1124 IEEE TRANSACTIONS ON ACOUSl'ICS. SPEECH, AND SIGNAL PROCESSIKG, VOI.. ASSP-34, N O .

5 , OCTOBER 1986

Adaptive Comb Filtering for Harmonic Signal

Enhancement

Abstract-A new algorithm is presented for adaptive comb filtering The filter appearing in the algorithm consists of a cas-
and parametric spectral estimation of harmonic signals with additive cade of second-order infinite impulse response (IIR) sec-
white noise. The algorithm is composed of two cascaded parts. The first
estimates the fundamental frequency and enhances the harmonic com-
tions,andhas a comb-typefrequencyresponse.These
ponent in the input, and the second estimates the harmonic amplitudes second-order sections are parametrized by a single vari-
and phases. Performance analysis provides new results for the asymp- able:theestimatedfundamentalfrequency of thehar-
totic Cramer-Rao bound (CRB) on the parameters of harmonic signals monic signal. This yields an estimation algorithm for the
with additive white noise. Resultsof simulations indicate that the vari- fundamental frequency which is more efficient and accu-
ances of the estimates are of the same order of magnitude as the CRB
for sufficiently large data sets, and illustrate the performance in en-
rate than others (for example, [l] and [2]) which view the
hancing noisy artificial periodic signals. sinusoidal frequencies as independent. In addition, simi-
larly to [SI, special constraints are imposed on the filter
coefficients, giving arbitrarily narrow-band-pass filters for
I. INTRODUCTION each harmonic. This improves the performance in com-

S IGNALS that consist of a sum of sine waves whose parison to other comb filters of finite impulse response
frequenciesareintegralmultiples of thelowestfre-
quency (so-called fundamental) are said to be harmonic.
(FIR) type (for example, [6] and [7]), which typically re-
quire a large number ofcoefficients to obtain narrow pass-
Many physical signals are approximately harmonic. Ex- bands. The harmonic amplitudes and phases are estimated
amplesincludevoicedspeechandotherbiologicalsig- separately, conditioned on the estimated fundamental fre-
nals,musicalwaveforms,helicopterand boat sound quency.Thealgorithm is computationally efficient; the
waves, and outputs of nonlinear systems excited by a sin- number of operations it requires per time sample is pro-
usoidal input. To filter noise corrupted harmonic signals portional to the squared numberof the filtered harmonics.
whose parameters are unknown and possibly time vary- Section I11 is devoted to performance analysis of the
ing, it is desirable to apply adaptive filtering. Most exist- algorithm. We first derive the Cramer-Rao bound (CRB)
ing adaptive filters (forinstance,in [l] and [ a ] ) do not for estimating the parameters of harmonic signal embed-
accountforthespecialstructure o f the harmonic spec- ded in white noise, Then results of Monte Carlo simula-
trum, thus, their performance is not likely to be optimal tions are presented, indicating that the variances of pa-
for such signals. rameter estimates are of the same order of magnitude as
In Section I1 of this paper, we develop a new adaptive the CRB for a sufficiently large number of data, but the
algorithm,speciallydesigned to enhanceharmonicsig- algorithm is not fully efficient in general. The algorithm
nals measured with additive white noise. It can also be is alsotested against periodic signalswith an infinite num-
used as an adaptive notch filter for eliminating harmonic ber of harmonics. The results demonstrate the applicabil-
interference from a measurement broad-band process. The ity of the algorithm to the filtering of artificial periodic
parameters of the harmonic signal, such as the fundamen- waveforms, such as square waves, saw-tooth, triangular
tal frequency, and the harmonic amplitudes and phases, waves, and others. Section 1V summarizes the paper.
are assumed unknown and are estimated by the algorithm.
The proposed algorithm is of recursive prediction error
(RPE) or recursive maximum likelihood (RML) type (see 11. THE ADAPTIVECOMBFILTER
[ 3 ] and [4]) and uses several nonstandard features to im-
prove its performance. This section derives the proposed adaptive comb filter
(ACF) for harmonic signal enhancement and spectral es-
timation. The subsections below consider the following
ManuscriptreceivedFebruary 5 , 1985;reviscdMarch29,1986.The subjects: I ) the special model and parameterization of the
work of A. Nehorai was supported in part by the National Science Foun- harmonic signal with additive white noise, and a general
dation under Grant DCI-8604351.
A. Nehorai was with Systems Control Technology, Inc., Palo Alto, CA description of the algorithm; 2) the error regression and
94304. He is nowwiththeDepartment of ElectricalEngineering, Yale gradient; 3) therecursivealgorithmforestimatingthe
University, New Haven, CT 06520-2157. fundamental frequency and enhancing the harmonic sig-
B. Porat is with the Department of Electrical Engineering, Technion-
Israel Institute of Technology, Haifa 32000, Israel. nal; and 4) the recursive algorithm for estimating the am-
IEEE Log Number 8609626. plitudes and phases of the harmonic components.

0096-3518/86/1000-1124$01.00 @ 1986 IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 NEHORAI AND PORAT: ADAPTIVE COMB FILTERING 1125

x(t) inspecial
our
case,
the zeros of A ( q integral
- ' )atare
mul-
tiples of the fundamental frequency, this polynomial can
WE Wo't)
be-written as
Y(t) __j ACF 3 RLS
n
A(q-') = (1 + akq-' + q-2) (5a)
Fig. 1. Block diagram of the
algorithm.
proposed k= 1

where
A . The Model
(IIk = -2 COS ko,. (5b)
Let x ( t ) be the harmonic signal whose parameters are
summetry
the
of estimated.
to
toDue
Thus,
be ') A(q-
n
A(q-') = 1 + q q - ' + - - - + anqWn
X(t) = c ck Sin (kw,t + 4 k ) (1)
k= 1
+ .. . + + q-?
alq-2n+l
(6)
where w, is the fundamental frequency, andc k and c $ ~ are
the amplitude and phase of the kth harmonic component The whitening filter of y (t) is required to be stable, and
of x ( t ) , respectively. The number n is the assumed num- its output has to be u (t) when excited by y (t). By inspec-
ber of harmonics in x ( t ) . In cases where theactual signal tion, we find from (4)that the whitening filter can be ap-
consists of an infinite number of harmonics, we truncate proximated by
the infinite sum at n harmonics where n is chosen so that
the energyin the remainingharmonics is sufficiently
small. The remaining harmonicswill be considered as part
of the, noise, cf. (2) below. n

The white noise corrupted measurement at time t is as-

sumed to be
Y( 0 = + u (t)
n
= C ck sin (kw,t
k=l
+ &) + u ( t ) (2) This function satisfies the stability requirements when p
< 1 . It is characterized by 2n zeros on the unit circle at
{ei j k o n , 1 Ik 4 n) and 2n poles having the same phase
where u (t) is a zero-mean white noise with variance u 2 ,
angles at the zeros and their radii are all p , that is, the
We assume that the parameter vector
poles are at (pe'jkoO, 1 5 k In]. The parameter p is
e = c1 - . cn,4' . - 4~~ (3) chosen by the user; typical values are 0.95-0.995.
From (4),( 3 , and (7), we observe that the error signal
is unknown. A maximum likelihood estimation of 0 would
require a nonlinear algorithm of dimension 2n 1. To + 4= w q - 9 Y (t) (8)
simplify this situation, we shall divide our algorithm into
approximates the noise u ( t ) when p is sufficiently close to
two cascaded parts, as is illustrated in Fig. 1 . As shown
one.
in the figure, the first part of the algorithmis the recursive
Note that H(q-') can be used as a notchfilter for elim-
prediction error adaptive comb filter, whichestimates the
inating harmonic disturbance added to a desired broad-
fundamentalfrequency w, andenhancestheharmonic
band process. The reverse operationof extracting thehar-
component x ( t ) of y ( t ) . Based on these results, the am-
monic signal x ( t ) from the noisy measurement y (t) is ob-
plitudes { Ck} and phases { &} are estimated (after param-
tained by
eter transformation, see later) using a linear recursive least
squares (RLS) algorithm.
Let us discuss the whitening filter of y ( t ) . For this we
recall that for n sine waves (not necessarily harmonics)
with additive white noise,y ( t ) can be shownto satisfy the where K is the zero frequency gainof H(q-').
relationship [SI One of the main advantages of our algorithm can be
observed already at this point by noting that the whitening
4 - l ) Y ( t ) = A(q--])u(t) (4) filter H ( q - ' ) is modeled by a single parameter, namely,
where A(q-') is the 2nth-order polynomialin the unit de- the fundamental frequency of x@). This leads to a com-
lay operator q-', whose zeros are on the unit circle at the putationally efficient algorithmfor estimating w,, asis
sine wave frequencies. We should stress that since the discussed in the sequel.
nulls of A(q-') are at the sine wave frequencies, it does
not follow from (4)that y ( t ) = u (t).Also note thatA(q-') B. f i e Error Regression and Gradient
has 2n poles at the originwhich are not important for the The proposed algorithm uses the model (7), (8) and ad-
present discussion as they are cancelled out in (4).Since justs the estimateof w, so as to minimize the cost function

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 1126 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-34,NO. 5, OCTOBER 1986

N
v, = rc
=1
E2(t)

where N is the number of data samples.

To compute the error ~ ( twe
) need to find its regression
expression. Let
a = [al - a,] T (1 1) and let H be the Hankel matrix

[ j.
*

superscript
where T denotes
transpose.
From (7) and (8) 1
one obtains the difference equation
E(t) = y (t) + y (t - 2n) - p2"c(t - 212) - (p'(t>a (12) H = (20)
... .... .
~..:-::::~1

where .
1 p]' * - .P m - 1
CP(~)[ ' ~ l ( t )+ * * ~n(t)l'
(13a) Assumethat { X i , 1 5 i 5 m} arealldistinct.Then
and

i = n.
The gradient of ~ ( twith
) respect to u, can be found as
follows. Let

We now apply the results of lemma 1 to our case. Due

The derivatives -aE(t)/aai can be shown to be given by to the symmetry of the polynomial A(q-'), we only need
(see also [ 5 ] ) to consider the derivativesof the n first coefficients. Using
(20) and (21), we get
_- sew -- Pi ( 0
___
aai
A
A(P~-*)
= 'PFi (t),
i.e., -&(t)/aai are filtered versions of {pi( t ) } .
To compute the derivatives aai/au, in (14), we shall
use the relationship
(15)
I;;; 1
- . I . -
ah 2n
- -
= -[OiHIV (22)

where [0 i 7?] is an n X 2n matrix whose n leftmost col-

umns are zero and

where { X i , 1 5 i 5 2n) are the zeros of the polynomial

z Z n ~ ( z - ' )namely,
,
h2k- = eikuo = e-jkmo , 1 5 k 5 n. (17) The matrix 7 in (22) is similar to (19), but here m = 2n
and X i are given in (17). Now let
The expression in the right-hand sideof (16) can be com-
puted using the following lemma proven in Appendix A.
Lemma I : Let P(z) be a polynomial of order m in z ,
viz.,
P(2) = z m + plzm-' + - - + pm
*

= (Z - X,)(Z - X,) * * (Z - Xm). (18)

Let V be the Vandemonde matrix Then

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 NEHORAI AND PORAT: ADAPTIVE COMB FILTERING 1127

C. The Adaptive Algorithm for 2(t) and Go@)

With the results above, we can now apply the general
n
RML or RPE approach to our case. The resulting recur-
sions of the algorithm for enhancing the harmonic signal
and estimating thefundamentalfrequency are summa-
n rized below. Note that in each recursion, the latest avail-
=2 k sin kio, able estimates of a, and A(q-') replace their value in the
k= 1 expressions of the previous subsection. The explanations
of the algorithm and its special features are discussed be-
(n ,+ 1) cos i (2nl + l)iw,
2 1 low.
-
-
sin [i(n + l)wO] - 1) The RPE Adaptive Comb Filter:
2 sin2 (?) sin (2) DesignVariables: n , X(l), X,, y(l), r ( l ) , p ( l ) , pot
P(W).
Initialization: G,(O), pi(0) = pFi(0)= 0 , i = 1 , * - ,
n , y ( - i ) = 0 , i = 1, * -
, 2n.
Nominal Values:
Combining (22)-(24), we now have
n = number of band-pass filters

or
I= I 1 X(l) = 0.45 - 0.65, X, = 0.98, y(1) = 1 , r(1) z
Ey2(t)/100
p(1) = 0.8, po = 0.98, p ( ~ =) 0.95 - 0.995 (see also
comments below).

Main Loop:

where a, = 1. Thus, $(t) can be computed by first eval- G0(t) = Go(t -

uating the sequence { x i } in (25), then computing { aai/
aoo)by (27), and finally,

where

v = [V, * vny

n- 1

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 1128 IEEETRANSACTIONSONACOUSTICS,SPEECH, AND SIGNALPROCESSING, VOL. ASSP-34, NO. 5, OCTOBER 1986

sin [i(n + 1)&,(t)]

fi(t) =
(T)
2 sin2 i&,(t) sin (y)
i- I
Vi(t + 1) = Ic=O
dl@) ai-&) nominal values of X( 1) and X, were foundto yield the best
results in our simulation experiments. For SNR = 0 dB,
p(t + 1) =' [p& + 1) * p,(t + l)lT it was useful to take X(1) = 0.45 and for SNR 2 4 dB
X(1) = 0.65.
VFQ + 1) = [PFl(f + 1) * PFnO + 1>IT With the above recursions and for stationary signals,
V(t + 1) = [V,(t + 1) V,(t + l)lT
* * *
the resulting step size y(t)/r(t) asymptotically approxi-
mates the inverse second derivative of the cost function
$(t + 1) = p g t + 1) V(t + 1) evaluated at &,(t - 1).Forthisreason, the algorithm
can be viewed as a recursive Gauss-Newton type.
X ( t + 1) = X,X(t) + (1 - X,) Toallowtrackingoftime-varyingparameters, it is
common to choose y ( t ) = yo, where yo is a small positive
y(t + 1) = y(t)/[y(t) + +
1)1
number. This corresponds to an exponentially weighted
r(t + 1) = r ( t ) + y(t + 1)[g2(t + 1) - r(t)] (30s) cost function characterized by time constant lly, (see [4]).
4) Stability Monitoring: The analysis of [3] and [4]
P(t + 1) = POP@) + (1 - P o ) P(W). (300
shows that for RPE algorithms, one must check at each
We now discuss some of the special features which were time update whether the current parameter estimates cor-
used to improve the performance of the algorithm. Some respond to a stable filter. Otherwise, the parameter esti-
of these features were used also in the algorithm of [SI, mates must be projected back into the stable model set.
and therefore, they are discussed only briefly. However, in ourcase, the estimated whitening filter
2) Time-Varying Pole Moduli: The recursion (30q) for H(q-') is stable by construction, and therefore, stability
the pole moduli yields a time-varying p(t) instead of a monitoring is not needed.
constant value of the pole moduli. Note that p(t) grows 5) Estimating the Number of Harmonics: For practical
exponentially from p(1) to p ( m ) with a time constant To cases where the number of harmonics n is unknown, we
= 1/(1 - p,). Thus, at the beginning of the data process- note that n can be estimated as follows. Apply the algo-
ing, the notches of the estimated whitening filter are wide rithm given above and the cascaded algorithm of Section
and they become narrower as time goes on. This improves 11-D-1 with an overestimated number of harmonics based
the sensitivity of the algorithm to the presence of har- on prior information on n. Then n can be estimated on
monics and increases its convergence rate. To see this, line using the number of harmonics yielding significant
note that when the initial conditions of the algorithm are amplitudes, i.e., larger than some predetermined thresh-
poor, and if the notches are too narrow, the input har- old. The use of an underestimated n usually adds a dis-
monics are likely to appear at the flat part ofthe estimated tortion to the filtered signal. The faster the spectrum en-
whitening filter transfer function. In such cases, the gra- velope of x ( t ) decreases, the smaller the distortion will
dient of the algorithm vanishes and the algorithm may not be. The use of an overestimated n usually adds somenoise
converge to the desired transfer function. to the output.
The nominal initial value p(1) = 0.8 was chosen as the 6) A ModifiedVersion: The algorithmabove directly
smallest value of pole modulus which still has significant estimates the fundamental frequency ofx (t).Another ver-
effect on the notch bandwidths. The value of p ( m ) is cho- sion of the algorithm first estimates the coefficient a1[see
sen asa tradeoff betweenaccurateasymptotic perfor- (5b)l and then the fundamental frequency can be evalu-
mance ( p ( m ) as close as possible to one) and robustness ated using the relationship a, = cos-' ( -cu1/2). This
as well as tracking capability of time-varying frequencies modified version was found to be more robust than the
(smaller values of p(00)). original algorithm, especially when most of the energy is
3) Updating the Gain and Tracking Time-Varying Pa- concentrated in the first harmonic of x ( t ) . (See the ex-
rameters: The recursion (30r) for the gain y(t) and the amples bf the next section.) This result can be explained
related recursions (30q) and (30s) are adapted from [4] heuristically by the fact that in such cases, the error gra-
for stationary signals. Equation (30q) of updating X ( t ) in- dient tends to be more linear with respect to al than with
fluences the cost function at the transient phase and im- respkct to w, (consider, for instance, the special case of a
proves the convergence rate of the algorithm. The above single sine wave). The modified algorithmis obtained by

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 NEHORAI AND PORAT: ADAPTIVE COMB FILTERING 1129

multiplying +(t) by Initialization: C(0) = 0 , P(0) = (l0O/SNR)Z2,.

Main Loop:
dG,(t)
- = 1
d&,(t) 2 sin G,(t) r(t) = [sin Got * * - sin n&t, cos Got - * * cos nij,t]
7) ConvergenceAnalysis: The convergence of RPE- (354
typealgorithmshasbeenanalyzedin [3] and [4]. The e(t) = a ( t ) - r*(t) C(t - 1 ) (35b)
proof that the assumptions of this analysis hold for our
case is 'technical, and we omit its detail. Among the re-
sults of the canvergence analysis, webriefly mention that
in the model complete case (i.e., when the model is con-
sistent with the data) and white Gaussian noise, the pa- (35c)
rameter estimates are asymptotically unbiased, normally
distributed, and achieve the Cramer-Rao bound (i.e., sta-
C(t) = $(t - 1) P(t) c(t)
e(t). + (354
tistically efficient). In the above algorithm, the variable X(t) is updated sim-
Finally, we note that the initialization r ( 1 ) = Ey2(t)/ ilarly to the algorithm in (30). The amplitudes andphages
100 is only an approximation implicitly based on recom- can'be evaluated by
mendations in [4, p. 2991. As noted in [4], in practice, it
is common to taker(1) = 1/C where C is a large number.
As C becomes' larger, the value of c3,(0) becomes only
marginalfy important. (35f)
D. Estimating the Harmonic Amplitudes and Phases
In the next section, we will show that the expected rel-
In some applications, it i s desired to estimate the am- ative error of 2, is significantly smaller than those of the
plitudes and'phases of the harmonic components in a$$- amplitudes and phases when obtained by a maximum like-
tion to the fundamental frequency. The algqrithm below lihood estimator. This implies that by replacing w, by its
is suggested for this purpose. estimate &, the accuracy of the amplitude and phase es-
To derive the algorithm, we shall first assume that w, timates is not expected to deteriorate significantly.
is known, and apply proper transformation to, the desired To improve the estimation accuracy of this procedure,
parameters. We have /I
it is also useful to replace theraw data y ( t ) by the filtered
n data a ( t ) [see (35b)l. Here it should noted
be that the noise
y (t) = C ck sin (kw,t + &) + u (t)
k= 1
in a ( t ) is colored, which may cause some bias in the es-
timates of the amplitudes and phases. However, since this
n noise is much smaller than in the raw data y ( t ) ,the overall
= x
k= 1
( g k COS kW,t + hk Sin k0,t) + U(t) (32) accuracy of this procedure is still better than if y ( t ) were
used.
where The initialization P(0) = (100/SNR) Z2n, where ZZn is
gk = ck sin (334 the 2n X 2n identitymatrix,isonlyanapproximation
based on recommendations in [4, p. 2991. In practice it is
hk = ck COS &. (33b) common to take P(0) = C ZZn where C is a large number
This can be written as (see r41).
2) Amount of Computations: The overall algorithm for
y (t) = [sin w,t * - sin nwot, adaptive estimation of the harmonic signal parameters re-
quires a number of operations proportional to n2 per time
cos w,t --- cos nw,t] q + u(t) (34a) sample.Most of thecomputationloadistaken by the
where computation of A ( q - ' , t )in (~OC), computation of the gra-
+
dient vector V ( t 1) in (301), and updating the RLS gain
= [SI * * gn, hl * (34b) vector P(t) c(t) in (35c, d). To save computations, the
Equation (34a) implies that if w, were known, an RLS algorithm (35)of amplitude and phase estimation does not
algorithm could be used to estimate the sequences { g k } havetobeimplementedwhenonly c3
, and R(t) are
and (hk}. Then the desired amplitudes { Ck} and phases needed.
{ &} could be foundby a simple transformation from rect-
angular to polar coordinates. Since the fundamental fre- 111. PERFORMANCEEVALUATION
quency w, is unknown, we can replace it by its estimate In this section we evaluate the performance of the pro-
obtained from the algorithm of Section II-C. This proce- posed algorithm by several Monte Carlo simulation runs,
dure is summarized below. and make comparisons to the Cramer-Rao bound (CRB).
I ) The RLS Algorithm for the Harmonic Amplitude and Then we illustrate .the convergence of the algorithm for
Phase Estimation: signals with a large number of harmonics. Finally, we

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 1130 IEEE TRANSACTIONS
ACOUSTICS,
ON SPEECH, SlGNAL
ANDPROCESSING, VOL. ASSP-34, NO. 5 , OCTOBER 1986

quency f , = 0.08, and u(t) is a zero mean unit variance

illustrate the algorithm’s performance for artificial signals
with an infinite number of harmonics. whiteGaussiannoise. The harmonicamplitudes {ck)
were chosen to yield both the desired signal-to-noise ratio
A . The Asymptotic CRB for Harmonic Signals given by SNR(dB) = 10 log ( X i Z l C ~ I 2 o 2and
) the de-
Asymptotic formulas for theCramer-Rao bounds of the sired spectrum. To make the examples as close as possi-
parameters in the model (1) are derived in Appendix B. bletopractical
situations,
thespectrumenvelope was
These formulas are valid when N, thenumber of data chosen such that its decrease in gain was -6 dB per oc-
samples, satisfies N >> l/uO.The main results of Ap- tave, or harmonic powers proportional to 114’. This type
pendix B are summarized as follows:
f r n
phases are unknown and
amplitudes are known or not
phases are known and
amplitudes are known or not
20
Var (&) 2 - in all cases (36c)
N
fundamental frequency is
unknown and amplitudes are
known or not
Var (&) I
fundamental frequency is
known and amplitudes are
known or not
where
n of envelopeis oftenused to model the voiced speech spec-
y = C k2Ci. (360 trum envelope generated by the vocal tract.
k= 1 The algorithm was applied, using a DEC VAX com-
Remarks: Notethequantity y / 2 a 2 = E t = , k2Ci/202 puter with double precision arithmetic, to estimate the pa-
appearing in the CRB of the fundamental frequency, is rameters of the signal (37) with design variables p(]) =
0.8, po = 0.98, p ( m ) = 0.96, and X, = 0.98. For SNR
not the SNR, but an “enhanced SNR,” in which the var-
= 0 dB, we used X(l) = 0.45 and for largerSNR’s values
ious harmonics are weighted both by their energies and
by theirrelativefrequencies. In particular,higherfre- X(1) = 0.65. As was explained earlier, since in this case
quency harmonics have larger enhancement than low-fre- the number of harmonics is relatively small ( n = 5 ) , and
quency harmonics. most of the energy of x(t) is contained in its first har-
It is of interest to note that the results (36a) and (36b) monic, it was was useful to update the parameter& 1 rather
can also be written as
f
I
12 phases are unknown and

Var (Go)I SNR I 3

w:ffN3
amplitudes are known or not

phases are known and

SNR w&N3 amplitudes are known or not
whereSNR = X,”= C i / ( 2 0 2 )andthe“effectiveband-
than &. This was found to improve the robustness of the
width” = E;=, k2C$Z,“=, C:. The factor 1/(SNR
algorithm.
w$f) appears also in the CRB on time delay estimation
The algorithm was tested under different data lengths
(see e.g., [IO]).
and signal-to-noise ratios. Each of the statistics below was
B. Monte Carlo Simulations computedfrom 100 independentMonteCarloexperi-
ments. The parameter estimates G l ( t ) were set to values
The following examples illustrate the bias and variance such thatL(0) was equal to 0.05 at the beginning of each
of the estimated parameters, compared to the CRB. The experiment.
input data were Table I summarizes the resulting sample statistics of the
5 normalized fundamental frequency estimates for the sig-
y(t) =
k= 1
c
C, sin 2~0.08kt u(t), + (37) nals in (37). In the simulations, we encountered some re-
alizations which gave outlier performance, i.e., the esti-
i.e., the number of harmonics n = 5 , fundamentalfre- mates were not close to the typical behavior. Out of 100

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 NEHORAI AND PORAT: ADAPTIVE COMB FILTERING 1131

TABLE I
RESULTSFOR THE FUNDAMENTAL FREQUENCY
STATISTICAL ESTIMATE
OF
(37)
THE SIGNAL

fo

N SNR Standard
(Samples)
(dB) Bias X10-6 Deviation XlO-’ CRB XlO-’

200 0 -98.7(7) 42.3 10.5

8 -26.0(3) 14.3 4.20
16 1.67 1.20 5.05
500 0 -22.6(6) 14.4 2.67
0.40(3) 8 1.06 2.03
16 0.42
-0.99 0.78

TABLE 11
STATISTICAL RESULTSFOR THE AMPLITUDE
ESTIMATES
OF THE SIGNAL (37)

C1 c
2 c
3 c
4 c5
Standard N SNR
Standard Standard
(Samples)
(dB)
True Bias
Deviation True
Bias
Deviation
True
Bias
Deviation
True
Bias
Deviation
True
Bias Deviation CRB

200 0 1.17 -0.08 0.15 0.58 -0.13 0.13 ’ 0.39 0.13 0.10 0.29 -0.14 0.09 0.23 -0.08 0.08 0.1
-0.13
2.948 0.32 1.47 -0.12 0.24 -0.15
0.98 0.19 0.73 -0.13 0.17 0.59 -0.15 0.15 0.1
16 7.38, -0.06 0.22 3.69 -0.07 2.46
0.18 -0.05 0.18 1.84 -0.07 0.21 1.48 -0.09 0.19 0.1
500 0 1.17 -0.04 0.10 0.58 -0.05 0.09 -0.04
0.39 0.09 0.29 -0.06 0.08 0.23 -0.04 0.07 0.063
2.948 -0.02 0.73
-0.03
0.07
0.98
-0.02
0.08
1.47
0.08 -0.04 0.59
0.08 -0.04 0.08 0.063
16 7.38 -0.02 3.69
0.12 -0.02 0.11 2.46 -0.02 0.11 -0.03
1.84 0.11 1.48 -0.03 0.10 0.063

TABLE 111
RESULTSFOR THE PHASEESTIMATES
STATISTICAL OF THE SIGNAL
(37)

61 $2 63 64 45
Standard Standard Standard Standard Standard
N SNR Bias
Deviation CRB Bias
Deviation CRB Bias Deviation
CRB
Bias
Deviation
CRB
Bias Deviation CRB
(Samples)
(dB) x ~ O - ~x 1 0 - l x10-l X10-2
x10-l
x10-l x ~ O - ~x1O-I
x10-’ x ~ O - ~~ 1 0 - ’ x 1 0 - l x ~ O - ~x10-’ x10-‘

200 01.08
4.48
6.79 9.84 7.64 2.16 -0.21 3.25 9.04 4.80 9.21 4.33 8.26 8.66 5.42
8 -0.02 1.25 0.43 0.56 2.88 0.86 0.88 ,1.29
4.49 7.6121.3
1.72
5.42 6.21 2.15
16 -0.81 0.56 0.17 0.11
-1.66 0.34 -3.29 0.17 0.51 -3.66 0.21 2.42
1.94
0.69 0.86
500 0 0.62 2.42 0.68 2.99 4.29 1.37
2.056.332.22 -5.162.56
2.74
7.27 8.70 3.42
8 0.09 0.51 0.27 0.56 0.55
1.05 0.06 0.821.45 -0.40 2.01 1.09 -1.23 2.71 1.36
16 0.34 0.25 0.11 0.65 0.47 0.22 -0.17 0.65 0.33 0.17 0.78 0.43 0.32 0.97 0.54

experiments in each frequency estimate, the number of Using the estimates of the fundamental frequency, the
outliersisindicatedinparenthesesinthetable.These second part of the algorithm was used to estimate the har-
cases were eliminated from the statistical computations. monic amplitudes and phases for the same signals above.
This is justified by the fact that the Cramer-Rao bound is Table I1 presentstheresultingsamplestatisticsforthe
derived under small error assumption. The decision on amplitude estimates. The results of the table show that the
outlier was based on an arbitrary threshold of 0.0015 in amplitude estimates have a negative bias. This bias be-
theerror of thefundamentalfrequencyestimates.The comes larger for the higher frequency harmonics whose
number of cases in which this happened decreased with power in these examples is relatively small. The negative
the data length and signal-to-noise ratio. bias is due to the error in the fundamental frequency es-
Table I presents also the CRB of the fundamental fre- timates,whichtendstoattenuatethesinewaveampli-
quency for the above examples. The ratio of the actual tudes. However, the bias decreases with the data length
standard deviation to the CRB appears to decrease as N and signal-to-noise ratio.
increases, at least for high SNR’s. However, this ratio is Table I11 presents the sample statistics of the phase es-
not close to one, so the estimates are not fully efficient. timates for the above signals, and the CRB computed by
We applied significance tests to the bias (using the “t” the square root of (36d). Similarly to the previous esti-
statistic) and found that w, is unbiased with confidence mates,weobservethatthestandarddeviations of the
level 0.95, in all cases except for N = 200 and SNR = 0 phases decrease with the data length and the signal-to-
dB. The CRB was computed using (36a). noise ratio, but they do not approach the CRB in general.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 1132 IEEE TRANSACTIONSONACOUSTICS, SPEECH, ANDSIGNALPROCESSING, VOL. ASSP-34,NO. 5, OCTOBER 1986

.0851
.1
L -2

NormalizedFrequency
.3
4

I
.5 .o .1 .2 .3
Normalized Frequency
.4 .5

,060 1 -60. -

.055
-uO50.
t
0. 500.
'
1000. 15O
:. 2000.
-70.

-80, I
.o .1 .2
.
.3
I

.4
1
-
.5
Tlme (samples) i,(tj NormalizedFrequency

"I
v.
.o .I .2 .3 .4 .5
NormalizedFrequency
Fig. 2. Estimation results for harmonic signal with additive white noise.
n = 5 , SNR = 0 dB. (a) Power spectral density of the noise-free har-
monic signal. (bj Power spectral density of the measurement. (cj The
fundamental frequency estimate. (dj Transfer function of the convergent
comb filter at f = 2000. (e) Power spectral density of the algorithm's
output.

Application of significance tests confirmed that the phase Fig. 2(a)-(e) presents the results for the signal with five
estimates were unbiased with confidence level 0.95 for all harmonics and fundamental frequency f, = and SNR 6
cases, except for N = 200 and SNR = 8 dB, where 3, = 0 dB. The harmonic amplitudes decrease by 6 dB/oc-
was biased. tave. Fig. 2(a) and (b) depicts the power spectral density
of the noise-free signal x ( t ) and the noisy measurement
C. ConvergenceExamples y ( t ) , respectively. The algorithm was applied with initial
The following examples demonstrate the convergence conditionfo(0) = 0.05 and design variables as described
of the algorithm by two simulation runs for signals with above in the Monte Carlo simulations, except that p(00)
a different number of harmonics. was set to 0.99. This larger value was chosen to achieve
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 NEHORAI AND PORAT: ADAPTIVE COMB FILTERING 1133

higher reduction of noise. Fig. 2(c) shows the fundamen- with five passbands, design variables p ( 1 ) = 0.8, po =
tal frequency estimate of the algorithm versus time. We 0.98, p ( a ) = 0.995, X ( l ) = 0.45, X, = 0.98, and &,(O)
note that for larger signal-to-noise ratios, the convergence = -27r cos 0.05. The above relatively large value of p ( - )
was usually faster and smoother. Fig. 2(d) illustrates the was chosen to achieve high reduction of noise. Fig. 4(a)-
magnitude of the corresponding comb filter transfer func- (c), respectively, shows the original triangular wave sig-
tion after convergence given by nal,the noisy measurement,andthealgorithm filtered
output after convergence, from t = 1900 to t = 2000. At
G(ej., t) = 1 - t 2 ( e i w , t ) / ~ ( t ) (38) t = 2000,thefundamentalfrequency bias was1.75 X
at t = 2000, and where t ) is the estimated whit-
ening filter of themeasurement,and K(t) its zerofre- We have also .applied the algorithm to filter a square
quency gain. At t = 2000, the bias in the fundamental wave signal plus white noise. The resulting filtered signal
frequency was 8.00 X Fig.2(e) depicts the power was slightly worse than in the triangular wave. This is
spectral density of the filtered output R(t) after conver- explained by the fact that thespectrum envelope of a
gence. This figure was drawn by applying FFT to the last square wave decreases by 6 dB/octave, while that of a
1024 data points of R(t) from t = 977 to t = 2000. triangular wave decreases by 12dB/octave.Thus,the
The next example illustrates the ability of the algorithm truncation of higher harmonics causes larger distortion in
to filter noisy signals with a relatively large number of the square wave than in the triangular wave. We omit the
harmonics. For this purpose, the harmonicsignal was with results due to space limitation.
20 harmonics, fundamental frequency & and SNR = 0 Signals such as the above square wave and triangular
dB. Similarly to the above example, the harmonic ampli- wave which have so-called symmetric half-wave (i.e., x(t)
tudes were adjusled to yield a spectrum envelope of -6 = -x(t - T/2) where T is the period) are characterized
dB/octave. The algorithm was run with the same design by zero odd indexed harmonics. Whenit is a prioriknown
variables as before and initial condition A,(O) = 0.015. that the harmonic component of the input is of this type,
Since, in this example, the number of harmonics was rel- it is desirable to apply aspecial comb filter with only even
atively large, the algorithm that updates ;,(t) was used. index passbands. The algorithm of Section I1 can be easily
The results are presented in Fig. 3. The bias in the fun- modified for such applications.
damentalfrequencyestimatewas -2.5 1 X at t =
2000. IV. CONCLUSION
The result of the last example is interesting as it dem- In this paper, we have presented a new adaptive algo-
onstrates that the algorithm is applicable for signals with rithm for harmonic signal enhancementandparametric
a large number of harmonics and large orders of A ( q - ’ ) spectral estimation.Itscomputational efficiency advan-
and A ( p q - ’ ) . This is remarkable in comparison to the be- tage stems from the separation of the solution into two
havior of most existing system identification algorithms cascaded parts, as is illustrated in Fig. 1. The first part
whose usual largest possible order is between 5-10 for enhances the harmonicsignalsand estimates its funda-
these types of SNR conditions and unknown input. The mentalfrequency.Thesecond part estimates the har-
reason for this good behavior of the algorithm is our use monic amplitudes and phases. In this way, the nonlinear
of a single parameter model in the algorithm for estimat- part of the algorithm involves only one parameter-the
ing the fundamental frequency. signal fundamental frequency. This enables the algorithm
The power spectral densitiesof Figs. 2(b) and 3(b) may to workwith significantly larger order ARMA polyno-
lead one to think that FFT can be used to estimatef, eas- mials than in general system identification schemes. An-
ily, as they clearly show the harmonic distribution of the other improvement compared to general RPE algorithms
signal. However, this is only because&, was chosen to be is the stability of the filter. Thus, stability monitoring,
a submultiple of the number of FFT points (X, = &) in which is usually necessary for general RPE algorithms, is
this simulation. In practice,usuallyf, does not satisfy this not needed in our scheme. Simulationresults indicate that,
condition and then the “leakage” between the FFT bins for sufficiently large data sets, the variances of the esti-
makes it prohibitive toestimate L), especially in low mated parameters aregenerally of the same orderof mag-
SNR’s. Moreover, the FFTis a batch method, which can- nitude as the Cramer-Rao bound, but the algorithm is not
not be applied adaptively in contrast to the proposed al- fully statistically efficient in general.
gorithm. Other variants of the algorithm presented in this paper
can be derived. For example, instead of using constant
D. Enhancement of Periodic Signals bandwidth passbands, one can use so-called constant-Q
As mentioned in Section 11, the proposed adaptive comb passbands, i.e., let the bandwidth of each passband be
filter is also useful for enhancement of periodic signals proportional to its central frequency. Another possibility
with an infinite number of harmonics by truncating the is to include only certain harmonics, say, the odd ones,
low-energy high-frequency harmonics. This is illustrated based on a priori information about the signal.
by the following example. The ubiquity of harmonic signals both in nature and in
This example is of a triangular wave whose period is artificial environments suggests a wide variety of poten-
12 samples and additive white noise with SNR = 0 dB. tial applications of the proposed algorithm. These include
The modifiedalgorithmthatupdates & , ( t )was applied enhancement of noisy biological signals, such as voiced
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 1134 IEEETRANSACTIONSONACOUSTICS,SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-34, NO. 5 , OCTOBER 1986

-.
n -
.o .1 .2 .3 .4 .5 .@ .1 .2 .3 .4 .5
Normalized Frequency Normalized Frequency
(a) (b)

500. 1000. 1520. 2000.

::I
-50.
.o
,

.1
,

.2
,

.3
,

.4
1
.5
Time (samples) f , ( t ) Frequency
Normalized

V.

.o .1 .2 .3 .4 .5
NormalizedFrequency
(e)
Fig. 3. Estimation results for harmonic signal with additive white noise.
n = 20: SNR = 0 dB.

speech and heart wave forms. Another important appli- noise, e.g., to estimate a radar’s modulated pulse repeti-
cation is tracking of propeller-engine vehicles,for in- tive frequency by a passive radar. Further research has to
stance, submarines and helicopters. The alternative use, be carried out to investigate the performance in such ap-
as an adaptive harmonic notch filter, is useful for har- plications.
monic noise cancellation. This may be used, for example, Our analysis has provided new results for the CRB of
to reduce the noise in helicopter cockpit communication. the parameter estimates of harmonic signals in additive
Other applications are when there is a need to estimate white noise. One interesting result is that the power of
artificial harmonic signal parameters in thepresence of each harmonic appearing in the CRB of the fundamental

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 NEHORAI AND PORAT: ADAPTIVE COMB FILTERING 1135

tively low. Clearly this is due to the fact that the high-
frequency harmonics yield more information on the fun-
damental frequency. The CRB results of this paper are
general and potenfially useful in evaluating the perfor-
mance of other harmonic spectral estimation algorithms.
An interesting topic of furtherresearch is theevaluation
of the optimal valuesof p,, p ( l ) , p ( m ) , X(l), and X,. This
subject is currently being investigated.

APPENDIX A
PROOFOF LEMMA1
Introduce the polynomials
-3. k
1900. 1950. 2000.
Time (samples) Pk(z) = pizk-i OIkSWZ (All
i=O
(a)
where po = 1. Note that P,(z) = P(z) and that
3.I

2.

Hence,
1.
m- I
0.

- 1.

-2.

m-1
-3.
= zPm-l(z) - X? - cpixy
-4.
1980. 1960.
'
1900. 1940. 1920.
I
2000.
i=l

Time (samples) = ZP, - l(2) + p m = P(z). (A3)

3.
Dividing both sides by (z - X,) we have
I

2. i

(A51
-3.1
2000. 1900. 1950. This can also be written
Time (samples)
(c)
Fig. 4. Enhancement results for triangular wave with additive white noise.
n = 5, SNR = 0 dB. (a) The noise-free signal. (b) The noisy measure-
ment. (c) The filtered output.

where I/ and H were defined in (19) and (20), respec-

frequencyestimate is multiplied by its squared relative tively. Next, note from (18) that
frequency. Thus,higherfrequencyharmonicshaveen-
hanced SNR in the CRB. This mathematical result gives
a new explanation to a known physical phenomenon by
which high-frequency, harmonics in speech are important
to its intelligibility although their energy is usually rela- Hence, substituting z = X1, X2, * , X, we get +

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 1136 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND
SIGNAL PROCESSING, VOL. ASSP-34, NO, 5, OCTOBER 1986

The CRB for 8 is given by the inverse of J . It should be

noted that an expression similar to (B3) was used in [9]
for multiple sinewaveswithindependent frequencies
using complex signal formulation.
Let O(x) denote a quantity that is asymptotically linear
in x, i.e., such that

L o 1
( 4 Also let O( 1) denote a bounded quantity. Then we have
for all kw, such that ko,/27r is not an integer and N >>
where the nonzero entry at the right-hand side appears at
1 /w,
its kth row. Hence,
N
E
t= 1
ti sin kw,t = O(Ni); i = 0, 1 , 2, . * 0-35)
..
VT N
2 t i cos kw,t
t= I
= O(Ni) i = 0, 1, 2 * -. (B6)

Assume further that w, and n in (1) are such that kw,/2n

is noninteger for all 1 Ik 5 2n. (Note that this and the
= -diag II (A, - A,>; 1 Ik Im above assumption on kw, are always satisfied for a finite
( i i+=k 1 number of harmonics which are sampled at a rate larger
than the Nyquist rate.) Then it can be shown that
(A9)
Premultiplying by ( V T ) - ’ andmakinguse of (A6), we
obtain
the desired expression (2 1).

APPENDIX B
THE ASYMPTOTIC CRAMER-RAO BOUND FOR THE
PARAMETERS OF HARMONIC SIGNALS I N WHITE NOISE
The Cramer-Rao is a general lower bound on the error
covariance of unbiased estimators. This appendix derives
the asymptotic Cramer-Rao bound for the parameter es-
timates of harmonic signals in additive white noise.
The assumed model is as in (1) parameterized by 8 de-
fined in (2). To derive the CRB for this model, we will
first derivethecorrespondingbound for the alternative
model (32) parameterized by
8 = [w, g1 * g,, h , . - h,lT
* (B1)
where { g k } and {hk} were defined in (33a) and (33b), re-
spectively. The desired CRB for (1) and 0 will then be
evaluated using the relationship between the two models.
Under the above model, the jointprobability density of
measurements { y(1)
the y ( N ) } is

From (B2) the general expression for the ( k , Z)th element Hence, the Fisher information matrix for the transformed
of the Fisher information matrix is parameter vector 8 of (Bl) can be written as

1
J=-(J
22 1 + J2)

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 NEHORAIANDPORAT:ADAPTIVE COMB FILTERING 1137

where In our case, from (33) these derivatives are

where 6kl istheKroneckerdeltafunction.Substituting

(B23) and (B21) into (B22), we obtain, after straightfor-
ward but lengthy computations, the result
(B16)

where
jl i
(B17a) CRB(@ = 2u2 0
(B17b)

yN2 IN
(B17c)

-
The matrix J;' can be easily computed to yield
1

Now from (B15) we get

where
J-' = 2u2[J;' - J&.p + J7'J2J;'J*J;1 - * ]
u = [l, 2 * * nIT (B25a)
(B 19)
providedtheinfinitesumcoverages.From (B16) and D = diag (l/C:}. (B25b)
(B18), we have for sufficiently large N
The Cramer-Rao bounds on the fundamental frequency
w,, the amplitudes (ck}, and phases (4k})can now be
found from the diagonal entries in (B24) as summarized
in (36).
Hence, for large N , J;'J2J;' is negligible compared to Thestructure of CRB(0) impliesthattheasymptotic
CRB of w; and (4R) are independent of whether the am-
5; Similarly, we can prove for the rest of the terms in plitudesareknown or unknown,andviceversa.From
(B19); hence, for large enough N , the CRB for 8 is
(B24) we can also find that the asymptotic CRB of o,
CRB@ = 2 ~ ~ 5 ; ' (B21) given the phases is
where J;' was given in (B18).
To evaluate now the CRB for the original parameter
vector 8 in (3), we will use the following general rela-
tionship:

CRB(8) = E] CRB($) [$] T

(B22)
Using a similar notation we also find that

where the (k, Z)th entry of [a19la8] is aOk/a8,. CRB (4(w0)= 202 D. (B27)
N

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.
 1138 IEEE TRANSACTIONS
ACOUSTICS,
ON SPEECH,
AND
SIGNAL
PROCESSING,
VOL. ASP-34, NO. 5 , OCTOBER 1936

REFERENCES
r11 B . Widrow and J. R. Glover, Jr. et al., “Adaptive noise cancelling:
Principlesandapplications,” Proc. IEEE, vol. 63, pp. 1692-1716,
Dec.1975.
B. Freidlander,“Arecursivemaximumlikelihoodalgorithm for
ARMAlineenhancement,” IEEE Trans.Acoust.,Speech, Signal
Processing, vol. ASSP-30, pp. 651-657, Aug. 1982.
131 L. Ljung, “Analysis of a general recursive prediction error identifi-
cation algorithm,” Autornatica, vol. 17, no. 1, pp. 89-100, Jan. 1981.
141 L. Ljung and T. Soderstrom, T h e o v and Practice of Recursive Iden-
tljication. Cambridge,MA:M.I.T.Press,1983.
151 A. Nehorai, “A minimal parameter adaptive notch filter with con-
strained poles and zeros,” IEEE Trans. Acoust., Speech, Signal Pro-
cessing, vol. ASSP-33, pp. 983-996, Aug. 1985.
J . A. Moorer, “The optimum comb method of pitch period analysis
of continuous digitized speech,” IEEE Trans. Acoust., Speech, Sig-
nal Processing, vol. ASSP-22, pp. 330-338, Oct. 1974.
J. S . Lim, A . V. Oppenheim, and L. D. Braida, “Evaluation of an
adaptive comb filtering method of enhancing speech degraded by white
noise addition,” IEEETrans.Acoust.,Speech,SignalProcessing,
V O ~ASSP-26,
. pp. 354-358, Aug. 1978.
181 S. M. Kay and S. L. Marple, Jr., “Spectrum analysis-A modem
perspective,” Proc. IEEE, vol. 69, pp. 1380-1419, Nov. 1981.
191 D. C. Rife and R. R. Boorstyn, “Multiple tone parameter estimation
from discrete-time observations,” Bell Syst. Tech. J . , vol. 55, no. 9,
pp.1389-1410,NOV.1976.
C. W. Helstrom, Sfatistical Theory of Signal Detection. Elmsford, Boaz Porat (S’79-M’82), for a photograph and biography, see p. 130 of
NY: Pergamon, 1968, pp. 274-319. the February 1986 issue of this TRANSACTIONS.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on August 23,2023 at 03:55:23 UTC from IEEE Xplore. Restrictions apply.