IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, SIGNAL AND

PROCESSING, VOL. ASSP-34, NO.

2, APRIL 1986

33 1

The Statistical Performance of the MUSIC and the Minimum-Norm Algorithms in Resolving Plane Waves in Noise
Abstract-This paper presents an asymptotic statistical analysis O f the null-spectra of two eigen-assisted methods, MUSIC [I] and Minimum-Norm [2], for resolving independent closely spaced plane waves in noise. Particular attention is paid to the average deviation of the null-spectra from zero at the true angles of arrival for the plane waves. These deviations are expressed as functions of signal-to-noise ratios, number of array elements, angular separation of emitters, and the number of snapshots. In the case of MUSIC, an approximate expression is derived for the resolution threshold of two plane waves with equal power in noise. This result is validated by Monte Carlo sirnulations.

well as other spectrum estimation approaches to this parameter estimation problem, the directions of arrival (DOA) are given by the positions of the spectral peaks or by locations Of the Of the inverse Of the spectrum (called the null-spectrum here). Thus, sources are resolved if the estimated null-spectrum contains minima (nulls) at or in the immediate neighborhoods of the true directions of arrival. When the exact ensemble spatial covariance matrix is used, MUSIC and Min-Norm result in unbiased values (i.e., zero) forthenull-spectrum of uncorrelated plane I. INTRODUCTION waves at the true DOA irrespective of the signal-to-noise IGENDECOMPOSITION-based methods have found ratios and angular separations of the sources. This is in new prominence in array spectral analysis of plane contrast to schemes as such the autoregressive (AR) signals received in noise.Examples of recent reported methodwhichproducesabiasednull-spectrum that dework in this area include those of Schmidt [ l ] , Kumaresan pends on the signal-to-noise ratios even when the exact and and Tufts [2], Johnson and DeGraff [3], Bohme [4], covariance matrix is used. For a given angular separation Bienvenu [6] among others. A more extensive list of re- of sources and AR model order, there is a signal-to-noise search on this subject can be found in the abovemen-ratiobelowwhich the null-spectrum, at an angle away tioned references.These references also contain detailed from (usually between) those oftwoclosesources, is explanations of thephilosophies behind the various eigen- smaller than at either of the two true directions of arrival analysis-based techniques. The aim of this paper is, there- resulting in failure to resolve the two plane waves. fore, neither to introduce a new array spectral estimator When the narrow-band spatial covariance matrix is esnor to review, in any detail, the known ones. Rather, this timated from a finite number of independent snapshots, paper presents a statistical analysis oftwoofthemoretheeigen-assistedmethods also exhibit deviations from popular methods as reported in [ l ] and [2], namely, the zero in their null-spectra at the true angles, resulting in a MUSIC and the Minimum-Norm (Min-Norm) methods, loss of resolution. This deviation is due to the statistical with the aim of determining their resolving properties. sampling perturbation of the signal and noise subspaces. A common feature of the two methods discussed here This perturbation depends on the signal-to-noise ratios, is the decomposition of an estimate of the received signal signal parameters, and array specifications, which tocovariance matrix into orthogonal signal and noise gether determine the resolving capability of the estimation subspacesandformulating the direction-of-arrival esti- method used. Inthispaperwe will examinethe finitemator in one or the other subspace. For these methods, as sample bias in the null-spectra of the two eigen-assisted methods mentioned earlier. Manuscript received July 30, 1985; revised November 1, 1985. This The paper is organized as follows.First,thesignal, workwas supported by theDepartmentofDefense.Theviewsexpressed noise, and array models are formulated and the MUSIC are those of the authors and do not reflect the officialpolicy or position of and M ~ ~ methods are presented. N ~ first-order , - N ~ ~ ~ ~

the U . S . Government. The work of M . Kaveh was supported inpart by LincolnLaboratoryunderAirForceContractF19628-85-C-0002and in part by the National Science Foundation under Grant ECS-8105962. The workofA. J. Barabellwassupported by LincolnLaboratoryunderAir Force Contract F19628-85-C-0002. M. Kaveh is with the Department of Electrical Engineeiing, University of Minnesota, Minneapolis, MN 55455. A. J . Barabell was with M.I.T. Lincoln Laboratory, Lexington, MA. He is now with M/A-Corn Linkabit, Lexington, MA 02173. Number 8406980. Log IEEE

approximations to the mean and variance of the MUSIC null-spectrum. are derived. The expression for the MUSIC bias is used to obtain a resolution threshold for two uncorrelated, equipowered plane waves in white noise. This threshold is compared to the threshold signal-to-noise ratios obtained by Monte Carlo simulations. Finally, a firstorder approximation for the mean of the Min-Norm null-

.OO 0096-35 18/86/0400-033~1$01

0 1986 IEEE

332

IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-34, NO. 2 , APRIL 1986

spectrum is derived. One- andtwo-signal-in-noise models are used to assess the dependence of the spectral bias on the various measurement parameters. An important issue that is not addressed here is the determination of the number of sources. Throughout the paper it will be assumed that the number of sources are known.

11. FORMULATION H E METHODS OF T L Assume M incoherent plane waves are incident on a linR = X,S,S? i= I ear equispaced array of L sensors. For the kth observation period (snapshot), the spatial samples of the signal plus whereX1 IX 2 I 2 XM > X M + , = = XL -noise are given by o: are the eigenvalues of R , and Si are its orthonormal M 1 eigenvectors. X i ( S i ) , i 5 M are the signal-space eigenvalues (eigenvectors) and remaining are the ones the noise-space eigenvalues (eigenvectors). The null-spectra where for plane waves at a temporal frequency o f f Hz, for the different methods are now examined. wi = 2nDf/C sin Oi, C is the velocity of propagation, D A. MUSIC [I] is the element spacing, and Oi is the angle of incidence, Under the planewavemodel, D ( w ) forthe MUSIC with respect to broadside, for the ith wave. Nk is assumed method is given alternatively in terms of noise- and sigto be a complex, zero-mean, circular Gaussian vector with orthogonal elements, and plk) are the complex amplitudes nal-space quantities by
= of the plane waves with(pjk)I2 Pi. These amplitudes are assumed to be jointly circular Gaussian and jointly independent and independent of Nk. Throughout the paper we will use the superscript - and E [ interchangeably to and denote statistical expectation. The observation covariance matrix R , and its estimate R are given by

e ]

D ( w ) and B ( w ) , its estimate for the algorithms interest, of since these functions are substantially easier to analyze than their inverses. Throughout this presentation * will denote the estimate of the quantity over which it appears. This estimate, in turn, is a result of using R in place of R . We begin by writing the eigendecomposition of the covariance matrix R as

i=M+l

R = E[XkXF]

,Z Pi Rs, +
1 =

u:Z

(1) where the steering vector V ( w )is given by

where H denotes Hermitian transpose, u i is the element noise variance, and

e -jai

...

When V ( w )coincides with a signal direction vector at angular frequency wi, D ( w i ) = 0 as desired.

= -

k=l

c XkXf

B. Minimum-Norm [2] This technique finds the vector A with a unit first element which is entirely in the noise-space and has the minimum Euclidean norm. The null-spectrum is given by
D(w) = IVH(w)AI2.

(8)

A particularly useful expression for A is given in [2]. R is the statistic on which the angular spectral estimates We reproduce this formulation below and use itin the discussed in this paper are based. In the following, specsubsequent analysis of the Min-Norm technique. LetE be tral modelsof interest are first reviewed.These techconstructed as niques are based on an eigendecomposition of the covariance matrix into signal and noise subspaces followed by E = [SI, s , - SMl an associated spectrum, based on the signal-space or the noise-space information of the form
9

* We emphasize herethat S(w) is, in general, not the spatial where g = [s,(I), * * , SM(l)]. Then power spectrum of the process. We call D ( w ) the nullAT = [ l , - f l E T / ( l - gHg)]. spectrum. An ideal spectral model is one for whichD(wi) = 0, and D ( w ) > 0, w # wi, i = 1, * , M . In the It can be shown that in this case, as well, D ( q ) * , M , as desired. following analysis, we concentrate on the properties of = 1,

(10)
=

0, i

KAVEH AND BARABELL: STATISTICAL PERFORMANCE OF MUSIC AND MINIMUM-NORM ALGORITHMS

333

In the following, we introduce sampling errors in the estimationof R resulting in statisticalerrors in b ( w ) . Some statistics of b ( w ) are then obtained and related to the resolving capabilities of the two methods.

111. FIRST- AND SECOND-ODER MOMENTS OF THE MUSICNULL-SPECTRUM In this section, approximate statistical behavior of the MUSIC method is examined. Unfortunately, a theoretical analysis of thespectral estimators leading to a direct measure of resolution is extremely difficult at best. However, k#i especially in the statistical bias and variance of b(<w), neighborhood of w i , can beinterpreted as indicators of the Clearly,theexpressionfor q in (14) satisfies the conresolving capabilities of these techniques. Thus, in this straint in (17). We now proceed to derive the 'first- and ( section, we evaluate E [b w ) ] and var [d( w ) ] for MUSIC second-order statistics of the null-spectrum of MUSIC. and relate them to the angular separation and dynamic The estimated null-spectrum for this method is given rangeof thesources, array signal-to-noise ratios, and by number of snapshots. We closethe section by deriving an expression for the resolution threshold of MUSIC in the b ( w ) ' = 1 - VH(o) SiS?)V(,). (18) case of two equipowered sources. In the following analysis, we make use of the asymptotic statistics for the eigenvaluesand eigenvectors of the The expected value of b ( w ) , using the definition of S , is sample covariance matrix R of a complex Gaussian pro- given by cess as derived partially in [ 5 ] and partially in Appendix N A. Asymptotic is in the sense of large and our analysis is based on the first-order perturbation of D ( w ) and D ( w ) ~ in the neighborhoods of the signal angles. One- and twosignal examples are specifically treated to obtain illustrative results. We begin by summarizing the first- and second-order statistics of the estimation errorsin f i i and S i as given in Appendix A. Let Si= Si v i and = hi + pi, where hiis the ith distinct eigenvalue of R . Then q and Substituting for the expectations in (19) gives pi have the following asymptotic properties: B ( w ) = D(w) - V"(w) E [ p i p j ] = - 6ij. I N

Taking the expectation of (16) andretaining order N terms results in

'

(jl

xi

x2

Using (A. 15) and (A.13)

E[qiqfl =
from (A. 16)
E[qiqf1

x.

kzl
k#i

I -

Ak

(hi - hk)z

sksr6ij,

(12)

hi Aj -N(Ai - hj)z SjST(1 - 6ij)7

3) (A. and

gives

hi E [ ? j ] = -2N

k= 1
k#i

" c (Ai -

S . = aisi.

(20) Since our primary objective is the investigation of the resolving capability for these methods, we will concentrate on the statistics of b ( w ) in the neighborhood of q , i = l , . . . , M . The results for M = 1 and M = 2 follow (I3)immediately, noting that V ( o i )E span ( S i > ,i = 1, * * , M and, therefore, V ( w i )is orthogonal to Si, = M i 1, . . . , L. It follows that M = 1: (14) hp2,(L - 1) D(w1) = N(A1 s

In theabove, 6 . . is the Kroneckerdeltaand the approximations are o ($-I). It is worth noting that is (14) consistent with (12) and the normalization constraint on ISi[ and 1 Si\ canbe seen as in the following. The constraints result in

M = 2:

b (wk)

0; V H ( W k )

I$('

s?si + srqi + q r S i +

(15)

334

IEEE TRANSACTIONS SIGNAL ACOUSTICS, ON SPEECH, AND PROCESSING,

VOL. ASSP-34, NO.

2, APRIL 1986

The next quantity of interestis the variance of6 (w) for which we again obtain the order of N - ' approximation. It is convenient to define D ( w ) as

resulting in

D(w) = 1

D,(w) This is not surprising as the one signal case is a rather stable one. 2) M = 2: Using the second-order statistics (12) and (13) and simplifying the result at wk,k = 1, 2, leads to

where D,(w) is the null-complement spectrum. The variance of D ( o ) is then given by var

[d(w)] var [B,(w>l. =

Therefore, we first derive an expression for @(a). From the definition of d(w)(18), it follows that

s;(w)
where

VH(W)TV(W)

(23)

'

V*(k) V(1)Qk)

q?( I ) ST(n) $;(rn)

(24) Therefore,

and V ( k ) ( S ; ( k ) ) is the kth element of V(w)( S , ) . Substituting (k) = & ( k ) 7; ( k ) and neglecting higher-thansecond-order moments of 7; ( k ) , we obtain

si

T(m, n) =

k=l I=l

c c V*(k) V ( I ) c c [Si(k) Siyc(n) w,(m, I )

+ T(I) W;(k, + S;(k) ST(1) Wi(rn, n) 6,j $(m) n) + S;(k) S,(m) Uf,(n, I ) + S?(n) q y I ) U;,,(k, m) + S?(n) S,(m) Fv.(k, I) + Si@) s.( I ) ST (n) S,(rn) j
where
4- 2 (a;

Uj)

S;(k)sj. (I) q m ) ST cn,l)

Y(I,m) = (I, m)th element of E[v j q,f]

and

Ui,,(I, m) = ( I , m)th element of E[q ;q;].

The expression in (25)isquiteunwieldy with doubtful general informational value. For one and two signals in white noise, fi:(wk) can be simplified using the statistics of v i and the orthogonality of V ( w k )and the noise eigenvectors. These results are briefly examined as follows. I) M = I:

Therefore,

KAVEH AND BARABELL: STATISTICAL PERFORMANCE OF MUSIC AND MINIMUM-NORM ALGORITHMS

335

We now evaluate var [d(wk)] B ( w k ) in terms of the and signal, noise, and array parameters of interest. The deviation of B ( w k ) from zero, the null-spectral bias, signifies a loss in resolution of the spectral estimator. We first give approximate expressions for this bias, for later comparison to Min-Norm. Subsequent analysis approximates the resolution threshold ,for equipowered, two closely spaced uncorrelated sources as that signal-to-noise ratio at which
b(wk),

1, 2,

is equal to

r i ( W 1

; ").
ARRAY SIGNAL-TO-NOISE RATIO (dB)

[B(wl)]

For one signal in noise it was shown in (27) that var 2 0. b(q) : for LPI >> a t is given by
D(q) =

Fig. 1. Probability of resolution as a function of the ASNR for two equipowered emitters spaced 0.1 BW apart. Ten looks, 100 trials, five-element array.

( L - 1)o; N L PI

Thus, the bias at w1 is directly proportional to the noiseto-signal ratio and inversely proportional to the number of snapshots as expected. The above expressions for B ( w k ) indicate the general For closely spaced equal-power signals, using (B-7) and behavior of the null-spectrumbias as a functionof source (B-8) from Appendix Bin (30) and retaining the two dom- and array parameters. In order to obtaina quantitative inant terms gives measure of the resolution threshold for twoclosely spaced equipoweredsources,we now introduceaplausible,non3(L - 2)A4(l 2ASNR) var [B(wl)l 2: (32) probabilistic approach based on f i ( w ) . We propose that 160N(ASNR)2 the signal-to-noise ratio at which B ( w J = B ( w 2 ) = where ASNR = LP/o: is the array signal-to-noise ratio B(w,), where w, = (wl w2)/2, is approximatelythis and A = L ( o l - w 2 ) / 2 h is proportional to the normal- threshold. The reasons for this conjecture are as follows. ized difference between sines of the DOA. It is clearthat Resolution is achieved when b(wl)and B ( w 2 ) are both << 1/2 lessthan D(wm).' When the above equality is true, the this variance decreases as (ASNR)-2 for ASNR region probability of resolution ranges from approximately 0.33, and as (ASNR)-' for ASNR >> 1/2. For the latter (the practical one) and large L , the variance is actually for the case of totally independent variations of b (wl j , determined by the inverse of the elemental signal-to-noise 6 ( w 2 ) , and b (w,), to nearly 0.5 for the situation when ratio, P/o$. b ( w I ) and B ( w 2 ) are maximally correlated. Again using the results of Appendix B in (22) and reFig. 1 showsthe resolution probabilities for MUSIC taining thelargesttwoterms, we obtain approximate and Min-Norm of a typical array obtained by Monte Carlo expressions for B ( u k ) For equal power sources, we get . simulations. It is clear that the ASNR for probabilities of 0.33 and 0.5 are'within 1-2 dB of each other. Thus, the proposed method should give the approximate ASNR for the 0.3-0.5 probability of resolution threshold region. of B and It is clear from (32) and (33) that, for reasonable array Note also that equating the averages 6 (ol), (a2), B ( w m j is justified due to the small standard deviation of parameters, the bias in D ( w k ) is substantially larger than relative to its mean. We again substitute the expresits standard deviation (SD). For example, for L = 10, A2 B(ok) sions from Appendix B in (22) and retain all terms to ob= 0.003, and N = 100, d . ( w k ) / S D [ b ( w k ) 2 lo4. Thus, ]: and tain b ( w J and B(u2), in (20) to find b ( w , ) . Thresh~ _ we expect that the resolution threshold is mostly deter= ) j old occurs at the ASNR for which f i ( ~ ~ ( ~d(w,) with mined by the behavior of B ( w ) , For two sources with unequal powers as given also in the resolution probability increasing as B ( W ~ becomes (~)) Appendix B, we obtain the following approximations to increasingly less than d(w,). Let the threshold ASNR be D(w& denoted by $+, Then

where ASNR in this case refers to the weaker source array signal-to-noise ratio and A2 << 6 = P2/P1<< 1. A similar expression may be obtained for B ( w 2 ) .

(34)

'In each simulation trial, the two sources were considered resolved if the spectral output of the algorithm under test had two peaks, each within abeamwidth of the respective true source location, and also the power estimate [ I ] for each source wasnot more than 10 dB above the true power of the source.

336
""

IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING,

80

VOL. ASSP-34, NO. 2, APRIL 1986

\.
c
v

-. - 100

LOOKS

\.

\.

- 1000

m n

\.

__
\.
I

\. \.

LOOKS

x.\.

10000 Simulation Data

100 - 1000 --- 10000

Simulation Data

z
cn
c3

LT

J I
0

r
W

40-

v)

cn
W

LT

20 I I I I I 1 1 1 1 1

LT

I -

20 -

' .

' .
1. 0 '
1

10-2

10-1

10-2

ANGLE SEPARATION (BW) Fig. 2. Resolution threshold versus angular separation in beamwidths for two equipowered emitters. Five-element array.
80
L
. x.

ANGLE SEPARATION (BW) Fig. 4. Resolution threshold ,versus angular separation in beamwidths for two equipowered emitters. Twenty-element array.

LOOKS
1 .

100

\.

- 1000
-- 10000 I Simulation Data

0 40-

I I 0 w ILI I

IV. THEMEANOF THE MIN-NORM NULL-SPECTRUM The evaluation of the moments of the null-spectrum for this method general in considerably complex is more than the previous one. In the following, an approximate expression for b(q)is derived. The approach is again based on a perturbation of the exact quantities with small random disturbances as discussed before. We use (10) in conjunction with the estimate A where
A T = [ l , -f8'T/(1 The null-spectral estimate is

gHg)].

(36)

I - -

then
I
I
I 1 b 1 1 1 1 1 1 1 1

B ( w ) = ( V H ( w ) A = VH(W)A V(w).(37) l2 AH
10-2

10-1

Let

A =A +

a . d ( w ) given is then

by

ANGLE SEPARATION

(BW)

Fig. 3 . Resolution threshold versus angular separation in beamwidths for two equipowered emitters. Ten-element array.
f
tT
L

B ( w ) = D(w)

+ V(w)" a

+ 2Re [VH(w)a
a

AH V ( w ) ]
(38)

V(w).

r
L

It is of interest again to consider the behavior of D(a)at signal frequencies wk. Using fact the that D(wk) = IA"V(w;)12 = 0, b ( w k ) reduces to

(35)
Several features of ,$?-canbe observed from (35). First, Thus, we have to find an expression for the correlation for large L , ET is approximately proportional to L. This matrix of a . Denote the random perturbation in E ' and means that one can define a threshold element SNR which g by E ' and y , respectively. Then does not change significantly when L changes. Second, E' = E ' E' and g =g y. (40) for N << 5LA-2, tTvariesas A d 4 N - ' whereasfor N >> 5LA-2, it varies as A-3N-1'2. We now findan approximation to a and then derive an Finally, to verify the'accuracy of this approximate analexpression for aa". Note that the first- and second-order ysis, ET was calculated for several combinations of array statistics of E' and y are those of the appropriate elements and signal parameters according to (35) and compared to of the set { q i i = 1, . * * , M , [see (9)]. By definition 1, the resolution thresholds from Monte Carlo simulations. we h.ave Figs. 2-4 show examples of these comparisons for three array lengths. The threshold is ET in decibels and the angle separation is 2wd. In the figures, the simulation data is plotted as I . The bottom to the top of the I approximately indicates the 33-50 percent threshold ASNR's, reTherefore, @wk) = V' *(ak) ' a' H V ( ~ k ) .V is a vector a spectively. The close agreement between the theoretical formed from the secondto the Lth elements of V . We now predictions and simulation results is clearly evident.

KAVEH AND BARABELL: STATISTICAL PERFORMANCE OF MUSIC AND MINIMUM-NORM

ALGORITHMS

331

develop an approximate expression for ' in terms of the Then a lowest order perturbations as follows:
1

1 -gHg
where

1 + P

1 - gHg

LPlN

(49)

closely spaced For M = 2 we considerthecaseof sources with equal powers. Again, the second termin (45) dominates, resulting in
1

and

=
Consequently,

-1
I

[pE'g*

1 - g g

+ E'Y * + e r g * ]

(rllV'H(Uk) , 2+ r21VfH(Uk) 2 l 2 l (50) F1 F where T i = [ y * y and Fi is defined in Appendix C as the vector of the second to 'Lth elements of S i . Following the same procedure as for M = 1, we obtain
*

where

+ E' y * g T C f H .
Thequantitiesunder the expectation operations are evaluated in Appendix C. In the following, these results are used to obtain d.(Uk) for M = 1 and M = 2. Unfortunately, @(corn) is analytically intractable for Min-Norm. Therefore, no expression for athreshold ASNR is derived for this method. For M = 1 we have g = S1(l) = l/& and
1

Substituting the values from Appendix B, we obtain

1
fi(Uk)

(1 -

;+V' [ ) I

(ASNR)

1 1 .
(l -

V' H ( W 1 ) E'E' HV'(WI) 1. =

1\2
L \

(47)

E)

We now make a further assumption that L >> 1. This assumption is not crucial but allows for a more manage. able expression for d . ( ~ , ) Substituting the various quantities fromAppendixCinto (45) and noting thedominance of the second term in (45) results in

The above results on d.(w1) are somewhat optimistic for Min-Norm.There are other, smaller terms in (45) that contribute to the null-spectral bias. Nevertheless, the dominant makeup of this quantity does indicate a smaller bias (by nearly a factor of L ) as compared to MUSIC. This might be expected2 tocarry over to the relative resolution thresholds for the two methods, making the MinNorm threshold lower than that of MUSIC. This difference in the resolution thresholds has been verified in numerous simulations, an exampleof which is shown in Fig. 1.

V. CONCLUSION This paper presented an asymptotic evaluation of the resolving capability of two eigen-assisted spectral domain 1 estimators of the directions of arrival of closely spaced, &w) = V' H(o) * TE'H V ' ( ~ ) E' * narrow-band plane waves. The mean and variance and the 11 - mean, respectively, of thenull-spectra of the MUSIC and Minimum-Norm algorithms, including O(N- I ) errors, were derived. These were approximately related to the resolving power of the two methods and their dependence on such parameters as the relative angular separation of of snapshots, and emitters, number of sensors, number signal-to-noise ratios were delineated. For the MUSIC al(48) gorithm, an expression for a plausible detection threshold The sum in (48) can be found using the orthonormalityof c s k > as *Of course,.we have not analyzed &urn) here,,nor have we shown that

'I2

thestandarddeviation of the Min-Norm null-spectrum is smallaswe showed for MUSIC. Hence, the analytical comparison of the methods is conjectural.

338

TRANSACTIONS IEEE

ON ACOUSTICS, SPEECH, SIGNAL AND PROCESSING,

VOL. ASSP-34. NO. 2, APRIL 1986

was derived that showed close agreement results from with Monte Carlo simulations. Our results indicate a smaller bias in the Min-Norm null-spectrum, at a source angle, compared to the MUSIC null-spectrum. This suggests (but does not imply) a resolution threshold which is at a lower signal-to-noise ratio for Min-Norm, a fact supported by simulation results. Above below and these thresholds simulations have shown the two methods to behave similarly in terms of resolution probability, bias, and variance of the estimated angles.

+ I= c

tlypSj

APPENDIX A The expressions for tg) are obtained by substitution of In this appendix, we derive the o ( N - l ) approximations to the first- and second-order moments of i jand Sibased (A.8) into the perturbed equation for the eigenvalue probon the perturbation formulations in [7] and the statistical lem. Following [7], we have results in [5]. It is understood that the statistics that are giveninthefollowing areforthedistincteigenvalues (signal-space eigenvalues) and their associated eigenvecand S j , i, j = and tors.areasymptoticallynormal with 1, * , L , asymptoticallyindependent.Furthermore[5],

I j#i

tg!p2Sj.
.j = 1 j # i

(A.8)

xi

xi

X;

+ O(N-1)

(A. 1) (A. 2)

cov

(Ai, i j ) G,X?/N. =

No explicit expression for S j is given in [5]. Moreover, all the existing derivations of the statistics of the eigenvector estimates neglect the numerical normalization of $. In the following, we derive the o(N-l) approximation to the first- and second-order momentsof Siassuming that \ S i \ = 1. Following Wilkinson's approach [7, p. 681, we define R in terms of a random perturbation to R with a perturbation factor p , 0 < p << 1. Thus,
R=R+(R-R)=R+~B

-c

(S:BSk)
-

k=I (Xj k#i

(SFBSJ X;)(& - Xi)'

(A. 10)

The expected value of in (A. 8) only involves E[It:! 1'1 and E[t$]], as E[tt)] = 0 from (A.9). Using (A.4) and the fact that Si are eigenvectors of R (i.e., SYRS, = X i a j j ) , it follows that

si

(A.3) and where B is a Hermitian, zero-mean random matrix with elementsthatareasymptoticallyjointlyGaussian,obviously defined by (A.3). We will need the following re- Thus, sult later. Let A;, i = 1, 2, 3, 4, be four complex vectors with appropriatedimensions. The followingexpression results (see [ 5 , p. 1141):

E[t$'] = 0.

(A. 12)

Let Sj denote the unnormalized eigenvector given perturbation expansion as in [7] by

L
/ m
\

The approximation of order N-l to the central secondorder moments of S i are the same as those given by BrilThis in a linger ([5, p. 4541) for {Si>,. can be seen from the following. Using (A.7) in the definition of S i , we obtain cov

(Si, Sj) = cov (Si, 3;)

f

( , are the coefficients of the perwhere tkj0 , k = 1, 2, * along Sj. The following steps turbation expansion of develop the lowest order perturbation expression for the Using the normalized estimate of the ith eigenvector orthonormality of Si and keeping the term with the lowest order of p , we have

si

si.

= 1

+ jc =
I
j#i

1tjyp2

(A.6)

KAVEH AND BARABELL: STATISTICAL PERFORMANCE

OF MUSIC AND MINIMUM-NORM ALGORITHMS

339

Careful examination oflast three terms the inabove the equation shows that they involve third or higher order moments of elements of the zero-mean complex Gaussian matrix B . Thus, these terms contribute zero (for odd or5 N - , . It follows that ders) or quantities of order cov

X!,(,) = PL(1 + 1 4 )
(-)

03.5)

u + 1
S1m = J . &

u 2

(B .6)
(e)

(Si, Sj) = cov Sj)

=

(si,sj)+ o(N-,).
XiXk

(A.14)

Thus ( [ 5 , p. 343]), cov and cov

(Si,
(Si,

Let V be the direction vector for u, = (a1 , A2 = L2w213. Then

+ @,)I2 and

k = l (Xi k f i

SkSf6, N

(A. 15)

Xi IvH(ai)S I \ , = __ 2PL

A, 3 = 1 - - + - A4 4 80
'

ST) = - (Xi

XiXj S,Si'<l - 6,). - X,)2 N

(A.16)

A2 3 Xi IVH(uJ S2I2 = - = - - - A 4 80 2PL

and

Formulas (12) and (13) follow from (A.15) and (A.16) by W 2or neglecting terms of order N less.

A4 B APPENDIX (VES[I2= 1 - 80 This appendix develops expressions for the eigenvalues, eigenvectors, and several associated inner products Iv:s212 = 0. (B. 10) that are needed for the approximate evaluation of the mosignal ments of the null-spectra for one and two signal cases. We next consider an extreme case for which one For one signal in ,noise SI = V(wl) and X I = LPI u2,. dominates the other. For two uncorrelated signals, define centered direction 2) A2 << (P2/PIj = 6 << 1: To obtainmanageable V(ui), i = 1 , 2. We consider expressions in this case, we only retain terms of the lowvectors by Ui = e-j'(Ltwo cases: equal and unequal powersources. In both cases est possible order in 6 and A,. Expanding the square root the sources will be assumed to be closely spaced with re- in (B.3), we get spect to the array beamwidth. Thus, letd = VH(uI) v(~,) Xi = L(l X; L ~ A ~ and a d = (ul- w2)/2. We have 6) and - = - (B. 11) PI P I 1 6'

From now on assume we panded as

<< 1 . (B. 1) can be

Substituting (B. 11) in (B.4) and carrying out the normalization of Si the following approximate expressions ex- develop:

d = e-j(L-I)wc/[l-

&2W2

+ .12,4W2+ i!&

. . . 1.

(B.2)
and ei-

I S ~ V ( W= (1~- A2) ~) ~ I S ~ V ( W2: (1~- A2)(1 ~) ~

For two uncorrelated sources, the eigenvalues genvectors are given by [8]

+ 6)2,

(B.13) (B. 14) (B. 15)

and

where oc is used to indicate that the expression is not nor: malized and X; = X i - u . The two cases, P I = P2 and P I >> P2, are now considered. I ) PI = P2 = P: Forthiscondition,simple expressions for the eigenvalues and eigenvectors result. From (B.3) and (B.4)

APPENDIX C In this appendix, we derive approximate expressions for the expectations in equation (45). Again, we only keep the first-order terms. 1) Let [A], denote the ijth element of the matrix A . Then the second-order statistics of Y are given by
[Y*YT1ij =

[rr"]g

340

IEEE TRANSACTIONS ACOUSTICS, ON

SPEECH, AND

SIGNAL PROCESSING,
M

VOL. ASP-34, NO. 2, APRIL 1986

and

..
(C. 13)

-&Xj
N(hi hj)2

S ( l ) Sj(l)(l - 6q).

(C.2)

6) The final term of interest is y * g T e r H . First, consider the jth element of gTe
M

[gTcH]j =
Simplifying, we obtain Then

k=I

Sk(1) q f ( j

+ 1).

(C. 14)

(C.4)
3 ) E r g * g T d H derived by noting that is 1 ) and g ( i ) = S i ( l )

[Iij

= qj(i

+
(C. 15)

where FT = [Sj(2), Sj(3),

---

ACKNOWLEDGMENT

,Sj(L)].

The authors are indebted to Drs. K. D. Senne, L. Horowitz, and D. DeLong of Lincoln Laboratory for many of this investigahelpful discussions during the course tion.

From (C.2), we get

REFERENCES
R . 0. Schmidt, Multiple emitter location and signal parameter estimation, in Proc. RADCSpectral Est. Workshop, G r i e s s AFBS, NY, 1979. R. Kumaresan and D. W. Tufts, Estimating the angles of arrival of multiple plane waves, IEEE Trans. Aerosp. Electron. Syst., vol. AES19, Jan. 1983. D. H.Johnson and S . R. DeGraaf, Improving the resolution of bearing in passivesonararrays by eigenvalueanalysis, IEEETrans. vol. ASSP-30, pp. 638-647, Aug. Acoust., Speech, Signal Processing, 1982. J. F. Bohme, On parametric methods for array processing, in Proc. EUSIPCO-83, Sept.1983. D. R. Brillinger, Time Series: Data Analysis and Theory. New York: Holt, Rinehart and Winston, 1975. G. Bienvenu and L. Kopp, Optimality of high resolution array processing using eigensystem the approach, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-31, pp. 1235-1248, Oct. 1983. J. H.Wilkinson, The Algebraic Eigensalue Problem. New York: Oxford University Press, 1965. J. E. Hudson, Adaptive Array Principles. Stevenage, U.K.: Peter Peregrinus Ltd., 1981.

The second term follows from (C. 1) as

Again, we first consider the first term in the brackets.

g H T =

i= I

c
M

S7(1) q j ( l ) r H .

(C. lo)

Therefore,

Mostafa Kaveh (S73-75%SM83) received the

j = 1,

* * *

, L - 1. (C.11)

The second term is given by

YHg@ =

i= 1

c Si(1)q?(l)

ErH

(C. 12)

KAVEH AND BARABELL:PERFORMANCE STATISTICAL OF MINIMUM-NORM AND MUSIC

ALGORITHMS

34 1