196

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO. 3, MARCH 1981

Continuous Phase Modulation-Part Signaling

TOR AULIN,
MEMBER, IEEE, AND

I: Full Response

CARL-ERIK

w. SUNDBERG, M E M B E R , IEEE

Abstract-The continuous modulation phase (CPM) signaling scheme has gained interest in recent years because of its attractive spectral properties. Data symbol pulse shaping has previously been studied with regard to spectra, for binary data and modulation index 0.5. In this paper these results have been extended to theM-ary case, where the pulse shaping is over a one symbol interval, the so-called fullresponsesystems. Resultsaregivenformodulationindexes of practical interest, concerning performance both and spectrum. Comparisons are made with minimum shift keying (MSK) and systems have been found which are significantly better in E b / N O for a large signal-to-noise ratio(SNR)withoutexpandedbandwidth.Schemes withthesamebit error probability as MSK but withconsiderably smaller bandwidth have also been found. Significant improvement in both power and bandwidth are obtained by increasing the number of levels M from 2 to 4.

error probability at large signal-to-noise ratio (SNR) and spectrum is achieved. This trade off is studied for modulation indexes of practical interest and also for systems where the instantaneous frequency is not constant over each symbol interval. The channel noise is assumed to be additive,w h t e Gaussian throughoutthepaper.Thesymbolerrorprobabilityforan optimum detector at large SNR is calculated using the minimum Euclidean distance between any two signals in the signal space [3] . Theoptimumdetectoroperatescoherently,and due to the continuous phase, the detector must observe the received signal for more than one symbol interval to make a decision about aspecific symbol [3]

I. INTRODUCTION
OR digital transmission over bandlimited channels, the demand for bandwidth efficient constant envelope signaling schemes with good reliability has increased in recent years. A system often used in practice is multilevel phase shift keying, M-aryPSK,whichhasthedrawback that, although, for M equal 2 or 4, the receiver sensitivity is acceptable, thesignal is too wide-bandbecauseofdiscontinuousphase.Thus,RFfdteringhastobeperformedbeforetransmissioncausinga
nonconstant e n v e l o p e s i g n a l a n d a decreased receiver sensitiv-

11. GENERAL SYSTEM DESCRIPTION

For CPM systems, the transmittedsignal is ~ ( tCU) , =

IE

-COS (2nf0t + cp(t,a) + cpo)

where the idormation carryingphase is

ity. The so-called minimum shift keying (MSK), or fast frequency shift keying (FFSK), binary signaling schemes opened new prospects since the error probability performance is the same as coherent 2- or 4-ary PSK but the spectrumis narrower for large frequencies. Choosing an M larger than 4 (e.g., M = 8 or M = 16) in the MPSK system makes the main lobe of the spectrum narrower, but the sensitivity to noise is considerably increased. A general definition of continuous phase modulation (CPM) systems isgiven in the next section. Assume that each data symbol only affects the instantaneous frequency of the transmitted signal in one symbol interval and that the phase is a This definesthesubclassfull continuousfunctionoftime. response CPM systemsconsideredinthispaper.InPart I1 more general CPM schemes are considered. In some cases the phase is allowed to bediscontinuouswhilemaintainingthe coupling between the phasein successive symbol intervals. By increasing M , an interesting tradeoff between symbol
Manuscript received March 19, 1980; revised September 19, 1980. This work was supported by the Swedish Board of Technical Development under Grant79-3594. The authors are with the Department Telecommunication Theory, of University of Lund, Fack, S-220 07 Lund, Sweden.

cp(t,a) = 2nh

It 5 -f=--m

aig(r-iT)dr;

-<t<=

(2) and a = ... a-2 a P 1a a1 ... is an infinitely long sequence of . uncorrelated Mary data symbols, each taking oneof the values
a = +1, + 3 , ..., ?(M - 1); i

i = 0 , + I , +2,...

(3)

with equal probability1/ M . ( M is assumed even.) E is the symbol energy, T i s the symbol time,fo.is the carrier frequency,and cpo is anarbitraryconstantphaseshift which without loss of generality can be set tozero in the case of coherent transmission. The variable h is referred to as the modulation index, and the amplitude of the baseband pulse g(t) is chosen to give the maximum phase change olhn radians over each symbol interval when all the data symbols in the sequence a take the same value a . Forimplementationreasons,rational values of the modulation index h are used. This is discussed in some detail in Part 11. To a have CPM signal, the information carrying phase q(t, a) is a continuous function of time t , w h c h implies that the frequency baseband pulse g(t) does not contain any im-

0090-6778/81/0300-0196$00.75 0 1981 IEEE

AULIN AND SUNDBERG: CONTINUOUS PHASE MODULATION-PART I

Phase 4 4 hn

197

~Fig. 1.
~

2nh

fo

3 hn

Schematicmodulatorfor CPM.
Zhn

pulses. A schematic modulator is shown in Fig. 1 . Note that a CPM signal always has a constant envelope. Defining the baseband phase response (phase pulse)

hn

.q
- hn

it is seen that the phase of the CPM signal is formed by

-2hn

-3hn

-4hn

A causal CPM system is obtained if the frequency pulse g(t) satisfies

Fig. 2.

Phase trajectories for a binary fullresponse CPFSK system. Four bit time intervals are shown.

&)SO;
g(t)+O;

t<O,

t>LT

O<t<LT

(6)

where the pulse length L measured in symbol intervals T may be infinite. L = 1 yields full response schemes considered in this part. The normalizing constraint for the frequency pulse g(t) can be expressed as

infinitely long sequence 6 which minimizes the error probability. This is referred t o as maximum likelihoodsequence estimation (MLSE). In order to be able to study the performance of an optimum MLSE detector, a suboptimum detector is studied instead. The limiting case of this suboptimum detector is the MLSE detector [ 7 ] , [ l l ] . This suboptimum detector observes the received signal r(t) for N symbol intervals t o make a decision about a specific data symbol, say 01,. Thus;the receiver observes the signal

The CPFSK modulation schemes [ 5 ] , [7] , [13] are a subclass of theCPM signaling scheme wherethe instantaneous frequency is constant over each symbol interval. Thus, for a full response and if we let N + 03, an MLSE detector is obtained. An optimum detector maximizes the likelihood function [3] , [5] CPFSK modulation scheme,we have

t OGtGT
+;

t>T

and since the quantities SoNT ?(t)dt and independent of CN = Co, Z l , maximize
e . . ,

SoNT s2(t, 6)dt are

one can, as well,

which corresponds to linear phase trajectories over each symbol interval (see Fig. 2 ) . Note that although the scheme is full response, the actual phase in any specific symbol interval depends upon the previous data symbols. The CPM signal is assumed t o be transmitted over an additive, white,and Gaussian channel having a one-sided noise power spectral density N o . Thus, the signal available for observation is

or the log likelihood function

log ( A N ' [ r ( t ) l =

iNT
'''9

r(t>s(t,6)d t .

(13)

r(t) = s(t, a) + n(t);

--Oo

<t <

00

(9)

Since there are M N sequences, GN = Go, Zi1, C N P l , but the detector is onlyinterestedin finding an estimate 6 , of , 010,the M N sequences can be formed intoM groups:
.-e,

where n(t) is a Gaussian random process having zero mean and one-sided power spectral density N o . A detector which minimizes the probability of erroneous decisions must observe the received signal r(t) over the entire time axis and choose the

Q-I,N, a-3,N>

a-(M-l),N

(14)

198
where

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO. 3, MARCH 1981

can

written Equation (19)

be

andit is not necessary for the detector to find the specific sequence GN which maximizes (13) and to choose Go as an estimate of q,. Instead, the detector must find which group of sequences 6 k , N k = + I , +3, *.-,+(M - 1) jointly maximizes , (13), and take Zo as the group belonging. It is believed, however, that for large SNR the two detectors have the same performance [7]. Theprobabilityofanerroneous decision canbeupper bounded by using the union bound [3], [5] , [ l o ]

Thus, it is sufficient to consider the difference sequence

instead of the pair of sequences The approximation

(Yk

and

p e = r o - Q [ D m in,N -]

rn

where

and the summation is taken over all pairs of sequences defined k # 1, k , 1 = k l , + 3 , .--, by (15), with the restriction that +(M - 1). D [(Yk . N , (YI,N] is the Euclidean distance between the signals i ( t , a k , N ) and s ( t , ala). The squared Euclidean distance can be written

of (16) is good for large E/No. It is now assumed that E/No is sufficiently large forthisapproximation t o be valid. The limitations of this assumption is considered in detail in [20] . ro is a positive constant, independent of E/No, and D m i n is the minimum of D ( a k , N , a13) with respect to the pair of sequences and a I 3 withtherestrictionthat k # 1. This quantity can also be calculated using the difference sequence Y N through

(24)
with the restriction that
70

=2,4,6,

..., 2(M - 1) ..., +2(M - 1); i = 1 , 2 ,

.a .,

yi = 0, +2, +4,

N - 1. (25)

In Part 11, we will only deal with squared Euclidean distances normalized by bit energy Assuming 27rf0T S 1 , this can be written
n2

= 2 E (N -

lNT
COS [

N- 1

2~h
i= 0

( f f i k-

ffi)

Note that

- -in
q(t

I)
dt

The superscript denotes the value of the first symbol in a sequence of N symbols, i.e.,

where Eb is the bit energy. Thus, error probability comparisons for large SNR can be made directly in Eb/fVo between systems even if they have different M . Only values that are powers of two (M = 2, 4, 8, -) will beconsidered. As a reference point in the following, notethat d k i n = 2 for MSK, binary PSK (BPSK), andquaternary phase shiftkeying (QPSK). Forthe case of fullresponseCPFSK systems, calculations of d k i n , Nhas been considered for both the bi-

AULIN AND SUNDBERG: CONTINUOUS PHASE MODULATION-PART I

199

nary case [ 1 ] , [5] - [7], [I31 and also for M t o some extent [ l o ] , [ I l l , [ 1 3 ] .

=4

and M = 8

111. BOUNDS ON THE MINIMUM EUCLIDEAN DISTANCE

/-

An important tool for the analysis of CPM systems is the so-called phase tree. This tree is formed by all phase trajectories q(t, a) having a common start value zero at t = 0 . The ensemble is over the sequence Q! and Fig. 2 shows a part of the phase treefor abinary full response CPFSK system.A more general case is shownin Fig. 3 where two phase trees for a quaternary CPM system having different frequency baseband pulsesg(t) are shown. To calculate the minimum squared Euclidean distance for an observation length of N symbols, all pairs of phase trajectories in the phase' tree over N symbol intervals must be considered. The phase trajectories must not coincide over the first symbol interval however. The Euclidean distance is calculated according t o (21) for all these pairs, and'the minimum of these Euclidean distances is the desired result. It is of great importance to remember thatthe phase must always be viewed modulo 2n in conjunction with distance calculations. A practical method to do this is to form a cylinder by folding the phase tree [16] [17] . Trajectories which seem to be far apart in the phase tree might actually be very close or even coincide Fig. 3 . Phasetrees for M = 4 CPM schemeswithtwodifferent basewhen viewed modulo 2n. band pulses g(t). The a, p function [see(56)] with a = 0.25, is p = It is clear from (18) that, for a futed pair of phase trajec0 is shown by a solid lineand the HCS, half cycle sinusoid [see(SS)] tories, the Euclidean distance is a nondecreasing function of is shown by a dashed line. the observation length N . If just a few pairs of infinitely long sequences are chosen, an upper bound on the minimum Euthat instead of using the pair of sequences clidean distance at alj values of the observation interval N is It can be noted could obtained. Good candidates for these infinitely long pairs are (28), the single difference sequence 7 = +2, -2,0,0, be used together with (24) for calculation of the normalized pairs that merge as soon as possible. Twophasetrajectories merge at a certain time if they coincide all the. time thereafter. squared Euclidean distance. Turning to the quaternary case, we can find pairs of phase These merges are called 'inevitable if they occur independently of h . Thus, an upper bound on the minimum Euclidean dis- trajectories merging at t = 2T. Fig. 3 shows two examples of tance is obtained as a. function of the modulation index h phase trees for this case. These merges occur at the points labeled A , B , C , D ,and E in the phase tree. Unlike the binary for all N . Applying this method to the scheme i Fig. 2;it is seen that case, there is more than one merge point, and two different n pairs of phase trajectories can have the same merge point. if a pair of sequences is chosen as There are only three phase differences however, namely, those a + 1 =+l,-l,a2,a3;having phase difference +2hn7+4hn,.and +6hn at t = T . It is easily seen that an upper bound on the minimum Eua-1 =-1,+8 ) l , a 2 , a 3 , (2 clideansquareddistance forthe M-ary case is obtainedby using the difference sequences the two phase trajectories coincide for all t > 2T. Thus, the upper bound on the normalized minimum squared Euclidean distance is

d~~(h)=2--

;CT

COS

[2nh(24(t)- 2q(t - Q)] dt (29)

and taking the minimum of the resulting Euclidean distances, 1.e.,

dB2(h)= log,?

where (21) was used. For thebinary linear phase trajectories, the result is

CPFSK system, i.e.,

min
IQkGM-1

{ [
2-

2T

cos [2nh

200 which, for theM-ary CPFSK system, specializes to

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO. 3, MARCH 1981

d 2 (hl

2.5

N-4

Fig. 4 shows the minimum distanced2 for binary CPFSK (h) as a function of h and N . The upper bound dB2 is shown dashed. Note the peculiar behavior at h = 1. This will be discussed later. It canbenoticedthatfor a l fullresponse CPM systems l with the property

2.0

(34) a merge can occur at t = T and thus the difference sequences

1.5

yield the upper bound

dB2 (h) = log2 0
*

min
1 4k4M- 1

{ [
1-

T
COS

[2nh * 2q(t)] dt

1.o

/.'

1.0'

(36) This class of pulses is caged weak [20] ,,and is not considered because distance properties are poor. Only positive pulses g(t) will be considered below. Furthermore, they are assumed t o be symmetric with
g(t) = g(T - t);

Fig. 4. Normalized squared Euclideandistance versus modulationindex for binary CPFSK: upper bound (dashed line) and &(h) for N = 1 , 2, 3 and 4 bit decision intervals.

The modulation indexes [defined by (38)]

0 < t < T.

(37)

n = 1 , 2 , ...
Weak Modulation Indexes,h, For the construction of the upper bound dB2(h) on the minimum squaredEuclideandistance, pairs of phase trajectories giving merges at t = 2 T were used. These merges occur independently of the value of h. Forall pulses g(t) except weak ones the first inevitable (for all h), merge occurs at t = 2T. For specific values of the modulation index h , however, other merges are also.possible. In the binary case (see Fig.2), a merge can occur at t = T if the difference sequence is chosen to be -y = +2, 0, '0, ... andthemodulationindex h is an integer. This is because the two points labeled A and B are a multiple of 2n apart, and thuscoincide modulo 2n,. For an M-aryfullresponse system, the phase trajectories take the values +2nhq(T), *6nhq(T), *2(M - l)hnq(T), -.., which for positive g(t) pulses reduces to+hn,*3hn, ?(A4 - l)hn., Thus, there areM - 1 phase differences between the nodes in the phase tree at t = T , and merges occur at t = T for h-values given by
.e-,

(39)

are called weak modulationindexes of thefirstorder.For these modulation indexes, the minimum Euclidean distance is normally below the upper bound for all values of N ; see Fig. 4 for h = 1. Sometimes the minimum distance for weak modulation index values is considerably below dB2(h,). For such cases, h, is called a catastrophic modulation index value [I71 -[20]. Merges of the weak (catastrophic) type occur at any observation length. Thus, weak h-values of the second order are defined by

nh.Y,-,=2n'n;

~0=2,4,6;..,2(M-l) n = 1 , 2 , ...
(38)

As we will see later, the effect of weak modulation indexes of a higher order than one are of minor importance, since the correspondingEuclideandistance for the pairs of phase trajectories causing the merge is above the upper bound dB2(h) '[20] . For weak modulation indexes of the second order, the

AULIN AND SUNDBERG: PHASE MODULATION-PART CONTINUOUS

20 1

corresponding Euclidean distance might be on upper the bound dB2(h); forthe third orderand higher it is strictly above. But from (39) it can be seen that the number, of firstorder weak modulation indexes grows rapidly with M. For weak modulation indexes of the first order, it is sufficient to calculate the corresponding Euclidean distance over the firstsymbolinterval. Since the calculation of the upper bound uses the two first symbol intervals, it can be concluded that d2(h,) can be,and normally is, smaller than dB2(h,). Thus, for weak modulation indexes of the first order, dLin(h,) might be smaller than dB2(h,) (for details, see [20]). Furthermore, it is shown in detail [ 1 7 ] ,[20] that only of the first order can influence the minimum distance calculation for full response CPM systems.

The tightness of the upper bound and the behavior of (43) is illustrated in all the minimum distance figures.
Optimization of the Upper Bound dB2 (h)

For frequency pulses g(t) having the symmetry property

(37), the expression for the upper bound can be written

+,

(45)

Tightness of the Upper Bound dB2(h)

A powerful property of the upper bound on the minimum Euclidean distance is that except for weak modulation indexes of the first order, minimum the Euclidean distance itself equals this bound if the observationinterval is long enough [ 1 8 ] , 2 0 ] . Denoting minimum [ the normalized squared N symbol observation interval Euclidean distance for an d &in .,,(h), we have that

and since cos (.) 2 -1, dB2(h)can never exceed 4 10g2g. Thus, at most, an improvement of 3 dB in E/No for a large SNR might be obtained for the binarycase compared to MSK. Furthermore, (45) can also be written

1G k 4 M - 1

where

%,

COS

[4nhkqo(t> nhk] - dt

)]

(46)

if

N2
h

(h) (42)

+ h,: (first order).

The binary case (M = 2 ) will be considered first. In this case,

NB(h) is the number of symbol intervals required t o reach the upper bound for the specific modulation index h . If aspecific pair of phase trajectories never merge, the Euclideandistance will grow without limit. This is true because the Euclidean distance calculated over each symbol interval is positive. Since the minimum distance was previously shown t o be upperbounded,the pair of phase trajectories giving the minimum distance must eventually merge. For a modulation index near.a first-order weak modulation index, the difference sequence (35) gives the smallest growing Euclidean distance with N :

dB2(h) = 2 1 - 2

COS nh

min
14k4M-1

{ -$
1

cos [2nhkyoq(t)]dt

It can at once be observed that cosnh = 0 for h = 1 / 2 , 3 / 2 , 5 / 2 , etc. Thus, the upper bound equals two, independent of the shape of the frequency pulse g(t). Actually, this is true for all pulses g(t) (except weak ones) [16]. T h s is of particular interest since much attention has been devoted to the case of h= [ a , ~ 1 2 1 ~ 1 4 1~, 1 5 1 . , To maximize dB2(h) for the binary case, the last term in (48) must be minimized.Two different cases can be distinguished :
Case I:

+ (N - 1)(1 -COS

[nhkyo])

(43) OGh<$ wherecosnh20. Case 1 : 1 (44) (49)

and the smallest growth rate per symbol interval is log, M

min
1G k 4 M - 1

{ 1 - cos [nhkyo]}.

3 S h < 1 where cos nh < 0.

(50)

For sufficiently large N in (43), the minimum distance for the considered h is given dB2 (h) d2 (h) exceed dB2(h). by since will

To maximize dB2 (h), integral in (48) must be minimized the for case I and maximized for case 11. It is 'also clear that the

202

IEEE T R A N S A C T I O N S ON COMMUNICATIONS, VOL. COM-29,

NO. 3 , MARCH 1981

pulse, which maximized d B 2 @ ) case I, minimizes d B 2 @ ) for for case I1 and vice versa. To make the integral in (48) as small as possible, the argument inside the cosine must be as close to n as possible. This yields for the interval 0 < h < 1

the beginning and the end of the symbol interval, but remains constant (forms a plateau) in themiddle ofthe symbolinterval. Notethat q 1( t ) and q 2 ( t ) give systemswith discontinuous phase. However, the ensemble of possible phase trajectories is completely known to the receiver just as for all CPM systems. T h s gives a slight generalization of the considered class of systems. It is interesting to note that the plateau function with 6 = 1/2 gives the upper bound [ 171

with the.resulting phase responses . .

. .

and
(0;
t=O

which,for small values of a , approaches a value of 4 near h = 1. Since h = 1 is a f i r s t a d e r weak modulation index, it is concludedthat large values of N are required to make dii,,(h)'equal hg2(h) near h = 1. In practice, the phase responses q1 ( t ) and q 2 ( t ) are not attractive because of their spectra. This will be discussed more in Section VI. From a spectral point of view, the phase during a symbol interval should change slowly and smoothly, and the following frequency pulses with corresponding phase responses are of interest. The first one is

From the sign symmetry, to minimize d B 2 @ )the phase responses above have to be interchanged with each otherfor the . . respective cases. The upperbound d ~ ' ( h )forthetwo phase responses is now found as
.'

g(t) =

1,

nt sin-; T

O<t<T

with the corresponding phase response

Thus, for any binary scheme, (h) i s bounded by d~~

(5 9)

1 -COS 2nh < < ~ ~ ( h ) < - cos nh); 0 < h < 2(1
. .

3
(55)

t>T

2(1-c0snh)<d~2(h)~1-c0s2~hh; h < l . i<

An analogoustechnique can be to derive bounds for 2 1. used h The two phase responses q1 (t-> q 2 ( t ) are'members of and the class called a', functions [ 161-, [ 17.1 defind by 0

0 -

-.t

2a T '
a t 1-20

O<t<aT CYT<t < (1 -a)T

and like CPFSK this pulse At) has no continuous derivatives attheendpoints[14]'.The pulse in itself is continuous, however, unlike CPFSK: Since the frequency pulse 'is a half cycle sinusoid, this scheme 'will be referred to as halfcycle sinusoid (HCS) [see Fig. 3 (dashed tree) for the'quaternary . : case] . Another pulse of interest is

0 + -; T 1-2201 2

with the corresponding phase response

t2T.
Hence, when h = 1/2, q 1(t) corresponds to binary PSK but

I 0;

t<O

with a T/2 time offset. The phase response q 2 ( t ) corresponds

to' a so-called plateau function; Le., the phase changes only in

t 2 T.

AULlN AND SUNDBEKG: CONTINUOUS PHASE MODULATION -PAKT I

MSK, BPSK e t c .

203

Fig. 5. Upper bound comparison for M = 2, 0 S h < 0.5. The bound for q l ( t ) is the upper bound on all upper bounds d s 2 ( h ) , and the 01 = 0 plateau function is the lower bound on all the upper bounds d&h) in the considered interval.

This pulse g(t) has one continuous derivative at the end points t = 0 and t = T. Since this frequency pulse is a raised cosine function, it will be referred to as raised cosine (RC).When h = 1/2, this scheme has previously been referred to as SFSK [12] . Fig. 5 shows the upper and lower bounds on d B 2 @ ) all for binary schemes in the region 0 < h < 1/2, computed with (55). Fig. 5 also shows d ~ ~ (for CPFSK, HCS (with formula h ) in [ 181 ), and a, functions with a = 0.25, 0 = 0.5 and a = 0.25, p = 0 . The bound for the RC scheme is in between that of QI = 0.25,p = 0 and HCS. The problem of finding frequency pulsesg(t) that optimize d B 2 @ ) given h and M , is far more complicated in the general , M-ary case than for the binary case. This general problem has not been solved. The reason is that the upper bound is constructed from the minimum of more than one function, and h varies with fixed M , differentfunctionstaketheminimum value. This is also true for futed h and M when the frequency pulse g(t) is varied. For anM-ary scheme with 0 < h < 1/M the binary bounds (55) on dB2( h ) still apply after multiplication with log2M. Of course, also for M-ary schemes d B 2 @ )< 410g2M. However, duetothefactthatthebound d B 2 @ )inthis case is formed bytaking theminimum of several componentfunctions (32), this maximum value can never be reached [ 171, [201 . It is previously known [5] , [7] , that the h-value maximizing theminimum Euclidean distance (N > 3) for binary

CPFSK is h = 0.715. This value of h also maximizes dB2(h) for this scheme. The same h was also shown by Kotelnikov [ 11 to maximize the Euclidean distance when N = 1. It is possible tofindthe values of h whichmaximize d B 2 @ )forM-ary CPFSK [see (33)] and they are given in Table I together with the maximumvalue of dB2( h ) for M = 2 , 4 , 8 , 1 6 , and 32. Theoptimum occurs for h-values slightly below h = 1. Unfortunately, h = 1 is a first-order weak modulation index for all M , but if N is made large enough (N > NB)d$i,(ho) equals dB2 (ho). For the quaternary case, the CPFSK scheme gives a maximum of d s 2 ( h o )= 4.232. In [17] it is shown that a scheme based on the a, 0 function with a = 0 , = 0.17 for h = 0.62 gives a minimum distance of 4.62. This value came out of a nonexhaustivesearch forquaternary schemes yielding large distance values. However, better schemes may exist. IV. NUMERICAL RESULTS ON THE MINIMUM EUCLIDEAN DISTANCE In thissection numerical results ontheminimumnormalized squaredEuclideandistance will be given in form of graphs. These graphs present the minimum Euclidean distance versus the modulation index h for specific schemesand different values number the of N , of received signal intervals observed. Thus, these graphs will show what is below the upper bound dB2(h),and also how large N has t o be made in a

204

TRANSACTIONS IEEE

O N COMMUNICATIONS, VOL. COM-29, NO. 3, MARCH 1981

TABLE I OPTIMUM h-VALUES ANDCORRESPONDING NORMALIZED EUCLIDEAN DISTANCES FORM-ARY CPFSK SCHEMES

3.5

specific situation, to make the minimum Euclidean distance d$in(/z) equal to d B 2 @ ) .

Plateau Functions

3.0

As an example of a plateau function, a binary scheme with a phase response very similar to that of QI = 0.05 and /3 = 1/2 wdl be chosen. The difference between the chosen phase functionandthe a, /3 function is thatthephasedoesnot vary linearly in the intervals 0 < t < aT and ( 1 - a) T < t < T . Instead,thephase varies like a raised cosine(foranexact definition, see b-functions, b = 0.05 in [16]). Fig. 6 shows d$in(h) for this binary system. The number 7. of observed bit intervals N = 1 , 2 , --, The upper bound on is the minimum Euclidean distance is also shownbyadashed linewheretheminimumEuclideandistancestilldoes not equal dB2(h).
CPFSK and Mary PSK

2.5

2.0

Fig. 4 shows the well-known minimum distance for binary

CPFSK for N = 1 , 2, 3 , and 4 observedbitintervals. Also shown in Fig. 4 is the upper bound dB2(h).It can be noted that h = 1/2 corresponds to MSK and gives dkin(1/2) = 2,

1.5
I v

which is the same as antipodal signaling, e.g., BPSK. The required observation interval for PSK is one bit interval, and for detectors making bit by bit decisions, PSKis optimum [3] . The required observation interval for MSK is two bit intervals, and the asymptotic performance in terms of error probability is the same as that for PSK. The optimum modulation index for CPFSK is h = 0.715 when the number of observed symbol intervals is 3. This gives theminimumEuclideandistance d$i,(O.715) = 2.43 and thus a gain of 0.85 dB in terms of E,/No is obtained compared to MSK or PSK. The minimum normalized squared Euclidean distance versus the modulation index h i s shown in Fig. 7 for the quaternary CPFSK system also (see [lo]). Notethattheupper bound dB2 ( h ) (shown by a dashed line where it not reached) is is twice the minimum distance for areceiver observation interval ofthe N = 1symbol.This is becausetherectangular frequency pulse g(t) hasthesymmetryproperty(37).The maximum value of d B 2(h) is approximately reached forN = 8 observed symbol intervals (compareto Table I). It is clear from (39) that the first-order weak modulation indexesintheintervalO<h<2areh,=1/3,1/2,2/3,1,4/3, 3/2, 5/3, and 2, and the effect of some of these early merges can be clearly seen in Fig. 7. Note that most of these weak

.5

1.o

Fig. 6 . Normalized squared minimum distances &h) versus modulation index for a b-function with b = 0.5. This phase function is verysimilar to the a-p function with a = 0.5, fi = 1/2 (see [16] or [181).

indexes are catastrophic. The minimum Euclidean distance for these is no better than 2. It is interesting to compare the minimum distance for the quaternary CPFSKsystem to QPSK (phaseresponse 41(t), h = 1/4) A indicatedin Fig. 7, theminimumsquareddiss tance for QPSK is dLi, = 2, and for the quaternary CPFSK system it is slightlybelow this value for h = 1/4. This is a different relative performance level than that for M = 2. For M = 2, h = 1/2 all schemes have the minimum squared distance d&in = 2, CPFSK and PSK included. The minimum distance for the octal(M = 8) CPFSK system is given in Fig. 8 N = 1 , 2 , 3 and in some interval for N = 4 and 5. The upper bound dB2(h), which, as usual, is shown by a dashed line where it is not reached, is like the quaternary case reached with N = 2 observed symbol intervals for low

AULIN AND SUNDBERG: CONTINUOUS PHASE MODULATION -PART I

205

Fig. 7.

'V
oJ

Fig. 8.

.5 1 .o Minimum normalized squared Euclidean distance versus modulation index for M = 8, CPFSK.

Minimum normalized squared distance versus modulationindex for M = 4 CPFSK.

HCS, RC, and Mary PSK

The HCS system yields a phase tree where the phase trajectories arealways raised cosine shaped over eachsymbol interval. Fig. 3 shows the quaternary case. The upper bounddB2(h)for HCS,M= 2 is given by [18]
dB (h) = 2( 1 - cos nh Jo (nh))

modulationindexes. Compared tothequaternary case, the number of firstarder weak (catastrophic) modulation indexes has increased in the interval of 0.3 < h < 1. Larger values of N are required to reach d B 2 @ ) ,compared to the quaternary and especially the binary case. The scheme 8PSK (ql(t),h = 1/8), previously shown to maximize dB2(l/8) is also indicated in Fig. 8 , and it is&en that octal CPFSK yields the same minimum distance if h is chosen slightly larger than h = 1/8 and if N = 2. Much larger distances can be obtained for the CPFSK system by choosing, for instance,h 0.45 and N > 5 . Distance properties of CPFSK schemes with larger values of Mhave been investigated i [17] , [20] . The maximum attainn able minimum distance value grows with M , but the number of first-order weak (castastrophic) modulation index values also grows with M , as does the length of the observation interval necessary for reaching the upper bound d B 2 @ ) .However, for h 5 0.3, N = 2 is sufficient for all M . As an illustration to the behavior discussed above, Table I shows NB which is a lower bound the on observation interval for reaching the upper bound.

(62)

where .To(.) is the Bessel function of the first kind and zero order. The maximum value of d B 2 @ ) smaller than that for is binary CPFSK, but this maximum value is still reached with N = 3 observed symbols (d&i, = 2.187 for h = 0.626, [18]). For h = 1/2, dkin(h) equals that of MSK and PSK, as in all binary full response CPM systems. In the region of 0 < h < 1/2, the upper bound is reached with N = 2 observed symbols as in binary CPFSK. Fig. 5 shows that HCS gives a larger minimum distance than binaryCPFSK in this region. The minimum Euclidean distances for the quaternary HSC system are given in Fig. 9 when N = 1, 2, 3, and 4 observed symbol intervals, and QPSK is also indicated. dB2(1/4) is still smaller than the minimum distance for QPSK of course, but since the upper bound forHCS is reached with N = 2 observed symbol intervals in theregion of 0 <h ,< 0.3, HCS is better than CPFSK. Note that the value of the minimum Euclidean distance at h = h, no longer equals 2 as for CPFSK. This k due to

206

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL.NO. COM-29,

3, MARCH 1981

Fig. 10.
0

.5

1 .o

.5 1 .o Minimum normalized squared Euclidean distance modulation index h for M = 8, HCS system.

versus

Fig. 9.

Minimum normalized squared distance versus modulationindex h for M = 4, HCS system.

commonly used function [ 141, [15]

the pulse shaping. The results for the octal case (see Fig. 10) follow the same trend as for the CPFSK system; i.e., the upper bound is reached with N = 2 for low h-values and the number of catastrophic modulation indexes has increased compared to the quaternary case. V. POWER SPECTRUM

g
pOb(B>=

Rodndf

which gives the fractional out of band power at the one-sided bandwidth B,will also be given. The power spectral density for the responseCPM schemes full Fig. 11 shows the power spectra (double-sided) for M-ary considered in this paper can be calculated with formulas given CPFSK with h = 1/M, M = 2 , 4, 8. The corresponding fracin [2]. The data symbols are assumed to be independent and tional out of band power plots are shown inFig. 12 (for other identically distributed.'For the case of full response CPFSK values of h , see [17], [20]). It is well known that for f x e d M systems, the spectrum can be expressed directlyin terms of and g(t), the spectrum widens for increasing h. For certain h elementaryfunctions [ 2 ] andfor RC systems intermsof values discrete componentsoccur. Fig. 11andthedistance Bessel functions [ 121 . figures illustrate the fact that for a roughly fixed distance, the Spectra for systems with different values of M should be spectralmain lobe is decreasing with increasing M. The becompared at the same bit rate. The bit ratenormalized variable havior of the spectra for large frequencies (i.e., the spectral f.Tb is used where tails) depends only on the number c of continuous derivatives of the instantaneous phase. It is shown in [8] that the tails decrease withfas I f I-'(' '1. For CPFSK, c equals 0. Fig. 13 shows the spectra for the quaternary CPFSK, HCS, and RC schemes for h = 1/4. Note that for increased values of Hence, the power spectra R o o are plotted against the bit rate c the main lobe becomes larger. The main lobe widens intuinormalized frequencyseparationfrom carrier.Plotsof the tively due to thepresence of higher phase slopes over a portion
' p
+

AULIN AND SUNDBERG: CONTINUOUS PHASE MODULATION-PART I

10.lolog

207

Ro(fl

[dB1
1

-10

-20

-30

-40

-50

-60

-70

-8C

Fig. 11. Normalized power spectral density in Sxibels for M-ary (M= 2, 4 , and 8) CPFSK with modulation indexesH = 1/2, 1/4, and 1/8, respectively.

1 .o

1.5

-50

1
Fig. 12. Fractional out of band power in decibels for M-ary (M= , 4 , 2 and 8) CPFSK with modulation indexes h = 1/2, 1/4, 1/8, respectively.

of the pulse for non-CPFSK schemes.Thespectraltail of HCS d f in (close to 4) are reached with plateau functions with behaves like f 6 , and like f 8 for RC. Further spectra for M = 2 and h close to 1. For h = 1, the power spectrum con[20] . tains spectral lines however. these schemes are plotted [19], [17], in The spectra of schemes with plateau functions are investiVI. DISCUSSION AND CONCLUSIONS gatedin [16], [18]. As might be expected, the rapid phase change in the beginning and the end of each symbol interval From the distance and spectrum results above and in [17] , gives wide spectra. The previously mentioned spectral tail [ 191 , [20] , it is evident that M-ary full response CPM schemes behavior versus c is alsoapplicabein this case. However, f have both bandwidth compaction properties and yield gain in must be impractically large before this asymptotic behavior is Eb/No as compared to MSK. Schemes within this class of CPM dominating. Furthermore, it was concluded above that large systems can also be designed to give a large gain in Eb)No with

208

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO. 3 , MARCH 1981 TABLE I1 BANDWIDTH/DISTANCE TRADEOFF FOR SOMEM-ARY CPFSK SYSTEMS
b
I

0 t

0.5

1
*

1.5 .

'

-10

P i
I

i CPFSK
scheme

Bandwidth 2E.T

Gain
MSK, dB

90%

99%

99.9%
2.78 1.20 0.78 2.0 1.45

syrlbols

\h

M=2 h=.5 M-4 h=.Z5

M=8 h=.125
M=4 h=.40

0.42

0.80 1.42
0.54
1.18 1.00 1.40

-1 .?8

0.30

0.96
2.08 2.20 1.76 1.04 2.36 1.08

.60
3.0P
3.60
3.0

0.68
0.76
0.70

M=4 h=.45

N=8 h=.30
M=8 h=.45

2
4.31

5.40

Gain i n dB r e l a t i v e t o PSK

"1

I
Fig. 13. Normalizedpower spectral densities in decibles for quaternary CPFSK with modulation index h = 1/4. The schemes are CPFSK (solid line), HCS (dashed line), and RC (dash-dotted line).

4 -

3 2 .

P
0.5

n
n

1 0

-1

ll.0
E

,.
Y

k ,
NSK

1.5

2.0

the same bandwidth as MSK, or considerably smaller bandwidth at the expense of an increased Eb/No. This holds, for example, for M-ary CPFSK. The same also holds for systems like HCS and RC, and in these cases the tradeoffbetween bandwidthand gain in E,/No at large SNR is even more attractive. In the binary case, plateau functions are a way t o achieve considerable gains intermsof Eb/No, whichunfortunately gives poor spectra. In Table I1 comparisons between various CPFSKschemes are. made, both concerning bandwidth and gain in terms of Eb/No (dB) at large SNR. The'reference system is MSK. Three different definitions of bandwidth will be used. The normalized bandwidth (double-sided) .is defined at 2BTb, for which 90, 99, or 99.9 percentofthetotal signal power is within the frequency bandIf - fo. I <B . Table I1 also gives thenumber of observed symbols NB required t o reach the given .&nimurn squared Euclidean distance value. The quaternary scheme with h = 0.45 has approximately the same bandwidth as MSK (99 percent bandwidth) and yields' a gain in Eb/No of 2.56 dB. The octal'scheme with h = '0.45 gives a slight bandwidth expansion when compared to M-SK (at 99 percent bandwidth), but gives the gain 4.31 dB in terms of , ? / O !,N. A more exhaustivecomparison betweendifferent M-ary CPFSK systems can be found in Fig. 14. In this figure the gain in decibles of various schemes is shown versus the 99 percent bandwidth. The schems are binary (indicated by x), quaternary (indicated by 0 ) and octal (indicated by 0 ) . Note the supe-

-8 -7

x: M=2
m : M=4
0:

M=8

Fig. 14. Bandwidth/performance comparison relative to MSK for various CPFSK systems (99 percent fractional out of band power bandwidth, see Table 11).

rior performance of quaternaryandoctal schemes. Inthe binary case i t was possible to find the frequency pulses g(t) maximizing the upper bound the on minimum Euclidean distance. This was not the case for multilevelsystems, and it is believed that the optimum frequency pulse-depends on M and h. In the interval 0 < h < l/M,M-ary PSK was shown to yield the largest minimum Euclidean distance, but the HCS and RC schemesare notfarfromthisoptimum. However, the two latter schemes have much smaller spectral tails than M-ary PSK. It was shown that the number of f i r s t a d e r weak modulation indexes grows with M , thus putting a practical limit on how large a n M should be chosen.

PART TINUOUS AND UNDBERG:

AULIN

209
T. A. Schonhoff, Bandwidth vs performance considerations for CPFSK. in Proc. IEEE NationalTelecommun.Conf.Record, 1975. pp. 38.1-38.5. F. Amoroso.Pulseandspectrummanipulation in theminimum (frequency) shift keying (MSK) format, IEEE Trans. Commun.. vol. COM-24, pp. 381-384, Mar.1976. T. A. Schonhoff, Symbol error probabilities, for M-ary CPFSK: Coherent and noncoherent detection. IEEE Trans. Commun.. vol. COM-24, pp. 64-52. June 1976. M. Simon, K. A generalization the of minimum-shift-keying IMSK)-type signaling based upon ,input symbol data pulse 845-856. shaping, IEEE Trans.Commun.. vol.COM-24,pp. Aug.1976. M . Rabzel and S . Pasupathy. Spectral shaping in minimum shift keying IMSK) type signals. IEEE Trans. Commun., vol. COM26.pp, 189-195, Jan.1978. T . Aulinand C-E.Sundbern,BinaryCPFSKtype of signaling with data input symbol pulse shaping-Error probability and spectrum. Telecommunication Theory, Techn. TR-99, Rep. Univ. Lund, Lund, Sweden, July 1978. __ , M-aryCPFSKtypeofsignaling with inputdatasymbol pulse shaping-Minimum distanceandspectrum,Telecommunication Theory, Techn. Rep. TR-I 11, Univ. Lund, Lund. Sweden, Aug. 1978. -. Bounds on the performance of binary CPFSK type of signaling with input data symbol pulse shaping, in Proc. IEEE Nut. Telecommun. Conf. Record, Birmingham, AL, 1978. pp. 6.5.1-6.5.5. , M-aryCPFSKtypeofsignaling with input datasymbol pulse shaping-Minimum distance and spectrum, in Pioc. IEEE I n t . Conf. Commun. Conf. Record. Boston. MA, 1979, pp. 42.3. I42.3.6. T. Aulin,CPM-A power and bandwidth efficient digital constant envelope modulation scheme, Techn. Dr. dissertation, Telecommunication Theory. Univ. Lund. Lund. Sweden. Nov. 1979. T . Aulin, N. Rydbeck, and C.-E. W . Sundberg, Continuous phase modulation-Part 11: Partial response signaling, this issue, pp. 210-225.

It is interesting to note that for the schemes considered in this paper, a gain in terms of Eb/No is obtained without expanded bandwidth, compared to MSK. This is different from the case with a channel coded MSK system, where the spectrum must be expanded by a factor of 1/Rwhere R is the code rate [3] , [9] . For the CPM systems, no paritysymbolsare transmitted, and the totalsignal energy is devoted to the information symbols. This paper explores the distance and bandwidth properties of full response CPM systems. In spite of the restriction that the schemes must be afull response type (i.e., the instantaneous frequency only dependson one data symbol), we have found considerableimprovements. However, larger improvements are obtainable with partial response systems (the instantaneous frequency depends more one on than data symbol). This class of system is considered in part 11. We have intentionallyomitted all problems dealing withtransmitter and receiver considerations. These problems will be treated in a unified manner in Part 11.

REFERENCES
V. A. Kotelnikov, The Theory of Optimum Immunity. Noise New York: Dover, 1960. R. R. Anderson and J. Salz, Spectra of digital FM, Bell Syst. Tech. J . , vol. 44. pp. 1165-1 189, July-Aug. 1965. J . M. Wozencraft and 1.. M., Jacobs, Principles of Communication Engineering. New York: Wiley. 1965. R. W. Lucky, J. Salz. and E. J. Weldon.Jr. Principles of Data Communication. New York: McGraw-Hill, 1968. M. G . Pelchat. R. C,. Davis, M.Luntz. andB. Coherent demodulation of continuous phase binary FSK signals. i n Proc. Inr. Telemetering Conf., Washington, DC, 1971, pp. 181-190. R . deBuda, Coherent demodulation of frequency-shift keying with low deviation ratio, IEEE Trans. Commun., vol. COM-20, pp. 429-436, June 1972. P. and B. Luntz, Coherent noncoherent and r71 W. Osborne M. detection of CPFSK. IEEE Trans. Commun., vol. COM-22, pp. 1023-1036,Aug.1974. T. J. Baker, Asymptotic behaviour of digital FM spectra. IEEE Trans. Commun.. vol.COM-22,pp, 1585-1594. Oct. 1974. W. Lindsey M. Simon, C. and K. Telecommunication Systems Engineering. Englewood Cliffs. NJ: Prentice-Hall,1974. T. A. Schonhoff. Symbol error probabilities for M-ary coherent continuous phase frequency-shift keying (CPFSK). in Proc. IEEE Int. Conf. Commun. Conf. Record. San Francisco, CA, 1975. pp. 34.5-34.8.

*
Tor Aulin (S77-M80). for a photograph and biography, see this issue. p . 195.

Carl-Erik W. Sundberg (S69-M75), for a photograph and biography, see this issue, p. 195.