Beamforming for Antenna Array

www.ietdl.org

Published in IET Signal Processing
Received on 23rd August 2007
Revised on 19th April 2008
doi: 10.1049/iet-spr:20070162

ISSN 1751-9675

Efficient implementation of robust adaptive
beamforming based on worst-case
performance optimisation
A. Elnashar
Emirates Integrated Telecommunications Company (du), Mobile Access Network, Technology Department, Dubai Media City,
P.O. Box 502666, Dubai, UAE
E-mail: nashar_eg@yahoo.com

Abstract: Traditional adaptive beamforming methods undergo serious performance degradation when a
mismatch between the presumed and the actual array responses to the desired source occurs. Such a
mismatch can be caused by desired look direction errors, distortion of antenna shape, scattering due to
multipath, signal fading as well as other errors. This mismatch entails robust design of the adaptive
beamforming methods. Here, the robust minimum variance distortionless response (MVDR) beamforming
based on worst-case (WC) performance optimisation is efficiently implemented using a novel ad hoc adaptive
technique. A new efficient implementation of the robust MVDR beamformer with a single WC constraint is
developed. Additionally, the WC optimisation formulation is generalised to include multiple WC constraints
which engender a robust linearly constrained minimum variance (LCMV) beamformer with multiple-beam WC
(MBWC) constraints. Moreover, the developed LCMV beamformer with MBWC constraints is converted to a
system of nonlinear equations and is efficiently solved using a Newton-like method. The first proposed
implementation requires low computational complexity compared with the existing techniques. Furthermore,
the weight vectors of the two developed adaptive beamformers are iteratively updated using iterative
gradient minimisation algorithms which eliminate the estimation of the sample matrix inversion. Several
scenarios including angle-of-incidence mismatch and multipath scattering with small and large angular spreads
are simulated to study the robustness of the developed algorithms.

1 Introduction directions errors, uncertainty in array sensor positions,
mutual coupling, imperfect array calibration, multipath
Adaptive beamforming is a versatile approach to detect and propagation due to local and remote scattering and limited
estimate the signal of interest (SOI) at the output of a sample support.
sensor array with applications in wireless communications,
radar, sonar, astronomy, seismology, medical imaging and Many approaches have been proposed during the last two
microphone array speech processing. Unfortunately, decades to improve the robustness of the traditional
traditional adaptive array algorithms are known to be beamforming methods. A survey on these approaches can
extremely sensitive even to slight mismatch between the be found in [2, 3] and references therein. Among those
presumed and the actual array responses to the desired approaches, the worst-case (WC) performance optimisation
signal [1]. Whenever a mismatch occurs, the adaptive has been shown as a powerful technique which yields a
beamformer inclines to misconstrue the SOI components beamformer with robustness against an arbitrary signal
in the array observations as interference and hence steering vector mismatch, data non-stationarity problems
suppressing these components is most likely expected. The and small sample support [3 – 10]. The WC approach
errors in array response to SOI can take place due to look explicitly models an arbitrary (but bounded in norm)

IET Signal Process., 2008, Vol. 2, No. 4, pp. 381 – 393 381
doi: 10.1049/iet-spr:20070162 & The Institution of Engineering and Technology 2008

www.ietdl.org

mismatch in the desired signal array response and uses the beamformer with MBWC constraints entails solving a set
WC performance optimisation to improve the robustness of nonlinear equations. As a consequence, a Newton-like
of the minimum variance distortionless response (MVDR) method is mandatory to solve the ensuing system of
beamformer [5]. A theoretical analysis for this class of nonlinear equations which yields a vector of Lagrange
robust beamformers in terms of signal-to-interference-plus- multipliers. It is worthwhile to note that the approaches in
noise ratio (SINR) in the presence of random steering [15, 17] adopt ad hoc techniques to optimise the
vector errors is presented in [11, 12]. In addition, the beamformer output power with spherical constraint on the
closed-form expressions for the SINR are derived therein. steering vector. Unfortunately, the adaptive beamformer
developed in [17] is apt to noise enhancement at low SNR
Unfortunately, the natural formulation of the WC and additional constraint is required to bear the ellipsoidal
performance optimisation involves the minimisation of a constraint [15].
quadratic function subject to infinity non-convex quadratic
constraints [5]. The approaches in [5, 6] reformulated the The rest of the paper is organised as follows. In Section 2,
WC optimisation as a convex second-order cone program the standard MVDR and LCMV beamformers with single
(SOCP) and solved it efficiently via the well-established and multiple constraints are summarised in the context of a
interior point method [13]. Regrettably, the SOCP method single point source and a source with multipath rays,
does not provide a closed-form solution for the respectively. In Section 3, the WC optimisation formulation
beamformer weights and even it cannot be implemented is introduced by summarising general and special
online, whereas the weight vector needs to be recomputed formulations for the steering vector uncertainty set. Efficient
completely with the arrival of a new array observation implementations of single and multiple WC formulations
[14 – 16]. are derived and analysed in Section 4 and Section 5,
respectively. Moreover, a geometric illustration for the single
Attractive approaches based on eigendecomposition of the WC implementation is presented. Simulations and
sample covariance matrix have been introduced in [7 – 10]. performance analysis are provided in Section 6. Conclusions
These approaches developed a closed-form solution for a and points for future work are encapsulated in Section 7.
WC robust detector using the Lagrange method which
incorporates the estimation of the norm of the weight
vector and/or the Lagrange multiplier. A binary search
2 Standard beamforming
algorithm followed by a Newton-like algorithm is proposed methods
in [4, 8] to estimate the norm of the weight vector after Consider an array comprising M uniformly spaced sensors
dropping the Lagrange multiplier. Although these receives a narrowband signal sd (k). Initially, it is assumed
approaches have provided closed-form solutions for the that the desired signal is a point source with time-invariant
WC beamformer, they, unfortunately, incorporate several wavefront, and the M Â 1 vector of array observations can
difficulties. First, eigendecomposition for the sample be modelled as [2 – 5]
covariance matrix is required with the arrival of a new array
observation. Second, the inverse of diagonally loaded x(k) ¼ ad (w)sd (k) þ i(k) þ n(k) (1)
sample covariance matrix is required to estimate the weight
vector. Third, some difficulties are encountered during where k is the time index, sd (k) the complex signal waveform
algorithm initialisation and a stopping criterion is necessary of the desired signal and ad (w) its M Â 1 steering vector
to prevent negative solution of the Newton-like algorithm. where w is the angle of incidence (AOI) and i(k) and n(k)
the statically independent components of the interference
In this paper, two efficient ad hoc implementations of the and the noise, respectively.
WC performance optimisation problem are adopted. First,
the robust MVDR beamformer with a single WC A generalised model with multipath propagation can be
constraint is implemented using an iterative gradient expressed as follows
minimisation algorithm with an ad hoc technique to
estimate the Lagrange multiplier instead of the Newton- X
L
like algorithm. The proposed algorithm exhibits several x(k) ¼ sd (k) gn ad (w þ fn ) þ i(k) þ n(k) (2)
merits including simplicity, low computational load and no n¼1
need for either sample-matrix inversion or
eigendecomposition. A geometric interpretation of the where L is the number of multipaths with each path has a
proposed implementation is introduced to supplement the random complex gain gn and an angular deviation fn from
theoretical analysis. Second, a robust linearly constrained the nominal AOI w. The scattered signals associated with
minimum variance (LCMV) beamformer with multiple- the multipath propagation from a single source arrive at the
beam WC (MBWC) constraints is developed using a novel base station (BS) from several directions within an angular
multiple WC constraints formulation. The Lagrange region called the angular spread. The angular spread arises
method is exploited to solve this optimisation problem, due to the multipath, both from local scatters near to the
which reveals that the solution of the robust LCMV source and near to the BS and from remote scatters.

382 IET Signal Process., 2008, Vol. 2, No. 4, pp. 381– 393
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070162

www.ietdl.org

It varies according to the cell morphological type (i.e. dense The minimum variance beamformer with single-beam
urban, urban and rural), cell radius, BS location and constraint (SBC) can be formulated as follows
antenna height. It can vary from few degrees at rural road
_
cells to 3608 in microcellular and indoor environment due min w H Rw
to the reflecting surfaces surround the BS antenna. It is w
(6)
assumed that the time delays of the different multipath subject to H
w ad ¼ 1
components are small compared with the inverse of the
signal bandwidth (i.e. narrowband channel model) and The solution of (6) engenders the standard MVDR
therefore the delay can be modelled as a phase shift in the beamformer with SBC and can be easily derived as
complex gain gn [18]. The angular spread is used here to
describe the angular region associated with the entire _ À1
multipaths. Notwithstanding each of the rays itself may be R ad
wSBC ¼ _ À1
(7)
composed of a large number of ‘mini-rays’ with roughly
aH R ad
d
equal angles and delays but with arbitrary phases due to
scattering close to the source [19]. In this paper, the model
is simplified by using the nominal AOI of each ray group Considering the generalised received signal model in (2), the
and multipath delays are modelled as a small angle in the optimum MVDR beamformer can be obtained using
complex gain. multiple constraints to provide multiple-beam constraint
(MBC) beamformer, that is
The beamformer output signal can be written as _
min wH Rw
w
(8)
H H
y(k) ¼ w (k)x(k) (3) subject to w L(u0 ) ¼ v
Â Ã Â Ã
Â ÃT where u0 ¼ u1 Á Á Á uL and L(u0 ) ¼ a(u1 ) Á Á Á a(uL ) is
where x(k) ¼ x1 (k), . . . , xM (k) Â is an M Â 1 complex vector
ÃT the M Â L spatial constraint matrix consists of the steering
of the array observations, w(k) ¼ w1 (k), . . . , wM (k) is an vectors corresponding to the AOIs of the multipath rays
M Â 1 complex vector of the beamformer weights and (Á)T associated with the desired source and v is a vector of
and (Á)H stand for the transpose and Hermitian transpose, the constrained values (i.e. gain vector) which can all be
respectively. set to one for equal gain combining or alternatively it
can be optimised using maximal ratio combining
Consider the simplified model in (1) with the point source. (MRC) technique. Accordingly, the optimal weight
The optimal weight vector seeks maximisation of the output vector of (8), termed as the LCMV beamformer, is given
SINR [3– 5, 10, 11] by [10]

2 À1
_ À1 _ À1
s2 wH ad
d wMBC ¼ R L LH R L n (9)
SINR ¼ (4)
wH Riþn w

n o
D
where Riþn ¼ E ði(k) þ n(k)Þði(k) þ n(k)ÞH is the 3 Robust adaptive MVDR
interference-plus-noise covariance matrix and s2
d is the beamformer with single WC
desired source power. The optimal solution of w which
maximises the output SINR in (4) can be obtained by
constraint
maintaining distortionless response to the desired source The beamforming formulations in (6) and (8) assume that
while minimising the output interference-plus-noise power the array response to the desired source (i.e. the steering
(i.e. wH Riþn w). In practical applications, the interference- vector ad of the point source or the spatial matrix L of a
plus-noise covariance matrix can be replaced by the sample source with multipath rays) is precisely known. However,
covariance matrix [2– 10], which can be estimated using practically, the knowledge of the desired source steering
the first-order recursion vector or spatial matrix may be imprecise. In this paper, the
recently emerged rigorous approach to robust MVDR
beamforming based on the WC performance optimisation
_ X
n _
[3 – 10] is considered.
R(n) ¼ hnÀi x(i)xH (i) ¼ hR(n À 1) þ x(n)xH (n) (5)
i¼1
First consider the formulation of the standard MVDR
beamformer in (6) with SBC, and following the approaches
where h is a forgetting factor which satisfies 0 ( h 1. in [5 – 10], to add robustness to the standard MVDR

doi: 10.1049/iet-spr:20070162 The Institution of Engineering and Technology 2008

www.ietdl.org

beamformer in (7), the WC weighted power output of the respectively, [10] and [5]
array is minimised in the presence of uncertainties in the
steering vector, that is _
min wH Rw
w

_
min wH R w subject to wH a d ! A H w þ 1 (17)
w
(10) _
H min wH Rw
subject to w z ! 1 8z [ 1 w

subject to wH ad ! zkwk þ 1 (18)
where 1 is an ellipsoid that covers the possible range of the
imprecise steering vector z. Assuming 1 is centred at the The constraints in (17) and (18) are called second-order cone
presumed steering vector ad [10], that is constraints. Two SOCP approaches are proposed in [6] and
[5] for real and complex formulations, respectively.
È É
1 ¼ Au þ ad jkuk 1 (11)

where the matrix A determines the size and shape of the 3.1 Lagrange approach
ellipsoid 1. If A ¼ zI [5– 8] is set, the following special First form the following Lagrange function
case of 1 is obtained
_
È É J (w, l) ¼ wH Rw À lt wH ad À zkwk À 1 (19)
1 ¼ e þ ad jkek z , e ¼ zu (12)

_ where t(Á) is a step function guarantees that
Assuming that R in (10) is a positive definite matrix and then wH ad ! zjwj þ 1 and l is the Langrage multiplier. The
the optimisation problem in (10) along with the generalised inequality constraint in (18) is satisfied by equality if the
ellipsoid in (11) can be converted to the following form [10] cost function in (19) is minimised. This fact can be proved
by contradiction [4, 8] and hence the step function in (19)
_ is dispensable. By differentiating (19) and equating the
min w H Rw
w result to zero, one has [4]
(13)
H H
subject to w a d ! A w þ 1 _ w
Rw þ lz ¼ lad (20)
kwk
Likewise, the optimisation problem in (10) with the WC
constraint in (12) can be expressed as [5] By solving for w, the following closed-form solution is
obtained
_
min wH R w !À1
w lz_
(14) wWC ¼l Rþ
w I ad (21)
H WC
subject to w ad ! zkwk þ 1

The WC Robust MVDR beamformer in (21) encompasses
Unfortunately, the nonlinear constraints in (13) and (14) are three difficulties as follows: the estimation of the weight
non-convex due to the absolute value function on the left- vector norm, the estimation of the Lagrange multiplier
hand side. Indeed, the cost functions in (13) and (14) are which achieves (21) and the computational load of
unchanged when w undergoes an arbitrary phase rotation computing the inverse of the diagonally loaded sample
[5 – 10]. As a consequence, with the optimal solutions of covariance matrix. In the following two sections, two
(13) and (14), it can always rotate without affecting the techniques are summarised for computing wWC .
cost function optimisation. Therefore the optimal solution
may be chosen, without loss of generality, such that
3.1.1 Eigendecomposition method: Several eigen
n o decomposition approaches have been developed to solve the
Re wH ad ! 0 (15) WC performance optimisation problem. The optimisation
n o problems in (17) and (18) have been solved, respectively, in
Im wH ad ¼ 0 (16) [10] and [4, 8] using eigendecomposition methods. For the
sake of comparison, the approach in [4, 8] is briefly
reviewed. Using the fact that multiplying wWC in (20) by
Using (15) and (16), the optimisation problems in (13) and any arbitrary constant does not affect the bit error rate
(14) can be converted to the following convex formulations, performance of the beamformer [4, 8], a scaled version of

The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070162

www.ietdl.org

the WC beamformer can be obtained as follows [4] determined at snapshot k and m(k) an adaptive step size
which determines the convergence speed of the algorithm.
!À1
_ z
wWC ¼ R þ
~
I ad (22) The gradient vector of the cost function in (19) is given by
wWC
~

@J (w, l) _ a d À zw
A binary search algorithm followed by Newton– Raphson r¼ ¼ Rw À l (26)
@w kwk
iterations is proposed in [8] to compute the norm of the

WC beamformer wWC . The eigendecomposition
~
The step function is dropped due to ad hoc adaptive
approach accurately estimates the norm of the robust
implementation. Hence, the adaptive weight vector can be
detector wWC and hence the optimal weight vector can
~
obtained by substituting (26) into (25), which yields
be obtained using the closed form in (22).
_
3.1.2 Taylor series approximation method: By w(k þ 1) ¼ w(k) À m(k)R(k)w(k)

applying the Taylor series expansion to (22) analogous to a À zw(k)
the approach in [20], the following þ m(k)l d (27)
kw(k)k
!À1 _
z _ À1 _ À1 For simplicity, two new vectors are introduced: w(k þ 1) ¼
wWC ¼ I þ
~ w R
R ad _
~ WC w(k) À m(k)R(k)w(k) (referred to as unconstrained MV
(23)
_ À1
weight vector) and p(k) ¼ ad À zw(k)=kw(k)k. Therefore
z
’ wSBC À
w
R wSBC
the weight vector of the robust WC adaptive beamformer
~ WC can be updated as follows
_ À1
is obtained, where wSBC ¼ R ad is a biased version of the
_
w(k þ 1) ¼ w(k þ 1) þ m(k)lp(k) (28)
standard MVDR beamformer in (7). By introducing a new
_ À1
vector wSBC ¼ R wSBC and substituting into (23) yield
~ 4.1 Lagrange multiplier estimation
wWC ’ wSBC À kwSBC
~ ~ (24) We assume that the weight vector w(k) satisfies the WC
constraint in (18) and then, w(k þ 1) should also satisfy the
_
where k ¼ z=wWC is a parameter related to the weight
~ WC constraint. The weight vector w(k þ 1) represents the
vector norm of the WC robust beamformer and can be minimisation of the unconstrained MV cost function (i.e.
_
estimated by plugging (24) into the WC constraint in (18). wH Rw) which leads to trivial zero solution if the additional
This approach is almost similar to the eigendecomposition WC constraint is not imposed. In order to fully satisfy the
approach where low complexity is introduced at the inequality constraint in (18), first the weight vector
_ _
expense of the weight vector norm estimation accuracy w(k þ 1) is computed and then it is verified if w(k þ 1)
owing to Taylor series approximation. achieves the WC constraint in (18). Consequently, if
_
w(k þ 1) satisfies the WC constraint, the weight vector is
accepted and the algorithm continues with a new array
4 Efficient implementation of observation. Otherwise, (28) is substituted into the
robust adaptive MVDR beamformer inequality constraint in (18) to estimate the Lagrange
with single WC constraint multiplier as follows
H '
In this section, efficient adaptive implementations of robust _
adaptive MVDR beamformer based on WC performance Re w(k þ 1) þ m(k)lp(k) ad
optimisation is developed. The WC performance (29)
_
optimisation MVDR beamforming is efficiently ! z w(k þ 1) þ m(k)lp(k) þ 1
implemented using iterative gradient minimisation
algorithm with ad hoc technique to satisfy the WC constraint. where Ref†g is inserted to make sure that (15) and (16) are
always guaranteed during adaptive implementation. After
The adaptive beamformer can be found by searching for a arranging and boosting both sides of (29) to the power of two
weight vector w that minimises the cost function (19). In
order to find the target beamformer in an iterative manner, H ' 2
_
the weight vector can be updated as follows Re w(k þ 1) þ m(k)lp(k) ad À 1
(30)
w(k þ 1) ¼ w(k) À m(k)r(k) (25) 2
2 _
! z w(k þ 1) þ m(k)lp(k)
where k is the snapshot index, r(k) the gradient vector of the
Lagrange function J (w, l) in (19) with respect to w is obtained.


www.ietdl.org

By rearranging (30), one has (34), respectively. The positive root is selected to make sure
_ À1
n_ o n o2 that R þ lz=wRMV I is a positive definite matrix.
Re w(k þ 1)H ad À 1 þ m(k)lRe p(k)H ad
_ H _ † A , 0 and B . 0:
! z2 w(k þ 1) þ m(k)lp(k) w(k þ 1) þ m(k)lp(k) Equation (33) has only one real positive root resulting from
(31) positive sign in (34) if B2 . AC.

The Lagrange multiplier l which achieves the WC constraint † A . 0 and B , 0:
in (18) needs to be estimated. During the ad hoc Therefore (33) has two positive real solutions. In this case,
implementation, (31) will be solved only if the WC the smaller root is selected to guarantee algorithm stability.
constraint is not met. Following this fact, the inequality in
(31) is replaced by equality and after some manipulations to † A , 0 and B , 0:
(31), one has Equation (33) is guaranteed to have one real positive solution
if B2 . AC.
2
x2 þ 2m(k)lxp(k)H ad þ m(k)2 l2 p(k)H ad 4.2 Recursive implementation
2 n o _
_ _
¼ z2 w(k þ 1) þ2m(k)z2 lRe p(k)H w(k þ 1) The optimum step size of minimising wH Rw is the best
estimate to the optimum step size which minimises (19).
þ m(k)2 z2 l2 kp(k)k2 (32) As a consequence, the optimum step size can be obtained
by substituting (25) into (19) and differentiating with
n_ o
respect to the adaptive step size, then equating the result to
where x ¼ Re w(k þ 1)H ad À 1.
zero, the following optimum step size is obtained [15, 21, 22]
Therefore the Lagrange multiplier l can be computed as _ _

the solution to the following quadratic equation arH (k)r(k)
mopt (k) ¼ _ _ _ (36)
n o2 rH (k)R(k)r(k)_ _
r¼R(k)w(k)
m(k)2 l2 Re p(k)H ad Àz2 kp(k)k2
n o n o The parameter a is added to improve the numerical stability
_
þ 2m(k)l xRe p(k)H ad À z2 Re p(k)H w(k þ 1) of the algorithm. For a practical system, it should be adjusted
during initial tuning of the system and it should satisfy
2
_ 0 , a , 1 [21, 22].
þ x2 À z2 w(k þ 1) ¼ 0 (33)
To summarise, the proposed WC robust adaptive
Therefore the value of l which achieves the WC constraint in beamformer algorithm consists of the following steps.
(18) has the following form
_
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Step 0. Initialise: R(0) ¼ I , w(0) ¼ ad , a ¼ 0:1, h ¼ 0:97
B2 À AC
l ¼ ÀB + (34)
A Step 1. Pick a new sample from array observations and
compute _
_ the sample covariance matrix:
where R(k) ¼ hR(k À 1) þ x(k)xH (k); M 2 .
n o2
2 H 2 2 Step 2. Compute the optimum step-size using (36); M 2 þ 2M.
A ¼ m(k) Re p(k) ad Àz kp(k)k
n o n _
o Step 3. Update the unconstrained MV weight vector:
_
_
B ¼ m(k) xRe p(k)H ad À z2 Re p(k)H w(k þ 1) w(k þ 1) ¼ w(k) À m(k)R(k)w(k); the matrix vector
_
2 multiplication R(k)w(k) is computed in step 2.
2
2 _
C ¼ x À z w(k þ 1) (35)
_
Step 4. If x , zw(k þ 1) compute l using (34); 5M.
_
The Lagrange multiplier estimation is merely executed when
Else l ¼ 0 and w(k þ 1) ¼ w(k þ 1) ! go to step 1.
_
the WC constraint is not achieved, that is, x , zw(k þ 1).
Therefore C , 0 and the roots of (33) fall under one of the Step 5. Update the WC weight vector as:
_
following categories. w(k þ 1) ¼ w(k þ 1) þ m(k)lp(k) ! go to step 1.

† A . 0 and B . 0: As shown in the above implementation, the total
Equation (33) has two real roots: one positive root and one multiplications complexity of the proposed algorithm is
À Á
negative root resulting from positive and minus signs in about O 2M 2 þ 7M . More interestingly, the WC


www.ietdl.org

optimisation step with the ad hoc implementation requires WC constraint. In a nutshell, the WC constraint prevents
O(M) complexity while it requires O(M 3 ) using SOCP, the weight vector from reaching the trivial zero solution by
and in the eigendecomposition method, the estimation of maintaining distortionless response to a set of possible
the norm vector of the WC weight vector requires O(M 3 ) steering vectors which is controlled by the WC constraint.
alone [3 – 8].
5 Robust LCMV beamforming
4.3 Geometric interpretation of the with MBWC constraints
proposed WC beamformer
The majority of the developed robust techniques in
In order to further illustrate the proposed algorithm, it is beamforming literature are based on single constraint in the
exemplified using a geometric interpretation. Fig. 1 desired look direction [3– 10, 14– 16]. Therefore if the
represents the geometric illustration for the proposed WC desired source experiences multipath propagation and
adaptive beamforming implementation using simple 2D impinging on the antenna array from different angles
ƒ!
case similar to the approach in [23]. The vector OA associated with the dominant multipath rays, the robust
ƒ!
represents the presumed steering vector ad . The vector OB technique with single constraint is not capable of gathering
represents the WC robust beamformer at snapshot k. The all multipath components, especially with large angular
concentric ellipses represent the unconstrained MV cost spread. Alternatively, the robust technique may concentrate
_
function, that is, wH Rw and the centre of these ellipses is only on the nominal AOI and neglect other components
the minimal point (i.e. trivial zero solution) that minimises scattered in different multipaths which is not optimal in
this cost function. Assuming that the WC weight vector terms of optimising the output SINR. As a consequence, it
w(k) satisfies the WC constraint, that is, is worthwhile to generalise the WC robust technique to
ƒ! H ƒ! ƒ! include multiple constraints to form the robust LCMV

OB OA ! z OB þ 1. The forthcoming update of the beamformer with MBWC constraints analogous to the
unconstrained MV weight vector is computed as
_
standard LCMV beamformer with MBC in (9). A
_
w(k þ 1) ¼ w(k) À m(k)R(k)w(k), that is, generalisation for (18) with MBWC constraints can be
ƒ! ƒ! ƒ! ƒ!
OC ¼ OB þ BC . As depicted in Fig. 1, the vector BC expressed as
represents the gradient of the MV cost function, that is,
_
Àu(k)R(k)w(k), which is perpendicularly inward inside the ^H_ ^
min w Rw
contours and towards the centre of the ellipsis. When the w
ƒ! (37)
subsequent vector OC does not satisfy the WC constraint ^H ^
ƒ! H ƒ! ƒ! subject to w L ! y w þ i

(i.e. OC OA , z OC þ 1), the condition in the step 4
in the algorithm is met, and subsequently, the vector where L is an M Â N spatial matrix of the desired source, y a
ƒ! ƒ!
AE ¼ Àzw(k)=kw(k)k, which parallel to the vector BO, 1 ÂN vector consisting of the WC constrained values and i a
ƒ! ƒ! ƒ! ƒ! 1 Â N all-one vector where N is the number of WC
is added to OA to estimate p (k) (i.e. OE ¼ OA þ AE ).
ƒ! constraints (i.e. dominant multipath components, N L).
Then, the WC weight vector OD ¼ w(k þ 1) is generated Then, a generalised cost function corresponding to (19) can
ƒ!
by adding the vector CD ¼ m(k)lp(k), which is parallel to be expressed as
ƒ! ƒ! ƒ! ƒ! ƒ!
the vector OE , to the vector OC , (i.e. OD ¼ OC þ CD ). ^H
ƒ! ^H_ ^ ^
Consequently, the ensuing weight vector OD satisfies the Q(w, t) ¼ w Rw À w L À yw À i t (38)

where t is an N Â 1 vector of Lagrange multipliers. The step
function is dropped due to ad hoc implementation. The
following equations are corresponding to (26) and (28),
respectively

^ ^
!!
@Q(w, t) _^ w
r¼ ^ ¼ Rw À L À y ^ t (39)
@w kwk
^ ¯
w(k þ 1) ¼ w(k þ 1) þ m(k)Lt (k)
~ (40)
^
¯ ^
where L ¼ L À y w(k)=w(k) is an M Â N matrix and
^ _ ^ _
w(k þ 1) ¼ w(k) À m(k)R(k)w(k) is similar to w(k þ 1).
~

Figure 1 Geometric interpretation of the proposed ad hoc The vector of Lagrange multipliers is obtained by
implementation substituting (40) into the set of WC constraints in (37)


www.ietdl.org

which engenders the following set of nonlinear equations 608. In the last two scenarios, a source with multipath
propagation is considered. However, the main multipath
¯
w(k þ 1) þ m(k)Lt (k)H L
~ component with dominant power is considered at 08. The
(41) desired source is 5 dB power and each interference-to-noise
~ ¯
À yw(k þ 1) þ m(k)Lt (k) À i ¼ 0 ratio is equal to 10 dB. The noise power at each antenna
element is equal to 0 dB to model a low SNR environment.
where 0 is the all-zero vector. Unfortunately, a closed-form
solution similar to (21) cannot be obtained because (41) is
a system of nonlinear equations. In this case, a Newton-like 6.1 AOI mismatch scenario
method is obligatory to find the optimum vector of In this scenario, the performance of the standard MVDR
Lagrange multipliers t (k) that satisfies the set of WC beamformer in (6) (referred to as standard MVDR), the
constraints in (37). The trust region method [24, 25] is robust MVDR beamformers with WC constraint which
adopted to solve the system of nonlinear equations in (41). implemented using SOCP [5, 6] and eigendecomposition
A minor drawback of this technique is that all WC [4, 8] approaches (referred to, respectively, as robust
constraints have to be solved via reducing them to equality MVDR-WC/SOCP and robust MVDR-WC/EigDec),
if any of the WC constraints is not achieved. and the proposed robust adaptive beamformer outlined in
Section 4.2 (referred to as robust MVDR-WC/proposed)
The algorithm of the robust LCMV beamforming with are compared. The aforementioned beamformers are
MBWC consists of the following steps. simulated using a mismatched steering vector of the desired
_ source where the presumed AOI equals 58. The robust
Step 0. Initialise R(0) ¼ I , w(0) ¼ ad , a ¼ 0:5, h ¼ 0:97. MVDR-WC/EigDec beamformer is computed using (22)
and its norm is obtained using a Newton-like algorithm
Step 1. Pick a new sample from array observations and [4, 8]. In addition, the benchmark MVDR beamformer at
compute _ the
_ sample covariance matrix as (7) is simulated with the actual steering vector of the
R(k) ¼ hR(k À 1) þ x(k)xH (k); M 2 . desired source. The benchmark MVDR beamformer is
implemented using the well-known RLS algorithm. The
Step 2. Compute the optimum step size using (36) where
_ _ update of the sample covariance matrix in (5) is used with
_
r ( R(k)w(k); M 2 þ 2M. all beamformers with h ¼ 0:97. The WC constrained
parameter z ¼ 1:8 is chosen for both robust MVDR-WC/
Step 3. Update the unconstrained MV weight vector as
_ EigDec and robust MVDR-WC/proposed beamformers,
_ ^
w(k þ 1) ¼ w(k) À m(k)R(k)w(k); the matrix vector
~ _ whereas z ¼ 3 is chosen for robust MVDR-WC/SOCP
multiplication R(k)w(k) is computed in step 2. beamformer. This is because the SOCP method is
initialised with normalised weight vector [5, 6]. The WC
Step 4. If any w(k þ 1)H L ! ykw(k þ 1)k þ iÞ; ðMN þ M Þ
~ ~ constrained parameter is selected based on the best
Compute t(k) by solving (41); ð2MN þ M ÞðM þ 1ÞR, where performance achieved from several simulation runs. In
R is the required number of iterations for the trust region practical, it is selected based on some preliminary (coarse)
method convergence. knowledge about wireless channels or using Mont Carlo
^
Else t (k) ¼ 0 and w(k þ 1) ¼ w(k þ 1)! go to step 1.
~ simulation. Fig. 2 shows the output SINR of the

Step 5. Update the weight vector of robust LCMV
beamformer using (40) ! Go to step 1.

As demonstrated in the above implementation, the
robust LCMV beamformer with MBWC constraints
À Á
requires O M 2 ð2NR þ R þ 2Þ þ M ð2NR þ N þ R þ 3Þ
complexity. Indeed, it requires higher computational load;
however, it cannot be compared with the single WC
beamformers. Several simulation scenarios demonstrated
that the trust region algorithm requires 4 to 12 iterations
for convergence.

6 Numerical examples
A uniform linear array of M ¼ 5 omnidirectional sensors
spaced half-wavelength apart is considered. All results are
obtained by averaging 100 independent simulation runs.
Through all examples, it is assumed that there is one Figure 2 Output SINR against snapshot index for the first
desired source at 08 and two interfering sources at 458 and scenario


www.ietdl.org

abovementioned beamformers against snapshot and beam
patterns against AOI is illustrated in Fig. 3. The proposed
algorithm offers the best SINR compared with other robust
approaches and even faster convergence speed over the
benchmark MVDR beamforming with the RLS algorithm.
The eigendecomposition and SOCP methods are
considered as batch algorithms where the weight vector of
the robust MVDR-WC/EigDec beamformer is computed
using the closed-form in (22) and the weight vector of the
robust MVDR-WC/SOCP beamformer is recomputed
completely with each snapshot [5]. Finally, the proposed
algorithm is the best at eliminating sidlobes and
interference compared with other robust approaches as
evident from Fig. 3 where it is ranked after the Benchmark
MVDR beamformer.

Fig. 4 shows the output SINR against noise power using Figure 4 Output SINR against noise power for the first
50 fixed training sample size (i.e. low sample support). The scenario with training data size N ¼ 50
figure conspicuously demonstrates the superiority of the
proposed beamformer especially at low noise power (i.e.
high SNR) thanks to its optimality at low snapshot index
as observed from Fig. 2.

In order to analyse the Lagrange multiplier in (34), the
parameters of (33) which are given in (35) are investigated.
These parameters and the Lagrange multiplier l (referred
to as WC parameters) are plotted against snapshot index in
Fig. 5 at z ¼ 1:8. The figure illustrates that A . 0, B . 0
and C , 0. Therefore (33) has one real positive root as
explained in Section 4.1. It has to be noted that the
algorithm commenced into the WC optimisation from the
first snapshoot as shown in Fig. 5.

It is worthwhile to investigate the sensitivity of the
proposed algorithm against the WC constrained value
(i.e. z). Fig. 6 demonstrates the performance of the
proposed algorithm at several z values. It reveals that the

Figure 5 WC parameters of the robust MVDR-WC/
proposed beamformer at z ¼ 1.8

algorithm performs well at a reasonable window of
z ¼ [1.4:2.2] with optimality at z ¼ 1:8 in terms of start-
up performance. Indeed, z is a crucial factor for any WC
performance optimisation algorithm and it should be
properly selected. As shown in Fig. 6, the algorithm starts
to degrade when z is decreased because the algorithm is no
longer capable of handling the mismatch degree. For
clarity, the WC parameters at z ¼ 1:2 are illustrated in
Fig. 7, which are almost analogous to Fig. 5 where
A . 0, B . 0 and C , 0. However, the algorithm delays
executing the WC optimisation because z is very low and
subsequently the algorithm performance is degraded. More
Figure 3 Steady-state array beam patterns against AOI (in preciously, the algorithm executes the unconstrained MV
radian) for the first scenario minimisation without WC optimisation more than


www.ietdl.org

Figure 6 Effect of WC parameter z on the output SINR of
the first scenario
Figure 8 WC parameters of the robust MVDR-WC/
proposed beamformer at z ¼ 2.4

implementation to guarantee that A . 0 and therefore the
complex solution of (33) could be prevented. However,
during initial iterations of some runs, one may have A , 0
while B2 . AC and therefore the algorithm can continue
without adjusting z. Consequently, the best practice is to
verify if A , 0 and B2 , AC are met, and if so, z is
decreased. The WC algorithm in Section 4.2 is revised by
amending step 4 as follows.

_
Step 4 If x , zw(k þ 1), compute l using (34); 5M.
À À ÁÁ
If A , 0 B2 , AC , z ¼ z À 0:1, end Else l ¼ 0 and
_
w(k þ 1) ¼ w(k þ 1) go to step 1.

Another simulation is conducted to evaluate the
performance of the above modified robust MVDR-WC/
proposed beamformer. The modified algorithm is initialised
Figure 7 WC parameters of the robust MVDR-WC/ with the same parameters of the first scenario except that
proposed beamformer at z ¼ 1.2 z ¼ 3. The WC parameters and the output SINR for the
modified robust MVDR-WC/proposed beamformer are
demonstrated in Figs. 9 and 10 respectively. Fig. 9
necessary and hence a part of the interested signal is indicates that the algorithm starts with
suppressed and it could not be recovered again with A , 0, B . 0, C , 0 and B2 , AC and then z starts to
adaptive implementation. decrease until an acceptable value which prevent complex
solution of (33).
Finally, the algorithm performance is seriously degraded
when z is increased to 2.4 as shown in Fig. 8. The WC
parameters at z ¼ 2:4 are shown in Fig. 8 which
6.2 Small angular spread scenario
demonstrate that A , 0, B . 0, C , 0 and, in turn, (33) In this scenario, a desired source with small angular spread
has one real positive root if B2 . AC. Regrettably, the emerging from multipath propagation as in rural cells is
preceding condition could not be achieved where jC j . 0
. simulated. The same parameters of first scenario are used
and hence the solution of (33) has two complex roots and except that the SOI is impinging on the array from three
therefore the algorithm performance is seriously degraded. directions associated with three multipath rays. There is a 58
The plot of parameter l in Fig. 8 is only for real part. In mismatch with the dominant multipath ray. The other two
order to avoid complex solution to (33), the WC rays amplitudes are 40% of the main component and they are
constrained value z can be adjusted during adaptive impinging on the array from the directions 48 and 238. The


www.ietdl.org

Figure 11 Output SINR against snapshot index for the
second scenario

Figure 9 WC parameters of the modified robust MVDR- beamformers by picking up some signal components from the
WC/proposed beamformer at z ¼ 3 multipath signals scattered inside the mismatch region
bounded by z.
maximum angular spread (i.e. 48) associated with the multipath
components in this scenario is less than the AOI mismatch of 6.3 Large angular spread scenario
the dominant multipath ray (i.e. 58). In addition to the five
beamformers simulated in the first scenario, the benchmark In this scenario, a large angular spread as in the cellular indoor
LCMV beamformer in (9) with MBC is simulated, which is environment is simulated. The simulation system is similar to
imposed towards the three actual AOIs (i.e. 08, 48, 238) of the previous scenario except that the three multipath
the multipath rays (referred to as benchmark LCMV). The components are impinging on the array from directions 08,
multipath components in the benchmark LCMV 2308 and 2808. The dominant ray impinges on the array
beamformer are combined using MRC. The benchmark from 08 direction and there is a 58 look direction mismatch. It
MVDR beamformer in (7) is simulated using only the actual is assumed that the phases of the multipath rays are
steering vector of the dominant multipath ray. The WC independently and uniformly drawn from the interval
parameter z of the robust beamformers is selected as in the [À p, p] in each run. The phases associated with multipaths
first scenario. The performance of the aforesaid beamformers vary from run to run and stay constant during adaptive
in terms of SINR is illustrated in Fig. 11. The proposed implementation of each run. In this scenario, the
algorithm offers about 2 dB improvement over other robust beamformers in the previous experiments are simulated in
approaches as evident from Fig. 11. The performance of the addition to the proposed robust LCMV beamformer with
benchmark MVDR beamformer is degraded below the robust MBWC constraints (referred to as robust LCMV-MBWC).
approaches, whereas the WC constraint bears the robust

Figure 10 Output SINR against snapshot index for the
modified robust MVDR-WC/proposed beamformer with Figure 12 Output SINR against snapshot index for the third
the parameters of the first scenario scenario


www.ietdl.org

Moreover, the standard LCMV beamformer with multiple [2] GERSHMAN A.B.: ‘Robust adaptive beamforming in sensor
constraints in (9) (referred to as standard LCMV) is simulated arrays’, AEU, Int. J. Electron. Commun., 1999, 53, (6),
using a mismatched steering vector of the dominant multipath pp. 305– 314
ray while there are no mismatches with the other two rays (i.e.
58, 2308, 2808). The multipath rays of the benchmark [3] LI J., STOICA P.: ‘Robust adaptive beamforming’ (John
LCMV and the standard LCMV beamformers are combined Wiley Sons, Inc, Hoboken, NJ, 2006)
using MRC. The phases of multipath rays are unknown to all
beamformers except the benchmark beamformers. The [4] GERSHMAN A.B., LUO Z.-Q., SHAHBAZPANAHI S.: ‘Robust adaptive
beamformers of the first scenario are simulated using the same beamforming based on worst-case performance
parameters, whereas theÂrobust LCMV-MBWC beamformer
Ã optimization’, in LI J. , STOICA P. (EDS): ‘Robust adaptive
is simulated using y ¼ 1:6 0:2 0:2 . The selection of y beamforming’ (John Wiley Sons, Inc, Hoboken, NJ,
is obtained practically using several simulation runs. It is 2006), pp. 49 – 89
somehow embodying the amplitude distribution of multipath
rays. However, in-depth analysis for tuning this vector and [5] VOROBYOV S.A., GERSHMAN A.B., LUO Z.-Q.: ‘Robust adaptive
even optimal estimation is a good candidate for future beamforming using worst-case performance optimization:
research. The SINR performance of the seven beamformers is a solution to the signal mismatch problem’, IEEE Trans.
demonstrated in Fig. 12. First of all, the benchmark LCMV Signal Process., 2003, 51, (2), pp. 313– 324
is considerably degraded, despite tracing the dominant
multipath ray. This is because the large angular spread [6] CUI S., KISIALIOU M. , LUO Z.-Q., DING Z. : ‘Robust blind
deforms the effective steering vector of the SOI. The multiuser detection against signature waveform mismatch
performances of the robust beamformers with the single WC based on second order cone programming’, IEEE Trans.
constraint resemble their performances in the first scenario. Wirel. Commun., 2005, 4, (4), pp. 1285 – 1291
The robust LCMV-MBWC beamformer offers about 1 dB
improvement over the single WC constraint beamformers due [7] SHAHBAZPANAHI S., GERSHMAN A.B.: ‘Robust blind multiuser
to efficient multipath handling using multiple WC constraints. detection for synchronous CDMA system using worst-case
performance’, IEEE Trans. Wirel. Commun, 2004, 3, (6),
pp. 2232 – 2245

7 Conclusions [8] ZARIFI K., SHAHBAZPANAHI S., GERSHMAN A.B., LUO Z.-Q.: ‘Robust
blind multiuser detection based on the worst-case
In this paper, the robust adaptive beamforming using WC performance optimization of the MMSE receiver’, IEEE
performance optimisation is implemented using novel ad hoc Trans. Signal Process., 2005, 53, (1), pp. 295 – 205
approaches. Two efficient implementations are developed
using single and multiple WC constraints. The proposed [9] ELNASHARA.: ‘Robust adaptive beamforming’. ACE2 Network
implementations are based on iterative gradient minimisation. of Excellence Workshop on Smart Antennas, MIMO Systems
In contrast to the existing single WC robust approaches, the and Related Technologies, Mykonos, Greece, 8 June 2006
proposed single WC implementation requires very low
computational load and it engenders the best performance, [10] LORENZ R.G. , BOYD S.P.: ‘Robust minimum variance
especially at low sample support. In addition, the proposed beamforming’, IEEE Trans. Signal Process., 2005, 53, (5),
algorithm eliminates the covariance matrix inversion pp. 1684 – 1696
estimation. The WC performance optimisation is generalised
to include multiple WC constraints which produce a robust [11] BESSON O. , VINCENT F.: ‘Performance analysis of
LCMV beamformer with MBWC constraints. An efficient beamformers using generalized loading of the
solution for the LCMV-MBWC beamformer is introduced covariance matrix in the presence of random steering
by solving a system of nonlinear equations. Simulation results vector errors’, IEEE Trans. Signal Process., 2005, 53, (2),
demonstrated the superiority of the proposed beamformers pp. 452– 459
over the existing robust approaches. Future research may
include fine-tuning of the constrained vector of the LCMV- [12] VINCENT F., BESSON O.: ‘Steering vector errors and diagonal
MBWC beamformer and developing low complexity adaptive loading’, IEE Proc., Radar Sonar Navig., 2004, 151, (6),
implementations. pp. 337– 343

[13] STURM J.F.: ‘Using SeDuMi 1.02, a MATLAB toolbox for
optimization over symmetric cones’, Optim. Methods
8 References Softw., 1999, 11– 12, pp. 625 – 653

[1] FELDMAN D.D., GRIFFITHS L.J.: ‘A projection approach for [14] SHAHBAZPANAHI S., GERSHMAN A.B., LUO Z.Q., WONG K.M.: ‘Robust
robust adaptive beamforming’, IEEE Trans. Signal Process., adaptive beamforming for general-rank signal model’, IEEE
1994, 42, (4), pp. 867– 876 Trans. Signal Process., 2003, 51, (9), pp. 2257–2269


www.ietdl.org

[15] ELNASHAR A., ELNOUBI S., ELMIKATI H.: ‘Further study on robust constraint’, IEEE Trans. Signal Process., 2001, 49, (6),
adaptive beamforming with optimum diagonal loading’, IEEE pp. 1138 – 1145
Trans. Antennas Propag., 2006, 54, (12), pp. 3647– 3658
[21] ATTALLAH S., ABED-MERAIM K.: ‘Fast algorithms for subspace
[16] LI J., STOICA P., WANG Z.: ‘On robust Capon beamforming tracking’, IEEE Trans. Signal Process. Lett., 2001, 8, (7),
and diagonal loading’, IEEE Trans. Signal Proces., 2003, pp. 203– 206
51, (7), pp. 1702– 1715
[22] ELNASHAR A., ELNOUBI S., EL-MAKATI H.: ‘Performance analysis
[17] ELNASHAR A., ELNOUBI S., ELMIKATI H. : ‘Robust adaptive of blind adaptive MOE multiuser receivers using inverse
beamforming with variable diagonal loading’. Proc. 6th QRD-RLS algorithm’, IEEE Trans. Circuits Syst. I, 2008, 55,
IEE Int. Conf. 3G Beyond (3G 2005), London, UK, (1), pp. 398– 411
November 2005, pp. 489 – 493
[23] CHOI S., SHIM D.: ‘A novel adaptive beamforming
[18] ERTEL R.B., CARDIERI P., SOWERBY K.W., RAPPAPORT T.S., REED J.H.: algorithm for smart antenna system in a CDMA mobile
‘Overview of spatial channel models for antenna communication environment’, IEEE Trans. Veh. Technol.,
communication systems’, IEEE Pers. Commun., 1998, 5, 2000, 49, (5), pp. 1793– 1806
(1), pp. 10– 22
[24] POWELL M.J.D.: ‘A Fortran subroutine for solving systems
[19] VAN DER VEEN A.-J.: ‘Algebraic methods for deterministic of nonlinear algebraic equations’, in RABINOWITZ P. (ED.):
blind beamforming’, Proc. IEEE, 1998, 86, (10), ‘Numerical methods for nonlinear algebraic equations’
pp. 1987 – 2008 (Gordon and Breach Science Publishers, New York, 1988)

[20] TIAN Z., BELL K.L., VAN TREES H.L.A.: ‘Recursive least squares [25] CONN N.R., GOULD N.I.M., TOINT PH.L.: ‘Trust-region methods’
implementation for LCMP beamforming under quadratic MPS/SIAM series on optimization’ (SIAM and MPS, 2000)


Beamforming for Antenna Array

More Related Content

Beamforming for Antenna Array