IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 42, NO.

5, MAY 1994 2165

Robust Feature Detection for Signal Interception

Chad M. Spooner, Member, IEEE, and William A. Gardner, Fellow, IEEE

Abstract-The problem of detecting the presence of direct Section 11) for the regenerated spectral line at the pulse-rate
sequence spread-spectrumsignals in noise is considered, and the frequency [8]. This detector is the special case of the DM
detector consisting of a filter followed by a delay-and-multiply detector for which the prefilter is the matched filter and the
device is optimized with respect to the filter bandwidth and
the delay to maximize robustness to errors in assumed values delay is zero. This detector, however, is shown to be not as
of pulse rate and carrier frequency. Output SNR is used for robust to errors in pulse rate and carrier frequency as that
performance optimization, and the results are corroborated by which is optimized for robustness.
evaluating receiver operating characteristics. In [IO] the problem of detecting the presence of a distor-
tionless perfectly down-converted BPSK signal (viz., a real
PAM signal whose pulse transform is real) by using a DM
device is studied. An output SNR is evaluated there and used
to determine several optimum prefilter-and-delay structures,
I. INTRODUCTION as well as to determine optimum bandwidths for rectangular
prefilters for a small set of delays. The effect of an error
T HE PURPOSE of this paper is to present the results
of a study of the robustness of the quadratic detector
consisting of a filter followed by a delay-and-multiply device
in the knowledge of the pulse rate on the performance of
these detectors is found in terms of output SNR and some
that multiplies the filtered signal by a delayed and conjugated approximate formulas for the probabilities of detection and
replica of itself. This delay-and-multiply (DM) detector is false alarm are derived.
commonly used to regenerate a spectral line at a frequency This paper’ generalizes on the work in [IO] by 1 ) jointly
equal to the pulse rate of a PCM signal for purposes of optimizing the delay and rectangular prefilter bandwidth pa-
detection of the signal’s presence [1]-[5], [lo] or synchro- rameters with respect to two distinct optimality criteria that are
nization to the signal’s pulse timing phase 161-[9]. Optimizing aimed at maximizing robustness to error in signal parameter
the DM detector with respect to the filter bandwidth and values; 2) considering the general bandpass (or complex base-
the delay requires knowledge of the pulse rate and carrier band) case as well as the real baseband case; 3) considering
frequency. When the pulse rate or carrier frequency assumed the effect of errors in the knowledge of the demodulated
for optimization is in error, the bandwidth and/or delay that carrier offset (or carrier frequency); and 4) performing simu-
maximizes detection performance for the erroneous pullse rate lations to estimate the detection and false-alarm probabilities
or carrier frequency is suboptimum for the actual pulse rate and corresponding to the various detectors. The relationship be-
carrier frequency. This suggests maximizing the robustness of tween detection by spectral line generation and the theory of
the detector, as measured by the degree of tolerance to error cyclostationary time-series is also clarified.
in pulse rate and carrier frequency. In (31-151 the related problem of designing the signal (i.e.,
In this paper, both output SNR and receiver operating char- designing the pulse shape and the distribution of the symbol
acteristics (ROC’s) are used as measures of performance for sequence) to reduce detectability by reducing the strength
studying the robustness of the DM detector for baseband PCM of the spectral line that can be generated at the pulse-rate
signals, modeled as real PAM, and passband PCM signals frequency is addressed.
(digital QAM, BPSK, QPSK), modeled as complex PAM, in In Section 11, the signal models are defined, and the output
white Gaussian noise (WGN). Since the primary application SNR for a regenerated spectral line is defined and expressed
motivating this work is detection of spread-spectrum signals explicitly in terms of the parameters of the signal, noise, and
in noise, only low-SNR conditions are considered. Further- detector. This SNR is maximized in the Appendix for both the
more, low-SNR conditions give rise to some simplifications in real PAM and complex PAM models. In Section 111, the SNR
the theory. is used to study the optimization of the filter bandwidth and
To put the DM detector in perspective, it is pointed out that the delay, and to evaluate the effects of errors in the values of
the weak-signal likelihood-ratio detector for PAM in WGN pulse rate and carrier frequency used in the optimization. In
is a matched filter followed by a magnitude squarer [I]. Section IV the conclusions regarding detection performance
This detector also maximizes the output SNR (as defined in are corroborated using ROC’s obtained from simulations.

Paper approved by C. N. Georghiades, the Editor for Synchronization and

Optical Detection of the IEEE Communications Society. Manuscript received
March 15, 1991; revised July 28, 1992.
The authors are with the Department of Electrical and Computer Engineer- ‘The work reported in this paper was conducted prior to the appearance
ing, University of California at Davis, Davis, CA 95616. of [ I O ] in the literature, which occurred almost simultaneously with the
IEEE Log Number 9213529. submission of this paper for review for publication.

0090-6778/94$04.00 0 1994 IEEE

2166 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 42, NO. 5 , MAY 1994

11. OUTPUTSNR FORMULAS where w ( t ) is the filtered input

A . Signal Models w ( t ) = z ( t )@ h ( t ) (9)
The model for a real baseband PAM signal is and h ( t ) is the impulse response of the filter preceding the
cc
delay-and-multiply device. From (6)-(9), we see that the
S(t) = anp(t - 7lTo - e) (1) output
n=-m

where p ( l ) is the pulse, {a,} is a zero-mean, independent,

g ( t ) = [ s ( t )@ h ( t ) ] [ s (-t d ) 63 h(t - d)]*
and identically distributed symbol sequence, and 1/7h and 0 + [ ~ (@t )h ( t ) ] [ ~-~d( )t 8 h(t - d ) ] *
are the pulse rate and phase, respectively. For spread-spectrum + [ 7 1 ( t ) @ h(t)][s(t - d ) @ h(t - d)]*
signals, this model for { a n } is admittedly only approximate.
For example, for a binary direct-sequence spread-spectrum
+ [ 7 L ( t ) @ h(t)][n(t - d ) @ h(t - d ) ] * (10)

signal, {(A7%} is actually the modulo-two sum of the high-rate

spreading code sequence and the low-rate message symbol se- consists of signal-only, noise-only, and signal-cross-noise
quence. Also, for spread-spectrum signals, unlike bandwidth- terms. For the case of a weak signal, we can ignore the cross
efficient PCM, p ( t ) has large excess bandwidth (bandwidth terms in defining the output SNR; that is, we can compute the
in excess of the minimum bandwidth for zero intersymbol output signal strength from the signal-only term and the output
interference) and can be usefully approximated as a rectangu- noise strength (approximately) from the noise-only term.
lar pulse
B. Output SNR
Since the sequence {a,} has zero mean value, then s ( t )
In this case, the complex counterpart of model (1) applies to does not contain any spectral lines. However, the output signal-
the complex envelope of both digital QAM and PSK signals only term will in general contain spectral lines at the pulse rate
(for which the phase-keying envelope is rectangular). For this and its harmonics k/To, IC = f l ,f 2 , . . .. The output signal
passband signal model, y ( t ) and { u n } are both complex-valued strength is defined to be the time-averaged power of one of
in order to accommodate any carrier offset, say fo, and lack these regenerated spectral lines
of Hermitian symmetry of the baseband pulse transform, as
well as to accommodate the in-phase and quadrature symbol
I', 2 I([@) @ h ( t ) ] [ s (-t d ) @ h(t - d ) ] * p - L 2 q 2 ( 1 I )
sequences. For example, for the real digital QAM signal where (.) denotes average over all time t , and YI = IC/To.
The output-noise strength is defined to be the time-averaged
s(t)= c,,y(t - nTo - 0) cos(2Tf0t - p)
power in a very narrow band of width B,, centered at
71
frequency a, due to the noise-only term
+ s,,g(t - 6) sin(2.lrjot -
- ni~o v) (3)
m
pn =
A 1 a+I3/2

n-B/2
Sm(f)df
we have the complex envelope (1) with
= B,Sn,(Q)
where m ( t ) is the noise-only term
and A
m ( t ) = [ 7 L ( t ) 8 h(t)][TL(t- d ) @ /L(t - d)]* (12)
and Sm(f)is its spectral density of the time-averaged power.
It should be clarified that the model for s ( t ) can be complex- The output SNR is then given by
valued even when the carrier frequency fo is known (which
implies that perfect downconversion is possible) because, a Ps
SNRz-. (13)
for instance, the channels's impulse response can force the Pn
spectrum to be asymmetrical about the carrier frequency, For the real signal and noise model (1) (with { a n } ,p ( t ) , and
which results in a complex-valued baseband pulse in the 7 1 ( t ) real), it is shown in the Appendix that
complex-envelope model.
The complex envelope of the input to the detector consists
of the signal ( 1 ) plus complex WGN
"= $04
I .iI, H(f +f" -

x(1,) = s ( t ) + 71(t) ,
x P (f + a / 2 )P*( f - 4 2 ) P f d df
The cutput of the detector is given by
74t) =7 4 t ) +i71S(t).
y(t) = 7u(t)w*(t - d )
SPOONER AND GARDNER: DETECTION FOR SIGNAL INTERCbPTIOh 2167

For the complex signal and noise model (1) (with { a n } ,p ( t ) , 2, I 1

+
and n(t) = n,(t) in,(t) complex), it is also shown in the
Appendix that

Fig. I . Robustness to error in TO for peak-optimal and width-optimal

bandwidth B and delay d in the delay-and-multiply (DM) detector, and for
the matched-filter-squarer(MFS) detector, for real PAM.

In (14) and (15) crf is the temporal variance of the zero-mean

strength (Fourier coefficient) of the sine-wave component with
sequence {a,} :
frequency Q in the output, then the detection statistic becomes
h'

It is further shown in the Appendix that the SNR (14) ior the
real signal is maximum with respect to H ( f ) if a solution to

H(f + (2/2)H*(:f - Q / 2 ) Ix This statistic can be re-expressed as

I P o 3 I

where Sz(f)T is an estimate of the spectral correlation

exists; for d = 0, the solution is function or cyclic spectrum of .c(t) obtained by Fourier
transforming the estimate of the cyclic autocorrelation of r ( t )

This is recognized as the matched filter for the pulse p ( t ) in

X;(7)T = ;L T
:c(t + 7/2)z*(t,- 7/2)e,-62*%lt (22)

noise with power spectral density S,,(f). This optimum DM and where W a ( f )is given by
detector- the matched-filter squarer (MFS)-is identical to
the optimum (maximum SNR for regenerated spectral line) ""(f) =H ( f + a/2)H*(f - u/2) (23)
among all quadratic detectors-not just those of the filter- which plays the role of a spectral smoothing window in (21).
and-delay-and-multiply form [8], [9]. When d = 0 and H ( f ) is the matched filter (17) or (19),
Similarly, it is shown in the Appendix that the SNR (15) then (21) is identical to the optimum feature detector described
for the complex signal is maximum with respect to H ( f ) if in I I].
a solution to

H(f + U / 2 ) H * (f - Q / 2 ) Ix
EVALUATION
111. ROBUSTNESS AND OPTrMlZATlON

The usual approach to optimizing a detector, such as the

DM detector, is to ignore possible errors in parameters such as
the pulse rate and to simply maximize a performance measure
such as SNR for some nominal parameter value. Following
exists and again, for d = 0, we obtain the matched filter this approach first, we adopt an idealized filter with rectangular
solution passband and one-sided bandwidth 8 , and we maximize the
SNR (14) or (15) with respect to both the delay d and the
bandwidth B . The resultant peak-optimal detector exhibits
the SNR versus pulse-rate error performance shown in Fig. I
Before proceeding, the relationship between the DM detec- (cf. [lo], Figs. 5 and 61 for the real PAM model and in Fig. 4
tor and the more general class of cyclic feature detectors that for the complex PAM model. The peak-optimal values of
exploit the spectral correlation property that is characteristic bandwidth and the resultant peak SNR versus delay are shown
of cyclostationary signals, as described in [I], is clarified. If in Fig. 2 for real PAM and in Fig. 5 for complex PAM.
we try to detect the presence of a regenerated spectral line at As an altemative to the preceding approach, to obtain a
frequency CY in the output of the DM detector by measuring the detector that is more robust to errors in the assumed value
2168 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 42, NO. 5, MAY 1994

I 1 I I I
0

Ped-Optimal
Width-Optimal
-
-
-8
SNR
(W
-16

-24
-10 -8 -6
iO\og(T~~/T~)

Fig. 2. Peak-optimal bandwidth B and peak SNR for the DM detector versus Fig. 4. Robustness to error in To for the DM detector with peak-optimal
delay d for real PAM. The vertical axis is in decibels for peak SNR and is and width-optimal bandwidth E and delay d , and for the MFS detector, for
dimensionless for BTo. complex PAM.

Peak SNR (dB) -

-4
-5
0' _I
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.8
d/To d/To O''

Fig. 3. Width-optimal bandwidth B and pulse-rate tolerance AT0 for the Fig. 5. Peak-optimal bandwidth B and peak SNR for the DM detector versus
DM detector versus delay d for real PAM. AT0 is the -3dB-width of the delay d for complex PAM.
peak-SNR versus pulse-rate-error curve (see, for example, Fig. 1).

of pulse rate, we maximize the -3-dB width of the SNR

versus pulse-rate error performance curve with respect to the
delay d and bandwidth B. The resultant width-optimal detector
exhibits the SNR versus pulse-rate error performance shown
in Fig. 1 for real PAM and in Fig. 4 for complex PAM.
The width-optimal values of bandwidth B and the resultant
tolerance (-3-dB width of SNR versus pulse-rate curve) to
pulse-rate error versus delay d are shown in Fig. 3 for real
0' I
PAM and in Fig. 6 for complex PAM. 0 0.2 0.4 0.6 0.8 1
d/To
The results in Figs. 1-6 were obtained for a signal and noise
model with the following parameters: Fig. 6. Width-optimal bandwidth B and pulse-rate tolerance AT0 for the
DM detector versus delay d for complex PAM.
ua = 1
S,(f) 3 I or S n C ( f E
) 1
are the jointly peak-optimal detector parameters and, from
f o = 1/4T, Fig. 1, these parameter values result in an SNR that is only
To = 16T, (24) 0.6 dB below that of the optimum (MFS) detector, which
is used as a reference (0-dB SNR). From Fig. 3, it can be
seen that
where T, is the sampling increment for a discrete-time im-
plementation, and fo is the carrier-frequency offset in the d = 0.557To
downconverted complex PAM model. The spectral line fre- B = 1.452/To
quency a used was the pulse rate l/To.(Note: The jagged
appearance of the curves in Figs. 2,3, and 6 is a result of the are the joint width-optimal detector parameters for real PAM,
fact that the optimization search was over discretized versions and from Fig. 1 it follows that the width-optimal detector, by
of the parameters d, B, Test). comparison with the peak-optimal, provides up to 2 dB more
For real PAM, it can be seen from Fig. 2 that SNR for pulse rates l/To that are lower than the expected
d = 0.45To rate l/Test[log(Test/T~) < 01, but also results in up to 2 dB
less SNR for some pulse rates that are higher than expected.
B = 1.445/To Although the peak value of width-optimal SNR is 1.7 dB
SPOONER AND GARDNER: DETECTION FOR SIGNAL INTERCEWION 2169

below that of the optimum MFS detector, both the peak- 0-.

optimal and width-optimal detectors are far superior to the

MFS for some pulse rates that are higher than expected, -
-6
whereas they are comparable to the MFS for pulse rates that
are lower than expected. SNR MFS -
-
A practical compromise between the peak-optimal and (dB)
-12 -
Peak-Opltmal
width-Optimal -
width-optimal parameter values is d = To/2 and B = 1.5T0.
For complex PAM, the results are a little more interesting. It
follows from Fig. 5 that the peak-optimal parameter values are -18 -

d=O
B = 0.85/To
and we can see from Fig. 4 that this results in an SNR that
is 1.55 dB lower than that provided by the optimum MFS
detector. From Fig. 4, it follows that specified in (24). Thus in the case of real signals, a baseband
d = 0.52To binary PAM signal with To = lGT, is simulated, and in the
B = 1.4o/To case of complex signals, the same real signal is simulated
and then modulated by a complex exponential with fre-
are the width-optimal parameter values, and Fig. 4 shows quency 1/4Ts.
that the width-optimal detector provides up to 10 dB more For each ROC curve, 500 independent statistics for each
SNR than the peak-optimal detector for pulse rates that are hypothesis are computed. Each statistic is computed in the
lower than expected and even more than 10-dB improvement following way. The signal is generated (on the signal-present
for higher pulse rates. However, this results in 2.65-dB loss hypothesis only) and added to WGN that is simulated using an
relative to the peak-optimal and 4.2-dB loss relative to the IMSL library routine. These data are filtered by the appropriate
optimal MFS detector when there is no error in the expected H ( f ) and then multiplied by delayed versions of themselves.
pulse rate. On the other hand, the MFS detector can perform The ideal rectangular filter is implemented in the frequency
quite poorly compared with the width-optimal detector when domain by simply zeroing the appropriate bins in the FFT
the actual pulse rate is higher than expected. of the signal. The delay and the bandwidth of the filter are
The robustness optimization for complex signals was done assumed to be specified in terms of the estimated value of the
for the case of a real-valued pulse and a perfect match between pulse rate and are therefore determined by Test.An FFT is
the filter’s center frequency and the signal’s center frequency. applied and the bin corresponding to 1/To, the true pulse rate,
In this case, the pulse-rate line appears in the real portion is selected. The magnitude of this bin is the detection statistic.
of the delay-and-multiply product. The complex statistic is The value of the spectral density height for the noise changes
nevertheless used because it provides a best case complex- from case to case (but remains constant for all the ROC’s in
signal robustness evaluation: for any other case, a specific a single figure).
complex pulse (channel impulse response) or carrier mismatch Figs. 8 and 9 show ROC’s for the MFS and peak-optimal
must be chosen. detectors for real signals with TeSt/To= 0.5 ( - 3 dB) and
In addition, Fig. 7 shows the robustness of the three de- 1.6 (1.9 dB), respectively. These ROC’s show that the per-
tectors to errors in the assumed value of carrier offsel for formance difference between the two detectors is correctly
complex PAM signals. The data in this figure were generated predicted by the separation between the SNR curves in Fig. 1.
by evaluating the SNR’s for the case of a mismatch between Figs. 10 and 11 show ROC’s for the case of complex signals,
the center frequency of the filter and the center frequency using the same two values of !!‘,,,/To as in the previous two
of the signal’s power spectrum (the center of P ( f ) does not figures. These figures validate the performance ordering given
match the center of H ( f ) ) . Although the peak-optimal and by the SNR curves in Fig. 4.
MFS detectors exhibit similar degradations due to this type of Figs. 12 and 13 show ROC’s for all three complex-signal
error, the width-optimal detector is substantially more tolerant. optimal detectors for the case of a carrier-offset error. That is,
In summary, by accepting a few decibel loss in SNR there is a mismatch between the signal’s carrier offset and the
for the best case of no error in the assumed pulse rate or center frequency of the pre-filter H(f). In Fig. 12, this offset
carrier frequency for complex PAM, we can obtain subvtan- is 1% of the pulse rate 1/To and in Fig. 13, it is 2% of 1/To.
tial improvements in tolerance to both of these errors when These ROC’s corroborate the SNR curves in Fig. 7.
they are relatively large by using the width-optimal parame-
ters values. V. CONCLUSIONS
For real PAM signals, both the peak-optimal and width-
IV. SIMULATIONS optimal DM detectors offer enhanced tolerance, relative to
To corroborate the results and conclusions drawn in the MFS detector, to error in the assumed value of the pulse
Section I11 on the basis of SNR, we show here a sampling of rate of a signal to be detected when the actual value exceeds
ROC’S obtained for a BPSK signal with parameter values as the assumed rate by more than 1 dB. However, for complex
2170 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 42, NO. 5, MAY 1994

0.9

0.8 0.6
PD
0.7

0.6

0.5
MFS -
Peak-Optimal - n3
0.4 Y MFS
Peak-Optimal
Width-Optimal
-
-
-
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8
PFA PFA

Fig. 8. Receiver operating characteristics (ROC’s) for a real signal: Fig. 12. ROC’s for a complex signal: carrier offset error of lOlog
lOlog(T,,st/To) = -3 dB. (Tofoffsfset) = -8 dB.

PD

Peak-Optimal
Width-Optimal -
-

Fig. 9. ROC’s for a real signal: lOlog(T,.,t/To) = 1.9 dB. Fig. 13. ROC’s for a complex signal: carrier offset error of lOlog
(TofotTeet) = -5 dB.

that are lower than the assumed rate, the width-optimal de-
tector-compared with the MFS detector-can provide up to
2 dB higher SNR for real PAM but can be nearly 2 dB worse
for complex PAM, and is as much as 4 dB worse when the
PD
pulse-rate error is between -1 and +1 dB for complex PAM.
However, to offset this, the width-optimal detector performs
well for carrier-offset errors (normalized by the pulse rate)
Peak-Optimal -
as large as -1 dB, whereas the MFS detector fails for errors
exceeding -5 dB.
In conclusion, by comparison with the optimum MFS de-
tector, the optimally robust DM detector offers enhanced
Fig. IO. ROC’s for a complex signal: lOlog(T,,t/To) = -3 dB.
tolerance to errors in assumed values of pulse rate and carrier
offset. However, if the objective is to design a detector that
performs as well as possible for all possible carrier frequencies
0.8 - and pulse rates, then the prefilter/delay/multiply followed by
a spectral-line detector is not necessarily the most appropriate
quadratic device. The cyclic spectrum analyzer (also called the
PD
spectral correlation analyzer) [ 11, [9] provides more flexibil-
ity for search-type detection and, with linear postprocessing
MFS -
Peak-Oplimal - (weighted integration over frequency), can implement the
optimum quadratic detectors for all possible carrier frequencies
and pulse rates. Moreover, it can implement the optimum
OO’ 0.2 0.4 0.6 0.8
PFA quadratic detector for any cyclostationary signal-not just the
PSK signals considered in this paper-simply by choosing the
Fig. 1 1 . ROC’s for a complex signal: lOlog(T,,t/To) = 1.9 dB. appropriate linear postprocessing.

PAM signals, the peak-optimal detector performs quite poorly, APPENDIX

whereas the width-optimal (or optimally robust) detector again SNR OPTIMIZATION
provides enhanced tolerance to actual pulse rates that exceed In this Appendix the performance measure SNR (13) is
the assumed rate by more than 1 dB. For actual pulse rates derived and maximized for a PAM signal in white Gaussian
SPOONER AND GARDNER: DETECTION FOR SIGNAL INTERCEPTION 2171

noise. Recall that the SNR is given by Since the symbol sequence {a,} is independent and identically
distributed, then
ps
SNR= K

where Sm(cy) is the PSD of the output of the DM with noise where p = ', *', *2, ' ' '. Therefore
only at the input. The calculation of ( A l ) is split into numerator (b(t)e-iZ"Qt) =
and denominator calculations.

Numerator Calculation
The signal model for s ( t ) is given by (1). Let b ( t ) denote x ~ * (-tnTo - 8 - d)ePiaTatd t , (A9)
the output of the DM with s ( t ) at the input
b ( t ) = [ s ( t )8 h ( t ) ] [ s (-t d ) 8 h(t - d ) ] * (A2) for cy = p / T , . Expressing y(t) in terms of its Fourier transform
c c m

= a,n;g(t-nT" -0)

In (A3), g ( t ) is the filtered pulse

in (A9) yields
d t ) = P ( t ) 8 Il(t>' ('44)

The value of (b(t)e-i2"at) is sought. Starting with

where a = p / T o , and P ( f ) and H ( f ) are the Fourier

A change of summation indices for n and m and subsequent
transforms of p ( t ) and h ( t ) ,respectively. Note that for even
simplification yields
G ( f ) = P ( f ) H ( f ) ,(A12) reduces to
2172 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 42, NO. 5, MAY 1994

Denominator Calculation
The noise is bandlimited and Gaussian, either lowpass (in
the case of real signals) or bandpass (in the caFe of complex
signals) and has a constant spectral height in its passband. Complex Signal: Here R n c ( ~ ) Rns(7),which implies
In either case it is assumed that the noise 7 i ( t ) has the that (A 19) vanishes. The spectral density therefore becomes
representation
7 4 ) = 7 1 , ( t ) cos(27rf c t )- n s ( t )sin(27rfct) . (A14)
In the case of real signals, f c = 0 and n ( t ) n,(t). whereas
in the case of complex signals f c can be taken to be the center
frequency of the signal of interest. In either case, if we assume
that the spectrum of the noise is symmetric about f , , then the Combining (AI), (A12), and (A20) yields the SNR (14) for
crosscorrelation between n,(t) and n,(t) is zero. real signals. Combining (AI), (A12), and (A21) yields the
The input to the DM receiver is the complex envelope SNR (15) for complex signals.
7L,(t) = n,(t) + ms(t) (A151
The Ratio
and the corresponding output is m(t).To find the PSD of m(t),
the autocorrelation is computed and then Fourier transformed. To obtain the optimum SNR in (15), the numerator, which
With u ( t ) = n,(t) @I h ( t ) , the autocorrelation function for is the squared magnitude of (A12), is written as

'(f + + ) P * ( f - 4 )

An application of Isserlis' formula [9] to (A16) yields

Rm(7) =
x H ( f + :)H*(f- G) df /2
I&(d)12 + lRU(.)I2 + L L ' ( 7 + d)R;,, (7 - 4. and the Schwarz inequality is applied. This leads to the
(A171 sufficient optimal-filter design equation

It is straightforward to show that

H(f + :)H*(f - %) 'x
RU(.) = I[ h(X)\z(r)
x [ R n , (-~ X + 7 ) + %,(. - + r)]d X d r
( ~ 1 8 ) for complex signals. The maximum SNR for real signals
(14) can be obtained similarly. This results in the sufficient
optimal-filter design equation

H(f++)H*(f-%) K

P * ( f + $ ) P ( f - +)e--12=fd
('423)
cos2(27rfd)Sn(f$4)S,(f- 4 )

for real signals. Equations (A22) and (A23) are only suffi-
The spectral density ST,L(cy) for 0 # 0 is given by the Fourier cient, not necessary, because the Schwartz inequality yields a
transform of the last two terms in (A17). The special cases of maximum Only when a solution to (A22) or (A23) exists. For
real and complex signals are treated separately. example, for d = 0, solutions exists, but for rl # 0, solutions
Real Signal: Here 7 1 , ( t ) E 0, which implies that R n S ( r E
) do not necessarily exist. However, if G ( f ) = p ( f ) H ( f )
0. Thus the integrals in (A18) and (A19) are identical. The turns out to be even, then using (A13) in Place of (AI21
spectral density therefore becomes results in the replacement of e--L2?Tfd in (A22) and (A23)

cos2(27rf$)l~*( f - " ) ~ ( f 4.
2
c)1' with ros(27rfd). In this case, solutions do exist for some
d # 0, e.g., d = r M / 2 c ~ for odd integers T , since then
+
('OS ( 2 T f d ) = 2cos(7r(f IY/2)d) cos(7r(f - 4 2 ) d ) .
SPOONER AND GARDNER: DETECTION FOR SIGNAL INTERCEPTION 2173

REFERENCES William A. Gardner (S’64- M’67 - SM’84 -F’9 I )

was bom in Palo Alto, CA, on November 4, 1942.
W. A. Gardner, “Signal interception: A unifying theoretical frame- He received the M.S. degree from the Stanford
work for feature detection,” IEEE Trans. Commun., vol. 36, no. 8, University, in 1967, and the Ph.D. degree from the
pp. 897-906, Aug. 1988. University of Massachusetts, Amherst, in 1972, both
I21 R. E. Ziemer and R. L. Peterson, Digital Communicutions and Spread in electrical engineering.
Spectrum Systems. New York: Macmillan, 1985. He was a member of the Technical Staff at
(31 D. E. Reed and M. A. Wickert, “A performance comparison of opti- Bell Laboratories in MA, from 1967 to 1969. He
mum and sub-optimum receiver structures for rate-line detection of has been a faculty member at the University of
digitally modulated carriers,” in Proc. IEEE Region Five Con$, 1988, California, Davis, since 1972, where he is Professor
pp. 177-181. of Electrical Engineering and Computer Science.
[41 -, “Minimization of detection of symbol-rate spectral lines by Since 1982, he has also been President of the engineering consulting firm
delay and multiply receivers,” IEEE Trans. Commun., vol. 36, no. I , Statistical Signal Processing, Inc., Yountville, CA. His research interests are
pp. 188-120, Jan. 1988. in the general area of statistical signal processing, with primary emphasis on
151 -, “Spread spectrum signals with low probability of chip-rate the theories of time-series analysis, stochastic processes, and signal detection
detection,” IEEE J. Selected Areas Commun., vol. 7, no. 4, May 1989. and estimation and applications to communications and signals intelligence.
L6l J. Imbeaux, “Performances of the delay-line multiplier circuit for clock Dr. Gardner is the author of Introduction of Random Processes with
and carrier synchronization in digital satellite communications,” IEEE Applications to Signals and Systems, Macmillan, New York, 1985, second
J. Selected Area,s Commun., vol. SAC-I, no. 1, pp. 82-95, Jan. 1983. edition, McGraw-Hill, New York, 1990, The Random Processes Tutor: A
[71 A.N. D’Andrea and U. Mengali, “Performance analysis of the delay- Comprehensive Solutions Manual for Independent Study, McGraw-Hill, New
line clock regenerator,” IEEE Trans. Commun., vol. COM-34, no. 4, York, 1990, and Statistical Spectral Analysis: A Nonprobabilistic Theory,
pp. 321-328, Apr. 1986. Prentice-Hall, Englewood Cliffs, NJ, 1987. He holds several patents and is the
L81 W. A. Gardner, “The role of spectral correlation in design and perfor- author of numerous research-journal papers. He received the Best Paper of the
mance of synchronizers,” IEEE Trans. Commun., vol. COM-34, no. 1 I. Year Award from the European Association for Signal Processing in 1986 for
pp. 1089- 1095, Nov. 1986. the paper entitled “The spectral correlation theory of cyclostationary signals,”
-, Statistical Spectral Analysis: A Nonprobabilistic Theoty. En-
the 1987 Distinguished Engineering Alumnus Award from the University of
glewood Cliffs, NJ: Prentice-Hall, 1987.
J. F. Kuehls and E. Geraniotis, “Presence detection of binary-phase-shift- Massachusetts, and the Stephen 0. Rico Prize Paper Award in the Field of
keyed and direct-sequence spread-spectrum signals using a prefilter- Communication Theory from the IEEE Communications Society in 1988 for
the paper entitled “Signal interception: A unifying theoretical framework for
delay-and-multiply device,” IEEE J. Selected Areas Commun.. vol. 8,
feature detection,” He organized and chaired the NSF/ONR/ARO/AFOSK-
no. 5, pp. 915-933, June 1990.
sponsored workshop on Cyclostationary Signals, 1992. He is a member of
the American Association for the Advancement of Science, the European
Association for Signal Processing, and a member of the honor societies Sigma
Xi, Tau Beta Pi, Eta Kappa Nu, and Alpha Gamma Sigma.
Chad M. Spooner (S’88-M’92) was born in Des
Moines, IA, on October 5, 1963. From 1981 to 1984
he attended Santa Rosa Junior College, and was
awarded the A.S. degree with high honors in 1984.
He received the B.S. degree from the University
of California at Berkeley in 1986, and the M.S. and
Ph.D. degrees from University of California at Davis
in 1988 and 1992, all in electrical engineering.
He was a Teaching Assistant in the Department of
Electrical and Computer Engineering ut UC Davis
from 1986 to 1988. From 1988 to 1992 hs was
a Research Assistant at UC Davis and a Consultant to Statistical Signal
Processing, Inc. He is currently a postgraduate researcher at UC Davis, and
is coauthor (with Stephan Schell) of a unique commercially available cyclic
spectral analysis software package. His research interests include statistical
signal processing, weak-signal detection and parameter estimation, and the
1 heory of higher order cyclostationary signals.