Compressive Sensing in Electromagnetics A Review PDF

This article has been accepted for inclusion in a future issue of this journal.
Content is final as presented, with the exception of pagination.
Compressive Sensing in
ElectromagneticsVA Review
Andrea Massa, Paolo Rocca, and Giacomo Oliveri
ELEDIA Research Center, Department of Information Engineering and Computer Science,
University of Trento, 38123 Trento, Italy,
E-mails: andrea.massa@unitn.it; paolo.rocca@disi.unitn.it; giacomo.oliveri@disi.unitn.it
Abstract
Several problems arising in electromagnetics can be directly formulated or suitably recast for an effective solution
within the compressive sensing (CS) framework. This has motivated a great interest in developing and applying CS
methodologies to several conventional and innovative electromagnetic scenarios. This work is aimed at presenting,
to the best of the authors knowledge, a review of the state-of-the-art and most recent advances of CS formulations
and related methods as applied to electromagnetics. Toward this end, a wide set of applicative scenarios comprising
the diagnosis and synthesis of antenna arrays, the estimation of directions of arrival, and the solution of inverse
scattering and radar imaging problems are reviewed. Current challenges and trends in the application of CS to the
solution of traditional and new electromagnetic problems are also discussed.
Keywords: Antenna and array diagnosis; antenna arrays; compressive sensing (CS); direction-of-arrival (DoA) estimation;
inverse problems; radar imaging; sparse problems
1. Introduction and Scenario
he Compressive Sensing (CS) paradigm has enabled the

development of completely new approaches in several
research areas of signal processing, information theory, computer science, and electrical engineering [1, 2]. The promise to
overcome the common wisdom in data acquisition based on
Shannons celebrated theorem [1] and to allow one to recover
(certain) signals/phenomena from far fewer measurements than
traditional techniques has attracted considerable interest (see
Figure 1). Therefore, its study/application to those systems
usually severely constrained by the Nyquists sampling rate [1, 2]
(e.g., imaging [3, 4], audio/video capture [5], and communications [6]) has been strongly motivated.
CS-based techniques build upon the assumption that many
physical quantities are intrinsically or extrinsically sparse and
they can be represented by few nonzero expansion coefcients,
with respect to suitable expansion bases [1]. Indeed, CS approaches essentially look for an approximate solution x of a
linear system y Ax, while requiring that x has the minimum
number of nonzero entries [see (8)]. If the acquisition process
(i.e., the input/output transformation matrix A) ts suitable working conditions, a high-dimensional signal can be exactly retrieved
from a small set of measurements through efcient deterministic/
Bayesian search strategies [1, 712].
IEEE Antennas and Propagation Magazine, Vol. 57, No. 1, February 2015
In addition to their fundamental advantages over Shannons

theorem-based methodologies, the success of CS formulations is
also related to their intrinsic exibility and generality [1, 11, 12],
the effectiveness and the numerical efciency of the corresponding retrieval techniques [10], and the wide availability of
software (SW) libraries implementing state-of-the-art CS algorithms [1315] for effectively dealing with complex engineering
problems [10].
In this framework, Electromagnetics represents a relatively
new eld of application. While CS methods have been early
adopted in some specic electromagnetic-related applicative domains such as radar imaging [16], their exploitation has been
initially limited to those recovery problems that naturally t the
standard CS requirements (i.e., linearity and unknowns sparseness [1, 16]). However, starting from the consideration that
several conventional electromagnetic problems can be properly
reformulated (e.g., using suitable approximations or if some a
priori knowledge is available [1720]) to still lie within the set of
the CS-tractable ones, CS has been recently extended to applications beyond intrinsically CS-compliant problems, and interesting
results have been obtained, ranging from antenna array diagnosis
and synthesis [2123] or direction-of-arrival (DoA) estimation
[18, 20] until inverse scattering and microwave imaging [2426].
The aim of this work is then that of reviewing, to the best
of the authors knowledge, the state-of-the-art and the more
1045-9243/15/$26.00 2015 IEEE
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
with Lfg being the linear observation operator. Let us assume that the signal yr is acquired through the following
measurement operation:
ym h yr; m ri m 1; . . . ; M
(3)
where h; i is the inner product, and fm r; m 1; . . . ; Mg

are the M sensing waveforms (e.g., Dirac delta functions) being
M N:
(4)
By virtue of the linearity of the operators at hand, one can substitute (2) and (1) in (3) to yield the following relationships:
Figure 1. Number of CS-related papers published each year

(based on IEEE Xplore databases).
recent advances of CS techniques, as applied to electromagnetics, focusing on three main elds of research, i.e., antenna arrays, inverse scattering, and radar imaging. The most widely
adopted solution strategies, current challenges, and limitations
will be discussed by also envisaging future trends of the CS research topic.
Toward this end, the paper is organized as follows: After
a short resume of CS problem statements and fundamental
theorems (see Section 2), the CS reconstruction strategies usually
adopted in electromagnetics are surveyed in Section 3. Section 4
presents a review of the applications of CS in the electromagnetics framework starting from the diagnosis and the synthesis
of antenna arrays (see Section 4.1) and the DoA estimation (see
Section 4.2) to the solution of inverse scattering (see Section 4.3)
and radar imaging (see Section 4.4) problems. Some conclusions are eventually drawn (see Section 5).
2.1 Definition of the Signal Model

Let us consider a real-valued signal f r [without loss
of generality, space-dependent signals will be assumed hereinafter. However, the same formulation applies to other scenarios, by replacing r with the suitable variable (e.g., time,
frequency, wavenumber, etc.)] represented in a suitable expansion basis n r by means of an N -dimensional vector
f ffn 2 R; n 1; . . . ; N g complying with
N
X
fn n r
n1
N
X
fn Lfn rg; m r
n1
N
X
fn hLfn rg; m ri
n1
N
X
mn fn
n1
whose matrix form looks as follows:

y f
(5)
with y fym 2 R; m 1; . . . ; Mg being the vector of observations, and

fmn hLfn rg; m ri;
m 1; . . . ; M; n 1; . . . ; N
o
(6)
is the sensing matrix.
f x
(7)
P
where fn Nk1 nk xk , and x fx
Pn 2 R; n 1; . . . ; N g has
only S nonzero entries (i.e., kxk0 Nn1 jxn j0 S, with k k0
being the 0 -norm). Under such hypotheses, (5) and (7) can be
combined to give the following observation equation [7]:
y x Ax
(8)
where A fAmn 2 R; m 1; . . . ; M ; n 1; . . . ; N g is the observation matrix whose (m, n)th entry is equal to
Amn
N
X
hLfk rg; m rikn :
(9)
k1
whose image yr is related to f r through
(1)
n1
yr Lf f rg
Moreover, let us suppose that f is S-sparse [1], with respect

to the signal basis fnk 2 R; k 1; . . . ; K; n 1; . . . ; N g,
so that it can be expressed as
2. CSVProblem Statement
f r
ym hLf f rg; m ri
* (
)
+
N
X
fn n r ; m r m 1; . . . ; M
L
(2)
Equation (8) points out two fundamental requirements of the

standard CS paradigm: (a) the relation between unknowns
(x) and data (y) is linear (i.e., it can be represented by

means of a matrix multiplication) and (b) the unknown vector (x) is sparse (i.e., a suitable signal matrix is known so
that x has only few nonzero entries).
it is mainly dictated by the physics of the problem at hand (i.e.,

the form of the operator Lfg) and it can be only slightly controlled in an indirect way (see Equation (4)). Since no general
rules exist or have been yet proposed in such a latter direction, the CS sampling problem (i) will not be further discussed
in the following.
2.2 Sampling and Recovery Problems

Equation (8) can be used to formulate two different CS
problems: (i) the CS sampling problem and (ii) the CS recovery problem.
The CS sampling problem is mainly concerned with the
design of the signal acquisition system, and it requires specifying
the type [i.e., m r] and the minimum number of measurements
M that allows one the exact recovery of the unknown S-sparse
vector x for a given combination of Lfg, fn r; n 1; . . . ; N g,
and . Mathematically, the sampling problem can be stated
as follows.
As concerns the retrieval of an S-sparse vector from a

set of measurements (CS recovery problem), it is usually formulated in its canonical form as follows.
CS Recovery Problem
Given y 2 RM , nd x 2 RN complying with (8) and
such that x is S-sparse, whose unique solution is guaranteed
under the hypotheses of Theorem 1. Nevertheless, real-world
problems cannot be exactly formulated in the form (8) because of the unavoidable measurement noise. Therefore, the
noise formulation considers the form [10]
e Ax z
y
CS Sampling Problem
Given Lfg, fn r; n 1; . . . ; N g, and , nd M and
fm r, m 1; . . . ; Mg, such that (7) is a well-posed problem
when x is S-sparse.Such a problem can be exactly solved (i.e.,
a unique and exact solution can be found) due to the following
theorem.
(11)
with z 2 R M being the stochastic or deterministic unknown

noise term. Accordingly, the recovery problem turns out to
be generally formulated as follows.
Noisy CS Recovery Problem

e 2 RM , nd x 2 R N complying with (11) and
Given y
such that x is S-sparse.
Theorem 1 [7]
A necessary and sufcient condition for the well-posedness
of (8), when x is S-sparse, is that A complies with the Restricted
Isometry Property (RIP) of order 2 S.
The observation matrix A satises the RIP of order S with
constant 0 G G 1 if, for all S-sparse v 2 RN , the following
condition holds true [1, 7]:
1
kAvk2
1
kvk2
(10)
with k k2 being the 2 -norm. Intuitively, such a condition

guarantees that A preserves the length (i.e., the corresponding k k2 value) of every S-sparse signal v after its projection
in the lower (M N ) M -dimensional space of the observations.
This does not verify if x is not S-sparse, and the equation y
Ax turns out to be still ill posed since the kernel of a rectangular
(M G N ) matrix A is not empty.
On the other hand, it is worthwhile to notice that checking (10) is, in practice, numerically unfeasible even for small
A matrices [7], while observation matrices that a priori t the
RIP (e.g., random Gaussian matrices [27]) are usually considered in CS signal processing/compression problems [27]. Unfortunately, electromagnetic applications do not usually allow
the direct user-denition of the observation matrix [21] since
3. CS Recovery Algorithms
This section is aimed at reviewing the basic ideas behind
CS techniques usually adopted in electromagnetics rather than
providing an exhaustive discussion of the existing methodologies to address the Noisy CS Recovery Problem (the interested
reader is referred to [10] and the references therein for an indepth introduction on computational methods for solving CS
problems). Toward this end, CS recovery methods belonging
to both deterministic (see Section 3.1) and Bayesian (see
Section 3.2) classes [10] will be briey recalled hereinafter.
3.1 Deterministic CS Strategies

From a deterministic point of view, the solution of the
CS recovery problem turns out to be [10]
n
h
io
b arg min kxk0
ek2 :
x
subject to kAx y
x
(12)
Since the minimization of the 0 -norm functional in (12) is an

NP-hard problem [7, 11, 28], it cannot be easily/protably
(computationally) directly addressed. Therefore, greedy pursuit
methods (e.g., orthogonal matching pursuit (OMP) [29], stagewise OMP (StOMP) [30], and compressive sampling matching
pursuit (CoSaMP) [31]) aimed at nding the sparsest x through
an iterative search have been proposed. They are usually based
b (i.e.,
on the iterative renement of the estimated solution x
bi ; i 0; . . . ; I) as schematized in the following [10, 29, 32].
x
1.
Initialization: Initialize the guess solution (b

x0 0),
0
e), the set of nonzero coeffithe residual ( y
cients (0 ;), and the iteration index (i 1).
the 1 -regularized least-square method [34]

n

o
b arg min kAx y
ek2 kxk1
x
x
(17)
with being a regularization parameter; the least absolute

shrinkage and selection operator (LASSO) [35]
n

o
b arg min kAx y
ek2 subject to kxk1 G
x
x
(18)
2.
3.
Identification: Find the nth column of A,

an fAmn 2 R; m 1; . . . ; M g, such that n 6
i1 and an is most strongly correlated with
i1 (i.e., n argfmaxn i1 an g). Update
i (i
i1 [ n).
bi , such that
Estimate: Find the ith guess solution x
i
b
x
n
h
io
e Ai x
arg min y
2
x
(13)
with Ai fAmn 2 R; m 1; . . . ; M; n 1; . . . ; N g
being the ith observation matrix whose generic entry is given by
Ai
mn
Amn ;
if n 2 i
0;
elsewhere
; m 1; . . . ; M ; n 1; . . . ; N :
(14)
4.
e Ai x
bi , update the iteration
Update: Set i y
index (i
i 1), and repeat steps 24 until a suitable stopping criterion (e.g., based on the maximum
number of iterations) is satisfied.
Alternatively, convex relaxation (CR) approaches based

on the Basis Pursuit (BP) [33] have been widely used by
substituting the 0 -norm functional in (12) with an 1 -norm
functional (relaxation) to yield [10, 11]
n

o
b arg min kxk1 subject to kAx y
ek2 :
x
x
n
h
io
b arg min kxk1 krAxk1
x
x
ek2 :
subject to kAx y
(19)
It is worth pointing out that a shared key-feature of these approaches is the exploitation of numerically efcient local search
techniques [10, 38], whose implementation is often available
[1315] because of their nature of linear- or quadratic-programming
problems. Toward this end, let us recall interior-point methods [13]
and gradient techniques [10, 39]. As an example and with referbi1 acence to (17), this latter computes the next estimate x
cording to the following rule [10]:
n
h
8

0 i
i
>
bi A0 Ab
e
x arg minc c x
x y
>
>
<
io

2
bi kck1
12 i c x
2
>

>
>
i
:x
i1
i
i
i
b
b x x
b
x
where i and i are scalar user-dened parameters [10]. In
such a case, the convergence to an optimal or quasi-optimal solution holds true when A satises the RIP [10], and the convergence rate can be signicantly improved for a warm starting,
b0 is available [10]. Unfortui.e., when a good initial estimate x
nately, both previous assumptions are rarely satised in CS-based
electromagnetics.
(15)
3.2 BCS Techniques
The diffusion and success of such a latter formulation is

mainly motivated by the fact that the unique solution of the
linear problem associated to (15) exactly coincides with that
of (12) in the noiseless case and if the RIP holds true. Moreover, due to the convexity of (15), the arising optimization
problem enables the use of local search algorithm [10], and
efcient-solution SW packages have been developed (e.g.,
the well-known L1-Magic tool [13]).
Still within the CR framework, alternative formulations
to (15) have been then employed/proposed. Let us consider
the minimum 1 -error technique [28]
n

o
e k 1
b arg min kAx y
x
x
and the Total Variation (TV) [36, 37]
(16)
Whether deterministic CS recovery algorithms provide reliable and computationally efcient solutions to CS under RIP
conditions [11], these latter techniques cannot be generally satised or a priori efciently veried in several electromagnetic problems [21, 40, 41]. Moreover, since deterministic
approaches do not usually provide any estimation on the conb, their exploitation is not
dence level of the estimated solution x
advisable whenever some sort of (a priori) reliability assessment of the CS result is mandatory [42]. Therefore, alternative
CS recovery algorithms have been studied in CS electromagnetics literature that, on the one hand, do not rely on the RIP of
A to yield accurate and stable results, and, on the other hand,
b [8]. Let us refer
naturally provide the degree of condence of x
to a Bayesian perspective resulting in the following formulation

of the CS retrieval [8, 4346]:
n
o
b arg maxP xje
x
y subject to x Px
x
(20)
where stands for distributed as, and Px is a suitable

prior used to enforce the sparsity of x [8]. In such a way, the
RIP of A is not required, and the arising Bayesian CS (BCS)
approach implicitly calculates the full posterior density funcb, i.e., Pxje
tion for x
y, rather than a single-point estimation
b, giving the condence level for each nth component of x
b, i.e.,
x
bxn , as a function of the covariance matrix C of Pxje
y [8]. Indeed, it is possible to prove that if Px complies with the
following zero-mean Gaussian density function [8]:
Px

QN p
hn x2n
n1 hn exp 2
N
2 2
with 0 being the transpose operator, and covariance

1
b
C
b1 A0 A h
(23)
where the auxiliary parameters h;

b are determined through
dedicated algorithms [8] solving (24), shown at the bottom of
the page.
8
>
>
<
6
b
b arg max6
h;
>
h;
4
>
:
h
i1 39

>
e >
e0
I Adiagh1 A0 y
N log 2 log
I Adiagh1 A0 y
7=
7 :
5>
2
>
;
(24)
b, the knowlIn addition to the information on the reliability of x
edge of C can be also used to extend the capabilities of the retrieval algorithm, with respect to deterministic CS strategies. As
a matter of fact, since C yields a direct evaluation on the level
of uncertainty of the estimation, it indirectly acts as an indicator about the enhancement or usefulness of a set of measurements depending on the higher/lower level of uncertainty of the
resulting estimation [8]. Following this line of reasoning, an
adaptive BCS algorithm can also be envisaged in terms of an
iterative application of the BCS, where, at each step, the choice
of additional measurements is based on the information content
of the added information (i.e., uncertainty reduction) through
the computation of (23).
Within the BCS formulation (20), several variations have
been proposed to include a priori information in the inversion
process through suitable prior denitions [9, 47, 48]. For instance, hierarchical-Laplace priors on x have been included
in [8, 9] by replacing (21) with
(25)
Otherwise, the inversion of K correlated CS tasks (e.g.,

repeated MRI images of the same scene [47]) has been also
considered by statistically linking the Pxk , k 1; . . . ; K as
follows:
Z
Pxk
(22)

exp kxk1
2
2
where is a hyperprior usually distributed according to the

Gamma density function [8]. Such a formulation has the advantage of being mathematically equivalent to (17), thus enabling a direct link between Bayesian and deterministic CS
retrieval strategies [9]. However, the BCS inference cannot be
carried out in closed form [8] since (25) is not conjugate to
the Gaussian likelihood function usually considered to model
the noise distribution.
(21)
where hn is a Gamma-distributed hyperparameter, then

Pxje
y turns out to be a multivariate Gaussian distribution
with mean
b
e
x
b1 CA0 y
Px
Pxk hPhdh
(26)
with h being the shared hyper-prior, to derive the so-called

multitask BCS (MT-BCS) strategy [47]. Furthermore, temporal correlation between successive CS inversion processes has
been also considered, by introducing a block-sparse Bayesian
framework [48]. It is also worth noticing that greedy-like algorithms to solve the CS problems formulated in a Bayesian
framework have been developed [49].
Due to their effectiveness, which often outperforms deterministic CS strategies in terms of reconstruction accuracy and
robustness [8, 9, 47, 48, 51], as well as the availability of standard implementations of BCS and MT-BCS techniques [50],
BCS with Laplace prior algorithms [51], and fast Bayesian
matching pursuit [52] techniques, Bayesian approaches have
been widely adopted in electromagnetics (see Section 4).
4. Application of CS to Electromagnetics
This section is aimed at providing a review, to the best of
the authors knowledge, on the use of CS strategies to electromagnetic problems, by focusing on four main applicative domains: diagnosis and synthesis of antenna arrays (see Table 1
in Section 4.1), DoA estimation (see Table 2 in Section 4.2),
electromagnetic inverse scattering (see Table 3 in Section 4.3),
and radar imaging (see Table 4 in Section 4.4).
4.1 CS in Antenna Array Analysis

and Synthesis
In antenna arrays, the excitation coefcients are linearly
related to the radiated eld y through the M N observation
matrix A3 , whose (m, n)th entry is given by
Amn exp j2 dn cosm
(27)
Table 1. List of scientic publications on CS as applied to

array diagnosis/synthesis.
where dn is the nth element position (in wavelengths), and m

is the mth observation angle (in radians) [for notation simplicity, far-eld measurements in a 1-D arrangement are assumed.
However, both near-eld data and more complex array layouts
can be dealt with, and they have been already addressed in the
CS literature]. Starting from the linearity of such a relationship,
two standard array problems can be formulated within the CS
framework. The former is related to the diagnosis of isolated element failures (see Section 4.1.1), while the other is concerned
with the synthesis of sparse arrangements (see Section 4.1.2).
4.1.1 Array Diagnosis

Let us consider the problem of detecting the failed elements in an N -sized linear array [indicated as array under
DoA estimation.

electromagnetic inverse scattering.
test (AUT)] with excitation coefcients fwtn ; n 1; . . . ; N g

starting from a set of M far-eld measurements [40, 53]
F t m
N
X
wtn exp j2 dn cosm zm
m 1; . . . ; M :
n1
(28)
More in detail, the problem at hand consists in nding the array
elements for which wtn 6 wrn , with wrn being the nth excitation
coefcient of the undamaged array (denoted as reference array). Despite the linearity of (28) as pointed out in Section 4.
1, the problem at hand does not t CS-applicability hypotheses
because of the dense nature of the unknown vector x
fwtn ; n 1; . . . ; N g due to the generally small number of failures
in realistic structures. Nevertheless, the array diagnosis can be
casted as a CS one by exploiting a differential approach [40, 53].
Instead of determining the AUT coefcients fwtn ; n 1; . . . ; N g
from the measured data fym F t m ; m 1; . . . ; M g, the set
of differential excitations x fxn wrn wtn ; n 1; . . . ; N g that
radiate the differential eld
eym F r m ym F r m F t m
(29)
is looked for, with F r m being the mth far-eld pattern measurement of the reference array [40]. Due to this formulation,
the unknown vector x turns out to be now sparse since the number of failed elements S is much more smaller than N [40],
while the relationship between data y and unknowns x is still
linear (27). Thus, CS retrieval techniques can be applied, provided that (27) is transformed in a real-valued form to enable the
exploitation of state-of-the-art algorithms (see Section 3). This
can be easily done by introducing the ctitious matrix

RfAg
A
I fAg
I fAg
RfAg

(30)

radar imaging.
point method called log-barrier algorithm [13]. The proposed

approach proved to outperform standard fast Fourier transform
(FFT) or matrix inversion methods, when highly undersampled
data (i.e., M =N 0:1) affected by noise (e.g., SNR
35 dB)
are at hand and a very small number of failures is present
within the array layout (i.e., S=N 0:01) [40].
The main limitation of such an approach is that it cannot
a priori guarantee its effectiveness since a proof of the RIP for
the observation matrix is generally not available. To overcome
such a drawback, BCS methodologies have been recently applied [53]. More specically, a BCS strategy [8] has been considered in [53] to diagnose linear arrays from near-/far-eld
measurements with nonnegligible improvements in terms of detection accuracy over 1 -based minimization strategies (e.g., the
diagnosis error reduced of above 50%) and truncated singular
value decomposition (SVD) techniques and an enhanced robustness to the noise with good performance also when
SNRG30 dB. Moreover, suitable guidelines for the nonuniform
selection of the data measurement (i.e., the angular sample
points m , m 1; . . . ; M ) have been derived to yield a good
accuracy (e.g., failure retrieval error below 1%) with a moderate
number of failures (S=N
0:1) even when limited measurements (M =N
0:3) are at disposal. Furthermore, the capability
of CS strategies to detect partial failures (i.e., amplitude errors
rather than element shutdown) has been assessed [53].
4.1.2 Sparse Array Synthesis

The design of unequally spaced arrays through CS strategies has attracted a great attention within the electromagnetics
community, as conrmed by the recent list of publications (e.g.,
[2123, 5558]).
and rearranging the differential excitation and differential eld

vectors in double-length vectors comprising their real and imaginary components [53], i.e., x fxn 2 R; n 1; . . . ; 2N g, where

R wrn wtn

xn
I wrnN wtnN
n 1; . . . ; N
n N 1; . . . ; 2N ;
e feym 2 R; m 1; . . . ; 2Mg, where

and y

eym
RF r m F t m
I F r mM F t mM
m 1; . . . ; M
m M 1; . . . ; 2M.
Accordingly, deterministic CS strategies have been rst applied

for detecting the array failures. In [40], the diagnosis of planar
arrays from a small number (M N ) of near-eld measurements has been carried out by means of a reweighted 1 minimization algorithm [54] based on the solution of a sequence
of convex problems formulated as (15) through an interior
Let us consider the Pattern Matching problem [21],

where the positions and the excitations of S array elements
are computed, so that (a) S is minimized, (b) the element locations belong to a (user-dened) candidate set of positions
fdn ; n 1; . . . ; N g, and (c) the corresponding array pattern
F t matches a reference one F r in a set of user-dened
directions f m ; m 1; . . . ; M g [21]. Mathematically, such
a synthesis problem can be directly formulated as a CS one
by setting (a) the nth unknown entry xn of the unknown vector
x to the excitation coefcient of the nth candidate antenna element located at dn , (b) the observation matrix as (27), and (c)
the observation/measurement vector y to the reference pattern
samples, i.e., fym F r m ; m 1; . . . ; M g [21]. As a matter
of fact, x turns out linearly linked to y and, if a suitable sampling of the desired antenna aperture is used (i.e., N is chosen
to be sufciently large), also sparse [21]. Accordingly, the reb provides both the actual element positions (i.e., the
trieved x
candidate locations dn for which bxn 6 0; n 1; . . . ; N ) as
well as the excitations of the array elements (bxn 6 0).
Following this line of reasoning, the synthesis of maximally sparse arrays has been rst solved as a fully real variable
problem in [21, 55]. Symmetric linear [21, 55] and planar [55]
arrangements with real excitations have been synthesized

by considering:
Amn cos2 dn cosm 1-D Case
(31)
and symmetric real-valued patterns as references [i.e.,

F r m F r m , F r m 2 R]. When applied to such a
framework, BCS implementations [8] proved to be a convenient solution, as compared to state-of-the-art strategies, in
terms of accuracy, array sparseness, and computational efciency
[21, 55]. More in detail, a satisfactory matching (i.e., normalized
integral errors below 104 ) has been reached with a 35%40%
element reduction with respect to the corresponding fully populated arrangements affording the same patterns [21, 55]. Similar
results have been also yielded by combining the BCS method with
other techniques such as the matrix pencil method (MPM) [58].
Although efcient, these strategies [21, 55] cannot be effectively employed when complex excitations (e.g., those associated to asymmetrically shaped beams) are of interest [56]
since they rely on the real-valued nature of the synthesis problem. To overcome such a constraint, the pattern matching problem has been recently formulated for asymmetric linear [56]
and planar [57] arrangements within the MT-BCS framework
[47]. Since complex arrangements comprise weights that frequently exhibit nonnegligible real and imaginary components at
the same spatial locations [56], the correlation between the two
components (i.e., the real part and the imaginary one) of the array excitations has been directly included into the prior denition, Px. Due to this approach, performance close to those of
the fully-real synthesis has been reached despite the higher
complexity of the problem at hand [56, 57].
A further extension of CS formulations has been more recently proposed [22, 23] to address the Mask-Constrained
syntheses, as well. Toward this end, the following mathematical
formulation has been considered [22, 23]:
t
n

o
F m 1
b arg min kxk1 subject to
x
x
F t m m
m1
m 2; . . . ; M
(32)
where 1 is the steering angle, m is an angular direction belonging to the sidelobe region S (m 2 S, m 2; . . . ; M ), and
is the user-dened target mask. By suitably casting (32) into a
second-order cone problem, a CS-based approach has been implemented by applying a reweighted 1 -based algorithm [54] for
the minimization [22, 23]. Effective performance in terms of
computation time (e.g., several orders in magnitude smaller than
procedures involving global optimization methods), exibility
(i.e., arbitrary user-dened geometries/constraints), and easy calibration (i.e., very few parameters to be tuned) has been yielded [22].
4.2 DoA Estimation

Unlike CS applications to antenna array synthesis, DoA
estimation through CS has been widely investigated in the
literature [1720]. As a matter of fact, one of the earliest applications of CS theory in electromagnetics [17, 19] was
concerned with the relationships between the achievable performance of DoA retrieval techniques when applied to random sensor arrays and CS. Nevertheless, it is worth noticing
that the exploitation of CS strategies to DoA problems is not
as direct as for the diagnosis and the synthesis of antenna arrays (see Section 4.1.2).
With reference to a linear array composed of M isotropic sensors located at dm , m 1; . . . ; M and measuring the
incident eld due to S monochromatic plane waves coming
from (unknown) directions s , s 1; . . . ; S, the mth received
voltage at the kth temporal snapshot is given by [20]
eykm
S
X
Esk exp j2 dm coss zkm
s1
m 1; . . . ; M ; k 1; . . . ; K
(33)
where Esk is the (unknown) amplitude of the sth incident wave

at the kth snapshot, and zkm is the additive noise term at the mth
array element and kth instant. Since the measurement vector at
the kth instant, i.e., yk fykm ; m 1; . . . ; M g, is not linearly
dependent on the incident directions fs ; s 1; . . . ; Sg of the
incoming signals (33), a suitable reformulation of the problem
is mandatory to enable the use of CS-based methods [20]. Toward this end and similarly to Capon or multiple signal classication (MUSIC) algorithms [59], the angular range is oversampled
with N S samples to rewrite (33) as follows [20]:
eykm
N
X

xkn exp j2 dm cosn zkm ;
n1
m 1; . . . ; M ; k 1; . . . ; K
(34)
where

xkn
Esk ;
0;
if n s
otherwise
k 1; . . . ; K
(35)
is the (sparse) vector whose nth entry is the (unknown) amplitude of the signal impinging from the direction n . Accordingly,
the DoA problem is that of recovering the sparse signal vector
xk fxkn ; n 1; . . . ; N g linearly related to the measurement
vector yk through the observation matrix A of entries [20]

Amn exp j2 dm cosn ; m 1; . . . ; M ; n 1; . . . ; N :
(36)
Chronologically, the rst application of CS to DoA signal detection [18, 60] was concerned with the exploitation of an
1 -regularized least square and CR [39, 61]. In order to enhance the robustness and the angular resolution accuracy, an
iterative CS-based approach has been adopted in [18]. At each
step, the DoA problem is formulated as the minimization of
an 1 -SVD functional corresponding to an even ner grid following a multiresolution scheme similar to that in [62] for inverse
scattering. Still dealing with an 1 -regularized formulation, a
truncated Newton technique belonging to the gradient-based

strategies [39] has been adopted in [60]. Both methods proved
to be a competitive alternative to classical DoA techniques, including Estimation of Signal Parameters via Rotational Variance
Techniques (ESPRIT) and MUSIC, in terms of angular resolution, robustness to noise, and tradeoff between acquisition time
(i.e., number of snapshots) and estimation reliability [18].
As for multisnapshot DoA architectures, customized approaches have been recently developed [20, 48, 59], as well.
In [59], the multisnapshot DoA estimation has been formulated as the minimization of the following mixed 2;0 norm functional:
(
"
#)
" PK 2 #
xk
2
n
k1
b arg min
X
exp
kYe AX kF
X
2
n1
N
X
least square (S-TLS) method has been proposed to solve the

following single-snapshot problem:
8

39
2
k e k 2 >
>
e Ax =
y
<
6
2 7
bk arg min4kxk k1
x
5
k
k k2
>
>
x
1
x
k
;
:
2
where Ae A Z is the perturbed observation matrix, and

Zfzmn ; m 1; . . . ; M ; n 1; . . . ; N g is the noise term. Despite the single-snapshot data at hand [66], the proposed approach was shown to overcome LASSO strategies [35]. A
similar perturbed CS problem has been also dealt with in
[67] through the following CR formulation:

e k

bk arg min kxk k1
ek
x
y
subject to Ax
k
x
(37)
where X fxk ; k 1; . . . ; Kg, Ye fe
yk ; k 1; . . . ; Kg, k k2F is
the Frobenius norm,
and are control parameters, and K is
the total number of snapshots [59]. Toward this end, a joint
0 -approximation (JLZA) algorithm [64] has been adopted by
also comprising the grid renement strategy proposed in [18].
As a result, the JLZA-DoA performances turned out to be
signicantly better than those of existing strategies for the
minimum/optimal number of snapshots, K, the accuracy in
estimating correlated sources, and the suppression of aliasing
effects due to widely spaced (i.e., up to three wavelengths)
sensors [59].
The use of Bayesian formulations (20) has been proposed in [48] to deal with the same multisnapshot processing,
but including the temporal correlation of the sources within
the estimation strategy through a block-sparse prior over X
[48]. The arising evidence maximization problem has then
been solved through an ad-hoc BCS strategy called temporal
multiresponse sparse Bayesian learning (T-MSBL) and derived from the empirical Bayesian technique in [65]. The numerical results have shown that temporally correlated sources
can be more effectively resolved than throughout state-of-the-art
DoA techniques [48].
Still dealing with multiple snapshots, the DoA problem
has been also recently addressed by exploiting hierarchical
sparseness hyperpriors [20]. The multitask version of the BCS
formulation [47] has been adopted by setting the prior as in
(26) and using an ad-hoc relevance vector machine (RVM)
solver for optimizing the arising MT-BCS functional [20]. Numerical comparisons with the single-task BCS (ST-BCS) [8]
implementation of the same approach and state-of-the-art DoA
algorithms proved the effectiveness and robustness of the approach with a reduction of the root-mean-square error of more
than one order in magnitude (K 20) [20].
When the observation matrix A is affected by noise,
modeling for instance the nonidealities of the receivers or the
angular grid mismatches, etc., suitable CS-based strategies have
been recently discussed also [6669]. In [66], the sparse total
(38)
(39)
successively solved with the Alternating Algorithm for Perturbed Basis Pursuit Denoising (AA-P-BPDN) [67]. Beyond
the effectiveness of the DoA estimation, it is worth noticing
that theoretical conditions for the a priori estimate of the recovery error have been deduced [67]. In addition, BCS approaches
[9] have been investigated for solving the perturbed CS problem [68]. For example, the off-grid DoA detection has been
formulated with Laplace priors in [9] with enhanced performances with respect to [18], but at the expense of a slower
speed, particularly when a dense angular grid (i.e., a large N ) is
at hand [68]. Within the same line of reasoning, the use of an
expected likelihood approach [70] in a two-step BCS procedure has been considered to mitigate the bias on A [69]. Once
again, the resulting approach outperformed the spatially
smoothed MUSIC for the accuracy, but with a higher computational complexity [69].
The use of CS for DoA has not been limited to simple
test cases, but different operative conditions have been taken
into account, ranging from narrow-/wideband signals, linear/
planar arrays, up to xed/dynamic sensors, assessing the exibility of the CS tool besides its effectiveness. For instance, the
use of CS to deal with dynamic sensor arrays (i.e., maneuvering receiver) has been addressed in [63], by using a CR strategy [10]. The arising spatial CS (SCS) technique turned out
to overcome several state-of-the-art methods [e.g., MUSIC, focal undetermined system solver (FOCUSS), Wagstaff (WS),
and spatial steered covariance matrix (SSTCM)] in terms of angular resolution and ambiguity resolution [63].
4.3 Inverse Scattering

Electromagnetic inverse scattering problems have been
widely investigated in the CS literature [2426], [41, 42],
[7173]. Although the standard framework of an imaging problem cannot be directly tackled with CS strategies because of its
intrinsic nonlinearity [74], several alternative formulations have
been proposed either within the fully-nonlinear framework
(see Section 4.3.1) or assuming some approximations (see

Section 4.3.2) in modeling the relationships between the data
(usually, the scattered eld) and the unknowns.
4.3.1 Fully Nonlinear Formulations

Dealing with the reconstruction of dielectric proles
of penetrable objects through CS strategies and within the
fully nonlinear framework (i.e., without approximations), the
so-called contrast source formulation [74] has been usually
used [25, 75]. With reference to a 2-D scenario and under
transverse-magnetic (TM) illumination, the scattered electric
eld E k x; y in correspondence with the kth probing source
illumination (k 1; . . . ; K) complies with the following
Data equation:

e k rk
E
m

J k r0 G rkm =r0 drkm z rkm rkm 2 O
More recently, such a strategy has also been extended to

localize sparse metallic scatterers [76] by combining the localshape-function (LSF) formulation of the inverse scattering
problem with the MT-BCS retrieval tool. The arising two-step
procedure proved its effectiveness also when a low number of
illuminations/measurements were available [76].
4.3.2 Approximate Formulations
where frkm ; m 1; . . . ; M g are the M measurement points lying in the observation domain O external to the investigation
domain D, Gr=r0 is the 2-D free-space Greens function,
zr is a zero-mean additive Gaussian noise, and J k r is the
(unknown) Contrast Source. The inversion procedure is then
aimed at solving (40) by looking for the unknown contrast
source J k r starting from the knowledge of the incident electric
eld and the samples of the scattered electric eld at each illumination (the dielectric prole r is then retrieved by means
of the State equation [74]). Toward this end, the method-ofmoments-based pulse-discretized version of (40) is considered
[under such assumption, D is discretized in N subdomains Dn
(D [Nn1 Dn ) centered at rn ; n 1; . . . ; N , and the nth pulse
function is dened as n rf1 if r 2 Dn ; 0 otherwiseg [25]]
and a set of K matrix equations (one for each kth illuminae k rk , xk
tion) equivalent to (11) is derived being eykm E
m
n
k
J rn and where the (m, n)th entry of the observation matrix
is given by
Z
Dn

G rkm =r0 drkm :
(41)
Doing so, the problem can be still casted into the linear CSbased framework, provided that xk fxkn ; n 1; . . . ; N g is
sparse [25].
Following such a guideline, Bayesian strategies [8] have
been applied to qualitatively image sparse (i.e., composed of
few pixels) dielectric scatterers in 2-D scenarios [25]. By exploiting a fast RVM technique [8] to solve (24) for retrieving
the unknown contrast source, the corresponding dielectric contrast in D [25] resulted faithfully reconstructed with a reduction
of one order in magnitude of the inversion error when compared to the state-of-the-art techniques. Moreover, the method
turned out to be more than three orders in magnitude faster
than deterministic techniques when dealing with sparse proles.
10
Still using the contrast source BCS [75], transverse-electric

(TE) data have been effectively processed. Toward this end, the
multitask version of the Bayesian retrieval technique [8] has
been taken into account to mathematically model the relationships among the contrast currents induced by each kth illumination. Despite the increased problem complexity due to the
vectorial nature of the data [75], the MT-BCS-TE outperformed
the method in [25].
(40)
Akmn
Furthermore, it proved to be more robust in high-noise conditions (i.e., SNR 5 dB) [25], as well.
In order to recast the data and the state equations to

protable forms for applying CS retrieval tools, several approximate formulations have been considered for linearization
purposes. Early developments included the application of CS
strategies to imaging problems linearized through the Born approximation (BA) and comprising Laplace priors [9], as envisaged in [7779]. In such a framework, a 1-D inverse problem
has been solved in [80], by means of a greedy-pursuit strategy
(i.e., the subspace pursuit technique [81]), while the use of
BCS strategies with hierarchical priors [8, 47] has been also investigated in [42] and [72], where, besides the reconstruction of
the contrast prole, an estimate of its condence level [42]
has been provided also.
Linearized formulation alternatives to BA have been recently considered [41, 71, 72]. A representative approach is
that discussed in [71], where the inversion strategy has been
developed within the Rytov approximation (RA) [82].
The application of CS retrieval strategies to phaseless data
has been introduced [41], as well. In such a case, a two-step reformulation of the problem has been adopted to retrieve pointlike scatterers by solving of a linear system of equations then
effectively yielded through an 1 -based CS minimization approach (i.e., the Dantzig selector [83]) [41].
4.4 Radar Imaging

A wide set of the scientic literature on the application of
CS strategies to electromagnetics is concerned with radar imaging problems [16]. Since an exhaustive list of all the existing approaches and implementations is almost impossible also due to
the limited space, a summary, to the best of the authors knowledge, of the CS techniques as applied to different radar imaging
modalities will be provided in the following.
4.4.1 Monostatic Radar Imaging

Conventional radar imaging problems [84] are usually
formulated as the retrieval of a reectivity map given a set
of measurements of the scattered electric eld [16]. They
comprise a wide variety of scenarios differing in the measurement setup, the sensing objectives, the processing techniques, the propagation media, and the sensor congurations
[16, 84].
Analogously to inverse scattering, the simplest model of
the reected radar signal for adapting the nonlinear formulation at hand to the CS framework is based on the BA applied
to the free-space propagation of narrowband plane waves impinging on slow targets located in the far eld [16]. Within
such a framework, a regularized OMP strategy [29] has been
applied in [85], showing improved discrimination capabilities
with respect to matched ltering algorithms [85] when detecting close targets.
Otherwise, the retrieval of sparse scatterers located in
the Fresnel region has been carried out in [24], where the
CR formulation (15) of the arising CS problem has been addressed through the subspace pursuit approach in [81].
In order to effectively treat localized extended scatterers
[26], these techniques have been more recently extended
to also deal with multishot single-inputsingle-output sampling schemes.
mitigate the gridding error instabilities, yielding better image reconstructions.
4.4.3 Tomo-SAR Imaging

In tomographic SAR (i.e., Tomo-SAR, which extends the
synthetic aperture principle by using acquisitions from slightly
different viewing angles in elevation to yield 3-D reconstructions [88]), CS approaches have been considered to minimize
the number of signal measurements, while keeping the desired
accuracy in image reconstruction [8890]. Toward this aim,
CR formulations have been widely adopted [8890] to obtain
a superior resolution, an improved robustness to noise, and a
higher computational efciency with respect to state-of-the-art nonlinear least square techniques [88] and truncated SVD (TSVD)
methods [89]. Different greedy/local minimization solvers have
been employed, ranging from BP [33, 90], BP DeQuantizer
[88, 91], DouglasRachford splitting method [89, 92], up to the
iterative shrinkage/thresholding method [89, 93].
Hybrid approaches based on CS retrieval tools have
been also analyzed to further improve Tomo-SAR imaging
[94. 95]. For instance, spectral estimation algorithms (e.g., the
SL1MMER technique) have been proposed by combining 1 minimization CS steps with a model order reduction and a
maximum-likelihood parameter selection [94]. Moreover, combining the NDOF-TSVD approaches [96] and CS techniques has
been discussed [95] by formulating the sparse problem at hand
as the second-order cone problem [97, 98].
4.4.2 Spotlight SAR Imaging
4.4.4 ISAR Imaging
The solution of Synthetic Aperture Radar (SAR) problems through CS strategies has attracted great attention because, on the one hand, most targets can be easily modeled as
sparse distributions in suitable representation bases (e.g., wavelet or complex wavelet [37]) and, on the other hand, it is possible to linearly (approximated) describe the relation between the
received demodulated signal and the unknown reectivity eld
[37]. Typical applications are mainly concerned with spotlight
SAR (where the traveling radar sensor continuously steers the
antenna beam to illuminate the terrain patch being imaged)
[37, 86, 87] where both TV (19) [37] and BCS (20) [86] formulations have been assessed in undersampling SAR data, increasing the robustness to noise, and reducing the sidelobes in
retrieved images despite the use of simple, but efcient, reconstruction algorithms [8], [36].
Dealing with CS applications, inverse SAR (ISAR) problems (in which a xed radar monitors a moving target to yield
high-resolution images [37]) have been widely discussed also
[99, 100], because of their intrinsically sparse nature related to
the representation of the targets as few strong scatterers whose
number is much more smaller than that of the pixels of the image under analysis [100]. Moreover, the ISAR signal model turns
out to be linear with respect to the (complex) amplitude of the
scattering bins [100] in the rangeDoppler domain. Due to these
properties, both Bayesian [99] and 1 -regularized formulations
[100104] have been adopted to yield a high range resolution
with fewer data samples of stepped-frequency chirp signals [99]
and to enhance the ISAR antijamming capabilities [102]. More
specically, BCS approaches [8] proved in [99] to be able to
give an image detailed as those from state-of-the-art FFT reconstruction techniques, while more effectively suppressing the
image sidelobes.
More recently, the presence of errors in the matrix

A caused by misalignments between actual targets and processing grid (called gridding errors) has been addressed
[87], as well. Similarly to the perturbed CS problem discussed
in Section 4.2 for the DoA estimation, a CR formulation has
been presented in [87]. A greedy technique, the supportconstrained OMP (SCOMP) [87], has then been introduced to
An improvement of the ISAR image quality has also been

reached through greedy-pursuit strategies [33, 101] and local
search techniques (e.g., half-quadratic regularization [105], conjugate gradient [106], or convex programming [38] methods)
when applied to 1 -based functionals [101103].
11
Further enhancements in signal recovery and noise suppression have been yielded by solving the following weighted
1 -norm problem [100]:
n

o
b arg min kW xk1 subject to kAx y
e k 2
x
x
wall backscattering [112] and effectively dealing with moving targets when integrated with change-detection approaches [113].
5. Final Remarks and Future Trends
(42)
by means of OMP techniques [29]. In (42), W fwnp ; n
1; . . . ; N ; p 1; . . . ; N g is the weighting matrix [100], where
small weights allow one to detect strong signal components,
while large coefcients mitigate the noise components [100].
As for the phase errors due to unexpected target motions in 1 -norm regularized ISAR imaging [104], local
search strategies based on quasi-Newton solvers have been
applied, showing better compensation features with respect
to state-of-the-art gapped-data amplitude and phase estimation (GAPES) techniques [104].
4.4.5 GPR Imaging

Data acquisition and imaging methods based on CS have
been also applied to stepped-frequency continuous-wave ground
penetrating radars (SFCW-GPR) mainly to mitigate/overcome
technological issues related to the required high data acquisition
speed [107110]. In such a framework, a popularly adopted
formulation is based on the Dantzig selector [83] dealt with
convex optimization tools [38] because of its higher stability
when processing incomplete and noisy data as those of GPR
applications [107109]. Due to the CS-based imaging features
and the intrinsic sparsity of GPR imaging problems, robust reconstructions with better resolutions than standard back-projection
algorithms have been achieved also experimentally [107].
In order to minimize the acquisition time in SFCW-GPR
prototypes [110], TV formulations solved with interior-point
methods [98] have been investigated also.
4.4.6 TWRI
As far as through-the-wall radar imaging (TWRI) is
concerned, there has been an increasing interest within the
scientic community [111113] to simplify high-resolution
ultrawideband-TWRI imaging systems, in terms of acquisition
time and hardware (HW) complexity, by leveraging on the features of CS [111]. As a matter of fact, the use of CS strategies
based on Dantzig selector formulations [83] has been shown
to outperform conventional delay-and-sum beamforming (DSBF)
algorithms in both imaging accuracy and data reduction for the
inversion. As for this latter issue, let us consider that the CS requires only 7.7% of the DSBF data [111]. Moreover, the use of
CR formulations solved through greedy techniques (e.g., OMP
[29] and CoSaMP [31]) has been introduced, for mitigating the
12
The CS paradigm has enabled a wide range of new applicative scenarios to be investigated and developed, due to its unique features. This paper was aimed at giving a short review of
the applications of CS to electromagnetics starting from the CS
signal model denition and including the formulation of the
sampling and recovery problems. While pointing out the
mandatory hypotheses for the CS exploitation, some indications
on the reliability of CS algorithms (mainly concerned with the
properties of the observation matrix) have been also recalled.
Popular Bayesian and deterministic CS recovery algorithms
have shortly summarized, to point out the main features of each
solution strategy as well as their advantages and limitations in
terms of efciency and exibility. Successively, the potentialities
of CS strategies in solving sparse formulations naturally arising
in a broad class of electromagnetic problems (e.g., array synthesis and diagnosis) have been illustrated by reporting, to the best
of the authors knowledge, the most recent advances on these
topics. Moreover, the possibility to apply CS techniques to nonlinear problems concerned with DoA estimation, inverse scattering, and radar imaging through suitable reformulations and
approximations has been discussed also mentioning the most diffused retrieval strategies usually adopted in these cases.
For the interested readers and potentially future practitioners, the key motivations and main advantages (e.g., numerical efciency, robustness to noise, exibility, and accuracy) of
applying CS in electromagnetics have been pointed out remarked in the illustrated scenarios. Although, certainly, CS
does not outperform all existing retrieval techniques, it has been
shown that it is able to yield enhanced performances with respect
to several state-of-the-art strategies whenever a suitable sparse
description of the problem is at hand. According to the results
in the leading-edge researches, CS represents a reliable, effective, and efcient paradigm/tool for properly addressing several
conventional electromagnetic problems as well as for envisaging
new applicative elds of research.
On the other hand, it cannot be neglected that both theoretical issues, mathematical implementations, and numerical
features of CS applications to electromagnetics are still at the
beginning, and they have still to be carefully studied and addressed in ongoing and future research activities. For instance,
the solution of sampling problems arising in electromagnetics
(e.g., the minimization of the number of required measurements in inverse scattering and array diagnosis) through a
strategy that a priori guarantees the CS observation matrix to
comply with the RIP represents an interesting and challenging
approach to minimize HW and processing costs. Moreover,
the extension of CS formulations to fully nonlinear problems,
whether possible, could enable signicant enhancements in a
wider set of applicative areas within electromagnetics.
6. References
[1] E. J. Candes and M. B. Wakin, An introduction to compressive
sampling, IEEE Signal Process. Mag., vol. 25, no. 2, pp. 2130,
Mar. 2008.
[2] R. Baraniuk, More is less: Signal processing and the data deluge,
Science, vol. 331, no. 6018, pp. 717719, Feb. 2011.
[3] M. Lustig, D. L. Donoho, J. M. Santos, and J. M. Pauly, Compressed sensing MRI, IEEE Signal Process. Mag., vol. 25, no. 2,
pp. 7282, Mar. 2008.
[4] W. L. Chan, M. Moravec, R. Baraniuk, and D. Mittleman, Terahertz imaging with compressed sensing and phase retrieval, Opt.
Lett., vol. 33, no. 9, pp. 974976, May 2008.
[5] M. Cossalter, G. Valenzise, M. Tagliasacchi, and S. Tubaro, Joint
compressive video coding and analysis, IEEE Trans. Multimedia, vol. 12,
no. 3, pp. 168183, Apr. 2010.
[6] C. R. Berger, Z. Wang, J. Huang, and S. Zhou, Application of
compressive sensing to sparse channel estimation, IEEE Commun.
Mag., vol. 48, no. 11, pp. 164174, Nov. 2010.
[7] R. G. Baraniuk, Compressive sampling, IEEE Signal Process. Mag.,
vol. 24, no. 4, pp. 118124, Jul. 2007.
[8] S. Ji, Y. Xue, and L. Carin, Bayesian compressive sensing, IEEE
Trans. Signal Process., vol. 56, no. 6, pp. 23462356, Jun. 2008.
[9] S. D. Babacan, R. Molina, and A. K. Katsaggelos, Bayesian compressive sensing using laplace priors, IEEE Trans. Image Process.,
vol. 19, no. 1, pp. 5363, Jan. 2010.
[10] J. A. Tropp and S. J. Wright, Computational methods for sparse solution of linear inverse problems, Proc. IEEE, vol. 98, no. 6, pp. 948958,
Jun. 2010.
[11] R. G. Baraniuk, V. Cevher, M. F. Duarte, and C. Hegde, Modelbased compressive sensing, IEEE Trans. Inf. Theory, vol. 56, no. 4,
pp. 19822001, Apr. 2010.
[12] M. F. Duarte and Y. C. Eldar, Structured compressed sensing: from
theory to applications, IEEE Trans. Signal Process., vol. 59, no. 9,
pp. 40534085, Sep. 2011.
[13] E. Candes and J. Romberg, L1-Magic Code. [Online]. Available:
http://users.ece.gatech.edu/~justin/l1magic/.
[14] V. Stodden et al., SparseLab Code. [Online]. Available: http://sparselab.
stanford.edu/.
[15] K. Koh, S.-J. Kim, and S. Boyd, ell-1LS Code [Online]. Available:
http://www.stanford.edu/~boyd/l1_ls/.
[16] L. C. Potter, E. Ertin, J. T. Parker, and M. Cetin, Sparsity and
compressed sensing in radar imaging, Proc. IEEE, vol. 98, no. 6,
pp. 10061020, Jun. 2010.
[17] L. Carin, On the relationship between compressive sensing and random sensor arrays, IEEE Antennas Propag. Mag., vol. 51, no. 5,
pp. 7281, Oct. 2009.
[18] D. Malioutov, M. Cetin, and A. S. Willsky, A sparse signal reconstruction perspective for source localization with sensor arrays, IEEE
Trans. Signal Process., vol. 53, no. 8, pp. 30103022, Aug. 2005.
[19] L. Carin, L. Dehong, and G. Bin, Coherence, compressive sensing,
and random sensor arrays, IEEE Antennas Propag. Mag., vol. 53, no. 4,
pp. 2839, Aug. 2011.
[20] M. Carlin, P. Rocca, G. Oliveri, F. Viani, and A. Massa, Directionsof-arrival estimation through Bayesian compressive sensing strategies,
IEEE Trans. Antennas Propag., vol. 61, no. 7, pp. 38283838, Jul. 2013.
[21] G. Oliveri and A. Massa, Bayesian compressive sampling for pattern
synthesis with maximaly sparse non-uniform linear arrays, IEEE
Trans. Antennas Propag., vol. 59, no. 2, pp. 467481, Feb. 2011.
[22] B. Fuchs, Synthesis of sparse arrays with focused or shaped beampattern via sequential convex optimizations, IEEE Trans. Antennas Propag., vol. 60, no. 7, pp. 34993503, Jul. 2012.
[23] G. Prisco and M. DUrso, Maximally sparse arrays via sequential
convex optimizations, IEEE Antennas Wireless Propag. Lett., vol. 11,
pp. 192195, Feb. 2012.
[24] A. Fannjiang, P. Yan, and T. Strohmer, Compressed remote sensing
of sparse objects, arXiv, vol. 0904.3994v2, pp. 122, 2009.
[25] G. Oliveri, P. Rocca, and A. Massa, A Bayesian compressive samplingbased inversion for imaging sparse scatterers, IEEE Trans. Geosci. Remote
Sens., vol. 49, no. 10, pp. 39934006, Oct. 2011.
[26] A. C. Fannjiang, Compressive inverse scattering II. SISO measurements with born scatterers, Inverse Probl., vol. 26, no. 3, pp. 117,
Mar. 2010.
[27] R. Baraniuk, M. Davenport, R. DeVore, and M. Wakin, A simple
proof of the restricted isometry property for random matrices, Construct. Approx., vol. 28, no. 3, pp. 253263, Dec. 2008.
[28] E. J. Candes and T. Tao, Decoding by linear programming, IEEE

Trans. Inf. Theory, vol. 51, no. 12, pp. 42034215, Dec. 2005.
[29] J. Tropp and A. Gilbert, Signal recovery from random measurements
via orthogonal matching pursuit, IEEE Trans. Inf. Theory, vol. 53,
no. 12, pp. 46554666, Dec. 2007.
[30] D. L. Donoho, Y. Tsaig, I. Drori, and J. L. Starck, Sparse solution
of underdetermined systems of linear equations by stagewise orthogonal matching pursuit, IEEE Trans. Inf. Theory, vol. 58, no. 2,
pp. 10941121, Feb. 2012.
[31] D. Needell and J. A. Tropp, CoSaMP: Iterative signal recovery from
incomplete and inaccurate samples, Appl. Comput. Harmon. Anal.,
vol. 26, no. 3, pp. 301321, May 2009.
[32] S. G. Mallat and Z. Zhang, Matching pursuits with time-frequency dictionaries, IEEE Trans. Signal Process., vol. 41, no. 12, pp. 33973415,
Dec. 1993.
[33] S. Chen, Basis pursuit, Ph.D. dissertation, Dept. Stat., Stanford
Univ., Stanford, CA, USA, 1995.
[34] S.-J. Kim, K. Koh, M. Lustig, S. Boyd, and D. Gorinevsky, An
interior-point method for large-scale 1 -regularized least squares, IEEE J.
Sel. Topics Signal Process., vol. 1, no. 4, pp. 606617, Dec. 2007.
[35] R. Tibshirani, Regression shrinkage and selection via the Lasso,
J. Roy. Stat. Soc. B, vol. 58, no. 1, pp. 267288, 1996.
[36] J. Shi and S. Osher, A nonlinear inverse scale space method for a
convex multiplicative noise model, Comput. Appl. Math., Univ. California, Los Angeles, CA, USA, Tech. Rep. 0710, 2007.
[37] V. M. Patel, G. R. Easley, D. M. Healy, and R. Chellappa, Compressed synthetic aperture radar, IEEE J. Sel. Topics Signal Process.,
vol. 4, no. 2, pp. 244254, Apr. 2010.
[38] M. Grant and S. Boyd, Graph implementations for nonsmooth convex programs, in Recent Advances in Learning and Control, Berlin,
Germany: Springer-Verlag, 2008, pp. 95110.
[39] S. Nash, A survey of TruncatedNewton methods, Journal on Computational and Applied Mathematics, vol. 24, no. 1/2, pp. 4559, Dec. 2000.
[40] M. D. Migliore, A compressed sensing approach for array diagnosis
from a small set of near-eld measurements, IEEE Trans. Antennas
Propag., vol. 59, no. 6, pp. 21272133, Jun. 2011.
[41] L. Pan, X. Chen, and S. P. Yeo, A compressive-sensing-based phaseless imaging method for point-like dielectric objects, IEEE Trans.
Antennas Propag., vol. 60, no. 11, pp. 54725475, Nov. 2012.
[42] L. Poli, G. Oliveri, and A. Massa Microwave imaging within the
rst-order born approximation by means of the contrast-eld Bayesian
compressive sensing, IEEE Trans. Antennas Propag., vol. 60, no. 6,
pp. 28652879, Jun. 2012.
[43] M. E. Tipping, The relevance vector machine, in Advances in Neural Information Processing Systems, vol. 12, S. A. Solla, T. K.
Leen, and K.-R. Muller Eds., Cambridge, MA, USA: MIT Press,
2000, pp. 652658.
[44] M. E. Tipping, Sparse Bayesian learning and the relevance vector machine, J. Mach. Learn. Res., vol. 56, no. 6, pp. 211244, Jun. 2001.
[45] M. E. Tipping and A. C. Faul, Fast marginal likelihood maximization for sparse Bayesian models, in Proc. 9th Int. Workshop Artif. Intell. Stat., C. M. Bishop and B. J. Frey, Eds., 2003, pp. 36.
[Online]. Available: http://citeseer.ist.psu.edu/611465.html.
[46] A. C. Faul and M. E. Tipping, Analysis of sparse Bayesian learning,
in Proc. Adv. NIPS, T. G. Dietterich, S. Becker, and Z. Ghahramani,
Eds., 2002, pp. 383389. [Online]. Available: http://citeseer.ist.psu.edu/
faul01analysis.html.
[47] S. Ji, D. Dunson, and L. Carin, Multi-task compressive sampling,
IEEE Trans. Signal Process., vol. 57, no. 1, pp. 92106, Jan. 2009.
[48] Z. Zhang and B. D. Rao, Sparse signal recovery with temporally correlated source vectors using sparse Bayesian learning, IEEE J. Sel.
Topics Signal Process., vol. 5, no. 5, pp. 912926, Sep. 2011.
[49] P. Schniter, L. C. Potter, and J. Ziniel, Fast Bayesian matching pursuit, in Proc. Inf. Theory Appl. Workshop, Jan. 27Feb. 1, 2008,
pp. 326333.
[50] S. Ji, Y. Xue, and L. Carin, Bayesian Compressive Sensing Code.
[Online]. Available: http://people.ee.duke.edu/~lcarin/BCS.html.
[51] R. R. McCormick, Bayesian Compressive Sensing Using Laplace Priors
Code. [Online]. Available: http://ivpl.eecs.northwestern.edu/node/1169.
[52] P. Schniter, L. C. Potter, and J. Ziniel, Fast Bayesian Matching Pursuit Code. [Online]. Available: http://www2.ece.ohio-state.edu/~zinielj/
fbmp/download.html.
[53] G. Oliveri, P. Rocca, and A. Massa, Reliable diagnosis of large linear arrays-Bayesian compressive sensing approach, IEEE Trans. Antennas Propag., vol. 60, no. 10, pp. 46274636, Oct. 2012.
[54] E. J. Candes, M. B. Wakin, and S. P. Boyd, Enhancing sparsity by
reweighted 1 minimization, J. Fourier Anal. Appl., vol. 14, no. 5/6,
pp. 877905, Oct. 2008.
13
[55] W. Zhang, L. Li, and F. Li, Reducing the number of elements in
linear and planar antenna arrays with sparseness constrained optimization, IEEE Trans. Antennas Propag., vol. 59, no. 8, pp. 31063111,
Aug. 2011.
[56] G. Oliveri, M. Carlin, and A. Massa, Complex-weight sparse linear
array synthesis by bayesian compressive sampling, IEEE Trans. Antennas Propag., vol. 60, no. 5, pp. 23092326, May 2012.
[57] F. Viani, G. Oliveri, and A. Massa, Compressive sensing pattern
matching techniques for synthesizing planar sparse arrays, IEEE
Trans. Antennas Propag., vol. 61, no. 9, pp. 45774587, Sep. 2013.
[58] L. F. Yepes, D. H. Covarrubias, M. A. Alonso, and J. G. Arceo,
Synthesis of two-dimensional antenna array using independent compression regions, IEEE Trans. Antennas Propag., vol. 61, no. 1,
pp. 449453, Jan. 2013.
[59] M. M. Hyder and K. Mahata, Direction-of-arrival estimation using a
mixed 20 norm approximation, IEEE Trans. Signal Process., vol. 58,
no. 9, pp. 46474655, Sep. 2010.
[60] D. Model and M. Zibulevsky, Signal reconstruction in sensor arrays using sparse representations, Signal Process., vol. 86, no. 3,
pp. 624638, Mar. 2006.
[61] D. M. Malioutov, A sparse signal reconstruction perspective for
source localization with sensor arrays, M.S. thesis, Dept. Elect. Eng.
Comput. Sci., MIT, Cambridge, MA, USA, Jul. 2003. [Online]. Available: http://ssg.mit.edu/~dmm/publications/malioutov_MS_thesis.pdf.
[62] S. Caorsi, M. Donelli, D. Franceschini, and A. Massa, A new methodology based on an iterative multiscaling for microwave imaging, IEEE
Trans. Microw. Theory Tech., vol. 51, no. 4, pp. 11621173, Apr. 2003.
[63] I. Bilik, Spatial compressive sensing for direction-of-arrival estimation
of multiple sources using dynamic sensor arrays, IEEE Trans. Aerosp.
Electron. Syst., vol. 47, no. 3, pp. 17541769, Jul. 2011.
[64] M. M. Hyder and K. Mahata, An improved smoothed approximation
algorithm for sparse representation, IEEE Trans. Signal Process., vol. 58,
no. 4, pp. 21942205, Apr. 2010.
[65] D. P. Wipf and B. D. Rao, An empirical Bayesian strategy for solving the simultaneous sparse approximation problem, IEEE Trans. Signal Process., vol. 55, no. 7, pp. 37043716, Jul. 2007.
[66] H. Zhu, G. Leus, and G. B. Giannakis, Sparsity-cognizant total
least-squares for perturbed compressive sampling, IEEE Trans. Signal
Process., vol. 59, no. 5, pp. 22022016, May 2011.
[67] Z. Yang, C. Zhang, and L. Xie, Robustly stable signal recovery in
compressed sensing with structured matrix perturbation, IEEE Trans.
Signal Process., vol. 60, no. 9, pp. 46584671, Sep. 2012.
[68] Z. Yang, L. Xie, and C. Zhang, Off-grid direction of arrival estimation using sparse Bayesian inference, IEEE Trans. Signal Process.,
vol. 61, no. 1, pp. 3843, Jan. 2013.
[69] E. T. Northardt, I. Bilik, and Y. I. Abramovich, Spatial compressive
sensing for direction-of-arrival estimation with bias mitigation via expected
likelihood, IEEE Trans. Signal Process., vol. 61, no. 5, pp. 1183106,
Mar. 2013.
[70] Y. Abramovich and B. Johnson, GLRT-based detection-estimation
for under-sampled training conditions, IEEE Trans. Signal Process.,
vol. 56, no. 8, pp. 36003612, Aug. 2008.
[71] G. Oliveri, L. Poli, P. Rocca, and A. Massa, Bayesian compressive
optical imaging within the Rytov approximation, Opt. Lett., vol. 37,
no. 10, pp. 17601762, May 2012.
[72] F. Viani, L. Poli, G. Oliveri, F. Robol, and A. Massa, Sparse scatterers imaging through approximated multi-task compressive sensing
strategies, Microw. Opt. Technol. Lett., vol. 55, no. 7, pp. 15531558,
Jul. 2013.
[73] A. C. Fannjiang, Compressive inverse scattering I. High frequency
SIMO measurements, Inverse Probl., vol. 26, no. 3, pp. 130,
Mar. 2010.
[74] P. M. van Den Berg and A. Abubakar, Contrast source inversion method:
State of the art, Progr. Electromagn. Res., vol. 34, pp. 189218, 2001.
[75] L. Poli, G. Oliveri, P. Rocca, and A. Massa, Bayesian compressive
sensing approaches for the reconstruction of two-dimensional sparse
scatterers under TE illuminations, IEEE Trans. Geosci. Remote Sens.,
vol. 51, no. 5, pp. 29202936, May 2013.
[76] L. Poli, G. Oliveri, and A. Massa, Imaging sparse metallic cylinders
through a local shape function Bayesian compressive sensing approach,
J. Opt. Soc. Amer. A, Opt. Image Sci., vol. 30, no. 6, pp. 12611272,
Jun. 2013.
[77] E. A. Marengo et al., Compressive sensing for inverse scattering,
Proc. XXIX URSI Gener. Assem., Chicago, IL, USA, Aug. 716, 2008.
[78] E. A. Marengo, Compressive sensing and signal subspace methods
for inverse scattering including multiple scattering, in Proc. IEEE Int.
Workshop Comput. Adv. Multi-Sensor Adapt. Process., Boston, MA,
USA, Jul. 711, 2008, pp. 109112.
14
[79] E. A. Marengo, Subspace and Bayesian compressive sensing methods

in imaging, in Proc. Progr. Electromagn. Res. Symp., Cambridge,
MA, Jul. 26, 2008.
[80] A. Fannjiang, Compressed imaging of subwavelength structures,
SIAM J. Imag. Sci., vol. 2, no. 4, pp. 12771291, 2009.
[81] W. Dai and O. Milenkovic, Subspace pursuit for compressive sensing signal reconstruction, IEEE Trans. Inf. Theory, vol. 55, no. 5,
pp. 22302249, May 2009.
[82] A. J. Devaney, Inverse-scattering theory within the Rytov approximation, Opt. Lett., vol. 6, no. 8, pp. 374376, Aug. 1981.
[83] E. Candes and T. Tao, The Dantzig selector: Statistical estimation
when p is much larger than n, Ann. Stat., vol. 35, no. 6, pp. 23132351,
Dec. 2007.
[84] G. Franceschetti, R. Lanari, Synthetic Aperture Radar Processing.
Boca Raton, FL, USA: CRC Press, 1999.
[85] M. Tello Alonso, P. Lopez-Dekker, and J. J. Mallorqui, A novel
strategy for radar imaging based on compressive sensing, IEEE Trans.
Geosci. Remote Sens., vol. 48, no. 12, pp. 42854295, Dec. 2010.
[86] J. Xu, Y. Pi, and Z. Cao, Bayesian compressive sensing in synthetic
aperture radar imaging, IET Radar Sonar Navig., vol. 6, no. 1, pp. 28,
Jan. 2012.
[87] A. Fannjiang and H.-C. Tseng, Compressive radar with off-grid
targets: A perturbation approach, Inverse Probl., vol. 29, no. 5,
pp. 054008-1054008-23, May 2013.
[88] X. X. Zhu and R. Bamler, Tomographic SAR inversion by L1 -norm
regularizationThe compressive sensing approach, IEEE Trans.
Geosci. Remote Sens., vol. 48, no. 10, pp. 38393846, Oct. 2010.
[89] A. Budillon, A. Evangelista, and G. Schirinzi, Three-dimensional
SAR focusing from multipass signals using compressive sampling, IEEE
Trans. Geosci. Remote Sens., vol. 49, no. 1, pp. 488499, Jan. 2011.
[90] S. Xilong, Y. Anxi, D. Zhen, and L. Diannong, Three-dimensional
SAR focusing via compressive sensing: The case study of angel stadium, IEEE Geosci. Remote Sens. Lett., vol. 9, no. 4, pp. 759763,
Jul. 2012.
[91] L. Jacques, D. K. Hammond, and J. M. Fadili, Dequantizing compressed sensing: When oversampling and non-Gaussian constraints combine, IEEE Trans. Inf. Theory, vol. 57, no. 1, pp. 559571, Jan. 2011.
[92] P. L. Combettes and J. C. Pesquet, A Douglas-Rachford splitting approach to nonsmooth convex variational signal recovery, IEEE J. Sel.
Topics Signal Process., vol. 1, no. 4, pp. 564574, Dec. 2007.
[93] M. Figueiredo and R. Nowak, An EM algorithm for wavelet-based image restoration, IEEE Trans. Image Process., vol. 12, no. 8, pp. 906916,
Aug. 2003.
[94] X. X. Zhu and R. Bamler, Super-resolution power and robustness of
compressive sensing for spectral estimation with application to spaceborne tomographic SAR, IEEE Trans. Geosci. Remote Sens., vol. 50,
no. 1, pp. 247258, Jan. 2012.
[95] R. Solimene, F. Ahmad, and F. Soldovieri, A novel CS-TSVD strategy to perform data reduction in linear inverse scattering problems,
IEEE Geosci. Remote Sens. Lett., vol. 9, no. 5, pp. 881885, Sep. 2012.
[96] F. Soldovieri, R. Solimene, and F. Ahmad, Experimental validation
of a microwave tomographic approach for through-the-wall radar
imaging, in Proc. SPIE Defense Security Symp., vol. 7699, 2010,
p. 76990C.
[97] Y.-S. Yoon and M. G. Amin, Through-the-wall radar imaging using
compressive sensing along temporal frequency domain, in Proc.
IEEE ICASSP, Mar. 2010, pp. 28062809.
[98] E. J. Candes, J. Romberg, and T. Tao, Robust uncertainty principles:
Exact signal reconstruction from highly incomplete frequency information, IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489509, Feb. 2006.
[99] X.-C. Xie and Y.-H. Zhang, High-resolution imaging of moving
train by ground-based radar with compressive sensing, Electron.
Lett., vol. 46, no. 7, pp. 529531, Apr. 2010.
[100] L. Zhang et al., Resolution enhancement for inversed synthetic aperture
radar imaging under Low SNR via improved compressive sensing, IEEE
Trans. Geosci. Remote Sens., vol. 48, no. 10, pp. 38243838, Oct. 2010.
[101] F. Ye, D. Liang, and J. Zhu, ISAR enhancement technology based
on compressed sensing, Electron. Lett., vol. 47, no. 10, pp. 620621,
May 2011.
[102] L. Zhang, Z.-J. Qiao, M. Xing, Y. Li, and Z. Bao, High-resolution
ISAR imaging with sparse stepped-frequency waveforms, IEEE Trans.
Geosci. Remote Sens., vol. 49, no. 11, pp. 46304651, Nov. 2011.
[103] H. Wang, Y. Quan, M. Xing, and S. Zhang, ISAR imaging via
sparse probing frequencies, IEEE Geosci. Remote Sens. Lett., vol. 8,
no. 3, pp. 451455, May 2011.
[104] L. Zhang et al., High-resolution ISAR imaging by exploiting sparse apertures, IEEE Trans. Antennas Propag., vol. AP-60, no. 2, pp. 9971008,
Feb. 2012.
Paolo Rocca (S08; M09; SM13) received
the M.S. degree (summa cum laude) in telecommunications engineering and the Ph.D.
degree in information and communication
technologies from the University of Trento,
Trento, Italy, in 2005 and 2008, respectively.
He is currently an Assistant Professor with
the Department of Information Engineering and
Computer Science, University of Trento, where
he is also a member of the ELEDIA Research
Center. He has been a Visiting Student at the
Pennsylvania State University, State College,
PA, USA, and at the University Mediterranea of
Reggio Calabria, Reggio Calabria, Italy, and a Visiting Researcher at the cole
suprieure dlectricit (SUPELEC), Paris, France. He is the author/coauthor of
over 200 peer-reviewed papers on international journals and conferences. His
main interests are in the framework of antenna array synthesis and design, electromagnetic inverse scattering, and optimization techniques for
electromagnetics.
Dr. Rocca was a recipient of the best Ph.D. thesis award IEEE-GRS Central
Italy Chapter from the IEEE Geoscience and Remote Sensing Society and the
Italy Section. He serves as an Associate Editor of the IEEE ANTENNAS AND
WIRELESS PROPAGATION LETTERS.
[105] M. Cetin, W. C. Karl, and A. S. Willsky, Feature-preserving regularization method for complex-valued inverse problems with application to
coherent imaging, Opt. Eng., vol. 45, no. 1, pp. 017003-1017003-11,
Jan. 2006.
[106] O. Axelsson, Iterative Solution Methods. Cambridge, U.K.: Cambridge
Univ. Press, 1994.
[107] A. C. Gurbuz, J. H. McClellan, and W. R. Scott, A compressive
sensing data acquisition and imaging method for stepped frequency
GPRs, IEEE Trans. Signal Process., vol. 57, no. 7, pp. 26402650,
Jul. 2009.
[108] M. A. C. Tuncer and A. C. Gurbuz, Ground reection removal in
compressive sensing ground penetrating radars, IEEE Geosci. Remote
Sens. Lett., vol. 9, no. 1, pp. 2327, Jan. 2012.
[109] L. Qu and T. Yang, Investigation of air/ground reection and antenna beamwidth for compressive sensing SFCW GPR migration imaging, IEEE Trans. Geosci. Remote Sens., vol. 50, no. 8, pp. 31433149,
Aug. 2012.
[110] A. B. Suksmono, E. Bharata, A. A. Lestari, A. G. Yarovoy, and
L. P. Ligthart, Compressive stepped-frequency continuous-wave
ground-penetrating radar, IEEE Geosci. Remote Sens. Lett., vol. 7,
no. 4, pp. 665669, Oct. 2010.
[111] Q. Huang, L. Qu, B. Wu, and F. Guangyou, UWB through-wall
imaging based on compressive sensing, IEEE Trans. Geosci. Remote
Sens., vol. 48, no. 3, pp. 14081415, Mar. 2010.
[112] E. Lagunas, M. G. Amin, F. Ahmad, and M. Najar, Joint wall mitigation and compressive sensing for indoor image reconstruction, IEEE
Trans. Geosci. Remote Sens., vol. 51, no. 2, pp. 891906, Feb. 2013.
[113] F. Ahmad and M. G. Amin, Through-the-wall human motion indication using sparsity-driven change detection, IEEE Trans. Geosci. Remote Sens., vol. 51, no. 2, pp. 881890, Feb. 2013.
Andrea Massa (M96) received the laurea

degree in electronic engineering and the Ph.D.
degree in electronics and computer science
from the University of Genoa, Genoa, Italy,
in 1992 and 1996, respectively.
From 1997 to 1999, he was an Assistant Professor in electromagnetic elds with the
Department of Biophysical and Electronic
Engineering, University of Genoa, teaching the
university course of Electromagnetic Fields 1.
From 2001 to 2004, he was an Associate Professor with the University of Trento, Trento,
Italy. Since 2005, he has been a Full Professor
in electromagnetic elds with the University of Trento, where he currently teaches electromagnetic elds, inverse scattering techniques, antennas and wireless
communications, and optimization techniques. He is also the Director of the
ELEDIA Research Center at the University of Trento. Moreover, he is an
Adjunct Professor at Pennsylvania State University, State College, PA, USA, and
holder of a Senior DIGITEO Chair developed in co-operation between the
Laboratoire des Signaux et Systmes in Gif-sur-Yvette and the Department
Imagerie et Simulation for the Contrle of CEA LIST in Saclay (France) from
December 2014, and he has been a Visiting Professor at the Missouri University
of Science and Technology, Rolla, MO, USA, at the Nagasaki University,
Nagasaki Japan, at the University of Paris Sud, Orsay, France, and at the
Kumamoto University, Kumamoto, Japan. His research work since 1992 has
been principally on electromagnetic direct and inverse scattering, microwave
imaging, optimization techniques, wave propagation in presence of nonlinear
media, wireless communications and applications of electromagnetic elds
to telecommunications, medicine, and biology.
Prof. Massa is currently a member of the Progress in Electromagnetics
Research Symposium (PIERS) Technical Committee and of the InterUniversity Research Center for Interactions Between Electromagnetic Fields
and Biological Systems (ICEmB), and he has served as Italian representative in the general assembly of the European Microwave Association
(EuMA). He serves as an Associate Editor of the IEEE TRANSACTIONS ON
ANTENNAS AND PROPAGATION.
Giacomo Oliveri (S07; M09; SM13) received the B.S. and M.S. degrees in telecommunications engineering and the Ph.D.
degree in space sciences and engineering from
the University of Genoa, Genoa, Italy, in 2003,
2005, and 2009, respectively.
He is currently an Assistant Professor with
the Department of Information Engineering
and Computer Science, University of Trento,
where he is also a member of the ELEDIA
Research Center. In 2012, 2013, and 2015, he
was a Visiting Researcher at the Laboratoire
des signaux et systmes (L2S), cole suprieure dlectricit (SUPELEC), Paris, France. Moreover, in 2014, he was an
Invited Associate Professor at the University of Paris Sud, Paris, France. He is
an author/coauthor of over 200 peer-reviewed papers on international journals
and conferences. His research work is mainly focused on electromagnetic direct
and inverse problems, system-by-design and metamaterials, and antenna array
synthesis. He is the Chair of the IEEE AP/ED/MTT North Italy Chapter.
Dr. Oliveri serves as an Associate Editor of the International Journal
of Distributed Sensor Networks, of the International Journal of Antennas
and Propagation and of the journal Microwave Processing.
15

Compressive Sensing in Electromagnetics A Review PDF

Diunggah oleh

Informasi Dokumen

Deskripsi Asli:

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Compressive Sensing in Electromagnetics A Review PDF

Diunggah oleh

Hak Cipta:

Format Tersedia

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

1. Introduction and Scenario

he Compressive Sensing (CS) paradigm has enabled the

In addition to their fundamental advantages over Shannons

1045-9243/15/$26.00 2015 IEEE

where h; i is the inner product, and fm r; m 1; . . . ; Mg

Figure 1. Number of CS-related papers published each year

2.1 Definition of the Signal Model

whose matrix form looks as follows:

with y fym 2 R; m 1; . . . ; Mg being the vector of observations, and

is the sensing matrix.

hLfk rg; m rikn :

whose image yr is related to f r through

Moreover, let us suppose that f is S-sparse [1], with respect

Equation (8) points out two fundamental requirements of the

(x) and data (y) is linear (i.e., it can be represented by

it is mainly dictated by the physics of the problem at hand (i.e.,

2.2 Sampling and Recovery Problems

As concerns the retrieval of an S-sparse vector from a

with z 2 R M being the stochastic or deterministic unknown

Noisy CS Recovery Problem

with k k2 being the 2 -norm. Intuitively, such a condition

3.1 Deterministic CS Strategies

Since the minimization of the 0 -norm functional in (12) is an

Initialization: Initialize the guess solution (b

the 1 -regularized least-square method [34]

with  being a regularization parameter; the least absolute

Identification: Find the nth column of A,

Alternatively, convex relaxation (CR) approaches based

3.2 BCS Techniques

The diffusion and success of such a latter formulation is

and the Total Variation (TV) [36, 37]

to a Bayesian perspective resulting in the following formulation

where stands for distributed as, and Px is a suitable

with 0 being the transpose operator, and covariance

where the auxiliary parameters h;

Otherwise, the inversion of K correlated CS tasks (e.g.,

where is a hyperprior usually distributed according to the

where hn is a Gamma-distributed hyperparameter, then

with h being the shared hyper-prior, to derive the so-called

4.1 CS in Antenna Array Analysis

Table 1. List of scientic publications on CS as applied to

where dn is the nth element position (in wavelengths), and m

4.1.1 Array Diagnosis

Table 3. List of scientic publications on CS as applied to

test (AUT)] with excitation coefcients fwtn ; n 1; . . . ; N g

wtn exp j2 dn cos m zm

Table 4. List of scientic publications on CS as applied to

point method called log-barrier algorithm [13]. The proposed

4.1.2 Sparse Array Synthesis

and rearranging the differential excitation and differential eld

e feym 2 R; m 1; . . . ; 2Mg, where

Accordingly, deterministic CS strategies have been rst applied

Let us consider the Pattern Matching problem [21],

arrangements with real excitations have been synthesized

and symmetric real-valued patterns as references [i.e.,

4.2 DoA Estimation

Esk exp j2 dm cos s zkm

where Esk is the (unknown) amplitude of the sth incident wave

truncated Newton technique belonging to the gradient-based

least square (S-TLS) method has been proposed to solve the

with being a regularization parameter; the least absolute

wtn exp j2 dn cosm zm

Esk exp j2 dm coss zkm