Exploitation of Higher=ChderCyclostationarity for Weak-Signal
Detection and Time-Delay Estimation
Chad M. Spooner William A. Gardner
Department of Electrical and Computer Engineering
University of California
Davis, CA 95616
Abstract
TIie cumulant theory of cyclostationary time-series
iS applied to a e v e d tVpes of weak-signal detection
problems that arise an the a m of signal intemption,
and to the problem of estimating the dative tame-
delay of a heady corrupted slgnd that w received at
two locations. The cumutant theory of cyclostationar-
aty (CS) U the theory of hagher-der tempoml and
spectml cumulants and moments of CS time-series.
Specifically, the theory characterizes the additive sine
waves present in the output of nonlinear tnansforma-
tiona of CS time-series. The detection and time-delay
atimation problems that ate posed a n daficult to solve
bemuse the sigral iS weak, the nowe and interference w
nonatationaty and non-Gawsian, and the signal does
not ezhibit sewnd-order CS.
1 INTRODUCTION
The problems considered in this paper can be for-
mulated in terms of the following two-sensor signal
model
M
z(t) = 8(t) + m,(t) + C i i ( t ) (1)
k=l
M
P(t) = &s(t + do) + mv(t)+ Aki&(t+ dk),
k 1
where 8(t) is the signal of interest (SOI),m.(t) and
my(t) are independent white Gaussian noises, the
{ik(t)}are signals not of interest (SNOIs), { A L }are
attenuation factors, and {dk} are the relative delays
between the signal components in z(t) and y ( t ) . For
example, suppose that M = 2 and the power levels
of m. and m yare time-varying. hrther, assume that
the SO1 is weak and fourth-order cyclostationary(CS)
*This work waa jointly supported by the National Science
Foundation under grant MIP-91-12800and the Army Research
Office under contract DAAM3-91-C-0018.
-,197
- but not second-order CS [8] - with period TO, and
the two interferers are second-order CS with periods
+
TI and Ts such that TI Ts = TO.The first prob-
lem is to detect the presence of the SO1 given a finite
segment of z(t). The next problem is to determine
the parameter 4 given finite segments of both ~ ( t )
and y(t). These problems are ditlicult to solve using
the stationary models of the various signals (which
leads to radiometry [4] and generalized cross correla-
tion methods [SI)becausefor detection the nonstation-
ary noise and interference complicates the threshold
setting, and for time-delay estimation the interference
corrupts the relevant phase information in the cross
spectrum. The theory of higher-order statistics is not
helpful because it also is not signal selective: all of
the signals contribute to higher-order cumulants for
a stationary signal model [8]. The theory of second-
order CS (SOCS) is also not helpful because the SO1
has no second-order cyclic features [3].However, these
two tasks (and other similar tasks) can be handled by
using the higher-order CS (HOCS) of the SOI.Before
showing how thia can be done, we first review the defi-
nitions of higher-order moments and cumulants for CS
time-series.
2 THE PARAMETERS OF HOCS
For the time-series z ( t ) for -00 < t < 00, we define
the nth-order lag-product time-series by
n
j=1
where t denotes transposition. The nth-order cyclic
tempoml moment fvnction (CTMF) is defined by the
limiting time average
e(T)n =
A Lim /v+T'2 L,(t, T)" dt
T-rooT v - T / 2
-
and is simply the Fourier coefficient associated with for many timeseries of interest and, therefore, it is
the component ei21crt in the theseries L#(t,T)n. It strictly Fourier tramformable [SI. The RD-CTCF
can be seen that the CTMF is a Fourier CoefIiCient is simply the CTCF aimdated with the n variables
of a moment function because the nth-order fraction- +
{z(t r ~ ) ) $ ~ 7" = 0. We use! the notation
with
of-time probabilistic moment (the tempnal moment
function) sssociated with L#(t,T)n c(uL be expreseed ct(U)n Ct(T)n T = [U 019 (7)
as 111
where U is an (n-l)-dimensional vector. The (n-1)-
&(t, T)n = K ( T ) n Ciarat, (4)
dimension$ Fourier transform of (7) is denoted by
a

where the sum is Over all real numbers a,called nth- P/(f')n:
order impun cycle jhquencies, for which e ( T ) n f 0.
The functions (3) and (4) exist and are well-behaved
J-00
for appropriate models of many timeseries including
amplitude modulated, pulse-amplitude-modulated, where f' = [fi fn-11'.
* * a

and phase!-shift-keyed and frequency-shirtkeyed sig- Coneider the n complex-demodulate timeseries

nals [8]. X~(t,fj)for j = l , . . * , n , see~cistedwith natrow
The nth-order temponal cumulont function (TCF) bandpasa ftltered versions of z(t), where
for the set of timeseries translates {z(t Tj)}&l is +
defined by (cf. [q and [8])
f) =
XT(~, J t+T/2
z(u)e-iar'u du. (9)
- t-T/1

T)n =
c=(t,
p={"b):,i
[t@)gR=(tsTuj)nj
P

The sum in (5) is over the set P of all distinct parti-

1 (5)
The limit as T + 00 of the limiting timeaverage of
the product of these spectral componentsis called the
spectml moment function
tions of the index set {1,2,--.,n}, where each par- I n \
tition { ~ k } i = has
~ p elements, 1 s p s n, and
k@) = (-l)P-l@- l)!. The vector Tvj consists of the
nj lags with indices in the set Vj. The cyclic tempoml
and it can be shown that Dirac deltas (denoted by
cumulont function (CTCF) is the Fourier coefficient
6(.)) can be factored out as follows:
of the TCF:

Ct(r)n = (c=(t,
A
7 ) n e-i2"fi')o , (6) Sz(t)n = C s:(t')nb(ttl
a
- a). (11)

where /3 is an nth-order pure cycle liequency for z(t)

[2, 81. Combining (4)-(6) reveals that the CTCF is
However, the factor s$(
j')n typically contains addi-
tional Dirac deltas [8]. The spectml cumdun: function
given by the following explicit function of lower-order is given by
CTMFs:
1
ct<T>n=
p=(vb):,I
)..[ ficj(Tuj)nj
a t i = f i j=l

where 1 = [ l - . .l]t, and a = [a1..-a#. An nth-

1 9

where f vj is the vector of frequenees with subscripts

order pure cycle frequency, as first defined in (21, is
the frequency of a finite-strength additive sine-wave
componentin L,(t, T ) that~ is free of all contributions
from products of sine-wave components from lower-
order lag products obtained by factoring L,(t, T)"
The CTMF and the CTCF are not in general ab-
solutely integrable due to the presence of sinusoidal
components in 7. These components formally result
in Dirac deltas in the n-dimensional Fourier trans-
form of the CTCF. However, a reduced-dimension
(RD)version of the CTCF is absolutely integrable
in the set Vj, and it follows from (8) and (11) that
Dum deltas can be factored out:
where P,B(f')n is called the cyclic plyspactrum (CP)
and does not contain Dirac deltas. As first shown in
[2], the CP is the (n-1)-dimensional Fourier transform
of the RD-CTCF (8). This is a generalization of the
Wiener relation between the power spectrum and au-
tocorrelation from second-order stationary time-series

- 198-
to nth-order CS time-series (d.111). For dtrictw s t e
tionary time-series, the only cycle frequene ( p u b or
impure) is a = I/ = 0. For a strictly stationary
stochastic process, the same is true for probabilistic
versions of the temporal momedh and cumula,nt~d o
fined in this section, which form the basis for the t h o
ory of higher-order statistics. Such parameters are not
signal selective, as noted in the next section, whereas
the cyclic parameters (p # 0, a # 0), which form the
basis for the theory of HOCS, can be signal selective.
3 SIGNAL SELECTIVITY
An important advantage of exploiting CS in sig-
nal processing tasks is that the cyclic parameters are
signal selective in that the parameters associated with
the SO1 can be estimated from data that also contains
noise and SNOIs; and as the amount of data becomes
infinite, the effects on the estimate of the noise and
interference vanish. The nature of the signal selec-
tivity properties of higher-order cyclic moments and
cumulants is examined next.
Because the signals and noises in (1) are assumed
to be mutually independent, the TCF for z(t) is given
by the simple formula [8]
M presence or absence of this signal.
cz(t,r)n = ca(t,T)n + Cm,(t,T)n + Ccih(t,7)n,
k=1
which implies that if p is an nth-order pure cycle fre-
quency for s ( t ) , and is not for any of the other signals
(it is unique to e ( t ) ) , then Cf(r)n = Cf'(r)n. Simi-
larly, if p k is a unique nth-order pure cycle frequency
for i k ( t ) , then C ~ : ( T )=
, , Ct(T)n.
For moments, the signal selectivity property de-
pends on the cycle frequencies for all of the signals
present for all orders m 5 n. This can be seen by
expressing the CTMF in terms of CTCFs [a]:
r 1
If there is a vector fl such that at least one of its el-
ements is a cycle frequency that is associated with a
SNOI, then the CTMF for z ( t ) will not be equal to
the CTMF for e(t). Nevertheless, it can happen that
the contributions to the CTMF from SNOL do not
affect the phase of the CTMF but only its magnitude.
For moments, then, there are two kinds of signal selec-
tivity, depending on what information in the moment
is considered useful (magnitude or phase). Because of
this potentially troublesome complication, the signal-
selectivity properties of cumulants are deemed more
- 199 -
useful than t h o a for moments. Thus, the following
llectiollll describe algorithms in terms of cumulants,
but in moet cases, an alternative algorithm can be
created by Simply replacingcumulanta with moments.
The aeefuloewr of the resulting alternatives depends
on the cycle frequencies sesociated with the SNOLe,
which must be evaluated on a case-by-caeebasis.
4 WEAK-SIGNAL DETECTION
In this h i o n , the problem of detecting the pres-
ence of the signal r(t) in a received data set z(t) as in
(1) is considered. There are several versions of this de-
tection problem that are of interest. The first is called
the g e w d r a d ptoblem, in which a data set is ana-
lyzed to determine if there are any CS signals present.
No information about the received data is assumed to
be known in the general search problem. In the second
problem, called the know-cycle-lnquency pmblem, a
specific pure cycle freguency/order pair (P,no) is of
interest, and it is desired to determine if there is a sig-
nal present in the data correspondingto this pair. In
the third problem, called the known-modulation pmb-
lem, the modulation format of the signal of interest is
known, and hence the cyclic cumulants of the signal
are known (in principle); it is desired to determine the
The General Search Problelq
In this problem there is a maximum order N of non-
linearity that is to be used for processing. The goal
of the processing is to produce a list of pure cycle fre-
quencies {&} for each order n IN. This list {&}
for each n characterizes the detectable CS of order n
(and only n) that is associated with z(t). Thus, these
lists are not contaminated by entries that are due to
lowersrder sine wave interactions. To accomplish this
task, the TCF is estimated for z(t) for each order n.
&om the estimate of the TCF of order n, the cycle
frequencies {/3n}, which are needed for the estimate of
+
the TCF for order n 1, can be found. More explic-
itly, the general search problem can be tackled using
the following algorithm:
0.
1.
2.
3.
4.
5.
6.
In step 4, the interval over which the average is
performed is determined by the amount of data avail-
able. If any of the detected cycle frequencies are of
particular interest, a cyclic polppectral anal* C(UL
be performed from which the modulation type can
possibly be determined [a].
The Known-Cycle-Frequency Problem
In this problem, one or more of the signal's modulation
frequencies, such (LB a symbol rate or c a d e r frequency,
is assumed to be known,but the exact functional form
of the CTCF is unknown. The environment in still
assumed to be unknown and, therefore, the general
search algorithm in still of interest. However,it can
be improved for the known-cycle-frequency problem
by combining it with a least-squares estimation tech-
nique. Let (@,%) be the cycle frequency/order pair of
interest. Use tke general search algorithm up to order
-
n~ 1. Form C : ( ~ , T ) , and
, ~ , use a least-squarea esti-
mator to detect the presence of the signal of interest
using the statistic
where
and where w is the least-squares weight vector
W
( -
w = arg min ~ W ~ C L ( ~ ei2rp112)
,T)~~
0
. (15)
The solution to (15) is w = R - ' C ~ ( T ) ,where
,~,
in which H denotes conjugate transpose. Thus, the
detection statistic is
Y =ct(T)foR-lct(T)no,
which is obtained by forming the particular linear
..I
---
combination of data sets C (t, T 1 ) n o , ,C'(t,T K ) n o
that optimally combines the regenerated sine waves
with frequency @ present in each set, and then cor-
relates this composite regenerated sine wave with the
stored sine wave ei2lPt.
The Known-Modulation Problem
In this problem, it is desired to determine if a signal
with known modulation type is present. In particular,
the CTCF of s ( t ) for n = no and pure cycle frequency
@ is known. The general search algorithm can be used
- 200 -
~~
to remove all her-order sine wwes up to order no 1. -
Then, from &;(t,r)- the CTCF estimate CT(U)no
for cycle frequency /3 can be determined by computing
the Fourier coefacient M in (6). The proposed detec-
tion statistic in
Y = f C&(u>,Cf(u):, du.
-do
The primary jwtilication for thia particular statistic
is that when no eignaI in present with nth-order pure
cycle frequency B, then C&.(U)no+ 0, which implies
that Y -D 0; when the signal of interest is present,
then
Y -+ ICt(U)nr du. (16)
--oo
Thus, 8 is an asymptotically noise-free statistic on
both the signal-present and signal-absent hypotheses.
Furthermore, the integral (16) ia finite [8]. Hence, this
statistic in the natural generalizationof the single cy-
cle detector that exploits SOCS [4].
The detection statistic Y CBD be generalized to in-
clude only a portion of u-space, denoted by G c R"0,
y= ar(u)n,w4:o du-
Choices for G might include those values of u for which
the RD-CTCF C'!(U)no is particularly large, or for
which the coefficient of variation (variance divided by
squared mean) of the estimator C&,(u)n, of the RD-
CTCF is particularly small [8].
5 TIMEDELAY ESTIMATION
Conventional approaches to the problem of esti-
mating the time-delay (or time difference of arrival
(TDOA)) between signal components in data from
two sen" can be collectively referred to (LB gener-
alized cm88 comlotion (GCC) methods [SI. In the
GCC methods, filtered versions of the sensor outputs
z ( t ) and ~ ( tare
) cross correlated, and the estimate
of do is taken to be the location of the peak in the
cross-correlation estimate. These methods suffer when
interferers are present (M 2 1 in (l)), because each
interferer contributes a peak of its own to the cross cor-
relation function. This causes two problems. The first
is a resolution problem which, to be solved, requires
that the differences in the TDOAs for each of the sig-
nals be greater than the widths of the cross correlcl
tion functions so that the peaks can be resolved. The
second problem is that it is difficult to correctly as-
sociate each peak with its correspondingsignal. Both
of these problems arise because the GCC methods are
not signal selective; they produce TDOA peaks for
all the signals in the received data unless they are
spectrally disjoint and can, therefore, be separated
by filtering. Signal-selective methods that exploit the
SOCS of the desired signal, whiah is assumed to be
unique to that signal, are studied in [SI. These meth-
ods have been shown to outperform the GCC methods,
and have been shown to produce unbiased TDOA esti-
mates with variance that is smaller than the Cramer-
Rao lower bound on the d a n c e of TDOA estimators
that are based on the assumption that the signal and
its environment are stationary. However, these meth-
ods fail when there is no SOCS to exploit. In this case,
the theory of HOCS can be used to develop signal-
selective TDOA estimatore. Following the approach
in [SI for SOCS, the methodology considered here for
HOCS is based on least-squares estimation. The fol-
lowing two examples illustrate the methodology [8].
Define a cross cumulant between n-1 time-
translates of z(t) and one translate of y(t) as follows
A
+ +
Cz,(t, 7 ) n = Cumuiant { ~ ( tTn), ~ ( tTj)}&1
The Fourier coefficient of this cross cumulant for the
cycle frequency p for the signal model (1) (assuming
that the noise and interference do not exhibit nth-
a)
order CS with pure cycle frequency is given by
Cgy(7)n =
A (Czu(t,r)ne-i2r'c)o
= + ando),,
~gct(7
where 6, is the unit vector along the nth coordinate.
It is easy to show that the following relations involving
RD-CTCFS hold:
C'g,(U)n = A&t(u-i&)nei2rBdo; C;(U), = C'!(U)n.
This suggests a least-squares fit of a measurement of
e&, to a measurement of over a region G of u-
space of interest:
which leads to the following estimator of the delay 4:
This estimator is a higher-order generalization of the
SPECtral Coherence Alignment algorithm for TDOA
estimation [5), which exploits SOCS, and has been
shown to possess several optimality properties.
As an alternative, cross-sensor measurements can
be avoided entirely by noting that
Cz ( U)n = c!(u>n,
'
C,"<U),= ~C~(u)nei2*'do;
-201
which suggests the following least-squares approach:
The estimator for 4 is given explicitly by
which is a higher-order generalization of the second-
order Cyclic Phase Merence algorithm for TDOA e%
timation without cross-sensor measurements [SI.
6 CONCLUSIONS
The higher-order cumulants and moments of cyclo-
stationary time-series can be used to perform diffi-
cult detection and estimation tasks. The basic signal-
selectivity properties of cumulants yield estimators
that are tolerant to noise and cyclostationary inter-
ference. This admittedly terse presentation will be
expanded on in a forthcomingjournal paper (see [a]).
References
[l] W. A. Gardner. Statistical Spectral Analysis: A Non-
pmkbilistic Theory. Englewood CWs, New Jersey:
PrenticaHd 1987.
[2]W.A. Gardner and C. M. Spooner. "Highex-Order
Cyclostationarity, Cyclic Cumulants, and Cyclic
Polyspectra," in Proceedings of the Intemationd
Symposium on Information Theory and its Applica-
tions (ISITA), Honolulu, Hawaii, November 1990.
[3] W. A. Gardner. "Exploitation of Spectral Redun-
dancy in Cyclostationary Signals," IEEE Signal Pro-
cessing Magazine, Vol. 8, pp. 14-36, April, 1991.
[4]W. A. Gardner. "Signal Interception: A Unifying
Theoretical Ramework for Festure Detection", IEEE
Zhm. Comm., Vol. 36,pp. 897-906,August 1988.
[5] W. A. Gardnex and C. K. Chen. "Signal-Selective
Time-Difterence-of-AmvalEstimation for Psssive Lo-
cation of Man-made Signal Sources in Highly Cor-
ruptive Environments: Parts I and 11," IEEE Zhns.
Signd Pm.,Vol. 40,pp. 1168-1197,1992.
161 C. H. Knapp and G. C. Carter. "The Generalized
Correlation Method for Estimation of Time Delay,"
IEEE lknsaciions on AcowticJ, Speech, and Signal
Processing, Vol. 24,August 1976.
[I V. P. Leonov and A. N. Shiryaev. "On a Method of
Calculation of Semi-Invariants," Theory of Pmbabd-
i f y and Its Applications, Vol. 4,pp. 319-328,1959.
181 C. U. Spooner. "Theory and Application of Higher-
Order Cyclostationarity," Ph.D. Dissertation, Dept.
of Electrical and Computer Engineering, University
of California, Davis, CA, June 1992.
-