Performance Analysis of the MUSIC Algorithm t

DANIELSPIELMAN, A . PAULRAJ and THOMASKAILATH

Information Systems Laboratory
StanfordUniversity
Stanford, CA 94905

Abstract-The MUSIC algorithm is one of t h e m o r e irn- (a(0)) for 0 E [0,2n] define the action manifold; and the
portant high resolution approaches for direction finding goal is to collect multiple snapshots XI, x2,... ,XN and
and spectral estimation that have been developed in re- estimate the specific 01,62,.. . ,BK that gave rise to the
centyears.Asymptotically (i.e., infinitedataor SNR) data. For example, in the direction finding problem, the
t h e MUSIC algorithm has been shown to yield efficient
6"s represent angles of arrival, the action vectors are the
unbiased estimates. I-Iowever, the performance of t h e al-
gorithm for the non-asymptotic situation of high noise and
direction vectors for the wavefronts, the action manifold
l i m i t e d d a t a h a s n o t b e e n fully addressed. In this paper, is referred to as the array manifold and s is a vector of
we study the performance of t h e M U S I C a l g o r i t h m w h e n the impinging signals.
only finite noise corrupted data is available. TheMUSICalgorithm ([111 [Z])forestimating these
We focus on the role of array design in the performance quantities proceeds as follows. N ( > M ) snapshotsare
of MUSIC algorithm for direction finding and introduce collected in a large data matrix
certainmeasurestocharacterizeitsperformance. We
show that in the single target situation, these measures x = (x1. XN] (2)
c a n be describedinterms of thefamiliarconventional
bearnpatterns.Results of computersimulationscarried and this is used to estimate the array covariance matrix
out to check the usefulnessof such measures are also pre- bv
sented.

i= 1
I . INTRODUCTION a spectral
This Hermitian matrix is then decomposed via
decomposition into
We considerthe following data model. Let x be a
M X 1 complex vector of sensor outputs that consist of M
impinging wavefront signals and additive noise. The sig-
nal component of the received vector is a weighted linear i=l
combination of K (where K < M ) action vectors. These
vectors lie on a one-D continuum traced by a(6) (which where A 1 2 X2 . . . 2 .AM is the set of ordered real eigen-
is parameterized nonlinearly by 0 ) hnd is referred to as values and {?I I 9 2 . . . ,O M } the corresponding eigenvec-
the actionmanifold. Theactionvector a(0) is defined tors.
as the vector received by the array when a single unit From (l),it is easily shown that the true covariance
amplitude signal from direction0 impinges on the array. matrix
Thus R =E [=*I = ASA' +
a21 (5)

x = [ a ( s l ). . .a ( O ~ ) ] s+ n=As 4n (1)
where S is the covariance of the sources, 'a is the noise
variance and is equal to the repeated ( M - K times)
minimumeigenvalue of R. The space spanned by the
where n is a M X 1vector of uncorrelated white gaussian eigenvectors associated with the non-minimaleigenvalues
noises and s is a K X 1 vector of the signals as sensed
is referred to as the signal subspace. The other M -
at a reference point (say tap/sensor 1). A is a M X K K eigenvectors associated with the minimal eigenvalues
matrix of the action vectors. The set of action vectors span the noise subspace.
Since the array manifold vectors {.(e,), . . .,.(OK)}
'This work was sponsored in part by the U.S. Army Research
also span the signal subspace, the solutions for {e; i =
Office Fellowship Program under grant numberDAAG29-83-G-
0020, the Office of Army Research under contract DAAG29-83-
1,.. . , K } are given by the K intersections of the array
K-0028, and by the SDI/IST Program managedby the Office of manifold with the signal subspace. In practice, since we
Naval Research under Contract N00014-85-K-0550. do not have a perfect estimate of the array covariance

35.24.1
ICASSP 86, TOKYO CH2243-4/86/0000-1909 $1.00 0 1986 IEEE 1909
matrix, statistical tests are needed to estimate the num- its intersections with estimated signal subspace provide
ber of sources (i.e., M - {multiplicity of the minimum the parameter estimates.
eigenvalue }). This can be estimated either via multiple We consider two types of estimation errors; namely lo-
hypothesis testing (see [I],[SI), or information theoretic cal and global. Asymptotically, the high-resolution mea-
criterions (AIC or MDL) (see 141). After determining the sure given in (6) will produce infinite peaksat the truedi-
number of sources, the intersections of the array mani- rection angles. However, with finite data, the peaks will
fold with the signal subspace can be estimatedusing the be rounded and perhaps shifted. That is, the desired ar-
inverse squared magnitude of the angles between the ar- ray manifold vectors are not quite orthogonal to the noise
ray manifold vectors and the eigenvectors spanning the subspace. Thus local estimation errors can be described
noise subspace: by the deviation of the shape of the observed peak from
the peak thatwould have resulted in the asymptotic case.
Global errors are defined as “false” peaks, that is peaks
in (6) that are not within some small neighborhood of
the true directions.
A one-D search is carried out over 0 E [ O , 271-1 and the We shall first consider the local error performance. If
the estimated noise eigenvectors are the perturbed ver-
K dominantpeaksarethedesiredestimatesforthe
directions-of-arrival. sions of the true eigenvectors. i.e.,
In general, an arbitrary array manifold may not have
K uniqueintersectionswiththeK-dimensionalsignal
subspace. In order to avoid this situation, the manifold
should not have any ambiguities. Specifically, any set of for i = K f 1,... ,M , then the observed peak will be
K , where K < M , unique array manifold vectors must
span a unique subspace. Ambiguities are an important
considerationin choosing thearray geometry.For ex-
ample, in the direction finding problem, if we assume
narrowband signals, far field sources and an uniform lin- Defining A0 as the width of the ideal response at the
ear array with X/2 element spacing, the action vectors amplitude corresponding the the height of the observed
have the following Vandermonde form response, it can be shown that

If Avi isassumed to consist of independent equi-

variance components, the numerator in this expression
It has been shown that a manifold t r x e d by such vectors becomes independent of the array geometry and is of no
does not have the ambiguities of any order. interest in the minimization. Thus to minimize AO, the
denominator must be maximized. Rewriting the denom-
inator of (10) yields
11. ANALYSISOF PERFORMANCE
Asymptotically, i.e., as the number of snapshots be-
comes very large,theMUSICalgorithm yields unbi-
ased estimates whose variances match the Cramer Rao
bound ([11). However, with a limited number of snap- where u is a unit vector in the direction of da/aO and
shots and/or high noise, the behavior of the algorithm is P, is the projection matrix onto the noise subspace.
not well understood. Of particular interest is the rela-
tionship between the error performance and the design P,=P - A(A*A)-~A* ( 12)
of the array manifoldor equivalently the array geometry.
. To addressthesequestions,somegeneralcomments For K sources, maximizing (11)is difficult because P,
about the MUSIC algorithm are necessary. For simplic- is a function of the K unknown directions. However, for
ity, let us consider a one dimensional problem (Le., the a single source at direction Bo, we have
array manifold is function of a single parameter) where
the parameter t o be estimated is denoted by 8 . Perfor- P, = I - a(B,)a*(O,) (13)
mance, or more specifically the ability to accurately esti-
mate {e, for k = 1,.. . , K } , for the MUSIC algorithm is Now, considering a line array with elements at positions
clearly based on the structureof the array manifoldsince . . . ,z ~ - 1with zo = O and 3;M-1 = L, (;.e., the
zo, zl,

35.24. 2
ICASSP 86, TOKYO
1910
array has a fixed aperture L), the minimization of A6 is Again considering the linear array mentioned above,
achieved when the distance measure becomes

xi = L for i = 1,..., M-1 (14)

The above solution maximizes the magnitude of a a / d 6

and thus makes the array manifold as long as possible. where x0 = 0. This function is easily recognized as the
However, placing all but one of the antenna elements at beam pattern for the antenna array. Traditional array
zi = L is not an practical solutionsince it would result in designs strive for beam patterns with the same proper-
what is conventionally referred to as aliasing or grating ties that are desirable for the high resolution MUSIC
lobes; and in the context of the MUSIC algorithm, such algorithm;namely,anarrowmainbeamand low side
grating lobes represent unresolvable ambiguities in the lobes. It is, however, important to stress the difference
arraymanifold.Sourcesfromtwo different directions between the interpretation of this beam pattern for con-
produce the same response at the output. ventional processing and for the MUSIC algorithm. In
To resolve this issue, the probabijity of global errors, conventional beamforming, resolution is limited to the
that is ambiguities or near ambiguities, also need to be width of the main lobe; and signals leaking in though
examined. A global error, namely a peak in (6) which is side lobes can mask weaker sources. However, the MU-
not within some small neighborhood of a true direction, SIC algorithm is able to resolve sources much closer than
occurs if there exits some 6 outside the neighborhoods the width of the main beam; and there is no masking
around {01, 02,. . . ,e,} such that of weaker signals. Side lobes are only important in the
threshold phenomenon of global errorswhich occurs with
few snapshots and high noise.
Unfortunately,theprecedinganalysisdoesnot indi-
cate how to choose the array geometry for the MUSIC
where P, is a projection matrix th;tt projects onto the formulation of the direction finding problem. For multi-
signal subspace. i.e.,
ple targets it is clear that the optimum array geometry is
a function of the number of signals and their directions-
of-arrival. One approach would be t o optimize the ar-
P, = A ( A * A ) - ~ A * (16)
ray geometry for the worst case signal scenario. Such a
Again P, is a function of all of the K unknown directions search procedure should also include any practical con-
straints on the array geometry etc. For the single source
and thus (15) is not a very useful performance measure
though it does propose a clear criterion for maximizing case, the problem reduces to conventional beampattern
global error performance. In order to get a better insight design. This should be no surprise since the MUSIC al-
gorithm itself reduces to conventional beamforming.For
into the problem,we consider a single source ata(&)and
normalizing all a(6) to have unit ncrm, (15) reduces to conventional beamforming, considerable research has be
an easily interpretable result - it is the magnitude of the done on designing arrays; and while there are no clear
solutions to the problem of optimum element spacing,
cosine of the angle between two array manifold vectors
all such work is directly applicable to the MUSIC prob-
lem for the one-Dcase. The main results use dynamic
l'pya(s)l' = a*(O)a(Oo)= cos(a(O),a(&)) (17) programing, perturbation analysis, as well as statistical
1140) I1 techniques; and we refer the interested reader to ([SI, [6],
(71, [81). for details on the design of so called aperiodic
This performance measure yields considerable insight
arrays.
intothebehavior of theMUSICalgorithm for differ-
entarraygeometries.Consider (17) asafunction of
( 6 - eo}. For good performance, we would like this func- 111. SIMULATIONS
tion to fall off rapidly as 0 moves away from 60 A small
Icos(a(6),a(Oo))lfor some pair (6,Oo) implies that the two Several computer simulations were performed to jus-
vectors are widely separated in space. Since the MUSIC tify the use of (15) as a performance measure for the
algorithm looks for array manifold vectors which lie is MUSICalgorithm.Thedirectionfindingproblemwas
some subspace, a wide separation results in two vectors simulatedusing a 4 elementlineararraywith a fixed
which are easily distinguished.Furthermore,this dis- interelement spacing of 0 , X, 1.5 X, and 6.5 X, The sim-
tance function gives additional insight into the shape of ulationsconsisted of 5000 trialswhere,for each trial,
the manifold. Sidelobes, that is lartye peaks other than 10 snapshots were generated and the MUSIC algorithm
the main peak, represent near ambiguities in the array was used to produce an estimated direction for the single
manifold where sections of the manifold for widely sep- source located at 6 = 0'. The SNR for the first simula-
arated 6 become close together.Thesecorrespondto tion was -10 dB. The 5000 estimates were plotted in the
regions with high probability of global errors. form of a histogram.

35.24.3
ICASSP 86, TOKYO 1911
 Both the antenna array beam pattern and histogram 151 M. G. Andreasen,Lineararrayswithvariable in-
are shown in figure 1. These plots ihow that the beam terelementspacings, IEEE Trans. onAntennas and
pattern is closely related to the probability density func- Propagation, AP-10:131-136, March 1962.
tion of the MUSIC estimate for a single source. As the
sidelobe heights increase so does the probabilityof global [6] A. Ishimaru,Theory of unequally-spacedarrays,
errors. As the SNR is increased to -3 dB, see figure 2 , IEEE Trans. onAntennas and Propagation, AP-
the number of global errors decrease:;. This i s consistent 10:691-702, November 1962.
with the previous asymptotic analyses of the MUSIC al-
gorithm. [7] Y. T. Lo and S. W. Lee, Astudy of space-tapered
arrays, IEEE Trans. on Antennas and Propagation,
AP-14(1):22-30, January 1966.
I V . CONCLUSIONS
[8]Bernard D. Steinberg, Principles of Aperture and
In this paper, we addressed the relationship between Array System Design, John Wiley andSons, New
the performance of the MUSIC algorithm and the design York, 1976.
of the array manifold or equivalently the array geome-
try. We derived a performancecriterionforlocaland
global errors that also involved the unknown directions-
of-arrival. Thus resultingin the optimum array geometry 0.9

being a function of the signal environment. This suggests 0.8

a worst case approach to arraydesign where, we optimize -.
20.7
array geometry to maximize performance under the pre- +
dicted worst case signal environment. Additional work 20.6

is necessary to explore optimum array designin multi V;0 . 5

target environments. $0.4
..
In the single sourcecase, we showed that the angle E
k0.3
between two array manifold vectors is a reliable perfor- c
mance measure, valideven for non-asymptotic situations. 0.2

These results indicated that the traditional work done on 0.1

antenna array design is applicable to the MUSIC formu-
0
lation at least in a single target problem and possibly - 1 0- 3 0- 5 0- 7 0- 9 0 10 30 50 ?0 90
also provides a basis for multi target situations. Also, ongle

the performance analysis for the MUSIC algorithm pre-

sented here may provide further insights into the classical
array design problem of finding good beampatterns. Figure 1: Histogram of MUSIC Estimates for SNR -10 dB

. 1 . . ~ 1 , - . , . 1 1 ( 1 . .

REFERENCES 0.9 -
R. 0. Schmidt, A Signal Subspace Approach to Mul- ~ 0*8 4
tiple Emitter Location andSpectral Estimation, PhD 20.7
thesis, Stanford University, Stanford, CA, November 2 0.6
1981. i\
Ir
G. Bienvenu and L. Kopp, Adaptivity to background
noise spatial coherence for high resolution passive 0.3 2
methods, In IEEEICASSP, pages 307-310, Denver, '
CO, 1980. 1

D. N. Simkins, Multichannel
Angle-of-Arrival Esti- . i . . .
I . . . 3.'. - . ? . . . ' . . . , . . . , . . .
- 3 0- 5 0- 7 0- 9 0 9 0 7 0 5-01 03 0 10
mation, PhDthesis,StanfordUniversity,Stanford, ongl e
CA, 1980.

M. Wax and T. Kailath, Detection of signals byinfor- Figure 2: Histogram of MUSIC Estimates for SNR -3 dB
mation theoretic criteria, IEEE Trans. on Acoustics,
Speech, and Signal Processing, ASSP-33(2):387-392,
April 1985.

35.24. 4
ICASSP 86, TOKYO
1912