5 84 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-35. NO. 6 .

JUNE 1987

Spectral Correlation of Modulated Signals:

Part I-Analog Modulation

Abstract-The importance of the concept of cyclostationarity in design are of significant

practical
value: 1) differenttypes
of
andanalysis
of signal
detectors,
synchronizers,
and
extractors in modulated signals (such as BPSK, QPSK, and SQPSK) that
communicationsystems is brieflydiscussed,andthecentralroleof have identical powerspectral density. functionscanhave
spectral correlation, in the characterization of random processes that are highlydistinctspectralcorrelationfunctions;2)stationary
cyclostationaryinthewidesense,isexplained. A spectralcorrelation noiseandinterferenceexhibitnospectralcorrelation (the
function that is a generalization of the power spectral density functionis spectral correlation function is identically zero); 3) the spectral
described, and a corresponding generalization of the Wiener-Khinchine correlation function contains phase and frequency information
relation and several other fundamental spectral correlation relations also related to timing parameters in modulated signals; and 4) the
are described. Explicit formulas for the spectral correlation function for existence of spectral correlation in a signal means that some
various types of analog-modulated signals are derived. This includespulse spectralcomponentscan be estimatedusingotherspectral
and carrier amplitude modulation, quadrature amplitude carrier modula- componentsofthesignal.Furthermore, all optimumsignal
tion, andphaseandfrequencycarriermodulation. To illustrate.the processors (such as estimators and detectors) that are specified
differingspectralcorrelationcharacteristics of differentmodulation in terms of the cyclostationary autocorrelation function canbe
types, the magnitudesof the spectral correlation functions are graphed or interpreted and implemented in terms of the spectral correla-
describedingraphicalterms as theheights of surfaces above abifre- tion function. For example, the low-SNR likelihood ratio for
quency plane. detection and the low-SNR likelihood function for parameter
estimation for random signals in white Gaussian noise are, in
I . INTRODUCTION general,completelyspecifiedbythesignalautocorrelation
function (cf. [lo], [ 191, [34]). Similarly, the optimum periodi-
A cyclostationary
probabilistic
process
parameters, suchas
the
is random
a process
autocorrelation
function,thatvaryperiodicallywithtime.Cyclostationary
with cally time-variant filter for a cyclostationary signal in station-
arynoiseiscompletelyspecifiedbythesignalandnoise
autocorrelation functions [13], [19], [35]. Also, the optimum
processes are appropriate probabilistic models for signals that quadratictransformationforgenerationofmaximum SNR
have undergone periodic transformations, such as sampling, spectral lines from a cyclostationary process, for purposes of
scanning,modulating,multiplexing,andcodingoperations, detection and synchronization is completely specified by the
provided that the signal
is
appropriately modeledasa spectralcorrelationfunctionforthesignalandthepower
stationary process before undergoing the periodic transforma- spectraldensityforthe
noise [lo], [19]. Consequently,
tion. In response to the increasing demands on communication properties 1)-4) can be exploited for detection, classification,
system performance, systems analysts are beginning to dis- parameter estimation, and extraction of signals buried in noise
cover the importance of recognizingthecyclostationary and h i t h e r masked by interference.Althoughthespectral
character of most communications signals, andto abandon the correlation function is a second-order (quadratic) statistic or
simpler more approximate stationary models that have been probabilistic parameter like the power spectral density func-
used in the past. The growing awareness of the relevance of tion, these four properties enable it to be used to accomplish
the concept of cyclostationarity is illustrated by recent work in tasks that are impossible to accomplish with the power spectral
synchronization [ 11-[ lo], crosstalk interference and modula- density function [1!2]. This includes synchronization [lo] and
tion transfer noise [ 111, [ 121, transmitter and receiver filter noiseandinterferencerejectionforsignalextractionand
design [ 131-[ 161, adaptive filtering and system identification detection [19]. Forexample, modulated signals that are
[17]-[19], coding [20], [21], queueing [22], [23], and detec- severely masked by other interfering signals as well as noise
tion [19], [24], [25]. In addition,the
growing role of can, in someapplications,bemoreeffectivelydetected by
cyclostationarity in other signal processing areas is illustrated detectionofspectralcorrelationratherthandetectionof
by recent work in biomedical engineering [26], climatology energy. This is so, for instance, when the energy level of the
[27], and hydrology [28], and by recent developments in basic backgroundnoiseorinterferencefluctuatesunpredictably
theory for prediction [29], extraction [ 131, [ 191,. detection [25]. Also, there are situations in which cyclostationarity is
[19], [25], and signal modeling and representation [30]-[33]. problematic, rather than being a property to be exploited. For
It is shown in thispaperthatspectralcorrelationisa example, nonlinearities in transmission systems (e.g., travel-
characteristic property of wide-sense cyclostationarity (cyclo- ing-wave tube amplifiers and noise limiters) can inadvertently
stationarity of the autocorrelation), and a spectral correlation generate spectral lines from cyclostationary signals, and these
function that is a generalization of the power spectral density spectral lines can cause severe interference effects (cf. [121).
functionisdescribed.Thespectralcorrelationconcepthas As explained in [lo], the characteristics of spectral lines (their
associated withit four fundamental properties of processes that frequencies and strengths) that can be generated with quadratic
nonlinearities are
determinedbythe
spectral
correlation
Paper approved by theEditor for Analog Communication of the IEEE function.
Communications Society.Manuscript received October 2 1, 1985; revised Since constraints on the length of this paper do not allow for
October 30, 1986. This work was supported in part by the Naval Postgraduate explicitdiscussionofthevariousapplicationsofspectral
School and in part by ESL, Inc.
correlation, only one particular application is described here
The author is with the Signal and Image Processing Laboratory, Depart-
ment of Electrical andComputerEngineering, University of California, andthedescriptionisbrief.Thisapplicationconcernsthe
Davis, CA 95616. problem of extracting signal,
a of interest from a background of
IEEE Log Number 8714584. both
broad-band noiseand
band-limited
interference. If

0090-6778/87/0600-0584$01.OO 0 1987 IEEE

GARDNER: SPECTRAL CORRELATION OF MODULATED SIGNALS-I
interference in some portions of the signal band is so strong
that itswamps thesignal in thosepartialbands,thenthe
optimumtime-invariant filter(Wiener filter)simplyrejects
those bands. However, if the spectral components of the signal
in those bands are correlated with spectral components of the
signal in otherbands,then the optimum periodicallytime-
variant filter will use those correlated components to estimate
and replace the rejected components 1191. If the correlation is
high and the
noise
and interference in the bands with
correlated components are low, the replaced components can
be of high fidelity. If an adaptive periodicallytime-variant
filter isused, the appropriate weighting of the frequency-
shifted components is learned through the process of adapta-
tion [ 191. In order togain an understanding of the potential for
this approachto signal extraction, thespectralcorrelation
characteristics of signals of interest must be determined.
Theprimary objective of this paperis,therefore,to
introduce a general approach and several general formulas for
calculatingthespectral correlationfunction, and to obtain
explicit formulasfor thespectral correlation functions for
many of the most commonly used types of modulation. As a
visual aid, themagnituaes of thesecalculatedfunctions
graphed (or described in graphical terms) as theheights of
surfaces above bifrequency
a plane. In this Part I, the
modulation types considered are all analog and include pulse
and carrier amplitude modulation, quadrature amplitude car-
rier modulation, and phase and frequency carrier modulation.
In Part I1 [43], varioustypes of digitalpulseand carrier
modulation are considered.
Althoughtheprobabilistic
values,which
approach basedonexpected
is conventional forpower spectraldensity
calculations, could be adopted here, as in 1351, the nonproba-
bilistic approach based on limiting time-average values is used
insteadbecause of its closer conceptualties with empirical
methods for measurement ofspectral correlation 1411. The
link between these two approaches, which is briefly discussed
herein, is explained in more detail in [42].
11. SPECTRALCORRELATION
Thissection is averybriefreview of thefundamental
concepts, definitions, and a few basic properties of spectral
correlation. An in-depth introductory treatment is given in the
tutorial paper 1421 which is prerequisite for a full understand-
ing of the notion of spectral correlation.
In order to measurethe local spectral contentof a waveform
x ( t ) over the time interval [ t - T/2, t + T / 2 ] , we use the
finite-time Fourier transform
nz
X7(t, u) 4 S l+
I- T/2
x(u)e-i2*vu du,
evaluated at frequency Y. The correlation between two such
local spectral components at frequencies Y = f + a!/2 and v
= f - a!/2 (where f is the midpoint and a! is the separation of
the two frequencies)normalized by e,
measured Over an
interval of length A t , is given by
1 cyclostationarity. In modulated signals, the periods of cyclo-
. -X ; ( t , f - d 2 ) dt (2)
JT
(where the superscript asterisk denotes complexconjugation.)
An idealized measure of spectral correlation for a persistent
waveform is then given by the double limit
As At -+ 03, the measure of correlation becomesideal
585
random effects areaveragedaway), and as T 03, the
-+
bandwidth of the spectralcomponentsbeing correlated be-
comesinfinitesimal. Theorderofthetwo limits in (3) is
crucial sincefor anytruly random waveform (one whose
spectral density function contains continuous components,not
just Dirac deltas), the limit as T + 00 does not exist for finite
A t 1191. It isshownin 1421 thattheidealizedspectral
correlationfunction (3) can be characterized asthe Fourier
transform
of the limit function’
To show this, one first shows that the product of normalized
finite-time Fourier transformsin (2) is itself characterized by a
Fourier transform
are
e-i2nol(t+u) du
1 e - i 2 ~ f r dr
and then the limit as At .+ 03 in (3) andtheintegral
(6)
with
respect to t in (2) are interchanged with theintegrals with
respect to r and u in (6) to obtain
which can be shown to yield (4).
If a! is set equal to zero, then (2) becomesthetemporal
mean-squarevalue (averagepower) of the local spectral
component, and therefore (3) beccmes the average‘power
spectral density function S:( f ) E Sx(f )._Also, in _thiscase,
( 5 ) becomes the autocorrelation function R:(r) E R x ( r ) ,and
(4) thereforebecomes the Wiener relation 1361 (which is
commonly calledthe Wiener-Kh_inchine relation in the
probabilistic setting). For a # 0, R;(r) in ( 5 ) is the strength
of any finite additive sine wave component with frequency a!
that might be contained in the lag product waveform y r ( t ) 4
(1) x(t + 7/2)x(t - r/2). Relation (4) reveals that there is a one-
to-one correspondence between such periodicity in lag prod-
ucts, which is called cyclostationarity, and spectral correla-
tion. Thiscorrespondence,(4),is calledthe cyclic Wiener
relation [42]. The function Z?;(T) is called the cyclic autocor-
relation, and the spectral correlation func_tion S ; ( f ) is also
called the cyclic spectral density 1421. If R ; ( r ) 3 0 for all a!
# _O and Rx(r)& 0,then x ( t ) is said to be purely stationary.
IfR;(7) & 0 only for a! = integer/To for someperiod To,then
xlt) is said to be purely cyclostationary with period To. If
R:(r) & 0 for values of a! that are not all integer multiples of
some fundamental frequency l / T o ,then x ( t ) is said to exhibit
stationarity correspondtocarrierfrequencies, pulse rates,
spreadingcode repetition rates, time-divisionmultiplexing
rates, and so on.
If x ( t ) itself contains no finite additive sine wave compo-
nent, then the temporal mean of X T ( t ,f ) is zero for f # 0.
Consequently, S : ( f ) is actually a spectral covariance function
The relationship between the limit function ( 5 ) and the radar ambiguity
(all function is discussed in [42].
 586 IEEE TRANSACTIONS ON COMMUNICATIONS. VOL.
NO. COM-35. 6, JUNE 1987

$,( f ) is actually a
in this case, and the power spectral density ticprocessiscycloergodic [19], [311,1351, then R ^ ; ( T ) =
variance. Therefore, the normalized function R ; ( T ) and these two dual theories merge into one.
A particularly convenient method for calculating the spec-
tral correlation function for many types of modulated signals is
to model thesignal as apurelystationarywaveformtrans-
formed by a linear periodically time-variant transformation.
is a complex correlation coefficient which satisfies Specifically,consideravector-input,scalar-output, linear
periodically time-variant (LPTV) system with period TO, and
withimpulse-responsefunction h , expanded in a Fourier
This function e;(f ) is called the spectral autocoherence series:
function [42]. The waveform x ( t ) is said to be completely
coherent at f and CY if I C ; ( f )I = 1 , and it is 'said to be
completelyincoherent at f and CY if C:( f ) = 0. The
y(t)= Sm
-m
h ( t , u ) x ( u ) du (17)

parameter CY is called thecycle frequency, and the parameterf m

is called the spectrum frequency [42]. h ( t , u)= g,(t-u)ei2*nu'To
In the probabilistic approach, a stochastic process x ( t ) is ( 1 8)
n= -m
defined to be cyclostationary (in the wide sense) with period
Toif the probabilistic autocorrelation where h(t, u)is a row vector andx(#) is a column vector. The
system function, which is defined by the row vector
R,(t, 7) i?
E{x(t+7/2)x(t-r/2)} (10)

(in which E { - } denotesexpectedvalue)isperiodicwith ~ ( tf), 4 S m

-m
h ( t , t--T)e-i2TfTd7 9 (19)
period To.R,(t +
To, 7 ) = R,(t, 7 ) [35]. The coefficients of
the sine wave components of this autocorrelation is then given by the Fourier series

(where CY = integer/T,)arecalledthe cyclicautocorrela-
tions. For a stochastic processthat exhibits cyclostationarity at
'more than one fundamental frequency, R,(t, 7 ) consists of a
denumerable sum of finite additive periodic components and
has the Fourier series representation
for which G,, is the Fourier transform of g,,. It can be shown by
straightforward (but tedious) calculation that the cyclic auto-
correlations and cyclic spectra of the input x ( t ) and outputy ( t )
of the LPTV system are related by the formulas2 [42]

where

R;(T) 4 lim -
T-tm T
ST/2- T/2
R,(t,r)e-i2aaf dt, (13)
n,m= -w
and the sum in (12) ranges over all values ofa for which (13)
isnotidentically zero. If only oneperiodiccomponentis G;(f-a/2)*
(22)
present in R,(t, 7 ) (i.e., if R,(t, 7 ) isperiodic),then(13)
reduces to (1 1). The nonprobabilistic counterpart of R,(t, 7 ) in in which R; is the matrix of cyclic cross correlations for the
either (10) or (12)' is given by vector-valued waveform x ( t ) , S; is the corresponding matrix
of Fourier transforms, and r;m is the matrix of finite cyclic
cross correlations of theFourier-coefficientfunctions gn,
Q which is defined by
m
where_R^;(T) i s defined by (5). For a purely cyclostationary r;,(7) 2 g I : ( t + 7 / 2 ) g , ( t - r / 2 ) e - j 2 " " ' dt. (23)
x(t), R,(t,7 ) is also given by [42] -m

1 Formula (22) is used in the following sections for calculating

R,(t, 7 )= lim - the spectral correlation for various types of modulated signals.
N + m 2N-k 1
Thisformulacan be generalizedforlineartime-variant
systems with more than one periodicity,in which case the sum
x ( i + 7 / 2 + n T o ) x ( t - r / 2 + n T o ) , (15) in (18) is expanded to include all harmonics of each of the
n= - N
fundamental frequencies correspondingto each periodicity,by
simply expanding the double sum in (22) similarly [ 191.
whereasforan x ( t ) thatexhibitscyclostationaritywiththe A special case of interest is a time-invariant filter, for which
periods T = T I , T = T2, T = T,, . -,T = TMforsome all n # 0 terms in (18) and (20) are zero. In this case, (22)
integer M (or M = m), we have [19], 1421 reduces to
M

i= I and for CY = 0, this is recognized as the conventional formula

where R,(t, 7 ; T , ) is given by (15) with To = Ti. Thus, there
is a duality between the more common probabilistic theory of
stochastic processes that exhibit cyclostafionarity [35] and the
nonprobabilistic theory 1191, [42] adopted here. If the stochas-
for the power spectral density at the output of a multiinput
filter.
2 In (21) and (22), the prime denotes matrix transpositionand tr{ .} denotes
the matrix trace operation.
 588 IEEE TRANSACTIONS
ON COMMUNICATIONS, VOL.
NO. COM-35. 6 , JUNE 1987

wave, with period TOanddutycycle of, say, 100 percent:

AS an example of (36), if x ( t ) is a stacked carrier AM signal,

then P p is aconstant for afiniteset of values of 6, such as 0 =
fo+nAfforn=0,1,2,3;--,N- 1,andPpiszerofor
all other values of p. Let us consider some other examples in
A
-4f0/3
- f
4f0/3
I."

more detail. for surface

Fig.
magnitude
correlation
1. Spectral an amplitude-modulated
sine wave with carrier frequency fo.
Example (Amplitude Modulation): If p ( t ) is given by

P ( t ) = cos ( 2 r f o t + $ 0 ) (38) Example

(Pulse-Amplitude
Modulation): If p ( t ) is given
thenwe have conventional AM and (34) and (35) reduceto by
m

then Pa = 1/ To for p = m/ TOfor allintegers m , and

+-1 l?a-2fo(T)e+i2+o (39) therefore (35) yields
4 ,
l m
and s^;(f) = $-m'To(f- n/To+ m / 2 To). (45)
Tn n,m=- m
If the product time series (32) and (44) is filtered using an
impulse-response function q ( t ) ,then y ( t ) = x ( t ) 8 q ( t ) is the
pulse-amplitude modulation (PAM) signal
m
This result can be used to obtain the cyclic spectra for other
types of modulationthatinvolve an amplitude-modulated u(O= a(nTo)q(t-nTo). (46)
carrier. Examples for which a ( t ) is cyclostationary are binary n= -m

phase-shiftkeying and amplitude-shiftkeying,which are Application of the input-output spectral correlation relation
treated in [43]. For the special case inwhich the amplitude a(t) for filters (24) (with Go there replaced by Q) to (45) yields
is purely stationary, x ( t ) is purely cyclostationary with period
To = 1/2f o , and (40) reduces to
r

2
n,m = - 0
s ^ ~ - m ' T ~ ( f - - / T o + m / 2 T (47)
o)

for this PAM signal. The spectral correlation aliasing formula

otherwise. (27) [with x ( t ) there replaced by a(t)]applied to (47) yields the
alternative formula for PAM:
A-typical graph of the spectral correlation magnitude surface
I S,"(f )I is shown in Fig. 1. It follows from (41) that the 1 Q ( f + d 2 ) Q * ( f - ~ ~ / 2 ) 3 ; ( f ) .(48)
s^;(f) = -
magnitude of the spectralautocoherence of x ( t ) (8)is given by TO

for a = +2f0. Consequently, if a ( t ) is band limited such that If a ( t ) is purely stationary, then (48) reduces to
S,(f)=OforIfI > B a n d S , ( f ) # O f o r ( f l < B , w i t h B <
f o , then Sya(f 1

and x ( t ) iscompletely coherentfor a = &2f0 andall

frequencies f for which-it is not completely incoherent because
there is no power density. But if B > f o , then I C;( f )I < 1
(49)
for 1f I > 2f o - B: For example, if a ( t ) is white ( S , ( f ) = A . (for all integers k ) by using
for '-'w < f < w), then I f ) l = 1/2 for - 03 < f < 00
and la1 = 2f0. S;+k'Tq f ) = S;(f+k/2 To).
GARDNER: SPECTRAL CORRELATION OF MODULATED SIGNALS-I 589

It follows from (49) that the spectral autocoherence magnitude

(8) is given by
I C y ( f ) l =1, (Y=k/To (50)
for all f for which $,,( f f a / 2 ) # 0. Therefore, the PAM
signal is completely coherent at CY = k/To for all integers k
and all f for which the signal is not completely incoherent
f
because there is no power densjty. A graph of the spectral
correlation magnitude surface 1 S,"(f )I for a white amplitude Fig. 2. Spectral correlation magnitude surface for
a pulse-amplitude
sequence {a(nTo)}and a rectangle pulse q ( t ) of width To is modulated pulse train with pulse rate I/To.
shown in Fig. 2 .
Example(Jittered PAM): Themore realistic model of
for all integers k and is zero for all other values of C Y .The
PAM that incorporates pulse-timing jitter
function is the characteristic function for the sequence { E , } ,
m
defined by
~(t)= anq(f-nTo-En) (5 1 )
n= -m 1 N

can be reexpressed as

Y ( o = [ a ( t )w ( 0 l @ q ( t ) (52) In general, the larger the variance of { E , } is, the narrower q,

where w ( t ) is the jittered impulse train will be and the more the spectral correlation in y ( t ) will be
attenuated. As an example, if E , has Gaussian fraction-of-time
m
density with zero mean and variance uf , then [ 191
w(t)= 6(t-nTo-E,) (53)
n= - m

and a, = a(nTo - e n ) . This model is appropriate for either

jitter that occurs in the process of forming the PAM signal or
jitter that occursafterthe PAM signalhas been formed, Example (Phase-Deviated AM): The more realistic model
provided in this latter case that a ( t ) is sufficiently narrow band for AM that incorporates carrier phase deviation
(low pass) or broad band andthatthestatistics of an are X(f)=a(t)cos[2nfot++(t)] (59)
independent of the jitter. The general formulas (24) and (29)
applied to (52) (assuming { E , } and a(t) are statistically is still an example of product modulation,andtherefore
independent) yield formula (29) applies (assuming that + ( t ) and a(t) are statisti-
cally independent) to yield
S^y(f)= Q ( f + d 2 ) Q * ( f - d 2 )

where
and for a purely stationary amplitude a ( t ) , this reduces to

As explained in Section V , the spectral correlation functionfor

If the jitter sequence { E , } is purely stationary, then w ( t ) is the phase-modulated sine wave (61) is given by the Fourier
purely cyclostationary with period To, and y ( t ) is therefore transform of the cyclic autocorrelation function
also purelycyclostationary with period To. However, the
strength of spectral correlation at CY = k / Towill be attenuated
by the convolution in ( 5 5 ) . On the other hand, if { E , } is an
independent-increment sequence [35] (in the fraction-of-time
sense [19]), then w(t) will be purely stationary and so too will
y ( t ) . Nevertheless, y ( t ) can still exhibit reliably measurable
spectral correlation locally in time. Theformer (purely
stationary { E , } ) model is appropriate for jitter relative to a
tracking clock synchronized (imperfectly) toy ( t ) , whereas the ff= -2f0
latter (independent-increment { E , } ) model is appropriate for
absolute jitter and drift.Itcanbe shown [43] that for an
independentsequence of jitter variables { E , } , thespectral a # +2f0, 0
correlation function is given by
1 where Re{ * } denotes the real part and the function qris the
s^.,(f)=--; E- \k,(2nn/To)42(2a[n+k]/To) joint characteristic function for +(t + 7/2)and d ( t - 712):
' 0 n= -m

1
6(f+n/To+k/2To)+- {\k327rk/TO)
TO + + ( t - 7 / 2 ) ~ 2 ]dt. (63)
- \k 3 2 n [f+k/2 To])9,(27r [f-k/2 T o ] ) ) ,
In general, the larger the varianceof + ( t ) is, the smaller q,(l,
1) will be in magnitude, and the more the spectral correlation
in x ( t ) will be attenuated. For example, if + ( t ) is a purely
 590 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-35. NO. 6 , JUNE 1987

stationary Gaussian waveform with zero mean, then [19] and

*J, 1)=exp{ -[R^,(o)+R^,(7)1}. (64)
IV. QUADRATURE-CARRIER AMPLITUDE MODULATION
Consider
the
quadrature-carrier
amplitude modulation with g ; = 0 for a # 0 and a # f 2fo.(In these formulas, the
(QAM) waveform subscripts r and i denotetherealandimaginaryparts,
respectively.) Thus, the cyclic spectra1 density of QAM is of
x(t)=c(t)cos(27rfot)-s(t)sin(27rfot) the same general form as that shown in Fig. 1 for AM, except
(654 =a(t)cos[27rfot++(t)] that thefull symmetry shown there is typically not exhibited by
QAM. Only symmetry about the f axis and symmetry about
for which the a axis is always exhibited by QAM. Moreover, it follows
from the autocoherence inequality (8)-(9) that the heights of
a(t)=[c(t)2+s(t)2I1'2
the surfaces centered at (f,a ) = (0, .t 2f0) are lower than or
+(t)=tan-' [s(t)/c(t)] equal to the heights at (f,a ) = (+-fa, 0). If c ( t ) and s ( t ) are
band limited such that &(f) = S , ( f ) = 0 for If1 2 fo, then
= a(t)cos[+(t)l (68)-(69) yield

s(t)=a(t)sin[+(t)]. (65)
This is a particular LPTV transformation of the two-dimen-
sional vector of waveforms[c(t),s(t)] ' for which the vector of
impulse-response functions is (704
for the spectral autocoherence magnitude ofx ( f ) .Moreover, it
h ( t , ~ ) = [ ~ 0 ~ ( 2 7 r f ~ t ) 6 ( t --sin(27rfot)6(t-u)]
~), is shown in [35] that any process x ( t ) that is
purely
and the vector of corresponding system functions is cyclostationary with period 1/2f0 and is band limited to f E
( - 2foj 2f0) can be represented in the form of (65) (Rice's
G ( t ,f)= [cos(27rfot), - sin(27rfot)]. representation,see [32]) for which c(t) and s ( t ) are purely
stationary and band limited to f E ( -fo, fo). This includes
Application of formula (21) yields many band-limited analog-modulated sine wave carriers, with
1 purelystationarymodulatingsignals,used in conventional
7)+ RY( 7)]cos(2 7 r f o 7 )
I?,"(7)= - [I?,*( communications systems and essentially all modulated peri-
2 odic pulse trains with excess bandwidth (beyond the, Nyquist
1 bandwidth).of 100 percent or lessandpurelystationary
-- [R^,nc(7)-R^~,(7)lsin(27rfo.r) modulatingsignals. Thus,for all suchsignals,thespectral
2 autocoherencemagnitudeisgivenby(70a).Furthermore,
(70a) can be reexpressed as
+-1 ([R^,*+2"fo(7)-R^~+2"f0(7)]
4 nE{-l,l}

-ni[R^,qc+2"f~(.r)+R^~s+2~f0(7)]} (66) (70b)

andapplicationofformula(22)yieldsthecyclicspectral where &( f ) is the cross coherence between c(t) and s ( t )

[19], [35], and the denominator can be reexpressed as
density
1 [ s c , < f )+ g s ( f ) 1 ~ - 4 [ & ( ~ ) i 1 ~ = 16[&(~-~0)2
s;(f)=, {[Q~f+nfo)+$;(f+nfo)l
-~x(f-fO>:l, If1 < f a (71)
fl€{-l,l}

- n@,nc(f+ nfo) - qs(f+

nfo>l> where the subscripts e and o denote the even and odd parts,
respectively, of SI(f ) about the point f = fo for f > 0.
1 Consequently,forgivenspectrum S, andspectralproduct
+- C {[~~+2"f0df)-~sQ+2"f0(f)] Sr S,, the autocoherence magnitude of x ( t ) increases as the
4 nE{-1,1} magnitudeof thecrosscoherencebetween c(t) aFd s(t)
increases. Furtherpore, for given cross coherence 1 C,,l and
- ni[s^;c+2'2"fo(f) +p + w o spectral product S, S,, the autocoherence magnitude of x(t)
cs (f)I> * (67)
increases as the dominance of the even part (aboyt fo) of the
Examplesforwhich c(t) and s(t) arecyclostationaryare
spectrum S, over the odd part increases (assuming Sr( f f) )ss(
f 0).
quaternary-phase-shift keying and some amplitude-phase-shift It followsfrom(70a)that x ( t ) is completelyincoherent
keying, which are treated in [43].
For the special case in which the in-phase and quadrature
(e;( f) = 0) at a = k 2f0 and at anyf if and only if c(t) and
s(t) are balanced at f in the sense that
components c(t) and s(t) are jointly purely stationary, x ( t ) is
purely cyclostationary with period To = 1/2fa, and we have 1) SCc<f>= W )
and
1
S A f ) = 4 [Sc(f+fo) + w -fo) + W + f o ) + W - f o ) l 2) g c s < f) r = 0.
It $so follows .from (70b) that x ( t ) is completely coherent
1
-- [3cs(f+fo)i- 3cs(f-.f0)i1(68) ( I C ; ( f ) l = 1 ) a t . a = -t2foandatany If1 < f o i f a n d o n l y i f
2 either
GARDNER: SPECTRAL CORRELATION OF MODULATED SIGNALS--I 59 1

3) the
supports
of Sc and ss
are disjoint
such
that for which 9, is the jointcharacteristicfunctionforthe
9A.f > 9 S ( f 10 or= variables +(t + 7/2) and $(t - 7/2) defined by (63). But (75)
4) the waveforms c(t)and s(t) are completely cross coherent is simplytheconditionunder which thephase-modulated
at f such that waveform

I C S ( f ) I =1 for If1 <fo X(t)=COS[2Tfot+~(t)]

and either 1) or 2) (or both) is violated. is completely incoherent. This i s discussed further in the next
It follows from (70a) and(71) that for a time series x ( t ) that section.
is completelycoherentat If1 < fa, thecyclicspectrum v. PHASE AND CARRIER MODULATION
FREQUENCY
magnitude is characterized by the symmetry of the conven-
tional spectrum: Consider the phase-modulated (PM) sine wave
X(t)=COS[2Tfot+4(t)] (76)

for la1 =2fo. (72) for which'+(t)contains noperiodicity (of any order). It can be
shown'by direct calculation that the cyclic autocorrelation for
When 4) holds, c(t) and s(t) are related (at least in the x()t%$r given by
temporal mean square sense) by an LTI transformations(t) =
h ( t ) 8 c(t) (cf. [19], [35]). But there do exist LTI transforma- Re{Q7(1, - l)ei2?rf07}, a=O
tions for which neither 1) nor 2) is violated, namely, those for
which the transfer functions are unity in magnitude and purely a=2fo
imaginary with arbitrary signs at arbitrary frequencies f , that
is, H ( f ) = +-i. R,*( 7)=
a = -2fo
Example (SSB, DSB, VSB): TheHilberttransform(for
which N(f ) = + i for f < 0 and H ( f ) = - i forf > 0 ) ,
which yields a single-sideband (SSB) signal x ( t ) ,

W ) =o,
If1 < h
results in a waveform thatis completely incoherent for all f . In and the cyclic mean of x ( t )
contrast,atransferfunctionthat is arealconstantyieldsa
double-sideband (DSB) signal x ( t ) that is completely coher-
ent for all If1 < fo, as established in Section 111. Similarly a
vestigial-sideband (VSB) signal, which is obtained by sub-
jecting aDSBsignaltoalow-passfilteringoperationwith is given by
bandwidth, say, B = fo +
b, is completely coherent for1f I <
b and completely incoherent (for an ideal low-pass filter) for
If1 > 6. For the DSB signal, there is no phase modulation,
+ ( t ) = constant. However, if the DSB signal is filtered and
the filter transfer functionis asymmetric about the point f =f o
(for f > 0) andnonzero,then +(t) in (65) isnolonger
constant, but x ( t ) is still completely coherent for all If1 < fo where 9,is the joint characteristic function for
+(t + 7/2) and
(since the autocoherence magnitude is invariant to linear time- +(f - 7/2)
invariant transformations [19], [35], provided that the transfer
function does not equal zero).
In ordertodeterminewhattype o f phasemodulation
annihilates coherence, one can use the fact that the necessary
and sufficient conditions 1) and 2) for complete incoherence
for all f are equivalent to the pair of conditions
and !P is the characteristic function for +(t)

O(w) 2 lim
T-m
-
T
ST'*
-T/2
exp(ib(t)w} dt. (78b)
cos[q5(t+7/2)+q5(f-~/2)] dt =0 (73a)
Thus, x ( t ) is purely cyclostationary with period 1/2f0 except
T/2 when *,(I, 1) = 0 for all 7 , and x ( t ) exhibits spectral lines at
lim - a(t+7/2)a(t-7/2) ff o except when 9(1) = 0.Neither qr(1, 1) = 0 nor 9 ( l ) =
T-m T - T/2
0 can hold for a purely stationary Gaussian waveform+(t)[cf.
sin[c$(t+~/2)+c$(t-~/2)]dt = 0. (73b) (58) and (a)]. However, there are some waveforms d ( t ) that
do satisfy ! P T ( l , 1) = 0 and 9 ( l ) = 0. Anexampleisthe
It follows from (73) that an equivalent condition is balanced quaternary-valued waveform
c$(t)=tan-' [ s ( t ) / c ( t ) ]
for which c(t)and s(t) are statistically identical, uncorrelated,
binary-valued (+.1)waveformswithstationarytransition
e x p { i [ 4 ( t + 7 / 2 ) + 4 ( t - ~ / 2 ) ] }dt 0. (74) times (e.g., a Poisson point processes [35]). Another example
is the balanced quaternary-valued PAM waveform + ( t ) , which
For example, if a ( t ) is statistically independent of $ ( t ) (e.g., yields the QPSK signal discussed in [43]. However, since the
a(t) = constant), then (74) reduces to transition times in this + ( t ) are periodic with period, say TO,
then x ( t ) is purely cyclostationary with periodTo,even though
q7(l, 1)=0, --03<7<00 (75) it is not cyclostationary with period 1/2f o . Furthermore, both
 592 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL.
NO. COM-35. 6, JUNE 1987

q7(1, 1) = 0 and T(l) = 0 are satisfied by some physical frequencymodulatormust be lossyand therefore
nonstationary waveforms, such as those thatarise from should be modeled as
frequency modulation. Specifically, let + ( t ) be given by4
+(t)= S: e-Y(l-u) z ( u ) du (83)
4(t)= ib z ( u ) du (79a)
for someappropriately small positive value of y. Unlibe (79a),
(83) is inde_eda stable Lransformakion, and the_reforeR+(m)=
for which z(t) is a purely stationary waveform. Then [M6I2if R,(oo) = [MZ]*, and M+ = 0 if M, = 0.
z(t)=d$(t)/dt, t>O (79b) If (83) is used as themodel for FM, thenthe preceding
statements about the distinction between PM and FM need to
and z(t) is the instantaneous frequency deviation (in radians bemodified.Specifically, FM modeled with (83) will, in
per unit o,f time) of themodulated sine wave (76). If the general, be purely cyclostationary regardless of the low-pass
spectrum S,( f ) is not high pass (or bandpass) in the sense that spectral content of z(t). However, the strength of the spectral
it does not approach zero faster than linearly in f as f --t 0, correlation will be weak when the spectrum of z ( t ) does not
then + ( t ) can satisfy 97(1,1) = 0 and 9 ( l ) = 0. For approach zero faster than linearly in f , and the magnitude of
example, if z(t) iswhite (gz(f ) = k ) , then +(t) isan the spectral correlation will, in fact, approach zero asr_ -+ 0 in
independent-increment waveform (a diffusion), for which it (83). This can beseen by using (64)and showingthat R6(0) -+

is well known that x(t) is stationary and contains no spectral 00 as y -+ 0.

lines [37]. Thus, frequency-modulated (FM) sine waves with Beforeconcludingthis discussion, let us consider the
low-pass (or all-pass)
modulation are purely stationary, relationshipbetween thespectral densityandthe spectral
whereas those with high-pass (or bandpass) modulation can be correlation for PM and FM. From (77a), we have
purely cyclostationary. For example,if z ( t )is defined by (79b)
for a purely stationary + ( t ) , then x ( t ) can be purely cyclosta- 1
tionary . Thus, we see that some FM sine waves are purely RX(7)=- Re{ $ r ( l , - I)e'2rfor} (84)
2
stationary, but most PM sine waves of practical interest are
purely cyclostationary. Hence, there is a fundamental distinc- where q T isgiven by (77b).For azero-meanGaussian
tion to be made between the statistical properties of FM sine waveform, qTis given explicitly by [I91
waves and PM sine waves. This is often not recognized, since
it iscommonpracticetouse probabilisticmodelsand to q r ( l , - 1) = exp{ - [R+(o) - R+(7)]}. ( 8 5 )
introduce a random time-invariant phase variable to renderall Thus, it follows from (77a), (64),(84), and (85) that R^:'o(o(7)
modulated sine waves stochasticallystationary(cf. E331 and and R J T ) differ by only a cosine factor, a factor of 1/2, and a
[38]), and to adopt relatively arbitrary conventions (e.g., [39]- sign. However, this sign different: is very significant.5 For
[40]) for distinguishingbetweenphasemodulationand fre- example, it follows that although R,(T) reaches its maximum
quencymodulation.Itshouldbe emphasized, however, that value of 1/2 at T = 0, R^:fo(7)reaches its maximum value of
even frequency-modulated sine waves that are stationary in the
long run 'can exhibit reliably measurable properties of cyclo- 1
stationarity locally in time. max R . f o ( ~ ) = - exp { - R,(o)}
It is also worth clarifying that the transformation (79a) from r 4 (86a)
z ( t ) to +(t)is only marginally stable and can therefore be an at 7 c*, for R^,(T) 2 0 and &(m)
-+ = 0. Furthermore, if
inappropriate model. For example,, it can be shown that even R^+(T) can be negative, then
though z(t) might have a zero temporal mean, the temporal
mean of +(t)need not be zero. This iseasily demonstrated for 1
thesimplecase z(t) = cos (ut -+ e). MoJe importantly, max k ; f o ( ~ ) =
4
exp- { -[R+(O)+minR+(~)1}. (86b)
although the temporal covariance RZ(7)- ( M , ) 2of z(t) may
T 7

approach zero as 7 + 00, the temporal covariance of + ( t ) can Thus, for a nonoscillatory phase time series (&(TI) 2 0 ) , we
approach a positive constant, which can attenuate the spectral have
correlation in the model of the FM signal x ( t ) at CY = f 2fo.
For example_, R+(7) in, (64), whic_h must be replaced by the max R : ~ o ( < T )< max RX(7) (87)
r 7
covariance R+(r)- [M6I2when M+ # 0, can be expressed as
the sum of its constant asymptote for &(O) > > 1, but for an oscillatory or narrow-band time
series for which
c g R4(O3)-[A2+1* (80)
and its residual, R,(O)+minR+(T)<
t <R,(o)
R(T) & R^+(7)-&,(a). (81) (87) need not hold, even if &(O) > 1. In fact, R^:fo(o<7) will
%>

itself be highly oscillatory in this case, and this canresult in a

The exponential in (64)can then be factored to obtain highly oscillatory transform s:fo(f).
Consequently,even
q 7 ( l , l)=exp{ - c } exp{ -[R(O)+R'(T)]} (82) though the area of S^O':( f ), which is given by

for apurelystationary Gaussianwaveform r#~(t)_. Thus, the 1

asymptote c attenuates thecyclic autocorrelation R ; ( T )at CY = R:f0(0) =- exp{ - ~ R , ( o ) } \
4
-t 2fo (77a). The factthat c # 0 represents anomalous
behavior can be seen from the fact that the strengt_h of the ?ill be very small for R^+(O) > > 1 compared to the- area of
spectral line at f = 0 for such a time series is not [M6I2,but S,( f ), its peaks can be comparable to thepeaks of SX(f ).
rather it is c + [&?+I2. This nonphysical behavior canbe As an example to illustrate the difference between PM and
avoided by realizingthattheintegrationoperation in any
Because of thissign difference, thetechniqueused to develop Wood-
ward's theorem (cf. [%]) for approximating the spectrum of FM for the case
For the one-sided ( t 2 0) model, all limit averages over I t I < m must be of highmodulationindexcannotbeused to approximatethespectral
multiplied by 2 to produce the appropriate result. correlation function.
GARDNER: SPECTRAL CORRELATlON OF MODULATED SIGNALS-I 593

n n nnnnt I

(d)
Fig. 3. Spectral correlation surfaces for a unity-power phase-modulated sine
wave. (a) Magnitude surface forfa = 9000, modulating-phase passband =
[300, 33001, rms phase = 1. (Thespectral-line-pair power = 0.37.) (b)
Magnitude surface for fo = 36 000, modulating-phase passband = [300,
33001, rms phase = 2.5.(The spectral-line-pair power = 0.082.)(c)
Magnitude surface for fo = 636 000, modulating-phase passband =
[30300, 333001, rms phase = 2.5.(The spectral-line-pair power =
0.082.) (d) Phase surface forlo = 636 000, modulating-phase passband =
[30 300, 333001, rms phase = 2.5 (range: 0-n).
(d)
Fig. 4. Spectral correlation surfaces for a unity-power frequency-modulated
sine wave. (a) Magnitude surfacefor fo = 9000, modulating-frequency
passband = [300. 33001. rms phase = 1. (The spectral-line-pair power =
0.37.) (b) Magnitude surface for fo = 36 000, modulating-frequency
passband = [300. 33001, rms phase = 2.5. (The spectral-line-pair power
= 0.082.) (c) Magnitude surface forfo = 636 000, modulating-frequency
passband = [30300, 33 3001, rms phase = 2.5.(The spectral-line-pair
power. = 0.082.) (d) Phase surface for fo = 636 000, modulating-
frequency passband = [30 300. 33 3001, rms phase = 2.5 (range: 0-T).

FM and the dependence on mean-square phase deviation and signals. These differing spectral correlation characteristics can
phasebandwidth,webegin by consideringa model for the be used to empirically classify modulated signals, even when
modulating signaly ( t ) that is typical of speech, namely,y ( t )is the signals are buried in noise, and they also can be exploited
zero-mean
a purely
stationary
Gaussian waveform with for signal detection, synchronization, and extraction (cf. [lo],
spectral density 1191, 1251, [35]). A signalcan be synchronized to using a

and
we
$(f>= Io,
consider typical
a
So, 300 6
otherwise
If1
Q 3300

value of mean-square phase

quadratic synchronizer if andonly

signal of interestmakes
if it exhibitsspectral
correlation. Similarly,the existence of spectral correlation in a
it possibletoreject ,noise and
interference for purposes of signal detection and extraction in
d_eviation Rm(0)= ( 2 . 5 ) * . We then consider a smaller value ways that would be impossible for signalswithoutspectral
R+(O) = 1 and a smaller relative bandwidth 30 300 < If\ < correlation.The results in this paperare essential forthe
33 300. For PM, 4(t) = y ( t ) , and for FM, 4(t)is given by designandanalysis of signalprocessing systems, such as
(83) with z(t) = y ( t ) and y < < 300. The spectral correlation detectors,
classifiers,
synchronizers, and extractors, that
magnitude surfaces for these various cases are shown in Figs. exploit spectral correlation since one of the first steps in such
3 and 4. design or analysis istodetermine thespectralcorrelation
characteristics of the signals of interest.
VIII. CONCLUSIONS Since
the spectralcorrelation function reduces to the
A new characteristic of modulated signals, thespectral conventional powerspectral densityfunction for a cycle
correlationfunction,is calculated for avariety of analog frequency of zero, a = 0, the spectral correlation formulas
modulationtypes. Theresultsclarifythe ways in which derived here yield the well-known formulas for powerspectral
cyclostationarity is exhibited by different types of modulated density as special
a case.InPart I1 [43], the spectral
correlation function is calculated for a variety of digital
modulation types.
ACKNOWLEDGMENT
The author gratefully acknowledges the interest of Prof. H .
H. Loomis, Jr., and the resultant partial support of this work
through the Naval Postgraduate School, and the interest of Dr.
C . W . Scott, and the resultant partial support through ESL,
Inc. Thanks are also due C.-K. Chenfor producing the
computer-generated graphs of spectral correlation surfaces
presented in this paper.
REFERENCES
[I] L. E. Franks and J. Bubrouski, “Statistical properties of timing jitter in
a PAM timing recovery scheme,’‘ IEEE Trans.Commun., vol.
COM-22, pp. 913-930, July 1974.
[2] U. Mengali andE.Pezzani,“Trackingproperties of phase-locked
loops in optical communicationsystems,” IEEE Trans. Commun.,
vol. COM-26, pp. 1811-1818, Dec. 1978.
[3] L.E.Franks.“Carrierand bit synchronization in data communica-
tion-A tutorial review,” ZEEE Trans. Commun., vol. COM-18, pp.
1107-1121, 1980.
[4] M. Moeneclaey, “Linear phase-locked loop theory for cyclostationary
input disturbances,” IEEE Trans. Commun., vol. COM-30,pp.
2253-2259, Oct. 1982.
[5] -, “Comment on ‘Tracking performance of the filter and square bit
synchronizer,’ ” IEEE Trans. Commun., vol.COM-30,pp. 407-
410,Feb.1982.
[6] ?- “The optimum closed-loop transfer function of a phase-locked
loop used for synchronization purposes,” IEEE Trans. Commun.,
vol. COM-31, pp. 549-553, Apr. 1983.
[7] -, “A fundamental lower bound on the performance of practical
jointcarrierand bit synchronizers,” ZEEE Trans. Commun., vol.
COM-32,pp. 1007-1012, Sept.1984.
[8] S. Pupolin and C. Tomasi, “Spectral analysis of line regenerator time
jitter,” IEEE Trans. Commun., vol. COM-32. pp.561-566, May
1984.
[9] J. J. O’Reilly, “Timing extraction for baseband digital transmission,”
’ in Problems of Randomness in Communication Engineering, K. W.
Cattermole and J. J. O’Reilly, Ed. London:Plymouth,1984.
[IO] W.A.Gardner,“Therole. of spectralcorrelation in design and
performanceanalysis of synchronizers,” IEEE Trans. Commun.,
VOI. COM-34, pp. 1089-1095, NOV. 1986.
[ I l l J. C. Campbell, A. J. Gibbs, and B. M. Smith, “The cyclostationary
nature of crosstalk interference fromdigital signals in multipair cable-
Part I: Fundamentals,” IEEE Trans. Commun., vol. COM-31, pp.
629-637, May 1983.
[I21J.P.A.Albuquerque, 0. Shimbo, and L.N.Ngugen, “Modulation
transfer noise effects from a continuous digital carrier to FDM/FM
carriers in memoryless nonlinear devices,” IEEE Trans. Commun.,
vol. COM-32, pp. 337-353, Apr. 1984.
[13] W. A. Gardner and L. E. Franks, “Characterization of cyclostationary
random signal processes,” IEEE Trans. Znform. Theory, vol. IT-21,
pp. 4-14, 1975.
[14] M. F. Mesiya, P. J. McLane, and L. L. Campbell, “Optimal receiver
filters for BPSK transmission over a bandlimited nonlinear channel,”
IEEE Trans. Commun., vol. COM-26, pp. 12-22, Jan.1978.
[15] T. H. E. Ericson, “Modulation by means of linear periodic filtering,”
IEEE Trans. Znform. Theory, vol. IT-27, pp. 322-327, May 1981.
[16] F.K.Graef,“Joint optimization oftransmitter and receiverfor
cyclostationary random signal processes,” in Proc.NATOAdv.
Study Inst. Nonlinear Stochastic Problems, Algarve, Portugal, May
1982.Dordrecht,The Netherlands: Reidel,1983.
[I71 E.R.Ferrara,Jr.and B. Widrow,“The time-sequenced adaptive
filter,” IEEE Trans.Acoust., Speech, . Signal Processing, vol.
ASSP-29, pp. 679-683, June 1981.
[18] W.A.Gardner, “Optimization and adaptation of linear periodically
time-variant digital systems,” Signal and Image Processing Lab., Dep.
Elec. Comput. Epg., Univ. California, Davis, Tech. Rep. SIPL-85-9,
1985.
[I91 -, StatisticalSpectralAnalysis:ANon-ProbabilisticTheory.
Englewood Cliffs, NJ: Prentice-Hall, 1987, to be published.
[20] G. L. Cariolaro, G. L. Pierobon, and G. P. Tronca, “Analysis of codes
and spectra calculations,” Int. J . Electron., vol. 55, pp. 35-79, 1983.
[21] C.A.French and W.A.Gardner,“Spread-spectrumdespreading
without the code,” IEEE Trans. Commun., vol. COM-34, pp. 404-
407, Apr. 1986.
[22] M.H.Ackroyd,“Stationary and cyclostationary finite buffer beha-
viour computation via Levinson’s method,” AT&T Bell Lab. Tech.
J . , vol. 63, pp. 2159-2170, Dec. 1984.
M.Kaplan,“Single-server queue with cyclostationary arrivalsand
arithmetic service,” Oper. Res., vol. 31, pp. 184-205, 1983.
W. A. Gardner, “Signal interception: A unifying theoretical frame-
work for featuredetection,” in Proc.OnziemeColloque sur le
~
Traitement du Signal et des Images, Nice, France, June 1-5, 1987.
W. A. GardnerandL.Paura, “Signal interception: Performance
advantages of cycle detectors,” in Proc. Onzieme Colloque sur le
Traitement du Signal et des Images, Nice, France, June 1-5, 1987.
W. K . Johnson, “The dynamic pneumocardiogram: An application of
coherent signal processing tocardiovascularmeasurement,” IEEE
Trans. Biomed. Eng., vol. BME-298, pp. 471-475, 1981.
K . Hasselmann and T . P. Barnett, “Techniques of linear prediction for
systems with periodic statistics,” J. Atmos. Sci., vol. 38, pp. 2275-
2283, 1981.
A. V. Vecchia,“Periodicautoregressive-movingaverage(parma)
modeling with applications towaterresources,” Water Resource
Bull., VOI.21, pp. 721-730, 1985.
A. G. Miamee andH.Salehi,“On the prediction of periodically
correlated stochastic processes,” in Multivariate Analysis- V , P. R.
Krishnaiah, Ed. Amsterdam:North-Holland,1980,pp. 167-179.
M. Paganoand E. Parzen,“Anapproachto modeling seasonally
stationary time-series,” J. Econometrics, vol. 9, pp. 137-153, 1979.
R. A. Boyles andW. A. Gardner,“Cycloergodicproperties of
discrete-parameter non-stationary stochastic processes,” ZEEE Trans.
Znform. Theory, vol. IT-29, pp. 105-114, 1983.
1321 W. A. Gardner, “Rice’s representation for cyclostationary processes,”
IEEE Trans. Commun., vol. COM-35, pp. 74-78, Jan. 1987.
-, “Common pitfalls in the application of stationary process theory
to time-sampled and modulated signals,” IEEE Trans. Commun.,
vol. COM-35, pp. 529-534, May 1987.
-, “Structuralcharacterization of locally optimumdetectors in
terms of locally optimum estimators and correlators,” ZEEE Trans.
Inform. Theory, vol. IT-28, pp. 924-932, Nov. 1982.
-, Introduction toRandom Processes withApplications to
Signals and Systems. New York: Macmillan, ,1985.
N. Wiener, “Generalized harmonic analysis,” Acta Math., vol. 55,
pp. 117-258, 1930.
R.L.Stratonovich, Topics in the Theory of RandomNoise,
Volumes I and ZI, revised English ed. translated by R.A.Silver-
man. New York:Gordon and Breach,1967.
W.A.Gardner,“Stationarizablerandomprocesses,” IEEE Trans.
Inform. Theory, vol. IT-24, pp. 8-22, Jan. 1978.
A. Papoulis, Probability, Random Variables and Stochastic Proc-
esses, 2nd ed. New York: McGraw-Hill,1984.
-, “Random modulation: A review,” ZEEE Trans.Acoust.,
Speech, Signal Processing, vol. ASSP-31, pp. 96-105, 1984.
W. A. Gardner, “Measurementof spectral correlation,” ZEEE Trans.
Acoust., Speech, Signal Processing, vol. ASSP-34, pp. 1111-1123,
Oct.1986.
-, “The spectral correlationtheory of cyclostationary time-series,’’
Signal Processing, vol. 11, pp. 13-36, July 1986.
W. A. Gardner, W. A. Brown, and C.-K. Chen, “Spectral correlation
of modulated signals:Part 11-Digital modulation,” IEEE Trans.
Commun., this issue, pp. 595-601.
’
*
William A. Gardner (S’@-M’67-SM’84) was
born in Palo Alto, CA, on November 4, 1942. He
received the M.S. degree from Stanford University,
Stanford, CA, in 1967, and the Ph.D. degree from
the University of Massachusetts, Amherst, in 1972,
both in electrical engineering.
He was a member of the Technical Staff at Bell
Laboratories in Massachusetts from 1967 to 1969.
He has been a faculty member at the University of
California,Davis, since 1972, where he is a
Professor of Electrical and Computer Engineering.
His research interests are in the general area of statistical signal processing,
with primary emphasisonthetheories of time-series analysis, stochastic
processes, and signal detection and estimation.
Prof. Gardner is the author of Introduction to Random Processes with
Applications to Signals and Systems (New York: Macmillan, 1985) and
Statistical Spectral Analysis:ANonprobabilisticTheory (Englewood
Cliffs, NJ: Prentice-Hall, 1987).