Вы находитесь на странице: 1из 4



171 C. S . Burms and T. W . Parks, D F T I F F T a n d Contdution AIXoriihrn.7.

New York: Wiley, 1985.
[8] D. M. W . Evans, "A second improved digit-reversal permutation algorithm for fast transforms." l Trans. Acousr.. Speech, Signcl/
Processing, vol. 37, pp. 1288-1291, 1989.

z axis

On the Application of Uniform Linear Array Bearing

Estimation Techniques to Uniform Circular Arrays
A . H. Tewfik and W . Hong

Abstract-The problem of estimating the directions of arrivals of

narrow-band plane waves impinging on a uniform circular array with
M identical sensors uniformly distributed around a circle is considered. It is shown that if the number of sensors M is large enough then
a reordering of the inverse discrete Fourier transform of the sensor
outputs yields a sequence of measurements z(q) = ZAA,Jq(2nR
sin $ " / I )
exp (-&On),
where 0, and q5" are the azimuth and elevation angles of
arrival of the nth plane wave, respectively. For each candidate elevation angle 6. this sequence is processed using ROOT MUSIC or any
other modern line spectral estimation technique as if it came from a
uniform linear array. Any root estimated via ROOT MUSIC which is
on or close to the unit circle then indicates the presence of a source at
the elevation under consideration and an azimuth equal to the phase
of the root. Experimental results are provided to demonstrate the advantages of processing the transformed data.

y ax1\

Fig. 1. Circular array

ular, for each candidate elevation angle I$,l the sequence z(q) is
processed using ROOT MUSIC or any other modem line spectral
estimation technique as if it came from a uniform linear array. Any
root estimated via ROOT MUSIC which is on o r close to the unit
circle then indicates the presence of a source at the elevation under
consideration and an azimuth equal to the phase of the root.



Consider first the case where a single narrow-band plane wave

of wavelength h is impinging in the absence of noise on the array
at an azimuth 8 and an elevation I$ as shown in Fig. 1. Using the
complex notation for narrow-band signals [4] we may write the
output of sensor m as

Circular arrays are known to be equivalent to nonuniform linear

arrays [ l ] . Unfortunately, some of the modem array processing
algorithms (e.g., ROOT MUSIC [ 2 ] )can only be applied to uniform linear arrays. Furthermore, other algorithms have been observed to perform better when applied to uniform linear arrays than
when applied to circular arrays (e.g., in terms of signal-to-noise
ratio resolution thresholds) [3].
In this correspondence we show that it is possible to use ROOT
MUSIC and other procedures which were designed to work with
linear arrays to determine the azimuths and elevations of N sources
using a circular array which consists of M ( M > N ) identical sensors uniformly distributed around a circle of radius R (Fig. I ) . Our
proof is based on an expansion of a plane wave in an infinite series
of Bessel functions of the first kind. The series may be considered
to be the discrete time Fourier transform of a discrete time signal.
The sensors provide samples of that Fourier transform at the frequencies 27rm/M, m = 0 , M - 1. By computing the inverse discrete Fourier transform of the sensor outputs we obtain a temporally aliased version of the discrete time signal. We show that if
the number of sensors M is larger than 47rR/h where h the wavelength of the plane wave, then a reordering of the inverse discrete
Fourier transform of the sensor outputs yields a sequence of measurements z(q) = E A,,Jq(2*R sin I $ ~ / A ) exp (-jqO,), where O,, and
O,, are the azimuth and elevation angles of arrival of the nth plane
wave and A , is a complex constant proportional to the nth plane
wave magnitude and phase and to the sensor response. To find 0,,
and d,, a search is conducted over possible elevations I$n. In partic-

x ( m , r) = A(r) exp j


sin I$ cos




where the complex parameter A(r) denotes the complex amplitude

of the plane wave at time t measured at the center of the array.
Using the addition theorem for Bessel function [ 5 ] we can rewrite
(1) as

x ( m , t) = A(?)k = - r n


sin 0) exp ( j k ( , -



where J k ( . ) denotes a Bessel function of the first type and order k

[5] and 8 = 8 - 7r/2. Observe that according to ( 2 ) ,x ( m , r ) may
be interpreted as providing the value of a sample of the Fourier
transform of the sequence y ( k ) = A(r)Jk(27rRsin $,,/A) exp ( - j k O )
at the frequency 2 7 r k / M . Using this interpretation we find that the
inverse discrete Fourier transform of x ( m , r ) , x ( q , t ) 0 5 q 5 M
- 1 , is given by [6]

Manuscript received May 9, 1990; revised February 15. 1991.

The authors are with the Department of Electrical Engineering, University of Minnesota, Minneapolis, M N 55455.
IEEE Log Number 9 1060i 2 ,

The above equation may also be found in [ 7 ] .

If the number of sensors M is much larger than 47rR/h, then
sin I$/h) = 0 for 1 5 k and k 5 -2 [ 5 ] .Under those

1053-587X/92$03.00 0 1992 IEEE




data correlation matrix when the background noise is white or a

generalized eigendecomposition of that matrix and the correlation
matrix of the transformed noise process when the noise is colored
with a known correlation matrix. For each candidate elevation 6
we compute the roots of the equation:

conditions (3) may be rewritten as

ZHDH(4)E,E"D(4)Z= 0
where we have used the fact that J - & . ) = ( . ) [ 5 ] .Note
that if M - q is large enough then J M - , ( 2 a R sin 4 / X ) = 0 and
x ( q , t ) = A(t)J,(2nR sin d/X) exp ( - j q 0 ) . On the other hand, if
q is large enough then Jq(2nR sin 4/A) = 0 and x ( q , t ) =
A ( t ) ( - 1 ) M - 4 J M - y ( 2 ~sin
R $/A) exp ( - j ( q - M ) @ . Hence, by
properly reordering the sequence x ( q , r ) we can obtain another sequence that is proportional to A(t)J,(2aR sin 4/A) exp ( - j q 0 ) . In
particular, using the fact that M is much larger than 4 a R / X we can
define a new sequence z(q, t ) as follows:

where z is an M


1 vector given by





and D ( 4 ) is the diagonal M x M matrix

A root close to the unit circle would then indicate the presence of
a source at the elevation currently under consideration and an azimuth equal to the phase of the root.

where L M / 2 J denotes the largest integer that is smaller than or

equal to M / 2 .
If the array response is linear then (5) implies that the transformed array outputs z(q, t ) corresponding to N narrow-band plane
waves of wavelength X impinging on the array with angles 8,, 1 I
n 5 N is of the form

Equation (6) with 4" = a / 2 (i.e., for the case where all the
sources and the array are coplanar) appears already in [ 8 ] . However, the derivation of [8] is incorrect as it claims that (6) (with 4,,
= a / 2 ) may be obtained by computing a discrete Fourier transform of the sensor outputs. As mentioned above, a discrete Fourier
transform of the sensor outputs will in fact yield (3) and not (6).
Note that since J,(x) = 0 when q >> 2x, ( 6 ) also implies that
the effective number of sources that may be resolved by a circular
array at a fixed elevation is a function of the radius R of the array
and the elevation angle. In particular, that number may be smaller
than M , the number of sensors. The effect of this implication on
the procedure presented here and other high resolution bearing estimation techniques when applied to uniform circular arrays is currently under investigation.
Now observe that z ( q , t ) is the sum of exponentials that are modulated by the sequence ofA,(t)Jq(2aR sin &/A). To determine the
elevation and azimuth angles of the sources we use a ROOT MUSIC based approach as follows. W e begin by discretizing the range
of possible elevation angles q5 and estimating a matrix E, whose
columns form an orthogonal basis for the "noise-only" subspace
[ 2 ] . As in other eigendecomposition techniques, E, may be estimated either by using an eigendecomposition of the transformed

Consider now the case where the array operates in the presence
of a background noise process only and observe that the array samples the noise process at the points 7, = ( R , 27rm/M, 7r/2), 0 5
m 5 M - 1. Denote by K,(m, m') = E { x ( m , t ) x * ( m ' , t ) } the
autocorrelation of the sensor outputs x ( m , t ) .In general, K,(m, m ' )
will not be a function of m - m ' , i.e., the process x ( m , t ) will not
be stationary in the m variable, even if the background noise field
is spatially homogeneous. However, if the underlying noise process is isotropic, isotropic and homogeneous, or cylindrically symmetric [ 9 ]then x ( m , t ) will be stationary in the m variable. Specifically, if the underlying field is isotropic and stationary then

K,(m, m ' )

K(I ?,


K(2R sin (a(m - m ' ) / M )


where K(r) is the rotationally invariant correlation of the underlying isotropic and homogeneous background noise field. In particular, K,(m, m ' ) = 026(m - m ' ) when the background noise field
is spatially white and homogeneous with a variance of u 2 .
The covariance of the inverse discrete Fourier transform x ( q , t )
of x ( m , t ) can be directly computed from that of x ( m , t ) . If we
denote by Rk and R, the correlation matrices of the sensor outputs
and the inverse discrete Fourier transform of the sensor outputs in
the presence of the noise field only respectively, we can relate Rk
and R, through the equation

R, = UR, U H


where denotes U H the complex conjugate transpose of U and U =

[U,,] is the M x M discrete Fourier transform matrix with

1 exp ( j


Now observe that since K,(m, m ' ) = K(2R sin ( a ( m - m ' ) / M )

for a homogeneous and isotropic background noise field, the matrix
Rk is circulant [ l o ] when the background noise field is homogeneous and isotropic. Since a circulant matrix is diagonalized by the
discrete Fourier transform matrix, we conclude that R, will be a
diagonal matrix. Thus, an isotropic and homogeneous background
noise is mapped into a possibly nonstationary white noise process.
In particular, a spatially white and homogeneous noise process is
mapped into a stationary white noise with variance 0 2 / M .
If we assume that the real and imaginary parts of x ( m , t ) are




Trans-Root MUSIC
---e-Conv-MIN NORM

















Fig. 2. Probability of resolution of MUSIC, proposed approach and MinNorm applied to the array data and transformed array data.
uncorrelated at any given time t and have identical covariance functions then E { x ( n , t ) x ( m , t ) } = 0. Combining this fact with the
above discussion and ( 5 ) we find by direct substitution that for a
homogeneous and isotropic background noise field the transformed
process z ( q , t ) is a white noise process which is possibly nonstationary .
Combining the results of this section with those of Section 11,
we find that the transformed output z ( q , 1 ) corresponding to a circular array that is operating in the presence of M narrow-band plane
waves and a background noise field is

where the O,,'s are the azimuths of the angles of arrival of the plane
waves, q5,, the elevations, and w ( q , r ) is generally a nonstationary
process whose covariance is related to that of the background noise
field as explained above and which is white if the background noise
field is homogeneous and stationary.

Simulation experiments were conducted to compare the results
of applying ROOT MUSIC and Minimum Norm [ 111 to the transformed outputs of a 23 element circular array of radius 5 h / 2 a with
those that are obtained by applying MUSIC [I21 and Minimum
Norm to the array outputs directly. The sensors were equally spaced
along the circumference of the array. Two sources of equal power
were assumed to be in the far field of the array at azimuths and
elevations (e,q5) of (60", 40") and (80", 45"). The probability of
resolution of MUSIC and Minimum Norm applied to the array outputs and transformed array outputs as well as the average meansquare errors in the elevations and azimuths estimated via those
procedures were computed from 100 independent trials at each of
the signal-to-noise ratios that we considered. The corresponding
results for ROOT MUSIC applied to the transformed array outputs
were computed from 50 independent trials at each signal-to-noise
ratio. In each trial the covariance matrices of the array outputs and
the transformed array outputs were estimated using 100 snapshots.
The signal-to-noise ratio was defined in terms of the ratio of a single source power to the variance of the additive white noise process. The sources were considered to be resolved if two estimates
of the direction of arrivals were obtained and each was located
within k2' of the true elevation and azimuth angles. Fig. 2 shows
the probability of resolution at various signal-to-noise ratios for the




S N R ( dB )

S N R ( dB )

Fig. 3. Average mean-square error in elevation estimates





Trans-Root MUSIC






SNR ( d B )

Fig. 4. Average mean-square error in azimuth estimates

ROOT MUSIC based technique described in Section 111, Minimum
Norm applied to the transformed array output data and for MUSIC
and Minimum Norm applied to the sensors output data. Figs. 3 and
4 show the average mean-square errors in the estimated elevations
and azimuths. Note that both figures confirm the improvement in
performance gained by using ROOT MUSIC with the transformed

In this correspondence we have studied a transformed data sequence that is equal to a reordering of the inverse discrete Fourier
transform sequence corresponding to the outputs of a circular array
with M elements uniformly distributed around the array circumference. We have shown that the transformed sequence may be processed using a ROOT MUSIC based approach to estimate the elevations and azimuths of the observed sources. A computationally
less expensive search over the elevation angles q5 may also be done
using the multichannel Schur algorithm [13] and will be reported
in the near future together with a statistical performance analysis
of the proposed approach.
A. Monzingo and T. W. Miller, Introduction to Adaptive Arrays.
New York: Wiley. 1980.

[ 11 R .

101 I


[2] A. J. Barabell, Improving the resolution performance of eigenstructure-based direction-finding algorithms, in Proc. 1983 IEEE Con5
Acoust., Speech, Signal Processing (Boston, MA), 1983, pp. 336339.
[3] K. M. Buckley and X. L. Xu, A comparison of element and beam
space spatial-spectrum estimation for multiple source clusters, in
Proc. 1990 IEEE Conf: Acoust., Speech, Signal Processing (Albuquerque, NM), 1990, pp. 2643-2646.
[4] S. Haykin, Communications Systems, 2nd ed. New York: Wiley,
[5] W. Magnus, Formulas and Theorems f o r the Special Functions of
Mathematical Physics. New York: Springer, 1966.
[6] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1989.
[7] J. N. Maksym, Directional accuracy of small ring arrays, J .
Acousr. Soc. Amer., vol. 61, no. 1, pp. 105-109, Jan. 1977.
[8] M. P. Moody, Resolution of coherent sources incident on a circular
array, Proc. IEEE, vol. 68, no. 2, pp. 276-277, Feb. 1980.
[9] M. I. Yadrenko, Spectral Theory of Random Fields. New York:
Optimization Software, 1986.
[IO] J. H. McClellan and T. W. Parks, Eigenvalue and eigenvector decomposition of the discrete Fourier transform, IEEE Trans. Audio
Electroacoust., vol. AU-20, pp. 66-74, Mar. 1972.
[ l l ] R. Kumaresan and D. W. Tufts, Estimating the angles of arrival of
multiple plane waves, IEEE Trans. Aerosp. Electron. S y s t . , vol.
AES-19, no. 1, pp. 134-139, Jan. 1983.
[12] R. 0. Schmidt, Multiple emitter location and signal parameter estimation, IEEE Trans. Antennas Propag., vol. AP-34, pp. 276-280,
March 1986.
[ 131 J. Rissanen, Algorithms for triangular decomposition of block Hankel and Toeplitz matrices with applications to factoring positive matrix polynomials, Mat. Comput., vol. 27, no. 121, pp. 147-154,

H ( z ) is the transfer function of the underlying noncausal ARMA

model (see ( I C ) and ( I d ) in the paper), and the symbol * denotes
convolution. This claim is false. The correct maximum orderp, is
given by

The p + causal poles of H(z) yield the p(p

1 ) / 2 causal poles
of H2(z, n), and similarly for the anticausal poles. The interaction between the causal and the anticausal poles of H(z) does not
contribute any poles to H2(z, n). We refer the reader to [ l ] for
details. Here we will illustrate this with a simple example of a
noncausal AR(2) model that has one causal pole and one anticausal
pole, i.e., p + = 1, p - = 1, a n d p = 2. (Note that since this is an
AR model, we have B(z) = 1 so that there is no cancellation between B(z) and A2(z-, n) for any n.)
We consider an AR(2) model with the transfer function

It has a causal pole at 0.5 and an anticausal pole at 2.5. We have


H(u) = [u][(u

- 0 . 5 ) ( ~- 2.5)]-


H(zt4-I) = [ZZ][1.25(U- 2z)(u - 0.4z)I-

Counterexamples to On Estimating Noncausal

Nonminimum Phase ARMA Models of
Non-Gaussian Processes

The closed contour in (3) is the unit circle running counterclockwise since H ( z ) is analytic in an annular region enclosing the unit
circle. The poles of the integrand in (3) encircled by the unit circle
are located at


Jitendra K. Tugnait


and 0

if n

< -1

and at
Abstract-In the above-paper, an order selection procedure has been
proposed for parameter estimation for noncausal, nonminimum phase
ARMA models of non-Gaussian processes. We will show that it has
been derived under an erroneous assumption, and we also give a counterexample to show that it does not yield a consistent order estimate in
general. Two linear approaches for parameter estimation have also
been presented in the paper. We point out that an existing counterexample to an earlier version of one of the algorithms also applies to
both the approaches of the above paper.




We first
on the order selection method presented in
Section IV of the above paper, I Throughout the section the authors
exploit the claim that for a noncausal ARMA model with p + causal
poles and p - anticausal poles (p = p + + p - ), the maximum order
p2 of the polynomial A 2 ( z - l , n) is p 2 = p ( p + 1 ) / 2 , where A , ( z ,
n) is the denominator polynomial of H 2 ( z , n) : = ~3~ H ( z ) * H ( z ) ,
Manuscript received May 8, 1990; revised April 15, 1991.
The author is with the Department of Electrical Engineering, Auburn
University, Auburn, AL 36849.
IEEE Log Number 9106003.
G. B. Giannakis and A. Swami, IEEE Trans. Acoust., Speech, Signal
Processing, vol. 38, no. 3, pp. 478-495, Mar. 1990.


and 0.42

if n 2 -1.

By the Cauchy residue theorem, we have (n 2 - 1)

-0 . 5 ~
Hz(z n, = 4(z - 0.25)(z - 1.25)

+ 0.8(z

- 1.25)(z - 6.25)

From (6) one may be tempted to conclude that for n 2 - I , H ~ ( z ,

n) has poles at 0.25, 1.25, and 6.25. In reality, there are only two
poles: 0.25 and 6.25, obtained by interaction of causal poles
with causal poles, and anticausal poles with anticausal Poles, respectively. There are no poles produced by interaction between
causal poles and anticausal poles: details are in [l]. We illustrate
this by examining (6) a bit further.
The denominator polynomial of H2(z, n) is given by

From (6) and (7), the numerator polynomial Of H2(Z,n) is given by

N(z) = -5(4.0z)z2(z - 0.25)

- 0.25)

1053-587X/92$03.00 0 1992 IEEE

- (O.5)z2(z - 6.25)

+ ( 1 . 2 5 ~ - ) (-~ 6.25)l.