IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, rio.

9, SEPTEMBER 1992 2353

TABLE VI
PERFORMANCE OF THE PROPOSED REALIZATION
I N FIG. 1 FOR r = M/2 + 1 (C,, = C, = 1, f,, = 1, = 1
TIMEUNIT)

M = N 2 4 6 8 10 14 20 30
r 2 3 4 5 6 8 11 16

N, 27 69 13c 20 1 29 1 519 981 207 1

S,, 16.5 48.5 9C.5 160.5 240.5 448.5 881.5 1920.5
E, 0.611 0.703 C.710 0.799 0.826 0.861 0.899 0.927

TABLE VI1
PERFORMANCE OF DA REALIZATIJN 11 FOR r = M/2 + 1 (C,, = C, = 1, f,, = f, = 1 TIMEUNIT)

M = N 2 4 6 8 10 14 20 30
r 2 3 ’I 5 6 8 11 16

N, 34 93 180 290 438 808 1573 3408

S,, 16.5 48.5 96 5 160.5 240.5 448.5 881.5 1920.5
E, 0.485 0.522 0 53ti 0.553 0.549 0.555 0.560 0.564

izations of the denominator of H ( z I , z 2 ) are performed in the DA decomposition theorem and its implications to the realization of two-
realization I1 and these multiple realizations are absent in the pro- dimensional digital filters,” IEEE Trans. Acoust., Speech, Signal Pro-
posed realization and only vector multiplications are involved. It cessing, vol. 33, pp. 694-711, 1985.
[9] B . P. George and A . N . Venetsanopoulos, “Design of two-dimen-
is necessary to make a detailed investigation of the rolmdoff noise sional digital filters on the basis of quadrantal and octagonal symme-
properties of all the realizations for a proper comparison. try,” Circuits, Sysr., Signal Processing, vol. 3, pp. 59-78, Jan. 1984.

V. CONCLUSION
In this correspondence, an efficient realization of 2-D denomi-
nator-separable digital filters is given based on mirimizing the
throughput delay and maximizing the parallelism and pipelining for
real time processing and multiprocessor implementation using the An Analysis of ESPRIT Under Random Sensor
basic primitive structure of [4]. The proposed realization (referred Uncertainties
to as realization V for a multiprocessor implementation of denom-
V . C. Soon and Y. F. Huang
inator-separable 2-D filters in continuation of the realizations of
[4]) has an as good efficiency as the realization I and 1 V of [4] for
the implementation of symmetric fan filters. Further, the proposed Abstract-In this correspondence, a general expression for the mean-
realization can be advantageously used in the case tha the rank of square error (MSE) of the ESPRIT direction-of-arrival (DOA) esti-
the matrix [ p,,] of the numerator polynomial of the transfer func- mation of narrow-band, far-field sources under random sensor pertur-
tion is less than or equal to half of the order of the transfer function. bations is derived. Explicit solutions for the case of an arbitrary ES-
PRIT array geometry with one and two sources are given. Solutions
It is also suitable for vector processing. for the case of a uniform linear array (ULA) and arbitrary number of
sources are found provided that the number of sensors is large. For
REFERENCES this case, it is found that the MSE of ESPRIT with maximum aperture
(i.e., maximum-overlapping subarrays) is lower than that of ESPRIT
R. C. Gonzalez and P. Wintz, Digital Image Prcressinq. Reading, with nonoverlapping subarrays. A comparison with MUSIC suggests
MA: Addison-Wesley, 1977. that the MSE for MUSIC is lower than that for ESPRIT. Furthermore,
S . G. Tzafestas, Ed., Multidimensional Systems, Techn,qucs and Ap- for the cases studied here, it is shown that the MSE of ESPRIT depends
plications. New York: Marcel Dekker, 1986. only on sensor phase errors while that of MUSIC is dependent on both
H. T. Kung, “Why systolic architectures,” Computer Mag., vol. 15, sensor gain and phase errors.
pp. 37-46, Jan. 1982.
M. Y . Dabbagh and W. E. Alexander, ‘iMultiprocessor implementa-
tion of 2-D denominator-separable digital ,filters for real-time process- I. INTRODUCTION
ing,” IEEE Trans. Acoust., Speech, Sign41 Processing vol. 37, pp.
872-881, June 1989. \
Eigenstructure-based algorithms for array signal processing have
J . H. Kim and W. E. Alexander, “A multiprocessor aichirecture for attracted considerable interest in recent years. Algorithms such as
two dimensional digital filters,” IEEE Trans. Comput., 7.01.C-36, pp. MUSIC [ 11 and ESPRIT [2] have been proposed generally for the
876-884, 1987.
H. C. Andres and C. L. Patterson, “Singular value d6:compositions
and digital image processing,” IEEE Trans. Acoust., Sveech, Signal Manuscript received September 2, 1990; revised July 24, 1991. This
Processing, vol. ASSP-24, pp. 26-53, Feb. 1976. work was supported in part by the Office of Naval Research under Contract
A. N. Venetsanopoulos and B. G . Mertzios, “A decomposition theo- N00014-9 1-J- 1461.
rem and its implications to the design and realization c>f two-dimen- The authors are with the Laboratory for Image and Signal Analysis
sional filters,’’ lEEE Trans. Acoust., Speech, Signal Pr<icessing,vol. (LISA), Department of Electrical Engineering, University of Notre Dame,
ASSP-33, pp. 1562-1574, 1985. Notre Dame, IN 46556.
C. L. Nikias, A. P. Chrysalis, and A. N. Venetsanopou os, “The LU IEEE Log Number 9201603.

1053-587X/92$03.00 0 1992 IEEE

I1
2354 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, NO. 9, SEPTEMBER 1992

DOA estimation problem. Much work on the performance analysis The individual steering vector at DOA angle On is defined as
of these algorithms has been directed to studying the effects of noise
and finite number of snapshots, with no errors assumed on the sen-
sors gaindphases, locations etc., see, e.g., [3]-[8].
[ [
ak = exp -j
2a
7 ( p , sin ek + vI cos e,)
1 ...
This correspondence presents analysis results on the MSE of ES-
PRIT DOA angle estimates with errors in the array smsors. Such
errors include random errors in the sensor gains and phases, ran-
dom errors in sensor locations and random errors in sensor pair Furthermore, Q, = diag {exp ( - j ( 2 a d ) / h ) sin 0 1 ) , . . , exp
alignments. Both MUSIC and ESPRIT perform well mhen the sen- (-j(2ad/h d ) sin e,)}, X is the wavelength of the narrow-band
sors in the array are either calibrated (i.e., the array manifold is signals, s is the source vector and q x , q, denote the additive noise
known) as in MUSIC or the sensors in a sensor pair are matched vectors at the X and Y subarrays, respectively.
perfectly as in ESPRIT. In practice, however, this ma:, not always It is assumed that the matrix A has full column rank, the narrow-
be possible due perhaps to external environmental effects on the band sources are incoherent and the noise vectors are uncorrelated
array, deterioration of electronic components, measurement errors, zero-mean processes with covariance aZI.
etc. As a consequence, there has been interest in the 3erformance Under small array model errors and by first-order approximation,
analysis of these methods under model errors. Analysis of MUSIC a general formulation of the array signal processing problem under
under various model errors can be found in [9]-[13]. Analysis of model errors can be derived, namely,
ESPRIT where the steering matrix errors are modeled as uncorre-
lated, zero-mean Gaussian distributed can be found in [12], [13].
X subarray: x = (TA + aA,)s + qx
In [13], [14], the analysis of ESPRIT under model errors are per- Y subarray: y = (TA + aA,)Q,s + qy. (4)
formed using its formulation as the minimization of a certain cost
function. The MSE expression obtained therein are very compli- The aA,, aA, terms for the various model errors types are as fol-
cated due to its parameterization in terms of not only the DOA lows.
angles but also other "nuisance" parameters [8], sucn as the real Errors in Sensor Gains and Phases: aA, = AT,A and aA, =
and imaginary parts of the steering matrix, the magn tudes of the A T, A where
elements of a diagonal matrix, etc. In this correspondence, a direct AT, = diag { ( A a , + jalA41)e'", . . * , (Aa, + ja,A&,)e'""'}
approach towards the analysis of ESPRIT under moiel errors is
taken where the DOA angles are the sole parameters considered. (5)
This approach yields simple MSE expressions that art: related ex- AT, = diag {(AZ, + j a l A ~ l ) e ' @ .' ,. * , (AZ, + ja,,,A&,)e'~"'}.
plicitly to the model errors. As such, it provides interesting insight
into the performance sensitivity of ESPRIT to model 8:rrors. (6)
Here { ( A a , , A+,)} and {(AZ,, ATl)} denote errors in the sensor
gains and phases of the X and Y subarrays, respectively.
11. PROBLEM
FORMULATION
Errors in Sensor Locations: aA, = TAA, and aA, = TAA, where
Throughout this correspondence, the superscripts d,T, *, and 2a .
+ denote the Hermitian, AAx = - j - (diag { A p l , . * . , Ap,,,}A5
transpose, complex conjugate, and the
Moore-Penrose pseudoinverse of a matrix, respectively. Also, n
x
denotes the number of narrow-band far-field sources, I;Z denotes the + diag { A V , , . . . 1 AVmJA,). (7)
number of sensors within a subarray, mo denotes the :otal number
of sensors (i.e., mo 5 2m) and { p , , vl} and { E , , V , } denote the 2a .
sensor locations in the X and Y (possibly overlapping) subarrays,
AA, = -j- (diag {Ajil, ..* , ALm}A,
h
respectively.
Here, we will be concerned with the situation where the covari- + diag {AVl, . . . , AF,,,}A,). (8)
ance matrix is assumed known (i.e., the finite sample effects on
data covariance matrix are not investigated here). Under nominal Here { c a p i , A V [ ) } and {(Aji,, AT,)} denote errors in the sensor
conditions, the sensor pairs within the X and Y sibarrays are locations of the X and Y subarrays, respectively, A, = [al sin e l ,
matched perfectly, with a distance spacing d between them [2]. By . . . , a, sin e,] and A, = [a, cos e,, . . * , a, cos e,].
appropriate choice of coordinates, one can rewrite the sensor lo- Errors in Sensor Pair Alignments: In this case, the sensor pairs
cations within the X and Y subarrays to be { p l , vI} and { p, + d, within the X and Y subarrays are perturbed by small alignment er-
rors such that they are rotated by small angles from the nominal
vl}, respectively. Then, the observation vectors from X and Y
subarrays can be written as aligned direction. Hence, aA, = TAA, and aA, = FAA, where
ad
X subarray: x = TAs + qx and Y subarray: y = I A h + qs. AA, = -j -
h diag {PI, . . * , P,}A, (9)

ad .
AA, = +j - diag {PI, . . . , &,}A,. (10)
X
Here T denotes the diagonal array of nominal sensm gains and
phases, namely, r = diag { a l e i @ ' , * * . , awei@"'},
m d A is the Here {Pi)denotes the small angle rotations from the normal of the
steering matrix defined by the set of distinct DOA angles {e,} (the sensor pairs due to misalignment errors and A, is as defined pre-
DOA angles are measured with reference to the normiil of the sen- viously.
sor pairs): Remark: Note that the random sensor error assumptions here en-
ter directly into the sensor parameters such as gaindphases, loca-
A = [a, ... a,,]. (2) tions, and alignments as opposed to the uncorrelated Gaussian error
I
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, PIO. 9, SEPTEMBER 1992 2355

assumptions imposed on the steering matrix (i.e., on d A X , aA,,) in where

[12], [I31 which may not be physically justifiable.
U, = [e'@"(A(YI- A a , ) + joc,e'"(h& - A+,)
-
111. ANALYSIS
OF ESPRIT UNDERSMALLARRAY . . . e'"(AE, - Aa,) + ja,e'"(A+, - A+,)lT. (18)
PERTURBATIONS
Let the covariance matrix be R = E { z z H } ,where B. Errors in Sensor Locations
MO, = r D k E { v k u F } ( r D k ) Hand M I , = T D k E { u k u l (rDk)T
}
(19)

and let its eigendecomposition be R = E , A , E t + U ' E , E ; . The where

ESPRIT algorithm solves for Y from = E? and estimates the
DOA angles from its eigenvalues. Here E, and E? are the signal
subspace eigenvector matrices for the X and Y subarriys, respec-
v, = -j
(3
-

. . . (Aji,
[ ( A y , - A p , ) sin

- Ap,) sin 19,

ek + (AY, - A V , )cos 0,

+ ( A i , - AV,) cos (20)

tively, [2]. Under the general model of (4), the application of the
ESPRIT algorithm will result in perturbation of the eigenvalues of C. Errors in Sensor Pair Alignments
Y.
Proposition 1: Given the general formulation of the perturbed MO, and Mlkare as defined in (19) where
array signal processing problem of (4), the perturbed kth eigen-
value of Y can be written as
lk = [k + A[, where A[, = ( , b l ( a A , - a/kJe,. (11) In general, the matrix terms MO,, M I , are not diagonal. However,
Here, F k = exp ( - j (27rd sin O,/A)), b; = e l ( I A ) + irnd ek is de- in the case where the X and Y subarrays do not overlap and the
j n e d as the n x l unit vector with unity value at the kth element sensor errors (i.e., gain/phase, location, or alignment errors) have
position and zero elsewhere. zero means and are uncomelated, these terms are found to be di-
Pro08 The signal subspace eigenvector matrix $4 spans the agonal and more explicit expressions for MO,, M , , can be derived.
perturbed steering matrix Suppose that the subarrays do not overlap and that
1 ) The errors in sensor gains and phases are zero mean, uncor-
related with variance U ; and U ; , respectively. Then
MOA = 2 diag (U: + U:.:, . . . , U: + u i a : } (22)
Thus, there exists a nonsingular matrix T such that
and
M I , = 2 diag {e'2"(ui - uia:), . . ., e'2"(u: - uiai)}. (23)

2) The errors in sensor locations are zero mean, uncorrelated

Solving for the least squares solution of Y and using a first-order
+
approximation for the pseudoinverse of ( T A aA,), [ 151, yields
with variance U ; , and 03,
respectively. Then

Y = T-'{Z + (TA)+(aA,- aA,)}@T. (13) (U: sin2 0, + U : COS* e,)rrH (24)

Since the eigenvalues of a matrix are invariant under a similarity
transformation, the eigenvalues of Y are the same as t i e eigenval- and
ues of
{I + (TA)+(aA, - aA,)}@. (14)
Using first-order perturbation approximation [ 161 for the perturbed 3) The angle errors in sensor pair alignments are zero mean,
k th eigenvalue yields uncorrelated with variance U ; . Then

fk = i, + {ke;(rA)'(aA, - aA,)ek. (15)

The proof is done.
Using proposition 1 and standard first-order perturbation tech-
niques as in [6], a general expression can be derived 'or the MSE
of the ESPRIT DOA under the various model errors:

IV. DISCUSSIONS
In the subsequent discussion, it will be assumed that the nominal
(16)
sensor gains and phases are unity and zero, respectively.
Let D, = diag { a k } ,a diagonal matrix formed from the elements
of uk and the matrix terms MO,,MIAbe specified as fo l o w . A. Arbitrary Array Geometry-One and Two Source Case
For nonoverlapping subarrays, simple closed form solutions for
A. Errors in Sensor Gains and Phases
the one source case and approximate solutions for the two source
MOk= D,E{u,uF}D; and M I , = DkE{uAu:} D l (17) case can be found for the errors as in (22)-(27) and (16). Note that

I
I1
2356 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, NO. 9, SEPTEMBER 1992

for the one source case, it is easily shown that of 2mo than when the subarrays are nonoverlapping. This is intui-
tively satisfying as in this case, the subarrays have a larger aperture
1 1
lb,I2 = ;and Re { b r D : b , } = - than in the nonoverlapping subarrays case. Again, the MSE are
m affected only by the phase errors.
For the two source case, the MSE is expressed approximately in a The MSE expressions for MUSIC under random model errors
simple form when the number of sensors in a subarrsy rn is large can be found in [ l l ] . As before, by arguments similar to those
enough such that lapa212<< m 2 , i.e., found in [SI, for the case of zero-mean, uncorrelated errors in the
sensor gains and phases, the MUSIC MSE is
1 1
lbkI2 = - and Re {b%D:b,} = ,; k = 1, 2. (29)
m
Applying (29) to the case of errors in sensor gains a i d phases of
(22)-(23), we get

E{lA8k(2}= f (m
x
y 0;- k = 1, 2. (30) Remarks: It can be seen that the MUSIC MSE is affected by both
errors in gains and phases. In contrast ESPRIT is affected only by
Remark: Thus, for the one and two source case and large m, the phase errors. Furthermore, from (32), (34), and (35) it is clear that
ESPRIT DOA MSE under errors in sensor gains and phases are the MUSIC MSE is smaller when the number of sensors mo is large.
affected only by the sensor phase errors. The fact that the MSE of MUSIC is less than that for ESPRIT for
large mo is also consistent with the results in [7]. Similar obser-
B. Uniform Linear Array ( U L A - n Sources vations for ESPRIT and MUSIC can be made for the other model
error scenarios.
We now turn to the special case of a uniform linear m a y (ULA)
and where the total number of sensors mo is assumed to be large.
There are two cases for the selection of the ESPRIT sitbarrays dis- V. SIMULATION
EXAMPLES
cussed here. The subarrays are chosen from the ULA to be either
nonoverlapping (i.e., m = mo/2, the so-called interkaved array, Two uncorrelated sources are given at DOA angles -20" and
- 10" both with unit power. The X and Y subarrays are taken from
[8]) or to have maximum overlapping (i.e., m = m0 - l ) , with the
distance spacing between sensors in a sensor pair kept at d. We a uniform linear array with distance spacing between the sensors
consider the nonoverlapping subarrays case first. in sensor pairs taken to be d = ( h / 2 ) . The nominal sensor gains
I) Nonoverlapping Subarrays: For the case of zero-mean, un- and phases are unity and zero respectively. Zero-mean, uncorre-
correlated errors in sensor gains and phases, and by using argu- lated errors in the sensor gains and phases are introduced with stan-
ments similar to those found in [ 5 , appendix GI, it c.in be shown dard deviations of and 4",respectively. The total number of
that sensors mo is vaned from 6 to 20 and a hundred trial runs are used
to average the MSE of the TLS-ESPRIT (see [2]) DOA estimate
for both the nonoverlapping and maximum overlapping case. Their
respective Root-MSE are plotted in Figs. 1 and 2 where the solid
line depicts the approximate Root-MSE for the nonoverlapping case
as computed from (32), the dashed line gives the approximate Root-
MSE for the maximum-overlapping case as computed from (34)
while the symbols * and + denote the simulation Root-MSE of
Hence, we find ESPRIT for the nonoverlapping and maximum-overlapping cases,
respectively. The approximate Root-MSE for MUSIC, as com-
puted from (35) shown by a dotted line, and the simulation Root-
MSE for MUSIC, denoted by the symbol 0, are plotted in these
figures.
2) Subarrays with Maximum Overlapping: Simi1ar.y , for zero-
mean, uncorrelated errors in sensor gains and phases, it can be VI. CONCLUSIONS
shown, when the subarrays are selected such that the) have maxi-
mum overlapping (i.e., m = mo - l ) , that The MSE expressions of the ESPRIT DOA estimates under three
random model error scenarios are derived assuming that the co-
2 variance matrix is known. Solutions for the ESPRIT MSE with an
bfkfokb: = ?(U:
m0
+ U:) and aribtrary ESPRIT array geometry and one and two sources are
given. Approximate solutions for a uniform linear array (ULA) and
arbitrary number of sources with a large number of sensors are also
given. Solutions for both cases suggest that ESPRIT is affected
only by phase errors. Furthermore, it is shown that ESPRIT with
Hence, maximum-overlapping subarrays will exhibit lower MSE (which

E { J A 8 k ( 2=} 4 (-) X 2 U: k = 1, .. ,n. (34)

decreases the rate [ l /mi]) than ESPRIT with nonoverlapping sub-
arrays (which decreases at the rate [l/mo]). When compared to
mo 2rdCOS 8k MUSIC estimates (whose MSE decreases at the rate [ l /mi]), it is
found that MUSIC estimates will generally give a lower MSE than
Remarks: Thus, when the subarrays are chosen with maximum ESPRIT. The simulation MSE are well in accordance with the an-
overlapping, the MSE of ESPRIT estimates are smaller by a factor alytical results.
I1
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, ?io. 9, SEPTEMBER 1992 2351

-. (Theoruical) (Simu.)
. - . ( T h d c a l ) + (Simu.)
ESPRIT, Mn-OVdP.
ESPRIT. max-overlap.
....... F.eonueal) o (Simu.) MUSIC

-0 ’ I
6 8 10 12 14 16 18 m
Total number of sensors
Fig. 1. The Root-MSE of the DOA estimate at -20” for ESPRIT and MUSIC. Here, ua = IO-* and u6 = 4”

1.2

-
..- -
meorwcal)
(Thmncal)+
(Simu.)
(Simu.)
ESPRIT. non-ovmhp.
ESPRIT, max-overlap.
....... (Thmrwcal) o(Simu.) MUSIC

w
0.6

0.4

I
6 8 10 12 14 16 18 20

Total number of sensors

Fig. 2. The Root-MSE of the DOA estimate at - 10” for ESPRIT and MUSIC. Here, ua = IO-’ and U+ = 4”

ACKNOWLEDGMENT SIC,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, pp.
1939-1949, Dec. 1989.
The authors wish to thank Dr. R. Liu of the Department of Elec- [5] P. Stoica and A. Nehorai, “MUSIC, maximum likelihood, and Cra-
trical Engineering, University of Notre Dame, and Dt. L. Tong of m&-Rao bound,” IEEE Trans. Acoust., Speech, Signal Processing,
the Department of Electrical Engineering, West Viginia Univer- vol. 37, pp. 720-741, May 1989.
sity, for their suggestions and valuable discussions in the course of [6] 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,”
this research. IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, pp. 1990-
1995, Dec. 1989.
REFERENCES [7] P. Stoica and A. Nehorai, “Performance comparison of subspace ro-
tation and MUSIC methods for direction estimation,” in Proc. 5th
[l] R. 0. Schmidt, “Multiple emitter location and signal parameters es- ASSP Workshop Spectrum Estimation Modeling, Oct. 1990, pp. 357-
timation,” IEEE Trans. Antennas Propagat., vol. AP-34, pp. 276- 361.
280, Mar. 1986. [8] B. Ottersten, M. Viberg, and T. Kailath, “Performance analysis of
[2] R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters the total least squares ESPRIT algorithm,” IEEE Trans. Signal Pro-
via rotational invariance techniques,” IEEE Trans. Acsust., Speech, cessing, vol. 39, pp. 1122-1135, May 1991.
Signal Processing, vol. 37, pp. 984-993, July 1989. [9] K. M. Wong, R. S. Walker, and G. Niezgoda, “Effects of random
[3] M. Kaveh and A. J. Barabell, “The statistical performance of MU- sensor motion on bearing estimation by the MUSIC algorithm,” Proc.
SIC and the minimum-norm algorithms in resolving plan$ waves in Inst. Elec. Eng., vol. 135, pt. F, pp. 233-250, June 1988.
noise,” IEEE Trans. Acoust., Speech, Signal Processing, vol. [IO] B. Friedlander, “A sensitivity analysis of the MUSIC algorithm,” in
ASSP-36, pp. 331-340, Apr. 1986. Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing (Glasgow,
[4] B. D. Rao and K. V. S . Hari, “Performance analysis of Root MU- Scotland), 1989, pp. 2811-2814.

1
I
2358 IEEE TICANSACTIONS ON SIGNAL PROCESSING, VOL 40, NO. 9. SEPTEMBER 1992

[ l l ] A. Swindlehurst and T. Kailath, “A performance analysis of sub- and shown to be very useful for computing the globally optimal
space-based methods in the presence of model errorr-Part I: The values of rational model parameters.
MUSIC algorithm,” IEEE Trans. Signal Processing, \ol. 40, no. 7,
pp. 1758-1774, July 1992. Homotopy continuation methods are numerical methods for
[12] F. Li and R. J . Vaccaro, “Statistical comparison of siibspace based solving systems of nonlinear equations and have previously been
DOA estimation algorithms in the presence of sensor emrs,” in Proc. applied to nonlinear problems in areas such as applied mathematics
5rh ASSP Workshop Spectrum Esrimation Modeling, Oct. 1990, pp. [2], [3], engineering [4], [5], and circuit design and analysis [6],
327-33 1. [7]. The application of these methods to nonlinear optimization
[13] A. L. Swindlehurst and T. Kailath, “An analysis of the subspace
fitting algorithm in the presence of sensor errors,’’ in P-oc. IEEE Inr. problems in signal processing is not yet very widely developed.
Conf. Acoust., Speech, Signal Processing (Albuquerquc:, NM), 1990, These methods are advantageous as they are globally convergent
pp. 2647-2650. and can be constructed to be exhaustive. Globally convergent al-
[14] A. L. Swindlehurst and T. Kailath, “On the sensitivity of the ESPRIT gorithms always converge to a solution from any initial point, even
algorithm to nonidentical subarrays,” Proc. Indian AcGd. Sci., 1991. initial points far from any solution, and exhaustive algorithms com-
[ 151 S. D. Hodges and P. G. Moore, “Data uncertainties and least squares
regression,” Appl. Srar., vol. 21, pp. 185-195, 1972. pute all solutions to a given system of equations. For the rational
[16] G. H. Golub and C. F. VanLoan, Matrix CompirrarionJ. Baltimore, approximation problem, these properties theoretically guarantee
MD: Johns Hopkins University Press, 1983. that the globally optimum model parameters can be found. How-
ever, computing these optimal parameter values requires comput-
ing the set of all stationary points, and then selecting the optimal
values from this set. Although this procedure may be computation-
ally intensive, the computation time required for this task can be
Globally Optimal Rational Approximation [Jsing significantly reduced as the solutions can be computed simulta-
Homotopy Continuation Methods neously using the particular approach described in this correspon-
dence.
Virginia L. Stonick and S . T. Alexander
In this correspondence, a specific homotopy continuation method
for solving the nonlinear equations arising in the mean-square ra-
Abstract-Homotopy continuation methods are applied to the nonlin- tional approximation of higher order pole/zero systems is devel-
ear problem of approximating a higher order system by a lower order oped. The specific homotopy function and continuation methods
rational model, such that the mean-square modeling etror is mini- used are described in Section 11. The use of this procedure to solve
mized. First, a homotopy function is constructed which creates distinct rational approximation problems is developed in Section 111, and a
paths from each of the known solutions of a simple problem to each of
the solutions of the desired nonlinear problem. This homotopj function numerical example is presented in Section IV. The attention given
guarantees that the globally optimum rational approximation solution in this correspondence to solving these theoretical nonlinear equa-
may be determined by finding all the solutions. A simple numerical tions is similar to the attention given to using the steepest descent
continuation algorithm is described for following the paths to the op- method to solve the theoretical normal equations for all-zero mod-
timum solution. Finally, a numerical example is included which dem-
eling [8]. In both cases the system is assumed to be known so that
onstrates that the globally optimum model will he ohtainec, by applying
this homotopy continuation method. the performance of the algorithm in different data environments can
be evaluated. Demonstrating that homotopy methods successfully
solve these theoretical equations is a prerequisite to the future use
I. INTRODUCTION of homotopy methods for problems involving actual data environ-
ments.
The problem addressed in this correspondence is coriputation of
the optimal coefficients for a rational model such that the mean-
square error (MSE) between the impulse response of a specific sys- 11. HOMOTOPY
CONTINUATION
METHODS
tem and that of the model is minimized. Problems of this type oc- In this correspondence, the objective is to compute w , the
cur, for example, in the optimal design of fixed order infinite-im- N-component vector of rational model parameters, that solves the
pulse-response filters used to model only the dominant poles and set of N nonlinear equations
zeros of higher order autoregressive moving average (ARMA) pro-
cesses [l]. In general, the equations defining the minimum MSE f ( w ) = 0. (1)
parameter estimates for the rational model are nonlinear and the
MSE surface is potentially multimodal. Consequently, gradient- Specifically, for the rational approximation problems considered
based methods may converge to local minima or diverge, and equa- in this correspondence, ( 1 ) is defined by the polynomial forms of
tion error approaches are suboptimal. In this correspondence a new the equations ( d / d w ) { ( w ) = 0 defining the stationary points of the
approach based upon homotopy continuation methods IS presented MSE surface { ( w ) . These equations are discussed in Section 111.
The homotopy approach uses (1) as defined by the application to
Manuscript received June 25, 1990; revised September 2f, 1991. This create the homotopy vector function h(w, 7) as
work was funded in part by the Bell Communications Reserch Graduate
Fellowship Program and the National Science Foundation undc r Grant MIP-
8552571. h(w, 7) = ( 1 - r ) g ( w ) + r f ( w ) = 0, 0 5 r 5 1. (2)
V. L. Stonick was with the Department of Electrical and C’omputer En-
gineering, North Carolina State University, Raleigh, NC 276‘95-7911, She In (2), 7 is the homotopy parameter and g ( w ) = 0 is a simpler set
is now with the Department of Electrical and Computer Engi leering, Car- of equations for which the solution is known or is easily calculable.
negie-Mellon University, Pittsburgh, PA 15213-3890. The form for g ( w ) used in this correspondence is [2]
S . T. Alexander is with the Department of Electrical and C’omputer En-
gineering, North Carolina State University, Raleigh, NC 27695-791 1.
IEEE Log Number 9201607.

.OO 0 1992 IEEE