IM, TRANSACTIONS ON SIGNAL PROCESSING, VOL 44, NO in, O C T O B ~ K1996 2531

Asymptotic Performance Analysis

of ESPRIT, Higher Order ESPRIT,
and Virtual ESPRIT Algorithms
Norman Yuen and Benjamin Friedlander, Fellow, ZEEE

Abstract-In this paper, we present an asymptotic performance the variance of estimated DOA’s for the ESPRIT algorithm
analysis of three subspace-based methods for direction of arrival using the data matrix.
(DOA) estimation-the ESPRIT algorithm using second-order The performance of ESPRIT has also been analyzed by Rao
statistics, the higher order ESPRIT algorithm using fourth-order
cutrmlants, and the virtual ESPRIT (VESPA) algorithm using and Hari [ 1I], who derived the mean squared error (MSE)
fourth-order cumulants. We examine the least-squares version for estimated DOA’s of a second-order ESPRIT algorithm.
of these algorithms, derive the expressions for the asymptotic They assumed a linear equispaced array and jointly complex
variance of the estimated DOA’s, and use specific examples to Gaussian sources in additive white Gaussian noise. Their
compare the relative performance of the algorithms. Finally, analysis is done for the least-squares version of ESPRIT.
we present Monte Carlo simulations to validate the theoretical
anallysis. However, they show that the asymptotic MSE’s of the LS-
ESPRIT and TLS-ESPRIT are the same. Our work in this
paper follows that of Rao and Hari.
I. INTRODUCTION While the analysis in [ 111 is valid for only the second-order
ESPRIT and for only Gaussian sources, the development in
I N THE last 15 years, subspaced-based methods for es-
timating the directions of arrival (DOA’s) of far-field
sources impinging on an array of sensors have become very
this paper is valid for Gaussian and non-Gaussian sources in
the second-order case. Furthermore, the analysis here encom-
passes the fourth-order ESPRIT [2] and the recently developed
popular. In particular, the MUSIC algorithm and its deriva-
fourth-order VESPA [3] algorithms.
tives have received much attention. Extensive simulations
This paper is organized as follows. Section I1 summa-
and performance analyses have been presented by several
rizes the least-squares version of the second-order ESPRIT,
authors showing the benefits of MUSIC in array processing
fourth-order ESPRIT, and VESPA algorithms. The asymptotic
estimation problems. One disadvantage of these subspaced-
variance of the DOA estimates is then derived in Section 111.
bawd methods, including MUSIC, is that information (e.g.,
The analysis common to all three ESPRIT algorithms is first
gain, phase, relative location) about the array of sensors must
presented and then the analysis for each specific algorithm is
be known a priori. The ESPRIT algorithm, developed by Roy
given. In Section IV, Monte Carlo simulations and numerical
et al. 191, [12], lessens the a priori information requirement
evaluations of the equations derived in the previous section are
by using two identical arrays of sensors.
provided to show the match between analysis and simulation.
The performance of second-order ESPRIT and its derivative
This paper concludes in Section V. Some relevant equations
algorithms has been analyzed by several authors. Ottersten
are derived in the Appendix.
et al. [SI derived the asymptotic distribution of the error in
estimated DOA’s for the total least-squares ESPRIT (TLS-
ESFRIT) and compared the variance to the Cramer-Rao 11. SUMMARY
OF ESPRIT ALGORITHMS
Bound. Mathews and Zoltowski [7] derived the asymptotic In this section, we summarize the least-squares (LS) ES-
variance for the uniform circular array ESPRIT, which is a 2-D PRIT algorithms using second-order statistics and fourth-
angle estimation algorithm. Gavish and Weiss [5] analyzed the order cumulants. To avoid ambiguity, we refer to the second-
virtual interpolated array ESPRIT algorithm, which is based order ESPRIT as ESPRIT2, the higher order ESPRIT as
on interpolated arrays [4]. Li et al. [6] derived expressions for ESPRIT4, and the virtual ESPRIT using fourth-order cu-
mulants as VESPA. The ESPRIT2 and ESPRIT4 algorithms
Manuscript received July 8, 1995; revised April 9, 1996. This work was require two subarrays of sensors that are identical but separated
supported by the National Science Foundation under Grant NSF MIP-90- by a known displacement vector A with magnitude A. The
17221, and in part by the University of California MICRO program and
Applied Signal Technology, Inc., and thc Office of Naval Research under outputs of the subarrays are modeled as
Contracts N00014-9 1-J- 1602 and N00014-95- 1-09 12. The associate editor
coordinating the review of this manuscript and approving it for publication x(t) = A s ( t ) + n,(t) (1)
was 11. A. Lee Swindlehurst.
The authors are with the Department of Electrical and Computer En- Y(L) = A W t ) + n,(t) (2)
gineering, University of Califomia, Davis, CA 95616 USA (e-mail: fried-
Ian @ece.ucdavis.edu). where A is the m x d array response matrix associated with
Publisher Item Identifier S 1053-587X(96)07127-9. each of the (identical) subarrays, s ( t ) is the d x 1 vector of
1053-587X/96$05.00 0 1996 IEEE
2538 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. IO, OCTOBER 1996

narrowband zero-mean signals, n,(t) and n,(t) are m x 1 Decomposing E, into two m x d matrices E, and E,, we get
Gaussian noise vectors of the X and Y subarrays, respectively,
and @ is a d x d diagonal matrix of phase delays between the
doublet sensors for the d signals. The matrix $ is given by
(1 1)
$42 ...
(3) with
... e34d
R{E,) =R{E,)
where =R{A}. (12)
WOA sin ek. Since the range of E,Tand the range of E, are equal, the rank
4% = C
(4) of the matrix [E, E,] is d, which implies that there exists
a 2d x d matrix
tlk is the direction of arrival of the kth source, W O is the carrier
frequency, and c is the speed of propagation. (13)
Now we combine the subarray outputs to get

z(t) = [;I of rank d, which spans the null-space of [E,

O=[E, E,]P
E,]. That is,

= E,P,+EY P Y

where =ATP, + A@TP, (14)

-ATP,Pyl = A$T. (15)

and
We define a matrix
(7)
4
F = -P,P,I
AL. Second-Order ESPRIT From (11), (15), and (16), we see that
The least-squares (LS) ESPRIT algorithm using second-
order statistics is presented by Roy and Kailath [12]. Forming E,F = E , (17)
tlhe covariance matrix of z ( t ) , we get or

-
=A R , , A ~ + g2xn (8)
where E: denotes the pseudoinverse of E,. Combining (15)
and (16), we get
where R,, is the source covariance matrix. The 2m - d ATF = A+T. (19)
smallest generalized eigenvalues of (Rzz,E,) are equal to
0'. The d generalized eigenvectors corresponding to the d Since T is invertible and A has full rank, it follows from
largest generalized eigenvalues are used to form ( 19) that

E, =C,[el e2 . . . ed] = TFT-~. (20)

This says that F and @ are similar and, hence, they have
the same eigenvalues, which are the diagonal elements of @.
Because C, must be a known matrix, we simply assume that Using (3) and (4), the DOA's can be computed from the
X"=,= I in this paper. This assumption also implies that the
eigenvalues of F.
sensors are not shared between the two subarrays, in which
case E:, would not be diagonal, even if the noise is white. The
B. Fourth-Order ESPRIT
case of overlapping subarrays is examined in [ 8 ] . However, for
fourth-order ESPRIT and VESPA algorithms, the arrays can In recent years, much attention has been given to signal
be overlapping and the noise can be colored, as long as it is processing techniques using higher order statistics. One of
Gaussian. We shall call S I . . . . , sd the signal eigenvectors. The the main motivations is that a higher order (>a) cumulant is
range of E,5is equal to that of [12] (i.e., R{E,} = R{A}).insensitive to additive Gaussian noise regardless of whether
Thus, there exists a nonsingular matrix T such that it is white or colored. Thus, it is not necessary to model
the noise when it is additive Gaussian. For signal processing
E, = AT. (10) using higher order statistics, the model used in (1) and (2)
YUEN AND FRIEDLANDER: ESPRIT ALGORITHMS 2519

has an extra requirement that the signals in s ( t ) are non- In a manner similar to the second-order ESPRIT, we can
Gaussian. Because the signals are assumed to have zero mean, formulate a combined output as in (5)-(7). Doing this, we can
the fourth-order cumulant is given by get the fourth-order cumulant matrix
&4(kl,
A
-
k2;l1, l 2 )
c,, = -
ArP
- Cum rq, zlc, , .T2, Zk21

- E{z,*l%1}E{~,*2Zn,}
= E{z,*lzZIilzf2ZkL}
- E { z , * l ~ k 2 } ~ { 4 2 ~ k-l lE { Z k , Z k 2 } E { Z , * l ~ ; 2 ) . (21)
From this point, the fourth-order ESPRIT algorithm proceeds
We shall use the notation ~ ( k . 1 k.2; , 11, I z ) , where as in the second-order case, with the covariance matrix RZz
k l , k2. 11. 12 are the subscripts of the array sensor elements, replaced by the cumulant matrix C,, in (8).
to denote the fourth-order cumulant whenever there is no
conl’usion about which sensor element (of which subarray) C. Virtual ESPRIT
is being referred.
A fourth-order ESPRIT algorithm has been proposed by Recently, Dogan and Mendel [3] have given a novel inter-
Chizing and Nikias [2] for the case of equispaced sensors and pretation to cumulants for narrowband array processing. Basi-
statistically independent sources. The steering matrix is given cally, they have shown that, for independent sources, a fourth-

1
order cross-cumulant can be interpreted as a (second-order)
by
cross-covariance. The interesting insight of this interpretation
1 1 ...
is that for some cumulants, the interpreted cross-covariance is
A= “’w’ eJW2 ... plWd

. (22)
a “virtual” cross-covariance. That is, some cross-covariances
are between data received at real sensors and “virtual” sensors
pl(m-1) p 2 ( 7 7 - 1 ) . .i . pd(m-1)
1 1 (i.e., sensors that don’t physically exist). Dogan and Mendel
have applied the idea of virtual sensors to several array
In the second-order ESPRIT algorithm, it is not necessary to
processing problems, such as implementing a virtual ESPRIT
know the form of the steering matrix, only that there are two
algorithm (VESPA) and suppressing non-Gaussian additive
identical arrays and that the displacement vector A is known.
noise. Here, we give a brief summary of the interpretation.
However, for the fourth-order ESPRIT algorithm, when the
See [3] for more details.
cteeiing matrix is of the form (22), then the cumulant matrices
Single Source: First assume that we have ,a single (;on-
(23) and (24) (shown at the bottom of the page) have the forms
Gaussian) source with a propagation vector k, where k =
C,, =AI’AH (25) A-&, + k&, (a’, and SY are unit vectors along the z and y
C,, =AraHAH (26) axis, respectively), with variance af and kurtosis ys.
Referring to Fig. 1, the shaded circles denote real sensors
and the clear circles denote a virtual sensor. Let the reference
doublet be and 211 (both real sensors), then the signals
received at the real sensors are, for example

where yi is the kurtosis of the ith source. In the above

equations, c(zT, xj,:I$, :cl) represents the cumulant of the
arguments 2 : ; xj,xi,51.The matrix C,, in (23) is not iden-
tical to the one proposed in [2], although they are equivalent and at the virtual sensor
[i.e., they both have the equality in (25)]. In the case of finite
data, the sample estimate of the matrix in (23) is Hermitian
while the one proposed in [2] is not.
 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 10, OCTOBER 1996

/ and v(t) is

where a;j is the response of the ith sensor to the jth source.
Now consider the fourth-order cross-cumulant

//
Fig. I. Virtual sensor interpretation of higher order cumulants. The shaded .;l.T(t) exp [ j ( G ' &k)]
circles represent real sensors and the clear circles represent virtual sensors.
The source approaches the anay with propagation vector 1;. + ak;(t) exp [ j ( G 2 . ,I)&
UllSl(t) exp [-j(21 . I),&
where a; is the response (i.e., gain and phase) of the ith sensor. + a l 2 s z ( t ) exp [-&2 . &,)I
The noiseless cross-covariance between T , ( t )and vl ( t ) is = lall12aElallr.sl exp [-.j(G1 . A)]
exp [-&
. CXP [j(h L)] . &,)I
E{r;(t)w/(t)}= E{nZs*(t)cxp [.j(G & k ) ] '

-t /a12I2&al2Ys2 exp [-j(G2 . A)]

. a,s(t) exp [-.j(L.4,111
= ai,,,: exp [ j ( G . Z l k ) l exp [-j(G.A)] . CXP [ j 6 2 ' &k)l exp 1 4 2' &,)I. (35)

. exp [ - j ( G . dll)]. (31) Comparing (34) and ( 3 3 , we see that the cumulant and
covariance are no longer proportional as is the case in (33).
Now consider the fourth-order cross-cumulant However, we can still intemret (35) as a covariance with
I ~,

sources from the same directions. The only difference is that

Cum [TT ( t )
SUI ( t ) , ( t ) ,T I ( t ) ]
the sources have different "power" levels. However, the DOA
1 cum{aTs*(t); a l s ( t ) exp [ - j ( G . A)], information is not altered. Note that ys,, being a cumulant,
n i s * ( t ) exp [j(g. a,,)], als(t) exp [-j(G.a,,)]} may be negative, whereas a true power is always positive.
= I a l / 2 a i n l y , exp [ j ( G . iil,)]
This case is representative of the case of multiple sources
because of the cumulant property that if the random vari-
s exp [ - j ( G . A)]exp [-j(L.21,)l (32) ables x ~2 ,2 , . . . , x, are independent of the random variables
Comparing (31) and (32), we have y l i y 2 , . . . , yT1, then

This means that for several sources, the cumulant of the sum
'This equation shows that a fourth-order cumulant of the output of sources is simply the sum of the cumulant of each source.
from real physical sensors is proportional to a cross-covariance Given an array of sensors and one additional sensor, we
of output from a real and a virtual sensor. In other words, we can have, using this interpretation, a virtual copy of the first
have an interpreted array that has a larger aperture than the array, allowing us to implement the VESPA algorithm. The
physical array. additional sensor has identical characteristics to the reference
Multiple Sources: Assume now that we have two statisti- sensor and the pair forms a doublet. For example, if we have an
cally independent (non-Gaussian) sources impinging on an array of four sensors and one additional sensor (five total), the
array. In this case, the noiseless cross-covariance between ~ ( t )cumulant matrix used to implement the least-squares VESPA
Y U t N AND FRIEDLANDER ESPRIT ALGORITHMS 2541

in this case would be Furthermore, the left and right eigenvectors are orthogonal
[13] and we can normalize the left eigenvectors so that
(37)

where we obtain (38) and (39) (shown at the bottom of the Equations (40) and (41) imply that
page), where c ( Z , xj,xi, 21) represents the cumulant of
the arguments z:, x j , xi, 21. It is shown in [3] that, with A j = q;Fvi. (42)
five physical sensors, VESPA can be implemented with two
Under most circumstances, the matrix F has to be estimated
subiirrays of five sensors in each subarray instead of only four
using finite data, so that, from (18)
sensors in each subarray. In this case, V,, and V,, would
be 5 x 5 matrices instead of the 4 x 4 matrices shown in (38)
F = E$Ey. (43)
and (39). This larger virtual subarray is used for the examples
in Section IV. The VESPA algorithm now proceeds as in the An error SF = F - F in estimating F will cause an error
second-order case, with R,, replaced by V,, in (8). SA; = & - A,. As a first-order approximation, the error SA
From a different standpoint, we can view VESPA as using can be written as
the available cumulants in a more efficient way. For instance,
in an m-element array of sensors, there are m4 possible SA; N q;SFV;. (44)
cumlulants. ESPRIT4 requires two rn-element arrays, which
means there are 16m4possible cumulants. ESPRIT4 uses only It follows from (18) that we have the first-order approxima-
4m" out of 1 6 7 ~ 1possible
~ cumulants. VESPA, on the other tions
hand, uses 4m2 out of m4 possible cumulants.
(E, + SE,)(F + SF) = E, + SE ?I> (45)
E,SF z bE, - SE,F (46)
I [I. ASYMPTOTIC
PERFORMANCE
ANALYSIS
OF ESPRIT
and
In this section, we derive the asymptotic variance of the 6F zE$SE, - EFGE,F. (47)
estimated DOA's. The derivation in this paper is along the
line:; of the first-order analysis done by Rao and Hari [ 111. In Using (47) in (44) and noting that I A , L / =
~ 1, we get
that paper, Rao and Hari derived the asymptotic variance of the
estimated DOA's for the second-order LS-ESPRIT algorithm SA, RZ q;E,#(SE,v;- SE,FV;)
under the assumption of Gaussian signals. Furthermore, they = q;E,#(SE,v; - A;SE,v;)
show that the asymptotic variance of LS-ESPRIT is equal to
that of the total LS-ESPRIT algorithm, whether the signals = qzE$(W2 - ALW1)~ESVZ
are Gaussian or not. The assumption of this paper is that = - Aiq;E,#(Wl - ATW~)SE,?V; (48)
the signals may be non-Gaussian and, hence, the variance
expressions of [ l l ] in general do not apply. However, the where
methodology of [ 111 is followed here.
In LS-ESPRIT, we compute a matrix F using (18) and then w1 = [I, On,]; (49)
perform an eigendecomposition to get its associated eigenval- wz = [ O m In,] (50)
ues. Let A, be an eigenvalue of F, v, be the corresponding
eigenvector, and q, be the corresponding left eigenvector, such and
that
= Es- E,s. (51)

The matrix E, is defined in (9), I, is the m x vi identity

matrix, and 0 , is the m x m zero matrix. In the following
2542 IEEE TRANSACTIONS ON SIGNAL PROCESSING. VOL 44, NO. 10, OCTOBER 1996

ainalysis, we use the "=" sign even though many of the ex-
pressions are simply approximations. From (48), the variance
of A, 1s
This expression is common to all three algorithms consid-
E{/6X,/2}=q7E,#(W1- X:W~)E{SE,V,V,NSE$}
ered in this paper. The difference among the three ESPRIT
. (W, - ~:W2)H(q7E$)H algorithms is in the statistic used for estimation. ESPRIT2
=q,E$(W1 - X,*W,) uses a covariance matrix Rz2;ESPRIT4 uses a fourth-order
cumulant matrix C z z ; VESPA uses a different fourth-order
cumulant matrix Vzz.Each of these matrices produces a set
of eigenvectors {SI,. . . , sd} that has different statistical prop-
erties from the eigenvectors produced by the other matrices. In
other words, the covariances of the eigenvectors, E{SsgSsf}
and E{bs,SsZ}, which appear in (52) and (53), are unique
to each algorithm. Therefore, in the following subsections, we
Lg=1 h = l
derive these covariances for each algorithm.
. ( W i - A:W2)H(qtE$)H (52)
A. Covariance of Eigenvectors-ESPRIT2
where 6s, is the gth column of the matrix bE, and v, is
The first-order Taylor series expansion of the eigenvectors
t'he gth element of the vector v,. The quantities sil and v, h
(corresponding to distinct eigenvalues) of a 2m x 2m matrix
are similarly defined.
Rzz [14] is
Following the above steps in an analogous manner, we can
ad50 show that the unconjugated variance is

E{(SAz)2} = q,E$(W2 - ALW1)

1

Lg=1 R=1

' (W2 - ~iW1)T(qiE$)T. (53)

For DOA estimation, the quantity of interest is Qi, which where a, is an eigenvalue of the matrix R,, and s ~ , is~ the ,
is related to A; by a;th element of the eigenvector s g . The notation ( .)mn refers to
A . - ,j27r(A/L) sin 0% the ( m ,n ) element of a matrix. The expansion in (60) assumes
7 - (54) that the matrix (R,, - RZz)is conjugate symmetric and that
The separation between the two subarrays is A and the the eigenvectors S I , sa, . . , szm are orthonormal. That is
3

wavelength of the impinging wavefronts is L. Using a first-

1 i=j
order 'Taylor series expansion, we have si sj =
{o i#j.
SA,
soi = (55)
a
j27r - (cos Bi)Xi
When s, corresponds to a simple eigenvalue, the eigenvector
' L s , is uniquely determined up to a complex constant. When
The DOA 8,is a real quantity, but (55) is not necessarily real. there are repeated eigenvalues, the corresponding eigenvectors
'Therefore, we let are not uniquely determined. This will be the case for the
eigenvectors s d + l , . . , szm, which correspond to a single
so, = ;(so; + SO,*). (56) eigenvalue of multiplicity 2m - d. Also, all the eigenvectors
Skipping some algebraic steps, we get have to be normalized to have unity norm. It follows from (60)
that the covariance matrix of the signal eigenvectors Gg and G,,
is, to the first order, shown in (62), at the bottom of the next
page. We note that this expression is greatly simplified when
the signals are Gaussian, as is assumed in [ 111. The covariance
of sample covariances, needed to compute (62) is derived in
and
Appendix B, (88), and repeated here for convenience.

- real{E{(SA,)2}(A5)2}]. (58)
If B, is given in degrees instead of radians, then
r The second-order and fourth-order moments, denoted by p2 ( )
and pug( ), respectively, of the received data z ( t ) are given by
YUI:N AND FRIEDLANDER ESPRIT ALGORITHMS 2543

to the row and column of a matrix, do not in general have

a direct mapping to the indices ( k l , kz, 11, 1 2 ) . The exact
mapping depends on which cumulants are used to form the
matrix C,, and how the cumulants are ordered within Czz.
The same comment applies to ( b l , ha) and ( m ~m2, , 721, n2).
Furthermore, the unconjugated covariance of the sample
eigenvectors sg and S t l , to the first-order, is
E{Gs,Ss;f} =
2rn 2m 2m 2m 2m 2171

S ? ~ a I S g , n 2 S ~ . b l S hbz
Ifs n f h ni=l n a = l b l = l by=l
1=1 n=l

' q ( C z , ?- c , z ) n l n 2 ( c z z- C z z ) b 1 b 2 }
Because of the inherent symmetry in the covariance matrix szsnT
R,,, (63) can be used to compute E{(R,, - R,Z)alu2(Rzz - (69)
Rzz)ljlb2}by swapping the indices bl and b2.
(ay - a l ) ( a h- "1 '

Because of the inherent symmetry in the fourth-order

B. Covariance of Eigenvectors-ESPRIT4 cumulant, (81) can also be used to compute E { ( C z z -
Czz)~ln,(Czz-C,,)61b2} by swapping the indices (mi, m2)
Since (60) is valid for eigenvectors of any Hermitian matrix, with (nl, n2).
we can use it as an estimate of the eigenvectors of a cumulant
malrix. We simply replace R,, with C,, and R,, with C,,. C. Covariance of Eigenvectors-VESPA
The, covariance matrix of the signal eigenvectors sy and sl,
for a cumulant matrix, to the first order, is shown in (67) at From a mathematical point of view, there is very little
the bottom of the next page. The sample cumulant of (21) is difference between the performance equations used for the
given by ESPRIT4 and VESPA algorithms. The only difference is
that the cumulants used for the two algorithms are different.
b ( h >k 2 : [ I , l 2 ) = ( c z z ) a I ~ , , Therefore, the covariance of sample cumulants
W(Czz - C,,),, U2 ( C Z Z - CzZ)X,ba 1 (70)
in (67)-and the unconjugated counterpart in (69)-will in-
volve different cumulants, which are given in the matrix
V,, in (37). Therefore, the conjugated and unconjugated
covariances of the eigenvectors for VESPA are given by

E { 6s, ssf }

(68)
and its asymptotic covariance is given in (81) of the Appendix.
That is, (81) is used to compute E{(C,, - C,z)nlc,2(Czt-
C-IL);),b2}in (67). Note that the indices (al.a ~ )which
, refer
2544 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. IO, OCTOBER 1996

subarrays. Each subarray has four omnidirectional sensors.

Each sensor is spaced 0.5 wavelengths apart and the subarrays
are spaced 0.5 wavelengths apart. The subarrays are depicted
in Fig. 2, where the angle-of-arrival of a single source is
shown. As explained in [ 3 ] , when given an M element
subarray and one guiding sensor, the resulting subarrays used
for VESPA have M + 1 elements. The steering vector for
subarray 1 (excluding rs) has the form

Note that m for VESPA is the total number of sensors and

is one more than m for ESPRIT4 or ESPRIT2, where m,
is the number of sensors in each subarray. We note that the
covariances in (71) and (72) can still be computed using (81).
From ti practical point of view, the advantage of VESPA is
The ESPRIT2, ESPRIT4, and VESPA algorithms are used
that the cumulants can be computed from data collected from
for DOA estimation. There are a total of eight sensors for
fewer sensors.
ESPRIT4 and 13SPRIT2. We use two versions of the VESPA
algorithms, which we denote by VESPA5 and VESPAS. For
VESPA5, we use a four-element linear array and one more
I v . SIMULATION AND NUMERICAL
EVALUATION sensor as the guiding sensor (i.e., five sensors total). To make
In this section, we examine several cases of interest and a comparison of the algorithms using the same number of real
evaluate the asymptotic results derived in the previous sec- sensors, we use the array shown in Fig. 3 for VESPA8. The
tions. In addition, we use Monte Carlo simulations to verify DOA of source 2 is fixed at 40" while the DOA of source
the theoretical results for these cases. In all cases, we assume 1 is varied from 0" to 180". The signals are corrupted by
that the number of sources is perfectly known. additive white Gaussian noise with input signal-to-noise ratio
We note that the theoretical results for ESPRIT4 and VESPA (ISNR) of 10 dB. The results are shown in Figs. 4 and 5.
in Section 111 are valid for spatially colored (Gaussian) noise, The lines are theoretical values while the discrete points are
while the result for ESPRIT2 is valid for spatially white noise simulation values. We use 500 independent runs with 5000
only. It is straightforward to extend the results of ESPRIT2 snapshots per run. The simulation and theoretical results show
to the colored noise case where the covariance matrix of the good agreement. Notice, in the figures, the standard deviation
nioise 11s known. However, this is not done in the previous increases as the DOA of source I approaches the DOA
analysiis because ESPRIT4 and VESPA do not require the of source 2 (i.e., 40"). Furthermore, the standard deviation
noise covariance matrix to be known. To make a reasonable for source 1 increases as source 1 approaches 90'. In the
comparison of the three algorithms, we will use in all the region of small DOA separation between the two sources,
examples a white Gaussian noise scenario. Furthermore, as ESPRIT2 and VESPA5 are comparable in performance. It
noted in Section 11-A,the analysis presented for ESPRIT2 must is somewhat surprising that in this region, VESPA8 has
I-9e modified when overlapping subarrays are used. However, slightly worse performance than VESPAS. However, because
the analysis for ESPRIT4 and VESPA are valid even when the the two versions of VESPA and ESPRIT4 are implemented
subarrays are overlapping, given that the noise is Gaussian. with different cumulants, the results demonstrate that some
Example I-Standard Deviation Versus DOA Separation: In cumulants may provide more information. For larger DOA
this example, we let two equal power statistically independent separation or when the DOA is near 90°, ESPRIT2 performs
4-QAM sources impinge on two uniformly spaced linear the best.
YU EN AND FRIEDLANDER: ESPRIT ALGORITHMS 2545

\%SPA5
Subarray 1
r2 3 0 v2

WSPA5
Subarray 2
VEsPA8
rT
Subarray 1
r2

r3 a
"\ v5

VESPAB
Subarray 2

r3 0 0 V3

r4 0 0 v4
0
vs 1
Fig. 3 . Subarrays of sensors used in Examples 1 to 2 for VESPA8. Subarray
1 consists of rl -rs while subatray 2 consists of VI-VK. The shaded circles
represent real sensors while the clear circles represent virtual sensors. There
is an overlap of four real sensors in subarray 2. A source approaches the
array at an angle 8 .

Fig. 2. Subarrays of sensors used in Examples 1 and 2. For ESPRIT2 and

ESPRIT4, subarray 1 consists of rl-rq while subarray 2 consists of V I - V , I
Std Dev of Estimated DOA for Source 1 - Linear Array
and all circles represent real sensors except for v:,, which is a virtual sensor
used only in VESPAS. For VESPAS, subarray 1 consists of rl-rs while
subarray 2 consists of v1-v~ and the shaded circles reprcsent real sensors
while the clear. circles represent virtual sensors. A source approaches the
array at an angle 0.
ESPRIT2, dot / xxx

Example 2-Standard Deviation Versus ISNR: We repeat

the experiment using the parameters in Example I , except that
we vary the input SNR level, and the simulations are done with
500 snapshots. The DOA's of the two sources are fixed. Source
1 impinges on the array at 20" and source 2 impinges on the
array at 40". The results for source 1 are shown in Fig. 6. The
results for source 2 look very similar and, hence, we have not
included them in this paper. The simulation and theoretical
results show good agreement, even at 500 snapshots. In this
example, the difference in performance between VESPAS and
VESPA8 is more clearly shown. As in Example I, ESPRIT2 0 20 40 60 80 100 120 140 160 180
DOA of Source 1 (Degrees)
again has the best performance.
Example 3-Standard Deviation Versus DOA Separation Fig 4 Results for Example 1. The standard deviation of estimated DOA for
for T-Array: In this example, we apply the same signal bource 1, plotted as a function of the DOA of source 1 for two linear subarrays
with four sensors each. The DOA of source 2 is hxed at 40". The dash and
parameters as Example 1 to an array used in [ 3 ] . The ar- solid lines are the theoretically predicted results and the discrete points are
ray configuration is shown in Fig. 7. We shall refer to this simulation iesults Five thousand snapshots are used.
con figuration as the (inverted) T-array. Furthermore, to make
a comparison for VESPA using the same number of sensors
as ESPRIT2, we also use the configuration in Fig. 8. As in The response of each sensor takes the form - cos (e)
previous examples, we call the resulting algorithm VESPA8. and if it is rotated clockwise by Q, then its response
Because the ESPRIT4 algorithm in [2] is formulated only foi is cos(8 +
a ) . The sensors used in this example
linear arrays, we compute the results only for the ESPRIT2, have orientations { 0" , -5", 3" -loo} with locations
VESPA5, and VESPA8 algorithms. ((0, 01, (0, -X/2), -A), (A/2, -A)) on the 2-Y
2546 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 10, OCTOBER 1996

Std Dev of Estimated DOA for Source 2 - Linear Array

ESPRIT4: dash / +++

VESPA5: solid I '*'

i
VESPAB: dash-dot IO00
ESPRITP: dot ixxx
\
Subarr Subarray 2

0
'
IY x
20 40 60 80 100 120 140 160
I
180
DOA of Source I (Degrees)

Fig. S . Results Tor Example I. The standard deviation of estimated DOA for Subarr
source 2, plotted as a function of the DOA of source 1 for two linear subarrays
with four sensors each. The DOA of source 2 is fixed at 40'. The dash and
solid lines are the theoretically predicted results and the discrete points are
simulation results. Fivc thousand snapshots are used.

Std Deviation for Estimated DOA of Source 1 -Linear Array

+
4r
1

(b)
Fig. 7. Array configuration used i n [ 3 ] . Configuration in (a) is used for
VESPAS and (b) i s used for ESPRIT2. The shaded triangles and X's represent
real sensors, and the clear triangles represent virtual sensors.

Subarray 1
/ I . / . \

Fig. 6. Results for Example 2. The standard deviation o f estimated DOA for
source I , plotted as a function of the input SNR. Source 1 and 2 impinge on
the array at 20" and 40'. respectively. Two linear arrays with four sensors
each. SO0 snapshots.
Fig. 8. Array configuration used for VESPAS in [ 3 ] .Five sensor elements
plane. Thus, a steering vector has the form' arc common to both subarrays. The clear triangles represent virtual sensors.

r COS (e) 1 Example 4-Standard Deviation Versus ISNR for T-Array:

We repeat the experiment using the parameters in Example 3,
except that we vary the input SNR level and the simulations
are done with 500 snapshots. The DOA's of the two sources
(The frame of reference for 0 used in this paper is not the are fixed. Source 1 impinges on the array at 20" and source
same as in [ 3 ] ,so that the sensor response looks analytically 2 impinges on the array at 40". The results for source 1 are
different from that given in [3].) We plot the standard deviation shown in Fig. 11. The results for source 2 look very similar
for both sources in Figs. 9 and 10. The dash and solid lines and, hence, we have not included them in this paper. The dash
are the theoretically predicted results and the discrete points and solid lines are the theoretically predicted results and the
are simulation results. discrete points are simulation results.
It is clear that in these examples, where the noise is additive
'Equation (74) is for ESPRIT2 only. The steering vector has to be modified
white Gaussian and the are 4-QAM, the
slightly for VESPAS and VESPA8 to reflect the larger number of sensors for
each subarray. order ESPRIT algorithm performs the best among the three
YUEN AND FRIEDLANDER: ESPRIT ALGORITHMS 2547

Std Deviationfor Estimated DOA of Source 1, T-Array Std Deviation for Estimated DOA of Source 1, T-Array

-
g14
e!
a
Q12
c
5 1
g
n
$08
V

go6

04

02

0 5 10 15 20 25 30
Input SNR (dB)

Fig. 9. Results for Example 3. The theoretical standard deviation of esti- Fig. 11. Results for Example 4. The standard deviation of estimated DOA
maled DOA for source 1, plotted as a function of the DOA of source 1. The for source 1, plotted as a function of the input SNR for the inverted T-array
DCIA of source 2 is fixed at 40’. ESPRIT2 uses two inverted T-arrays with configuration. The DOA of source 1 and 2 are fixed at 20’ and 40°,
four sensors each. VESPAS uses five real sensors and VESPAI uses eight rcal respectively. The dash and solid lines are the theorctically predicted rcsults
sensors. The dash and solid lines are the theoretically predicted results and and the discrete points are simulation results. Five hundred snapshots are used.
the discrete points are simulation results. Five thousand snapshots are used.
performance of an algorithm. However, it is unclear which
Std Deviation for Estimated DOA of Source 2, T-Array set of cumulants would yield the best performance.
2.5
I It is noteworthy that the simulation results in [3] show that
1 the performance of VESPA exceeds that of ESPRIT2 for the
/ I
case of BPSK signals in white Gaussian noise. It is possible
that BPSK signals, which takes on fewer possible values than
a 4-QAM signal, can be viewed as being “less Gaussian” than
VESPAB: solid I )(yx 4-QAM signals. Hence, BPSK signals may yield better results
VESPA8: dot I ooo for the higher order techniques, which require non-Gaussian
ESPRIT2 dash / *’* signals.

V. CONCLUSION
In this paper, we derived expressions for the asymptotic
variance of DOA estimates for three LS-ESPRIT algorithms
under a non-Gaussian signal assumption. The formulas derived
are used to evaluate the performance of the second-order
ESPRIT algorithm, the fourth-order ESPRIT algorithm, and
the VESPA algorithm. Monte Carlo simulations are presented
Fig. IO. Results for Example 3. The theoretical standard deviation of esti-
to verify the theoretically predicted performance. From the
mated DOA for source 2, plotted as a function of the DOA of source I . The examples we used, where the additive noise is white Gauss-
DOA of source 2 is fixed at 40’. ESPRIT2 uses two inverted T-arrays with ian, it appears that the second-order ESPRIT has the best
four sensors each. VESPAS uses five real sensors and VESPAR uses eight real
sen:,ors. The dash and solid lines are the theoretically predicted results and
performance.
the discrete points are simulation results. Five thousand snapshots are used.
APPENDIXA
COVARIANCE OF FOURTH-ORDER
SAMPLE CUMULANTS
algorithms. We emphasize that when the noise is colored with In this appendix, we derive approximate expressions for the
unknown covariance matrix, it is better to use the higher covariance of fourth-order sample cumulants. The expressions
order methods. It is interesting that the VESPAS algorithm for zero-mean and symmetrically distributed signals are de-
performs just as well as ESPRIT2 and better than ESPRIT4 rived in [IO]. We extend that derivation to zero-mean and
in 1he region of small DOA separation, even though the latter possibly nonsymmetrically distributed signals.
algorithms require more sensors. While the performance of The fourth-order cumulant and sample cumulant are given
VESP.48 in most cases is better than VESPAS, VESPA8 still by
trails ESPRIT2. Because ESPRIT4, VESPAS, and VESPA8
use different cumulants to form their respective cumulant 64(kl? k2; 11, l 2 )
matrices, the difference in performance suggests that using = I*.4(k1? k 2 ; 11, l 2 ) - VZ(k1; ll)!-LZ(k2; 12)
a different set of cumulants will improve or degrade the - P 2 ( h ; h ) / L z ( k z ; 11) - P2(h, b ) P 2 ( h , l2) (75)
2548 IEEE TKANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. in, OCTOBER 1996

2 2

i=l 7=1

n n

Hence

n n

To complete the derivation, we need the expressions for the

covariance of even-order moments. However, for the sake
of continuity, the covariance of even-order moments will be
derived in the next section. Substituting (87) into (80), we get
YUEN AND FRIEDLANDER: ESPRIT ALGORITHMS 2549

2=1 J=1 p=l q=l

.m(m3-,; n 3 - q ) ~ 4 ( kmp;
, I,, n4)
- [P4(kl, k 2 ; 11, 1 2 ) - 2P2(kl; / 1 ) p 2 ( h ; l 2 )
- 2Pz(h; / 2 ) P z ( h ; 1111
. [1*.4(ml,m2; n l , n ~-)2 ~ 2 ( m 1n; 1 ) ~ 2 ( m 2.z);
- 2 P 2 ( m ; nz)CLz(m2; n1)] (83)
which is the equation derived in [lo]. Given the true model
of the measurements, all the moments can be computed. A
method for computing the moments for statistically indepen-
dent QAM signals in spatially white Gaussian noise is given
in [lo]. Alternatively, these moments can be estimated when
only measurements are given.

APPENDIXB
p=l q=l
COVARIANCE OF EVEN-ORDER
MOMENTS
' [P4(kl, k2, m p ; n q ) We assume that the received signals have zero first-order
moments. We also assume that the signals emitted at dif-
- P2(kl> k2)P2(mp; %)I ferent times by the same source or by different sources are
f p2(llr /2)pZ(ml, m2)[p4(k1, k2; 121, n2)
statistically independent. The even-order moments are given
- P2(h, kz)P2(nlr m)] by
+ P2(11, l2)p2(nl, n2)[P4(kl, k 2 , ml, m2)

- b2(k1, k2)I12(m1, ma)]. (81)

The. indices (kl, k z , 11, Z2) of the first sample cumulant refer (84)
to the elements in the received data vector z ( i ) . Likewise, the
indices ( m l ,m2, n l , n2) also refer to the elements of ~ ( t ) . ICLzn(k1, '", ku; Zl, '", Ib)
For most QAM signals, the symbols are symmetrically
distributed in such a way that the fourth-order cumulant is *} (85)
given by

&4(k1,k2; ll, k?)=p4(kl, k2; ll, 12)

- p2(k1; 11)Pz(k2; l 2 )
- p2(k1; la)p2(k2; Zl). (82)
That is, the unconjugated second-order moments are zero.
+
where the sum of the subscripts is a b = 2n and n is an
integer. The covariance of even order sample moments is, in
P2(kl, k 2 ) =E{xnlzn,} general
=O E{Spz,(h, '", CI ;, 11, . . . , Lb)
2550 IEEl TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 10, OCTOBER 1996

[3] M. C. Dogan and J. M. Mendel, “Applications of cumulants to array

processing-Part I: Aperture extension and array calibration,” IEEE
b 1* i y , vol. 43, no. 5 , pp. 1200-1216, May 1995.
[4] B. Friedlander, “Direction finding with an interpolated array,” in Proc.
ICASSP, Albuquerque, NM, Apr. 1990, pp. 295 1-2954.
[5] M. Gavish and A. J. Weiss, “Performance analysis of the VIA-ESPRIT
algorithm,” in IEE Proc. F (Rudur Signal Processing), vol. 140, pp.
123-128. Apr. 1993.
[h] F. Li, R. J. Vaccaro, and D. W. Tufts, “Performance analysis of the state-
space realization (TAM) and ESPRIT algorithms for DOA estimation,”
IEEE Trans. Anfenna.s Propugat., vol. 39, pp. 418-423, Mar. 1991.
[7]C. P. Mathews and M. D. Zoltowski, “Performance analysis of the
UCA-ESPRIT algorithm for circular ring arrays,” IEEE Truns. Signal
Processing, vol. 42, no. 9, pp. 2535-2539, Sept. 1994.
[8] B. Ottersten, M.Viberg, and T. Kailath, “Performance analysis of the
total least squares ESPRIT algorithm,” IEEE Trans. Signal Processing,
vol. 39, pp. 1122-1 135, May 1991.
[9] A. Paulraj, R. Roy, and T. Kailath, “A subspace rotation approach to
signal parameter estimation,” Proc. IEEE, vol. 74, no. 7, pp. 1044-1045,
July 1986.
[IO] B. Porat and B. Friedlander, “Direction finding algorithms based on
high-order statistics,’’ IEEE Truns. Signal Proce,ving, vol. 39, no. 9, pp.
2016-2024, Sept. 1991.
[ I l l B. D. Rao and K. V. S. Hari, “Performance analysis of ESPRIT and
TAM in determining the direction of arrival of plane waves in noise,”
IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, no. 12, pp.
1990-1995, Dec. 1989.
[I21 R. Roy and T. Kailath, “ESPRIT-Estimation of signal parameters via
rotational invariance techniques,” IEEE Trans. Acoust., Speech, Signal
Processing, vol. 37, no. 7, pp. 984-995, July 1989.
[I31 R. E. Skelton and D. A. Wagie, “Minimal root sensitivity in linear
systems,” J . Guidance, vol. 7, no. 5,pp. 570-574, Sept. 1984.
[ 141 J. H.Wilkinson, The Algebraic Eigenvalue Pmblem. Oxford: Claren-
don, 1965.

Norman Yuen grew up in Oakland, CA He re-

ceived the B S and M S degrees m electrical engi-
neering from the University of Califomia, Berkeley,
CA, in 1982 dnd 1986, respectively He received
the Ph D degree 111 electrical engineering from the
University of California, Davis (UCD), in 1995.
In betwecn academic degrees, he has worked as
an engineer at National Semiconductor, Elcctronic
Support Systems, and Sun Microsystems As a re
search assiqtant at UCD, he actively studicd array
processing methods using higher order stati3tics
He is currently working as a portdoctoral researcher at UCD His research
interests include rensor array processing, application of higher order stati?tics,
blind signal yeparation, channel equali7ation, and performance analysis of
andy processing methods
.p2n(m1, . . . , me;n l ; ’ ” , m) (87) Dr Yuen is a member of Tau Beta Pi and Eta Kappa Nu

where n + b = 2a and c + d = 2p. In particular, the covariance

of samlple covariances (i.e., a = p = 1) is given by
Benjamin Friedlander (S’74-M’76-SM’82-F’87)
received the B S c . and M.Sc. degrees in electrical
engineering from Technion-Israel Institute of Tech-
nology in 1968 and 1972, respectively. He received
the Ph.D. degree in electrical engineering and the
Recall that we have assumed zero-mean processes in this M.Sc. degree in statistics from Stanford University
paper, so that the second order moment ,Q(.) is equal to its in 1976.
From 1976 to 1985, he was with Systems Control
corresponding covariance. Technology, Inc., Palo Alto, CA. From November
1985 to July 1988, he was with Saxpy Computer
REFERENCES Comoration. Sunnvvale, CA. Currentlv, be is with
the University of Califomia, Davis, CA, and Signal Processing Technology,
[I] D. R. Brillinger. Time Series: Data Anulysis and Theory. New York: Ltd., Palo Alto.
Holt, Rhinehart and Winston, 1975. Dr. Friedlander is the recipient of the 1983 ASSP Senior Award, the 1985
[2] H.H.Chiang and C. L. Nikias, “The ESPRIT algorithm with higher- Award for the Best Paper of the Year from the European Association for
order statistics,” in Proc. Vail Workshop Higher-Order Spectral Anal., Signal Processing (EURASIP), and the 1989 Technical Achievement Award
Vail, CO, June 1989, pp. 163-168. of the Signal Processing Society.