96 IEEE TRANSACTIONS ON ACOUSTICS,SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-31, NO.

1 , FEBRUARY 1 9 8 3

Random Modulation: A Review

ATHANASIOS PAPOULIS, FELLOW, IEEE

Abstract-A unifiedtreatment of variousaspects of random modula- Basic Concepts: Given two real random processes a ( t ) and
tion is presented with emphasis on the spectral properties of FM signals. b ( t )and a constant w o , we form the complex process
The study includes reexamination of Rice’s representation, Woodward’s
theorem, FSK, andprocesseswithnormal, Markov, and cyclostation-
z ( t ) = [a(t) f j b ( t ) ] e j w o t = x ( 9 iu ( t )
f (1)
ary frequencies.
where
I. INTRODUCTION x ( t ) = a ( t ) cos w 0 t - b ( t ) sin m o t
Y ( t ) = b ( t ) COS w o t + a ( t ) sin wet.
M ODULATION isfundamental
a topic
with applications
in many areas. A basic topic in modulation is the rep-
resentation of a stochastic process as an AM or FM signal and
(2)
It is well known [ 11 that, if the processes o(t) and b ( t ) are
thedeterminationofitspowerspectrum. In this paper, we wide-sense stationary (WSS) with zero mean and their correla-
review and clarify anumber of conceptsunderlyingsucha tions are such that
representation,
and we discuss in some detail the
spectra of R,,(T) =Rbb(T) Rub(T)= -Rba(T) (3)
stochastic FM processes. Thepaper is essentially tutorial. then the processesx(t)andy(t) are also wss with
However, the material is developed with some originality and
it includes several new results. Rxx ( T ) = R , (7)= R,, ( T ) cos w oT f Rub( T ) sin w oT
In the first section, we present for easy reference the funda-
mentals of randommodulation [ l ] - [ 4 ] . In Section 11, we
determine the class of envelopes and carrier frequencies that
can be used in the representation of random processes as AM
or FM signals and we derive Rice’s representation [5] as a
special case satisfying certainoptimality criteria [6]. In the
last three sections, we consider the problem of determining the
power spectra of FM signals. This problem is, in general, dif-
ficult. We show,however, that three important special cases
lead t o reasonableresults. InSection 111, wegive a simple
proof of Woodward‘s theorem [7], [8] and we analyze normal
processesdevelopingexplicitexpressions fortheasymptotic
form of their spectra. In Section IV, we prove that, if the in-
stantaneous frequency c ( t ) of a signal is a finite stateMarkov
process, then its power spectrum is a rational function that can
be determined in terms of the transition probabilities of ~ ( t ) .
The result is based on the well-known Chapman-Kolmogoroff
equation [ l ] , however, its derivation is not obvious. The last
section deals with PAM, FSK,and PSK processes [3]. This
topic has been treated extensively in recent years [lo] -[12]
but the known results are rathercomplex. In ourapproach,
we derive relatively simple expressions based on stationary
models obtained with a random shift of the origin. The anal- where
ysis involves the properties of strict-sense cyclostationary pro-
cesses [13], [ 141. We review briefly the underlying concepts
extending the results to the strict-sense case.
Hence,
Manuscript received December 30, 1981; revised July 13, 1982. This
material is taken from the forthcoming second edition of the author’s x ( t ) = Re [ w ( t ) eiwot] = r ( t ) cos [ m o t t ~ ( t ) .] (13)
book Probability, Random Variables, and Stochastic Processes. It was
supported by DARPA under Contract N00014-76-C-0144, and in P a t we have, thus, represented x(t) as a modulated signal with
by JSTAC under Contract F49620-80-C-0077.
The author is with PolytechnicInstitute of New York, Farmingdale, real envelope r ( t ) ,complex envelope w(f), and instantaneouS
NY 11735. [2] frequency

0096-3518/83/0200-0096$01.OO 0 1983 IEEE

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 PAPOULIS:RANDOMMODULATION 91

Clearly, S,, ( w ) = S,, (a+ w,), hence,

r m

rn = J -m
( w - wo)2 ~ , ( o ) d o

Thus, our problem is to determine wo a n d y ( t ) so as to min-

imize the moment of inertia rn with respect to wo of the spec-
trum S , (w).
q ( t )= wo +cp'(t). (14) Suppose, first, that y ( t ) is specified. In this case, S , ( w ) is
also specified and, aswe know, m is minimum if wo is the
As we shall presently see, this representation is not unique. center of gravity
We note that
z'(t) = r'(t) t j q ( t ) r ( t ) ]e i [ w o t + ~ p ( t ) l= x ' ( t )+&'(t).
From this it follows that
of S , (a).Since
U i ( t )= [x(t)y'(t)- x'(t)v( 9 1 / r 2 ( t ) (15)

sm
S z z (a)= 2SXX ( w ) + 2 4 , (0)
1 -
E{r2(t) id&)} = - j E { z ' ( t ) z " ( t ) }= - wS,,(w)dw S,, (- w ) = S , ( w ) and B,, (- w ) = - Bxy ( w ) (20)
2n
we conclude that the optimumw o is given by
(16)
and
0 0 = ~ B , y ( W )dw/J- S, ( 0 )dw. (21)
E{r(t)r'(t)}= 0.
11. OPTIMUM ENVELOPEAND RICE'S REPRESENTATION The resulting rn is the central momentof inertia
In the preceding section, we were given the carrier frequency m

wo and the two processes a(t) and b(t). We consider now the
converse problem: Given a WSS process x ( t ) with zero mean
and an arbitrary constant w o ,find two WSS processes a(t) and
b(t) such that =4 (w2 - a;) S,, ( w ) dw i-
~ ( t= u) ( t ) cos m o t - b ( t )sin w o t . (1 7)
of Szz (a).
To solve this problem, we construct a process y ( t ) with zero With wo so determined, it remains to findy(t) so as to min-
mean and such that imize m. Since S , ( w ) is given, m is minimum if wo is maxi-
mum. This is the case if the integral
R,, (7)= R y y (7)

that the processes

R x , ( 7 ) = - R x, ( 7 ) .
From the discussion of Section I, it follows mutatismutandis

a ( t ) = x ( t ) cos oott y ( t ) sin w o t

(1 8)

I= sm m

aBxy(4

is maximum. Aswe know, IBxy(w)I < S,, (w), hence, I is

b(t) = y ( t )cos o o t -x ( t ) sin mot (1 9) maximum if

are WSS andx(t) is given by (1 7). B,, ( 0 )= S,, (0) for w > 0. (22)
The representation of x ( t ) as a modulated signal of the form We, thus, conclude that rn is minimum ify(t) is so chosen that
(17) is not unique. First,thefrequency wo is chosenarbi- (Fig. 2)
trarily. Second, even if wo is specified, the processesu(t) and
b ( t ) are not unique because they depend on the processy(t) -jSxy (a)= S , (w) sgn w.
and t h ~ sprocess can be chosen arbitrarily subject to the con- Inserting into (20), we obtain
dition (18). We could,forexample, select y ( t ) suchthat
R,, ( 7 ) = 0.
Among all possible representations, we shall use the one that
minimizes the integral
The optimum w,, is, thus, the center of gravity of S, (a)*
-
.
,

rn = I_ wzs,,(w) do
U(w)and the resulting rn is its central moment of inertia.
We note that
m
subject to the constraint that S , (a)is given. At the end of
thesection, we comment on the reason for thischoice [6].
2nE{lw'(t)12} = J -m
o2Sw,(w)do (24)

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 98 IEEE
TRANSACTIONS ON ACOUSTICS,
SPEECH,
ANDSIGNAL
PROCESSING, VOL. ASSP-31,
NO. 1, FEBRUARY 1983

In Rice’s representations theinstantaneousfrequency is

given by
q ( t )= [ ~ ( it’ () t ) - x’(t) $(t)] /r2(t). (29)
In general, it is difficult todeterminethe statisticalprop-
erties of wi(t). However, if the process x(t) is normal, then
the first order densityf(o) of q ( t )is given by [ 161
Fig. 2. Spectra of a signal with optimum envelope.

because the power spectrum of w ’ ( t ) equals w 2 S w w ( w ) .

From this and (23) it follows that the minimum m equals
where a = m / 2 ~ R , ,(0).

MODULATION
111. FREQUENCY
We shall represent a frequency modulated signal in the form
= - 2 [R:, ( 0 ) + R,, (O)] . (25 1
~ ( t= )cos [mot+ hcp(t)+pol = Re z ( t ) (3 1)
Rice’s Representation and Hilbert Transfoms: The Hilbert
transform ?(t) of a process x(t) is the output of asystem where
[15] (quadrature filter) with input x(t) and system function t
- j sgn o. Hence, p(t)= ~ ( o ldol.
)
s;, ( w ) = sxx (a) 0

In the above, hv(t)t po is the phase, w o + h c ( t ) the instan-

;S
, ( w ) = -S;, ( w ) = jS,, (a)sgn w. (26)
taneous frequency, andh the modulation indexof x ( t ) .
This shows that the signal y ( t ) = $(t)satisfies (18), therefore, We maintain that, if ~ ( tis) astrict-sense stationary (SSS)
the processes process and cpo is a random variable independent of c ( t ) and
a(t) =x(t)cos oott $(t) sin w o t such that
E(JlP0) =E{ei2Po} = 0 (33)
b(t) = ?(t) cos oot- x(t) sin w o t (27)
thenthe process x(t) is WSS withzeromeanandautocor-
are WSS and x(t) is given by (1 7). The resulting expression is
relation
known as Rice’s representation [5] of x(t). The correspond-
ing complex process z ( t ) = x(t) + j?(t) is the analytic signal
(34)
associated withx(t).
From ( 6 ) and (26), itfollows that Since

S
,, (w) = 4Sxx (a) U(w). x(t) = 4 [z(t)+z*(t)]
This equals the spectrum of the signal that minimizes m [see it suffices to show that the processesz(t) andz*(t) are jointly
(23)] and it shows that Rice’s representation is optimum in wss.
the above sense. From (33) it follows that
Wegive nextthree reasons for selecting as optimality cri-
E ( z ( t ) } = E { ej[wot+lP(t)I 1EceilPo} = 0
terion the minimization of m.
j[wo(zt+~)+Xlp(t+7)+hq(t)]
1) In the representation o f x ( t ) as a modulated signal, it is E{z(t+T)z(t)}=E{e } E{ejZPo1 = o
reasonable to choose as carrier frequency the center of gravity
Furthermore, for anyt , t + T > 0
of S, ( w ) U ( w ) and as envelope a process whose bandwidth
is as small as possible. Our optimum meets these requirements
if we use a measure of the bandwidth of w ( t ) the normalized
second moment of, ,S (0).
2) A desirable property of the envelope of x(t) is the mini-
mization of its average rate of change 161. This is consis- and (34) results because
tent with (23) if we use as measure of this rate the mean of R, (7)= 3Re R , (7).
I w ’ ( 6 l2 *
3) The optimum carrier frequency w o equals the weighted The last equality in (35) is a consequence of the stationarity
average of the instantaneous frequency w i ( t )ofx(t): of c(t). The angle po is introduced to establish thestation-
arity of x(t).
0 0 = E { r 2 ( t ) oj(t)}/E{‘2(t)}. (28) From the above we conclude that, for the determination of
This follows readily from (15) and (16)because the spectral properties of the real bandpass process x(t), it
suffices to find the mean

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 PAPOULIS: RANDOM MODULATION 99

of the complex low-pass process From this it follows that

w(t> = e j h d t ) . c ( t ) = c(O), I t I <70, (44)
This process is the solution of the first order stochastic differ- Hence,
ential equation
w'(t) - jhc(t) w ( t ) = 0 p(t) = J r C(.) dor = tc(O), I tl < T o . (45)
0
and it has the property that its mean equals its autocorrelation
Denoting by f c ( c ) the first order density of c ( t ) and by
(37)
The autocorrelation of w ( t ) can be expressed in terms of the
first order characteristic function
a,+,
(A ; t ) = E {ejh9(t)) (3 8) thecorrespondingcharacteristicfunction, we concludefrom
(45) that the characteristic functionof q(t) is given by
of the phase p(t). indeed, as we see from (37)
@,+,(A; t ) = E{ejhtc(o)}= Q.,(A~), I tl < T ~ . (47)
Rww(~)=@~(A;~). (39)
This leads to the conclusion that[see (39)]
For the determination of the spectrum ofx(t), it suffices,
therefore, to evaluate theFouriertransformwith respect t o R w w ( 7 ) = @c(A7), 171 <70. (48)
7 of the function @,+,(A; 7). The solution of this problem de-
This result cannot be used to determine the power spectrum
pends on the form of the available.data. of ~ ( tbecause
) it holds for small 7 only. Suppose, however,
We shall say that a process ~ ( tis) phase modulated if the that
statistics of p(t) are known. In this case, Rww(7)is obtained
directly from (39) and the remaining problem is the determi- @c(p)=O for IPI>PO (49)
nation of its transform. (Fig. 3). In this case, the two sides of (48) are nearly equal for
We shall say that ~ ( t is) frequency modulated if the available every 7 if the modulation index h is larger than po / T ~ .Under
information is expressedin termsoftheinstantaneous fre- this condition, the power spectrum ofw(t) equals
quency of x(t). The resulting problem is no longer simple be-
cause there is, in general, no simple relationship between the
characteristic function of p(t) and the statistics of c(t). As
we show presently, however, the normal case is an exception.
This follows if we take transforms of both sides of (48) and
Since c ( t ) is SSS and
use the inversion formula
t
dt)= j-44
0
da
fc(c) =
1
I_ -
@c(P) e-ipc dr-l
we conclude that
with p = AT. Since [see (34) and (38)]
E I d t ) ) = t E { c ( t ) ) = q&) (40)
and
R,,(T) cv 3 Re [eiwor@,+,(h;7)] (5 1)
we conclude that the power spectrum ofx(t) equals
h I i ( t ) } = ~ -' tR c ( u ) ( / i / - a ) d o r = n ~ ( f ) f o : ( f(41)
)
where
Rc(7) = E { c ( t f 7 ) c ( t ) } (42) This result is known as Woodward's theorem. It holds for any
FM signal of the form (31) under the condition that the mod-
is theautocorrelation of c(t). If ~ ( tis) normal,then this ulation index is sufficiently large.
informationdetermines Qq(A; t ) because then p(t) is also Normal Processes: We now assume that c(t) is a normal
normal and process with zero mean. In this case, ~ ( tis )also normal with
@,+,(A; t ) = exp ( j k q q ( t ) - 3A' o i ( t ) ) . (43) variance

We now discuss several special cases starting with the asymp-

totic solution of thegeneral problem.
Woodward's Theorem/ 7 / , /8/ : Suppose thatc ( t ) is an arbi-
o;(t) = 2

Thisfollows
I' &(a) (t - a) da,

from(41)
t > 0.

because E ( q ( t ) } = 0 and RC(-7)=

(53)

trary real SSS process with autocorrelation Rc(7). IfRc(7) is

-
RC(7).
continuous, then Inserting into (39) andusing (43), we obtain
E { [c(t + 7)- C(t)]2} = 2 [ R c ( 0 )- Rc(7)] 0.
710

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 100 IEEE TRANSACTIONS ON ACOUSTICS,
SPEECH,
ANDSIGNAL
PROCESSING,
VOL.
ASSP-31, NO. 1, FEBRUARY
1983

Fig. 3. Correlation and characteristic function of instantaneous f r e Fig. 4, Significant time constant of R,(T).
quency.

In general, the transform of the above can be determined only

numerically. However, the followinglimiting cases yield ex-
plicit results.
We shall determine the form of S,,(w) for small and large
values of X . For this purpose, we introduce the constant
r o

lm
0 W

Fig. 5. Asymptotic forms of corrclation and spectrum of an FM signal.

7, = Rc(T) dT/P, P =R,(O).

This constant can be interpreted as the duration ofR,(T), and

its inverse gives somemeasure of thebandwidth of the fre-
quency c(t).
We further select two constants ro and T~ such that (Fig. 4)

R,(T)=O for 171 > r l >>T,

n. N a r r o w - b a n d FM: Rational

Wtde-band FM: Gausstan

R,(T) = p for 171 < ro << 7,. W

P C 1'
Inserting into (53), we obtain
Fig. 6. Normal FM spectra for small and large h.

because c(t) is a normal r.v. with zero mean, variance p , and

density function

2
-ph TcT
, This agrees with Woodward's theorem.
R,, (7) =e r>r1.
(56) Narrow-Band FM: If X is sufficiently small, then [see (54)]
As we know [17], the behavior of R,,(T) for small (large) T
determines the behavior of S, ( w ) for large (small) w. Since R,,(T)=Rww(O)= 1
(Fig. 5)

,-PA~T~/~ - 16
X
e - ~ 2 / 2 p 2h (57)
From this it follows that, if ph2r1T, <<the
1, then second
approximationin (56) holds for every r. Hence, S,,(w)
equals therationalspectrum in (58) anditsbandwidth is of
order ofprCh2.
2
-PA TclT1 ti; 2prc X2 IV. MARKOV PROCESSES[9]
e
w2 + p 2 r 3 4 Incorporating X into c(t), we obtain
we conclude that S,,(w) is normal near the origin and it ap-
proaches a first order rational curve as w -.+a. We show next w(t) = , h ( t ) (5 9)
that these limiting cases give an adequate description of
S, ( w ) for large or small X.
Wide-Band FM: If X is sufficiently large, then [see (54)] where now we assume that c ( t ) is a continuous-time Markov
chain taking the N values yk with
R,,(T)=~ for IT~>T~.
P { c ( t >= Y k 1= P k
From this it follows that, if ph27; >> 1, then the first approx-
imation in (56) holds for every r. Hence, S,,(w) equals the P{C(t T) = yk I C ( t ) = y j } = T j k ( T ) , 7 > 0.
normal curve in (57) anditsbandwidth is of theorderof Our objective is to find the function
h 6 (Fig. 6).
R,,(T)=E{W(t+T)W*(t)}=E{W(7)}.
We note that in this case
For this purpose, we introduce the conditional correlations
R i k ( 7 ) = E { W ( 7 ) 1 C ( O ) = Y i , C ( r ) = Y k } 71ik(7)-

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 PAPOULIS: RANDOM MODULATION 101

and fori # k

Inserting into (65), we obtain

R : k ( Tt)Q=!( kj Yk k) R
&im
k (k7R) ti r n ( 7 ) . (67)
m#k

This is a system of differential equations yielding R i k ( 7 ) . In-

serting into (60), we obtain R,,(T). Since (67) has constant
coefficients, we conclude that the spectrum of the FM signal
x ( t ) is rational.
Illustration: Supposethat c(t) is random telegraph signal
taking the values + A . Such a process is completely specified
in terms of the transition probabilities
A = A } 2: 1 - & A t
n 1 ( A t ) = P { c ( A t ) =Ic(0)
~21(At)=P{c(At)=A Ic(O)=-A)2:pAt
a,, = - a a12 =Q! a21 =p a22 = -0.
Fromthestationarity of c ( t ) it follows thatthelast term In this case,
above equals

With i = 1, the system (67) yields

R ' , l ( T ) = ( i A- ~ ! ) R 1 1 ( 7 ) + 0 R 1 2 ( 7 )R 1 , ( 0 ) = 1
R i 2 ( 7 ) = a R 1 1 ( 7 )(-j A + f l ) R 1 2 ( 7R) 1 2 ( O ) = O .
Thefunctions R Z 1(7) and R z 2 (7) satisfya similar system.
and (6 1) results. Denoting by Si+,(s) the unilateral Laplace transform of R i k (I-),
We next differentiate (61) with respect to v and set v = 0. we conclude that
This yields

R:*k(7) =E
m
R i m (7)R A k (O+)- (65)

Thus, to complete the determination of Rik(t-), we must find

Rik(O+) and R:.k (0').
Initial Conditions: As we know Hence [see (60)]
1, i=k
"ik (0') =
t 0, ifk.
Finally
, ,S (G) = S:, ( j w ) t SL, (-io).
for i f k . For small 7,the probability that the state of c ( t ) If, a = P, then D(s) = s2 t 2as t A 2 . In Fig. 7, we show the
changes once in the interval (0, 7) is of the order of 7 and the location of the roots of D ( s ) for various values of the ratio
probability that it changes more than once is of the order of a / A , and the corresponding spectrum, ,S (0).

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 102 IEEE
TRANSACTIONS
ONACOUSTICS,
SPEECH,
AND
SIGNAL
PROCESSING,
VOL.
ASSP-31, NO. 1 , FEBRUARY
1983

a < <A a = 0.6A a>A

t
Fig. 7. Spectrum of a two-state Markov FM signal.

V. CYCLOSTATIONARY PROCESSES, PAM, AND FSK Furthermore, since 8 is independent of x ( t )

In this section we show that the determination of the spec- p ( 1~0 = e ) = p { x ( t l + E - e < x l , . . . ,x(t, + E - e)
tra of FSK signals is simplified if their instantaneous frequency
is modeled as a randomly shifted
cyclostationary process. As = ~ ( x , , . . - , x , ; +t E~ - e , - , t , + E - e).
we recall [ I ] , [2], a random process x ( t ) is called strict-sense Inserting into (721, we obtain
cyclostationaryifit has the same statistics as the process
x(t + T). From this it follows that the distribution function
of x ( t ) of anyorder does not change if the origin is shifted = F(xl ' ' '1 3 'E -3 .' ' t,
+ ' 9 9 E- '1
by T
and (71) results [see ( 6 8 ) ] .
Theorem 2: If x ( t ) is WS cyclostationary, then i ( t ) is WS
(68) stationarywith mean
Hence, the mean ~ ( tof) x ( t ) is periodic and
its
autocorrela- T
tion R ( t l, t z )is invariant to a diagonal shift of theorigin (73)
R ( t l + nT, tz + n T ) = R ( t l, t z ) . (69)
and autocorrelation
Aprocess x ( t ) is called wide-sense cyclostationary if the
functions ~ ( tand) R(t + r , t ) are periodic in t. T
If the origin of a cyclostationary process is subjected to a ~ (= ~ 1 R ( +~r, t ) dt. (74)
randomshift e ( t ) and 8 ( t ) varies sufficientlyslowly, then
the resulting process x [t - O(t)] is nearly stationary. The fol-
lowing theorems [ 131, [ 141 give a precise modelfor this
Proof: Since 0 is independent ofx(t)
equivalence. ~ { ~ e( ) tl =- E { E { ~ ( te)le)]
-
Theorem: If x ( t ) is a SS cyclostationary process and 8 is a
T
r.v. uniform in the interval (0, Tindependent
then the process
) and of x ( t ) , 1
=E{q(t-e)}=T J
0
v(t- e)de
.ap(t)= x(t - e) (70) results
and (73) because
is SSS and its nth order distribution
equals rl(t + T ) = Q(0.
Similarly

. rT

Pro05 To prove the theorem, it suffices to show that the

probability of theevent and (74) results because

A = { i ( t l + E ) < X I , . . . , i ( t , t E ) <x,} R(t+T+7,t+T)=R(t+~,t).

is independent of e and it equals the integral in (71). Aswe Pulse-Amplitude Modulation (21, [3]: Animportantex-
know ample of a cyclostationary process is the PAM signal

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 PAPOULIS: RANDOM MODULATION 103

Fig. 8. PAM signal and correlation.

where h ( t ) isanarbitraryfunctionand 4n is a stationary dis- b(t)= b(t - 0) = v ( t - 0 ) * h ( t ) = V ( t ) 8 h ( t ) (81)

Crete-time process with autocorrelation
is the output of a linear system with input G(t) and impulse
Rm = E C q n + m q n 1. response h (t).
it this
From follows that
Using the preceding results, we shall determine the autocor-
relation of the stationary process 6(t) = b(t - e). We assume,
first, that h ( t ) is a rectangular pulse of area 1 and duration e
as in Fig. 8. The resulting process be(t)is a sequence of rect- where p ( f ) is the inverse of the energy spectrum IH(jw)I2 of
angular pulses of area q n . We shall determine its autocorrela- the system. This yields
tion Re(t + r, t ) for 0 < t < T. For this purpose, we write r
1
in the form Sb (0)= - S, (e'"T) I~ ( j wl2). (83)
T
T = w z T + T ~ ,1 ~ ~ 1 < T / m:integer.
2
FrequencyShift Keying /1OJ-[12]: Suppose now that the
As we see from the figure, if 0 < T~ < e and t < e - r l , or instantaneous frequency c ( t ) in (59) equals
if - e < r1 < 0 and - r l < t < e thentheautocorrelation
Re ( t + r, t ) of be(t)equals ~ ( t= )8(t)= b(t - e)
1 1 where b(t) is the cyclostationary process in (75). To simplify
R e ( t + r , t ) = ~e E { 4 0 ~ m 1 = ~ R m (76) the analysis, we assume that the r.v. qn are independent and
we denote by @(p) their characteristic function
and it is 0 otherwise. In both cases, the length of the t-interval
for which (76) holds equals e - I T~ I. Hence, the autocorrela- = E { e j p p n 1. (84)
tion of the stationaryprocess be (t)is given by As we see from (37), to determine the spectrum of the pro-
cessx(t), it suffices to find the mean
T

for 171 I < e, and it is 0 otherwise. Thus, Re(r) is a train of

E { w ( t ) }= E { E { w ( t ) I O } } = $J E{w(t)IB}dO
0
(85)
triangular pulses of area R , / T .
The limit of be (t)as e + 0 is a random impulse train of w(t). We consider next various special cases where for con-
venience we write the variable r in the form
"
(78) T = pT + TO 0 < ro < T m: integer
and we assume that T > 0.
Theautocorrelation R v ( 7 ) of thecorrespondingstationary
Continuous-Phase FSK: Suppose, first, that h(t) is arect-
process u(t - 0 ) is the limit
angular pulse of duration T as in Fig. 9. In this case
1 " (79) b(t)=q, for n T - T < t < n T
R,(T) = lim &(r) = - R , S(T- m T ) .
e+O T m=--m
hence, the phase
This is an impulse train with Fourier transform

1
is continuous and it consists of line segments with slope qn.
Furthermore, if T > T , then
Thus, Ts,(w) is the DFT of R,.
Returning to the general PAM process ( 7 9 , we observe that
03

b(t) = t )-= v(t) * h ( t )

~ ~ h (n T
n=-m

Hence,

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 104
IEEE
TRANSACTIONS ON ACOUSTICS,
SPEECH, AND SIGNAL
PROCESSING, VOL. ASSP-31, NO. 1, FEBRUARY
1983

+T-+

T'mT+To 4 Fig. 10. Discontinuous-phase FSK: c(t)-frequency, q(t)-phase.

Fig. 9. Continuous-phase FSK: c(t)-frequency,&)-phase.

andifO<r<T
sa(r)= {;;;,+ (7 - 8 ) Q1, 0 <T
e >r.
The above leads to the following conclusion. If T > T, then
Fig. 11. DPSK: &)-frequency, &)-phase.

If 0 < r < T, then

I Y
Inserting into (85), we obtain
Fig. 12. (a) Correlation of DPSK; (b) correlation of PSK.

Differential Phase-Shift Keying: Suppose, finally, that p(t)

is a staircase function (DPSK) as in Fig. 11, where

Tqn = q n - p n - 1 -

T>T We now have

[lT
@(e) @(r- 19)dB + ( T - I-) @ ( T ) , r < T.

Discontinuous-Pkase FSK: We now assume that p ( t ) con-

sists again of line segments but it is discontinuous at the points
t = nT + 6 with jumps equd toTq,. In this case, c(t)contains
impulses at these points (Fig. 10). As we see from the figure,

Hence [Fig. 12(a)]

hence, as in the previous case,

Pkase-Skift Keying: In the above, q, is a sequence of inde-

pendent r.v., but the r.v. pn are not independent. If the se-
quence qn is specified directly and it consists of independent
r.v. with characteristic function
TE{w(r)} =
-t J @(e- T ) @ ( T t r O-S)dB, r>T
\k ( p ) = E {ej'+'a

then

I r < T.

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS
ON ACOUSTICS,
SPEECH, AND
SIGNAL PROCESSING, VOL. ASSP-31,
NO. 1 , FEBRUARY 1983 10s

[ 121 T. Aulin and C. E. W. Sundberg, “Continuous phase modulation,”

IEEE Trans. Comniun., vol. COM-29, Mar. 1981.
[ 131 W. A. Gardner and L. E. Franks, “Characterization of cyclosta-
tionary random signal processes,” IEEE Trans. Inform. Theory,
vol. IT-21, Jan. 1975:
[ 141 W. A. Gardner, “Stationarizable random processes,” IEEE Trans.
Inform. Theory, IT-24, Jan. 1978.
[ 151 A. Papoulis, SignalAnalysis. New York: McGraw-Hill, 1977.
[ 161 H. Broman, “The instantaneous frequency of a Gaussian signal:
REFERENCES Theone-dimensionaldensityfunction,” IEEE Trans. Acoust.,
A. Papoulis, Probability Random Variables and Stochastic Pro- Speech, Signal Processing, vol. ASSP-29, Feb. 1981.
cesses. New York: McGraw-Hill, 1965. [ 171 A. Papoulis, The Fourier Integral anditsApplications. New
L. E. Franks, Signal Theory. EnglewoodCliffs, NJ:Prentice- York: McGraw-Hill, 1962.
Hall, 1969.
J. Proakis, Introduction to Digital Communication. New York:
McGraw-Hill, 1983.
W. R. Bennett, Introduction to Signal Transmission. McGraw-
Hill: New York, 1970.
S . 0. Rice, .“Mathematical analysisof random noise,” in Selected
Papers on Noise and StochasticProcesses, N. Wax,Ed.New
York: Dover, 1954.
L. Mandel,“Com,plex representation of optical fieldsin coher-
ence theory,” J. Opt. Soc. Amer.,vol. 57, pp. 613417, 1967.
P. M. Woodward, “The spectrum of random frequency modula-
tion,”Telecommun. Res. Establish., Great Malvern, Worcs.,
England, Memo. 666, Dec. 1952.
N. M. Blachman, G. A. McAlpine, “The spectrum of a high-index
FM waveform:Woodward’s theoremrevisited,” IEEE Trans.
Commun. Technol.,vol. COM-17, Apr. 1969.
A. Papoulis,, “Spectra of stochastic FM.signals,” in Proc. Trans.
9th Prague Conf. Inform Theory, June 1982.
R. R. Anderson and J. Salz, “Spectra of digital FM,” Bell Syst.
Tech. J., July-Aug. 1965.
H. E.Roweand V. K. Prabhu, “Power spectrum of a digital,
frequency-modulation signal,” Bell Syst.Tech. J., July-Aug.
1975.

Thinning Digital Filters: A Piecewise-Exponential

Approximation Approach

Absfruct-An algorithm is presented for designing fiiite impulse re- I. INTRODUCTION

sponse (FIR) recursive digital filters that require few multiplies to pro-
duce good frequency response. The process of reducing the number of
multiplies needed to implement a digital filter is called thinning. This
thinning algorithm uses dynamic programming techniques to find the
T HE transfer function of a digital fiiter

best least-squares piecewiseexponential approximationto a desired im-

pulse response of length P. Because these filters are implemented re-
cursively, the number of arithmetic. operations is independent of the
model filter length P and is dependent only on the number of pieces or can be expressed as a rational function with a numerator poly-
segments S used in the approximation (S<< P). Examples of thinned
narrow-band, broad-band, lowpass, and bandpass filters are given. Sev- nomial N ( z ) and a denominator polynomial D ( z ) . Frequently,
eral of these thinned filters require fewer than one-thirdthe number of a digital filter
multiplications required for the correspondbig model filter, while still
P-1
retainingdesirablefrequencyresponsecharacteristics.Theeffects
G(z) = gnz-,
of coefficientquantizationand finite precisionarithmeticarealso n=O
discussed.
with a finite impulse response (FIR) g,, is desired that is com-
ManuscriptreceivedMarch 11,1981; revised June4, 1982.This putationally fast, yet still has a freqeuncy response
work was supported in part by NSF Grant ENG-7809033, and in part
by a Zonta InternationalAmelia Earhart Fellowship.
The authors are with the Department of Electrical Engineering, Rice
University, Houston, TX 7725 1. n=O

0096-3518/83/0200-Ol05$01.00 0 1983 IEEE

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 25,2023 at 12:13:01 UTC from IEEE Xplore. Restrictions apply.