You are on page 1of 13

1

Sparse Array Design for Azimuthal


Direction-of-Arrival Estimation
Michael R ubsamen and Alex B. Gershman
AbstractWe propose a novel approach to the design of array
geometries for azimuthal direction-of-arrival (DOA) estimation.
The proposed array design is related to that of minimum
redundancy arrays, but the array sensors are not required to
lie on a uniform grid. Based on the proposed array geometry
design, we develop a subspace-based DOA estimation technique,
which allows to estimate the DOAs of more sources than sensors,
using only second-order statistics of the received data. This DOA
estimation technique is related to the covariance augmentation
technique, but in contrast to the latter technique, it provides non-
ambiguous DOA estimates for the full 360

azimuth eld-of-view.
Index TermsSparse planar arrays, direction-of-arrival esti-
mation
I. INTRODUCTION
Array signal processing is used in several application areas
such as radar, sonar, wireless communications, radio astron-
omy, seismology, acoustics, and medical imaging [1], [2].
Coherent receive arrays simultaneously measure a spatial eld
at different locations. The received data is processed, for ex-
ample, to estimate the source directions-of-arrival (DOAs). A
fundamental aspect of any array processing system is the array
geometry as it strongly affects the system performance [3].
A number of array geometry designs have been investigated
in the literature. One popular approach is to optimize the
beampattern, see [4], [5] and the references therein. For
instance, the width of the mainlobe may be minimized while
enforcing an upper bound on the sidelobe height. Then, it is
manifest to estimate the DOAs using beamformers. However, a
number of estimators often outperform the beamforming based
DOA estimators [2]. Alternatively, the sensor locations may
be determined by optimizing a measure of the performance
of the DOA estimator or the Cram er-Rao bound (CRB) [6].
This method requires that the source parameters (their DOAs,
powers, etc.) are approximately known a-priori. Since this
information is often not available, this method is of limited
Manuscript received May 7, 2011; revised August 25, 2011.
(c) 2011 IEEE. Personal use of this material is permitted. Permission from
IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional
purposes, creating new collective works, for resale or redistribution to servers
or lists, or reuse of any copyrighted component of this work in other works.
M. R ubsamen is with the Communication Systems Group, Technische
Universit at Darmstadt, Germany, e-mail: michael@rubsamen.net.
A. B. Gershman, deceased, was with the Communication Systems Group,
Technische Universit at Darmstadt, Germany.
This work was supported in part by the European Research Council
(ERC) Advanced Investigator Grants Program under Grant 227477-ROSE and
German Research Foundation (DFG) under Grant GE 1881/4-1.
Preliminary results of this work were presented at the IEEE Sensor
Array and Multichannel Signal Processing Workshop (SAM) 2010, Maale
Hahamisha, Israel.
practical value. In the sequel, we focus on array geometry
designs, which are not tailored to a specic DOA estimator,
and which do not require accurate knowledge of the source
parameters.
Suppose that only second-order statistics of the received
data are used. In general, the spatial covariance function
depends on the two locations at which the eld is measured.
However, in many applications the spatial eld is wide-sense
stationary. Then, the covariance function depends only on
the displacement between the points at which the eld is
measured, but not on their absolute locations. Such wide-
sense stationary spatial elds occur when the sources are
uncorrelated and located in the far-eld.
Uniform linear arrays of N sensors contain N m realiza-
tions of the displacement (lag) md
z
, where d
z
is the baseline.
Hence, these arrays contain N m1 redundancies for the
lag md
z
. With linear minimum redundancy arrays (MRAs), the
number of redundant lags is minimized [7]. This leads to larger
apertures as compared to uniform linear arrays (ULAs) with
the same number of sensors and the same baseline d
z
. Such
increased apertures result in an improved spatial resolution
capability at nearly the same hardware costs.
In the literature, the so-called general and restricted linear
MRAs have been studied [7]. In the case of general linear
MRAs, the geometry is optimized so that it contains the lags
d
z
, 2d
z
, 3d
z
and so forth for as many contiguous integer
multiples of the baseline d
z
as possible. The aperture of
general linear MRAs may be larger than the extent of the
contiguous sequence of lags. In the case of restricted linear
MRAs, the additional constraint that the set of lags does not
contain any holes (missing integer multiples of d
z
) is imposed.
Hence, the aperture of restricted linear MRAs is equal to the
extent of the contiguous sequence of lags.
Linear arrays of omnidirectional sensors provide only 180

angular coverage, and the spatial resolution capability is poor


for directions close to the endre direction. To achieve 360

angular coverage using omnidirectional sensors, planar array


geometries are required. The MRA principle can be extended
straightforwardly to planar geometries. However, linear and
planar MRAs are particularly suitable to estimate the one-
dimensional (1D) and two-dimensional (2D) source electric
angles [8]. Neither linear nor planar MRAs are optimized for
the estimation of azimuth angles.
Since linear arrays provide good spatial resolution capability
for limited angular sectors close to the broadside direction,
arrays for azimuth estimation often consist of several linear ar-
rays, which cover different angular sectors. Popular geometries
are T-shaped, L-shaped, +-shaped, Y-shaped, and -shaped
arrays [2], [9]. Clearly, one drawback of this approach is that
2
the sensor locations are restricted to a shape, which is chosen
in an ad-hoc way.
In this paper, we propose a novel approach to the design
of array geometries for azimuthal DOA estimation. Our array
geometries are appropriate for approximately planar systems
such as terrestrial wireless communication systems or air
trafc control systems. The proposed array design can be
interpreted as a modication of the MRA concept for azimuth
estimation.
Based on the proposed array geometry design, we develop
an associated subspace-based method to estimate the DOAs
of more sources than there are sensors (the case referred to
in [10], [11] as superior case DOA estimation). It is similar
to the covariance augmentation (CA) technique [10][15], but
in contrast to the latter technique, it provides non-ambiguous
DOA estimates for the whole 360

azimuth eld-of-view.
The proposed DOA estimation method rst computes an
augmented Toeplitz covariance matrix with corresponding
Vandermonde steering vectors that contain in their elements
the conventional Fourier basis functions. Subsequently, the
source DOAs are estimated by applying a root-MUSIC-type
algorithm to the Fourier-domain augmented covariance ma-
trix. Hereafter, we refer to this technique as Fourier-domain
covariance augmentation (FDCA) technique.
The proposed array geometry design leads to sparse pla-
nar arrays, for which manifold ambiguities can occur [16]
[20]. Such manifold ambiguities can lead to a breakdown of
subspace-based DOA estimation methods such as the MUSIC
algorithm. We also demonstrate that the proposed FDCA
technique enables to resolve manifold ambiguities.
II. BACKGROUND
Let us assume L narrowband far-eld point sources that
impinge on an array of N omnidirectional sensors. The
baseband snapshot vector at the kth time instant is given by
x(k) = A(, )s(k) +n(k) C
N1
,
where
A(, ) = [a(
1
,
1
), . . . , a(
L
,
L
)] C
NL
is the steering matrix, a(
l
,
l
) is the steering vector of the lth
source,
l
[, ) and
l
[/2, /2] are the azimuth
and elevation angles of the lth source, respectively,
= [
1
, . . . ,
L
]
T
R
L1
= [
1
, . . . ,
L
]
T
R
L1
are the vectors of source azimuth and elevation angles, respec-
tively,
s(k) = [s
1
(k), . . . , s
L
(k)]
T
C
L1
is the source waveform vector, s
l
(k) is the complex envelope
of the lth source, n(k) C
N1
is the noise vector, ()
T
denotes the transpose, and R
mn
and C
mn
are the sets of
real and complex m n matrices, respectively. The steering
vector of the lth source can be expressed as
a(
l
,
l
) =
_

_
e
j
2

u
T
(
l
,
l
)p
1
.
.
.
e
j
2

u
T
(
l
,
l
)p
N
_

_
C
N1
,
where
u(
l
,
l
) =
_

_
u
x,l
u
y,l
u
z,l
_

_
=
_

_
cos
l
cos
l
cos
l
sin
l
sin
l
_

_
is a vector of length , which points in the direction of the
lth source,
p
n
= [p
xn
, p
yn
, p
zn
]
T
is the location of the nth sensor, and is the wavelength. The
components of u(
l
,
l
) are referred to as the electric angles
of the lth source. In the sequel, the array manifold and the
number of sources are assumed to be known.
The baseband signal and noise waveforms are modeled as
unknown zero-mean stationary uncorrelated random processes.
Then, the N N array covariance matrix can be expressed as
R
x
= Ex(k)x
H
(k) =
L

l=1
P
l
a(
l
,
l
)a
H
(
l
,
l
) + I
N
,
(1)
where P
l
= E[s
l
(k)[
2
is the power of the lth signal, E
denotes the statistical expectation, is the sensor noise power,
I
N
is the N N identity matrix, and ()
H
denotes the
Hermitian transpose. The (n
1
, n
2
)th component of R
x
can
be expressed as
[R
x
]
n1,n2
=
L

l=1
P
l
e
j
2

u
T
(
l
,
l
)(p
n
1
p
n
2
)
+
n1,n2
,
where
n1,n2
is the Kronecker delta. Hence, the components
of the array covariance matrix depend on the lags of the array.
The set of all lags is denoted by
o(p) =
_
p
n1
p
n2

n
1
, n
2
N
N
1
_
,
where
p =
_
p
T
1
, . . . , p
T
N

T
stacks the sensor locations, and N
N
1
is the set of natural
numbers from 1 to N. Since there are N 1 redundancies for
the zero-lag, the cardinality of o(p) is at most N
2
N + 1.
The geometry corresponding to o(p) is commonly referred to
as the co-array.
In practice, the array covariance matrix can be estimated as

R
x
=
1
K
K

k=1
x(k)x
H
(k),
where K is the number of snapshots.
3
A. Linear minimum redundancy arrays
Unless specied otherwise, we assume for linear arrays that
the sensors lie on the z-axis of the coordinate system. In this
case, we drop any dependencies on the azimuth angle for
notational brevity.
The optimization problem for restricted linear MRAs can
be formulated as
max
M,p
z
M s.t. o
z
(p
z
) = d
z
Z
M
M
, (2)
where
o
z
(p
z
) =
_
p
zn
1
p
zn
2

n
1
, n
2
N
N
1
_
describes the set of lags,
p
z
= [p
z1
, . . . , p
zN
]
T
stacks the z-coordinates of the sensor locations, Z
n
m
is the
set of integer numbers from m to n, and the product of a
set with a scalar is dened in the element-wise sense. Hence,
M is maximized subject to (s.t.) the constraint that the array
geometry contains the lags Md
z
, (M + 1)d
z
, . . . , Md
z
,
and no other lags. The optimization variables are M and p
z
.
Let M
RL
denote the solution of (2), where the subscript refers
to restricted linear. In [21] and [22], it has been shown that
2.434 lim
N
N
2
M
RL
3. (3)
Thus, M
RL
increases quadratically with the number of sensors.
The optimum array geometry of (2) is not unique. It can
be easily seen that the solution of (2) does not change if we
introduce the constraints
p
z1
< p
z2
< . . . < p
zN
,
p
z1
= 0, p
z2
= d
z
, p
zN
= Md
z
,
(4)
where the constraints in the second line of (4) imply that the
lag (M 1)d
z
is between the second and the Nth array
sensors. The solution of (2) may be attained for several
different geometries, even if the constraints in (4) are taken
into account. Ambiguity-reducing constraints such as the ones
in (4) are not essential for our approach, and therefore we do
not specify them in the following.
The optimization problem for general linear MRAs can be
formulated as
max
M,p
z
M s.t. o
z
(p
z
) d
z
Z
M
M
, (5)
where c
1
c
2
means that c
1
is a superset of c
2
. The
constraint in (5) implies that the array geometry contains
the lags Md
z
, (M + 1)d
z
, . . . , Md
z
, but it may contain
also other lags. Let M
GL
denote the solution of (5), where
the subscript refers to general linear. Since the constraint
in (5) is less restrictive than the constraint in (2), we have
that M
GL
M
RL
for the same number of sensors. It has been
shown that [21], [23]
2.434 lim
N
N
2
M
GL
2.6646.
The complexity to compute linear MRA geometries in-
creases exponentially with the number of sensors. Therefore,
these geometries are known only for small numbers of sensors.
In [24], restricted linear MRAs with N 17 are listed. A list
of general linear MRAs with N 11 has been published
in [7].
The array covariance matrices of linear MRAs contain the
entries
R
z
(m) =
L

l=1
P
l
e
j
2

mdzu
z,l
+
m,0
, m Z
M
M
, (6)
where M = M
RL
for restricted linear MRAs, and M = M
GL
for general linear MRAs. In the following, we refer to linear
MRAs with the baseline d
z
= /2 as standard linear MRAs.
For such arrays, (6) can be reformulated as
R
z
(m) =
_

uz
(u
z
)e
jmuz
du
z
+
m,0
, m Z
M
M
,
where

uz
(u
z
) =
L

l=1
P
l
(u
z
u
z,l
)
is the angular power density expressed as a function of u
z
, and
() is the Dirac delta function. Hence, the off-diagonal entries
of the array covariance matrices of standard linear MRAs are
Fourier series coefcients of
uz
(u
z
) [25]. This means that
for standard linear MRAs, the number of available contiguous
Fourier series coefcients of
uz
(u
z
) is maximized. Due to
the orthogonality of the Fourier basis functions [25], standard
linear MRAs are particularly suitable to estimate the 1D source
electric angles u
z,l
.
B. Covariance augmentation technique
The array covariance matrices of linear MRAs contain the
entries R
z
(m), m Z
M
M
. Therefore, these arrays enable
to estimate the (M + 1) (M + 1) augmented covariance
matrix [12]
C
CA

L

l=1
P
l
a
CA
(
l
)a
H
CA
(
l
) + I
M+1
=
_

_
R
z
(0) R
z
(1) R
z
(M)
R
z
(1) R
z
(0)
.
.
.
.
.
.
.
.
. R
z
(1)
R
z
(M) R
z
(1) R
z
(0)
_

_
,
where
a
CA
(
l
)
_
1, e
j
2

dz sin
l
, . . . , e
j
2

Mdz sin
l
_
T
is the (M+1)1 augmented steering vector of the lth source.
Applying the MUSIC algorithm to the nite-snapshot estimate
of the augmented covariance matrix enables to estimate the
DOAs of up to M sources. Since M > N if N 4, the CA
technique allows estimating the DOAs of more sources than
there are sensors. For comparison, the conventional MUSIC
algorithm allows estimating the DOAs of at most N 1
sources [26], [27].
4
Generally, if the aperture of a linear array is larger than
(N1)/2, then manifold ambiguities exist [17]. This means
that there exist DOAs
1
, . . . ,
L
such that
!(A()) = !(A(

)),
where

= [

1
, . . . ,

L
]
T
is another vector of L < N DOAs with at least one DOA that
is not contained in , and !() is the column-space operator.
The ambiguity of the column-space of the steering matrix
is critical for subspace-based DOA estimation methods. For
example, the noise-free MUSIC null-spectrum function will
have nulls at all angles
l
and

l
, l N
L
1
. The spurious
nulls lead to outliers among the DOA estimates of the MUSIC
algorithm applied to

R
x
.
Since the apertures of standard linear MRAs are larger
than (N 1)/2, manifold ambiguities exist for such arrays.
However, M+1 pairwise different augmented steering vectors
are always linearly independent due to the Vandermonde
structure of these vectors [28]. For standard linear MRAs, the
vector function a
CA
() is one-to-one for [/2, /2).
Assuming that all elevation angles are in [/2, /2) and
L < M + 1, we have
!(A
CA
()) ,= !(A
CA
(

)),
where
A
CA
() = [a
CA
(
1
), . . . , a
CA
(
L
)]
is the (M +1) L augmented steering matrix. Thus, the CA
technique allows to resolve manifold ambiguities [19].
C. Sparse planar arrays
In the case of planar arrays, we assume without loss of
generality that the sensors are located in the xy-plane of the
coordinate system. The baselines in x- and y-directions are
denoted by d
x
and d
y
, respectively.
The MRA concept can be generalized straightforwardly to
planar geometries. Similar to (2), the optimization problem for
restricted planar MRAs can be formulated as
max
M,p
xy
M s.t. o
xy
(p
xy
) =
_
d
x
Z
M
M
_

_
d
y
Z
M
M
_
, (7)
where
o
xy
(p
xy
) =
_
_
p
xn
1
p
xn
2
, p
yn
1
p
yn
2

n
1
, n
2
N
N
1
_
describes the set of lags,
p
xy
= [p
x1
, p
y1
, . . . , p
xN
, p
yN
]
T
contains the x- and y-coordinates of the sensor locations,
and denotes the Cartesian product. Hence, the constraint
in (7) enforces that the co-array has a uniform rectangular
geometry with the baselines d
x
and d
y
in x- and y-directions,
respectively. Let M
RP
denote the solution of (7), where the
subscript refers to restricted planar. It has been shown
that [29]
2.207 lim
N
N
M
RP
3. (8)
Hence, M
RP
increases only linearly with the number of sen-
sors.
The optimization problem for general planar MRAs is given
by
max
M,p
xy
M s.t. o
xy
(p
xy
)
_
d
x
Z
M
M
_

_
d
y
Z
M
M
_
. (9)
The solution M
GP
(the subscript refers to general planar) of
this problem is bounded asymptotically by [29]
2 lim
N
N
M
GP
2.6646,
where the lower bound follows from the fact that the cardi-
nality of o
xy
(p
xy
) is at most N
2
N + 1.
The array covariance matrices of planar MRAs contain the
spatial covariances
R
xy
(m
1
, m
2
) =
L

l=1
P
l
e
j
2

(m1dxu
x,l
+m2dyu
y,l
)
+
m1,0

m2,0
, m
1
, m
2
Z
M
M
,
(10)
where M = M
RP
for restricted planar MRAs, and M = M
GP
for general planar MRAs. For standard (d
x
= d
y
= /2)
planar MRAs, the spatial covariances (10) can be written as
R
xy
(m
1
, m
2
) =
__
I

ux,uy
(u
x
, u
y
)e
j(m1ux+m2uy)
du
x
du
y
+
m1,0

m2,0
, m
1
, m
2
Z
M
M
,
where

ux,uy
(u
x
, u
y
) =
L

l=1
P
l
(u
x
u
x,l
) (u
y
u
y,l
)
is the 2D angular power density expressed as a function of u
x
and u
y
, and the interval of integration is J = [, ]
2
. Hence,
the off-diagonal entries of the array covariance matrices of
standard planar MRAs are 2D Fourier series coefcients of

ux,uy
(u
x
, u
y
). This means that for standard planar MRAs,
the number of available contiguous 2D Fourier series coef-
cients of
ux,uy
(u
x
, u
y
) is maximized. Standard planar MRAs
are therefore particularly suitable to estimate the 2D source
electric angles.
The complexity to compute planar MRA geometries is even
higher than for linear MRAs with the same number of sen-
sors [30]. For this reason, sparse easy-to-compute geometries
without holes among the set of lags have been proposed.
Examples of such arrays are the Greene-Wood arrays of [31]
and the cross-product arrays discussed in [32]. The sparsity of
such arrays can be assessed using (8).
Restricted planar MRAs have uniform rectangular co-arrays.
Intuitively, arrays for azimuth estimation should have an
approximately circularly symmetric co-array. Uniform circular
arrays (UCAs) are a popular class of planar arrays, which have
this property [33]. However, UCAs sample the spatial eld
uniformly, and therefore they can be considered as the planar
counterparts of ULAs.
Another array geometry design, which leads to approxi-
mately circularly symmetric co-arrays, has been proposed by
Cornwell in [34]. The sensors are distributed inside a circle
5
such that the product of the distances between the points
of the co-array is maximized. This prevents small distances
between the points of the co-array, and leads to a rather
uniform sampling of the spatial covariance function. The array
geometry optimization problem of [34] can be formulated as
max
p
xy
d(p
xy
) s.t. p
2
xn
+ p
2
yn
1, n N
N
1
,
where
d(p
xy
) =

n1,n2,n3,n4
(n1,n2)=(n3,n4)
n1=n2
log |(p
n1
p
n2
) (p
n3
p
n4
)|
2
,
the cases (n
1
, n
2
) = (n
3
, n
4
) and n
1
= n
2
are excluded from
the summation to prevent zero-arguments of the logarithm,
and ||
2
denotes the Euclidean vector norm. For example,
instead of the product of the distances between the points of
the co-array, the minimum nearest neighbor distance between
the points of the co-array can be maximized.
Planar MRAs as well as Cornwell arrays are designed to
achieve a uniform planar sampling of the spatial covariance
function. Since the cardinality of o
xy
(p
xy
) increases at most
quadratically with the number of sensors, the apertures of
planar MRAs and Cornwell arrays increase only linearly with
the number of sensors. The uniform planar sampling of the
spatial covariance function is well-justied for the estimation
of 2D source electric angles, but we will see in the following
that it is non-optimum for the estimation of azimuth angles.
Since linear arrays provide rather good azimuth estimation
performance for limited angular sectors, practically used arrays
for azimuth estimation often consist of several linear arrays.
The linear arrays are assembled to cover different angular sec-
tors. Popular geometries are T-shaped, L-shaped, +-shaped,
Y-shaped, and -shaped arrays [2], [9]. If these arrays are
built using linear MRAs, then the array apertures increase
quadratically with the number of sensors due to (3). However,
this approach is obviously ad-hoc and does not satisfy an
optimality criterion such as MRAs.
III. PROPOSED ARRAY GEOMETRY DESIGN
In Section II-A, we showed that for standard linear MRAs
the number of available contiguous Fourier series coefcients
of
uz
(u
z
) is maximized. Therefore, standard linear MRAs
are particularly suitable to estimate the 1D source electric
angles u
z,l
. For the similar reason, standard planar MRAs are
particularly suitable to estimate the 2D source electric angles
(u
x,l
, u
y,l
). In this section, we propose a novel approach to
the design of planar arrays for 360

azimuth estimation.
Towards this end, we maximize the number of available
contiguous Fourier series coefcients of the azimuthal power
density function

() =
L

l=1
P
l
(
l
).
Assuming that the sources and sensors are located in the
xy-plane of the coordinate system, the (n
1
, n
2
)th component
of R
x
can be expressed as
[R
x
]
n1,n2
=
L

l=1
P
l
a
n1
(
l
)a

n2
(
l
) +
n1,n2
, (11)
where for notational brevity we dropped the dependency on
the elevation angle, and ()

denotes the complex conjugate.


Equation (11) can be reformulated as
[R
x
]
n1,n2
=

a
n1
()a

n2
(),

()
_
+
n1,n2
, (12)
where
f(), g()
_

f()g

()d
is the scalar product for the Hilbert space L
2
([, ]) of
Lebesgue measurable functions with nite norm [25]
|f()|
L
2
__

[f()[
2
d
_
1/2
.
The mth Fourier series coefcient of

() can be expressed
as
F

[m] =
_

()e
jm
d =

e
jm
,

()
_
. (13)
It can be easily seen that it is not possible to choose the sensor
locations such that (12) equals a Fourier series coefcient of

() or a scaled version thereof.


We use linear functions of

R
x
to estimate the Fourier series
coefcients of

(). To eliminate the disturbing effect of the


non-vanishing (for K ) sensor noise components in the
diagonal entries of

R
x
, we dene vec as the vectorization
operator for a square matrix, which leaves out the entries on
the main diagonal. Then, the estimator for the mth Fourier
series coefcient of

() can be expressed as

[m] = w
H
m
vec

R
x
,
where the (N
2
N) 1 weight vector w
m
is yet to be
determined. We can write

[m] = w
H
m
vecR
x
+w
H
m
vec
R
x
,
where

R
x
=

R
x
R
x
denotes the covariance matrix estimation error. Equation (1)
gives
vecR
x
=
L

l=1
P
l
b
p
xy
(
l
) =
_
b
p
xy
(),

()
_
, (14)
where
b
p
xy
() = veca()a
H
()
is an (N
2
N) 1 vector, the subscript p
xy
emphasizes
the dependency on the array geometry, and the scalar product
between b
p
xy
() and

() is dened in the element-wise


sense. From (14), we obtain

[m] =
_
w
H
m
b
p
xy
(),

()
_
+w
H
m
vec
Rx
. (15)
Due to (13), we choose w
m
such that
w
H
m
b
p
xy
() = e
jm
+
m
(), (16)
6
where
m
() is a small approximation error. Substituting (16)
in (15) yields

[m] = F

[m]+
m
(),

()+w
H
m
vec
R
x
. (17)
Note that
m
() does not depend on the received data. Hence,
it leads to biased estimates of the Fourier series coefcients
of

(). The variance of the estimation errors is solely due


to the nite sample effect.
The error terms in (17) depend on

() and
Rx
. Since
these quantities are not known a-priori, we use the constraints
_
_
_w
H
m
b
p
xy
() e
jm
_
_
_
L
2
, m N
M
1
(18)
and
1
M
M

m=1
|w
m
|
2
(19)
to achieve small estimation errors for the Fourier series coef-
cients of

(), where and are user-dened parameters.


For the choice of , an analogy to standard linear MRAs can
be exploited. For such arrays, the Fourier series coefcients
of
uz
(u
z
) can be estimated as

F
uz
[m] = w
H
m
vec

R
x
= F
uz
[m] + w
H
m
vec
Rx
,
(20)
where w
m
is an (N
2
N) 1 selection vector, assuming
that there are no redundancies. Hence, the nite snapshot
estimation errors in (17) can be expected to be comparable
to those in (20) if = 1.
Since

() is real-valued, its (m)th Fourier series co-


efcient is equal to the complex conjugate of the mth Fourier
series coefcient. Therefore, the constraints in (18) and (19)
support estimation of the Fourier series coefcients for all
m Z
M
M
0.
To maximize the number of available contiguous Fourier
series coefcients of

(), we jointly optimize the weight


vectors and sensor locations as
max
M,p
xy
,{wm}
M
s.t.
_
_
_w
H
m
b
p
xy
() e
jm
_
_
_
L
2
, m N
M
1
1
M
M

m=1
|w
m
|
2
.
(21)
In the sequel, we assume that <

2 to avoid the trivial


solution.
The Fourier series coefcients of the entries of b
p
xy
() de-
cay super-exponentially [35]. exponentially, see Gilbert Strang,
Computational Science and Engineering, Chapter 4. Doron and
Doron proof that the Fourier series coefcients decay super-
exponentially. Therefore, these functions can be accurately ap-
proximated by truncated Fourier series expansions. To evaluate
the L
2
-norms in (21), we sample the functions at

k
= 2k/N

, k Z
N1
0
,
where N

has to be sufciently large so that the Nyquist sam-


pling criterion is satised for exp(jM) and all components
of b
p
xy
(). Using the notations
g
m
=
_
e
jm0
. . . e
jmN1
_
T
B =
_
b
p
xy
(
0
) . . . b
p
xy
_

N1
_
_
,
the problem in (21) can be formulated as
max
M,p
xy
,{wm}
M
s.t.
_
_
_B
H
w
m
g
m
_
_
_
2

_
N

2
, m N
M
1
1
M
M

m=1
|w
m
|
2
.
(22)
Remark 1: Necessary conditions for the feasibility of (22)
are given by
_
_
(I
N

B
H)g
m
_
_
2

_
N

2
, m N
M
1
, (23)
where
B
H is the orthogonal projection matrix onto the
column-space of B
H
.
Remark 2: Let M
FD
denote the solution of (22), where the
subscript refers to Fourier-domain. We show in Appendix A
that
M
FD

N
2
N
2
_
1

2
_
2
.
Hence, the solution of (22) increases at most quadratically
with the number of sensors.
Remark 3: The more Fourier series coefcients of

()
are available, the better is the angular resolution capability. As
the angular resolution capability is closely related to the array
aperture, maximizing M implicitly leads to large apertures [2].
Remark 4: Due to the constraints in (7) and (9), the array
covariance matrices of standard planar MRAs contain the
entries

e
j(m1 cos +m2 sin )
,

()
_
,
m
1
, m
2
Z
M
M
, (m
1
, m
2
) ,= (0, 0).
Using truncated Fourier series expansions, it can be easily
veried that the functions
e
j(m1 cos +m2 sin )
, m
1
, m
2
Z
M
M
, (m
1
, m
2
) ,= (0, 0)
(24)
are practically linearly dependent for M 3. Therefore, even
if a standard planar MRA does not contain any non-trivial
redundancies, the off-diagonal entries of the array covariance
matrix are still redundant for azimuth estimation. In other
words, the constraints in (7) and (9) enforce that the spatial
covariance function is estimated for a set of lags, which is
redundant for azimuth estimation.
7
A. Implementation
The problem in (22) is a mixed integer non-linear program-
ming problem. We approach this problem by solving
min
p
xy
,{wm}
1
M
M

m=1
|w
m
|
2
s.t.
_
_
_B
H
w
m
g
m
_
_
_
2

_
N

2
, m N
M
1
(25)
for different values of M. Then, M
FD
is the maximum value
of M such that the solution of (25) is less than or equal to .
The problem in (25) is convex with respect to the weight
vectors, but non-convex with respect to the sensor coordinates.
We divide this problem into an outer optimization of the sensor
coordinates, and an inner optimization of the weight vectors.
Let f
M
(p
xy
) denote the solution of the inner optimization
problem
min
{wm}
1
M
M

m=1
|w
m
|
2
s.t.
_
_
_B
H
w
m
g
m
_
_
_
2

_
N

2
, m N
M
1
.
(26)
To determine the sensor locations that minimize f
M
(p
xy
),
we generate a large number of random array geometries p
xy
.
For each of these random arrays, we determine the optimum
scaling, and subsequently perform a local optimization of the
best scaled arrays.
In Appendix B, we describe a technique to solve (26).
The computational complexity of this technique is O(N

N
4
),
assuming that N

N
2
N. In comparison, the computation
of the number of contiguous lags for a candidate MRA geom-
etry requires only O(N
2
) operations. Moreover, the evaluation
of d(p
xy
) for a candidate Cornwell array requires O(N
4
)
operations. Hence, the computational complexity to evaluate
a candidate array geometry is signicantly higher for the pro-
posed method as compared to the MRA and Cornwell meth-
ods. Compared to the MRA methods, the proposed method
has the advantage that the sensor coordinates are continuous
optimization variables, and therefore continuous optimization
techniques (such as sequential quadratic programming) can be
used to obtain locally optimum arrays.
IV. FOURIER-DOMAIN COVARIANCE AUGMENTATION
TECHNIQUE
Similar to the augmented covariance matrix C
CA
, we dene
the (M
FD
+ 1) (M
FD
+ 1) covariance matrix
C
FDCA

L

l=1
P
l
a
FDCA
(
l
)a
H
FDCA
(
l
) + I
MFD+1
,
where
a
FDCA
()
_
1, e
j
, . . . , e
jMFD

T
is an (M
FD
+ 1) 1 Vandermonde vector. Since C
FDCA
is
Toeplitz and Hermitian, it is fully determined by the entries
of its rst column
[C
FDCA
]
m,1
=
L

l=1
P
l
e
j(m1)
l
+
m,1
, m N
MFD+1
1
.
Hence, the off-diagonal entries of C
FDCA
are Fourier series
coefcients of

().
The rst entry of the rst column of C
FDCA
can be estimated
as _

C
FDCA
_
1,1
= tr

R
x
/N,
where the estimation error is given by
_

C
FDCA
C
FDCA
_
1,1
= tr
R
x
/N, (27)
and tr denotes the trace operator. Using (17), the remaining
entries of the rst column of C
FDCA
can be estimated as
_

C
FDCA
_
m+1,1
= w
H
m
vec

R
x
, m N
MFD
1
,
where the estimation errors are given by
_

C
FDCA
C
FDCA
_
m+1,1
=
m
(),

()+w
H
m
vec
R
x
.
(28)
The essence of the FDCA technique is to estimate the DOAs
by applying the root-MUSIC algorithm to the Toeplitz Hermi-
tian matrix

C
FDCA
. Note that this matrix can be indenite due
to the estimation errors in (27) and (28). A positive denite
Toeplitz Hermitian covariance matrix estimate can be obtained
from

C
FDCA
by a straightforward modication of the technique
of [10], [36]. For the sake of simplicity, we do not consider
this technique in the following.
The vector function a
FDCA
() is one-to-one on [, ),
and M
FD
+ 1 pairwise different Fourier-domain augmented
steering vectors are always linearly independent due to the
Vandermonde structure of these vectors [28]. Similar as in
Section II-B, we can therefore conclude that two different sets
of L < M
FD
+ 1 source azimuth angles cannot lead to the
same column-space of the Fourier-domain augmented steering
matrix
A
FDCA
() = [a
FDCA
(
1
), . . . , a
FDCA
(
L
)] C
(MFD+1)L
.
Thus, there are no manifold ambiguities for a
FDCA
(). For this
reason, the FDCA technique allows to unambiguously estimate
up to M
FD
source DOAs for the full 360

eld-of-view,
assuming that

C
FDCA
is a sufciently accurate estimate of
C
FDCA
. In particular, the FDCA technique allows to estimate
the DOAs of more sources than there are sensors if M
FD
> N.
The rst term on the right hand side of (28) leads to biased
DOA estimates, as it does not depend on the received data.
Therefore, the DOA estimates of the FDCA technique are in-
consistent. Furthermore, we cannot conclude that the proposed
array geometries allow identifying M
FD
source DOAs due
to the approximation errors
m
(). However, our simulation
results suggest that these drawbacks are often irrelevant in
practice, because the nite sample errors usually dominate (28)
(see Section V).
The DOA estimation MSE performance of the CA technique
is often signicantly worse than the stochastic CRB [10]. Our
simulation results also show that the FDCA technique does not
achieve the stochastic CRB. However, the DOA estimates of
both the CA and FDCA techniques can be rened efciently
by a local maximization of the likelihood function as suggested
in [10] in application to the CA approach.
8
(a) (b) (c)
(d)
Fig. 1. Proposed array geometries for N = 9, . . . , 12 sensors, = 10
2

2, and = 1.25.
V. SIMULATION RESULTS
To obtain good feasible points of (22), we generated 10
5
independent random arrays by distributing the sensors uni-
formly inside a sphere of radius one. For each of these arrays
and each value of M, we determined the scaling of p
xy
that minimizes f
M
(p
xy
). Subsequently, for each value of M,
we performed a local optimization of the best 10
3
scaled
arrays. We chose = 10
2

2, so that the L
2
-norms of the
approximation errors
m
() are 40 dB smaller than the L
2
-
norms of exp(jm). Moreover, we chose = 1.25. Fig. 1
depicts the arrays obtained for N = 9, . . . , 12. In Table I, we
list the sensor locations. Fig. 2 shows the corresponding co-
arrays. In these examples, the cardinalities of o
xy
(p
xy
) are
equal to the upper bound, that is N
2
N + 1.
locally optimum points of f
M
(p
xy
) occurred several times.
Hence, the arrays depicted in Fig. 1 are probably not globally
optimum points of (25).
The apertures of the proposed arrays are substantially larger
than the apertures of standard planar MRAs with the same
number of sensors. For example, the aperture of the standard
restricted planar MRA of N = 12 sensors lies between 2
and 2

2, depending on the line of sight [29]. In contrast,


Fig. 1(d) shows that the aperture of the proposed array of
N = 12 sensors is larger than 6. Due to their large apertures,
the proposed arrays provide better azimuth resolution.
For the arrays of Fig. 1, we have
N = 9, M
FD
= 23
N
2
MFD
3.522
N = 10, M
FD
= 29
N
2
MFD
3.448
N = 11, M
FD
= 35
N
2
MFD
3.457
N = 12, M
FD
= 42
N
2
MFD
3.429.
(29)
The latter results suggest that M
FD
increases quadratically with
the number of sensors. It can therefore be expected that also
the apertures of the proposed arrays increase quadratically with
the number of sensors.
Next, we evaluate the performance of the FDCA technique.
Thereto, we consider the array of N = 12 sensors depicted in
Fig. 1(d). The following simulation results are averaged over
the sources.
In our rst example, we consider the case when there
are more sources than sensors. Specically, we assume that
L = 15 equal-power signals with SNR = 0 dB impinge
on the array from the DOAs
l
= (l 1)10

(/180

).
In this case, it is not possible to estimate the signal DOAs
by applying the MUSIC algorithm to

R
x
. In Fig. 3, the
RMSEs and standard deviations of the FDCA estimates are
depicted. Furthermore, we show the RMSE performance of
the rened FDCA estimates, obtained by a local maximization
of the likelihood function. As a reference, we also plot the
stochastic CRB [2].
Fig. 3 shows that the FDCA technique reliably estimates all
signal DOAs, but it does not achieve the stochastic CRB. How-
9
(a) (b)
(c) (d)
Fig. 2. Co-arrays corresponding to the arrays of Fig. 1.
Fig. 1(a) Fig. 1(b) Fig. 1(c) Fig. 1(d)
n pxn
/ pyn
/ pxn
/ pyn
/ pxn
/ pyn
/ pxn
/ pyn
/
1 1.608 1.049 2.521 0.693 -0.405 2.518 -0.016 3.336
2 -1.793 -1.256 -1.650 1.623 2.611 1.328 -1.776 -2.892
3 -0.683 -1.916 1.018 -2.393 -1.568 -2.979 -1.825 3.313
4 -1.136 1.035 0.959 2.003 -1.608 1.153 3.811 0.300
5 -0.566 1.615 -1.974 1.442 -2.277 -0.089 -2.859 -0.170
6 1.789 -1.308 2.563 -0.150 0.224 -3.049 3.849 -0.051
7 -0.816 1.915 -1.800 -1.094 -1.238 -2.666 3.001 2.069
8 2.187 0.355 0.111 2.081 3.541 -0.722 -2.372 2.114
9 -0.591 -1.489 -1.106 -1.736 -1.212 2.049 2.204 -2.123
10 -0.641 -2.468 3.169 -0.001 0.195 -3.250
11 -1.238 2.458 -2.142 0.535
12 -2.069 -3.183
TABLE I
SENSOR LOCATIONS OF THE ARRAYS OF FIG. 1.
10
Fig. 3. DOA estimation performance for more sources than sensors versus
the number of training snapshot vectors.
ever, the RMSE performance of the rened FDCA estimates
obtained by a local maximization of the likelihood function
is close to the stochastic CRB. Moreover, the RMSEs and
standard deviations of the FDCA estimates are close to each
other. Thus, the approximation errors
m
() do not lead to
signicantly biased DOA estimates. Also, note that the FDCA
technique requires rather large numbers of snapshots to resolve
all sources.
In the next example, we vary the angular separation between
the sources. The DOA of the lth source is
l
= (l 1).
The number of snapshots has been set to K = 1000. All other
parameters are chosen as before.
Fig. 4 shows that the FDCA technique resolves all the
sources if their angular separation is above 8

. The azimuth
resolution can be improved by increasing the number of
sensors. Since the results in (29) suggest that M
FD
increases
quadratically with the number of sensors, it can be expected
that a two times larger number of sensors leads to a four
times better angular resolution. points of (21) is prohibitively
high for arrays with large numbers of sensors. However,
it is rather straightforward to develop computationally less
expensive techniques to obtain good feasible points of (21) for
large numbers of sensors. For example, sensor locations may
be optimized one at a time. We keep the study of techniques
as a future research topic.
In our third example, we consider a scenario with manifold
ambiguity. We assume L = 7 equal-power sources with
the DOAs 176.45

, 132.50

, 24.74

, 29.53

, 77.57

,
125.74

, and 153.89

. A manifold ambiguity occurs, because


the steering vector for the DOA 173.24

lies approximately
within !(A()). Hence, the noise-free MUSIC null-spectrum
function is approximately zero for = 173.24

/180

. Thus,
choosing the L smallest minima of the MUSIC null-spectrum
function leads to outliers among the DOA estimates. There are
Fig. 4. DOA estimation performance for more sources than sensors versus
the angular separation between the sources.
several alternatives to the FDCA technique, which allow to
resolve manifold ambiguities. For example, the performance
breakdown of the MUSIC algorithm due to manifold ambi-
guities can be avoided by determining those minima of the
MUSIC null-spectrum function that correspond to the actual
sources by means of compressed sensing (CS) methods [19].
Also, the Capon estimator does not suffer from manifold
ambiguities [2]. In Fig. 5, we compare the FDCA technique
with the Capon estimator and the modied FD root-MUSIC
technique [37], which uses the CS method of [38] to determine
the source roots. The array covariance matrix is estimated
using K = 200 independent snapshot vectors.
Fig. 5 demonstrates that the FDCA technique reliably re-
solves the manifold ambiguity. However, Fig. 5 also shows
that the modied FD root-MUSIC technique and the Capon
estimator outperform the FDCA technique. Therefore, the
main benet of the FDCA technique is that it can estimate
the DOAs of more sources than sensors.
VI. CONCLUSIONS
We have proposed a novel approach to the design of sparse
arrays for azimuthal DOA estimation. The proposed array
design maximizes the number of available contiguous Fourier
series coefcients of the azimuthal power density function.
Our simulation results suggest that the number of available
contiguous Fourier series coefcients increases quadratically
with the number of sensors. Therefore, it can be expected that
the apertures of the proposed arrays also increase quadratically
with the number of sensors.
Based on the proposed array geometry design, we have
developed a subspace-based DOA estimation technique, which
can estimate the DOAs of more sources than sensors, using
only second-order statistics of the received data. The pro-
posed DOA estimation technique provides coverage for the
11
Fig. 5. DOA estimation performance in the case of a manifold ambiguity
versus the SNR.
360

azimuth eld-of-view, and it allows to resolve manifold


ambiguities.
APPENDIX A
PROOF OF REMARK 2
To derive an upper bound on M
FD
, we show that the
feasibility constraints in (23) cannot be satised for arbitrary
large values of M. These constraints imply that
_
_
(I
N

B
H)H
_
_
2
F
2M
2
N

2
, (30)
where
H =
_
g
M
, . . . , g
1
, g
1
, . . . , g
M

C
N(2M)
,
and ||
F
denotes the Frobenius norm. Furthermore, the trian-
gle inequality gives
_
_
(I
N

B
H)H
_
_
F
|H|
F
|
B
HH|
F
. (31)
It can be easily veried that
|H|
F
=
_
2MN

. (32)
Due to the Nyquist sampling criterion, N

is chosen larger
than 2M. Therefore, the columns of H/
_
N

are orthonor-
mal. Consequently,
|
B
HH|
F

_
N

|
B
H|
F
. (33)
Since B is an (N
2
N) N

matrix, we have that


|
B
H|
F

_
N
2
N. (34)
The relations (31)(34) yield
_
_
(I
N

B
H)H
_
_
F

_
2MN

_
N

(N
2
N). (35)
Then, (23) immediately follows from (30) and (35).
APPENDIX B
SOLUTION OF (26)
Since the sensor locations are xed in (26), the weight
vectors can be optimized independently from each other.
Hence, for the mth weight vector we solve
min
wm
|w
m
|
2
2
s.t.
_
_
_B
H
w
m
g
m
_
_
_
2
2

2
N

2
.
(36)
The problem in (36) can be solved efciently by means of a
1D search over the Lagrange parameter, similar as it has been
explained in [39]. The Lagrangian of (36) can be written as
/(w
m
,
m
) = |w
m
|
2
2
+
m
_
_
_
_B
H
w
m
g
m
_
_
_
2
2

2
N

2
_
,
where
m
0 is the Lagrange parameter. Taking the deriva-
tive of the Lagrangian with respect to the weight vector and
setting it to zero yields [40]
w
m
(
m
) =
m
_
I
N
2
N
+
m
BB
H
_
1
Bg
m
. (37)
The latter weight vector minimizes the Lagrangian for a xed
value of
m
.
The rows of B form conjugate pairs. Hence, there is a
unitary matrix Q such that
G = QB (38)
is real-valued, where each row of Q contains exactly two non-
zero entries. The singular value decomposition (SVD) of G
can be expressed as
G = U
G

G
V
T
G
, (39)
where the diagonal matrix

G
= diag
1
, . . . ,
rG
,
1

2
. . .
rG
> 0
contains the singular values in non-ascending order, r
G
is the
rank of G, and
U
G
R
(N
2
N)rG
V
G
R
NrG
contain the orthonormal left- and right-singular vectors, re-
spectively.
Using (38) and (39), we obtain
w
m
(
m
) =
m
Q
H
U
G
_
I
rG
+
m

2
G
_
1

G
V
T
G
g
m
. (40)
Hence, the norm of w
m
(
m
) can be evaluated efciently as
|w
m
(
m
)|
2
=
m
_
rG

k=1
[z
m,k
[
2

2
k
(1 +
m

2
k
)
2
_
1/2
,
where z
m,k
is the kth component of the r
G
1 vector
z
m
= V
T
G
g
m
.
To obtain the optimum Lagrange parameter

m
, we exploit
that the norm of any weight vector w
m
, which satises the
constraint in (36) with strict inequality, can be reduced without
violating the constraint. Therefore, the constraint in (36) has
12
to be satised with equality for the optimum weight vector.
Consequently,

m
is a null of
h
m
(
m
) =
_
_
_B
H
w
m
(
m
) g
m
_
_
_
2
2

2
N

2
. (41)
Equations (38)(40) allow simplifying (41) as
h
m
(
m
) =
rG

k=1
[z
m,k
[
2
(1 +
m

2
k
)
2

m
, (42)
where

m
=
2
N

2
|(I
N

B
H)g
m
|
2
2
is non-negative, assuming that (36) is feasible.
From (37) and (41), we obtain
h
m
(0) = |g
m
|
2
2

2
N

2
= N

_
1

2
2
_
> 0.
Equation (42) shows that h
m
(
m
) monotonically decreases for

m
0. Furthermore,
lim
m
h
m
(
m
) =
m
0.
Hence, h
m
(
m
) has a unique positive null. Since
h
m
(

m
) =
rG

k=1
[z
m,k
[
2
(1 +

2
k
)
2

m
= 0,
we have
rG

k=1
[z
m,k
[
2
(1 +

m

2
1
)
2

m
0

m

1

2
1
_
|z
m
|
2

m
1
_
rG

k=1
[z
m,k
[
2
(1 +

m

2
rG
)
2

m
0

m

1

2
rG
_
|z
m
|
2

m
1
_
rG

k=1
[z
m,k
[
2
(

m

2
k
)
2

m
0

_
1

m
rG

k=1
[z
m,k
[
2

4
k
.
(43)
The bounds in (43) limit the uncertainty in the optimum
Lagrange parameter. This parameter can be computed ef-
ciently using, for example, Newtons method. Our simulation
results suggest that the complexity for computing f
M
(p
xy
)
is dominated by the SVD of G. If N

N
2
N, the
computational complexity of the SVD of Gis O(N

N
4
) [41].
REFERENCES
[1] D. H. Johnson and D. E. Dudgeon, Array Signal Processing: Concepts
and Techniques, A. V. Oppenheim, Ed. Prentice Hall, Englewood Cliffs,
NJ, USA, 1993.
[2] H. L. van Trees, Optimum Array Processing. Wiley, New York, NY,
USA, 2002.
[3] B. Steinberg, Principles of Aperture and Array System Design. Wiley,
New York, NY, USA, 1976.
[4] R. M. Leahy and B. D. Jeffs, On the design of maximally sparse
beamforming arrays, IEEE Trans. Antennas Propag., vol. 39, no. 8,
pp. 11781187, 1991.
[5] S. Holm, A. Austeng, K. Iranpour, and J.-F. Hopperstad, Sparse sam-
pling in array processing, in Sampling Theory and Practice, F. Marvasti,
Ed. Plenum, New York, NY, USA, 2001, pp. 787831.
[6] X. Huang, J. P. Reilly, and M. Wong, Optimal design of linear array
of sensors, in Proc. IEEE Int. Conf. Acoustics, Speech and Signal
Processing (ICASSP), 1991, pp. 14051408.
[7] A. Moffet, Minimum-redundancy linear arrays, IEEE Trans. Antennas
Propag., vol. 16, no. 2, pp. 172175, 1968.
[8] A. R. Thompson, J. M. Moran, and G. W. Swenson, Interferometry and
Synthesis in Radio Astronomy. Wiley-Interscience, New York, 2001.
[9] S. W. Ellingson, Design and evaluation of a novel antenna array
for azimuthal angle-of-arrival measurement, IEEE Trans. Antennas
Propag., vol. 49, no. 6, pp. 971979, 2001.
[10] Y. I. Abramovich, D. A. Gray, A. Y. Gorokhov, and N. K. Spencer,
Positive-denite Toeplitz completion in DOA estimation for nonuni-
form linear antenna arrays. I. Fully augmentable arrays, IEEE Trans.
Signal Process., vol. 46, no. 9, pp. 24582471, 1998.
[11] Y. I. Abramovich, N. K. Spencer, and A. Y. Gorokhov, Positive-denite
Toeplitz completion in DOA estimation for nonuniform linear antenna
arrays. II. Partially augmentable arrays, IEEE Trans. Signal Process.,
vol. 47, no. 6, pp. 15021521, 1999.
[12] S. U. Pillai, Y. Bar-Ness, and F. Haber, A new approach to array
geometry for improved spatial spectrum estimation, Proc. IEEE, vol. 73,
no. 10, pp. 15221524, 1985.
[13] Y. I. Abramovich, N. K. Spencer, and A. Y. Gorokhov, Detection-
estimation of more uncorrelated Gaussian sources than sensors in
nonuniform linear antenna arrays. I. Fully augmentable arrays, IEEE
Trans. Signal Process., vol. 49, no. 5, pp. 959971, 2001.
[14] , Detection-estimation of more uncorrelated Gaussian sources than
sensors in nonuniform linear antenna arrays. II. Partially augmentable
arrays, IEEE Trans. Signal Process., vol. 51, no. 6, pp. 14921507,
2003.
[15] , Detection-estimation of more uncorrelated Gaussian sources than
sensors in nonuniform linear antenna arrays. III. Detection-estimation
nonidentiability, IEEE Trans. Signal Process., vol. 51, no. 10, pp.
24832494, 2003.
[16] A. Manikas, A. Alexiou, and H. R. Karimi, Comparison of the ultimate
direction-nding capabilities of a number of planar array geometries,
IEE Proceedings Radar, Sonar and Navigation, vol. 144, no. 6, pp.
321329, 1997.
[17] A. Manikas and C. Proukakis, Modeling and estimation of ambiguities
in linear arrays, IEEE Trans. Signal Process., vol. 46, no. 8, pp. 2166
2179, 1998.
[18] A. Manikas, C. Proukakis, and V. Lefkaditis, Investigative study of
planar array ambiguities based on hyperhelical parameterization,
IEEE Trans. Signal Process., vol. 47, no. 6, pp. 15321541, 1999.
[19] Y. I. Abramovich, N. K. Spencer, and A. Y. Gorokhov, Resolving
manifold ambiguities in direction-of-arrival estimation for nonuniform
linear antenna arrays, IEEE Trans. Signal Process., vol. 47, no. 10, pp.
26292643, 1999.
[20] A. Manikas, A. Sleiman, and I. Dacos, Manifold studies of nonlinear
antenna array geometries, IEEE Trans. Signal Process., vol. 49, no. 3,
pp. 497506, 2001.
[21] J. Leech, On the representation of 1, 2, . . . , n by differences, J.
London Math. Soc., vol. 31, pp. 160169, 1956.
[22] D. Pearson, S. U. Pillai, and Y. Lee, An algorithm for near-optimal
placement of sensor elements, IEEE Trans. Inf. Theory, vol. 36, no. 6,
pp. 12801284, 1990.
[23] L. R edei and A. R enyi, On the representation of 1, 2, . . . , N by
differences, Mat. Sb. (N.S.), vol. 24 (66), no. 3, pp. 385389, 1949,
(in Russian).
[24] D. A. Linebarger, I. H. Sudborough, and I. G. Tollis, Difference bases
and sparse sensor arrays, IEEE Trans. Inf. Theory, vol. 39, no. 2, pp.
716721, 1993.
[25] W. Rudin, Principles of Mathematical Analysis. McGraw-Hill, Tokyo,
Japan, 1964.
[26] R. O. Schmidt, Multiple emitter location and signal parameter estima-
tion, in Proc. RADC Spectral Estimation Workshop, Rome, NY, USA,
Oct. 1979, pp. 243258.
[27] G. Bienvenu and L. Kopp, Adaptivity to background noise spatial
coherence for high resolution passive methods, in Proc. IEEE Int. Conf.
Acoustics, Speech and Signal Processing (ICASSP), vol. 5, Denver, CO,
USA, April 1980, pp. 307310.
[28] L. Mirsky, An Introduction to Linear Algebra. Oxford University Press,
Oxford, UK, 1955.
[29] Y. Meurisse and J.-P. Delmas, Bounds for sparse planar and volume
arrays, IEEE Trans. Inf. Theory, vol. 47, no. 1, pp. 464468, 2001.
[30] S. Haykin, J. P. Reilly, V. Kezys, and E. Vertatschitsch, Some aspects
of array signal processing, IEE Proc. F Radar Signal Process., vol.
139, no. 1, pp. 126, 1992.
[31] C. R. Greene and R. C. Wood, Sparse array performance, J. Acoust.
Soc. Amer., vol. 63, no. 6, pp. 18661872, 1978.
13
[32] H. C. Pumphrey, Design of sparse arrays in one, two, and three
dimensions, J. Acoust. Soc. Amer., vol. 93, no. 3, pp. 16201628, 1993.
[33] J.-W. Liang and A. J. Paulraj, On optimizing base station antenna array
topology for coverage extension in cellular radio networks, in Proc.
IEEE 45th Vehicular Technology Conf, vol. 2, 1995, pp. 866870.
[34] T. J. Cornwell, A novel principle for optimization of the instantaneous
Fourier plane coverage of correlation arrays, IEEE Trans. Antennas
Propag., vol. 36, no. 8, pp. 11651167, 1988.
[35] M. A. Doron and E. Doron, Waveeld modeling and array processing,
Part I Spatial sampling, IEEE Trans. Signal Process., vol. 42, no. 10,
pp. 25492559, 1994.
[36] K. M. Grigoriadis, A. E. Frazho, and R. E. Skelton, Application
of alternating convex projection methods for computation of positive
Toeplitz matrices, IEEE Trans. Signal Process., vol. 42, no. 7, pp.
18731875, 1994.
[37] M. R ubsamen and A. B. Gershman, Direction-of-arrival estimation for
nonuniform sensor arrays: From manifold separation to Fourier domain
MUSIC methods, IEEE Trans. Signal Process., vol. 57, no. 2, pp. 588
599, 2009.
[38] P. Stoica, P. Babu, and J. Li, SPICE: A sparse covariance-based
estimation method for array processing, IEEE Trans. Signal Process.,
vol. 59, no. 2, pp. 629638, 2011.
[39] 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.
[40] D. H. Brandwood, A complex gradient operator and its application
in adaptive array theory, IEE Proc. F Communications, Radar and
Signal Processing, vol. 130, no. 1, pp. 1116, Feb. 1983.
[41] G. H. Golub and C. H. van Loan, Matrix Computations. The Johns
Hopkins University Press, Baltimore, MD, USA, 1996.
Michael R ubsamen (S05) received the Dipl.-Ing.
degree from RWTH Aachen University, Germany,
in 2006, and the Dr.-Ing. degree from Technische
Universit at Darmstadt, Germany, in 2011, both in
electrical engineering.
In 20042005, he held a research scholarship from
Bell-Labs, Crawford Hill, USA, where he worked
on signal processing techniques for ber optic com-
munication systems. In June 2006, he joined the
Communication Systems Group at Technische Uni-
versit at Darmstadt, Darmstadt, Germany, where he
is currently working as a postdoctoral researcher. In 2009, he was a visiting
scientist at Temasek Labs at Nanyang Technological University (NTU),
Singapore. His research interests are focussed on array signal processing,
signal processing for communication systems, beamforming, and direction-
of-arrival estimation.
Dr. R ubsamen received the Student Best Paper Award at the IEEE Sensor
Array and Multichannel Signal Processing (SAM) Workshop in 2010. He was
student contest nalist at the IEEE International Conference on Acoustics,
Speech, and Signal Processing (ICASSP) in 2008. He was awarded a grant
from the German National Academic Foundation in 2003. In 2002 and 2003,
he received the Philips Student Award and the Henry-Ford-II Student Award,
respectively. He served as local organization co-chair for the IEEE SAM
Workshop in Darmstadt in 2008.
Alex B. Gershman (M97SM98F06), deceased,
received his Diploma and Ph.D. degrees in Radio-
physics from the Nizhny Novgorod State University,
Russia, in 1984 and 1990, respectively.
From 1984 to 1999, he held several full-time and
visiting research appointments in Russia, Switzer-
land, and Germany. In 1999, he joined the De-
partment of Electrical and Computer Engineering,
McMaster University, Hamilton, Ontario, Canada,
where he became a professor in 2002. From April
2005, he has been with the Technische Universit at
Darmstadt, Darmstadt, Germany, as a professor of Communication Sys-
tems. His research interests were in the area of signal processing and
communications with the primary emphasis on array processing, statistical
signal processing, beamforming, multi-antenna, multi-user, and cooperative
communications, and estimation and detection theory.
Prof. Gershman was the (co-)recipient of several awards including the 2004
IEEE Signal Processing Society (SPS) Best Paper Award; the IEEE Aerospace
and Electronic Systems Society (AESS) Barry Carlton Award for the best
paper published in 2004; the CrownCom 2010 Best Paper Award, the 2002
Young Explorers Prize from the Canadian Institute for Advanced Research
(CIAR); the 2001 Wolfgang Paul Award from the Alexander von Humboldt
Foundation, Germany; and the 2000 Premiers Research Excellence Award,
Ontario, Canada. He also co-authored the paper that received the 2005 IEEE
SPS Young Author Best Paper Award.
Prof. Gershman was a member of the IEEE SPS Board of Governors.
He was Editor-in-Chief for the IEEE Signal Processing Letters (2006-2008);
Chair of the Sensor Array and Multichannel (SAM) Technical Committee of
the IEEE SPS (2007-2008); Associate Editor of the IEEE Transactions on
Signal Processing (1999-2005); Technical Co-Chair of the IEEE International
Symposium on Signal Processing and Information Technology (ISSPIT),
Darmstadt, Germany, December 2003; Technical Co-Chair of the 4th IEEE
Sensor Array and Multichannel Signal Processing Workshop, Waltham, MA,
USA, June 2006; General Co-Chair of the 1st IEEE Workshop on Compu-
tational Advances in Multi-Sensor Adaptive Processing (CAMSAP), Puerto
Vallarta, Mexico, December 2005; General Co-Chair of the 5th IEEE Sensor
Array and Multichannel Signal Processing Workshop, Darmstadt, Germany,
July 2008; and Tutorial Chair of EUSIPCO, Florence, September 2006.
Prof. Gershman died in 2011.

You might also like