Anda di halaman 1dari 17

EE531 (Semester II, 2010)

5. Fourier analysis

empirical transfer function estimate (ETFE)


discrete Fourier Transform of a finite-length signal
properties of ETFE

5-1

Empirical transfer function estimate


consider a linear system with the representation
Y () = G()U ()
we extend the frequency analysis to the case of multifrequency inputs
an estimate of the transfer function:

G()
= Y ()U ()1
is proposed and is called the empirical transfer-function estimate (ETFE)
Estimates of Y (), U () are given by the Fourier transform of the finite
sequences:
N
1 X
y(t)eit,
YN () =
N t=1

Fourier analysis

N
1 X
UN () =
u(t)eit
N t=1

5-2

Discrete Fourier Transform (DFT)


the DFT of the length-N time-domain sequence x[n] is defined by
N 1
1 X
X[k] =
x[n]ei2kn/N ,
N n=0

0k N 1

the sequences X[k] are, in general, complex numbers even x[n] are real

the inverse DFT is given by


N 1
1 X
X[k]ei2kn/N ,
x[n] =
N k=0

Fourier analysis

0nN 1

5-3

Matrix relation of DFT


Define W = ei2/N , we can write the DFT in a matrix form as

1
1
1
X[0]
1
2
1
X[1]
W
W

2
4
X[2] = 1 1
W
W

..
..
..
N
..

X[N 1]
1 W N 1 W 2(N 1)

or

1
x[0]
x[1]

W N 1

2(N 1)

W
x[2]

..
..
...

x[N 1]
W (N 1)(N 1)

X = Dx,
where D is called the DFT matrix
The inverse DFT is given by x = D1X and using the fact that
D1 = D
(D is called an orthogonal matrix, i.e., DD = I)
Fourier analysis

5-4

Orthogonality of DFT matrix


the columns of DFT matrix are of the form:

T
k
2k
k(N
1)
k = (1/ N ) 1 W
,
W
W

or equivalently


k = (1/ N ) 1 ei2k/N

i2k2/N


i2k(N 1)/N T

use hl, k i = k l and apply the sum of geometric series:

N 1
1 X i2(kl)n/N
1 ei2(kl)
1
hl, k i =

e
=
N n=0
N 1 ei2(kl)/N

the columns of DFT matrix are therefore orthogonal


(
1, for k = l + rN, r = 0, 1, 2, . . .
hl, k i =
0, for k =
6 l
Fourier analysis

5-5

Frequency sampling of the Fourier transform


the Fourier transform of the length-N sequence x[n] is given by
X() =

x[n]ein =

n=

N
1
X

x[n]ein

n=0

if we uniformly sampling X() on the -axis between [0, 2) by


k = 2k/N,

0 k N 1,

then we have
X() |=2k/N =

N
1
X
n=0

x[n]ei2kn/N ,

0k N 1

frequency samples of X() of length-N sequence x[n] at N equally spaced


frequencies is precisely the N -point DFT X[k]
Fourier analysis

5-6

Computing DFT on MATLAB


y[k]
3

2.5

1.5

0.5

0
0

20

40

60

80

100

120

k
y[k] = a|k64|/(1 a2), with a = 0.8 (an autocorrelation sequence)
Fourier analysis

5-7

Computing DFT on MATLAB


Sampling Fourier transform

25

25

20

20

15

15

Y (k )

Y [k]

DFT

10

0
0

10

20

40

60

80

100

120

0
3

0
1
Frequency (rad)

Use fft command


y[k] is symmetric about 0, and so is Y [k]
Compare with Y () = 1/(1 + a2 2a cos )
Fourier analysis

5-8

Transformation of DFT
let y(t) and u(t) are related by a strictly linear SISO system:
y(t) = G(q)u(t)
where q is the forward shift operator and G(q) is the transfer function
assume that |u(t)| C for all t and let
N
1 X
YN () =
y(t)eit,
N t=1

N
1 X
UN () =
u(t)eit
N t=1

then the DFTs of (windowed) y(t) and u(t) are related by


YN () = G()UN () + RN ()
where

Fourier analysis

2KC
|RN ()|
N

(Ljung 1999, THM 2.1)

5-9

Proof. by definition
N
N X

X
1 X
1
YN () =
y(t)eit =
g(k)u(t k)eit
N t=1
N t=1 k=1

N
k
X
1 X
g(k)eik
u( )ei
=
N k=1
=1k

the last term on RHS is deviated from UN () by


N k
1 X
u( )ei |
|UN ()
N =1k

0
1 X
1
i
|
u( )e
| + |
N =1k
N

Fourier analysis

N
X

=N k+1

u( )e

2Ck
|
N

5-10

therefore,

X

N
k
X


i
ik 1

|YN () G()UN ()| =


u( )e UN ()
g(k)e
N =1k
k=1

2 X
|kg(k)Ceik |

N k=1

If we define
K=

k=1

k|g(k)| <

then the inequality is bounded by


2KC
|YN () G()UN ()|
N

Fourier analysis

5-11

Properties of ETFE
consider a linear model with disturbance
y(t) = G(q)u(t) + v(t)
from the previous page, we found that
RN () VN ()

G() = G() +
+
UN () UN ()
where VN () denotes the Fourier transform of the disturbance term
if we assume that v(t) has zero mean, then
RN ()

E G()
= G() +
UN ()

the estimate has a bias term which decays as 1/ N


Fourier analysis

5-12

Properties of ETFE
It can be shown that (Ljung 1999,6.3)

E[(G()
G())(G()
G())]

Sv () + N

2
|UN ()|
=
N

U ()U (),
N
N

if

if

| | =

2k
N ,k

= 1, 2, . . .

with |N | is bounded by 1/N (up to a constant factor)

Fourier analysis

5-13

Conclusions
Periodic inputs

G()
is defined only for a fixed number of frequencies
at these frequencies the ETFE is unbiased and its variance decays like
1/N
Nonperiodic inputs

G()
is an asymptotically unbiased estimate of G() at many
frequencies

the variance of G()


does not decay with N but is given as the
noise-to-signal ratio at the frequency in question as N increases
this property makes the empirical estimate a crude estimate in most
cases in practice
Fourier analysis

5-14

Example
No smoothing
True
ETFE

0.8

Frequency Response

10

0.7

10

0.6

10

0.5

10

G(z) =
0.25

5
z0.2 ,

0.2

0.4
0.6
0.8
Normalized Frequency (x rad)

white noise input with power 1, additive noise variance is

Use etfe command in System Identification Toolbox


Fourier analysis

5-15

Example
Smoothing with a window of length 32
True
ETFE (N=512)
ETFE (N=32768)

0.77

Frequency Response

10

0.74

10

0.71

10

0.68

10

0.65

10

0.62

10

Fourier analysis

0.2

0.4
0.6
0.8
Normalized Frequency (x rad)

5-16

References
Chapter 6 in
L. Ljung, System Identification: Theory for the User, Prentice Hall, Second
edition, 1999
Chapter 5 in
S. K. Mitra, Digital Signal Processing, McGraw-Hill, International edition,
2006

Fourier analysis

5-17

Anda mungkin juga menyukai