5914 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO.

11, NOVEMBER 2010

Further work involves extension to nonseparable, nonaxisymmetric A Robust Adaptive Beamformer Based on Worst-Case
functions. Semi-Definite Programming
Zhu Liang Yu, Zhenghui Gu, Jianjiang Zhou, Yuanqing Li,
Wee Ser, and Meng Hwa Er
REFERENCES
[1] J. P. Antoine and C. B. Torresani, “Wavelets on the 2-sphere: A group-
theoretical approach,” Appl. Comput. Harmon. Anal., vol. 7, pp. 1–30, Abstract—In this correspondence, a novel robust adaptive beamformer
1999. is proposed based on the worst-case semi-definite programming (SDP). A
[2] J. Antoine and P. Vandergheynst, “Wavelets on the 2-sphere and related recent paper has reported that a beamformer robust against large steering
manifolds,” Rep. Math. Phys., vol. 43, no. 1–2, pp. 13–24, 1999. direction error can be constructed by using linear constraints on magni-
[3] B. Yeo, W. Ou, and P. Golland, “On the construction of invertible filter tude response in SDP formulation. In practice, however, array system also
banks on the 2-sphere,” IEEE Trans. Image Process., vol. 17, no. 3, pp. suffers from many other array imperfections other than steering direction
283–300, Mar. 2008. error. In order to make the adaptive beamformer robust against all kinds of
[4] J. Wang, K. Xu, K. Zhou, S. Lin, S. Hu, and B. Guo, “Spherical har- array imperfections, the worst-case optimization technique is proposed to
monics scaling,” Vis. Comput. vol. 22, no. 9, pp. 713–720, Sep. 2006 reformulate the beamformer by minimizing the array output power with
[Online]. Available: http://dx.doi.org/10.1007/s00371-006-0057-8 respect to the worst-case array imperfections. The resultant beamformer
[5] I. Dokmanic and D. Petrinovic, “Convolution on the n-sphere with ap- has the mathematical form of a regularized SDP problem and possesses
plication to pdf modeling,” IEEE Trans. Signal Process., vol. 58, no. 3, superior robustness against arbitrary array imperfections. Although the
pp. 1157–1170, Mar. 2010. formulation of robust beamformer uses weighting matrix, with the help
[6] P. Hall, G. S. Watson, and J. Cabrera, “Kernel density-estimation with of spectral factorization approach, the weighting vector can be obtained
spherical data,” Biometrika, vol. 74, no. 4, pp. 751–762, Dec. 1987. so that the beamformer can be used for both signal power and waveform
[7] J. R. Driscoll and J. D. M. Healy, “Computing Fourier transforms and estimation. Simple implementation, flexible performance control, as well
convolutions on the 2-sphere,” Adv. Appl. Math., vol. 15, no. 2, pp. as significant signal-to-interference-plus-noise ratio (SINR) enhancement,
202–250, 1994. support the practicability of the proposed method.
[8] D. Healy, D. Rockmore, P. Kostelec, and S. Moore, “FFTs for the Index Terms—Adaptive array, constraints on magnitude response, ro-
2-sphere-improvements and variations,” J. Fourier Anal. Appl. vol. 9, bust adaptive beamforming, semi-definite programming, spectral factor-
no. 4, pp. 341–385, Jul. 2003 [Online]. Available: http://dx.doi.org/10. ization, worst-case optimization.
1007/s00041-003-0018-9
[9] I. Gradshteyn and I. Ryzhik, Table of Integrals, Series, and Products,
A. Jeffrey and D. Zwillinger, Eds., 7th ed. New York: Academic I. INTRODUCTION
Press, 2007.
[10] N. Y. Vilenkin, Special Functions and the Theory of Group Represen- Adaptive arrays [1] are widely used in wireless communications, mi-
tations, 1st ed. Providence, RI: AMS, 1968. crophone array, radar, sonar and medical imaging. It has high perfor-
[11] J. Avery, Hyperspherical Harmonics: Applications in Quantum mance in interference suppression if the array steering vector (ASV)
Theory, ser. Reidel Texts in the Mathematical Sciences, 1st ed. New corresponding to the target is accurately known. Due to the existence
York: Springer, 1989, vol. 5.
[12] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal of array imperfections, the key problem in the application of adaptive
Processing, 2nd ed. Upper Saddle River, NJ: Prentice-Hall, 1999. array is its robustness against these imperfections. Many approaches,
such as multiple-point constraints [2], derivative constraints [3], eigen-
vector constraints [4], Bayesian approach [5], as well as techniques re-
straining excess coefficients growth [6], [7], have been proposed for
adaptive beamformers to achieve robustness. However, they have a
common drawback that the degrees of freedom (DOFs) for interfer-
ence rejection are consumed by additional constraints, which causes
drastic output SINR degradation especially when array is designed to

Manuscript received March 16, 2010; accepted July 01, 2010. Date of publi-
cation July 15, 2010; date of current version October 13, 2010. The associate ed-
itor coordinating the review of this manuscript and approving it for publication
was Dr. Konstantinos Slavakis. This work was supported in part by the National
Natural Science Foundation of China under Grants 60802068 and 60825306,
Guangdong Natural Science Foundation under Grants 8451064101000498, Pro-
gram for New Century Excellent Talents in University under Grant NCET-10-
0370, and the Fundamental Research Funds for the Central Universities, SCUT
under Grant 2009ZZ0055.
Z. L. Yu, Z. Gu, and Y. Li are with the College of Automation Science and En-
gineering, South China University of Technology, Guangzhou, China, 510640
(e-mail: [email protected]; [email protected]; [email protected]).
J. Zhou is with the College of Information Science and Technology, Nanjing
University of Aeronautics and Astronautics, Nanjing, China 210016 (e-mail:
[email protected]).
W. Ser and M. H. Er are with the School of EEE, Nanyang Technological
University, Singapore, 639798 (e-mail: [email protected]; [email protected]).
Color versions of one or more of the figures in this correspondence are avail-
able online at http://ieeexplore.ieee.org.
This paper has supplementary downloadable multimedia material available at
http://ieeexplore.ieee.org provided by the authors. The Matlab simulation code
of the proposed method can be found in the supplementary multimedia material.
Digital Object Identifier 10.1109/TSP.2010.2058107

1053-587X/$26.00 © 2010 IEEE

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 11, NOVEMBER 2010 5915

tolerate a large steering direction error. The eigenspace-based beam- beamformer based on semidefinite programming and worst-case op-
former [8] is also a good candidate for robust beamforming. It does not timization (RAB-SDP-WC). It has a mathematical form of regularized
suffer from the problem of DOF consumption. However, it is inefficient SDR problem with only a regularization parameter to control the ro-
when the input signal-to-noise ratio (SNR) is low, or the dimension of bustness of array. After our numerical evaluation, we find out that the
signal-plus-interference subspace is high or not exactly known. output SINR of RAB-SDP-WC is not very sensitive to the selection of
Some new approaches in [9]–[11] are robust against arbitrary ASV regularization factor if it is in a suitable range.
errors within an assumed uncertainty set. If the uncertainty set is mod-
elled as an ellipsoid centered at the nominal ASV, it has been proved II. THE PROPOSED ROBUST ADAPTIVE BEAMFORMER
Considering a linear array consisting of N isotropic sensors, without
in [11] that these beamformers [9]–[11] are equivalent and belong to
the class of diagonal loading approach [6], [7], [12]. The approaches in
[10] and [11] have lower computation complexity of O(N 3 ) than that loss of generality, we assume that the coordinate of the first sensor is
d0 = 0. The coordinates of other elements are di , i = 1; 1 1 1 ; N 0 1. In
this correspondence, we assume that di is a multiple of a base distance,
of the second-order cone programming (SOCP) based approach in [9],
O(N 3:5 ), where N is the number of sensors. Moreover, the method d. As a special case, a ULA has the coordinates of di = i 2 d, i =
0; 1 1 1 ; N 0 1. A plane wave with wavelength  is incident on the array
in [11] also provides accurate estimate on the signal power while the
from an angle,  . The sensor-sampled signals, xi (n), i = 0; 1 1 1 ; N 0 1
others do not. The robust beamformers [9], [13] also exploit the idea
are linearly weighed by wi , i = 0; 1 1 1 ; N 0 1 to generate the output
of worst-case optimization [14] to deal with arbitrary ASV errors. Al-
though they are robust against array imperfections, they do not give
accurate estimate of signal power.
y(n).
We have discussed in [17] that the constraints on magnitude response
Many robust adaptive beamformers have been proposed, but few
of them are able to flexibly control the beamwidth and the response L2  j G() 2 U 2 ;
j  8  [ ; 
2 l u] (1)
ripple of robust response region. Some recently proposed beamformers
can be used in adaptive beamforming against large steering direction
error. l and u respectively denote the lower and upper bounds of the
[15]–[17] provide functions on controlling the beamwidth of the ro-
bust response region. However, the controllable beamwidth of Chen’s
interested robust response region where target signal arrives from with
a high probability to prevent target signal cancellation. L and U are the
method [16] is limited and inversely proportional to the array aperture.
The adaptive matrix approach (AMA) [15] uses semi-definite relax-
lower and upper limits of the magnitude response, respectively. The
square of array magnitude response function jG()j2 is given as
ation (SDR) [18] formulation and constrains the peak sidelobe level
as well as the lower response of array in the 3-dB mainbeam to pre-
vent mainbeam splitting. Due to the tradeoff between the beamwidth G() 2 = tr R ()W
j j f s g (2)

where R () = s()s (), W = ww , tr

and peak side lobe level, the width of 3-dB mainbeam cannot be ar- H H
bitrarily set with a given peak sidelobe level. On the other hand, the s , and (1)H denote the
f1g

trace operator and Hermitian transposition operator, respectively. The

AMA [15] is mainly designed for signal power estimation. Only for a
matrix W is rank-one constrained generally. In the SDR formulation,
uniform linear array (ULA), it has been proved [15] that the optimal
array weight vector can be obtained by spectral factorization [19]. The it is just semi-definite and denoted as W
 0. Generally, the SDR

resultant optimal beamformer can be used not only for signal power problem has lower optimal value than the SDP problem. However, for a
ULA, is was proved in [17] that the SDR problem has the same solution
estimation but also for signal waveform estimation.
as the SDP problem. The array steering vector () and array weight s
In many practical applications, due to limited computational re-
sources, continuous updating of the beamformer’s weights may be
w
vector are defined as
prohibitive. Instead, frozen weights are used, which are designed using s() = [e e j j
111 ej
]
T

array data collected over a time interval and then applied over the next
time interval despite changes in the scenario [20], [21]. If the direction
w = [w0 w1 w 01 ] 111 N
T
(3)

of source changes during this interval, the output power of the target where i = 2d cos( )=, i = 0;
i ; N 1.
111 0

source may change drastically if the response ripple in the robust By expressing the array response function and array output power in
response region is large. In wireless communications, when signal SDP formulation, the RAB-SDP [17] is expressed as
power information is used (e.g., amplitude modulation), the variation
of output power may significantly degrades the system performance. In W
min trf RW g (4a)
order to overcome it, a robust adaptive beamformer based on semidef- s.t. tr fR ()W U 2 ;  [ ; 
s g  8 2 l u] (4b)
inite programming (RAB-SDP) [17] uses constraints on upper and tr f R ()W L2;  [ ;  ]
s g  8 2 l u (4c)
lower responses. In other words, it restricts the array response ripple
in the mainbeam. The RAB-SDP has also been formulated as an SDR W 0:  (4d)
problem. Unlike the AMA being a matrix weighting based method, where R is the covariance matrix of array snapshot. The constraints
the RAB-SDP is a vector weighting based method. The weight vector in this beamformer design are suitable for all possible  in the range
of RAB-SDP is obtained by the extended spectral factorization [22]. [l ; u ].
From our numerical study, we find out that arbitrary array imperfec- The RAB-SDP in (4) works well when the ideal array covariance
tions, such as sensor location errors, sensor phase and gain mismatches, R
matrix, , is used. However, in practice, this is difficult to achieve due
degrade the performance of the RAB-SDP. In order to improve the to the following reasons. Firstly, with limited number of array snap-
robustness of RAB-SDP against arbitrary array imperfections and for shots, the sample-averaging method may cause errors in the estima-
the ease of robustness control, in this correspondence, worst-case op- R
tion of . For many practical applications, it is more desirable yet
timization [9], [13], [14], [23], [24] is combined with the RAB-SDP. challenging to guarantee the array performance with a small number
The idea is to find out the worst-case effect of the array imperfections of snapshots. Secondly, there exist many array imperfections, such as
on the array output power. The adaptive beamformer is optimized ac- sensor mismatches, sensor location errors, wavefront distortion, etc.
cording to this worst-case array imperfection. As a result, the perfor- The estimated covariance matrix may deviate from the ideal one in a
mance of the resultant beamformer can be insensitive to any kind of significant manner. These sources of errors may seriously degrade the
array imperfections. This new beamformer is called robust adaptive performance of the designed adaptive beamformer.
5916 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 11, NOVEMBER 2010

Most of the array imperfections are difficult to be formulated in a method, the optimal weighting vector can be generated not only
simple mathematical model. It is also hard to find a mathematical model for ULA but also for non-uniform linear array whose inter-element
R
for the effects of these errors on the estimation of . In this correspon- spacings are multiple of a base distance.
E
dence, we define an error matrix, , as the difference between the es-
R
timated array covariance matrix, ^ , and the ideal one, , i.e., R III. DISCUSSIONS ON RAB-SDP-WC
E = R 0 R^ : (5) A. Existence of Solution of RAB-SDP-WC
E
The matrix is Hermitian and the only constraint on is its limited E As stated above, the RAB-SDP [17] is a special case of
E
Frobenius norm, i.e., k kF  ", where " is a positive real number. RAB-SDP-WC with " = 0. When the array has imperfections,
Substituting (5) into (4), the robust beamforming problem is refor- the RAB-SDP may suffer from robustness problem. In the worst case,
mulated as the RAB-SDP method does not have a solution [18]. In this section,

W tr (R^ + 2E)W
we show that the RAB-SDP-WC is guaranteed to have a solution if
min (6a)
the constraints are feasible and the parameter " is large enough even if
s.t. tr fRs ( )Wg  8 2 [ l u] the true norm of error matrix is unknown.
R R E
(6b)
0 . Since
 U ;   ;
According to (5), the true covariance matrix ^ =
tr fRs ( )Wg  2 8 2 [ l u ]
 L ;   ; (6c)
E W EW E W
0k kF k kF  trf g  k kF k kF , the cost function in
W0 (6d) (9a) follows
kEkF  ": (6e) RW W
trf ^ g + "k kF = trf RW EW
g 0 trf g + "k kF W
To guarantee the performance of beamformer against all kinds of  ( 0 kEkF ) kWkF
" (10)
array imperfections, in this correspondence, we use the idea of the
worst-case optimization [9], [13] to minimize the array output power
using the fact that the true output power of the array trf RW
g is non-
E
when the error matrix, , brings the worst-case effect on the objective
negative. If " is chosen larger than the norm of error matrix, for any
E
error matrix , the cost function of problem (9) is positive and lower
function. The optimization problem is expressed as
bounded. The optimal solution can be found using numerical methods.
W E R EW
min max tr ( ^ + ) (7a) Some suggestions on selection of " can be found in Section IV-A-2.
s.t. tr fRs ( )Wg 

2
U 8 2 [ l u]
;   ; (7b) B. Relationship Between RAB-SDP-WC and RAB-CMR-WC
tr fRs ( )Wg 

2
L ; 8 2 [ l u]   ; (7c) A robust beamformer based on constraints on magnitude responses
W0 (7d) and worst-case optimization (RAB-CMR-WC) was derived in [23] as
kEkF  ": (7e) a regularized linear programming problem. In [17], we have proved
that the RAB-SDP is equivalent to RAB-CMR [23] in the case of a
With this new optimization formulation, the performance of the beam- ULA. In this section, we will show that RAB-SDP-WC also has close
former is guaranteed. relationship with the RAB-CMR-WC for a ULA.
In (7a) In the cost function of RAB-SDP-WC, there is a new term "k kF W
E
R E W = trfRW
max tr ( ^ + ) ^ g + max trfEWg
E
(8) compared with RAB-SDP. The RAB-CMR-WC also differs from
r
RAB-CMR in a term "k w k2 , where
i.e., the error matrix E only affects the second term. Since E and W r w = [rw (0N + 1) 1 1 1 rw (0) 1 1 1 rw (N 0 1)]
T (11)
are both Hermitian matrices, it is straightforward to prove that the
0 Wi;i+m , rw (0m) = rw (m), m  0 is
trace of their product is a real value, i.e., trfEWg = trfWH EH g, and rw (m) = N jmj01 3
i=0
and trfEWg  kEkF kWkF . Therefore, the maximum value of the autocorrelation function of wi , Wi;j is the ith row and j th column
trfEWg is achieved when the error matrix E is a positive multiple of element of W .
WH , i.e., E = WH , where is a positive real number. Considering For any given m, we have
the norm constraint on E, the worst-case error matrix E is obtained as
2
N 0jmj01 N 0jmj01
E = (WH kWkF ).
" = jrw (m)j =
2
Wi;i+m  (N 0 jmj) jWi;i+m j2 ;
Hence, the objective function in (7a) is simplified and the proposed i=0 i=0
RAB-SDP-WC is given as (12)
RW W
therefore,
min trf ^ g + "k kF (9a) N 01 N 0jmj01
W
s.t. tr fRs ( )Wg 

2
U 8 2 [ l u]
;   ; (9b)
W
k kF2 = jWi;i+m j2
m=0N +1 i=0
tr fRs ( )Wg  2
8 2 [ l u] p p p
1 1 1 2 1g 1rw 2 (13)
(9c) 2
 diagf1; 2; 1 1 1 ;
 L ;   ;

W0
0
N; ; ; :
: (9d)
Compared with the RAB-SDP in (4), the RAB-SDP-WC in (9) only From (13), the second-order term k kF of RAB-SDP-WC has sim-
W

differs in the objective function, where a regularization term is added. ilar effect in optimization as the term krw k2 in RAB-CMR-WC. Since
Its objective function in (9a) is an affine second-order cone function. k kF is greater than the weighted norm of rw as shown in (13), in
W

Meanwhile, the RAB-SDP is a special case of the RAB-SDP-WC when practice, to achieve similar robustness of the beamformer, the param-
" = 0. The parameter " is called regularization factor in the following eter " for RAB-SDP-WC can be selected smaller than " for RAB-
context. CMR-WC. However, the exact relationship between these two param-
Similar as [17], the semi-infinite constraints (9b), (9c) are ap- eters is difficult to reveal. We suggest to select this parameter by exper-
proximated in a straightforward way by sampling or discretizing the iment in practice.
angle. Although the RAB-SDP-WC is derived using matrix weighting
form, its optimal weighting vector can be obtained from the optimal C. Advantages and Disadvantages of RAB-SDP-WC
semi-definite matrix W by using the extended spectral factorization The robustness problem of RAB-SDP against arbitrary array imper-
method [17], [22]. With the help of the extended spectral factorization fection is solved by the proposed RAB-SDP-WC method. Unlike the
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 11, NOVEMBER 2010 5917

RAB-CMR-WC can only work for a ULA, the RAB-SDP-WC can deal
with more kinds of arrays, like linear array whose inter-element spac-
ings are multiple of a base distance, with the help of the spectral fac-
torization for non-uniform sampled sequences [22].
As a worst-case extension of the RAB-SDP, compared with the
other robust beamformers in [9]–[11], [13], [15], and [16], the
RAB-SDP-WC inherits from RAB-SDP the capability of producing
a robust response region with controlled beamwidth and response
ripple. When the desired robust response region is large, the output
SINR improvement of RAB-SDP-WC is significantly higher than the
other approaches (see numerical results in Section IV).
RAB-SDP-WC also has some disadvantages. For example, it does
not provide a closed-form solution. The optimal solution must be ob-
tained using numerical method [18]. The current method is derived
based on linear array, for an array with arbitrary geometry, it is difficult
to obtain the optimal array weight vector. However, like the approach
in [15], it can still be used in the application of power estimation. Com-
pared with the conventional adaptive beamformers, the RAB-SDP-WC
ON M
has computational load higher than the order of ( 4 2:5 ) [25],
where M denotes the number of constraints.
Fig. 1. Beampatterns of RAB-SDP-WC with various regularization factors.

IV. NUMERICAL RESULTS

The adaptive beamformers are usually derived based on the criterion
of minimization of the output power while its performance is always
evaluated according to the improvement of output SINR. Except some
special cases as well as the adaptive beamformers with closed-form
solution, theoretical performance analysis of general adaptive beam-
formers is difficult. The proposed RAB-SDP-WC also has no closed-
form solution. Like most of the state-of-the-art adaptive beamformers
[9]–[11], [13], [15], [16], we use numerical experiments to evaluate
the performance of RAB-SDP-WC against various randomly gener-
ated array imperfections [23].
The performance of the RAB-SDP-WC for a ULA is studied
in Section IV-A. Some numerical results on a non-uniform linear
array (NULA) are presented in Section IV-B. In all the simulations,
the array beampatterns are obtained from one experiment. The output
power and SINR are calculated by the average of 200 Monte Carlo
results.

A. Performance Study on RAB-SDP-WC for ULA

Fig. 2. Output SINR of the RAB-SDP-WC versus regularization factors with
A ULA consisting of 10 sensors with half-wavelength inter-element different number of snapshots ( = 25).
spacing is used. In all the experiments, the non-directional noise is
assumed to be a spatially white Gaussian noise with a unit covari-
r
ance. For a predetermined ripple, db , its upper and lower limits are response ripple are well controlled by the RAB-SDP-WC even if the
given by U = 10
r =20
and = 100r =20 . Since the parameter
L " regularization factor changes significantly.
bounds the norm of the error matrix, E, it changes with the power of 2) Sensitivity of Output SINR Versus Regularization Factor: The
input signal. To remove this ambiguity, we use a relative regulariza- regularization factor should be chosen to be above the norm of the error
"
tion factor r = "N=trf g R^ , i.e., the normalized regularization factor matrix E caused by the tolerable array imperfections. The selection
to the power of array received signals. Two directional interferences of the regularization factor affects the performance of the proposed
have the power of 20 dB and the directions-of-arrival (DOAs) 70 and beamformer. However, to obtain an accurate estimate of the norm of
140 , respectively. The power and DOA of the target signal are 20 dB E is quite difficult in practice. It is desirable if the output SINR is
and 93 . In the following context, we use such simulation settings if "
inert to r in a wide range. In this experiment, we study the sensi-
not explicitly indicated. tivity of output SINR to regularization factor in the RAB-SDP-WC.
1) Beamwidth and Ripple Control Versus Regularization Factor: The array is assumed to have random phase errors with standard devi-
In this experiment, we evaluate the performance on the accuracy of  "
ation  = 5 . The relative regularization factor r varies from 0.0 to
beamwidth and ripple control of RAB-SDP-WC with various regular- 5. The output SINR of RAB-SDP-WC with various numbers of array
ization factors. The RAB-SDP is a special case of the RAB-SDP-WC snapshots versus regularization factor are shown in Fig. 2. It is clear
"
with zero regularization factor, i.e., r = 0. The allowable response that even with different number of array snapshots, the performance
ripple in the robust region is fixed as 0.3 dB. The beamwidth of robust of RAB-SDP-WC is not very sensitive to the selection of regulariza-
response region is 20 . The assumed direction of the target signal is tion factor if it is in a reasonable range above a threshold determined
"
90 . With the relative regularization factor, r , varying from 0.0 to 0.8, by the norm of the tolerable error matrix E. In this experiment from
the resultant array beampatterns are shown in Fig. 1. Examining the re- " :
Fig. 2, the threshold can be r = 0 25. If the chosen regularization
sponses in the configured region, we find out that the beamwidth and factor increase to 20 times larger than the threshold, the performance
5918 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 11, NOVEMBER 2010

Fig. 4. Comparison of output SINR among the robust methods with a

Fig. 3. Output SINR of RAB-SDP-WC versus arbitrary ASV error with various beamwidth of robust response region 20 , versus ASV error (SC: sidelobe
regularization factors. control).

TABLE I
of RAB-SDP-WC degrades slightly. This is an attractive property of COMPARISON OF ESTIMATED OUTPUT POWER AND SINR AMONG
the RAB-SDP-WC in practical applications. ROBUST ADAPTIVE BEAMFORMERS (THE TRUE POWER OF THE
3) Output SINR Versus Various Kinds of Array Imperfections: In the TARGET SIGNAL IS 20 dB)
following experiments, we evaluate the performance of RAB-SDP-WC
considering different kinds of array imperfections. Without loss of gen-
erality, we only show the SINR output against arbitrary ASV error in
Fig. 3. It indicates that the RAB-SDP-WC is robust against these array
imperfections if the regularization factor is properly selected. In the
case of small array imperfection, the output SINR of RAB-SDP-WC
with large regularization factor is slightly below the others. However,
a large regularization factor is necessary to guarantee the robustness of
beamformer against significant array imperfections.
4) Performance Comparison With Other Robust Methods: In this number of snapshots, the output SINRs of RAB-SDP-WC and AMA
section, the performance of RAB-SDP-WC is compared with the other degrade significantly ("r = 0:2). Although the output SINR of RCB
state-of-the-art robust adaptive beamformers, including RCB [11],
and Shah’s method change little in such case, for an array designed to
AMA [15], Shahbazpanahi’s method [13] and Chen’s method [16] tolerate a large steering direction error, their output SINR is lower than
considering various kinds of array imperfections. RCB method is
that of the proposed method.
selected for comparison because it is a good representative of the
recently proposed robust beamforming methods [9]–[11]. B. Performance Study of RAB-SDP-WC for NULA
In Fig. 4, the SINR outputs of these methods are compared when ar-
bitrary ASV error occurs. The RAB-SDP-WC with various regulariza- Regarding NULA beamforming, a potential candidate is the RAB-
tion factors achieves superior SINR improvement. The performance of SDP-WC with the extended spectral factorization method [22]. In this
RAB-SDP-WC is guaranteed if a suitable regularization factor is used. section, we show some results of the RAB-SDP-WC applied to non-
Since the performance of the proposed method is not very sensitive to uniform integer-interval spaced linear array. The array is similar to the
the change of "r as discussed, it is very suitable for practical applica- ULA used in above experiments except that the inter-element space
tions. The AMAs with/without sidelobe control are included in com- between the last two sensors is 2d instead of d. Some simulation results
parison in Fig. 4. It is clear that the robustness of AMA can be improved shown in Figs. 5 and 6. reveal that the RAB-SDP-WC also possesses
by imposing sidelobe control. However, the sidelobe control requires excellent performance on robust region control and SINR improvement
heavy computational load. The insensitivity of the other methods to- in the case of a NULA.
wards array imperfections results from the fact that the uncertainty set
for a large DOA error already covers the variations of the ASV error V. CONCLUSION
caused by these array imperfections. In this correspondence, a novel robust adaptive beamformer in the
We also compare the accuracy of signal power estimation using dif- form of a semi-definite programming (SDP) problem is proposed
ferent number of snapshots. The beamwidths of robust response region using the idea of worst-case optimization. With the SDP formulation,
are set as 20 . The target signal has steering direction error of 8 and constraints on magnitude response are easily imposed on the adaptive
its true power is 20 dB. The power of the interference at 140 is 30 dB beamformer to achieve a large robust response region. The resultant
which is higher than that of the target signal. The estimated output beamformer also demonstrates outstanding robustness against arbi-
powers and SINRs of the robust beamformers are shown in Table I. trary array imperfections. The robustness of the adaptive beamformer
These results show that the RAB-SDP-WC gives accurate estimate of is controlled by a regularization factor that can be chosen from a large
signal power and produces high output SINR ("r = 0:001). With small range above a certain threshold. The significant SINR improvement
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 11, NOVEMBER 2010 5919

[3] M. H. Er and A. Cantoni, “Derivative constraints for broadband

element space antenna array processors,” IEEE Trans. Acoust.,
Speech, Signal Process., vol. ASSP-31, no. 6, pp. 1378–1393, Dec.
1983.
[4] M. H. Er and A. Cantoni, “A new set of linear constraints for broadband
time domain element space processor,” IEEE Trans. Antennas Propag.,
vol. AP-34, no. 3, pp. 320–329, Mar. 1986.
[5] K. L. Bell, Y. Ephraim, and H. L. Van Trees, “A Bayesian approach to
robust adaptive beamforming,” IEEE Trans. Signal Process., vol. 48,
no. 2, pp. 386–398, Feb. 2000.
[6] H. Cox, R. M. Zeskind, and M. M. Owen, “Robust adaptive beam-
forming,” IEEE Trans. Acoust., Speech, Signal Process., vol. ASSP-35,
no. 10, pp. 1365–1376, Oct. 1987.
[7] B. D. Carlson, “Covariance matrix estimation errors and diagonal
loading in adaptive arrays,” IEEE Trans. Aerosp. Electron. Syst., vol.
24, pp. 397–401, Jul. 1988.
[8] L. Chang and C. C. Yeh, “Performance of DMI and eigenspace-based
beamformers,” IEEE Trans. Antennas Propag., vol. 40, no. 11, pp.
1336–1347, Nov. 1992.
[9] S. A. Vorobyov, A. B. Gershman, and Z. Q. Luo, “Robust adaptive
Fig. 5. Performance of beamwidth and response ripple control of beamforming using worst-case performance optimization: A solution
RAB-SDP-WC with varying regularization factors and a fixed beamwidth 20 to the signal mismatch problem,” IEEE Trans. Signal Process., vol. 51,
for a NULA. no. 2, pp. 313–324, Feb. 2003.
[10] R. G. Lorenz and S. P. Boyd, “Robust minimum variance beam-
forming,” IEEE Trans. Signal Process., vol. 53, no. 5, pp. 1684–1696,
May 2005.
[11] J. Li, P. Stoica, and Z. Wang, “On robust Capon beamforming and
diagonal loading,” IEEE Trans. Signal Process., vol. 51, no. 7, pp.
1702–1715, Jul. 2003.
[12] N. K. Jablon, “Adaptive beamforming with the generalized sidelobe
canceller in the presence of array imperfection,” IEEE Trans. Antennas
Propag., vol. 34, no. 8, pp. 996–1012, Aug. 1986.
[13] S. Shahbazpanahi, A. B. Gershman, Z. Q. Luo, and K. M. Wong,
“Robust adaptive beamforming for general-rank signal models,”
IEEE Trans. Signal Process., vol. 51, no. 9, pp. 2257–2269, Sep.
2003.
[14] S. Verdú and H. V. Poor, “Minimax robust discrete-time matched fil-
ters,” IEEE Trans. Commun., vol. COM-31, no. 2, pp. 208–215, Feb.
1983.
[15] J. Li, Y. Xie, P. Stoica, X. Zheng, and J. Ward, “Beampattern synthesis
via a matrix approach for signal power estimation,” IEEE Trans. Signal
Process., vol. 55, no. 12, pp. 5643–5657, Dec. 2007.
[16] C. Y. Chen and P. P. Vaidyanathan, “Quadratically constrained beam-
forming robust against direction-of-arrival mismatch,” IEEE Trans.
Signal Process., vol. 55, no. 8, pp. 4139–4150, Aug. 2007.
[17] Z. L. Yu, M. H. Er, and W. Ser, “Novel adaptive beamformer based
on semidefinite programming (SDP) with constraints on magnitude re-
Fig. 6. Comparison of output SINR among the robust methods with a
beamwidth of robust response region 20 , versus sensor phase error for a sponse,” IEEE Trans. Antennas Propag., vol. 56, no. 5, pp. 1297–1307,
NULA. May 2008.
[18] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge,
U.K.: Cambridge Univ.Press, 2004.
[19] A. H. Sayed and T. Kailath, “A survey of spectral factorization
of the proposed method results from the fact that the inequality methods,” Numer. Linear Algebra Appl. vol. 8, no. 6–7, pp. 467–496,
magnitude constraints do not consume many DOFs for interference 2001 [Online]. Available: citeseer.ist.psu.edu/sayed01survey.html
suppression. Compared with the state-of-the-art beamformers, the [20] J. Liu, A. B. Gershman, Z.-Q. Luo, and K. M. Wong, “Adaptive beam-
proposed method has flexibility in robust response region control, forming with sidelobe control: A second-order cone programming ap-
superior performance of SINR improvement, and robustness against proach,” IEEE Signal Process. Lett., vol. 10, no. 11, pp. 331–334, Nov.
2003.
arbitrary array imperfections. With these attractive properties, it is a [21] J. Riba, J. Goldberg, and G. Vázquez, “Robust beamforming for in-
promising adaptive beamformer for practical applications. terference rejection in mobile communications,” IEEE Trans. Signal
Process., vol. 45, no. 1, pp. 271–275, Jan. 1997.
ACKNOWLEDGMENT [22] Z. L. Yu, M. H. Er, W. Ser, and Z. Gu, “Spectral factorization for in-
The authors acknowledge all the anonymous reviewers for their teger-interval sampled sequence with applications in array processing,”
constructive comments that helped to improve the quality of this Signal Process., vol. 88, no. 7, pp. 1715–1724, Jul. 2008.
[23] Z. L. Yu, W. Ser, M. H. Er, Z. Gu, and Y. Li, “Robust adaptive
correspondence. beamformers based on worst-case optimization and constraints on
magnitude response,” IEEE Trans. Signal Process., vol. 57, no. 7, pp.
REFERENCES 2615–2628, Jul. 2009.
[1] J. E. Hudson, Adaptive Array Principles. Stevenage, U.K.: Pere- [24] Z. L. Yu, M. H. Er, and W. Ser, “Novel adaptive antenna array based on
grinus, 1981. robust semidefinite programming,” in Proc. ICC, Beijing, China, May
[2] K. Takao, M. Fujita, and T. Nishi, “An adaptive antenna array under 2008, pp. 825–829.
directional constraint,” IEEE Trans. Antennas Propag., vol. 24, no. 5, [25] L. Vandenberghe and S. Boyd, “Semidefinite programming,” SIAM
pp. 662–669, Sep. 1976. Rev., vol. 38, pp. 49–95, Mar. 1996.