// - rm)]}
U sin g (1 - a m) am < 1 Pe = = - /
w e o b ta in a n u p p e r b o u n d fo r th e to ta l e r r o r p o w e r as - a me ^ - ' ^ n ]\2dco
\X(u>)[ejwT - (1 -
(10.8.8)
Sec. 10.9
Applications of Multirate Signal Processing
821
T his result indicates that the error m agnitude is inversely proportional to I 2. H en ce w e call the approxim ation using linear interpolation a s e c ond- or d e r appr ox i ma t i o n. U sin g (10.8.8) and (10.8.1), the ratio o f signal-to-distortion d u e to a sam pling tim e error for the secon d-ord er approxim ation, d en oted by S D ,R 2 , is b ou n ded from b elo w as S D jR 2 = P Pe 80/ 4 > arx (10.8.9)
T h erefore, th e signal-to-distortion ratio is p roportional to the fourth p ow er o f the num ber o f subfilters. Exam ple 10.8.2 D eterm ine the num ber of subfilters required to meet the specifications given in E x ample 10.8.1 when linear interpolation is employed. Solution obtain To achieve S D ,R > 105, we set S D iR 2 = 80 I A/w* equal to 10s. Thus we
From this exam ple we see that the required num ber o f subfilters for the secon d-ord er approxim ation is reduced by a factor o f ab ou t 15 com pared to the first-order approxim ation. H ow ever, w e now n eed to co m p u te tw o interpolated sam p les in this case, instead o f o n e for the first-order approxim ation. H en ce w e have d ou b led the com putational com plexity. L inear interpolation is the sim plest case o f the class o f approxim ation m eth od s based o n Lagrange p olynom ials. It is also p ossib le to use higher-order L a grange polyn om ial approxim ations (interp olation) to further reduce the num ber o f subfilters required to m eet specifications. H ow ever, th e secon d-ord er approx im ation seem s sufficient for m ost practical applications. T h e in terested reader is referred to the paper by R am stad (1984) for higlier-order L agrange interpolation m ethods.
10.9 APPLICATIONS OF MULTIRATE SIGNAL PROCESSING

T here are num erous practical applications o f m ultirate signal processing. In this sectio n w e d escribe a few o f these applications.
10.9.1 Design of Phase Shifters

S u p p ose that w e w ish to design a netw ork that delays th e signal x ( n ) by a fraction o f a sam ple. L et us assum e that the d elay is a rational fraction o f a sam pling
822
x(n) Ft 71 If I
Chap. 10
Lowpass filler
IF,
Delay by k samples
1 j
Figure 10.29
Method for generating a delay in a discrete-time signal.
in te rv a l Tx [i.e.. d = (k / I ) T x, w h e re k a n d / a re relativ ely p rim e po sitiv e in teg ers]. In th e freq u e n cy d o m ain , th e d elay c o rre sp o n d s to a lin e a r p h a s e shift o f th e fo rm @(a>) = kio (10.9.1)
T h e d esign o f an all-p ass lin e a r-p h a s e filter is relativ ely difficult. H o w e v e r, w e can use th e m e th o d s o f sa m p le -ra te co n v e rsio n to ach iev e a d elay o f (k / l ) T x, ex actly , w ith o u t in tro d u c in g an y significant d isto rtio n in th e sig n al. T o be specific, let us co n sid e r th e sy stem sh o w n in Fig. 10.29. T h e sa m p lin g r a te is in crease d by a fa c to r / using a s ta n d a rd in te rp o la to r. T h e lo w p ass filter e lim in a te s th e im ages in th e sp e ctru m o f th e in te rp o la te d signal, a n d its o u tp u t is d e la y e d by k sa m p le s a t th e sam p lin g r a te I F x . T h e d e la y e d sig n a l is d e c im a te d by a fa c to r D = I. T h u s we h av e ach iev ed th e d e sire d d e la y o f (k / l ) T x . A n efficient im p le m e n ta tio n o f th e in te r p o la to r is th e p o ly p h a s e filter illu s tra te d in Fig. 10.30. T h e d e la y o f k sa m p le s is ach iev ed by p lacing th e initial p o sitio n o f th e c o m m u ta to r a t th e o u tp u t o f th e Jtth su b filter. S in ce d e c im a tio n by
Figure 10.30 Fig. 10.29.
Polyphase filter structure for implementing the syslem shown in
Sec. 10.9
823
D = I m e a n s th a t w e ta k e o n e o u t o f ev ery / sa m p le s fro m th e p o ly p h a s e filter, th e c o m m u ta to r p o sitio n c a n b e fixed to th e o u tp u t o f th e fcth su b filter. T h u s a d e la y in k / I can b e a c h iev ed by using o n ly th e k th su b filter o f th e p o ly p h ase filter. W e n o te th a t th e p o ly p h a se filter in tro d u c e s an a d d itio n a l d e la y o f (Af l ) /2 sam ples, w h ere M is th e le n g th o f its im p u lse re sp o n se . F in ally , w e m e n tio n th a t if th e d e s ire d d elay is a n o n ra tio n a l fa c to r o f th e sa m p le in te rv a l Tx , e ith e r th e firs t-o rd e r o r se c o n d -o rd e r a p p ro x im a tio n m e th o d d e s c rib e d in S ectio n 10.8 can b e u se d to o b ta in th e delay.
10.9.2 Interfacing of Digital Systems with Different Sampling Rates

In p ra c tic e w e fre q u e n tly e n c o u n te r th e p ro b le m o f in terfacin g tw o d igital sy stem s th a t a re c o n tro lle d b y in d e p e n d e n tly o p e ra tin g clocks. A n an a lo g so lu tio n to th is p ro b le m is to c o n v e rt th e signal fro m th e first sy stem to a n a lo g form an d th e n re sa m p le it a t th e in p u t to th e se c o n d sy stem u sin g th e clock in th is system . H o w e v e r, a sim p le r a p p ro a c h is o n e w h e re th e in te rfa c in g is d o n e by a digital m e th o d u sin g th e b asic sa m p le -ra te c o n v ersio n m e th o d s d e sc rib e d in this c h a p te r. T o b e specific, le t us c o n s id e r in te rfa c in g th e tw o sy stem s w ith in d e p e n d e n t clocks as sh o w n in F ig. 10.31. T h e o u tp u t o f system A a t ra te Fx is fed to an in te rp o la to r w hich in c re a se s th e sa m p lin g ra te by I . T h e o u tp u t o f th e in te rp o la to r is fed a t th e ra te I Fx to a d igital sa m p le -a n d -h o ld w hich se rv es as th e in te rfa c e to sy stem B a t th e h ig h sa m p lin g r a te 1FX. S ignals fro m th e d igital sa m p le -a n d -h o ld a re re a d o u t in to sy stem B at th e clock ra te D F y of sy stem B. T h u s th e o u tp u t r a te fro m th e sa m p le -a n d -h o ld is n o t sy n c h ro n iz e d w ith th e in p u t rate. In th e sp ecial case w h e re D = I a n d th e tw o clock ra te s a re c o m p a ra b le b u t n o t id en tical, so m e sa m p le s a t th e o u tp u t o f th e sa m p le -a n d -h o ld m ay be r e p e a te d o r d ro p p e d a t tim es. T h e a m o u n t o f signal d isto rtio n re su ltin g from th is m e th o d can b e k e p t sm a ll if th e in te rp o la to r/d e c im a to r fa c to r is large. B y using lin e a r in te rp o la tio n in p la c e o f th e d ig ital sa m p le -a n d -h o ld , as w e d e sc rib e d in S ectio n 10.8, w e can f u rth e r re d u c e th e d isto rtio n a n d th u s re d u c e th e size o f th e in te rp o la to r facto r.
Figure 1031
Interfacing of two digital systems with different sampling rates.
824
Chap. 10
10.9.3 Implementation of Narrowband Lowpass Filters

In S ectio n 10.6 w e d e m o n s tra te d th a t a m u ltista g e im p le m e n ta tio n o f sam plingra te co n v ersio n o fte n p ro v id es fo r a m o re efficient re alizatio n , esp ecially w h en th e filter sp ecificatio n s a re v ery tig h t (e.g., a n a rro w p a s sb a n d a n d a n a rro w tra n sitio n b a n d ). U n d e r sim ilar c o n d itio n s, a low pass, lin e a r-p h a s e F I R filte r m ay b e m o re efficien tly im p le m e n te d in a m u ltista g e d e c im a to r-in te rp o la to r co n fig u ratio n . T o b e m o re specific, w e can e m p lo y a m u ltistag e im p le m e n ta tio n o f a d e c im a to r o f size D, fo llo w ed by a m u ltista g e im p le m e n ta tio n o f an in te r p o la to r o f size / , w h ere I = D. W e d e m o n s tra te th e p ro c e d u re b y m ean s o f a n ex a m p le fo r th e desig n o f a lo w p ass filter w hich h as th e sa m e sp ecificatio n s as th e filter th a t is given in E x a m p le 10.6.1. Example 10.9.1 Design a linear-phase FIR filter that satisfies the following specifications: Sampling frequency: Passband: Transition band: Stopband Passband ripple: Stopband ripple: 8000 Hz 0 < F < 75 75 < F < 80 80 < F < 4000 5] = 10-2 & 2 = 10~4
Solution If this filter were designed as a single-rate linear-phase F IR filter, the length of the filter required to m eet the specifications is (from Kaisers form ula) M 5152
Now, suppose that we employ a multirate im plem entation of the lowpass filter based on a decimation and interpolation factor of D = / = 100. A single-stage implem entation of the decim ator-interpolator requires an FIR filter of length
^ ^ - O l o g W ;/ 2 ) - 1 3 + l a 5 4 a )
14.6 A /
However, there is a significant savings in computational complexity by implementing the decimator and interpolator filters using their corresponding polyphase filters. If we employ linear-phase (symmetric) decimation and interpolation filters, the use of polyphase filters reduces the multiplication rate by a factor of 100. A significantly more efficient im plementation is obtained by using two stages of decimation followed by two stages of interpolation. For example, suppose that we select D\ = 50, D2 = 2, h = 2, and I2 = 50. Then the required filter lengths are -lO lo g C JA /4 ) - 1 3 _____ W l.......... 1 4 .6 V ---------+
-101oglo( W 4 ) - 1 3 . , _ ---------- u Z T f ----------+ 1 ^ 233 Thus we obtain a reduction in the overall filter length of 2(5480)/2(177+233) ~ 13.36In addition, we obtain further reduction in the multiplication rate by using polyphase
Sec. 10.9
825
filters. For the first stage of decimation, the reduction in multiplication rate is 50, while for the second stage the reduction in multiplication rate is 100. F urther reductions can be obtained by increasing the num ber of stages of decim ation and interpolation.
10.9.4 Implementation of Digital Filter Banks

F ilte r b a n k s a re g en erally c a teg o rized as tw o types, analysis filter b a n k s a n d s y n thesis filter ban k s . A n analysis filter b a n k co n sists o f a se t o f filters, w ith system fu n c tio n s {#;(&)}, a rra n g e d in a p a ra lle l b a n k as illu s tra te d in Fig. 10.32a. T h e f re q u e n c y re sp o n s e c h a ra c te ristic s o f th is filter b a n k sp lits th e signal in to a c o r re sp o n d in g n u m b e r o f su b b a n d s. O n th e o th e r h a n d , a sy n th e sis filter b a n k co n sists o f a se t o f filters w ith system fu n ctio n s {G *(')}, a rra n g e d as sh o w n in Fig. 10.32b, w ith c o rre sp o n d in g in p u ts {v*(/i)}- T h e o u tp u ts o f th e filters a re su m m e d to fo rm th e sy n th e siz e d signal {*(/;)) F ilte r b a n k s a re o fte n u se d fo r p e rfo rm in g sp e c tru m analysis a n d signal syn th esis. W h en a filter b a n k is em p lo y ed in th e c o m p u ta tio n o f th e d isc re te F o u rie r
Analysis fitter bank (a)
Synthesis filter bank (b) Figure 10.32 A digital filter bank.
826
Chap. 10
tra n sfo rm (D F T ) o f a se q u e n c e {*()}, th e filter b a n k is called a D F T filter b an k . A n an aly sis filter b an k co n sistin g o f N filters {Hk(z), Jt = 0 , 1 , ___ N 1} is called a u n ifo rm D F T filter b a n k if Hk (z ), it = 1, 2 , . . . , N 1, a re d e riv e d fro m a p ro to ty p e filter H q(z ), w h ere (10.9.2) H e n c e th e fre q u e n c y re sp o n se c h a ra c te ristic s o f th e filters { //* ( '), Jt = 0, 1 , , / / l} a re sim ply o b ta in e d b y u n ifo rm ly sh iftin g th e fre q u e n c y re sp o n s e o f th e p ro to ty p e filter b y m u ltip le s of 2 n / N . In th e tim e d o m a in th e filters a re c h a ra c te riz e d by th e ir im p u lse resp o n ses, w hich can b e e x p re ss e d as h k (n) = h 0( n)ej2nnk/N k = 0 , l ____ N - l (10.9.3)
w h e re {/iq()} is th e im pulse re sp o n s e o f th e p ro to ty p e filter. T h e u n ifo rm D F T analysis filter b a n k can be re a liz e d as sh o w n in Fig. 10.33a, w h e re th e fre q u e n c y c o m p o n e n ts in th e se q u e n c e {jr(n)} a re tra n s la te d in freq u e n cy to lo w p ass by m u ltip ly in g x( n) w ith th e co m p lex e x p o n e n tia ls e x p ( j l n n k j N ) , k = 1___ N - 1, an d th e re su ltin g p ro d u c t signals a re p assed th ro u g h a low pass filter w ith im p u lse re sp o n se {ho(n)). Since th e o u tp u t o f th e low pass filter is relativ ely n a rro w in b an d w id th , th e signal can b e d e c im a te d by a fa c to r D < N . T h e resu ltin g d e c im a te d o u tp u t signal can b e e x p re ss e d as X k(m) n m = 0 , 1,... w h e re {X*(m)} a re sa m p les o f th e D F T a t fre q u e n c ie s a> k = 2nk/N. T h e c o rre sp o n d in g sy n th e sis filte r fo r ea c h e le m e n t in th e filter b a n k can b e v iew ed as show n in Fig. 10.33b, w h e re th e in p u t signal se q u e n c e s {l*(m ), k = 0, 1 , . . . , N 1} a re u p sa m p le d by a fa c to r o f I = D, filte re d to re m o v e th e im ag es, a n d tra n sla te d in fre q u e n c y by m u ltip lic a tio n by th e c o m p lex ex p o n e n tia ls { e x p i j l n n k / N ) , k = 0, 1, . . N 1). T h e re su ltin g fre q u e n c y -tra n sla te d signals fro m th e N filters are th e n su m m ed . T h u s w e o b ta in th e se q u e n c e h o ( mD n)x(/i)e', -jlxnk/N k = 0, 1, . . . , N - 1 (10.9.4)
^
i n
ej2"klN
Y , y^ m ^8o(n - m l )
_
= !> ( *
w n
- m l)
Jlnnk/N 1 Y d m ) e ^ nklN 1* I__ n
(10.9.5)
w h e re th e fa c to r 1 / N is a n o rm a liz a tio n fa c to r, {y(m)} re p re s e n t sa m p le s o f th e in v e rse D F T se q u e n c e c o rre sp o n d in g to {}*(m)}, {go()} is th e im p u lse resp o n se o f th e in te rp o la tio n filter, a n d I ~ D.
Sec. 10.9
Applications of Muttirate Signal Processing
627
e-jain n
Analysis (a)
e)<w
Synthesis
(b)
Figure 1033
w* = N
A uniform D FT filter bank.
T h e re la tio n sh ip b e tw e e n th e o u tp u t (X*(n)} o f th e analysis filter b a n k a n d th e in p u t {y*(m )} to th e sy n th e sis filter b a n k d e p e n d s o n th e a p p lic a tio n . U s u ally, {y*(rn)| is a m o d ified v ersio n o f {X*(m)), w h e re th e specific m o d ifica tio n is d e te rm in e d b y th e ap p licatio n . A n a lte rn a tiv e re a liz a tio n o f th e an aly sis a n d sy n th e sis filter b a n k s is illu s tr a te d in Fig. 10.34. T h e filters a re re a liz e d as b a n d p a s s filters w ith im p u lse r e sp o n ses hk(n) = h0(n)eJ7* nk/N k = 0,1,..., N - 1 (10.9.6)
828

p-jtiitfnD
Chap. 10
Analysis (a)
ejm<jnD
Synthesis (b) Figure 10.34 Alternative realization of a uniform DFT filter bank.
T h e output o f each bandpass filter is d ecim ated by a factor D and m ultiplied by e x p i j l n m k / N ) to p roduce the D F T seq u en ce {X*(m)}. T h e m odu lation by the com plex exp on en tial allow s us to shift the spectrum o f the signal from = 2nk/N to oo = 0. H en ce this realization is equ ivalen t to the realization given in Fig. 10.33. T h e filter bank output can b e written as "-jlnmkD/N
X k( m) =
j c ( n ) * o ( m D - n ) e j2nk(mD~n)/N
(10.9.7)
T h e corresponding filter bank syn th esizer can be realized as show n in Fig. 10.34b, w h ere the input seq u en ces are first m ultiplied by the exp on en tial factors [ e x p ( j 2 n k m D / N) ] , upsam pled by the factor I = D, and the resulting se-
Sec. 10.9
829
q u e n c e s a re filtered by th e b a n d p a s s in te rp o la tio n filters w ith im p u lse re sp o n se s

g k ( n) = g 0( n ) e j2j!nk/N
(10.9.8)
w h e re {o()) is th e im p u lse re sp o n se o f th e p r o to ty p e filter. T h e o u tp u ts o f th e se filters are th e n su m m e d to yield N-1 1 < " ) = j j i .

k=0
(10.9.9)
w h e re I D. In th e im p le m e n ta tio n of d igital filters b a n k s, c o m p u ta tio n a l efficiency can be ac h iev ed by u se o f p o ly p h a se filters fo r d e cim atio n a n d in te rp o la tio n . O f p a rtic u la r in te re st is th e case w h e re th e d e c im a tio n fa c to r D is se le c te d to be e q u a l to th e n u m b e r N o f fre q u e n c y b an d s. W h en D = N, w e say th a t th e filter b a n k is critically sampled. F o r th e an aly sis filter b an k , let us define a se t o f N = D p o ly p h a se filters w ith im p u lse resp o n ses p t ( n ) = ho ( n N k) k = 0. 1.........jV 1 (10.9.10)
a n d th e c o rre sp o n d in g se t o f d e c im a te d in p u t se q u e n c e s x k (n) = x ( n N + k) * = 0 . 1 .........N - 1 (10.9.11)
N o te th a t th is d efin itio n o f {/;*()) im plies th a t th e c o m m u ta to r fo r th e d e c im a to r r o ta te s clockw ise. T h e s tru c tu re o f th e analysis filter b an k b ased o n th e use o f p o ly p h ase filters can b e o b ta in e d by su b stitu tin g (10.9.10) a n d (10,9.11) in to (10.9.7) a n d re a rra n g in g th e s u m m a tio n in to th e fo rm
-jZ n n k / N
k - 0 , 1 ____ D - 1
(10.9.12)
w h e re N = D. N o te th a t th e in n e r su m m a tio n r e p re s e n ts th e co n v o lu tio n o f i p n(l)} w ith {*(/)). T h e o u te r su m m atio n re p re s e n ts th e W -point D F T o f th e filter o u tp u ts. T h e filter s tru c tu re c o rre sp o n d in g to th is c o m p u ta tio n is illu strated in Fig. 10.35. E a c h sw eep o f th e c o m m u ta to r re su lts in N o u tp u ts, d e n o te d as ( r ( m ), n = 0, 1.........A ' - l ) from th e N p o ly p h ase filters. T h e / V -p o in t D F T o f th is se q u e n c e yield s th e sp e c tra l sa m p le s (Jft(m )]. F o r la rg e valu es o f N , th e F F T alg o rith m p ro v id e s an efficient m e a n s fo r co m p u tin g th e D F T . N o w su p p o se th a t th e sp e ctral sa m p le s { ^ ( m ) } a re m o d ified in so m e m a n n e r, p re sc rib e d by th e ap p lic a tio n , to p ro d u c e |y t.(m)}. A filter b a n k sy n th e sis filter b a s e d o n a p o ly p h a se filter s tru c tu re can b e re a liz e d in a sim ilar m a n n e r. F irst, w e d efin e th e im p u lse re sp o n se o f th e N {D = I N ) p o ly p h a se filters fo r th e in te rp o la tio n filter as qk(n) = go( nN + k) k = 0 ,1 ........ N - 1 (10.9.13)
830
Chap. 10
Figure 1035
Digital filter bank structure for the computation of (10.9.12).
a n d th e c o rre sp o n d in g set o f o u tp u t signals as vt () = v(n N + k) * = 0 , 1 .........A ' - l
(10.9.14)
N o te th a t th is d efin itio n o f {qt(n)} im plies th a t th e c o m m u ta to r fo r th e in te rp o la to r r o ta te s co u n terclo ck w ise . B y su b stitu tin g (10.9.13) in to (10.9.5), w e can ex p ress th e o u tp u t v/(n) o f th e /th p o ly p h a se filter as
j N -\
V i(n) = ^ < ? /( m )
-]T n(m )e
j2Kkl/H
i =0,1...., N - 1
(10.9.15)
T h e te rm in b ra c k e ts is th e /V -point in v e rse D F T o f {% ("*)}, w h ich w e d e n o te as {y; (m)}. H en ce Vi (tt) = ^ qt(n m) yi {m) I 0, 1, . .. , N 1 (10.9.16)
T h e sy n th esis s tru c tu re c o rre sp o n d in g to (10.9.16) is sh o w n in Fig. 10.36. It is in te re stin g to n o te th a t by d efin in g th e p o ly p h a s e in te rp o la tio n filte r as in (10.9.13), th e stru c tu re in Fig. 10.36 is th e tra n sp o s e o f th e p o ly p h a se an aly sis filter show n in Fig. 10.35. In o u r tr e a tm e n t o f d ig ital filter b a n k s w e c o n s id e re d th e im p o rta n t case o f critically sa m p le d D F T filter b a n k s, w h e re D ~ N . O th e r ch o ic e s o f D a n d N c an b e e m p lo y ed in p ra c tic e , b u t th e im p le m e n ta tio n o f th e filters b e c o m e s m o re c o m p lex . O f p a rtic u la r im p o rta n c e is th e o v e rsa m p le d D F T filte r b a n k , w h ere N = K D, D d e n o te s th e d e c im a tio n f a c to r a n d K is an in te g e r th a t specifies the
Sec. 10.9
831
Figure 10,36
Digital filter bank structure for the computation of (10.9.16).
o v e rsam p lin g facto r. In this case it can be sh o w n t hat th e p o ly p h ase filter b a n k s tru c tu re s for th e an alysis an d sy n th e sis filters can be im p le m e n te d by use o f N su b filters a n d /V-point D F T s an d in v erse D F T s.
10.9.5 Subband Coding of Speech Signals

A v ariety o f te c h n iq u e s have b e e n d e v e lo p e d to efficiently r e p re s e n t sp e ech signals in dig ital fo rm fo r e ith e r tran sm issio n o r sto ra g e . S ince m o st o f th e sp e ech en e rg y is c o n ta in e d in th e lo w er freq u e n cies, w e w ould like to e n c o d e th e lo w er-freq u en cy b a n d w ith m o re b its th a n th e h ig h -fre q u e n c y b a n d . S u b b a n d coding is a m e th o d , w h ere th e sp e ech signal is su b d iv id ed in to se v e ra l fre q u e n c y b an d s a n d each b a n d is d ig itally e n c o d e d se p a ra te ly . A n e x am p le o f a fre q u e n c y su b d iv isio n is show n in Fig. 10.37a. L e t us a s su m e th a t th e sp eech signal is sa m p le d a t a ra te Fs sa m p le s p e r se co n d . T h e first fre q u e n c y su b d iv ision splits th e signal sp e c tru m in to tw o eq u al-w id th se g m e n ts, a lo w p ass sig n al (0 < F < Fxj 4) a n d a h ig h p ass signal (Fs/ 4 < F < Fs / 2). T h e se co n d fre q u e n c y subdivision sp lits th e low pass signal from th e first stag e in to tw o e q u a l b an d s, a low pass signal (0 < F < Fs/ 8) an d a h ig h p ass signal ( Fs/& < F < F J 4). F inally, th e th ird fre q u e n c y su b d iv isio n splits th e low pass signal fro m th e se co n d sta g e in to tw o e q u a l b a n d w id th signals. T h u s th e sig n a l is su b d iv id e d in to fo u r fre q u e n c y b an d s, co v erin g th re e o ctav es, as sh o w n in Fig. 10.37b. D e c im a tio n by a fa c to r o f 2 is p e rfo rm e d a fte r fre q u e n c y su b division. By a llo catin g a d iffe re n t n u m b e r o f b its p e r sa m p le to th e signal in th e fo u r su b b a n d s, w e can a c h iev e a re d u c tio n in th e b it ra te o f th e d ig italized sp e ech signal.
832
Chap. 10
(a)
j 1
0 JT
2
JT
3
7T
4 2 (b)
[
JT
Figure 10.37
Block diagram of a subband speech coder.
F ilte r d esign is p a rtic u la rly im p o rta n t in ach iev in g goo d p e rfo rm a n c e in s u b b a n d coding. A liasin g re su ltin g fro m d e c im a tio n o f th e su b b a n d signals m u st be n eg ligible. It is c le a r th a t w e c a n n o t use brickw all filter c h a ra c te ristic s as sh o w n in Fig. 10.38a, since such filters a re ph y sically u n re a liz a b le . A p a rtic u la rly p ractical so lu tio n to th e aliasin g p ro b le m is to u se quadrat ur e mi r r or filters (Q M F ), w hich h a v e th e fre q u e n c y re sp o n se c h a ra c te ristic s show n in Fig. 10.38b. T h e se filters are d e sc rib e d in th e fo llo w ing sectio n . T h e syn th esis m e th o d for th e su b b a n d e n c o d e d sp e ech signal is b asically th e rev erse o f th e e n c o d in g p ro cess. T h e signals in a d ja c e n t lo w p ass a n d h ig h p ass freq u e n cy b an d s a re in te rp o la te d , filte re d , a n d c o m b in e d as sh o w n in Fig. 10.39. A p a ir o f Q M F is u se d in th e signal sy n th e sis fo r ea c h o ctav e o f th e signal. S u b b a n d co d in g is also an effe ctiv e m e th o d to ach iev e d a ta co m p re ssio n in im ag e signal p ro cessin g . B y c o m b in in g su b b a n d co d in g w ith v e c to r q u a n tiz a tio n fo r each su b b a n d sig n al, S a fra n e k e t al. (1988) h av e o b ta in e d c o d e d im ag es w ith a p p ro x im a te ly ^ b it p e r pixel, c o m p a re d w ith 8 b its p e r p ix el fo r th e u n c o d e d im age. In g e n eral, s u b b a n d co d in g o f signals is an effe ctiv e m e th o d fo r achieving b an d w id th c o m p ressio n in a d ig ital r e p re s e n ta tio n o f th e signal, w h en th e signal e n erg y is c o n c e n tra te d in a p a rtic u la r re g io n o f th e fre q u e n c y b a n d . M u ltira te signal p ro cessin g n o tio n s p ro v id e efficient im p le m e n ta tio n s o f th e su b b a n d e n co d er.
Sec. 10.9
833
(a)
Q M F
(b) Figure 10.38 Filter characteristics for subband coding.
10.9.6 Quadrature Mirror Filters

T h e b asic b u ild in g b lo c k in a p p licatio n s o f q u a d r a tu re m irro r filters (Q M F ) is th e tw o -c h a n n e l Q M F b a n k show n in Fig. 10.40. T h is is a m u ltira te digital filter s tru c tu re th a t em p lo y s tw o d e c im a to rs in th e signal an a ly sis se c tio n a n d tw o in te r p o la to r s in th e sig nal sy n th e sis sectio n . T h e lo w p ass a n d h ig h p ass filters in th e an aly sis se c tio n h a v e im p u lse resp o n ses ho(n) a n d h \ ( n ) , resp ec tiv ely . S im ilarly, th e lo w p ass a n d h ig h p ass filters c o n ta in e d in th e sy n th e sis se c tio n h a v e im p u lse re sp o n s e s go(n) a n d g i( n ), respectively.
834
Chap. 10
Outpa
Figure 10J9
Synthesis o f subband-encoded signals.
Analysis section Figure 10.40
Synthesis section Two-channel QM F bank.
T h e F o u rie r tra n sfo rm s o f th e sig n a ls a t th e o u tp u ts o f th e tw o d ecim ato rs
X a0(co) = -
(10.9.17)
X aX(co) = -
If Xsq(m) and Xlt(<u) represent the two inputs to the synthesis section, the output
Sec. 10.9
835
is sim ply X(co) = Xfo(2(o)Go(co) + X s\(2(o)G\(co) (10.9.18)
N o w , su p p o s e th a t we c o n n ect th e analysis filter to th e c o rre sp o n d in g sy nthesis filter, so th a t X ao{(o) = a n d X a\(oo) = -X*i(ew). T h e n , by su b stitu tin g fro m (10.9.17) in to (10.9.18), w e o b ta in X(co) = ^ \ H q{(o ) G q((o ) + H\(a))G\{( d ) ]X {w) (10.9.19)
+ \[H0(u > 7 t)Go ( < o) + Hi(co 7 t)G\(u ))]X(o } t t )
T h e first te rm in (10.9.19) is th e d e sire d signal o u tp u t fro m th e Q M F b an k . T h e seco n d te rm re p re s e n ts th e effect o f aliasing, w hich w e w o u ld like to elim in ate. H e n c e w e re q u ire th a t H 0((o - n)Gu(a>) + Hi(co -
7t)Gx(co) =
(10.9.20)
T h is co n d itio n can be sim ply satisfied by selectin g Go(co) a n d G i M as G q(cd) = Hiico tt)
G x(a>) = - H a( o > - n )
(10.9.21)
T h u s th e se co n d term in (10.9.19) vanishes. T o e la b o ra te , let us assum e th a t Hu(co) is a lo w p ass filter an d H x(io) is a m irro r-im a g e h ig h p ass filter. T h e n w e can e x p ress Hti(u>) a n d H\(u>) as H0(a>) = H(to) (10.9.22) H\ (w) = H ( oj tt) w h e re H( w) is th e fre q u e n c y re sp o n se o f a lo w p ass filter. In th e tim e d o m ain , th e c o rre sp o n d in g re la tio n s a re hnin) = h(n) (10.9.23) h\ ( n) = ( - 1 )nh( n) A s a c o n s e q u e n c e , Hn(a>) a n d H x(u>) h ave m irro r-im a g e sy m m e try a b o u t th e fre q u e n c y a) tt/2, as sh o w n in Fig. 10.38b. T o b e c o n siste n t w ith th e c o n s tra in t in (10.9.21), w e se lect th e low pass filter G0(to) as G 0(co) = 2H(co) a n d th e h ig h p ass filter G x(a>) as G x (co) = - 2 H(co - n ) In th e tim e d o m ain , th e s e re la tio n s b eco m e go(n) = 2h( n) (10.9.26) g\( n) = 2 (1 )nh{n) T h e scale fa c to r o f 2 in g0(n) a n d g x(n) c o rre sp o n d s to th e in te rp o la tio n fa c to r u se d to n o rm a liz e th e o v erall fre q u e n c y re sp o n s e o f th e Q M F . W ith th is choice o f (10.9.25) (10.9.24)
836
Chap. 10
th e filter c h aracteristics, th e c o m p o n e n t d u e to aliasin g v an ish es. T h u s th e aliasing resu ltin g fro m d e c im a tio n in th e analysis se ctio n o f th e Q M F b a n k is p e rfe c tly can ce le d by th e im ag e signal sp e c tru m th a t arises d u e to in te rp o la tio n . A s a resu lt, th e tw o -ch an n el Q M F b e h a v e s as a lin ear, tim e -in v a ria n t system . If we su b s titu te fo r Ho(a>), H-\(co), Go(co), a n d Gi(a>) in to th e first te rm o f (10.9.19), w e o b ta in X(co) = [ H 2(co) - H 2(u) tt) ] X ( co)
(10.9.27)
Id eally , th e tw o -c h a n n e l Q M F b a n k sh o u ld h ave u n ity gain, \ H 2{oj) H 2(co n ) \ = 1 fo r all co (10.9.28)
w h ere H(co) is th e fre q u e n c y re sp o n se o f a low pass filter. F u rth e rm o r e , it is also d e s ira b le fo r th e Q M F to h av e lin e a r p h ase. N ow , let u s c o n sid e r th e use o f a lin e a r p h a se filter H(co). H e n c e H(co) m ay be ex p ressed in th e form H ( w ) = Hr (co)e~Jai<N- h/2 w h ere N is th e filter len g th . T h en
H 2 (co) = H 2( w ) e - j,' N - h
(10.9.29)
(10.9.30)
=
\ H ( w ) \ 2e~J(N- ])
an d H 2( a>- tt) = H ? ( o > - j r ) e - jia,- ,niN- ' ) (10.9.31) = ( - l ) N~ l \H(co - jz)\2e - JwiN- l) T h e re fo re , th e o v erall tra n s f e r fu n ctio n o f th e tw o -ch an n el Q M F w hich em p lo y s lin e a r-p h a se F IR filters is = [\H(a>)\2 - ( - l f - ]\ H( c o - 7 T ) \ 2] e - JU ,'N- ]) (10.9.32)
X(co)
N o te th a t th e o v erall filter h a s a d elay o f N 1 sa m p le s an d a m a g n itu d e c h a ra c te ristic A(co) = \H(co)\2 - { - \ ) N- ' \ H { c o - n ) \ 2 (10.9.33)
W e also n o te th a t w h en N is o d d , A ( n f l ) = 0, b e c a u se \H{7ijT)\ = |// ( 3 j t/ 2 ) |. T h is is an u n d e sira b le p r o p e rty fo r a Q M F design. O n th e o th e r h a n d , w h e n N is ev en , A((D) = \H{co)\2 + \ H { c o - n ) \ 2 (10.9.34)
w h ich av o id s th e p ro b le m o f a z e ro at a) = n f l . F o r N ev en , th e id e a l tw o -ch an n el Q M F sh o u ld satisfy th e c o n d itio n A(<d) = \ H( w) \ 2 -I-1 H(C0 - n ) I2 = 1 for all to (10.9.35)
Sec. 10.9
837
w h ich fo llo w s fro m (10.9.33). U n fo rtu n a te ly , th e only filter fre q u e n c y re sp o n se fu n c tio n th a t satisfies (10.9.35) is th e trivial fu n ctio n \H(u>)\2 = co s2 aw. C o n s e q u e n tly , a n y n o n triv ia l lin e a r-p h a se F IR filter H(a>) in tro d u c e s so m e a m p litu d e d isto rtio n . T h e a m o u n t o f a m p litu d e d isto rtio n in tro d u c e d by a n o n triv ia l lin e a r p h a s e F IR filter in th e Q M F c a n be m in im ized by o p tim izin g th e F IR filter coefficients. A p a rtic u la rly effe ctiv e m e th o d is to select th e filter coefficients o f H(<o) such th a t A ((d) is m a d e as flat as p o ssib le w hile sim u ltan eo u sly m in im iz in g ( o r c o n strain in g ) th e s to p b a n d en erg y o f H(u>). T his a p p ro a c h lead s to th e m in im iz atio n o f th e in te g ra l sq u a re d e r ro r (10.9.36) w h e re w; is a w eig h tin g fa c to r in th e ran g e 0 < w < 1. In p e rfo rm in g th e o p ti m izatio n , th e filter im p u lse re sp o n se is c o n s tra in e d to b e sy m m e tric (lin e a r p h a se ). T h is o p tim iz a tio n is easily d o n e n u m erically on a d igital c o m p u te r. T h is a p p ro a c h h as b e e n u se d by J o h n s to n (1980), an d Ja in a n d C ro c h ie re (1984) to design tw och a n n e l Q M F s. T a b le s o f o p tim u m filter coefficients h ave b e e n ta b u la te d by J o h n sto n (1980). A s an a lte rn a tiv e to lin e a r-p h a se F IR filters, w e can design an I I R filter th a t satisfies th e all-p ass c o n s tra in t given by (10.9.28). F o r th is p u rp o s e , e llip tic filters p ro v id e esp ecially efficient designs. Since th e Q M F w o u ld in tro d u c e so m e p h ase d isto rtio n , th e signal a t th e o u tp u t o f th e Q M F can be p a sse d th ro u g h an all-pass p h a s e e q u a liz e r d esig n e d to m inim ize p h ase d isto rtio n . In a d d itio n to th e s e tw o m e th o d s fo r Q M F design, o n e can also design th e tw o -c h a n n e l Q M F s to e lim in a te co m p letely b o th a m p litu d e a n d p h a s e d isto rtio n as w ell as c a n c e lin g aliasin g d isto rtio n . Sm ith a n d B a rn w e ll (1984) h a v e show n th a t su ch p er f ect reconstruct i on Q M F can be d esig n ed by relax in g th e lin e a r-p h a se co n d itio n o f th e F IR lo w pass filter H{(o). T o ach iev e p e rfe c t re c o n stru c tio n , w e b eg in by d esig n in g a lin e a r-p h a se F IR h a lfb a n d filter o f le n g th 2 N 1. A h a lf-b a n d filter is defin ed as a z e ro -p h a se F I R filter w h o se im p u lse resp o n se {b(n}} satisfies th e c o n d itio n b(2n) = co n stan t, 0, n = 0 n 7^0 (10.9.37)
H e n c e all th e e v e n -n u m b e re d sa m p le s are z e ro e x c e p t a t n = 0. T h e z e ro -p h a se r e q u ire m e n t im p lies th a t b( n) = b( n). T h e fre q u e n c y re sp o n s e o f such a filter is K (10.9.38) w h ere K is o d d . F u rth e rm o r e , B(co) satisfies th e c o n d itio n B( w) + B(7i ~ co) is e q u a l to a c o n s ta n t fo r all freq u e n cies. T h e typical fre q u e n c y re sp o n s e c h a ra c te ristic o f a h a lf-b a n d filter is sh o w n in Fig. 10.41. W e n o te th a t th e filter re sp o n s e is sym m etric w ith re sp e c t to n j 2, th e b a n d e d g es freq u e n cies cup a n d a)s a re sy m m etric a b o u t
838
Chap. 10
w = n / 2 , a n d th e p e a k p a s s b a n d a n d s to p b a n d e r r o r s a r e e q u a l. W e a ls o n o te th a t th e f ilte r c a n b e m a d e c a u s a l b y in tr o d u c in g a d e la y o f K s a m p le s .
N o w , s u p p o s e th a t w e d e s ig n an F I R h a lf- b a n d filte r o f le n g th 2jV 1, w h e r e
N is e v e n , w ith f r e q u e n c y r e s p o n s e a s s h o w n in F ig . 1 0 .4 2 ( a ) .
c o n s t r u c t a n o th e r h a lf-b a n d filte r w ith fr e q u e n c y r e s p o n s e
F ro m
B(a>) w e
B+(w) = B(co) +
as s h o w n in F ig . 1 0 .4 2 ( b ) . s p e c tr a l fa c to r iz a tio n fl+ ( z ) = H { z ) H { z - \ - (N~ h o r , e q u iv a le n tly ,
( 1 0 .9 .3 9 )
N o t e th a t B +(a> ) is n o n n e g a tiv e a n d h e n c e it h a s th e
( 1 0 .9 .4 0 )
B+(a>) = \H(co)\2e~ ja>(N~1)

w h e r e H (to) is th e f r e q u e n c y r e s p o n s e o f a n F I R f ilt e r o f le n g th N
(1 0 .9 .4 1 ) w ith re a l
c o e f f ic ie n ts . D u e t o th e s y m m e tr y o f B+(co ) w ith r e s p e c t to co = n / 2 , w e a ls o h a v e
B+(z) + ( - 1 ) " - ' B + i - z ) = c t z - ^ - V

o r , e q u iv a le n tly ,
( 1 0 .9 .4 2 )
B+{a>)
-I- ( - D ^ f l + O w -
n) = a
( 1 0 .9 .4 3 )
w h e r e a is a c o n s t a n t . T h u s , b y s u b s t itu tin g ( 1 0 .9 .4 0 ) in to ( 1 0 .9 .4 2 ) , w e o b ta in
H(z)H(z~l) + H ( - z ) H ( ~ z ~ l ) = a
( 1 0 .9 .4 4 )
S in c e H ( z ) s a tis fie s ( 1 0 .9 .4 4 ) a n d s in c e a lia s in g is e lim in a te d w h e n w e h a v e G o (z ) =
H \ ( z) a n d G i ( z ) =
H q ( z ) , it fo llo w s t h a t t h e s e c o n d itio n s a r e s a tis fie d b y
Sec. 10.9
839
Bffft)
amplitude response of G(z)
-S
BJta) amplitude response of

* (*)
Figure 10.42 Frequency response characteristics of half-band filters B(a>) and B + ( ). (From Vaidyanathan (1987))
840
c h o o s in g H \ ( z ) , G o ( z ) , a n d G \ ( z ) as
Chap. 10
H o (z) = H { z) H \{ z ) = - z - iN - h H 0( - Z ~ ')
( 1 0 .9 .4 5 )
G 0{z) = z - (N- u H q {z - 1)
G ,( z ) = z ^ - ^ H ^ r 1 ) = - H o ( - z ) T h u s a lia s in g d is to r tio n is e lim in a te d a n d s in c e X { a > ) /X { a ) ) is a c o n s t a n t , t h e Q M F p e r f o r m s p e r f e c t r e c o n s tr u c tio n s o t h a t x ( n ) = a x ( n N + 1 ). H o w e v e r , w e n o te th a t H ( z ) is n o t a l i n e a r -p h a s e f ilte r . T h e F I R filte r s H o (z), H ] ( z ), G o (z ), a n d G i ( z ) in th e tw o - c h a n n e l Q M F b a n k a r e e f fic ie n tly r e a liz e d as p o ly p h a s e filte r s . S in c e I = D = 2 , tw o p o ly p h a s e filte r s a r e im p le m e n te d f o r e a c h d e c im a t o r a n d tw o f o r e a c h in te r p o la to r . H o w e v e r , if w e e m p lo y lin e a r -p h a s e F I R filte r s , th e s y m m e tr y p r o p e r tie s o f th e a n a ly s is filte r s a n d s y n th e s is f ilte r s a llo w u s to s im p lify th e s tr u c tu r e a n d r e d u c e th e n u m b e r o f p o ly p h a s e filte r s in th e a n a ly sts s e c tio n t o tw o filte r s a n d t o a n o t h e r tw o f ilte r s in th e s y n th e s is s e c tio n . T o d e m o n s tr a t e th is c o n s t r u c tio n , le t u s a s s u m e th a t th e filte r s a r e lin e a r p h a s e F I R filte r s o f le n g th N o f 2 , ca n b e ex p ressed as
( N e v e n ) , w h ich h a v e im p u ls e r e s p o n s e s g iv e n by
( 1 0 .9 .2 3 ) . T h e n th e o u tp u ts o f th e a n a ly s is f ilte r p a ir , a f t e r d e c im a tio n b y a fa c to r
X a k (m ) =
Y ^ ( l ) knh ( n ) x ( 2 m - n )
f l = OC 1 OO
* = 0 ,1
=
i'= 0
( - 1 )k<l2J+l)h { 2 l + l) x ( 2 m - 2 1 - i)
N \
l= oc h ( 2 i) x ( 2 m - H ) + ( - 1 ) *
N -1
J2
1=0
Y i^ (2 !
1 =0
+ l) x ( 2 m - 2 1 - 1 )
( 1 0 .9 .4 6 )
N o w le t u s d e fin e th e im p u ls e r e s p o n s e o f tw o p o ly p h a s e filte r s o f le n g th N ( 2 as P i i m ) = h {2m + i) T h e n ( 1 0 .9 .4 6 ) c a n b e e x p r e s s e d as

N/2-
i = 0 ,1
( 1 0 .9 .4 7 )
X ak{m) =
Y2
1=0
P o (m )x (2 (m - /))
( 1 0 .9 .4 8 )
AT/2-1
+ ( 1 )*
1=0
P \{ m ) x ( 2 m 2 1 1 )
Jt = 0 , 1
T h is e x p r e s s io n c o r r e s p o n d s t o th e p o ly p h a s e f ilt e r s tr u c tu r e f o r th e a n a ly s is s e c t i o n sh o w n in F ig . 1 0 .4 3 . N o t e t h a t t h e c o m m u t a t o r r o t a t e s c o u n te r c lo c k w is e
Sec. 10.9
841
A nalysis____________ section
F ig u re 10.43
11
___________ Synthesis section
Polyphase filter structure for the Q M F bank.
a n d th a t th e filte r w ith im p u ls e r e s p o n s e (p o (ro )} p r o c e s s e s th e e v e n - n u m b e r e d s a m p le s o f th e in p u t s e q u e n c e a n d th e f ilte r w ith im p u ls e r e s p o n s e {/ ?i(m )} p r o c e s s e s th e o d d -n u m b e r e d s a m p le s o f th e in p u t s ig n a l. In a s im ila r m a n n e r , b y u sin g ( 1 0 .9 .2 6 ) , w e c a n o b ta in th e s tr u c tu r e f o r th e p o ly p h a s e s y n th e s is s e c tio n , w h ich is a ls o sh o w n in F ig . 1 0 .4 3 . T h is d e r iv a tio n is le ft a s an e x e r c is e fo r th e r e a d e r (P r o b le m 1 0 .1 6 ) , N o te t h a t th e c o m m u t a t o r a ls o r o ta te s c o u n te r c lo c k w is e . F in a lly , w e o b s e r v e th a t th e p o ly p h a s e filte r s tr u c tu r e s h o w n in F ig . 1 0 ,4 3 is a p p r o x im a te ly fo u r tim e s m o r e e f fic ie n t th a n th e d ir e c t -fo r m F I R f ilte r r e a liz a tio n .
10.9.7 Transmultiplexers
A n o t h e r a p p lic a t io n o f m u ltir a te s ig n a l p r o c e s s in g is in t h e d e s ig n a n d im p le m e n t a tio n o f d ig ita l tr a n s m u lt ip le x e r s w h ic h a r e d e v ic e s f o r c o n v e r tin g b e tw e e n tim e d iv is io n -m u ltip le x e d s ig n a ls . In a tr a n s m u lt ip le x e r fo r T D M - t o - F D M c o n v e r s io n , t h e in p u t s ig n a l {jr(/?)} is a tim e -d iv is io n m u ltip le x e d s ig n a l c o n s is tin g o f L s ig n a ls , w h ich a r e s e p a r a te d b y a c o m m u t a t o r s w itc h . E a c h o f th e s e L s ig n a ls a r e th e n m o d u la te d o n d if f e r e n t c a r r i e r f r e q u e n c ie s t o o b t a in an F D M s ig n a l f o r tr a n s m is s io n . In a t r a n s m u lt ip le x e r fo r F D M - t o - T D M c o n v e r s io n , th e c o m p o s ite s ig n a l is s e p a r a te d b y filte r in g in to t h e L s ig n a l c o m p o n e n ts w h ic h a r e th e n tim e -d iv is io n m u ltip le x e d . I n te le p h o n y , s in g le - s id e b a n d tr a n s m is s io n is u s e d w ith c h a n n e ls s p a c e d a t a n o m in a l 4 - k H z b a n d w id th . T w e lv e c h a n n e ls a r e u s u a lly s ta c k e d in f r e q u e n c y L a r g e r b a n d w id th t o f o r m a b a s ic g ro u p c h a n n e l, w ith a b a n d w id th o f 4 8 k H z . ( T D M ) s ig n a ls a n d f r e q u e n c y - d iv is io n - m u ltip le x e d (F D M )
F D M s ig n a ls a r e f o r m e d b y f r e q u e n c y tr a n s la tio n o f m u ltip le g ro u p s in to a d ja c e n t f r e q u e n c y b a n d s . W e s h a ll c o n fin e o u r d is c u s s io n to d ig ita l t r a n s m u lt ip le x e r s f o r 1 2 - c h a n n e l F D M a n d T D M s ig n a ls . L e t us firs t c o n s id e r F D M - t o - T D M c o n v e r s io n . T h e a n a lo g F D M s ig n a l is p a s s e d th r o u g h an A / D c o n v e r t e r a s sh o w n in F ig . 1 0 .4 4 a . T h e d ig ita l s ig n a l is th e n
842
Chap. 10
TD M
(a)
(b) Figure 10.44 Block diagram of FDM-to-TDM transmultiplexer.

d e m o d u la te d t o b a s e b a n d b y m e a n s o f s in g le -s id e b a n d d e m o d u la to r s . T h e o u tp u t o f e a c h d e m o d u la to r is d e c im a te d a n d fe d to c o m m u t a t o r o f t h e T D M s y s te m . T o b e s p e c ific , le t us a s s u m e t h a t th e 1 2 -c h a n n e l F D M s ig n a l is s a m p le d a t th e N y q u is t r a t e o f 9 6 k H z a n d p a s s e d th r o u g h a f ilte r - b a n k d e m o d u la to r . The b a s ic b u ild in g b lo c k in th e F D M d e m o d u la to r c o n s is ts o f a f r e q u e n c y c o n v e r te r , a lo w p a s s filte r , a n d a d e c im a to r , a s illu s tr a te d in F ig . 1 0 .4 4 b . F r e q u e n c y c o n v e r s io n ca n b e e ff ic ie n tly im p le m e n te d b y th e D F T f ilte r b a n k d e s c r ib e d p r e v io u s ly . T h e lo w p a s s filte r a n d d e c im a to r a r e e ff ic ie n tly i m p le m e n te d b y u s e o f th e p o ly p h a s e filte r s tr u c tu r e . T h u s th e b a s ic s tr u c tu r e f o r th e F D M - t o - T D M c o n v e r t e r h a s th e f o r m o f a D F T f ilte r b a n k a n a ly z e r . S in c e th e s ig n a l in e a c h c h a n n e l o c c u p ie s a 4 - k H z b a n d w id th , its N y q u is t r a t e is 8 k H z , a n d h e n c e th e p o ly p h a s e f ilte r o u t p u t c a n b e d e c im a te d b y a f a c t o r o f 1 2 . C o n s e q u e n tly , th e T D M c o m m u t a t o r is o p e r a tin g a t a r a t e o f 1 2 x 8 k H z o r 9 6 k H z . In T D M - t o - F D M c o n v e r s io n , th e 1 2 -c h a n n e l T D M s ig n a l is d e m u ltip le x e d in to th e 1 2 in d iv id u a l s ig n a ls , w h e r e e a c h s ig n a l h a s a r a t e o f 8 k H z . T h e sig n a l
Sec. 10.9
843
Figure 10.45 Block diagram of TDM-to-FDM transmultiplexer.

in e a c h c h a n n e l is in te r p o la te d b y a fa c t o r o f 1 2 a n d -fre q u e n c y c o n v e r te d b y a s in g le -s id e b a n d m o d u la to r , as s h o w n in F ig . 1 0 .4 5 . T h e s ig n a l o u tp u ts fro m th e 1 2 s in g le -s id e b a n d m o d u la to r s a r e s u m m e d an d fe d to th e D / A c o n v e r te r . T h u s w e o b ta in th e a n a lo g F D M s ig n a l fo r tr a n s m is s io n . A s in th e c a s e o f F D M - t o - T D M c o n v e r s io n , th e in t e r p o la t o r an d th e m o d u la to r f ilte r a r e c o m b in e d a n d e f f ic ie n tly im p le m e n te d b y u se o f a p o ly p h a s e filte r . T h e f r e q u e n c y tr a n s la tio n c a n b e a c c o m p lis h e d b y th e D F T . C o n s e q u e n tly , th e T D M - t o - F D M c o n v e r t e r e n c o m p a s s e s th e b a s ic p r in c ip le s in tr o d u c e d p r e v io u s ly in o u r d is c u s s io n o f D F T f ilte r b a n k s y n th e s is .
10.9.8 Oversampling A/D and D/A Conversion

O u r tr e a t m e n t o f o v e r s a m p lin g A / D a n d D / A c o n v e r te r s in C h a p te r 9 p r o v id e s a n o t h e r e x a m p le o f m u ltir a te s ig n a l p r o c e s s in g . R e c a l l t h a t a n o v e r s a m p lin g A /D c o n v e r t e r is im p le m e n te d b y a c a s c a d e o f a n a n a lo g s ig m a - d e lta m o d u la to r ( S D M ) fo llo w e d b y a d ig ita l a n tia lia s in g d e c im a tio n filte r a n d a d ig ita l h ig h p a s s f ilte r a s s h o w n in F ig . 1 0 .4 6 . T h e a n a lo g S D M p r o d u c e s a 1 -b it p e r s a m p le o u tp u t a t a v e r y h ig h s a m p lin g r a t e . T h is 1 -b it p e r s a m p le o u tp u t is p a s s e d th ro u g h a d ig ita l lo w p a ss f ilte r , w h ic h p r o v id e s a h ig h -p r e c is io n ( m u ltip le -b it) o u tp u t th a t is d e c im a te d to a lo w e r s a m p lin g r a t e . T h is o u tp u t is th e n p a s s e d to a d ig ita l h ig h p a s s f ilte r th a t s e r v e s to a t te n u a te th e q u a n tiz a tio n n o is e a t th e lo w e r fr e q u e n c ie s . T h e r e v e r s e o p e r a tio n s t a k e p la c e in a n o v e r s a m p lin g D / A c o n v e r te r , a s s h o w n in F ig . 1 0 .4 7 . A s illu s tr a te d in th is fig u re , th e d ig ita l s ig n a l is p a s s e d th ro u g h a h ig h p a s s f ilt e r w h o s e o u tp u t is fe d t o a d ig ita l in t e r p o l a t o r (u p s a m p le r a n d a n t i im a g in g f ilte r ) . T h is h ig h -s a m p lin g -r a te s ig n a l is th e in p u t t o th e d ig ita l S D M th a t p r o v id e s a h ig h -s a m p lin g -r a te 1 -b it p e r s a m p le o u tp u t. T h e 1 - b it p e r s a m p le o u tp u t
FigDre 10.46 Diagram of oversampling A/D converter
844
Chap. 10
precision digital signal
per sam ple
Figure 10.47 Diagram of oversampling D/A converter

is th e n c o n v e r te d to a n a n a lo g s ig n a l b y lo w p a s s filte r in g a n d f u r t h e r s m o o th in g w ith a n a lo g filte r s . F ig u r e 1 0 .4 8 illu s tr a te s t h e b lo c k d ia g ra m o f a c o m m e r c ia l ( A n a l o g D e v ic e s A D S P - 2 8 m s p 0 2 ) c o d e c ( e n c o d e r a n d d e c o d e r ) f o r v o ic e -b a n d s ig n a ls b a s e d o n s ig m a -d e lta A / D a n d D / A c o n v e r t e r s a n d a n a lo g fr o n t- e n d c ir c u its n e e d e d a s a n in t e r f a c e to th e a n a lo g v o ic e -b a n d s ig n a ls . T h e n o m in a l s a m p lin g r a t e ( a f t e r d e c i m a tio n ) is 8 k H z a n d th e s a m p lin g r a t e o f th e S D M is 1 M H z . T h e c o d e c h a s a 6 5 - d B S N R a n d h a r m o n ic d is to r tio n p e r f o r m a n c e .

T h e n e e d f o r s a m p lin g r a t e c o n v e r s io n a r is e s fr e q u e n t ly in d ig ita l s ig n a l p r o c e s s in g a p p lic a tio n s . I n th is c h a p t e r w e firs t t r e a t e d s a m p lin g r a t e r e d u c tio n ( d e c im a tio n ) a n d s a m p lin g r a t e in c r e a s e ( in t e r p o l a t i o n ) b y in t e g e r fa c to r s a n d th e n d e m o n s tr a t e d h o w th e tw o p r o c e s s e s c a n b e c o m b in e d t o o b t a i n s a m p lin g r a t e c o n v e r s io n b y a n y r a t io n a l f a c t o r . L a t e r , in S e c t i o n 1 0 .8 , w e d e s c r ib e d a m e t h o d to a c h ie v e s a m p lin g r a t e c o n v e r s io n b y a n a r b itr a r y f a c t o r . I n g e n e r a l, th e im p le m e n ta tio n o f s a m p lin g r a t e c o n v e r s io n r e q u ir e s t h e u s e o f a lin e a r tim e - v a r ia n t filte r . W e d e s c r ib e d m e t h o d s f o r im p le m e n tin g s u c h filte r s , in c lu d in g th e c la s s o f p o ly p h a s e f ilte r s tr u c tu r e s , w h ic h a r e e s p e c ia lly s im p le to im p le m e n t. W e a ls o d e s c r ib e d th e u s e o f m u ltis ta g e im p le m e n ta tio n s o f m u ltir a te c o n v e r s io n as a m e a n s o f s im p lify in g th e c o m p le x ity o f th e f ilte r r e q u ir e d t o m e e t t h e s p e c ific a tio n s . In th e s p e c ia l c a s e w h e r e t h e s ig n a l t o b e r e s a m p le d is a b a n d p a s s s ig n a l, w e d e s c r ib e d tw o m e t h o d s f o r p e r f o r m in g th e s a m p lin g r a t e c o n v e r s io n , o n e o f w h ich in v o lv e s fr e q u e n c y c o n v e r s io n , w h ile th e s e c o n d is a d ir e c t c o n v e r s io n m e t h o d th a t d o e s n o t e m p lo y m o d u la tio n . F in a lly , w e d e s c r ib e d a n u m b e r o f a p p lic a t io n s t h a t e m p lo y m u ltir a te s ig n a l p r o c e s s in g , in c lu d in g th e im p le m e n ta tio n o f n a r r o w b a n d filte r s , p h a s e s h ifte r s , fil t e r b a n k s , s u b b a n d s p e e c h c o d e r s , a n d tr a n s m u lt ip le x e r s . T h e s e a r e ju s t a fe w o f th e m a n y a p p lic a t io n s e n c o u n t e r e d in p r a c t ic e w h e r e m u ltir a te s ig n a l p r o c e s s in g is u se d . T h e firs t c o m p r e h e n s iv e t r e a t m e n t o f m u ltir a te s ig n a l p r o c e s s in g w a s g iv e n in th e b o o k b y C r o c h i e r e a n d R a b in e r ( 1 9 8 3 ) . In t h e te c h n ic a l l i t e r a t u r e , w e c ite th e p a p e r s b y S c h a f e r a n d R a b i n e r ( 1 9 7 3 ) , a n d C r o c h i e r e a n d R a b i n e r ( 1 9 7 5 , 1 9 7 6 , 1 9 8 1 , 1 9 8 3 ) . T h e u s e o f i n te r p o la tio n m e t h o d s t o a c h i e v e s a m p lin g r a t e c o n v e r s io n
Serial
output
D igital control select Sclk
Serial input
Figure 10.48 Diagram of Analog Devices ADSP-28 codec
845
846
Chap. 10
b y a n a r b itr a r y f a c t o r is t r e a t e d in a p a p e r b y R a m s t a d ( 1 9 8 4 ) . A th o r o u g h tu to r ia l tr e a t m e n t o f m u ltir a te d ig ita l filte r s a n d f ilt e r b a n k s , in c lu d in g q u a d r a tu r e m ir r o r filte r s , is g iv e n b y V e t t e r l i ( 1 9 8 7 ) , a n d b y V a id y a n a t h a n ( 1 9 9 0 , 1 9 9 3 ) , w h e r e m a n y r e f e r e n c e s o n v a r io u s a p p lic a tio n s a r e c ite d . A c o m p r e h e n s iv e s u r v e y o f d ig ita l tr a n s m u lt ip le x in g m e t h o d s is fo u n d in t h e p a p e r b y S c h e u e r m a n n a n d G o c k l e r ( 1 9 8 1 ) . S u b b a n d c o d in g o f s p e e c h h a s b e e n c o n s id e r e d in m a n y p u b lic a tio n s . T h e p io n e e r in g w o r k o n th is to p ic w a s d o n e b y C r o c h i e r e ( 1 9 7 7 , 1 9 8 1 ) a n d b y G a r la n d a n d E s t e b a n ( 1 9 8 0 ) . S u b b a n d c o d in g h a s a ls o b e e n a p p lie d t o c o d in g o f im a g e s . W e m e n tio n th e p a p e r s b y V e t t e r l i ( 1 9 8 4 ) , W o o d s a n d O N e il ( 1 9 8 6 ) , S m ith a n d E d d in s ( 1 9 8 8 ) , a n d S a f r a n e k e t a l. ( 1 9 8 8 ) a s ju s t a fe w e x a m p le s . I n c lo s in g , w e w ish t o e m p h a s iz e th a t m u ltir a te s ig n a l p r o c e s s in g c o n tin u e s t o b e a v e r y a c tiv e re s e a r c h a re a .
PROBLEMS
10.1 An analog signal x a(t) is bandlimited to the range 90 0 < F < 1 1 0 0 Hz. It is used as an input to the system shown in Fig. P10.1. In this system, H(&) is an ideal lowpass filter with cutoff frequency Fc = 125Hz.
jr = 2500
Tx
= 1 = 250
Ty
Fignre P10.1 (a) Determine and sketch the spectra for the signals x(n), w(n), v(n), and y(n).
(b) Show that it is possible to obtain y(n) by sampling x*(t) with period T = 4
milliseconds. 10.2 Consider the signal x(n) = a*u(rt), \a\ < 1. (a) Determine the spectrum X(w). (b) The signal x(n) is applied to a decimator that reduces the rate by a factor of 2. Determine the output spectrum. (c ) Show that the spectrum in part (b) is simply the Fourier transform of x(2n). 10-3 The sequence x(n) is obtained by sampling an analog signal with period T. From this signal a new signal is derived having the sampling period T/2 by use of a linear interpolation method described by the equation x{n j 2), n even
Chap. 10
Problems
847
(a) Show that this linear interpolation scheme can be realized by basic digital signal processing elements. (b) Determine the spectrum of y(n) when the spectrum of x(n) is W f 1, 1 0, 0 < M < 0.2* otherwise
(c) Determine the spectrum of y(.n) when the spectrum of x(n) is
* ( ,) = ! ''
10,
0.7* < M <0.9*

otherwise
10.4 Consider a signal x(n) with Fourier transform X (co) = 0 for w, < |o| < n
fm < I/I < \
(a) Show that the signal x( n ) can be recovered from its samples x(mD) if the sampling frequency ws = 2tt/ D 2uim( f s = 1/ D > 2f m). (b) Show that x(n) can be reconstructed using the formula
O C
x(n) =
w here
x( k D) h r(n kD)
hr(n) = --------2n n
sin (2^/ rn)
fm CO m < (t)c < Cl)J co
(c) Show that the bandlimited interpolation in part (b ) can be thought as a two-step process: First, increasing the sampling rate by a factor of D by inserting (> 1) zero samples between successive samples of the decimated signal xa(rt) = x(nD) and second, filtering the resulting signal using an ideal lowpass filter with cutoff frequency cuc. 10.5 In this problem we illustrate the concepts of sampling and decimation for discretetime signals. To this end consider a signal x ( n ) with Fourier transform X (w) as in Fig. P10.5. X(o>)
L m
...
- 2 - 1 0 1 2
Figure P10.5 (a) Sampling x(n) with a sampling period D = 2 results to the signal , , [ x(n), 'W - l o ,
n = 0 , 2 , 4 , . ..
n = 1 , 3, 5 . . . .
Compute and sketch the signal x,(n) and its Fourier transform X s(w). Can we reconstruct jc(n) from xt (n)? How?
848
Chap. 10
(b)
D ecim atin g x(n) by a facto r o f D = 2 produ ces the signal
x<i(n) = x(2n)
all
Show th at X d{a)) = X x(a>/2). P lo t the signal x j( n ) and its transfo rm X d(cj). D o we lose any in form ation w hen we d ecim ate th e sam pled signal xr (n)? 1 0 .6 * D esign a d ecim ato r th at dow nsam ples an input signal x(n) by a fa cto r D = 5. U se th e R e m ez algorithm to d eterm in e th e co efficien ts o f the F I R filter that has 0.1 -d B ripple in th e passband (0 < w < 7t/5) and is dow n by at least 3 0 d B in th e stopband. A lso determ in e the correspond ing polyphase filter stru ctu re fo r im plem enting th e decim ator. 1 0 .7 * D esig n an in terp o lato r th at in creases th e input sam pling rate by a fa cto r o f / = 2. U se the R e m ez algorithm to determ in e th e coefficien ts o f the F I R filter th at has a 0.1-d B ripple in th e passband (0 < w < n /2 ) and is down by at least 3 0 d B in the stopband. A iso , determ in e the correspond ing polyphase filter stru ctu re fo r im plem enting the in terp olato r. 1 0 .8 * D esign a sam p le-rate co n v e rte r th at redu ces the sam pling rate by a fa cto r \ . U se the R e m ez algorithm to d eterm in e th e co efficien ts o f the F I R filter th at has a 0 .1 -d B ripple in the passband and is dow n by at least 30 d B in th e stopband. Specify the sets o f tim e-v ariant co efficien ts g (n ,m ) and th e co rresp o nd ing co efficien ts in the polyphase filter realization o f the sam p le-rate co n v erter. 10.9 C onsid er the two d ifferen t ways o f cascad ing a d ecim ato r with an in terp o la to r shown in Fig. P I 0.9.
x(n)
>' i() (a)
x(n)
ID (b)
17
---------------- -
y2(n)
Figure P10.9
(a) (b)
If D = I , show th at th e outpu ts o f th e tw o con figu rations are different. H en ce, in g en eral, the two system s are n o t identical. Show th at th e tw o system s are id en tical if and only if D and I a re relatively prim e.
10.10 Prove th e eq u iv alen ce o f th e two d ecim ato r and in terp o lato r con fig u rations shown in Fig. P 10.10. T h e se equivalen t relation s are called the n ob le id en tities (see V aid y anathan , 1990). 10.11 C onsid er an arb itrary digital filter with tran sfer fun ction O O
H (z) =
n*-00
h (n )z~n
(a)
P erfo rm a tw o-com p on en t polyphase d eco m p osition o f H ( z ) by grouping the even -nu m bered sam ples Ao(n) = h(2n ) and th e odd-n um bered sam ples h \(n ) =
Chap. 10
Problems
849
(a)
x(n)
(b)
Figure P1G.10
h(2n + 1). T hu s show that H ( z ) can b e exp ressed as
H ( i ) = H o ( z 2) + r 1H ,(zJ ) and d eterm in e N 0(z) and N ,(~).
(b)
G en era liz e the result in part (a ) by showing that H ( z ) can be decom p osed in to an >-com ponent polyphase filter structure with tran sfer function
D ete rm in e //*(;). (c) F o r th e I I R filter with transfer function
H {z) =
1 a z~ l
d eterm in e Ho(z) and H \ (z) fo r the tw o-com p on en t d ecom p osition . 10.12 D esign a tw o-stage d ecim ato r for th e follow ing specification s
Passband: T ran sitio n band: In pu t sam pling rate: R ip p le:
D = 100 0 < F < 50 5 0 < F < 55 1 0,000 H z 8i = 1 0 " ', 5 2 = 1 0 " 3
1 0 ,1 3 D esign a lin ear ph ase F I R filter th a t satisfies th e follow ing sp ecification s based on a single-stage and a tw o-stage m ultirate stru ctu re. Sam pling rate: Passband: T ran sitio n band: R ip p le: 1 0,000 H z 0 < F <60 60 < F < 65 5, = 1 0 - 1,S 2 = 1 0 - 3
1 0 .1 4 Prov e th at th e h alf-ban d filter th at satisfies (1 0 .9 .4 3 ) is alw ays odd and the even coefficien ts are zero.
850
Chap. 10
10.15 D esign o n e-stag e and tw o-stage in terp olato rs to m ee t the follow ing specification :
/
Input sam pling rate: Passband: T ran sitio n band: R ip p le:
=20
10.000 H z 0 < F < 90 90 < F < 100 a, = lO "2. ^ = lO "3
10.16 B y using (1 0 .9 .2 6 ) derive the equations correspond ing to the stru ctu re fo r the poly phase synthesis section shown in Fig. 10.43. 10 .1 7 Show th at th e transpose o f an L-stage in terp o lato r fo r increasin g th e sam pling rate by an in teg er facto r I is equivalen t to an L -stag e d ecim ato r th at d e cre a ses the sam pling rate by a facto r D = I. 1 0.18 Sk etch the polyphase filter structure fo r achieving a tim e ad v an ce o f ( k / l ) T , in a seq u en ce x(n). 10.19 Prov e th e follow ing exp ressions fo r an in terp olato r o f o rd e r /. (a) T h e im pulse response h{n) can be expressed as
where
n = 0. I . 2 1 , . . .
o therw ise
(b) H (z)
may be expressed as
( ;) =
i-i
1 0 .2 0 * Zoom -frequertcy an alysis
C o nsid er the system in Fig. P I0 .2 0 a .

jy i( n )
(a) Sketch th e spectrum o f the signal _v(n) = y*(/i) +

has the spectru m shown in Fig. P I 0.20b.
if th e input signal jc(n)
(b)
Suppose th at we are in terested in th e analysis o f the fre q u e n cies in th e band
fo 5 / fo + A / , w here /o = jt/6 and A / = tt/3. D e te rm in e th e cu to ff of a lowpass filter and the decim ation fa cto r D requ ired to re ta in th e inform ation con tain ed in this band o f frequ en cies. (c) A ssum e th at
w here p = 4 0 and /* = k/ p, Jt = 0 , 1 , , D F T o f x(n ).
p 1. C om pute and p lo t th e 1024-point
Chap. 10
Problems
851
(a)
X(a> )
(b)
Figure P10.19
(d ) R e p e a t p art (b ) for the signal x(n) given in part (c ) by using an appropriately designed low pass lin ear phase F I R filter to d eterm in e th e decim ated signal *(> =
**{")
+ j s i (n ).
( e ) C o m p u te th e 1024-p o in t D F T o f s(n ) and in v estig ate to see if you h av e obtain ed th e exp ected results.
11
Linear Prediction and Optimum Linear Filters
T h e d e s ig n o f filte r s to p e r f o r m s ig n a l e s tim a tio n is a p r o b le m
th a t f r e q u e n tly
a r is e s in th e d e s ig n o f c o m m u n ic a tio n s y s te m s , c o n t r o l s y s te m s , in g e o p h y s ic s , a n d in m a n y o th e r a p p lic a t io n s a n d d is c ip lin e s . In th is c h a p t e r w e t r e a t th e p r o b le m o f o p tim u m f ilte r d e s ig n fr o m a s ta tis tic a l v ie w p o in t. T h e filte r s a r e c o n s tr a in e d to b e lin e a r a n d th e o p tim iz a tio n c r ite r io n is b a s e d o n th e m in im iz a t io n o f th e m e a n -s q u a r e e r r o r . A s a c o n s e q u e n c e , o n ly th e s e c o n d - o r d e r s ta tis tic s ( a u t o c o r r e la tio n a n d c r o s s c o r r e la t io n f u n c t io n s ) o f a s ta tio n a r y p r o c e s s a r e r e q u ir e d in th e d e te r m in a tio n o f th e o p tim u m filte r s . In c lu d e d in th is t r e a t m e n t is th e d e s ig n o f o p tim u m filte r s f o r lin e a r p r e d ic tio n . L i n e a r p r e d ic t io n is a p a r tic u la r ly im p o r ta n t to p ic in d ig ita l s ig n a l p r o c e s s in g , w ith a p p lic a t io n s in a v a r ie ty o f a r e a s , s u c h as s p e e c h s ig n a l p r o c e s s in g , im a g e p r o c e s s in g , a n d n o is e s u p p r e s s io n in c o m m u n ic a tio n s y s te m s . A s w e s h a ll o b s e r v e , d e te r m in a tio n o f t h e o p tim u m lin e a r f ilt e r f o r p r e d ic t io n r e q u ir e s th e s o lu tio n o f a s e t o f lin e a r e q u a tio n s th a t h a v e s o m e s p e c ia l s y m m e tr y . T o s o lv e th e s e lin e a r e q u a tio n s , w e d e s c r ib e tw o a lg o r ith m s , t h e L e v in s o n - D u r b in a lg o r it h m a n d th e S c h u r a lg o r ith m , w h ich p r o v id e th e s o lu tio n to th e e q u a tio n s th r o u g h c o m p u ta tio n a lly e ff ic ie n t p r o c e d u r e s t h a t e x p lo it th e s y m m e tr y p r o p e r tie s . T h e la s t s e c tio n o f th e c h a p t e r t r e a t s a n im p o r ta n t c la s s o f o p tim u m filte rs c a lle d W i e n e r filte r s . W ie n e r filte r s a r e w id e ly u se d in m a n y a p p lic a t io n s in v o lv in g th e e s tim a tio n o f s ig n a ls c o r r u p te d w ith a d d itiv e n o is e .
11.1 INNOVATIONS REPRESENTATION OF A STATIONARY RANDOM PROCESS

In th is s e c tio n w e d e m o n s tr a t e t h a t a w id e -s e n s e s ta tio n a r y r a n d o m p r o c e s s c a n b e r e p r e s e n te d a s th e o u tp u t o f a c a u s a l a n d c a u s a lly in v e r t ib le l i n e a r s y s te m e x c ite d b y a w h ite n o is e p r o c e s s . T h e c o n d itio n t h a t th e s y s t e m is c a u s a lly i n v e r t ib le a ls o a llo w s us to r e p r e s e n t t h e w id e -s e n s e s ta tio n a r y r a n d o m p r o c e s s b y th e o u tp u t o f th e in v e r s e s y s te m , w h ic h is a w h ite n o i s e p r o c e s s .
852
Sec. 11.1
Innovations Representation of a Stationary Random Process
853
L e t us c o n s id e r a w id e -s e n s e s ta tio n a r y p r o c e s s {jc ( ) } w ith a u t o c o r r e la t io n seq u en ce
{yxx(m)}
a n d p o w e r s p e c tr a l d e n s ity r ( / ) , \ f \ <
W e a s s u m e th a t
!%.<(/) is r e a l a n d c o n tin u o u s f o r a ll |/| < j . T h e z -tr a n s fo r m o f th e a u t o c o r r e l a tio n s e q u e n c e {y xx(m )} is ^ (
2) =
( 1 1 .1 .1 )
fr o m w h ic h w e o b ta in th e p o w e r s p e c tr a l d e n s ity b y e v a lu a tin g 1 % ,^ ) o n th e u n it c ir c le [i.e . b y s u b s t itu tin g z = e x p ( y 2 jr / ) ] . N o w , l e t u s a s s u m e th a t lo g I V , (z) is a n a ly tic (p o s s e s s e s d e r iv a tiv e s o f a ll o r d e r s ) in a n a n n u la r r e g io n in th e z -p la n e th a t in c lu d e s th e u n it c ir c le ( i .e ., n <
\z\ < r 2 w h e r e n < 1 a n d r 2 > 1 ). T h e n , l o g r ^ z ) c a n b e e x p a n d e d in a L a u r e n t

s e r ie s o f th e fo r m lo g
r (z ) =
v(m )z~
( 11. 1.2 )
w h e r e th e (u (m )} a r e th e c o e f f ic ie n t s in th e s e r ie s e x p a n s io n . W e c a n v ie w {u (m )} a s th e s e q u e n c e w ith z -tr a n s fo r m V ( z ) = l o g T ( z ) . E q u iv a le n tly , w e c a n e v a lu a te lo g T J r (z ) o n th e u n it c ir c le , lo g r (/ )=
J2
v ( m ) e - J23lfm
( 1 1 .1 .3 )
s o t h a t th e { r ( m ) } a r e th e F o u r i e r c o e f f ic ie n ts in th e F o u r i e r s e r ie s e x p a n s io n o f th e p e r io d ic fu n c t io n l o g r ^ ( / ) . H e n c e
v( m) =
W e o b s e r v e th a t v {m ) =
[ \ g T xx{ f ) ] e j l f md f
m = 0 , d b l------
( 1 1 .1 .4 )
s in c e T x l ( f ) is a re a l a n d e v e n fu n c tio n o f / .
F r o m ( 1 1 .1 .2 ) it fo llo w s th a t r ^ f z ) = exp
_m= oc
^2 u t -,\ w h e r e , b y d e f in it io n , a 2 = e x p [i? (0 )] a n d
( 1 1 .1 .5 )
H ( z ) = exp
T . v(m )z~
( 11. 1.6 )
I f ( 1 1 .1 .5 ) is e v a lu a te d o n th e u n it c i r c l e , w e h a v e th e e q u iv a le n t r e p r e s e n ta tio n o f th e p o w e r s p e c tr a l d e n s ity a s
r (/) = < r2 \ H ( f ) \ 2
(11.1.7)
//* (/)
We note that
lo g Txx{ f )
= lo g a i+
O C
lo g H ( f ) + lo g
v ( m ) e - j2* fm
854
Chap. 11
F r o m th e d e fin itio n o f H ( z ) g iv e n b y ( 1 1 .1 .6 ) , it is c l e a r t h a t th e c a u s a l p a r t o f th e F o u r i e r s e r ie s in ( 1 1 .1 .3 ) is a s s o c ia t e d w ith H ( z ) a n d th e a n tic a u s a l p a r t is a s s o c ia t e d w ith H ( z ~ l ). T h e F o u r i e r s e r ie s c o e f f ic ie n t s { v ( m ) } a r e th e ce p stra l c o e fficie n ts a n d t h e s e q u e n c e { u (m ) } is c a lle d th e c e p s tr u m o f th e s e q u e n c e {y xx(m )}, a s d e fin e d in S e c t i o n 4 .2 .7 . T h e filte r w ith s y s te m fu n c t io n H ( z ) g iv e n b y ( 1 1 .1 .6 ) is a n a ly tic in t h e r e g io n | z| > r \ < 1. H e n c e , in th is r e g io n , it h a s a T a y l o r s e r ie s e x p a n s io n a s a c a u s a l s y s te m o f th e f o r m
OC
H ( z ) = J 2 h in ) z ~ n
m=0
( 1 U '8 )
T h e o u tp u t o f th is filte r in r e s p o n s e t o a w h ite n o is e in p u t s e q u e n c e w (n ) w ith p o w e r s p e c tr a l d e n s ity a 2 is a s ta tio n a r y r a n d o m p r o c e s s { * ( ) } w ith p o w e r s p e c tr a l d e n s ity r xx( f ) = o 2 \ H ( f ) \ 2 . C o n v e r s e ly , th e s ta tio n a r y r a n d o m p r o c e s s |a-(/j)} w ith p o w e r s p e c tr a l d e n s ity T xx( f ) c a n b e tr a n s f o r m e d in to a w h ite n o is e p r o c e s s b y p a s s in g (jc ( ti)} th r o u g h a lin e a r f ilte r w ith s y s te m f u n c t io n 1 / H ( z ) . W e c a ll th is f ilte r a n o is e w h ite n in g filte r. It s o u tp u t, d e n o te d as (w ( n ) } is c a lle d th e in n o
v a tio n s p ro c e s s a s s o c ia t e d w ith th e s ta tio n a r y r a n d o m p r o c e s s { * ( ) } . T h e s e tw o

r e la tio n s h ip s a r e illu s tr a te d in F ig . 1 1 .1 . T h e r e p r e s e n t a t i o n o f th e s ta tio n a r y s t o c h a s t ic p r o c e s s {.v fn )} a s t h e o u tp u t o f a n I I R filte r w ith s y s te m fu n c tio n H ( z ) g iv e n b y ( 1 1 .1 .8 ) a n d e x c ite d b y a w h ite n o is e s e q u e n c e (u ^ n )} is c a lle d th e W o ld re p re se n ta tio n .
11.1.1 Rational Power Spectra

L e t u s n o w r e s t r ic t o u r a t t e n t io n t o th e c a s e w h e r e th e p o w e r s p e c tr a l d e n s ity o f th e s ta tio n a r y r a n d o m p r o c e s s { * ( ) ) is a r a t io n a l fu n c t io n , e x p r e s s e d a s r ^ ( z ) = g u,7 7\ J
A {z)A (z~i )
r\ < \ z \ < r 2
( 1 1 .1 .9 )
w h e r e th e p o ly n o m ia ls B ( z ) a n d A ( z ) h a v e r o o t s t h a t fa ll in s id e th e u n it c ir c le in th e s - p la n e . T h e n th e l in e a r f ilt e r H ( z ) f o r g e n e r a tin g th e r a n d o m p r o c e s s U ( ) }
w(rr) W hile noise
L inear causal filter
x(n ) = Y , h(i)w(n - Jc)
k= 0
H{z)
(a)
-r(n)
L inear causal filter VH(z)
w(n) W hite noise
(b)
Figure 11.1 Filters for generating (a) the random process Jt(n) from white noise and (b) the inverse filter.
Sec. 11.1
Innovations Representation of a Stationary Random Process
855
fr o m th e w h ite n o is e s e q u e n c e {iu (n )} is a ls o r a t io n a l a n d is e x p r e s s e d a s
(11.1 .10)
w h e r e {bk } a n d [ak } a r e th e filte r c o e f f ic ie n ts t h a t d e te r m in e th e lo c a t io n o f th e z e r o s a n d p o le s o f H ( z ) , re s p e c tiv e ly . T h u s H { z ) is c a u s a l, s ta b le , a n d m in im u m p h ase. s y s te m . I t s r e c ip r o c a l 1 / H { z ) is a ls o a c a u s a l, s ta b le , a n d m in im u m - p h a s e lin e a r T h e r e f o r e , t h e ra n d o m p r o c e s s {;c (n )} u n iq u e ly r e p r e s e n t s t h e s ta tis tic a l
p r o p e r tie s o f t h e in n o v a tio n s p r o c e s s {u j(/ i)}, a n d v ic e v e r s a . F o r t h e li n e a r s y s t e m w ith th e r a t io n a l s y s t e m f u n c t io n H (z ) g iv e n b y ( 1 1 .1 .1 0 ) , i h e o u tp u t x ( n ) is r e la te d to t h e in p u t u>(n) b y th e d if f e r e n c e e q u a tio n ( 1 1 .1 .1 1 ) W e w ill d is tin g u is h a m o n g t h r e e s p e c ific c a s e s .
Autoregressive (AR) process.

in p u t - o u t p u t r e la tio n s h ip is
b0 =
1. bk = 0 , k > 0 .
I n th is c a s e , t h e
lin e a r f ilte r H ( z ) = 1 / A ( z ) is a n a ll-p o le filte r a n d th e d if f e r e n c e e q u a t io n f o r th e
(11.1 . 12)
k= 1
In tu r n , th e n o is e -w h ite n in g filte r f o r g e n e r a tin g th e in n o v a tio n s p r o c e s s is a n a ll- z e r o filte r .
Moving average (MA) process.

r e la tio n s h ip is
a k = 0 , k > 1. In th is c a s e , th e l i n e a r f ilte r
H { z ) = B { z ) is a n a ll-z e r o filte r a n d th e d if f e r e n c e e q u a tio n f o r th e i n p u t- o u tp u t
( 1 1 .1 .1 3 ) T h e n o is e -w h it e n in g filte r f o r th e M A p r o c e s s is a n a ll- p o le f ilte r .
Autoregressive, moving average (ARMA) process.

c o r r e s p o n d in g d if f e r e n c e e q u a tio n is g iv e n b y ( 1 1 .1 .1 1 ) . 1 / H ( z ) = A ( z ) /B { z ) .
I n th is c a s e , th e lin
e a r f ilte r H ( z ) = B { z ) / A { z ) h a s b o th fin ite p o le s a n d z e r o s in t h e z - p la n e a n d th e T h e i n v e r s e s y s te m f o r g e n e r a tin g t h e in n o v a tio n p r o c e s s f r o m x ( n ) is a ls o a p o l e - z e r o s y s te m o f th e fo r m
11.1.2 Relationships Between the Filter Parameters and the Autocorrelation Sequence
W h e n th e p o w e r s p e c tr a l d e n s ity o f th e s ta tio n a r y r a n d o m p r o c e s s is a r a t io n a l f u n c t io n , t h e r e is a b a s ic r e la tio n s h ip b e tw e e n t h e a u t o c o r r e l a t i o n s e q u e n c e
856
Chap. 11
{XjcjcC"t)} a n d th e p a r a m e te r s {a * } a n d {>*} o f th e l i n e a r f ilte r H ( z ) th a t g e n e r a te s th e p r o c e s s b y filte r in g th e w h ite n o is e s e q u e n c e w (n ). T h is r e la tio n s h ip c a n b e o b ta in e d b y m u ltip ly in g th e d if f e r e n c e e q u a tio n in ( 1 1 .1 .1 1 ) b y x * (n m ) a n d ta lcin g th e e x p e c te d v a lu e o f b o th s id e s o f th e re s u ltin g e q u a tio n . T h u s w e h a v e
p E [ x { n ) x * { n m )] = ^ f l t [ . x ( n k )x * (n m )]
( 1 1 .1 .1 4 ) + ^T^f>i[u>(n k )x * (n m )] H en ce
p Yxx ( m ) =
1=1
< ?
~ f t + Y 2 btYwx^w ~ f t ( H . 1 . 1 5 )
i= 0
w h e r e y w.r (m) is th e c r o s s - c o r r e la t io n s e q u e n c e b e tw e e n u;(/j) a n d x ( /i) . T h e c r o s s c o r r e la tio n y wx(m ) is r e la te d to th e f ilte r im p u ls e r e s p o n s e . T h a t is,
ywxOn) = [ A * ( f i) w ( + m )]
= E
h (k )w * (n k ) w ( n -i- m )
( 1 1 .1 .1 6 )
= a lh ( -m ) w h e r e , in th e la s t s te p , w e h a v e u se d th e fa c t th a t th e s e q u e n c e w (n) is w h ite . H en ce I < 0 B y c o m b in in g ( 1 1 .1 .1 7 ) w ith ( 1 1 .1 .1 5 ) , w e o b ta in th e d e s ir e d r e la tio n s h ip
p ~ Y 2 , a tYxx{m - k ), k=\ m > q

( 1 1 .1 .1 8 )
Yxxim) =
~
- Yxx
kYx* (m ~ k ^ k=l
h tt' lbk+m , k= 0
0 <m <q m < 0
T h is r e p r e s e n ts a n o n lin e a r r e la tio n s h ip b e tw e e n yJX (m ) a n d th e p a r a m e t e r s {a * } , {**} T h e r e la tio n s h ip in ( 1 1 .1 .1 8 ) a p p lie s , in g e n e r a l, t o th e A R M A p r o c e s s . F o r a n A R p r o c e s s , ( 1 1 .1 .1 8 ) s im p lifie s to
p ~ 2 2 akY ( m ~ k ) ,
k =i
m > 0
Yxx(m) =
p - 2 2 a k Yx*(m ~ W + a l
k=i
(11.1.19)
m = 0 m <o
y A (~ m )'
Sec. 11.2
Forward and Backward Linear Prediction
857
T h u s w e h a v e a lin e a r r e la tio n s h ip b e tw e e n y xx (m ) a n d th e { a t } p a r a m e te r s . T h e s e e q u a tio n s , c a lle d th e fo rm
Y u le - W a lk e r e q u a tio n s , c a n b e e x p r e s s e d in th e m a tr ix
~Yxx(0)
y (l)
-Yxx(p)
y , , ( - 1) K(0)
YxA
p
Yxx( ~ 2 ) YxA ~ l) Y x x i p - 2)
Yxx(-P) Yxx(~P + l) - YxAO)
- 1 a\ -a p _
( 11. 1.20)
- 1)
_0 _
T h i s c o r r e l a t i o n m a tr ix is T o e p lit z , a n d h e n c e it c a n b e e ff ic ie n tly in v e r t e d b y u s e o f t h e a lg o r it h m s d e s c r ib e d in S e c t i o n 1 1 .3 . F in a lly , b y s e tt in g a k 0 , 1 < k < p , a n d h (k ) bk, 0 < k < q , in ( 1 1 .1 .1 8 ) , w e o b ta in th e r e la tio n s h ip fo r th e a u t o c o r r e la t io n s e q u e n c e in th e c a s e o f a M A p r o c e s s , n a m e ly , < ?
YxAm ) =
0,
cru , 2 2 bkbk+m' n *=<> y*x (m ),
- m - ?
m > q m < 0
( 1 1 - 1 .2 1 )
11.2 FORWARD AND BACKWARD LINEAR PREDICTION

L i n e a r p r e d ic t io n is an i m p o r ta n t to p ic in d ig ita l s ig n a l p r o c e s s in g w ith m a n y p r a c tic a l a p p lic a t io n s . tim e . In th is s e c tio n w e c o n s id e r t h e p r o b le m o f lin e a r ly p r e d ic tin g th e v a lu e o f a s ta tio n a r y r a n d o m p r o c e s s e it h e r fo r w a rd in tim e o r b a c k w a r d in T h is fo r m u la tio n le a d s t o la t t ic e filte r s tr u c tu r e s a n d t o s o m e in te r e s tin g c o n n e c t io n s t o p a r a m e tr ic s ig n a l m o d e ls .
11.2.1 Forward Linear Prediction

L e t u s b e g in w ith th e p r o b le m o f p r e d ic t in g a fu tu r e v a lu e o f a s ta tio n a r y r a n d o m p r o c e s s fr o m o b s e r v a t io n o f p a s t v a lu e s o f t h e p r o c e s s . I n p a r tic u la r , w e c o n s id e r th e o n e -ste p f o r w a r d lin e a r p r e d ic to r , w h ic h fo r m s th e p r e d ic t io n o f t h e v a lu e x ( n ) b y a w e ig h te d lin e a r c o m b in a tio n o f th e p a s t v a lu e s x ( n 1 ), x ( n - 2 ) , . . . , x ( n p ). H e n c e t h e lin e a r ly p r e d ic t e d v a lu e o f x ( n ) is
p x (n ) = ^ a p( i ) x (n k)
*= l
( 1 1 .2 .1 )
w h e r e th e { a p {k)} r e p r e s e n t t h e w e ig h ts in t h e lin e a r c o m b in a tio n . T h e s e w e ig h ts a r e c a lle d th e p r e d ic t io n c o e fficie n ts o f t h e o n e - s t e p fo r w a r d l i n e a r p r e d ic t o r o f
o r d e r p . T h e n e g a tiv e s ig n in th e d e fin it io n o f x ( n ) is f o r m a th e m a tic a l c o n v e n ie n c e

a n d c o n f o r m s w ith c u r r e n t p r a c tic e in th e te c h n ic a l lite r a tu r e .
858
Chap. 11
F ig u re 1 L 2
F o r w a r d lin e a r p re d ic tio n
T h e d if f e r e n c e b e tw e e n th e v a lu e x ( n ) a n d th e p r e d ic t e d v a lu e x ( n ) is c a lle d th e f o r w a r d p r e d ic t io n e r r o r , d e n o te d a s f p (n):
f p (n) = x ( n ) - x ( n )
p
( 11. 2 .2 )
a p (k )x (n - k)
= x (n ) + ^
W e v iew li n e a r p r e d ic tio n as e q u iv a le n t t o li n e a r filte r in g w h e r e t h e p r e d ic t o r is e m b e d d e d in th e lin e a r filte r , as s h o w n in F ig . 1 1 .2 . T h is is c a lle d a p r e d ic t io n -
e rro r filt e r w ith in p u t s e q u e n c e {jc(ai)} a n d o u tp u t s e q u e n c e { f r (n )}. A n e q u iv a le n t

r e a liz a tio n f o r th e p r e d ic t io n - e r r o r filte r is s h o w n in F ig . 1 1 .3 . T h i s r e a liz a tio n is a d ir e c t -f o r m F I R f ilte r w ith s y s te m fu n c tio n
A p (z) =
i-=0 w h e r e , b y d e f in it io n , a ^ (0 ) = 1.
( 1 1 .2 .3 )
A s sh o w n in S e c tio n 7 .2 .4 , th e d ir e c t -f o r m F I R f ilte r is e q u iv a le n t t o a n a llz e r o la ttic e f ilte r . T h e l a t t i c e filte r is g e n e r a lly d e s c r ib e d b y t h e fo llo w in g s e t o f
p( P)
1 ----/,(")
Figure 11-3 Prediction-error filter
Sec. 11.2
859
o r d e r -r e c u r s iv e e q u a tio n : fo ( n ) = go(n) - x {n ) fm (n ) = / m -i(n ) + K mg m~ i( n - 1) g m(n) = A r*/m_ i ( n ) + g m- i ( n - 1) m = l,2 ,...,p m = 1 . 2 ........... p

( 1 1 .2 .4 )
w h e r e { m) a r e th e r e f le c tio n c o e f f ic ie n ts a n d g m(n) is th e b a c k w a r d p r e d ic tio n e r r o r d e fin e d in th e fo llo w in g s e c tio n . N o te th a t f o r c o m p le x - v a lu e d d a ta , th e F ig u r e 1 1 .4 illu s tr a te s a c o n ju g a t e o f K m is u se d in th e e q u a tio n f o r g m(n). th e c o m p u ta tio n s g iv e n b y ( 1 1 .2 .4 ) . A s a c o n s e q u e n c e o f th e e q u iv a le n c e b e tw e e n th e d ir e c t -f o r m p r e d ic t io n e r r o r F I R filte r a n d th e F I R la t t ic e filte r , th e o u tp u t o f th e />-stage l a t t i c e filte r is e x p r e s s e d as
/7-stage l a t t i c e filte r in b lo c k d ia g ra m fo r m a lo n g w ith a ty p ic a l s ta g e sh o w in g
f p (n) -
a p(f:^x (n ~ f t
a p (0) = l
( 1 1 .2 .5 )
S in c e ( 1 1 .2 .5 ) is a c o n v o lu tio n su m , th e z - tr a n s fo r m r e la t io n s h ip is
F r (z) = A p ( z ) X ( z )
o r , e q u iv a le n tly .
( 11 .2 .6 )
A (z) =
'
( 1 1 .2 .7 )
X (z)
F 0(z)
T h e m e a n -s q u a r e v a lu e o f th e fo r w a rd lin e a r p r e d ic t io n e r r o r f p {n) is
f p = E [ \ f p ( n ) \2] p = > 'jj(0 ) + 2 R e a * ( l) a p( k ) y xx(l - k )
(11. 2. 8)
Figure 1L4
p -stage lattice filter
860
Chap. 11
<?/ is a q u a d r a tic fu n c t io n o f th e p r e d ic t o r c o e f f ic ie n t s a n d its m in im iz a tio n le a d s t o th e s e t o f lin e a r e q u a tio n s
p
yx A d
= - Y
*= i
, a r {k )y * * v ~ k)
l = l ' 2 .......... p
( n -2 -9 )
T h e s e a r e c a lle d th e n o r m a l e q u a tio n s f o r th e c o e f f ic ie n ts o f th e lin e a r p r e d ic t o r . T h e m in im u m m e a n - s q u a r e p r e d ic t io n e r r o r is s im p ly m in [ / ] s
E { = y ( 0 ) + Y , a p { k ) YxA - k )
*= i
( 1 1 .2 .1 0 )
In th e fo llo w in g s e c tio n w e e x t e n d th e d e v e lo p m e n t a b o v e to th e p r o b le m o f p r e d ic tin g th e v a lu e o f a tim e s e r ie s in th e o p p o s ite d ir e c t io n , n a m e ly , b a c k w a r d in tim e .
11.2.2 Backward Linear Prediction

L e t u s a s s u m e th a t w e h a v e th e d a ta s e q u e n c e x ( n ) , x ( n - 1 ) .......... x(rt p + 1) fro m a s ta tio n a r y ra n d o m p r o c e s s a n d w e w ish t o p r e d ic t th e v a lu e x ( n p ) o f th e p r o c e s s . In th is c a s e w e e m p lo y a o n e -ste p b a c k w a rd lin e a r p re d ic t o r o f o r d e r p . H e n c e
p
-i
( 11 .2 . 11 )
x ( n p ) = Y 2 bp (k )x {n k) k=0 b a c k w a r d p re d ic tio n e r r o r , d e n o te d a s g p( n ): P- \ g p (n) = x ( n - p ) + 2 2 bp (k)x (n ~ f t

p
T h e d if f e r e n c e b e tw e e n th e v a lu e x ( n p ) a n d th e e s tim a te x ( n - p ) is c a lle d th e
i=
bp (p ) = 1
( 11.2 .12)
' Y l bP ^ x ^n ~ f t
t=o
T h e b a c k w a r d lin e a r p r e d ic t o r c a n b e r e a liz e d e i t h e r b y a d ir e c t -f o r m F I R f ilte r s tr u c tu r e s im ila r t o th e s tr u c tu r e s h o w n in F ig . 1 1 .2 o r a s a la ttic e s tr u c tu r e . T h e l a t t ic e s tr u c tu r e sh o w n in F ig . 1 1 .4 p r o v id e s t h e b a c k w a r d lin e a r p r e d ic t o r as w e ll a s th e fo r w a rd lin e a r p r e d ic to r . T h e w e ig h tin g c o e ffic ie n ts in th e b a c k w a r d li n e a r p r e d i c t o r a r e th e c o m p le x c o n ju g a te s o f th e c o e f f ic ie n ts f o r th e fo r w a r d l i n e a r p r e d ic t o r , b u t th e y o c c u r in re v e rse o rd e r. T h u s w e hav e
bp(k) a * ( p k )
k = 0 ,l,...,p
( 1 1 .2 .1 3 )
In th e z -d o m a in , th e c o n v o lu tio n s u m in ( 1 1 .2 .1 2 ) b e c o m e s
G p(z) = Bp (z)X (z)
(11.2.14)
or, equivalently,
G p (z) G p (z)
Sec. 11.2
861
filte r w ith c o e f f ic ie n ts
w h e r e B p {z) r e p r e s e n ts th e s y s te m fu n c t io n o f th e F I R
bP (k).
S in c e bp (k) = a * ( p k ), G r (z) is r e la te d to A r (z)
B p (z) =
= X > t=o
>
-*)*"* ( 1 1 .2 .1 6 )
= z - ' fl; ( * ) z * t=0 = Z-p A*p (z - ') T h e r e la tio n s h ip in ( 1 1 .2 .1 6 ) im p lie s th a t th e z e r o s o f th e F I R f ilte r w ith s y s te m fu n c t io n B p(z) a r e s im p ly th e ( c o n ju g a t e ) r e c ip r o c a ls o f th e z e r o s o f A ,,(z ). H e n c e
B p (z) is c a lle d th e r e c ip r o c a l o r re v e rse p o ly n o m ia l o f A p (z).

N o w th a t w e h a v e e s ta b lis h e d t h e s e in te r e s tin g r e la tio n s h ip s b e tw e e n th e f o r w a r d a n d b a c k w a r d o n e - s te p p r e d ic t o r s , le t us re tu r n to th e re c u r s iv e l a t t ic e e q u a tio n s in ( 1 1 .2 .4 ) a n d t r a n s fo r m th e m to th e z -d o m a in . T h u s w e h a v e
F 0(z) = G q ( z ) = X ( z ) F m(z) = F m- i ( z ) + /rm; - ' G ffl_ , W

G m(z ) =
m = 1 ,2 ,..., p m = 1, 2 , . . . , p
( 1 1 .2 .1 7 )
K " F m- i { z ) + z _ 1 G m_ i ( ; )
I f w e d iv id e e a c h e q u a tio n b y X ( z ) , w e o b ta in th e d e s ir e d r e s u lts in th e f o r m 4 0 (z ) = 5 o (z ) = 1
A m{z) = A m^ ( z ) + K mz ~ l B m^ { z ) B m(z) = K*mA m^ ( z ) + z ' 1 B m- i ( z )
m = 1 , 2 .......... p
in = 1 , 2 .......... p
( 1 1 .2 .1 8 )
T h u s a la t t i c e filte r is d e s c r ib e d in th e z -d o m a in b y th e m a tr ix e q u a tio n
A m( z ) ' _ B m(z) _
'1
K mZ
A m- i ( z ) B m -l( z )
( 1 1 .2 .1 9 )
T h e r e la t io n s in ( 1 1 .2 .1 7 ) f o r A m(z) a n d B m(z) a llo w u s t o o b ta in t h e d ir e c t -f o r m F I R f ilte r c o e f f ic ie n ts (a m(A)] fr o m t h e r e f le c tio n c o e f f ic ie n t s ( K m), a n d v ic e v e r s a . T h e s e r e la tio n s h ip s w e r e g iv e n in S e c t io n 7 .2 .4 b y ( 7 .2 .5 1 ) th r o u g h ( 7 .2 .5 3 ) . T h e l a t t i c e s t r u c tu r e w ith p a r a m e t e r s K i , K 2........ K p c o r r e s p o n d s t o a c la s s o f p d ir e c t -f o r m F I R f ilte r s w ith s y s te m f u n c t io n s A i ( z ), A i ( z ) , . . , A p (z). I t is in te r e s tin g t o n o t e t h a t a c h a r a c te r iz a tio n o f th is c la s s o f p F I R filte r s in d ir e c t f o r m r e q u ir e s p ( p + l ) / 2 f ilte r c o e f f ic ie n ts . I n c o n t r a s t , th e l a t t i c e - f o r m c h a r a c te r iz a tio n r e q u ir e s o n ly t h e p r e f le c t i o n c o e f f ic ie n t s T h e r e a s o n th e l a t t i c e p r o v id e s a m o r e c o m p a c t r e p r e s e n t a t io n f o r t h e c la s s o f p F I R filte r s is b e c a u s e a p p e n d in g
862
Chap. 11
s ta g e s to th e la t t ic e d o e s n o t a l t e r th e p a r a m e t e r s o f th e p r e v io u s s ta g e s . O n th e o th e r h a n d , a p p e n d in g th e /?th s ta g e t o a la t t ic e w ith (p 1 ) s t a g e s is e q u iv a le n t t o in c r e a s in g th e le n g th o f a n F I R filte r b y o n e c o e f f ic ie n t . T h e r e s u ltin g F I R f ilte r w ith s y s te m fu n c t io n A p (z) h a s c o e f f ic ie n ts t o ta lly d if f e r e n t f r o m th e c o e f f ic ie n ts o f th e lo w e r - o r d e r F I R f ilte r w ith s y s te m f u n c tio n T h e fo r m u la f o r d e te r m in in g th e f ilt e r c o e f f ic ie n ts {ap (k ) \ r e c u r s iv e ly is e a s ily d e riv e d fr o m p o ly n o m ia l r e la tio n s h ip s ( 1 1 .2 .1 8 ) . W e h a v e
A m(z) = m y ' a m(k )z k *=0
- f K mz ~ ] B m-
1( ' )
-i ( 1 1 .2 .2 0 ) i= 0 a n d r e c a llin g th a t a m( 0 ) = 1 1 - k)z a+ 1)
m -\
Y ' a m_ i ( i ) c A=0 * + K m
B y e q u a tin g th e c o e f f ic ie n t s o f e q u a l p o w e r s o f fo r m = c o e f f ic ie n ts in th e fo r m Am(0 ) = 1
1, 2 , ___ p , w e o b ta in th e d e s ir e d r e c u r s iv e e q u a tio n f o r th e F I R filte r
a m(m) = K m a,(k) = a m~ \{ k ) + K ma*n_ } (m - k)

= a m~ \( k ) + a m (m )a* ( 1 1 .2 .2 1 ) 1 < k < m 1
(m - k)
m = 1 .2 ,.... p
T h e c o n v e r s io n fo r m u la f r o m th e d ir e c t -fo r m F I R filte r c o e f f i c i e n t s (a^(/:)} to th e l a t t ic e r e fle c tio n c o e f f ic ie n t s |Af, } is a ls o v e r y s im p le . F o r th e />-s ta g e la ttic e w e im m e d ia te ly o b ta in th e r e fle c tio n c o e f f ic ie n t K p = a n (p ). T o o b t a i n K p~ i , w e n e e d th e p o ly n o m ia ls A m(z) f o r m = p 1 . ___ 1. F r o m ( 1 1 . 2 . 1 9 ) w e o b ta in
A m- i( z ) = ------ ------
1
. x
A m(z) K mB m(z)
~ 2
------
m = p ........1
(11.2.22)
... ,
I& m I
w h ich is ju s t a s te p -d o w n r e c u r s io n . T h u s w e c o m p u te a ll lo w e r - d e g r e e p o ly n o m ia ls f r o m t h e r e la tio n K m = as | A T mi ^ 1 fo r m =
A m(z) b e g in n in g w ith A p- i ( z ) a n d o b ta in t h e .d e s ir e d l a t t ic e r e f l e c t i o n c o e f f ic ie n ts a m(m ). W e o b s e r v e th a t th e p r o c e d u r e w o rk s a s lo n g

1, 2 , . . . , p 1. F ro m th is s te p -d o w n r e c u r s io n f o r th e
p o ly n o m ia ls , it is r e la tiv e ly e a s y to o b t a in a fo r m u la f o r r e c u r s iv e ly a n d d ir e c tly co m p u tin g K m, m = p 1 , . . . , 1. F o r m p - 1 , . . . , 1, w e h a v e
K m = a m(m )
/r
Qm(k) K mbm(k) = _
l - l
i t
iU 2 2 3 )
a m(k) - a m{m)a*m(m - k)
1 -
\am(m )\2
w h ic h is ju s t t h e r e c u r s io n in th e S c h u r - C o h n s ta b ility t e s t f o r t h e p o ly n o m ia l
A m(z).
Sec. 11.2
863
A s ju s t in d ic a te d , th e r e c u r s iv e e q u a tio n in ( 1 1 .2 .2 3 ) b r e a k s d o w n i f a n y o f t h e la t t ic e p a r a m e t e r s \ K m \ = 1. I f th is o c c u r s , it is in d ic a tiv e th a t th e p o ly n o m ia l Am_ i ( z ) h a s a r o o t l o c a t e d o n th e u n it c ir c le . S u c h a r o o t c a n b e f a c t o r e d o u t f r o m A m_ i ( z ) a n d th e ite r a tiv e p r o c e s s in ( 1 1 .2 .2 3 ) c a r r ie d o u t f o r th e r e d u c e d - o r d e r s y s te m . F in a lly , le t us c o n s i d e r th e m in im iz a tio n o f t h e m e a n - s q u a r e e r r o r in a b a c k w a rd lin e a r p r e d ic t o r . T h e b a c k w a r d p r e d ic t io n e r r o r is
p -1
g p (n) x (n - p ) - f 2 2 bP (k )x (n - k)
( 1 1 .2 .2 4 )
= j(h - p) + ^
k=i
a n d its m e a n -s q u a r e v a lu e is
a* ()*(/1 ~ p + k)
b p = [ M ) 1 2]
( 1 1 .2 .2 5 )
T h e m in im iz a tio n o f b p w ith r e s p e c t to th e p r e d ic t io n c o e f f ic ie n t s y ie ld s th e s a m e s e t o f lin e a r e q u a tio n s a s in ( 1 1 .2 .9 ) . H e n c e th e m in im u m m e a n - s q u a r e e r r o r is m in [ { ] = E j = E ; w h ic h is g iv e n b y ( 1 1 .2 ,1 0 ) . ( 1 1 .2 .2 6 )
11.2.3 The Optimum Reflection Coefficients for the Lattice Forward and Backward Predictors
I n S e c t io n s 1 1 .2 .1 a n d 1 1 .2 .2 w e d e r iv e d th e s e t o f lin e a r e q u a t i o n s w h ic h p r o v id e t h e p r e d i c t o r c o e f f ic ie n ts t h a t m in im iz e t h e m e a n - s q u a r e v a lu e o f th e p r e d ic t io n e r r o r . In th is s e c t io n w e c o n s id e r th e p r o b le m o f o p tim iz in g t h e r e f le c tio n c o e f f i c ie n ts in th e l a t t ic e p r e d ic t o r a n d e x p r e s s in g th e r e f le c tio n c o e f f ic ie n t s in te r m s o f t h e fo r w a r d a n d b a c k w a r d p r e d ic t io n e r r o r s . T h e fo r w a r d p r e d ic t io n e r r o r in th e l a t t ic e filte r is e x p r e s s e d as
f m(n) = f m^ ( n ) + K mgm- X(n - 1 )
( 1 1 .2 .2 7 )
T h e m in im iz a t io n o f E[|/m(n)|2] w ith r e s p e c t t o th e r e f le c t io n c o e f f i c i e n t K m y ie ld s th e re s u lt - [ / m- i ( n ) g ; _ ! ( n - l ) ] rTi '( TTm [ | g _ i ( n - 1)|2]
m
o r , e q u iv a le n tly ,
( 1 1 ./ .Z o ;
K m --
E [ f m- \( n ) g *
i-.
A n 1)1
( 1 1 .2 .2 9 )
, gm~ 1 -------- --
w h ere E f m_ x = E b m_ x = E [ \ g m^ ( n -
1)|2] = E[|/m_ i(n )| 2].
W e o b s e r v e t h a t th e o p tim u m c h o ic e o f th e r e f le c t io n c o e f f ic ie n ts in th e l a t t i c e p r e d i c t o r is th e n e g a tiv e o f t h e ( n o r m a liz e d ) c r o s s c o r r e la t io n c o e f f ic ie n ts
864
Chap. 11
b e tw e e n th e fo r w a r d a n d b a c k w a r d e r r o r s in t h e la t t ic e .* S i n c e it is a p p a r e n t fr o m ( 1 1 .2 .2 8 ) th a t | A T, | < 1 . it fo llo w s t h a t t h e m in im u m m e a n - s q u a r e v a lu e o f th e p r e d ic tio n e r r o r , w h ich c a n b e e x p r e s s e d r e c u r s iv e ly as
E { = {1 - |
is a m o n o to n ic a lly d e c r e a s in g s e q u e n c e .
( 11. 2. 30)
11.2.4 Relationship of an AR Process to Linear Prediction

T h e p a r a m e te r s o f a n A R (/ ? ) p r o c e s s a r e i n tim a te ly r e la te d t o a p r e d ic t o r o f o r d e r
p f o r th e s a m e p r o c e s s .
T o s e e th e r e la tio n s h ip , w e r e c a ll th a t in a n A R ( p )
p r o c e s s , th e a u t o c o r r e la t io n s e q u e n c e { y ^ { m ) } is r e la te d to t h e p a r a m e te r s {a *} b y t h e Y u l e - W a l k e r e q u a tio n s g iv e n in ( 1 1 .1 .1 9 ) o r ( 1 1 .1 .2 0 ) . T h e c o r r e s p o n d in g e q u a tio n s f o r th e p r e d ic t o r o f o r d e r p a r e g iv e n b y ( 1 1 .2 .9 ) a n d ( 1 1 .2 .1 0 ) . A d ir e c t c o m p a r is o n o f t h e s e tw o s e ts o f r e l a t i o n s r e v e a ls t h a t t h e r e is a o n e t o - o n e c o r r e s p o n d e n c e b e tw e e n th e p a r a m e te r s { a * } o f th e A R ( p ) p r o c e s s a n d th e p r e d ic t o r c o e ff ic ie n ts {ap (k)} o f th e p t h - o r d e r p r e d ic t o r . In f a c t , i f th e u n d e rly in g p r o c e s s l x ( n ) } is A R ( p ) , th e p r e d ic t io n c o e f f i c ie n t s o f th e p t h - o r d e r p r e d ic t o r a r e id e n tic a l to {ak}. F u r t h e r m o r e , th e m in im u m M S E in th e / ?th -o rd e r p r e d ic t o r / is id e n tic a l to crj,, th e v a r ia n c e o f th e w h ite n o is e p r o c e s s . s e q u e n c e (u ;{n )). In th is c a s e , th e p r e d ic t io n - e r r o r filte r is a n o is e -w h ite n in g f ilte r w h ic h p r o d u c e s th e in n o v a tio n s
11.3 SOLUTION OF THE NORMAL EQUATIONS

In th e p r e c e d in g s e c tio n w e o b s e r v e d th a t th e m in im iz a t io n o f th e m e a n - s q u a r e v a lu e o f th e fo r w a r d p r e d ic t io n e r r o r r e s u lte d in a s e t o f l in e a r e q u a tio n s f o r th e c o e f f ic ie n ts o f th e p r e d ic t o r g iv e n b y ( 1 1 .2 .9 ) . T h e s e e q u a t i o n s , c a lle d t h e n o r m a l e q u a tio n s , m a y b e e x p r e s s e d in th e c o m p a c t fo r m
Y 2 ,ap ^ y ^ 1
* =0
= 0
1 = 1 , 2 .......... p
0 ,( 0 ) = 1
(n .3 .1 )
T h e r e s u ltin g m in im u m M S E ( M M S E ) is g iv e n b y ( 1 1 .2 .1 0 ) . I f w e a u g m e n t ( 1 1 .2 .1 0 ) t o t h e n o r m a l e q u a tio n s g iv e n b y ( 1 1 .3 .1 ) w e o b ta in t h e s e t o f a u g m en te d
n o r m a l e q u a tio n s, w h ic h m a y b e e x p r e s s e d as
; : 2 .......... P (1 U .2 ,
W e a ls o n o te d t h a t i f th e r a n d o m p r o c e s s is a n A R ( p ) p r o c e s s , t h e M M S E E p = a 2 .
The normalized crosscorrelation coefficients between the forward and backward error in the lattice (i.e., are often called the partial correlation (PARCOR) coefficients.
Sec. 11.3
Solution of the Normal Equations
865
In th is s e c tio n w e d e s c r ib e tw o c o m p u ta tio n a lly e f f ic ie n t a lg o r it h m s f o r s o lv in g th e n o r m a l e q u a tio n s . O n e a lg o r it h m , o r ig in a lly d u e t o L e v in s o n ( 1 9 4 7 ) a n d m o d ifie d b y D u r b in ( 1 9 5 9 ) , is c a lle d th e L e v in s o n - D u r b in a lg o r ith m . T h i s a l g o r ith m is s u ita b le f o r s e r ia l p r o c e s s in g a n d h a s a c o m p u ta tio n c o m p le x ity o f 0 ( p 2 ). T h e s e c o n d a lg o r it h m , d u e t o S c h ii r ( 1 9 1 7 ) , a ls o c o m p u te s t h e r e f le c tio n c o e f f i c ie n ts in 0 ( p 2) o p e r a tio n s b u t w ith p a r a lle l p r o c e s s o r s , th e c o m p u ta tio n s c a n b e p e r f o r m e d in O ( p ) tim e . B o t h a lg o r it h m s e x p lo it th e T o e p lit z s y m m e tr y p r o p e r ty in h e r e n t in th e a u t o c o r r e la t io n m a tr ix . W e b e g in b y d e s c r ib in g th e L e v in s o n - D u r b in a lg o r ith m .
11.3.1 The Levinson-Durbin Algorithm

T h e L e v in s o n - D u r b in a lg o r it h m is a c o m p u ta tio n a lly e f fic ie n t a lg o r it h m f o r s o lv in g th e n o r m a l e q u a tio n s in ( 1 1 .3 .1 ) f o r t h e p r e d ic t io n c o e f f ic ie n ts . e x p lo its th e s p e c ia l s y m m e tr y in th e a u t o c o r r e la t io n m a tr ix > x ,( 0 ) y ,,( D
k; , (
T h i s a lg o r ith m
i)
* A p ~ 1)
( 1 1 .3 .3 )
Yxx(O)
Y*x i P ~ 2 )
. Y x x i p - 1)
N o te th a t The
Y x x (p ~ 2)
yÔ )
m a trix . S in c e
rp(i,j) = j) , rp(i,j) = T*(j, i ) ,

to th e
s o t h a t t h e a u t o c o r r e la t i o n m a tr ix is a T o e p lit z th e m a tr ix is a ls o H e r m itia n . m eth o d o f s o lu tio n , t h a t e x p lo its th e
key
L e v in s o n - D u r b in
T o e p l it z p r o p e r ty o f th e m a tr ix , is to p r o c e e d r e c u r s iv e ly , b e g in n in g w ith a p r e d ic t o r o f o r d e r m = 1 ( o n e c o e f f i c i e n t ) a n d th e n t o in c r e a s e th e o r d e r r e c u r s iv e ly , u s in g th e lo w e r - o r d e r s o lu tio n s to o b ta in t h e s o lu tio n t o th e n e x t- h ig h e r o r d e r . T h u s th e s o lu tio n t o th e f ir s t- o r d e r p r e d ic t o r o b t a in e d b y s o lv in g ( 1 1 .3 .1 ) is
o i(l) = -
Yxx( 1)
y*.t ( 0 )
( 1 1 .3 .4 )
a n d th e re s u ltin g M M S E is e{
= yj (0 )+ f l i ( l ) y ( - l )
( 1 1 .3 .5 )
= y ^ (0)[l - jfli (1)I2]

R e c a l l th a t a j ( l ) = K \ , th e first r e fle c tio n c o e f f i c ie n t in th e l a t t ic e f ilte r . T h e n e x t s te p is t o s o lv e f o r th e c o e f f ic ie n t s {o 2( I ) , ^ 2 (2 )} o f th e s e c o n d - o r d e r p r e d ic t o r a n d e x p r e s s th e s o lu tio n in te r m s o f f l i ( l ) . T h e tw o e q u a tio n s o b ta in e d fr o m ( 1 1 .3 .1 ) a r e
a 2 ( l ) y xx( 0 ) + a 2 (2)y * x ( l ) ~ Y x x ( 1)
(11.3.6)
02( l ) y ^ ( l )
+ a 2 ( 2 ) y xx(0) = - y xx( 2 )
866
Chap. 11
B y u s in g th e s o lu tio n in ( 1 1 .3 .4 ) to e lim in a te yJ r ( l ) , w e o b ta in th e s o lu tio n
a 2 (2) = -
y xx(2) + a i ( 1 ) ^ ( 1 )
X r*(0 )[l | fll(l)i2] ( 1 1 .3 .7 )
Y jcjc( 2 ) - f a \ { l ) y xx( l )
^2(1)
fli d )
+ a 2(2)aUl)
T h u s w e h a v e o b ta in e d th e c o e ffic ie n ts o f th e s e c o n d - o r d e r p r e d ic t o r . A g a in , w e n o t e t h a t ct2(2) = K 2 , th e s e c o n d r e f le c tio n c o e f f ic ie n t in th e la t t i c e f ilte r . P r o c e e d in g in th is m a n n e r , w e c a n e x p r e s s th e c o e f f i c i e n t s o f th e m th - o r d e r p r e d ic t o r in te r m s o f th e c o e f f ic ie n ts o f th e (m l ) s t - o r d e r p r e d ic t o r . T h u s w e c a n w rite th e c o e f f ic ie n t v e c t o r
am as
th e su m o f tw o v e c to r s , n a m e ly .
~ a m( 1) * a m(2)
am = _ f l m( m) w h ere .
1
+ 0 .
" d m_ , " ( 1 1 .3 .8 )
_Km _
a,_i
is th e p r e d ic t o r c o e f f ic ie n t v e c t o r o f th e (m ~ l ) s t - o r d e r p r e d ic t o r a n d
th e v e c t o r d m_ j a n d th e s c a la r K m a r e to b e d e te r m in e d . L e t u s a ls o p a r titio n th e
m x m a u t o c o r r e la t io n m a tr ix T r* as
r m=
w h ere = [y xx(m 1)
r#n-l
7m-1
y ^ (0 )
(1 1 .3 .9 )
yxx(m -
2)
y xx( 1 )] = ( 7 * _ ] ) r , th e a s te r is k ( * ) d e y xx(2 ) y xx (m - 1 ) ] w ith e l e
n o te s th e c o m p le x c o n ju g a t e , a n d y r m d e n o te s th e tr a n s p o s e o f y m. T h e s u p e r s c r ip t
b on
7 m_ ,
d e n o t e s th e v e c t o r
7^
, = [y (l)
m e n ts ta k e n in r e v e r s e o r d e r . T h e s o lu tio n to th e e q u a tio n r ma m = y m c a n b e e x p r e s s e d a s 1 f .1 a m- i _0
um
k
] ) [ * > ]
< im o>
T h is is th e k e y s te p in th e L e v in s o n - D u r b in a lg o r it h m . F r o m ( 1 1 .3 .1 0 ) w e o b ta in tw o e q u a tio n s , n a m e ly , + K my b*_x 7 m - i a m -1 + 7 t '_ i d m - i + K myxx{0 ) = S in c e r m_ i a m_ i = - y m_ : , ( 1 1 .3 .1 1 ) y ie ld s th e s o lu tio n
- y m_ x - y xx(m )
( 1 1 .3 .1 1 ) ( 1 1 .3 .1 2 )
dm_!
But
= - K mT - \ xy b ;_ ,
( 1 1 .3 .1 3 )
7 **_ j
is ju s t
7 m_ !
w ith e le m e n t s ta k e n in r e v e r s e o r d e r a n d c o n ju g a te d . T h e r e -
Sec. 11.3
867
f o r e , th e s o lu tio n in ( 1 1 .3 .1 3 ) is sim p ly '< C _ ] ( " J - 1 ) d m_ , = K ma * * .! = K m flm i -2 ) ( 1 1 .3 .1 4 )
T h e s c a la r e q u a tio n ( 1 1 .3 .1 2 ) c a n n o w b e u s e d to s o lv e f o r K m. I f w e e lim in a te d m_ i in ( 1 1 .3 .1 2 ) b y u s in g ( 1 1 .3 .1 4 ) , w e o b ta in
Km [Yxx(Q) +
H en ce
+ 'Ymîa m -l = ~ Y x x ( )
T h e r e f o r e , b y s u b s titu tin g th e s o lu tio n s in ( 1 1 .3 .1 4 ) a n d ( 1 1 .3 .1 5 ) i n t o ( 1 1 .3 .8 ) , w e o b ta in th e d e s ir e d r e c u r s io n f o r th e p r e d ic t o r c o e f f i c i e n t s in th e L e v i n s o n D u r b in a lg o r ith m as , , ^ Yxx(m) + a m_ , y xx(m ) + a m(m ) = K m = --------------------- j- ------ T -. = --------------------- --------------( 1 1 .3 .1 6 )
K,,<0)
a ( k ) = a m- i ( k ) + K ma*m_ x(m - k) = a m_ i( k ) + a m{ m ) a ^ ( m ~ k ) k -\2 ,...,m tn = i , z , . . . , p

1 ^1 1 3
T h e r e a d e r s h o u ld n o t e t h a t th e r e c u r s iv e r e la tio n in ( 1 1 .3 .1 7 ) is id e n tic a l to th e r e c u r s iv e r e la t io n in ( 1 1 .2 .2 1 ) f o r th e p r e d ic t o r c o e f f ic ie n t s , o b ta in e d f r o m th e p o ly n o m ia ls A m( z ) a n d B m(z). F u r t h e r m o r e , K m is t h e r e f le c t io n c o e f f ic ie n t in th e m th s ta g e o f t h e l a t t ic e p r e d ic t o r . T h i s d e v e lo p m e n t c le a r ly illu s tr a te s th a t th e L e v in s o n - D u r b in a lg o r it h m p r o d u c e s th e r e f le c tio n c o e f f i c ie n t s f o r th e o p tim u m l a t t i c e p r e d ic t io n f ilte r a s w e ll a s th e c o e f f ic ie n ts o f th e o p tim u m d ir e c t -f o r m F I R p r e d ic t o r . F in a lly , le t u s d e te r m in e th e e x p r e s s io n f o r th e M M S E . F o r t h e m th -o r d e r p r e d ic t o r , w e h a v e

m
E f m = yjrjr(O) + 2 2 a >(k ) Y x x ( -lc )
yxx(O ) + ^ [ o m^ i ( / c ) + f l m ( m ) f l * _ 1( r n - * ) ] y ( - / c ) i= l
( 1 1 .3 .1 8 )
E f m_ l [ \ - \ a m(m )\2\ = E f tn^ \ - \ K
m\2)
m = 1,2,...,
w h e r e E ^ = y ( 0 ) . S in c e th e r e fle c tio n c o e f f ic ie n t s s a tis fy t h e p r o p e r ty t h a t | A T m| < 1 , th e M M S E f o r t h e s e q u e n c e o f p r e d ic t o r s s a tis fie s th e c o n d itio n
E{ > E { > E f 2 >-- -> Ep f
(11.3.19)
868
Chap. 11
T h is c o n c lu d e s th e d e r iv a tio n o f th e L e v in s o n - D u r b in a lg o r it h m f o r so lv in g th e lin e a r e q u a tio n s T ma m = y m, f o r m = 0 , 1 a p p e a r s as a v e c t o r in
p . W e o b s e r v e th a t th e lin e a r
e q u a tio n s h a v e th e s p e c ia l p r o p e r ty th a t th e v e c t o r o n th e r ig h t- h a n d s id e a ls o
rm .
In th e m o r e g e n e r a l c a s e , w h e r e th e v e c t o r o n th e
rig h t-h a n d s id e is s o m e o t h e r v e c to r , s a y c m, th e s e t o f li n e a r e q u a tio n s c a n b e s o lv e d r e c u rs iv e ly b y in tr o d u c in g a s e c o n d r e c u r s iv e e q u a t io n t o s o lv e th e m o r e g e n e r a l lin e a r e q u a tio n s
rm bm = cm. T h e
re s u lt is a g e n e r a liz e d L e v in s o n - D u r b i n
a lg o r ith m ( s e e P r o b le m 1 1 .1 2 ) . T h e L e v in s o n - D u r b in r e c u r s io n g iv e n b y ( 1 1 .3 .1 7 ) r e q u ir e s 0 ( m ) m u ltip lic a tio n s a n d a d d itio n s ( o p e r a t i o n s ) to g o fr o m s ta g e m to s ta g e m + 1. T h e r e f o r e , f o r
p s ta g e s it ta k e s o n th e o r d e r o f 1 + 2 + 3-1-------- (- p { p + 1)/ 2, o r 0 ( p 2 ), o p e r a tio n s to

s o lv e f o r th e p r e d ic t io n f ilte r c o e ff ic ie n ts , o r th e r e f le c tio n c o e f f i c i e n t s , c o m p a r e d w ith O ( p ^ ) o p e r a tio n s if w e d id n o t e x p lo it th e T o e p lit z p r o p e r ty o f th e c o r r e la tio n m a tr ix . I f th e L e v in s o n - D u r b in a lg o r ith m is im p le m e n te d o n a s e r ia l c o m p u te r o r s ig n a l p r o c e s s o r , th e r e q u ir e d c o m p u ta tio n tim e is o n th e o r d e r o f 0 ( p 2 ) tim e u n its . O n th e o t h e r h a n d , i f th e p r o c e s s in g is p e r fo r m e d o n a p a r a lle l p r o c e s s in g m a c h in e u tiliz in g a s m a n y p r o c e s s o r s a s n e c e s s a r y t o e x p lo it th e fu ll p a r a lle lis m in th e a lg o r ith m , th e m u ltip lic a tio n s a s w e ll a s th e a d d itio n s r e q u ir e d to c o m p u te ( 1 1 .3 .1 7 ) c a n b e c a r r ie d o u t s im u lta n e o u s ly . T h e r e f o r e , th is c o m p u ta tio n c a n b e p e r f o r m e d in O ( p ) tim e u n its . H o w e v e r , th e c o m p u ta tio n in ( 1 1 .3 .1 6 ) f o r th e r e f le c tio n c o e ff ic ie n ts ta k e s a d d itio n a l tim e . C e r ta in ly , th e in n e r p r o d u c ts in v o lv in g th e v e c to r s a m_ i a n d p ro cesso rs. c a n b e c o m p u te d s im u lta n e o u s ly b y e m p lo y in g p a r a lle l H e n c e , th e c o m p u ta tio n s in th e H o w e v e r , th e a d d itio n o f th e s e p r o d u c ts c a n n o t b e d o n e s im u lta n e
o u s ly , b u t in s te a d , r e q u ir e 0 ( \ o g p ) tim e u n its . c o m p lis h e d in 0 ( p \ o g p ) tim e .
L e v in s o n - D u r b in a lg o r it h m , w h e n p e r f o r m e d b y p p a r a lle l p r o c e s s o r s , c a n b e a c In th e f o llo w in g s e c tio n w e d e s c r ib e a n o t h e r a lg o r it h m , d u e to S c h iir ( 1 9 1 7 ) , th a t a v o id s th e c o m p u ta tio n o f in n e r p r o d u c ts , a n d t h e r e f o r e is m o r e s u ita b le f o r p a r a lle l c o m p u ta tio n o f th e r e f le c tio n c o e ff ic ie n ts .
11.3.2 The Schur Algorithm

T h e S c h u r a lg o r ith m is in tim a te ly r e la te d to a r e c u r s iv e t e s t f o r d e te r m in in g th e p o s it iv e d e fin it e n e s s o f a c o r r e la t io n m a tr ix . T o b e s p e c ific , l e t u s c o n s id e r th e a u t o c o r r e la tio n m a tr ix T ^ + i a s s o c ia t e d w ith th e a u g m e n t e d n o r m a l e q u a tio n s g iv e n b y ( 1 1 .3 .2 ) . F r o m th e e le m e n ts o f th is m a tr ix w e f o r m t h e fu n c t io n
y(l)z_I +Y*A2)z- 2 + --- + yxAp)z-p

^ (0 ) + r ( l ) z - V - + y ( P n-<> ( 1 U '2 0 ) a n d th e s e q u e n c e o f fu n c t io n s R m(z) d e fin e d r e c u r s iv e ly as
R m(z) -
D i \
z *[1 - / ?*_ ! ( o o ) ( z ) J
Rm - 1 ( z ) R m - l ( o o ) ------ rr
m = 1,2,...
( 1 1 .3 .2 1 )
/1 1 o - i n
Sec. 11.3
869
S c h i i r s t h e o r e m s t a t e s t h a t a n e c e s s a r y a n d s u ffic ie n t c o n d itio n f o r th e c o r r e l a t i o n m a t r ix to b e p o s itiv e d e f in it e is t h a t |/?m(oo)| < 1 f o r m = 1, 2 , . . . , p . L e t u s d e m o n s tr a t e th a t th e c o n d itio n f o r p o s itiv e d e fin ite n e s s o f th e a u t o c o r r e la t io n m a t r ix I ^ + i is e q u iv a le n t t o th e c o n d itio n t h a t th e r e f le c tio n c o e f f ic ie n ts in t h e e q u iv a le n t la t t ic e f ilt e r s a tis f y t h e c o n d itio n | K m \ < 1, m = 1, 2 , . . . , p . F ir s t , w e n o t e th a t R o(oo) = 0 . T h e n , fr o m ( 1 1 .3 .2 1 ) w e h a v e
R \ ( 2) ---------^
Yxx(l) +
Kr*(0) + y (l)z -1 + + Yxx(p)z~p
X x x (2 )z _1 + + Yxx(p)z~p+l -----------------------------------
^ ( 1 1 .3 .2 2 )
H e n c e r t i( o o ) = yxx( l ) / y Xx(0 ). W e o b s e r v e t h a t R \( o o ) = K \ . S e c o n d , w e c o m p u te R j( z ) a c c o r d in g to ( 1 1 .3 .2 1 ) a n d e v a lu a te th e r e s u lt a t
z = 00 . T h u s w e o b ta in
Yxx(2) + K ^ x x d )
x(0)(l - lîi2)
A g a in , w e o b s e r v e t h a t # 2( 00) = K j . B y c o n tin u in g th is p r o c e d u r e , w e fin d th a t
R m(oo) = K m, f o r m = 1, 2 , . . . , p . H e n c e th e c o n d itio n |J?m(oo)| < 1 f o r m = 1,

2 , . . . , p , is id e n tic a l to th e c o n d itio n | | < 1 f o r m = 1, 2 , . . . , p , a n d e n s u r e s th e p o s it iv e d e f in it e n e s s o f th e a u t o c o r r e l a t i o n m a tr ix r p+i. S i n c e th e r e fle c tio n c o e f f ic ie n t s c a n b e o b ta in e d fr o m th e s e q u e n c e o f f u n c tio n s R m( 2), m = 1, 2 , . . . , p , w e h a v e a n o t h e r m e th o d f o r s o lv in g th e n o r m a l e q u a tio n s . W e c a ll th is m e t h o d th e S c h u r a lg o rith m .
Schur algorithm.
L e t us firs t r e w r ite R m(z) as
*<z) = 7 frr Q m (z)

w h ere
m= l ' P
( 1 1 .3 .2 3 )
--------- h Y x x (p )z ~ p P o (z) = Y x x O ) Z ~ X + Y x x ( 2 ) z ~ 2 H ( 1 1 .3 .2 4 )
Q a (z) = Y xxi 0 ) + Y x x ( l) z ~ l + -------h Y x x (p )z ~ p

S in c e
^0
0 a n d K m R m(oo) f o r m =
1 , 2 , . . . , p , th e r e c u r s iv e e q u a tio n
( 1 1 .3 .2 1 ) im p lie s th e fo llo w in g r e c u r s iv e e q u a t io n s f o r th e p o ly n o m ia ls P m(z) a n d
Qm(z):
[Q m (z)j
T h u s w e have Z_1
r p m_ i ( z ) [_ Q m \ (z )
J
( 1 1 .3 .2 6 )
P l( z ) =
P o (z) = X t x ( l ) z _1 -I- Y x x (2 )z ~ 2 + + Y x x (p )z ~ p
f i i ( z ) = z -1 Q o (z ) = Y xx(0 )z - 1 + y ( l ) z - 2 + - + y xx( p ) z ~ p- 1
and
K\ =
P i( z ) Q iiz )
Y xx(1) Vxx(0)
(11.3.27)
870
Chap. 11
N e x t th e r e fle c tio n c o e f f ic ie n t K i is o b ta in e d b y d e te r m in in g P 2 (z) a n d Q 2 U ) fro m ( 1 1 .3 .2 5 ) , d iv id in g P 2 (z) b y Q i( z ) a n d e v a lu a tin g th e re s u lt a t z = o o . T h u s w e fin d th a t
P i( z ) =
=
P \( z ) + K xQ x{z) [Yxx( 2 ) + K i y xx( l ) \ z ~ 2 +

+ [Y x A p ) + * 1 Y x x ip 1 )]z~ p ( 1 1 .3 .2 8 )
Q i( z ) =
=
+ [x rJf(0 ) + Js:1 v ( i ) ] z - 2 + --+ [Y x x ip 2 ) + K * y xx( p - 1 ) } z ~ p T h u s w e o b s e r v e th a t th e
w h e r e th e te r m s in v o lv in g z ~ p~ l h a v e b e e n d r o p p e d .
re c u r s iv e e q u a tio n in ( 1 1 .3 .2 5 ) is e q u iv a le n t to ( 1 1 .3 .2 1 ) . B a s e d o n th e s e r e la tio n s h ip s , th e S c h u r a lg o r ith m is d e s c r ib e d b y th e fo llo w in g re c u r s iv e p r o c e d u r e .
In it ia liz a t io n . F o r m th e 2 x ( p + 1) g e n e r a t o r m a tr ix
Gn = 0 y ^ (0 ) Kv-r(l) yÔ )
Yxx( 2 ) y xx (2 )
^ ( Z 7) ]
y xx{ p ) \
w h e r e th e e le m e n t s o f th e firs t ro w a r e th e c o e f f i c i e n t s o f P q( z ) a n d th e e le m e n ts o f th e s e c o n d ro w a r e th e c o e f f ic ie n ts o f ?o(z).
Step 1.
S h ift th e s e c o n d ro w o f th e g e n e r a t o r m a tr ix t o th e rig h t by o n e
p la c e a n d d is c a r d th e la s t e le m e n t o f th is ro w . A z e r o is p la c e d in th e v a c a n t p o s itio n . T h u s w e o b ta in a n e w g e n e r a to r m a tr ix , G, = 0 .0
Yxx( 1) y xx (0 )
Y xxO ) y xx( 1)
Yx A 2 ) yxx{ 1)
y xA p ) y xx(p - 1 )
( 1 1 .3 .3 0 )
T h e (n e g a tiv e ) r a t io o f t h e e le m e n ts in th e s e c o n d c o lu m n y ie ld th e r e f le c tio n c o e f f ic ie n t K \ = - y xA l ) / y xx {0).
Step 2. M u ltip ly th e g e n e r a t o r m a tr ix b y th e 2 x 2 m a tr ix
Vi = T h u s w e o b ta in 1
Ki
1
(1 1 .3 .3 1 )
V!G1=
0 0
0 y (0 ) + ^ * y ( l)
y < 2 ) + K xYxx{ \ )
Y x x ip ) + K \ y xx{ p - 1) Y x x ip 1 ) + K * y xx(p )_ ( 1 1 .3 .3 2 )
Step 3. S h ift th e s e c o n d ro w o f V i G i b y o n e p la c e t o th e r ig h t a n d th u s fo rm
th e n e w g e n e r a t o r m a tr ix G: = 0 L0 0 0
yxx(2) + K i y xx( l ) y xA 0 ) + K ; y xx( \ )
y xx(p ) + K \ y X I (p - 1) Yxx( p - 2 ) + K \ y xx( p - l )

( 1 1 .3 .3 3 )
T h e n e g a tiv e r a t io o f th e e le m e n ts in th e th ir d c o lu m n o f G
y ie ld s K
2.
Sec. 11.3
871
S te p s 2 a n d 3 a r e r e p e a t e d u n til w e h a v e s o lv e d f o r a ll p r e f le c tio n c o e f f ic ie n ts . In g e n e r a l, th e 2 x 2 m a tr ix in s te p 2 is 1
K
1
K*
( 1 1 .3 .3 4 )
a n d m u ltip lic a tio n o f V m b y G m y ie ld s V mG m. In s te p 3 w e s h ift th e s e c o n d ro w o f V mG m o n e p la c e t o th e rig h t a n d o b ta in th e n e w g e n e r a t o r m a tr ix G m+i . W e o b s e r v e th a t th e s h iftin g o p e r a tio n o f th e s e c o n d ro w in e a c h ite r a tio n is e q u iv a le n t t o m u ltip lic a tio n b y th e d e la y o p e r a t o r z ~ x in t h e s e c o n d re c u r s iv e e q u a tio n in ( 1 1 .3 .2 5 ) . W e a ls o n o te t h a t th e d iv is io n o f th e p o ly n o m ia l P m(z) b y th e p o ly n o m ia l Q m(z) a n d th e e v a lu a tio n o f t h e q u o tie n t a t z = oo is e q u iv a le n t to d iv id in g th e e le m e n ts in th e (m + l ) s t c o lu m n o f G m. T h e c o m p u ta tio n o f th e p r e fle c tio n c o e f f ic ie n ts c a n b e a c c o m p lis h e d b y u se o f p a r a lle l p r o c e s s o r s in
0 { p ) tim e u n its . B e lo w w e d e s c r ib e a p ip e lin e d a r c h it e c t u r e f o r p e r f o r m in g t h e s e

c o m p u ta tio n s . A n o t h e r w ay o f d e m o n s tr a t in g th e r e la tio n s h ip o f th e S c h u r a lg o r it h m t o th e L e v in s o n - D u r b in a lg o r ith m a n d th e c o r r e s p o n d in g l a t t ic e p r e d i c t o r is to d e te r m in e th e o u tp u t o f th e la ttic e filte r o b ta in e d w h e n th e in p u t s e q u e n c e is th e c o r r e la tio n s e q u e n c e {y.r.v(m ), m = 0 . 1 , . . . ) , T h u s , th e first in p u t to th e l a t t i c e f ilte r is yx_ v( 0 ) , th e s e c o n d in p u t is y ( l ) , a n d s o o n [i.e ., / ( ) = th e firs t s ta g e , w e h a v e g o(n 1) =
yxx(n )].
A f t e r th e d e la y in 1, th e r a t io
yxx(n ~ \ ) .
H e n ce , fo r n =
./ ii(l)/ g o(0) = K r,(l)/ y jt.v (0 K w h ich is th e n e g a tiv e o f th e r e f le c tio n c o e f f ic ie n t K \ . A lt e r n a t iv e ly , w e c a n e x p r e s s th is r e la tio n s h ip a s
/o(D + îgo(O ) = X n ( l ) + ^ iyjjc(O ) = 0

F u r t h e r m o r e , # o (0 ) = yxx( 0 ) = is, a c c o r d in g to ( 1 1 .2 .4 ) , / ,( 2 ) = M 2 ) + K ^ g o G ) = y xx( 2 ) + K lY x x ( l ) a n d a f t e r th e u n it o f d e la y in th e s e c o n d s ta g e , w e h a v e
g,(l) ^
. A t tim e n = 2 , th e in p u t to th e s e c o n d s ta g e
K * M l ) + g o ( 0 ) = K l y xx(1)
y ( 0 )
N o w th e r a t io f \ { 2 ) / g \ ( \ ) is / i(2 ) si(l) H en ce /i (2 ) + K i g \ ( \ ) = 51 (1 ) = 0
y xx(2) + K lY x x ( l )
y x x(0) + ^ K ( l )
yxx(2) + K \ y xx{ \ )
e
= - k 2
E{
B y c o n tin u in g in th is m a n n e r , w e c a n s h o w t h a t a t t h e in p u t t o th e m th l a t t i c e s ta g e , th e r a t io f m- \ ( m ) / g m- i ( m 1) = - K m a n d g m- i ( m
~ 1) = E } m_ v
C o n s e q u e n tly ,
th e la t t ic e f ilt e r c o e f f ic ie n ts o b ta in e d f r o m th e L e v in s o n a lg o r it h m a r e id e n tic a l to t h e c o e f f i c i e n t s o b ta in e d in th e S c h u r a lg o r it h m . F u r t h e r m o r e , th e l a t t ic e filte r
872
Chap. 11
s tr u c tu r e p r o v id e s a m e t h o d f o r c o m p u tin g th e r e f le c t io n c o e f f i c i e n t s in th e la ttic e p r e d ic to r .
A pipelined architecture for implementing the Schur algorithm.
Kung
a n d H u ( 1 9 8 3 ) d e v e lo p e d a p ip e lin e d la ttic e - ty p e p r o c e s s o r f o r im p le m e n tin g th e S c h u r a lg o r ith m . T h e p r o c e s s o r c o n s is ts o f a c a s c a d e o f p l a t t i c e - t y p e s ta g e s , w h e r e e a c h s ta g e c o n s is ts o f tw o p r o c e s s in g e le m e n t s ( P E s ) , w h ic h w e d e s ig n a te a s u p p e r P E s d e n o te d a s A \ , A z . . . . , A p , a n d lo w e r P E s d e n o te d a s S j , B 2 .......... B p, a s sh o w n in F ig . 1 1 .5 . T h e P E d e s ig n a te d a s A \ is a s s ig n e d t h e ta s k o f p e r f o r m in g d iv isio n s. T h e r e m a in in g P E s p e r fo r m o n e m u ltip lic a tio n a n d o n e a d d itio n p e r i te r a tio n ( o n e c lo c k c y c le ) . I n itia lly , t h e u p p e r P E s a r e lo a d e d w ith th e e le m e n ts o f t h e firs t ro w o f th e g e n e r a t o r m a tr ix G o . as illu s tr a te d in F ig . 1 1 .5 . T h e lo w e r P E s a r e lo a d e d w ith th e e le m e n ts o f th e s e c o n d ro w o f th e g e n e r a t o r m a tr ix G o- T h e c o m p u t a t io n a l p r o c e s s b e g in s w ith th e d iv is io n P E , A \, w h ic h c o m p u te s t h e firs t r e f l e c t i o n c o e f f ic ie n t as
K \ = yJ J :( l ) / K t j( 0 ) . T h e v a lu e o f K \ is s e n t s im u lta n e o u s ly to a ll th e P E s in th e
u p p e r b r a n c h a n d lo w e r b r a n c h . T h e s e c o n d s te p in th e c o m p u ta tio n is to u p d a te th e c o n t e n t s o f a ll p r o c e s s in g e le m e n ts s im u lta n e o u s ly . T h e c o n te n ts o f th e u p p e r a n d l o w e r P E s a r e u p d a te d a s fo llo w s ; P E A m: A m < A m + K i B m P E B m: B m * B m + K \ A m
m = 2 , 3 .......... p m = 1,2,.... p
T h e th ird s te p in v o lv e s th e s h iftin g o f th e c o n te n ts o f th e u p p e r P E s o n e p la c e t o th e le ft. T h u s w e h a v e P E A m\ A - i + - A m
m = 2 , 3 .......... p B \ c o n ta in s y ( 0 ) +
A t th is p o in t, P E A \ c o n ta in s yxx( 2 ) + K \ y xx( \ ) w h ile P E
K * y xx( \ ) . H e n c e th e p r o c e s s o r A \ is r e a d y t o b e g in th e s e c o n d c y c le b y c o m p u tin g
Figure 11.5 Pipelined parallel processor for computing the reflection coefficients
Sec. 11.4
Properties of the Linear Prediction-Error Filters
873
th e s e c o n d r e f le c tio n c o e f f i c i e n t K b e g in n in g w ith th e d iv is io n A \ / B \ e a c h i te r a tio n .
-A \/B \.
T h e t h r e e c o m p u ta tio n a l s te p s
a r e r e p e a t e d u n til all p r e f le c tio n c o e f f ic ie n ts
a r e c o m p u te d . N o te th a t P E B \ p r o v id e s th e m in im u m m e a n - s q u a r e e r r o r e L f o r I f t< j d e n o t e s th e tim e f o r P E A \ to p e r f o r m a ( c o m p l e x ) d iv is io n a n d rma is th e tim e r e q u ir e d to p e r fo r m a ( c o m p le x ) m u ltip lic a tio n a n d an a d d itio n , th e tim e r e q u ir e d t o c o m p u te a ll p r e fle c tio n c o e f f ic ie n ts is p (z^ + zma) f o r th e S c h u r a lg o r ith m .
11.4 PROPERTIES OF THE LINEAR PREDICTION-ERROR FILTERS

L i n e a r p r e d ic t io n filte r s p o s s e s s s e v e r a l im p o r ta n t p r o p e r tie s , w h ich w e n o w d e s c r ib e . W e b e g in b y d e m o n s tr a t in g th a t th e f o r w a r d p r e d ic t io n - e r r o r f ilte r is m in im u m p h a s e .
Minimum-phase property of the forward prediction-error filter.

e f f ic ie n ts , a n d c o n s e q u e n tly , |A", | < = (1 \ K , c a n 1 fo r all i .
We
h a v e a lr e a d y d e m o n s tr a t e d th a t th e r e fle c tio n c o e f f ic ie n ts { AT,} a r e c o r r e la tio n c o T h is c o n d itio n a n d th e re la tio n b e u se d to s h o w th a t th e z e r o s o f th e p r e d ic t io n - e r r o r
f ilt e r a r e e it h e r all in s id e th e u n it c ir c le o r th e y a r e all o n th e u n it c ir c le . F ir s t , w e s h o w th a t i f e J, > 0 , th e z e r o s |z,| < 1 f o r e v e r y i. T h e p r o o f is by in d u c tio n . C le a r ly , fo r p = 1 th e s y s te m fu n c t io n f o r th e p r e d i c t i o n - e r r o r f ilte r is
A A z ) = 1 + K i z -1
( 1 1 .4 .1 )
H e n c e z \ = K \ a n d e { = (1 - | K \ \ 2) E l > 0 . N o w , s u p p o s e th a t th e h y p o th e s is is tr u e f o r p 1. T h e n , i f n is a r o o t o f A p { z ) , w e h a v e f r o m ( 1 1 .2 .1 6 ) a n d ( 1 1 .2 .1 8 ) ,
A p (Zi) = A p - \ ( z i ) + K p z ~ l Bp
( 1 1 .4 .2 )
H en ce
1
Kp
z - pA ; _ x( i / z , )
A p ^ i (Zi )
s
Q U i)
( 1 1 .4 .3 )
W e n o t e th a t th e fu n c tio n Q( z) is a ll p a s s . In g e n e r a l, a n a ll-p a s s fu n c t io n o f th e fo rm p ( z ) = n ^ u *i 1 f o r [zl < 1, [/ (z )! = 1 fo r | z| = 1, a n d
|/*(z)| < 1 f o r | z| > 1. S i n c e Q ( z ) = - P { z ) f z , it fo llo w s th a t |z,| < 1 i f |f?(z)| > 1. C le a r ly , th is is th e c a s e s in c e j2 ( z ,) = 1/ATP a n d e {, > 0 .
874

O n th e o t h e r h a n d , s u p p o s e th a t
Chap. 11
> 0 and
= 0 . I n th is c a s e \ K p \ - = \
a n d \Q(Zi)\ = 1. S in c e th e M M S E is z e r o , th e r a n d o m p r o c e s s jr ( n ) is c a lle d p r e
d ic t a b le o r d e te rm in istic .
fo rm
S p e c ific a lly , a p u r e ly s in u s o id a l r a n d o m p r o c e s s o f th e
M x ( n ) = 2 2 a ke jinoi*+et) k=l
(0 , 2 n ) , h a s th e a u t o c o r r e la t io n
( 1 1 .4 .5 )
w h e r e th e p h a s e s { 8 k] a r e s ta tis tic a lly in d e p e n d e n t a n d u n ifo r m ly d is tr ib u te d o v e r
M YxA m ) = 2 2 a h Jmwi
*= i a n d th e p o w e r d e n s ity s p e c tr u m ( H - 4 .6 )
r xA f ) =
*= 1
- /*>
h = ?
Z7r
(114-7)
T h is p r o c e s s is p r e d ic t a b le w ith a p r e d ic t o r o f o r d e r p > M . T o d e m o n s tr a t e th e v a lid ity o f th e s t a te m e n t, c o n s id e r p a s s in g th is p r o c e s s th r o u g h a p r e d ic t io n e r r o r f ilt e r o f o r d e r p > M . T h e M S E a t th e o u tp u t o f th is f ilte r is M/2 / = /
J -\n
r xA f ) \ A p ( f ) \ 2d f
r 1/2 [ =
j-i/2
* (/ -/ *)
\A p{ f ) \ 2d f
( 1 1 .4 .8 )
= * 2i v / * ) i 2 k=i
B y c h o o s in g M o f t h e p z e r o s o f t h e p r e d ic t io n - e r r o r filte r t o c o in c id e w ith th e fr e q u e n c ie s {/ * } , th e M S E / c a n b e fo r c e d to z e r o . T h e r e m a in in g p M z e r o s c a n b e s e le c t e d a r b itr a r ily to b e a n y w h e r e in s id e th e u n it c ir c l e . F in a lly , t h e r e a d e r c a n p r o v e t h a t i f a r a n d o m p r o c e s s c o n s is ts o f a m ix tu r e o f a c o n tin u o u s p o w e r s p e c tr a l d e n s ity a n d a d is c r e te s p e c tr u m , th e p r e d ic t io n e r r o r filte r m u s t h a v e a ll its r o o t s in s id e th e u n it c ir c le .
Maximum-phase property of the backward prediction-error filter.

s y s te m fu n c t io n f o r t h e b a c k w a r d p r e d ic t io n e r r o r f ilt e r o f o r d e r p is
The
B p (z) = z - pA ; ( z ~ 1)
( 1 1 .4 .9 )
C o n s e q u e n tly , th e r o o t s o f B p ( z ) a r e th e r e c ip r o c a ls o f th e r o o t s o f th e fo r w a rd p r e d ic t io n - e r r o r f ilt e r w ith s y s te m fu n c t io n A p(z ). a ll th e r o o t s o f B p ( z ) lie o n t h e u n it c ir c le . H e n c e i f A p (z) is m in im u m p h a s e , th e n B p (z) is m a x im u m p h a s e . H o w e v e r , i f th e p r o c e s s x ( ) is p r e d ic t a b le ,
Sec. 11.4
Properties of the Linear Prediction-Error Filters
875
Whitening property.
S u p p o s e th a t th e ra n d o m p r o c e s s x(rt) is a n A R (p )
s ta tio n a r y r a n d o m p r o c e s s th a t is g e n e r a te d b y p a s s in g w h ite n o is e w ith v a r ia n c e
a th r o u g h a n a ll-p o le f ilt e r w ith s y s te m fu n c t io n H( Z) = --------- y ----------k= 1

T h e n th e p r e d i c t i o n - e r r o r f ilte r o f o r d e r p h a s th e s y s te m fu n c t io n
p
( 1 1 .4 .1 0 )
A p (z) 1 + 2 2 a p ( f t z ~ k
*= i w h e r e th e p r e d ic t o r c o e f f ic ie n ts a p (k) = a f ilte r is a w h ite n o is e s e q u e n c e { to (n ) }. in S e c t i o n 1 1 .2 .
( 1 1 .4 .1 1 )
T h e r e s p o n s e o f th e p r e d i c t i o n - e r r o r I n th is c a s e th e p r e d ic t io n - e r r o r f ilte r
w h ite n s t h e in p u t r a n d o m p r o c e s s x ( n ) a n d is c a lle d a w h ite n in g filte r , a s in d ic a te d M o r e g e n e r a lly , e v e n i f th e in p u t p r o c e s s x ( n ) is n o t a n A R p r o c e s s , th e p r e d i c t io n - e r r o r f ilt e r a tte m p ts t o r e m o v e th e c o r r e l a t i o n a m o n g th e s ig n a l s a m p le s o f th e in p u t p r o c e s s . A s th e o r d e r o f th e p r e d ic t o r is in c r e a s e d , th e p r e d ic t o r o u tp u t x ( n ) b e c o m e s a c lo s e r a p p r o x im a tio n to x ( n ) a n d h e n c e th e d if f e r e n c e
f ( n ) = x ( n ) x ( n ) a p p r o a c h e s a w h ite n o is e s e q u e n c e .
Orthogonality of the backward prediction errors.

T h a t is,
T h e backw ard p re
d ic t io n e r r o r s { # ( ) ) fr o m d iff e r e n t s ta g e s in t h e F I R la t t i c e f ilte r a r e o r th o g o n a l.
E [ g m(n )g i (n)] =
, t i
t m
( 1 1 .4 .1 2 )
T h is p r o p e r ty is e a s ily p r o v e d b y s u b s titu tin g f o r gm(n) a n d g j( n ) in to ( 1 1 .4 .1 2 ) a n d c a r r y in g o u t th e e x p e c ta t io n . T h u s
m E [g m (n )g/*(n )] = i > m ( * )
i-0
I
> ,* ( ;) [ * ( - k )x * (n - j ) ] ( 1 1 .4 .1 3 )
J= 0
m b * a ) ] C bm ^ Yxx a ~ f t
j =0
B u t t h e n o r m a l e q u a tio n s f o r th e b a c k w a r d l in e a r p r e d ic t o r r e q u ir e th a t
k=0
T h e re fo re ,
> m (*)X U - *) =
f
[
m'
Jj Z 2 .......... m ~ 1
( 1 1 .4 .1 4 )
E [ g m( n )? g n= )}= !( ; *( )} { \ J! " ; - E '
( 1 1 .4 .1 5 ) 0 < I < m 1
876
Chap. 11
Additional properties.
T h e r e a r e a n u m b e r o f o t h e r in t e r e s t in g p r o p e r tie s la ttic e f ilte r .
r e g a r d in g th e fo r w a rd a n d b a c k w a r d p r e d ic t io n e r r o r s in t h e F I R th e re a d e r. ( a ) E [ f m(n )x (n - /')] = 0 , ( b ) E [ g m{n )x (n - /)] = 0 , 1 < i < m 0 < i < m 1
T h e s e a r e g iv e n h e r e f o r r e a l-v a lu e d d a ta . T h e i r p r o o f is l e f t a s a n e x e r c is e f o r
( c ) [ / m( n ) x ( n ) ] = [ g m(n )jr (n - m )] = E m ( d ) E [ f i( n ) f j ( n ) ] = EtnajO 'i j ) (e ) E [ f ,( n ) f j{ n - f)] = 0 , fo r
j
{
>
i < j
J ^ J :
( f ) E [ g i ( n)g j (n - 0 ] = 0 ,
for
J. + ^
! *
J.
(g ) E [ M n + i ) f j ( n + j ) } = ^
) = j
(h) E[gi(n + i ) gj ( n + ; ) ] = max {iJ)

f f l [ / < < > < > ] - { '* ' ( j ) E [ f i( n ) g i { n - 1 )] = - K i + \ E i ( k ) [ g ,( r t - l ) jt ( n ) ] = E [ f j { n + 1 )jc(u 1 )] = - * , + ] ,
i J ~ 0'
* "
(I) [ / , ( B > a < - D J -
j \ +l(
11.5 AR LATTICE AND ARMA LATTICE-LADDER FILTERS

In S e c t io n 1 1 .2 w e s h o w e d th e r e la tio n s h ip o f t h e a ll-z e r o F I R p r e d ic t io n . T h e lin e a r p r e d i c t o r w ith t r a n s f e r f u n c t io n ,
p
l a t t i c e t o lin e a r
A P(z) = 1 + ' ^ a p ( k) z ~ k
*= l
( 1 1 .5 .1 )
w h e n e x c ite d b y a n in p u t ra n d o m p r o c e s s {jc ( n ) } , p r o d u c e s a n o u tp u t th a t a p p r o a c h e s a w h ite n o is e s e q u e n c e a s p -> o o. O n th e o t h e r h a n d , i f th e in p u t p r o c e s s is a n A R ( p ) , th e o u tp u t o f A p {z) is w h ite . S in c e A p{z) g e n e r a t e s a M A ( p ) p r o c e s s w h e n e x c it e d w ith a w h ite n o is e s e q u e n c e , t h e a ll- z e r o l a t t i c e is s o m e tim e s c a lle d a M A la t t ic e . I n th e fo llo w in g s e c t i o n , w e d e v e lo p th e la t t i c e s tr u c tu r e f o r t h e in v e r s e filte r
1 / A p (z),
p rocess.
c a lle d th e A R
l a t t ic e , a n d th e l a t t i c e - l a d d e r s tr u c tu r e f o r a n A R M A
Sec. 11.5
AR Lattice and ARMA Lattice-Ladder Filters
877
11.5.1 AR Lattice Structure

L e t us c o n s id e r a n a ll-p o le s y s te m w ith s y s te m fu n c tio n
H ( z ) = --------- j l ------------i= l
( 1 1 .5 .2 )
T h e d if f e r e n c e e q u a tio n f o r th is I I R s y s te m is
y(n) = - ^ P p ( ) y ( n - k ) + x ( n )
k= l
( 1 1 .5 .3 )
N o w s u p p o s e t h a t w e in te r c h a n g e th e r o le s o f t h e in p u t a n d o u tp u t [ i.e ., i n t e r c h a n g e x (n ) w ith y ( n ) in ( 1 1 .5 .3 ) ] o b ta in in g th e d i f f e r e n c e e q u a tio n
*(n) = k= 1
o r, e q u iv a le n tly ,
~ft +
p
v ( ) = *(/?) + 2 2 , a p ( f t x (n ~ f t *= i
( 1 1 .5 .4 )
W e o b s e r v e th a t ( 1 1 .5 .4 ) is a d if f e r e n c e e q u a tio n f o r an F I R s y s te m w ith s y s te m fu n c t io n A p(z ). T h u s a n a ll-p o le I I R s y s te m c a n b e c o n v e r te d to an a ll z e r o s y s te m b y in te r c h a n g in g th e r o le s o f th e in p u t a n d o u tp u t. B a s e d o n th is o b s e r v a t io n , w e c a n o b ta in th e s tr u c tu r e o f a n A R { p ) la ttic e f r o m a M A ( p ) la t t ic e b y in te r c h a n g in g t h e in p u t w ith th e o u tp u t. S in c e th e M A ( p ) l a t t i c e h a s v (n ) = f P (n) as its o u tp u t a n d x ( n ) = / o(n ) is t h e in p u t, w e le t
x { n ) = f p {n)
( 1 1 .5 .5 )
y (n ) =
fo (n )
T h e s e d e fin itio n s d ic t a te t h a t th e q u a n t it ie s { f m{n )} b e c o m p u te d in d e s c e n d in g o r d e r. T h is c o m p u ta tio n c a n b e a c c o m p lis h e d b y r e a r r a n g in g th e r e c u r s iv e e q u a tio n f o r { / * ( ) } in ( 1 1 .2 .4 ) a n d s o lv in g f o r /m_ j( n ) in t e r m s o f f m{n ). T h u s w e o b ta in
f m~ \( n ) -
f m{n) - K mg m- } (n - 1 )
m = p , p - 1 .......... 1
T h e e q u a tio n f o r g m(n) r e m a in s u n c h a n g e d . T h e r e s u lt o f th e s e c h a n g e s is th e s e t o f e q u a tio n s *(n ) =
f p (n)
f m- i ( n ) = f m(n) - K mg m- i ( n - 1)
(1 1 .5 .6 )
gm(n) =
y(n) = M n ) = gain)
T h e c o r r e s p o n d in g s tr u c tu r e f o r t h e A R ( p ) l a t t i c e is s h o w n in F ig . 1 1 .6 . N o te th a t th e a ll-p o le la t t ic e s tr u c tu r e h a s a n a ll-z e r o p a th w ith in p u t g o ( ) a n d o u tp u t g p(n ),
878
Chap. 11
Figure 11.6 Lattice structure for an all-pole system

w h ic h is id e n tic a l to th e a ll-z e r o p a th in th e M a ( p ) la t t ic e s tr u c tu r e . T h is is n o t
s u r p ris in g , s in c e th e e q u a tio n f o r g m (n ) is id e n tic a l in th e tw o l a t t i c e s tr u c tu r e s . W e a ls o o b s e r v e th a t th e A R ( p ) a n d M A (/ ?) la t t ic e s t r u c tu r e s a r e c h a r a c te r iz e d b y th e s a m e p a r a m e te r s , n a m e ly , th e r e fle c tio n c o e f f i c i e n t s {/if,). C o n s e q u e n tly , th e e q u a tio n s g iv e n in ( 1 1 .2 .2 1 ) a n d ( 1 1 .2 .2 3 ) f o r c o n v e r t in g b e tw e e n th e s y s te m p a r a m e te r s {o^ O t)} in th e d ir e c t -fo r m r e a liz a tio n s o f th e a ll- z e r o s y ste m A p(z ) a n d th e l a t t ic e p a r a m e te r s { ; ] o f th e M A ( p ) la t t ic e s t r u c t u r e , a p p ly a s w e ll to th e a ll- p o le s tr u c tu r e s .
11.5.2 ARMA Processes and Lattice-Ladder Filters

T h e a ll-p o le l a t t ic e p r o v id e s th e b a s ic b u ild in g b lo c k f o r la t t ic e - t y p e s tr u c tu r e s th a t im p le m e n t I I R s y s te m s th a t c o n ta in b o th p o le s a n d z e r o s . T o c o n s t r u c t th e a p p r o p r ia te s tr u c tu r e , le t u s c o n s id e r a n I I R s y s te m w ith s y s te m fu n c t io n
c 9(*)z _*
H ( z ) = ------------------- =
, v *= i W ith o u t lo s s o f g e n e r a lity , w e a s s u m e th a t p > q. T h is s y s te m is d e s c r ib e d b y th e d if f e r e n c e e q u a tio n s , *
( 1 1 .5 .7 )
A p { z)
v (n ) = - ' ^ T a p (k )v (n - k) + x (n ) k= 1
< 7
y (n ) =
( 1 1 .5 .8 )
J 2 c ^ ( f t v (n
o b ta in e d b y v ie w in g th e s y s te m a s a c a s c a d e o f a n a ll-p o le s y s t e m f o llo w e d b y a n a ll- z e r o s y s te m . F r o m ( 1 1 .5 .8 ) w e o b s e r v e th a t th e o u tp u t y ( n ) is s im p ly a lin e a r c o m b in a tio n o f d e la y e d o u tp u ts fr o m th e a ll- p o le s y s te m . S in c e z e r o s r e s u lt fr o m fo r m in g a lin e a r c o m b in a tio n o f p r e v io u s o u tp u ts , w e c a n c a r r y o v e r th is o b s e r v a t io n t o c o n s t r u c t a p o l e - z e r o s y s t e m u s in g t h e a ll- p o le
Sec. 11.5
AR Lattice and ARMA Lattice-Ladder Fitters
879
l a t t ic e s t r u c tu r e a s th e b a s ic b u ild in g b lo c k . W e h a v e c le a r ly o b s e r v e d th a t gm(n) in th e a ll-p o le l a t t ic e c a n b e e x p r e s s e d a s a lin e a r c o m b in a tio n o f p r e s e n t a n d p a s t o u tp u ts . In f a c t , th e s y s te m ( 1 1 .5 .9 )
is a n a ll-z e r o s y s te m . a ll- z e r o f ilte r .
T h e r e f o r e , a n y lin e a r c o m b in a tio n o f { m(n ) } is a ls o a n
L e t u s b e g in w ith a n a ll- p o le l a t t ic e filte r w ith c o e f f ic i e n t s K m, 1 < m < p , a n d a d d a la d d e r p a r t by ta k in g a s th e o u tp u t, a w e ig h te d l i n e a r c o m b in a tio n o f { m (n )}. T h e r e s u lt is a p o l e - z e r o filte r t h a t h a s th e la t t ic e -la d d e r s t r u c tu r e s h o w n in F ig . 1 1 .7 . I t s o u tp u t is
v (n ) = 2 2 p k g k ( n )
( 1 1 .5 .1 0 )
w h e r e { / y a r e th e p a r a m e te r s th a t d e te r m in e th e z e r o s o f th e s y s te m . T h e s y s te m fu n c tio n c o r r e s p o n d in g t o ( 1 1 .5 .1 0 ) is
( 1 1 .5 .1 1 )
Inpul x(n) Stage
fp-M )
fp -iW
f iW
/o(") Stage
o(n)
Stage
8P- ()_
P~ 1 gP- 2< ) |( )
/>()
'p
O utput
(a) P o le -z e ro system
f m - 1<")
.-I (b) m1 * 1stage of lattice
Figure 11.7 Lattice-ladder structure for pole-zero system
880
Chap. 11
S in c e X ( z ) = F p (z) a n d F q ( z ) = G o (z ), ( 1 1 .5 .1 1 ) , c a n b e e x p r e s s e d a s -2 P*
G t i z ) F o (z)
k= 0
G0(z) Fp{z)
( 1 1 .5 .1 2 )
a p ( z) *=o
T h e re fo re , < ? C 9 (z) = f t * ( z ) t=o ( 1 1 .5 .1 3 )
T h is is th e d e s ire d r e la tio n s h ip th a t c a n b e u se d t o d e te r m in e th e w e ig h tin g c o e f f ic ie n ts {fik} as p r e v io u s ly s h o w n in S e c t i o n 7 .3 .5 . G iv e n th e p o ly n o m ia ls C q ( z ) a n d A p (z), w h e r e p > q , t h e r e f le c tio n c o e f f ic ie n ts { , } a r e d e te r m in e d firs t fr o m th e c o e f f ic ie n t s { ^ ( / t ) } . B y m e a n s o f th e s te p -d o w n r e c u r s iv e r e la tio n g iv e n b y ( 1 1 .2 .2 2 ) , w e a ls o o b t a i n th e p o ly n o m ia ls
B k ( z ) , k = 1 , 2 , . . . , p . T h e n th e la d d e r p a r a m e t e r s c a n b e o b t a i n e d f r o m ( 1 1 .5 .1 3 ) ,
w h ic h c a n b e e x p r e s s e d as
m -1
C , (z ) = ^
to
P k B k iz ) + fimB m(z)
( 1 1 .5 .1 4 )
= C m- i ( z ) + fimB m(z) o r , e q u iv a le n tly ,
C m- i ( z ) = C m(z) - P mB m{z)
m = p, p - 1 , . . . , 1
( 1 1 .5 .1 5 )
B y ru n n in g th is r e c u r s iv e r e la tio n b a c k w a r d , w e c a n g e n e r a t e a ll th e lo w e r - d e g r e e p o ly n o m ia ls , C m( z ), m =
p 1 .......... 1. S in c e bm(m ) = 1, t h e p a r a m e t e r s f5m a r e
d e te r m in e d fr o m ( 1 1 .5 .1 5 ) b y s e tt in g
p m = c m(m )
m = p , p - 1 ------, 1 , 0
W h e n e x c ite d b y a w h ite n o is e s e q u e n c e , th is la t t i c e - l a d d e r f i l t e r s tr u c tu r e g e n e r a t e s a n A R M A ( p , ^ ) p r o c e s s th a t h a s a p o w e r d e n s ity s p e c tr u m
\C q( f ) \
IV / ) l a n d a n a u t o c o r r e la tio n fu n c t io n th a t s a tis fie s ( 1 1 .1 .1 8 ) , w h e r e a 2 in th e v a ria n c e o f t h e in p u t w h ite n o is e s e q u e n c e .
11.6 WIENER FILTERS FOR FILTERING AND PREDICTION

I n m a n y p r a c tic a l a p p lic a t io n s w e a r e g iv e n a n in p u t s ig n a l { * ( ) } , c o n s is tin g o f th e s u m o f a d e s ir e d s ig n a l {.s(/i)} a n d a n u n d e s ir e d n o is e o r i n t e r f e r e n c e {u>(n)}, a n d w e a r e a s k e d t o d e s ig n a f itte r th a t s u p p r e s s e s th e u n d e s ir e d i n te r f e r e n c e
Sec. 11.6
Wiener Filters for Filtering and Prediction

d (n )
881
----------- t { + )-------< ()
Figure l l j t problem c o m p o n e n t.
Model for linear estimation
In s u c h a c a s e , th e o b je c t i v e is t o d e s ig n a s y s te m t h a t f ilte r s o u t
th e a d d itiv e in t e r f e r e n c e w h ile p r e s e r v in g th e c h a r a c t e r is t ic s o f th e d e s ir e d s ig n a l {* ( )} In th is s e c t io n w e t r e a t th e p r o b le m o f s ig n a l e s tim a tio n in th e p r e s e n c e o f a n a d d itiv e n o is e d is tu r b a n c e . T h e e s tim a to r is c o n s t r a in e d to b e a li n e a r f ilte r F ig u r e 1 1 .8 illu s tr a te s th e lin e a r e s tim a tio n w ith im p u ls e r e s p o n s e {/ i(n )}, d e s ig n e d s o th a t its o u tp u t a p p r o x im a te s s o m e s p e c ifie d d e s ire d s ig n a l s e q u e n c e \d ( n ) } . p r o b le m . T h e in p u t s e q u e n c e t o th e f ilte r is x ( n ) = s ( n ) + w (n ), a n d its o u tp u t s e q u e n c e is v (n ). T h e d iffe r e n c e b e tw e e n t h e d e s ir e d s ig n a l a n d th e filte r o u tp u t is th e e r r o r s e q u e n c e e(n ) = d (n ) y (n ) . W e d is tin g u is h t h r e e s p e c ia l c a s e s : 1 . I f d {n ) = s ( n ) , th e lin e a r e s tim a tio n p r o b le m is r e f e r r e d to as, filte rin g . 2 . I f d (n ) = s {n + D ), w h e r e D > 0 , th e li n e a r e s tim a tio n p r o b le m is r e f e r r e d to a s s ig n a l p re d ic t io n . N o te t h a t th is p r o b le m is d if f e r e n t th a n th e p r e d ic tio n c o n s id e r e d e a r l i e r in th is c h a p te r , w h e r e d {n ) = x i n + D ) , D > 0 . 3 . I f d {n ) s ( n D ) , w h e r e D > 0 , th e li n e a r e s tim a tio n p r o b le m is r e f e r r e d t o a s s ig n a l s m o o th in g . O u r tr e a t m e n t w ill c o n c e n t r a t e o n filte r in g a n d p r e d ic t io n . T h e c r ite r io n s e le c te d f o r o p tim iz in g t h e - f i l t e r im p u ls e r e s p o n s e {/t(n)} is th e m in im iz a t io n o f th e m e a n -s q u a r e e r r o r . T h is c r it e r i o n h a s th e a d v a n ta g e s o f s im p lic ity a n d m a th e m a tic a l tr a c ta b ility . T h e b a s ic a s s u m p tio n s a r e t h a t th e s e q u e n c e s { $ ( ) } , { a n d {/(n)} a r e z e r o m e a n a n d w id e -s e n s e s ta tio n a r y . T h e lin e a r f ilte r w ill b e a s s u m e d to b e e it h e r F I R o r I I R . I f it is I I R , w e a s s u m e th a t th e in p u t d a ta (jc (n )} a r e a v a ila b le o v e r th e in f in it e p a s t. W e b e g in w ith t h e d e s ig n o f t h e o p tim u m F I R filte r . T h e o p t i m u m lin e a r f ilte r , in t h e s e n s e o f m in im u m m e a n - s q u a r e e r r o r ( M M S E ) , is c a lle d a W ie n e r filte r.
11.6.1 FIR Wiener Filter

S u p p o s e th a t t h e filte r is c o n s t r a in e d t o b e o f le n g th M w ith c o e f f ic ie n ts {A*, 0 <
k <
M - l ) .
H e n c e its o u tp u t y ( n ) d e p e n d s o n th e fin ite d a ta r e c o r d x ( n ) ,
882 x (n 1 ) , , x (n M + 1 ),
Linear Prediction and Optimum Linear Fitters
Chap. 11
Af -1
y (n ) = t=o
h ( k )x (n - k)
( 1 1 .6 .1 )
T h e m e a n -s q u a r e v a lu e o f th e e r r o r b e tw e e n th e d e s ir e d o u tp u t d ( n ) a n d y in ) is
Z M = E \e { n ) \2 ^
= ( 1 1 .6 .2 )
d (n ) k =o
h ( k ) x ( n k)
S in c e th is is a q u a d r a tic f u n c t io n o f th e filte r c o e f f ic ie n t s , th e m in im iz a t io n o f u y ie ld s th e s e t o f lin e a r e q u a tio n s M -l 5 > ( * W *=o / - ) = ^ ,(/ ) / = 0 , 1 .......... M
- l
( 1 1 .6 .3 )
w h e r e yxx(k) is th e a u t o c o r r e l a t io n o f th e in p u t s e q u e n c e { * ( ) } a n d ydx(k) =
E [ d ( n ) x * ( n fc)] is th e c r o s s c o r r e l a t i o n b e tw e e n th e d e s ir e d s e q u e n c e [d (n )} a n d
th e in p u t s e q u e n c e { x {n ), 0 < n < M 1 ). T h e s e t o f lin e a r e q u a t i o n s th a t s p e c ify th e o p tim u m filte r is c a lle d th e W ie n e r - H o p f e q u a tio n . T h e s e e q u a tio n s a r e a ls o c a lle d th e n o r m a l e q u a tio n s , e n c o u n te r e d e a r l ie r in th e c h a p t e r in th e c o n t e x t o f lin e a r o n e - s te p p r e d ic t io n . In g e n e r a l, th e e q u a tio n s in ( 1 1 .6 .3 ) c a n b e e x p r e s s e d in m a tr ix f o r m a s r wh = y d w h ere r ( 1 1 .6 .4 )
is a n M x M ( H e r m i t i a n ) T o e p lit z m a tr ix w ith e l e m e n t s T/* = Y * A l ~ k )
a n d y d is th e M x 1 c r o s s c o r r e l a t io n v e c t o r w ith e l e m e n t s Yd x(0 , i = 0 , 1 , . . . , M - \ . T h e s o lu tio n f o r th e o p tim u m f ilte r c o e f f ic ie n ts is hopt = Y M y d a n d th e r e s u ltin g m in im u m M S E a c h ie v e d b y th e W i e n e r f ilte r is ( 1 1 .6 .5 )
M -l
M M S E M = m in M = a 2 d h o r , e q u iv a le n tly ,
M M S E M = a 2 - y J T ^ y d
ôpt( * ) ? ( ) 5
( 1 1 .6 .6 )
( 1 1 .6 .7 )
w h ere a j = E \d ( n ) \2 . L e t us c o n s id e r s o m e s p e c ia l c a s e s o f ( 1 1 .6 .3 ) . I f w e a r e d e a lin g w ith filte rin g , t h e d (n ) = s ( n ) . F u r t h e r m o r e , i f s ( n ) a n d w (n ) a r e u n c o r r e la te d r a n d o m s e q u e n c e s , a s is u s u a lly th e c a s e in p r a c t ic e , th e n
Yxx(k) =
y ss(k) + Yww(k)
(11.6.8)
Ydx(k)
Yss(k)
Sec. 11.6
Wiener Fitters for Filtering and Prediction
883
a n d th e n o r m a l e q u a tio n s in ( 1 1 .6 .3 ) b e c o m e
M -l
Y , W
i= 0
) M
- V
+ Y w w V - k ) ] = y ss(l)
1 = 0 , 1 .......... M
-l
( 1 1 .6 .9 )
I f w e a r e d e a lin g w ith p r e d ic t io n , th e n d (n ) = s (n + D ) w h e r e D s u m in g th a t s ( n ) a n d w (n ) a r e u n c o r r e la te d ra n d o m s e q u e n c e s , w e h a v e
> 0. A s
Y d ,(k ) = y ss(l + D )
H e n c e th e e q u a tio n s f o r th e W ie n e r p r e d ic t io n f ilt e r b e c o m e
( 1 1 .6 .1 0 )
M-l
* ( * ) [ y ( / - * ) + yv.tt. ( / - * ) ] = y</ + z>) i= 0 /= 0,1...., w - l ( 1 1 .6 .1 1 )
In all th e s e c a s e s , th e c o r r e la t io n m a tr ix t o b e in v e r t e d is T o e p lit z . H e n c e th e ( g e n e r a liz e d ) L e v in s o n - D u r b in a lg o r ith m m a y b e u s e d to s o lv e f o r th e o p tim u m f ilte r c o e ffic ie n ts .
Example 11.6.1 Lei us consider a signal x( n ) = j(n) + w(n), where s(n) is an AR(1) process that satisfies the difference equation
s(n ) = 0.6.v(n 1) + u(n)
where lu(n)l is a white noise sequence with variance <tl 2 = 0.64, and (u>(n)) is a white noise sequence with variance en = 1. We will design a Wiener filter of length M = 2 to estimate (j(h)}. Solution Since (j(n)) is obtained by exciting a single-pole filter by white noise, the power spectra! density of is
r ( / ) = < r ;\H { f ) \2
0.64
| 1 0.6e~-i*nf \ 2'
0.64 1.36 1.2cos2jt/ The corresponding autocorrelation sequence (y(m)} is

Y,Am ) = (0.6)|m |
The equations for the filter coefficients are

2h(0) + 0.6A (1) = 1
0 .6 A (0 )+ 2 A (1 ) = 0 .6
Solution of these equations yields the result
h(0) = 0.451
h(l) = 0.165
884
Chap. 11
The corresponding minimum MSE is MMSE2 = 1 - A(0)y(0) - /i(l)>v,(l) = 1 - 0.451 - (0.165X0.6) = 0.45 This error can be reduced further by increasing the length of the Wiener filter (see Problem 11.27).
11.6.2 Orthogonality Principle in Linear Mean-Square Estimation

T h e n o r m a l e q u a tio n s f o r th e o p tim u m filte r c o e f f ic ie n t s g iv e n b y ( 1 1 .6 .3 ) c a n b e o b ta in e d d ir e c tly b y a p p ly in g th e o r th o g o n a lity p r in c ip le in lin e a r m e a n - s q u a r e e s tim a tio n . S im p ly s ta te d , th e m e a n - s q u a r e e r r o r m in ( 1 1 .6 .2 ) is a m in im u m if th e f ilte r c o e f f ic ie n ts {/ ;()) a r e s e le c t e d s u ch th a t th e e r r o r is o r th o g o n a l t o e a c h o f th e d a ta p o in ts in th e e s tim a te ,
E \e ( n ) x * ( n /)] = 0
w h ere
M -l
/ = 0 , 1 .......... M 1
( 11.6 .12)
( 1 1 .6 .1 3 )
e (n ) = d (n )
i =o
h ( k ) x ( n k)
C o n v e r s e ly , i f th e f ilte r c o e f f ic ie n t s s a tis fy ( 1 1 .6 .1 2 ) , th e r e s u ltin g M S E is a m in i m um . W h e n v ie w e d g e o m e tr ic a lly , th e o u tp u t o f t h e filte r , w h ic h is th e e s tim a te

M -l
d (n )
*=o
h (k )x (n k)
(1 1 .6 .1 4 )
is a v e c to r in th e s u b s p a c e s p a n n e d b y th e d a ta { * ( & ) , 0 < k < M 1 }. T h e e r r o r
e (n ) is a v e c to r f r o m d (n ) to <?(n) [ i.e ., d (n ) = e ( n ) + d (n )], a s s h o w n in F ig . 1 1 .9 .

T h e o r th o g o n a lity p r in c ip le s ta te s th a t th e le n g th M = d a ta p o in t x ( k ) , 0 < k < M 1 ], W e n o te t h a t th e s o lu tio n o b t a in e d fr o m th e n o r m a l e q u a t i o n s in ( 1 1 .6 .3 ) is u n iq u e i f th e d a ta { * ( ) } in th e e s t im a t e d (n ) a r e lin e a r ly in d e p e n d e n t. In th is c a s e , th e c o r r e la t io n m a tr ix is n o n s in g u la r . O n th e o t h e r h a n d , i f th e d a ta a r e lin e a r ly d e p e n d e n t, th e r a n k o f T M is le s s th a n M a n d t h e r e f o r e th e s o lu tio n is n o t u n iq u e . In th is c a s e , th e e s t im a t e d ( n ) c a n b e e x p r e s s e d a s a lin e a r c o m b in a tio n o f a r e d u c e d s e t o f lin e a r ly in d e p e n d e n t d a ta p o in t s e q u a l to th e r a n k o f T w . S in c e th e M S E is m in im iz e d b y s e le c tin g t h e f ilte r c o e f f i c i e n t s to s a tis fy th e o r th o g o n a lity p r in c ip le , th e r e s id u a l m in im u m M S E is s im p ly E M / i) !2 is a m in im u m w h e n e (n ) is p e r p e n d ic u la r t o th e d a ta s u b s p a c e [ i.e ., e(rt) is o r th o g o n a l to e a c h
MMSE*f = E[e(n)d*()]
(11.6.15)
which yields the result given in (11.6.6).
Sec. 11.6
885
Figure 11.9
Geometric interpretation
of linear M S E problem
11.6.3 IIR Wiener Filter

In th e p r e c e d in g s e c tio n w e c o n s t r a in e d th e f ilte r t o b e F I R a n d o b ta in e d a s e t o f
M lin e a r e q u a tio n s f o r th e o p tim u m filte r c o e f f ic ie n ts . In th is s e c tio n w e a llo w th e

f ilte r to b e in fin ite in d u ra tio n ( I I R ) a n d th e d a ta s e q u e n c e to b e in fin ite a s w e ll. H e n c e th e f ilte r o u tp u t is y ( ) = / 2 ( ) * ( n - k)
( 11 .6 . 16 )
T h e f ilt e r c o e f f ic ie n ts a r e s e le c te d to m in im iz e th e m e a n -s q u a r e e r r o r b e tw e e n th e d e s ir e d o u tp u t d (n ) a n d v ( n ). th a t is,
x =
E \ e { n ) \2
2 ( 1 1 .6 .1 7 )
A p p lic a tio n o f th e o r th o g o n a lity p r in c ip le le a d s to th e W i e n e r - H o p f e q u a
tion.
( 1 1 .6 .1 8 ) T h e r e s id u a l M M S E is s im p ly o b t a in e d b y a p p lic a t io n o f th e c o n d itio n g iv e n b y ( 1 1 .6 .1 5 ) . T h u s w e o b ta in M M S E s c = m in ^ = aj f c opt ( * ) > ( ) ( 1 1 .6 .1 9 )
T h e W i e n e r - H o p f e q u a tio n g iv e n b y ( 1 1 .6 .1 8 ) c a n n o t b e s o lv e d d ir e c tly w ith z - tr a n s f o r m te c h n iq u e s , b e c a u s e th e e q u a tio n h o ld s o n ly f o r 1 > 0 . W e s h a ll s o lv e
886
Chap. 11
f o r th e o p tim u m I I R W ie n e r filte r b a s e d o n th e in n o v a tio n s r e p r e s e n t a t io n o f th e s ta tio n a r y r a n d o m p r o c e s s { * ( ) } . R e c a l l th a t a s ta tio n a r y r a n d o m p r o c e s s U ( n ) ) w ith a u t o c o r r e l a t i o n yxx(k) a n d p o w e r s p e c tr a l d e n s ity r ( / ) c a n b e r e p r e s e n te d b y a n e q u iv a le n t in n o v a tio n s p r o c e s s , ( (n )} b y p a s s in g { * ( ) } th r o u g h a n o is e -w h ite n in g f ilt e r w ith s y ste m fu n c t io n 1 / G ( z ) , w h e r e G ( z ) is th e m in im u m -p h a s e p a r t o b t a in e d f r o m th e s p e c tr a l fa c to r iz a tio n o f r xJt(z ): r (z ) = < T ,2 G ( z ) G ( z - 1) H e n c e G i z ) is a n a ly tic in t h e r e g io n jz| > n , w h e r e n < \ . N o w , th e o p tim u m W i e n e r f i lt e r c a n b e v ie w e d a s th e c a s c a d e o f th e w h ite n in g filte r 1 / G (z ) w ith a s e c o n d f ilte r , s a y Q ( z ) , w h o s e o u tp u t v ( ) is id e n tic a l to th e o u tp u t o f t h e o p tim u m W i e n e r filte r . S in c e
OC
( 1 1 .6 .2 0 )
j( n ) = *=o
<?(*)/( - k)
( 1 1 ,6 .2 1 )
a n d e(rt) = d { n ) > ( n ), a p p lic a t io n o f th e o r th o g o n a lity p r in c ip le y ie ld s th e n e w W i e n e r - H o p f e q u a tio n as

OC
22<}(k'>y"V - k ) = ydi(D
/>
( 11.6. 22)
B u t s in c e { i( n ) ) is w h ite , it fo llo w s th a t y(/ k ) = 0 u n le s s / = k. T h u s w e o b ta in th e s o lu tio n as
q d ) = ZZLLL =
Ya(0)
Yd, (I)
Ydi(0
/> 0
( 1 1 .6 .2 3 )
T h e z - tr a n s f o r m o f t h e s e q u e n c e l g ( i ) } is
Q iz ) = Y l q ( k ) z ~ k k 0
=
( 1 1 .6 .2 4 )
\
a <
Y lr d i( lc ) z ~ k
t=o
I f w e d e n o te th e z -tr a n s fo r m o f th e tw o -s id e d c r o s s c o r r e la tio n s e q u e n c e Y d iik ) by r\ ,,(z ):

OC
r di( z ) = a n d d e fin e [r\/,-(z)]+ a s
Y i
* = -0 0
W *( * ) * *
( 1 1 .6 .2 5 )
[ I \ ,,( z ) ] + = 2 2 Y * W z ~ k *=o
( 1 1 .6 .2 6 )
then
Q iz ) = ~ [ r d iiz ) ) + of
( 1 1 .6 .2 7 )
Sec. 11.6
887
T o d e te r m in e [r</;(z)]+ , w e b e g in w ith th e o u tp u t o f th e n o is e - w h it e n in g f ilte r , w h ic h c a n b e e x p r e s s e d a s

OC
i{ n ) = ^
*=o
v ( );c (n k)
( 1 1 .6 .2 8 )
w h e r e (u (Jt), k > 0 } is t h e im p u ls e r e s p o n s e o f th e n o is e -w h it e n in g filte r , ( 1 1 .6 .2 9 ) Then
Yd i{k) = E [ r f ( n ) i * ( n - * ) ]
OC
2 2 v ( m ) E [ d { n ) x * ( n m Jt)]
( 1 1 .6 .3 0 )
22 v (m ')ydAk
m=0
+m )
T h e z - tr a n s fo r m o f th e c r o s s c o r r e la tio n Y d iik ) is
oc oc
r ji( z ) =
2 2
' Y 2 v ^ Y d x { k + m)
k= oo m=0
= u ( m ) 2 2 Y d x (k + m )z~ m=0 *= -o c
oc oc
( 1 1 .6 .3 1 )
u (m )z m 2 2
Y * * (k )z~ k r dx(z)
=
T h e re fo re ,
V iz -'W iA z ) =
G i z - 1)
Q iz ) = A
r^ (z ) I
( 1 1 .6 .3 2 )
F in a lly , th e o p tim u m I I R W ie n e r f ilt e r h a s th e s y s te m f u n c t io n
H op\ iz ) =
Q iz ) G iz )
( 1 1 .6 .3 3 )
In s u m m a r y , th e s o lu tio n f o r th e o p tim u m I I R W i e n e r f ilte r r e q u ir e s t h a t w e p e r fo r m a s p e c t r a l fa c to r iz a tio n o f e x a m p le illu s tr a te s th e p r o c e d u r e . (z ) t o o b ta in G i z ) , th e m in im u m -p h a s e c o m p o n e n t , a n d th e n w e s o lv e f o r th e c a u s a l p a r t o f T dxi z ) / G i z ~ 1 ). T h e f o llo w in g
888
Chap. 11
Example 11.6.2
L e t us d eterm in e the optim um I I R W ien e r filter fo r th e signal given in E x a m p le 11.6.1.
Solution
F o r this signal we have r ^ ) - r w 1.8(1 - k - ' X l - h ) + i = - _ ^ . T . ( 1 - (|^ .
w here a f = 1.8 and 1 - |z-1 ca) = rro fe ^ T h e z-transform o f th e crossco rrelation ydx (m ) is r dJ(z) = r ( z ) H en ce 0 .6 4 (1 - 0 . 6 z " > ) ( l - 0 . 6 z ) 0 .64 z X l - 0 . 6 ; - 1) 0 .2 6 6 ; i _ i zj
Lg(z-)J+ L a - 1
rrâ)! =r__
0.8
[ 1 - 0.6z~'
0.8
1 - 0 .6 z - ]
T h e optim um I I R filter has th e system function
H -'w = r8(TTTFrj(rro^j
and an im pulse response
* o p t( n ) =
/ l-0.6r'\
0. 8
5( j ) "
n >
W e c o n c lu d e th is s e c t io n b y e x p r e s s in g th e m in im u m M S E g iv e n b y ( 1 1 .6 .1 9 ) in te r m s o f th e f r e q u e n c y -d o m a in c h a r a c te r is tic s o f th e filte r . F i r s t , w e n o t e th a t
a j = E\ d{ n) \ 2 is s im p ly th e v a lu e o f th e a u t o c o r r e l a t i o n s e q u e n c e \ydd(k)} e v a lu
a te d a t it = 0. S in c e
Vdd(k) = ^ y ^ r ^ ( z ) z * ^ z
it fo llo w s th a t
( 1 1 .6 .3 4 )
a ] = ydd(0 ) =
( 1 1 .6 .3 5 )
w h e r e t h e c o n t o u r in te g r a l is e v a lu a te d a lo n g a c lo s e d p a th e n c i r c l i n g th e o rig in in th e r e g io n o f c o n v e r g e n c e o f
Sec. 11.6
889
T h e s e c o n d te r m in ( 1 1 .6 .1 9 ) is a ls o e a s ily tr a n s f o r m e d t o t h e f r e q u e n c y d o m a in b y a p p lic a t io n o f P a r s e v a ls t h e o r e m . S in c e h opl( k ) = 0 f o r k < 0 , w e h a v e
cc
fto p ,( * ) } ( * ) =
^ S fio p d z W .A z - ^ d z
J *
( 1 1 .6 .3 6 )
w h e r e C is a c lo s e d c o n to u r e n c ir c lin g th e o r ig in t h a t lie s w ith in t h e c o m m o n r e g io n o f c o n v e r g e n c e o f / / optU ) a n d th e M M S E c c in th e fo r m M M S E ^ = ~ < ^ [ r dd(z) E x a m p le 1 1 .6 3
r dx(z~l).
B y c o m b in in g ( 1 1 .6 .3 5 ) w ith ( 1 1 .6 .3 6 ) , w e o b ta in th e d e s ir e d e x p r e s s io n f o r
H opd z W dA z - l ) ] z - l d z
( 1 1 .6 .3 7 )
For the optimum Wiener filter derived in Example 11.6.2, the minimum MSE is
m m sem =
2 x iJ f
i fi r
-J)(l-0.fe)J
o 35ss
There is a single pole inside the unit circle at z = j. By evaluating the residue at the pole, we obtain MMSE^ - 0.444 We observe that this MMSE is only slightly smaller than that for the optimum two-tap Wiener filter in Example ] 1.6.1.
11.6.4 Noncausal Wiener Filter

I n th e p r e c e d in g s e c tio n w e c o n s t r a in e d th e o p tim u m W i e n e r f ilte r t o b e c a u s a l [ i.e ., h opt(n) = 0 f o r n < 0 ]. In th is s e c t io n w e d r o p th is c o n d itio n a n d a llo w th e f ilte r t o in c lu d e b o th th e in fin ite p a s t a n d th e in fin ite f u tu r e o f th e s e q u e n c e { * ( ) } in f o r m in g th e o u tp u t y ( n ) , th a t is, O O y( n) =
2 2
oc
h ( k ) x (n k)
( 1 1 .6 .3 8 )
T h e re s u ltin g filte r is p h y s ic a lly u n r e a liz a b le . I t c a n a ls o b e v ie w e d a s a s m o o th in g
filt e r in w h ic h th e in fin ite fu tu r e s ig n a l v a lu e s a r e u s e d t o s m o o th t h e e s tim a te d (n ) = y ( n ) o f th e d e s ir e d s ig n a l d { n ) .

A p p lic a tio n o f th e o r th o g o n a lity p r in c ip le y ie ld s th e W i e n e r - H o p f e q u a tio n f o r t h e n o n c a u s a l filte r in th e fo r m O O
2 2 k= oc
a n d t h e r e s u ltin g M M S E nf as
- k ) = y dx( l )
- oo < I <
00
( 1 1 .6 .3 9 )
O O M M S E nc =
22
Jt=-00
h (* ) r lc ^
d 1 -6 -4 )
890
Chap. 11
S in c e ( 1 1 .6 .3 9 ) h o ld s f o r oo < I < oo, th is e q u a tio n c a n b e tr a n s f o r m e d d ir e c tly to y ie ld th e o p tim u m n o n c a u s a l W i e n e r filte r as ^ (z ) = p ^ ~ ( 1 1 .6 .4 1 )
I jrjrv)
T h e M M S E ( c a n a ls o b e s im p ly e x p r e s s e d in t h e z -d o m a in a s M M S E nf = J - ^ [ r dd(z) -
H nc( z ) r dA t - l ) ] z - l d z
( 1 1 .6 .4 2 )
In th e fo llo w in g e x a m p le w e c o m p a r e th e fo r m o f th e o p tim a l n o n c a u s a l filte r w ith th e o p tim a l c a u s a l f ilt e r o b ta in e d in th e p r e v io u s s e c tio n .
Example 11.6.4 The optimum noncausal Wiener filter for the signal characteristics given in Exam ple 11.6.1 is given by (11.6.41), where 0.64
r^ a> F-"(
and
(1 - 0 .6 ; - )(l -0.6c)
r ( s ) = r (;) + l 2(1 - 0 .3 --1 - 0.3;)

(1 - 0 . 6 - - ' )(1 - 0.6c)
Then. 0.3555
H ", U ~ ( l - i c - ' K l - i c )
This filter is clearly noncausal. The minimum MSE achieved by this filter is determined from evaluating (11.6.42). The integrand is
The only pole inside the unit circle is c = |. Hence the residue is 0.35551 0.3555 = 0.40
Hence the minimum achievable MSE obtained with the optimum noncausal Wiener filter is MMSEnr = 0.40 Note that this is lower than the MMSE for the causal filler, as expected.

T h e m a jo r f o c a l p o in t in th is c h a p t e r is th e d e s ig n o f o p tim u m l i n e a r s y s te m s fo r lin e a r p r e d ic t io n a n d f ilte r in g . f ilte r o u tp u t. T h e c r i t e r i o n f o r o p tim a lity is t h e m in im iz a tio n o f th e m e a n -s q u a r e e r r o r b e t w e e n a s p e c ifie d d e s ir e d f ilt e r o u tp u t a n d t h e a c tu a l
Sec. 11.7
891
I n th e d e v e lo p m e n t o f lin e a r p r e d ic t io n , w e d e m o n s tr a t e d th a t th e e q u a tio n s f o r th e fo r w a r d a n d b a c k w a r d p r e d ic t io n e r r o r s s p e c ifie d a la ttic e f ilte r w h o s e p a r a m e te r s , th e r e fle c tio n c o e f fic ie n ts [ K m), w e r e s im p ly r e la te d to th e f ilte r c o e f fic ie n ts (a m(A:)} o f t h e d ir e c t fo r m F I R lin e a r p r e d ic t o r a n d th e a s s o c ia t e d p r e d ic t io n e r r o r filte r . T h e o p tim u m f ilte r c o e f f ic ie n ts { K m) a n d (a m( ) } a r e e a s ily o b ta in e d f r o m th e s o lu tio n o f th e n o r m a l e q u a tio n s . W e d e s c r ib e d tw o c o m p u ta tio n a lly e ff ic ie n t a lg o r it h m s f o r s o lv in g th e n o r m a l e q u a tio n s , th e L e v in s o n - D u r b in a lg o r ith m a n d th e S c h u r a lg o r it h m . B o t h a lg o H ow rith m s a r e s u ita b le f o r s o lv in g a T o e p lit z s y s te m o f li n e a r e q u a tio n s a n d h a v e a c o m p u ta tio n a l c o m p le x ity o f 0 ( p 2 ) w h e n e x e c u te d o n a s in g le p r o c e s s o r . e v e r , w ith fu ll p a r a lle l p r o c e s s in g , th e S c h u r a lg o r ith m s o lv e s th e n o r m a l e q u a tio n s in O ( p ) tim e , w h e r e a s th e L e v in s o n - D u r b in a lg o r ith m r e q u ir e s O ( p l o g p ) tim e . In a d d itio n t o th e a ll- z e r o la ttic e filte r r e s u ltin g f r o m lin e a r p r e d ic t io n , w e a ls o d e riv e d th e A R la t t ic e ( a ll- p o le ) filte r s tr u c tu r e a n d th e A R M A l a ttic e - la d d e r ( p o l e - z e r o ) filte r s tr u c tu r e . F in a lly , w e d e s c r ib e d th e d e s ig n o f th e c la s s o f o p t i m u m lin e a r filte r s , c a lle d W ie n e r filte r s . L in e a r e s tim a tio n th e o r y h a s h a d a lo n g a n d r ic h h is to r y o f d e v e lo p m e n t o v e r th e p a s t fo u r d e c a d e s . K a ila th (1 9 7 4 ) p r e s e n ts a h is to r ic a l a c c o u n t o f th e firs t t h r e e d e c a d e s . T h e p io n e e r in g w o rk o f W ie n e r ( 1 9 4 9 ) o n o p tim u m lin e a r filte r in g f o r s ta tis tic a lly s ta tio n a r y s ig n a ls is e s p e c ia lly s ig n ific a n t. b y K a lm a n ( 1 9 6 0 ) a n d K a lm a n a n d B u c y ( 1 9 6 1 ) . T h e g e n e r a liz a tio n o f th e W i e n e r filte r t h e o r y t o d y n a m ic a l s y s te m s w ith ra n d o m in p u ts w a s d e v e lo p e d K a lm a n filte r s a r e tr e a t e d in The Tu th e b o o k s b y M e d itc h ( 1 9 6 9 ) , B r o w n ( 1 9 8 3 ) , a n d C h u i a n d C h e n ( 1 9 8 7 ) . m o n o g r a p h b y K a ila t h ( 1 9 8 1 ) tr e a ts b o th W i e n e r a n d K a lm a n f ilte r s . T h e r e a r e n u m e r o u s r e f e r e n c e s o n lin e a r p r e d ic t io n a n d la ttic e f ilte r s . t o r ia l t r e a t m e n t s o n t h e s e s u b je c t s h a v e b e e n p u b lis h e d in th e jo u r n a l p a p e r s b y M a k h o u l ( 1 9 7 5 , 1 9 7 8 ) a n d F r ie d la n d e r (1 9 8 2 a , b ) . T h e b o o k s b y H a y k in ( 1 9 9 1 ) , M a r k e l a n d G r a y 1 9 7 6 ) , a n d T r e t t e r ( 1 9 7 6 ) p r o v id e c o m p r e h e n s iv e tr e a t m e n t s o f th e s e s u b je c t s . A p p lic a tio n s o f li n e a r p r e d ic t io n t o s p e c tr a l a n a ly s is a r e fo u n d in th e b o o k s b y K a y ( 1 9 8 8 ) an d M a r p le ( 1 9 8 7 ) , to g e o p h y s ic s in th e b o o k R o b in s o n a n d T r e i t e l ( 1 9 8 0 ) , a n d to a d a p tiv e filte r in g in th e b o o k b y H a y k in ( 1 9 9 1 ) . T h e L e v in s o n - D u r b in a lg o r ith m f o r s o lv in g th e n o r m a l e q u a tio n s r e c u r s iv e ly w a s g iv e n b y L e v in s o n ( 1 9 4 7 ) a n d la t e r m o d ifie d b y D u r b in ( 1 9 5 9 ) . b y D e l s a r t e a n d G e n i n ( 1 9 8 6 ) a n d b y K r is h n a ( 1 9 8 8 ) . o f 2 in th e n u m b e r o f m u ltip lic a tio n s . T h e S c h u r a lg o r it h m w as o r ig in a lly d e s c r ib e d b y S c h u r ( 1 9 1 7 ) in a p a p e r p u b lis h e d in G e r m a n . n o m ia ls A n E n g lis h tr a n s la tio n o f th is p a p e r a p p e a r s in th e b o o k e d ite d b y G o h b e r g ( 1 9 8 6 ) . T h e S c h u r a lg o r ith m is in tim a te ly r e la te d t o th e p o ly w h ich c a n b e i n te r p r e te d a s o r th o g o n a l p o ly n o m ia ls . A t r e a t m e n t o f o r t h o g o n a l p o ly n o m ia ls is g iv e n in th e b o o k s b y S z e g o ( 1 9 6 7 ) , G r e n a n d e r a n d S z e g o ( 1 9 5 8 ) , a n d G e r o n im u s ( 1 9 5 8 ) . T h e th e s is o f V i e i r a ( 1 9 7 7 ) a n d th e p a p e r s b y K a ila t h e t a l. ( 1 9 7 8 ) , D e l s a r t e e t a l. ( 1 9 7 8 ) , a n d Y o u l a a n d K a z a n jia n ( 1 9 7 8 ) V a r ia t io n s o f th is c la s s ic a l a lg o r it h m , c a lle d s p lit L e in s o n a lg o r it h m s , h a v e b e e n d e v e lo p e d T h e s e a lg o r it h m s e x p lo it a d d itio n a l s y m m e tr ie s in th e T o e p l it z c o r r e la tio n m a tr ix a n d s a v e a b o u t a f a c t o r
892
Chap. 11
p r o v id e a d d itio n a l re s u lts o n o r th o g o n a l p o ly n o m ia ls .
K a il a t h ( 1 9 8 5 , 1 9 8 6 ) p r o
v id e s t u t o r ia l t r e a t m e n t s o f th e S c h u r a lg o r it h m a n d its r e la tio n s h ip to o r th o g o n a l p o ly n o m ia ls a n d t h e L e v in s o n - D u r b in a lg o r ith m . T h e p ip e lin e d p a r a lle l p r o c e s s in g s tr u c tu r e f o r c o m p u tin g t h e r e fle c tio n c o e f f ic ie n t s b a s e d o n t h e S c h u r a lg o r ith m a n d th e r e la te d p r o b le m o f s o lv in g T o e p l it z s y s te m s o f lin e a r e q u a t i o n s is d e s c r ib e d in th e p a p e r b y K u n g a n d H u ( 1 9 8 3 ) . F in a lly , w e s h o u ld m e n t i o n t h a t s o m e a d d itio n a l c o m p u ta tio n a l e f fic ie n c y c a n b e a c h ie v e d in th e S c h u r a lg o r it h m , b y fu r t h e r e x p lo itin g s y m m e tr y p r o p e r tie s o f T o e p l it z m a t r ic e s , a s d e s c r i b e d b y K r is h n a ( 1 9 8 8 ) . T h is le a d s t o t h e s o -c a lle d s p lit- S c h iir a lg o r it h m , w h ic h is a n a lo g o u s to th e s p lit- L e v in s o n a lg o r ith m .
PROBLEMS
11.1 T h e pow er density spectrum o f an A R process [x (n )) is given as r jr M = g- |A{oj)|2 25 | 1 - e~>w + i e - J '2" ! 2 w h ere al, is th e v arian ce o f th e input seq uen ce. (a) D ete rm in e th e d ifferen ce equation fo r g en eratin g the A R p ro ce ss w hen the e x citatio n is w hite noise. (b) D eterm in e th e system function fo r th e w hitening filter. 1 L 2 A n A R M A p rocess has an au to co rrelatio n [yIX (m)} w hose z-tran sfo rm is given as
(a) (b)
D ete rm in e th e filter H ( z ) for g en eratin g (x (n )) fro m a w hite n o ise input seq uen ce. Is H ( z ) un iqu e? E xp lain . D ete rm in e a stab le lin ear w hitening filter for th e seq u en ce {*())
1 L 3 C o n sid er th e A R M A p ro cess g en erated by th e d ifferen ce eq u a tio n
x(n ) = 1.6x(n 1) 0 .6 3 * ( - 2) +
+ 0.9w (n - 1)
(a) (b)
D eterm in e th e system fun ction o f th e w hitening filter and its p o les and zeros. D ete rm in e th e pow er density spectrum o f {jc(n )}.
11.4 D ete rm in e the lattice coefficien ts correspond ing to th e F I R filter w ith system function
H(z) = Ai(z) = 1 + j j z ' 1 + f z~2 + jz~ 3

1 L 5 D ete rm in e th e reflection co efficie n ts { K m} o f the lattice filter co rresp o n d in g to the F I R filter described by the system fun ction
Chap. 11
Problems
893
1 1 .6 (a) D ete rm in e the zeros and sk etch the zero p attern fo r the F I R la ttice filter with reflectio n co efficien ts
Ki = \
K2 = - \
*3 = 1
(b ) R e p e a t part (a ) but with Ky = 1. (c ) Y o u should have found that th e zeros lie on th e unit circle. C an this result be g en eralized ? H ow ? 1 1 .7 D ete rm in e th e im pulse response o f th e F I R filter th at is described by the lattice co efficien ts K \ = 0.6, K 2 = 0.3, K 3 = 0 .5 , and K A = 0.9. 1 1 .8 In S ec tio n 11.2.4 we indicated th at th e n oise-w hitening filter A p{z ) for a causal A R ( p ) process is a forw ard lin ear p red iction -erro r filter o f o rd e r p. Show th a t th e backw ard lin ear p red ictio n -erro r filter o f o rd er p is the n oise-w hiten in g filter o f th e co rresp o nd ing an ticau sal A R (p) process. 1 1 .9 U se th e o rth og o nality principie to d eterm in e the norm al eq u atio ns and the resulting m inim um M S E fo r a forw ard p red ictor o f ord er p th at p red icts m sam ples {m > 1) into th e futu re (m -step forw ard p red icto r). S k etc h th e p red iction erro r filter. 11 .1 0 R e p e a t P rob lem 11.9 fo r an m -step backw ard predictor. 11.11 D ete rm in e a L e v in so n -D u rb in recu rsive algorithm fo r solving fo r the coefficien ts o f a backw ard pred iction -erro r filter. U se th e result to show th at co efficien ts o f th e forw ard and backw ard predictors can be exp ressed recu rsively as
1 1 .1 2 T h e L e v in so n -D u rb in algorithm d escrib ed in S ectio n 11.3.1 solved the lin ear eq u a tions r a m= - y m where the right-hand side o f this eq u atio n has elem en ts o f th e au toco rrelatio n se qu en ce th a t are also elem en ts o f th e m atrix T . problem o f solving th e lin ear eq uations r mb m w here cm is an arb itrary v ector. (T h e v ecto r bm is n ot related to th e coefficien ts o f th e backw ard p red icto r.) Show th at th e solu tion to r mb = c can b e obtain ed from a generalized L e v in s o n -D u rb in algorithm which is given recursively as L e t us co n sid er th e m ore general
bm(k) =
- bm{m )a' ^(m - k)
m = 1 ,2 ........p
'" "
w here M l ) = c ( l) / y j,( 0 ) = c(1)/q and a(k) is given by (11,3.17). T hu s a second recu rsion is required to solve th e eq u atio n I^ b * , = c * . 1 L 1 3 U se th e g eneralized L e v in so n -D u rb in algorithm to solve th e norm al eq uations recu r sively fo r th e m -step forw ard and backw ard p redictors.
894
Chap. 11
11.14 Show that the transform ation
in the Sch ur algorithm satisfies the special property V ,JV = (1 - \ K J 2) i w here
T hu s V m is called a 7 -ro ta tio n m atrix. Its ro le is to ro tate o r h y p erbolate th e row o f G m to lie alon g the first co o rd in ate d irectio n (K a ila th , 1985). 11.15 Prov e the additional p rop erties (a ) through (1) o f th e p red iction - e rro r filters given in Sectio n 11.4. 11.16 E x te n d the add ition al prop erties (a ) through (1) o f th e p red iction erro r filters given in Sectio n 11.4 to com plex-valued signals. 1 1.17 D eterm in e th e reflection co efficie n t K 3 in term s o f th e a u to co rrela tio n s {yTr (m )} from th e Schur algorithm and co m p are your result with the ex p ressio n fo r K 3 o btain ed from th e L e v in so n -D u rb in algorithm . 1 1.18 C onsid er a in finite-length (p = oo) o n e-step forw ard p red ictor fo r a statio nary random process (x(n )) with a pow er density spectru m o f Show th a t th e m ean -squ are erro r o f th e p red ictio n -erro r filter can b e expressed as
1 1 .1 9 D ete rm in e th e output o f an in finite-len gth (p = oo) m -step forw ard p red ictor and th e resulting m ean -squ are e rro r w hen th e input signal is a first-o rd er autoregressive process o f the form
x(n ) = ax(n 1) + w(n)

1 L 2 0 A n A R (3 ) process {x(n)} is ch aracterized by the au toco rrelatio n seq u en ce yJX (0) = 1, (a ) U se the Schtir algorithm to d eterm in e the th ree reflection co efficien ts K u K 2, and (b ) Sk etch th e lattice filter fo r synthesizing {x (n )} from a white n oise ex citation . 1 L 2 1 T h e purpose o f this p roblem is to show th at the polynom ials (A (z )}, which a re the system functions o f th e forw ard p red ictio n -erro r filters o f o rd e r m, m = 0 , 1 , . . . , p, can be in terp reted as o rth og o n al on th e unit circle. T ow ard th is end, suppose that r j x ( f ) is the pow er spectral density o f a zero-m ean rand om p ro cess (* { n ) } and let M m (z)}, m = 0, 1 . . . . , p ), b e the system functions o f th e co rresp o n d in g prediction erro r filters. Show th at th e polyn o m ials {A m(z)} satisfy the o rth og o n ality property
1 L 2 2 D eterm in e th e system fun ction o f th e all-p ole filter d escrib ed by th e la ttice co efficien ts J fi = 0.6, K 2 = 0 .3 , K } = 0 .5 , and K A = 0.9.
Chap. 11
Problems
895
11.23 Determine the parameters and sketch the lattice-ladder filter structure for the system with system function
H{Z)
11.24 Consider a signal
jc ( )
1 + 0 .1 ;- i- 0 .7 2 ;- 2
= s(n) + u ;(n), where s(n) is an A R ( 1 ) process that satisfies

s (n ) = Q .8s(n 1) + v(n)
the difference equation
where {u(n)) is a white noise sequence with variance er2 = 0.49 and {w(n)) is a white noise sequence with variance er2 = 1. The processes {t>{n)) and {u)(n)J are uncorrelated. (a) Determine the autocorrelation sequences {y(m)) and {yXJ(m)]. ( b ) Design a Wiener filter of length M = 2 to estimate |j{n)}. (c) Determine the MMSE for M = 2. 11.25 Determine the optimum causal IIR Wiener filter for the signal given in Problem 11.24 and the corresponding MMSE^. 11.26 Determine ihe system function for the noncausal IIR Wiener filter for the signal given in Problem 11.24 and the corresponding MMSEnr. 11.27 Determine the optimum FIR Wiener filter of length M 3 for the signal in Ex ample 11.6.1 and the corresponding MMSEj. Compare MMSEj with MMSEi and comment on the difference. 11.28 An AR(2) process is defined by the difference equation = x(n 1) 0.6jr(n 2) + u>(ri) where {u>(n)} is a white noise process with variance <t2, Use the Yule-Walker equa tions to solve for the values of the autocorrelation y(0), ^ x(l). and yxx(2). 11.29 An observed random process (*()) consists of the sum of an AR{j>) process of the form
p
and a white noise process {u>{n)} with variance tr2. The random process {u()} is also white with variance er2. The sequences {u(n)} and fio(n)} are uncorrelated. Show that the observed process = j() + u>()} is ARM A(p, p) and de termine the coefficients of the numerator polynomial (MA component) in the corre sponding system function.
a 12
th e f r e q u e n c y d o m a in .
Power Spectrum Estimation
In th is c h a p te r w e a r e c o n c e r n e d w ith th e e s tim a tio n o f th e s p e c t r a l c h a r a c te r is tic s o f s ig n a ls c h a r a c te r iz e d as r a n d o m p r o c e s s e s . M a n y o f th e p h e n o m e n a th a t o c c u r in n a tu r e a r e b e s t c h a r a c te r iz e d s ta tis tic a lly in te r m s o f a v e r a g e s . a r e b e s t c h a r a c te r iz e d s ta tis tic a lly a s ra n d o m p r o c e s s e s . s ig n a ls th a t a r e w e ll m o d e le d a s ra n d o m p r o c e s s e s . D u e to th e r a n d o m f lu c tu a tio n s in s u ch s ig n a ls , w e m u s t a d o p t a s ta tis ti c a l v ie w p o in t, w h ich d e a ls w ith th e a v e r a g e c h a r a c t e r is t ic s o f r a n d o m s ig n a ls . In p a r tic u la r , th e a u t o c o r r e la tio n fu n c t io n o f a r a n d o m p r o c e s s is th e a p p r o p r ia te s ta tis tic a l a v e r a g e th a t w e w ill u se fo r c h a r a c te r iz in g ra n d o m s ig n a ls in th e tim e d o m a in , a n d th e F o u r i e r tr a n s fo r m o f th e a u t o c o r r e la t io n f u n c t io n , w h ic h y ie ld s t h e p o w e r d e n s ity s p e c tr u m , p r o v id e s th e t r a n s fo r m a tio n fr o m t h e tim e d o m a in t o P o w e r s p e c tr u m e s tim a tio n m e t h o d s h a v e a r e la tiv e ly lo n g h is to r y . For a F o r e x a m p le , m e t e o r o lo g ic a l p h e n o m e n a s u c h a s th e flu c tu a tio n s in a ir t e m p e r a t u r e a n d p r e s s u re T h e r m a l n o is e v o lta g e s g e n e r a te d in r e s is to r s a n d e le c t r o n i c d e v ic e s a r e a d d itio n a l e x a m p le s o f p h y s ic a l
h is to r ic a l p e r s p e c tiv e , th e r e a d e r is r e f e r r e d to th e p a p e r b y R o b i n s o n ( 1 9 8 2 ) a n d t h e b o o k b y M a r p le ( 1 9 8 7 ) . O u r t r e a t m e n t o f th is s u b je c t c o v e r s t h e c la s s ic a l p o w e r s p e c tr u m e s tim a tio n m e th o d s b a s e d o n th e p e r io d o g r a m , o r ig in a lly in tr o d u c e d by S c h u s t e r ( 1 8 9 8 ) , a n d b y Y u l e ( 1 9 2 7 ) , w h o o r ig in a te d t h e m o d e m m o d e l- b a s e d o r p a r a m e t r ic m e th o d s . T h e s e m e t h o d s w e r e s u b s e q u e n tly d e v e lo p e d a n d a p p lie d b y W a lk e r (1 9 3 1 ), B a r tle tt (1 9 4 8 ), P a rz e n (1 9 5 7 ), B la c k m a n an d T u k e y (1 9 5 8 ), B u rg ( 1 9 6 7 ) , a n d o th e r s . W e a ls o d e s c r ib e t h e m e t h o d o f C a p o n ( 1 9 6 9 ) a n d m e t h o d s b a s e d o n e ig e n a n a ly s is o f th e d a ta c o r r e l a t io n m a tr ix .
12.1 ESTIMATION OF SPECTRA FROM FINITE-DURATION OBSERVATIONS OF SIGNALS

T h e b a s ic p r o b le m th a t w e c o n s id e r in th is c h a p t e r is th e e s t im a t io n o f th e p o w e r d e n s ity s p e c tr u m o f a s ig n a l fr o m th e o b s e r v a t io n o f th e s ig n a l o v e r a fin ite tim e in te r v a l. A s w e w ill s e e , th e fin ite r e c o r d le n g th o f t h e d a ta s e q u e n c e is a m a jo r
896
Sec. 12.1
Estimation of Spectra from Finite-Duration Observations of Signals
897
lim ita tio n o n th e q u a lity o f th e p o w e r s p e c tr u m e s t i m a t e . e s t im a t e t h a t c a n b e e x t r a c t e d fr o m th e d a ta . e s tim a te t h e s p e c tr u m .
W h e n d e a lin g w ith
s ig n a ls t h a t a r e s ta tis tic a lly s ta tio n a r y , th e lo n g e r th e d a ta r e c o r d , t h e b e t t e r th e O n t h e o t h e r h a n d , i f th e s ig n a l s ta tis tic s a r e n o n s ta tio n a r y , w e c a n n o t s e le c t a n a r b itr a r ily lo n g d a ta r e c o r d to In s u c h a c a s e , th e le n g th o f t h e d a ta r e c o r d t h a t w e s e l e c t is d e te r m in e d b y th e ra p id ity o f th e tim e v a r ia tio n s in th e s ig n a l s ta tis tic s . U l t i m a t e l y , o u r g o a l is t o s e le c t a s s h o r t a d a ta r e c o r d a s p o s s ib le t h a t s till a llo w s u s t o r e s o lv e t h e s p e c tr a l c h a r a c te r is tic s o f d if f e r e n t s ig n a l c o m p o n e n ts in th e d a ta r e c o r d t h a t h a v e c lo s e ly s p a c e d s p e c tr a . O n e o f t h e p r o b le m s th a t w e e n c o u n t e r w ith c la s s ic a l p o w e r s p e c tr u m e s t im a tio n m e t h o d s b a s e d o n a fin ite -le n g th d a ta r e c o r d is t h e d is to r tio n o f th e s p e c tr u m t h a t w e a r e a tte m p tin g t o e s tim a te . T h i s p r o b le m o c c u r s in b o th th e c o m p u ta tio n o f t h e s p e c tr u m f o r a d e te r m in is tic s ig n a l a n d th e e s t im a t io n o f th e p o w e r s p e c tr u m o f a r a n d o m s ig n a l. S in c e it is e a s ie r to o b s e r v e t h e e f f e c t o f th e fin ite le n g th o f th e d a ta r e c o r d o n a d e te r m in is tic s ig n a l, w e t r e a t th is c a s e firs t. T h e r e a f t e r , w e c o n s id e r o n ly r a n d o m s ig n a ls a n d th e e s tim a tio n o f t h e i r p o w e r s p e c tr a .
12.1.1 Computation of the Energy Density Spectrum

L e t us c o n s id e r th e c o m p u ta tio n o f th e s p e c tr u m o f a d e te r m in is tic s ig n a l f r o m a f in ite s e q u e n c e o f d a ta . T h e s e q u e n c e jc (n ) is u s u a lly t h e re s u lt o f s a m p lin g a c o n tin u o u s - tim e s ig n a l x a (t) a t s o m e u n ifo rm s a m p lin g r a t e F s . O u r o b je c t i v e is to o b t a in a n e s tim a te o f t h e tr u e s p e c tr u m fr o m a f in ite - d u r a tio n s e q u e n c e x ( n ) . R e c a l l t h a t if x ( t ) is a fin ite -e n e r g y s ig n a l, t h a t is,
CC
\xa (t)\2d t < oo

-00
th e n its F o u r i e r t r a n s f o r m e x is ts a n d is g iv e n a s
X a(F ) =
F r o m P a r s e v a ls t h e o r e m w e h a v e oo
f
J -OO
Xo( t ) e - j2 * F ,d t
\xa ( t ) \2dt =
roc / \ X a ( F ) l 2d F
J oc
( 1 2 .1 .1 )
00
T h e q u a n t it y |Xa (F )| 2 r e p r e s e n ts th e d is tr ib u tio n o f s ig n a l e n e r g y a s a f u n c tio n o f fr e q u e n c y , a n d h e n c e it is c a lle d th e e n e r g y d e n s ity s p e c tr u m o f th e s ig n a l, t h a t is,
SXA F ) = \ X e ( F ) \ 2
( 1 2 .1 .2 )
a s d e s c r ib e d in C h a p te r 4 . T h u s t h e to t a l e n e r g y in t h e s ig n a l is s im p ly th e in te g r a l o f SXX{ F ) o v e r a ll F [ i.e ., th e to t a l a r e a u n d e r SJtJt( F ) ] . I t is a ls o in te r e s tin g t o n o te t h a t S XX( F ) c a n b e v ie w e d a s t h e F o u r i e r t r a n s fo rm o f a n o t h e r fu n c t io n ,
R IX { t ) ,
c a lle d
th e
a u to c o rr e la tio n fu n c t io n o f th e
898
fin ite - e n e r g y s ig n a l x a (t), d e fin e d as
Chap. 12
RxA*)= f
I n d e e d , it e a s ily fo llo w s th a t
J OC
x * { t) x a { t + x ) d t
( 1 2 .1 .3 )
R x x ( r ) e - J2jrFrd x = SXA F ) = l * ( F ) | 2 J C s o t h a t R x x (?) a n d SXX( F ) a r e a F o u r i e r t r a n s fo r m p a ir.
( 1 2 .1 .4 )
N o w s u p p o s e th a t w e c o m p u te th e e n e r g y d e n s ity s p e c tr u m o f th e s ig n a l x a (t ) f r o m its s a m p le s t a k e n a t th e r a t e F s s a m p le s p e r s e c o n d . T o e n s u r e th a t th e r e is n o s p e c tr a l a lia s in g re s u ltin g f r o m th e s a m p lin g p r o c e s s , th e s ig n a l is a s s u m e d to b e p r e filte r e d , s o th a t, f o r p r a c tic a l p u r p o s e s , its b a n d w id th is lim ite d to B h e r tz . T h e n th e s a m p lin g fr e q u e n c y F s is s e le c te d s u c h th a t F s > 2 B . T h e s a m p le d v e r s io n o f x a (t) is a s e q u e n c e * { ) , - o c < n < o c , w h ich h a s a F o u r i e r tr a n s fo r m ( v o lta g e s p e c tr u m )
OC
X { a i) =
o r, e q u iv a le n tly ,
ns-cc
22
A! = CC
22
x (n )e ~ J<"
X (f) =
X ( n ) e - j2xfn
( 1 2 .1 .5 )
R e c a l l th a t X { f ) c a n b e e x p r e s s e d in te r m s o f th e v o lta g e s p e c tr u m o f th e a n a lo g s ig n a l jco ( 0 as * ( ) =
X ( F ~ k F ^
( l 2 1 '6)
w h e r e / = F / F s is th e n o r m a liz e d f r e q u e n c y v a r ia b le . I n th e a b s e n c e o f a lia s in g , w ith in th e f u n d a m e n ta l ra n g e l^ l < F J 2 , w e h a v e
X a ( F ) , |Fj < F J 2
( 1 2 .1 .7 )
H e n c e th e v o lta g e s p e c tru m o f th e s a m p le d s ig n a l is id e n tic a l to t h e v o lta g e s p e c tr u m o f th e a n a lo g s ig n a l. A s a c o n s e q u e n c e , th e e n e r g y d e n s it y s p e c tr u m o f th e s a m p le d s ig n a l is

2
= F } \X a ( F ) \2
( 1 2 .1 .8 )
W e c a n p r o c e e d f u r t h e r b y n o tin g t h a t th e a u t o c o r r e la t io n o f t h e s a m p le d s ig n a l, w h ich is d e fin e d as O O
rxx ( k ) =
2 2
# *= 00
* * (
+ *)
(1 2 - 1 .9 )
h a s th e F o u r i e r tr a n s fo r m ( W i e n e r - K h i n t c h i n e t h e o r e m )
S (/)=
j t
r*Ak )e~j2*kf
(12.1.10)
Sec. 12.1
899
H e n c e th e e n e rg y d e n s ity sp e c tru m c a n b e o b ta in e d by th e F o u rie r tra n s fo rm o f th e a u to c o rre la tio n o f th e se q u e n c e {Jt(n)}. T h e r e la tio n s a b o v e le a d us to d istin g u ish b e tw e e n tw o d istin ct m e th o d s fo r co m p u tin g th e en e rg y d e n s ity sp e c tru m o f a signal x(t) fro m its sa m p le s Jt(n). O n e is th e direct method, w hich involves co m p u tin g th e F o u rie r tra n s fo rm o f {*()], a n d th e n
SxA f ) = l * ( / ) l 2
2 O O 2 2 X( n) e~j2*fn
( 12.1. 11)
T h e se co n d a p p ro a c h is c a lle d th e indirect method b e c a u s e it re q u ire s tw o ste p s. F irst, th e a u to c o r re la tio n rxx(k) is c o m p u te d fro m x(rt) a n d th e n th e F o u rie r tr a n s fo rm o f th e a u to c o rre la tio n is c o m p u te d as in (12.1.10) to o b ta in th e e n e rg y d e n sity sp e ctru m . In p ra c tic e , h o w e v e r, only th e fin ite -d u ra tio n se q u e n c e x (n ), 0 < n < N - 1 , is a v ailab le fo r c o m p u tin g th e sp e c tru m o f th e signal. In effe ct, lim itin g th e d u ra tio n o f th e se q u e n c e x ( n ) t o N p o in ts is e q u iv a le n t to m u ltip ly in g x ( n ) b y a re c ta n g u la r w ind o w . T h u s w e h av e 0 < n < N 1 o th e rw ise
( 12. 1. 12)
F ro m o u r d iscu ssio n o f F I R filter design b ased o n th e u se o f w indow s to lim it th e d u ra tio n o f th e im p u lse re sp o n s e , w e recall th a t m u ltip lic a tio n o f tw o se q u e n c e s is e q u iv a le n t to co n v o lu tio n o f th e ir v o ltag e sp e c tra . C o n s e q u e n tly , th e freq u e n cy d o m a in r e la tio n c o rre sp o n d in g to (12.1.12) is
X i f ) = * ( / ) * W{ f )
(12.1.13)
- 1/2
R e call fro m o u r d iscussion in S ectio n 8.2.1 th a t c o n v o lu tio n o f th e w indow f u n c tio n W ( f ) w ith X ( f ) s m o o th s th e sp e c tru m X ( f ) , p ro v id e d th a t th e sp e c tru m W ( f ) is re la tiv e ly n a rro w c o m p a re d to X ( / ) . B u t th is c o n d itio n im plies th a t th e w in d o w w i n ) b e su fficiently long (i.e., N m u st b e sufficiently la rg e ) such th a t W ( f ) is n a rro w c o m p a re d to X ( / ) . E v e n if W ( f ) is n a rro w c o m p a re d to X ( f ) , th e co n v o lu tio n o f X ( f ) w ith th e sid e lo b e s o f W ( f ) re su lts in sid e lo b e en erg y in X ( f ) , in fre q u e n c y b a n d s w h e re th e tr u e signal sp e c tru m X ( / ) = 0. T h is sid e lo b e en e rg y is called leakage. T h e follow ing ex a m p le illu stra te s th e le a k a g e p ro b le m . Example 12.1.1 A signal with (voltage) spectrum l/l< 0 .1 otherwise is convolved with the rectangular window of length N = 61. D eterm ine the spectrum
o f X ( f ) g iv e n b y (1 2 .1 .1 3 ).
900
Chap. 12
Figure 1Z1 Spectrum obtained by convolving an M = 61 rectan g u lar window with the ideal lowpass spectrum in Exam ple 12.1.1.
Solution The spectral characteristic W ( f ) for the length N = 61 rectangular window is illustrated in Fig. 8.2(b). Note that the width of the main lobe of the window function is Aw = 4^/61 or A/ = 2/61, which is narrow compared to X ( f ) . The convolution of X ( f ) with W ( f ) is illustrated in Fig. 12.1. We note that energy has leaked into the frequency band 0.1 < | / | < 0.5. where X ( f ) = 0. A part of this is due to the width of the main lobe in VV(/), which causes a broadening or smearing of X ( f ) outside the range \ f \ < 0.1. However, the sidelobe energy in X ( f ) is due to the presence of the sidelobes of W (/), which are convolved with X ( f ) . The smearing of X ( f ) for | / | > 0.1 and the sidelobes in the range 0.1 < | / | < 0.5 constitute the leakage. Ju s t as in th e case o f F IR filter d esig n , w e can re d u c e sid e lo b e le a k a g e by selectin g w in d o w s th a t h av e low sid e lo b e s. T h is im p lies th a t th e w in d o w s h av e a sm o o th tim e -d o m a in cu to ff in ste a d o f th e a b ru p t cu to ff in th e re c ta n g u la r w indow . A lth o u g h such w in d o w fu n ctio n s re d u c e sid e lo b e le a k a g e , th ey re s u lt in an in crease in sm o o th in g o r b ro a d e n in g o f th e s p e c tra l c h a ra c te ristic X ( f ) . F o r ex am p le, the use o f a B lac k m a n w in d o w o f le n g th N = 61 in E x a m p le 12.1.1 re su lts in th e sp e c tra l c h a ra c te ristic X ( / ) sh o w n in F ig. 12.2. T h e sid e lo b e le a k a g e h as c e rta in ly b e e n re d u c e d , b u t th e sp e c tra l w id th h a s b e e n in c re a se d by a b o u t 5 0 % . T h e b ro a d e n in g o f th e s p e c tru m b e in g e s tim a te d d u e to w in d o w in g is p a rtic u larly a p ro b le m w h e n w e w ish to reso lv e signals w ith closely sp a c e d fre q u e n c y co m p o n e n ts . F o r ex a m p le , th e signal w ith sp e ctral c h a ra c te ristic X ( / ) = { f ) + X i { f ), as sh o w n in Fig. 12.3, c a n n o t b e reso lv ed as tw o se p a ra te sig n a ls u n le ss th e w idth o f th e w in d o w fu n c tio n is significantly n a r ro w e r th a n th e fre q u e n c y se p a ra tio n A f . T h u s w e o b se rv e th a t u sing sm o o th tim e -d o m a in w indow s re d u c e s le ak ag e a t th e ex p en se o f a d e c re a s e in fre q u e n c y re so lu tio n .
Sec. 12.1
901
Figure 12.2 Spectrum obtained by convolving an M = 61 B lackm an window w ith the ideal lowpass spectrum in Exam ple 12.1.1.
I t is c le a r fro m th is discussion th a t th e e n e rg y d en sity s p e c tru m o f th e w in d o w ed se q u e n c e {*()} is an a p p ro x im a tio n o f th e d e s ire d sp e c tru m o f th e se q u en ce {jc(n)J. T h e sp e c tra l d en sity o b ta in e d fro m {Jc()} is
2
Sisif) = W ) \ 2 =
j2nfn
(12.1.14)
T h e sp e c tru m given by (12.1.14) can be c o m p u te d n u m erically a t a set of N fre q u e n c y p o in ts by m e a n s o f th e D F T . T h u s X ( k ) = 2 2 , x ( n ) e - J2nkn/N ji=Q (12.1.15)
Then
|X(fc)|2 = Si-Af)\j-k/N = Sri
(12.1.16)
902
Chap. 12
and hence
j2 n k n /N
(12.1.17)
w hich is a d isto rte d v ersio n o f th e tru e s p e c tru m Sxx( k / N ) .
12.1.2 Estimation of the Autocorrelation and Power Spectrum of Random Signals: The Periodogram
T h e fin ite-en erg y signals c o n s id e re d in th e p re c e d in g se c tio n possess a F o u rie r tra n sfo rm an d a re c h a ra c te riz e d in th e sp e c tra l d o m a in by th e ir e n e rg y d en sity sp e ctru m . O n th e o th e r h a n d , th e im p o rta n t class o f signals c h a ra c te riz e d as s ta tio n a ry ra n d o m p ro c e sse s d o n o t h a v e finite e n e rg y a n d h e n c e d o n o t po ssess a F o u rie r tran sfo rm . S uch signals h a v e finite a v e ra g e p o w e r a n d h en c e a re c h a ra c te riz e d by a p o w e r density spect rum. If x ( t ) is a s ta tio n a ry ra n d o m p ro cess, its a u to c o rre la tio n fu n ctio n is yxx(r) = E [ x * ( t ) x ( t +
t )]
(12.1.18)
w h ere [] d e n o te s th e sta tistical a v e ra g e . T h e n , via th e W ie n e r-K h in tc h in e th e o re m , th e p o w e r d e n sity sp e c tru m o f th e sta tio n a ry ra n d o m p ro c e ss is th e F o u rie r tra n sfo rm o f th e a u to c o rre la tio n fu n c tio n , th a t is, C / O yxA T ) e ~ j2* FTd t (12.1.19) -oc In p ra c tic e , w e d e a l w ith a single re a liz a tio n o f th e ra n d o m p ro cess fro m w hich w e e stim a te th e p o w e r sp e c tru m o f th e p ro cess. W e d o n o t k n o w th e tru e a u to c o rre la tio n fu n ctio n yxx( t ) a n d as a c o n se q u e n c e , w e c a n n o t c o m p u te th e F o u rie r tra n sfo rm in (12.1.19) to o b ta in rjx(F). O n th e o th e r h a n d , fro m a single re a liz a tio n o f th e ra n d o m p ro c e ss w e c a n c o m p u te th e tim e -a v e ra g e a u to c o rre la tio n fu n ctio n R x A r ) = r ^ r f x *{ t ) x ( t + x) dt J - T0 (12.1.20)
w h ere 27o is th e o b se rv a tio n in te rv a l. I f th e sta tio n a ry r a n d o m p ro cess is ergodic in th e first a n d se co n d m o m e n ts (m e a n a n d a u to c o rre la tio n fu n c tio n ), th e n X ut(r) = lim R x x i r )
To-*oo
lim
To-+oc
27b /_ jc*(r)jc (r + r )dt

To
7b
(12.1.21)
T h is re la tio n ju stifies th e u se o f th e tim e -a v e ra g e a u to c o rre la tio n Rxx( r ) as a n e stim a te o f th e sta tistical a u to c o r re la tio n fu n c tio n yxx( r ) . F u rth e rm o r e , th e F o u rie r tra n sfo rm o f R xx( r ) p ro v id e s a n e s tim a te PXX( F ) o f th e p o w e r d en sity
Sec. 12.1
903
s p e c tru m , th a t is,
7b
Pxx(F) = [ R xx{ T ) e - j:brFrd r J- T q 7 b

i frTo T r r rr7b * .
1
e-pxFx
= t t / / x ^ x ^ + T^d t -^ o J-T q U - T q
1 I f T
dT
(12.1.22)
= r / x ( t ) e - * F' d t -iio \J-To T h e a c tu a l p o w e r d e n sity sp e c tru m is th e e x p e c te d v a lu e o f PXX( F) in th e lim it as 7b - oo, r (F ) = lim [ / > ( ) ] T il- oo = lim j J I/ To-ôc a i
[ / :;
(12.1.23)
x( t ) e j27rFdt
F ro m (12.1.20) an d (12.1.22) w e again n o te th e tw o p o ssib le a p p ro a c h e s to co m p u tin g PXX(F) , th e d ire c t m e th o d as given by (12.1.22) o r th e in d ire c t m e th o d , in w hich w e o b ta in ^ ( r ) first a n d th e n c o m p u te th e F o u rie r tra n sfo rm . W e sh all c o n s id e r th e e stim a tio n o f th e p o w e r d e n s ity sp e c tru m fro m sa m p le s o f a sin g le re a liz a tio n o f th e ra n d o m process. In p a rtic u la r, w e assu m e th a t x a ( t ) is sa m p le d a t a ra te Fs > 2 B, w h e re B is th e h ig h est fre q u e n c y c o n ta in e d in th e p o w e r d e n sity sp e c tru m o f th e ra n d o m p ro cess. T h u s w e o b ta in a fin ite -d u ra tio n s e q u e n c e x ( n ), 0 < n < N 1, by sa m p lin g x a(t). F ro m th e s e sa m p le s w e can co m p u te th e tim e -a v e ra g e a u to c o rre la tio n se q u e n c e
j A '- m - l
r' (m) = ---------N m

x
Y x*(n)x{n+m) n=Q
N- 1
m = 0 ,1 ,
1 (12.1.24)
A A
= T, -----r 7 A/ lm I
Jl
x (n)x(n+m)
m = -1 , - 2 ,..., 1 - N
a n d th e n c o m p u te th e F o u rie r tra n sfo rm

A '- l
P^ ( f ) =
E r'x x { m ) e - M * m= N+l
(12.1.25)
T h e n o rm a liz a tio n fa c to r N |m | in (12.1.24) re su lts in an e s tim a te w ith m e a n v a lu e ^ N- m- l
E [r x ' A m )] = r; r~,
N ~
22
to
E[x*(n)x{n + m)]
(12.1.26)
= Yxx(m) w h e re yxx( m ) is th e tr u e (statistical) a u to c o rre la tio n s e q u e n c e o f x ( n) . H e n c e rxx(m) is a n u n b ia s e d e s tim a te o f th e a u to c o rre la tio n fu n c tio n yxx( m). T h e v a ria n c e
904
Chap. 12
o f th e e stim a te rxx(m) is a p p ro x im a te ly N v a r [r i j ( m )] ^ ^
_ |m |]2
22
[\Yx*(n)\2 + y * A n - m ) y xx(n + m) ]
(12.1.27)
w hich is a resu lt given by J e n k in s a n d W a tts (1968). C learly,

N oc
lim v a r [r ' (m)] = 0
(12.1.28)
p ro v id e d th a t 3C 22
n -oc
\Yxx(n)\2 < oo
S ince E[r'xx{m)\ = y xx(m) a n d th e v a ria n c e o f th e e stim a te c o n v e rg e s to z e ro as N oc, th e e stim a te r'xx(m) is said to b e consistent. F o r larg e v alu es o f th e lag p a r a m e te r m, th e e s tim a te r'xx(m) g iv en by (12.1.24) h as a larg e v arian ce, esp ecially as m a p p ro a c h e s N . T h is is d u e to th e fact th a t few er d a ta p o in ts e n te r in to th e e stim a te fo r larg e lags. A s an a lte rn a tiv e to (12.1.24) w e can u se th e e s tim a te I w - -i rxx{m) = Y jt*(n)jt(ii + m) N t=o | N-1 ^ r(m ) = y x *( n ) x ( n + m) N n=\m\ 0 < m < JV 1 (12.1.29) m = 1. 2 ........ 1 N
w hich h a s a b ias o f \m\yxx( m ) / N , since its m e a n v alu e is

| A '- m - l
E[rIX(m)] -
22
E[ x* ( n) x ( n + m)] (12.1.30)
_ _
n=0 W -_M yjrjr(m) , , = ^ = _
| m |j\ yx A m )
H o w e v e r, th is e stim a te h a s a sm a ller v arian ce, given a p p ro x im a te ly as 1 1 X 1 v a r[rJJt(m )] 2 2 [Xr.r(n)|2 + Yx x (n - m ) y xx(n + m )]

n ~oo
(12.1.31)
W e o b se rv e th a t rxx(m) is asympt ot i cal l y u n b i a s e d , th a t is,
H-+x
lim E[rxx{m)] = yxx(m)
(12.1.32)
an d its v arian ce c o n v erg es to z e ro as N -* oo. T h e re fo re , th e e s tim a te rxx(m) is also a consi st ent est imate o f yxx(m). W e sh all u se th e e stim a te rxx( m ) given by (12.1.29) in o u r tr e a tm e n t o f p o w e r sp e c tru m e s tim a tio n . T h e c o rre sp o n d in g e s tim a te o f th e p o w e r d e n s ity sp e c tru m is
N -1
/ (/)=
22
m = - ( N - 1)
r* A m ) e - M m
(12.1.33)
Sec. 12.1
905
I f w e su b s titu te fo r rxx(m) fro m (I2 .I.2 9 ) in to (12.1.33), th e e s tim a te Pxx( f ) can also b e e x p re ss e d as 2 Pxx ( / ) N
-jln fn
(12.1.34)
w h e re X ( / ) is th e F o u rie r tra n sfo rm o f th e sam p le se q u e n c e x{n) . T h is w ell kno w n fo rm o f th e p o w e r d en sity sp e c tru m e s tim a te is called th e periodogram. I t w as o rig in ally in tro d u c e d by S c h u ste r (1898) to d e te c t a n d m e a s u re h id d e n p e rio d ic itie s in d a ta . F ro m (12.1.33), th e av e ra g e v alu e o f th e p e rio d o g ra m e s tim a te PxA f ) is
22
m = - ( N - 1)
rx A m ) e ,2jrfn
J2
m= (/V 1)
E[rxA m ) ] e - * " '
E [ p xx( f ) ] =
22
f 1 J r )
'
Y xim )e
J2nfm
(12.1.35)
m = - ( N - 1) ^
T h e in te r p re ta tio n th a t w e give to (12.1.35) is th a t th e m e a n o f th e e stim a te d sp e c tru m is th e F o u rie r tra n sfo rm o f th e w in d o w ed a u to c o rre la tio n fu n ctio n Yxx(m)
Y xx(m )
(12.1.36)
w h e re th e w in d o w fu n c tio n is th e (tria n g u la r) B a rtle tt w in d o w . H e n c e th e m e a n o f th e e s tim a te d sp e c tru m is E[PxA f ) } = J 2 YA m ) e - i2nfm m= oc 0
(12.1.37)
1/2 if*
r xx( a ) W B( f - a ) d a
1/2
w h e re W B( f ) is th e sp e c tra l ch a ra c te ristic o f th e B a rtle tt w indow . T h e re la tio n (12.1.37) illu stra te s th a t th e m e a n o f th e e s tim a te d sp e c tru m is th e co n v o lu tio n o f th e tr u e p o w e r d en sity sp e c tru m T ( / ) w ith th e F o u rie r tra n sfo rm W B( f ) o f th e B a rtle tt w in d o w . C o n se q u e n tly , th e m e a n o f th e e s tim a te d s p e c tru m is a sm o o th e d v e rsio n o f th e tr u e s p e c tru m a n d su ffers fro m th e sa m e sp e c tra l le a k a g e p ro b le m s w h ich a re d u e to th e fin ite n u m b e r o f d a ta p o in ts. W e o b s e rv e th a t th e e stim a te d sp e c tru m is asy m p to tically u n b ia s e d , th a t is,
N -1
lim E N-t-cc
2 2
rx x ( m ) e ' j2nfm
Yxx{m)e-J2* fm = TxA f )
H o w e v e r, in g e n e ra l, th e v a ria n c e o f th e e s tim a te Pxx( f ) d o e s n o t d ec a y to zero as N - oo. F o r ex a m p le , w h e n th e d a ta se q u e n c e is a G a u s sia n ra n d o m process,
906
Chap. 12
the variance is easily shown to be (see Problem 12.4)

(12.1.38) w hich , in th e lim it as N > oo, b ec o m e s
N - + oo
lim v a r [ P ( / ) ] = ^ l x ( f )
(12.1.39)
H e n c e w e c o n clu d e th a t th e p e r i o d o g r a m is n o t a consi st ent est imate o f the true p o w e r density spe ct r um (i.e., it d o e s n o t co n v e rg e to th e tru e p o w e r d en sity sp e c tru m ). In su m m ary , th e e s tim a te d a u to c o rre la tio n rxx(m) is a c o n s iste n t e stim a te o f th e tr u e a u to c o rre la tio n fu n ctio n yxx(m). H o w e v e r, its F o u rie r tra n sfo rm Pxx( f ) , th e p e rio d o g ra m , is n o t a c o n siste n t e s tim a te o f th e tru e p o w e r d e n sity sp e ctru m . W e o b se rv e d th a t Pxx( f ) is an a s y m p to tically u n b ia s e d e stim a te o f Vxx{ f ) , b u t fo r a fin ite -d u ra tio n se q u en ce, th e m e a n v alu e o f Pxx{ f ) c o n ta in s a bias, w hich from (12.1.37) is e v id e n t as a d isto rtio n o f th e tru e p o w e r d en sity sp e c tru m . T h u s th e e s tim a te d sp e c tru m su ffers fro m th e sm o o th in g effe c ts a n d th e le a k a g e e m b o d ie d in th e B a rtle tt w ind o w . T h e sm o o th in g an d le a k a g e u ltim ately lim it o u r a b ility to reso lv e closely sp a ced sp e ctra. T h e p ro b le m s o f leak ag e an d fre q u e n c y re so lu tio n th a t w e h ave ju st d e sc rib ed as w ell as th e p ro b le m th a t th e p e rio d o g ra m is n o t a c o n s iste n t estim a te o f th e p o w e r sp e c tru m , p ro v id e th e m o tiv a tio n fo r th e p o w e r sp e c tru m e s tim a tio n m e th o d s d e sc rib e d in S ectio n s 12.2, 12.3, a n d 12.4. T h e m e th o d s d escrib ed in S ectio n 12.2 a re classical n o n p a ra m e tric m e th o d s, w hich m a k e n o assu m p tio n s a b o u t th e d a ta se q u en ce. T h e em p h asis o f th e classical m e th o d s is on o b ta in in g a c o n s iste n t e stim a te o f th e p o w e r sp e c tru m th ro u g h so m e av e ra g in g o r sm o o th in g o p e ra tio n s p e rfo rm e d d irectly o n th e p e rio d o g ra m o r o n th e a u to c o rre la tio n . A s w e will see, th e effe ct o f th e se o p e ra tio n s is to re d u c e th e fre q u e n c y re so lu tio n fu rth e r, w hile th e v a ria n c e o f th e e s tim a te is d e c re a se d . T h e sp e c tru m e stim a tio n m e th o d s d e sc rib e d in S ectio n 12.3 a re b a se d on so m e m o d e l o f h o w th e d a ta w ere g e n e ra te d . In g e n e ra l, th e m o d e l-b a se d m e th o d s th a t h av e b een d e v e lo p e d o v e r th e p a s t tw o d e c a d e s p ro v id e significantly h ig h e r re so lu tio n th a n d o th e classical m e th o d s. A d d itio n a l m e th o d s a re d e sc rib e d in S ectio n s 12.4 a n d 12.5. O n e o f th ese m e th o d s, d u e to C a p o n (1969), is b a s e d o n m in im iz in g th e v a ria n c e in th e sp e c tra l e stim a te . T h e m e th o d s d e sc rib e d in S ectio n 12.5 a re b a s e d on an e ig e n v a lu e /e ig e n v e c to r d e c o m p o sitio n o f th e d a ta c o rre la tio n m atrix .
12.1.3 The Use of the DFT in Power Spectrum Estimation

A s giv en b y (12.1.14) a n d (12.1.34), th e e s tim a te d en e rg y d e n s ity sp e c tru m , , ( / ) a n d th e p e rio d o g ra m Pxx( f ), resp ec tiv ely , can b e c o m p u te d b y u se o f th e D F T , w hich in tu rn is efficiently c o m p u te d by a F F T a lg o rith m . I f w e h a v e N d a ta p o in ts,
Sec. 12.1
907
w e c o m p u te as a m in im u m th e A ppoint D F T . F o r ex am p le, th e c o m p u ta tio n yields sa m p le s o f th e p e rio d o g ra m * = 0 ,1 ,..., JV -1 (12.1.40)
a t th e fre q u e n c ie s /* = k / N . In p ra c tic e , h o w ev er, such a sp a rs e sa m p lin g o f th e sp e c tru m d o e s n o t p ro v id e a v ery g o o d re p re s e n ta tio n o r a g o o d p ic tu re o f th e c o n tin u o u s sp e c tru m estim ate * > ( /). T h is is easily re m e d ie d b y e v a lu a tin g Pxx( f ) a t a d d itio n a l freq u e n cies. E q u iv a le n tly , w e can effectiv ely in c re a se th e le n g th o f th e se q u e n c e by m e a n s o f z e ro p a d d in g a n d th e n e v a lu a te Pxx{ f ) at a m o re d en se se t o f freq u e n cies. T h u s if w e in crease th e d a ta se q u e n c e le n g th to L p o in ts by m e a n s o f z e ro p a d d in g an d e v a lu a te th e L -p o in t D F T , we h av e (12.1.41) W e em p h a siz e th a t z e ro p a d d in g an d e v a lu a tin g th e D F T at L > d o e s n o t im p ro v e th e fre q u e n c y re so lu tio n in th e sp e c tra l estim a te . p ro v id e s us w ith a m e th o d fo r in te rp o la tin g th e valu es o f th e m e a s u re d a t m o re fre q u e n c ie s. T h e fre q u e n c y re so lu tio n in th e s p e c tra l e stim a te d e te rm in e d by th e len g th N o f th e d a ta reco rd . Exam ple 12.1.2 A sequence of N = 16 samples is obtained by sampling an analog signal consisting of two frequency components. The resulting discrete-time sequence is x( n) = sin2jr(0.135)w + cos2;r(0.135 + Af ) n n = 0 , 1 , . . . , 15 N p o in ts It sim ply sp e ctru m PXI{.f ) is
where A/ is the frequency separation. Evaluate the power spectrum P ( f ) = (1/7V )|X (/)|2 at the frequencies f k = k/ L, k = 0, 1 ,_ _ L 1, for L = 8, 16, 32, and 128 for values of A/ = 0.06 and A/ = 0.01. Solution By zero padding, we increase the data sequence to obtain the power spec trum estim ate PXI(k/L). The results for A/ = 0.06 are plotted in Fig. 12.4. Note that zero padding does not change the resolution, but it does have the effect of interpo lating the spectrum Pxl ( f ) . In this case the frequency separation A/ is sufficiently large so that the two frequency components are resolvable. The spectral estimates for A/ = 0.01 are shown in Fig. 12.5. In this case the two spectral com ponents are not resolvable. Again, the effect of zero padding is to provide more interpolation, thus giving us a better picture of the estim ated spectrum. It does not improve the frequency resolution. W h e n o n ly a few p o in ts o f th e p e rio d o g ra m a re n e e d e d , th e G o e rtz e l alg o rith m d e s c rib e d in C h a p te r 6 m ay p ro v id e a m o re efficient c o m p u ta tio n . Since th e G o e rtz e l a lg o rith m h a s b e e n in te r p re te d as a lin e a r filterin g a p p ro a c h to c o m p u t in g th e D F T , it is c le a r th a t th e p e rio d o g ra m e s tim a te can b e o b ta in e d by passing
908
Chap. 12
_S___ B ___ S_
L I ..
T 32
^mTTTTnK/f
Figure 12.4
I^WrffTllnaa
128 Spectra of two sinusoids with frequency separation A f 0.06.
th e signal th ro u g h a b a n k o f p a ra lle l tu n e d filters a n d sq u a rin g th e ir o u tp u ts (see P ro b le m 12.5).
12.2 NONPARAMETRIC METHODS FOR POWER SPECTRUM ESTIMATION

T h e p o w e r sp e c tru m e stim a tio n m e th o d s d e sc rib e d in th is se c tio n a re th e classical m e th o d s d e v e lo p e d by B a rtle tt (1948), B la c k m a n a n d T u k e y (1958), an d W elch (1967). T h e se m e th o d s m a k e n o a ssu m p tio n a b o u t h o w th e d a ta w ere g e n e ra te d a n d h e n c e a re called nonparamet ri c.
Sec. 12.2
Nonparametric Methods for Power Spectrum Estimation
909
F igure 1X5
Spectra of two sinusoids with frequency separation A/ = 0.01.
S ince th e e stim a te s a re b ased e n tire ly on a fin ite re c o rd o f d a ta , th e fre q u e n c y re so lu tio n o f th e s e m e th o d s is, at b est, e q u a l to th e sp e c tra l w id th o f th e re c ta n g u la r w in d o w o f le n g th N , w hich is a p p ro x im a te ly 1( N a t th e - 3 - d B p o in ts. W e shall b e m o re p recise in sp ecify in g th e fre q u e n c y re so lu tio n o f th e specific m e th o d s. A ll th e e s tim a tio n te c h n iq u e s d e sc rib e d in th is section d e c re a s e th e fre q u e n c y re so lu tio n in o r d e r to re d u c e th e v a rian ce in th e sp e c tra l e stim a te . F irst, w e d e sc rib e th e e stim ates an d d e riv e th e m e a n a n d v aria n c e o f e ach . A c o m p a riso n o f th e th re e m e th o d s is given in S ectio n 12.2.4. A lth o u g h th e sp e ctral e stim a te s a re e x p ressed a s a fu n ctio n o f th e c o n tin u o u s fre q u e n c y v a ria b le / , in p ra c tic e , th e e stim a te s a re c o m p u te d a t d isc rete fre q u e n c ie s via th e F F T alg o rith m . T h e F F T -b a se d c o m p u ta tio n a l re q u ire m e n ts a re c o n s id e re d in S ectio n 12.2.5.
910
Chap. 12
12.2.1 The Bartlett Method: Averaging Periodograms

B a r tle tts m e th o d fo r re d u c in g th e v arian ce in th e p e rio d o g ra m involves th re e step s. F irst, th e N -p o in t se q u e n c e is su b d iv id e d in to K n o n o v e rla p p in g se g m en ts, w h e re ea c h se g m e n t h as le n g th M, T h is re su lts in th e K d a ta se g m e n ts x,(n) = x( n + i M ) i = 0, I, . . . . K 1 n = 0 , 1 .........M - l F o r ea c h se g m en t, w e c o m p u te th e p e rio d o g ra m p i 'v " M-l = _L E * . '< n ) e - J2nfn\ M (12.2.1)
i= 0 ,l,...,J f - l
(12.2.2)
F in ally , w e av erag e th e p e rio d o g ra m s fo r th e K se g m e n ts to o b ta in th e B a rtle tt p o w e r sp e c tru m e stim a te [B a rtle tt (1948)] 1 * -1 p xB A f^ = T Z y ( / )

K ,=v>
( 12-2 -3 >
T h e statistical p ro p e rtie s of this e s tim a te a re easily o b ta in e d . F irst, th e m ean v alu e is (1 2 2 4 ) = [* (/)] F ro m (12.1.35) an d (12.1.37) w e h ave th e e x p e c te d v a lu e fo r th e single p e rio d o g ra m as Af-1
E f1 -
y ^ e~s2nfm
(12.2.5)
M J_ ]/2 -i rw h ere
s in ;r ( / - a) J . _ . ( 12. 2. 6)
1 / s in jr/M V B(f) = t ; : M V s ni f f / / is th e fre q u e n c y c h aracteristics o f th e B a rtle tt w indow m, w B(n) =

] ------ ,
0,
11 ~ " o th e rw ise
|m|
< M 1
(12.2.7)
F ro m (12.2.5) w e o b se rv e th a t th e tru e sp e c tru m is no w c o n v o lv e d w ith th e fre q u e n c y c h a ra c te ristic W a i f ) o f th e B a rtle tt w indow . T h e e ffe c t o f re d u c in g th e le n g th o f th e d a ta fro m N p o in ts to M = N / K re su lts in a w in d o w w hose sp e c tra l
Sec. 12.2
911
w id th h a s b e e n in c re a se d b y a fa c to r o f K . C o n s e q u e n tly , th e fre q u e n c y re so lu tio n h as b e e n re d u c e d b y a fa c to r K. In r e tu rn fo r th is re d u c tio n in re so lu tio n , w e h a v e r e d u c e d th e v arian ce. T h e v a ria n c e o f th e B a rtle tt e s tim a te is var[ / ( / ) ] =
K2 I=U
( 12.2 .8 )
= ^ v a r [/> (/)] If w e m a k e u se o f (12.1.38) in (12.2.8), w e o b ta in v ar [ / > * ( / ) ] = j T l x { f )
1+
/ sin I n f M ' V \M s in 2 ;r/ J
(12.2.9)
T h e re fo re , th e v a rian ce o f th e B a rtle tt p o w e r sp e c tru m e s tim a te h a s b e e n re d u c e d by th e fa c to r K.
12.2.2 The Welch Method: Averaging Modified Periodograms

W elch (1967) m ad e tw o b asic m o d ificatio n s to th e B a rtle tt m e th o d . F irst, h e a l lo w ed th e d a ta se g m en ts to o v erlap . T h u s th e d a ta se g m e n ts c a n b e r e p re s e n te d as
Xj ( n) = x( n + i D )
n = 0 ,1 , . . . , M 1 i = 0 . 1 .........L - l
(12.2.10)
w h e re i D is th e sta rtin g p o in t fo r th e /th se q u e n c e . O b s e rv e th a t if D = M , th e se g m en ts d o n o t o v e rla p an d th e n u m b e r L o f d a ta se g m e n ts is id e n tic a l to th e n u m b e r K in th e B a rtle tt m e th o d . H o w e v e r, if D = M /2 , th e r e is 50% o v e rla p b e tw e e n su ccessive d a ta se g m e n ts a n d L = 2 K se g m e n ts a re o b ta in e d . A lte rn a tiv e ly , w e can fo rm K d a ta se g m en ts ea c h o f le n g th 2 M . T h e se c o n d m o d ifica tio n m a d e by W elch to th e B a rtle tt m e th o d is to w indow th e d a ta se g m e n ts p rio r to c o m p u tin g th e p e rio d o g ra m . T h e re su lt is a m o d ifie d p e rio d o g ra m 1 * * (/) = ~MU n=0 ( 12.2. 11)
w h e re U is a n o rm a liz a tio n fa c to r f o r th e p o w e r in th e w in d o w fu n c tio n a n d is se le c te d as 1 M-l u /(n ) ( 12. 2 . 12) T h e W elch p o w e r sp e c tru m e stim a te is th e a v e ra g e o f th e se m o d ified p e rio d o g ra m s, th a t is, 1 L~x ^ ( /) = (12.2.13)
912
Chap. 12
The mean value of the Welch estimate is
^ . =o = * ( /)] B u t th e e x p e c te d v alu e o f th e m od ified p e rio d o g ra m is

i M -l M-\
(12.2.14)
[ ^ ( / ) ] = T777 E E '( " ) w ( * ) k ( " K ( ) ] < A T /l* ' " ) =0 m=0

M - l Af-1
(12.2.15)
Since
Vjcx(n) = j
r ia
r xx(a)el27!anda
(12.2.16)
J - 1/2 s u b s titu tio n fo r y ^ n ) fro m (12.2.16) in to (12.2.15) yields
* ? < /> ]=
1/2
M - 1 M -l n = 0 m =0
m
1 /2
j- 1/2 l r w
da (12.2.17)
f = / r (a )W (/-a )rfa '-1/2 w h e re , by d efin itio n , E u>(n)e n=0

-jZnfn
(12.2.18)
3 # 17
T h e n o rm a liz a tio n fa c to r U e n s u re s th a t /
1/2
W (/W / = 1
(12.2.19)
J 1/2
T h e v arian ce o f th e W elch e s tim a te is v a r l O / ) ] = Tr E E ^ ~ ( f ) P g ( f ) ] ~ {E[P?Af)])2 (12. 2. 20)
In th e case o f n o o v e rla p b e tw e e n successive d a ta se g m e n ts ( L = K ) , W elch has sh o w n th a t v a r |P * ( / ) J = i v a r [ P ' ( / ) l
(12.2.21)
* zr(/)
Sec. 12.2
913
In th e case o f 5 0% o v e rla p b e tw e e n successive d a ta se g m en ts (L = 2 AT), th e v a ria n c e o f th e W elch p o w e r sp e c tru m e s tim a te w ith th e B a rtle tt (tria n g u la r) w indow , also d e riv e d in th e p a p e r by W elch, is v a r [ / > ( / ) ] * ^ r * T( / ) (12.2.22)
A lth o u g h w e c o n s id e re d only th e tria n g u la r w in d o w in th e c o m p u ta tio n o f th e v a rian ce, o th e r w in d o w fu n c tio n s m ay be used. In g e n e ra l, th ey will yield a d iffe re n t v arian ce. In ad d itio n , o n e m ay v ary th e d a ta se g m e n t o v e rla p p in g b y e ith e r m o re o r less th a n th e 50% co n sid e re d in th is se c tio n in a n a tte m p t to im p ro v e th e re le v a n t c h a ra c te ristic s o f th e estim ate.
12.2.3 The Blackman and Tukey Method: Smoothing the Periodogram

B lac k m a n a n d T u k e y (1958) p ro p o se d an d an aly zed th e m e th o d in w hich th e sa m p le a u to c o rre la tio n se q u e n c e is w in d o w ed first a n d th e n F o u rie r tra n sfo rm e d to yield th e e s tim a te o f th e p o w e r sp e ctru m . T h e ra tio n a le fo r w in d o w in g th e estim a te d a u to c o rre la tio n se q u e n c e rxx(m) is th a t, fo r larg e lags, th e e s tim a te s are less re lia b le b ecau se a sm a ller n u m b e r (N m) o f d a ta p o in ts e n te r in to th e estim ate. F o r v alu es o f m a p p ro a c h in g N , th e v a rian ce o f th e s e e stim a te s is very h igh, an d h en c e th e se e s tim a te s sh o u ld be given a sm a lle r w eig h t in th e fo rm a tio n o f th e e stim a te d p o w e r sp e c tru m . T h u s th e B la c k m a n -T u k e y e s tim a te is
A f-1
E r , A m ) w { m ) e - J2*fm m= (M 1)
(12.2.23)
w h e re th e w in d o w fu n c tio n w( n) h as length 2 M 1 a n d is z e ro fo r \m\ > M . W ith th is d e fin itio n fo r w( n) , th e lim its o n th e sum in (12.2.23) can b e e x te n d e d to ( oo, oo). H e n c e th e fre q u e n c y -d o m a in e q u iv a le n t ex p re ssio n fo r (12.2.23) is th e c o n v o lu tio n in teg ral
yl/2
* * //(/)= /
J- 1/2
P*A<x)W{f-a)da
(12.2.24)
w h e re Pxx( f ) is th e p e rio d o g ra m . It is clear from (12.2.24) th a t th e effe ct o f w in d o w in g th e a u to c o r re la tio n is to sm o o th th e p e rio d o g ra m e s tim a te , th u s d e creasin g th e v a rian ce in th e e s tim a te a t th e e x p en se o f re d u c in g th e re so lu tio n . T h e w in d o w s e q u e n c e w( n) sh o u ld be sy m m e tric (e v e n ) a b o u t m = 0 to e n s u re th a t th e e stim a te o f th e p o w e r sp e c tru m is re a l. F u rth e rm o r e , it is d e sirab le to se lect th e w in d o w sp e c tru m to b e n o n n eg ativ e, th a t is, W (f) > 0 | / | < 1 /2 (12.2.25)
T h is c o n d itio n e n s u re s th a t PXJ ( f ) > 0 fo r | / | < 1/2, w hich is a d e s ira b le p ro p e rty fo r an y p o w e r sp e c tru m e s tim a te . W e sh o u ld in d ic a te , h o w ev er, th a t so m e o f th e w in d o w fu n c tio n s w e h a v e in tro d u c e d d o n o t satisfy th is c o n d itio n . F o r exam p le,
914
Chap. 12
in sp ite o f th e ir low sid e lo b e levels, th e H a m m in g a n d H a n n ( o r H a n n in g ) w indow s d o n o t satisfy th e p ro p e rty in (12.2.25) a n d , c o n s e q u e n tly , m ay re su lt in n eg ativ e sp e c tru m e stim a te s in so m e p a rts o f th e fre q u e n c y ran g e. T h e e x p e c te d v alu e o f th e B la c k m a n -T u k e y p o w e r s p e c tru m e s tim a te is
, 1/2
[* * //( /) ]= / E [ P xA a ) ] W ( f - a ) d a J - 1/2 w h e re fro m (12.1.37) w e h av e ,1/2 [/> ( )]= / r xA 0 ) W B( a - 0 ) d e J -1/2 a n d WB{ f ) is th e F o u rie r tra n sfo rm o f th e B a rtle tt w in d o w . (12.2.27) in to (12.2.26) yields th e d o u b le c o n v o lu tio n in teg ral r 1/2 ,1/2 E[P?rT ( / ) ] = / / r ( e ) W JJ( a - 0 ) W ( / - a ) r f a r f e / 1/2 J \[2 (12.2.27) S u b stitu tio n of (12.2.26)
(12.2.28)
E q u iv a le n tly , by w o rk in g in th e tim e d o m a in , th e e x p e c te d v alu e o f th e B la c k m a n -T u k e y p o w e r sp e c tru m e stim a te is M- 1 = m = w h ere th e B a rtle tt w in d ow is
22
(iW ll Af-1
E [ r * A m ) ] w ( m ) e - J 27lfm (12.2.29) y , A m ) w B( m ) w ( m ) e - j 2nfm
22
m= (M-l)
wB(m)
0,
o th e rw ise
C learly , w e sh o u ld select th e w in d o w le n g th fo r w( n) such th a t M << N , th a t is, w( n) sh o u ld b e n a rro w e r th a n w B(m) to p ro v id e a d d itio n a l sm o o th in g o f the p e rio d o g ra m . U n d e r th is c o n d itio n , (12.2.28) b e c o m e s , 1/2 E lP f/if)}* J - 1/2 since VxA 0 ) W { f - d ) d d (12.2.31)
,
/
1/2
,
WB(a - d ) W { f - a ) d a = /
1/2
W B( a ) W ( f - 9 - a)doc (12.2.32)
J - 1/2
J - 1/2
^ w (f-e)
T h e v arian ce o f th e B la c k m a n -T u k e y p o w e r sp e c tru m e s tim a te is v a r [ / > / / ( / ) ] = { [ P / / ( / ) 2]} - { [ / > / / ( / ) ] } 2 (12.2.33)
Sec. 12.2
915
w h e re th e m e a n can b e a p p ro x im a te d as in (12.2.31). (11.2.33) is

rl/2
T h e se c o n d m o m e n t in
fl/2
E{[PxB xT ( f ) ] 2) = /
E [ P xx( a ) P xx( . e ) ] W( f
-6)dade
(12.2.34)
J - 1/2 J-\f2
O n th e a s su m p tio n th a t th e ra n d o m p ro c e ss is G a u ssia n (see P ro b le m 12.5), w e find th a t E [ P xA a ) P xA&)] =
r (a )r (0) 1 +
sin7r(0 + a ) N _Ns i n7 r ( 8 + a)_
sin7r(0 a ) N N s i n n ( 6 a) _ (12.2.35)
S u b stitu tio n o f (12.2.35) in to (12.2.34) yields -1/2 E { [ P ? / ( f ) ] 2} =

- 1/2
-.2
r xA 0 ) W ( f - e ) d e
,1 [2
,
/
1/2
+ /
r xA<*)rxA e ) W ( f - a ) W ( f - d ) si nn( & a ) N N sin7r(8 or)
J - 1/2 J -l/2
s in jr(# -l-oON N s in jr(0 + a)_
d a d$
(12.2.36)
T h e first te rm in (12.2.36) is sim ply th e sq u a re o f th e m e a n o f P f J ( / ) , w hich is to b e s u b tra c te d o u t a c co rd in g to (12.2.33). T h is le a v e s th e se c o n d te rm in (12.2.36), w h ich c o n s titu te s th e v arian ce. F o r th e case in w hich N Af, th e fu n c tio n s s i n n ( 9 + a ) N / N s i n ^ ( 0 + a ) a n d s m n ( 8 a ) N / N sin jr ( 0 a ) a re relativ ely n a rro w c o m p a re d to W ( f ) in th e vicinity o f 0 = a a n d 9 = a , resp ec tiv ely . T h e re fo re ,
y _ i /2
{ [L N s m n ( 8
+ a;)J
r [ _ ^ sin 7 r(0 ~ a j
N
-a)N
de (12.2.37)
% r xA - a ) W ( f + c c ) + r xA<x)W( f - a)
W ith th is a p p ro x im a tio n , th e v a ria n c e o f PX J i f ) b ec o m e s

1
ar[ / " ( / ) ] -
" J - 1/2 1 /2
-
Cxn
r ( a ) W ( / - a ) [ r ( - a ) W ( / + a) + r (o )W (/-o ()]rfa
r lfl
-
J i L 1/2 rZ ^
w^ f ~ a)da
(12.2.38)
w h e re in th e last s te p w e m a d e th e a p p ro x im a tio n
1/2
r z, ( a ) r ( - c r ) V ( / - a ) W ( f + a ) d a 0
1/2
(12.2.39)
916
Chap. 12
W e shall m a k e o n e a d d itio n a l a p p ro x im a tio n in (12.2.38). W h e n W ( f ) is n a rro w c o m p a re d to th e tru e p o w e r sp e c tru m r ( / ) , (12.2.38) is f u rth e r a p p ro x im a te d as v a r[/> //(/)] *
r2 x x (f)
H A f)
-
f '* W 2 (0)d0j
1/2
M-l
w 2( m )
(12.2.40)
(Af 1)
12.2.4 Performance Characteristics of Nonparametric Power Spectrum Estimators

In th is sectio n w e co m p a re th e q u ality o f th e B a rtle tt, W elch, a n d B la c k m a n an d T u k e y p o w e r sp e c tru m estim ates. A s a m e a s u re o f q u ality , w e u se th e ra tio o f its v arian ce to th e s q u a re o f th e m e a n o f th e p o w e r sp e c tru m e s tim a te th a t is,
Q
a
iE[P?Af)}}2 t >ar [ P ( f ) ]
(12.2.41)
w h e re A = B , W , o r B T fo r th e th re e p o w e r s p e c tru m estim ates. T h e recip ro c al o f th is q u a n tity , called th e variability , can also be u se d as a m e a s u re o f p e rfo rm a n c e . F o r re fe re n c e , th e p e rio d o g ra m has a m e a n a n d v arian ce
E[P*Af)] = f r(0)WW - 8)de

J- l / 2
(12.2.42)
v a r [/> ( /) ]= r ,V /)[ l + ( | ^ ) 2' \N sin2nf/ w h ere 1 / sin n f N \ 2 W B( f ) = ~ ~ Lr N V sin n f J i-l/2 E [ P * i ( f ) ] -+
(12.2.43)
(12.2.44)
r (/) /
J - 1/2
w B( d ) d e =
u>,(0) r ( / ) = r ( / )
(12.2.45)
var[/>(/)] - r l x( f )
H e n c e , as in d ic a te d p rev io u sly , th e p e rio d o g ra m is an asy m p to tically u n b ia se d e stim a te o f th e p o w e r sp e c tru m , b u t it is n o t c o n siste n t becau se its v arian ce d o es n o t a p p ro a c h z e ro as N in c re a se s to w a rd infinity. A sy m p to tically , th e p e rio d o g ra m is c h a ra c te riz e d by th e q u a lity fa c to r
Qp
r\
= z.1 -,
2 (f )} _ r xx^J
rlA f)
= 1
(12.2.46)
T h e fact th a t Qp is fixed a n d in d e p e n d e n t o f th e d a ta le n g th N is a n o th e r in d ic a tio n o f th e p o o r q u ality o f th is e stim a te .
Sec. 12.2
917
Bartlett power spectrum estimate.

p o w e r s p e c tru m e s tim a te a re E[P?Af)]=
T h e m e a n a n d v a ria n c e o f th e B a rtle tt
J-
1/2
r xxm w B( f - 6) dd / s in 2 jr /A f V
(12.2.47)
v a r[/ (/)] = - r 1 ( f ) an d
^ \ M sin 2 n f )
(12.2.48)
1 / sin n f M \ 2 W B( f ) = . J . M \ s in ^ / / A s N y oo a n d M y oo, w hile K = N / M rem ain s fixed, w e find th a t

r 1/2
(12.2.49)
[ * * < /) ]
r ( /)
/
7 - -1/2 1 /2
w ,(/)r f/
= r xx( f ) w B(0) =
r (/)
(12.2.50)
v a r [P /x( / ) ]
rL (/)
W e o b se rv e th a t th e B a rtle tt p o w e r sp e c tru m e s tim a te is asy m p to tically u n b i ased a n d if K is allo w ed to in crease w ith an in crease in N , th e e s tim a te is also c o n sisten t. H e n c e , asy m p to tically , th is e stim a te is c h a ra c te riz e d b y th e q u a lity fa c to r N Qb = K = (12.2.51) M T h e fre q u e n c y re so lu tio n o f th e B a rtle tt e s tim a te , m e a s u re d by tak in g th e 3-dB w id th o f th e m ain lo b e o f th e re c ta n g u la r w indow , is 0.9 A/ = J M H e n c e , M = 0 .9 /A / a n d th e q u ality fa c to r b eco m es N Qb = 0 .9 /A / = l.IN A f (12.2.52)
(12.2.53)
Welch power spectrum estimate.

p o w e r sp e c tru m e stim a te are
[ * ( / ) ] =
T h e m e a n a n d v a ria n c e o f th e W elch
f r^mw(f-B)de
2 2 w ( n ) e - j2*f *
1/2
(12.2.54)
w h ere (12.2.55)
and
fo r no o v e rla p
v a r [ j ( /)] =
9 , r 2 (f),
8L xx
fo r 50% o v e rla p a n d tria n g u la r w indow
(12.2.56)
918
Chap. 12
As N
oo an d M - * oo, th e m e a n co n v e rg e s to [* (/> ]-> r ( / ) (12.2.57)
an d th e v arian ce c o n v erg es to z e ro , so th a t th e e s tim a te is c o n s iste n t. U n d e r th e tw o co n d itio n s given by (12.2.56) th e q u ality fa c to r is N L = , Qw = S L = 16N 9 9M ' fo r n o o v e rla p fo r 5 0 % o v e rla p a n d tria n g u la r w indow (12.2.58)
O n th e o th e r h a n d , th e sp e ctral w idth o f th e tr ia n g u la r w indow a t th e 3-dB p o in ts is A/ = M C o n se q u e n tly , th e q u ality fa c to r e x p re ss e d in te rm s o f A f a n d N is 0.78N A /, = 1 39NAf fo r n o o v e rla p fo r ^ 0/ o v e rla P a n d tr ia n g u la r w indow (12.2.60) (12.2.59)
Blackman-Tukey power spectrum estimate.

th is e s tim a te a re a p p ro x im a te d as
- 1/2
T h e m e a n a n d v a ria n c e o f
E[Pf / ( f ) ] * f
J --M l 1/2
r xxm w ( f - d ) d 8
v a r[P //(/)] *
r;x(f)
L
- r
1 N
>7 1 w (m)
(12.2.61)
m=-(W -1)
w h e re u>(m) is th e w in d ow se q u e n c e u se d to ta p e r th e e s tim a te d a u to c o rre la tio n se q u e n c e . F o r th e re c ta n g u la r an d B a rtle tt (tria n g u la r) w in d o w s w e h av e 1 N r * '1 2/ s (2 M/N, < ")= { 2U/3N. re c ta n g u la r w in d o w tr ia n g u la r w in d o w ^ (12'2 '62)
I t is c le a r fro m (12.2.61) th a t th e m e a n v a lu e o f th e e s tim a te is asym ptotically u n b ia se d . Its q u a lity fa c to r fo r th e tr ia n g u la r w in d o w is Q bt = 1 - 5 ^ M (12.2.63)
S ince th e w in d o w le n g th is 2 M 1, th e fre q u e n c y re so lu tio n m e a s u re d a t th e 3-dB p o in ts is A/ = ^ = 2M M (12.2.64)
and hence
Q b t = x r i N & f = 2 .3 4 W A / U.64 (12.2.65)
Sec. 12.2
N onparametric Methods for Power Spectrum Estimation TABLE 12.1 QUALITY OF POWER SPECTRUM ESTIMATES
Estimate Quality Factor
919
Bartlett
Welch (50% overlap) Blackman-Tukey
l.UNAf
1.39NA/
2.34A fA /
T h e se re su lts a re su m m a riz e d in T a b le 12.1. It is a p p a re n t fro m th e resu lts w e h av e o b ta in e d th a t th e W elch a n d B la c k m a n -T u k e y p o w e r sp e c tru m e stim a te s a re so m e w h a t b e tte r th a n th e B a rtle tt e stim a te . H o w e v e r, th e d iffe re n c e s in p e r fo rm a n c e a re relativ ely sm all. T h e m ain p o in t is th a t th e q u a lity fa c to r in crease s w ith an in c re a se in th e le n g th N o f th e d a ta . T h is c h a ra c te ristic b e h a v io r is n o t s h a re d by th e p e rio d o g ra m estim a te . F u rth e rm o re , th e q u a lity fa c to r d e p e n d s on th e p ro d u c t o f th e d a ta len g th N a n d th e fre q u e n c y re so lu tio n A f . F o r a d e sire d level o f q u ality , A f can b e d e c re a s e d (fre q u e n c y re so lu tio n in c re a se d ) by in creasin g th e len g th N o f th e d a ta , a n d vice versa.
12.2.5 Computational Requirements of Nonparametric Power Spectrum Estimates

T h e o th e r im p o rta n t a s p e c t o f th e n o n p a ra m e tric p o w e r sp e c tru m e stim a te s is th e ir c o m p u ta tio n a l re q u ire m e n ts . F o r th is c o m p a riso n w e assu m e th e e stim a te s a re b a s e d o n a fixed a m o u n t o f d a ta N a n d a specified re so lu tio n A f . T h e radix2 F F T a lg o rith m is a ssu m ed in all th e c o m p u ta tio n s. W e shall c o u n t only th e n u m b e r o f c o m p lex m u ltip lic a tio n s re q u ire d to c o m p u te th e p o w e r sp e c tru m e sti m a te .
Bartlett power spectrum estimate

F F T len g th = M = 0 .9 / A / N u m b e r o f F F T s = = 1.11W A f M J N u m b e r o f c o m p u ta tio n s = log2 M j ~ log2
Welch power spectrum estimate (50% overlap)

F F T len g th = M 1 .2 8 /A / N u m b e r o f F F T s = = 1. 56N A f M
Number of computations = I log2 (y 'o feM *)
920
Chap. 12
In a d d itio n to th e 2 N /Af F F T s, th e r e a re a d d itio n a l m u ltip lic a tio n s re q u ire d fo r w in d o w in g th e d a ta . E a c h d a ta re c o rd re q u ire s M m u ltip lic atio n s. T h e re fo re , th e to ta l n u m b e r o f c o m p u ta tio n s is 1 28 5 12 T o ta l c o m p u ta tio n s = 2 N + N log2 = N lo g , Af - Af
Blackman-Tukey power spectrum estimate. In th e B ia c k m a n -T u k e y m e th o d , th e a u to c o rre la tio n rxx{m) can b e c o m p u te d efficiently via th e F F T a l g o rith m . H o w e v e r, if th e n u m b e r o f d a ta p o in ts is larg e, it m ay n o t b e p o ssib le to c o m p u te o n e A p p oint D F T . F o r ex a m p le , we m ay h ave N = 105 d a ta p o in ts b ut o nly th e cap ac ity to p e rfo rm 1024-point D F T s. S ince th e a u to c o rre la tio n se q u e n c e is w in d o w ed to 2 M 1 p o in ts w h e re M <: N , it is p o ssible to c o m p u te th e d esired 2 M 1 p o in ts o f rxx(m) by se g m e n tin g th e d a ta in to K = N f l M re c o rd s a n d th e n c o m p u tin g 2 A /-p o in t D F T s a n d o n e 2 M -p o in t I D F T via th e F F T alg o rith m . R a d e r (1970) h as d e sc rib e d a m e th o d fo r p e rfo rm in g th is c o m p u ta tio n (se e P ro b le m 12.7). If w e b ase th e c o m p u ta tio n a l co m p lex ity o f th e B la c k m a n -T u k e y m e th o d on th is a p p ro a c h , we o b ta in th e follow ing c o m p u ta tio n a l re q u ire m e n ts .
F F T len g th = 2 M = 1 .2 8 /A / / N \ N N um ber o f FFT s = 2 K + 1 = 2 { ) + 1 ^ \2M J M N u m b e r o f c o m p u ta tio n s = log2 2 M ) = N log2
W e can neg lect th e a d d itio n a l M m u ltip lic a tio n s re q u ire d to w in d o w th e a u to c o r re la tio n se q u e n c e rxx(m), since it is a relativ ely sm all n u m b e r. F in ally , th e re is th e a d d itio n a l c o m p u ta tio n re q u ire d to p e rfo rm th e F o u rie r tra n s fo rm o f th e w in d o w ed a u to c o rre la tio n se q u en ce. T h e F F T a lg o rith m can b e used fo r th is co m p u ta tio n w ith so m e zero p a d d in g fo r p u rp o se s o f in te rp o la tin g th e s p e c tra l e stim a te . A s a re su lt o f th e se a d d itio n a l c o m p u ta tio n s, th e n u m b e r o f c o m p u ta tio n s is in crease d by a sm all a m o u n t. F ro m th e se resu lts w e c o n c lu d e th a t th e W elch m e th o d r e q u ire s a little m o re c o m p u ta tio n a l p o w e r th a n d o th e o th e r tw o m e th o d s. T h e B a rtle tt m e th o d a p p a r en tly re q u ire s th e sm a llest n u m b e r o f c o m p u ta tio n s. H o w e v e r, th e d ifferen ces in th e c o m p u ta tio n a l re q u ire m e n ts of th e th re e m e th o d s a re re la tiv e ly sm all.
12.3 PARAMETRIC METHODS FOR POWER SPECTRUM ESTIMATION

T h e n o n p a ra m e tric p o w e r sp e c tru m e stim a tio n m e th o d s d e s c rib e d in th e p re c e d in g se ctio n a re re lativ ely sim p le, w ell u n d e rsto o d , an d easy to c o m p u te using th e F F T alg o rith m . H o w ev er, th e s e m e th o d s re q u ire th e av ailab ility o f lo n g d a ta re c o rd s in o r d e r to o b ta in th e n ecessa ry fre q u e n c y re so lu tio n re q u ire d in m a n y ap p licatio n s. F u rth e rm o re , th e s e m e th o d s su ffer fro m sp e c tra l le a k a g e effects, d u e to w indow -
Sec. 12.3
Parametric Methods for Power Spectrum Estimation
921
ing, th a t a re in h e re n t in fin ite-len g th d a ta reco rd s. O fte n , th e sp e c tra l leak ag e m ask s w eak sig n als th a t a re p re s e n t in th e d a ta . F ro m o n e p o in t o f view , th e basic lim ita tio n o f th e n o n p a ra m e tric m e th o d s is th e in h e re n t a ssu m p tio n th a t th e a u to c o rre la tio n e stim a te rxx(m) is z e ro fo r m > N, as im p lied by (12.1.33). T h is a ssu m p tio n se v ere ly lim its th e fre q u e n c y re so lu tio n an d th e q u a lity o f th e p o w e r sp e c tru m e s tim a te th a t is ach iev ed . F ro m a n o th e r v iew p o in t, th e in h e re n t a ssu m p tio n in th e p e rio d o g ra m e s tim a te is th a t th e d a ta are p e rio d ic w ith p e rio d N . N e ith e r o n e o f th e s e a s su m p tio n s is realistic. In th is se ctio n w e d esc rib e p o w e r sp e c tru m e s tim a tio n m e th o d s th a t d o n o t re q u ire such a s su m p tio n s. In fact, th e s e m e th o d s ext rapolate th e v alu es o f th e a u to c o rre la tio n fo r lags m > N. E x tra p o la tio n is p o ssib le if w e h av e so m e a pri ori in fo rm a tio n o n h o w th e d a ta w ere g e n e ra te d . In such a case a m o d e l fo r th e signal g e n e ra tio n can b e c o n s tru c te d w ith a n u m b e r o f p a ra m e te rs th a t can b e e s tim a te d fro m th e o b se rv e d d a ta . F ro m th e m o d e l an d th e e s tim a te d p a ra m e te rs , w e can c o m p u te th e p o w e r d e n sity sp e c tru m im plied by th e m odel. In effect, th e m o d e lin g a p p ro a c h e lim in a te s th e n e e d fo r w indow fu n ctio n s an d th e a s su m p tio n th a t th e a u to c o rre la tio n se q u e n c e is z e ro fo r \m\ > N . A s a c o n s e q u e n c e , paramet ri c (m o d e l-b a s e d ) p o w e r sp e c tru m e s tim a tio n m e th o d s avoid th e p ro b le m o f le a k a g e a n d p ro v id e b e tte r freq u e n c y re so lu tio n th a n d o th e F F T b ased , n o n p a ra m e tric m e th o d s d e sc rib e d in th e p re c e d in g se ctio n . T h is is e s p e cially tru e in a p p lic a tio n s w h e re sh o rt d a ta re c o rd s a re a v a ila b le d u e t o tim e -v a ria n t o r tra n sie n t p h e n o m e n a . T h e p a ra m e tric m e th o d s c o n s id e re d in th is section a re b ased o n m o d elin g th e d a ta se q u e n c e * ( ) as th e o u tp u t o f a lin ear system c h a ra c te riz e d by a ra tio n a l sy stem fu n ctio n o f th e fo rm
(12.3.1)
T h e c o rre sp o n d in g d iffe re n c e e q u a tio n is p (12.3.2) w h e re w( n) is th e in p u t se q u e n c e to th e system a n d th e o b se rv e d d a ta , x( n ) , re p re s e n ts th e o u tp u t se q u e n c e . In p o w e r sp e c tru m e stim a tio n , th e in p u t se q u e n c e is n o t o b se rv a b le . H o w e v e r, if th e o b se rv e d d a ta a re c h a ra c te riz e d as a s ta tio n a ry ra n d o m p ro cess, th e n th e in p u t se q u e n c e is also assu m ed to b e a s ta tio n a ry ra n d o m process. In such a case th e p o w e r d e n s ity sp e c tru m o f th e d a ta is r ( / ) = \ n ( f ) \ 2r ww( f ) w h e re r ,**,(/) is th e p o w e r d en sity sp e c tru m o f th e in p u t se q u e n c e a n d H ( f ) is th e fre q u e n c y re sp o n s e o f th e m o d el.
9 22
Chap. 12
S ince o u r o b jectiv e is to e stim a te th e p o w e r d en sity s p e c tru m r ( / ) , it is c o n v e n ie n t to assu m e th a t th e in p u t se q u e n c e w( n ) is a z e ro -m e a n w h ite noise se q u e n c e w ith a u to c o rre la tio n
Yww( m) = ^ S ( m )
w h e re a 2 is th e v a ria n c e (i.e., a 2 = [|u>()|2]). T h e n th e p o w e r d en sity sp e c tru m o f th e o b se rv e d d a ta is sim ply T (/) = ol\H {f)\2 = (12.3.3)
In S e ctio n 11.1 w e d e sc rib e d th e r e p re s e n ta tio n o f a sta tio n a ry ra n d o m p ro c e ss as given b y (12.3.3). In th e m o d e l-b a se d a p p ro a c h , th e sp e c tru m e s tim a tio n p ro c e d u re consists o f tw o step s. G iv e n th e d a ta se q u e n c e x ( n) , 0 < n < JV - 1, w e e stim a te th e p a r a m e te r s {a*} a n d (6*) o f th e m o d el. T h e n fro m th e s e e s tim a te s , w e c o m p u te th e p o w e r sp e c tru m e s tim a te acco rd in g to (12.3.3). R e c a ll th a t th e ra n d o m p ro c e ss x ( n ) g e n e ra te d b y th e p o le - z e r o m o d e l in (12.3.1) o r (12.3.2) is called an aut or e gr es s i ve- movi ng average (A R M A ) pr oc es s o f o r d e r ( p , q) a n d it is u sually d e n o te d as A R M A ( p , q). I f q = 0 a n d fro = th e re su ltin g system m o d e l h as a system fu n c tio n H ( z ) = 1 /A (z) a n d its o u tp u t x( n) is called a n autoregressive (A R ) process o f o r d e r p. T h is is d e n o te d as A R(/>). T h e th ird p o ssib le m o d e l is o b ta in e d by se ttin g A ( z ) = 1, so th a t H( z ) B{z). Its o u tp u t x ( n ) is called a m o v i n g average (M A ) p r o c e s s o f o r d e r q a n d d e n o te d as M A (^ ). O f th e se th re e lin e a r m o d els th e A R m o d e l is by fa r th e m o st w idely used. T h e re a so n s a re tw o fo ld . F irst, th e A R m o d e l is su ita b le fo r re p re s e n tin g sp e ctra w ith n a rro w p e a k s (re so n a n c e s). S eco n d , th e A R m o d e l re su lts in v ery sim ple lin e a r e q u a tio n s fo r th e A R p a ra m e te rs . O n th e o th e r h a n d , th e M A m o d e l, as a g e n e ra l ru le , re q u ire s m an y m o re co effic ie n ts to r e p re s e n t a n a rro w sp ectru m . C o n s e q u e n tly , it is ra re ly u se d by itse lf as a m o d e l f o r s p e c tru m e s tim a tio n . By co m b in in g p o le s a n d zero s, th e A R M A m o d e l p ro v id e s a m o re efficien t re p re s e n ta tio n , fro m th e v iew p o in t o f th e n u m b e r o f m o d e l p a ra m e te rs , o f th e sp e c tru m o f a ra n d o m p ro cess. T h e d e c o m p o sitio n th e o re m d u e to W o ld (1938) a s se rts th a t an y A R M A o r M A p ro c e ss can b e re p re s e n te d u n iq u ely by an A R m o d e l o f p o ssib ly infinite o rd e r, a n d a n y A R M A o r A R p ro c e ss c a n b e r e p re s e n te d by a M A m o d e l o f possibly in fin ite o rd e r. In v iew o f this th e o re m , th e issue o f m o d e l se le c tio n re d u c e s to selectin g th e m o d e l th a t re q u ire s th e sm a llest n u m b e r o f p a r a m e te r s th a t a re also easy to c o m p u te . U su ally , th e choice in p ra c tic e is th e A R m o d e l. T h e A R M A m o d e l is u sed to a le s se r ex ten t. B e fo re d escrib in g m e th o d s fo r e s tim a tin g th e p a r a m e te r s in a n A R ( p ) , MA(<?), a n d A R M A ( p , q) m o d els, it is useful to e sta b lish th e b asic re la tio n sh ip s b etw een th e m o d e l p a r a m e te r s a n d th e a u to c o rre la tio n s e q u e n c e yxx( m). In a d d itio n , we re la te th e A R m o d e l p a ra m e te rs to th e co efficien ts in a lin e a r p r e d ic to r fo r the p ro c e ss x( n) .
Sec. 12.3
9 23
12.3.1 Relationships Between the Autocorrelation and the Model Parameters

In S ectio n 11.1.2 w e esta b lish e d th e b asic re la tio n sh ip s b e tw e e n th e a u to c o rre la tio n { ^ ( m ) } a n d th e m o d e l p a r a m e te r s {a*} a n d {bk}. F o r th e A R M A ( p , q) process, th e re la tio n sh ip g iv en by (11.1.18) is
~22Q kY xx > jn Yxx (m) =

p
k),
m > q (12.3.4) + <*122 *=o 0 <m <q m < 0
~ 2 2 ot Yxxi m - ^ *=l y*x (m) ,
T h e re la tio n sh ip s in (12.3.4) p ro v id e a fo rm u la fo r d e te rm in in g th e m o d el p a ra m e te rs {ak} by re stric tin g o u r a tte n tio n to th e case m > q. T h u s th e se t o f lin e a r e q u a tio n s ~Yxx(q) Yxxiq + 1) Yxxiq ~ 1) Yxxiq) Yxxiq ~ p + 1 ) ' Yxx (q + P + 2)
-Yxx(q + / ? - ! )
Yxx(q + p - 2)
Yxxiq) Yxxiq + 1) yxx(q + 2) (12.3.5) - Yx xi q + P ) -
can b e u se d to so lv e fo r th e m o d e l p a ra m e te rs {a k} by u sin g e s tim a te s o f th e a u to c o rre la tio n se q u e n c e in place o f Yxx ) fo r m > q. T h is p ro b le m is d iscussed in S e ctio n 12.3.8. A n o th e r in te r p re ta tio n o f th e re la tio n sh ip in (12.3.5) is th a t th e v alu es o f th e a u to c o rre la tio n Yxx ) iox m > q a re u n iq u e ly d e te rm in e d fro m th e p o le p a ra m e te rs {a k} a n d th e v alu es o f y xx(m) fo r 0 < m < p. C o n s e q u e n tly , th e lin e a r sy stem m o d e l a u to m a tic a lly e x te n d s th e v alu es o f th e a u to c o rre la tio n se q u e n c e Yxx(m ) fo r m > p. If th e p o le p a ra m e te rs {a*} a re o b ta in e d fro m (12.3.5), th e re su lt d o e s n o t h e lp u s in d e te rm in in g th e M A p a r a m e te r s {i*}, b e c a u se th e e q u a tio n
q-m p
l 2 2 h ( f t bk+m = Yxx(m) + ' Y 2 a kYxx{m - k)
0 < m < q
d e p e n d s o n th e im p u lse re sp o n s e h( n). A lth o u g h th e im p u lse re sp o n s e can be e x p re ss e d in te rm s o f th e p a ra m e te rs {*} by lo n g division o f B(z) w ith th e kn o w n A( z ) , th is a p p ro a c h re su lts in a se t o f n o n lin e a r e q u a tio n s fo r th e M A p a r a m eters.
924
Chap. 12
I f w e a d o p t a n A R ( p ) m o d e l fo r th e o b se rv e d d a ta , th e re la tio n s h ip b e tw e e n th e A R p a ra m e te rs a n d th e a u to c o rre la tio n se q u e n c e is o b ta in e d by se ttin g q = 0 in (12.3.4). T h u s w e o b ta in

p
~ Y 2 akYxx(m - k), *=i

Y xx(m ) =
p
m> 0 (12.3.6) m = 0 m < 0
- J 2 akYxx(m - k ) + ct^, Jt=i Yxx ( ~ m )-
In th is case, th e A R p a r a m e te r s {a*) a re o b ta in e d fro m th e so lu tio n o f th e Y u le W a lk e r o r n o rm a l e q u a tio n s Yxx (0)

Yxx(l) Y x x ( ~ 1) Y x x ( 0) Y x x i - P + 1) Yxx(~ P
~a\ " a2
Y x x i 1) " Yxx
+ 2)
(2)
(12.3.7)
- Y x x i p - 1)
Y x x ip -
2)
Krjr (0)
. Y xx( P ) .
a n d th e v aria n c e a 2 can b e o b ta in e d fro m th e e q u a tio n

p
= Yxxi ) + Y 2 a k Y x x i ~ k )
(12.3.8)
T h e e q u a tio n s in (12.3.7) a n d (12.3.8) a re u su ally co m b in e d in to a single m atrix e q u a tio n o f th e fo rm

Yxx ( 0 ) Y xx(l) Yx x ( 1) Y x x ( - p + 1)
- 1 a\ -ap_ =
Yxx ( 0 )
~a l ~ 0 (12.3.9)
L y(p)
Y x x ip -
1)
y (0 )
_0 _
S ince th e c o rre la tio n m atrix in (12.3.7), o r in (12.3.9), is T o e p litz , it can b e effi cien tly in v e rte d b y u se o f th e L e v in s o n - D u rb in a lg o rith m . T h u s all th e sy stem p a r a m e te r s in th e A R (/?) m o d e l a r e easily d e te rm in e d fro m k n o w le d g e o f th e a u to c o rre la tio n se q u e n c e yx x (m) fo r 0 < m < p. F u rth e r m o re , (12.3.6) can b e u se d to e x te n d th e a u to c o rre la tio n s e q u e n c e f o r m > p, o n ce th e {at) a re d e te rm in e d . F in ally , f o r c o m p le te n e s s, w e in d ic a te th a t in a M A (^ ) m o d e l f o r th e o b se rv e d d a ta , th e a u to c o rre la tio n se q u e n c e yjjr(m ) is re la te d to th e M A p a r a m e te r s [bk] by th e e q u a tio n 9 < ? l Y ^ b kbk+m, 0 <m <q
Yxx(m )
1=0 0, y*x( - m ) ,
m > q m < 0
(12.3.10)
which was established in Section 11.1.
Sec. 12.3
925
W ith th is b a c k g ro u n d esta b lish e d , w e no w d esc rib e th e p o w e r sp e c tru m e s tim a tio n m e th o d s fo r th e A R (/?), A R M A (/?,< ?), a n d MA(<?) m o d els.
12.3.2 The Yule-Walker Method for the AR Model Parameters

In th e Y u le - W a lk e r m e th o d we sim ply e stim a te th e a u to c o rre la tio n fro m th e d a ta an d u se th e e stim a te s in (12.3.7) to solve fo r th e A R m o d e l p a ra m e te rs. In th is m e th o d it is d e s ira b le to use th e b iased fo rm o f th e a u to c o rre la tio n e sti m a te ,
^ N - m 1
rxx(m) = N
y ' n=0
jc*(n)jc( + m)
m> 0
(12.3.11)
to e n s u re th a t th e a u to c o rre la tio n m atrix is p o sitiv e se m id e fin ite. T h e re su lt is a sta b le A R m o d el. A lth o u g h sta b ility is n o t a critical issue in p o w e r sp e c tru m e s tim a tio n , it is c o n je c tu re d th a t a sta b le A R m o d e l b e st re p re s e n ts th e d ata. T h e L e v in s o n -D u rb in alg o rith m d e sc rib e d in C h a p te r 11 w ith rxx( m ) su b sti tu te d fo r yxx(m) yield s th e A R p a ra m e te rs. T h e c o rre sp o n d in g p o w e r sp e c tru m e s tim a te is 2 P r w ( f ) = --------------- ^ ---------------(12 3 12) " II + L i a P( * ) e - J2*/ l l2 w h e re a p(k) a re e s tim a te s o f th e A R p a ra m e te rs o b ta in e d fro m th e L e v in s o n D u rb in re c u rsio n s a n d < 7l p = / = rxx (0) f [ [ l - M *=i ) | 2] (12.3.13)
is th e e s tim a te d m in im u m m e a n -sq u a re v alue fo r th e p th - o rd e r p re d ic to r. A n ex a m p le illu stra tin g th e fre q u e n c y re so lu tio n ca p a b ilitie s o f th is e s tim a to r is given in S ectio n 12.3.9. In e stim a tin g th e p o w e r sp e c tru m o f sin u so id al signals via A R m o d els, L acoss (1971) sh o w ed th a t sp e c tra l p e a k s in an A R sp e c tru m e s tim a te a re p ro p o rtio n a l to th e s q u a re o f th e p o w e r o f th e sin u so id al signal. O n th e o th e r h a n d , th e a re a u n d e r th e p e a k in th e p o w e r d en sity sp e c tru m is lin early p ro p o rtio n a l to th e p o w e r o f th e sin u so id . T h is c h a ra c te ristic b e h a v io r h o ld s fo r all A R m o d e l-b a se d estim atio n m e th o d s.
12.3.3 The Burg Method for the AR Model Parameters

T h e m e th o d d ev ised b y B u rg (1968) fo r e stim a tin g th e A R p a r a m e te r s can be v iew ed as an o rd e r-re c u rsiv e le a st-sq u a re s lattice m e th o d , b a s e d on th e m in im iz a tio n o f th e fo rw a rd a n d b ac k w a rd e rro rs in lin e a r p re d ic to rs , w ith th e c o n s tra in t th a t th e A R p a r a m e te r s satisfy th e L e v in s o n -D u rb in re c u rsio n .
926
Pow er Spectrum Estimation
Chap. 12
T o d e riv e th e e s tim a to r, su p p o s e th a t w e a re given th e d a ta * (n ), n = 0, 1......... N 1, a n d let us c o n sid e r th e fo rw a rd a n d b a c k w a rd lin e a r p re d ic tio n e sti m a te s o f o r d e r m , given as m
x(n) = y
*=1
m
am(fc)x (rt fc)

(12.3.14) (fc)x(n + k m)
x ( n - m) =
k= 1
a n d th e c o rre sp o n d in g fo rw a rd a n d b a c k w a rd e rro rs f m{n) a n d gm(n) defin ed as f m{n) = x(n) x(n) a n d gm(n) = x(n m) x(n m ) w h e re am(k), 0 < k < m 1, m = 1, 2 , . . . , p, are th e p re d ic tio n coefficients. T h e le a st-sq u a re s e r ro r is
A'-l
+ \ g M \ 2] n=/n
(12.3.15)
T h is e r r o r is to b e m in im ized by se lectin g th e p re d ic tio n coefficients, su b je ct to th e c o n s tra in t th a t th e y satisfy th e L e v in s o n -D u rb in re c u rsio n given by am(k) = a m- i( * ) -I-k) 1 < fc < m - 1 1 < m < p w h ere K m = a m(m) is th e m th reflec tio n c o effic ie n t in th e la ttic e filter re a liz a tio n o f th e p re d ic to r. W h e n (12.3.16) is s u b s titu te d in to th e ex p re ssio n s fo r f m(n) a n d gm(n), th e re su lt is th e p a ir o f o rd e r-re c u rsiv e e q u a tio n s fo r th e fo rw a rd a n d b ac k w a rd p re d ic tio n e rro rs given by (11.2.4). N ow , if w e su b s titu te fro m (11.2.4) in to (12.3.16) a n d p e rfo rm th e m in im iza tio n o f m w ith re sp e c t to th e co m p lex -v alu ed re fle c tio n co effic ie n t K m, w e o b ta in th e resu lt N-1 (12.3.16)
fm-i(n)g*m- i ( n - D
Km =
n=m
------------------------------------
m = 1 ,2 , . . . , p
(12.3.17)
T h e te rm in th e n u m e r a to r o f (12.3.17) is a n e s tim a te o f th e c ro ssc o rre la tio n b e tw e e n th e fo rw a rd a n d b ac k w a rd p re d ic tio n e rro rs. W ith th e n o rm a liz a tio n facto rs in th e d e n o m in a to r o f (12.3.17), it is a p p a r e n t th a t |AT,J < 1, so th a t th e all-pole m o d e l o b ta in e d fro m th e d a ta is sta b le . T h e r e a d e r sh o u ld n o te th e sim ilarity o f (12.3.17) to its sta tistical c o u n te rp a rts given b y (11.2.29). W e n o te th a t th e d e n o m in a to r in (12.3.17) is sim ply th e le a s t-s q u a re s e s tim a te o f th e fo rw a rd a n d b a c k w a rd e rro rs, a n d E ^ _ v re sp e c tiv e ly . H e n c e (12.3.17)
Sec. 12.3
927
can b e e x p re ss e d as
- 1) Km = m = 1, 2 , . . . , p (12.3.18)
w h e re + Eb m- \ is an e stim a te o f th e to ta l sq u a re d e r ro r E m. W e leave it as a n ex ercise fo r th e r e a d e r to verify th a t th e d e n o m in a to r te r m in (12.3.18) can be c o m p u te d in an o rd e r-re c u rsiv e fash io n a c co rd in g to th e re la tio n E m = (1 - \ Km |2) m_i - |/m -] (m - 1)[2 - |gm_ ,(m - 2 ) |2 (12.3.19)
w h e re E, = E ^ + E b is th e to ta l le a st-sq u a re s e rro r. T h is re su lt is d u e to A n d e rse n (1978). T o su m m arize, th e B u rg a lg o rith m c o m p u te s th e reflec tio n coefficients in th e e q u iv a le n t lattice s tru c tu re as specified by (12.3.18) a n d (12.3.19). a n d th e L e v in s o n -D u rb in a lg o rith m is u se d to o b ta in th e A R m o d e l p a ra m e te rs. F ro m th e e s tim a te s o f th e A R p a ra m e te rs, w e fo rm th e p o w e r sp e c tru m e s tim a te
(12.3.20)
1 + j 2 a p( k ) e - W
T h e m a jo r a d v a n ta g e s o f th e B u rg m e th o d fo r e stim a tin g th e p a ra m e te rs o f th e A R m o d e l a re (1) it resu lts in high fre q u e n c y re so lu tio n , (2) it yields a sta b le A R m o d el, an d (3) it is c o m p u ta tio n a lly efficient. T h e B u rg m e th o d is k n o w n to h av e se v e ra l d isa d v a n ta g e s, h o w ev er. F irst, it ex h ib its sp e c tra l lin e sp littin g a t high sig n a l-to -n o ise ratio s, [see th e p a p e r by F o u g e re e t al. (1976)]. B y line sp littin g , w e m e a n th a t th e sp e c tru m o f x( n) m ay h a v e a sin g le sh a rp p e a k , b u t th e B u rg m e th o d m ay re s u lt in tw o o r m o re closely sp a ced p e a k s. F o r h ig h -o rd e r m o d els, th e m e th o d also in tro d u c e s sp u rio u s p eak s. F u rth e rm o r e , fo r sin u so id al signals in n o ise , th e B u rg m e th o d ex h ib its a sensitivity to th e in itial p h a se o f a sinusoid, esp ecially in s h o rt d a ta reco rd s. T h is sensitivity is m a n ife st as a fre q u e n c y shift fro m th e tru e freq u e n cy , re su ltin g in a p h a se d e p e n d e n t fre q u e n c y b ias. F o r m o re d etails o n so m e o f th e s e lim ita tio n s th e r e a d e r is r e fe rr e d to th e p a p e rs o f C h e n a n d S te g e n (1974), U lry c h a n d C lay to n (1976), F o u g e re e t al. (1976), K ay a n d M a rp le (1979), S w ingler (1979a, 1980), H e rrin g (1980), a n d T h o rv a ld s e n (1981). S ev eral m o d ifica tio n s h av e b e e n p ro p o se d to o v e rc o m e so m e o f th e m o re im p o rta n t lim ita tio n s o f th e B u rg m e th o d : n a m ely , th e lin e sp littin g , sp u rio u s p e a k s , a n d fre q u e n c y bias. B asically, th e m o d ifica tio n s inv o lv e th e in tro d u c tio n o f a w eig h tin g (w in d o w ) se q u e n c e o n th e sq u a re d fo rw a rd a n d b ack w a rd e r ro rs . T h a t is, th e le a st-sq u a re s o p tim iz a tio n is p e rfo rm e d o n th e w e ig h te d s q u a re d
928
Chap. 12
e rro rs
N-1
b =
wm( n) [ \ f m(n)\2 + |gm(n)[2]
(12.3.21)
w hich , w h en m in im iz ed , re su lts in th e reflec tio n coefficient e s tim a te s

A '- l
']Tl Wm--i(n)fm- i ( n ) g Z _ l (n - 1) Km = ~1 ~
1 n=m
L -------------------------------------------------+ |g m- i ( n - 1 )|2]
(12.3.22)
In p a rtic u la r, w e m e n tio n th e use o f a H a m m in g w in d o w u se d by S w ingler (1979b), a q u a d ra tic o r p a ra b o lic w in d o w u sed by K av eh a n d L ip p e rt (1983), th e e n erg y w eig h tin g m e th o d u se d by N ik ias a n d S co tt (1982), a n d th e d a ta -a d a p tiv e en e rg y w eig h tin g u se d by H e lm e an d N ik ias (1985). T h e se w in d o w in g a n d en erg y w eig h tin g m e th o d s h av e p ro v e d effe ctiv e in re d u cin g th e o c c u rre n c e o f line sp littin g a n d sp u rio u s p eak s, a n d a re also effective in re d u c in g fre q u e n c y bias. T h e B u rg m e th o d fo r p o w e r sp e c tru m e s tim a tio n is u su a lly asso ciated w ith m a x i m u m ent r opy spe ct r um est i mati on , a c rite rio n u sed by B u rg (1967, 1975) as a basis f o r A R m o d elin g in p a ra m e tric s p e c tru m e s tim a tio n . T h e p ro b le m co n sid e re d by B u rg w as h o w b e st to e x tra p o la te fro m th e given v alu es o f th e a u to c o rre la tio n se q u e n c e yxx{m), 0 < m p, such th a t th e e n tire a u to c o r re latio n se q u e n c e is p o sitiv e se m id efin ite. S ince an in fin ite n u m b e r o f e x tra p o la tio n s a re p o ssib le, B u rg p o s tu la te d th a t th e e x tra p o la tio n s b e m a d e o n th e b asis o f m ax i m izing u n c e rta in ty (e n tro p y ) o r ra n d o m n e ss , in th e se n se th a t th e sp e c tru m Txx{ f ) o f th e p ro cess is th e fla tte st o f all sp e c tra w hich h a v e th e g iv en a u to c o rre la tio n valu es yxx(m), 0 < m < p. In p a rtic u la r th e e n tro p y p e r sa m p le is p ro p o rtio n a l to th e in te g ra l [see B u rg (1975)]
r 1/2
/ In J-l/2 rlfl / J - 1/2
r xx( f ) d f
(12.3.23)
B u rg fo u n d th a t th e m ax im u m o f th is in te g ra l su b je ct to th e ( p + 1) c o n stra in ts = Yxx(m) 0 < m < p (12.3.24)
is th e A R ( p ) p ro c e ss fo r w hich th e given a u to c o rre la tio n se q u e n c e yxx( m), 0 < m < p is re la te d to th e A R p a ra m e te rs by th e e q u a tio n (12.3.6). T h is so lu tio n p ro v id e s an a d d itio n a l ju stificatio n fo r th e u se o f th e A R m o d e l in p o w e r sp e c tru m e stim a tio n . In view o f B u rg s b asic w o rk in m ax im u m e n tro p y s p e c tra l es tim a tio n , th e B u rg p o w e r sp e c tru m e s tim a tio n p ro c e d u re is o fte n called th e m a x i m u m ent ropy m e t h o d (M E M ). W e sh o u ld em p h asize, h o w ev er, th a t th e m a x im u m e n tro p y sp e c tru m is id en tical to th e A R -m o d e l sp e c tru m o n ly w h e n th e e x a c t a u to c o rre la tio n
Sec. 12.3
929
yx x (m) is k n o w n . W h e n o n ly an e s tim a te o f yxx(m) is a v a ilab le fo r 0 < m < p, th e A R -m o d e l e s tim a te s o f Y u le -W a lk e r a n d B u rg a re n o t m ax im u m e n tro p y sp e c tra l e s tim a te s. T h e g e n e ra l fo rm u la tio n fo r th e m ax im u m e n tro p y sp e c tru m b a se d on e stim a te s o f th e a u to c o rre la tio n se q u e n c e resu lts in a se t o f n o n lin e a r e q u a tio n s. S o lu tio n s fo r th e m ax im u m e n tro p y sp e c tru m w ith m e a s u re m e n t e rro rs in th e c o rre la tio n se q u e n c e h a v e b e e n o b ta in e d by N ew m an (1981) a n d S c h o tt a n d M cC lellan (1984).
12.3.4 Unconstrained Least-Squares Method for the AR Model Parameters

A s d e sc rib e d in th e p re c e d in g se ctio n , th e B u rg m e th o d f o r d e te rm in in g th e p a r a m e te rs o f th e A R m o d e l is basically a le a st-sq u a re s la ttic e a lg o rith m w ith th e a d d e d c o n s tra in t th a t th e p re d ic to r coefficients satisfy th e L e v in so n re c u rsio n . A s a re su lt o f th is c o n s tra in t, a n in crease in th e o r d e r of th e A R m o d e l re q u ire s only a sin g le p a r a m e te r o p tim iz a tio n a t e a c h stage. In c o n tr a s t to th is a p p ro a c h , w e m ay use an u n c o n s tra in e d le a st-sq u a re s alg o rith m to d e te rm in e th e A R p a r a m e ters. T o e la b o ra te , w e fo rm th e fo rw a rd an d b ac k w a rd lin e a r p re d ic tio n e stim a te s an d th e ir c o rre sp o n d in g fo rw a rd a n d b ac k w a rd e r ro r s as in d ic a te d in (12.3.14) a n d (12.3.15). T h e n w e m in im ize th e sum o f sq u a re s o f b o th e rro rs , th a t is,
N -1
p -
n= p
Y l ^ f p ^ 2+
p
2"
x ( n ) + ' Y \ ap ^ x ^n ~ k ) *=i
+ x ( n - p) + 2 2 Op(k ) x( n + k - p) k=1
(12.3.25) w h ich is th e sa m e p e rfo rm a n c e in d ex as in th e B u rg m e th o d . H o w e v e r, w e d o n o t im p o se th e L e v in s o n - D u rb in re c u rsio n in (12.3.25) fo r th e A R p a ra m e te rs. T h e u n c o n s tra in e d m in im iz atio n o f p w ith re sp e c t to th e p re d ic tio n co effic ie n ts yields th e se t o f lin e a r e q u a tio n s M * ) ^ ( U ) = - r ( l , 0) k=l l = l,2,...,p (12.3.26)
w h e re , by d efin itio n , th e a u to c o rre la tio n rxx( l , k ) is JV-1 r ( / , k) = n p T h e re su ltin g re sid u a l le a st-sq u a re s e r r o r is p ep L s = r ( 0 ,0 ) + t f , ( * ) r ( 0 , k ) t-i (12.3.28) ~ k ) x *(n 0 + x ( n - P + l)x*(n p + )] (12.3.27)
930
Chap. 12
H e n c e th e u n c o n s tra in e d le a st-sq u a re s p o w e r s p e c tru m e s tim a te is ls P ^ i f ) = ----------------E --------------- T 1 + i 2 a p( k ) e - J^ k=l k (12.3.29)
T h e c o rre la tio n m a trix in (12.3.27), w ith e le m e n ts rxx(l, k) , is n o t T o e p litz , so th a t th e L e v in s o n -D u rb in alg o rith m c a n n o t b e a p p lie d . H o w e v e r, th e c o rre la tio n m atrix h a s su fficien t s tru c tu re to m a k e it p o ssib le to d ev ise c o m p u ta tio n a lly effi cien t a lg o rith m s w ith c o m p u ta tio n a l co m p lex ity p ro p o rtio n a l to p 2. M a rp le (1980) d ev ised such an a lg o rith m , w hich h a s a lattice s tru c tu re a n d em p lo y s L e v in so n D u rb in -ty p e o r d e r re c u rsio n s a n d a d d itio n a l tim e recu rsio n s. T h is fo rm o f th e u n c o n s tra in e d le a st-sq u a re s m e th o d d e s c rib e d h a s also b e e n called th e u n w i n d o w e d data le a st-sq u a re s m e th o d . I t h a s b e e n p ro p o s e d fo r sp e c tru m e s tim a tio n in se v era l p a p e rs, in clu d in g th e p a p e rs b y B u rg (1967), N u tta ll (1976), an d U lry ch an d C la y to n (1976). Its p e rfo rm a n c e c h a ra c te ristic s h av e b e e n fo u n d to b e s u p e rio r to th e B u rg m e th o d , in th e se n se th a t th e u n c o n s tra in e d leastsq u a re s m e th o d d o e s n o t e x h ib it th e sa m e sen sitiv ity to such p ro b le m s as line sp lit ting, fre q u e n c y b ias, a n d sp u rio u s p eak s. In view o f th e c o m p u ta tio n a l efficiency o f M a rp le s a lg o rith m , w hich is c o m p a ra b le to th e efficiency o f th e L e v in s o n -D u rb in alg o rith m , th e u n c o n s tra in e d le a st-sq u a re s m e th o d is very a ttr a c tiv e . W ith this m e th o d th e re is n o g u a ra n te e th a t th e e stim a te d A R p a r a m e te r s yield a sta b le A R m o d el. H o w e v e r, in sp e c tru m e s tim a tio n , th is is n o t c o n s id e re d to be a p ro b le m .
12.3.5 Sequential Estimation Methods for the AR Model Parameters

T h e th r e e p o w e r sp e c tru m e s tim a tio n m e th o d s d e sc rib e d in th e p re c e d in g se ctio n s fo r th e A R m o d e l can b e classified as b lo ck p ro cessin g m e th o d s. T h e se m e th o d s o b ta in e stim a te s o f th e A R p a ra m e te rs fro m a b lo ck o f d a ta , say jr(n), n = 0, I , . , N I. T h e A R p a ra m e te rs , b a s e d o n th e b lo ck o f N d a ta p o in ts, a re th e n u sed to o b ta in th e p o w e r sp e c tru m estim a te . In situ a tio n s w h e re d a ta a re av ailab le on a c o n tin u o u s b asis, w e can still se g m en t th e d a ta in to b lo ck s o f N p o in ts an d p e rfo rm sp e c tru m e stim a tio n on a b lo ck -b y -b lo ck basis. T h is is o fte n d o n e in p ra c tic e , fo r b o th re a l-tim e an d n o n -re a l-tim e ap p licatio n s. H o w e v e r, in such a p p lic a tio n s, th e r e is an a lte rn a tiv e a p p ro a c h b ased o n se q u e n tia l (in tim e) e s tim a tio n o f th e A R m o d e l p a ra m e te rs as each n ew d a ta p o in t b ec o m e s av ailab le. B y in tro d u c in g a w e ig h tin g fu n ctio n in to p a s t d a ta sam p les, it is p o ssib le to d e e m p h a siz e th e effect o f o ld e r d a ta sam ples as n ew d a ta a re receiv ed . S e q u e n tia l la ttic e m e th o d s b ased o n recu rsiv e le a s t sq u a re s can b e em p lo y ed to o p tim a lly e s tim a te th e p re d ic tio n a n d re fle c tio n c o effic ie n ts in th e lattice re alizatio n o f th e fo rw a rd a n d b ac k w a rd lin e a r p re d ic to rs . T h e re c u rsiv e e q u a
Sec. 12.3
931
tio n s fo r th e p re d ic tio n coefficients re la te directly to th e A R m o d el p a ra m e te rs. In a d d itio n to th e o rd e r-re c u rsiv e n a tu re o f th e s e e q u a tio n s , as im p lied by th e la ttic e stru c tu re , w e can also o b ta in tim e -re c u rsiv e e q u a tio n s fo r th e reflectio n co effic ie n ts in th e lattice an d fo r th e fo rw a rd a n d b a c k w a rd p re d ic tio n coeffi cien ts. S e q u e n tia l rec u rsiv e le a st-sq u a re s a lg o rith m s a re e q u iv a le n t to th e u n c o n stra in e d le a st-sq u a re s, b lo ck p ro cessin g m e th o d d e sc rib e d in th e p re c e d in g se c tio n . H e n c e th e p o w e r sp e c tru m e stim a te s o b ta in e d by th e s e q u e n tia l recursive le a s t-s q u a re s m e th o d re ta in th e d e s ira b le p r o p e rtie s o f th e b lo ck p ro cessin g alg o rith m d e s c rib e d in S ectio n 12.3.4. S ince th e A R p a ra m e te rs a re b ein g co n tin u o u sly e s tim a te d in a se q u e n tia l e stim a tio n a lg o rith m , p o w e r sp e c tru m e stim a te s can b e o b ta in e d as o fte n as d e s ire d , fro m o n ce p e r sa m p le to o n c e ev ery N sam ples. By p ro p e rly w e ig h tin g p a s t d a ta sam p les, th e s e q u e n tia l e s tim a tio n m e th o d s a re p a r ticu larly su ita b le fo r e stim a tin g a n d tra c k in g tim e -v a ria n t p o w e r s p e c tra resu ltin g fro m n o n s ta tio n a ry signal statistics. T h e c o m p u ta tio n a l co m p lex ity o f s e q u e n tia l e s tim a tio n m e th o d s is g en erally p ro p o rtio n a l to p, th e o r d e r o f th e A R p ro cess. A s a c o n se q u e n c e , se q u e n tia l e s tim a tio n a lg o rith m s a re c o m p u ta tio n a lly efficient an d , fro m this v iew p o in t, m ay o ffe r so m e a d v a n ta g e o v e r th e b lo ck p ro cessin g m e th o d s. T h e re a re n u m e ro u s re fe re n c e s on se q u e n tia l e stim a tio n m e th o d s. T he p a p e rs by G riffith s (1975), F rie d la n d e r (1982a, b ), a n d K a lo u p tsid is a n d T h e o d o rid is (1987) a re p a rtic u la rly re le v a n t to th e sp e c tru m e s tim a tio n p ro b le m .
12.3.6 Selection of AR Model Order

O n e o f th e m o st im p o rta n t asp ects of th e use o f th e A R m o d e l is th e se lectio n o f th e o r d e r p. A s a g e n e ra l ru le , if w e se lect a m o d e l w ith to o low an o r d e r, w e o b ta in a highly sm o o th e d sp e c tru m . O n th e o th e r h a n d , if p is se le c te d to o hig h , w e r u n th e risk o f in tro d u c in g sp u rio u s low -level p e a k s in th e sp e c tru m . W e m e n tio n e d p rev io u sly th a t o n e in d ic a tio n o f th e p e rfo rm a n c e o f th e A R m o d e l is th e m e a n -sq u a re v a lu e o f th e re sid u a l e rro r, w hich, in g en eral, is d iffe re n t fo r each o f th e e stim a to rs d e sc rib e d a b o v e . T h e c h a ra c te ristic of th is re sid u a l e r r o r is th a t it d e c re a s e s as th e o r d e r o f th e A R m o d e l is in crease d . W e can m o n ito r th e r a te o f d e crease an d d e cid e to te r m in a te th e p ro cess w h en th e ra te o f d e c re a s e b e c o m e s re la tiv e ly slow . I t is a p p a re n t, h o w ev er, th a t th is a p p ro a c h m ay b e im p re cise an d ill-defined, an d o th e r m e th o d s sh o u ld b e in v esti g ated . M u ch w o rk h as b e e n d o n e by v a rio u s re s e a rc h e rs o n th is p ro b le m a n d m any e x p e rim e n ta l re su lts h a v e b e e n given in th e lite ra tu r e [e.g., th e p a p e rs by G ersch a n d S h a rp e (1973), U lry ch an d B ish o p (1975), T o n g (1975, 1977), J o n e s (1976), N u tta ll (1976), B e rry m a n (1978), K a v e h a n d B ru z z o n e (1979), a n d K ash y ap (1980)]. T w o o f th e b e tte r k n o w n c rite ria fo r se lectin g th e m o d e l o rd e r h av e b e e n p r o p o s e d b y A k a ik e (1969, 1974). W ith th e first, c a lle d th e f i na l pr edi ct ion error
932
Chap. 12
( FPE) criterion, th e o rd e r is se le c te d to m in im iz e th e p e rfo rm a n c e in d ex
F P E ( ',= <(^373t )
A I C ( p ) = In a l p + 2 p / N
d 2 '3 3 0 *
(12.3.31)
w h e re a 2p is th e e s tim a te d v a rian ce o f th e lin e a r p re d ic tio n e r ro r . T h is p e rfo r m a n ce in d ex is b a s e d o n m in im izin g th e m e a n -s q u a re e r ro r fo r a o n e -ste p p re d ic to r. T h e seco n d c rite rio n p r o p o s e d by A k a ik e (1974), called th e A k a i k e i n f o r ma t ion criterion (A IC ), is b a se d o n se lectin g th e o r d e r th a t m in im iz es
N o te th a t th e te rm a 2p d e c re a s e s a n d th e r e fo r e In er2p also d e c re a s e s as th e o rd e r o f th e A R m o d e l is in crease d . H o w e v e r, 2 p / N in c re a se s w ith an in crease in p. H e n c e a m in im u m v alu e is o b ta in e d fo r so m e p. A n a lte rn a tiv e in fo rm a tio n c rite rio n , p ro p o s e d by R is sa n e n (1983), is b ased o n se lectin g th e o r d e r th a t m i n i mi z e s the description l ength (M D L ), w h e re M D L is d efin ed as M D L (/j) = N \ r \ a 2p + p i n N (12.3.32)
A fo u rth c rite rio n has b e e n p ro p o s e d by P a rz e n (1974). T h is is called th e criterion aut oregressive transf er (C A T ) fu n ctio n a n d is d e fin ed as C A T (p) = ( I w h ere (12.3.34) T h e o rd e r p is se le c te d to m in im ize C A T (/?). In ap p ly in g th is crite ria , th e m e a n sh o u ld b e re m o v e d fro m th e d a ta . Since a 2k d e p e n d s o n th e ty p e o f sp e c tru m e s tim a te w e o b ta in , th e m o d e l o rd e r is also a fu n ctio n o f th e c riterio n . T h e e x p e rim e n ta l re su lts given in th e re fe re n c e s ju s t c ite d in d ic a te th a t th e m o d e l-o rd e r se lectio n c rite ria d o n o t y ield d efin itiv e re su lts. F o r ex am p le, U lry ch an d B ish o p (1975), J o n e s (1976), a n d B e rry m a n (1978), fo u n d th a t th e F P E ( p ) c rite rio n te n d s to u n d e re s tim a te th e m o d e l o rd e r. K a s h y a p (1980) sh o w ed th a t th e A IC c rite rio n is sta tistically in c o n s is te n t as N oo. O n th e o th e r h a n d , th e M D L in fo rm a tio n c rite rio n p ro p o s e d b y R issan n is statistically c o n sisten t. O th e r e x p e rim e n ta l re su lts in d ic a te th a t fo r sm a d a ta len g th s, th e o r d e r o f th e A R m o d e l sh o u ld b e se le c te d to b e in th e ra n g e N /3 to N f l fo r g o o d resu lts. I t is a p p a r e n t th a t in th e a b se n c e o f any p r io r in fo rm a tio n r e g ard in g th e p h y sical p ro c e ss th a t re su lte d in th e d a ta , o n e sh o u ld tr y d iffer e n t m o d e l o r d e rs a n d d iffe re n t c rite ria an d , finally, c o n s id e r th e d iffe re n t resu lts. _L ) - J (12.3.33)
Sec. 12.3
933
12.3.7 MA Model for Power Spectrum Estimation

A s show n in S ectio n 12.3.1, th e p a r a m e te r s in a M A (^ ) m o d e l a re re la te d to th e sta tistical a u to c o rre la tio n yxx(m) by (12.3.10). H o w e v e r,
(12.3.35)
w h e re th e co efficien ts [dm] a re re la te d to th e M A p a r a m e te r s by th e ex p ressio n
g-\ m \
(12.3.36)
(12.3.37)
a n d th e p o w e r sp e c tru m fo r th e MA(< 7 ) p ro cess is (12.3.38) It is a p p a r e n t fro m th e se e x p ressio n s th a t w e d o n o t h av e to solve fo r th e M A p a ra m e te rs {>*} to e stim a te th e p o w e r sp e c tru m . T h e e stim a te s o f th e a u to c o r re latio n yxx( m) fo r |?| < q suffice. F ro m such e stim a te s w e c o m p u te th e e s tim a te d M A p o w e r sp e ctru m , g iven as
(12.3.39)
w h ich is id e n tic a l to th e classical (n o n p a ra m e tric ) p o w e r sp e c tru m e stim a te d e sc rib ed in S ectio n 12.1. T h e re is an a lte rn a tiv e m e th o d fo r d e te rm in in g {bk} b a se d o n a h ig h -o rd e r A R a p p ro x im a tio n to th e M A p ro cess. T o b e specific, let th e MA(<?) p ro cess b e m o d e le d by an A R ( p ) m o d el, w h e re p > > q. T h e n B( z ) = 1 / A( z ) , o r eq u iv alen tly , B ( z ) A ( z ) = 1. T h u s th e p a ra m e te rs {bk} a n d {ak} a re r e la te d by a co n v o lu tio n sum , w h ich can b e e x p ressed as n = 0 n tÔ
(12.3.40)
w h ere {a} a re th e p a r a m e te r s o b ta in e d by fittin g th e d a ta to an A R ( p ) m odel. A lth o u g h th is set o f e q u a tio n s c a n be easily solved fo r th e {*}, a b e tte r fit is o b ta in e d b y u sin g a le a st-sq u a re s e r r o r c rite rio n . T h a t is, w e fo rm th e sq u a re d e r ro r 2
ao = l,
ak = 0,
k <0
(12.3.41)
which is minimized by selecting the MA(?) parameters {*}. The result of this
9 34
Chap. 12
minimization is
b = - R ~i J r aa w h ere th e e le m e n ts o f R aa(\i ~ y |) = a n d raa a re given as (12.3.42)
p-tf-jl
22
n=0
i,j = l,2,...,q (12.3.43) n+ ' I = 1, 2, . . . ,
p-i
raa 0 ) = 2 2
T h is least sq u a re s m e th o d fo r d e te rm in in g th e p a r a m e te r s o f th e MA(<j) m o d el is a ttr ib u te d to D u rb in (1959). I t h as b e e n sh o w n by K a y (1988) th a t this e s tim a tio n m e th o d is a p p ro x im a te ly th e m ax im u m lik e lih o o d u n d e r th e a ssu m p tio n th a t th e o b se rv e d p ro c e ss is G au ssian . T h e o r d e r q o f th e M A m o d e l m a y b e d e te rm in e d e m p iric a lly b y se v era l m e th o d s. F o r e x a m p le , th e A IC fo r M A m o d els h a s th e sa m e fo rm as fo r A R m o d els, A lC (tf) = l n < 4 + ^ (12.3.44)
w h ere a 2q is an e s tim a te o f th e v arian ce o f th e w h ite noise. A n o th e r a p p ro a c h , p ro p o se d by C h o w (1972b), is to filter th e d a ta w ith th e in v e rse MA(<y) filter an d test th e filtered o u tp u t fo r w h iten ess.
12.3.8 ARMA Model for Power Spectrum Estimation

T h e B u rg alg o rith m , its v a riatio n s, a n d th e le a s t-s q u a re s m e th o d d e sc rib e d in th e p rev io u s sectio n s p ro v id e re lia b le h ig h -re so lu tio n sp e c tru m e s tim a te s b a s e d o n th e A R m o d el. A n A R M A m o d e l p ro v id e s us w ith an o p p o r tu n ity to im p ro v e on th e A R sp e c tru m e s tim a te , p e rh a p s, by using fe w e r m o d e l p a ra m e te rs . T h e A R M A m o d e l is p a rtic u la rly a p p r o p ria te w h e n th e sig n a l h as b e e n co r ru p te d b y no ise. F o r e x am p le, su p p o s e th a t th e d a ta x ( n ) a re g e n e r a te d by an A R sy stem , w h ere th e sy stem o u tp u t is c o rru p te d by ad d itiv e w h ite n o ise . T h e z-tra n sfo rm o f th e a u to c o rre la tio n o f th e re s u lta n t signal can b e e x p re sse d as
r Az) = J u m ? T )+ "
_ a l +<t2A ( z ) A ( z j)
(12.3.45)
w h ere cr2 is th e v a ria n c e o f th e a d d itiv e noise. T h e re fo re , th e p ro c e ss x ( n ) is A R M A (/j, p), w h e re p is th e o rd e r o f th e a u to c o rre la tio n p ro c e ss . T h is re la tio n sh ip p ro v id e s so m e m o tiv a tio n fo r in v estig atin g A R M A m o d e ls fo r p o w er sp e c tru m estim a tio n . A s w e h a v e d e m o n s tra te d in S ectio n 12.3.1, th e p r a m e te rs o f th e A R M A m o d el a re re la te d to th e a u to c o rre la tio n b y th e e q u a ' on in (12.3.4). F o r lags
Sec. 12.3
935
|m | > q, th e e q u a tio n involves o n ly th e A R p a r a m e te r s {a*}. W ith e stim a te s s u b s titu te d in p lace o f yxx(m), w e can solve th e p e q u a tio n s in (12.3.5) to o b ta in o*. F o r h ig h -o rd e r m o d e ls, h o w ev er, th is a p p ro a c h is likely to yield p o o r e stim a te s o f th e A R p a ra m e te rs d u e to th e p o o r e stim a te s o f th e a u to c o rre la tio n fo r larg e lags. C o n s e q u e n tly , th is a p p ro a c h is n o t re c o m m e n d e d . A m o re re lia b le m e th o d is t o c o n s tru c t an o v e rd e te rm in e d se t o f lin ear e q u a tio n s fo r m > q, a n d to u se th e m e th o d o f le a st s q u a re s o n th e se t o f o v e r d e te r m in e d e q u a tio n s , as p ro p o se d by C a d zo w (1979). T o e la b o ra te , su p p o se th a t th e a u to c o rre la tio n se q u e n c e can b e a c c u ra te ly e s tim a te d u p to lag M, w h ere M > p + q. T h e n w e can w rite th e follow ing se t o f lin e a r e q u a tio n s: rxx(q) rxx(q + 1) ,rxx( M - 1) o r eq u iv a le n tly ,
R r r S --Tr
rxx(q ~ 1) rxx(q) rxx( M - 2 )
rXx(q - p + 1) -] ~a\ o2 rxx(q - p + 2)
= _
~rxx(q + 1)~ rxx(q + 2 )
rxx( M - p)
. r xx( M )
_ (12.3.46)
(12.3.47)
Since is o f d im en sio n ( M q) x p, an d M q > p w e c a n use th e le a st-sq u a re s c rite rio n to so lv e fo r th e p a r a m e te r v e c to r a. T h e re su lt o f th is m in im iz atio n is a = -(R ' R r (12.3.48)
T h is p ro c e d u re is called th e least-squares mo d i f i e d Y u l e - W a l k e r me t hod. A w eig h tin g fa c to r can also b e a p p lie d to th e a u to c o rre la tio n se q u e n c e to d e e m p h a size th e less re lia b le e s tim a te s for larg e lags. O n c e th e p a ra m e te rs fo r th e A R p a r t o f th e m o d e l h av e b e e n e stim a te d as in d ic a te d ab o v e , w e h a v e th e system p A( z ) = 1 + ^ t= i * (12.3.49)
T h e se q u e n c e x ( n ) can n o w b e filte re d by th e F IR filter A ( z ) to yield th e se q u e n c e p v(n) = x (n ) + 2 2 & tx (n k) *=i n = 0 ,1 ,..., N 1 (12.3.50)
T h e c a scad e o f th e A R M A ( p , q) m o d e l w ith A ( z ) is a p p ro x im a te ly th e MA(<?) p ro c e ss g e n e r a te d by th e m o d e l B{z). H e n c e w e can ap p ly th e M A e s tim a te given in th e p re c e d in g se ctio n to o b ta in th e M A sp e c tru m . T o b e specific, th e filtered se q u e n c e v(n) f o r p < n < N 1 is u se d to fo rm th e e s tim a te d c o rre la tio n se q u e n c e s ruv( m), fro m w h ich w e o b ta in th e M A sp e c tru m
^ A( / ) = L
rw ( m ) e ~ ^
(12.3.51)
936
Chap. 12
F irst, w e o b se rv e th a t th e p a ra m e te rs {>*} a re n o t r e q u ire d to d e te r m in e th e p o w e r sp e ctru m . S eco n d , w e o b se rv e th a t r(m ) is an e s tim a te o f th e a u to c o r re la tio n fo r th e M A m o d e l g iv en by (12.3.10). In fo rm in g th e e stim a te rvv( m) , w eig h tin g (e.g., w ith th e B a rtle tt w in d o w ) m ay b e u se d to d e e m p h a s iz e c o rre la tio n e s tim a te s fo r larg e lags. In a d d itio n , th e d a ta m ay b e filte re d by a b a c k w a rd filter, th u s cre a tin g a n o th e r se q u e n c e , say v b(n), so th a t b o th v(n) a n d v b(n) can b e u se d in fo rm in g th e e stim a te o f th e a u to c o rre la tio n r(m ), as p ro p o s e d by K a y (1980). F inally, th e e s tim a te d A R M A p o w e r sp e c tru m is pMA/ P * RMA( f ) = ----------- v j L- l l l -----(12.3.52)
1 + E
T h e p ro b le m o f o rd e r se lectio n fo r th e A R M A ( p , g ) m o d e l h as b e e n in v es tig a te d by C how (1972a, b) a n d B ru z z o n e a n d K a v e h (1980). F o r th is p u rp o s e th e m in im u m o f th e A IC in d ex A I C ( p , q) = I n a l " + ~ ~ (12.3.53)
can b e u sed , w h e re 6 l pq is a n e s tim a te o f th e v a ria n c e o f th e in p u t e rro r. A n a d d itio n a l te s t o n th e a d e q u a c y o f a p a rtic u la r A R M A ( p , q) m o d e l is to filter th e d a ta th ro u g h th e m o d e l a n d te s t fo r w h iten ess o f th e o u tp u t d a ta . T h is re q u ire s th a t th e p a ra m e te rs o f th e M A m o d e l b e c o m p u te d fro m th e e s tim a te d a u to c o rre la tio n , u sing sp e ctral fa c to riz a tio n to d e te rm in e B(z) fro m D( z ) = F o r a d d itio n a l re a d in g o n A R M A p o w e r s p e c tru m e s tim a tio n , th e r e a d e r is re fe rre d to th e p a p e rs by G ra u p e e t al. (1975), C a d zo w (1981, 1982), K ay (1980), an d F rie d la n d e r (1982b).
12.3.9 Some Experimental Results

In th is se ctio n w e p re s e n t so m e e x p e rim e n ta l re su lts on th e p e rfo rm a n c e o f A R a n d A R M A p o w e r s p e c tru m e s tim a te s o b ta in e d b y u sin g artificially g e n e ra te d d ata. O u r o b jectiv e is to c o m p a re th e s p e c tra l e stim a tio n m e th o d s o n th e basis o f th e ir fre q u e n c y re so lu tio n , bias, a n d th e ir ro b u stn e s s in th e p re se n c e o f ad d itiv e noise. T h e d a ta c o n sist o f e ith e r o n e o r tw o sin u so id s a n d a d d itiv e G a u s sia n noise. T h e tw o sin u so id s a re sp a c e d A f a p a rt. C lea rly , th e u n d e rly in g p ro cess is A R M A ( 4 ,4). T h e re su lts th a t a re sh o w n em p lo y a n A R ( p) m o d e l fo r th ese d a ta . F o r h ig h sig n a l-to -n o ise ra tio s (S N R s) w e e x p e c t th e A R ( 4 ) to b e a d e q u a te . H o w e v e r, fo r low S N R s, a h ig h e r-o rd e r A R m o d e l is n e e d e d to a p p ro x im a te th e A R M A ( 4 ,4) p ro cess. T h e re su lts given b e lo w a re c o n s iste n t w ith th is sta te m e n t. T h e S N R is d efin ed as 101og10 w h e re a 2 is v a ria n c e o f th e ad d itiv e n oise a n d A is th e a m p litu d e o f th e sinusoid. In F ig. 12.6 w e illu stra te th e re su lts fo r N = 20 d a ta p o in ts b a s e d o n an A R (4 ) m o d e l w ith a S N R = 20 dB a n d A f = 0.13. N o te th a t th e Y u le - W a lk e r m e th o d gives an ex tre m e ly sm o o th ( b ro a d ) sp e c tra l e s tim a te w ith sm a ll p e a k s . I f A f is
Sec. 12.3
937
-1 0
Burg
//II
II
LS
20 points 4 poles SNR = 20 dB
E
3 5
N -------- Yule-Walker ^
r I i i i i m j i i i i l i i i i j j n j i l - I i l l n i i i i i . i j l i i l i i - L l i i i x i j . j . i l i j i l j i i i i i i j i l i m i i i i i l n m i n | J | n t u n i ___ f
.20
.22
.24
.26
.28 .30 .32 .34 Frequency (cycles/sample)
.36
.38
.40
Figure 12.6
Comparison of A R spectrum estimation methods.
d e c re a s e d to A f = 0.07, th e Y u le -W a lk e r m e th o d n o lo n g e r reso lv es th e p e a k s as illu stra te d in Fig. 12.7. S o m e b ias is also e v id e n t in th e B u rg m e th o d . O f c o u rse , by in c re a sin g th e n u m b e r o f d a ta p o in ts th e Y u le -W a lk e r m e th o d e v e n tu a lly is able to reso lv e th e p eak s. H o w e v e r, th e B u rg a n d le a st-s q u a re s m e th o d s a re clearly s u p e rio r fo r s h o rt d a ta reco rd s. T h e effe ct o f ad d itiv e n oise o n th e e s tim a te is illu stra te d in Fig. 12.8 fo r th e le a s t-s q u a re s m e th o d . T h e effect o f filter o rd e r o n th e B u rg a n d le a st-sq u a re s m e th o d s is illu stra te d in F igs. 12.9 a n d 12.10, resp ec tiv ely . B o th m e th o d s ex h ib it s p u rio u s p e a k s as th e o r d e r is in crease d to p = 12. T h e effe ct o f in itial p h a se is illu stra te d in Figs. 12.11 a n d 12.12 fo r th e B u rg an d le a s t-s q u a re s m e th o d s. I t is clear th a t th e le a st-sq u a re s m e th o d ex h ib its less se n sitiv ity to in itial p h a s e th a n th e B u rg alg o rith m . A n e x a m p le o f lin e sp littin g fo r th e B u rg m e th o d is sh o w n in Fig. 12.13 w ith p = 12. It d o e s n o t o c c u r fo r th e A R (8 ) m o d el. T h e le a st-sq u a re s m e th o d d id n o t e x h ib it lin e sp littin g u n d e r th e sa m e co n d itio n s. O n th e o th e r h a n d , th e line sp littin g o n th e B u rg m e th o d d isa p p e a re d w ith an in crease in th e n u m b e r o f d a ta p o in ts N . F ig u re s 12.14 a n d 12.15 illu stra te th e re so lu tio n p r o p e rtie s o f th e B u rg an d le a st-s q u a re s m e th o d s f o r A f = 0.07 a n d N = 20 p o in ts a t low S N R (3 dB ). S ince th e ad d itiv e n o ise p ro c e ss is A R M A , a h ig h e r-o rd e r A R m o d e l is re q u ire d t o p ro v id e a g o o d a p p ro x im a tio n a t low S N R . H e n c e th e fre q u e n c y re so lu tio n im p ro v e s as th e o r d e r is in crease d .
938
Chap. 12
Normalized power spectrum (dB)
Frequency (cycles/sample)
Figure 12.7 Comparison of A R spectrum estimation methods.
Normalized power spectrum
(dB)
Frequency (cycles/sample) Figure 1Z8 Effect of additive noise on LS method.
Sec. 12.3
939
Figure 12.9
Effect of filler order of Burg method.
(dB)
Figure 12.10
Effect of filter order
od
LS method.
Chap. 12
/
Frequency (cycles/sample) Figure 12.11 Effect of initial phase on Burg method.
-6
20 points 4 poles SNR = 15 dB / , = -23 f l = -36
16 sinusoid phases
-1 2
_ J4 r Hi ! I1111 1l 1 .20 .22
.26
.28 .30 .32 .34 Frequency (cycles/sample)
.36
.38
.40
Figure 12.12
Effect of initial phase on LS m ethod.
Sec. 12.3
941
(dB)
Frequency (cycles/sample) Figure 12.13 Line splitting in Burg method.
FigHre 12.14
Frequency resolution o f B urg m ethod w ith N = 20 points.
942
Chap. 12
Frequency (cycles/sample) Figure 12,15 Frequency resolution of LS method with N 20 points.
T h e F P E fo r th e B u rg m e th o d is illu stra te d in Fig. 12.16 f o r a n S N R = 3 dB . F o r th is S N R th e o p tim u m v a lu e is p = 12 a c co rd in g to th e F P E crite rio n . T h e B u rg a n d le a st-sq u a re s m e th o d s w ere also te s te d w ith d a ta fro m a n a r ro w b a n d p ro cess, o b ta in e d by exciting a fo u r-p o le (tw o p a irs o f c o m p le x -c o n ju g a te p o les) n a rro w b a n d filter a n d se lectin g a p o rtio n o f th e o u tp u t se q u e n c e fo r th e d a ta re c o rd . F ig u re 12.17 illu stra te s th e su p e rp o s itio n o f 20 d a ta re c o rd s o f 20 p o in ts e ach . W e o b se rv e a relativ ely sm all v ariab ility . In c o n tra st, th e B u rg m e th o d e x h ib ite d a m u ch la rg e r v ariab ility , a p p ro x im a te ly a fa c to r o f 2 c o m p a re d to th e le a st-sq u a re s m e th o d . T h e re su lts sh o w n in F igs. 12.6 th ro u g h 12.17 a re ta k e n fro m P o o le (1981). F in ally , w e sh o w in Fig. 12.18 th e A R M A ( 1 0 ,10) sp e c tra l e s tim a te s o b ta in e d by K ay (1980) fo r tw o sin u so id s in n o ise using th e le a s t-s q u a re s A R M A m e th o d d e sc rib e d in S ectio n 12.3.8, as a n illu stra tio n o f th e q u a lity o f p o w e r sp e c tru m e s tim a tio n o b ta in e d w ith th e A R M A m odel.
12.4 MINIMUM VARIANCE SPECTRAL ESTIMATION

T h e sp e c tra l e s tim a to r p r o p o s e d by C a p o n (1969) w as in te n d e d fo r u se in large seism ic a rra y s fo r fre q u e n c y -w a v e n u m b e r e stim a tio n . I t w as la te r a d a p te d to sin g le-tim e-series s p e c tru m e s tim a tio n b y L aco ss (1971), w h o d e m o n s tr a te d th a t
Sec. 12.4
Minimum Variance Spectral Estimation
943
l.l
1.0
0.9
0.8
o
c
|3 0.6
D
t 0.7 fe.0.5
| 0.4 0.3
0.2
0.1 0
6 8 1 0 1 2 1 4 1 6 1 8 N u m b e rof p o le s
F ig u re 12.16 F in al p red ictio n erro r fo r B u rg estim ate.
20p o in ts 4p o le s
;1 .;t = .95f^ - 23'
Zi
20d a tase q u e n ce s
= ,95e *
-12
.20
.22
.2 4
.26
.28 .30 .32 .3 4 F req u e n cy(cy cle s/ sa m p le )
.36
.3 8
.40
Figure 12.17
Effect of starting point in sequence on LS m ethod.
944
Chap. 12
Frequency (cycles/sample)
(b)
Figure 12.18 A R M A (10, 10) pow er spectrum estim ates from p ap er by Kay (1980). R eprin ted w ith perm ission from the IE E E .
th e m e th o d p ro v id es a m in im u m v a ria n c e u n b ia s e d e stim a te o f th e sp e c tra l c o m p o n e n ts in th e signal. F o llo w in g th e d e v e lo p m e n t o f L aco ss, le t us c o n s id e r an F I R filter w ith co ef ficients a*, 0 < k < p, to b e d e te rm in e d . U n lik e th e lin e a r p re d ic tio n p ro b le m , w e d o n o t c o n stra in ao to b e u n ity . T h e n , if th e o b se rv e d d a ta * ( ) , 0 < n < N 1, a re p assed th ro u g h th e filter, th e re sp o n s e is
p
y(n) = 21<ikx(n - k ) = X '(n )a k=0 (12.4.1)
Sec. 12.4
Minimum Variance Spectral Estimation
9 45
X l (n) = [jr(n ) x ( n 1) x ( n - p) ] is th e d a ta v e c to r a n d a is th e filter c o e f ficien t v e c to r. If w e assu m e th a t [ * ( )] = 0, th e v arian ce o f th e o u tp u t se q u e n c e is a 2 = [ |y ( n ) |2] = [a * 'X * (n )X ' (n)a] (12.4.2) = a*T a w h e re T xx is th e a u to c o rre la tio n m a trix o f th e se q u e n c e jc(n), w ith e le m e n ts yxx(m). T h e filter co efficien ts a re se le c te d so th a t a t th e fre q u e n c y f s, th e fre q u e n c y re sp o n s e o f th e F I R filter is n o rm a liz e d to u n ity , th a t is, ; a ke - j2*k}' = 1
k= 0
T h is c o n s tra in t can also b e w ritte n in m a trix fo rm as E * '( //) a = 1 w h e re E '( / , ) = [ l e*2** ej2*pf' ] (12.4.3)
B y m in im izin g th e v a ria n c e a 2 su b je c t to th e c o n s tra in t (12.4.3), w e o b ta in an F IR filter th a t p a sse s th e fre q u e n c y c o m p o n e n t / / u n d is to rte d , w hile c o m p o n e n ts d is ta n t fro m / / a re se v e re ly a tte n u a te d . T h e re su lt o f th is m in im iz a tio n is sh o w n by L aco ss to le a d to th e co efficien t v e c to r a = r - 1E* E* ( /,) (12.4.4)
If a is su b s titu te d in to (12.4.2), w e o b ta in th e m in im u m v arian ce ? 1 C T min = ------------- i---------E 'o n r - 'E 'a , ) (12.4.5)
T h e ex p re ssio n in (12.4.5) is th e m in im u m v aria n c e p o w e r sp e c tru m e s tim a te a t th e fre q u e n c y f s. B y ch an g in g / / o v e r th e ra n g e 0 < f t < 0.5, w e can o b ta in th e p o w e r s p e c tru m e s tim a te . I t sh o u ld b e n o te d th a t a lth o u g h E ( / ) c h an g es w ith th e ch o ice o f fre q u e n c y , F ^ 1 is c o m p u te d o n ly o n ce. A s d e m o n s tr a te d by L acoss (1971), th e c o m p u ta tio n o f th e q u a d ra tic fo rm E ' c a n be d o n e w ith a sin g le D F T . W ith an e s tim a te R o f th e a u to c o rre la tio n m a trix su b s titu te d in p la c e o f T j j , w e o b ta in th e m in im u m v a ria n c e p o w e r s p e c tru m e s tim a te o f C a p o n as p W ( f ) = ----------------------P (/)R ^ E * (/) (12.4.6)
I t h a s b e e n sh o w n b y L aco ss (1971) th a t th is p o w e r sp e c tru m e s tim a to r yields e s tim a te s o f th e sp e c tra l p e a k s p ro p o rtio n a l to th e p o w e r a t th a t fre q u e n c y . In c o n s tra s t, th e A R m e th o d s d e sc rib e d in S ectio n 12.3 re s u lt in e s tim a te s o f th e sp e c tra l p e a k s p ro p o rtio n a l to th e s q u a re o f th e p o w e r a t th a t freq u e n cy . T h is m in im u m v a ria n c e m e th o d is basically a filter b a n k im p le m e n ta tio n fo r th e sp e c tru m e s tim a to r. I t d iffers basically fro m th e filter b a n k in te r p re ta tio n o f th e p e rio d o g ra m in th a t th e filter coefficients in th e C a p o n m e th o d a re o p tim ized .
94 6
Chap. 12
E x p e rim e n ts o n th e p e rfo rm a n c e o f this m e th o d c o m p a re d w ith th e p e r fo rm a n c e o f th e B u rg m e th o d h av e b e e n d o n e by L acoss (1971) an d o th e rs. In g e n e ra l, th e m in im u m v a ria n c e e stim a te in (12.4.6) o u tp e rfo rm s th e n o n p a ra m e tric sp e c tra l e s tim a to rs in fre q u e n c y re so lu tio n , b u t it d o e s n o t p ro v id e th e high fre q u en cy re so lu tio n o b ta in e d fro m th e A R m e th o d s o f B u rg a n d th e u n c o n s tra in e d least sq u a res. E x te n siv e c o m p a riso n s b e tw e e n th e B u rg m e th o d a n d th e m ini m u m v arian ce m e th o d h a v e b e e n m a d e in th e p a p e r by L aco ss. F u rth e rm o re , B u rg (1972) d e m o n s tra te d th a t fo r a k n o w n c o rre la tio n s e q u e n c e , th e m in im u m v a rian ce sp e c tru m is re la te d to th e A R m o d e l sp e c tru m th ro u g h th e e q u a tio n (12.4.7) w h ere r*x R(f, k) is th e A R p o w e r sp e c tru m o b ta in e d w ith an A R (it) m o d el. T h u s th e recip ro c al o f th e m in im u m v a ria n c e e stim a te is e q u a l to th e a v e ra g e o f th e recip ro c als o f all sp e c tra o b ta in e d w ith A R ( k ) m o d e ls fo r 1 < k < p. Since lo w -o rd e r A R m o d els, in g e n e ra l, d o n o t p ro v id e g o o d re so lu tio n , th e av erag in g o p e ra tio n in (12.4.7) re d u c e s th e fre q u e n c y re so lu tio n in th e sp e c tra l estim a te . H e n c e w e c o n clu d e th a t th e A R p o w e r sp e c tru m e stim a te o f o r d e r p is s u p e rio r to th e m in im u m v a ria n c e e stim a te o f o rd e r p + 1. T h e re la tio n sh ip given by (12.4.7) re p re se n ts a fre q u e n c y -d o m a in re la tio n sh ip b e tw e e n th e C a p o n m in im u m v a ria n c e e s tim a te an d th e B u rg A R e stim a te . A tim e -d o m a in re la tio n sh ip b e tw e e n th e se tw o e s tim a te s also can b e e s ta b lis h e d as show n by M u sicu s (1985). T h is has led to a c o m p u ta tio n a lly efficien t alg o rith m fo r th e m in im u m v a ria n c e e s tim a te . A d d itio n a l re fe re n c e s to th e m e th o d o f C a p o n a n d c o m p a riso n s w ith o th e r e s tim a to rs can b e fo u n d in th e lite ra tu re . W e cite th e p a p e rs o f C a p o n a n d G o o d m a n (1971), M a rz e tta (1983), M a rz e tta a n d L a n g (1983, 1984), C a p o n (1983), a n d M c D o n o u g h (1983).
12.5 EIGENANALYSIS ALGORITHMS FOR SPECTRUM ESTIMATION

In S ectio n 12.3.8 w e d e m o n s tra te d th a t an A R ( p ) p ro c e ss c o r ru p te d by ad d itiv e (w h ite) n o ise is e q u iv a le n t to a n A R M A (/? , p ) p ro cess. In th is se c tio n w e c o n sid e r th e sp ecial case in w hich th e signal c o m p o n e n ts a re sin u so id s c o r ru p te d by ad d itiv e w h ite n o ise . T h e a lg o rith m s a re b a se d o n an e ig e n -d e c o m p o sitio n o f th e c o rre la tio n m atrix o f th e n o ise -c o rru p te d signal. F ro m o u r p re v io u s d isc u ssio n o n th e g e n e ra tio n o f sin u so id s in C h a p te r 4, w e recall th a t a re a l sin u so id al signal can b e g e n e ra te d via th e d iffe re n c e e q u a tio n , x (n) = d]X(n 1) 02x ( n 2) (12.5.1)
w h e re a\ = 2 c o s 2 jt/* , 02 = 1, a n d initially, x ( - l ) = 1, x ( 2) = 0. T h is system h as a p a ir o f c o m p le x -c o n ju g a te p o le s ( a t f = f a n d f = f t ) a n d th e re fo re g e n e ra te s th e sin u so id x(n) cos 2 n f kn, fo r n > f
Sec. 12.5
Eigenanalysis Algorithms for Spectrum Estimation
947
In g e n e ra l, a signal co n sistin g o f p sin u so id al c o m p o n e n ts satisfies th e d iffe r e n c e e q u a tio n

2P
x( n) = - 2 2 a x (n ~ m )
m=l
(12.5.2)
a n d c o rre sp o n d s to th e sy stem w ith sy stem fu n ctio n H{ z ) = --------------------
2p
(12.5.3)
..tZ
1 T h e p o ly n o m ia l
2p
A( z ) = 1 +
m=l
(12-5 '4 >
h a s 2 p ro o ts o n th e u n it circle w hich c o rre sp o n d to th e fre q u e n c ie s o f th e sinusoids. N ow , su p p o s e th a t th e sinusoids a re c o rru p te d by a w h ite n o ise seq u en ce ic(rt) w ith [ t i f ( n ) |2] = <r2. T h e n w e o b se rv e th a t y( n) = x ( n ) + w( n ) If w e s u b s titu te x( n ) = y ( n ) w( n) in (12.5.2), w e o b ta in
2p
( 12.5.5)
y ( n ) - w( n) ~ - ^ [ . v ( n - m) - w( n - m) ]am m=1 o r, eq u iv a le n tly ,
2p
T
2p
amy( n m ) Y J omw( n m) (12.5.6)
m = 0
m = 0
w h e re , by d efin itio n , a0 = 1. W e o b se rv e th a t (12.5.6) is th e d ifferen ce e q u a tio n fo r a n A R M A ( p , p ) p r o cess in w h ich b o th th e A R a n d M A p a ra m e te rs a r e id e n tic a l. T h is sy m m etry is a c h a ra c te ristic o f th e sin u so id al signals in w h ite n o ise. T h e d iffe re n c e e q u a tio n in (12.5.6) m ay b e e x p re sse d in m atrix fo rm as Y 'a = W 'a (12.5.7)
w h e re Y r = [y (n ) y ( n 1) y ( n - 2 p ) ] is th e o b s e rv e d d a ta v e c to r o f d i m e n s io n (2 p 4 - 1 ), W ' = [ wi n ) w( n 1) w( n - 2 p ) ] is th e n o ise vector, a n d a = [ 1 a\ ct2P ] is th e c o effic ie n t v e cto r. If w e p re m u ltip ly (12.5.7) by Y a n d ta k e th e e x p e c te d v a lu e , w e o b ta in E ( Y Y > = ( Y W ') a = [ ( X -I- W )W r]a (12.5.8) r yya = < a w h e re w e h a v e u se d th e a ssu m p tio n th a t th e s e q u e n c e w( n ) is z e ro m e a n a n d w h ite , a n d X is a d e te rm in is tic signal.
9 48
Chap. 12
T h e e q u a tio n in (12.5.8) is in th e fo rm o f a n e ig e n e q u a tio n , th a t is, ( I V - < r 2I ) a = 0 (12.5.9)
w h e re c 2 is an eig e n v a lu e o f th e a u to c o rre la tio n m a trix T >7. T h e n th e p a ra m e te r v e c to r a is an e ig e n v e c to r a s so ciated w ith th e eig e n v a lu e cr2. T h e e ig e n e q u a tio n in (12.5.9) form s th e basis fo r th e P isa re n k o h a rm o n ic d e c o m p o s itio n m e th o d .
12.5.1 Pisarenko Harmonic Decomposition Method

F o r p ra n d o m ly -p h a se d sin u so id s in ad d itiv e w h ite n o ise , th e a u to c o r re la tio n val u es are y ( 0 ) = l + Y 2 Pi (12.5.10) Yyy(k) = 2 2 COS^ n f , k k it 0
w h e re Pj = A ] / 2 is th e a v e ra g e p o w e r in th e ith sin u so id an d A, is th e c o rre sp o n d in g am p litu d e. H e n c e w e m ay w rite

c o s 2 jt/i c o s 2 tt f i c o s 4 j t /2
cos 4 ;r/i
cos 2 n f p n r p r cos 4 n f p P2
r y .w d ) n Yyy( 2) (12.5.11)
- Y y y (p ) -
.c o s 2 n p f i
cos 2 ttp /2
cos 2:
-Pp-
If w e k n o w th e fre q u e n c ie s f j , 1 < i < p , w e can u se this e q u a tio n to d e te rm in e th e p o w e rs o f th e sin u so ids. In p la c e o f yxx(m), w e use th e e s tim a te s rxx(m). O n c e th e p o w e rs a re k n o w n , th e n o ise v a ria n c e c a n b e o b ta in e d fro m (12.5.10) as
= ''v v ( O )
y.p,
1=1
(12.5.12)
T h e p ro b le m th a t re m a in s is to d e te rm in e th e p fre q u e n c ie s /, 1 < 1 < p, w hich , in tu rn , re q u ire k n o w led g e o f th e e ig e n v e c to r a c o rre sp o n d in g to th e e ig e n v a lu e cr2. P isa re n k o (1973) o b se rv e d [see also P a p o u lis (1984) a n d G re n a n d e r an d S zego (1958)] th a t fo r an A R M A p ro c e ss co n sistin g o f p sin u s o id s in ad d itiv e w h ite n o ise , th e v a ria n c e a 2 c o rre sp o n d s to th e m in im u m e ig e n v a lu e o f r vv w hen th e d im en sio n o f th e a u to c o rre la tio n m a trix e q u a ls o r e x cee d s (2 p + 1 ) x (2 p + 1 ) . T h e d e s ire d A R M A co efficien t v e c to r c o rre sp o n d s to th e e ig e n v e c to r asso ciated w ith th e m in im u m eig en v alu e. T h e re fo re , th e fre q u e n c ie s / , , 1 < i < p are o b ta in e d fro m th e ro o ts o f th e p o ly n o m ia l in (12.5.4), w h e re th e c o effic ie n ts are th e e le m e n ts o f th e e ig e n v e c to r a c o rre sp o n d in g to th e m in im u m eig e n v a lu e <r2. In su m m ary , th e P isa re n k o h a rm o n ic d e c o m p o s itio n m e th o d p ro c e e d s as follow s. F irst w e e s tim a te r yy fro m th e d a ta (i.e., w e fo rm th e a u to c o rre la tio n m a trix R yy). T h e n w e find th e m in im u m e ig e n v a lu e a n d th e c o rre sp o n d in g m in im u m e ig en v ecto r. T h e m in im u m e ig e n v e c to r y ield s th e p a r a m e te r s o f th e
Sec. 12.5
949
A R M A ( 2 /? ,2 p ) m o d el. F ro m (12.5.4.) w e can c o m p u te th e ro o ts th a t c o n s titu te th e fre q u e n c ie s {/,}. B y using th e se fre q u e n c ie s, w e can solve (12.5.11) fo r th e sig n al p o w e rs {/,} by su b stitu tin g th e e stim a te s r yy(m) fo r y vy(m). A s w ill b e se e n in th e follow ing ex am p le, th e P is a re n k o m e th o d is b a s e d o n th e u se o f a n o ise su b sp ace e ig e n v e c to r to e stim a te th e fre q u e n c ie s o f th e sinusoids. Exam ple 125.1 Suppose that we are given the autocorrelation values yv v ,(0) = 3, yv,( l) = 1, and Yyy (2) = 0 for a process consisting of a single sinusoid in additive white noise. D eter mine the frequency, its power, and the variance of the additive noise. Solution The correlation matrix is r3 1 -0 -3 -A
g ( k ) = 1 3 -
i o3 1 1 3J
The minimum eigenvalue is the smallest root of the characteristic polynomial 1

X
0
1
= (3 - k)(k2 - 6X + 7} = 0
3- k j
T herefore, the eigenvalues are A .) = 3, k2 = 3 + ->/2. ^ = 3 - \ f l . The variance of the noise is a] = Amjn = 3 - V 2 The corresponding eigenvalue is the vector that satisfies (12.5.9), that is.
-V 2 1 . 0 1 y/2 1 0 1 y /2 . -02' - 1 = -o0 .0 .
The solution is ai = - y / 2 . and a2 = \. The next step is to use the value aj and a2 to determ ine the roots of the polynomial in (12.5.4). W e have
s2 - n /2 z +
1= 0
Thus Zl Z2 1 . . 1 V2 J V2
Note that |zi| = |z2| = 1, so that the roots are on the unit circle. The corresponding frequency is obtained from Z; =*'" 1 s / 2 + y V2
which yields / i = | . Finally, the pow er of the sinusoid is

P \ COS 2 n f l = Yyy ( 1 ) = 1
/>! = V 2
and its amplitude is A = J 2P X = \/2v^.
950
Chap. 12
As a check on our computations, we have a l = y,.v(0) - Pi = 3 - -Jl which agrees with An,jn.
12.5.2 Eigen-decomposition of the Autocorrelation Matrix for Sinusoids in White Noise

In th e p rev io u s d iscu ssio n w e assu m ed th a t th e sin u so id al signal co n sists o f p real sinusoids. F o r m a th e m a tic a l co n v e n ie n c e w e shall n o w a ssu m e th a t th e signal co n sists o f p co m p lex sin u so id s o f th e fo rm
p
x(n) =
22
;=i
A ,eJ(2,t/n+,f,
(12.5.13)
w h ere th e a m p litu d e s {A;} a n d th e fre q u e n c ie s j / ;- } a re u n k n o w n a n d th e p h ases {< f> , } a re statistically in d e p e n d e n t ra n d o m variab les u n ifo rm ly d is trib u te d o n (0, 2 n ) . T h e n th e ra n d o m p ro cess x ( n ) is w id e-sen se s ta tio n a ry w ith a u to c o r re la tio n fu n c tion y ,,( m ) =
22
/=!
Piei2nf'm
(12.5.14)
w h ere, fo r co m p lex sin u so id s, P, = A ? is th e p o w e r of th e / th sin u so id . Since th e se q u e n c e o b se rv e d is y( n) = x ( m ) + w ( n ), w h e re w{n) is a w h ite n o ise se q u en ce w ith sp e c tra l d e n sity erj, th e a u to c o rre la tio n fu n c tio n fo r v (n) is Yyy(m) = yIX( m ) + a l S ( m ) m = 0, 1 , . . . , { M - 1) (12.5.15)
H e n c e th e M x M a u to c o rre la tio n m a trix fo r y ( n ) can be e x p re ss e d as r V v= r +

c t u 2i
(12.5.16)
w h ere r fJt is th e a u to c o rre la tio n m atrix fo r th e signal x ( n ) a n d a l l is th e a u to c o r re la tio n m atrix fo r th e n o ise. N o te th a t if select M > p , T IX w hich is o f d im en sio n M x M is n o t o f full ra n k , b e c a u se its ra n k is p. H o w e v e r, T vv is full ra n k b ecau se <7*1 is o f ra n k M. In fact, th e signal m a trix T** can be re p re s e n te d as
p
T xx =
22
1=1
(12.5.17)
w h ere H d e n o te s th e c o n ju g a te tra n sp o s e an d s, is a signal v e c to r o f d im e n sio n M d efin ed as s,- = f l , ei7* f \ e iA*{\ l!* ] (12.5.18)
Since ea c h v e c to r ( o u te r p ro d u '0 s ,s f is a m atrix o f ra n k 1 a n d sin ce th e re a re p v e c to r p ro d u c ts, th e m a trix T *, is o f ra n k p. N o te th a t if th e sin u so id s w ere real, th e c o rre la tio n m a trix T JJf h a s ra n k 2 p.
Sec. 12.5
951
N o w , le t u s p e rfo rm an eig en -d e c o m p o sitio n o f th e m atrix L e t th e eig e n v a lu e s {A/} b e o r d e re d in d e creasin g v a lu e w ith k \ > Aj > A ,3 > - > k M an d le t th e c o rre sp o n d in g e ig e n v e c to rs b e d e n o te d as {v,, i = 1 , . . . , M}. W e assu m e th a t th e e ig e n v e c to rs a re n o rm a liz e d so th a t vj 1 v, = . In th e a b se n c e o f n oise th e eig e n v a lu e s A ., , i = 1, 2 , . . . , p, a re n o n z e ro w hile k p+i = k p+2 = = = 0. F u rth e rm o re , it follow s th a t th e signal c o rre la tio n m a trix can b e e x p ressed as
p
r = /=i
(12.5.19)
T h u s, th e e ig e n v e c to rs v(, i = 1, 2 , . . . , p sp a n th e sig n a l su b sp ace as do th e sig n al v ecto rs s, , i = 1, 2 , . , . , p. T h e se p e ig en v ecto rs fo r th e signal su b sp ace a re called th e pr i nc i pal eigenvectors a n d th e c o rre sp o n d in g e ig e n v a lu e s a re called th e pri nc i pal eigenvalues. In th e p re se n c e o f n o ise , th e n o ise a u to c o rre la tio n m a trix in (12.5.16) can b e re p re s e n te d as M o]\ = al 2 2 W ? (12.5.20) /=i B y s u b s titu tin g (12.5.19) a n d (12.5.20) in to (12.5.16), w e o b ta in
r w = 2 2 k *iVil + 2 2 ^ y y ^ 1=1 1=1 (12.5.21) = 2 2 (k + 1=1 +

i=p+i
2 2 fT- v' v/ /
T h is e ig e n -d e c o m p o sitio n se p a ra te s th e eig e n v e c to rs in to tw o sets. T h e se t {v,, i = 1, 2 , . . . , p], w h ich a re th e p rin cip al eig en v ecto rs, span th e signal su b sp ace, w hile th e se t {v,, i = p + \ .........M] , w hich a re o rth o g o n a l to th e p rin c ip a l e ig en v ecto rs, a re said to b e lo n g to th e n o ise su b sp ace. S ince th e signal v e c to rs {s,, / = 1 ,2 , , p] a re in th e sig n al su b sp ace, it follow s th a t th e {s, } a re sim p ly lin e a r co m b in a tio n s o f th e p rin cip al e ig e n v e c to rs a n d a re also o rth o g o n a l to th e v e c to rs in th e n o ise su b sp ace. I n th is c o n te x t w e se e th a t th e P isa re n k o m e th o d is b a s e d o n an e s tim a tio n o f th e fre q u e n c ie s by u sin g th e o rth o g o n a lity p ro p e rty b e tw e e n th e signal v ecto rs a n d th e v ecto rs in th e n o ise su b sp ace. F o r com plex sin u so id s, if w e se lect M = p + 1 (fo r re a l sin u so id s w e se lect M = 2 p + 1), th e r e is only a single e ig e n v e c to r in th e n o ise su b sp a c e (c o rre s p o n d in g to th e m in im u m e ig e n v a lu e ) w hich m u st b e o rth o g o n a l to th e signal vectors. T h u s w e h av e p s f vp+i = 2 2 i p + i t t + l ) e ' j2nf'k = 0 i=0 i = 1,2,..., p (12.5.22)
But (12.5.22) implies that the frequencies {/;} can be determined by solving for
9 52
Chap. 12
th e z e ro s o f th e p o ly n o m ial
p
V ( z ) = J 2 vp + ^ k + 1^ ~ k n=0 all o f w hich lie o n th e u n it circle.
(12.5.23)
T h e an g les o f th e s e r o o ts a re 2 n f , , i = 1,
2....... PW h e n th e n u m b e r o f sin u so id s is u n k n o w n , th e d e te rm in a tio n o f p m ay p ro v e to b e d ifficult, esp ecially if th e signal level is n o t m u ch h ig h e r th a n th e noise level. In th e o ry , if M > p + 1, th e re is a m u ltip lic ity (M - p ) o f th e m in im u m eig en v alu e. H o w ev er, in p ra c tic e th e (Afp) sm all e ig en v alu es o f R vv will p ro b a b ly b e d ifferen t. B y co m p u tin g all th e eig en v alu es it m ay b e p o ssib le to d e te rm in e p by g ro u p in g th e M p sm all (n o ise) e ig en v alu es in to a se t a n d a v e rag in g th e m to o b ta in an e stim a te o f cr*. T h e n , th e a v erag e v alu e can b e u se d in (12.5.9) along w ith R yj, to d e te rm in e th e c o rre sp o n d in g eig en v ecto r.
12.5.3 MUSIC Algorithm

T h e m u ltip le sig n al classification (M U S IC ) m e th o d is also a n o ise su b sp ace fre q u en cy estim ato r. T o d e v e lo p th e m e th o d , le t us first c o n s id e r th e w e ig h te d sp e c tra l estim ate M P (f)= 2 2 w*ls " ( / ) v*i2 *=/>+i (12.5.24)
{u;*)
w h ere {v*, jt = p + 1 , ___ M } are th e e ig e n v e c to rs in th e n o ise su b sp ace, a se t o f p o sitiv e w eig h ts, a n d s ( / ) is th e co m p lex sin u so id al v e c to r s( / ) =
a re
(12.5.25)
N o te th a t at / = / ; , s ( / ,) = s,, so th a t a t an y o n e o f th e p sin u so id a l fre q u e n c y c o m p o n e n ts o f th e signal, w e h av e /> (/,) = 0 i = 1,2,...,/? (12.5.26)
H e n c e , th e recip ro c al o f P ( f ) is a sh a rp ly p e a k e d fu n c tio n o f fre q u e n c y a n d p r o v id es a m e th o d fo r e stim a tin g th e fre q u e n c ie s o f th e sin u so id al c o m p o n e n ts . T h u s 1 22 *=p + i 1 I ( / ) n | (12.5.27)
A lth o u g h th e o re tic a lly 1 / P ( f ) is infinite a t / = / , , in p ra c tic e th e e s tim a tio n e rro rs re su lt in fin ite v a lu e s fo r 1 / P ( f ) a t all fre q u e n c ie s. T h e M U S IC sin u so id al fre q u e n c y e s tim a to r p ro p o s e d by S ch m id t (1981, 1986) is a sp ecial case o f (12.5.27) in w h ich th e w eig h ts wk 1 f o r all k. H e n c e
^ m u s ic ( / ) = j j --------------------
(12.5.28)
22
k=p+1
|s H( / ) * |2
Sec. 12.5
Eigenanaiysis Algorithms for Spectrum Estimation
9 53
T h e e s tim a te o f th e sin u so id al fre q u e n c ie s a re th e p e a k s o f P m u sic C/)- O n c e th e sin u so id a l fre q u e n c ie s a re e s tim a te d , th e p o w e r o f ea c h o f th e sin u so id s can b e o b ta in e d by so lv in g (12.5.11).
12.5.4 ESPRIT Algorithm

E S P R IT (e s tim a tio n o f signal p a ra m e te rs via ro ta tio n a l in v a ria n c e te c h n iq u e s ) is y e t a n o th e r m e th o d fo r e stim a tin g fre q u e n c ie s o f a su m o f sin u so id s by use o f a n e ig e n -d e c o m p o sitio n a p p ro a c h . A s w e o b se rv e fro m th e d e v e lo p m e n t th a t follow s, w h ich is d u e to R o y e t al. (1986), E S P R IT ex p lo its an u n d e rly in g r o t a tio n a l in v a ria n c e o f sig nal su b sp a c e s sp a n n e d by tw o te m p o ra lly d isp lac ed d a ta v ecto rs. W e ag ain c o n sid e r th e e stim a tio n o f p c o m p lex -v alu ed sin u so id s in ad d itiv e w h ite n o ise . T h e re c e iv e d se q u e n c e is given by th e v e c to r y( n) = [?( ). y( n + 1 ) , . . . , y ( n + M - 1)]' (12.5.29)
= x(n) + w (n)
w h e re x (n ) is th e signal v e c to r a n d w (n) is th e n oise v ecto r. T o e x p lo it th e d e te rm in is tic c h a ra c te r o f th e sinusoids, w e d efin e th e tim e-d isp laced v e c to r z( n) = y (n + 1). T h u s z( n) = [z(n), z(n + 1).........z( n + M - 1)]' (12.5.30) = [;y(n + 1), y( n + 2 )......... y( n + Af)]' W ith th e s e d efin itio n s we can ex p ress th e v ecto rs y(n) a n d z(n ) as y (rt) = S a + w (n) (12.5.31) i (n ) = S $ a + w (n) w h e re a = [ a \ , a i , . . ap]r, cii = A , , a n d $ is a d ia g o n a l p x p m a trix c o n sist in g o f th e re la tiv e p h a s e b e tw e e n a d ja c e n t tim e sa m p le s o f each o f th e co m p lex sin u so id s, $ = d ia g [ e '2,r/l, ej2jrf2, . . . , e i 2nf> ] (12.5.32)
N o te th a t th e m a trix & re la te s th e tim e-d isp laced v e c to rs y( n) a n d z( n) a n d can b e called a r o ta tio n o p e ra to r. W e also n o te th a t < l> is u n ita ry . T h e m a trix S is th e Af x p V a n d e rm o n d e m a trix specified b y th e co lu m n v ecto rs s, = [1, e}2*fl , ejAnf , . . . , ] i = 1 , 2 ......... p (12.5.33)
N o w th e a u to c o v a ria n c e m a trix fo r th e d a ta v e c to r y (n) is r y, = E [y (n )y " (n )]
(12.5.34)
= S P S W+ a l l
954
Chap. 12
w h e re P is th e p x p d ia g o n a l m atrix co n sistin g o f th e p o w e rs o f th e co m p lex sin u so id s, P = d ia g [|a i |2, \a2\2, - . . , \ap |2] (12.5.35) = d ia g [ f i, P2.........Pp] W e o b se rv e th a t P is a d iag o n al m a trix since co m p lex sin u so id s o f d iffe re n t fre q u e n c ie s a re o rth o g o n a l o v e r th e in fin ite in terv al. H o w ev er, w e sh o u ld e m p h a size th a t th e E S P R IT a lg o rith m d o e s n o t re q u ire P to b e d ia g o n a l. H e n c e th e alg o rith m is a p p lic a b le to th e case in w hich th e c o v a ria n c e m a tr ix is e s tim a te d fro m fin ite d a ta reco rd s. T h e cro ssco v arian ce m a trix o f th e signal v e c to rs y(n) a n d z (tt) is T v. = [ y ( n ) z ff(n)] = S P $ " S " + T w w h ere I V = [ w (n ) w w(n + 1)] (12.5.36)
(12.5.37)
0
r yw (0) y ;v(D
0 0
yyy(l) Yyy(0) yvv(A/ - 1) "1 Yyyi M - 2 ) (12.5.38)
T h e a u to an d cro ssco v arian ce m a tric e s r vv a n d T v; a re given as
rv v=
r
- y ; y ( M - 1) K w (i) yyy(0)
Yyy(M - 2) Yyy (2) yyy(l)
Yyy( 0)
Yyv(M) -1 Yyy(M -l) (12.5.39)
- Yyy ( M 2)
Y 'yy( M - 3)
yyy(l)
w h ere Yyy(m) = [y * (n )y (n + m )]. N o te th a t b o th T yy a n d 1%, a re T o e p litz m atric es. B a se d o n th is fo rm u la tio n , th e p ro b le m is to d e te rm in e th e fre q u e n c ie s {ft) a n d th e ir p o w ers {/*,} fro m th e a u to c o rre la tio n se q u e n c e [yyy(m)}. F ro m th e u n d e rly in g m o d e l, it is c le a r th a t th e m a trix S P S W h a s ra n k p. C o n se q u e n tly , T vv given by (12.5.34) h a s (M - p) id en tical e ig e n v a lu e s e q u a l to a 2. H e n c e T yy - a l l = S P S W = C yy F ro m (12.5.36) w e also h av e - a 2T w = S P * HS H m C (12.5.41) (12.5.40)
N o w , le t us c o n s id e r th e m a trix C yy XCyz, w h ich can b e w ritte n as
Cyy - k C yi = SP(I - \ * H)S H
(12.5.42)
Sec. 12.5
955
C lea rly , th e c o lu m n sp a ce o f S P S W is id en tical to th e co lu m n space o f S P $ HS W. C o n s e q u e n tly , th e ra n k o f C vv - XCy: is e q u a l to p. H o w e v e r, w e n o te th a t if X = ex p (;'2 7 r/(), th e rth ro w o f (I is z e ro an d , h e n c e th e ra n k o f [I ex p O 2 ^ - /) ] is p 1. B u t A, = e x p { j 2 n f i ) , i = 1, 2 , . . . , p, a re th e g en eralized e ig en v alu es o f th e m a trix p a ir (C yy, C n ). T h u s th e p g e n e ra liz e d eig en v alu es {A, } t h a t lie o n th e u n it circle c o rre sp o n d to th e e le m e n ts o f th e r o ta tio n o p e ra to r 3>. T h e re m a in in g M p g e n e ra liz e d eig en v alu es o f th e p a ir {Cvv, C v;} w hich c o rre sp o n d to th e c o m m o n null sp a ce o f th e se m a tric e s, a re z e ro [i.e., th e ( M p) e ig e n v a lu e s a re a t th e o rig in in th e co m p lex p lan e]. B a sed o n th e s e m a th e m a tic a l re la tio n sh ip s w e can fo rm u la te an alg o rith m ( E S P R IT ) fo r e s tim a tin g th e fre q u e n c ie s {/]. T h e p ro c e d u re is as follow s: 1. F ro m th e d a ta , c o m p u te th e a u to c o rre la tio n valu es r yy(nt), m = 1 , 2 ........ M , a n d fo rm th e m a tric e s R vv a n d R v- c o rre sp o n d in g to e stim a te s o f r vv an d
rv :
2. C o m p u te th e e ig en v alu es o f R vv. F o r M > p, th e m in im u m eig en v alu e is an e stim a te o f a 2. 3. C o m p u te C vv = R vv a * I an d C v- = R vc <7*Q, w h e re Q is defin ed in (12.5.37). 4. C o m p u te th e g e n e ra liz e d eig en v alu es o f th e m a trix p a ir (C TT, C v- |. T he p g e n e ra liz e d e ig en v alu es o f th e s e m atric es th a t lie o n (o r n e a r) th e u n it circle d e te rm in e th e ( e stim a te ) e le m e n ts of $ a n d h e n c e th e sin u so id al freq u en cies. T h e re m a in in g M p eig e n v a lu e s will lie at ( o r n e a r) th e origin. O n e m e th o d fo r d e te rm in in g th e p o w e r in th e sin u so id al c o m p o n e n ts is to solve th e e q u a tio n in (12.5.11) w ith ryy{m) s u b s titu te d fo r y yy(m). A n o th e r m e th o d is b a se d on th e c o m p u ta tio n o f th e g e n e ra liz e d eig en v ecto rs {v,} c o rre sp o n d in g to th e g e n eralized e ig en v alu es {A, }. W e h av e (C vv - A ., c v:)v, = S P (I - X/ = 0 (12.5.43)
S in ce th e co lu m n sp a ce o f ( Cyyk ( - Cyz) is id en tical to th e co lu m n space sp a n n e d by th e v e c to rs {S;, j ^ /} g iven by (12.5.33), it follow s th a t th e g e n e ra liz e d eig e n v e c to r v, is o rth o g o n a l to sj5 j ^ i. S ince P is d iag o n al, it follow s fro m (12.5.43) th a t th e sig n al p o w e rs a re
yh C v Pi = ~ n r i r
|V " S ,|2
i =
( 12 .5 .44)
12.5.5 Order Selection Criteria

T h e eig e n a n a ly sis m e th o d s d e sc rib e d in th is se ctio n fo r e stim a tin g th e freq u e n cies a n d th e p o w ers o f th e sin u so id s, also p ro v id e in fo rm a tio n a b o u t th e n u m b e r o f sin u so id al c o m p o n e n ts . If th e r e a re p sin u so id s, th e e ig e n v a lu e s a s so ciated w ith th e
956
Chap. 12
signal su b sp ace a re [kj+cr*, i 1 , 2 . . . , / >} w hile th e re m a in in g {M p ) eig en v alu es are all eq u al to a 2 . B a sed o n th is e ig e n v a lu e d e c o m p o sitio n , a te s t can b e d esigned th a t c o m p a re s th e eig en v alu es w ith a specified th re sh o ld . A n a lte rn a tiv e m e th o d also u ses th e e ig e n v e c to r d e c o m p o s itio n o f th e e s tim a te d a u to c o rre la tio n m atrix of th e o b se rv e d sig n al a n d is b a se d o n m a trix p e rtu rb a tio n analysis. T h is m e th o d is d e c rib e d in th e p a p e r by F u c h s (1988). A n o th e r a p p ro a c h b a se d o n an e x te n sio n a n d m o d ifica tio n o f th e A I C crite rio n to th e e ig e n -d e c o m p o sitio n m e th o d , h a s b e e n p ro p o se d by W ax an d K ailath (1985). If th e eig en v alu es o f th e sa m p le a u to c o rre la tio n m a trix a r e ra n k e d so th a t î > k 2 > > k M , w h e re M > p, th e n u m b e r o f sin u so id s in th e signal su b sp ace is e s tim a te d b y se lectin g th e m in im u m v alu e o f M D L (p ), given as M D L (p ) = log w h ere
g (p
G(p) M
p)
+ E(p) J
(12.5.45)
)=
p = 0 , 1 .........M - l
M p
A(p) =
1 M / />+] (12.5.46)
1 E ( P ) = 2 p(-2 M ~ P')loZ N N: n u m b e r o f sa m p le s u sed to e s tim a te th e M a u to c o rre la tio n lags S o m e resu lts o n th e q u ality o f th is o r d e r se le c tio n c rite rio n a re given in th e p a p e r b y W ax an d K ailath (1985). T h e M D L c rite rio n is g u a ra n te e d to b e co n sisten t.
12.5.6 Experimental Results

In th is sectio n w e illu stra te w ith an ex a m p le , th e re so lu tio n c h a ra c te ristic s o f th e eig en an aly sis-b ase d sp e c tra l e s tim a tio n a lg o rith m s a n d c o m p a re th e ir p e rfo rm a n c e w ith th e m o d el-b ased m e th o d s an d n o n p a ra m e tric m e th o d s. T h e signal se q u e n c e is
4
x ( n ) = 2 2 A iej a n f 'n+4,') + w( n) i=l w h e re A, = 1, i = 1, 2, 3, 4, {tp,} a re sta tistically in d e p e n d e n t ra n d o m v ariab les u n ifo rm ly d istrib u te d on (0, 2j t ), {w(rt)) is a z e ro -m e a n , w h ite n o ise se q u e n c e w ith v a ria n c e a 2, an d th e fre q u e n c ie s a re f \ = 0.222, f a 0.166, f s = 0.10, an d f i = 0.122. T h e se q u e n c e {jt(n), 0 < n < 1023} is u se d to e s tim a te th e n u m b e r o f fre q u e n c y c o m p o n e n ts a n d th e c o rre sp o n d in g v alu es o f th e ir fre q u e n c ie s for a 2 = 0 . 1 , 0.5, 1.0, a n d M 12 (le n g th o f th e e s tim a te d a u to c o rre la tio n ).
Sec. 12.5
957
Frequency Figure 12.19 Power spectrum estimates from Blackman-Tukey method.
F ig u re s 12.19, 12.20, 12.21, a n d 12.22 illu stra te th e e s tim a te d p o w e r sp e c tra o f th e signal u sin g th e B la c k m a n -T u k e y m e th o d , th e m in im u m v a ria n c e m e th o d o f C a p o n , th e A R Y u le -W a lk e r m e th o d , a n d th e M U S IC a lg o rith m , resp ec tiv ely . T h e re su lts fro m th e E S P R IT a lg o rith m a re given in T a b le 12.2. F ro m th e s e resu lts it is a p p a r e n t th a t (1) th e B la c k m a n -T u k e y m e th o d d o e s n o t p ro v id e sufficient
Figure H 2 0
Power spectrum estimates from minimum variance method.
958
Chap. 12
Figure 1X21
Power spectrum estimates from Yule-Walker A R method.
re so lu tio n to e s tim a te th e sin u so id s fro m th e d a ta ; (2) th e m in im u m v arian ce m e th o d o f C a p o n reso lv es o n ly th e fre q u e n c ie s f \ , f 2 b u t n o t / 3 a n d / 4; (3) th e A R m e th o d s reso lv e all fre q u e n c ie s fo r cr2 = 0.1 a n d cr2 = 0.5; a n d (4) th e M U S IC a n d E S P R IT a lg o rith m s n o t o n ly re c o v e r all fo u r sin u so id s, b u t th e ir p e rfo rm a n c e fo r d iffe re n t v a lu e s o f cr2 is esse n tia lly in d istin g u ish a b le . W e f u r th e r o b se rv e th a t
Frequency
Figure 12.22 Power spectrum estimates from MUSIC algorithm.
Sec. 12.6

TABLE 12.2 ESPRIT ALGORITHM
959
h
0.1 0.5 1.0 True values -0.2227 -0.2219 -0.222 -0.222
h
-0,1668 -0.167 -0.167 -0.166
h
-0.1224 -0.121 0.1199 0.122
Ia
0.10071 0.0988 0.1013 0.100
th e re so lu tio n p r o p e rtie s o f th e m in im u m v a ria n c e m e th o d a n d th e A R m e th o d is a fu n c tio n o f th e n o ise v a rian ce. T h e s e re su lts clearly d e m o n s tr a te th e p o w e r o f th e e ig e n a n a ly sis-b a se d a lg o rith m s in reso lv in g sin u so id s in a d d itiv e n oise. In co n clu sio n , w e sh o u ld em p h asize th a t th e h ig h -re s o lu tio n , eig en an aly sisb a s e d sp e c tra l e s tim a tio n m e th o d s d e sc rib e d in th is se ctio n , n a m ely M U S IC an d E S P R IT , a re n o t o n ly ap p licab le to sin u so id al signals, b u t ap p ly m o re g en erally to th e e stim a tio n o f n a rro w b a n d signals.

P o w e r sp e c tru m e s tim a tio n is o n e o f th e m o st im p o rta n t a r e a s o f re se a rc h a n d a p p lic a tio n s in d ig ita l sig n a l p ro cessin g . In th is c h a p te r w e h a v e d e sc rib e d th e m o st im p o rta n t p o w e r sp e c tru m e s tim a tio n te c h n iq u e s a n d a lg o rith m s th a t h av e b e e n d e v e lo p e d o v e r th e p a s t cen tu ry , b e g in n in g w ith th e n o n p a ra m e tric o r classical m e th o d s b a s e d o n th e p e rio d o g ra m a n d c o n clu d in g w ith th e m o re m o d e rn p a r a m e tric m e th o d s b a s e d o n A R , M A , a n d A R M A lin e a r m o d e ls. O u r tr e a tm e n t is lim ite d in sc o p e to sin g le-tim e-series sp e c tru m e s tim a tio n m e th o d s, b a se d on se co n d m o m e n ts (a u to c o rre la tio n ) o f th e sta tistical d a ta . T h e p a ra m e tric a n d n o n p a ra m e tric m e th o d s th a t w e d e s c rib e d h a v e b e e n e x te n d e d to m u ltic h a n n e l a n d m u ltid im e n sio n al sp e c tru m es tim a tio n . T h e tu to ria l p a p e r by M cC lellan (1982) tr e a ts th e m u ltid im e n sio n a l sp e c tru m e s tim a tio n p r o b lem , w h ile th e p a p e r by Jo h n s o n (1982) tr e a ts th e m u ltic h a n n e l s p e c tru m e s tim a tio n p ro b le m . A d d itio n a l sp e c tru m e s tim a tio n m e th o d s h a v e b e e n d e v e lo p e d fo r u se w ith h ig h e r-o rd e r c u m u la n ts th a t involve th e b isp e c tru m a n d th e trisp e c tru m . A tu to ria l p a p e r o n th e s e to p ics h a s b e e n p u b lish e d by N ik ias a n d R a g h u v e e r (1987). A s e v id e n c e d fro m o u r p rev io u s d iscussion, p o w e r sp e c tru m e s tim a tio n is an a r e a th a t h a s a ttr a c te d m a n y re s e a rc h e rs a n d , a s a re su lt, th o u s a n d s o f p a p e rs h av e b e e n p u b lish e d in th e te c h n ic a l lite ra tu re on th is su b je ct. M u c h o f th is w o rk h as b e e n c o n c e rn e d w ith n ew a lg o rith m s a n d te c h n iq u e s, a n d m o d ifica tio n s o f ex ist in g te c h n iq u e s . O th e r w o rk h as b e e n c o n c e rn e d w ith o b ta in in g a n u n d e rsta n d in g o f th e c a p a b ilitie s a n d lim ita tio n s o f th e v ario u s p o w e r s p e c tru m m e th o d s. In th is c o n te x t th e sta tistical p ro p e rtie s a n d lim ita tio n s o f th e classical n o n p a ra m e tric m e th o d s h av e b e e n th o ro u g h ly an aly zed a n d a re w ell u n d e rs to o d . T h e p a ra m e tric m e th o d s h a v e also b e e n in v e stig a te d b y m an y re se a rc h e rs , b u t th e an aly sis o f th e ir
960
Chap. 12
perform ance is difficult and, con sequ en tly, few er results are availab le. S om e o f the papers that have addressed the problem o f p erform ance characteristics o f param et ric m eth od s are th o se of K rom er (1969), L acoss (1971), B erk (1974), B aggeroer (1976), Sakai (1979), Sw ingler (1980), L ang and M cC lellan (1980), and Tufts and K um aresan (1982). In addition to the referen ces already given in this chapter on the various m eth od s for spectrum estim ation and their perform ance, w e sh ou ld include for referen ce som e o f th e tutorial and survey papers. In particular, w e cite the tutorial paper by K ay and M arple (1981), w hich in clu d es ab ou t 280 referen ces, the paper b y B rillinger (1974), and the Special Issue on Spectral E stim ation o f the I E E E P ro ceed in g s, Sep tem b er 1982. A n oth er indication o f the w id esp read interest in the subject o f spectrum estim ation and analysis is the recent p u b lication o f texts by G ardner (1987), K ay (1988), and M arple (1987), and the IE E E b ook s edited by Childers (1978) and K esler (1986). M any com puter program s as w ell as softw are packages that im plem ent the various spectrum estim ation m eth od s described in this chapter are available. O ne softw are package is available through the IE E E (P ro g ra m s f o r D ig ita l S ig n a l P ro cessing, IE E E Press, 1979); others are available com m ercially.
PROBLEMS 12.1 (a) By expanding (12.1.23), taking the expected value, and finally taking the limit as
To -+ oo, show that the right-hand side converges to
(b) Prove that
12 2 For zero mean, jointly Gaussian random variables, [see Papoulis (1984)] that
X 2, X 3, X*, it is well known
E W x X iX 3X t ) = (X iX 2 ) (X 3 X4) + (X 1 X3) (X 2 X4) + E (X iX t )E (X 2X 3) Use this result to derive the mean-square value of r^(m), given by (12.1.27) and the variance, which is v a r ^ fm )] = E[\r'xz(m)\2] - \E[r'xx(m)]{
1 2 3 By use of the expression for the fourth joint moment for Gaussian random variables,
show that (a) E [P (fx)P * A h )] = a* 1 +

+
sin jr(/i + f i) N
[Afsin7r(/i + / 2) 1
Chap. 12
Problems
961
(b) co v [/> (/,)/> (/2)] = +
sin 7T(fi + f 2)N N sin jt(/i + / 2) sin ^ (/i - f 2)N 1

A? s i n n - ( / i
h )\
(c) var[/> (/)] = crj
j+
|
/ s i n 2 j r / W \ 2] under the condition that the sequence x(n) y N sin 2jt/ 1 ) }
is a zero-mean white Gaussian noise sequence with variance a],

1 2 .4 Generalize the results in Problem 12.3 to a zero-mean Gaussian noise process with
power density spectrum r ( / ) . Then derive the variance of the periodogram Pxr( f ), as given by (12.1.38). {Hint: Assume that the colored Gaussian noise process is the output of a linear system excited by white Gaussian noise. Then use the appropriate relations given in Appendix A.) 12.5 Show that the periodogram values at frequencies /* = k /L , k = 0, 1 , . . . , L - 1, given by (12.1.41) can be computed by passing the sequence through a bank of N IIR filters, where each filter has an impulse response M ) = e -j2*nk,f/u(n) and then compute the magnitude-squared value of the filter outputs at n = N. Note that each filter has a pole on the unit circle at the frequency f t . 12.6 Prove that the normalization factor given by (12.2.12) ensures that (12.2.19) is satisfied. 12.7 Let us consider the use of the DFT (computed via the FFT algorithm) to compute the autocorrelation of the complex-valued sequence x (n ), that is,
1 N -m -l
i'u(m ) = N
x*(n)x(n + m), m > 0

/I sbO
Suppose the size M of the FFT is much smaller than that of the data length N. Specifically, assume that N = K M . (a) Determine the steps needed to section x(n) and compute rxx(m) for (M /2) + 1 < m < (M /2) - 1, by using 4 K A/-point DFTs and one M-point IDFT (b) Now consider the following three sequences x\(n), x2(n), and x^(n), each of du ration M. Let the sequences x\(n) and x2(n) have arbitrary values in the range 0 < n < (M/2) - 1, but are zero for (M /2) < n < M 1. The sequence xi(n) is defined as xi(n). x i(n) = xi
M
"2
Determine a simple relationship among the M-point DFTs X i(k), X 2(k), and
* 3 (J t).
H>
< rt < M - 1
(c) By using the result in part (b), show how the computation of the DFTs in part (a) can be reduced in number from AK to 2 K. 12J5 The Bartlett method is used to estimate the power spectrum of a signal x(n). We know that the power spectrum consists of a single peak with a 3-dB bandwidth of 0 .0 1 cycle per sample, but we do not know the location of the peak.
962
Chap. 12
(a) Assuming that N is large, determine the value of M = N / K so that the spectral
window is narrower than the peak.
(b) Explain why it is not advantageous to increase M beyond the value obtained in
pan (a). 12.9 Suppose we have N = 1000 samples from a sample sequence of a random process. (a) Determine the frequency resolution of the Bartlett, Welch (50% overlap), and Blackman-Tukey methods for a quality factor Q = 10. (b) Determine the record lengths (Af) for the Bartlett, Welch (50% overlap), and Blackman-Tukey methods. 12.10 Consider the problem of continuously estimating the power spectrum from a sequence x (n) based on averaging periodograms with exponential weighting into the past. Thus with P \ f ) = 0, we have
1 l M _1
C (/) =
l Y ' xm(n)e-
where successive periodograms are assumed to be uncorrelated and w is the (expo nential) weighting factor. (a) Determine the mean and variance of for a Gaussian random process. (b) Repeat the analysis of part (a) for the case in which the modified periodogram defined by Welch is used in the averaging with no overlap. 12.11 The periodogram in the Bartlett method can be expressed as
M -l
-J2n/n m=-<Af-1 ) V /
where r J ( *(m) is the estimated autocorrelation sequence obtained from the j'th block of data. Show that P ^ ( f ) can be expressed as
/ *> (/ ) = E ' ! ( f ) R ^ E ( f )
where (/) = [l and therefore, eila} ]'
12.12 Derive the recursive order-update equation given in (12.3.19). 12.13 Determine the mean and the autocorrelation of the sequence jcfn), which is the output of a ARMA (1,1) process described by the difference equation x(n) = (n 1) + w(n) - w(n - 1) (n) generated by the
where u>(n) is a white noise process with variance crj. 12.14 Determine the mean and the autocorrelation of the sequence M A(2) process described by the difference equation x(n) = w(n) - 2 w(n - 1) + w(n 2) where w(n) is a white noise process with variance a*.
jc
Chap. 12
Problems
963
12.15 An M A(2) process has the autocorrelation sequence

6cr2 4ct2 ,
Yxx (m ) = 2 a 2,
m = 0 m= 1 m = 2 otherwise
0, correlation.
(a) Determine the coefficients of the MA(2) process that have the foregoing auto (b) Is the solution unique? If not, give all the possible solutions. 12.16 An MA(2) process has the autocorrelation sequence
al 35 , - cr2. 62 m =0 m = 1
y(m) =
m = 2 7*al 62 (a) Determine the coefficients of the minimum-phase system for the MA(2) process. (b) Determine the coefficients of the maximum-phase system for the MA(2) process, (c) Determine the coefficients of the mixed-phase system for the MA(2) process. 12.17 Consider the linear system described by the difference equation v(n) = 0.8v(n 1) + x(n) + x(n 1) where x(n) is a wide-sense stationary random process with zero mean and autocorre lation * ,( > = G ) 1" 1
(a) Determine the power density spectrum of the output y(n). (b) Determine the autocorrelation yyy(m) of the output. (c) Determine the variance o\2 of the output. 12.18 From (12.3.6) and (12.3.9) we note that an A R (p) stationary random process satisfies
the equation p
Yxx(m) +
=
f 2
{ .q
= 0,
1 < m < p.
where ap(k) are the prediction coefficients of the iinear predictor of order p and a 2 is the minimum mean-square prediction error. If the (p + 1) x (p + 1) autocorrelation matrix Txx in (12.3.9) is positive definite, prove that: (a) The reflection coefficients |K| < 1 for 1 < m < p. (b) The polynomial p A p(z) = 1 + y ^ a p(k)z~k *i has all its roots inside the unit circle (i.e., it is minimum phase).
12.19 Consider the A R(3) process generated by the equation

x(n) = %x(n - 1) + %x(n - 2) - %x(n - 3) + w(n) where w(n) is a stationary white noise process with variance a*.
964
Chap. 12
(b)
(a ) D eterm in e th e co efficien ts o f the optim um p = 3 lin ear p red ictor. D eterm in e th e au toco rrelatio n seq u en ce y ( m ) , 0 < m < 5. (c ) D eterm in e th e reflection co efficien ts co rresp o nd ing to th e p = 3 lin ear predictor.
1 2 2 0 * A n A R (2 ) p rocess is d escrib ed by th e d ifferen ce eq u atio n
x (n) = 0.81jc (n 2) + w(n)

w here w(n) is a white n oise p rocess with variance er2 . (a ) D eterm in e th e p aram eters o f th e M A (2 ), M A (4 ), and M A (8 ) m od els which p ro vide a m inim um m ean -squ are e rro r fit to th e data x (n ). (b ) P lo t th e true spectrum and th ose o f the MA(<?), q = 2, 4, 8 , sp ec tra and com pare the results. C o m m en t on how well the M A (g ) m od els ap p ro x im a te th e A R ( 2 ) process. 1 2.21 A n M A (2 ) p rocess is described by th e d ifferen ce equation
x () = w(n) + 0.81u>(n 2)
w here w(n) is a w hite n oise p rocess with v arian ce a 2 . (a ) D ete rm in e th e p aram eters o f th e A R ( 2 ) , A R ( 4 ) , and A R ( 8 ) m od els that provide a m inim um m ean -squ are erro r fit to the data x(n). (b ) P lo t th e true spectra m and th ose o f th e A R ( p ) , p = 2. 4 , 8. and com pare the results. C o m m en t on how w ell the A R ( p ) m odels a p p roxim ate the M A (2 ) process.
1 2 2 2 T h e z-transform o f th e a u to co rrelatio n yT I(m )of an A R M A ( 1 ,1) process is TIX(z) = a l H ( z ) H ( z ~ ' )
(a) Determine the minimum-phase system function H(z).

(b ) Determine the system function H (z ) for a mixed-phase stable system.
12.23 Consider a FIR filter with coefficient vector

[1 2 r COS 6
t2 ]
(a) Determine the reflection coefficients for the corresponding FIR lattice filter.
(b ) Determine the values of the reflection coefficients in the limit as r
1.
1 2 2 4 An AR(3) process is characterized by the prediction coefficients
o3 (l) = 1.25.
(b ) Determine yxx(m) for 0 < m < 3.
a3(2) = 1.25,
a 3 (3) = - 1
(a) Determine the reflection coefficients. (c) Determine the mean-square prediction error.
1 2 2 5 The autocorrelation sequence for a random process is
1, 0.5, 0.625, 0.6875, 0
m = 0 m = 1 m = 2 m = 3 otherwise
Chap. 12
Problems
965
D ete rm in e th e system fun ction s A (z) for th e p red ictio n -erro r filters for m = 1, 2, 3, th e reflectio n coefficien ts { K m}, and the correspond ing m ean -squ are p red iction errors. 1 2 .2 6 D e te rm in e th e p o w er spectra fo r th e ran d om p rocesses g en erated by the follow ing d iffe re n ce eq uations. (1 ) x (n ) = 0.81jr(n 2) + ui(n) w(n 1) (2 ) x(n) = u >{n) w(n 2) (3 ) x (n ) = 0 .8 1 * (n 2) + ui(n) w here w{n) is a w hite noise process with v arian ce cr2 . (b) S k etc h th e sp ectra fo r th e p rocesses given in part (a ). (c) D ete rm in e th e au to co rrelatio n yX I(m ) fo r th e p rocesses in (2 ) and (3 ).
(a)
1 2 .2 7 T h e au to co rrelatio n seq u en ce fo r an A R process x(n ) is
(a) (b)
D ete rm in e th e d iffe re n ce eq u atio n for i ( n ) . Is y ou r answ er uniqu e? I f n o t, give any o th e r possible solutions.
1 2 .2 8 R e p e a t P ro b lem 12.27 fo r an A R p rocess with a u to co rrelatio n
Vi:r(ffi) = a' ' cos

w here 0 < a < 1.
nm
1 2 2 9 T h e B a r tle tt m ethod is used to estim ate th e pow er spectru m o f a signal from a s e q u en ce x (n) consisting o f N = 2 4 0 0 sam ples. (a) D ete rm in e the sm allest length M o f each segm ent in th e B a rtle tt m ethod th at yields a frequ en cy resolu tion o f A/ = 0.01. (b) R e p e a t part (a ) fo r A / = 0.02. (c) D ete rm in e th e qu ality factors Q b fo r parts (a ) and (b ). 1 2 3 0 P rov e th at a F I R filter with system function
p
A(Z) = *>1 and reflectio n co efficien ts | A T* | < 1 fo r 1 < k 1 is m axim um phase [all th e ro o ts o f A p ( z ) lie outside th e unit circle]. 1 2 3 1 A random p rocess x (n ) is ch aracterized by th e p o w er density spectrum
U ) ~
\e W -Q .9 ? _ JQ 9 |2|cjr2T/ + jQ 9l2
w h ere cr2 is a co n stan t (scale facto r). (a) I f w e view r ( / ) as th e pow er spectrum a t th e o utpu t o f a lin ear p o le -z e ro system H ( z ) driven by white n oise, d eterm in e H (z). (b) D ete rm in e th e system function o f a stab le system (noise-w hiten in g filter) that p rod u ces a w hite n oise output w hen ex cited by x(n). 1 2 3 2 T h e W -point D F T o f a random seq u en ce x(n) is
N -1
X (k ) = y '^ x ( n ) e - j2*nk/"
H kO
A ssu m e th at E [x {n )] = 0 and E [x (n )x (n + m ) ] = cr25(m ) [i.e., x(n) is a w hite noise process].
966
Chap. 12
(a) Determine the variance of X (k). (b) Determine the autocorrelation of X (k).
1233 Suppose that we represent an ARM A(p, q) process as a cascade of a MA(<?) followed by an AR(p) model. The input-output equation for the MA(<y) model is v(n) = 2 2 b t w(n - k) *=o where w(n) is a white noise process. The input-output equation for the AR(p) model is
x(n) + 22a/cX(n - k) = v(n ) *=i
(a) By computing the autocorrelation of v(n), show that

q m
Yw(m) = o l ' y ' b l b i+m
(b) Show that

p
J t=0
a = 1
Yw(m) = 2 2 aky vx(m + k)

*s=0
where yM(m) = [u{n + m)jr*(n)]: 1234 Determine the autocorrelation yxx(m) of the random sequence x(n) = A cos(ti>i + < p) where the amplitude A and the frequency w\ are (known) constants and < j> is a uni formly distributed random phase over the interval (0 , 2 t t ) . 1235 Suppose that the AR(2) process in Problem 12.20 is corrupted by an additive white noise process v(n) with variance crjf. Thus we have y(n ) = x(n) + v(n)
(a) Determine the difference equation for y(n) and thus demonstrate that y(n) is an
ARMA(2, 2) process. Determine the coefficients of the A R M A process.
(b) Generalize the result in part (a) to an AR(p) process

p
x(n) = - ^ 2 ak(xn ~ k) + w(n) t-i and >(ti) = x(n) + v(n) 1236 (a) Determine the autocorrelation of the random sequence K X (n ) = 2 2 A t COS(< ot n + < /> k) + ttf(/j) *l where (A*} are constant amplitudes, [cot] are constant frequencies, and {4>t) are mutually statistically independent and uniformly distributed random phases. The noise sequence w(n) is white with variance (b) Determine the power density spectrum of x(n).
Chap. 12
Problems
967
1237 The harmonic decomposition problem considered by Pisarenko can be expressed as the solution to the equation a"'rvva = tr2 a"a The solution for a can be obtained by minimizing the quadratic form a*'rvva subject to the constraint that a*'a = 1. The constraint can be incorporated into the performance index by means of a Lagrange multiplier. Thus the performance index becomes = a * T vva 4- A(1 a*'a) By minimizing with respect to a, show that this formulation is equivalent to the Pisarenko eigenvalue problem given in (12.5.9) with the Lagrange multiplier play ing the role of the eigenvalue. Thus show that the minimum of is the minimum eigenvalue a 2. 12.38 The autocorrelation of a sequence consisting of a sinusoid with random phase in noise is y(m ) = P cos2jr/im where f\ is the frequency of the sinusoidal, P is its power, and is the variance of the noise. Suppose that we attempt to fit an AR(2) model to the data. (a) Determine the optimum coefficients of the AR(2) model as a function of a* and fi. (b) Determine the reflection coefficients K\ and K 2 corresponding to the AR(2) model parameters. (c) Determine the limiting values of the AR(2) parameters and (K\, K 2) as <xu 2 -* 0 . 1239 This problem involves the use of crosscorrelation to detect a signal in noise and esti mate the time delay in the signal. A signal x (n) consists of a pulsed sinusoid corrupted by a stationary zero-mean white noise sequence. That is, x(n) = y(n n o) + ui(n) where w(n) is the noise with variance y(n) = A cos twon, = 0, 0 < n < N 1
and the signal is 0 < n < M 1 otherwise
The frequency a > o is known but the delay h0, which is a positive integer, is unknown, and is to be determined by crosscorrelating x(n) with _y(n). Assume that N > M + n0. Let N- 1
ri y (.m) = Y n* 0
y (n
- m )x (n )
denote the crosscorrelation sequence between x(n) and y(n). In the absence of noise this function exhibits a peak at delay m = n0. Thus n0 is determined with no error. The presence of noise can lead to errors in determining the unknown delay. (a) For m = no, determine E[riy(n(,)]. Also, determine the variance, var[riV(rtB )], due to the presence of the noise. In both calculations, assume that the double frequency term averages to zero. That is, M 3> 2n/wo. (b) Determine the signal-to-noise ratio, defined as SNR = l % < ^ var[r^(n0)j
968
Chap. 12
(c) What is the effect of the pulse duration M on the SNR?

1 2.40* Generate 100 samples of a zero-mean white noise sequence w(n) with variance a j =
j 2, (a) (b) (c)
by using a uniform random number generator. Compute the autocorrelation of u>(n) for 0 < m < 15. Compute the periodogram estimate Pz l ( f ) and plot it. Generate 10 different realizations of w(n) and compute the corresponding sample autocorrelation sequences n (m ), 1 < k < 10 and 0 < m < 15. (d) Compute and plot the average periodogram for part (c): r>v(m) = -L 2 * , r*(m) (e) Comment on the results in parts (a) through (d).
12.41* A random signal is generated by passing zero-mean white Gaussian noise with unit variance through a filter with system function
(1 + az~l + 0 .9 9 i- 2 ) ( l - a z - 1 + 0 . 9 8 ; - ; )
(a) Sketch a typical plot of the theoretical power spectrum r { / ) for a small value of the parameter a (i.e., 0 < a < 0.1). Pay careful attention to the value of the two spectral peaks and the value of Pzx(a>) for co = n/2. (b) Let a = 0.1. Determine the section length M required to resolve the spectral peaks of T ( / ) when using Bartletts method. (c) Consider the Blackman-Tukey method of smoothing the periodogram. How many lags of the correlation estimate must be used to obtain resolution compa rable to that of the Bartlett estimate considered in part (b)? How many data must be used if the variance of the estimate is to be comparable to that of a four-section Bartlett estimate? (d) For a 0.05, fit an AR(4) model to 100 samples of the data based on the Yule-Walker method and plot the power spectrum. Avoid transient effects by discarding the first 200 samples of the data. (e) Repeat part (d) with the Burg method. (f) Repeat parts (d) and (e) for 50 data samples and comment on similarities and differences in the results.
Appendix Random Signals, Correlation Functions, and Power Spectra
In this appendix w e provide a brief review o f the characterization o f random signals in term s o f statistical averages expressed in both the tim e dom ain and the frequency dom ain. T he reader is assum ed to have a background in probability theory and random p rocesses, at the level given in the b ook s o f H elstrom (1990) and P eeb les (1987).
Random Processes
M any physical p h en om en a en cou n tered in nature are best characterized in statis tical term s. For exam p le, m eteorological p h en om en a such as air tem perature and air pressure fluctuate random ly as a function o f tim e. Therm al n oise voltages gen erated in the resistors o f electron ic d evices, such as a radio or television receiver, are a lso random ly fluctuating p h en om en a. T h ese are just a few exam p les o f ran dom signals. Such signals are usually m odeled as infinite-duration infinite-energy signals. Suppose that w e take the set o f w aveform s corresponding to the air tem p er ature in differen t cities around the w orld. For each city there is a corresponding w aveform that is a function o f tim e, as illustrated in Fig. A .I . T h e set o f all p ossible w aveform s is called an en sem b le o f tim e functions or, equivalently, a ra n d o m p r o cess. T h e w aveform for the tem perature in any particular city is a single realization or a sa m p le fu n c tio n o f the random process. Sim ilarly, th e therm al n oise voltage gen erated in a resistor is a single real ization or a sam p le function o f the random p rocess con sistin g o f all n oise voltage w aveform s gen erated by the set o f all resistors. T h e set (en sem b le) o f all p ossib le noise w aveform s o f a random process is d en o te d as X (r, S ), w h ere t rep resents the tim e index and S represents the set (sam p le sp a ce) o f all p ossib le sam ple functions. A single w aveform in the en sem b le is d en o te d by x ( t , s). U sually, w e drop the variable s (or S ) for n otational co n v en ien ce, so that the random p rocess is d en oted as X (r) and a single realization is d en o ted as x ( t) .
A1
A2
Random Signals, Correlation Functions, and Power Spectra
App. A
Figure A.1
H aving defined a random process X ( t ) as an en sem b le o f sam p le functions, let us consider the v alu es o f the p rocess for any set o f tim e instants ti > t2 > > t, w here n is any p ositive integer. In general, the sam p les X,. = x(r, ), i = 1, 2 , . . . , n are n random variables characterized statistically by their joint probability density function (P D F ) d en o ted as p ( x tl, x,2, . . . , x , j for any n.
Stationary Random Processes

Suppose that w e have n sam ples o f th e random p rocess X ( t ) at t = i = 1,
2 , . . . , n, and another set o f n sam p les disp laced in tim e from th e first set by an
am ount z . Thus the secon d se t o f sam p les are X ,i+T = X (tj + z ) , i = 1, 2 , . . . , n , as
App. A
A3
show n in Fig. A .I . T h is secon d set of n random variables is characterized by the joint p robability density function p (x ,i+T, . . . , x lii+T). T he jo in t P D F s o f the tw o sets o f random variables m ay or m ay not be identical. W hen th ey are id en tical, then P (x h , x l2........ x ,J = p ( x lt+T, x,1+x, . . . . x,+z) ( A .l)
for all r and all n, then the random process is said to b e stationary in the strict sense. In other words, the statistical properties o f a stationary random p rocess are invariant to a translation o f the tim e axis. O n the other hand, w h en th e joint P D F s are different, the random process is nonstationary.
Statistical (Ensemble) Averages

Let us con sid er a random process X ( t ) sam pled at tim e instant t t,. Thus X (ti) is a random variable with P D F p { x ti). T he / th m o m e n t o f the random variable is d efined as the exp ected value o f X l (tj), that is, (A .2) In general, the value o f the /th m om en t dep en d s on the tim e instant tit if the P D F o f X,, dep en d s on t,. W h en the process is stationary, h ow ever, p { x u+T) p { x r.) for all x. H en ce th e P D F is in d ep en dent o f tim e and, con seq u en tly, the /th m om en t is in d ep en dent o f tim e (a constant). N ext, let us con sid er the two random variables X ,t = X (?,), i = 1 , 2, corre sponding to sam p les o f X (t) taken at t = t] and t = 12. T h e statistical (en sem b le) correlation b etw een X t] and X ,2 is m easured by the joint m om en t (A .3) Since the joint m om en t d epends on the tim e instants ti and ti, it is d en oted as
Y x x i h , h ) and is called the autocorrelation fu n c tio n o f the random process. W hen
the process X (/) is stationary, the joint P D F o f the pair (X tj, X ,2) is identical to the joint P D F o f the pair (X fl+r, X ,1+T) for any arbitrary r. T his im plies that the autocorrelation function o f X ( t ) dep en d s on the tim e d ifferen ce t \ ti = x. H en ce for a stationary real-valued random process the au tocorrelation function is yxx( r) = E [ X tl+TX h ] O n the other hand, y ( - r ) = ( * - , * ) = ( * , ; * , ;+r) = ^ ( r ) (A .5) (A .4)
T h erefore, yXI{x) is an even function. W e also n o te that yz x ( 0) = E ( X 2 ') is the a verage p o w e r o f the random process. There exist nonstationary processes with the property that the m ean value o f th e p rocess is a constant and the au tocorrelation function satisfies the property Yxz(t 1, *2) = Y x x i h h ) - Such a p rocess is called w ide-sense stationary. Clearly,
A4
App. A
w id e-sen se stationarity is a less stringent con d ition than strict-sen se stationarity. In our treatm ent w e shall require on ly that the p rocesses b e w id e-sen se stationary. R ela ted to the autocorrelation fun ction is the au tocovarian ce function, which is defined as C ( l i , l 2 ) = E { [ X tl - m ( f i ) ] [ X , j - m ( r 2)]} (A .6) = Y x x (h ,t2) m ( t\)m ( t2) w here m (t\) = E { X tl) and m (t2) = E (X ,2) are the m ean v alu es o f X h and X n , respectively. W hen the process is stationary, cXx ( h , h ) = cxx( fi - t2) =
cxx ( t )
= yxx( r) - m \
(A .7)
where r = h t2. Furtherm ore, the variance o f the p rocess is a } = <^,(0) = X(0 ) - m 2.
Statistical Averages for Joint Random Processes

Let X ( t ) and Y ( t ) be tw o random p rocesses and let X tj = X(r,-), i = 1, 2 , . . . , n , and Y,< = y{rj), j = 1, 2 , . . . , m , represent the random variables at tim es h > h > > / and t { > t'2 > > t'm, resp ectively. T h e tw o sets o f ran d om variables are characterized statistically by the joint P D F
......>,;)
for any set o f tim e instants {/,} and \t'} and for any p ositive in teger valu es o f m and n. T h e crosscorrelation fu n c tio n o t X ( t ) and Y (t), d en oted as yxv (t \ , t2)> is defined by the joint m om en t
/
and the crosscovariance is
OO
00
I
00 * ' 00
j:tly,2p ( x t], y,2) d x tldy,2
(A .8)
C x y (tu h ) = yxy( t \ , t 2) - m x (ti)m y (t2)
(A .9)
W hen the random p rocesses are jointly and individually stationary, w e have YxyVi, h ) = Yxyih - t2) and cxy(ti, t2) = cxy(h - t2). In this case ^ ( - r ) = E( XtlYll+r ) = E (X ,r r Yt) = yyx( r) (A .10)
T h e random processes X ( t ) and Y (t) are said to be statistically in d ep en dent if and on ly if p(xh x<., y ,, ......... y t> m) = p ( x h , . . . , x tt )p(y,> , - . . , y , m)
for all ch oices o f t- and for all p ositive in tegers n and m. T h e p rocesses are said to b e uncorrelated if Yxy(h, t2) = E (X ,t)E {Y ,t)
( A .ll)
so that cxy(ti, t2) = 0.
App. A
A5
A co m p lex-valu ed random p rocess Z( /) is defined as Z { t) = X ( t) + j Y ( t ) (A .12)
w here X( t ) and K(/) are random p rocesses. T he joint P D F o f the com plex-valued random variables Z,, = Z(r,), i = 1, 2 , , is given by the joint P D F o f the co m p o n en ts ( X l;, Yti), i = 1, 2 ........ n. Thus the P D F that characterizes Z,,, i = 1, 2 .........n is p ( x n , x l2.........x t ,y ,y l2, . . . , y , m ) A com p lex-valu ed random p rocess Z (t) is en cou n tered in the representation o f the in-phase and quadrature com p onents o f th e low p ass equivalent o f a nar row band random signal or noise. A n im portant characteristic o f such a process is its autocorrelation function, which is defined as = E { Z h Z* ) = E [ ( X t] + j Y tl)(X ,2 - j Y ,2)] = Y xxih, h ) + yvv('i, h )t+ j [ Y y A h , h ) Yxy(t\- h ) \ W hen the random p rocesses X (r ) and Y (t) are jointly and individually stationary, the autocorrelation function o f Z (t) b ecom es YzztiU*2) = Yzz0 1 - t2) = /;;(T ) w here
t
(A .13)
ti t2. T h e com plex conjugate of (A .1 3 ) is Y*z (r) = (Z * Z ,i_r) = y ( - r ) (A .14)
N ow , su p p ose that Z(r) = X ( t) + j Y ( t ) and W(/) = U (t) + j V ( t ) are tw o com p lex-valu ed random processes. T heir crosscorrelation function is defined as Y zv d u ti) =
e
( z w ;j . ( A .15)
= E [ ( X h + j Y li)(U l2 ~ j V t2)}
= Yxuihi h ) + Y y v U \,h ) + j[Y y u (h , h ) Yxvih* h)] W hen X ( t) , Y (t) , U (t), and V ( t ) are pairwise stationary, th e crosscorrelation func tion s in (A .1 5 ) b eco m e functions o f the tim e differen ce z t2. In addition, w e have y; j r) = E (Z * W ,t. T) = E (Z * + tWh ) = y wz( - T ) ( A .16)
Power Density Spectrum

A stationary random p rocess is an infinite-energy signal and h en ce its Fourier transform d o e s n o t exist. T h e spectral characteristic o f a random p rocess is o b tained according to the W ien er-K h in ch in e th eorem , by com puting the Fourier transform o f th e autocorrelation function. That is, the distribution o f pow er with
A6
App. A
frequency is given by the function
(A.17)
T he inverse F ourier transform is given as (A .18) W e observe that y ( 0 ) = f Z c r * * (F )d F (A .19) = (* ? ) > 0 Since E (X ? ) = yxx (0) rep resents the average p ow er o f the random p rocess, which is the area under r ^ F ) , it follow s that r ^ F ) is the distribution o f p ow er as a function o f frequency. For this reason, TXX(F ) is called the p o w e r d en sity sp e ctru m o f the random process. If the random process is real, y IX( z ) is real and even and h en ce r(F) is real and even. If the random p rocess is com p lex valued, yxx( r) = yx x ( z ) and, h ence
aoc a cc
n jF ) =
y*x (z )e j2nfrdT = /
y ; j - z ) e - ' 2" F'd z
T herefore, r xx(F) is always real. T h e definition o f the p ow er d en sity spectrum can be exten d ed to tw o jointly stationary random p rocesses X ( t ) and Y (t) , which have a crosscorrelation function yxv(z). The Fourier transform o f yxy(z) is
(A.20)
which is called the cro ss-p o w er d en sity sp e ctru m . It is easily sh ow n that T*V(F ) = rv ,( - F ) . For real random p rocesses, th e condition is P yx(F ) = I \ v( F).
Discrete-Time Random Signals

This characterization o f con tinu ou s-tim e random signals can b e easily carried over to discrete-tim e signals. Such signals are usually ob tain ed by uniform ly sam pling a con tinu ou s-tim e random process. A discrete-tim e random p rocess X {n) con sists o f an en sem b le o f sam p le s e q u en ces x ( n ). T h e statistical p rop erties o f X { n ) are sim ilar to th e characterization o f X (/), with the restriction that n is n ow an in teger (tim e) variable. T o be specific, w e state the form for the im portant m om en ts that w e use in th is text.
App. A
A7
T he /th m om en t o f X( n) is defined as E {X n) = f
J -O C
x'nP (xn)d x n
(A .21)
and the autocorrelation seq u en ce is
/
Sim ilarly, the autocovariance is
OC
/ 00
/
oc . / - 0
0
x nx kp {x nyx k)dx dxk
(A .22)
c ( n , t ) = Y x x in .k ) - E { X ) E ( X k) For a stationary p rocess, w e have the special form s (m = n k) Yxx(n - k ) = yxx(m) Cxx(n - k) = cxx(m) = - m]
(A .23)
(A .24)
w here m , = E (X ) is the m ean o f th e random process. T h e variance is defined as a2= = yjr ( 0 ) - m \. For a com p lex-valu ed stationary process Z(n) = X (n ) + j Y ( n ) , w e have y (m ) = yxx (m ) + y vv (m) + j [yyx (m ) - yxx (m )] (A .25)
and the crosscorrelation seq u en ce o f tw o com p lex-valu ed stationary seq u en ces is Y zA m ) = Yxu(m) -I- y yv(m) + j [ y yu{m) - yxv(m )] (A .26)
A s in the case o f a continuous-tim e random p rocess, a discrete-tim e random p rocess has infinite en ergy but a finite average p ow er and is given as E (X 2 n) = yxx (0) (A .27)
B y use o f the W ien er-K h in ch in e theorem , w e obtain the p ow er d en sity spectrum o f the discrete-tim e random process by com puting the F ourier transform o f the au tocorrelation seq u en ce yxx(m ), that is,
00
r,,( /)= T h e inverse transform relationship is
22
m = oc
Y xx{m )e~ M m
(A .28)
r i/2 Y x x (m )=
J-\a
1/2
T x x ( f ) e iln fm d f
(A .29)
W e ob serve that the average p ow er is f\a YxA 0) = / r xx( f ) d f J 1/2 (A .30)
so that r xx( f ) is the distribution o f p ow er as a function o f frequency, that is, Vxx( f ) is the pow er d en sity spectrum o f the random p rocess X (n ), T h e properties w e have stated for Trx(F ) also hold for Vxx( f ) .
A8
App. A
Time Averages for a Discrete-Time Random Process

A lth o u g h w e have characterized a random process in term s o f statistical averages, such as the m ean and the autocorrelation seq u en ce, in practice, w e usually have available a single realization o f th e random process. L et us con sid er the problem o f obtaining th e averages o f the random process from a sin gle realization. T o accom plish this, the random p rocess m ust be ergodic. B y definition, a random p rocess X ( n ) is ergod ic if, with p robability 1, all the statistical averages can b e d eterm in ed from a sin gle sam p le function of the process. In effect, the random process is ergod ic if tim e averages ob tain ed from a single realization are equal to the statistical (en sem b le) averages. U n d er this con d ition w e can attem pt to estim ate the en sem b le averages u sin g tim e averages from a single realization. T o illustrate this point, let us con sid er the estim ation o f th e m ean and the au tocorrelation o f the random process from a sin gle realization x ( n ). Since w e are in terested o n ly in these tw o m om en ts, w e define en god icity with respect to th ese param eters. F or additional details on the requirem ents for m ean ergodicity and autocorrelation ergodicity w hich are given b elo w , the reader is referred to the b o o k o f P apoulis (1984).
Mean-Ergodic Process
G iven a stationary random p rocess X ( n ) with m ean
let us form the tim e a erage (A .31) In general, w e view m x in (A .3 1 ) as an estim ate o f the statistical m ean w h ose value w ill vary w ith the different realizations o f the random p rocess. H en ce m x is a random variable with a P D F p ( m x). L et us com p u te the ex p ected valu e o f m , over all p ossible realizations o f X (n ). Sin ce the sum m ation and th e expectation are linear operations w e can interchange them , so that
Since th e m ean value o f the estim ate is equal to the statistical m ean , w e say that the estim a te m , is unbiased. N ex t, w e com p ute the variance o f rhx. W e have
v a r ( / n j = E(\mx \ 2) - \mx \ 2
App. A
A9
But ( |/ ,|2 ) = ( 2 /v'+ T )2 5 1

'
n= A/ k= N
T h erefore,
(A .33)
If var(m x) -* 0 as A' -* oo, the estim ate con verges with probability 1 to the statistical m ean m x. T h erefore, the p rocess X ( n ) is m ean ergod ic if (A .34) U n d er this con d ition , the estim ate m x in the lim it as N > oc b ecom es equal to the statistical m ean, that is, (A .35) T h u s th e tim e-average m ean, in the lim it as T V -* oo, is equal to the en sem b le m ean. A sufficient con d ition for (A .3 4 ) to hold is .if
00
(A .36) w hich im plies that cxx(m ) 0 as m -* oo. T h is condition h old s for m ost zero-m ean p rocesses en co u n tered in th e physical world.
Correlation-Ergodic Processes
N o w , let us con sid er the estim ate o f the autocorrelation yxx(m ) from a single realization o f the p rocess. F ollow in g ou r previou s n otation , w e d en ote the estim ate (for a com p lex-valu ed signal, in gen eral) as
(A.37)
A10
App. A
A g a in , w e regard rxx(m ) as a random variable for any given lag m, since it is a function o f the particular realization. T h e exp ected value (m ean valu e over all realizations) is E [r (m )] 1
2N + 1
N V
E [x * (n )x (n + m)] (A .38)
= 2 T h erefore, the exp ected value o f the tim e-average autocorrelation is eq u al to the statistical average. H en ce w e have an u nbiased estim ate o f yxx (m). T o determ ine the variance o f the estim ate rxx(m), w e co m p u te the exp ected value o f \rxx(m )\2 and subtract the square o f th e m ean value. T hus v ar[r(m )] = E [\rxx(m )\2] - |y (m ) |2 But
j
N N
(A .39)
E [\rXI(m ) |2] = t t t T Y E [x*{n)x(.n + m )x (k )x * (k + m )] (2 N -I-1)2 n ? Nkt ? N
(A .40)
T h e exp ected value o f the term ;c*(n)jc(n + m )x (k )x * (k + m ) is just the autocorre lation sequ en ce o f a random process defined as vm(n) = x*(n)x(n + m ) H en ce (A .4 0 ) m ay be exp ressed as
E E t f O - W ( 2 N + l ) 2 n=-Nk=-N
(A .41)
2N
T 2N + 1 and the variance is v ar[r (m )] = ^ (* ~

a v
( i - J s L W ) V 2 N + l ) rvv V '
+t )
< AAZ)
If va r[rxx(m)] -* 0 as N -* oo, the estim ate rxx(m) con verges with probability
1 to the statistical autocorrelation yxx(m). U n d er these con d ition s, the process is
correlation ergodic and th e tim e-average correlation is identical to the statistical average, that is, lim r Y ] x * ( n ) x ( n + m ) = yxx( m) N -C O O 2 N + 1 n A r
(A .43)
In our treatm ent o f random signals, w e assum e that the ran d om p rocesses are m ean ergodic and correlation ergod ic, so that w e can deal w ith tim e averages o f the m ean and the au tocorrelation ob tain ed from a sin gle realization o f the process.
Random Number Generators
In som e o f the exam p les given in the text, random num bers are generated to sim ulate the effect o f n oise on signals and to illustrate how the m ethod o f correlation can be used to d etect the presen ce o f a signal buried in n oise. In the case o f periodic signals, the correlation tech n iqu e also allow ed us to estim ate the period o f the signal. In practice, random num ber generators are often used to sim ulate the effect o f noiselik e signals and other random p h en om en a en cou n tered in the physical world. Such n oise is p resent in electron ic d evices and system s and usually limits our ability to com m u n icate over large distances and to be able to d etect relatively w eak signals. B y generating such n oise on a com puter, w e are able to study its effects through sim ulation o f com m unication system s, radar d etection system s, and the lik e and to a ssess the p erform ance o f such system s in the presen ce o f noise. M ost com puter softw are libraries include a uniform random num ber gen er ator. Such a random num ber gen erator gen erates a num ber b etw een zero and 1 with equal probability. W e call the output o f the random num ber generator a random variable. If A d en otes such a random variable, its range is the interval 0 < A < 1. W e know that the num erical output o f a digital com p uter has lim ited preci sion, and as a co n seq u en ce, it is im possible to rep resen t the continuum o f num bers in the interval 0 < A < 1. H ow ever, w e can assum e that our com puter represents each output by a large num ber o f bits in either fixed point or floating point. C on se quently, for all practical purposes, the num ber o f outputs in the interval 0 < A < 1 is sufficiently large, so that w e are justified in assum ing that any value in the interval is a possib le ou tp u t from th e generator. T he uniform probability density function for the random variable A , d en oted as p ( A ) , is illustrated in Fig. B .la . W e n ote that the average valu e or m ean value o f A , d en oted as m A, is m A = T h e integral o f th e probability density function, w hich represents the area under p ( A ) , is called the p robability distribution function
B1
B2
P(A)
App. B
2
(a)
(b) o f the random variable A and is defined as F (A ) = f

J OC
Figure B .l
p (x )d x
( B. l )
For any random variable, this area must always b e unity, w hich is the m axim um value that can be achieved by a distribution function. H en ce (B .2) and the range o f F( A) is 0 < F (A ) < 1 for 0 < A < 1. If w e wish to generate uniform ly distributed n oise in an interval (b , b + 1 ) it can sim ply be accom plished by using the output A o f the random num ber gen erator and shifting it by an am ount b. Thus a new random variable B can be defined as B = A + b w hich n ow has a m ean value m s = b + For exam p le, if b = (B .3) the random
variable B is uniform ly distributed in the interval ( | , | ) , as sh ow n in Fig. B.2a. Its probability distribution function F ( B ) is sh ow n in Fig. B.2b. A uniform ly distributed random variable in the range ( 0 , 1 ) can b e used to gen era te random variables with other probability distribution functions. For exam p le, suppose that w e wish to gen erate a random variable C with probability distribution function F ( C ) , as illustrated in Fig. B.3. Since the range o f F i C ) is the interval (0 , 1 ), w e begin by generating a uniform ly distributed random variable A in the range ( 0, 1) . If we set F (C ) = A (B .4)
then C = F~l (A) (B.5)
App. B
B3
Figure B.2
Figure B3 T hus w e so lv e (B .4 ) for C and the solution in (B .5) provides the value o f C for w hich F ( C ) = A. B y this m eans w e obtain a n ew random variable C with probability distribution F (C ). This inverse m apping from A to C is illustrated in Fig. B.3. Example B.1 Generate a random variable C that has the linear probability density function shown in Fig. B.4a, that is,
p(C)={2'
0,
05 C - 2
otherwise
Solution
This random variable has a probability distribution function

0,
F( C)
iC A1 - 2' 1,
C< 0 0< C < 2 C > 2
which is illustrated in Fig. B.4b. We generate a uniformly distributed random variable A and set F( C) = A. Hence
F( C) j C 1 = A
B4
App. B
Figure B.4
Upon solving for C, we obtain

c
Thus we generate a random variable C with probability function F (C ). as shown in Fig. B.4b. In E xam ple B .l the inverse m apping C = F - 1 (A) was sim ple. In som e cases it is not. This problem arises in trying to gen erate random num bers that h ave a norm al distribution function. N o ise encou n tered in physical system s is o ften characterized by th e norm al or G aussian probability distribution, which is illustrated in Fig. B.5. T h e probability
Figure B.5
App. B
B5
density function is given by oo < C < oo (B . 6 )
w here ct2 is the variance o f C , w hich is a m easure o f the spread o f the probability density function p {C ). T he probability distribution function F (C ) is the area under p (C ) over the range ( 0 0 , C ). Thus (B.7) U n fortu n ately, the integral in (B .7) can n ot be expressed in term s o f sim ple func tions. C onsequently, the inverse m apping is difficult to ach ieve. A w ay has b een found to circum vent this problem . F rom probability th e ory it is know n that a (R ayleigh distributed) random variable R , with probability distribution function
0, 1
R < 0 R > 0
(B.8)
is related to a pair o f G aussian random variables C and D , through the transfor m ation C = R cos D R sin (B .9) (B .10)
w here is a uniform ly distributed variable in the interval (0, 2 n ). T h e param eter ct2 is the variance o f C and D . Since (B . 8 ) is easily inverted, w e have F (R ) = 1 - e ~ R2f2n2 = A and h en ce (B.ll) w here A is a uniform ly distributed random variable in the interval ( 0, 1) . N ow if w e gen erate a secon d uniform ly distributed random variable B and define
s=
2nB
(B.12)
then from (B .9 ) and (B .10), w e obtain tw o statistically in d ep en d en t G aussian dis tributed random variables C and D . T he m eth od described ab ove is o ften used in practice to gen erate G aussian distributed random variables. A s sh ow n in Fig. B.5, these random variables have a m ean value o f zero and a variance a 2. If a n on zero m ean G aussian random variable is desired, then C and D can be translated by the addition o f the m ean value. A subroutine im plem enting this m eth od for generating G aussian distributed random variables is given in Fig. B .6 .
B6
C C C C C C C C C SUBROUTINE GAUSS PI=4.0*ATAN B=2.0*PI*B R=SQRT RETURN END C C C C C C NOTE: ( 1 .0)

SUBROUTINE GAUSS CONVERTS A UNIFORM RANDOM SEQUENCE XIN IN PARAMETERS XIN B SIGMA YOUT :U N I F O R M IN 0,1] [0, 1 ] RANDOM NUMBER RANDOM NUMBER :U N I F O R M I N [0,1] TO A GAUSSIAN RANDOM S E Q U E N C E W I T H G ( 0 , S I G M A * *2)
App. B
:S T A N D A R D D E V I A T I O N O F T H E G A U S S I A N :O U T P U T F R O M T H E G E N E R A T O R 9XIN,B,SIGMA,YOUT)
(2,0*(S I G M A * * 2 ) * A L O G ( 1 . 0 / (1.0-XIN) ) )
Y O U T = R * C O S (B }
TO USE THE ABOVE
SUBROUTINE
FOR A
GAUSSIAN RANDOM NUMBER GENERATOR YOU MUST PROVIDE AS INPUT TWO UNIFORM RANDOM NUMBERS XIN AND B XIN AND B MUST BE STATISTICALLY INDEPENDENT
Figure B.6
S u b ro u tin e fo r g e n e ra tin g G a u s s ia n ra n d o m v a ria b le s
Appendix Tables of Transition Coefficients for the Design of Linear-Phase FIR Filters
In Section 8.2.3 w e described a design m ethod for linear-phase F IR filters that in volved the specification of H r (a> ) at a set o f equally spaced frequencies (ok = 2 n ( k + cr)/M, w here a = 0 or a = 3 , k = 0, 1 , . . . , (M l ) / 2 for M od d and k = 0, 1 , 2 ___ _ (M /2 ) 1 for M even , w h ere M is the length o f the filter. W ithin the passband o f the filter, w e select Hr (a>k) = 1, and in the stopband, H r (a)k ) 0. For freq u en cies in the transition band, the values o f H r (cok) are op tim ized to m inim ize the m axim um sid elob e in the stopband. This is called a m in im a x o p tim iza tio n criterion. T h e optim ization o f the values o f Hr (cu) in the transition band has been p er form ed by R ab in er et al. (1970) and tables o f transition valu es have b een provided in the published paper. A selected num ber o f the tables for low p ass FIR filters are included in this appendix. Four tables are given. T ab le C .l lists the transition coefficients for the case a = 0 and o n e coefficient in the transition band for both M odd and M ev en . T ab le C.2 lists th e transition coefficients for the case a = 0, and tw o coefficients in the transition band for M od d and M ev en . T ab le C.3 lists the transition coefficients for the case a = \ , M even and o n e coefficient in the transition band. Finally, T able C.4 lists the transition coefficients for the case a = 5 , M ev en , and tw o coefficients in the transition band. T h e tables also in clude the level o f the m axim um sid elob e and a bandwidth param eter, denoted as BW . T o use the tables, w e begin with a set o f specifications, including (1) the bandw idth o f the filter, which can be defined as (2 jt/A /)(B W + a ) , w here B W is the num ber o f con secu tive frequencies at w hich H(a)k) = 1, (2) the width o f the transition region, w hich is roughly 2 n / M tim es the n um ber o f transition coeffi cients, and (3) the m axim um tolerab le sid elob e in the stopband. T he length o f the filter can be selected from the tables to satisfy th e specifications.
C1
TABLE C.1
TRANSITION COEFFICIENTS FOR a = 0 Af Odd Af Even Ti BW Minimax Af = 16 -39.75363827 -37.61346340 -36.57721567 -35.87249756 -35.31695461 -35.51951933 Af = 32 -42.24728918 -41.29370594 -41.03810358 -40.93496323 -40.85183477 -40.75032616 -40.54562140 -39.93450451 -38.91993237 M = 64 -42.96059322 -42.30815172 -42.32423735 -42.43565893 -42.55461407 -42.66526604 -43.01104736 43.28309965 -43.56508827 43.96245098 -44.60516977 -43.81448936 Af = 128 -43.15302420 -42.59092569 -42.67634487 -42.84038544 -42.99805641 -43.25537014 -43.52547789 -43.93180990 -44.18097305 -44.40153408 -44.67161417 -45.17186594 -46.92415667 -49.46298973 7i
BW
Minimax Af = 15 -42.30932283 -41.26299286 -41.25333786 -41.94907713 -44.37124538 -56.01416588 Af = 33 43.03163004 -42.42527962 -42.40898275 -42.45948601 -42.52403450 -42.44085121 -42.11079407 -41.92705250 -44.69430351 -56.18293285 Af = 65 -43.16935968 -42.61945581 -42.70906305 -42.86997318 -43.01999664 -43.14578819 -43.44808340 -43.54684496 -43.48173618 -43.19538212 -42.44725609 -44.76228619 -59.21673775 Af = 125 -43.20501566 -42.66971111 -42.77438974 -42.95051050 -43.25854683 -43.47917461 -43.63750410 -43.95589399 -44.05913115 -44.05672455 -43.94708776 -43.58473492 -42.14925432 -42.60623264 -44.78062010 -56.22547865
1 2 3 4 5 6 1 2 3 4 6 8 10 12 14 15 1 2 3 4 5 6 10 14 18 22 26 30 31 1 2 3 4 6 8 10 18 26 34 42 50 58 59 60 61
0.43378296 0.41793823 0.41047636 0.40405884 0.39268189 0.35766525 0.42994995 0.41042481 0.40141601 0.39641724 0.39161377 0.39039917 0.39192505 0.39420166 0.38552246 0.35360718 0.42919312 0.40903320 0.39920654 0.39335937 0.38950806 0.38679809 0.38129272 0.37946167 0.37955322 0.38162842 0.38746948 0.38417358 0.35282745 0.42899170 0.40867310 0.39868774 0.39268189 0.38579101 0.38195801 0.37954102 0.37518311 0.37384033 037371826 0.37470093 0.37797851 0.39086304 0.39063110 0.38383713 0.35263062
1 2 3 4 5 6 1 2 3 4 5 8 10 12 14
0.42631836 0.40397949 0.39454346 0.38916626 0.38840332 0.40155639 0.42856445 0.40773926 0.39662476 0.38925171 0.37897949 0.36990356 0.35928955 0.34487915 0.34407349
1 2 3 4 5 6 10 14 18 22 26 30
0.42882080 0.40830689 0.39807129 0.39177246 0.38742065 0.38416748 0.37609863 0.37089233 0.36605225 0.35977783 0.34813232 0.29973144
1 2 3 4 5 7 10 18 26 34 42 50 58 62
0.42889404 0.40847778 0.39838257 0.39226685 0.38812256 0.38281250 0.3782638 0.37251587 0.36941528 0.36686401 0.36394653 0.35902100 0.34273681 0.28751221
Source: Rabiner ct al. (1970); 1970 IEEE; reprinted with permission.
App. C
Transition Coefficients for the Design of Linear-Phase FIR Filters
C3
TABLE C.2 TRANSITION COEFFICIENTS FOR a = 0

M Odd M Even T\ Ti
BW
Minimax Af -70.60540585 -69.26168156 -69.91973495 -75.51172256 -103.45078300
BW
Minimax
M = 16
Tx
t2
1 2 3 4 5 1 2 3 5 7 9 11 13 14 1 2 3 4 5 9 13 17 21 25 29 30 1 2 3 5 7 9 17 25 33 41 49 57 58 59 60
= 15 0.09500122 0.10319824 0.10083618 0.08407953 0.05180206 M = 33 -70.60967541 0.09497070 -68.16726971 0.10585937 0.10937500 -67.13149548 -66.53917217 0.10965576 -67.23387909 0.10902100 0.10502930 -67.85412312 -69.08597469 0.10219727 -75.86953640 0.08137207 104.04059029 0,05029373 M = 65 70,66014957 0.09472656 68,89622307 0.10404663 -67.90234470 0.10720215 0.10726929 67.24003792 -66.86065960 0.10689087 0.10548706 -66.27561188 -65.96417046 0.10466309 0.10649414 -66.16404629 0.10701904 -66.76456833 -68.13407993 0.10327148 -75.98313046 0.08069458 -104.92083740 0.04978485 M = 125 -70.68010235 0.09464722 0.10390015 -68.94157696 -68.19352627 0,10682373 0.10668945 -67.34261131 0.10587158 -67.09767151 0.10523682 -67.058012% -67.17504501 0.10372925 -67.22918987 0.10316772 -67.11609936 0.10303955 -66.71271324 0.10313721 -66.62364197 0.10561523 -69.28378487 0.10061646 -70.35782337 0.09663696 -75.94707718 0.08054886 -104.09012318 0.04991760
0.58995418 0.59357118 0.58594327 0.55715312 0.49917424 0.58985167 0.59743846 0.59911696 0.59674101 0.59417456 0.58771575 0.58216391 0.54712777 0.49149549 0.58945943 059.476127 0,59577449 0.59415763 0.59253047 0.58845983 0.58660485 0.58862042 0.58894575 0.58320831 0.54500379 0.48965181 0.58933268 0.59450024 0.59508549 0.59187505 0,59821869 0.58738706 0.58358265 0.58224835 0.58198956 0.58245499 0.58629534 0.57812192 0.57121235 0.54451285 0.48963264
1 2 3 4 5 1 2 3 5 7 9 11 13
0.10703125 0.12384644 0.12827148 0.12130127 0.11066284 M = 32 -67.37020397 0.09610596 -63.93104696 0.11263428 -62.49787903 0.11931763 0.12541504 -61.28204536 -60.82049131 0.12907715 -59.74928167 0.12068481 -62.48683357 0.13004150 -70.64571857 0.11017914
M = 64
-65.27693653 -62.85937929 -62.96594906 -66.03942485 -71.73997498
0.60559357 0.62201631 0.62855407 0.61952704 0.60979204 0.59045212 0.60560235 0.61192546 0.61824023 0.62307031 0.60685586 0.62821502 0.60670943
1 2 3 4 5 9 13 17 21 25 29
-70.26372528 -67.20729542 -65.80684280 -64.95227051 -64.42742348 -63.41714096 -62.72142410 -62.37051868 -62.04848146 -61.88074064 -70.05681992
0.09376831 0.10411987 0.10850220 0.11038818 0.11113281 0.10936890 0.10828857 0.11031494 0.11254273 0.11994629 0.10717773
0.58789222 0.59421778 0.59666158 0.59730067 0.59698496 0.59088884 0.58738641 0.58968142 0.59249461 0.60564501 0.59842159
M -= 128
1 2 3 4 6 9 17 25 33 41 49 57 61
-70.58992958 -68.62421608 -67.66701698 -66.95196629 -66.32718945 -66,01315498 -65.89422417 -65.92644215 -65.95577812 -65.97698021 -65.67919827 -64.61514568 -71.76589394
0.09445190 0.10349731 0.10701294 0.10685425 0.10596924 0.10471191 0.10288086 0.10182495 0.10096436 0.10094604 0.09865112 0.09845581 0.10496826
0.58900996 0.59379058 0.59506081 0.59298926 0.58953845 0.58593906 0.58097354 0.57812308 0.57576437 0.57451694 0.56927420 0.56604486 0.59452277
Source: Rabiner et al. (1970); 1970 IEEE; reprinted with permission.
C4
App. C
TABLE C.3 TRANSITION COEFFICIENTS FOR a = \ BW 1 2 3 4 5 6 1 2 3 4 6 8 10 12 14 1 2 3 4 5 6 10 14 18 22 26 30 1 2 3 4 5 7 10 18 26 34 42 50 58 62 Minimax Af = 16 -51.60668707 -47.48000240 -45.19746828 -44.32862616 -45.68347692 -56.63700199 M = 32 -52.64991188 -49.39390278 -47.72596645 -46.68811989 -45.33436489 -44.30730963 -43.11168003 -42.97900438 -56.32780266 Af = 64 -52.90375662 -49.74046421 -48.38088989 -47.47863007 -46.88655186 -46.46230555 -45.46141434 -44.85988188 -44.34302616 -43.69835377 -42.45641375 -56.25024033 Af = 128 -52.96778202 -49.82771969 -48.51341629 -47.67455149 -47.11462021 -46.43420267 -45.88529110 -45.21660566 -44.87959814 -44.61497784 -44.32706451 -43.87646437 -42.30969715 -56.23294735 h 0.26674805 0.32149048 0.34810181 0.36308594 0.36661987 0.34327393 0.26073609 0.30878296 0.32984619 0.34217529 0.35704956 0.36750488 0.37810669 0.38465576 0.35030518 0.25923462 0.30603638 0.32510986 0.33595581 0.34287720 0.34774170 0.35859375 0.36470337 0.36983643 0.37586059 0.38624268 0.35200195 0.25885620 0.30534668 0.32404785 0.33443604 0.34100952 0.34880371 0.35493774 0.36182251 0.36521607 036784058 037066040 037500000 0.38807373 035241699
Source: Rabiner el al. (1970); 1970

IEEE; reprinted with permission.
App. C
05
TABLE C.4 TRANSITION COEFFICIENTS FOR

1
BW 1 2 3 4 5 1 2 3 5 7 9 11 13 1 2 3 4 5 9 13 17 21 25 29 1 2 3 4 6 9 17 25 33 41 49 57 61 Source;
Minimart
Ti
T2 0.41784180 0.49369211 0.51966134 0.51158076 0.46967784 0.40357383 0.49129255 0.52153983 0.54805908 0.56031410 0.56637987 0.56226952 0.48479068 0.40168723 0.48390015 0.51850058 0.53379876 0.54311474 0.56020256 0.56903714 0.57543691 0.58007699 0.57729656 0.48767025 0.40117195 0.47399521 0.51361278 0.53266251 0.54769675 0.55752959 0.56676912 0.57137301 0.57594641 0.57863142 0.58123560 0.57946395 0.48900685
M = 16 -77.26126766 0.05309448 -73.81026745 0.07175293 -73.02352142 0.07862549 -77.95156193 0.07042847 -105.23953247 0.04587402 M = 32 -80.49464130 0.04725342 -73.92513466 0.07094727 -72.40863037 0.08012695 -70.95047379 0.08935547 -70.22383976 0.09403687 -69.94402790 0.09628906 -70.82423878 0.09323731 -104.85642624 0.04882812 M = 64 -80.80974960 0.04658203 -75.11772251 0.06759644 -72.66662025 0.07886963 -71.85610867 0.08393555 -71.34401417 0.08721924 -70.32861614 0.09371948 -69.34809303 0.09761963 -68.06440258 0.10051880 -67.99149132 0.10289307 -69.32065105 0.10068359 -105.72862339 0.04923706 = 128 M= -80.89347839 0.04639893 -77.22580583 0.06295776 -73.43786240 0.07648926 -71.93675232 0.08345947 -71.10850430 0.08880615 -70.53600121 0.09255371 -69.95890045 0.09628906 -69.29977322 0.09834595 -68.75139713 0.10077515 -67.89687920 0.10183716 -66.76120186 0.10264282 -69.21525860 0.10157471 -104.57432938 0.04970703 Rabiner et al.
(1970); 1970 IEEE;
reprinted with permission.
C6
App. C
A s an illustration, the filter design for which Af = 15 and
corresponds to a = 0, B W = 4, sin ce H r (p>k) = 1 at th e four con secu tive fre quencies a>k = I n k 115, k 0, 1, 2, 3, and the transition coefficien t is T\ at the frequency a> k = 8 n / l 5 . T he valu e given in T ab le C .l for M = 15 and B W = 4 is 7\ = 0.40405884. T h e m axim um sid elo b e is at 41.9 d B , according to T ab le C .l.
Appendix D List of MATLAB Functions
In this A p p en d ix, w e list several M A T L A B functions that th e student can use to so lv e som e o f the problem s num erically. T h e list includes th e m ost relevant M A T L A B functions for each o f the chapters, but it is not exh au stive. H ow ever, this list is cum ulative in the sense that on ce a function is listed in any chapter, it is not re p eated in subsequent chapters. T h ese M A T L A B functions are obtained from two sources: (1) the student version o f M A T L A B and (2) the b o o k en titled D igital S ig n a l P ro cessin g U sing M A T L A B . (PW S K ent 1996), by V .K . Ingle and J.G . Proakis. O ur primary objective in listing these M A T L A B fun ction s is to inform the student w ho is not fam iliar with M A T L A B o f the existen ce o f these functions and to encourage the student to use them in the solution o f som e o f the hom ew ork problem s. CH APTER 1 sin (x ), c o s (x ), ta n (x ) a b s(x ) r e a l(x ) im a g (x ) c o n j(x ) e x p (z ) su m (x ) p ro d (x ) a n g le(x ) log(a;> trigonom etric functions absolute values o f a vector x with real or com plex com ponents. takes the real part o f each com p on en ts o f the vector x. takes the im aginary part o f each com p onent o f the vector x . com plex-conjugate o f each com p on en t o f x . ex(co s^ + j sin >), w here z = x + j y , sum o f the (real or com p lex) com p onents o f the vec tor x . product o f the (real or com p lex) com p on en ts of the vector x . com p utes the phase angles o f each com p onent o f the vector x . com p utes the natural logarithm o f each o f the e le m en ts o f x .
D1
D2
List of MATLAB Functions
App. D
loglO(x)
sq r t(x )
com p utes the logarithm to the base 1 0 o f the elem en ts of x. com p utes the square root o f the elem en ts o f x .
CH APTER 2
conv(x, h ) fliplr(x) filter(b, a , x )
con volu tion o f the tw o (vector) seq u en ce s x and h . folds the (vector) seq u en ce x . solves the differen ce eq u ation with coefficients a = [a0, a \ , . . . , a ] b = [fr0, b i , . . . , b m] x = input seq u en ce
filler(6.1, x )
r a n d (l, N ) r a n d n (l, N )
im plem ents an F IR filter with input x and coeffi cien ts b. gen erates a length N random seq u en ce that is uniform in the interval ( 0 , 1 ). gen erates a length N seq u en ce o f G uassian random variables with zero m ean and unit variance, com p utes the crosscorrelation o f th e tw o seq u en ces x and y. com p utes the autocorrelation o f the seq u en ce x .
xcorr(x, y ) xcorr(x)
CH APTER 3
roots(a)
com p utes the roots o f the polyn om ial w ith coefficients

a = [ao ,
, q n \
residuez(b, a )
com p utes the residues in a partial fraction expansion, where b = coefficients o f num erator polyn om ial [i>o, b \ , . . . b M] a = coefficients o f d en om in ator polyn om ial
[ a o , >i, -<3w]
deconv(i>, a) poly(r) pzplotz(b, a )
com p utes the result o f dividing b by a in a polyn om ial part p and a rem ainder r. com p utes the coefficients o f the p olyn om ial p with roots r. plots the p o les and zeros in the z-p lane given th e co efficient vectors b and a .
App. D
D3
fi)ter(b, a , x , x ic )
im plem en ts the filter given by a d ifferen ce eq u ation with coefficient vector b and a, input x and initial con d ition s x ir.
CHAPTER 4
freqz(6, a , N )
com p utes an //-p o in t com p lex frequency resp onse v ec tor and an Appoint frequency vector o j , uniform over the interval 0 < co < n , for filter with coefficient v e c tor & and a. sam e com p utation as freqz(, a, N), excep t that the freq u en ce range is 0 < co < 2n. com p utes the frequency resp on se o f the system at the freq u en cies specified by th e vector com p utes the group d elay o f the filter w ith num era tor polyn om ial having coefficients b and denom inator p olyn om ial with coefficients o , at N p oints over the interval (0 , jr). sam e as ab ove, ex cep t that the frequency range is 0 <
oj <
freqz(6, a , N , whole) freqz(b, a , u>) grpdelay(t>, a, N )
grpdelay(b, a , TV, whole)
In .
CHAPTER 5
dfs(x, N )
com p utes the discrete fourier series (D F S ) coefficient array for the period ic signal seq u en ce x with period
N.
Jdfs ( y , N ) rem (n. N ) m od(n , N ) dft(x, N ) idft(AT, N ) ovrlpsav(x, h, N )
com p utes the signal seq u en ce from the D F S co effi cient array y. determ in es the rem ainder after dividing n by N. com p utes n m od N. com p utes the Appoint D F T o f the data seq u en ce x. com p utes the N -point inverse D F T o f X . im plem ents the overlap -save m eth od to perform block con volu tion w here N is the block length.
CHAPTER 6
ffl(x , N ) i (X , N ) fftshift(x)
im plem ents a radix-2, W -point F F T algorithm , im plem en ts a radix-2, N -p oin t inverse F F T algorithm , rearranges the ou tputs o f fft so that the zero fre quency com p on en t is the cen ter o f the spectrum .
CHAPTER 7
dir2cas(b, a) converts a direct form structure to the cascade form.
D4
App. D
cas2dir(bo, B , A ) dir2par(b, a ) par2dir(C0B, A ) dir21atc(b) latc2dlr(AT) dir2ladr(b, a) Iadr2dir(AT, C ) casfiltr(60, B , A , x ) parfiltr(C, B , A , x ) latcfilt(i, x ) ladrfilt(K, C . x ) round(x) fix(x) sign(x) ss2tf(J 4 , jB, C , D , i u )
converts a cascade structure to the direct form struc ture. converts direct form to the parallel form structure, converts a parallel form to the direct form structure, converts a F IR direct form structure to an all-zero latice structure. converts an all-zero lattice structure to the direct form structure. converts direct form IIR structure to p ole-zero latticeladder structure. converts a lattice-ladder structure to the direct form IIR structure. im plem ents the cascade form IIR and F IR realization o f a filter with input seq u en ce x . im plem ents the parallel form IIR realization o f a dilter with input seq u en ce x . im plem ents the F IR lattice filter realization with input seq u en ce x . im plem ents the lattice-ladder realization o f a filter with input seq u en ce x . rounds the com p on en ts o f the vector x to the nearest integer rounds (truncates) the com p on en t o f th e vector x to the nearest in teger tow ard zero, each com p onent o f x is set to + 1 if it is positive and 1 if it is negative. com p utes the transfer function H ( x ) o f a system given the state-sp ace description o f the form x A x + Bu y = Cx + Du from the i'wth input.
ss2zp(J 4, D , C , D , tu )
com p utes the transfer function H (s ) and exp resses it in factored form , thus, giving the p o les and zeros o f H (s).
CH APTER 8
boxcar(M) bartlett(M ) hanning(M)
gen erates an A f-point rectangular w indow , gen erates an M -p oint B artlett w indow , gen erates an Af-point H an n ing w indow .
App. D
D5
ham m iitg(M ) blackm an(M ) kaiser(M ) buttap(iV)
gen erates an M -p oint H am m ing w indow , gen erates an M -point B lackm an w indow , gen erates an M -point kaiser w in d ow , p rovides th e coefficients o f an analog low pass B u tter w orth filter o f order N , w ith norm alized frequency, in cascade form. provides the coefficients o f an analog low p ass C heby shev filter o f order N , with norm alized frequency and passband ripple R p, in cascade form , p rovides the coefficients o f an an alog low pass elliptic filter o f order N , passband ripple R p, stopband atten u ation A 2, with norm alized frequency, in cascad e form, com p utes th e frequency resp onse o f an analog filter, with co in rad/sec. designs a digital low pass B utterw orth filter o f order N and cu toff frequency con. designs a digital low pass C h eb ysh ev filter o f order N passband ripple R p, and cu toff frequency con. designs a T ype 2 low pass C h eb ysh ev filter or order N , stopband ripple A and cu to ff frequency con. designs a digital low pass elliptic filter o f order N , passband ripple R p , stopband ripple A s , and cu to ff fre quency cun. uses the bilinear transform ation to convert an analog filter with zeros z, p o les p , and gain k, in to a digital filter, with fs being the sam ple frequency in H z. uses th e bilinear transform ation to con vert an ana log filter with num erator p olyn om ial coefficients num, den om in ator p olyn om ial coefficients den, and sam ple frequency fs, in to a digital filter, uses the R em ez algorithm to d eterm in e the coeffi cients o f an optim um equiripple, linear p h ase F IR fil ter o f length N + 1, from freq u en cy specifications / and gains m for each band. sam e description as ab ove with fty p e used to specify a H ilbert transform or differentiator, designs a highpass B utterw orth filter o f order N and 3-dB cu toff frequency con. designs a 2 N -order bandpass B utterw orth filter, with 3-dB passband col < co < co2, w h ere con = [twl, co2\. designs a highpass C hebyshev filter o f order N , pass band ripple R p , and cu toff frequency con.
chebiap(iV, R )
ellipap(7V, R P, A S)
freqs(fc>, a , u>)
butter(iV, u n ) chebyl(7V, R p , urn) cheby2(AT, A S i u>n)

e l l i p ( N , R p , A s ,u>n)
bilinear(z, p. fc, fs)
bilinear(num, den, fs)
remez(7V, / , m )
remez ( N , / , m , ftype') butter(7V, w n , high) botter(iV, w n , bandpass) chebyl(iV , i i,,, w n , high)
D6
App. D
ellilp(iV, R p , A s , w n )
designs an ellip tic bandpass filter o f order N , pass band ripple Rp , stopband atten u ation A s , and cu toff freq u en cies a>n = [a>l, w2], transform s an analog low pass filter to an analog band pass filter. transform s an analog low pass filter to an analog b and stop filter. transform s an analog low pass filter to an analog highpass filter. transform s an analog low p ass filter to an analog lowpass filter with cu toff frequency coo. finds a p olyn om ial p such that p ( x ) fits the data in a vector y in a least squares sense.
lp2bp(num, den, u>o, B u t) lp2bs(num, den,

u jo
.B u r)
lp2hp(num, den, u o ) Ip2lp(num, den, u>o) polyfit(x, y , n )
CH APTER 9
spline(x, y , x i ) spline(jTit, x , t)
cubic spline data interpolation, uses cubic spline interp olation , w h ere x and n t s are arrays containing sam ples x ( n ) at m s , and t is an ar ray that contains a fine grid at w hich the function is evaluated.
C H A P T E R 10
dnsam ple(x, M )
dow n sam p les the seq u en ce x by th e factor M .
L :_ References and Bibliography
A k a ik e , H. 1969. P o w er Spectru m E stim atio n T hrou gh A u toreg ressio n M o d el F ittin g ,"
A n n . Inst. Stat. M ath., V o l. 21, pp. 4 0 7 -1 4 9 .

A k a ik e , H. 1974. A New L o o k at the S tatistical M odel Id en tificatio n , I E E E Trans. A u to
m atic Control, V o l. A C -1 9 , pp. 7 1 6 -7 2 3 , D ecem b e r.

A n d e r s e n , N. O. 1978. C om m en ts on th e p erform an ce o f M axim um E n tro p y A lg o rith m .
Proc. I E E E , V o l. 66, pp. 1 5 8 1 -1 5 8 2 , N ovem ber.

A n t o n i o u , A . 1979. D ig ita l Filters: A n a ly s is and D e s ig n , M cG raw -H ili, New Y o rk . A u e r , E. 1987. A D igital F ilte r Structu re F re e o f L im it C y cles, Proc. 19 8 7 I C A S S P ,
pp. 2 1 .1 1 .1 -2 1 .1 1 .4 , D allas, T e x ., A pril.

A v e n h a u s , E ., and S c h u e s s l e r , H. W . 1970. O n the A p p ro xim atio n P ro b lem in th e D esign
o f D igital F ille rs with L im ited W o rd len g th , A rch . E le k . Ubertragung, V o l. 2 4 , pp. 571 572.
B a g g e r o e r , A . B . 1976. C onfid en ce In tervals fo r R egression (M E M ) S p ectra ! E stim a te s,
I E E E Trans. Inform atio n Th e o ry , V ol. IT -2 2 , pp. 5 3 4 -5 4 5 , S ep tem b er.

B a n d l e r , J. W ., and B a r d a k j ia n , B . J. 1973. L e ast p ih O ptim ization o f R e cu rsiv e D igital
F ilte rs, I E E E Trans. A u d io and Electroacoustics, V o l. A U -2 1 , pp. 4 6 0 4 7 0 . O cto b er.

B a r n e s , C. W and F a m , A . T . 1977. M inim um N orm R ecu rsiv e D ig ital F ilte rs T h a t A re
F re e o f O verflow L im it C y cles, I E E E Trans. C ircu its an d Systems, V o l. C A S -2 4 , pp. 5 6 9 5 7 4 , O cto b er.

B a r t l e t t , M. S. 1948. Sm oo th in g Periodogram s from T im e S eries with C ontin u ou s S p e c
tr a , N ature (L o n d o n ), V o l. 161, pp. 6 8 6 -6 8 7 , M ay.

B a r t l e t t , M . S. 1961. Stochastic processes, C am bridge U niversity P ress, C am bridge. B e r g l a n d , G. D. 1969. A G uided T o u r o f th e F ast F o u rie r T ra n sfo rm , I E E E Spectrum,
V o l. 6, pp. 4 1 - 5 2 , Ju ly . B e r k , K. N. 1974. C o nsisten t A u toreg ressiv e Sp ectral E stim a te s, A n n . Stat., \ o \ . 2, p p . 4 8 9 502.

B e r n h a r d t , P. A., A n t o n i a d is , D. A., and D a R o s a , A. V. 1976. L u n ar P ertu rb ation s in
C o lu m n ar E le c tr o n C o n ten t and T h e ir In te rp retatio n in T erm s o f D y n am o E le ctro sta tic F ie ld s, /. Geo ph ys. R es., V o l. 81, pp. 5 9 5 7 -5 9 6 3 , D ecem b e r.
B e r r y m a n , J. G. 1978. C h o ice o f O p e ra to r L en gth for M axim um E n tro p y S p e ctra l A n aly
sis, G e o p h ysics, V o l. 4 3 , pp. 1 3 8 4 -1 3 9 1 , D ecem b e r.

B l a c k m a n , R. B ., and T u k e y , J. W. 1958. The Measurem ent o f P o w er Spectra, D o v e r. New
Y o rk .
R1
R2
References and Bibliography
B l a h u t , R. E. 1985. Fast A lg o rithm s fo r D ig ita l S ig n a l Processing, A d d ison -W esley , R ead in g, M ass. B l u e s t e i n , L. I. 1970. A L in ear F ilterin g A p p ro ach to th e C o m p u tatio n o f the D iscrete F o u rier T ra n sfo rm , I E E E Trans. A u d io an d Electro acou stics, V o l. A U -1 8 , pp. 4 5 1 -4 5 5 , D ecem b e r. B o l t , B . A. 1988. Earth qu akes , W . H . F re em a n and C o ., N ew Y o rk . B o m a r, B. W. 1985. New Seco n d -O rd e r S ta te -S p a ce Stru ctu res for R e a liz in g Low R o u n d o ff N oise D igital F ilte r s, I E E E Trans. Acoustics, Speech, a n d Sig n a l Processing, V ol. A S S P 33, pp. 1 0 6 -1 1 0 , F eb ru ary . B r a c e w e l l , R. N. 1978. T he Fo u rie r Transform an d Its A p p licatio n s, 2nd ed ., M cG raw -H ill, New Y o rk . B r ig h a m , E. O. 1988. The Fast Fo u rie r Transform a n d Its A p p licatio n s. P ren tice H all. E n glew ood C liffs, N J . B r ig h a m , E. 0 . , and M o r r o w , R. E. 1967. T h e F a st F o u rie r T ra n sfo rm , I E E E Spectrum , V o l. 4. pp. 6 3 -7 0 , D ecem b er. B r i l l i n g e r , D. R. 1974. F o u rier A nalysis o f Statio n ary P ro cesses, Pro c. I E E E . V o l. 62, pp. 1 6 2 8 -1 6 4 3 , D ecem b e r. B r o p h y , F., and S a l a z a r , A. C. 1973, C o nsid eratio n s o f th e Pad e A p p ro xim an t T ech n iq u e in the Synthesis o f R ecu rsiv e D igital F ilte r s, I E E E Trans. A u d io an d Electroacoustics , V ol. A U -2 1 . pp. 5 0 0 -5 0 5 . D ecem b er. B r o w n . J. L J r ., 1980. F irst-O rd e r Sam plin g o f B an d p ass S ig n a ls -A N ew A p p ro a ch , I E E E Trans. Inform ation Theory. V o l. IT -2 6 , pp. 6 1 3 -6 1 5 , S ep tem b er. B r o w n , R. C. 1983. Introduction to R an d o m S ig n a l A n a ly s is an d K a lm a n Filte rin g , W iiey, New Y o rk . B r u b a k e r , T. A., and G o w d y , J. N. 1972. L im it C ycles in D ig ital F ilte r s, I E E E Trans. Autom atic Co n trol. V o l. A C -1 7 , pp. 6 7 5 -6 7 7 , O cto b er. B r u z z o n e , S. P., and K a v e h , M. 1980. O n S o m e Suboptim u m A R M A Spectral E stim a to rs . I E E E Trans. Acoustics, Speech, an d S ig n a l P ro cessing, V ol. A S S P -2 8 , pp. 7 5 3 -7 5 5 , D ecem b e r. B u r g , J. P. 1967. M axim um E n trop y Sp e ctral A n aly sis, Proc. 37th M eeting o f the S o c i
ety o f Exp lo ratio n Geophysicists, O k lah om a C ity, O k la ., O cto b er. R e p rin te d in M odern Spectrum A n a ly sis, D . G . C hildres, ed. I E E E press, N ew Y o rk .
B u r g , J. P. 1968. A New A nalysis T ech n iq u e fo r T im e S eries D a ta . N A T O A d vanced Study Institu te on Signal Processin g with E m p h asis on U n d erw ater A cou stics, A ugust 1 2 -2 3 . R ep rin ted in M o d e m Spectrum A n a ly s is , D . G . C h ilders, ed ., I E E E P ress, New Y o rk . B u r g , J. P. 1972. T h e R elatio n sh ip B e tw ee n M axim um E n tro p y and M axim u m L ik elih o od S p e ctra , G e o ph ysics, V o l. 37, pp. 3 7 5 -3 7 6 , A p ril. B u r g , J. P. 1975, M axim um E n tro p y S p ectral A n aly sis, P h .D . d issertatio n , D ep artm en t o f G eo p h ysics, Stan ford U niversity, Stan fo rd , C alif., M ay . B u r r u s , C. S ., and P a r k s , T. W . 1970. T im e -D o m a in D esig n o f R e cu rsiv e D igital F ilte rs. I E E E Trans. A u d io an d Electroacoustics, V o l. 18, pp. 1 3 7 -1 4 1 , Ju n e . B u r r u s , C. S. and P a r k s , T . W . 1985. D F T / F F T an d Co n v o lu tio n A lgo rith m s, W iley, New Y o rk .
R3
B u t t e r w e c k , H. J., v a n M e e r , A. C. P., and V e r k r o o s t , G. 1984. New Second-Order Digital
Filter Sections Without Limit Cycles," IE E E Trans. Circuits and Systems, Vol. CAS-31, pp. 141-146, February.
C a d z o w , J. A . 1979. ARM A Spectral Estimation: An Efficient Closed-Form Procedure,
Proc. R A D C Spectrum Estimation W orkshop , pp. 81-97, Rome, N.Y., October.

C a d z o w , J. A. 1981. Autoregressive-Moving Average Spectral Estimation: A Model Equa
tion Error Procedure, IE E E Trans. Geoscience Remote Sensing, Vol. GE-19, pp. 24-28, January.
C a d z o w , J. A. 1982. Spectral Estimation: An Overdetermined Rational Model Equation
Approach, Proc. IE E E , Vol. 70, pp. 907-938, September.

C a n d y , J. C . 1986. Decimation for Sigma Delta Modulation, IE E E Trans. Communica
tions, Vol. COM-34, pp. 72-76, January.

C a n d y , J. C , W o o le y , B. A., and B e n ja m in , D. J. 1981. A Voiceband Codec with Digital
Filtering, IE E E Trans. Communications, Vol. COM-29, pp. 815-830, June.

C a p o n , J. 1969. High-Resolution Frequency-Wavenumber Spectrum Analysis, Proc. IE E E ,
Vol. 57, pp. 1408-1418, August.

C a p o n , J. 1983. Maximum-Likelihood Spectral Estimation, in Nonlinear M ethods o f Spec
tral Analysis, 2nd ed., S. Haykin, ed., Springer-Verlag, New York.

C a p o n , J., and G o o d m a n , N. R. 1971. Probability Distribution for Estimators of the
Frequency-Wavenumber Spectrum, Proc. IE E E , Vol. 58, pp. 1785-1786, October.

C h a n , D. S. K and R a b i n e r , L. R. 1973a. Theory o f Roundoff Noise in Cascade Realiza
tions of Finite Impulse Response Digital Filters, Bell Syst. Tech. J., Vol. 52, pp. 329-345, March.
C h a n , D. S. K., and R a b i n e r , L. R. 1973b. An Algorithm for Minimizing Roundoff Noise
in Cascade Realizations of Finite Impulse Response Digital Filters, Bell Sys. Tech. J., Vol. 52, pp. 347-385, March.
C h a n , D. S. K., and R a b i n e r , L. R. 1973c. Analysis of Quantization Errors in the Direct
Form for Finite Impulse Response Digital Filters, IE E E Trans. A udio and Electroacous tics, Vol. AU-21, pp. 354-366, August.
C h a n g , T. 1981. Suppression of Limit Cycles in Digital Filters Designed w ith One
Magnitude-Truncation Quantizer, IE E E Trans. Circuits and Systems, Vol. CAS-28, pp. 107-111, February.
C h e n , C . T. 1970. Introduction to Linear System Theory, Holt, Rinehart and Winston, New
York.
C h e n , W. Y., and S te g e w , G. R. 1974. Experiments with Maximum Entropy Power Spectra
of Sinusoids, J. Geophys. Res., Vol. 79, pp. 3019-3022, July.

C h i l d e r s , D. G ed. 1978. M odem Spectrum Analysis, IEEE Press, New York. C h o w , J. C. 1972a. On the Estimation of the Order of a Moving-Average Process, IE E E
Trans. Autom atic Control, Vol. AC-17, pp. 386-387, June.

C h o w , J. C. 1972b. On Estimating the Orders of an Autoregressive-Moving Average Pro
cess with Uncertain Observations, IE E E Trans. A utom atic Control, Vol. AC-17, pp. 707709, October.
C h o w , Y., and C a s s ig n o l, E. 1962. Linear Signal Flow Graphs and Applications, Wiley, New
York.
R4
C h u i, C. K ., a n d C h e n , G. 1987. K a lm a n Filterin g , Sp rin g er-V erlag , N ew Y o rk . C la a S e n , T . A. C. M ., M e c k l e n b r a u k e r , W. F. G ., and P e e k , J. B. H. 1 973. S eco n d -O rd er D ig ital F ilte r with O n ly O n e M ag n itu d e-T ru ncation Q u an tizer and H aving P ractically No L im it C y cles, Electron. Lett., V o l. 9, N ovem ber. C o c h r a n , W . T., C o o l e y , J. W ., F a v in , D. L ., H e lm s, H. D., K a e n e l , R. A ., L a n g , W. W ., M a l i n g , G. C., N e l s o n , D. E ., R a d e r . C. E and W e lc h , P. D. 1967. W h a t Is the F ast F o u rier T ran sfo rm , I E E E Trans. A u d io an d Electroacoustics, V o l. A U -1 5 , pp. 4 5 -5 5 , Ju n e. C o n s t a n t i n i d e s , A. G. 1967. F req u en cy T ran sfo rm atio n s for D ig ital F ilte r s, Electron. Lett., V o l. 3, pp. 4 8 7 -4 8 9 , N ovem ber. C o n s t a n t i n i d e s , A. G. 1968. F req u en cy T ran sfo rm atio n s for D ig ital F ilte r s , Electron. Lett., V ol. 4 , pp. 1 1 5 -1 1 6 , A pril. C o n s t a n t i n i d e s , A. G. 1970. Sp ectral T ran sfo rm atio n s fo r D ig ital F ilte r s , Proc. I E E , V o l. 117, pp. 1 5 8 5 -1 5 9 0 . A ugust. C o o le y . J. W ., and T u k e y , J. W. 1965. A n A lgorithm for the M ach ine C om putation o f C om plex F o u rier S e rie s , Math. C o m p ., V ol. 19. pp. 2 9 7 -3 0 1 . A pril. C o o le y , J. W L e w is, P., and W e lc h , P. D. 1967. H istorical N otes on th e F ast F o u rier T ra n sfo rm , I E E E Trans. A u d io and Electroacoustics, V o l. A U -1 5 , pp. 7 6 -7 9 , Ju n e . C o o le y , J. W ., L e w is, P.. and W e lc h , P. D. 1969. T h e F ast F o u rier T ran sform and Its A p p licatio n s, I E E E Trans. Ed u catio n . V ol. E -1 2 pp. 2 7 -3 4 , M arch. C r o c h i e r e , R. E. 1977. O n the D esign o f Sub~band C od ers for Low B it R a te Speech C o m m u n ication ," B e l l S y s t. T e c h . J., V ol. 56, pp. 7 4 7 -7 1 1 , M a v -Ju n e . C r o c h i e r e , R. E. 1981. Sub-band C oding, B e ll Syst. Tech. J.. V o l. 6 0 . pp. 1 6 3 3 -1 6 5 4 , S ep tem b er. C r o c h i e r e . R. E., and R a b i n e r , L. R. 1975. O ptim um F I R D ig ital F ilte r Im p lem en tatio n s for D ecim atio n , In te rp o latio n , and N arrow band F ilterin g , I E E E Trans. A co ustics , Speech and S ig n a l Processing, V o l. A S S P -2 3 , pp. 444 456, O cto b er. C r o c h i e r e , R. E., and R a b i n e r , L. R. 1976. F u rth e r C o nsid eratio n s in th e D esign o f D ecim ato rs and In te rp o la to rs, I E E E Trans. A coustics, Speech, and S ig n a l Processing, V ol. A S S P -2 4 , pp. 2 9 6 -3 1 1 , August. C r o c h i e r e , R. E and R a b i n e r , L. R. 1981. In terp o lation and D ecim a tio n o f D ig ital S ig n a lsA T u to rial R ev iew , Proc. I E E E , V ol. 69, pp. 3 0 0 -3 3 1 , M arch . C r o c h i e r e , R. E., and R a b i n e r , L. R . 1983. M ultirate D ig ita l S ig n a l P ro cessing, P ren tice H all, E n gelw ood C liffs. N .J. D a n ie ls , R. W , 1974. A p p ro xim atio n M ethods fo r the D e sign o f Passive, A ctiv e an d D ig ita l Filters, M cG raw -H ill, New Y o rk .
D a v e n p o r t , W . B J r . 1970. P ro b ab ility and R an d o m Processes: A n Introduction fo r A p p lie d
Scientists an d Engineers. M cG raw -H ill, N ew Y o rk .

D a v is , H. F. 1963. F o u rie r Series an d O rtho go n al Functions. A llyn and B a c o n , B o sto n . D e c z k y , A . G. 1972. Sy n th esis o f R ecu rsiv e D ig ital F ilters U sing th e M inim um p -E r ro r
C riterio n , I E E E Trans. A u d io an d Electroacoustics, V o l. A U -2 0 , pp. 2 5 7 -2 6 3 , O cto b er.

D e l s a r t e , P., and G e n i n , Y . 1986. T h e Split Levin son A lg o rith m , I E E E Trans. Acoustics,
Speech, an d Sig n al P ro ce ssin g , V o l. A S S P -3 4 , pp. 4 7 0 -4 7 8 , Ju n e .

D e ls a rte ,
R5
P., G e n i n , Y., and K a m p , Y. 1978. Orthogonal Polynomial Matrices on the Unit Circle, IE E E Trans. Circuits and Systems. Vol. CAS-25, pp. 149-160. January.
D e R u s so , P. M., R o y , R. J., and C lo s e , C. M. 1965. State Variables fo r Engineers, Wiley, New
York.
D u h a m e l, P. 1986. Implementation of Split-Radix FFT Algorithms for Complex, Real, and
Real-Symmetric Data, IE E E Trans. Acoustics, Speech, and Signal Processing , Vol. ASSP34, pp. 285-295, April.
D u h a m e l, P., and H o l l m a n n , H. 1984. Split-Radix FFT Algorithm, Electron. Lett., Vol. 20,
pp. 14-16, January.

D u r b in , J. 1959. Efficient Estimation of Parameters in Moving-Average Models, Biome-
trika, Vol. 46, pp. 306-316.

D w i g h t, H. B. 1957. Tables o f Integrals and Other Mathematical Data , 3rd ed., Macmillan,
New York.
D ym , H., a n d M c K e a n , H . P. 1972. Fourier Series and Integrals, Academic, New York. E b e r t , P. M M a z o , J. E., and T a y l o r , M . G. 1969. Overflow Oscillations in Digital Filters,
Bell Syst. Tech . J., Vol. 48, pp. 2999-3020, November.

F am , A. T., and B a r n e s , C. W. 1979. Non-minimal Realizations of Fixed-Point Digital Filters
That Are Free of All Finite Wordlength Limit Cycles, IE E E Trans. Acoustics, Speech, and Signal Processing , Vol. ASSP-27, pp. 149-153, April.
F e t t w e i s , A. 1971. Some Principles of Designing Digital Filters Imitating Classical Filter
Structures, IE E E Trans. Circuit Theory , Vol. CT-18, pp. 314-316, March.

F l e t c h e r , R and P o w e l l , M. J. D. 1963. A Rapidly Convergent Descent Method for Minimization, Comput. J., Vol. 6 , pp. 163-168. F o u g e r e , P. F., Z a w a l ic k , E. J., a n d R a d o s k i, H. R. 1976. Spontaneous Line Splitting in
Maximum Entropy Power Spectrum Analysis, Phys. Earth Planet, Inter., Vol. 12, 201207, August.
F r i e d l a n d e r , B. 1982a. Lattice Filters for Adaptive Processing, Proc. IE E E , Vol. 70,
pp. 829-867, August.

F r i e d l a n d e r , B. 1982b. Lattice Methods for Spectral Estimation, Proc. IE E E , Vol. 70,
pp. 990-1017, September. Fuchs, J. J. 1988. "Estimating the Number of Sinusoids in Additive White Noise, IE E E Trans. Acoustics, Speech, and Signal Processing , Vol. ASSP-36, pp. 1846-1853, December.
G a n t m a c h e r , F. R. 1960. The Theory o f Matrices, Vol. I., Chelsea, New York. G a r d n e r , W. A. 1987. Statistical Spectral Analysis: A Nonprobabilistic Theory, Prentice Hall,
Englewood Cliffs, N.J.

G a r l a n , C., and E s te b a n , D. 1980. 16 Kbps Real-Time QMF Sub-band Coding Implemen
tation, Proc. 1980 International Conference on Acoustics, Speech, and Signal Processing, pp. 332-335, April.
G e r o n im u s ,
L. Y. 1958. Orthogonal Polynomials (in Russian) (English translation sultant's Bureau, New York, 1961).
by
Con
G e r s c h , W., and S h a r p e , D. R. 1973. Estimation of Power Spectra with Finite-Order A u
toregressive Models, IE E E Trans. A utom atic Control, Vol. AC-18, pp. 367-369, August.
G ib b s ,
A. J. 1969. An Introduction to Digital Filters, A u s t Tellcommun. Res., Vol. 3, pp. 3-14, November.
R6
G ib b s ,
A. J. 1970. The Design of Digital Filters, A use. Telecommun. Res., Vol. 4, pp. 29-34, May. 1968. An Algorithm for the Evaluation of Finite Trigonometric Series, A m . Math. Monthly, Vol. 65, pp. 34-35, January.
G o e rtz e l, G.
G o h b e rg ,
I., ed. 1986. 1. Schiir Methods in Operator Theory and Signal Processing, Birkhauser Verlag, Stuttgart, Germany.
G o ld ,
B., and J o r d a n , K. L J r . 1986. A Note on Digital Filter Synthesis, Proc. IE E E , Vol. 56, pp. 1717-1718, October. B., and - J o r d a n , K. L., Jr. 1969. A Direct Search Procedure for Designing Finite Duration Impulse Response Filters, IE E E Trans. A u d io and Electroacoustics, Vol. AU17, pp. 33-36, March.
G o ld ,
G o ld ,
B., and R a d e r , C. M . 1966. Effects of Quantization Noise in Digital Filters. Proc. AF1PS 1966 Spring Joint Computer Conference, Vol. 2 8 , pp. 2 1 3 -2 1 9 . B., and
R a d e r,
G o ld ,
C. M. 1969. Digital Processing o f Signals, McGraw-Hill, New York.
G o ld e n ,
R. M., and K a i s e r , J. F . 1964. Design of Wideband Sampled Data Filters, Bell Syst. Tech. J., Vol. 43, pp. 1533-1546, July. I. J. 1971. The Relationship Between Two Fast Fourier Transforms, IE E E Trans. Computers, Vol. C-20, pp. 310-317. J., ed. 1975. Frequency Synthesis: Techniques and Applications, IEEE Press,
G ood,
G o r s k i-P o p ie l,
New York.
G rau p e,
D., K r a u s e , D. J., and M o o r e , J. B. 1975. Identification of Autoregressive-Moving Average Parameters of Time Series, IE E E Trans. Autom atic Control, Vol. AC-20, pp. 104-107, February. A. H., and M a r k e l , J. D. 1973. Digital Lattice and Ladder Filter Synthesis, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-21, pp. 491-500, December.
G ray ,
G ra y ,
A. H., and M a r k e l , J. D. 1976. A Computer Program for Designing Digital Elliptic Filters, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-24, pp. 529 538, December. R. M. 1990. Source Coding Theory, Kluwer, Boston, MA. O,, and S z e g o , G . 1958. Teoplitz Forms and Their Applications, University of California Press, Berkeley, Calif. L. J. 1975. Rapid Measurements of Digital Instantaneous Frequency, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-23, pp. 207-222, April. E. A. 1957. Synthesis o f Passive Networks, Wiley, New York. S. C. 1966. Transform and State Variable M ethods in Linear Systems, Wiley, New R. W. 1962. Numerical Methods fo r Scientists and Engineers, McGraw-Hill, New
G ray ,
G re n a n d e r,
G r iff ith s ,
G u ille m in , G u p ta ,
York.
H a m m in g ,
York.
H a y k in ,
S. 1991. Adaptive Filter Theory , 2nd ed., Prentice Hall, Englewood Cliffs, N J.
H e l m e , B .,
and N i k i a s , C. S. 1985. Improved Spectrum Performance via a Data-Adaptive Weighted Burg Technique, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-33, pp. 903-910, August,

H e lm s , H .
R7
D. 1967. Fast Fourier Transforms Method of Computing Difference Equations and Simulating Filters, IE E E Trans. A udio and Electroacoustics, Vol. AU-15, pp. 85-90. June.
H e lm s , H . D .
1968. Nonrecursive Digital Filters: Design Methods for Achieving Specifi cations on Frequency Response, IE E E Trans. A u d io and Electroacoustics, Vol. AU-16, pp. 336-342, September.
H e ls tro m ,
C. W. 1990. Probability and Stochastic Processes fo r Engineers , 2nd ed., Macmil lan, New York. R. W. 1980. The Cause of Line Splitting in Burg Maximum-Entropy Spectral Analysis, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-28, pp. 692701, December. O. 1970. Design of Nonrecursive Digital Filters with Linear Phase, Electron. Lett., Vol. 6 , pp. 328-329, November.
H e r r in g ,
H erm an n ,
O., and S c h u e s s l e r , H . W. 1970a. Design of Nonrecursive Digital Filters with Minimum Phase, Electron. Lett., Vol. 6 , pp. 329-330, November. H e r r m a n n , O., and S c h u e s s l e r , H . W. 1970b. On the Accuracy Problem in the Design of Nonrecursive Digital Filters, Arch. Elek. Ubertragung, Vol. 24, pp. 525-526. H e r r m a n n , O., R a b i n e r , L. R ., and C h a n , D. S. K. (1973). Practical Design Rules for Optimum Finite Impulse Response Lowpass Digital Filters, Bell Syst. Tech. J., Vol. 52, pp. 769-799. July-August.
H erm an n , H ild e b ra n d ,
F. B. 1952. M ethods o f A pplied Mathematics, Prentice Hall, Englewood Cliffs,
N.J.
H o f s te tte r ,
E., O p p e n h e i m , A. V., and S i e g e l , J. 1971. A New Technique for the Design of Nonrecursive Digital Filters, Proc. 5th A nnual Princeton Conference on Information Sciences and Systems , pp. 64-72. 1964. The Theory o f Matrices in Numerical Analysis, Blaisdell, Waltham, Mass.
H o u s e h o l d e r , A . S.
S. Y, 1977. Minimum Uncorrelated Unit Noise in State Space Digital Filter ing. IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-25, pp. 273-281. August. J a c k s o n , L. B. 1969. An Analysis of Limit Cycles Due to Multiplication Rounding in Recursive Digital (Sub) Filters, Proc. 7th A nnual Allerton Conference on Circuit and System Theory, pp. 69-78.
H w ang, Jack so n ,
L. B . 1970a. On the Interaction of Roundoff Noise and Dynamic Range in Digital Filters, Bell Syst. Tech. J., Vol. 49, pp. 159-184, February.
L. B. 1970b. Roundoff Noise Analysis for Fixed-Point Digital Filters Realized in Cascade or Parallel Form, IE E E Trans. A udio and Electroacoustics, Vol. AU-18, pp. 107-122, June. J a c k s o n , L. B. 1976. Roundoff Noise Bounds Derived from Coefficients Sensitivities in Digital Filters, IE E E Trans. Circuits and Systems, Vol. CAS-23, pp. 481-485, August. J a c k s o n , L. B. 1979. Limit Cycles on State-Space Structures for Digital Filters, IE E E Trans. Circuits and Systems, Vol. CAS-26, pp. 67-68, January.
Jackson, Ja c k s o n , L.
B L i n d g r e n , A. G., and K jm , Y. 1979. Optimal Synthesis of Second-Order State-3pace Structures for Digital Filters, IE E E Trans. Circuits and Systems, Vol. CAS-26, pp. 149-153, March.
R8
J a h n k e . E .. a n d E m d e , F. J a in .
1945. Tables o f Functions , 4th ed.. Dover, New York.
V. K.. and C r o c h i e r e , R. E. 1984. Quadrature Mirror Filter Design in the Time Domain," IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-32, pp. 353361, April.
W a tts ,
J e n k in s ,
G. M., and San Francisco.
D. G. 1968. SpectraI Analysis and Its Applications, Holden-Day,
J o h n s t o n , J.
D. 1980. A Filter Family Designed for Use in Quadrature Mirror Filter Banks, IE E E International Conference on Acoustics, Speech, and Signal Processing, pp. 291-294, April. 1982. The Application of Spectral Estimation Methods to Bearing Estima tion Problems, Proc. IE E E , Vol. 70, pp. 1018-1028, September. R. H. 1976. Autoregression Order Selection, Geophysics , Vol. 41, pp. 771-773, August. 1964. Theory and Applications o f the z-Transform M ethod , Wiley, New York.
Jo h n so n , D. H.
Jo n es,
J u r y . E . I. K a ila th ,
T. 1974. A View of Three Decades of Linear Filter Theory." IE E E Trans. Infor mation Theory, Vol. IT-20, pp. 146-181, March. T. 1981. Lectures on Wiener and Kalman Filtering , 2nd Printing, Springer-Verlag, New York.
K a ila th ,
K a ila th .
T. 1985. Linear Estimation for Stationary and Near-Stationary Processes. in M odem Signal Processing, T. Kailath, ed.. Hemisphere Publishing Corp., Washington, D.C. T. 1986. A Theorem of I. Schur and Its Impact on Modern Signal Processing," in Gohberg (1986).
K a ila th ,
K a ila th ,
T.. V i e i r a , A. C. G, and M o r f , M. 1978. Inverses of Toeplitz Operators, Innova tions, and Orthogonal Polynomials. S IA M Rev., Vol. 20, pp. 1006-1019.
K a is e r . J. F. 1963. Design Methods for Sampled Data Filters." Proc. First Allerton Con ference on Circuit System Theory, pp. 2 2 1 -2 3 6 , November.
K a is e r,
J. F. 1966. Digital Filters, in System Analysis by Digital Computer, F. F. Kuo and J. F. Kaiser, Eds., Wiley, New York. N., and T h e o d o r i d i s , S. 1987. Fast Adaptive Least-Squares Algorithms for Power Spectral Estimation, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-35, pp. 661-670, May.
K a lo u p ts id is ,
K a lm a n ,
R. E. 1960. A New Approach to Linear Filtering and Prediction Problems." Trans. A SM E , J. Basic Eng., Vol. 82D, pp. 35-45, March. R. E., and B u c y , R. S. 1961. New Results in Linear Filtering Theory, Trans. ASM E, J. Basic Eng., Vol. 83, pp. 95-108.
K a lm a n ,
K ashyap.
R. L. 1980. Inconsistency of the AIC Rule for Estimating the Order of Autore gressive Models, IE E E Trans. Autom atic Control, Vol. AC-25, pp. 996-998, October J., and B r u z z o n e , S. P. 1979. Order Determination for Autoregressive Spectral Es timation." Record o f the 1979 R A D C Spectral Estimation W orkshop, pp. 139-145, Griffin Air Force Base, Rome, N.Y.
K aveh,
K aveh,
M., and L i p p e r t , G. A . 1983. An Optimum Tapered Burg Algorithm for Linear Prediction and Spectral Analysis, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-31, pp. 438-444, April.

K ay,
R9
S. M. 1980. A New ARMA Spectral Estimator, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-28, pp. 585-588, October. S. M. 1988. M odern Spectral Estimation , Prentice Hall, Englewood Cliffs, N. J.
K ay,
K ay , S, M ., and M a r p l e , S. L ., J r . 1979. Sources of and Remedies for Spectral Line Splitting in Autoregressive Spectrum Analysis, Proc. 1979 IC A SSP , pp. 1 5 1 -1 5 4 .
K ay,
S. M ., and M a r p l e , S. L., J r . 1981. Spectrum Analysis: A Modern Perspective, Proc. IE E E , Vol. 69, pp. 1380-1419, November. S. B., ed. 1986. M odem Spectrum Analysis II, IEEE Press, New York. B., and O l c a y t o , E . M. 1968. Coefficient Accuracy and Digital Filter R e sponse, IE E E Trans. Circuit Theory , Vol. CT-15, pp. 31-41, March.
K e s le r,
K n o w le s , J.
K rish n a ,
H. 1988. New Split Levinson, Schur, and Lattice Algorithms for Digital Signal Processing, Proc. 1988 International Conference on Acoustics, Speech, and Signal Pro cessing, pp. 1640-1642, New York, April.
K ro m er,
R. E. 1969. Asymptotic Properties of the Autoregressive Spectral Estimator, Ph.D. dissertation, Department of Statistics, Stanford University, Stanford, Calif. and Hu, Y . H. 1983. A Highly Concurrent Algorithm and Pipelined Architec ture for Solving Toeplitz Systems, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-31, pp. 66-76, January. S. Y., W h i t e h o u s e , H. J ., and K a i l a t h , T., eds. 1985. V L S I and Modern Signal Processing , Prentice Hall, Englewood Cliffs, N.J.
K u n g , S. Y
K ung,
L aco ss,
R. T. 1971. Data Adaptive Spectral Analysis Methods, Geophysics, Vol. 36, pp. 661-675, August. S. W ., and M c C l e l l a n , J. H. 1980. Frequency Estimation with Maximum Entropy Spectral Estimators, IE E E Trans. Acoustics, Speech, and Signal Processing , Vol. ASSP28, pp. 716-724, December.
L ang,
L e v in so n , N.
1947. The Wiener RMS Error Criterion in Filter Design and Prediction, J. Math. Phys., Vol. 25, pp. 261-278. H and
L essm an,
L evy, L iu ,
F. 1961. Finite Difference Equations, Macmillan, New York.
B. 1971. Effect of Finite Word Length on the Accuracy of Digital Filters A Review, IE E E Trans. Circuit Theory , Vol. CT-18, pp. 670-677, November. J. 1975. Linear Prediction: A Tutorial Review, Proc. IE E E , Vol. 63, pp. 561 580, April.
M a k h o u l,
1978. A Class of All-Zero Lattice Digital Filters: Properties and Applica tions, IE E E Trans. Acoustics, Speech, and Signal Processing , Vol. ASSP-26, pp. 304-314, August. M arkel, J. D., and G r a y , A. H., Jr. 1976. Linear Prediction o f Speech, Springer-Verlag, New York.
M a k h o u l, J. M a rp le ,
S. L ., J r . 1980. A New Autoregressive Spectrum Analysis Algorithm. IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-28, pp. 441-454, August.
M a rp le ,
S. L., J r . 1987. Digital Spectral Analysis with Applications, Prentice Hall, Englewood Cliffs, N J. a r z e t t a , T . L . 1983. A New Interpretation for Capons Maximum Likelihood Method of Frequency-Wavenumber Spectral Estimation, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-31, pp. 445-449, April.
R10
M a rz e tta ,
T. L., and Lang, S. W. 1983. New Interpretations for the MLM and DASE Spectral Estimators, Proc. 1983 IC A SSP , pp. 844-846, Boston, April.
M a rz e tta ,
T. L ., and L a n g , S. W. 1984. Power Spectral Density Bounds. IE E E Trans. Information Theory , Vol. IT-30, pp. 117-122, January. S. J., and New York.
Z im m e rm a n ,
M aso n ,
H. J. 1960. Electronic Circuits, Signals and Systems, Wiley,
M c C le lla n ,
J. H. 1982. Multidimensional Spectral Estimation, Proc. IE E E , Vol. 70, pp. 1029-1039, September.
M cD onough,
R. N. 1983. Application of the Maximum-Likelihood Method and the Maxi mum Entropy Method to Array Processing, in Nonlinear M ethods o f Spectral Analysis, 2nd ed., S. Haykin, ed., Springer-Verlag, New York.
M c G ille m , C .
D., and C o o p e r , G . R. 1984. Continuous and Discrete Signal and System Analysis, 2nd ed.. Holt Rinehart and Winston, New York. J. E. 1969. Stochastic Optimal Linear Estimation and Control. McGraw-Hill, New York.
M e d itc h ,
M ills , W .
L., M u l l i s , C . T ., and R o b e r t s , R . A . 1981. Low Roundoff Noise and Normal Realizations of Fixed-Point IIR Digital Filters. IE E E Trans. Acoustics, Speech, and Signal processing, Vol. ASSP-29, pp. 893-903, August.
M o r f , M V ie ir a ,
A., and L e e , D. T. 1977. Ladder Forms for Identification and Speech Processing, Proc. 1977 IE E E Conference Decision and Control, pp. 1074-1078, New Orleans, La., December. C. T ., and R o b e r t s , R . A . 1976a. Synthesis of Minimum Roundoff Noise Fixed* Point Digital Filters, IE E E Trans. Circuits and Systems , Vol. CAS-23, pp. 551-561, September.
M u llis ,
M u llis ,
C. T and R o b e r t s , R . A . 1976b. Roundoff Noise in Digital Filters: Frequency Transformations and Invariants, IE E E Trans. Acoustics, Speech and Signal Processing , Vol. ASSP-24, pp. 538-549, December
M u s ic u s , B .
1985. Fast M L M Power Spectrum Estimation from Uniformly Spaced Corre lations, IE E E Trans. Acoustics, Speech, and Signal Proc., Vol. ASSP-33, pp. 1333-1335, October. W. I. 1981. Extension to the Maximum Entropy Method III, Proc. 1st A SSP Workshop on Spectral Estimation, pp. 1.7.1 1.7.6, Hamilton, Ontario, Canada, August.
A
N ew m an,
N i k i a s , C . L .,
and R a g h u v e e r , M . R . 1987. Bispectrum Estimation: cessing Framework, Proc. IE E E , Vol. 75, pp. 869-891, July.
Digital Signal Pro
N ik ia s ,
C. L and S c o t t , P. D. 1982. Energy-Weighted Linear Predictive Spectral Es timation: A New Method Combining Robustness and High Resolution. IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-30, pp. 287-292, April. A. H. 1976. Spectral Analysis of a Univariate Process with Bad Data Points, via Maximum Entropy and Linear Predictive Techniques, N U SC Technical Report TR-5303 , New London, Conn., March.
N u tta ll,
N y q u is t,
H. 1928. Certain Topics in Telegraph Transmission Theory, Trans. A IE E , Vol. 47, pp. 617-644, April. A. V. 1978. Applications o f Digital Signal Processing, Prentice Hall, Englewood Cliffs, N J.
O p p e n h e im ,

O p p e n h e im ,
R11
A. V., and S c h a f e r , R. W. 1989. Discrete-Time Signal Processing, Prentice Hall, Englewood Cliffs, N J.
O p p e n h e im , A. V., and W e in s te in , C. W . 1972. Effects of Finite Register Length in Digital
Filters and the Fast Fourier Transform, Proc. IE E E , Vol. 60, pp. 957-976. August.
O p p e n h e im , A. V ., and W i l l s k y , A. S. 1983. Signals and Systems, Prentice Hall, Englewood
Cliffs, N J.
P a p o u lis , P a p o u lis ,
A. 1962 The Fourier Integral and Its Applications , McGraw-Hall, New York.
A. 1984. Probability, R andom Variables, and Stochastic Processes, 2nd ed., McGraw-Hill, New York. Circuit Theory, Vol. CT-18, pp. 687-696, November.
P a r k e r , S. R., and H ess, S. F. 1971. Limit-Cycle Oscillations in Digital Filters, IE E E Trans. P a r k s , T . W., and M c C l e l l a n , J. H. 1972a. Chebyshev-Approximation for Nonrecursive Digital Filters with Linear P h a s e , IE E E Trans. Circuit Theory, Vol. C T -1 9 , pp. 1 8 9 -1 9 4 , M a rc h .
P a rk s,
T. W., and M c C l e l l a n , J. H. 1972b. A Program for the Design of Linear Phase Finite Impulse Response Digital Filters, IE E E Trans. A udio and Electroacoustics, Vol. AU-20, pp. 195-199, August.
P a rze n ,
E. 1957. On Consistent Estimates of the Spectrum of a Stationary Time Series, A m . Math. Stat., Vol. 28, pp. 329-348. E. 1974. Some Recent Advances in Time Series Modeling, IE E E Trans. A u to matic Control, Vol. AC-19, pp. 723-730, December.
P a rze n ,
P eacock,
K. L, and T r e i t e l , S. 1969. Predictive Deconvolution-Theory and Practice, Geo physics, Vol. 34, pp. 155-169. Z., J r . 1987. Probability, Random Variables, and Random Signal Principles, 2nd ed., McGraw-Hill, New York.
P e e b le s , P.
P isa re n k o ,
V. F. 1973. The Retrieval of Harmonics from a Convariance Function, Geo phys. J. R. Astron. Soc., Vol. 33, pp. 347-366. 1981. Autoregressive M ethods o f Spectral Analysis, E.E. degree thesis, Depart ment of Electrical and Computer Engineering, Northeastern University, Boston, May.
P o o le , M . A.
P r ic e , R. 1990. Mirror FFT and Phase FFT Algorithms, unpublished work, Raytheon
Research Division, May.

P r o a k is ,
J. G. 1995. Digital Communications, 3rd ed., McGraw-Hill, New York.
R a b i n e r , L. R., and S c h a e f e r , R. W. 1974a. On the Behavior of Minimax Relative Error
FIR Digital Differentiators, Bell Syst. Tech. J., Vol. 53, pp. 333-362, February.
R a b i n e r , L. R ., and S c h a f e r , R . W . 1974b. On th e Behavior of Minimax F I R Digital Hilbert
Transformers, Bell SysL Tech. J., Vol. 53, pp. 363-394, February.
R a b i n e r , L. R ., and S c h a f e r , R. W. 1978. Digital Processing o f Speech Signals, Prentice Hall,
Englewood Cliffs, N J.
R a b i n e r , L. R., S c h a f e r , R . W ., and R a d e r , C. M. 1969. The Chirp ^-Transform Algorithm
and Its Applications, Bell S yst Tech. J., Vol. 48, pp. 1249-1292, May-June.
R a b i n e r , L . R ., G o l d , B., a n d M c G o n e g a l, C . A . 1970. An Approach to the Approxima
tion Problem for Nonrecursive Digital Filters, IE E E Trans. A u d io and Electroacoustics, Vol. AU-18, pp. 83-106, June.
R12
R a b i n e r , L . R ., M c C l e l l a n , J . H ., a n d P a r k s , T .
W. 1975. F I R Digital Filter Design Tech niques Using Weighted Chebyshev Approximation, Proc. IE E E , Vol. 63, pp. 595-610, April.
R a d e r, C. M .
1970. An Improved Algorithm f o r High-Speed Auto-correlation with Appli cations to Spectral Estimation, IE E E Trans. A udio and Electroacoustics, Vol. AU-18, pp. 439-441, December. C. M., and B r e n n e r , N. M. 1976. A New Principle for Fast Fourier Transformation, IE E E Trans. Acoustics, Speech and Signal Processing, Vol. ASSP-24, pp. 264-266, June. C. M., and G o l d , B. 1967a. Digital Filter Design Techniques in the Frequency Domain, Proc. IE E E , Vol. 55, pp. 149-171, February.
R a d e r,
R ad er,
R ad er,
C. M., and G o l d , B. 1967b. Effects of Parameter Quantization on the Poles of a Digital Filter, Proc. IE E E , Vol. 55, pp. 688-689, May.
R a m s ta d , T. A.
1984. Digital Methods for Conversion Between Arbitrary Sampling Fre quencies, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-32, pp. 577591, June. 1957. General Computational M ethods o f Chebyshev A pproxim ation, Atomic Energy Translation 4491, Kiev, USSR.
R e m e z, E. Y a.
R js s a n e n .
J. 1983. A Universal Prior for the Integers and Estimation by Minimum D e scription Length, Ann. Stat., Vol. 11, pp. 417-431. and
M u llis ,
R o b e r t s , R . A .,
C.
T,
1987. Digital Signal Processing, Addison-Wesley, Reading,
Mass.
R o b in s o n , R o b in so n ,
E. A. 1962. R andom Wavelets and Cybernetic Systems, Charles Griffin, London.
E. A. 1982. A Historical Perspective of Spectrum Estimation, Proc. IE E E , Vol. 70, pp. 885-907, September.
R o b in s o n ,
E. A., and T r e i t e l , S. 1978. Digital Signal Processing in Geophysics, in A p plications o f Digital Signal Processing, A. V. Oppenheim, ed., Prentice Hall, Englewood Cliffs, N.J. and wood Cliffs, N.J.
T r e i t e l , S.
R o b i n s o n , E . A .,
1980. Geophysical Signal Analysis, Prentice Hall, Engle
R o y , R ., P a u l r a j ,
A., and K a i l a t h , T. 1986. ESPRIT: A Subspace Rotation Approach to Estimation of Parameters of Cisoids in Noise, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-34, pp. 1340-1342, October.
S a fra n e k ,
R. J ., M a c K a y , K . J a y a n t , N. W., and K im , T. 1988. Image Coding Based on Selective Quantization of the Reconstruction Noise in the Dominant Sub-band, Proc. 1988 IE E E International Conference on Acoustics, Speech, and Signal Processing, pp. 765768, April.
S a k a i,
H. 1979. Statistical Properties of AR Spectral Analysis, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-27, pp. 402-409, August. I. W ., and K a i s e r , J. F. 1972. A Bound on Limit Cycles in Fixed-Point Im plementations of Digital Filters, IE E E Trans. A udio and Electroacoustics, Vol. AU-20, pp. 110-112, June. E. H,, and A l e x a n d e r J. T. 1978. High Resolution Spectral Analysis of Sinusoids in Correlated Noise, Proc. 1978 IC A SSP, pp. 349-351, Tulsa, Okla., April 10-12.
S a n d b e rg ,
S a to riu s ,

S c h a f e r , R.
R13
W., and R a b i n e r , L. R . 1973. A Digital Signal Processing Approach to Inter polation, Proc. IE E E , Vol. 61, pp. 692-702, June. H., and G o c k l e r , H. 1981. A Comprehensive Survey of Digital Transmul tiplexing Methods, Proc. IE E E , Vol. 69, pp. 1419-1450. R. D. 1981. A Signal Subspace Approach to Multiple Emitter Location and Spectral Estimation, Ph.D. dissertation, Department of Electrical Engineering, Stanford University, Stanford, Calif., November. R. D. 1986. Multiple Emitter Location and Signal Parameter Estimation, IE E E Trans. A ntennas and Propagation, Vol. AP 34, pp. 276-280, March.
S c h e u e rm a n n ,
S c h m id t,
S c h m id t,
S c h o tt,
J. P., and M c C l e l l a n , J. H. 1984. Maximum Entropy Power Spectrum Estimation with Uncertainty in Correlation Measurements, IE E E Trans. Acoustics, Speech, and Signal Processing , Vol. ASSP-32, pp. 410-418, April.
S ch u r,
I. 1917. On Power Series Which Are Bounded in the Interior of the Unit Circle, J. Reine Angew. Math., Vol. 147, pp. 205-232, Berlin. For an English translation of the paper, see Gohberg 1986. 1898. On the Investigation of Hidden Periodicities with Application to a Supposed Twenty-Six-Day Period of Meteorological Phenomena, Terr. Mag.. Vol. 3, pp. 13-41, March.
S c h u s te r , S ir A r t h u r .
S e d l m e y e r , A .,
and F e t t w e i s . A . 1973. Digital Filters with True Ladder Configuration. Int. J. Circuit Theory A ppL, Vol. 1, pp. 5-10. March.
S h a n k s , J. L. 1967. Recursion Filters for Digital Processing," Geophysics, Vol. 32, p p . 3 3 5 1 , February.
Shannon,
C. E. 1949. Communication in the Presence of Noise, Proc. IRE , pp. 10-21,
January.
S h e in g o ld ,
D. H., ed. 1986. Analog-Digital Conversion H andbook , Prentice Hall, Englewood Cliffs, N.J. W. M. 1986. Circuits, Signals and Systems , McGraw-Hill, New York.
S ie b e rt,
S in g le to n ,
R. C. 1967. A Method for Computing the Fast Fourier Transform with Auxil iary Memory and Limit High speed Storage. IE E E Trans. A udio and Electroacoustics, Vol. AU-15, pp. 91-98, June.
S in g le to n ,
R. C. 1969. An Algorithm for Computing the Mixed Radix Fast Fourier Trans form, IE E E Trans. A u dio and Electroacoustics, Vol. AU-17, pp. 93-103, June.
S m ith ,
M. J. T., and B a r w e l l , T. P. 1984. A Procedure for Designing Exact Reconstruc tion Filter Banks for Tree Structured Subband Coders, Proc. 1984 IE E E International Conference on Acoustics, Speech, and Signal Processing, pp. 27.1.1-27.1.4, San Diego, March.
S m ith ,
M. J. T., and E d d i n s , S. L. 1988. Subband Coding of Images with Octave Band Tree Structures, Proc. 1987 IE E E International Conference on Acoustics, Speech, and Signal Processing, pp. 1382-1385, Dallas, April. K. 1965. The Equivalence of Digital and Analog Signal Processing, Inf. Control, Vol. 8 , pp. 455-467, October.
S te ig litz ,
S te ig litz ,
K. 1970. Computer-Aided Design of Recursive Digital Filters, IE E E Trans. A u d io and Electroacoustics, Vol. AU-18, pp. 123-129, June.
R14
S to c k h a m ,
T. G. 1966. High Speed Convolution and Correlation, 1966 Spring Joint Com puter Conference, A F IP S Proc., Vol. 28, pp. 229-233. J. E. 1957. Passive Network Synthesis, McGraw-Hill, New York. P. 1986. Symmetric FFTs, Mathematics o f Computation, Vol. 47, pp. 323-
S to re r.
S w a r z tr a u b e r.
346. July.
S w in g le r,
D. N. 1979a. A Comparison Between Burgs Maximum Entropy Method and a Nonrecursive Technique for the Spectral Analysis of Deterministic Signals, J. Geophys. Res.. Vol. 84, pp. 679-685, February.
S w in g le r,
D. N. 1979b. A Modified Burg Algorithm for Maximum Entropy Spectral Anal ysis, Proc. IE E E , Vol. 67, pp. 1368-1369, September.
S w in g le r,
D. N. 1980. Frequency Errors in MEM Processing, IE E E Tram. Acoustics , Speech, and Signal Processing, Vol. ASSP-28, pp. 257-259, April. G. 1967. Orthogonal Polynomials, 3rd ed., Colloquium Publishers, no. 23, American Mathematical Society, Providence, R.l.
Szego,
T h o rv a ld s e n , T.
1981. A Comparison of the Least-Squares Method and the Burg Method for Autoregressive Spectral Analysis, IE E E Trans. A ntennas and Propagation. Vol. AP29, pp. 675-679, July. H. 1975. Autoregressive Model Fitting with Noisy Data by Akaike's Information Criterion," IE E E Trans. Information Theory, Vol. IT-21, pp. 4 7 6 - 4 8 0 , July. H. 1977. More on Autoregressive Model Fitting with Noise Data b y Akaike's Infor mation Criterion," IE E E Trans. Information Theory , Vol. IT-23, pp. 409-410. May. S. A. 1976. Introduction to Discrete-Time Signal Processing. Wiley, New York.
Tong,
Tong,
T re tte r. T u fts ,
D. W.. and K u m a r e s a n , R. 1982. Estimation of Frequencies of Multiple Sinusoids: Making Linear Prediction Perform Like Maximum Likelihood, Proc. IE E E , Vol. 70, pp. 975-989. September.
U lr y c h .
T. J., and B i s h o p , T. N. 1975. Maximum Entropy Spectral Analysis and Autore gressive Decomposition,7 ' Rev. Geophys. Space Phys., Vol. 13, pp. 183-200, February. T. J and C l a y t o n , R. W. 1976. Time Series Modeling and Maximum Entropy, Phys. Earth Planet. Inter., Vol. 12, pp. 188-200, August.
U lry c h ,
V a id y a n a th a n ,
P. P. 1990. Multirate Digital Filters, Filter Banks, Polyphase Networks, and Applications: A Tutorial, Proc. IE E E , Vol. 78, pp. 56-93, January. P. P. 1993. Multirate Systems and Filter B anks , Prentice Hall, Englewood Cliffs, N.J.
V a id y a n a th a n ,
V e tte r li,
J. 1984. Multi-dimensional Sub-band Coding: Some Theory and Algorithms, Signal Processing, Vol. 6 , pp. 97-112, April. J. 1987. A Theory of Multirate Filter Banks, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-35, pp. 356-372, March. A. C. G. 1977. Matrix Orthogonal Polynomials with Applications to Autoregressive Modeling and Ladder Forms, Ph.D. dissertation, Department of Electrical Engineering, Stanford University, Stanford, Calif., December.
V e tte rli,
V ie ir a ,
W a lk e r ,
G. 1931. On Periodicity in Series of Related Terms, Proc. R. Soc., Ser. A. Vol. 313, pp. 518-532. M., and K a i l a t h , T. 1985. Detection of Signals by Information Theoretic Criteria. IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-32, pp. 387-392, April.
W ax,

W e i n b e r g , L. 1962.
W e lc h ,
R15
Network Analysis and Synthesis, M cG raw-H ill, N ew York.
P. D. 1967. The Use of Fast Fourier Transform for the Estimation of Power Spectra: A Method Based on Time Averaging over Short Modified Periodograms, IE E E Trans. A u d io and Electroacoustics , Vol. AU-15, pp. 70-73, June. i e n e r , N. 1949. Extrapolation, Interpolation and Sm oothing o f Stationary Time Series with Engineering Applications , Wiley, New York.
N., and P a l e y , R. E. A. C. 1934. Fourier Transforms in the Complex Domain. American Mathematical Society, Providence, R.I. W i n o g r a d , S. 1976. "On Computing the Direte Fourier Transform, Proc. Natl. Acad. S e t, Vol. 73, pp. 105-106. W i n o g r a d , S. 1978. On Computing the Discrete Fourier Transform, Math. Comp., Vol. 32, pp. 1 7 7 -1 9 9 . W o l d , H. 1938. A Study in the Analysis o f Stationary Time Series, reprinted by Almqvist & Wiksell, Stockholm, 1954. W o o d , L. C., and T r e i t e l , S. 1975. Seismic Signal Processing, Proc. IE E E , Vol. 63, pp. 649661, April. W o o d s , J. W ., and O N e i l , S. D. 1986. Subband Coding of Images, IE E E Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-34, pp. 1278-1288, October. Y o u l a , D and K a z a n j i a n , N. 1978. Bauer-Type Factorization of Positive Matrices and the Theory of Matrices Orthogonal on the Unit Circle, IE E E Trans. Circuits and Systems, Vol. CAS-25, pp. 57-69, January. Y ule, G. U. 1927. On a Method of Investigating Periodicities in Disturbed Series with Special References to Wolfers Sunspot Numbers, Philos. Trans. R. Soc. London, Ser. A. Vol. 226, pp. 267-298, July. Z a d e h , L. A., and D e s o e r , C. A. 1963. Linear System Theory: The State-Space Approach, McGraw-Hill, New York. Z v e r e v , A. I. 1967. H andbook o f Filter Synthesis, Wiley, New York.
W ie n e r ,
Accumulator, 58 Akaike information criterion (AIC), 932 Algorithms, 4 Chiip-z, 482-486 FFT, 448-475 Goertzel, 4 8 0 -4 8 1 Remez, 645-647 Aliasing frequency-domain, 22, 276-279 time-domain. 276 Alternation theorem, 643 Amplitude, 16 Analog-to-digital (A/D) converter, 5, 21, 748-762 oversampling, 756-762 Analog signals (see Signals) Antialiasing filter, 746-747 Autocorrelation of deterministic signals, 118-133 of random signals, 327-329, A3 Autocovariance, A4 Autoregressive-moving average (ARMA) process, 855, 922 autocorrelation of, 856 Autoregressive (AR) process, 855, 922 autocorrelation of, 856 Averages autocorrelation, A3 autocovariance. A4 ensemble, A3-A5 for discrete-time signals, A6-A7 expected value, A3 moments, A3 power, A3 time-averages, A8-A10 Backward predictor, 515. 860 Bandlimited signals, 281 Bandpass filter, 331. 337-338 Bandpass signal, 280, 738-742 Bandwidth, 279-282. 619 Bartlett's method (see Power spectrum estimation) Bessel filter, 690-691 Bibliography, R1-R15 Bilinear transformation. 676-680 Binary codes, 752
Blackman-Tukey method (see Power spectrum estimation) Burg algorithm (see Power spectrum estimation) Butterworth filters, 682-683 Canonic form. 114 Capon method ( s e e Power spectrum estimation) Carrier frequency. 739-740 Cauchy integral theorem. 160 Causal signals, 87 Causal systems, 68-69, 86-87, 208-209 Causality, implications of, 615-618 Cepstral coefficients. 854 Cepstrum, 265-266 Characteristic polynomial, 101. 547 Chebyshev filters, 683-689 Chirp signal, 484 Chirp-z transform algorithm. 482-486 Circular convolution, 415-420 Coding, 22-38 Comb filter, 345-349 Complex envelope, 740 Constant-coefficient difference equations, 95-111 solution of, 100-111 Continuous-time signals exponentials. 19-20 sampling of, 2 1-24, 269-279 sampling theorem for, 29-31, 269-279 Convolution (linear), 75-82 circular, 415-420 properties, 82-85 sum, 75 Correlation, 1J 8 133. A3 autocorrelation, 122, 325-330, A3 computation, 130-131 cross-correlation, 120, 325-327 of periodic signals. 124, 130 properties, 122 124 Coupled-form oscillator, 352-354 Cross-power density spectrum, A6 Dead band, 584 Decimation, 784-787 See also Sampling rate conversion
Deconvolution. 266, 355-359, 363-365 homomorphic. 266. 365-367 Delta modulation, 758 Difference equations. 9 1 - 1 11 constant coefficient, 100-111 solution, 100-111 homogeneous solution. 100-103 particular solution. 103-104 for recursive systems. 95. I ll total solution. 105-111 from one-sided ;-translomi. 201-202 Differentiator, 652 design of, 652-657 Digital resonator. 340 Digital sinusoidal oscillator. 353-354 Digital-to-analog (D/A) converter. 5, 22, 3 8 .7 6 3 -7 7 4 oversampling. 774 Dirichlet conditions for Fourier series, 234 for Fourier transform. 243 Discrete Fourier transform (DFT). 399-402 computation, 4 49-473 butterfly, 460, 464, 466 decimination-in-frequency FFT algorithm, 461 464 decimination-in-time FFT algorithm, 456-461 direct, 449 450 divide-and-conquer method, 450-473 in-place computations, 461 radix-2 FFT algorithms, 456-464 radix-4 FFT algorithms, 465 469 shuffling of data, 461 split radix. 4 70 -47 3 via linear filtering. 479-486 definition, 401 IDFT, 401 implementation o f FFT algorithm, 473-479 properties, 4 09 -42 5 circular convolution, 415-420 circular correlation, 423 circular frequency shift, 422 circular time shift, 421-422 complex conjugate, 423 linearity, 410 multiplication, 4 15 -42 0
1 1
Index
Parseval's theorem, 424 periodicity, 410 symmetry, 413 415 table. 415 time reversal. 421 relationship to Fourier series. 407. 409 relationship to Fourier transform, 407 relationship to -transform, 408 use in frequency analysis, 433 -44 0 use in linear filtering, 425-433 Discrete-time signals, 9, 43-55 antisymmetric (odd), 5 1 correlation, 1 IS 133 definition, 9, 43 exponential, 46-47 frequency analysis of. 247-264 nonperiodic, 50 periodic. 14-19 random, 12 representation of, 44 sinusoidal, 16-18 symmetric (even), 51 unit ramp, 45 unit sample, 45 unit step, 45 Discrete-time systems, 56-71 causal, 6 8-6 9, 86-87 dynamic, 62 finite-duration impulse response. 90-91 finite memory, 62, 90-91 implementation of, 500-556 infinite-duration impulse response. 90-91 infinite memory, 62, 90-91 linear, 65 memoryless, 60 noncausal, 6 8-69 nonlinear, 67 nonrecursive, 94-95 recursive, 92-93 relaxed, 59 representation, 44 shift-invariant, 63-65 stability test for, 213-219 stability triangle, 216 stable (BIBO), 69-70, 87-90 static, 62 time-invariant, 63-65 unit sample (impulse) response, 76-82 unstable, 6 9-70, 87-90 Distortion amplitude, 317 delay. 332 harmonic, 378 phase, 317 Down sampling, 54, 55 See also Sampling rate conversion Dynamic range, 35, 561, 751 Eigenfunction, 307 Eigenvalue. 307, 547 Eigenvector, 547 Elliptic fitters, 689-690 Emigy definition, 47 density spectrum, 243-246, 260-264 partial. 391 signal. 47-49 Energy density spectrum, 243-246, 260-264 computation. 897-902 Ensemble, AI averages, A3-A3 Envelope, 740-741 complex, 740 Envelope delay, 332 Ergodic, A8 correlation-ergodic, A9-A10 mean-ergodic, A8-A9 Estimate (properties) asymptotic bias, 904 asymptotic variance, 904 bias, 904 consistent, 904 variance, 904 See also Power spectrum estimation Fast Fourier transform (FFT) algorithms, 448-475 application to correlation, 4 77-479 efficient computation of DFT, 448-475 linear filtering, 477-479 implementation. 473-475 minor FFT, 473 phase FFT, 473 radix-2 algorithm. 456-464 decimation-in-frequcncy. 461-464 decimation-in-time. 456-461 radix-4 algorithm, 465-469 split-radix, 470-473 Fibonacci sequence, 201, 548-549, 553 difference equation, 201-202 state-space form, 548-549, 553 Filter bandpass, 331, 337-338 definition, 317, 330-332 design of linear-phase FIR, 620-665 transition coefficient for, C1-C5 design of HR filters, 330-354, 666-692 all pass, 350-352 comb, 345-349 notch, 343-344 by pole-zero placement, 333-354 resonators (digital), 340-343 distortion. 317 distortionless, 332 frequency-selective, 331 highpass, 331. 333-334 ideal. 331-332 least squares inverse, 711-718 lowpass, 331, 333-334 nonideal, 332-333 passband ripple, 619 stopband ripie, 619 transition band, 619 prediction error filter, 512. 858 smoothing, 39 structures, 500-556 Wiener filter, 715, 880-890 n h e r banks, 825-831
1 2
critically sampled, 829 quadrature minor, 833-841 uniform DFT, 826-829 Filter transformations, 338-340, 692-700 analog domain. 693-698 digital domain, 698 -70 0 lowpass-lo-highpass, 338-340 Filtering of long data sequences, 430-433 overlap-add method for, 430-432 overlap-save method for, 430-431 via DFT, 430-433 Final prediction error (FPE) criterion, 931-932 Final value theorem, 200 FIR filters antisymmetric, 620-622 design, 620-665 comparison of methods, 662-665 differentiators, 652-657 equiripple (Chebyshev) approximation, 637-661 frequency sampling method, 630-637 Hilbert transformers, 657-662 window method, 623-630 linear phase property, 621-623 symmetric, 62(W>22 FIR filter structures, 502-519 cascade form, 504-505 direct form, 503-504 conversion to lattice form, 518-519 frequency sampling form, 506-510 lattice form, 511-561, 877 conversion to direct form, 516-517 transposed form, 5 25-526 FIR systems, 90, 94, 115 First-order hold, 768-769 Fixed-point representation, 557-560 Floating-point representation, 561-564 Flowgraphs, 521-524 Folding frequency, 28, 274-275 Forced response, 96-97 Forward predictor, 515, 857-858 Fourier series, 20, 232-240, 247-250 coefficients of, 234-235, 247-248 for continuous-time periodic signals, 232-240 for discrete-time periodic signals, 247-250 Fourier transform, 240-243, 253-256 of continuous-time aperiodic signals, 240-243 convergence of, 256-259 of discrete-time aperiodic signals, 253-256 inverse, 242 properties convolution, 297-298 correlation, 298 differentiation, 303-304 frequency shifting, 300 linearity, 294-295 modulation, 300-302 multiplication, 302-303 Paraevals theorem, 302
13
Fourier transform (continued) symmetry, 287-294 table, 304 time-reversal, 297 time-shifting, 296 relationship to ;-transform, 264-265 of signals with poles on unit circle, 267-268 Frequency, 14-18 alias, 18, 26 content, 29 folding, 28, 274-275 fundamental range, 19 highest. 19 negative. 15 normalized, 24 positive, 15 relative, 24 Frequency analysis continuous-time aperiodic signals, 240-243 continuous-time periodic signals, 232-240 discrete-time aperiodic signals, 253-256 discrete-time periodic signals, 247-250 dualities, 282-286 for LTI systems, 305-330 table of formulas for, 285 Frequency response. 311 computation. 321 325 to exponentials, 306-314 geometric interpretation of, 321-325 magnitude of, 311 phase of. 311 relation to system function, 319-321 lo sinusoids, 311-314 Frequency transformations (see Filter transformations) Fundamental period, 17 Gibbs phenomenon, 259, 629 Goertzel algorithm, 480-481 Granular noise, 753 Group (envelope) delay, 332 Harmonic distortion. 378, 779-780 High-frequency signal. 280 Hilbert transform, 618 Hilbert transformer, 657-662, 739 Homomorphic deconvolution, 365-367 system, 366 HR filters design from analog filters, 666-692 by approximation of derivatives, 667-671 by bilinear transformation, 676-680, 692 by impulse invariance, 671-676 by maiched-z transformation, 681 least-squares design methods, 706-724 frequency domain optimization, 719-724 least squares inverse, 711 718 Pronys (least squares), 706-708 Shanks' (least squares), 709-710 Padi approximation. 701-705 pole-zero placement, 333-354 IIR filter structures, 519-556 cascade form, 526-528 direct form, 519-521 lattice-ladder, 531-539, 8 78-880 parallel fonn, 529-531 second-order modules, 527 state-space forms, 539-556 transposed forms. 521-526 Impulse response. 108-110 Initial value theorems, 172 Innovations process, 852-854 Interpolation, 30-31, 273, 784, 787-790 first-order hold, 768-771 function, 30, 763 ideal, 3 0-31. 273 linear, 38, 768-774 See also Sampling-rate conversion Inverse filter, 355-356 Inverse Fourier transform, 242, 256 Inverse system, 355, 357 Inverse -transform, 160-172, 184-197 by contour integration, 160-161, 184-186 integral formula, 160 partial fraction expansion, 188-197 power series, 186-188 Lattice filters, 511-516, 531-539, 859. 876-880 AR structures, 877-878 ARMA structure, 878-880 MA structure, 857-859 Leakage, 434, 899 Least squares filter design, 706-724 inverse filter, 711-718 Levinson-Durbin algorithm, 716, 865-868 generalized, 868. 893 split Levinson. 891 Limit cycle oscillations, 583-587 Linear filtering based on DFT, 425-433 overlap-add method, 430-432 overlap-save method, 430-431 Linear interpolation, 768-774 Linear prediction. 512, 857-876 backward, 860-863 forward. 857-860 lattice filter for, 859-860 normal equations for, 864 properties of, 873-876 maximum-phase, 843-844 minimum-phase, 842-843 orthogonality. 844 whitening, 844 Linear prediction filter (see Linear prediction) LTI systems moving average, 115, 117 second order, 115-116 structures. 111-118 canonic form, 114 direct form I. 111-112 direct form II, 113-114 nonrecursive, 116-118 recursive, 116-118 weighted moving average, 115 Low-frequency signal, 279 Lowpass filter, 331
Index
Maximal ripple filters. 644 Maximum entropy method. 928 Maximum-phase system, 361-362 Mean square estimation, 882 orthogonality principle, 884-886 Minimum description length (MDL), 932 Minimum-phase system, 359-362 Minimum variance estimate, 942-946 Mixed-phase system, 361-362 Moving-average filter, 309 Moving-average (M A) process, 855, 922 autocorrelation of, 857 Moving-average signal, 115 Multichannel signal. 7 Multidimensional signal, 7 -8 Narrowband signal. 281 Natural response. 97 Natural signals. 282-283 Noise subspace, 951 Noise whitening filter, 854 Normal equations, 864 solution of, 864-873 Levinson-Durbin algorithm, 865-868 Schur algorithm. 868-872 Number representation, 556-564 fixed-point. 557-561 floating point, 5 61-564 Nyquist rate, 30 Ones complement. 558 One-sided z-transform, 197-202 Orthogonality principle, 884-885 Oscillators (sinusoidal generators) CORDIC algorithm for, 354 coupled-form, 353-354 digital. 352 Overflow, 588-589 Overtap-add method, 430 -43 2 Overlap-save method, 430-431 Overload noise, 418 Oversampling A/D, 756-762 Oversampling D/A, 774 Paley-Wiener theorem, 616 Parseval's relations aperiodic (energy) signals, 244, 260. 302 DFT. 424 periodic (power) signals, 236, 251 Partial energy, 363 Partial fraction expansion (tee Inverse r-transform) Periodogram, 9 02 -90 6 estimation of, 902 -90 6 mean value. 903
Index
variance, 903 Phase. 14, 741 maximum. 359-363 minimum, 359-363 mixed, 359-363 response, 311 Pisarenko method, 948-950 Poles, 172 complex conjugate, 193-194, 218-219 distinct, 178-179, 189-191. 217 location, 178-181 multiple-order, 179, 191-192 Polyphase filters, 797-800 for decimation, 800 for interpolation, 797 Power definition, 49 signal. 50 Power density spectrum, 235-240 definition. 236 estimation of (see Power spectrum estimation) periodic signals, 235-240, 250-253 random signals, A5-A7 rectangular pulse train, 237-240 Power spectrum estimation Capon (minimum variance) method. 942-945 direct method. 899 eigenanalysis algorithms, 950-959 ESPRIT. 953-955 MUSIC, 952 order selection, 955-956 Pisarenko, 948-950 experimental results, 936-942 from finite data, 902-908 indirect method, 899 leakage, 899 nonparametric methods, 908 -92 0 Bartlett, 910-911, 917 Blackman-Tukey, 913-916, 918-919 computational requirements, 919-920 performance characteristics, 916-919 Welch, 911-913, 917-918 parametric (model-based) methods, 920-942 AR model, 924 AR model order selection, 9 3 1-932 ARMA model, 924, 9 34-936 Burg method, 925-928 least-squares, 929-930 MA model, 924, 933-934 maximum entropy method. 928 model parameters, 923-924 modified Burg, 928 relation to linear prediction. 923-924 sequential least squares, 930-931 Yule-Walker. 925 use of DFT, 906-908 Prediction coefficients, 857 Prediction-error filter, 512, 858 properties of, 873-876 Principal eigenvalues, 951 Probability density function, A1-A3 Probability distribution function. B1-B2 Prony's method, 706-708 Pseudorandom sequences Barker sequence. 148 maximal-length shift register sequences, 148-149 Quadrature components, 740 Quadrature mirror filters for perfect reconstruction, 833-841 for subband coding, 832 Quality, 916-919 of Bartlett estimate, 917 of Blackman-Tukey estimate, 918-919 of Welch estimate, 917-918 Quantization, 21-22, 33-38, 750-753 in A/D conversion. 750-753 differential, 756 differential predictive. 757 dynamic range, 35, 561, 751 error. 37. 42, 582-598 in filler coefficients, 569-582 rounding. 35, 565-567 truncation. 35. 564-565 level, 35. 750 resolution. 35. 561 step size, 35. 5 6 1 Quantization effects in A/D conversion. 37-38, 753-756 in computation of DFT, 486-493 direct computation. 487-489 FFT algorithms. 489-493 in filter coefficients, 569-582 fixed-point numbers, 557-560 ones complement, 558-559 sign-magnitude. 558 table of bipolar codes, 752 Iwo's complement. 559-560 floating-point numbers, 561-564 limit cycles, 583-587 dead band, 584 overflow, 5 88-589 zero-input, 584 scaling to prevent overflow, 588 -58 9 statistical characterization, 590-598 Quantizer midrise, 750 midtread. 750 resolution. 750-752 uniform. 750 Random number generators. B I-B 6 Gaussian random variable. B4-B6 subroutine for. B6 Random processes. 327-330. A I-A 10 averages, A3-A8 autocorrelation. A3 autocovariance, A4 for discrete-time signals, A6-A7 expected value, A3 moments. A3 power. A3 cofielation-crgodic, A 9-A 10 discrete-time. A6-A7 ergodic, A8 jointly stationary, A2-A3
14
mean-ergodic, A8-A9 power density spectrum, A5-A6 response of linear systems, 327-330 autocorrelation, 327-329 expected value, 328 power density spectrum, 329-330 sample function, A 1 stationary, A3 wide-sense, A3 time-averages, A8-A9 Random signals (see Random processes) Rational z-transforms, 188-196 poles, 172-174 zeros, 172-174 Recursive systems, 116-118 References, R1-R15 Reflection coefficients, 512, 536. 863-864 Resonator (see Digital resonator) Reverse (reciprocal) polynomial. 515, 861 backward system function. 515, 861 Round-off error, 565-567, 590-598 Sample function. Al Sample-and-hold, 748-749, 765 Sampling, 9, 21. 23, 269-279, 742-746 aliasing effects. 2 7-28, 271-279 of analog signals, 23-33, 269-279. 742-746 of bandpass signals. 742-746 of discrete-time signals. 782-845 frequency, 23 frequency domain, 394-399 interval, 23 Nyquist rate, 30 period, 23 periodic, 23 rate, 23 of sinusoidal signals, 24-28 theorem, 29-30 time-domain, 2 4-2 8, 269-279 uniform, 23 Sampling-rate conversion, 782-845 applications of, 821-845 for DFT filter banks, 825-831 for interfacing, 823 for lowpass filters, 824 for oversampling A/D and D/A. 843-844 for phase shifters. 821-822 for subband coding, 831-832 for transmultipiexing, 841-843 by arbitrary factor. 815-821 of bandpass signals, 810-815 decimation, 784-787 filter design for, 792-806 interpolation, 784. 787-790 multistage, 806-810 polyphase filters for, 797-800 by rational factor, 790-792 Sampling theorem, 29-30, 269-279 Schur algorithm, 868-872 pipelined architecture for, 872-873 split-Schiir algorithm. 892
15
Schur-Cohn stability test. 213-215 conversion to lattice coefficients. 213-214 Shanks' method, 709-710 Sigma-delta modulation. 758 Sign magnitude representation, 558 Signal flowgraphs, 521-526 Signals, 2-3 analog, g antisymmetric, 51 aperiodic. 50 bandpass, 280. 738-742 complex envelope. 740 envelope, 741 quadrature components. 740 continuous-time. 8 deterministic. 11 digital. 11 discrete-time, 9, 43-55 electrocardiogram (ECG), 7 equivalent lowpass. 740 harmonically related. 19 multichannel. 7 multidimensional, 7 natural. 282 frequency ranges. 2K2-283 periodic, 15 random, 12, AI-A 10 correlation-ergodic. A9-A10 ergodic. A9 expected value of, A4 mean-ergodic. A9-A10 moments of. A4-A7 statistically independent. A4 strict-sense stationary'. A3 time-averages, A 8-A I0 wide-sense stationary. A3 unbiased. A8 unconelated. A4 seismic, 283 sinusoidal. 14 speech. 2-3 symmetric, 51 Signal subspace, 951 Sinusoidal generators {.tee Oscillators) Spectrum, 230-232 analysis. 232 estimation of. 232. 896-959 line, 237 Set also Power spectrum estimation Split-radix algorithms. 470-473 Stability of LTI systems, 208-217 of second-order systems, 215-217 Stability triangle. 216 State-space analysis, 539-566 definition of stale, 540 for difference equations. 540-542 LTI state-space model. 542 output equation. 542 relation to impulse response, 551-553 solution of state-space equations, 543-544 state equations, 542 state space, 541 state-space realizations cascade form, 555 coupled form, 556 minimal, 546 normal (diagonal) form, 555 parallel form, 555 state transition matrix, 544 state variables, 539 z-domain, 550-554 zero-input response, 544 zero-state response. 544 Steady-state response. 206-207, 314-316 Structures. 111-118 direct form I, 111-112 direct form II, 113-114 Subband coding, 831-833 Superposition principle, 65 Superposition summation, 76 System, 3, 56-59 dynamic, 62 finite memory, 62 infinite memory, 62 inverse, 356 invertible, 356 relaxed, 59 Syslem function 181-184, 319-321 of all-pole system, 183 of all-zero system, 182-183 of LTI systems, 182-183 relation to frequency response, 319-321 System identification, 355, 363-364 System modeling, 855 System responses forced, 96-97 impulse, 108-110 natural (free), 97, 204 o f relaxed pole-zero systems, 172-184 steady-stale, 206-207 of systems with initial conditions. 204-206 transient, 107, 206-207 zero-input, 97 zero-state, 96 Toeplitz matrix, 865, 883 Time averages, A8-A10 Time-limited signals, 281 Transient response. 107, 206-207. 314-315 Transition band, 619 Transposed structures, 521-526 Truncation error, 35, 564-565 Twos complement representation, 559 Uniform distribution, 487-488, 565-568, 755 Unit circle, 265, 267 Unit sample (impulse) response. 108-110 Unit sample sequence, 45 Variability. 916 Variance, 487-lXX, 591-593
Index
Welch method. 911-913, 917-918, 919-920 Wideband signal, 281 Wiener filters, 715, 880-890 for filtering, 8 8 1 FIR structure, 715, 881-884 IIR structure, 885-889 noncausal. 8 89-890 for prediction. 881 for smoothing. 881 Wiener-Hopf equation, 882 Wiener-Khintchine theorem. 299 Window functions, 626 Wold representation, 854 Wolfer sunspot numbers, 10 autocorrelation, 127-128 graph. 128 table, 127 Yule-Walker equations, 857 modified. 935 Yule-Walker method, 925 Zero-input linear, 98 Zero-input response, 97 Zero-order hold. 38. 765 Zero padding. 400 Zero-state linear, 98 Zeros, 172 Zoom frequency analysis, 850-851 Z - Irans forms definition, 151-152 bilateral (two-sided), 151-152 unilateral (one-sided), 197-202 inverse, 160-172. 184-197 by contour integration, 160-161, 184-186 by partial fraction-expansion, 188-197 by power series, 186-188 properties, 161-172 convolution. 168-169 correlation, 169-170 differentiation, 166-167 initial value theorem, 172 linearity. 161-163 multiplication, 170-171 Parsevals relation, 171-172 scaling. 164-165 table of, 173 time reversal, 166 time shifting, 163-164 rational, 172-184 region of convergence (ROC), 152-160 relationship of Fourier transform, 264-265 table of, 174

DSP Principles Algorithms and Applications Third Edition

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

DSP Principles Algorithms and Applications Third Edition

Diunggah oleh

Hak Cipta:

Format Tersedia

77un/ Edition

DIGITAL SIGNAL PROCESSING

J o h n G. Proakis Dimitris G. M anolakis

Digital Signal Processing

PRENTICE-HALL INTERNATIONAL, INC.

Printed in the United States of America 10 9 8 7 6 5

DISCRETE-TIME SIGNALS AND SYSTEMS

THE Z-TRANSFORM AND ITS APPLICATION TO THE ANALYSIS OF LTI SYSTEMS

FREQUENCY ANALYSIS OF SIGNALS AND SYSTEMS

4.2.2 4.2.3 4.2.4 4.2.5 4.2.6 4.2.7 4.2.8

vii S u m m ary a n d R e fe re n c e s P ro b le m s 368 367

THE DISCRETE FOURIER TRANSFORM: ITS PROPERTIES AND APPLICATIONS

EFFICIENT COMPUTATION OF THE DFT: FAST FOURIER TRANSFORM ALGORITHMS

7 IMPLEMENTATION OF DISCRETE-TIME SYSTEMS

DESIGN OF DIGITAL FILTERS

SAMPLING AND RECONSTRUCTION OF SIGNALS

10 MULTIRATE DIGITAL SIGNAL PROCESSING

M u ltistag e Im p le m e n ta tio n o f S a m p lin g -R a te C o n v e rs io n

11 LINEAR PREDICTION AND OPTIMUM LINEAR FILTERS

12 POWER SPECTRUM ESTIMATION

1.1 SIGNALS, SYSTEMS, AND SIGNAL PROCESSING

Signals, Systems, and Signal Processing

---------- ,|* y y y > v y y m w

' W W W .......... ............

1.1.1 Basic Elements of a Digital Signal Processing System

Analog input signal

Analog signal processor

Analog output signal Figure 1.2 A nalog signal processing.

Signals, Systems, and Signal Processing

Block diagram of a digital signal processing system.

1.1.2 Advantages of Digital over Analog Signal Processing

1.2 CLASSIFICATION OF SIGNALS

1.2.1 Multichannel and Multidimensional Signals

I____ _ J _______I _______ I----------- 1 -----------1 ----------- 1 -----------1 ----------- 1 -----------1

1.2.2 Continuous-Time Versus Discrete-Time Signals

Example of a two-dimensional signal.

bv sa m p lin g th e an a lo g signal x ( t ) 0 .8 ', t > 0 an d x ( t ) = 0. t < 0 once ev ery seco n d .

2. By accu m u latin g a v a ria b le o v e r a p e rio d o f tim e. F o r e x a m p le , c o u n tin g th e

1.2.3 Continuous-Valued Versus Discrete-Valued Signals

Year Figure 1.7 W olfer annual sunspot num bers (1770-1869).

Digital signal with four different amplitude values.

1.2.4 Deterministic Versus Random Signals

3 > 4 1 ---------------------------------------------------------- 1 ------- * ------------------- ----- ------ ---------------------------------------

400 f ----------------- ------------------- - -------------------------- ------------------- ------------------- 1

Figure 1.9 tograms.

----- ' 200

1.3 THE CONCEPT OF FREQUENCY IN CONTINUOUS-TIME AND DISCRETE-TIME SIGNALS

1.3.1 Continuous-Time Sinusoidal Signals

W e will use b o th fo rm s, (1.3.1) an d (1.3.3), in re p re s e n tin g sin u so id al signals.

Figure 1.10 Example of an analog sinusoidal signal.

Frequency Concepts in Continuous-Discrete-Tim e Signals

Figure 1.11 Representation of a cosine function by a pair of complex-conjugate exponentials (phasors).

1.3.2 Discrete-Time Sinusoidal Signals

Figure 1.12 Example of a discrete-time sinusoidal signal (w 7t / 6 and b 7r/3).

Frequency Concepts in Continuous-Discrete-Tim e Signals

Signal x ( n ) = c o s cu^n for various values of the frequency

Frequency Concepts in Continuous-Discrete-Time Signals

1.3.3 Harmonically Related Complex Exponentials

T h e basic signals for co n tin u o u s-tim e ,

A s in th e case o f c o n tin u o u s-tim e signals, it is o b v io u s th a t th e lin e a r co m b in atio n \-l

where 6 2 n q / N , where q and N are integers.

Analog-to-Digital and Digital-to-Analog Conversion

but different phase.

denote the sinusoidal signal sequence ( 2yrnk xk(n ) = sin I --------- ! V N

I____ _ J _I _ I----------- 1 -----------1 ----------- 1 -----------1 ----------- 1 -----------1