1

Design and Analysis of Compressive Antenna

Arrays for Direction of Arrival Estimation
Mohamed Ibrahim∗ , Venkatesh Ramireddy, Anastasia Lavrenko, Jonas König, Florian Römer, Markus Landmann,
Marcus Grossmann, Giovanni Del Galdo, and Reiner S. Thomä

Abstract — In this paper we investigate the design of level of estimation accuracy and robustness in the presence
arXiv:1611.03699v1 [cs.IT] 11 Nov 2016

compressive antenna arrays for direction of arrival (DOA) of multiple sources and/or multiple paths. Furthermore, the
estimation that aim to provide a larger aperture with a developed methods shall be applicable in practical applica-
reduced hardware complexity by a linear combination tions with realistic antenna arrays whose characteristics often
of the antenna outputs to a lower number of receiver significantly vary from commonly considered ideal models [8].
channels. We present a basic receiver architecture of In the last few decades, research on direction of arrival
such a compressive array and introduce a generic system (DOA) estimation using array processing has largely focused
model that includes different options for the hardware on uniform arrays (e.g., linear and circular) [2] for which many
implementation. We then discuss the design of the analog efficient parameter estimation algorithms have been developed.
combining network that performs the receiver channel Some well-known examples are ESPRIT [9], MUSIC [10]
reduction, and propose two design approaches. The first and Maximum Likelihood (ML)-based methods [6], [11]. Note
approach is based on the spatial correlation function which that ML-based methods are particularly suitable for realistic,
is a low-complexity scheme that in certain cases admits non-ideal antenna arrays since they can easily account for
a closed-form solution. The second approach is based the full set of parameters of the antenna array (e.g., antenna
on minimizing the Cramér-Rao Bound (CRB) with the polarization, non-ideal antennas and array geometries, etc.).
constraint to limit the probability of false detection of However, to perform well, the algorithms require to fulfill
paths to a pre-specified level. Our numerical simulations certain conditions on the sampling of the wavefront of the
demonstrate the superiority of the proposed optimized incident waves in the spatial domain. Namely, the distance
compressive arrays compared to the sparse arrays of the between adjacent sensors should be less than or equal to half
same complexity and to compressive arrays with randomly a wavelength of the impinging planar wavefronts, otherwise
chosen combining kernels. it leads to grating lobes (sidelobes) in the spatial correlation
function which correspond to near ambiguities in the array
Keywords: Compressive Sensing, DOA Estimation, Measure-
manifold. At the same time, to achieve DOA estimation with
ment Design
a high resolution, the receiving arrays should have a relatively
large aperture [2]. This implies that arrays with a large number
I. I NTRODUCTION of antennas are needed to obtain a high resolution, which is
Direction of arrival (DOA) estimation has been an active not always feasible.
field of research for many decades [1]. In general, DOA This limitation has triggered the development of arrays with
estimation addresses the problem of locating sources which inter-element spacing larger than half the impinging wave’s
are radiating energy that is received by an array of sensors wavelength combined with specific constraints to control the
with known spatial positions [2]. Estimated DOAs are used in ambiguity problem in DOA estimation. Such arrays are usually
various applications like localization of transmitting sources, called sparse arrays. In [12], it was proposed to constitute
for direction finding [3], [4], massive MIMO and 5G Networks a non-uniform sparse array with elements spaced at random
[4], channel sounding and modeling [5], [6], tracking and positions. However, using such random arrays will often result
surveillance in radar [7], and many others. A major goal in in an unpredictable behavior of the sidelobes in the array’s
research on DOA estimation is to develop approaches that spatial correlation function. As a result, it is necessary to
allow to minimize hardware complexity in terms of receiver optimize the positions of the antenna elements in order to
costs and power consumption, while providing a desired achieve a desired performance. An early approach towards that
goal was the Minimum Redundancy Linear Array (MRLA)
The authors M. Ibrahim, V. Ramireddy, A. Lavrenko, J. König, F. Römer,
G. Del Galdo and R. S. Thomä are with Ilmenau University of Technology, [13], where it is proposed to place the antenna elements such
P. O. Box 10 05 65, 98684 Ilmenau, Germany. The authors G. Del Galdo, that the number of pairs of antennas that have the same spatial
R. S. Thomä, M. Landmann, and M. Grossmann are with Fraunhofer Institute correlation properties are as small as possible. However, it
for Integrated Circuits IIS, Helmholtzplatz 2, 98683 Ilmenau, Germany. Parts
of this work have been presented as conference papers at the 40th International is very difficult to construct an MRLA when the number of
Conference on Acoustics, Speech and Signal Processing (ICASSP), Brisbane, elements is relatively large [14]. Some non-linear optimization
Australia in April 2015 (Theorem 1 and Corollary 1) and at the 23rd European methods like genetic algorithms [15] and simulated annealing
Signal Processing Conference (EUSIPCO), Nice, France in September 2015
(adaptive design approach from Section IV-C). [16] have been regularly used to find optimum configurations
∗ corresponding author for these sparse arrays. In more recent works, V-shaped arrays
2

[17], Co-Prime arrays [18], and Nested arrays [19] have been Depending on the frequency range, the components of the
proposed to extend the effective array aperture. analog combining network (power combiners, power splitters,
Recently, compressed sensing (CS) [20]–[22] has been phase shifters) will induce additional losses which also have
widely suggested for applications that exhibit sparsity in to be compensated by the LNAs. To name an example, some
time, frequency or space to reduce the sampling efforts. The typical commercially available phase shifters for phased array
application of sparse recovery to DOA estimation has been radar applications can induce insertion losses between 5 to
considered for applications like localization of transmitting 10 dB depending on the frequency range [41]. This motivates
sources [23], channel modeling [24], tracking and surveillance the need for the signal amplification prior the combining
in radar [25], and many others. It is highlighted in [26] that network.
if the electromagnetic field is modeled as a superposition Based on the generic system model we then discuss the
of a few plane waves, the DOA estimation problem can be design of the combining matrix, with the goal to obtain an
expressed as a sparse recovery problem. The main focus is to array that is suitable for DOA estimation (i.e., minimum
use the sparse recovery algorithms that became popular in the variance of DoA estimates and robustness in terms of low side
CS field for the DOA estimation problem as an alternative to lobe levels or low probability of false detections). We consider
existing parameter estimation algorithms [27]–[30]. two design approaches. The first approach is based on the
Compressed sensing has also been suggested to be applied spatial correlation function which is a low-complexity scheme
in the spatial domain (e.g., array processing and radar) with that in certain cases even admits a closed-form solution. The
the main goal to reduce the complexity of the measurement second approach is based on the minimization of the Cramér-
process by using fewer RF chains and storing less measured Rao Bound (CRB). CRB-minimizing array designs tend to
data without the loss of any significant information. Hence, result in high sidelobes in the spatial correlation function
the idea of sparse random arrays with increased aperture size which may lead to false estimates. In order to be able to
has been recently revisited and proposed to perform spatial constrain this effect, we analytically derive the probability
compressed sensing [31]–[34]. to detect a false peak (sidelobe) for a given array manifold.
An alternative approach that attempts to apply CS to the We then use this expression as an additional constraint in
acquisition of the RF signals that are used for DOA estima- our design, thus limiting the probability of false detection
tion has recently been proposed in [35], [36]. In particular, to a pre-specified level. Our numerical simulations demon-
the CS paradigm can be applied in the spatial domain by strate that both proposed design approaches have a significant
employing N antenna elements that are combined using an performance improvement compared to the state of the art,
analog combining network to obtain a smaller number of namely an array with a randomly chosen combining matrix and
M < N receiver channels. Since only M channels need to a sparse array with optimized sensor positions. Furthermore,
be sampled and digitized, the hardware complexity remains the compressive array is not only superior to the random and
comparably low while a larger aperture is covered which sparse arrays with respect to its estimation capabilities but also
yields a better selectivity than a traditional, Nyquist (λ/2) in terms of its ability to alter its weights on demand and thus
spaced L-channel antenna array. In baseband, the operation of facilitate signal-adaptive measurements. The comparison be-
the combining network can be described by complex weights tween the proposed designs demonstrates a trade-off between
applied to the antenna outputs with a subsequent combination the minimization of the CRB and the increase in the sidelobe
of the received signals from the antennas. The combining level. In the proposed design, the trade-off “CRB vs. sidelobe
(measurement) matrix that contains the complex weights and level” can be controlled by setting the parameters during the
the antenna array form an effective “compressive” array whose optimization. This provides an additional degree of freedom
properties define the DOA estimation performance. In the field for the system design that is unavailable in case of random
of “CS-DOA” it is usually advocated to draw the coefficients and sparse arrays.
of the measurement matrix from a random distribution (e.g., It is worth mentioning that similar efforts in spatial domain
Gaussian, Bernoulli) [35], [36]. Random matrices have certain processing exist in the context of beam space array processing
guarantees for signal recovery in the noise-free case and [42]–[49] and hybrid beamforming [50]–[52]. In contrast to
provide some stability guarantees in the noisy case [37]–[39]. the element space processing, where signals derived from each
However, since no criterion is used to design them, it is likely element are weighted and summed to produce the array output,
that they provide sub-optimal performance [40]. the beam space processing is a two-stage scheme. The first
In this paper, we discuss the design and the performance stage takes the array signals as an input and produces a set
of compressive arrays employing linear combinations in the of multiple outputs, which are then weighted and combined to
analog domain by means of a network of power splitters, phase form the array output. These multiple outputs may be thought
shifters, and power combiners. We present a basic receiver of as the output of multiple beams. The weights applied to
architecture of such a compressive array and introduce a different beam outputs are finally optimized according to a
generic system model that includes different options for the specific optimization criterion [53]. In hybrid beamforming,
hardware implementation. Importantly, the model reflects the the main idea is to apply beamforming and precoding tech-
implications for the noise sources. Particularly, a well-known niques in both, the radio-frequency (RF) and the baseband
source of the receiver noise is the low noise amplifier (LNA) (BB) [54]. This technique has attracted significant research
that is usually placed at the antenna outputs to account for the attention in millimeter wave (mmWave) applications [55] for
power losses of the following distribution/combining network. the next-generation indoor and mobile wireless networks [56],
 3

[57]. While the overall goal in these areas is similar (reducing domain (e.g. FPGA) and power consumption. In order to
the number of digitally processed receiver channels), the actual reduce the number of RF channels without a loss in the
design criterion for the antenna is entirely different from the array aperture, we apply the compressive approach, where
one we consider in this paper. We aim to obtain an array the antenna outputs are first linearly combined in the analog
that is ideally suited for DOA estimation in the sense that domain and then passed through a lower number of RF chains
it achieves an accurate estimate (by minimizing the CRB) to obtain the digital baseband signals as illustrated in Figure 1.
while controlling the sidelobe characteristics by a prespecified In this way, M RF receiver channels (fewer than the N antenna
probability of detecting a false direction. elements) are used for signal processing in the digital domain.
This manuscript is organized as follows. Section II intro- The signal combining can be done at different stages within
duces the data model for the compressive arrays we consider the receiver, e.g., on the RF (Radio Frequency) signal or at
and discusses the impact of the different sources of noise. the IF (Intermediate Frequency) stage. The particular choice
The proposed design is shown in Section III, where we on where to place the combining network highly depends on
discuss the general approach as well as the two specific design the application, especially the considered frequency. In any
methods based on the spatial correlation function and the CRB, case, additional signal losses will be introduced by the power
respectively. Section III also contains the derivation of the splitters and combiners as well as the phase shifters inside
analytical expression for the probability to detect a false peak. the combining network. The actual losses’ value will depend
A discussion is contained in Section IV, covering various pos- on multiple parameters including frequency, bandwidth and
sible extensions and an analysis of the achievable estimation adaptability of the phase shifters. However, these losses need
accuracy compared to sparse arrays. Numerical results are to be compensated by LNAs placed in each receiver chain as
presented in Section V before concluding in Section VI. shown in Figure 1.
To this end, let Φ ∈ CM×N denote the analog combining
II. S YSTEM MODEL FOR COMPRESSIVE ARRAYS matrix of a compressive array which compresses the output
of N antenna elements to M active RF channels. Then, the
A. Data model for narrowband DOA estimation complex (baseband) antenna output (2) after combining can
Consider K narrowband plane waves impinging on an array be expressed as
of N antenna elements. At the antenna output, the received
(baseband) signal can be expressed as ỹ(t) = Φ (A · s(t) + v(t)) + w(t), (3)
K
X where [Φ]m,n = αm,n eϕm,n , αm,n ∈ [0, 1], ϕm,n ∈
y(t) = a(γk ) · sk (t) + n(t), (1) [0, 2π], m = 1, 2, · · · M, n = 1, 2, · · · , N 1 , whereas v(t) ∈
k=1 CN ×1 and w(t) ∈ CM×1 are noise vectors with covariances
Rvv and Rww that represent additive noise sources which act
where y(t) ∈ CN ×1 is a vector of antenna outputs, n(t) ∈ before and after2 the combining network, respectively. For ex-
CN ×1 is an additive noise vector, t indicates the continuous ample, LNAs placed ahead the combining network contribute
time, and a(γ) denotes the antenna response as a function to v(t) (signal noise), whereas the ones placed behind the
of the parameter vector γ T = [θ, ψ, pT ] with θ and ψ combining network contribute to w(t) (measurement noise).
being the azimuth and the elevation angles, while p ∈ C2×1 Let à = Φ · A be the effective array steering matrix after
represents the Jones vector that describes the polarization combining, then (3) becomes
state of the incident plain wave at the receiver. Additionally,
sk (t) in (1) denotes the amplitude of the kth source, whereas ỹ(t) = Ã · s(t) + ñ(t), (4)
γkT = [θk , ψk , pTk ] is the vector containing its azimuth (θk ) where
and elevation (ψk ) angles of arrival along with its Jones vector
ñ(t) = Φ · v(t) + w(t) (5)
pTk = [pk,1 , pk,2 ]. It is often useful to write (1) in a matrix
from as is the effective noise vector with covariance Rnn =
y(t) = A · s(t) + n(t). (2) ΦRvv ΦH + Rww . Assuming that v(t) and w(t) are white
with elements that have variance σ12 and σ22 , respectively, the
Here, A = [a(γ1 ), a(γ2 ), · · · , a(γK )] ∈ CN ×K is the array covariance of ñ(t) becomes Rnn = σ12 ΦΦH + σ22 I.
steering matrix and s(t) = [s1 (t), s2 (t), · · · , sK (t)]T ∈ CK×1 Given (4), we aim to design Φ in such a way that it allows
is a vector containing the complex amplitudes of the K for a robust and efficient estimation of the DOAs of the K
sources. sources sk (t) from the set of measurements ỹ(t). Hence, our
main design goal includes the minimization of the number of
B. Compressive arrays the receiver chains while providing a minimum variance of the
DOA estimates and a reduced probability of spurious and/or
The model in (2) presumes a dedicated radio frequency (RF)
ghost path estimates.
receiver chain for each individual antenna element including a
low-noise amplifier (LNA), filters, down-conversion, analog- 1 Note that Φ does not need to be fully meshed, i.e., we do not need to

to-digital (ADC) conversion, etc. For specific applications, connect each of the N antennas to each of the M outputs. In case the nth
however, such separate RF chains for each antenna element antenna is not connected to the mth RF channel the corresponding entry in
the combining matrix is set to zero, e.g., [Φ]m,n = 0.
may come at a high cost in terms of the overall receiver 2 In the CS community, the former is often referred to as “signal noise” and
complexity, the amount of data to be processed in the digital the latter as “measurement noise”.
4

Figure 1. Compressive array hardware architecture

III. D ESIGN OF THE COMBINING MATRIX spatial correlation function (SCF) and the Cramér-Rao Lower
A. Generic design approach Bound (CRB).
For the SCF-based approach, we build our design on the
Consider the receiver architecture from Fig. 1 where the spatial correlation function defined as
combining network is realized by: (i) splitting the analog RF
signal of each of the N antennas into L ≤ M branches; (ii) ρ(γ1 , γ2 ) = ã(γ1 )H · ã(γ2 ), (8)
applying phase shifts in each of the branches; (iii) adding the where ã(γ) = Φ · a(γ) presents the effective array manifold
branches to form each of the M outputs, which are then passed after combining. The main idea is to design Φ such that
to the M RF chains. Mathematically, we model this structure the spatial correlation function ρ(γ1 , γ2 ) follows as close as
by a matrix Φ with elements given by possible some pre-specified target T (γ1 , γ2 ). By defining an
(
√1 · η · eϕm,n if (n, m) are connected appropriate target function, we can provide desired properties
[Φ]m,n = L (6) in the spatial correlation function as discussed in the following.
0 otherwise,
Although SCF is often useful for gaining initial insight
where the connections between antennas and ports are such into the array’s estimation capabilities, its ability to provide
that Φ has L nonzero elements per column. In (6), the factor quantitative evaluation of the achievable estimation quality is
√1 represents the power splitting of each antenna’s signal to L limited, especially in the case of multiple sources. Therefore,
L
branches and η ∈ (0, 1] is a scalar parameter that attributes for we propose a second approach that is based on the specific
the fact that each analog branch (consisting of a power splitter, requirements on the estimation accuracy. More specifically,
a phase shifter, and a combiner) is non-ideal and incorporates it aims at improving the accuracy of DOA estimation by
losses. A loss-less combining network would correspond to designing Φ such that it minimizes the CRB while keeping the
the special case η = 1. probability of detecting a false direction at a certain (desired)
From (6), the combining matrix Φ has M N elements that level.
provide M N degrees of freedom for its design. In the CS
literature, a typical approach for choosing Φ would be to draw B. Design based on the SCF
ϕm,n randomly. This, however, gives little control over the For the sake of simplicity, in the remainder of the paper it
array characteristics. Furthermore, it might result in unwanted is assumed that the sources are located in the azimuthal plane
effects as high sidelobes and blind spots [40]. of the antenna array and have an identical polarization of the
Here, we aim at a design of Φ that results in an effective impinging waves with perfectly matched antennas. Hence, the
array that has desired properties depending on the application effective array manifold depends on the azimuth angle θ only,
scenario, e.g., uniform sensitivity and low cross-correlation for i.e., ã(γ1 ) = ã(θ1 ). Note that an extension to a more general
direction finding, adaptive spatial selectivity for parameter es- case is straightforward and is sketched in Section IV-A.
timation during beam tracking, etc. Generally, the design task Under these assumptions, an ideal generic array for direction
can be formulated as the following constrained optimization finding would satisfy the conditions
problem (
H const θ1 = θ2
Φopt = arg minJ(Φ) s.t. c(Φ, α, β, · · · ), (7) ρ(θ1 , θ2 ) = ã(θ1 ) · ã(θ2 ) = . (9)
Φ 0 θ1 6= θ2
where J(Φ) is some objective function defined by the sce- The first condition guarantees that the array gain is constant
nario and c(Φ, α, β, · · · ) represents the set of optimization over all azimuth angles and makes the array uniformly sen-
constraints. In the following, we propose two particular for- sitive, whereas the second condition forces optimal cross-
mulations of (7) for direction finding applications: based on the correlation properties to tell signals from different directions
 5

apart. However, this is an example for a generic direction in (12) is not sufficiently selective since all row-orthogonal
finder. For particular applications, the design goal may differ, matrices achieve the same minimum of the cost function.
i.e., constraining on a certain sector of angles only or allowing The cost function (12) assigns an equal weight to the error
(G) (G)
certain values for the residual cross-correlation. We denote the for all pairs of grid points θ1 , θ2 , i.e., it tries to maintain
target function as T (θ1 , θ2 ), where T (θ1 , θ2 ) = const · δ(θ1 − a constant main lobe with the same weight as it tries to
θ2 ) represents the ideal generic array (9). minimize sidelobes everywhere. In practice it is often desirable
Due to the finite aperture of an N -element array, the target to have more control over the shape of the spatial correlation
in (9) can only be achieved approximately. This allows us to function, e.g., trading main lobe ripple against sidelobe levels
define a criterion for the optimization of Φ according to the or allowing for a transition region between the mainlobes and
cost function sidelobes that is not constrained. There are many ways such
constraints could be incorporated, e.g., maximum constraints
e(Φ, θ1 , θ2 ) = ã(θ1 )H · ã(θ2 ) − T (θ1 , θ2 ) (10) on the magnitude of cross-correlation in some region and
= a(θ1 )H · ΦH · Φ · a(θ2 ) − T (θ1 , θ2 ) . interval constraints on the autocorrelation inside the mainlobe.
For numerical tractability, we follow a simpler approach by
We can approximate the continuous variables θ1 and θ2 by
(G) introducing a weighting matrix W ∈ RP ×P into (12). The
considering the P -point sampling grid θp , p = 1, 2, . . . , P modified optimization problem is given by
and define the P × P matrices E and T according to E(i,j) =
(G) (G) (G) (G) 2
e(Φ, θi , θj ) and T(i,j) = T (θi , θj ). After insertion Φopt = arg min kE ⊙ W kF , (13)
Φ
into (10) we obtain
where ⊙ represents the Schur (element-wise) product. The
E = AH · ΦH · Φ · A − T . (11) weighting matrix allows to put more or less weight on the main
diagonal (controlling how strictly the constant mainlobe power
Based on (11), the quality of Φ can be assessed by a suitable shall be enforced), certain off-diagonal regions (controlling
norm of E. As a first step, let us consider the Frobenius norm, how strongly sidelobes in these regions should be suppressed),
i.e., we optimize Φ according to or even placing zeros for regions that remain arbitrary (such
Φopt = arg min kEkF .
2
(12) as transition regions between the mainlobe and the sidelobes).
Φ Thereby, more flexibility is gained and the solution can be
tuned to more specific requirements.
In the special case3 where A · AH = C · IM , with C being
The drawback of (13) is that it does not admit a closed-form
a constant, the optimization problem in (12) admits a closed-
solution in general. However, it can be solved by numerical
form solution as shown in the following theorem.
optimization routines that are available in modern technical
Theorem 1. Let S = A · T · AH and let SM be a rank-M - computing languages.
truncated version of S obtained by setting its P − M smallest
eigenvalues to zero. Then the set of optimal solutions to (12) C. Design based on the CRB
is given by the set of matrices Φ that satisfy ΦH Φ = SM .
For the case of a single source, a correlation-based DOA
Proof: cf. Appendix A. estimator amounts to finding the DOA θ0 that corresponds to
In other words, Theorem 1 states that we can find an the global maximum in the beamformer spectrum D(θ), i.e.,
optimal Φ by computing a square-root factor of the best rank-  H 
M approximation of S. Moreover, the following corollary can ã (θ) · R · ã(θ)
θ0 = arg max D(θ) ≡ arg max , (14)
be found from Theorem 1: θ θ kã(θ)k22

Corollary 1. Under the conditions of Theorem 1 any matrix Φ where R = E{ỹ(t) · ỹ H (t)} is the covariance matrix of the
is optimal in terms of the “ideal” target from (9) if and only if received signals. Note that in this case, (14) is equivalent to
the rows of Φ have equal norm and are mutually orthogonal. the maximum likelihood (ML) cost function, and therefore,
the correlation-based DOA estimator is equivalent to the ML
Proof: cf. Appendix B. estimator. We define then the false detection as the event
Corollary 1 agrees with the intuition that the measurements where the global maximum in the beamformer spectrum D(θ)
(i.e., the rows of Φ) should be chosen such that they are is outside the mainlobe area (either 3-dB or null-to-null
orthogonal in order to make every observation as informative beamwidth). The error probability Pd is hence given by
as possible. In addition, the corollary shows that this choice

also minimizes ΦH Φ − C · P · IN F which contains the Pd ≡ Prob D(θ0 ) < D(θ), ∀θ ∈ U , (15)
correlations between all pairs of columns in Φ as well as where U denotes the set of DOAs corresponding to the
the deviation of the columns’ norms (therefore, in a sense, directions in the beamformer spectrum outside the mainlobe
this choice minimizes the “average” mutual correlation). On area, and
the other hand, this also demonstrates that the optimization
2 2
D(θ0 ) = ãH
0 ỹ(t) = ãH H
0 ã0 s(t) + ã0 ñ(t) ,
3 This condition is, e.g., fulfilled for an ULA if the sampling grid is chosen
2 2
to be uniform in the spatial frequencies (direction cosines). Moreover, for D(θ) = ãH ỹ(t) = ãH ã0 s(t) + ãH ñ(t) (16)
many arrays the condition is approximately fulfilled (e.g., for UCAs). In this
case, the closed-form solution can still be applied as a heuristic method. with ãk ≡ Φ · a(θk ) and ã ≡ Φ · a(θ).
6

A direct evaluation of (15) is analytically intractable. In hence to improve the DOA estimation accuracy over the
order to proceed, we adopt the approach from [58] and full angular range (0, 2π], while lowering the false detection
approximate the continuous correlation function b(θ, θ0 ) = probability to a desired value ǫ0 for a given SNR threshold
|ãH ã0 |/(kãk kã0 k) of the antenna array by its discretized point ρth .
version Both the constraint and the objective in (22) are non-convex
L
X functions with respect to (Φ, θ0 ). The optimization problem
bd (θ, θ0 ) = b(θ, θ0 )δ(θ − θq ), (17) is thereby a non-convex problem exhibiting a multi-modal
q=0 cost function, where the optimal (global) solution can only be
where θ0 ∈ / U and θq ∈ U ∀q > 0 are the directions found by an exhaustive search strategy. Therefore, we apply a
corresponding to the mainlobe and the sidelobe peaks in the local minimizer to the above problem using an algorithm based
array’s correlation function, respectively, whereas L is the total on the interior-reflective Newton method [61], [62]. However,
number of the sidelobe peaks and δ(θ) denotes the Dirac delta by using this algorithm the obtained solution strongly depends
function. Using (17), the false detection probability can now on the initialization of the parameters (Φ, θ0 ). Moreover, there
be approximated by is no guarantee that the global optimum is found. One way of
addressing this issue is to apply the algorithm several times,
[ L 
 where for each run the initialization of the parameters (Φ, θ0 )
Pd ≈ Prob D(θ0 ) − D(θq ) < 0 . (18) is different. In doing so, the obtained solution to (22) is
q=1
likely to be sufficiently close to the optimal solution. However,
In order to simplify the calculation, we apply the union bound it might be time consuming due to the complexity of the
[59] on (18), and obtain optimization problem at hand. Another way of tackling this
L problem is to first obtain a solution for Φ by the SCF approach
X

Pd ≤ Prob D(θ0 ) − D(θq ) < 0 . (19) described above and then use it for the initialization in (22).
q=1

Now it remains to compute
the individual probabilities Pq ≡ IV. D ISCUSSION

Prob D(θ0 ) − D(θq ) < 0 , ∀q ∈ [1, L], where Pq denotes
the probability that the q-th sidelobe peak is higher than the A. Multi-dimensional DOA
mainlobe peak.
For simplicity, we have discussed the design of the com-
Theorem 2. Suppose v(t) and w(t) are independent zero- pression matrices only for 1-D case, i.e., the estimation of
mean complex white Gaussian noise vectors with covariances the azimuth angle, assuming that all the sources are located
σ12 I and σ22 I, respectively and let Ψq (s) be the moment on the plane of the array. However, it is straightforward to
generating function (MGF) of (D(θ0 ) − D(θq )) defined

by generalize the design to the 2-D case where both, azimuth and
(39) in App. C. Then, Pq = Prob D(θ0 ) − D(θq ) < 0 can elevation, are considered. For example, for the approach from
be computed as Section III-B we can define the spatial correlation function in
G
! the 2-D case as ã(θ1 , ψ1 )H ã(θ2 , ψ2 ) and define a criterion to
1 X (2g − 1)π
Pq ≈ Ψ̂q , (20) optimize Φ via
2G g=1 2G

e(Φ, θ1 , θ2 , ψ1 , ψ2 ) = ã(θ1 , ψ1 )H ã(θ2 , ψ2 )
where Ψ̂q (τ ) = (1−j tan(τ /2))Ψq −spq (1+j tan(τ /2)) , spq
is the so-called saddle point [60], and G is a natural number −T (θ1 , θ2 , ψ1 , ψ2 )| . (23)
that determines the accuracy of the approximation.
Here T (θ1 , θ2 , ψ1 , ψ2 ) is the target SCF which could for
Proof: cf. App. C. example be chosen as T (θ1 , θ2 , ψ1 , ψ2 ) = const · δ(θ1 − θ2 ) ·
Applying Theorem 2 to (19), we finally obtain δ(ψ1 − ψ2 ). To minimize the cost function we can introduce
L G
! a Pθ by Pψ 2-D sampling grid in azimuth and elevation and
1 XX (2g − 1)π
Pd ≤ Ψ̂q (21) then align the sampled cost function into a matrix E of size
2G q=1 g=1 2G Pθ Pψ × Pθ Pψ .
The analytic expression for the false detection probability A similar extension is possible to incorporate the polariza-
can now be used to optimize the combining matrix Φ with tion of the incoming wave. We can express the dependence of
the objective to improve the DOA estimation accuracy. For the array on the polarization of the incident wave via
detection of a single source, we can formulate it as  
cos(α)
a(θ, α, φ) = [aH (θ), aV (θ)] · (24)
Φopt = arg min max CRB(Φ, θ0 ) (22) sin(α) · eφ
Φ θ0 | {z }
s. t. Pd (Φ, θ0 , ρth ) < ǫ0 , p(α,φ)

where CRB(Φ, θ0 ) is given by the expression (44) in Ap- where aH (θ) and aV (θ) represent the array response to a
pendix D and ǫ0 is the desired false detection level. The purely horizontal and a purely vertical incident plain wave,
optimization strategy in (22) aims to minimize the CRB and respectively, and the parameters α, φ describe the polarization
 7

state4 of the wave. We then can optimize Φ by minimiz- C. Adaptive design

ing the error e(θ1 , α1 , φ1 , θ2 , α2 , φ2 ) = |aH (θ1 , α1 , φ1 ) · In (7), our target is a static combining matrix that yields an
a(θ2 , α2 , φ2 )−T (θ1 , θ2 )|. Note that the target does not depend array with certain properties, such as uniform sensitivity and
on the polarization parameters since our goal is to achieve a low sidelobe level, which is a good choice if no prior knowl-
high separation in the angular domain for any polarization edge of the targeted sources is available. However, we can
(and not a separation in the polarization domain). This cost extend this approach towards an adaptive design that makes
function can be minimized by defining a multidimensional use of the fact that for a slowly changing scene, the estimates
grid which leads to a corresponding error matrix E of size from the previous snapshots provide prior information about
Pθ Pα Pϕ × Pθ Pα Pϕ . the source locations in the next snapshots. This fact can be
The same extensions are also possible to the CRB-based utilized for adaptive focusing of the array’s sensitivity towards
approach from Section III-C. For example, we can incorporate regions of interest where the targets are expected [63]. In doing
both azimuth and elevation by defining the 2-D beamformer so, the SNR and the effective resolution in these directions of
spectrum as interest can be further improved, resulting in a superior DOA
D(θ, ψ) = ã(θ, ψ)H · R · ã(θ, ψ), (25) estimation performance. To achieve this, we adopt a sequential
measurement strategy which starts with a combining matrix
and writing (19) as designed for uniform sensitivity and then gradually refine it
L̄ towards the directions of interest that have been identified in
X

Pd ≤ Prob D(θ0 , ψ0 ) − D(θq , ψq ) < 0 , (26) the observations collected so far [63].
q=1 The adaptation mechanism proceeds as follows:
1) We begin by scanning the scene with a matrix Φ de-
where L is now the total number of sidelobe peaks in the
signed according to (12) or (13), designed for a uniform
discretized 2-D correlation function. The expression for the in-

target T or according to (23) for the full angular range
dividual probabilities Pq = Prob D(θ0 , ψ0 ) − D(θq , ψq ) < 0
θ0 ∈ (0, 2π].
remains the same as in (20) where the MGF is calculated for
2) Identify regions of interest based on an estimate of
D(θ0 , ψ0 ) − D(θq , ψq ) instead of D(θ0 ) − D(θq ). The same
the angular power spectrum obtained from the initial
holds for the total false detection probability given by (21).
observation(s).
3) Define a focusing region Θ as the union of all regions
B. Arbitrary number of sources of interest.
So far, we have discussed the case of a single source’s 4) Modify Φ by solving (12) or (13) for a target designed
wave impinging on the antenna array. However, we can easily for the focusing region Θ in the SCF-based approach or
extend the CRB design presented in Section III-C to account solving (23) with a restricted angular range.
for the presence of multiple signal sources by applying in 5) As the sources are assumed to change their position
(23) the full CRB given by (42) and modifying the false gradually, track sources by repeating steps (2) to (4)
detection probability expression in (21). Particularly, assuming sequentially, moving the regions of interest along with
a correlation-based DOA estimator, we need to compute the the currently identified source locations.
probability that one strongest source is falsely detected in the 6) Every S snapshots, rescan the scene with a matrix Φ
presence of K − 1 weaker ones. In light of that, (16) becomes designed for a uniform sensitivity in order to detect
2 newly appearing sources. If new sources are found,
K−1
2 X incorporate their location into the set Θ.
D(θ0 ) = ãH
0 ỹ(t) = ãH
0 ãk sk (t) + ãH
0 ñ(t) ,
The parameter S represents a design parameter that determines
k=0
2 how quickly the system reacts to sources appearing outside
K−1
X
2 the current direction of interest. Note that this adaptation
D(θq ) = ãH
q ỹ(t) = ãH
q ãk sk (t) + ãH
q ñ(t) (27)
mechanism allows for many degrees of freedom, e.g., in terms
k=0

of the rate of adaptation of Φ or the definition of the focusing
and Pq = Prob D(θ0 ) − D(θq ) < 0 can again be calculated regions. Some results on the performance of such an adaptive
by ! design based on an example of the SCF optimized combining
G
1 X (2g − 1)π matrix can be found in [63].
Pq ≈ Ψ̂q . (28)
2G g=1 2G
D. Estimation quality
The difference between (28) and (20) is that in case of multiple
In this section, we provide an analysis of the achievable
PK−1 the vector r in (40) in Appendix C becomes equal to
sources
performance of the proposed compressive arrays for DOA
k=0 ãk sk when calculating the non-centrality parameters
estimation. In the noisy case, there are two main estimation
of the MGF.
quality measures: the achievable estimation accuracy, and the
4 Values of φ = 0 and φ = π correspond to linear and circular polarized
2 resolution capabilities.
waves, respectively. All other values of φ imply an elliptical polarization. 1) Estimation accuracy: For a fixed aperture, the achiev-
The angle α describes the orientation of the polarization plane (e.g., φ = 0,
α = 0 corresponds to a horizontal and α = π2 to a vertical polarized wave, able accuracy is mainly determined by the SNR at the output
respectively). of the antennas. For this reason, we compare the SNR of a
8

compressive array and a sparse array at the same number of Therefore, the ratio of the SNRs becomes
active channels M . ρc η2 · N
We express the output signal for a single source via (3) as = η 2 ·N
(34)
ρs σ2
+ M
σ2
1+ 12 1+ 22
ỹ(t) = Φ · a(γ1 ) · s1 (t) + ñ(t), (29) σ2 σ1

Overall, this shows that for dominating signal noise (i.e.,

where the elements of Φ are given by (6) and the covariance σ12 ≫ σ22 ) we have ρc ≈ ρs and thus there is no SNR
of ñ(t) is given by Rnn = σ12 ΦΦH + σ22 IM . The SNR of the gain from using the compressed arrays. On the other hand,
compressed array can then be computed as for dominating measurement noise (i.e., σ22 ≫ σ12 ), the SNR
N
n o ratio approaches η 2 M which means an SNR improvement q if
2
E kΦ · a(γ1 ) · s1 (t)k the efficiency of the lossy components satisfies η > N
ρc = M.
trace {Rnn } In practice, the compression ratio M N
can be quite high and
2
kΦ · a(γ1 )k · Ps therefore, the SNR improvement of the compressive arrays can
= be very significant.
trace {ΦH Φ} σ12 + M σ22
 2) Resolution: The ability to distinguish closely spaced
trace Φ · a(γ1 ) · aH (γ1 ) · ΦH · Ps
= , (30) sources is an important characteristic of an antenna array.
trace {ΦH Φ} σ12 + M σ22
The achievable resolution of the array mainly depends on its
 aperture, i.e., the largest distance between pairs of antenna
where Ps = E |s1 (t)|2 is the source power.
elements. For ULAs, the aperture is equal to (N − 1)λ/2
As evident from (30), the SNR is dependent on the param-
since the elements are spaced in half-wavelength distance
eter vector γ, i.e., on the DOA. It is therefore meaningful to
from each other. For compressive arrays as well as sparse
consider the average SNR over all possible source directions.
2 arrays, this distance can be increased further. As a result, the
This requires to compute the average of g(γ) = kΦ · a(γ)k
array’s correlation function becomes sharper, at the price of
over γ which is not possible without further assumptions
an increase in sidelobes (grating lobes). We can control the
either about the array or aboutR Φ. Let ḡ be the average R of height of the grating lobes by proper design of the antenna
g(γ) over γ, i.e., ḡ = Γ−1 g(γ)dγ with Γ = 1dγ.
−1
R H placement (in the case of sparse arrays) as well as the analog
Moreover, let us define the matrix J = Γ a(γ)·a(γ) dγ
 H combining network (in the case of compressive arrays). In
so that ḡ = trace Φ ·J · Φ . To proceed, we would like
general, we expect that at the same covered aperture and the
to replace ḡ by trace ΦΦH . We can always do so when
same number of active RF chains, the compressive arrays will
J = IM which implies that the beam patterns of all antennas
have lower sidelobes (since the many degrees of freedom in
are orthogonal over the entire parameter space. This is, e.g.,
the analog combining network allow to suppress the sidelobes
fulfilled for an ULA if it is parametrized by spatial frequencies
significantly). As it is difficult to quantify the achievable
µ = cos(θ). Furthermore, for J 6= IM one can show that
  sidelobe suppression analytically, we will focus on this aspect
EΦ {ḡ} = EΦ trace ΦΦH for any random ensemble of
in the numerical results in Section V.
Φ where  its elements are i.i.d. Note that in our case, due to
(6), trace ΦΦH is not random but deterministic. Hence, the
expectation on the right-hand side is not needed. In light of V. N UMERICAL RESULTS

this assumption, we can express the average SNR ρ̄c (averaged In this section, we evaluate the performance of the compres-
over γ) as sive array with optimized combining network and compare it
 to its closest counterparts in terms of the aperture and hardware
trace ΦΦH · Ps kΦk2F · Ps complexity, namely random and sparse arrays. We perform
ρ̄c = = (31)
trace {ΦH Φ} σ12 + M σ22 2
kΦkF σ12 + M σ22 the numerical study based on a uniform circular array (UCA)
with N = 9 elements that are compressed to M = 5 receiver
 2
2
√η
channels (this amounts to ≈ 1.8 times reduction in the number
Using (6) it is easy to see that kΦkF = L
·N ·L = η 2 ·M .
of receiver channels). Note that for an UCA with isotropic
Therefore, the average SNR becomes
elements the response of the nth antenna element as a function
η 2 · N · Ps PS η2 · N of the azimuth angle θ can be written as
ρ̄c = = 2 · (32)
η2 · N σ12 + M σ22 σ1 + σ22 η 2 ·N
σ2
+ M
σ2
an (θ) = e2πR̃ cos(θ−ϑn ) , (35)
1+ 12 1+ 22
σ2 σ1
where ϑn = 2π(n − 1)/N with n = 1, 2, · · · , N and R̃ = R λ
To compare this SNR to the one that can be achieved with a is the array radius normalized to the wavelength. For both
sparse array we model the observed signal as as (γ1 ) · s1 (t) + proposed designs, the radius R̃ was fixed and set to 0.65.
ws (t), where, to make the comparison fair, the elements of The combining matrix Φ is chosen according to [Φ](m,n) =
ws (t) are i.i.d. with variance σ12 + σ22 . We then obtain for the eϕm,n , where ϕm,n are the optimization variables in the
SNR of a sparse array proposed approaches. To find an optimized design Φopt we
solve the weighted optimization problems (13) and (23) via
M · Ps Ps M ATLAB’s numerical optimization features. Since run-time is
ρ̄s = = 2 . (33)
M (σ12 + σ22 ) σ1 + σ22 not a concern for an off-line design, and in order to avoid local
 9

1 1
Random Random
0.9 Opt SCF 0.9 Opt SCF
Opt CRB Opt CRB
0.8 0.8

0.7 0.7

0.6 0.6
CCDF

CCDF
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9
CRB [rad2 ] Sidelobe Level

Figure 2. Comparison of the CRBs of the optimized compressive arrays Figure 3. Comparison of the sidelobe levels of the optimized linear
versus the random ones. The CCDF of the CRB for 5000 random realizations combining network versus the random ones. The CCDF of the mean sidelobe
are shown together with that of the optimized kernel. levels for 5000 random realizations are shown together with that of the
optimized kernel.

minima, we run fmincon and fminimax to solve (13) and

10 0
(23), respectively, with 100 random initializations and pick Rand
UCA N = 5
the solution with the smallest value of the cost function. In Opt SCF
the following, we refer to the design obtained by the SCF Opt CRB
10 -1 UCA N = 9
optimization approach from (13) as Opt SCF, whereas the Threshold
design obtained as a result of the CRB minimization from
(23) is referred to as Opt CRB. For the Opt SCF approach,
10 -2
Pd

we set T = AH ·A as a target which is the correlation function

we would achieve with an M -element (uncompressed) UCA.
For the Opt CRB approach, the threshold false detection 10 -3
probability is set to 0.05 to be achieved at an SNR of 0dB.

A. Performance analysis for a single source 10 -4

-4 -2 0 2 4 6 8 10 12 14
We begin by examining the performance of the optimized SNR [dB]
compressive arrays with respect to the attainable CRB and the
sidelobe level in the case of a single source as discussed in Figure 4. False detection probability of the uncompressed UCAs, compres-
sive arrays, and the random ones.
Sections III-B and III-C.
1) Comparison with random arrays: Figure 2 shows the
achievable CRB of the compressive arrays with the optimized has a lower sidelobe level at a specific CRB. This is confirmed
combining network and the random ones that have the same by the corresponding (analytic) probabilities of false detection
number of antennas and sampling channels while ϕm,n are depicted in Figure 4. For comparison Figure 4 also presents
drawn uniformly at random from (0, 2π]. At a fixed SNR the results for the uncompressed UCAs with N = 9 and
level of 0 dB, an estimate of the Complementary Cumulative N = 5 (i.e., with smaller aperture size) as well as the average
Distribution Function (CCDF) of the CRB obtained from 5000 Pd for the random arrays. As can be seen, both proposed
random realizations of Φ is shown. The optimized networks approaches achieve lower probability of false detection Pd
for both approaches have been designed to achieve the same than the uncompressed UCA with a lower number of antennas,
CRB. As evident from the figure, the CRB of the optimized whereas the sparse arrays on average are significantly inferior
compressive arrays is almost in every case lower than that to all the rest. It is worth nothing that for both proposed
of the random ones. In other words, the random kernel can optimization approaches, the CRB and the sidelobe level (and
potentially provide a performance comparable to the optimized hence the probability of false detection) can be controlled. In
ones but with a very low probability. The CCDFs of the the SCF-based design this can be done by a proper choice
average sidelobe levels for the same scenario are shown in of the weighting matrix in (13), whereas in the CRB-based
Figure 3. We observe that the random compressive arrays approach the false detection probability is chosen directly.
provide significantly higher sidelobe levels compared to both 2) Comparison with sparse arrays: Now we compare the
the SCF and the CRB-optimized ones. This supports the compressive array to a sparse array that has the same number
intuition that designing the combining matrix randomly results of receiver chains, i.e., for the considered scenario it means
in sub-optimal performance. that for a sparse array N = M = 5. According to [17], we
Comparing the sidelobe level of the SCF and CRB- design the sparse array such that the positions of its elements
optimized compressive arrays, we can notice that the latter are optimized towards obtaining a uniform sensitivity and
10

1 0.25 1
Sparse CRB
0.9 Opt SCF Sidelobe Level
Opt CRB
0.8 0.2 0.9
Spatial Correlation Function

0.7

Sidelobe Level
0.6 0.15 0.8

CRB
0.5

0.4 0.1 0.7

0.3

0.2 0.05 0.6

0.1

0 0 0.5
0 1 2 3 4 5 6 5 10 15 20 25 30
Azimuth DOA θ [rad] Number of Antennas (N)

Figure 5. The spatial correlation function at a specific DOA of an optimized Figure 7. The CRB and the sidelobe level for a compressive array with
sparse array and a compressive array with the same number of receiver chains different number of antenna elements and fixed number of channels
and an optimized combining network
optimized compressive array as a function of the number
0.7 of antennas N . It is clear that the compressive arrays not
only allow to control the CRB and the sidelobe level via an
0.65
optimization of the combining network, but also by adding
more antenna elements while the number of receiver channels
is kept fixed. In Figure 7, we can see that the CRB can be
Sidelobe Level

0.6
improved significantly when the number of antenna elements
is increased at the price of higher sidelobe levels. However the
0.55 network can then be re-optimized for the new scenario (e.g.,
with the OPT CRB approach), aiming at a better suppression
0.5
of the sidelobes while a certain level of CRB improvement is
Sparse
Opt SCF
maintained.
Opt CRB
0.45
0.08 0.085 0.09 0.095 0.1 0.105 0.11 0.115 0.12 B. Performance analysis for multiple sources
CRB [rad2 ]
Now we examine the performance of the proposed opti-
Figure 6. The CRB versus the sidelobe level for a compressive array with mized compressive array in the case of two sources impinging
an optimized combining network and a sparse array (single diamond marker) on the array from different DOAs. The power ratio between
the two sources is set to α = |s2 /s1 |2 = −6 dB, while their
DOAs are d radians apart, i.e., θ1 = θ, θ2 = θ + d where θ
desired CRB. Figure 5 shows the spatial correlation functions scans the whole angular space. The two sources are inphase.
at a specific DOA for the sparse array and an SCF-optimized Similar to Figure 2, Figure 8 shows the achievable CRB
compressive array (compression from N = 9 elements to of the strongest path using the compressive arrays with the
M = 5 receiver chains) that achieves the same CRB (0.113) optimized combining network and the random ones that have
at the fixed SNR level of 0 dB. It can be noted that the the same number of antennas and sampling channels for an
sidelobe level of the sparse array is relatively high compared to SNR level (with respect to the strongest source) of 12 dB. The
that of the compressive one (especially the Opt CRB design). CCDF of the CRB obtained from 5000 random realizations of
Particularly, the mean level of the sidelobes for the optimized Φ is shown for d = 0.2 and d = 0.4 radians. As discussed
compressive array is 0.53 compared to 0.68 of the sparse array. earlier, the Opt SCF and the Opt CRB designs both provide
Figure 6 shows the resulting trade-off between providing the same CRB and so only the Opt SCF is shown for clarity.
a good CRB and maintaining a low sidelobe level. One can The CCDF shows that the CRB of the optimized compressive
see that the compressive array (with either of the optimization arrays is again almost in every case lower than that of the
schemes) outperforms the sparse one and gives more degrees random ones. As the sources get closer (e.g., d = 0.2), the
of freedom to tune the array design with respect to some need for the optimized network increases as the probability to
desired properties (e.g., targeted CRB or sidelobe level). This achieve acceptable properties (e.g., low CRB) by the random
figure also confirms once again that the Opt CRB approach design gets lower.
outperforms the Opt SCF approach as it gives more control It has been proposed in Section IV-B, to extend the design
over the sidelobe level for a specific CRB. based on CRB for the case of multiple sources. Considering
The superiority of the compressive arrays over the sparse the same set-up with two sources, and for a specific DOA of
ones with respect to adaptability is further highlighted in the first source, to do so the sidelobes in the correlation func-
Figure 7. It presents the CRB and the sidelobe level of an tion have to be searched for all possible DOAs of the second
 11

1
to traditional array designs. We present an architecture of such
Random, d = 0.2 [rad] a compressive array and introduce a generic system model that
0.9 Opt SCF, d = 0.2 [rad]
Random, d = 0.4 [rad] includes different options for the hardware implementation.
0.8
Opt SCF, d = 0.4 [rad] We then focus on the choice of the coefficients in the analog
0.7
combining network. Instead of choosing them randomly, as
0.6 advocated by earlier work in this area, we propose a generic
CCDF

0.5 design approach for the analog combining network with the
0.4
goal to obtain an array with certain desired properties, e.g.,
uniform sensitivity, low cross-correlation, or low variance in
0.3
the DoA estimates. We exemplify the array design via two
0.2
concrete examples. Our numerical simulations demonstrate the
0.1 superiority of the proposed optimized compressive arrays to
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
compressive arrays with randomly chosen combining kernels,
CRB [rad2 ] as the latter result in very high sidelobes (which imply a higher
probability of false detection) as well as higher CRBs. We
Figure 8. Comparison of the CRBs of the optimized compressive arrays also compare our optimized compressive array to a sparse
versus the random ones for the case of two impinging sources. The CCDF array of the same complexity (i.e., same number of receiver
of the CRB for 5000 random realizations are shown together with that of the
optimized kernel. channels M ) and find that sparse arrays suffer from much
higher sidelobes at the same CRB level. Also our proposed
10 0 compressive array enjoys a high degree of adaptability since
the combining weights can be altered to adjust the array to the
current requirements, which is impossible for sparse arrays due
10 -1 to their static nature.

ACKNOWLEDGMENT
10 -2
Pd

This work was partially supported by the Deutsche

Forschungsgemeinschaft (DFG) projects CLASS (grant MA
Rand
UCA N = 5 1184/23-1) and CoSMoS (grant GA 2062/2-1) and the Carl-
-3
10 Opt SCF Zeiss Foundation under the postdoctoral scholarship project
Opt CRB
UCA N = 9 “EMBiCoS”.
Threshold
10 -4
0 2 4 6 8 10 12 14 16 18 20 A PPENDIX A
SNR [dB]
P ROOF OF T HEOREM 1 IN S ECTION III-B
Figure 9. False detection probability of the uncompressed UCAs, compres- To prove the theorem we use the fact that for a uni-
sive arrays, and the random ones with two source signals with power ratio tary matrix U and an arbitrary square matrix X we have
α = −6dB.
kX · U kF = kU · XkF = kXkF . Since A satisfies A·AH =
(M−N )×N
N · IM we √ can find a matrix Ā ∈ C such that
.
source. This leads to a very high computational complexity. V = 1/ N · [AT , ĀT ]T ∈ CN ×N is a unitary matrix.
√ T
The same search strategy has to be performed for all possible Therefore, we have V · AH = N · IM , 0M×N −M . The
DOAs of the first source. Therefore, for simplicity, we fix the cost function (12) can then be rewritten as
second source DOA and perform the optimization similar to 2 2
that of the single source case. Figure 9 shows the probabilities kEkF = V · E · V H F
of false detection Pd for two sources. It can be seen that  √  2
N IM √ 
the design based on CRB shows superior performance in = ΦH Φ N IM , 0M×N −M − V T V H
0N −M×M F
terms of lower false detection probability compared to that     2
H
of the uncompressed (N = 5)-element UCA, compressive NΦ · Φ 0M×N −M A  
= −N · T AH , ĀH
array with SCF based designed network, and the averaged 0N −M×M 0N −M×N −M Ā F
random ones. Although the SCF based design can not be  H  2
Φ · Φ − A · T · AH −A · T · ĀH
re-optimized for the multiple source case, it still provides a = N·
−Ā · T · AH −Ā · T · ĀH F
significantly lower probability of false detection compared to 2
the random arrays and is comparable to a non-compressed = N 2 · ΦH · Φ − S F
+ const, (36)
(N = 5)-element UCA. using the short-hand notation S = A · T · AH . Equation (36)
demonstrates that the optimization problem is equivalent to
VI. C ONCLUSIONS finding the best approximation of the matrix S by the matrix
In this paper we consider the design of compressive antenna ΦH · Φ. Since Φ is an m × M matrix, the rank of the M × M
arrays for direction of arrival (DOA) that aim to provide a matrix ΦH · Φ is less than or equal to m < M . Therefore,
larger aperture with a reduced hardware complexity compared (36) represents a low-rank approximation problem. According
12

to the Eckart-Young theorem, its optimal solution is given by where r = ã0 s0 .

truncating the M − m smallest eigenvalues of S. The PDF of Xq can now be obtained by computing the
inverse Laplace transform of ΨXq (s), whereas the probability
A PPENDIX B of false detection is then obtained by integrating the resulting
P ROOF OF C OROLLARY 1 IN S ECTION III-B PDF. In order to compute the integral we apply an iterative
The sampled version of (9) is given by a scaled identity approach for numerical integration from [68] that utilizes
matrix, i.e., T = C · IN . Since A is row-orthogonal it follows the saddle point technique from [60]. In this technique, the
that S = A · T · AH = C · N · IM . As all eigenvalues of S are integration path is chosen such that is passes through the
equal to C · N , its eigenvalue decomposition can be written as saddle point of the integrand on the real axis [65]. Since
S = U · (C · N · IM ) · U H , where U ∈ CM×M is an arbitrary the integrand is convex, a single saddle point s = sp exists
unitary matrix. Truncating the M − m “smallest” eigenvalues, in Re {s} > 0. Furthermore, it can be easily computed by
we obtain Sm = C · N · Um · Um H
, where Um ∈ CM×m Newton search method as sp ← sp − (Ψ′ (sp )/Ψ′′ (sp )), where
contains the first m columns of U . Invoking Theorem 1, we Ψ′ (sp ), Ψ′′ (sp ) are the first and second order derivatives of
have ΦH H
opt Φopt = C · N · Um · Um and therefore Φopt is a
Ψ(s) = ln(ΨXq (s)/s) evaluated at the saddle point sp [68].
H Therefore, using the Gauss-Chebyshev quadrature [68], the
scaled version of Um , which proves the claim.
probability of false detection can be obtained as
A PPENDIX C G
!
P ROOF OF T HEOREM 2 IN S ECTION III-C 1 X (2g − 1)π
Pq ≈ Ψ̂ (41)
2G g=1 2G
Denote by Xq (t) = D(θ0 ) − D(θq ), where D(θ) =
2 2
ã ỹ(t) and D(θq ) = ãH
H
q ỹ(t) . Then, Xq (t) is a random where Ψ̂(τ ) = (1 − j tan(τ /2))ΨXq − sp (1 + j tan(τ /2))


variable that we can write as and G is the number of steps that determines the accuracy of
2 2
Xq (t) = ãH
0 ỹ(t) − ãH
q ỹ(t)
the integration.
= ỹ H (ã0 ãH H H
0 − ãq ãq )ỹ = ỹ D ỹ (37)
A PPENDIX D
where ỹ = ã0 s(t) + ñ(t) is a complex random vector with a CRB FOR COMPRESSED ARRAY
non-zero mean and a covariance matrix Rnn . Assuming that In this section, we present results of the CRB derived in [2]
v(t) and w(t) are white with elements that have variance σ12 for the receiver model shown in Fig. 1 where the noise vectors
and σ22 , ỹ becomes a complex-Gaussian random vector with v(t) and w(t) are assumed to be white with covariances
mean equal to ã0 s0 and covariance Rnn = σ12 ΦΦH + σ22 I. Rvv = σ12 IN and Rww = σ22 IM , respectively. The associated
This said, Xq (t) is a chi-square random variable that, due to CRB matrix is then found to be
the structure5 of D and the fact that ỹ is non-zero mean, has 
a so-called non-central indefinite quadratic form [64]. CRB(Φ, θ) = σ12 2Re F ⊙ RT s )
−1
, (42)
In order to compute the probability Prob(Xq (t) < 0), we where ⊙ denotes Shur (element-wise) matrix product, Rs =
need to derive the distribution of Xq (t). Since the covariance s(t)sH (t) is the signal covariance matrix, and F is a matrix
matrix of ỹ is colored, it is convenient to express Xq (t) as that depends on the array beampattern and the combining
H matrix Φ as
Xq (t) = ỹw B ỹw , (38)
−1/2 F = D H ΦH ZΦD. (43)
where ỹw = Rnn ỹ contains pre-whitened observations   −1 
1/2 1/2
whose covariance is an identity matrix and B = Rnn DRnn . In (43), Z = Q IN − Ã ÃH QÃ ÃH Q ,
This, way the quadratic form (37) is reduced to a diagonal  
D = ∂a(θ0 )/∂θ0 , ∂a(θ1 )/∂θ1 , ..., ∂a(θK−1 )/∂θK−1 , and
form in independent random variables with unit variance [65]. σ 2
Q = (ΦΦH + βIN )−1 where β = σ22 . Since we consider
By representing B via its eigen value decomposition (EVD) 1

as B = Un λUnH with Un and λ being the unitary matrix only a single source for solving the optimization problem
and the diagonal matrix consisting of eigenvalues, respectively, in (23), F reduces to a scalar F and Rs to Rss = ks(t)k2 .
then we can compute the moment generating function (MGF) Therefore, for a single source (42) becomes
 −1 1
shown in [66], [67] as
P  CRB(Φ, θ0 ) = σ12 2F Rss = , (44)
R µ2r λr s 2F ρ
exp r=1 1−λr s Rss
ΨXq (s) = QR . (39) where ρ = σ12
is the input SNR.
r=1 (1 − λr s)
Here, λr is the rth eigenvalue of B, R denotes the rank D, and R EFERENCES
µr is the rth element of the vector of non-centrality parameters
[1] H. Krim and M. Viberg. Two decades of array signal processing
µ defined as research: the parametric approach. IEEE Signal Processing Magazine,
13(4):67–94, Jul 1996.
µ = UnH Rn−1/2 r, (40) [2] H. L. Van Trees. Detection, estimation, and modulation theory. Part IV.,
Optimum array processing. Wiley-Interscience, New York, 2002.
5 Note that since the matrix D in (37) results from the subtraction of the [3] S. Valaee, B. Champagne, and P. Kabal. Parametric localization of dis-
outer products of two vectors ã0 and ãq , it is a complex symmetric matrix tributed sources. IEEE Transactions on Signal Processing, 43(9):2144–
that whose maximum rank is 2 . 2153, Sep 1995.
 13

[4] J. C. Chen, K. Yao, and R. E. Hudson. Source localization and [29] P. Stoica, P. Babu, and J. Li. Spice: A sparse covariance-based estimation
beamforming. IEEE Signal Processing Magazine, 19(2):30–39, Mar method for array processing. IEEE Transactions on Signal Processing,
2002. 59(2):629–638, Feb 2011.
[5] A. Richter, D. Hampicke, G. Sommerkorn, and R. S. Thomä. Joint [30] D. Model and M. Zibulevsky. Signal reconstruction in sensor arrays
estimation of dod, time-delay, and doa for high-resolution channel using sparse representations. Signal Processing, 86(3):624 – 638,
sounding. In IEE Vehicular Technology Conference, volume 2, pages 2006. Sparse Approximations in Signal and Image ProcessingSparse
1045–1049 vol.2, 2000. Approximations in Signal and Image Processing.
[6] R. S. Thomä, M. Landmann, and A. Richter. Rimax-a maximum [31] K. Han, Y. Wang, B. Kou, and W. Hong. Parameters estimation using a
likelihood framework channel parameter estimation in multidimensional random linear array and compressed sensing. In International Congress
channel sounding. International Symposium on Antennas and Propaga- on Image and Signal Processing (CISP), volume 8, pages 3950–3954,
tion, pages 53–56, 2004. Oct 2010.
[7] W. D. Blair and M. B. Pearce. Monopulse doa estimation of two [32] M. Rossi, A. M. Haimovich, and Y. C. Eldar. Spatial compressive
unresolved rayleigh targets. IEEE Transactions on Aerospace and sensing in mimo radar with random arrays. In Annual Conference on
Electronic Systems, 37(2):452–469, Apr 2001. Information Sciences and Systems (CISS), pages 1–6, March 2012.
[8] M. Landmann and R. S. Thomä. Common pitfalls in multidimensional [33] S. Shakeri, D. D. Ariananda, and G. Leus. Direction of arrival estimation
high resolution channel parameter estimation. In IEEE Digital Signal using sparse ruler array design. In International Workshop on Signal
Processing Workshop and 5th IEEE Signal Processing Education Work- Processing Advances in Wireless Communications (SPAWC), pages 525–
shop, pages 314–319, Jan 2009. 529, June 2012.
[9] R. Roy and T. Kailath. Esprit-estimation of signal parameters via [34] M. B. Hawes and W. Liu. Compressive sensing-based approach to
rotational invariance techniques. IEEE Transactions on Acoustics, the design of linear robust sparse antenna arrays with physical size
Speech, and Signal Processing, 37(7):984–995, Jul 1989. constraint. IET Microwaves, Antennas Propagation, 8(10):736–746, July
[10] R. Schmidt. Multiple emitter location and signal parameter estimation. 2014.
IEEE Transactions on Antennas and Propagation, 34(3):276–280, Mar [35] Y. Wang, G. Leus, and A. Pandharipande. Direction estimation using
1986. compressive sampling array processing. In 2009 IEEE/SP 15th Work-
[11] P. Stoica, B. Ottersten, M. Viberg, and R. L. Moses. Maximum shop on Statistical Signal Processing, pages 626–629. IEEE, 2009.
likelihood array processing for stochastic coherent sources. IEEE [36] J. F. Gu, W. P. Zhu, and M. N. S. Swamy. Compressed sensing for
Transactions on Signal Processing, 44(1):96–105, Jan 1996. doa estimation with fewer receivers than sensors. In IEEE International
[12] Y. Lo. A mathematical theory of antenna arrays with randomly spaced Symposium of Circuits and Systems (ISCAS), pages 1752–1755, May
elements. IEEE Transactions on Antennas and Propagation, 12(3):257– 2011.
268, May 1964. [37] R. Baraniuk, M. Davenport, R DeVore, and M Wakin. A simple proof
[13] A. Moffet. Minimum-redundancy linear arrays. IEEE Transactions on of the restricted isometry property for random matrices. Constructive
Antennas and Propagation, 16(2):172–175, Mar 1968. Approximation, 28(3):253–263, 2008.
[14] W. K. Ma, T. H. Hsieh, and C. Y. Chi. Doa estimation of quasi- [38] T T. Cai, T. Jiang, et al. Limiting laws of coherence of random matrices
stationary signals via khatri-rao subspace. In International Conference with applications to testing covariance structure and construction of
on Acoustics, Speech and Signal Processing, pages 2165–2168, April compressed sensing matrices. The Annals of Statistics, 39(3):1496–1525,
2009. 2011.
[15] R. L. Haupt. Thinned arrays using genetic algorithms. IEEE Transac- [39] Y. C Eldar and G. Kutyniok. Compressed sensing: theory and applica-
tions on Antennas and Propagation, 42(7):993–999, July 1994. tions. Cambridge University Press, 2012.
[16] A. Trucco and V. Murino. Stochastic optimization of linear sparse arrays. [40] M. Ibrahim, F. Roemer, and G. Del Galdo. On the design of the
IEEE Journal of Oceanic Engineering, 24(3):291–299, Jul 1999. measurement matrix for compressed sensing based doa estimation. In
[17] H. Gazzah and K. Abed-Meraim. Optimum ambiguity-free directional International Conference on Acoustics, Speech and Signal Processing
and omnidirectional planar antenna arrays for doa estimation. IEEE (ICASSP), pages 3631–3635, April 2015.
Transactions on Signal Processing, 57(10):3942–3953, Oct 2009. [41] TriQuint Semiconductor. Phase shifters, August 2016.
[18] P. P. Vaidyanathan and P. Pal. Sparse sensing with co-prime samplers http://www.triquint.com/products/all/control-products/phase-shifters.
and arrays. IEEE Transactions on Signal Processing, 59(2):573–586, [42] H. B. Lee and M. S. Wengrovitz. Resolution threshold of beamspace
Feb 2011. music for two closely spaced emitters. IEEE Transactions on Acoustics,
[19] P. Pal and P. P. Vaidyanathan. Nested arrays: A novel approach to array Speech, and Signal Processing, 38(9):1545–1559, Sep 1990.
processing with enhanced degrees of freedom. IEEE Transactions on [43] S. Anderson. Optimal dimension reduction for sensor array signal
Signal Processing, 58(8):4167–4181, Aug 2010. processing. In Asilomar Conference on Signals, Systems and Computers,
[20] E. J. Candés and T. Tao. Near optimal signal recovery from random pages 918–922 vol.2, Nov 1991.
projections: universal encoding strategies. IEEE Transactions on Infor- [44] M. D. Zoltowski, G. M. Kautz, and S. D. Silverstein. Beamspace root-
mation Theory, 52:5406–5425, 2006. music. IEEE Transactions on Signal Processing, 41(1):344–, Jan 1993.
[21] E. J. Candés, J. Romberg, and T. Tao. Robust uncertainty principles: ex- [45] Guanghan Xu, S. D. Silverstein, R. H. Roy, and T. Kailath. Beamspace
act signal reconstruction from highly incomplete frequency information. esprit. IEEE Transactions on Signal Processing, 42(2):349–356, Feb
IEEE Transactions on Information Theory, 52(2):489–509, Feb 2006. 1994.
[22] D. L. Donoho. Compressed sensing. IEEE Transactions on Information [46] M. D. Zoltowski, M. Haardt, and C. P. Mathews. Closed-form 2-d angle
Theory, 52(4):1289–1306, 2006. estimation with rectangular arrays in element space or beamspace via
[23] V. Cevher, A. C. Gurbuz, J. H. McClellan, and R. Chellappa. Compres- unitary esprit. IEEE Transactions on Signal Processing, 44(2):316–328,
sive wireless arrays for bearing estimation. In International Conference Feb 1996.
on Acoustics, Speech and Signal Processing, pages 2497–2500, March [47] A. B. Gershman. Direction finding using beamspace root estimator
2008. banks. IEEE Transactions on Signal Processing, 46(11):3131–3135,
[24] C. Feng, S. Valaee, and Z. Tan. Multiple target localization using Nov 1998.
compressive sensing. In Global Telecommunications Conference, pages [48] A. Hassanien, S. A. Elkader, A. B. Gershman, and K. M. Wong. Convex
1–6, Nov 2009. optimization based beam-space preprocessing with improved robustness
[25] J. H. G. Ender. On compressive sensing applied to radar. Signal against out-of-sector sources. IEEE Transactions on Signal Processing,
Processing, 90(5):1402 – 1414, 2010. Special Section on Statistical 54(5):1587–1595, May 2006.
Signal Array Processing. [49] H. Hung and M. Kaveh. Focussing matrices for coherent signal-
[26] D. Malioutov, M. Cetin, and A. S. Willsky. A sparse signal recon- subspace processing. IEEE Transactions on Acoustics, Speech, and
struction perspective for source localization with sensor arrays. IEEE Signal Processing, 36(8):1272–1281, Aug 1988.
Transactions on Signal Processing, 53(8):3010–3022, Aug 2005. [50] V. Venkateswaran and A. J. van der Veen. Analog beamforming in
[27] A. C. Gurbuz, V. Cevher, and J. H. Mcclellan. Bearing estimation mimo communications with phase shift networks and online channel
via spatial sparsity using compressive sensing. IEEE Transactions on estimation. IEEE Transactions on Signal Processing, 58(8):4131–4143,
Aerospace and Electronic Systems, 48(2):1358–1369, APRIL 2012. Aug 2010.
[28] A. Gretsistas and M. D. Plumbley. A Multichannel Spatial Compressed [51] X. Huang, Y. J. Guo, and J. D. Bunton. A hybrid adaptive antenna array.
Sensing Approach for Direction of Arrival Estimation, pages 458–465. IEEE Transactions on Wireless Communications, 9(5):1770–1779, May
Springer Berlin Heidelberg, Berlin, Heidelberg, 2010. 2010.
14

[52] J. Nsenga, A. Bourdoux, and F. Horlin. Mixed analog/digital beamform-

ing for 60 ghz mimo frequency selective channels. In IEEE International
Conference on Communications, pages 1–6, May 2010.
[53] L. C. Godara. Smart Antennas. CRC Press LLC, 2004.
[54] T. E. Bogale and L. B. Le. Beamforming for multiuser massive
mimo systems: Digital versus hybrid analog-digital. In IEEE Global
Communications Conference, pages 4066–4071, Dec 2014.
[55] Z. Pi and F. Khan. An introduction to millimeter-wave mobile broadband
systems. IEEE Communications Magazine, 49(6):101–107, June 2011.
[56] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath.
Spatially sparse precoding in millimeter wave mimo systems. IEEE
Transactions on Wireless Communications, 13(3):1499–1513, March
2014.
[57] W. Roh, J. Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, and
F. Aryanfar. Millimeter-wave beamforming as an enabling technology
for 5g cellular communications: theoretical feasibility and prototype
results. IEEE Communications Magazine, 52(2):106–113, February
2014.
[58] F. Athley. Threshold region performance of maximum likelihood
direction of arrival estimators. IEEE Transactions on Signal Processing,
53(4):1359–1373, 2005.
[59] E. Kreyszig. Advanced engineering mathematics. John Wiley and Sons,
New York, 2005.
[60] C. Helstrom. Calculating error probabilities for intersymbol and cochan-
nel interference. IEEE Transactions on Communications, 34(5):430–435,
1986.
[61] T. Coleman and Y. Li. On the convergence of reflective newton methods
for large-scale nonlinear minimization subject to bounds. Mathematical
Programming, 67(2):189–224, 1994.
[62] T. Coleman and Y. Li. An interior trust region approach for nonlin-
ear minimization subject to bounds. SIAM Journal on optimization,
6(2):418–445, 1996.
[63] M. Ibrahim, F. Römer, and G. Del Galdo. An adaptively focusing
measurement design for compressed sensing based doa estimation. In
Signal Processing Conference (EUSIPCO), 2015 23rd European, pages
859–863. IEEE, 2015.
[64] A. M Mathai and S. B Provost. Quadratic forms in random variables:
theory and applications. M. Dekker New York, 1992.
[65] F. Athley. Threshold region performance of deterministic maximum
likelihood doa estimation of multiple sources. In Signals, Systems
and Computers, 2002. Conference Record of the Thirty-Sixth Asilomar
Conference on, volume 2, pages 1283–1287. IEEE, 2002.
[66] G. L. Turin. The characteristic function of hermitian quadratic forms in
complex normal variables. Biometrika, 47(1/2):199–201, 1960.
[67] D. Raphaeli. Distribution of noncentral indefinite quadratic forms in
complex normal variables. IEEE transactions on Information Theory,
42(3):1002–1007, 1996.
[68] Y. Ma, T. L. Lim, and S. Pasupathy. Error probability for coherent
and differential psk over arbitrary rician fading channels with mul-
tiple cochannel interferers. IEEE Transactions on Communications,
50(3):429–441, 2002.