Compressed Sensing in Radar Signal Processing
Compressed Sensing in Radar Signal Processing
Compressed Sensing in Radar Signal Processing
Learn about the most recent theoretical and practical advances in radar signal processing
using tools and techniques from compressive sensing. Providing a broad perspective
that fully demonstrates the impact of these tools, the accessible and tutorial-like
chapters cover topics such as clutter rejection, CFAR detection, adaptive beamforming,
random arrays for radar, space–time adaptive processing, and MIMO radar. Each chapter
includes coverage of theoretical principles, a detailed review of current knowledge, and
discussion of key applications, and also highlights the potential benefits of using
compressed sensing algorithms. A unified notation and numerous cross-references
between chapters make it easy to explore different topics side by side. Written by
leading experts from both academia and industry, this is the ideal text for researchers,
graduate students, and industry professionals working in signal processing and radar.
Yonina C. Eldar is a professor at the Weizmann Institute of Science. She has authored
and edited several books, including Sampling Theory: Beyond Bandlimited Systems
and Compressed Sensing: Theory and Applications (Cambridge University Press, 2015;
2012). She is a Fellow of the IEEE and EURASIP, and a member of the Israel National
Academy of Science and Humanities.
ANTONIO DE MAIO
University of Naples Federico II
YONINA C. ELDAR
Weizmann Institute of Science
ALEXANDER M. HAIMOVICH
New Jersey Institute of Technology
University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906
www.cambridge.org
Information on this title: www.cambridge.org/9781108428293
DOI: 10.1017/9781108552653
© Cambridge University Press 2020
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2020
Printed in the United Kingdom by TJ International Ltd, Padstow Cornwall
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: De Maio, Antonio, 1974– editor. | Eldar, Yonina C., editor. |
Haimovich, Alexander M., 1954– editor.
Title: Compressed sensing in radar signal processing / edited by Antonio De Maio,
University of Naples Federico II, Yonina C. Eldar, Weizmann Institute of Science,
Alexander M. Haimovich, New Jersey Institute of Technology.
Description: First edition. | Cambridge, United Kingdom ; New York, NY :
Cambridge University Press, [2020] | Includes bibliographical references and index.
Identifiers: LCCN 2019014859 | ISBN 9781108428293 (hardback)
Subjects: LCSH: Radar. | Compressed sensing (Telecommunication)
Classification: LCC TK6580 .C66 2020 | DDC 621.3848/3–dc23
LC record available at https://lccn.loc.gov/2019014859
ISBN 978-1-108-42829-3 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy
of URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
To my daughter Claudia: my light, my hope, my love – ADM
To my husband Shalomi and children Yonatan, Moriah, Tal, Noa, and Roei
for their boundless love and for filling my life with endless happiness – YE
To my students and collaborators for their contributions to my work
on radar – AH
Contents
3 RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 72
Tianyi Zhang, Jiaying Ren, Jian Li, David J. Greene, Jeremy A. Johnston, and Lam H. Nguyen
3.1 Introduction 72
3.2 RPCA for RFI Mitigation 75
3.3 CLEAN-BIC for RFI Mitigation 82
vii
viii Contents
5 Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 135
Haley H. Kim and Alexander M. Haimovich
5.1 Introduction 135
5.2 STAP Radar Concepts 137
5.3 STAP Detection Problem 145
5.4 Compressive Sensing CFAR Detection 148
5.5 Numerical Results 157
5.6 Summary 161
References 162
6 Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 165
Xiaopeng Yang, Yuze Sun, Xuchen Wu, Teng Long, and Tanpan K. Sarkar
6.1 Introduction 165
6.2 Signal Models 166
6.3 Sparsity Principle Analysis of STAP 168
6.4 Fast and Robust Sparsity-Based STAP Methods 172
6.5 Conclusions 190
References 190
9 Spectrum Sensing for Cognitive Radar via Model Sparsity Exploitation 257
Augusto Aubry, Vincenzo Carotenuto, Antonio De Maio, and Mark A. Govoni
9.1 Introduction 257
9.2 System Model and Problem Formulation 259
9.3 2-D Radio Environmental Map Recovery Strategies 263
9.4 Performance Analyses 270
9.5 Conclusions 280
References 280
11 Compressed Sensing Methods for Radar Imaging in the Presence of Phase Errors
and Moving Objects 321
Ahmed Shaharyar Khwaja, Naime Ozben Onhon, and Mujdat Cetin
11.1 Introduction and Outline of the Chapter 321
11.2 Compressed Sensing and Radar Imaging 322
x Contents
Index 355
Contributors
Laura Anitori
Netherlands Organisation for Applied Scientific Research (TNO)
Augusto Aubry
University of Naples Federico II
Vincenzo Carotenuto
University of Naples Federico II
Mujdat Cetin
University of Rochester; Sabanci University
Antonio De Maio
University of Naples Federico II
Yonina C. Eldar
Weizmann Institute of Science
David J. Greene
University of Florida
Nathan A. Goodman
University of Oklahoma
Mark A. Govoni
US Army Research Laboratory
Yujie Gu
Temple University
Alexander M. Haimovich
New Jersey Institute of Technology
xi
xii List of Contributors
Reinhard Heckel
Rice University
Jeremy A. Johnston
University of Florida
Haley H. Kim
New Jersey Institute of Technology
Bo Li
Qualcomm
Jian Li
University of Florida
Teng Long
Beijing Institute of Technology
Arian Maleki
Columbia University
Lam H. Nguyen
US Army Research Laboratory
Athina P. Petropulu
Rutgers, State University of New Jersey
Jiaying Ren
University of Florida
Tapan K. Sarkar
Syracuse University
Yuze Sun
Tsinghua University
List of Contributors xiii
Peter B. Tuuk
Georgia Tech Research Institute
Xuchen Wu
Beijing Institute of Technology
Xiaopeng Yang
Beijing Institute of Technology
Tianyi Zhang
University of Florida
Yimin D. Zhang
Temple University
Introduction
Digital signal processing (DSP) is a revolutionary paradigm shift that enables processing
of physical data in the digital domain, where design and implementation are consider-
ably simplified. The success of DSP has driven the development of sensing and pro-
cessing systems that are more robust, flexible, cheaper, and, consequently, more widely
used than their analog counterparts. As a result of this success, the amount of data gener-
ated by sensing systems has grown considerably. Furthermore, in modern applications,
signals of wider bandwidth are used in order to convey more information and to enable
high resolution in the context of imaging. Unfortunately, in many important and emerg-
ing applications, the resulting sampling rate is so high that far too many samples need to
be transmitted, stored, and processed. In addition, in applications involving very wide-
band inputs it is often very costly, and sometimes even physically impossible, to build
devices capable of acquiring samples at the necessary rate. Thus, despite extraordinary
advances in sampling theory and computational power, the acquisition and processing
of signals in application areas such as radar, wideband communications, imaging, and
medical imaging continue to pose a tremendous challenge.
Recent advances in compressed sensing (CS) and sampling theory provide a frame-
work to acquire a wide class of analog signals at rates below the Nyquist rate, and
to perform processing at this lower rate as well. Together with the theory, various
prototypes have been developed that demonstrate the feasibility of sampling and pro-
cessing signals at sub-Nyquist rates in a robust and cost-effective fashion. More specif-
ically, CS is a framework that enables acquisition and recovery of sparse vectors from
underdetermined linear systems. This research area has seen enormous growth over the
past decade and has been explored in many areas of applied mathematics, computer
science, statistics, and electrical engineering. At its core, CS enables recovery of sparse
high-dimensional vectors from highly incomplete measurements using very efficient
optimization algorithms. More specifically, consider a vector x of length n. The vector
is said to be k-sparse if it has at most k nonzero components. More generally, CS results
apply to signals that are sparse in an appropriate basis or overcomplete representation.
The main idea underlying CS is that the vector x can be recovered from measurements
y = Ax, where y is of length m n as long as A satisfies certain mathematical
properties that render it a suitable CS matrix. The number of measurements m can be
chosen on the order of k log n, which in general is much smaller than the length of
the vector x. A large body of work has been published on a variety of optimization
algorithms that can recover x efficiently and robustly when m ≈ k log n. Loosely
xiv
Introduction xv
speaking, the theory of CS deals with conditions under which the recovery of informa-
tion has vanishing or small errors. The mathematical framework of CS has inspired new
acquisition methods and new signal processing applications in a large variety of areas,
including image processing, analog to digital conversion, communication systems, and
radar processing. In many of these examples the basic ideas underlying CS need to be
extended to include, for example, continuous-time inputs, practical sampling methods,
other forms of structure on the input, computational aspects, noise affects, different
metrics for recovery performance, nonlinear acquisition methods, and more.
Two books devoted to this topic have been published recently, which focus on
many of these aspects, as well as on the underlying mathematical results [1,2]. Their
main emphasis is on the basic underlying theory and its generalizations, optimization
methods, as well as applications primarily to image processing and analog-to-digital
conversion. The latter is also covered in depth in [3].
Radar signal processing represents a fertile field for CS applications. By their very
nature, radars collect data about surveillance volumes (search radars), targets (tracking
radars), terrain and ground targets (imaging radars), or buried objects (radar tomogra-
phy). From radar’s early days in World War II, through the emergence of digital radar in
the 1970s, to today’s advanced systems, the amount of data a radar system has to handle
has increased by orders of magnitude. While early digital radars had to contend with 10s
and 100s of kbps, today’s radars may be faced with data rates in the Gbps range or more,
leading to demanding requirements in cost, hardware, data storage, and processing. The
implications of applying CS to radar are potentially enormous: sampling rates could
be lowered, the number of antenna elements in large arrays might be reduced and the
computers required to handle the data may be downsized.
This book aims to present the latest theoretical and practical advances in radar signal
processing using tools from CS. In particular, this book offers an up-to-date review of
fundamental and practical aspects of sparse reconstruction in radar and remote sensing,
demonstrating the potential benefits achievable with the CS paradigm. We take a wider
scope than previous edited books on CS-based radars: we do not restrict ourselves to
specific disciplines (such as earth observation as in [4]) or applications (such as urban
sensing as in [5]), but discuss a variety of diverse application fields, including clutter
rejection, constant false alarm rate (CFAR) processing, adaptive beamforming, random
arrays for radar, space–time adaptive processing (STAP), multiple input multiple output
(MIMO) systems, radar super-resolution, cognitive radar [6] applications involving sub-
Nyquist sampling and spectrum sensing, radio frequency interference (RFI) suppres-
sion, and synthetic aperture radar (SAR).
The book is aimed at postgraduate students, PhD students, researchers, and engi-
neers working on signal processing and its applications to radar systems, as well as
researchers in other fields seeking an understanding of the potential applications of
CS. To read and fully understand the content it is assumed that the reader has some
background in probability theory and random processes, matrix theory, linear algebra,
and optimization theory, as well as radar systems. The book is organized into eleven
chapters broadly cathegorized into five areas: sub-Nyquist radar (Chapter 1); detection,
clutter/interference mitigation, and CFAR techniques (Chapters 2–6); super-resolution
xvi Introduction
and beamforming (Chapters 7 and 8); radar spectrum sensing/sharing (Chapters 9 and
10); radar imaging (Chapter 11). Each chapter is self-contained and typically covers
three main aspects: fundamental theoretical principles, overview of the current state of
the art, and emerging/future research directions. Some chapters are also complemented
with analyses on real data. Since the chapters are independent, there is flexibility in
selecting material both for university courses and short seminars.
In Chapter 1, the authors review several sub-Nyquist pulse-Doppler radar systems
based on the Xampling framework. Contrary to other CS-based designs, their formu-
lations directly address the reduced-rate analog sampling in space and time, avoid a
prohibitive dictionary size, and are robust in the face of noise and clutter. The chapter
begins by introducing temporal sub-Nyquist processing for estimating the target loca-
tions using less bandwidth than conventional systems. This paves the way to cognitive
radars, which share their transmit spectrum with other communication services, thereby
providing a robust solution for coexistence in spectrally crowded environments. Next,
without impairing Doppler resolution, the authors reduce the dwell time by transmitting
interleaved radar pulses in a scarce manner within a coherent processing interval or
slow time. Then, they consider MIMO array radars and demonstrate spatial sub-Nyquist
processing, which allows the use of few antenna elements without degradation in
angular resolution. Finally, they demonstrate application of sub-Nyquist and cognitive
radars to imaging systems such as SAR. For each setting, the authors present a state-
of-the-art hardware prototype designed to demonstrate the real-time feasibility of
sub-Nyquist radars.
Chapter 2 discusses the problem of clutter mitigation, which has posed challenges to
radar designers and engineers since the early days of radar. Early techniques matured to
current approaches like STAP, which use a coherently processed data cube to estimate
clutter statistics and to perform adaptive filtering. This chapter examines CS techniques
for the mitigation of structured interference, such as clutter. The author first introduces
the relevant sensing model and describes results in uncompressed adaptive filtering.
This paves the way to the development of models for measurement compression of the
coherent data cube and of approaches to estimate and filter clutter from compressed
measurements. The chapter includes recent results showing how clutter second-order
statistics can be reliably estimated from compressed measurements if the clutter has
well-controlled eigenspectrum. Additionally, the covariance of the interference can be
incorporated into the CS estimation process to improve performance.
RFIs pose serious threats to the proper operations of ultra wideband (UWB) radar
systems due to severely degrading their imaging and target detection capabilities. RFI
mitigation is a challenging problem, since dynamic RFI sources utilize diverse mod-
ulation schemes, hence they are difficult to model precisely. Fortunately, RFI sources
possess certain unique properties that can be exploited for their mitigation. In Chapter 3
the authors propose several sparse signal recovery methods for effective RFI mitigation.
They first show that the RFI sources possess a low rank property and are sparse in the
frequency domain, while in contrast the desired UWB radar echoes are sparse in the time
domain. Therefore, robust principal component analysis (RPCA) can be used to simul-
taneously exploit these properties for effective RFI mitigation. RPCA, however, requires
Introduction xvii
verified that the developed methods can provide satisfactory performance with a small
training sample support in a practical complex nonhomogeneous environment.
Chapter 7 considers the use of CS techniques for the resolution of multiple targets.
Estimating the relative angles, delays, and Doppler shifts from the received signals
allows for the determination of the locations and velocities of objects. However, due to
practical constraints, the probing signals have finite bandwidth B, the received signals
are observed over a finite time interval of length T only, and in addition, a radar typically
has only one or a few transmit and receive antennas. Those constraints fundamentally
limit the resolution up to which objects can be localized: the delay and Doppler reso-
lution is proportional to 1/B and 1/T , and a radar with NT transmit and NR receive
antennas can only achieve an angular resolution proportional to 1/(NT NR ). The author
shows that the continuous angle-delay-Doppler triplets and the corresponding attenua-
tion factors can be resolved at much finer resolution, using ideas from CS. Specifically,
provided the angle-delay-Doppler triplets are separated either by factors proportional
to 1/(NT NR − 1) in angle, 1/B in delay, or 1/T in Doppler direction, they can be
recovered at significantly smaller scale or higher resolution.
Traditional adaptive beamformers are very sensitive to model mismatch, especially
when the training samples for adaptive beamformer design are contaminated by the
desired signal. In Chapter 8, the authors propose a strategy to reconstruct a signal-
free interference-plus-noise covariance matrix for adaptive beamformer design. Using
the sparsity of sources, the interference covariance matrix can be reconstructed as a
weighted sum of the tensor outer products of the interference steering vectors, and the
corresponding parameters are estimated from a sparsity-constrained covariance matrix
fitting problem. In contrast to classical CS and sparse reconstruction problems, the for-
mulated sparsity-constrained covariance matrix fitting problem can be effectively solved
by using the a priori information on array structure rather than using convex relaxation.
Simulation results demonstrate that the proposed adaptive beamformer almost always
provides near-optimal performance.
Chapter 9 deals with two-dimensional (2-D) spectrum sensing in the context of a
cognitive radar to gather real-time space–frequency electromagnetic awareness. Assum-
ing a sensor equipped with multiple receive antennas, a formal discrete-time sensing
signal model is developed, and two signal processing techniques capable of recovering
the space–frequency occupancy map via block sparsity exploitation are presented. The
former relies on the iterative adaptive algorithm (IAA) and incorporates a BIC-based
stage to foster block-sparsity in the recovery process. The latter resorts to the regularized
maximum likelihood (RML) estimation paradigm, which automatically promotes block-
sparsity in the 2-D profile evaluation. Some illustrative examples (both on simulated and
real data) are provided to compare the different strategies and highlight the effectiveness
of the developed approaches.
In Chapter 10, a cooperative spectrum-sharing scheme for a MIMO communication
system and a sparse sensing-based MIMO radar is presented. Both the radar and the
communication systems use transmit precoding. The radar transmit precoder, the radar
subsampling scheme, and the communication transmit covariance matrix are jointly
designed in order to maximize the radar SIR, while meeting certain communication
Introduction xix
rate and power constraints. The joint design is implemented at a control center, which
is a node with which both systems share physical layer information, and which also
performs data fusion for the radar. Efficient algorithms for solving the correspond-
ing optimization problem are presented. The cooperative design significantly improves
spectrum sharing performance, and the sparse sensing provides opportunities to control
interference.
Chapter 11 discusses applications of CS to radar imaging problems with reference
to SAR and inverse synthetic aperture radar (ISAR) sensors. The authors first provide
the relevant mathematical expressions for CS and SAR necessary to formulate the prob-
lem of CS SAR imaging. Thereafter, they consider the case where unknown motion
errors are present during the SAR acquisition process. Autofocusing, i.e., the blind
compensation of the aforementioned errors, is discussed, and general CS solutions are
presented. The chapter ends with a survey of CS methods for ISAR imaging of targets
with unknown motion.
References
[1] Y. C. Eldar and G. Kutyniok, Compressed Sensing: Theory and Applications. Cambridge
University Press, 2012.
[2] S. Foucart and H. Rauhut, A Mathematical Introduction to Compressive Sensing. Birkhäuser
Basel, 2013, vol. 1, no. 3.
[3] Y. C. Eldar, Sampling Theory: Beyond Bandlimited Systems. Cambridge University Press,
2015.
[4] C.-H. Chen, Compressive Sensing of Earth Observations. CRC Press, 2017.
[5] M. Amin, Compressive Sensing for Urban Radar. CRC Press, 2014.
[6] A. Farina, A. De Maio, and S. Haykin, The Impact of Cognition on Radar Technology. Scitech
Publishing, Radar, Sonar & Navigation, 2017.
Symbols
xx
List of Symbols xxi
Statement of restricted isometry property (RIP): a matrix A satisfies the RIP of order
K if
(1 − δk )x2 ≤ Ax2 ≤ (1 + δk )x2
for all x with x0 ≤ K.
1 Sub-Nyquist Radar: Principles
and Prototypes
Kumar Vijay Mishra∗ and Yonina C. Eldar∗∗
1.1 Introduction
Radar remote sensing has advanced tremendously since 1950 and is now applied to
diverse areas such as military surveillance, meteorology, geology, collision avoidance,
and imaging [1]. In monostatic pulse-Doppler radar systems, an antenna transmits a
periodic train of known narrowband pulses within a defined coherent processing interval
(CPI). When the radiated wave from the radar interacts with moving targets, the ampli-
tude, frequency, and polarization states of the scattered wave change. By monitoring this
change, it is possible to infer the targets’ size, location, and radial Doppler velocity. The
reflected signal received by the radar antenna is a linear combination of echoes from
multiple targets; each of these is an attenuated, time-delayed, and frequency-modulated
version of the transmit signal. The delay in the received signal is linearly proportional
to the target’s range or its distance from the radar. The frequency modulation encodes
the Doppler velocity of the target. The complex amplitude or target’s reflectivity is a
function of the target’s size, geometry, propagation, and scattering mechanism. Radar
signal processing is aimed at detecting the targets and estimating their parameters from
the output of this linear, time-varying system.
Traditional radar signal processing employs matched filtering (MF) or pulse com-
pression [2] in the digital domain, wherein the sampled received signal is correlated
with a replica of the transmit signal in the delay-Doppler plane. The MF maximizes
the signal-to-noise ratio (SNR) in the presence of additive white Gaussian noise. In
some specialized systems, this stage is replaced by a mismatched filter with a different
optimization metric such as minimization of peak-to-sidelobe ratio of the output. Here,
the received signal is correlated with a signal that is close but not identical to the
transmit signal [3–5]. While all of these techniques reliably estimate target parameters,
their resolution is inversely proportional to the support of the ambiguity function of the
transmit pulse, thereby restricting ability to super-resolve targets that are closely spaced.
The digital MF operation requires the signal to be sampled at or above the Nyquist
sampling rate, which guarantees perfect reconstruction of a bandlimited analog signal
[6]. Many modern radar systems use wide bandwidths, typically ranging from hundreds
∗ K.V.M. acknowledges partial support via the Andrew and Erna Finci Viterbi Postdoctoral Fellowship and
the Lady Davis Postdoctoral Fellowship.
∗∗ This work is supported by the European Union’s Horizon 2020 research and innovation program under
grant agreement no. 646804-ERC-COG-BNYQ.
1
2 Mishra and Eldar
of MHz to GHz, in order to achieve fine radar range resolution. Since the Nyquist sam-
pling rate is twice the baseband bandwidth, the radar receiver requires expensive, high-
rate analog-to-digital converters (ADCs). The sampled signal is then also processed at
high rates, resulting in significant power, cost, storage, and computational overhead.
Recently, in order to mitigate this rate bottleneck, new methods have been proposed that
sample signals at sub-Nyquist rates and yet are able to estimate the targets’ parameters
[6,7].
Analogous trade-offs arise in other aspects of radar system design. For example, the
number of transmit pulses governs the resolution in Doppler velocity. The estimation
accuracy of target parameters is greatly affected by the radar’s dwell time [1], i.e., the
time duration a directional radar beam spends hitting a particular target. Long dwell
times imply a large number of transmit pulses and, therefore, high Doppler precision.
But, simultaneously, this negatively affects the ability of the radar to look at targets in
other directions. Sub-Nyquist sampling approaches have, therefore, been suggested for
the pulse dimension or “slow-time” domain in order to break the link between dwell
time and Doppler resolution [8–10].
Finally, radars that deploy antenna arrays deal with similar sampling problems in the
spatial domain. A phased array radar antenna consists of several radiating elements
that form a highly directional radiating beam pattern. Without requiring any mechan-
ical motion, a phased array accomplishes beam-steering electronically by adjusting
the relative phase of excitation in the array elements. The operational advantage is
the agile scanning of the target scene, ability to track a large number of targets, and
efficient search-and-track in the regions of interest [11]. The beam pattern of individual
array elements, array geometry, and its size define the overall antenna pattern [12,13],
wherein high spatial resolution is achieved by large array apertures. As per the Nyquist
Theorem, the array must not admit fewer than two signal samples per spatial period (i.e.,
radar’s operating wavelength) [14]. Otherwise, it introduces spatial aliasing or multiple
beams in the antenna pattern, thereby reducing its directivity. Often an exceedingly
large number of radiating elements are required to synthesize a given array aperture in
order to enhance the radar’s ability to unambiguously distinguish closely spaced targets;
the associated cost, weight, and area may be unacceptable. It is therefore desirable to
apply sub-Nyquist techniques to thin a huge array without causing degradation in spatial
resolution [15–17].
Sub-Nyquist sampling leads to the development of low-cost, power-efficient, and
small-size radar systems that can scan faster and acquire larger volumes than traditional
systems. Apart from design benefits, other applications of such systems have been
envisioned recently, including imparting hardware-feasible cognitive abilities to the
radar [18,19], a role in devising spectrally coexistent systems [20], and extension to
imaging [21]. In this chapter, we provide an overview of sub-Nyquist radars, their
applications, and hardware realizations.
The outline of the chapter is as follows. In the next section, we overview various
reduced-rate techniques for radar system design and explain the benefits of our approach
to sub-Nyquist radars. In Section 1.3, we describe the principles, algorithms, and hard-
ware realization of temporal sub-Nyquist monostatic pulse-Doppler radar. Section 1.4
Sub-Nyquist Radar: Principles and Prototypes 3
presents an extension of the sub-Nyquist principle to slow time. We then introduce the
cognitive radar concept based on sub-Nyquist reception in Section 1.5 and show an
application to coexistence in a spectrally crowded environment. Section 1.6 is devoted to
spatial sub-Nyquist applications in multiple-input multiple-output (MIMO) array radars.
Finally, we consider sub-Nyquist synthetic aperture radar (SAR) imaging in Section 1.7,
followed by concluding remarks in Section 1.8.
There is a large body of literature on reduced-rate sampling techniques for radars. Most
of these works employ compressed sensing (CS) methods, which allow recovery of
sparse, undersampled signals from random linear measurements [7]. A pre-2010 review
of selected applications of CS-based radars can be found in [22]. A qualitative, system-
level commentary from the point of view of operational radar engineers is available
in [23], while CS-based radar imaging studies are summarized in [24]. An excellent
overview on sparsity-based SAR imaging methods is provided in [25]. The review
in [26] recaps major developments in this area from a nonmathematical perspective.
In the following, we review the most significant works relevant to the sub-Nyquist
formulations presented in this chapter.
On-Grid CS The earliest application of CS toward recovering time delays with sub-
Nyquist samples in a noiseless case was formulated in [27]. CS-based parameter esti-
mation for both delay and Doppler shifts was proposed in [28] with samples acquired
at the Nyquist rate. These and similar later works [29–31] discretize the delay-Doppler
domain, assuming that targets lie on a grid. Subsequently, these ideas were extended
to colocated [32,33] and distributed [34] MIMO radars where targets are located on an
angle-Doppler-range grid. In practice, target parameters are typically continuous values
whose discretization may introduce gridding errors [35]. In particular, [28] constructs
a dictionary that exhaustively considers all possible delay-Doppler pairs, thereby ren-
dering the processing computationally expensive. Noise and clutter mitigation are not
considered in this literature. Simulations show that such systems typically have poor
performance in clutter-contaminated noisy environments.
Off-Grid CS A few recent works [36,37] formulate the radar parameter estimation for
off-grid targets using atomic norm minimization [38,39]. However, these methods do
not address direct analog sampling, the presence of noise, and clutter. Further details
on this approach are available in Chapter 7 (Super-resolution radar imaging via convex
optimization) of this book.
Parametric Recovery A different approach was suggested in [40], which treated radar
parameter estimation as the identification of an underlying linear, time-varying system
[41]. The proposed two-stage recovery algorithm, largely based on [42], first estimates
target delays and then utilizes these recovered delays to estimate Doppler velocities and
complex reflectivities. They also provide guarantees for system identification in terms
4 Mishra and Eldar
of the minimum time-bandwidth product of the input signal. However, this method does
not handle noise well.
Matrix Completion In some radar applications, the received signal samples are pro-
cessed as data matrices, which, under certain conditions, are low rank. In this context,
general works have suggested retrieving the missing entries using matrix completion
methods [10,43]. The target parameters are then recovered through classic radar signal
processing. These techniques have not been exhaustively evaluated for different signal
scenarios and their practical implementations have still not been thoroughly examined.
consisting of P uniformly spaced known pulses h(t). The interpulse transmit delay τ
is the pulse repetition interval (PRI) or fast time; its reciprocal is the pulse repetition
frequency (PRF). The entire duration of P pulses in (1.1) is known as the CPI or slow
time. The radar operates at carrier frequency fc so that its wavelength is λ = c/fc ,
where c = 3 × 108 m/s is the speed of light.
In a conventional pulse-Doppler radar, the pulse h(t) = hNyq (t) is a time-limited
baseband
∞ function whose continuous-time Fourier transform (CTFT) is HNyq (f ) =
h (t)e −j 2πf t dt. It is assumed that most of the signal’s energy lies within the
−∞ Nyq
frequencies ±Bh /2, where Bh denotes the effective signal bandwidth, such that the
following approximation holds:
h /2
B
proportional to the target’s radial velocity vl = cνl /4πfc ; and complex amplitude αl .
The target locations are defined with respect to the polar coordinate system of the radar
and their range and Doppler are assumed to lie in the unambiguous time-frequency
region, i.e., the time delays are no longer than the PRI, and Doppler frequencies are up
to the PRF.
Typically, the radar assumes the following operating conditions, which leads to a
simplified expression for the received signal [49]:
A1 “Far targets”: assuming νl 2πfc τl /P τ, target radar distance is large compared
to the distance change during the CPI over which attenuation α l is allowed to be
constant.
A2 “Slow targets”: assuming νl 2πfc /P τBh , target velocity is small enough
to allow for constant τl during the CPI. In this case, the following piecewise-
constant approximation holds νl t ≈ νl pτ, for t ∈ [pτ,(p + 1)τ].
A3 “Small acceleration”: assuming dνl /dt c/2fc (P τ)2 , target velocity remains
approximately constant during the CPI allowing for constant νl .
A4 “No time or Doppler ambiguities”: The delay-Doppler pairs (τl ,νl ) for all l ∈
[1,L] lie in the radar’s unambiguous region of delay-Doppler plane defined by
[0,τ] × [−π/τ,π/τ].
A5 The pairs in the set (τl ,νl ) for all l ∈ [1,L] are unique.
Under these assumptions, the received signal can be written as
P −1 L−1
rRX (t) = α l h(t − τl − pτ)e−j νl t + w(t), (1.4)
p=0 l=0
for 0 ≤ t ≤ P τ, where w(t) is a zero mean wide-sense stationary random signal with
autocorrelation rw (s) = σ 2 δ(s). It is convenient to express rRX (t) as a sum of single
frames
P −1
p
rRX (t) = rRX (t) + w(t), (1.5)
p=0
where
L−1
α l h(t − τl − pτ)e−j νl pτ,
p
rRX (t) = (1.6)
l=0
Figure 1.1 Sub-Nyquist sampling methods: (a) direct sampling; (b) low frequencies only;
(c) multiband sampling.
There are several ways to implement a sub-Nyquist sampler [47,61] in order to obtain
a set of Fourier coefficients from low-rate samples of the signal. For simplicity, consider
|κ| = K such that q = N/K is an integer defining the sampling reduction factor.
In direct sampling (Figure 1.1a), the signal rRx (t) obtained after the anti-aliasing filter
is passed through as many analog chains as the number of sub-Nyquist coefficients
K. Each branch is modulated by a complex exponential, followed by integration over
τ and necessary digital signal processing (DSP). This technique provides the largest
flexibility in choosing the Fourier coefficients, but is expensive in terms of hardware.
Another approach is to limit the bandwidth of the anti-aliasing filter such that only the
lowest K frequencies are free of aliasing (Figure 1.1b). We then sample these lowest
K frequencies. Here, the measurements are correlated and a modification in the analog
hardware is also required so that the anti-aliasing filter has reduced passband. In the
multiband sampling method shown in Figure 1.1c, M disjoint randomly chosen groups
of consecutive Fourier coefficients are obtained such that the total number of sampled
coefficients is K. This translates to splitting the signal across M branches, passing the
downconverted signal through reduced-passband anti-aliasing filters, and then sampling
each band with a low-rate ADC. This method can be easily implemented but requires
M low-rate ADCs. The sub-Nyquist hardware prototypes developed in [45,49] adopt
multiband sampling using four groups of consecutive coefficients. In practice, the spe-
cific Fourier coefficients are chosen through extensive software simulations to provide
low mutual coherence [6] for CS-based signal recovery.
Our goal now is to recover {α l ,τl ,νl }L−1
l=0 from cp [k] for k ∈ κ and 0 ≤ p ≤ P − 1.
To that end, [49] adopts the Doppler focusing approach. Consider the DFT of the coef-
ficients cp [k] in the slow-time domain:
P −1
L−1
P −1
1
˜ ν [k] =
cp [k]ej ν pτ = H [k] α l e−j 2πkτl /τ ej (ν −νl )pτ . (1.8)
τ
p=0 l=0 p=0
Sub-Nyquist Radar: Principles and Prototypes 9
Figure 1.2 Sum of exponents |g(ν|νl )| for P = 200, τ = 1 s, and νl = 0 [20,49]. ©2018 IEEE.
Reprinted, with permission, from [20].
depends only on the number of targets present and not on Bh . This shows that a sub-
Nyquist radar breaks the link between range resolution and transmit bandwidth. In gen-
eral, only a few targets are present in the radar coverage region leading to a significant
reduction in sampling rate.
theorem 1.3.1 [49] The minimal number of samples required for perfect recovery of
{αl ,τl ,νl }L 2
l=0 in a noiseless environment is 4L . In addition, the number of samples per
period is at least 2L, and the number of periods P ≥ 2L.
PT
ρ= 2 , γ = Q−12 (1 − N 1 − Pfa )
σ |FR | χ 2 (ρ)
intensively. In the context of CS-based radars, [63] provides a general overview of clut-
ter rejection algorithms. In [64], Capon beamforming is used to reject clutter and then
the target is retrieved by exploiting sparse reconstruction methods. On the other hand, a
few works such as [65–67] utilize sparsity of the clutter in the mitigation process. Along
similar lines, [68] assumes sparse clutter and proposes a GLRT detector. However, they
obtain signal samples at the Nyquist rate.
Conventionally, clutter is modeled as a random process with Doppler frequency that
follows a colored Gaussian noise distribution [69–71]. A standard operation to remove
this correlated noise is to use receive filters that maximize the signal-to-clutter-plus-
noise (SCNR) ratio. This method is equivalent to first whitening the received signal
samples, and then performing matched-filtering with respect to a whitened pulse. Our
approach [52] to clutter removal in sub-Nyquist radar is based on this philosophy as it
fits well with our Fourier-based analysis.
In the presence of clutter and noise, the received signal rq (t) is
where rRX (t) is the target signal with noise as in (1.5) and
P −1
C
y(t) = α c h(t − pτ − τc )ej vc pτ, (1.15)
p=0 c=1
is the echo from C clutter targets. We assume that the mean clutter amplitude is
E[|αc |2 ] = σc2 . Further, the delays τc ∼ U (0,τ) and the clutter Doppler spectrum
vc ∼ N (vd ,σd2 ) are independent and identically distributed.
Analogous to the target signal in (1.7), the Fourier series representation of the clutter
signal is given by
1 C−1
2π
c̃p [k] = H [k] α c e−j τ kτc e−j vc pτ . (1.16)
τ
c=0
Let the Fourier series coefficients of the noise be w̃p [k]. We now form a P × K matrix
R with kth column given by the Fourier coefficients Rp [k] = cp [k] + c̃p [k] + w̃p [k],
k ∈ κ such that
N H + B,
R = X + Y + N = FP AFK (1.17)
2π
where B = Y + N, FP is the P × P Fourier matrix with (l,k)th element e−j P lk , FK
N is
2π
a submatrix formed by K rows of the N × N Fourier matrix with (l,k)th element ej N lk ,
H = diag(H [k]) is a K × K diagonal matrix, A is a P × K sparse matrix with complex
reflectivity α l at the L indices (rl ,sl ), and Y and N are P × K matrices with (p,k)th
elements c̃p [k] and w̃p [k], respectively. As mentioned previously, noise is white over
the indices k (all tones).
Our goal is to extract A from the measurements R. For simplicity, we assume that
|H [k]|2 is unity for all k. The whitening transformation requires information about the
statistics of clutter and noise, which are summarized in the following proposition.
Sub-Nyquist Radar: Principles and Prototypes 13
proposition 1.3.3 [52] The mean of the clutter Fourier coefficients is E[c̃p [k]] = 0,
and their correlation is given by
1 2 2 2
Rl1 [k1,k2 ] = E c̃p [k1 ]c̃p+l1 [k2 ] = Cσc2 δk1,k2 e−j vd l1 τ− 2 σd l1 τ . (1.18)
The mean and variance of the Fourier coefficients of the noise are, respectively,
1
E Np [k] = 0, E Np [k1 ]Np+l1 [k2 ] = σn2 δk1 k2 δl2 . (1.19)
τ
Our clutter mitigation technique is based on whitening all the tones of the measure-
ments R. It follows from Proposition 1.3.3 that the columns of B are uncorrelated and
identically distributed. The covariance matrix M of the columns of B is a Toeplitz matrix
with mth diagonal value
1 2 1 2
M(m) = Cσc2 e−j vd mτ− 2 σd m
2 τ2
+ σ δm . (1.20)
τ n
Therefore, the columns of R can be whitened by multiplying on the left by M−1/2 :
where M−1/2 B corresponds to white noise. From here, we proceed with Doppler focus-
ing on M−1/2 R by taking a Hermitian transpose of (1.21) and multiplying on the right
by M−1/2 FP :
= H(FK H H H −1 H −1
N ) A FP M FP + B M FP = + W, (1.22)
where W is white noise for each focused frequency. This equation represents a sparse
matrix recovery problem. For known matrices D1 = H(FK H −1
N ) and D2 = FP M FP , we
H
are given measurements = D1 XD2 , and the goal is to retrieve the sparse matrix X =
AH . These problems are solved by matrix-sketching algorithms, as described in [72]. It
has been shown in [52] that whitened Doppler focusing generally increases the SCNR.
Compared to other CS-based radars [27,28], this technique is robust to the presence of
clutter, despite sampling at low rates.
Figure 1.3 The four-channel Xampler board [45]. ©2014 IEEE. Reprinted, with permission,
from [45].
Figure 1.4 Sub-Nyquist hardware prototype showing connections between the Xampler board and
NI chassis [8,45,49].
Figure 1.5 NI chassis showing various signal generation and synchronization components.
a local oscillator (LO), which modulates the desired frequency band of the channel to
the central frequency of a narrow 80 KHz bandwidth band pass filter (BPF). A fifth
LO, common to all 4 channels, modulates the BPF output to a low-frequency band,
and sampled with a standard low-rate ADC operating at 250 kHz frequency. The digital
samples are acquired by the chassis controller and a MATLAB function is launched that
runs Doppler focusing. The digital reconstruction algorithm, performed at a low rate of
250 ksps, allows recovery of the unknown delays and Doppler frequencies of the targets.
A block diagram of the system is shown in Figure 1.6.
Sub-Nyquist Radar: Principles and Prototypes 15
fLO1 28.915MHz
1Æ4 Splitter
14dB 14dB
Df3dB = 80KHz fs = 125KHz
fLO3 Digital Processing
Input Signal LPF 14dB 28.915MHz
f3dB = 2.5MHz Crystal Filter LPF
14dB fc = 29MHz 14dB fp = 100KHz
Df3dB = 80KHz fs = 125KHz
fLO4 28.915MHz
Crystal Filter LPF
14dB
fc = 29MHz 14dB fp = 100KHz
Df3dB = 80KHz fs = 125KHz
Figure 1.6 Block diagram of a 4-channel solid-state receiver with 4 up-modulating local
oscillators with respective center frequencies of 28.375, 28.275, 27.65, and 27.391 MHz [45].
©2014 IEEE. Reprinted, with permission, from [45]
Figure 1.7 Sub-Nyquist prototype experiment [52]. Top left to right: Signal corresponding to
targets, clutter, noise, and all three combined. Bottom left to right: Low rate samples at receiver
and delay-Doppler map with true and recovered targets.
Real-time analog experiments show that the system is able to maintain good detection
capabilities, while sampling radar signals that require Nyquist rate of about 30 MHz at a
total rate of 1 MHz, i.e., 1/30th of the Nyquist rate. We conducted several experiments
in order to test the accuracy of our system under various conditions. For example,
Figure 1.7 shows results for a hardware experiment where the target scene has seven
scatterers with different delays and Doppler frequencies. A few cases of closely spaced
targets in the delay-Doppler plane are also included. Clutter is also added to the scene
and identified by the system. Our low-rate processing rejects the clutter and successfully
detects only targets in the delay-Doppler plane despite sampling at 1/30th of the Nyquist
rate. The digital recovery algorithm is efficient as it involves only solving 1D delay
recovery problems post FFT-based Doppler focusing and without increasing the size of
the dictionary.
The temporal sub-Nyquist processing in the fast-time domain described in the previous
section breaks the link between signal bandwidth, sampling rate, and range resolution.
16 Mishra and Eldar
2 −1
P
rTX (t) = h(t − mp τ), 0 ≤ t ≤ P1 τ. (1.23)
p=0
The received signal rRX (t) is accordingly expressed as a sum of single frames
2 −1
P
p
rRX (t) = rRX (t), (1.24)
p=0
where
L−1
α l h(t − τl − mp τ)e−j νl mp τ,
p
rRX (t) = (1.25)
l=0
for 0 ≤ t ≤ P1 τ, is the return signal from the pth pulse. Our goal is to recover the
p
targets range and Doppler frequency from the received signals rRX (t), with a reduced
number of transmit pulses P2 < P1 as well as low-rate samples per pulse.
1
L−1
Xp [k] = H [k] α l e−j 2πkτl /τ e−j νl mp τ, 0 ≤ k ≤ N − 1, (1.26)
τ
l=0
where N = Bh τ. From (1.26), the Fourier coefficients embody all the information about
the unknown parameters {α l ,τl ,νl }L−1
l−0 . The goal is then to recover these parameters
from Xp [k], 0 ≤ p ≤ P2 − 1. The low rate sampling technique is as described earlier
in Section 1.3.2, but the processing steps to recover the target parameters are different
Sub-Nyquist Radar: Principles and Prototypes 17
to account for sub-Nyquist sampling in Doppler. Let X be the K × P matrix with pth
column given by the Fourier coefficients Xp [k], k ∈ κ. Then X can be expressed as
P2 T
X = HFK
N A(FP1 ) , (1.27)
P2
where H = τ1 diag(H [k]), FKN is a K × N partial Fourier matrix, FP1 is a P2 × P1 partial
Fourier matrix indexed by the values of mp , 1 ≤ p ≤ P2 , and A is an N × P1 sparse
matrix with α l values at the L indices {sl ,τl }. We would like to recover A from the
measurements X.
The system of linear equations (1.27) can be solved by CS techniques. However, this
problem is different than the temporal sub-Nyquist formulation of Section 1.3.2, where
only the range sensing matrix FK N is a partial DFT. In Doppler sub-Nyquist radar, both
P2
range and Doppler sensing matrices (i.e., FK N and FP1 , respectively) are partial DFTs.
Analogous to Theorem 1.3.1, we have the following result for the Doppler sub-Nyquist
radar:
theorem 1.4.1 [8] The minimal number of samples required for perfect recovery of A
for L targets in noiseless settings is 4L2 . In addition, the number of samples per period
is at least 2L, and the number of periods P2 ≥ 2L.
Note that the number of periods P2 here is for nonuniform transmission while the
minimum number of periods in Theorem 1.3.1 pertain to uniformly spaced pulses in the
CPI. Theorem 1.4.1 indicates the lower limit of rate reduction in temporal and Doppler
domains.
To solve for the sparse matrix A in (1.27) one can use the matrix version of OMP or
1 minimization [6]. Alternatively, Doppler focusing is still approximately applicable.
The nonuniform discrete Fourier transform of the coefficients Xp [k] is
2 −1
P
L−1 2 −1
P
j ν mp τ 1 −j 2πkτl /τ
ν [k] = Xp [k]e = H [k] αl e ej (ν −νl )mp τ . (1.28)
τ
p=0 l=0 p=0
This can be approximated similar to (1.9) and solved by Algorithm 1, described earlier.
However, this is a poor approximation because the P2 points in the sum of exponents
P2 −1 j (ν −ν )m τ
p are not equally spaced over the unit circle.
p=0 e
l
Figure 1.8 RToT sub-Nyquist radar prototype [8]. Top left: Targets at two different azimuths.
Bottom left: Echoes from both directions are acquired via nonuniform pulses. Top and bottom
right: delay-Doppler maps showing reconstruction for both directions.
In the previous two sections, we focused on processing the received signal. The receiver
design in the sub-Nyquist framework can be exploited to also alter the behavior of
the radar transmitter. In this section, we discuss the opportunistic control of the trans-
mitter to impart cognition to the radar and leverage it for spectrum sharing applica-
tions. For alternative, non-sub-Nyquist approaches to cognition in radars, we refer the
reader to Chapters 9 (“Spectrum sensing for cognitive radar via model sparsity exploita-
tion”) and 10 (“Cooperative spectrum sharing between sparse sensing based radar”) of
this book.
The unhindered operation of a radar that shares its spectrum with communication
systems has captured a great deal of attention within the operational radar community
during the last decade [73]. The interest in such spectrum-sharing radars is largely due
to the electromagnetic spectrum being a scarce resource and almost all services having
a need for greater access to it.
Recent research in spectrum sharing radars has focused on S- and C-bands, where
the spectrum has seen increasing cohabitation by long-term evolution (LTE) cellu-
lar/wireless commercial communication systems. Many synergistic efforts by major
agencies are underway for efficient radio spectrum utilization. A significant recent
development is the announcement of the Shared Spectrum Access for Radar and
Communications (SSPARC) program [74] by the Defense Advanced Research Projects
Agency (DARPA). This program is focused on S-band military radars and views
spectrum sharing as a cooperative arrangement where the radar and communication
services actively exchange information. It defines spectral coexistence as equipping
existing radar systems with spectrum sharing capabilities and spectral codesign as
developing new systems that utilize opportunistic spectrum access [75]. For a review of
spectral interference from different services at IEEE radar bands, see [20].
A variety of system architectures have been proposed for spectrum-sharing radars.
Most put emphasis on optimizing the performance of either radar or communications
Sub-Nyquist Radar: Principles and Prototypes 19
while ignoring the performance of the other. The radar-centric architectures [20,76]
usually assume fixed interference levels from communication systems and design the
system for high probability of detection (Pd ). Similarly, the communications-centric
systems attempt to improve performance metrics, like the error vector magnitude and
bit/symbol error rate for interference from radar. With the introduction of the SSPARC
program, joint radar-communication performance is being investigated [77]. In nearly
all cases, real-time exchange of information between radar and communications hard-
ware has not yet been integrated into the system architectures. In a similar vein, our
proposed method, described later in this section, incorporates handshaking of spectral
information between the two systems.
Conventional receiver processing techniques to remove RF interference in radar
employ notch filters at hostile frequencies. Typically, spectrum sharing is achieved by
notching out the radar waveform’s bandwidth, causing a decrease in range resolution.
Our spectrum sharing solution departs from this baseline. The approach we adopt
follows the Xampling architecture on which the sub-Nyquist radar prototype described
earlier in Section 1.3 is based. We recall that the sub-Nyquist receiver samples and
processes only small narrow subbands of the received signal. Hence, we capitalize on
the simple observation that if only narrow spectral bands are sampled and processed,
then one can restrict the transmit signal to these bands. The concept of transmitting
only a few subbands that the receiver processes is one way to formulate a cognitive
radar (CRr) [60]. The delay-Doppler recovery is then performed as presented earlier
in Section 1.3. The range resolution obtained through this multiband signal spectrum
fragmentation can be the same as that of a wideband traditional radar. Furthermore, by
concentrating all the available power in the transmitted narrow bands rather than over a
wide bandwidth, the CRr increases SNR, as illustrated in Figure 1.9.
In the CRr system [60], the support of subbands varies with time to allow for
dynamic and flexible adaptation to the environment. Such a system also enables the
radar to disguise the transmitted signal as an electronic countermeasure or cope with
crowded spectrum by using a smaller interference-free portion. The CRr configuration is
key to spectrum sharing since the radar transceiver adapts its transmission to available
bands, achieving coexistence with communication signals. To detect vacant bands,
a communication receiver is needed that performs spectrum sensing over a large
Figure 1.9 A conventional radar with bandwidth Bh transmits in the band Bh . A cognitive radar
Nb
transmits only in subbands {Bi }i=1 , but with increased in-band power. The sub-Nyquist receiver
samples and processes only these subbands [19].
20 Mishra and Eldar
− − − − 0
2 2
PC–Matlab Based Controller Mixing Series Generator PC - Labview + Matlab Based Controller
Generates RF Input in Real Time
pi(t)
yi(t)
i = 1..3
XIUNX VC707 – HighSpeed FPGA
NI PXIe-1065 with DC Coupled 4-Channel ADC
PCIe ×4 to MXI ×4
x(t)
Signal Generators RF Signal
Signal ADC
2× NI° USRP-2942R RF Generator The Cog-Radio Card *Note: Allows real-time sampling
Let FC ⊂ F be the unknown support of xC (t). The goal of the CRo communication
receiver is to retrieve FC , while sampling and processing xC (t) at low rates in order
to reduce system cost and resources. A CRo system was developed earlier [56,80] for
blind sensing (see Figure 1.11). Next, we explain the details on combining this system
with the sub-Nyquist radar to implement SpeCX.
The input signal at the communication receiver of the SpeCX system is
where xR (t) = rTX (t) + rRX (t) is the radar signal sensed by the communication receiver,
composed of the transmitted and received radar signals defined in (1.1) and (1.4), respec-
tively. Since the frequency support of xC (t) is unknown, a classic processor would
sample such a signal at its Nyquist rate, which can be prohibitively high. In this work,
we instead use the modulated wideband converter (MWC) [56], a sub-Nyquist sampling
technique that achieves the lower sampling rate bound for perfect blind recovery of
multiband signals, namely twice the Landau rate, and is also practically feasible. The
MWC is composed of M parallel channels. In each channel, an analog mixing front
end, where xC (t) is multiplied by a mixing function pi (t), aliases the spectrum, such
that each band appears in baseband. The mixing functions pi (t) are periodic with period
Tp such that fp = 1/Tp ≥ B and have thus the following Fourier expansion:
∞
j T2π lt
pi (t) = cil e p . (1.31)
l=−∞
In each channel, the signal next goes through a lowpass filter (LPF) with cut-off
frequency is sampled at rate fs ≥ fp , resulting in samples zi [n]. Define
fs /2 and
fNyq +fs
N = 2 2fp and Fs = [−fs /2,fs /2]. Following the calculations in [56], the
relation between the known discrete-time Fourier transform of the samples zi [n] and
the unknown XC (f ) is given by
where z(f ) is a vector of length M with ith element zi (f ) = Zi (ej 2πf Ts ) and the
unknown vector xC (f ) is given by
xC i (f ) = XC (f + (i − N/2)fp ), f ∈ Fs , (1.33)
22 Mishra and Eldar
Figure 1.12 Schematic implementation of the MWC analog sampling front end and digital signal
recovery from low-rate samples [20]. The CRo inputs are the communication signal xC (t) and
radar support FR . The communication support output FC is shared with the radar transmitter.
where
Q= z(f )zH (f )df , Z = x(f )xH (f )df , (1.35)
f ∈Fp f ∈Fp
V = A(UC + UR ). (1.36)
The support of the unique sparsest solution of (1.36) is the same as the support of our
original set of equations (1.32) [56]. Therefore, the support of UC and UR are disjointed.
The frequency support FR of xR (t) is known at the communication receiver. From
FR , we derive the support SR of the radar slices xR (f ), which is identical to the support
of UR , such that
fRi fs + BRi
SR = n n − − N/2 < , (1.37)
fp 2fp
Sub-Nyquist Radar: Principles and Prototypes 23
for 1 ≤ i ≤ Nb . Our goal can then be stated as recovering the support of UC from V,
given the known support SR of UR . This can be formulated as a sparse recovery with
partial support knowledge, studied under the framework of modified CS [81]. Modified-
CS has been used to adapt CS recovery algorithms to exploit partial known support. In
particular, greedy algorithms, such as OMP, have been modified to OMP with partial
known support [82]. Instead of starting with an initial empty support set, one starts with
SR as the initial support. In the first iteration, we compute the estimate
and residual
V1 = V − ASR Û1 . (1.39)
and does not change during operation. The location and extent of the bands Bri and fri
are determined by the radar transmitter through an optimization procedure to identify
the least contaminated bands (see Section 1.5.2). The resulting transmitted radar signal
CTFT is
where FRi = [fri − Bri /2,fri + Bri /2] is the set of frequencies in the ith band, such that
b i
FR = N i=1 FR . The parameters β i > 1 are chosen such that the total transmit power
PT of the spectrum-sharing radar remains the same as that of the conventional radar:
Bh /2 Nb
|HNyq (f )| df = 2
|HR (f )|2 df = PT . (1.43)
−Bh /2 i=1
Fri
24 Mishra and Eldar
Delay-Doppler Recovery
In order to recover the delay-Doppler map from only Nb transmitted narrow bands,
CRr employs a sub-Nyquist receiver that we explained earlier in Section 1.3.2. The
radar receiver first filters the CRr subbands supported on FR and computes the Fourier
coefficients of the received signal. Our resulting spectrum sharing SpeCX framework is
summarized in Algorithm 3.
Sub-Nyquist Radar: Principles and Prototypes 25
For time-delay estimation, [19] compares the performance of conventional and cog-
nitive radars using the extended Ziv–Zakai lower bound (EZB). In a conventional radar,
the EZB for a single target delay estimate τˆ0 is
3/2 SN R
SN R 4
EZBR ( τˆ0 ) = στ20 · 2Q + 2
, (1.45)
2 SN R · F
where Q(·) denotes the right tail Gaussian probability function, a (b) is the incomplete
gamma function with parameter a and upper limit b, and F is the root-mean-square
26 Mishra and Eldar
(rms) bandwidth of the full-band signal. The bound for CRr is given in the following
theorem.
theorem 1.5.1 [19] The extended Ziv–Zakai lower bound (EZB) for delay estimation
in a cognitive radar is
SN R
⎛ ⎞ 3/2
SN R 4
EZBCRr ( τ̂0 ) = στ20 · 2Q ⎝ ⎠+ , (1.46)
2
Nb
SN Ri · Fi2
i=1
where SN Ri and Fi are the in-band SNR and rms bandwidth of the ith subband and
SN R is the total SNR.
b i
As noted in [19], since N
i=1 Br ⊂ Bh , we have SN R > SN R for given PT . There-
fore, the SNR threshold for asymptotic performance of EZBCRr is lower than EZBR .
As the noise increases and power remains constant for both radars, the asymptotic
performance of EZBCRr is more tolerant to noise than EZBR .
The multiband design strategy, besides allowing a dynamic form of the transmitted
signal spectrum over only a small portion of the whole bandwidth to enable spectrum
sharing, has two additional advantages. First, as we show in hardware experiments
(Section 1.5.3), our CS reconstruction achieves the same resolution as traditional
Nyquist processing over a significantly smaller bandwidth. Second, the entire transmit
power is concentrated in small narrow bands. Therefore, the SNR in the sampled bands
is improved, which leads to better parameter estimation, as indicated by Theorem 1.5.1.
Comm Display
Comm Analog Rx
Comm Digital Rx
Radar Display
Figure 1.13 Shared spectrum prototype [20]. The system is composed of a signal generator,
a CRo receiver based on the MWC, a communication digital receiver, and a CRr analog and
digital receiver.
Sub-Nyquist Radar: Principles and Prototypes 27
analog front-end receiver. The card first splits the wideband signal into M = 4 hardware
channels, with an expansion factor of q = 5, yielding Mq = 20 virtual channels after
digital expansion. In each channel, the signal is then mixed with a periodic sequence
pi (t), generated on a dedicated FPGA, with fp = 20 MHz. The sequences are chosen as
truncated versions of Gold Codes. These were heuristically found to give good detection
results [85], primarily due to small bounded cross-correlations within a set.
Next, the modulated signal passes through a Chebyshev LPF of 7th order with a cut
off frequency (−3 dB) of 50 MHz. Finally, the low-rate analog signal is sampled by
a National Instruments ADC operating at fs = (q + 1)fp = 120 MHz, leading to
a total sampling rate of 480 MHz. The digital receiver is implemented on a National
Instruments PXIe-1065 computer with DC-coupled ADC. Since the digital processing
is performed at the low rate 120 MHz, very low computational load is required in order
to achieve real time recovery. MATLAB and LabVIEW platforms are used for digital
recovery operations.
The prototype is fed with RF signals composed of up to Nsig = 5 real communication
transmissions, namely K = 10 spectral bands with total bandwidth occupancy of up to
200 MHz and varying support, with Nyquist rate of 6 GHz. To test the system’s support
recovery capabilities, an RF input is generated using vector signal generators, each
producing a modulated data channel with individual bandwidth of up to 20 MHz, and
carrier frequencies ranging from 250 MHz up to 3.1 GHz. The input transmissions then
go through an RF combiner, resulting in a dynamic multiband input signal that enables
fast carrier switching for each of the bands. This input is specially designed to allow
testing the system’s ability to rapidly sense the input spectrum and adapt to changes, as
required by modern CRo and shared spectrum standards, e.g., in the SSPARC program.
The system’s effective sampling rate, equal to 480 MHz, is only 8% of the Nyquist
rate and 2.4 times the Landau rate. The main advantage of the Xampling framework,
demonstrated here, is that sensing is performed in real-time from sub-Nyquist samples
for the entire spectral range.
Support recovery is digitally performed on the low rate samples. The prototype
successfully recovers the support of the CRo transmitted bands, as demonstrated in
Figure 1.14. The signal is then reconstructed in real-time. Reconstruction does not
Figure 1.14 SpeCX communication system display [20] showing (a) low rate samples acquired
from one MWC channel at rate 120 MHz, and (b) digital reconstruction of the entire spectrum
from sub-Nyquist samples.
28 Mishra and Eldar
require interpolation to the Nyquist rate and the active transmissions are recovered at
the low rate of 20 MHz, corresponding to the bandwidth of the slices z(f ) defined
in (1.32). By combining spectrum sensing and signal reconstruction, the MWC serves
as two separate communication devices. The first is a state-of-the-art CRo that performs
real time spectrum sensing at sub-Nyquist rates, and the second is a receiver that is
able to decode multiple data transmissions simultaneously, regardless of their carrier
frequencies, while adapting to real-time spectral changes.
The CRr system is based on the sub-Nyquist radar receiver board described in
Section 1.3.4. The prototype simulates transmission of P = 50 pulses towards L = 9
targets. The CRr transmits over Nb = 4 bands, selected according to the procedure
presented in Section 1.5.2, after the spectrum-sensing process has been completed by
the communication receiver. We compare the target detection performance of our CRr
with a traditional wideband radar with bandwidth Bh = 20 MHz. The CRr-transmitted
bandwidth is thus equal to 3.2% of the wideband.
Figure 1.15 shows windows from the graphical user interface (GUI) of our CRr
system. Figure 1.15a illustrates the coexistence between the radar transmitted bands
Figure 1.15 SpeCX radar display [20] showing (a) coexisting CRo and CRr (b) CRr spectrum
compared with the full-band radar spectrum. The range-Doppler display of detected and true
locations of the targets for the case of (a) CRr (four disjoint bands) and (d) all four transmit
subbands together forming a contiguous 320 kHz band.
Sub-Nyquist Radar: Principles and Prototypes 29
(thick curve) and the existing communication bands (thin curve). The gain in power
is demonstrated in Figure 1.15b, which plots the wideband radar spectrum, CRr, and
noise. The true and recovered range-Doppler maps for the CRr (whose transmit signal
consists of four disjoint subbands) are shown in Figure 1.15c. All 9 targets are perfectly
recovered and clutter is discarded. Figure 1.15d shows the performance when the four
subbands are joined together to result in a 320 kHz contiguous band for the radar
transmitter. There are many missed detections and false alarms in this case. Let the
true and estimated ranges of the ith target be di and d̂i , respectively. Then the rms
localization error (RMSLE) of L targets is given by
!
"
"1 L
RMSLE = # (di − d̂i )2 . (1.47)
L
i=1
In Figure 1.15c–d, the RMSLE is shown as follows: CRr (0.34 km), 320 kHz band
or 4 adjacent bands with same bandwidth (8.1 km), and wideband (1.2 km). The poor
resolution of the 4 adjacent bands scenario is due to its small aperture. The native range
resolution in case of 2 MHz wideband scenario is 75 m. In Figure 1.15c, the CRr is able
to detect 9 targets at locations 6.097, 31.764, 35.046, 35.451, 35.479, 81.049, 81.570,
121.442, and 120.922 km. Here, the distance between two closely spaced targets is less
than 75 m.
We now consider extending sub-Nyquist processing to the spatial domain for the
particular case of MIMO radar [86]. MIMO radars use an array of several transmit
and receive antenna elements, with each transmitter radiating a different, mutually
orthogonal waveform. Waveform orthogonality can be in time, frequency or code. Our
system is based on the collocated MIMO configuration [87], in which the elements
are close to each other so that the radar cross section of a target appears identical in
all elements. The MIMO receiver separates and coherently processes the target echoes
that correspond to each transmitter. The angular resolution of MIMO using the classic
virtual ULA is the same as a phased array with equivalent virtual aperture but many
more antenna elements.
Conventional MIMO radar’s spatial (angular) and range resolutions are limited by
the number of elements and the receiver sampling rate, respectively. Here, we extend
the Xampling framework for temporal sub-Nyquist radar in Section 1.3 to both space
and time by simultaneously thinning an antenna array and sampling received signals at
sub-Nyquist rates. This sub-Nyquist collocated MIMO radar (SUMMeR) recovers the
target range, azimuth, and Doppler velocity without loss of any of the aforementioned
radar resolutions. In SUMMeR, the radar antenna elements are randomly placed within
the aperture, and signal orthogonality is achieved by frequency division multiplexing
(FDM). The FDM-based sub-Nyquist MIMO mitigates the range-azimuth coupling by
randomizing the element locations in the aperture [88].
30 Mishra and Eldar
Figure 1.16 Location of transmit (diamonds) and receive (triangles) antenna elements within the
same physical aperture for (a) conventional MIMO array with T = 5 transmitters and R = 4
receivers, (b) virtual ULA with T R = 20 antenna elements, and (c) randomly thinned MIMO
array with M = 4 transmitters and Q = 3 receivers.
P
−1
sm (t) = hm (t − pτ)ej 2πfc t , 0 ≤ t ≤ P τ, (1.48)
p=0
∞
Hm (ω) = hm (t)e−j ωt dt. (1.49)
−∞
For simplicity, we assume that fc τ is an integer. The pulse time support is denoted
by Tp .
Consider a target scene with L non-fluctuating point targets following the Swerling-0
model [1] whose locations are given by their ranges Rl , Doppler velocity vl , and azimuth
angles θl , 1 ≤ l ≤ L. The pulses transmitted by the radar are reflected back by the
Sub-Nyquist Radar: Principles and Prototypes 31
targets and collected at the receive antennas. When the received waveform is downcon-
verted from RF to baseband, we obtain the following signal at the qth antenna,
P −1 M−1
L
α l hm (t − pτ − τl ) ej 2πβmq ϑl ej 2πfl
D pτ
xq (t) = , (1.50)
p=0 m=0 l=1
where α l denotes the complex-valued reflectivity of the lth target, τl = 2Rl /c is the
range-time delay the lth target, flD = 2vc l fc is the frequency in the Doppler spectrum,
ϑl = sin θl is the azimuth parameter, and βmq is governed by the array structure. We
express xq (t) as a sum of single frames
P −1
p
xq (t) = xq (t), (1.51)
p=0
where
M−1 L
α l h(t − τl − pτ)ej 2πβmq ϑl ej 2πfl
p D pτ
xq (t) = . (1.52)
m=0 l=1
Our goal is to estimate the time delay τl , azimuth θl , and Doppler shifts flD of each
target from low rate samples of xq (t), for 0 ≤ q ≤ Q − 1, and a small number of M
channels and Q antennas.
L
2π
α l ej 2πβmq ϑl e−j e−j 2πfm τl ej 2πfl
p D pτ
ym,q [k] = τ kτl , (1.53)
l=1
2π −j 2π fm n
Here, Am denotes the K × T N matrix whose (k,n)th element is e−j T N κk n e Bh T
theorem 1.6.1 [17] The minimal number of transmit and receive array elements, i.e.,
M and Q, respectively, required for perfect recovery of XD with L targets in a noiseless
setting are determined by MQ ≥ 2L. In addition, the number of samples per receiver is
at least MK ≥ 2L where K is the number of Fourier coefficients sampled per receiver
and the number of pulses per transmitter is P ≥ 2L.
Theorem 1.6.1 shows that the number of SUMMeR transmit and receive elements
as well as samples K depend only on the number of targets present. These design
parameters, therefore, can be substantially lesser than the requirements of a Nyquist
MIMO array. Similar results for temporal and Doppler sub-Nyquist radars were obtained
in Theorems 1.3.1 and 1.6.1.
P −1
ym,q [k]e−j 2πν pτ
p
νm,q [k] = (1.55)
p=0
L
P −1
2π
α l ej 2πβmq ϑl e−j τ (k+fm τ)τl
D −ν )pτ
= ej 2π(fl ,
l=1 p=0
P −1
Then, for each focused frequency ν, (1.55) reduces to a 2D problem, which can be
solved using CS recovery techniques, as summarized in Algorithm 4. Note that step 1
Sub-Nyquist Radar: Principles and Prototypes 33
with t (l,i) the (l,i)th element in the index set t at the tth iteration, and Et =
[et (1) . . . et (t)]. Once XD is recovered, the delays, azimuths, and Dopplers are esti-
mated as
τL (l,1) 2L (l,2) ˆD 1 L (l,3)
τ̂l = , ϑ̂ l = −1 + , fl = − + . (1.58)
TN TR 2τ Pτ
Since in real scenarios, targets delays, Dopplers, and azimuths are not necessarily
aligned to a grid, a finer grid can be used around detection points on the coarse grid
to reduce quantization error. This technique adds a step after support detection in each
iteration (step 4 in Algorithm 4).
P −1
(m, ν ) j 2π ν pτ
i,j = Zm
i+j K,p e .
p=0
(m, ν )
2: Initialization: residual R0 = (m, ν ) , index set 0 = ∅, t = 1
3: Project residual onto measurement matrices for 0 ≤ p ≤ P − 1:
ν = AH Rν B,
T T T T T T
where A &= [A0 A1 · · · A(M−1) ' ] , B = [B B · · · B
T 0 1 (M−1) ]T , and
(0, ν ) (M−1, ν )
R = diag [Rt−1 · · · Rt−1
ν ] is block diagonal
4: Find the three indices λ t = [λ t (1) λ t (2) λ t (3)] such that
[λt (1) λt (2) λt (3)] = arg maxi,j, ν νi,j
5: Augment index set t = t {λ t }
6: Find the new signal estimate
Doppler frequencies, and amplitudes, and array configurations, i.e., number of trans-
mitters and receivers and antenna locations, can be examined using the prototype. The
waveform generator board produces an analog signal corresponding to the synthesized
radar environment, which is amplified and routed to the MIMO radar receiver board.
The prototype then samples and processes the signal in real time. The physical array
aperture and simulated target response correspond to an X-band (fc = 10 GHz) radar.
A conventional 8 × 10 MIMO radar receiver would require simultaneous hardware
processing of 80 (or 160 I/Q) data streams. Since a separate sub-Nyquist receiver for
each of these 80 channels is expensive, we implement the 8-channel analog process-
ing chain for only 1 receive antenna element, and serialize the received signals of all
10 elements through this chain. This approach allows the prototype to implement a
Sub-Nyquist Radar: Principles and Prototypes 35
number of receivers greater than 10 as the 8-channel hardware only limits the number
of transmitters.
If we use the same pre-filtering approach as in Section 1.3.4 for each of the eight
channels of our sub-Nyquist MIMO prototype, then the hardware design would need a
total of 4 × 8 = 32 BPFs and ADCs excluding the analog filters to separate transmit
channels. We sidestep this requirement by adopting cognitive transmission wherein
the analog signal of each channel exists only in certain predetermined subbands and
consequently, a BPF stage is not required. More importantly, for each channel, a single
low-rate ADC subsamples this narrow-band signal as long as the subbands are coset
bands so that they do not alias after sampling [46]. This implementation needs only
eight low-rate ADCs, one per channel. Another advantage of this approach is flexibility
of the prototype in selecting the Xampling slices. Unlike in Section 1.3.4, the number
and spectral locations of slices are not permanently fixed, and they can be changed.
Table 1.2 lists detailed technical characteristics of the prototype. The system can
be configured to operate in various array configurations or modes. Mode 3 and 4 are
sub-Nyquist MIMO modes; the hardware switches off the inactive channels and does
not sample any data over the corresponding ADCs. Figure 1.18 shows the sub-Nyquist
MIMO prototype, user interface and radar display. As shown in Figure 1.19a, the cog-
nitive radar signal occupies only certain subbands in a 15 MHz band. Here, the sliced
transmit signal has eight subbands each of width 375 kHz with the frequency range of
1.63–2, 2.16–2.53, 3.05–3.42, 3.88–4.25, 5.66–6.03, 6.51–6.88, 8.64–9.01, and 12.32–
12.69 MHz before subsampling. The total signal bandwidth is 0.375 × 8 = 3 MHz.
This signal is subsampled at 7.5 MHz and the subbands locations were chosen so that
there is no aliasing between different subbands (Figure 1.19b). A noncognitive signal
would have occupied the entire 15 MHz spectrum requiring a Nyquist sampling rate of
30 MHz. Therefore, the use of cognitive transmission enables spectral sampling reduc-
tion by a factor of 4 (= 30 MHz/7.5 MHz) for each channel. Depending on whether
the guard-bands of the noncognitive transmission are included in the computation or
not, the effective signal bandwidth is reduced by a factor of 5 (= 15 MHz/3 MHz) or 4
36 Mishra and Eldar
Figure 1.17 Tx and Rx element locations for the hardware prototype modes over a 6 m antenna
aperture. Mode 4’s virtual array equivalent is the 20 × 20 ULA [18].
Figure 1.18 Sub-Nyquist MIMO prototype and user interface. The analog preprocessor module
consists of two cards mounted on opposite sides of a common chassis [18].
Figure 1.19 The normalized one-sided spectrum of one channel of a given receiver (a) before and
(b) after subsampling with a 7.5 MHz ADC. Each of the subbands spans 375 kHz and is marked
with a numeric label. In a noncognitive processing, the signal occupies the entire 15 MHz
spectrum before sampling [18].
(= 12 MHz/3 MHz) respectively for each channel. Mode 3 has 50% spatial sampling
reduction when compared with Mode 1 or 2. Table 1.3 summarizes the reduction of
various resources in Mode 3 when compared with Mode 1.
We evaluated the performance of all modes through hardware experiments. We trans-
mitted P = 10 pulses at a PRF of 100 μs and all modes were evaluated against identical
target scenarios. In the first experiment, when the angular spacing (in terms of the
sine of azimuth) between any two targets was greater than 0.025 and the signal SNR
= −8 dB, the recovery performance of the thinned 4 × 5 array in Mode 3 was not
worse than Modes 1 and 2. For this experiment, Figures 1.20 and 1.21 show the plan
Sub-Nyquist Radar: Principles and Prototypes 37
Figure 1.20 Plan position indicator (PPI) display of results for (a) Mode 1 (b) Mode 2 (c) Mode 3
and (d) Mode 4. The origin is the location of the radar. The dark dot indicates the north direction
relative to the radar. Positive (negative) distances along the horizontal axis correspond to the east
(west) of the radar. Similarly, positive (negative) distances along the vertical axis correspond to
the north (south) of the radar. The estimated targets are plotted over the ground truth [18,53].
Figure 1.21 Range-azimuth-Doppler map for the target configurations shown in Figure 1.20 for
(a) Mode 1 (b) Mode 2 (c) Mode 3 and (d) Mode 4. The lower axes represent Cartesian
coordinates of the polar representation of the PPI plots from Figure 1.20. The vertical axis
represents the Doppler spectrum [18,53].
38 Mishra and Eldar
Figure 1.22 PPI plots as in Figure 1.20 for (a) Mode 1 (b) Mode 2 (c) Mode 3 and (d) Mode 4.
Only Mode 3 is operating cognitively. All modes have the same overall transmit power per
transmitter. The inset plots show the selected region in each PPI display on a magnified scale.
Figure 1.23 Range-azimuth-Doppler maps as in Figure 1.21 for (a) Mode 1 (b) Mode 2 (c) Mode 3
and (d) Mode 4. Only Mode 3 is operating cognitively. All modes have the same overall transmit
power per transmitter. The inset plots show the selected region in each map on a magnified scale.
position indicator (PPI) plot and range-azimuth-Doppler maps of all the modes. Here, a
successful detection (circle with light fill and no boundary) occurs when the estimated
target is within one range cell, one azimuth bin, and one Doppler bin of the ground truth
(circle with dark boundary and no fill); otherwise, the estimated target is labeled as a
false alarm (circle with dark fill). When a target remains undetected, we label the ground
truth location as a missed detection (circle with hatched fill).
Finally, we considered a high-noise scenario with SNR = −15 dB. We operated only
Mode 3 cognitively and kept all other modes in noncognitive mode. We noticed that the
noncognitive Nyquist 8 × 10 Mode 1 array exhibits false alarms while cognitive sub-
Nyquist 4 × 5 Mode 3 array is still able to detect all the targets (Figures 1.22 and 1.23),
thereby demonstrating robustness to low SNR.
Sub-Nyquist Radar: Principles and Prototypes 39
Synthetic aperture radar (SAR) and other similar radar techniques were one of the
first applications of CS methods (see reviews in [22,25]). SAR imaging data are not
naturally sparse in the range-time domain. However, they are often sparse in other
domains, such as wavelets. Our motivation to apply sub-Nyquist methods here is to
address the following SAR processing challenge. Among the several algorithms that are
available to process SAR data, the range-Doppler algorithm (RDA) is most widely used
to obtain high-resolution images [90]. Its performance is, however, limited by the range
cell migration correction (RCMC) step, which requires oversampled data in order to
decouple range and azimuth axes.
Recently, [21] proposed a sub-Nyquist SAR that replaces RDA by a Fourier domain
method that achieves non-integer nonconstant shifts in the RCMC interpolation via
the Fourier series coefficients. This avoids the interpolation step in RCMC and fur-
ther allows sub-Nyquist sampling that follows the Fourier-domain analysis presented in
previous sections. A similar technique was earlier employed in ultrasound imaging [48]
to dramatically reduce sampling and processing rates.
In this section, we present this Fourier domain RDA processing as a framework
for sub-Nyquist sampling of SAR signals. The first part of the sub-Nyquist algorithm
exploits the relationship between the signals before and after RCMC in the Fourier
domain. We show analytically that a single Fourier coefficient after RCMC can be
computed using a small number of Fourier coefficients of the raw data, which translates
into low rate sampling as shown in Section 1.3.2. Having the partial Fourier samples
after RCMC, the second part of the algorithm is aimed at solving a 2D CS problem in
order to reconstruct the image from the low rate samples. Finally, we show that cognitive
transmission can also be extended to SAR. We end by demonstrating a prototype that
we designed and developed to realize concepts of cognitive SAR (CoSAR) [21].
where ||r − xm || is the distance from the radar to a scatter point and wa (xm,r) is the
antenna beam pattern, which varies depending on the SAR operation mode [90]. The
main goal of SAR data processing is to construct the scene’s reflectivity map, σ(r),
from the raw data. The reflected signal dm (t) at a point m requires sampling at least
at the bandwidth Bh , as per the Nyquist sampling theorem. The resulting discrete-time
40 Mishra and Eldar
signal is d[n,m] = dm (nT s), with 0 ≤ n < N = Tfs , where fs = 1/Ts is the
sampling rate.
RDA processing consists of the following steps. First the sampled raw data is com-
pressed in the range dimension:
where h[n] is the sampled transmit signal. This data is then transformed to the range-
Doppler domain using DFT along the azimuth:
M−1
S[n,k] = s[n,m]e−j 2πkm/M . (1.61)
m=0
For every Doppler frequency k, the range axis is scaled by 1 + ak 2 . The value of a
is predetermined depending on the observation mode. For example, in stripmap SAR,
a = 8|ν |Tλ 2 M 2 . This range-variant shift requires values that fall outside the discrete grid.
2
where Ka [n] is the range dependent azimuth chirp rate. Finally, an inverse DFT in the
azimuth direction yields the focused image:
1
M−1
I [n,m] = Y [n,k]ej 2πmk/M . (1.64)
M
k=0
There are two ways to implement RCMC: In the first option, RCMC is performed by
range interpolation in the range-Doppler domain. However, this interpolation is time-
consuming and computationally demanding. The second approach involves the assump-
tion that the range cell migration is range invariant, at least over a finite range block. In
this case, RCMC is implemented using a DFT, linear phase multiply, and inverse DFT
per block. However, this implementation’s disadvantage is that samples should overlap
in range, and the efficiency gain may not be worth the added complexity.
to obtain the same image quality and without any assumptions on the signal structure or
the invariance of range blocks.
We begin with the continuous version of (1.62):
where Sk (nTs ) = S[n,k]. The Fourier series coefficients of Ck (t) over the interval [0,T )
can be expressed as [21]
T
1
Ck [l] = Sk (t)qk,l (t), (1.66)
T
0
where qk,l (t) approximates the scaling operation in (1.62). The Fourier coefficients of
the continuous-time signals Sk (t) and qk,l (t) are, respectively, Sk [n] and
1 −j π(n+ 1 2 ) l
Qk,l [n] = e 1+ak sinc n + . (1.67)
1 + ak 2 1 + ak 2
It can be shown that most of the energy of the set Qk,l [n] is concentrated around
a specific component nk,l . Thus, for every Doppler frequency k, the Fourier series
coefficients of the scaled signal, Ck (t), can be calculated as a linear combination of
a local choice of Fourier series coefficients of Sk (t)
Ck [l] = Sk [n]Qk,l [−n], (1.68)
n∈ν (k,l)
where ν(k,l) is the set of indices that dictate the decay property of Qk,l [n].
Assuming the Fourier series coefficients Dm [l] of the raw data dm (t) can be acquired
directly, the range compression is achieved in the Fourier domain as
where H [l] are the Fourier series coefficients of the transmitted pulse h(t). Applying
azimuth DFT gives Sk [l] that can be used in (1.68) to perform Fourier domain RCMC.
The inverse DFT on the coefficients Ck [l] provides the corrected sampled signal after
RCMC. One could then proceed with the remaining steps of RDA, i.e., (1.63) and (1.64),
to complete the processing.
The number of Fourier coefficients required can be further reduced if a basis (e.g.,
wavelet) is found in which the desired image is sparse. Then, the relationship between
Ck [l] and the raw data samples Dm [l] can be exploited to solve for the coefficients
in the sparse basis using fewer Fourier coefficients. In [21], it was suggested that we
modify a fast iterative shrinkage-thresholding algorithm (FISTA) to solve this problem
and achieve full data reconstruction from the partial measurements with reasonable
computational load. Similar to cognitive pulse-Doppler radar (Section 1.5.2) and cog-
nitive SUMMeR (Section 1.6.5), sub-Nyquist SAR systems can also be modified to
fit cognitive radar requirements and allow for dynamic transmission and reception of
several narrow frequency bands. We present the hardware prototype of such a system in
the next subsection.
42 Mishra and Eldar
Figure 1.24 Cognitive SAR (CoSAR) prototype and (inset) analog preprocessor.
Figure 1.25 CoSAR GUI showing the cognitive and noncognitive chirp waveforms along with the
sampled subbands at top right.
than the noncognitive Nyquist processing in low SNR scenarios, despite sampling at
a significantly reduced sampling rate. Further, the prototype demonstrates operation
of SAR using narrow subbands that can be adaptively changed. This opens up the
possibility of spectral coexistence of SAR with other satellite-borne services.
1.8 Summary
References
[1] M. I. Skolnik, Radar Handbook, 3rd edn. McGraw-Hill, 2008.
[2] P. Z. Peebles, Radar Principles. Wiley-Interscience, 1998.
[3] J. E. Cilliers and J. C. Smit, “Pulse compression sidelobe reduction by minimization of
lp -norms,” IEEE Transactions on Aerospace and Electronic Systems, vol. 43, no. 3,
pp. 1238–1247, 2007.
[4] J. George, K. V. Mishra, C. M. Nguyen, and V. Chandrasekar, “Implementation of blind
zone and range-velocity ambiguity mitigation for solid-state weather radar,” in IEEE Radar
Conference, 2010, pp. 1434–1438.
[5] K. V. Mishra, V. Chandrasekar, C. Nguyen, and M. Vega, “The signal processor system for
the NASA dual-frequency dual-polarized Doppler radar,” in IEEE International Geoscience
and Remote Sensing Symposium, 2012, pp. 4774–4777.
[6] Y. C. Eldar, Sampling Theory: Beyond Bandlimited Systems. Cambridge University Press,
2015.
[7] Y. C. Eldar and G. Kutyniok, Compressed Sensing: Theory and Applications. Cambridge
University Press, 2012.
[8] D. Cohen and Y. C. Eldar, “Reduced time-on-target in pulse Doppler radar: Slow time
domain compressed sensing,” in IEEE Radar Conference, 2016, pp. 1–4.
[9] J. Akhtar, B. Torvik, and K. E. Olsen, “Compressed sensing with interleaving slow-time
pulses and hybrid sparse image reconstruction,” in IEEE Radar Conference, 2017, pp. 0006–
0010.
[10] K. V. Mishra, A. Kruger, and W. F. Krajewski, “Compressed sensing applied to weather
radar,” in IEEE International Geoscience and Remote Sensing Symposium, 2014, pp. 1832–
1835.
[11] R. P. Shenoy, “Phased array antennas,” in Advanced Radar Techniques and Systems,
G. Galati, Ed. Peter Peregrinus, 1993.
[12] E. T. Bayliss, “Design of monopulse antenna difference patterns with low sidelobes,” Bell
System Technical Journal, vol. 47, no. 5, pp. 623–650, 1968.
[13] D. K. Cheng, “Optimization techniques for antenna arrays,” Proceedings of the IEEE,
vol. 59, no. 12, pp. 1664–1674, 1971.
[14] R. L. Haupt, Timed Arrays: Wideband and Time Varying Antenna Arrays. John Wiley &
Sons, 2015.
[15] K. V. Mishra, I. Kahane, A. Kaufmann, and Y. C. Eldar, “High spatial resolution radar using
thinned arrays,” in IEEE Radar Conference, 2017, pp. 1119–1124.
[16] M. Rossi, A. M. Haimovich, and Y. C. Eldar, “Spatial compressive sensing for MIMO radar,”
IEEE Transactions on Signal Processing, vol. 62, no. 2, pp. 419–430, 2014.
[17] D. Cohen, D. Cohen, Y. C. Eldar, and A. M. Haimovich, “SUMMeR: Sub-Nyquist MIMO
radar,” IEEE Transactions on Signal Processing, vol. 66, no. 16, pp. 4315–4330, 2018.
[18] K. V. Mishra, E. Shoshan, M. Namer et al., “Cognitive sub-Nyquist hardware prototype of
a collocated MIMO radar,” in International Workshop on Compressed Sensing Theory and
its Applications to Radar, Sonar and Remote Sensing, 2016, pp. 56–60.
[19] K. V. Mishra and Y. C. Eldar, “Performance of time delay estimation in a cognitive radar,”
in IEEE International Conference on Acoustics, Speech and Signal Processing, 2017,
pp. 3141–3145.
Sub-Nyquist Radar: Principles and Prototypes 45
[20] D. Cohen, K. V. Mishra, and Y. C. Eldar, “Spectrum sharing radar: Coexistence via
Xampling,” IEEE Transactions on Aerospace and Electronic Systems, vol. 29, no. 3,
pp. 1279–1296, 2018.
[21] K. Aberman and Y. C. Eldar, “Sub-Nyquist SAR via Fourier domain range-Doppler
processing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 55, no. 11,
pp. 6228–6244, 2017.
[22] J. H. Ender, “On compressive sensing applied to radar,” Signal Processing, vol. 90, no. 5,
pp. 1402–1414, 2010.
[23] N. A. Goodman and L. C. Potter, “Pitfalls and possibilities of radar compressive sensing,”
Applied Optics, vol. 54, no. 8, pp. C1–C13, 2015.
[24] L. Zhao, L. Wang, L. Yang, A. M. Zoubir, and G. Bi, “The race to improve radar imagery: An
overview of recent progress in statistical sparsity-based techniques,” IEEE Signal Processing
Magazine, vol. 33, no. 6, pp. 85–102, 2016.
[25] M. Cetin, I. Stojanovic, O. Onhon, et al., “Sparsity-driven synthetic aperture radar imaging:
Reconstruction, autofocusing, moving targets, and compressed sensing,” IEEE Signal
Processing Magazine, vol. 31, no. 4, pp. 27–40, 2014.
[26] M. A. Hadi, S. Alshebeili, K. Jamil, and F. E. A. El-Samie, “Compressive sensing applied to
radar systems: An overview,” Signal, Image and Video Processing, vol. 9, no. 1, pp. 25–39,
2015.
[27] R. Baraniuk and P. Steeghs, “Compressive radar imaging,” in IEEE Radar Conference, 2007,
pp. 128–133.
[28] M. A. Herman and T. Strohmer, “High-resolution radar via compressed sensing,” IEEE
Transactions on Signal Processing, vol. 57, no. 6, pp. 2275–2284, 2009.
[29] Y.-S. Yoon and M. G. Amin, “Compressed sensing technique for high-resolution radar
imaging,” in Signal Processing, Sensor Fusion, and Target Recognition XVII, vol. 6968,
2008, p. 69681A.
[30] X. Tan, W. Roberts, J. Li, and P. Stoica, “Range-Doppler imaging via a train of probing
pulses,” IEEE Transactions on Signal Processing, vol. 57, no. 3, pp. 1084–1097, 2009.
[31] J. Zhang, D. Zhu, and G. Zhang, “Adaptive compressed sensing radar oriented toward cog-
nitive detection in dynamic sparse target scene,” IEEE Transactions on Signal Processing,
vol. 60, no. 4, pp. 1718–1729, 2012.
[32] C.-Y. Chen, “Signal processing algorithms for MIMO radar,” PhD dissertation, California
Institute of Technology, 2009.
[33] Y. Yu, A. P. Petropulu, and H. V. Poor, “Measurement matrix design for compressive
sensing–based MIMO radar,” IEEE Transactions on Signal Processing, vol. 59, no. 11,
pp. 5338–5352, 2011.
[34] Y. Yu, A. P. Petropulu, and H. V. Poor, “MIMO radar using compressive sampling,” IEEE
Journal on Selected Topics in Signal Processing, vol. 4, no. 1, pp. 146–163, 2010.
[35] Y. Chi, L. L. Scharf, A. Pezeshki, and A. R. Calderbank, “Sensitivity to basis mismatch in
compressed sensing,” IEEE Transactions on Signal Processing, vol. 59, no. 5, pp. 2182–
2195, 2011.
[36] R. Heckel, V. I. Morgenshtern, and M. Soltanolkotabi, “Super-resolution radar,” Information
and Inference: A Journal of the IMA, vol. 5, no. 1, pp. 22–75, 2016.
[37] R. Heckel, “Super-resolution MIMO radar,” in IEEE International Symposium on Informa-
tion Theory, 2016, pp. 1416–1420.
[38] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht, “Compressed sensing off-the-grid,” IEEE
Transactions on Information Theory, vol. 59, no. 11, pp. 7465–7490, 2013.
46 Mishra and Eldar
[39] K. V. Mishra, M. Cho, A. Kruger, and W. Xu, “Super-resolution line spectrum estimation
with block priors,” in Asilomar Conference on Signals, Systems and Computers, 2014, pp.
1211–1215.
[40] W. U. Bajwa, K. Gedalyahu, and Y. C. Eldar, “Identification of parametric underspread linear
systems and super-resolution radar,” IEEE Transactions on Signal Processing, vol. 59, no. 6,
pp. 2548–2561, 2011.
[41] W. Kozek and G. E. Pfander, “Identification of operators with bandlimited symbols,” SIAM
Journal on Mathematical Analysis, vol. 37, no. 3, pp. 867–888, 2005.
[42] K. Gedalyahu and Y. C. Eldar, “Time-delay estimation from low-rate samples: A union of
subspaces approach,” IEEE Transactions on Signal Processing, vol. 58, no. 6, pp. 3017–
3031, 2010.
[43] S. Sun, W. U. Bajwa, and A. P. Petropulu, “MIMO-MC radar: A MIMO radar approach
based on matrix completion,” IEEE Transactions on Aerospace and Electronic Systems,
vol. 51, no. 3, pp. 1839–1852, 2015.
[44] M. Vetterli, P. Marziliano, and T. Blu, “Sampling signals with finite rate of innovation,”
IEEE Transactions on Signal Processing, vol. 50, no. 6, pp. 1417–1428, 2002.
[45] E. Baransky, G. Itzhak, I. Shmuel et al., “A sub-Nyquist radar prototype: Hardware and
algorithms,” IEEE Transactions on Aerospace and Electronic Systems, vol. 50, pp. 809–822,
2014.
[46] K. M. Cohen, C. Attias, B. Farbman, I. Tselniker, and Y. C. Eldar, “Channel estimation in
UWB channels using compressed sensing,” in IEEE International Conference on Acoustics,
Speech and Signal Processing, 2014, pp. 1966–1970.
[47] K. V. Mishra and Y. C. Eldar, “Sub-Nyquist channel estimation over IEEE 802.11ad link,” in
IEEE International Conference on Sampling Theory and Applications, 2017, pp. 355–359.
[48] T. Chernyakova and Y. C. Eldar, “Fourier-domain beamforming: The path to compressed
ultrasound imaging,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency
Control, vol. 61, no. 8, pp. 1252–1267, 2014.
[49] O. Bar-Ilan and Y. C. Eldar, “Sub-Nyquist radar via Doppler focusing,” IEEE Transactions
on Signal Processing, vol. 62, pp. 1796–1811, 2014.
[50] R. Tur, Y. C. Eldar, and Z. Friedman, “Innovation rate sampling of pulse streams with
application to ultrasound imaging,” IEEE Transactions on Signal Processing, vol. 59, no. 4,
pp. 1827–1842, 2011.
[51] N. Wagner, Y. C. Eldar, and Z. Friedman, “Compressed beamforming in ultrasound
imaging,” IEEE Transactions on Signal Processing, vol. 60, no. 9, pp. 4643–4657, 2012.
[52] Y. C. Eldar, R. Levi, and A. Cohen, “Clutter removal in sub-Nyquist radar,” IEEE Signal
Processing Letters, vol. 22, no. 2, pp. 177–181, 2015.
[53] D. Cohen, K. V. Mishra, D. Cohen et al., “Sub-Nyquist MIMO radar prototype with Doppler
processing,” in IEEE Radar Conference, 2017, pp. 1179–1184.
[54] S. Na, K. V. Mishra, Y. Liu, Y. C. Eldar, and X. Wang, “TenDSuR: Tensor-based 3D sub-
Nyquist radar,” IEEE Signal Processing Letters, 2018, in press.
[55] S. M. Kay, Fundamentals of Statistical Signal Processing, Volume 2: Detection Theory.
Prentice Hall, 1998.
[56] M. Mishali and Y. C. Eldar, “From theory to practice: Sub-Nyquist sampling of sparse
wideband analog signals,” IEEE Journal on Selected Topics in Signal Processing, vol. 4,
no. 2, pp. 375–391, 2010.
[57] M. Mishali and Y. C. Eldar, “Sub-Nyquist sampling: Bridging theory and practice,” IEEE
Signal Processing Magazine, vol. 28, no. 6, pp. 98–124, 2011.
Sub-Nyquist Radar: Principles and Prototypes 47
[58] D. Cohen and Y. C. Eldar, “Sub-Nyquist sampling for power spectrum sensing in cognitive
radios: A unified approach,” IEEE Transactions on Signal Processing, vol. 62, no. 15, pp.
3897–3910, 2014.
[59] D. Cohen and Y. C. Eldar, “Sub-Nyquist cyclostationary detection for cognitive radio,” IEEE
Transactions on Signal Processing, vol. 65, no. 11, pp. 3004–3019, 2017.
[60] D. Cohen, A. Dikopoltsev, R. Ifraimov, and Y. C. Eldar, “Towards sub-Nyquist cognitive
radar,” in IEEE Radar Conference, 2016, pp. 1–4.
[61] K. Gedalyahu, R. Tur, and Y. C. Eldar, “Multichannel sampling of pulse streams at the rate
of innovation,” IEEE Transactions on Signal Processing, vol. 59, no. 4, pp. 1491–1504,
2011.
[62] L. L. Scharf and B. Friedlander, “Matched subspace detectors,” IEEE Transactions on Signal
Processing, vol. 42, no. 8, pp. 2146–2157, 1994.
[63] P. B. Tuuk and S. L. Marple, “Compressed sensing radar amid noise and clutter using
interference covariance information,” IEEE Transactions on Aerospace and Electronic
Systems, vol. 50, no. 2, pp. 887–897, 2014.
[64] Y. Yu, S. Sun, and A. P. Petropulu, “A Capon beamforming method for clutter suppression
in colocated compressive sensing based MIMO radars,” in SPIE Defense, Security, and
Sensing, 2013, pp. 87 170J.
[65] K. Sun, H. Zhang, G. Li, H. Meng, and X. Wang, “A novel STAP algorithm using sparse
recovery technique,” in IEEE International Geoscience and Remote Sensing Symposium,
vol. 5, 2009, pp. V–336.
[66] X. Yang, Y. Sun, T. Zeng, and T. Long, “Iterative roubust sparse recoery method based on
focuss for space-time adaptive processing,” in IET International Radar Conference 2015,
2015, pp. 1–6.
[67] Z. Ma, Y. Liu, H. Meng, and X. Wang, “Jointly sparse recovery of multiple snapshots in
STAP,” in IEEE Radar Conference, 2013, pp. 1–4.
[68] Z. Wang, H. Li, and B. Himed, “A sparsity based GLRT for moving target detection in
distributed MIMO radar on moving platforms,” in Asilomar Conference on Signals, Systems
and Computers, 2015, pp. 90–94.
[69] S. Kay, “Optimal signal design for detection of Gaussian point targets in stationary Gaussian
clutter/reverberation,” IEEE Journal of Selected Topics in Signal Processing, vol. 1, no. 1,
pp. 31–41, 2007.
[70] L. E. Brennan and L. Reed, “Theory of adaptive radar,” IEEE Transactions on Aerospace
and Electronic Systems, no. 2, pp. 237–252, 1973.
[71] L. E. Brennan and I. S. Reed, “Optimum processing of unequally spaced radar pulse
trains for clutter rejection,” IEEE Transactions on Aerospace and Electronic Systems, no. 3,
pp. 474–477, 1968.
[72] T. Wimalajeewa, Y. C. Eldar, and P. K. Varshney, “Recovery of sparse matrices via matrix
sketching,” arXiv preprint arXiv:1311.2448, 2013.
[73] H. Griffiths, L. Cohen, S. Watts, E. Mokole, C. Baker, M. Wicks, and S. Blunt, “Radar
spectrum engineering and management: Technical and regulatory issues,” Proceedings of
the IEEE, vol. 103, no. 1, pp. 85–102, 2015.
[74] G. M. Jacyna, B. Fell, and D. McLemore, “A high-level overview of fundamental limits
studies for the DARPA SSPARC program,” in IEEE Radar Conference, 2016, pp. 1–6.
[75] J. R. Guerci, R. M. Guerci, A. Lackpour, and D. Moskowitz, “Joint design and operation of
shared spectrum access for radar and communications,” in IEEE Radar Conference, 2015,
pp. 0761–0766.
48 Mishra and Eldar
[76] K. V. Mishra, A. Zhitnikov, and Y. C. Eldar, “Spectrum sharing solution for automotive
radar,” in IEEE 85th Vehicular Technology Conference, 2017, pp. 1–5.
[77] A. R. Chiriyath, B. Paul, G. M. Jacyna, and D. W. Bliss, “Inner bounds on performance of
radar and communications co-existence,” IEEE Transactions on Signal Processing, vol. 64,
no. 2, pp. 464–474, 2016.
[78] D. Cohen, S. Tsiper, and Y. C. Eldar, “Analog to digital cognitive radio: Sampling, detection
and hardware,” IEEE Signal Processing Magazine, vol. 35, no. 1, pp. 137–166, 2018.
[79] M. Mishali and Y. C. Eldar, “Blind multi-band signal reconstruction: Compressed sensing
for analog signals,” IEEE Transactions on Signal Processing, vol. 57, no. 3, pp. 993–1009,
2009.
[80] D. Cohen, S. Tsiper, and Y. C. Eldar, “Analog to digital cognitive radio,” in Handbook of
Cognitive Radio, W. Zhang, Ed. Springer Singapore, 2017.
[81] N. Vaswani and W. Lu, “Modified-cs: Modifying compressive sensing for problems with
partially known support,” IEEE Transactions on Signal Processing, vol. 58, no. 9, pp. 4595–
4607, 2010.
[82] V. Stankovi, L. Stankovi, and S. Cheng, “Compressive image sampling with side informa-
tion,” in IEEE International Conference Image Processing, 2009, pp. 3037–3040.
[83] J. Huang, T. Zhang, and D. Metaxas, “Learning with structured sparsity,” Journal of Machine
Learning Research, vol. 12, no. 11, pp. 3371–3412, 2011.
[84] M. Mishali, Y. C. Eldar, O. Dounaevsky, and E. Shoshan, “Xampling: Analog to digital at
sub-Nyquist rates,” IET Circuits, Devices & Systems, vol. 5, pp. 8–20, 2011.
[85] M. Mishali and Y. C. Eldar, “Expected RIP: Conditioning of the modulated wideband
converter,” in IEEE Information Theory Workshop, 2009, pp. 343–347.
[86] E. Fishler, A. Haimovich, R. Blum, D. Chizhik, L. Cimini, and R. Valenzuela, “MIMO radar:
An idea whose time has come,” in IEEE Radar Conference, 2004, pp. 71–78.
[87] J. Li and P. Stoica, “MIMO radar with colocated antennas,” IEEE Signal Processing
Magazine, vol. 24, no. 5, pp. 106–114, 2007.
[88] D. Cohen, D. Cohen, and Y. C. Eldar, “High resolution FDMA MIMO radar,” arXiv preprint
arXiv:1711.06560, 2017.
[89] K. V. Mishra, Y. C. Eldar, E. Shoshan, M. Namer, and M. Meltsin, “A cognitive sub-Nyquist
MIMO radar prototype,” arXiv preprint arXiv:1807.09126, 2018.
[90] C. F. Barnes, Synthetic Aperture Radar: Wave Theory Foundations: Analysis and Algo-
rithms. Barnes (self-published), 2014.
2 Clutter Rejection and Adaptive
Filtering in Compressed
Sensing Radar
Peter B. Tuuk
2.1 Introduction
Clutter returns have posed challenges to radar designers and engineers since the early
days of technology development and use. Indeed, as early as the Second World War,
attention was paid to mitigating unwanted detections from terrain [1]. Generally, clutter
returns were mitigated by constructing the observation geometry so that targets were
above the radar and sensed against the background of sky. Early techniques included a
simple notch filter at the transmitted frequency that removed returns with zero Doppler
shift. This was effective in some cases for stationary radar systems, but did not com-
pensate for the effects of platform motion that shifted the frequency of clutter returns.
A subsequent development, pioneered in the 1950s, was the displaced phase center
technique that phase-shifted the returns from a series of pulses to align them, allowing
for more effective cancellation [2]. The next major development in airborne radar was
a revolution: pulse-Doppler radar, enabled by high-accuracy timing circuitry and early
digital memory, which coherently processed a set of pulses for clutter rejection and other
purposes. These techniques allowed effective airborne early warning development and
look-down, shoot-down modes for fighter aircraft [3]. One such early pulse-Doppler
radar, the AWG-10, was employed on the McDonnell-Douglas F-4 Phantom.
In the late 1970s and 1980s, space–time adaptive processing (STAP) was introduced
[4]. STAP takes advantage of improvements in digital signal processing to extend the
clutter cancellation to the two-dimensional domain. It does so by introducing a spa-
tial or array channel dimension. These additional degrees of freedom allow improved
cancellation in the joint domain and extend work on clutter cancellation to that of
other structured interference sources. In the years since, STAP theory and practice
have improved with the introduction of more array channels, prior-knowledge-aided
processing, and the introduction of computationally expensive matrix decompositions.
As STAP becomes a more mature technology, it is migrating to smaller platforms.
But cost, size, weight, power, and other considerations make large, multi-channel, high-
bandwidth array antennas infeasible in these settings. Compressed sensing (CS) offers
the hope that lower sampling requirements and data volumes could simplify data acqui-
sition requirements and allow advanced techniques on lower-end platforms. In some
contexts the computational costs of CS reconstruction are prohibitive today. But in oth-
ers, the signal acquisition problem is intractable under traditional Nyquist-rate sampling.
49
50 Tuuk
The radar identifies objects within its field of regard by transmitting radio frequency
electromagnetic energy into the surrounding medium. This energy propagates through
the medium impinging on objects in that environment. The objects reflect some portion
of the energy back to the radar where it is processed to estimate characteristics of the
environment.
fast time sampling dimensions. Further define n = Nc Ns Nf . This data cube can be
vectorized as y ∈ Cn and approximated by a linear combination of basis elements in a
sensing operator:
y = Sx, (2.1)
where y contains the samples in time and space. The vector x is the unknown vector
that describes the target scene; it is a vector that gives radar a cross section of scatterers
at each location in the observation extent. Generally this vector is sparse because many
gridpoints contain no target.
for some scalar γi . In a digital receiver fi,k (t) is sampled in fast time every Ts =
1/β seconds.
52 Tuuk
To express this in the range basis, S r , each column is a shifted copy of the transmitted
waveform, w(t), with leading and trailing zeros. The first column is the response from
a target at the minimum range in the range window and in each subsequent column the
waveform is shifted by one entry.
⎡ ⎤
w(Ts ) 0 0 ··· 0
⎢ w(2 Ts ) w(Ts ) 0 ··· 0 ⎥
⎢ ⎥
⎢ w(3 Ts ) w(2 Ts ) w(Ts ) · · · 0 ⎥
Sr = ⎢ ⎥ (2.5)
⎢ .. ⎥
⎣ . ⎦
0 0 0 ··· w(nr Ts )
The angle and Doppler bases, S a and S d , are both frequency bases. The Doppler
frequency is evaluated across pulses in the coherent processing interval. The angle
basis is the relative phase delay introduced at the elements of the antenna array as an
incoming planar wavefront reaches each element sequentially. The time delay between
receivers becomes a simple phase shift for narrowband signals. For wideband signals,
the spatial phase pattern becomes frequency-dependent. To coherently process signals
in a wideband setting, other approaches can be used to minimize the losses due to
phase mismatch. These include true time delay units or corrections applied in the
digital domain to counteract the known phase errors introduced. For radar systems
made up of subarrays, true phase combining can be used at the subarray level if the
subarray is small enough or the bandwidth is small enough to adequately control phase
migration.
This linear model can be represented explicitly as a matrix, with each column of
the matrix S being the return from a target at the corresponding range-angle-Doppler
position in space. But the size of the matrix grows rapidly as the dimensions of the
sample space and search space increase. For instance, for a system with 8 channels, 128
pulses, and 512 range samples, the matrix S has 2.8 × 1011 elements. Therefore, the
most efficient way to implement this model is not by explicitly storing the matrix but by
performing discrete Fourier transforms along the angle and velocity dimensions and a
convolution in the range dimension. For computational gains, these can be implemented
using the fast Fourier transform (FFT). By this means the storage requirements and
processing time can be reduced considerably.
This linear model suffices for analysis and estimation. Of course targets move through
continuous space, and so any discretization will necessarily be only an approximation.
Finer discretization can reduce the associated errors, but at the cost of increasing the
correlation between columns of the matrix. With any processing there are diminishing
returns as the discretization becomes finer than the fundamental resolution of the sensor.
This chapter expresses many of these concepts in a single dimension assuming a
uniform linear array, but all these results are generalizable to planar and nonuniform
arrays with somewhat more complicated notation. In this book we neglect that fourth
dimension in the interest of clarity and computational tractability. Furthermore, any
polarization effects are neglected in this model. For systems that record dual polarization
an added dimension could be used to represent that variable.
Clutter Rejection and Adaptive Filtering 53
ϕ
Antenna Subarray Elements
... A
D
Figure 2.1 A simplified block diagram of a single subarray of a large airborne phased array radar.
The signal path includes a low-noise amplifier, analog filtering, phase shifting, RF combining,
mixing to IF, and sampling. The digital samples from this and all other subarrays feed into the
digital signal and data processor for pulse compression, beamforming, detection, association,
tracking, prediction, and other functions.
y = S(x + c) + n. (2.8)
The clutter is illuminated by the same waveform as the targets themselves and thus
signal processing that is used to improve the signal-to-noise ratio of the targets will
cause the clutter to be accentuated as well. So more energetic waveforms or more
integration will not be useful in improving the signal-to-clutter ratio (SCR). By contrast
RFI exibits strong spatial correlation but little or no correlation between pulses, so
for this interference source the signal-to-interference ratio can be improved by longer
integration times or higher power.
Clutter can result from land or sea surface reflections, and the statistics of the
interference depend strongly on the particularities of the terrain being surveilled. Much
research has focused on describing the expected returns from clutter as a function
of terrain, radar band, radar resolution, and other parameters. The simplest model is
to assume a Gaussian distribution; for low-bandwidth radars this assumption is often
sufficient. A common distribution for clutter amplitude is gamma function [6,7]. But
detailed analysis of several collected datasets shows significant skewness and kurtosis
that do not match either the normal or gamma distributions. Having examined K,
log-normal, Weibull, and Rayleigh distributions, work in [8] showed that the Weibull
distribution best described the data collected in a flight test over open farmland in
Saskatchewan. In [9] research led to a compound distribution composed of two
separate gamma distributions that describe the modulation and speckle observed in
high-resolution radar.
Clutter Rejection and Adaptive Filtering 55
Probability of Occurance
One Realization
0.8 Underlying Distribution
0.6
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Amplitude Value (Scaled to Mean)
Figure 2.2 The sampled distribution of the observed clutter in one realization, along with the
underlying gamma distribution, which is defined by a shape parameter of 10/3.
So with all this in mind, we have elected to use a constant gamma-distributed random
model variable as in [6], with a shape factor of 10/3 in our examples. This model
approximates a terrain with relatively open land and the absence of man-made scatterers.
Figure 2.2 shows the amplitude of clutter observed in a realization of this clutter model.
Other models may be more appropriate at very low grazing angles and for other types
of terrain [8,9].
Processing using only the matched filtering applies a nonadaptive technique to a tar-
get scene. However the inter-bin correlation structure can be used to improve target
detection performance. Several approaches to treating the correlated clutter problem are
described in this section.
R = R n + R c = ν 2 I Ncs + V DV H , (2.9)
R̂ = Nf−1 Y Y H . (2.10)
As more samples from this distribution are collected, Nf /Ncs → ∞, and this estimate
will converge to the true covariance R. This estimate of the interference statistics can
be inverted to form a signal-to-interference-plus-noise ratio (SINR)-maximizing linear
filter, [12] which leads to an estimate of the target scene for range bin i as
−1
x̂ i = κ S H R̂ yi, (2.11)
Figure 2.3 The matched filter estimate reshaped into the range-angle-Doppler cube and projected
along each of the three dimensions. The true target location is r = 290 m, θ = −30◦ , and
v = 1 m/s. The marker indicates the true target location in each view.
A
A
V
R R V
Figure 2.4 The STAP estimate reshaped into the range-angle-Doppler cube and projected along
each of the three dimensions. The true target location is r = 290 m, θ = −30◦ , and v = 1 m/s.
The marker indicates the true target location in each view.
58 Tuuk
The line of clutter appears as a slice in the third frame of the figures due to the interaction
of the sensor platform motion with the stationary ground clutter. In the matched-filter-
only estimate that clutter line is such a strong ridge in the angle-Doppler plane that it
renders the target invisible. In the STAP estimate, the clutter has been nulled by the filter
and the target can be easily located in all three projections. This is a high SCR example
to enhance the visibility of the target relative to the clutter. This same framework will be
used to generate experimental results in the latter portion of this chapter, although with
different parameters for signal, clutter, and noise power.
The accuracy of the estimate of the interference covariance decreases with the
number of degrees of freedom in the estimate (for a fixed training data set). To improve
the stability of the estimates at the expense of resolution, reduced-rank STAP is utilized.
These techniques work well because the interference information can be described
using relatively few basis vectors. In [16], it is shown that signal dependent rank
reduction can improve performance. The cross-spectral metric for designing the rank
reducer incorporates information on the desired signal steering vectors, not just the
interference statistics to improve performance in cases of limited training data. In [17],
it is shown that reduced-rank STAP techniques offer better performance than full-rank
estimation under constraints on training time and data. In [18], it is shown how one
may select a dimensionality reduction basis efficiently. In addition to a performance
improvement given limited training data, these reduced-rank techniques also offer
a reduced computational burden relative to a full-rank algorithm. The optimal rank
may be set by processing timeline constraints, but even in the unconstrained case,
learning the optimal number of dimensions to use for estimating the covariance is not
straightforward.
x̂ cmf = S H C H z. (2.15)
This method has low computational cost but does not necessarily yield a sparse solution.
A CS estimate of the target vector can be computed by solving a convex linear
optimization problem, such as an 1 -regularized least-squares
where the first term of the minimization objective is the Euclidean norm of the residual
that enforces fidelity to the measured data, and the second term is the 1 norm of the
estimate that promotes sparsity in the solution. The parameter λ enables a trade off
between these competing priorities.
Much attention has been paid to how the random sampling at the core of CS can be
realized in hardware. The exact content of the compression matrix C will depend on the
measurement process it describes. In pulse-Doppler radar, this process may introduce
incoherence in the following ways:
We have seen that interference covariance can be estimated from samples via the STAP
approach. The central question of this work is how well samples produced from a
compression operation C, as in (2.13), can be used to form an estimate of the covariance
matrix. The simplest approach is compressed sample matrix inversion (SMI), in which
the adjoint of the compression operator is used to bring the samples back to the full
ambient signal space before forming the covariance estimate:
that low-rank matrix recovery from few measurements is not ill-posed and is convex.
The convex optimization problem
min ||X||∗ s. t. X i,j = M i,j , (i,j ) ∈ (2.18)
will recover the matrix X exactly with high probability if m ≥ Cn1.2 r log n, for a
specified constant C.
In [25], the topic of matrix completion is surveyed and it is shown that n × n matrices
of rank r can be recovered from m noise-corrupted direct samples via nuclear norm
minimization with high probability if m ≥ Cnr log2 n with an error on the order of
the noise level. In [26], this idea is expanded to show that matrices can be recovered
from expansion coefficients with respect to a known matrix basis as long as that basis is
not coherent with the matrix being recovered. This is directly analogous to the required
incoherence between the spasifying basis and the sensing basis in CS theory. The work
in [27] provides information theoretic lower bounds on the number of samples needed
to recover certain types of low-rank matrices. One result is that m = Cnr log n is the
lower limit on the number of samples needed to recover a random n × n matrix with
rank r.
The work in [28] treats the topic of estimating simultaneously sparse and low-rank
matrices from rank-one measurements. This measurement model consists of a series of
sketches of the underlying matrix. In [29] this same “sketching” measurement model is
considered and shows specifically that the number of measurements required for stable
estimation scales linearly with the rank of the matrix and the sparsity of the matrix and
with the logarithm of the number of rows. Though this model differs from that which we
will develop in this chapter, the results provide a basis for confidence that by exploiting
the low-rank nature of the covariance, an improvement in performance can be expected.
In [30], it is shown that the low-rank assumption can be used to improve the esti-
mation accuracy of the covariance matrix using a standard STAP benchmark dataset in
the case of limited training data. Their proposed algorithm out-performs other lower
computational complexity algorithms by using a dictionary learning approach.
In [31,32], a CS–STAP technique is developed in which a small amount of training
data, in some cases one snapshot, can be used to estimate the covariance statistics and
build the whitening filter. This technique assumes direct sparsity in the interference
covariance matrix. The validity of this assumption depends on the type of interference
being described. For certain types of electromagnetic interference, or for certain types
of man-made clutter, a concentration in this domain can be very pronounced. For other
types of natural ground cover, the spectrum can be more distributed. This distribution is
especially pronounced for foliage being blown in the wind.
In [33] an approach is proposed to estimate the sample covariance matrix (that would
be obtained from the uncompressed data) from a set of compressed measurements.
The goal of estimating the sample covariance matrix is slightly different than that of
estimating the underlying interference statistics, but does isolate the two stages of sam-
pling limits: first, a sample covariance matrix deviates from the true covariance matrix
because it is based on a limited number of realizations, and second, the compressed
sample covariance matrix differs from the full sample covariance matrix because of the
Clutter Rejection and Adaptive Filtering 61
dimensionality reduction. The nature of the second limit depends on the manner in
which the samples are compressed and the estimator used to generate the sample
covariance matrix. Also, we note that approaches from the low-rank matrix approxi-
mation body of theory have been applied to the moving-target detection problem, for
example in [34].
0
10
Eigenvalue Mag.
10–1
Figure 2.5 Synthetic structured interference eigenvalue decay functions. These functions are
shown with width parameter of 10. For other values, of this parameter, the spectral structure is
stretched or compressed proportionally.
We test four different spectrally structured synthetic data matrices to better understand
the performance of the SVT algorithm under various conditions. For an input rank
width r:
4. The fourth decay funtion has eigenvalues decaying, with the ith eigenvalue
having magnitude
1, $ %
i≤r
σi = exp −(35(i−r)/r)7/10 (2.23)
exp(−(35/r)7/10 )
, otherwise.
Parameter Value
10 –2
10 –3
True Cov.
–4 Full SMI
10 Compr. SMI
Compr. SVT
0 1 2 3
10 10 10 10
Eigenvalue Index
Figure 2.6 The spectral decay structures of the various covariance estimates of the synthetic
structured interference. The SVT estimate more closely matches the truth than the compressed
SMI does, which uses the same data as input.
the CNR is 15 dB, and the rank width parameter is 10. The various covariance estimates
exhibit varying spectral structure and are shown in Figure 2.6.
To illustrate performance over a broader problem space we conducted a parametric
sweep over several relevant problem input variables: under-sampling factor, clutter-to-
noise ratio (CNR), number of range samples (snapshots), and clutter eigenvalue decay
function. These results are all generated using Na = 8 spatial channels and Ns = 128
pulses. The results of this evaluation are shown the following figures.
Figure 2.7 shows how performance varies as the under-sampling factor varies. It is
evident that for low and moderate under-sampling factors the SVT estimator is able to
perform as well as the full-data SMI, but under these same conditions the performance of
compressed SMI falls off sharply. If the experiment is limited to favorable interference
conditions (low rank, simple structure), the SVT estimate performs very well, out to an
under-sampling factor of 20, as shown in Figure 2.8.
Figure 2.9 shows how performance varies as the number of pre-compression range
bins varies. It shows that the SVT estimate is best able to take advantage of a longer
observation interval to improve the estimation accuracy, even if those range bins will
64 Tuuk
1
True Cov.
Full SMI
Probability of Detection
0.8
Compress SVT
Compress SMI
No Cov.
0.6
0.4
0.2
0
1 2 5 10 20
Under-Sampling Factor
Figure 2.7 Average probability of detection as a function of the data under-sampling factor for
cases of synthetic structured interference. As the under-sampling factor increases and less data is
available for the two compressed estimates, the accuracy of those estimates degrades. Notably,
the SVT estimate maintains much better performance than compressed SMI as the USF
increases.
1
Probability of Detection
0.8
0.6
0.4
True Cov.
Full SMI
Compress SVT
0.2
Compress SMI
No Cov.
0
1 2 5 10 20
Under-Sampling Factor
Figure 2.8 Average probability of detection as a function of the data under-sampling factor
for cases with favorable interference structure. The SVT estimate based on the compressed data
performs better than the diagonally loaded estimate based on the full data and nearly as well as the
true covariance matrix. The interference rank is 5 and the structure is simple (structure ID = 1).
be compressed to reduce the number samples, a larger ambient dimension improves the
estimation performance. This somewhat challenges at least one motivation for reducing
the number of samples: non-stationarity interference forcing a reduced range sample
space. But for cases in which acquiring the samples themselves is a greater challenge
Clutter Rejection and Adaptive Filtering 65
Probability of Detection
0.8
True Cov.
Full SMI
0.6 Compress SVT
Compress SMI
No Cov.
0.4
0.2
0
100 200 300 400 500
Number of Range Samples
Figure 2.9 Average probability of detection as a function of the number of range samples for
cases of synthetic structured interference. As the number of range samples (snapshots) increases,
the two compressed estimates, which use compressed data (compressed snapshots), both
improve in estimation accuracy.
1
Probability of Detection
0.8
True Cov.
0.6
Full SMI
Compress SVT
Compress SMI
0.4
No Cov.
0.2
Clutter Structure ID
Figure 2.10 Average probability of detection as a function of interference structure for cases of
synthetic structured interference. The performance of various estimators is tested with four
different clutter structures, having the four eigenspectra illustrated in Figure 2.5. The
performance of the estimators varies as a function of the spectra, with lower rank clutter being
easier to filter out and higher rank clutter being more difficult.
than inherent interference constraints, this result shows that the SVT estimator could
provide improved performance.
Figure 2.10 shows how performance varies over the different clutter eigenspectra
shown in Figure 2.5.
66 Tuuk
Estimating the statistics of clutter from compressed measurements is necessary, but not
sufficient for the detection of targets embedded therein. A variety of techniques exist
for estimation of the scene, including an approaches that uses an estimated interference
covariance matrix and other approaches that do not.
One of the first approaches to this problem, from [36], uses a mask over the presumed
clutter ridge of the angle-Doppler domain. This allows traditional CS-based solution
techniques to be applied outside that mask where targets are presumed sparse. The extent
of the clutter region may be reasonably estimated from platform motion and sensing
geometry. However, this limits minimum detectable velocity and does not readily extend
to electromagnetic interference (EMI) and other structured interference sources.
In [37], the clutter covariance is assumed known and a set of Capon beamforming
weights is built into the sparsity matrix to favor detection of targets and suppress detec-
tion of clutter. The objective vector is assumed to be block-sparse, with high response in
all snapshots. This is combined with an arbitrary random sensing matrix and tools from
convex optimization.
The approach in [38] uses the clutter covariance to modify the optimization norm
or, equivalently, to modify the system model. For a combined sensing and sampling
model, matrix A and interference covariance matrix R with inverse R −1 , which have
a Cholesky decomposition P H P , a modified model matrix Ā = P A, and modified
measurement vector ȳ = P y, are introduced. These can be used in a standard sparse
recovery formulation:
2
x̂ = arg min = Āx − ȳ 2 + λ ||x||1 .
x
This is solved with the FISTA algorithm in the published work, but any desired opti-
mization routine could be applied.
Using the statistics of the clutter to develop a whitening filter is also the approach
taken in [39]. In this case, the whitening filter is derived in the frequency domain without
any spatial degrees of freedom. The resultant technique is applied to an experimental
CS-based radar system and provides improvement up to a factor of the number of pulses.
None of the mentioned CS solution methods, as described, are equipped to handle
structured interference. If R̂ is some estimate of the space–time covariance matrix of the
fully sampled interference, and R̃ = R̂ ⊗ I Nf is its expansion into the full measurement
domain, then the covariance matrix of the interference in the compressed domain can be
expressed as R̂ c = C R̃C H . This use of the dimensionally reduced covariance matrix
is a non-data-adaptive reduced-dimension STAP formulation. The literature of reduced-
dimension STAP provides a natural way to obtain an estimate of the scene from the
compressed measurements while including the covariance information. Define
−1
x̂ cstap = S H C H R̂ c z. (2.24)
This is analogous to the x̂ stap solution in that the matched filter is post-multiplied by
the covariance matrix inverse to whiten the interference. It is identical to the STAP
Clutter Rejection and Adaptive Filtering 67
solution in the case that C is the identity matrix (i.e., the fully sampled case). To compute
this estimate, one incurs the cost of inverting the covariance matrix (or parametrically
estimating the inverse), but gains a good deal of clutter suppression as we will show in
our results. However, this technique does not necessarily favor sparse solutions.
Work in [40,41] proposes a covariance-aware CS (CA CS) that accounts for structured
interference in a CS framework. As in STAP methods, this is accomplished via the
interference covariance matrix inverse. The original 1 regularized least squares problem
of (2.16) can be modified as follows:
−1
x̂ cacs = arg min (z − CSx)H R̂ c (z − CSx) + γ ||x||1 . (2.25)
x
This generalization of (2.16) is identical (for some value of γ) in the clutter-free case.
−1
In that case, R̂ c is a scaled identity matrix. Here γ is set relative to the entries
. $in R̂ c %.
−1
Specifically γ = λ R̂ c , using the Frobenius norm where ||Q||F = Tr QQH
F
and Tr(·) gives the trace of a matrix. This relationship is used so as to provide similar
performance as in (2.16) for a selected value of λ.
The first term in the objective function penalizes deviations from the measurements
in interference-free regions, but less so in interference regions. This term is akin to a
Mahalanobis distance. The second term penalizes large entries in the solution. Thus
0.8
Probability of Detection
0.6
CA CS, usf = 20
Full STAP
0.4 Full Adjoint
CS, usf = 20
CSTAP, usf = 20
Comp. MF, usf = 20
0.2
0
–50 –40 –30 –20 –10 0
Input SCR (dB)
Figure 2.11 The CA CS method subsamples the data just as the standard CS method does,
however it takes into account the covariance matrix that describes the interference structure.
By doing so, it improves the probability of detection over the CS case as well as beyond the
fully sampled, matched filter case that does not use the covariance information. These results
are shown with a probability of false alarm of 0.005.
68 Tuuk
the entries in the interfered region are unconstrained by the first term and they are
allowed to be driven to zero by the second term. The first term maintains fidelity to
the measurements in the clear areas while the second term promotes sparsity. Here the
γ parameter serves the same purpose as λ in (2.16): balancing the weight of the sparsity-
promoting 1 norm against the fidelity-preserving 2 norm.
Figure 2.11 compares the performance of various estimation techniques on problems
with clutter as the dominant interference. Here CA CS shows success beyond that of
the CS estimates. By using the covariance information, CA CS can even surpass the
performance of the fully sampled (but covariance-ignorant) matched filter. Still, the fully
sampled adaptive STAP filter remains the gold standard. Of particular interest is the 20×
under-sampled CA CS solution that achieves nearly the same detection performance
as the STAP solution. Also notable is the fact that the performance of the classically
formulated estimation methods (compressed STAP and compressed adjoint) is equal to
that of the CS-formulated ones (CA CS and CS).
When compared to the compressed STAP (CSTAP) methods, there is no advantage to
using CA CS. It is also evident that the compressed adjoint performs as well as the CS
solution in the interference limited case. This result holds for a range of under-sampling
factors, u, as well as for both the covariance-aware (CA CS vs. CSTAP) and covariance-
unaware (CS vs. compressed adjoint) cases. Furthermore, this performance equivalence
holds over a range of tested probability of false alarm settings.
2.8 Summary
As has been shown, clutter can be of great detriment to terrestrial and airborne radar sys-
tems. And mitigation of clutter can be a significant driver of radar design requirements,
in terms of dynamic range and space–time degrees of freedom. These statements hold
true whether in a classical or CS radar system. A rich body of literature that describs
techniques for mitigating clutter has led to the STAP filter that estimates interference
statistics for a range bin of interest from samples of nearby bins. This estimate is used to
whiten the bin under test prior to subsequent processing. This approach, leavened with
ideas from sparse estimation and low-rank matrix approximation, can be extended to the
compressed domain.
For the estimation of interference covariance information, the mechanisms of low-
rank matrix approximation have given new perspective to prior work in rank-reduced
and dimensionality-reduced STAP estimation. And recent results show that even in the
presence of sampling compression the interference covariance matrix can be accurately
estimated if it has a compact eigen-structure. This structure can be used with singular
value thresholding or other low-rank approximation techniques to recover the full and
the compressed covariance matrices to high accuracy.
And for the covariance information can be used in the recovery of a sparse scene from
compressed measurements. The whitening approach taken in STAP can be extended into
the sparsity-favoring convex objective function by a reshaping of the distance norm. This
allows the full machinery of CS to be applied to the problem.
Clutter Rejection and Adaptive Filtering 69
References
[1] L. Brown, A Radar History of World War II: Technical and Military Imperatives. Institute
of Physics Publishing, 1999.
[2] F. R. Dickey, M. Labitt, and F. M. Staudaher, “Development of airborne moving target
radar for long range surveillance,” Aerospace and Electronic Systems, IEEE Transactions
on, vol. 27, no. 6, pp. 959–972, Nov. 1991.
[3] L. P. Goetz and J. D. Albright, “Airborne pulse-doppler radar,” IRE Transactions on Military
Electronics, vol. MIL-5, no. 2, pp. 116–126, Apr. 1961.
[4] L. Brennan and L. Reed, “Theory of adaptive radar,” Aerospace and Electronic Systems,
IEEE Transactions on, vol. 9, no. 2, pp. 237–252, Mar. 1973.
[5] M. A. Richards, Fundamentals of Radar Signal Processing. McGraw-Hill, 2005.
[6] D. Shnidman, “Radar detection in clutter,” Aerospace and Electronic Systems, IEEE
Transactions on, vol. 41, no. 3, pp. 1056–1067, July 2005.
[7] D. Shnidman, “Generalized radar clutter model,” Aerospace and Electronic Systems, IEEE
Transactions on, vol. 35, no. 3, pp. 857–865, July 1999.
[8] J. Billingsley, A. Farina, F. Gini, M. Greco, and L. Verrazzani, “Statistical analyses of
measured radar ground clutter data,” Aerospace and Electronic Systems, IEEE Transactions
on, vol. 35, no. 2, pp. 579–593, Apr. 1999.
[9] V. Anastassopoulos, G. Lampropoulos, A. Drosopoulos, and N. Rey, “High resolution radar
clutter statistics,” Aerospace and Electronic Systems, IEEE Transactions on, vol. 35, no. 1,
pp. 43–60, Jan. 1999.
[10] R. McAulay, “A theory for optimal MTI digital signal processing part i. receiver
synthesis,” Massachusetts Institute of Technology Lincoln Laboratory, Lexington, MA,
Tech. Rep. 1972-14, Feb. 1972. [Online]. Available: www.ll.mit.edu/mission/aviation/
publications/publication-files/technical_notes/McAulay_1972_TN-1972-14i_WW-18358
.pdf.
[11] J. Ward, “Space-time adaptive processing for airborne radar,” MIT Lincoln Laboratory,
Lexington, MA, Tech. Rep. 1015, 1994.
[12] W. Melvin, “A STAP overview,” Aerospace and Electronic Systems Magazine, IEEE, vol. 19,
no. 1, pp. 19–35, Jan. 2004.
[13] J. Guerci, Space-Time Adaptive Processing for Radar. Artech House, 2003.
[14] I. Reed, J. Mallett, and L. Brennan, “Rapid convergence rate in adaptive arrays,” Aerospace
and Electronic Systems, IEEE Transactions on, vol. 10, no. 6, pp. 853–863, Nov. 1974.
70 Tuuk
[15] B. D. Carlson, “Covariance matrix estimation errors and diagonal loading in adaptive
arrays,” Aerospace and Electronic Systems, IEEE Transactions on, vol. 24, no. 4, pp. 397–
401, July 1988.
[16] J. Guerci, J. Goldstein, and I. Reed, “Optimal and adaptive reduced-rank STAP,” Aerospace
and Electronic Systems, IEEE Transactions on, vol. 36, no. 2, pp. 647–663, Apr. 2000.
[17] C. Peckham, A. Haimovich, T. Ayoub, J. Goldstein, and I. Reid, “Reduced-rank STAP
performance analysis,” Aerospace and Electronic Systems, IEEE Transactions on, vol. 36,
no. 2, pp. 664–676, Apr. 2000.
[18] R. Fa, R. de Lamare, and L. Wang, “Reduced-rank STAP schemes for airborne radar based
on switched joint interpolation, decimation and filtering algorithm,” Signal Processing,
IEEE Transactions on, vol. 58, no. 8, pp. 4182–4194, Aug. 2010.
[19] J. Tropp, J. Laska, M. Duarte, J. Romberg, and R. Baraniuk, “Beyond Nyquist: Efficient
sampling of sparse bandlimited signals,” Information Theory, IEEE Transactions on, vol. 56,
no. 1, pp. 520–544, Jan. 2010.
[20] L. Zhen, W. Xizhang, and L. Xiang, “CS-based moving target detection in random PRI
radar,” in Geoscience and Remote Sensing Symposium (IGARSS), 2012 IEEE International,
July 2012, pp. 7476–7479.
[21] L. Carin, “On the relationship between compressive sensing and random sensor arrays,”
Antennas and Propagation Magazine, IEEE, vol. 51, no. 5, pp. 72–81, Oct. 2009.
[22] L. Carin, D. Liu, and B. Guo, “Coherence, compressive sensing, and random sensor arrays,”
IEEE Antennas and Propagation Magazine, vol. 53, no. 4, pp. 28–39, Aug. 2011.
[23] P. P. Vaidyanathan and P. Pal, “Sparse sensing with co-prime samplers and arrays,” Signal
Processing, IEEE Transactions on, vol. 59, no. 2, pp. 573–586, Feb. 2011.
[24] E. J. Candès and B. Recht, “Exact matrix completion via convex optimization,” Foundations
of Computational Mathematics, vol. 9, no. 6, pp. 717–772, Dec. 2009.
[25] E. J. Candès and Y. Plan, “Matrix completion with noise,” Proceedings of the IEEE, vol. 98,
no. 6, pp. 925–936, Apr. 2010.
[26] D. Gross, “Recovering low-rank matrices from few coefficients in any basis,” Information
Theory, IEEE Transactions on, vol. 57, no. 3, pp. 1548–1566, Mar. 2011.
[27] E. J. Candès and T. Tao, “The power of convex relaxation: Near-optimal matrix completion,”
Information Theory, IEEE Transactions on, vol. 56, no. 5, pp. 2053–2080, May 2010.
[28] Y. Chen, Y. Chi, and A. J. Goldsmith, “Exact and stable covariance estimation from
quadratic sampling via convex programming,” Information Theory, IEEE Transactions on,
vol. 61, no. 7, pp. 4034–4059, July 2015.
[29] S. Bahmani and J. Romberg, “Near-optimal estimation of simultaneously sparse and low-
rank matrices from nested linear measurements,” Information and Inference: A Journal of
the IMA, vol. 5, no. 3, p. 331, Sept. 2016.
[30] L. Bai, S. Roy, and M. Rangaswamy, “Compressive radar clutter subspace estimation using
dictionary learning,” in Radar Conference (RADAR), 2013 IEEE, Apr. 2013, pp. 1–6.
[31] K. Sun, H. Zhang, G. Li, H. Meng, and X. Wang, “A novel STAP algorithm using
sparse recovery technique,” in Geoscience and Remote Sensing Symposium, 2009 IEEE
International, vol. 5, July 2009, pp. V-336-V-339.
[32] K. Sun, H. Meng, Y. Wang, and X. Wang, “Direct data domain STAP using sparse
representation of clutter spectrum,” Signal Processing, vol. 91, no. 9, pp. 2222–2236, Sept.
2011.
[33] F. Pourkamali-Anaraki, “Estimation of the sample covariance matrix from compressive
measurements,” IET Signal Processing, vol. 10, no. 9, pp. 1089–1095, Dec. 2016.
Clutter Rejection and Adaptive Filtering 71
[34] S. Sen, “Low-rank matrix decomposition and spatio-temporal sparse recovery for stap
radar,” Selected Topics in Signal Processing, IEEE Journal of, vol. 9, no. 8, pp. 1510–1523,
Dec. 2015.
[35] J.-F. Cai, E. J. Candès, and Z. Shen, “A singular value thresholding algorithm for matrix
completion,” SIAM Journal on Optimization, vol. 20, no. 4, pp. 1956–1982, Mar. 2010.
[36] I. Selesnick, S. Pillai, K. Y. Li, and B. Himed, “Angle-doppler processing using sparse regu-
larization,” in Acoustics Speech and Signal Processing (ICASSP), 2010 IEEE International
Conference on, Mar. 2010, pp. 2750–2753.
[37] Y. Yu, S. Sun, and A. P. Petropulu, “A capon beamforming method for clutter suppression in
colocated compressive sensing based MIMO radars,” in SPIE Conference Proceedings, vol.
8717, May 2013, p. 87170J. [Online]. Available: https://doi.org/10.1117/12.2015635.
[38] J. T. Parker and L. C. Potter, “A Bayesian perspective on sparse regularization for STAP
post-processing,” in Radar Conference, 2010 IEEE, May 2010, pp. 1471–1475.
[39] Y. C. Eldar, R. Levi, and A. Cohen, “Clutter removal in sub-nyquist radar,” IEEE Signal
Processing Letters, vol. 22, no. 2, pp. 177–181, Feb. 2015.
[40] P. B. Tuuk and S. L. Marple, “Compressed sensing radar amid noise and clutter using
interference covariance information,” Aerospace and Electronic Systems, IEEE Transactions
on, vol. 50, no. 2, pp. 887–897, Apr. 2014.
[41] P. B. Tuuk and S. L. Marple, “Compressed sensing radar amid noise and clutter,” in Signals,
Systems and Computers, 2012 Conference Record of the Forty-Sixth Asilomar Conference
on, Nov. 2012, pp. 446–450.
3 RFI Mitigation Based on Compressive
Sensing Methods for UWB
Radar Imaging
Tianyi Zhang, Jiaying Ren, Jian Li, David J. Greene, Jeremy A. Johnston, and Lam
H. Nguyen
3.1 Introduction
Ultra-wideband (UWB) radar operating at low frequencies (for example, from under
100 MHz to several GHz) has been used in a wide range of applications, including
landmine and unexploded ordinance (UXO) detection using ground-penetrating radar
(GPR), imaging with foliage-penetrating (FOPEN) radar, as well as detecting hidden
humans or objects via through-wall imaging [1]. Examples of such radar systems,
including the US Army Research Laboratory (ARL)’s synchronous impulse reconstruc-
tion (SIRE) system, which is a forward-looking GPR (FLGPR) system, and the ARL’s
BoomSAR system, are shown in [1]. These UWB systems are important to both military
and civilian applications.
One significant challenge for the proper operations of an UWB radar system is deal-
ing with the severe radio frequency interference (RFI) they encounter, since there are
many competing users within the UWB frequency range in which they operate. Typical
RFI sources include FM radio transmitters, TV broadcast transmitters, cellular phones;
their operating frequency bands tend to overlap with that of UWB radar systems [2].
Figure 3.1 shows an example of the spectrum of a radar echo signal from a single pulse
repetition interval (PRI), the spectrum of an RFI source (containing strong narrowband
interferes), and the spectrum of the radar echo signal merged with the RFI. Note that
the presence of RFI causes significant distortions to the spectrum of the received sig-
nal. Therefore, the RFI signal poses a significant hindrance to the proper operation of
UWB radar in terms of reduced signal-to-noise ratio and degraded radar image quality.
Therefore, effective RFI mitigation is critically important for the proper functioning of
a UWB radar system.
RFI mitigation is a notoriously challenging problem because RFI signals are dif-
ficult to predict and model accurately due to their dynamic range and diverse mod-
ulation schemes. Many methods have been developed for RFI mitigation, including
RFI suppression via filtering techniques and RFI extraction based on RFI estimation.
The former suppression approaches, including notch-filtering, subband filtering and
adaptive filtering, though popular due to their simplicity, usually introduce sidelobes
in the time-domain and suffer from filter transients and reduced data length [2–6]. The
latter RFI extraction methods are composed of techniques based on, for example, para-
metric modeling [7], spectral decomposition [8], eigensubspace decomposition [9,10],
72
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 73
160
Original SAR Data
RFI data
140 SAR+RFI
120
Amplitude(dB)
100
80
60
Figure 3.1 Comparison of spectrum obtained from experimental data collected by ARL’s
BoomSAR.
and independent component analysis (ICA) [10,11]. Most of these methods can only
provide satisfying performance under certain assumptions or constraints; most assump-
tions and constraints are often invalid in practice due to the difficulties of modeling
complex RFI sources. For instance, principal-component-based techniques, such as ICA
and eigendecomposition, heavily depend on orthogonal subspaces and have difficulty in
distinguishing between RFI sources and UWB radar echoes when they have similar
power levels within the same subspace [12].
The recent development of compressive sensing (CS) theory has stimulated numerous
investigations on exploiting sparsity and low-rank properties for RFI mitigation. Early
sparsity-based recovery methods solve the RFI mitigation problem by modeling both
the desired UWB radar echo signal and the RFI sources as sparse with respect to well-
designed dictionaries [1,12]. These methods work well but suffer from a significant
drawback: they require an additional dictionary-learning step based on the measured or
estimated data. In [13,14], an improved method based on a joint sparse and low-rank
model is proposed. By simply exploiting the low-rank property of the RFI sources, the
method does not require any specific prior knowledge on the interferences. The robust
principal component analysis (RPCA) approach can be used to extract RFI sources from
the observed data [15]. More specifically, this approach models the RFI contamination
across multiple PRIs within a coherent processing interval (CPI) using a general, low-
rank structure while treating the UWB radar echo signals as sparse impulse outliers.
Unlike the previous CS-based approaches, this technique is completely adaptive to
highly time-varying operating environments and does not require any prior knowledge
on the dictionary for the desired UWB radar signals and the unwanted RFI sources.
Additionally, the RPCA-based RFI mitigation method can be easily incorporated into
diverse UWB radar systems as a preprocessing stage before further signal processing
74 Zhang, Ren, Li, Greene, Johnston, and Nguyen
and image formation [15]. However, RPCA requires the fine-tuning of a user parameter,
which is nontrivial in practical applications due to the lack of prior information on the
RFI sources and the desired radar echo signals.
Since the desired UWB radar echo signals are relatively weak with a flat spectrum and
RFI sources are typically strong, with sparse, narrow spectral peaks, the RFI sources can
be approximately modeled as sparse spectral lines. The parameters of RFI sources can
be estimated using the CLEAN algorithm, which is a conceptually and computationally
simple approach widely used in diverse applications [16–19]. After parameterization,
the RFI sources can be extracted out from the observed RFI-contaminated data. More-
over, the CLEAN algorithm can be used with the Bayesian information criterion (BIC)
[20–22] to estimate the number of spectral lines needed to model the RFI sources.
Unlike the aforementioned RPCA method, the CLEAN-BIC approach does not require
the selection of any user parameter and hence can be easily used in practical applica-
tions. However, the recovered UWB radar echoes that were obtained by the CLEAN-
BIC approach are not sparsified, and the resulting radar images appear noisy in the
presence of severe RFI and noise contaminations.
To take advantage of the merits of both RPCA and CLEAN-BIC, we consider a hybrid
method, referred to as HM. We first utilize the CLEAN-BIC algorithm to estimate the
signal-to-interference ratio (SIR) of the received RFI-contaminated data. If the estimated
SIR is above a certain threshold, we use CLEAN-BIC to recover the desired UWB radar
echo signals. Otherwise, we use RPCA with the user parameter recommended in [23]
for RFI mitigation.
Furthermore, we introduce a framework of choosing the proper user parameter for the
RPCA method for RFI mitigation via first using CLEAN-BIC to estimate the SIR. The
RPCA algorithm, with its user parameter determined by the estimated SIR, is then used
for RFI mitigation. We refer to this approach as RPCA-CB.
Finally, both simulated and experimental results are presented to evaluate the RFI mit-
igation performance of the aforementioned algorithms in the presence of various levels
of RFI contaminations. Specifically, our experiments are conducted using the measured
RFI-free radar echo data set with two different RFI data sets: a simulated RFI-only
data set and measured RFI-only data set. The measured UWB data set was collected by
the ARL using their impulse-based, low-frequency UWB BoomSAR system covering
a frequency band from approximate 50 MHz to 1150 MHz. The UWB BoomSAR was
mounted on a platform that emulated an airborne geometry. Two transmitters and two
receivers were used to collect data in different polarizations. The measured data used
in these experiments was configured in a horizontal transmit, horizontal receive (HH)
polarization. The simulated RFI-only data set has the RFI sources modeled as a sum
of sinusoids, whereas the measured RFI-only data set is collected by the ARL radar
receiver with the antenna pointing toward Washington, DC [1,12]. We show that RPCA-
CB outperforms the aforementioned RPCA, CLEAN-BIC, and HM algorithms.
Notation: Most of the notation used in this chapter is listed in the list of Symbols
in the Preface. For this chapter, we need to add some definitions. det (·) denotes the
determinant of a matrix. x k· and x ·k refer to the k-th row and k-th column of matrix
X, respectively. R ∈ CN×M denotes the complex-valued N × M matrix. For a matrix,
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 75
·p means the p element-wise norm of this matrix, ·F is the Frobenius norm of a
matrix. · ∞ means the infinite norm of a matrix and · ∗ denotes the nuclear norm of
a matrix, that is, the sum of the singular values of the matrix. The subscript of I denotes
/ to denote the estimate
the size of the identity matrix. To avoid confusion, we also use (·)
of a parameter. X,Y = tr(X Y ) denotes the inner product of two matrices X and Y .
T
Y = X + R, (3.1)
where each column y ·m of Y denotes a data vector within a PRI with N fast-time
samples, composed of the desired UWB radar echo signal x ·m and the RFI-only signal
r ·m , m = 1,2,. . .,M. Here, m is the PRI index, i.e., the slow-time index. Our goal is
to extract the desired UWB radar signal X and the RFI signal R from the observed
RFI-contaminated data Y .
We exploit two main assumptions to accomplish the RFI mitigation task: (1) the
desired UWB radar echo matrix X is sparse and (2) the RFI matrix R is low rank [15].
The sparse nature of the desired UWB radar echo matrix is confirmed and illustrated in
Figure 3.2, where a typical example of the UWB impulse radar echoes from multiple
PRIs, i.e., across multiple slow-time indices, within a CPI, as measured by an ARL
radar receiver in the absence of RFI, is depicted. These measured UWB radar echo
signals contain occasional narrow backscattered pulses, which indicate the distances
between the targets and the radar and the reflection coefficients of the targets. Due to
the sparsity of strong targets, X is in general quite sparse in the fast-time domain. It
is worth mentioning that this sparse property of the radar echo signal is also valid for
stepped-frequency or chirp UWB radar systems, since the echoes received by these radar
systems can be easily converted into sparse narrow pulses through pulse compression.
The low-rank property of the RFI sources has been observed and utilized in [13,14].
Figure 3.3 gives the fast-time RFI-only spectra from multiple PRIs within a CPI using
data measured by the ARL radar. RFI sources, such as FM radios, digital TV and cellular
phones, tend to have strong sparse peaks in the fast-time frequency domain, whereas, in
contrast, the entire frequency band of the UWB radar system is occupied by the radar
echoes. Figure 3.4 shows that the singular values of the measured RFI-only data matrix
R decreases rapidly and the bottom 90% of these singular values are zero or close to
zero, confirming again that R possesses the low-rank property. Within a small time
window, such as within a CPI, the low-rank property of the RFI sources appears to be
due to the sinusoidal carrier approximations resulting from various modulation schemes
popular in today’s wireless communications systems [15].
76 Zhang, Ren, Li, Greene, Johnston, and Nguyen
(dB)
0
200
–5
400
600 –10
Fast-Time Samples
800
–15
1000
1200 –20
1400
–25
1600
–30
1800
2000
–35
200 400 600 800 1000 1200 1400 1600 1800
Slow-Time Index
Figure 3.2 An example of a measured ARL UWB impulse radar echo signal in the fast-time vs.
slow-time domain in the absence of RFI.
10 8 (dB)
15 0
–5
–10
–15
10
Frequency (Hz)
–20
–25
–30
5
–35
–40
–45
0 –50
200 400 600 800 1000 1200 1400 1600 1800
Slow-Time Index
Figure 3.3 An example of fast-time RFI spectrum vs. slow-time index of the RFI-only data
measured by the ARL radar receiver.
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 77
12
10
8
Singular Value
0
0 200 400 600 800 1000 1200 1400 1600 1800
Figure 3.4 The singular values of the RFI-only matrix R measured by the ARL radar receiver.
3.2.2 RPCA
RPCA has been widely used to recover the low-rank and sparse components from their
mixtures [23]. A natural choice of extracting X and R from the collected data Y is to
use the following RPCA optimization metric [15]:
where · ∗ denotes the nuclear norm that promotes the low rank property, · 1
represents the 1 element-wise norm, which is a sparsity-enforcing metric, and ζ is a
user parameter used to balance the trade-off between the two components. The user
parameter ζ is recommended to be ζ 0 = √max{N,M}
1
in [24].
Following from [24], we employ the augmented Lagrange multiplier (ALM) method
to solve the RPCA problem efficiently. First, we rewrite the RPCA problem (3.2) as
follows:
min ||R||∗ + ζ||X||1
(3.3)
s.t. Y − X − R = 0.
There are two types of ALM algorithms that can be used to solve the RPCA problem:
the exact ALM algorithm and the inexact ALM algorithm [24]. Compared with the
inexact ALM algorithm, which updates X∗k and R ∗k once when solving the sub-problem:
the exact ALM method, though performing sightly better due to using the iterative
thresholding approach to solve this sub-problem, requires a much longer computation
time even for moderate N and M. Therefore, we focus herein on using the inexact
ALM method. The detailed steps of the inexact ALM algorithm are summarized in
Algorithm 1 [24]. More implementation details can be found in [24].
In this section we provide initial evaluations of the RFI mitigation performance of
the RPCA algorithm using the measured data collected by ARL’s BoomSAR system.
(Further performance evaluations are given in Section 3.5.) The measured data consists
of the measured RFI-free radar echo data (see Figure 3.2) and the measured RFI-only
data (see Figure 3.3). We scale the RFI-only data based on the desired SIR value, i.e.,
||X||2F
, and add the scaled version of the RFI-only data to the RFI-free data X to obtain
||R||2F
the RFI-contaminated data Y .
By using all the data within the CPI, we can form synthetic aperture radar (SAR)
images. Figure 3.5 shows the original RFI-free SAR image that was obtained from
the data in Figure 3.2. Figure 3.6 shows the RFI-contaminated SAR image that was
obtained from Y for SIR = −10 dB. Figures 3.7 and 3.8 compare the recovered
SAR images after RFI mitigation using the RPCA method with different ζ values for
SIR = −10 dB. Visually, for SIR at −10 dB, RPCA with ζ = 0.4ζ0 significantly
outperforms RPCA with the recommended ζ0 , since the recovered SAR image that
was obtained by the latter is much sparser than the original RFI-free SAR image, with
many small targets missing. Consider now the case of more severe RFI; Figure 3.9
shows the RFI-contaminated SAR image for SIR = −30 dB. Figures 3.10 and 3.11
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 79
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.5 Original RFI-free SAR image obtained by using the back projection algorithm on the
measured RFI-free ARL radar data.
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.6 SAR image of the measured SAR data set contaminated by measured RFI signals with
a SIR of −10 dB.
80 Zhang, Ren, Li, Greene, Johnston, and Nguyen
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.7 Recovered SAR image from the measured SAR data set contaminated by measured
RFI signals with a SIR of −10 dB obtained by using RPCA with ζ = 0.4ζ0 .
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.8 Recovered SAR image from the measured SAR data set contaminated by measured
RFI signals with a SIR of −10 dB, obtained by using RPCA with ζ = ζ0 .
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 81
(dB)
40 0
50 –5
60 –10
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.9 SAR image from the measured SAR data set contaminated by measured RFI signals
with a SIR of −30 dB.
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.10 Recovered SAR image from the measured SAR data set contaminated by measured
RFI signals with a SIR of −30 dB obtained by using RPCA with ζ = 0.4ζ0 .
82 Zhang, Ren, Li, Greene, Johnston, and Nguyen
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.11 Recovered SAR image from the measured SAR data set contaminated by measured
RFI signals with a SIR of −30 dB obtained by using RPCA with ζ = ζ 0 .
compare the recovered SAR images after RFI mitigation using the RPCA method with
different ζ values for SIR = −30 dB. Note that most of the target features are still
retained in the resulting SAR image that was obtained by RPCA with ζ = ζ 0 , whereas
the SAR image obtained by RPCA with ζ = 0.4ζ 0 has a fairly high noise level, with
only a few strong targets discernible. Therefore, it appears that the choice of the user
parameter ζ has a significant impact on the RFI mitigation performance of the RPCA
algorithm and the resulting SAR image quality. Thus, a fine-tuning of ζ based on the
SIR value is warranted. However, this parameter selection is not straightforward in
practical applications, since the SIR value is usually not known a priori.
with other uncertainties and approximations, are modeled as noise. Within a PRI, the
observed data, denoted by vector y ∈ RN , can be modeled as follows:
y = x + r, (3.6)
where r ∈ RN denotes the RFI signal, which can be modeled as a sum of sinusoids:
Kc
r(n) = ak cos(ωk n + φk )
k=1
Kc 0
ak ak −j (ωk n+φk ) 1
= ej (ωk n+φk ) + e (3.7)
2 2
k=1
2Kc
= ãk ej (ω̃k n+φ̃k ), n = 0,1,. . .,N − 1
k=1
with r(n) denoting the n-th element of r, ak ,ωk ,φk denoting the amplitude, frequency
and phase of the k-th RFI source, and Kc denoting the number of the RFI sources.
Since r(n) is real-valued, we state that ãk = ãKc +k = a2k , ω̃k = −ω̃Kc +k = ωk , φ̃k =
−φ̃Kc +k = φk , k = 1,. . .,Kc . In (3.6), x ∈ RN denotes the desired UWB radar echo
signal within one PRI plus other uncertainties and approximations, and x will be treated
as additive noise when estimating the sinusoidal parameters in r. Our goal is to recover
x by estimating the sinusoidal parameters in r and removing the estimated r̂ of r from
the RFI-contaminated data y.
To reduce approximation errors, the sinusoidal parameters are allowed to change from
one PRI to another. Hence the CLEAN-BIC algorithm presented in the next subsection
is applied to one PRI at a time.
3.3.2 CLEAN-BIC
To remove the RFI sources, modeled as a sum of sinusoids, from the observed RFI-
contaminated data, we first estimate the sinusoidal parameters of these RFI sources.
CLEAN, with its name meaning “cleaning the undesired component from the data,” is
a simple and robust sinusoidal parameter estimation algorithm [18,19], first introduced
in [16]. From as early as 1988, CLEAN has been used for the extraction of a single
interference signal [17]. Since then, several works have considered using extensions of
CLEAN for RFI mitigation [25–28].
The basic idea of CLEAN is to first estimate the parameters of the strongest sinusoid
by minimizing the following nonlinear least-squares (NLS) criterion:
wH ( ω̃1 )y
α̃ˆ 1 = ˆ1
. (3.10)
N ω̃1 =ω̃
Hence, ω̃ˆ 1 is obtained as the location of the dominant peak of the scaled periodogram,
|w ( ω̃1 )y|2 , which can be efficiently computed by using the FFT with the data sequence
H
where p(y, θ̂) denotes the likelihood function of y, θK = [γTK ,σK 2 ]T denotes the
unknown parameter vector of y, γ K = [a1,. . .,aK ,ω1,. . .,ωK , φ1,. . .,φK ]T denotes
the unknown sinusoidal parameter vector, and J K is the following Fisher information
matrix (FIM):
∂ 2 ln p(y,θK )
J K = −E . (3.14)
∂ θK ∂ θTK
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 85
||y − r̆ K ||22
− 2 ln p(y,θK ) = N ln 2π + N ln σK
2
+ , (3.15)
σK
2
where
2
1
N K
σ̂K
2
= y(n) − âk cos(ω̂k n + φ̂k ) . (3.19)
N
n=1 k=1
Note that in the case of sinusoidal signals, we have the following approximation for
sufficiently large values of N [20,31,32]:
K N Ĵ K K N ≈ K N J K K N = O(1), (3.20)
where
2 3
1
I
N 3/2 K
0
KN = 1 , (3.21)
0 I
N 1/2 2K+1
ln[det(Ĵ K )] = ln[det(K −2
N )] + ln[det(K N Ĵ K K N )]
= (2K + 1) ln N + 3K ln N + O{1} (3.22)
= (5K + 1) ln N + O{1}.
Hence, the BIC metric takes on the following form [20] for our problem of interest:
BIC(K) = N ln σ̂K
2
+ 5K ln N, (3.23)
Algorithm 2 (CLEAN-BIC)
104
4.9
4.8
4.7
4.6
BIC(K)
4.5
4.4
4.3
4.2
0 50 100 150 200 250 300
K
Figure 3.12 BIC curve obtained via CLEAN-BIC for one PRI using the RFI-contaminated data
measured by the ARL radar for SIR = −30 dB.
Note that we need to iterate the algorithm until the iteration number Kmax is reached,
since the BIC curve is usually not a smooth convex function of K. Figure 3.12 shows
BIC(K) versus K for one PRI via CLEAN-BIC on the RFI-contaminated data that was
measured by the ARL radar when SIR = −30 dB.
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 87
10 8 (dB)
15 0
–5
–10
–15
10
Frequency (Hz)
–20
–25
–30
5
–35
–40
–45
0 –50
200 400 600 800 1000 1200 1400 1600 1800
Slow-Time Index
Figure 3.13 Fast-time RFI spectrum vs. slow-time index of the noise-free, simulated RFI data.
We now evaluate the performance of the CLEAN-BIC algorithm using the same
measured data set as used in Section 3.2. First, we simulate a set of noise-free RFI
data using 10 sinusoids. Figure 3.13 shows the fast-time spectrum versus slow-time
index image of the simulated RFI sources. The RFI-contaminated data is generated by
combining the measured radar echo data with the scaled simulated RFI data based on
the desired SIR. Figure 3.14 shows the contaminated SAR image at SIR = −10 dB.
Figure 3.15 shows the recovered SAR image at SIR = −10 dB. Figure 3.16 shows the
contaminated SAR image at SIR = −30 dB. Figure 3.17 shows the recovered SAR
image at SIR = −30 dB. Comparing Figures 3.16 and 3.17 with the original RFI-free
SAR image in Figure 3.5, we note that for the noise-free simulated RFI data, most targets
in the recovered SAR images are discernible and the noise levels of the recovered SAR
images are low, meaning that CLEAN-BIC does an excellent job of eliminating the
simple simulated and noise-free RFI sources even at very low SIR levels.
Next, we use CLEAN-BIC on the UWB radar echo data contaminated by the mea-
sured RFI signal (see Figure 3.3). Figures 3.18 and 3.19 show the recovered SAR images
after RFI mitigation using CLEAN-BIC for SIR = −10 dB and −30 dB, respectively.
Visually, for SIR = −10 dB, CLEAN-BIC performs well because it removes most of
the RFI signals and retains most targets. However, with a SIR of −30 dB, the recovered
SAR image that was obtained by CLEAN-BIC suffers from a high noise level, as
compared to its RPCA counterpart in Figure 3.11. The high noise level is due to the
fact that the measured RFI data inevitably contains noise (including thermal noise,
model approximation errors, and other uncertainties), which is amplified when the SIR
88 Zhang, Ren, Li, Greene, Johnston, and Nguyen
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.14 SAR image of the measured SAR data set contaminated by simulated RFI signals
with a SIR of −10 dB.
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.15 Recovered SAR image of the measured SAR data set contaminated by simulated RFI
signals with a SIR of −10 dB, obtained with CLEAN-BIC.
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 89
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.16 SAR image of the measured SAR data set contaminated by simulated RFI signals
with a SIR of −30 dB.
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.17 Recovered SAR image of the measured SAR data set contaminated by simulated RFI
signals with a SIR of −30 dB obtained by using CLEAN-BIC.
90 Zhang, Ren, Li, Greene, Johnston, and Nguyen
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.18 Recovered SAR image of the measured SAR data set contaminated by measured RFI
signals with a SIR of −10 dB obtained by using CLEAN-BIC.
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.19 Recovered SAR image of the measured SAR data set contaminated by measured RFI
signals with a SIR of −30 dB obtained by using CLEAN-BIC.
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 91
is decreased. Moreover, unlike RPCA, CLEAN-BIC does not make any attempt to
sparsify the radar echo signal, and returns a recovered SAR image with a high noise
level, especially when the SIR is low.
that RPCA with the recommended ζ0 works well when the SIR is low. We have also
shown that CLEAN-BIC does a satisfying job at high SIR. Therefore, if the estimated
/ that was obtained via CLEAN-BIC
SIR is high, such as −10 dB or higher, we use the R
to recover the desired UWB radar echo signals via X/ = Y − R. / If the estimated SIR is
lower than −10 dB, we then use RPCA with the recommended ζ0 to obtain X. /
3.4.2 RPCA-CB
We further consider using CLEAN-BIC to determine the user parameter ζ for RPCA,
since it appears that ζ should be tuned based on the SIR. This method is referred to
as RPCA-CB (RPCA with its user parameter determined via CLEAN-BIC). Similar to
the HM method, we first use CLEAN-BIC to obtain R. / Then, the user parameter ζ
for RPCA is selected based on the estimated SIR. The process of RPCA-CB can be
illustrated by Figure 3.20. More details on the selection of ζ based on the estimated SIR
will be depicted in Section 3.5.
Figure 3.20 A flow graph depicting the connections between RPCA and CLEAN-BIC in
RPCA-CB.
92 Zhang, Ren, Li, Greene, Johnston, and Nguyen
We evaluate the RFI mitigation performance of the aforementioned four methods and
compare them to each other in this section, using both simulated and experimentally
measured data.
10 8 (dB)
15 0
–5
–10
–15
10
Frequency (Hz)
–20
–25
–30
5
–35
–40
–45
0 –50
200 400 600 800 1000 1200 1400 1600 1800
Slow-Time Index
Figure 3.21 Fast-time RFI spectrum vs. slow-time index for the noise-free realistically simulated
RFI data.
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 93
the CLEAN-BIC algorithm on the measured ARL RFI-only data. Note that the realis-
tically simulated RFI-only data utilized herein is much more complex than that used in
Section 3.3 and consists of about 80 sinusoids per PRI. We note that Figure 3.21 resem-
bles its counterpart, obtained from measured data. All data sets have N = 2048 fast-time
samples and M = 1892 slow-time samples.
The received RFI-contaminated radar data is obtained by adjusting the power of the
RFI-only data based on the desired SIR and adding it to the RFI-free UWB radar echo
signal. Since the measured RFI-only data set contains noise, as we lower the SIR, we
are increasing the noise level of the received RFI-contaminated radar data as well. Note
that we have no prior information on the noise in the measured data.
ZF
SNR Metric = 20log10 4 4 , (3.24)
4Z − /
Z4 F
Table 3.1 The SNR metric of the recovered SAR images obtained by using RPCA with data obtained from
the ARL RFI-free radar data and the scaled simulated RFI-only data.
Table 3.2 The SNR metric of the recovered SAR images obtained by using RPCA with data obtained from
the ARL RFI-free radar data and the scaled measured ARL RFI-only data.
1.2
1.1
1
Parameter Coefficient β
0.9
0.8
0.7
0.6
0.5
–30 –25 –20 –15 –10 –5 0
SIR
Figure 3.22 Exponential curve to fit the relationship between the SIR of the observed data and the
corresponding proper user parameter ζ. Data was obtained from combining the ARL RFI-free
radar data with the scaled measured ARL RFI-only data (ζ = βζ0 , where β is the parameter
coefficient).
Note that the relationship between the proper ζ and the SIR should be nonlinear
because the appropriate ζ should approach infinity as the SIR approaches negative
infinity, and zero as the SIR approaches infinity. A natural choice is to use the following
exponential equation to express the relationship between the proper ζ denoted as ζ =
βζ0 and the SIR:
β = ηexp(−γSIR), (3.25)
where η and γ are the parameters that can be obtained by using, for example, the MAT-
LAB Curve Fitting Toolbox. Figure 3.22 gives the curve fitting for the RFI-contaminated
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 95
Table 3.3 The SNR metric of recovered SAR images obtained by using the
CLEAN-BIC algorithm. Data was obtained from combining the ARL RFI-free radar
data with the scaled, realistically simulated RFI-only data.
Table 3.4 The SNR metric of recovered SAR images obtained by using the
CLEAN-BIC algorithm. Data was obtained from combining the ARL RFI-free radar
data with the scaled, measured ARL RFI-only data.
data that was obtained using the measured RFI-only data with the measured RFI-free
radar echo data, based on (3.25) (η = 0.5027, γ = 0.02761). It can be demonstrated that
the curve in Figure 3.22 also works well for the realistically simulated RFI-only data.
For the RPCA-CB algorithm, we use this curve to determine the RPCA user parameter
based on the estimated SIR that was obtained via CLEAN-BIC.
Table 3.5 The estimated SIR obtained via CLEAN-BIC for data obtained from combining the ARL RFI-free
radar data with the scaled realistically simulated and measured ARL RFI-only data.
True SIR (dB)
0 −10 −20 −30
Estimated SIR (dB)
Simulated RFI 1.6975 −8.7983 −18.8360 −26.5150
Measured RFI 1.8753 −8.3260 −16.1593 −19.6084
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.23 Recovered SAR image of the measured SAR data set contaminated by realistically
simulated RFI-only signals with a SIR of −10 dB obtained by using HM.
Figures 3.23–3.26 show the recovered SAR images using these two enhanced algo-
rithms with data obtained from combining the measured ARL radar data with realisti-
cally simulated RFI-only data for SIR = −10 dB and SIR = −30 dB, respectively. Com-
pared with the RFI-free case in Figure 3.5, we see that the enhanced algorithms work
well for both strong and weak realistically simulated RFI cases. Figure 3.27 compares
the SNR metric of the recovered SAR images for various SIR levels for the realistically
simulated RFI case. Clearly, RPCA achieves a relatively low SNR value with a SIR
value higher than −10 dB, and CLEAN-BIC has a low SNR value in the case of severe
RFI. However, the enhanced algorithms, which combine the merits of both RPCA and
CLEAN-BIC, has a satisfying performance over a wide range of SIR levels. Moreover,
the RPCA-CB algorithm has the best performance among all four algorithms.
Figures 3.28–3.31 present the recovered SAR images using these two enhanced algo-
rithms with the RFI-contaminated data that was obtained by combining the measured
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 97
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.24 Recovered SAR image of the measured SAR data set contaminated by realistically
simulated RFI-only signals with a SIR of −10 dB obtained by using RPCA-CB.
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.25 Recovered SAR image of the measured SAR data set contaminated by realistically
simulated RFI-only signals with a SIR of −30 dB obtained by using HM.
98 Zhang, Ren, Li, Greene, Johnston, and Nguyen
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.26 Recovered SAR image of the measured SAR data set contaminated by realistically
simulated RFI-only signals with a SIR of −30 dB obtained by using RPCA-CB.
12
10
4
SNR(dB)
RPCA
0
CLEAN–BIC
–2
HM
–4
RPCA–CB
–6
–8
–30 –25 –20 –15 –10 –5 0
SIR(dB)
Figure 3.27 Comparison of RFI mitigation performances for various SIR values using the
simulated RFI-only data and the measured ARL RFI-free UWB radar data.
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 99
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.28 Recovered SAR image of the measured SAR data set contaminated by measured-only
RFI signals with a SIR of −10 dB obtained by using HM.
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.29 Recovered SAR image of the measured SAR data set contaminated by measured
RFI-only signals with a SIR of −10 dB obtained by using RPCA-CB.
100 Zhang, Ren, Li, Greene, Johnston, and Nguyen
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.30 Recovered SAR image of the measured SAR data set contaminated by measured
RFI-only signals with a SIR of −30 dB obtained by using HM.
(dB)
40 0
50 –5
60 –10
Down Range (meters)
70 –15
80 –20
90 –25
100 –30
110 –35
–740 –720 –700 –680 –660 –640 –620 –600 –580 –560 –540
Cross Range (meters)
Figure 3.31 Recovered SAR image of the measured SAR data set contaminated by measured
RFI-only signals with a SIR of −30 dB obtained by using RPCA-CB.
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 101
12
10
4
SNR(dB)
2
RPCA
0
CLEAN
–2 HM
RPCA-CB
–4
–6
–8
–30 –25 –20 –15 –10 –5 0
SIR(dB)
Figure 3.32 Comparison of RFI mitigation performances for various SIR values using the
measured ARL RFI-only data and the measured ARL RFI-free UWB radar data.
RFI-free SAR data with the scaled measured RFI-only data for SIR = −10 dB and
SIR = −30 dB, respectively. Similar to the realistically simulated RFI case discussed
in the previous paragraph, the enhanced algorithms perform better than their RPCA
counterparts (see Figures 3.8 and 3.11) and their CLEAN-BIC counterparts (see Figures
3.18 and 3.19). Figure 3.32 compares the RFI mitigation performance using the SNR
metric of the recovered SAR images using all four methods for various SIR values for
the case of measured RFI data. Again, the enhanced algorithms outperform their RPCA
and CLEAN-BIC counterparts. Note also that RPCA-CB outperforms HM slightly at
high SIR values.
3.6 Conclusions
In this chapter, we proposed several sparse signal recovery methods for effective RFI
mitigation. We first demonstrated that the RFI sources are low-rank and sparse in the
frequency domain; in contrast, the desired UWB radar echoes are sparse in the time
domain. RPCA can be used to exploit these properties for effective RFI mitigation;
however, RPCA requires fine tuning of a user parameter that is nontrivial in practical
applications due to the lack of prior knowledge about the RFI sources and UWB radar
echoes. To avoid the user parameter tuning problem, we introduced the CLEAN-BIC
algorithm for RFI mitigation, via modeling the RFI sources within a PRI as the sum
102 Zhang, Ren, Li, Greene, Johnston, and Nguyen
of sinusoids. We have shown that CLEAN-BIC can be used to remove dominant RFI
sources effectively; however, since the sparse property of the desired UWB radar echoes
is not utilized by CLEAN-BIC, the resulting SAR images contain high noise levels,
especially at low SIR levels. To combine the merits of both RPCA and CLEAN-BIC, we
proposed a hybrid method, or HM, wherein the use of RPCA or CLEAN-BIC is based
on the estimated SIR value obtained via CLEAN-BIC. Furthermore, we have considered
determining the user parameter for the RPCA algorithm based on using the estimated
SIR value that was obtained via CLEAN-BIC; the resulting algorithm is referred to as
RPCA-CB. These enhanced algorithms were applied to RFI-contaminated data that was
obtained from combining the measured RFI-free radar echo data with either realistically
simulated or experimentally measured RFI-only data for performance evaluations. We
have shown that the estimated methods outperformed both RPCA and CLEAN-BIC.
Finally, RPCA-CB has been shown to outperform HM slightly, especially at high SIR
levels.
3.7 Acknowledgment
This material is based upon work supported in part by the US Army Research Labora-
tory and the US Army Research Office under grant number W911NF-16-2-0223.
References
[1] L. H. Nguyen, T. Tran, and T. Do, “Sparse models and sparse recovery for ultra-wideband
SAR applications,” IEEE Transactions on Aerospace and Electronic Systems, vol. 50, no. 2,
pp. 940–958, 2014.
[2] T. Koutsoudis and L. A. Lovas, “RF interference suppression in ultrawideband radar
receivers,” in Algorithms for Synthetic Aperture Radar Imagery II. International Society
for Optics and Photonics, 1995, pp. 107–119.
[3] D. O. Carhoun, “Adaptive nulling and spatial spectral estimation using an iterated principal
components decomposition,” in 1991 International Conference on Acoustics, Speech, and
Signal Processing, 1991, pp. 3309–3312.
[4] H. Subbaram and K. Abend, “Interference suppression via orthogonal projections: a
performance analysis,” IEEE Transactions on Antennas and Propagation, vol. 41, no. 9,
pp. 1187–1194, 1993.
[5] V. T. Vu, T. K. Sjögren, M. I. Pettersson, and L. Håkasson, “An approach to suppress
RFI in ultrawideband low frequency SAR,” in Radar Conference, 2010 IEEE. IEEE, 2010,
pp. 1381–1385.
[6] X. Luo, L. M. H. Ulander, J. Askne, G. Smith, and P. O. Frolind, “RFI suppression in ultra-
wideband SAR systems using LMS filters in frequency domain,” Electronics Letters, vol. 37,
no. 4, pp. 241–243, 2001.
[7] T. Miller, L. Potter, and J. McCorkle, “RFI suppression for ultra wideband radar,” IEEE
Transactions on Aerospace and Electronic Systems, vol. 33, no. 4, pp. 1142–1156, 1997.
[8] X. Wang, W. Yu, X. Qi, and Y. Liu, “RFI suppression in SAR based on approximated spectral
decomposition algorithm,” Electronics Letters, vol. 48, no. 10, pp. 594–596, 2012.
RFI Mitigation Based on Compressive Sensing Methods for UWB Radar Imaging 103
[9] F. Zhou, R. Wu, M. Xing, and Z. Bao, “Eigensubspace-based filtering with application
in narrow-band interference suppression for SAR,” IEEE Geoscience and Remote Sensing
Letters, vol. 4, no. 1, pp. 75–79, 2007.
[10] F. Zhou and M. Tao, “Research on methods for narrow-band interference suppression
in synthetic aperture radar data,” IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, vol. 8, no. 7, pp. 3476–3485, 2015.
[11] F. Zhou, M. Tao, X. Bai, and J. Liu, “Narrow-band interference suppression for SAR
based on independent component analysis,” IEEE Transactions on Geoscience and Remote
Sensing, vol. 51, no. 10, pp. 4952–4960, 2013.
[12] L. H. Nguyen and T. D. Tran, “Efficient and robust RFI extraction via sparse recovery,”
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 9,
no. 6, pp. 2104–2117, 2016.
[13] L. H. Nguyen, M. D. Dao, and T. D. Tran, “Estimating unknown sparsity in compressed
sensing,” in 2014 IEEE International Conference on Image Processing (ICIP), 2014,
pp. 116–120.
[14] L. H. Nguyen, M. D. Dao, and T. D. Tran, “Joint sparse and low-rank model for radio-
frequency interference suppression in ultra-wideband radar applications,” in 2014 48th
Asilomar Conference on Signals, Systems and Computers, 2014, pp. 864–868.
[15] L. H. Nguyen and T. D. Tran, “RFI-radar signal separation via simultaneous low-rank and
sparse recovery,” in 2016 IEEE Radar Conference (RadarConf), 2016, pp. 1–5.
[16] J. A. Högbom, “Aperture synthesis with a non-regular distribution of interferometer
baselines,” Astronomy and Astrophysics Supplement Series, vol. 15, p. 417, 1974.
[17] J. Tsao and B. D. Steinberg, “Reduction of sidelobe and speckle artifacts in microwave
imaging: The CLEAN technique,” IEEE Transactions on Antennas and Propagation,
vol. 36, no. 4, pp. 543–556, 1988.
[18] P. T. Gough, “A fast spectral estimation algorithm based on the FFT,” IEEE transactions on
Signal Processing, vol. 42, no. 6, pp. 1317–1322, 1994.
[19] J. Li and P. Stoica, “Efficient mixed-spectrum estimation with applications to target feature
extraction,” IEEE transactions on signal processing, vol. 44, no. 2, pp. 281–295, 1996.
[20] P. Stoica and Y. Selen, “Model-order selection: a review of information criterion rules,”
IEEE Signal Processing Magazine, vol. 21, no. 4, pp. 36–47, 2004.
[21] T. Yardibi, J. Li, P. Stoica, M. Xue, and A. B. Baggeroer, “Source localization and
sensing: A nonparametric iterative adaptive approach based on weighted least squares,”
IEEE Transactions on Aerospace and Electronic Systems, vol. 46, no. 1, pp. 425–443, 2010.
[22] P. Stoica and P. Babu, “On the proper forms of BIC for model order selection,” IEEE
Transactions on Signal Processing, vol. 60, no. 9, pp. 4956–4961, 2012.
[23] E. J. Candès, X. Li, Y. Ma, and J. Wright, “Robust principal component analysis?” Journal
of the ACM, vol. 58, p. 11, 2011.
[24] Z. Lin, M. Chen, and Y. Ma, “The augmented lagrange multiplier method for exact recovery
of corrupted low-rank matrices,” arXiv preprint arXiv:1009.5055, 2010.
[25] P. A. Fridman and W. A. Baan, “RFI mitigation methods in radio astronomy,” Astronomy &
Astrophysics, vol. 378, no. 1, pp. 327–344, 2001.
[26] A. Camps, J. Gourrion, J. Miguel Tarongí et al., “RFI analysis in smos imagery,” in
Geoscience and Remote Sensing Symposium (IGARSS), 2010 IEEE International. IEEE,
2010, pp. 2007–2010.
104 Zhang, Ren, Li, Greene, Johnston, and Nguyen
[27] X. Peng, F. Hu, F. He, D. Zhu, Y. Cheng, H. Hu, and T. Zheng, “An improved clean
algorithm for RFI mitigation in aperture synthesis radiometers,” in 2017 IEEE International
Geoscience and Remote Sensing Symposium (IGARSS). IEEE, 2017, pp. 3448–3451.
[28] F. Hu, X. Peng, F. He, et al., “RFI mitigation in aperture synthesis radiometers using a
modified clean algorithm,” IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 1,
pp. 13–17, 2017.
[29] G. Schwarz, “Estimating the dimension of a model,” The Annals of Statistics, vol. 6, no. 2,
pp. 461–464, 1978.
[30] J. Rissanen, “Modeling by shortest data description,” Automatica, vol. 14, no. 5,
pp. 465–471, 1978.
[31] P. Stoica and R. L. Moses, Introduction to Spectral Analysis. Prentice Hall, 1997, vol. 1.
[32] P. M. Djuric, “Asymptotic MAP criteria for model selection,” IEEE Transactions on Signal
Processing, vol. 46, no. 10, pp. 2726–2735, 1998.
4 Compressed CFAR Techniques
Laura Anitori and Arian Maleki
4.1 Introduction
In this chapter we study the problem of target detection from a set of compressive radar
measurements that are corrupted by additive white Gaussian noise. The complications in
the calculation of false alarm and detection probabilities that are caused by the nonlinear
nature of target recovery schemes in compressed sensing represent a major limitation in
the application of such techniques in real radar systems. In this chapter, we show how
recent advances in the asymptotic analysis of the recovery algorithms help us overcome
this challenge.
We first review some standard tools that are used in classical radar detection, such
as the matched filter and the constant false alarm rate (CFAR) processor. Then, we
clarify the main challenges one would face in implementing such schemes for target
detection from compressed measurements. We show how the recently developed
asymptotic analyses of recovery algorithms enable us to overcome these challenges, and
develop fully adaptive target detection schemes. Finally, we evaluate the performance
of these schemes through theoretical analyses and extensive simulations on synthetic
and experimental data.1
where C is a complex constant including the target radar cross section (RCS), phase
terms and propagation effects. The time delay variable t can be mapped to the range via
1 This chapter is based on chapters 2, 3, and 4 of [1] and ©[2013] IEEE. Reprinted, with permission,
from [2].
2 The extension to azimuth and Doppler domains can be easily obtained in a similar way; see for example
[3,4].
105
106 Anitori and Maleki
the equation t = 2rc , where c is the speed of light and r is range (or distance). If there are
k point targets located at ranges ri ,i = 1,. . .,k, corresponding to time delays τi = 2rc i ,
the target reflectivity distribution can be expressed as x(t) = ki=1 xi δ(t − τi ), where
xi is the ith target RCS [5–7]. Hence, the (complex) baseband, received signal can be
rewritten as
k
y(t) = xi a(t − τi ), (4.2)
i=1
where xi ,i = 1,. . .,k is a (complex) amplitude proportional to, among others, the target
RCS, target distance and transmitted power [8]. In the remainder of this chapter we
consider |xi |2 as the power received from a target at position i.
It is often convenient to discretize the range and represent signals in vector forms.
To this end, let the vector x represent the target response (or scene) at discretized range
bins,3 i.e., r = [r1,. . .,rn ], with r1 = cT2min , rn = cTmax 2 = r1 + n R, where R is
the range bin size. Furthermore, assume that targets can only be present at locations
corresponding exactly to discrete grid points. Using the Nyquist sampling theorem, the
received signal y(t) is sampled at a rate fs ≥ B, where B is the bandwidth of the
transmitted signal. Then, the sampled received signal y(tl ),l = 1,. . .,L in (4.2) can be
rewritten in vector form as
k
y = Ax = xi a i , (4.3)
i=1
where each column a i ,i = 1,. . .,n of the matrix A is a time delayed version of the
sampled transmitted waveform corresponding to the received signal from a target at
range bin i (the ith model), and x is a length n vector with amplitude xi at indices i
corresponding to target located at ranges ri and zero elsewhere. Taking the noise into
account, we obtain
y = Ax + n, (4.4)
where we consider n ∼ CN (0,σ2 I ).4
3 Because of the relation between time delay and range, we will use the two interchangably.
4 In case of clutter or correlated noise, prewhitening filters are often applied to obtain the formulation in
(4.4). In that case, the matrix A includes also the pre-whitening.
Compressed CFAR Techniques 107
the range resolution is given by δR = cT2 . Hence, improving the range resolution
requires shortening the pulse duration, and results in reduced transmitted energy (for
a given fixed peak power). A common way to improve resolution without reducing the
pulse length is to use frequency or phase modulated pulses such as linear frequency
modulation (LFM) or chirp waveform, Barker codes, and pseudorandom noise (PN)
sequences. In this case, the output of the MF is a compressed pulse with resolution
δR = 2B c
, where B is the bandwidth of the transmitted pulse. Because of this property,
the operation of matched filtering is referred to as pulse compression. The SNR gain of
the MF after pulse compression is given by the time bandwidth product BT . Although
pulse compression by matched filtering results in significant range resolution improve-
ment compared to unmodulated pulses, one of the issues that needs to be addressed is
the sidelobe level (SLL). For example, the autocorrelation function of an LFM pulse
exhibits large sidelobes (about −13 dB) with respect to the mainlobe [10]. Large side-
lobes of strong targets may result in masking of weaker targets in a multiple targets
scenario and in a severe increase of the false alarm rate.
where T is the received signal length. In other words, GLRT computes the envelope (or
power) of the MF output at all discrete time delays and compares it with a threshold
to determine the presence (declare hypothesis H1 ) or absence (declare hypothesis H0 )
of a target at time delay τi . Using the discrete linear model introduced in (4.4), we can
rewrite the test statistic in (4.5) as
x̂ MF = AH y. (4.7)
The envelope (or power) of each component of the vector x̂ MF is compared to the
threshold γ to decide upon the presence of a target. For the envelope detector, the
threshold γ should be set to
γ = −σ 2 ln α (4.8)
It can be seen in (4.9) that each entry x̂ i of the MF output is the sum of the true target
response at location i, xi , plus the interference caused by the presence of other possible
targets at locations j != i ( nj=1,j !=i xj a H
i a j ), plus Gaussian noise.
Note that the interference is proportional to the cross-correlation between the
time-delayed version of the transmitted waveform [14]; in the ideal case that the
time-delayed versions of the transmitted waveforms are orthogonal and have unit norm,
i.e., AH A = I , we obtain
x̂ MF = x + z, (4.10)
where z ∼ CN (0,σ2 I ).
Therefore, if each target is exactly on a grid point and the matrix A is orthogonal,
the components of the vector x̂ MF are independent of one another and each range bin
can be treated independently. One can then apply hypothesis tests to each time delay to
estimate simultaneously the number of targets and their time delays.
For practical frequency or phase-modulated waveforms, the orthogonality condition
is never met. Hence, each target interferes, through its sidelobe, with the detection of the
others. This phenomenon is known as target masking. A reduction of the SLL can be
accomplished by applying a weighting function during matched filtering [9]. However,
the weighted MF output is no longer matched to the transmitted signal, and therefore,
in addition to reducing the SLL, the output SNR is also reduced. Alternatively, one can
design waveforms with low SLLs [15–17] or use different pulse compression filters,
such as adaptive pulse compression (APC) [5,6] or mismatched filters [18–22]. A review
of mismatched filters can be found in [14]. Note that the design of mismatched filters is
based on iterative algorithms and the optimum filter weights have to be estimated sepa-
rately for each range bin. Therefore, such schemes are computationally more demanding
than the MF.
Performing an independent binary GLRT at each range bin does not take into
account the interaction between targets. Hence when the sidelobes are not sufficiently
Compressed CFAR Techniques 109
where ε is a threshold proportional to the noise variance. The relation between λ and ε
that makes the two problems equivalent is data-dependent, and does not have an explicit
form.
Standard techniques, such as interior point method, can be used for solving (4.12).
However, the computational costs of such methods have encouraged researchers to con-
sider iterative algorithms with inexpensive per-iteration computations. See for example
[43,44] and references therein. These iterative algorithms exploit the fact that the opti-
mization problem arg minx 21 u − x22 + λx1 has an explicit solution x = η(u;λ)
(|u|− λ)ej u 1(|u| > λ), where η(·;λ) is called the complex soft thresholding function
and 1 is the indicator function. The complex soft thresholding function acts (componen-
twise) on the amplitudes of the input vector u and produces a sparse signal by shrinking
to zero all the elements of u whose amplitude is below the threshold, or regularization
parameter, λ, thus enforcing sparsity on the solution. The components of u that are
above the threshold will be shrinked toward zero by an amount equal to the threshold
λ, and their phase is unchanged. The complex soft thresholding function is shown in
Figure 4.1.
Iterative soft thresholding (IST) uses the following iterations to solve (4.12). Starting
with x̂ 0 = 0, at each iteration t the estimate x̂ of the vector x is updated using
& '
x̂ t+1 = η x̂ t + AH (y − Ax̂ t );λ . (4.14)
Therefore, at each iteration, the current residual is projected along the waveforms, and
added to the previous solution. In other words, at each iteration of (4.14), the algorithm
Compressed CFAR Techniques 111
moves in the negative gradient direction (of the objective function) −AH (y − Ax̂ t ),
and then applies the soft thresholding function to enforce sparsity on the solution. At
convergence, the estimated sparse signal x̂ will contain many zero components, and a
few non–zeros. As λ increases, the solution to (4.14) will become more sparse and, if
λ > AH y∞ , the only feasible solution is x̂ = 0.
It is clear that 1 -minimization is implicitly performing target detection; we can inter-
pret the nonzero elements of x̂ as detected targets. However, note that if the recovered
amplitude at a target location, say x̂ i , is zero, then there is no way to recover the target,
not even in any subsequent processing stage. In other words, if the estimated amplitude
resulting from 1 -norm minimization is zero, then the target is simply lost, i.e., we have
a missed detection. Furthermore, at locations not containing the targets, we expect most
of the estimates resulting from 1 -norm minimization to have zero amplitude. However,
depending on the SNR and the ratio m/n, some of the noise samples are reconstructed
with non–zero amplitude, and therefore, if no further processing is applied, they will
produce false targets in the recovered range profile. Increasing the threshold λ will
reduce the number of reconstructed noise samples, but it may also result in suppressing
the targets. In this sense, the sparse estimate produced by 1 -norm minimization is
comparable to the output of a classical detector, with the parameters λ or ε used in
the recovery controlling both the detection and false alarm probabilities.
Despite this similarity with classical radar detection, we should emphasize that the
relation between λ and false alarm probability in each cell is extremely complicated. In
particular, especially for m < n, the interference among targets’ sidelobes is not negli-
gible and is not easy to model either due to the nonlinear nature of the 1 -minimization.
These complications make the design of CFAR schemes very challenging for CS radar
systems. In the rest of this chapter, we will focus on this problem and describe how the
asymptotic theory of compressed sensing can help us address this challenge.
detector in the CS framework, one has to deal with a number of issues that are related
mostly to the non-linearity of the 1 -recovery. As explained in the previous sections,
classical radar architectures (without CS) use well-established signal processing and
detection schemes, such as MF and CFAR processors. Unlike classical radars, in which
the relation between the threshold parameter and the false alarm is straightforward,
in CS the relation between λ (or ε) and the false alarm rate is data-dependent and
complicated. For the design of possible CS radar detectors, it is essential to determine
the (statistical) properties of the signal obtained from 1 -norm minimization algorithms.
To solve the CS detection problem, we employ some of the recent advances in the
asymptotic analysis of compressed sensing algorithms. We will show that a debiasing
step can make the problem of CS radar detection similar to that of classical radar
detection. We will then combine this debiasing step with classical tools such as CFAR to
build practical CS radar detection schemes. A key ingredient for obtaining the debiasing
step is the complex approximate message passing (CAMP) algorithm [44], which will
be introduced in the next section.
The complex approximate message passing [44] is an iterative algorithm for solving
(4.12). CAMP is an extension of the original approximate message passing (AMP)
algorithm, first proposed in [45] for real-valued signals, to the case of signals and
measurements in the complex domain. The AMP algorithm and its properties have been
thoroughly investigated in [46–52]. However, in radar it is most common to work with
signals in the complex domain. Therefore, in the remainder of this chapter we will
concentrate exclusively on CAMP. As we will see in the next sections, properties of
CAMP enable us to achieve the following objectives:
Table 4.1 Pseudocode for the ideal CAMP algorithm. ©[2013] IEEE.
Reprinted, with permission, from [2].
Input: y, A, τ, x, δ = m/n
Initialization x̂ 0 = 0, z0 = y
for t = 1 : maxiter
x̃ t = A† zt−1 + x̂ t−1
σt = std(x̃ t − x)
1 $ ∂ η (x̃ t ;τσ ) + ∂ η (x̃ t ;τσ )%
R I
zt = y − Ax̃ t−1 + zt−1 2δ ∂xR t ∂xI t
x̂ t = η(x̃ t ;τσt )
end
Output: x̃, x̂,σ∗
next subsection how the parameter τ of CAMP relates to the regularization parameter
λ in (4.12).
We now explain each variable in the CAMP algorithm:
(i) x̂ t : the estimate of x at iteration t. Under a proper tuning of τ, x̂ t → x̂(λ) as
t → ∞. The relation between τ and λ is given later in (4.17).
(ii) x̃ t : the non-sparse and “noisy” estimate of x. Define the vector wt = x̃ t − x,
which represents the “noise vector” at iteration t of CAMP. The distribution of wt
is “close” to a zero-mean Gaussian probability density function. This property of
wt will be clarified and proved in Section 4.5.1.
(iii) σt : the standard deviation of wt . Furthermore, we define σ∗ limt→∞ σt .
Using this terminology, the operations executed in CAMP can be explained as fol-
lows. First, CAMP calculates a noisy, non-sparse estimate of the signal x, which is
given by x̃ t = A† zt−1 + x̂ t−1 . Then, to make this estimate sparse, the soft thresholding
function is applied to x̃ t to obtain the sparse vector x̂ t = η(x̃ t ;τσt ). This algorithm is
referred to as ideal CAMP. The word ideal in the algorithm’s name refers to the fact
that the sought x is used inside the iterations of CAMP for estimating the noise standard
deviation σt . Therefore although this is not a practical algorithm, since it uses in its
iterations the vector that is trying to estimate, this assumption is only used here for the
clarity of presentation. A practical scheme for estimating σt is described in Section 4.7.
Theorem 4.2 has been proved for Gaussian measurement matrices. However, empiri-
cal studies have already confirmed that this theoretical prediction holds for other sensing
matrices with i.i.d. elements other than Gaussian [45]. Also, in Section 4.8.1 we study
the validity of Theorem 4.2 for partial Fourier matrices, which are of particular interest
in radar applications.
Using the Gaussianity of the noise vector wt , we can answer the second question
raised at the beginning of this subsection, regarding the theoretical performance of the
algorithm. In fact, given that the noise has a zero-mean Gaussian distribution, to predict
the performance of CAMP we only need to track the standard deviation of the noise σt
across the iterations of the algorithm. This is performed through what is called the state
evolution (SE). Under the asymptotic setting, the value of the standard deviation at time
t + 1 is calculated from σt according to the following equation:
σt+1
2
= (σt ), (4.15)
where
1 & 2 '
(σt ) = σ2 + E η(X + σt Z;τσt ) − X , (4.16)
δ
Z ∼ CN (0,1), σ2 is the input noise variance and the expectation is with respect to
the two independent random variables X ∼ pX and Z, where pX denotes the marginal
distribution of x. It has been proved in [44] that the function is a concave function
of σt2 , and therefore the iteration (4.15) has at most one stable fixed point, which we
refer to as σ∗2 . Also, CAMP converges to this fixed point exponentially fast (linear
convergence according to optimization literature). An example of how the function
can be calculated in closed form for a given distribution pX of X is provided in appendix
A of [2].
Finally, let us answer the third question we raised at the beginning of this subsection
regarding the connection between LASSO and CAMP. In [44] it is proved that if τ
satisfies
R
1 ∂η ∂ ηI
λ τσ∗ 1− E (X + σ∗ Z;τσ∗ )+ (X + σ∗ Z;τσ∗ ) , (4.17)
2δ ∂xR ∂xI
then CAMP with threshold τ solves the LASSO in (4.12) with parameter λ.
Compressed CFAR Techniques 115
Using the properties of the CAMP algorithm described earlier, in this section we propose
two CS target detection schemes and analyze their performance through their receiver
operating characteristic (ROC) curves. Let k be the number of targets, i.e., the number
of nonzero coefficients in x, and define G as the distribution of the nonzero elements
of x. In this section k, G, and σ are assumed to be known. A more realistic case in which
none of these parameters are known will be studied in Section 4.7.
The two architectures we consider are displayed in Figure 4.2. In Architecture 1,
the measurements y are given to a recovery algorithm (say CAMP or LASSO). This
algorithm returns a sparse vector x̂ that has a few nonzero values. In this case the
nonzero elements of x̂ can be considered as detected targets, where the soft thresh-
olding operation inside the recovery algorithm performs the detection function. Since
the estimated x̂ (support and amplitude of the nonzero entries) depends on threshold
parameter τ in CAMP or the regularization parameter λ in LASSO, this parameter
will also automatically control both the false alarm α (or Pf a ) and detection (Pd )
probabilities.
√
proposition 1 Consider the CAMP iteration with threshold τα = − ln α. If
A(n),x(n),w(n) is a converging sequence, then
1
n
lim lim 1{x̂ t !=0,x i =0} = α (4.18)
t→∞ n→∞ n − k i
i=1
almost surely, where 1{.,.} denotes the indicator function. Also, τα is the only value of τ
for which (4.18) holds.
Proof Define z ∼ CN (0,1). According to [53] we have
1
n
1{x̂ i (n)!=0,xi (n)=0} = Prob(|σ∗ z| > τα σ∗ ) = e−τα .
2
lim lim
t→∞ n→∞ n − k
i=1
Note that this proposition does not provide any information on the relation between τ
and the detection probability. This issue will be discussed later.
In contrast to Architecture 1, Architecture 2 is inspired by classical radar detection
schemes. In fact, classical radar detection usually comprises two sequential stages: in the
first stage, called the estimation stage here, a noisy estimate of the signal is computed
often through a matched filtering with the goal of maximizing the output SNR (and
therefore Pd ).5 In the second stage, the noisy estimate obtained from the first stage is fed
to a detection block, whose threshold parameter is set to achieve a predefined Pf a = α.
Using a similar approach, in Architecture 2 we propose to first use CAMP to obtain a
noisy, non-sparse estimate of the signal x̃ = x + w. Similarly to the matched filter in
5 As explained in Section 4.3.2, the MF is optimal (in terms of SNR) only for the case of a single target in
white Gaussian noise. For the case of multiple targets, the optimality is only satisfied if the matrix A is
orthogonal.
116 Anitori and Maleki
(a) Architecture 1
(b) Architecture 2
Figure 4.2 Block diagrams of the proposed architectures for radar detection in the CS framework.
©[2013] IEEE. Reprinted, with permission, from [2].
0.6
δ = 0.6
0.55
δ = 0.2
0.5
0.45
σ*
0.4
0.35
0.3
0.25
0.5 1 1.5 2 2.5 3
τ
Figure 4.3 Fixed point σ∗ versus threshold τ for ideal CAMP with σ = 0.23, δ = 0.2 (dashed
line), and δ = 0.6 (solid line), ρ = 0.1. The nonzero entries in x all have amplitudes equal to 1
and phase uniformly distributed between −π and π. The sensing matrix A has i.i.d. Gaussian
entries. These curves are obtained using the analytical equation derived in appendix A of [2].
©[2013] IEEE. Reprinted, with permission, from [2].
classical radar, in the first stage we aim to maximize the output SNR. This can be done
by setting the CAMP threshold τ to a value that minimizes σ∗ . Figure 4.3 exhibits the
dependence of σ∗ on τ for two distinct values of δ. As it is clear from the figure, there
is a value of τ, say τo , for which σ∗ is minimized. In Section 4.7, we will propose a
simple method for calculating τo .
√
In the second stage, a detection bock with fixed threshold κ = σ∗ − ln(α) is applied
to the noisy estimate x̃ to control the false alarm rate. From the Gaussianity of w, in the
asymptotic setting this choice of κ results in the false alarm probability α as derived in
Proposition 1.
With respect to Architecture 1, Architecture 2 has two properties that are very useful
for practical radar applications. Namely:
1. All the parameters can be optimized and estimated adaptively and efficiently, even
without prior knowledge of k, G, and σ. This will be clarified later.
Compressed CFAR Techniques 117
1
0.8
0.9998
0.6
0.2 0.4 0.6
d
P
0.4
Figure 4.4 ROC curves for Architectures 1 and 2 with δ = 0.6, ρ = 0.1 and σ 2 = 0.05. All the
nonzero coefficients of x have amplitude equal to 1 and phase uniformly distributed between −π
and π. The solid (Architecture 2) and dashed (Architecture 1) lines are obtained from the
theoretical predictions using the SE equation. The dots are obtained by Monte Carlo (MC)
simulations using Ideal CAMP. The sensing matrix for MC simulations is i.i.d. Gaussian.
©[2013] IEEE. Reprinted, with permission, from [2].
For a more general version of this result the reader may refer to [54]. Theorem 4.3 is
proved in [2]. Figure 4.4 exhibits ROC curves for Architectures 1 and 2. The solid and
dashed lines are obtained from the analytical equations derived in appendix A of [2].
The theoretical ROC curves are confirmed by Monte Carlo simulations (dots). In the
simulations, we run the ideal CAMP algorithm given in Table 4.1 for several values of
Pf a ranging from 10−1 to 10−5 . One interesting phenomenon that is observed in Figure
4.4 is that for Architecture 1, in the region around Pf a = 0.4 (the zoomed area), the
probability of detection, Pd,1 , decreases as Pf a increases. This is a counterintuitive
behavior, but can be explained in the following way. Recall that in Architecture 1
the CAMP threshold τα varies with Pf a . Therefore, σ∗ (and hence the CAMP recon-
struction SNR) also changes with Pf a , and is not constant along the ROC curve. This
explains why Pd,1 reaches its maximum at around Pf a = 0.4 and then decreases again
as the Pf a goes to 1 (the SNR is maximized at τ = τo and then it descreases again).
Instead, in Architecture 2, σ∗ is fixed to its minimum along the ROC curve (SNR is
118 Anitori and Maleki
constant), and therefore Pd,2 increases with increasing FAP. From this figure it can also
be observed that, as predicted by Theorem 4.3, Pd,1 = Pd,2 happens at only one value
of FAP (α
= 0.22).
In the ideal CAMP algorithm, we made a few assumptions that are not met in real
applications.
(i) x is assumed to be known. This is required for the calculation of the standard
deviation of the noise.
(ii) ρ, σ, and G are assumed to be known for calculating the theoretical fixed point
solution.
In this section, we show how these assumptions can be removed. The main objective
here is to propose a practical, fully adaptive scheme that does not require any prior
information about the signal and can adapt to the changes in the signal and the noise
level. More specifically, the three main issues that we would like to address are: (i) how
to obtain a good estimate of σt without knowing x; (ii) how to obtain a good estimate of
τo for Architecture 2, efficiently and accurately without using the SE (which depends
on the unknown parameters ρ, σ, and G); (iii) how to replace κ in Architecture 2 with
an adaptive threshold that is able to maintain CFAR property in the multiple targets
scenario.
Several different schemes can be used to answer the first question. For the moment,
suppose that x = 0. If this assumption is true, then according to Theorem 4.2, x̃ t
is a Gaussian random vector with mean zero and variance σt . Hence, the following
expression gives a good estimate of σt .
1
/
σt = median(|x̃ t |). (4.19)
ln 2
However, in the presence of targets, i.e., x != 0, this is a biased estimator. More specif-
i.i.d.
ically, for large values of n, we have x̃ t = x + wt , where xi ∼ G(x) + (1 − )δ(x)
with = δρ 1 and δ(x) denotes a point mass at zero, and the elements of wt look
like a complex Gaussian random variable with mean zero and standard deviation σt .
The goal is to estimate the median u∗ of |wt |. However, since we only have access to x̃,
we estimate the median of |wt | as the û that satisfies Prob(|x̃it | > û) = 12 . Note that
since the median is robust to outliers (i.e., multiple targets), we still expect (4.19) to
offer a good estimate of σt . The following proposition confirms this claim.
proposition 2 [2] The error of the estimated median is bounded from above by
|û − u∗ | | ln(1 − )|
≤ √ . (4.20)
σ∗ 2 ln 2
Note that when the scene is very sparse, i.e., is small, | ln(1 − )| ≈ and, hence,
the error is proportional to the sparsity level. In such cases, the proposed estimator
Compressed CFAR Techniques 119
provides a good estimate of σt . We should also emphasize that, when the number of
targets is large and the median estimator is not reliable, we can use more recent and
better estimates of σt that are proposed in [52]. The algorithm that uses the estimate in
(4.19) instead of σt will be referred to as CAMP or median CAMP in the rest of this
chapeter.
The second question we raised at the start of this section was concerned with estimat-
ing the optimal threshold τo in Architecture 2. Suppose that we know or can estimate
τmax such that τo < τmax . We will discuss how this parameter can be set later. Define a
sequence of thresholds τ = {τ }L =1 such that τ1 = τmax and τ = τ−1 − δ τ , and δ τ
is a user-defined parameter that controls the step size. Starting from τmax , at each new
iteration , CAMP is initialized with x̂ 0 = x̂ −1 and z0 = z−1 . Using the solution of
CAMP at the previous iteration − 1 as an initial value for the current iteration, CAMP
needs only a few iterations to converge to the solution,6 and therefore the entire process
is very fast. After L iterations, we have a matrix of solutions X̂ = [x̂ 1, x̂ 2,. . ., x̂ L ] of
size n × L, where each column contains the CAMP solution for a given τ . Also, we
have L estimates { σ̂∗ }L
=1 . The optimum estimated threshold τ̂o is chosen as the one that
minimizes the estimated CAMP output noise variance σ̂∗2 .
At the first iteration ( = 1 and t = 1) τmax can be set as τmax = AH y∞ / σ̂0 . In
fact, if the CAMP algorithm is initialized with x̂ 0 = 0 and z0 = y, then x̃ = AH y,
where σ̂0 is an estimate of the standard deviation of the noise. In this case, any value of
τ larger than τmax will lead to the same estimate x̂ 1 = 0.
So far we have resolved two of the issues we raised at the beginning of this section,
i.e., estimating σ̂t and τ̂o . We will refer to the resulting algorithm as adaptive CAMP,
since both the noise variance σ̂t and the threshold τ̂o are adaptively estimated. To clarify
the differences between ideal, median, and adaptive CAMP, Table 4.2 shows the input
and output variables for each of the three algorithms. Please recall that ideal CAMP is
not an algorithm that can be used in practice, as it requires the knowledge of the true
vector x.
Figure 4.5 Block diagram of the adaptive CAMP CFAR detector. ©[2013] IEEE. Reprinted, with
permission, from [2].
radar detectors, often a CFAR processor estimates the unknown background plus inter-
ference level. In Architecture 2, since the signal x̃ is modeled as the sum of targets plus
Gaussian noise (just as after a classical MF), this estimate can be directly input to a
conventional CFAR processor. A block diagram of the Adaptive CAMP CFAR detector
based on Architecture 2 is shown in Figure 4.5. Replacing the fixed threshold detector
with a CFAR detector in Architecture 2 provides similar results as in classical CFAR
without CS. All we need to determine is the input/output SNR relations of CAMP, so that
we can use the output SNR in the detector equation for the prediction of Pd . A method
for estimating the output SNR of CAMP for a given problem is provided in [2].
Please note that, since the sparse estimate x̂ contains many zeros, this vector could
not be used in a CFAR processor.
7 Recall that an m × n partial Fourier matrix can be obtained from an n × n discrete Fourier transform
matrix by preserving only a random subset m of the original n matrix rows.
8 Recall that, in a multiple target scenario, the MF SNR is optimum and independent of the number of
targets as long as each target is exactly on a grid point and the matrix A is orthogonal. We assume these
(ideal) conditions are satisfied when computing the MF SNR.
Compressed CFAR Techniques 121
0 0 0 0
−0.4 −0.2 0 0.2 0.4 −0.2 0 0.2 −0.5 0 0.5 −2 0 2
Data Data Data Data
(a) Histogram of the real part of w.
4
2 1 0.5
2
0 0 0 0
−0.2 0 0.2 −0.2 0 0.2 −0.5 0 0.5 −1 0 1
Data Data Data Data
(b) Histogram of the imaginary part of w.
Figure 4.6 Histograms (bars) of (a) the real and (b) imaginary parts of the noise signal w for
different combinations (δ,ρ) using CAMP with threshold τ = 1.8, σ = 0.1 and n = 4000. The
solid line shows a Gaussian distribution fitted to the histograms. The sensing matrix is partial
Fourier. ©[2013] IEEE. Reprinted, with permission, from [2].
δ = 0.05 δ = 0.1
0.3 0.3
Ideal CAMP, Gaussian
0.29 0.29 Ideal CAMP, Fourier
Theoretical SE
0.28 0.28
*
*
σ
σ
0.27 0.27
0.26 0.26
0.25 0.25
2 2.5 3 1.5 2 2.5 3
τ τ
δ = 0.2 δ = 0.5
0.3 0.3
0.29 0.29
0.28 0.28
*
*
σ
σ
0.27 0.27
0.26 0.26
0.25 0.25
1.5 2 2.5 3 1 1.5 2 2.5 3
τ τ
Figure 4.7 σ∗ versus τ using ideal CAMP for both complex Gaussian and partial Fourier sensing
matrices. The empirical curves are obtained for several values of δ by averaging over 100 MC
samples. σ 2 = 0.05,ρ = 0.05, n = 4000. The theoretical SE curve shows the analytical σ∗ .
©[2013] IEEE. Reprinted, with permission, from [2].
(i) SE correctly predicts the performance of CAMP for the Gaussian sensing matrix.
(ii) SE does not predict the performance of CAMP for partial Fourier matrices. How-
ever, for τ = τo , the value of σ∗ for Fourier and Gaussian matrices is very similar.
(iii) As δ → 0 the predictions of SE become more accurate for the partial Fourier
matrix. However, as δ → 1, i.e., as the number of measurements increases, the
columns of the partial Fourier matrix become deterministic and orthogonal, and
hence the true behavior deviates from the SE, which is derived assuming matrices
with i.i.d. entries.
(iv) For the partial Fourier matrix, the optimal threshold τo seems to be almost the
same for different values of δ. Interestingly, although the curves of σ∗ are dif-
ferent for different δ’s, for a fixed δ and for τ > τo the variation of the output
variance is much smaller in the partial Fourier case than in the Gaussian case.
This behavior will have an impact on the difference in performance between
Architecture 1 and Architecture 2 for the case of partial Fourier and Gaussian
matrices, as the SNR of Architecture 1 varies less along the ROC curves and it is
closer to the optimal SNR for the partial Fourier case.
δ = 0.05 δ = 0.1
0.3 0.3
Median CAMP, Gaussian
0.29 0.29 Median CAMP, Fourier
Ideal CAMP, Gaussian
0.28 0.28
Ideal CAMP, Fourier
σ*
σ*
0.27 0.27
0.26 0.26
0.25 0.25
2 2.5 3 1.5 2 2.5 3
τ τ
δ = 0.2 δ = 0.5
0.3 0.3
0.29 0.29
0.28 0.28
*
*
σ
σ
0.27 0.27
0.26 0.26
0.25 0.25
1.5 2 2.5 3 1 1.5 2 2.5 3
τ τ
Figure 4.8 Output noise standard deviation versus τ for both ideal (σ∗ ) and median (/
σ∗ ) CAMP
for complex Gaussian and partial Fourier sensing matrices. The curves are obtained for several
values of δ by averaging over 100 MC realizations. σ2 = 0.05,ρ = 0.05, n = 4000. ©[2013]
IEEE. Reprinted, with permission, from [2].
1 1
0.9
0.95
0.8
0.9
0.7
Pd
d
0.6 0.85
P
0.5 0.8
0.4
0.75
Median CAMP Median CAMP
0.3
Ideal CAMP Ideal CAMP
−4 −3 −2
10 −4
10 −3
10 −2
10 −1 10 10 10 10−1
Pfa Pfa
Figure 4.9 ROC curves for Architecture 1 using both ideal and median CAMP. n = 1000,
δ = 0.6, ρ = 0.1, a 2 = 1, and σ 2 = 0.05 (corresponding to a MF SNR = 13 dB).
©[2013] IEEE. Reprinted, with permission, from [2].
and median CAMP. As expected, we observe that σ̂∗ deviates from the ideal CAMP case
because of the bias introduced by the estimator. Furthermore, the deviation diminishes
as = δρ decreases, as predicted by the upper bound provided in (4.20).
Overestimating /σ∗ in Architecture 1 results in a loss of detection performance. This
is shown in Figure 4.9, where we plot the ROC for Architecture 1 using both ideal
and median CAMP. The loss of detection performance is explained by the fact that
the soft thresholding function in CAMP uses the parameter τα σ̂∗ , which, for a fixed
124 Anitori and Maleki
τα , increases with σ̂∗ . If the overall threshold increases, then the detection probability
will decrease. In Figure 4.9, we also observe that Architecture 1 performs better when
the sensing matrix is the partial Fourier. This has to do with the behavior of the noise
variance σ∗2 , and therefore the SNR, versus the threshold, that changes along the ROC
curve. As can be seen from Figure 4.8, in the partial Fourier case the variance curve
becomes flatter than the curve obtained in the Gaussian case as δ increases. As a result,
when the false-alarm probability decreases, in the partial Fourier case the SNR along
the ROC curves for Architecture 1 deviates much less from the optimum SNR that is
achieved for τ = τo .
1
0.9 0.95
0.8 0.9
0.7 0.85
0.6 0.8
Pd
Pd
0.5 0.75
0.4 Ideal CAMP CA-CFAR
0.7 Ideal CAMP CA-CFAR
Adaptive CAMP FT Adaptive CAMP FT
0.3 Adaptive CAMP CA-CFAR
0.65 Ideal CAMP FT
Ideal CAMP FT Adaptive CAMP CA-CFAR
Theoretical CA-CFAR
0.2 CMF FT 0.6
Theoretical CA-CFAR
CMF FT
CMF CA-CFAR CMF CA-CFAR
0.1 –4 –3 –2 –1
10 10 10 10 10 –3 10 –2 10 –1
Pfa Pfa
Figure 4.10 ROC curves for Architecture 2 with different combinations of CAMP algorithms and
detection schemes. Here n = 1000, δ = 0.6, ρ = 0.1, a 2 = 1, and σ 2 = 0.05 (corresponding to
a MF SNR = 13 dB). FT denotes the use of an (ideal) fixed threshold detector. ©[2013] IEEE.
Reprinted, with permission, from [2].
matrix, the threshold in Ideal CAMP is derived from a plot like the ones shown in Figure
4.7 for the case δ = 0.6, and is equal to τo = 1.85.
As expected, adaptivity imposes extra losses in performance. One loss is due to the
use of adaptive instead of ideal CAMP, which is caused by the error in the estimated
τo . Another loss is caused by the CFAR processor and its estimate of the noise standard
deviation, and this is the well-known CFAR loss.
From Figure 4.10 the following observations can be made. First, adaptive CAMP
introduces almost no loss in the detection performance of Architecture 2. This can be
seen by observing that the perfomance of adaptive and ideal CAMP combined with
the same detector (either fixed threshold or CA-CFAR) are almost indistinguishable.
The reason for this is that, although σ̂∗ estimated in adaptive CAMP is biased, the
value of τ̂o at which the minimum σ̂∗ occurs is very close to the true optimal τo (see
Figure 4.8) computed in ideal CAMP, resulting in an almost optimal SNR even in the
adaptive case. The main loss instead is introduced by the adaptive CFAR detector, as
can be seen by comparing the curves obtained using the fixed threshold detector against
the ones obtained using the CA-CFAR processor, both with ideal or adaptive CAMP.
This loss is however not introduced by CAMP, but it is the well-known CFAR loss
[26]. Furthermore, if one compares the curve of adaptive CAMP plus CFAR with the
theoretical curve of a CA-CFAR processor (without CS) computed analytically using
the same parameters (SNR, CFAR window length, and number of guard cells), it can
be noted that the performance of the CA-CFAR detector appears to be independent
of the fact that the input to the detector is obtained by running CAMP instead of a
conventional MF.
Second, we observe that Architecture 2 significantly outperforms the CMF. This
should be expected since the output of the MF using the subsampled waveform will
result in severe target sidelobes (interference), which in turn leads to both an increase of
the false alarm rate and possibly the masking of weaker targets, there by reducing their
detection probability.
126 Anitori and Maleki
4 4
3.5 3.5
3 3
fa
fa
P
2.5
P
2.5
10
10
−log
−log
2 2
1.5 1.5
1 Arch.1 1 Arch.1
Arch.2 + CA−CFAR Arch.2 + CA−CFAR
(a) (b)
Figure 4.11 Estimated FAP versus design FAP α for Architectures 1 and 2. n = 1000, and
δ = 0.6. (a) Complex Gaussian sensing matrix; (b) Partial Fourier sensing matrix.
By comparing Figures 4.9 and 4.10 it can be seen that, in the fixed threshold case,
Architecture 2 always outperforms Architecture 1, as predicted by Theorem 4.3. Also
in the adaptive case, Architecture 2 followed by a CA-CFAR processor outperforms
Architecture 1 using median-based CAMP. However, the difference between the two
schemes can vary significantly with the system parameters (δ,ρ,σ), sensing matrix
type, and CFAR window length. For instance, for the value of δ used in these figures,
we observe that Architecture 1 performs much better in the Fourier case than in the
Gaussian sensing matrix case. Also, the loss in detection performance is significantly
reduced compared to the adaptive detector. This again depends on the behavior of σ∗2
versus τ. In general, to predict how the two architectures will perform one should
observe the behavior of the output noise variance as a function of the threshold τ. If
the variation of σ∗ versus τ is small, in the ideal detector case the ROC curves of the
two architectures will be almost identical, with Architecture 2 always slightly better.
In Figure 4.11 the estimated FAP is shown for both Architecture 1 (which is nonadap-
tive, and uses median-based CAMP) and Architecture 2, which uses adaptive CAMP in
combination with a CA-CFAR processor. The desired FAP α, on the x-axis, is used to
obtain the threshold multiplier β for the CFAR processor in Architecture 2 and to derive
the value of the fixed CAMP threshold τα in Architecture 1.
As expected, Figure 4.11 shows that, in homogeneous Gaussian noise, the proposed
architectures posses the CFAR property. In simulating the FAP for Figure 4.11, accord-
ing to hypothesis H0 (target absent) we generated a measurement vector y with standard
Gaussian distribution and x = 0. However, in practical scenarios, where the noise level
may change across range, or in the presence of one or multiple targets located anywhere
in the signal x, Architecture 1 can not achieve CFAR. This is because the noise estimate
computed by median CAMP is not performed locally, as in a CFAR processor, but
is based on the whole received signal using the median, which is a biased estimator
if x != 0. In Architecture 2 instead, separating the reconstruction from the detection
stage gives more flexibility, and, e.g., in a multiple interfering targets scenario, the CA
processor can be replaced with a more robust CFAR scheme, such as OS-CFAR.
Compressed CFAR Techniques 127
10 Since we are interested in the detection problem from a single range measurements, we kept the targets
fixed and did not perform any Doppler measurements.
128 Anitori and Maleki
1000 1000
600 600
400 400
200 200
0 0
0 20 40 60 80 100 0 5 10 15 20 25
Time [us] Time [us]
(a) (b)
Figure 4.12 Spectrogram of TX waveform. (a) Nyquist waveform with n = 200. (b) CS waveform
with m = 50 and δ = 0.25.
11 In -norm minimization, as in classical MF, a clustering and interpolation step could be included after
1
detection to improve range estimation and reduce straddeling losses, see, e.g., [62].
Compressed CFAR Techniques 129
T5
104 CAMP Arch. 1 T1 T2
CAMP Arch. 2 T3
MF T4
103
Amplitude [au]
102
101
0 5 10 15 20 25 30 35
Range [m]
Figure 4.13 Estimated range profile using: CAMP Architectures 1 and 2, and MF. For the MF
n = 200; for CS m = 100 and δ = 0.5. The y-axis is in log scale. ©[2013] IEEE. Reprinted,
with permission, from [2].
1600 δ = 0.5
δ = 0.25
1400
*
Estimated σ
1200
1000
800
1.5 2 2.5
τ
measurements. Note that, for the same input noise variance, for δ = 0.25 the output
noise power is always higher than for δ = 0.5, also implying that for the same target
received power the SNR decreases with δ, as predicted by the SE. We also see that the
behavior of the estimated output noise standard deviation resembles the one shown is
Section 4.8.3 for the simulated data.
For obtaining the ROC and FAP curves, since the SNR is very high for all targets (in
all cases above 20 dB), to evaluate the performance of the detectors at medium SNR
values, we added white Gaussian noise to the raw frequency data samples, resulting in
an equivalent MF output SNR of 17.2, 16.6, 14, 10.2 and 26 dB respectively, from the
closest corner to the farthest one.
For estimating the ROC cuves, we adopted the following procedure. The FAP is
estimated by averaging the detections on all other cells, excluding the true targets’ cells
130 Anitori and Maleki
1 1
T1
0.9 0.9
T1
0.8 0.8
T3 T3
0.7 0.7
0.6 0.6 T4
d
d
P
P
0.5 0.5
0.4 0.4
0.3 T4 0.3
0.2
0.2
Arch. 2 CA-CFAR
0.1 Arch. 2 FT 0.1
Arch. 1
0
-3 -2 -1 0
10 10 10 10 10-3 10-2 10-1 100
Pfa Pf
Figure 4.15 ROC curves for Architecture 1 (a) and Architecture 2 (b) using both fixed threshold
(FT) and CA-CFAR detectors. The curves correspond to targets T1, T3, and T4 in Figure 4.13.
Adapted from [2], ©[2013] IEEE.
plus four guard cells (because of straddling). The detections instead are obtained for
each target separatly by counting the detections at the location of the corresponding
target highest peak. Figure 4.15 shows the ROC curves for three of the five targets
(T1, T3, and T4), having different SNRs, for δ = 0.5 and 0.25. The ROC curves are
estimated for each target using both Architecture 1 and Architecture 2. In Architecture
2, CAMP recovery is followed by either a fixed threshold (FT) detector or a CA-CFAR
processor.12 Please note that, since in Architecture 1 the estimated sparse signal x̂ can
never contain more than m out of n nonzero coefficients, FAPs higher than δ cannot be
estimated. Again, we emphasize that in CAMP the reconstruction SNR is the ratio of
the target power to the system plus reconstruction noise power (σ∗2 ). Therefore, while
the total transmitted power remains fixed for δ = 0.5 and 0.25, the reconstruction SNR
for each target depends on the architecture used in addition to the compression factor δ.
As observed in the simulated results, when the number of measurements is reduced the
reconstruction noise variance σ∗ increases, and CAMP SNR decreases. Since a loss in
SNR translates directly into a loss in detection probability, for a given FAP, CAMP will
perform better for larger δ.
In agreement with our theoretical findings, from Figure 4.15 we observe that the
detection probability of Architecture 2 with the fixed threshold detector is always higher
than the one of Architecture 1. Additionally, Architecture 2 followed by a CA-CFAR
processor exhibits a detection performance loss compared to the fixed threshold case.
This is again the CA-CFAR loss.
For the same figure we observe that the two proposed architectures perform very
similarly at very low or very high SNRs. However, at low FAPs and high Pd , which
is the most relevant case in practical situations, Architecture 2 always outperforms
Architecture 1. Also note that Architecture 1 is designed in a way that it is very similar
to an OS-CFAR detector, in which the CFAR window consists of the whole signal, i.e.,
12 For the CA-CFAR processor we used the same parameters as in the simulations, i.e., four guard cells and
a CFAR window of length 20.
Compressed CFAR Techniques 131
3 3
2.5 2.5
2 2
−log10Pfa
−log10Pfa
1.5 1.5
1 1
Figure 4.16 Estimated FAP versus design FAP α for CAMP Architecture 2 using fixed threshold
detector (a), and CA-CFAR detector (b). δ = 0.5 (dashed line) and δ = 0.25 (solid line).
2Nw = m, and it also includes the CUT. Because in these measurements is very small,
the bias introduced by the median estimate used in Architecture 1 is also very small and
therefore the overall noise estimate of median CAMP is better than the noise estimate
of the CA-CFAR detector that is using only 20 bins. However, a serious disadvantage
of Architecture 1 is that, since the entire signal is used in the noise estimation and the
threshold τα is fixed, this scheme can inherently not adapt to local variation of noise
level. This makes Architecture 1 unsuitable for many radar applications. In Architecture
2 instead, one has the flexibility to choose both the most appropriate CFAR processor
and CFAR window length, depending on the specific scenario. This of course will also
depend on the target distribution in range. In fact, if for example multiple targets are
present in the same CFAR window, then Architecture 2 will have a significant loss in
performance, unless more clever CFAR schemes are used. Instead Architecture 1 is
insensitive to the targets’ locations, and therefore would have the same performance.
As we only performed measurements with targets, we are unable to evaluate the
CFAR property of Architecture 1. However, for CAMP Architecture 2, we can demon-
strate that our model x̃ = x + σ∗ w is correct by estimating the FAP from the recon-
structed noisy signal x̃ by excluding the range bins corresponding to the target locations
plus four guard cells. If our model is correct, and the noise in the signal x̃ is Gaussian,
then the estimated FAP should correspond to the design FAP used to set the detector
threshold. This should be true for both the fixed threshold and the CFAR detector. This
is demonstrated in Figure 4.16, where the estimated Pf a is plotted versus the design FAP
α for CAMP Architecture 2 using both the CA-CFAR and the fixed threshold detector.
From this figure we observe that, as expected, the estimated FAP matches the design
one, confirming that our model is correct.
4.10 Conclusions
In this chapter, we studied the problem of CS radar detection. The nonlinearity of target
reconstruction algorithms in CS makes the calculation of false alarm and detection
132 Anitori and Maleki
References
[1] L. Anitori, “Compressive sensing and fast simulations: Applications to radar detection,”
PhD dissertation, Technical University of Delft, 2013.
[2] L. Anitori, A. Maleki, M. Otten, R. Baraniuk, and P. Hoogeboom, “Design and analysis of
compressive sensing radar detectors,” IEEE Trans. Signal Process., vol. 61, no. 4, pp. 813–
827, Feb. 2013.
[3] M. A. Herman and T. Strohmer, “High-resolution radar via compressed sensing,” IEEE
Trans. Signal Process., vol. 57, no. 6, pp. 2275–2284, Jun. 2009.
[4] L. Anitori, M. Otten, and P. Hoogeboom, “Compressive sensing for high resolution radar
imaging,” in Proc. IEEE Asia-Pacific Microwave Conf. (APMC), 2010.
[5] S. M. Song, W. M. Kim, D. Park, and Y. Kim, “Estimation theoretic approach for radar pulse
compression processing and its optimal codes,” Electronic Letters, vol. 36, no. 3, pp. 250–
253, Feb. 2000.
[6] S. D. Blunt and K. Gerlach, “Adaptive pulse compression via MMSE estimation,” IEEE
Trans. Aerosp. Electron. Syst., vol. 42, no. 2, pp. 572–583, Apr. 2006.
[7] 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. 1006–1020, Jun. 2010.
[8] M. I. Skolnik, Radar Handbook. McGraw-Hill, 1970.
[9] M. A. Richards, Fundamentals of Radar Signal Processing. McGraw-Hill, 2005.
[10] L. R. Varshney and D. Thomas, “Sidelobe reduction for matched filter range processing,” in
Proc. IEEE Radar Conf., 2003.
[11] J. V. DiFranco and W. L. Rubin, Radar Detection. Artech House, 1980.
[12] S. M. Kay, Fundamentals of Statistical Singnal Processing: Detection Theory. Prentice-
Hall, 1998.
[13] H. L. V. Trees, Detection, Estimation and Modulation Theory: Part III. John Wiley & Sons,
2001.
[14] P. Stoica, J. Li, and M. Xue, “Transmit codes and receive filters for radar,” IEEE Signal
Process. Mag., vol. 25, no. 6, pp. 94–109, Nov. 2008.
[15] F. F. J. Kretschmer and K. Gerlach, “Low sidelobe radar waveforms derived from orthogonal
matrices,” IEEE Trans. Aerosp. Electron. Syst., vol. 27, no. 1, pp. 92–102, Jan. 1991.
[16] R. L. Frank, “Polyphase codes with good nonperiodic correlation properties,” IEEE Trans.
Inf. Theory, vol. 9, no. 1, pp. 43–45, 1963.
[17] A. Divito, A. Farina, G. Fedele, G. Galati, and F. Studer, “Synthesis and evaluation of phase
codes for pulse compression radar,” Rivista Tecnica Selenia, vol. 9, no. 2, pp. 12–24, 1985.
[18] Y. I. Abramovich and M. B. Sverdlik, “Synthesis of a filter which maximizes the signal-
to-noise radio under additional quadratic constraints,” Radio Eng. Electron. Phys., vol. 15,
pp. 1977–1984, Nov. 1970.
Compressed CFAR Techniques 133
[19] M. H. Ackroyd and F. Ghani, “Optimum mismatched filters for sidelobe suppression,” IEEE
Trans. Aerosp. Electron. Syst., vol. 9, no. 2, pp. 214–218, Mar. 1973.
[20] S. Zoraster, “Minimum peak range sidelobe filters for binary phase-coded waveforms,” IEEE
Trans. Aerosp. Electron. Syst., vol. 16, no. 1, pp. 112–115, Jan. 1980.
[21] A. Zejak, E. Zentner, and P. Rapajic, “Doppler optimised mismatched filters,” IET Electron-
ics Letters, vol. 27, no. 7, pp. 558–560, Mar. 1991.
[22] C. Candan, “On the design of mismatched filters with an adjustable matched filtering loss,”
in Proc. IEEE Radar Conf., 2010.
[23] B. Steenson, “Detection performance of a mean-level threshold,” IEEE Trans. Aerosp.
Electron. Syst., vol. 4, no. 4, pp. 529–534, Jul. 1968.
[24] H. M. Finn and R. S. Johnson, “Adaptive detection mode with threshold control as a function
of spatially sampled clutter-level estimates,” RCA Review, vol. 29, pp. 414–464, Sept. 1968.
[25] G. M. Dillard and C. E. Antoniak, “A practical distribution-free detection procedure for
multiple-range-bin radar,” IEEE Trans. Aerosp. Electron. Syst., vol. 6, no. 5, pp. 629–635,
Sept. 1970.
[26] P. P. Gandhi and S. Kassam, “Analysis of CFAR processors in homogeneous background,”
IEEE Trans. Aerosp. Electron. Syst., vol. 24, no. 4, pp. 427–445, Jul. 1988.
[27] A. D. Vito and G. Moretti, “Probability of false alarm in CA-CFAR device downstream from
linear-law detector,” IET Electronics Letters, vol. 25, no. 25, pp. 1692–1693, Dec. 1989.
[28] R. S. Raghavan, “Analysis of CA-CFAR processors for linear-law detection,” IEEE Trans.
Aerosp. Electron. Syst., vol. 28, no. 3, pp. 661–665, Jul. 1992.
[29] G. B. Goldstein, “False alarm regulation in log-normal and Weibull clutter,” IEEE Trans.
Aerosp. Electron. Syst., vol. 9, no. 1, pp. 84–92, Jan. 1973.
[30] E. Conte, M. Lops, and A. M. Tulino, “Hybrid procedure for CFAR in non-Gaussian clutter,”
IEEE Proc. Radar, Sonar, and Navig., vol. 144, no. 6, pp. 361–369, Dec. 1997.
[31] R. Ravid and N. Levanon, “Maximum-likelihood for Weibull background,” IEEE Proc. Part
F (London), vol. 139, no. 3, pp. 256–264, Jun. 1992.
[32] H. Rohling, “Radar CFAR thresholding in clutter and multiple target situations,” IEEE
Trans. Aerosp. Electron. Syst., vol. 19, no. 4, pp. 608–621, Jul. 1983.
[33] M. A. Khalighi and M. H. Bastani, “Adaptive CFAR processor for nonhomogeneous
environments,” IEEE Trans. Aerosp. Electron. Syst., vol. 36, no. 3, pp. 889–897, Jul. 2000.
[34] M. Sekine, T. Musha, Y. Tomita, and T. Irabu, “Suppression of Weibull-distributed clutters
using a cell-averaging LOG/CFAR receiver,” IEEE Trans. Aerosp. Electron. Syst., vol. 14,
no. 5, pp. 823–826, Sept. 1978.
[35] R. Nitzberg, “Constant-false-alarm-rate signal processors for several types of interference,”
IEEE Trans. Aerosp. Electron. Syst., vol. 8, no. 1, pp. 27–34, Jan. 1972.
[36] S. R. Babu and R. Srinivasan, “Analysis of envelope detected mean level CFAR processors
using importance sampling,” in Proc. IEEE Radar Conf., 2000.
[37] E. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruc-
tion from highly incomplete frequency information,” IEEE Trans. Inf. Theory, vol. 52, no. 2,
pp. 489–509, Feb. 2006.
[38] E. Candès, J. Romberg, and T. Tao, “Stable signal recovery from incomplete and inaccurate
measurements,” Comm. Pure Appl. Math., vol. 59, no. 8, pp. 1207–1223, Aug. 2006.
[39] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1289–
1306, Apr. 2006.
[40] R. Tibshirani, “Regression shrinkage and selection via the LASSO,” J. Roy. Stat. Soc., Series
B, vol. 58, no. 1, pp. 267–288, 1996.
134 Anitori and Maleki
[41] S. Chen, D. Donoho, and M. Saunders, “Atomic decomposition by basis pursuit,” SIAM J.
on Sci. Computing, vol. 20, no. 1, pp. 33–61, 1998.
[42] E. Candès and T. Tao, “Near optimal signal recovery from random projections: Universal
encoding strategies?” IEEE Trans. Inf. Theory, vol. 52, no. 12, pp. 5406–5425, Dec. 2006.
[43] A. Maleki and D. L. Donoho, “Optimally tuned iterative thresholding algorithm for
compressed sensing,” IEEE J. Sel. Topics Sig. Proc., vol. 4, no. 2, pp. 330–341, Apr. 2010.
[44] A. Maleki, L. Anitori, Y. Zai, and R. Baraniuk, “Asymptotic analysis of complex LASSO
via complex approximate message passing (CAMP),” IEEE Trans. Inf. Theory, vol. 59, no.
7, pp. 4290–4308, July 2013.
[45] D. L. Donoho, A. Maleki, and A. Montanari, “Message passing algorithms for compressed
sensing,” Proc. Natl. Acad. Sci., vol. 106, no. 45, pp. 18914–18919, 2009.
[46] D. L. Donoho, A. Maleki, and A. Montanari, “Message passing algorithms for compressed
sensing: I. Motivation and construction,” in IEEE Proc. Inform. Theory Work. (ITW), 2010.
[47] A. Maleki and A. Montanari, “Analysis of approximate message passing algorithm,” in Proc.
IEEE Conf. Inform. Science and Systems (CISS), 2010.
[48] D. L. Donoho, A. Maleki, and A. Montanari, “The noise-sensitivity phase transition in
compressed sensing,” IEEE Trans. Inf. Theory, vol. 57, no. 10, pp. 6920–6941, Oct. 2011.
[49] A. Maleki, “Approximate message passing algorithm for compressed sensing,” PhD disser-
tation, Stanford University, 2010.
[50] M. Bayati and A. Montanari, “The dynamics of message passing on dense graphs, with
applications to compressed sensing,” IEEE Trans. Inf. Theory, vol. 57, no. 2, pp. 764–785,
Feb. 2011.
[51] C. A. Metzler, A. Maleki, and R. G. Baraniuk, “From denoising to compressed sensing,”
IEEE Transactions on Information Theory, vol. 62, no. 9, pp. 5117–5144, 2016.
[52] A. Mousavi, A. Maleki, R. G. Baraniuk et al., “Consistent parameter estimation for lasso
and approximate message passing,” Ann. Statist., vol. 45, no. 6, pp. 2427–2454, 2017.
[53] M. Bayati and A. Montanari, “The LASSO risk for Gaussian matrices,” IEEE Trans. Inf.
Theory, vol. 58, no. 4, pp. 1997–2017, Apr. 2012.
[54] S. Wang, H. Weng, and A. Maleki, “Which bridge estimator is optimal for variable
selection?” arXiv preprint arXiv:1705.08617, 2017.
[55] J. H. G. Ender, “On compressive sensing applied to radar,” J. Signal Process., vol. 90, no. 5,
pp. 1402–1414, May 2010.
[56] L. Anitori, A. Maleki, W. van Rossum, R. Baraniuk, and M. Otten, “Compressive CFAR
radar detection,” in Proc. IEEE Radar Conf., 2012.
[57] S. Shah, Y. Yu, and A. Petropulu, “Step-frequency radar with compressive sampling
SFR-CS,” in Proc. IEEE Int. Conf. Acoust., Speech, and Signal Process. (ICASSP), 2010.
[58] F. J. Massey, “The Kolmogorov–Smirnov test for goodness of fit,” J. Am Statist. Assoc.,
vol. 46, no. 253, pp. 68–78, 1951.
[59] M. A. Davenport, M. B. Wakin, and R. Baraniuk, “The compressive matched filter,” Rice
Univ., ECE Dept., Tech. Rep. TREE-0610, Nov. 2006.
[60] D. P. Meyer and H. A. Mayer, Radar Target Detection: Handbook of Theory and Practice.
Academic Press Inc., 1973.
[61] L. Anitori, M. Otten, and P. Hoogeboom, “Detection performance of compressive sensing
applied to radar,” in Proc. IEEE Radar Conf., 2011.
[62] D. Beker, W. van Rossum, S. Jacobs et al., “On pre-whitening and accuracy in doa estimation
by sparse signal processing on beamformed data,” in Proc. Int. Work. on Compressed
Sensing on Radar, Sonar, and Remote Sensing (CoSeRa), 2016.
5 Sparsity-Based Methods for CFAR
Target Detection in STAP
Random Arrays
Haley H. Kim and Alexander M. Haimovich
5.1 Introduction
Ground moving target indicator (GMTI) radar is an airborne radar mounted on an air-
craft that detects the presence of targets on the ground. One of the main challenges faced
by GMTI radars is the detection of slow-moving targets in the presence of ground clut-
ter interference. Space–time adaptive processing (STAP) implementation with antenna
arrays has been a classical approach to clutter cancellation in airborne radar [1]. One of
the challenges with STAP is that the minimum detectable velocity (MDV) of targets is
a function of the baseline of the antenna array: the larger the baseline (i.e., the narrower
the beam), the lower the MDV. Unfortunately, increasing the baseline of a uniform linear
array (ULA) entails a commensurate increase in the number of elements.
Instead of using a large ULA and localizing targets by beamforming [2], one
may consider a smaller ULA, but use more sophisticated localization algorithms,
such as Capon’s method [3], MUSIC [4], or ESPRIT [5]. All three methods are
capable of resolving targets within the Rayleigh resolution limit, whereas conventional
beamformering cannot. MUSIC and ESPRIT, however, require knowledge of the
number of targets. This information is rarely known to a radar and must be obtained
by other means, such as using the Akaike information criteria (AIC) or the minimum
description length (MDL) [6,7]. Unfortunately, methods such as the AIC or the MDL
do not allow one to control the false alarm rate, a basic requirement in radar. In addition,
all three methods require a large number of snapshots, which usually are not available
in STAP applications.
An alternative approach to increasing the resolution of a radar, but without using
a large number of sensors is to use a large, but sparsely populated array. In a sparse
array, the sensors are placed across a large array with interelement spacing greater than
half a wavelength in a nonuniform manner to avoid grating lobes. There are typically
two methods used to determine the sensor positions of a sparse array. The first method
involves solving an optimization problem to determine the positions of the sensors such
that the resulting beampattern meets some specifications [8,9]. Another method is to
simply decide the sensor positions randomly. In random arrays [10,11], sensors are
randomly placed across a large array aperture. Since the resolution of the radar depends
mostly on the size of the aperture [10], a radar utilizing a sparse array may achieve a high
angular resolution with significantly fewer sensors than a ULA. Unfortunately, sparse
arrays do not come without drawbacks. Due to the spatial under-sampling, the array
135
136 Kim and Haimovich
beampattern suffers from high sidelobes. During the beamforming stages of STAP, these
high sidelobes may cause a significant increase in false alarms [12]. It was shown in [13]
that sparse arrays for which sensor locations are determined either by optimization or
randomly yield similar peak sidelobe levels. In this chapter, we focus our attention only
to random arrays. Moreover, one may envision applications in which the array elements
cannot be controlled, for example, an array constituted of unmanned aerial vehicles
(UAV) [14].
In [15], Carin demonstrates that measurements from random arrays are consistent
to projection measurements that can be utilized by compressive sensing (CS) [16]. This
suggests that the user may reap the full benefits of a large random array without worrying
that the high sidelobes unnecessarily increase the false alarm rate. The goal of CS is to
recover the signal of interest x, given the received data vector y and a linear model
y = Ax + e, where A is a measurement matrix and e is an interference vector. If the
signal x is known to be sparse (i.e., contains K nonzero elements where K is much
smaller than the number of entries in x), the K sparse solution (a solution with at most
K nonzero entries) may be found solving the nonconvex optimization problem
miny − Ax22
x
(5.1)
subject to x0 ≤ K,
where x0 counts the numbers of nonzero elements in x.
The optimization problem in (5.1) is nonconvex, and therefore only approximate
solutions can be obtained. One approach to obtaining an approximate solution is to
first convert the optimization problem (5.1) into the following optimization problem
To transform (5.2) into a convex optimization problem, the nonconvex term x0 is
replaced by the convex term x1 . Here, λ is a regularization parameter that controls the
sparsity of the solution in x. For a specific choice of λ, numerous algorithms in literature,
such as [16–21], are capable of solving (5.2) in polynomial time. These algorithms are
often referred to as basis pursuit (BP). In [22,23] the authors use this approach to solve
the sparse localization problem. Nevertheless, without the ability to control false alarms,
this approach does not lend itself to radar applications. The authors in [24] argue that
a constant false alarm rate (CFAR) radar may be obtained by properly designing the
regularization parameter λ. However, in [25], the authors point out that the output noise
distribution is unknown and unpredictable, which makes BP unsuitable for designing
CFAR radars.
Another approach to solving (5.1) is to use matching pursuit (MP) algorithms
[26–30]. MP belong to the class of greedy algorithms, which search iteratively one
by one for components of the unknown vector x. Components of x detected by MP
iterations are removed from subsequent iterations to reduce interference to components
of x yet to be detected. In this sense, MP implements a form of successive intereference
cancellation (SIC). Although the MP approach generally has weaker guarantees than BP,
it has been shown empirically that it often performs similarly, and in some applications
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 137
In this section, we introduce the STAP radar signal model and discuss properties of
random arrays in STAP radar. In particular, we discuss the average sidelobe and, more
importantly, the average peak sidelobe levels exhibited by a random array. We also
discuss the clutter rank of the random array in STAP.
up
Np pulses
Aperture Z
The mainbeam is defined as the array patten in the region |ω| ≤ 1/Z, while the sidelobe
region is |ω| > 1/Z. The peak sidelobe is defined μ = max|ω|>1/Z β (ω).
The Doppler shift caused by a target moving at velocity vt relative to the normal
to the array is fd = 2vt /λ. The normalized Doppler frequency v is the Doppler shift
fd normalized to the sampling frequency 1/Tr , where Tr is the pulse repetition interval,
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 139
where ⊗ represents the Kroneckor product. The N × 1 baseband y signal received at the
array from a target with steering vector a and complex amplitude x is given by
y = ax + e, (5.7)
where ξ = 4vp Tr /λ and s(u) is the power of a clutter patch at spatial frequency u and
normalized Doppler frequency ξu.
The signal model for K targets is given by
y = Ax + e. (5.10)
Here, A is the N ×G measurement matrix whose columns are steering vectors associated
with a grid of possible target locations on the angle-Doppler map, and x is a G × 1
vector of complex target amplitudes. The vector x contains only K G nonzeros. In
later sections, we apply optimization algorithms that operate on a grid. To this end, we
2
discretize the angle-Doppler map into G = G grid points, where G is the number
of grid points in each of the two domains. The G grid points serve as resolution cells.
Typically the dimensionality of the signal space N is much smaller than the number of
resolution cells, N G. The G × 1 vector of target gains x is assumed to be sparse, in
the sense that it has K G nonzero entries.
In STAP, the covariance matrix R is typically unknown, but can be estimated from
secondary data. The secondary data is assumed to consist of independent identically
distributed vectors with a covariance matrix common with the cell under test. Let L be
the number of secondary data vectors and q(l) a secondary data vector, the maximum
likelihood estimate (MLE) of the covariance matrix is the sample covariance matrix
140 Kim and Haimovich
1
L
/
R= q(l)q(l)H . (5.11)
L
l=1
In subsequent sections of this chapter, we will make use of the inverse of the sample
covariance matrix. In order to ensure that /
R−1 exists, we make the assumption that
L > N.
Na
φ (ω) = (1/Na ) E ej 2πzn ω = E ej 2πzω ,
n=1
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 141
0
ULA – 10
ULA – 15
RA – 15
–5
Beampattern (dB)
–10
–15
–0.25 –0.2 –0.15 –0.1 –0.05 0 0.05 0.1 0.15 0.2 0.25
Spatial frequency
Figure 5.2 Beampattern of a small ULA with 20 elements (10λ array), a random array with
an array aperture of Z = 15λ with N = 20 elements, and a large ULA with 30 elements
(15λ array).
and variance
2
var |b (ω)|2 = E |b (ω)|2 − φ (ω) .
Thus, the sidelobes of a random array are dominated by the term 1/Na rather than the
sidelobes of the associated filled array.
μ = max β (ω) .
|ω|>1/Z
Viewed as a function of ω, the array pattern β (ω) is a stochastic process. In the sidelobe
region, the stochastic process is approximately ergodic, meaning that statistical averages
may be gleaned from averages across the spatial frequency variable ω [10]. Further-
more, values of the stochastic process β (ω) can be approximated as independent when
the values of the spatial frequency ω are separated by a sidelobe or more [10].
As previously discussed, the number of sidelobes is approximately 2Z, where Z
is the aperture size of the random array. To find the CDF of the peak sidelobe, let
142 Kim and Haimovich
β (ω) 2Na β (ω) = 2Na |b (ω)|2 and μ 2Na μ. Since b (ω) ∼ CN (0,1/Na ), it
follows that β (ω) is a chi-square random variable with 2 degrees of freedom. Recall
that the sidelobes are approximated to be independent from each other [10]. It is easy to
verify that the cumulative distribution function (CDF) of
β (ω) is given by
−t/2
β (t) = 1 − e
. (5.12)
0
Random array – 15
Average sidelobe level
Average peak sidelobe level
–5
Beampattern (dB)
–10
–15
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.
Spatial frequency
Figure 5.3 Beampattern of a random array using Na = 15 elements filling a large aperture of
Z = 20λ. The computed average sidelobe level and average peak sidelobe level is also shown.
Taking ln of both sides, and noting that the expected result is such that Na η " 1, we
approximate the ln function with the first term in its Taylor expansion
$ %
ln 1 − e−Na η ≈ −e−Na η . (5.19)
Note that this approximation holds true because η is reasonably higher than the mean
value. Using (5.19) and after a little algebra, obtain
1$ %
Na = ln 2Z − ln ln α −1 ,
η
which links between the number of elements and confidence level that the sidelobes do
not exceed a set value. A similar result without proof has been presented in [38].
–0.5 –0.5
–0.4 –0.4
–0.3 –0.3
–0.2 –0.2
Normalized Doppler
Normalized Doppler
–0.1 –0.1
0 0
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
–0.5 –0.4 –0.3 –0.2 –0.1 0 0.1 0.2 0.3 0.4 0.5 –0.5 –0.4 –0.3 –0.2 –0.1 0 0.1 0.2 0.3 0.4 0.5
Spatial frequency Spatial frequency
Figure 5.4 (Left figure): Clutter map using a ULA with N = 10 elements, P = 20 pulses, and
β = 1. (Right figure): Clutter map using a random array with N = 10 elements, P = 20 pulses,
and β = 1. The elements of the sparse random array are spread across an array of size 15λ.
same number of elements (10) spread over 15λ (rather than the 5λ ULA aperture), both
the clutter maps are generated with β = 1. It is noticed that the clutter ridge of the
random array is narrower, which leads to a lower MDV. Note that the clutter map of
the sparse array also exhibits multiple, spurious clutter ridges due to higher sidelobes of
the beampattern.
The clutter rank of a filled ULA can be computed from Brennan’s rule
where recall that ξ = 4vp Tr /λ. Now, given a random array with aperture size Z, let
f ull
Na represent the number of sensors in a filled ULA configuration. From [39], the
clutter rank of a random array is
f ull
r c ≈ Na + (Np − 1)ξ. (5.21)
It is noticed that the clutter rank of a random array depends on the aperture size Z,
f ull
since Na = 2Z. This means the random array will require the same number of
degrees of freedom to suppress the clutter as a large ULA. However, the number of
degrees of freedom available to the random array is less than that of the large ULA.
Therefore, fewer degrees of freedom are left to supply gain for the target than for the
filled ULA.
The MDV depends on the width of the clutter ridge and therefore, the MDV depends
on the aperture size of the array [40]. To illustrate this, the signal-to-interference-plus-
noise ratio (SINR) vs. normalized Doppler for three array configurations: a small ULA
with 10 elements (Z = 5λ), a large ULA with 30 elements (Z = 15λ), and a random
array with 10 elements randomly positioned across an array of size Z = 15λ at u = 0
is shown in Figure 5.5.
The SNR is defined as |x|2 /σ2 and the clutter-to-noise ratio (CNR) is defined as
s(u)/σ2 , where s(u) is the power of the clutter at spatial frequency u and Doppler ξu.
For simplicity, it is assumed that the clutter power is constant for all clutter patches and
the power is given by σc2 . Then the CNR is also given by σc2 /σ2 . The SINR is defined as
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 145
20
18
16
14
12
SINR (dB)
10
8
ULA – 5
ULA – 15
6 RA – 15
0
–0.1 –0.08 –0.06 –0.04 –0.02 0 0.02 0.04 0.06 0.08 0.1
Normalized Doppler
Figure 5.5 SINR vs. normalized Doppler for three array configurations: a small ULA with 10
elements (Z = 5λ), a large ULA with 30 elements (Z = 15λ), and a random array with 10
elements randomly positioned across an array of size Z = 15λ. The SNR is 10 dB and the CNR
is 30 dB and the spatial frequency u is set to 0.
SINRi = aH −1
i R ai . From Figure 5.5, we see that both the large ULA and the random
array achieve a higher SINR at lower Doppler than the small ULA. For example, at
normalized Doppler = 0.02, the small ULA is about 4 dB SINR, but the random array is
about 9 dB. This disparity points to a lower MDV.
In this section, we introduce the detection problem and summarize a popular detection
algorithm, the adaptive beamformer (also known as the adaptive matched filter in [41]).
We also analyze the performance of the adaptive beamformer applied to a random array
to motivate the need for an alternative algorithm.
To recap, we are posing the problem of testing a STAP resolution cell for the presence
of a target of unknown amplitude observed in the presence of Gaussian colored noise
with unknown covariance matrix, when secondary data is available for estimating the
covariance matrix, this problem has been solved by Kelly in [42]. Kelly was able to
show that the probability of false alarm for the GLRT detector does not depend on
interference covariance matrix [42]. In fact it was shown that the probability of false
alarm depends on the dimensionality of the signal N , the number of training samples
used to estimate the interference covariance matrix R, and the threshold parameter γ.
Since the probability of false alarm does not depend on any unknown parameters (γ is
a design parameter chosen to achieve a desired false alarm probability), the detector is
described as a CFAR detector.
A simpler approach is suggested in [41], where the likelihood of the observation is
maximized only over the unknown amplitude (separately for each hypothesis). In this
approach, the covariance matrix is assumed known through the derivation of the test
statistic. It is later substituted with the sample covariance matrix of the secondary data
in the final expression of the test statistic since the covariance matrix is unknown. While
this procedure is ad hoc, it is argued in [41] that the resulting test statistic differs from
GLRT statistic in [42] only by a term that vanishes when the set of secondary data
is large.
The test for deciding H1 for a resolution cell defined by the steering vector a is given
by [41]
|aH /R−1 y|2
T = H −1 ≥ γ. (5.22)
a / R a
It is noted that the test statistic is essentially a beamfomer aH applied to whitened
observations /R−1 y and normalized by the product aH / R−1 a, hence we refer to this
approach as adaptive beamforming (ABF).
An alternative form of the test statistic that is used for performance evaluation is
found by observing that according to Robey et al.[41], the test (5.22) for deciding H1
may be expressed as a ratio of two independent random variables
|ζ|2
T = ≥ γ. (5.23)
hψ
It $
% in [41] that ζ is distributed CN (0,1) when a target is not present and
is shown
CN hρ,1 when a target with steering vector at is present. The effect of estimating
the covariance matrix (the fact that R and R̂ are not the same) is captured in the loss
factor h discussed in (5.25). The target’s SNR is given by
H −1 2
xa R at
ρ= . (5.24)
aH R−1 a
The test statistic (5.23) tests for the presence of a target by applying a steering vector
a, but the actual steering vector of the target is at . The denominator of (5.23), ψ, is a
chi-squared random variable with 2(L + 1 − N ) degrees of freedom. Since the factor
(L + 1 − N ) appears in several expressions in the sequel, for notational brevity, let
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 147
where F (·,·) denotes the CDF of the F distribution F (·,·). To obtain the probability of
false alarm we integrate (5.29) over the random variable h
1
PF A = 1 − F (2,2M) (hM γ|h)p(h)dh, (5.30)
0
where p(h) is given by (5.25). Note that the probability of false alarm depends only on
the dimensionality of the signal N and the term M, which depends on the number of
secondary samples used to obtain the estimate of the covariance matrix R̂.
We now turn our attention to the probability of detection again. For convenience
we seek to compute the probability Pr{T # ≥ hγ} instead of Pr{T ≥ γ}. Similar to
the probability of false alarm, we begin by fixing the random variable h and find the
probability of detection conditioned on h. When a target is present, the test statistic
follows the noncentral F distribution with parameters 2 and 2M, and noncentrality
148 Kim and Haimovich
$ %
parameter hρ, denoted F 2,2M,hρ . The conditional probability of detection of a
target with SNR ρ (5.24) is given by
Pr{T ≥ hγ|h} = 1 − F (2,2M,hρ) (hM γ|h). (5.31)
To obtain the probability of detection we integrate (5.32) over the random variable h
1
PD = 1 − F (2,2M,hρ) (hM γ|h)p(h)dh, (5.32)
0
where p(h) is given by (5.26).
Unfortunately, to solve (5.34) one requires the knowledge of the number of targets K,
which of course is unknown a priori. To implement a CFAR radar that exploits target
sparsity, we propose a two-stage MP-CFAR detection algorithm. Candidate targets are
localized in the first phase; in the second phase, they are tested for detection. A detected
target is then canceled from the data. The cancellation of detected targets from the data
is intended to remove mutual interference between targets and thus address one of the
flaws of detection by ABF. A block diagram of the MP-CFAR algorithm is shown in
Figure 5.6.
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 149
If test passes:
Set: k = k + 1
Sk If test fails:
Initialize:
Output: S0 = 0, k = 0
S0 = 0, k = 1
5.4.1 MP-CFAR
Stage 1: MP Localization
The first pass of the MP localization algorithm uses whitened data z = / R−1/2 y and
whitened steering vectors bj = / R−1/2 aj , j = 1,. . .,G. The first candidate target is
localized by the index m1 of the vector bj that has the largest data projection,
|bH
j z|
2
m1 = arg max (5.35)
j bH
j bj
for j = 1,. . .,G. The index m1 localizes the target in the angle-Doppler domains. This
information is subsequently used by the detection stage, as described in relation with
Stage 2 later in this section.
Next, we describe the localization of the k-th candidate target, given that k − 1 targets
have already been localized and passed the detection test. The observed and whitened
data z is processed to cancel the contribution of targets detected previously. Let a matrix
B be formed with the columns bj . Let Sk−1 be the set of indices of columns of B
associated with detected targets, and let BSk−1 be the matrix formed by the columns
indexed by Sk−1 . The $ Hprojection% matrix orthogonal to the detected targets is given by
−1
P⊥BS = I − B S k−1 BSk−1 BSk−1 BSk−1 . Similarly, steering vectors orthogonal to the
k−1
detected targets are formed as follows: wj = P⊥
BSk−1 bj , for all j ∈
/ Sk−1 . The k-th target
is localized according to
|wH
j z|
2
mk = arg max . (5.36)
j wH
j wj
This process continues until a candidate target fails the detection test.
Stage 2: Detection
We now derive a CFAR detector that is applied to candidate targets localized in Stage 1.
The first candidate target is detected according to (5.22), rewritten here for convenience:
|aH /−1 2
m1 R y|
T = ≥ γ, (5.37)
aH /−1
m R am1
1
150 Kim and Haimovich
where m1 is the index found in Stage 1. Note that the test (5.37) may also be expressed
in terms of the whitened steering vectors bm1 = /
R−1/2 am1 ,
|bH
m1 z|
2
T = ≥ γ. (5.38)
bH
m1 bm1
Next we describe the detection of candidate target k, given that k − 1 targets have
already been localized and passed the detection test. The signal model is given by the
expression
where mk is the index of the resolution cell of the k-th candidate target found in
Stage 1 (5.36), Sk is formed by adding mk to the set Sk−1 , Sk = Sk−1 mk , the
matrix BSk = [bmk ,BSk−1 ] is the matrix formed by columns with indices in Sk ,
T
xSk = xmk ,xTSk−1 , and n = / R−1/2 e. This signal model leads to the following
detection test:
H0 : xmk = 0
H1 : xmk != 0.
Here, the following problem is posed: detect a target located at a specified whitened
steering vector bmk and having unknown amplitude, observed in the presence of inter-
ference and noise. The interference is of unknown gain xSk−1 , but belonging to a known
subspace BSk−1 . The noise is Gaussian colored noise for which the covariance matrix is
unknown, but secondary data is available for its estimation.
To develop the test statistic for the detection problem, we start by expressing the
likelihoods of the observations under the two hypotheses. As in the discussion leading
to (5.22), the detector is a generalized likelihood ratio detector only in the sense that the
liklihood under H1 is maximized over the unknown target amplitude. To simplify the
detector, as in [41], it is assumed that the PDF’s of the test statistic under each hypothesis
are based on the true covariance matrix. It is noted that the subsequent analysis relies
on the properties of the estimated covariance matrix. Thus, z = / R−1/2 y is modeled as
having a covariance matrix equal to the identity matrix. It follows that under H0 , the
likelihood is
& 'H & '
$ % 1 − z−BSk−1 xSk−1 z−BSk−1 xSk−1
p z|H0 = N e ,
π
while under H1 the likelihood is
1 −$z−BS xS %H $z−BS xS %
p (z|H1 ) = e k k k k .
πN
The GLRT for deciding H1 is given by
maxxSk p(z|xSk )
T = ln ≥ γ. (5.40)
maxxSk−1 p(z|xSk−1 )
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 151
To obtain a more convenient form of the test, we note that under hypothesis H0 , the
MLE of the gain vector xSk−1 is found from
/
xSk−1 = min z − BSk−1 xSk−1 22 . (5.41)
xSk−1
Minimizing (5.41) with respect to the vector of complex gains xSk yields
−1 H
/
xSk−1 = (BH
Sk−1 BSk−1 ) BSk−1 z. (5.42)
Similarly,
−1 H
/
xSk = (BH
Sk BSk ) BSk z. (5.43)
T = z − BSk−1/
xSk−1 22 − z − BSk/
xSk 22
& '
= zH PBSk − PBSk−1 z, (5.44)
$ %−1
where PB = B BH B B is a projection matrix that projects onto the subspace
spanned by B. Note that the decision statistic is a difference between two quadratic
forms, where the quadratic form zH PBSk−1 z is an interference term that is canceled.
The test statistic (5.44) may be further simplified, which will be useful to obtain
expressions for performance evaluation of the MP-CFAR detector. We make use of the
following result from [46,47]. Let D and E be two subspaces, and let PD and P[D,E]
be projection matrices that project onto the subspaces spanned by the matrices D and
[D,E], respectively. Let F = P⊥ D E, then the difference between projection matrices
P[D,E] − PD is given by [47]
P[D,E] − PD = PF . (5.45)
From this expression, fk is the projection of the whitened steering vector bmk orthogonal
to the previous k − 1 targets. We use this vector to remove the potential interference the
other k − 1 targets provides through the sidelobes. Applying (5.45), we obtain
where Pfk = fk fH k /fk fk . Noting that BSk = [BSk−1 ,bmk ] and substituting (5.47)
H
into (5.44), the test for deciding H1 on the detection of the k-th target can be expressed as
|fH
k z|
2
T = ≥ γ. (5.48)
fH
k fk
Algorithm 1 CFAR-MP
1: Input: y,A, R̂, γ.
2: Initialize: S0 = ∅, r = R̂−1/2 y, B = R̂−1/2 A, W = B, k = 1.
3: Find: Search for the index l that maximizes the metric maxj |wH 2 H
j r| /wj wj .
4: Update set of targets: Sk = Sk−1 l.
5: Check: If Tsi ≥ γ (test statistic to decide if xsi is nonzero) for all si ∈ Sk continue.
Otherwise output Sk−1 as solution and terminate.
& '−1
6: Generate: P⊥ BS = I − BS k BHB
Sk S k BH
Sk .
k
7: Remove found targets: W = P⊥ BSk B.
8: Renormalize: If wi 2 = 0, set wi = 0.
9: Return to step 3.
The test statistic (5.48) is applied to every candidate target included in the set Sk . If
any of the k tests fails to exceed the threshold γ, the algorithm terminates and outputs
the set Sk−1 , the set of k − 1 target locations. Otherwise, MP-CFAR increments the
number of targets k by one and reruns MP with the new value of k. The psuedocode for
the MP-CFAR algorithm is listed in Algorithm 1.
Other than the max operator, the test statistic in (5.49) is of the form (5.22), hence it can
be reduced to the form (5.28),
T = max j , (5.50)
j
where
|ζ j |2 /2
j = . (5.51)
hψj /2M
The probability of false alarm is given by
5 6
PF A = 1 − Pr max j ≤ hM γ .
j
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 153
Using the assumption that the random variables j (5.51) are independent and identi-
cally distributed,
$ %G
PF A = 1 − Pr j ≤ hM γ .
The probability of detection of the first target is given by the same expression as for
the ABF (5.32). Note that it is assumed that the target is assumed to be recovered in
the first pass of the MP-CFAR detector. If this is the case, the probability of detection
follows that of (5.32). Otherwise, the probability of detection will decrease due to the
orthogonal projection. When a target is present, the probability of false alarm includes
the event that a target is detected at the incorrect resolution cell. That would increase the
PF A in (5.52). An algorithm that mitigates this type of false alarm is presented next.
5.4.3 MBMP-CFAR
The MP-CFAR algorithm localizes the first target according to (5.35), namely, it finds
the column of the whitened measurement matrix with the largest projection on the
whitened data z. A false alarm (localizing the target in the wrong resolution cell)
increases the chance of further false alarms downstream, since according to (5.36),
localizing subsequent candidate targets depends on the location of the first target (5.35).
A more robust approach is to hedge bets by finding multiple candidates for the loca-
tion of the first target. Each such candidate target serves as seed to the localization and
detection of subsequent targets. When the process is completed, a metric is used to select
the set of targets that provides the best fit to the data. This algorithm, which generalizes
MP-CFAR, is referred to as MBMP-CFAR.
We introduce some notation that facilitates the presentation of MBMP-CFAR.
A localization solution is referred to as a branch. The set D = {d1,d2,. . .,dk } contains
the number of branches per target. A path is a sequence of branches specified by
their index numbers. For example, the path (i1,i2,. . .,ik ), 1 ≤ i1 ≤ d1,. . .,1 ≤
(i ,i ,...,i )
ik ≤ dk . A localization solution 1 ≤ mk 1 2 k ≤ G, where G is the number of
resolution cells (see [5.10]), consists of a path and the index number of the resolution
(i ,...,i ) (i ) (i i ) (i ...,i )
cell. The set Sk 1 k = m1 1 ,m2 1, 2 ,ldots,mk 1, k contains the localization
solution associated with path (ı1,i2,. . .,ik ) . For k candidate targets, MBMP maintains
d1 × d2 × · · · × dk such sets. The matrix BSk was defined to consist of the whitened
steering vectors bj indexed by Sk . Similarly, we define the matrix B (i1,...,ik ) to consist
Sk
(i ,...,i ) (i ,...,i )
of whitened steering vectors indexed by the set Sk 1 k . The vector fk 1 k is defined
analogous to (5.46)
= P⊥
(i ,...,ik )
fk 1 B bk . (5.53)
Sk (i1 ,...,ik ) \mk
154 Kim and Haimovich
(1)
|bH
j z|
2
m1 = arg max . (5.54)
j bH
j bj
(i)
|bH
j z|
2
m1 = arg max . (5.55)
/ {m1,...,mi−1 }
j∈ bH
j bj
To generate the d1 d2 branches associated with the second target, define the modified
steering vectors wj = P⊥
(i)
b (i) bj , for 1 ≤ i ≤ d1 . The orthogonal projection prevents
m1
(i)
interference from a target at m1 . The resolution cell index associated with the first
branch of the second target is given by
(1)H
(1,1)
|wj z|2
m2 = arg max (1)H (1)
, (5.56)
j wj wj
whereas the index of branch i2 , 1 ≤ i2 ≤ d2 of the second target, given the path (i1,i2 ),
(i )H
(i ,i )
|wj 1 z|2
m2 1 2 = arg max (i )H (i )
. (5.57)
j∈
/ m1,...,mi2 −1 wj 1 wj 1
(i ,...,ik )
Generalizing to k targets and the path (i1,i2,. . .,ik ), define the vector wj 1
= P⊥
B bj . The index associated with the k-th target is given by
(i ,...,ik )
Sk 1
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 155
S0
S1(1) { m1(1) }
S1(2) {m(2)
1 }
S2(1,1) {m1(1),m2(1,1) }
S2(2,2) {m1(2) ,m(2,2)
2 }
(2) (2,1)
S2(1,2) {m1(1) ,m2(1,2) } S2(2,1) { m1 ,m 2 }
Figure 5.7 Graph of MBMP algorithm for a branch vector d = [2, 2]T .
(i1,...,ik−1 )H 2
wj z
(i ,...,ik )
mk 1 = arg 5 max 6 . (5.58)
(i ,...,i )H (i ,...,i )
j∈
/ m1,...,mik−1 wj 1 k−1 wj 1 k−1
Stage 2: Detection
The MBMP localization processing yielded d1 candidate locations for the first target.
The largest score relative to the objective function |bH 2 H
j z| /bj bj is obtained by the
(1)
steering vector index m1 , because
max |bH
j z| /bj bj ≥
2 H
max |bH 2 H
j z| /bj bj ,
j / {m1,...,mi−1 }
j∈
(1)
(see [5.54] and [5.55]). Note that the choice of m1 also minimizes the residual of
the objective function z − Bx22 (see [5.34]), m1(1) = arg minj P⊥ 2
bj z2 . The test to
(1)
determine whether a target is present in the resolution cell m1 is given by (5.48)
156 Kim and Haimovich
|bH(1) z|2
m1
T = ≥ γ. (5.59)
bH(1) bm(1)
m1 1
(i )
If the test (5.59) is met, d1 target sets are updated as follows S1 1
(i )
= m1 1 , 1 ≤ i1 ≤ d1 . If the test fails, then MBMP-CFAR declares that no targets
exist, and the algorithm terminates.
To test for the detection of the k-th target, assume that k−1 targets have been detected.
The residual along the path (i1,i2,. . .,ik ) is computed from
4 42
4 ⊥ 4
R (i1,...,ik ) = 4 P z 4
4 B i ,...,i 4 . (5.60)
(
Sk 1 k)
2
(i ...,i )
The test to determine whether a target is present in the resolution cell mk 1, k is given
by (see definition of vectors f in [5.53]).
(i ,...,ik )
|fk 1 z|2
T = (i ,...,ik )H (i1,...,ik )
≥ γ. (5.62)
fk 1 fk
(i ) (i i ) (i ...,i )
All k resolution cells m1 1 ,m2 1, 2 ,. . .,mk 1, k are tested. In other words, the
(i ) (i ,...,i )
test (5.62) is performed for the vectors f1 1 ,. . .,fk 1 k .
(i ,...,i )
If all k tests (5.62) are met, d1 × . . . × dk target sets are updated as follows Sk 1 k
(i1 ) (i1, i2 ) (i ...,i )
= m1 ,m2 ,. . .,mk 1, k , 1 ≤ i1 ≤ d1,. . .,1 ≤ ik ≤ dk . The MBMP-CFAR
algorithm proceeds to the localization and detection of the (k + 1) target. If the detection
test fails, MBMP-CFAR outputs as solution the path
the first iteration of MBMP-CFAR produces the same localization solution as the MP-
CFAR and also performs the same detection test. Hence when no targets exist, the
performance of both MBMP-CFAR and MP-CFAR is the same. Also note the path that
(1) (1,...,1)
corresponds to the set of resolution cells (m1 ,. . . mk ) is the MP-CFAR solution.
(1) (1,...,1)
However, MBMP-CFAR does not always test the set (m1 ,. . . mk ), because it may
not minimize the residual (5.60). For example, using Figure 5.7, in the first iteration
(1)
MBMP-CFAR will always test m1 , however in the next iteration MBMP-CFAR will
(1,1) (2,1)
test either the set of resolution cells S2 or the set of resolution cells S2 . Note
(1,2) (2,2)
that the algorithm will never test the paths S2 or S2 because they present higher
(1,1) (2,1)
objective functions than S2 and S2 , respectively.
In this section, we present numerical results on the MP-CFAR and MBMP-CFAR algo-
rithms and compare them with ABF. Unless stated otherwise, in figures presented in
this section, the aperture of the random arrays is 12λ (Z = 12, where Z is expressed
in units of wavelength). The number of elements in the random array is Na = 16,
thus the mean spacing between elements of the random array is 12λ/16. The number
of coherent pulses used by all arrays is Np = 25. The SNR, defined as |x|2 /σ2 , is
set to SNR = 15.5 dB unless stated otherwise. The CNR is set to 30 dB. It was seen
that the SINR of the random array, defined as SINRi = aH −1
i R ai , is roughly 15 dB
with these parameters. The number of training samples used to estimate the covariance
matrix for the random array is L = 2N . Reduced-rank methods may be applied to
reduce the size of the training set [48], but this is not the emphasis of this work. The
number of resolution cells on the angle-Doppler map is given by G = (2Z + 1)2 = 625.
A random realization of a random array is generated and remains fixed throughout the
Monte Carlo simulations for all figures unless otherwise stated. Let St be the true set of
resolution cells that contain targets, and let Ŝ be the set of resolution cells found by a
detector to have targets. A false alarm event occurs if Ŝ\St != ∅, and a detection event
7
occurs if Ŝ St != ∅. The threshold γ was computed by selecting a desired false alarm
probability, the equations in [49] were then used to find the appropriate value of γ.
The probabilities of false alarm of the MP-CFAR and ABF detectors are studied in
Figure 5.8, which plots the empirical probability of false alarm against the SINR of
a target present with the angle-Doppler pair (5/Z,0). The detection threshold for the
ABF detector is set using (5.30), such that PF A = 10−3 . Applying (5.52), the detection
threshold for the MP-CFAR detector is also set to PF A = 10−3 . In this figure, the arrays
compared are a 12λ ULA and an 12λ random array. The random array has Na = 16
sensors; the resolution cells for this experiment was spaced apart by 1/12λ. For each
curve (excluding the line PF A = 10−3 ), the results of 104 Monte-Carlo experiments
were averaged to obtain the curves, and the ABF tested every resolution cell on the
angle-Doppler map. The probability of false alarm of a true CFAR detector should not
change as a function of SNR of a target present somewhere in the search area. It is
158 Kim and Haimovich
10–1
ABF
MP-CFAR
ABF (large ULA)
PF
10–2
10–3
10 11 12 13 14 15 16 17 18 19 20
SINR
Figure 5.8 Probability of false alarm vs. SINR of a target for the ABF with a random array,
MP-CFAR with a random array, and the ABF with a large ULA.
observed from the figure that the 12λ ULA ABF and the random array MP-CFAR
detectors have probabilities of false alarm that are little changed as a function of the
SNR of a target. More specifically, at low SNR the MP-CFAR experiences a probability
of false alarm of about 2 × 10−3 instead of PF A = 10−3 . This slight increase in the
probability of false alarm occurs because at low SNR, the probability of correct recovery
(the probability that MP-CFAR recovers the correct resolution cell to test) is less than
one. As the SNR of the target increases, the probability of correct recovery increases,
and the false alarm probability of MP-CFAR decreases to PF = 10−3 as intended. It is
noticed that the ABF using a 12λ ULA experiences a slight increase in the probability
of false alarm as the SNR of the interfering target increases. This is because although the
peak sidelobe of a ULA is relatively small (roughly −13 dB), it is not zero and therefore
will ultimately affect the probability of false alarm. In contrast, a random array using
ABF cannot cope with energy leaked by high sidelobes, and as the strength of the target
increases, the probability of false alarm increases.
In Figure 5.9, shown are the receiver operating characteristic (ROC) curves of the
ABF using a 12λ ULA, the ABF using a random array, and MP-CFAR using a random
array, and a single target in the field of view. The target again has the angle-Doppler pair
(5/Z,0). From the figure, the large ULA using ABF performs well as expected. Since
the ULA array does not exhibit large sidelobes, the target does not significantly increase
the probability of false alarm. In contrast, it is seen that the ABF with the random array
performs considerably worse. The random array has large sidelobes, and since the ABF
does not account for the large sidelobes, the radar experiences a high false alarm rate.
The MP-CFAR with a random array on the other hand performs similarly to the ABF
with a large ULA. The MP-CFAR unlike the ABF, accounts for detected targets and
removes the targets before detecting more targets. Note that the MP-CFAR performs
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 159
0.95 ABF
CFAR-MP
ABF (Large ULA)
0.9
0.85
0.8
PD
0.75
0.7
0.65
0.6
10–4 10–3 10–2 10–1
PF
Figure 5.9 ROC curve for a single target for the ABF with a random array, MP-CFAR with a
random array, and the ABF with a large ULA. Parameters SNR = 15.5 dB and CNR = 30 dB.
ABF
MP-CFAR
10–1 MBMP-CFAR – [3 1 1 ..]
–3
P F = 10
PF
10–2
10–3
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9
Sparsity Ratio
Figure 5.10 Probability of false alarm vs. the sparsity ratio of an array for the ABF with a random
array, the MP-CFAR with a random array, and the MBMP-CFAR with a random array with
branch vectors D = [3,1,1,. . .]. Parameters: SNR = 15.5 dB and CNR = 30 dB.
similarly to the ABF with the large ULA using about 3/4s of the number of elements
compared to the large ULA. This demonstrates the savings without loss of performance
that are gained by random arrays and the proposed MP-CFAR detector.
In Figure 5.10, shown is the probability of false alarm vs. the sparsity ratio of a
f ull f ull
random array. The sparsity ratio is defined as sparsity ratio = Na /Na , where Na
is the number of elements required to fill a ULA for an array of the same size. There are
160 Kim and Haimovich
10–1 ABF
MP-CFAR
MBMP-CFAR – [6 1 1 ..]
PF
10–2
10–3
Number of Targets
Figure 5.11 Probability of false alarm vs. the number of targets in the field of view for the ABF
with a random array, the MP-CFAR with a random array, and the MBMP-CFAR with a random
array with branch vectors D = [3,1,1,. . .]. Parameters: SNR = 15.5 dB and CNR = 30 dB.
two targets in the field of view, with the angle-Doppler pairs (5/Z,0) and (−5/Z,0). The
threshold parameter γ is set so that the desired probability of false alarm is PF = 10−3
for all methods. As expected, the false alarm rate decreases as the sparsity ratio increases
for all three curves, since using more elements in a random array decreases the average
sidelobe and average peak sidelobe levels. Although the false alarm rate for the ABF
decreases as the sparsity ratio increases, the false alarm rate is still much larger than
10−3 as intended. In contrast, MP-CFAR experiences significantly less false alarms at
all sparsity ratios, in addition, at sparsity ratio 0.9, the MP-CFAR detector experiences
a false alarm rate of about 10−3 , as intended. MBMP-CFAR further decreases the false
alarms at low sparsity ratios, and as expected also experiences a false alarm rate of about
10−3 as intended. MBMP-CFAR experiences lower false alarm rates at lower sparsity
ratios, which can be attributed to the additional resolution cells that the algorithm tests.
In Figure 5.11, shown is the probability of false alarm vs. the number of targets for
the ABF, MP-CFAR, and the MBMP-CFAR with branch vector D = [6,1,1,. . .] using
a random array. The targets have the following angle-Doppler pairs s1 = (5/Z,0),
s2 = (5/Z,1/Z), s3 = (−4/Z,2/Z), s4 = (−3/Z,2/Z), s5 = (−6/Z,−1/Z),
s6 = (6/Z,−1/Z). For K = 1, the target with angle-Doppler pair s1 is placed on the
map; for K = 2, the targets s1,s2 are on the map, etc. Again, the threshold parameter
γ is set so that the desired probability of false alarm is PF = 10−3 for all methods.
From the figure it is seen that the ABF performs significantly worse than the MP-CFAR
and the MBMP-CFAR as the number of targets increases. The false alarm rate for the
MP-CFAR on the other hand increases slightly from PF = 10−3 for K = 0 to about
PF = 1.4 × 10−3 for K = 3. For K = 4, it is noticed that targets s3 and s4 are placed in
adjacent resolution cells in angle. This causes significant interference between the two
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 161
1
ABF
0.95 MP-CFAR
MBMP-CFAR [3 1 1 ..]
0.9
0.85
0.8
PD
0.75
0.7
0.65
0.6
0.55
0.5
10–3 10–2 10–1 100
PF
Figure 5.12 ROC curve for two closely spaced targets for the ABF with a random array,
MP-CFAR with a random array, and the MBMP-CFAR with a random array. Parameters
SNR = 15.5 dB and CNR = 30 dB.
targets and the false alarm rate increases to about PF = 7 × 10−3 . The false alarm rate
for MP-CFAR does not change significantly for K ≥ 4. The remaining targets s5 and
s6 are far apart from each other and do not significantly interfere with each other or the
other 4 targets and hence does not drastically impact the false alarm rate. MBMP-CFAR
performs identically to the MP-CFAR for K ≤ 3, it also experience a increase in the
false alarm rate for K = 4 from PF = 1.4 × 10−3 to PF = 4 × 10−3. Note that the false
alarm rate for the MBMP-CFAR at K ≥ 4 is about 4 × 10−3 instead of 7 × 10−3 . This
decrease in the false alarm rate stems from the increase in the probability of correct
recovery in the MBMP-CFAR algorithm.
In Figure 5.12, the ROC curves for the ABF, MP-CFAR, and the MBMP-CFAR
are shown with branch vector D = [3,1,1,. . .] using a random array for two targets
spaced closely together. The two targets are placed in adjacent resolution cells, and
the angle-Doppler pairs of the targets are (−5/Z,0) and (−4/Z,0). From the figure, the
ABF performs poorly because it cannot cope with the high sidelobes of a random array
and both sparsity-based radars significantly outperform the ABF. Comparing the two
sparsity-based CFAR algorithms, the MBMP-CFAR algorithm sees a slight performance
gain compared to the MP-CFAR. This increase in performance is due to the increase in
the probability of correct recovery that the MBMP-CFAR algorithm enjoys by testing
more resolution cells.
5.6 Summary
In this chapter we propose using a random array with the MP-CFAR and MBMP-CFAR
algorithms to solve the target detection problem in a STAP setting. The random array
162 Kim and Haimovich
is a large undersampled array that achieves high resolution due to the large aperture at
the cost of high sidelobes. Although conventional beamforming cannot cope with the
high sidelobes introduced by the random array, the proposed sparsity-based algorithms
can cope with the high sidelobes, allowing one to enjoy the high resolution of the
random array without the consequences of the high sidelobes. This was achieved using
the proposed algorithms by iteratively detecting targets one by one and removing their
contributions from the data. Numerical simulations show that the proposed algorithms
outperform beamforming methods when a random array is employed.
We show in simulations that both MP-CFAR and MBMP-CFAR outperform the pop-
ular beamformer when a random array is employed. In particular, we show that the
beamformer experiences significantly higher false alarms compared to the proposed
methods, and is not compatible with a random array. In contrast, the MP-CFAR and
MBMP-CFAR algorithms are shown to be able to cope with high sidelobes and are
compatible with a random array.
References
[1] J. Ward, “Space-time adaptive processing for airborne radar,” DTIC document, Tech. Rep.,
1994.
[2] B. D. Van Veen and K. M. Buckley, “Beamforming: A versatile approach to spatial filtering,”
IEEE ASSP Magazine, vol. 5, no. 2, pp. 4–24, 1988.
[3] J. Capon, “High-resolution frequency-wavenumber spectrum analysis,” Proceedings of the
IEEE, vol. 57, no. 8, pp. 1408–1418, 1969.
[4] R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Transactions
on Antennas and Propagation, vol. 34, no. 3, pp. 276–280, 1986.
[5] R. Roy and T. Kailath, “Esprit-estimation of signal parameters via rotational invariance
techniques,” IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 37, no. 7,
pp. 984–995, 1989.
[6] J. Rissanen, “A universal prior for integers and estimation by minimum description length,”
The Annals of Statistics, pp. 416–431, 1983.
[7] M. Wax and T. Kailath, “Detection of signals by information theoretic criteria,” IEEE
Transactions on Acoustics, Speech, and Signal Processing, vol. 33, no. 2, pp. 387–392,
1985.
[8] M. G. Bray, D. H. Werner, D. W. Boeringer, and D. W. Machuga, “Optimization of thinned
aperiodic linear phased arrays using genetic algorithms to reduce grating lobes during
scanning,” IEEE Transactions on Antennas and Propagation, vol. 50, no. 12, pp. 1732–
1742, 2002.
[9] L. Cen, Z. L. Yu, W. Ser, and W. Cen, “Linear aperiodic array synthesis using an improved
genetic algorithm,” IEEE Transactions on Antennas and Propagation, vol. 60, no. 2,
pp. 895–902, 2012.
[10] Y. Lo, “A mathematical theory of antenna arrays with randomly spaced elements,” IEEE
Transactions on Antennas and Propagation, vol. 12, no. 3, pp. 257–268, 1964.
[11] B. D. Steinberg, Principles of Aperture and Array System Design: Including Random and
Adaptive Arrays. Wiley Interscience, vol. 1, 1976, p. 374.
Sparsity-Based Methods for CFAR Target Detection in STAP Random Arrays 163
[12] F. Athley, C. Engdahl, and P. Sunnergren, “On radar detection and direction finding using
sparse arrays,” IEEE Transactions on Aerospace and Electronic Systems, vol. 43, no. 4,
pp. 1319–1333, 2007.
[13] B. D. Steinberg, “Comparison between the peak sidelobe of the random array and algo-
rithmically designed aperiodic arrays,” IEEE Transactions on Antennas and Propagation,
vol. 21, no. 3, pp. 366–370, 1973.
[14] S. Tonetti, M. Hehn, S. Lupashin, and R. D’Andrea, “Distributed control of antenna array
with formation of UAVs,” IFAC Proceedings Volumes, vol. 44, no. 1, pp. 7848–7853, 2011.
[15] L. Carin, “On the relationship between compressive sensing and random sensor arrays,”
IEEE Antennas and Propagation Magazine, vol. 51, no. 5, pp. 72–81, 2009.
[16] D. L. Donoho, “Compressed sensing,” IEEE Transactions on Information Theory, vol. 52,
no. 4, pp. 1289–1306, 2006.
[17] R. Tibshirani, “Regression shrinkage and selection via the lasso,” Journal of the Royal
Statistical Society, Series B (Methodological), pp. 267–288, 1996.
[18] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decomposition by basis pursuit,”
SIAM Review, vol. 43, no. 1, pp. 129–159, 2001.
[19] J. M. Bioucas-Dias and M. A. Figueiredo, “A new twist: Two-step iterative shrink-
age/thresholding algorithms for image restoration,” IEEE Transactions on Image Process-
ing, vol. 16, no. 12, pp. 2992–3004, 2007.
[20] D. L. Donoho, A. Maleki, and A. Montanari, “Message-passing algorithms for compressed
sensing,” Proceedings of the National Academy of Sciences, vol. 106, no. 45, pp. 18914–
18919, 2009.
[21] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear
inverse problems,” SIAM Journal on Imaging Sciences, vol. 2, no. 1, pp. 183–202, 2009.
[22] D. Malioutov, M. Çetin, and A. S. Willsky, “A sparse signal reconstruction perspective for
source localization with sensor arrays,” IEEE Transactions on Signal Processing, vol. 53,
no. 8, pp. 3010–3022, 2005.
[23] I. W. Selesnick, S. U. Pillai, K. Y. Li, and B. Himed, “Angle-doppler processing using sparse
regularization,” in 2010 IEEE International Conference on Acoustics, Speech and Signal
Processing 2010, pp. 2750–2753.
[24] L. Anitori, A. Maleki, M. Otten, R. G. Baraniuk, and P. Hoogeboom, “Design and analysis
of compressed sensing radar detectors,” IEEE Transactions on Signal Processing, vol. 61,
no. 4, pp. 813–827, 2013.
[25] N. A. Goodman and L. C. Potter, “Pitfalls and possibilities of radar compressive sensing,”
Applied Optics, vol. 54, no. 8, pp. C1–C13, 2015.
[26] J. A. Tropp and A. C. Gilbert, “Signal recovery from random measurements via orthogonal
matching pursuit,” IEEE Transactions on Information Theory, vol. 53, no. 12, pp. 4655–
4666, 2007.
[27] W. Dai and O. Milenkovic, “Subspace pursuit for compressive sensing: Closing the gap
between performance and complexity,” DTIC document, Tech. Rep., 2008.
[28] D. Needell and J. A. Tropp, “Cosamp: Iterative signal recovery from incomplete and
inaccurate samples,” Applied and Computational Harmonic Analysis, vol. 26, no. 3,
pp. 301–321, 2009.
[29] M. Rossi, A. M. Haimovich, and Y. C. Eldar, “Compressive sensing with unknown
parameters,” in 2012 Conference Record of the Forty Sixth Asilomar Conference on Signals,
Systems and Computers (ASILOMAR), 2012, pp. 436–440.
164 Kim and Haimovich
[30] M. E. Davies and Y. C. Eldar, “Rank awareness in joint sparse recovery,” IEEE Transactions
on Information Theory, vol. 58, no. 2, pp. 1135–1146, 2012.
[31] M. Elad, Sparse and Redundant Representations: From Theory to Applications in Signal
and Image Processing. Springer, 2010.
[32] J. A. Tropp and S. J. Wright, “Computational methods for sparse solution of linear inverse
problems,” Proceedings of the IEEE, vol. 98, no. 6, pp. 948–958, 2010.
[33] H. H. Kim, M. A. Govoni, and A. M. Haimovich, “Cost analysis of compressive sensing for
mimo stap random arrays,” in 2015 IEEE Radar Conference (RadarCon), 2015, pp. 0980–
0985.
[34] M. Rossi, A. M. Haimovich, and Y. C. Eldar, “Spatial compressive sensing in MIMO radar
with random arrays,” in 46th Annual Conference on Information Sciences and Systems
(CISS), 2012, pp. 1–6.
[35] J. Ward, “Space-time adaptive processing for airborne radar,” in IEE Colloquium on Space-
Time Adaptive Processing (Ref. No. 1998/241), 1998, pp. 2/1–2/6.
[36] H. L. Van Trees, Detection, Estimation, and Modulation Theory: Optimum Array process-
ing. John Wiley & Sons, 2004.
[37] B. Eisenberg, “On the expectation of the maximum of IID geometric random variables,”
Statistics & Probability Letters, vol. 78, no. 2, pp. 135–143, 2008.
[38] B. D. Steinberg, “The peak sidelobe of the phased array having randomly located elements,”
IEEE Transactions on Antennas and Propagation, vol. 20, no. 2, pp. 129–136, 1972.
[39] J. Ward, “Space-time adaptive processing with sparse antenna arrays,” in Conference Record
of the Thirty-Second Asilomar Conference on Signals, Systems and Computers 1998,
pp. 1537–1541.
[40] R. Klemm, Principles of Space-Time Adaptive Processing. IET, 2002, no. 159.
[41] F. C. Robey, D. R. Fuhrmann, E. J. Kelly, and R. Nitzberg, “A CFAR adaptive matched filter
detector,” IEEE Transactions on Aerospace and Electronic Systems, vol. 28, no. 1, pp. 208–
216, 1992.
[42] E. J. Kelly, “An adaptive detection algorithm,” IEEE Transactions on Aerospace and
Electronic Systems, no. 2, pp. 115–127, 1986.
[43] I. S. Reed, J. D. Mallett, and L. E. Brennan, “Rapid convergence rate in adaptive arrays,”
IEEE Transactions on Aerospace and Electronic Systems, no. 6, pp. 853–863, 1974.
[44] E. Kelly, “Performance of an adaptive detection algorithm; rejection of unwanted signals,”
IEEE Transactions on Aerospace and Electronic Systems, vol. 25, no. 2, pp. 122–133, 1989.
[45] D. M. Boroson, “Sample size considerations for adaptive arrays,” IEEE Transactions on
Aerospace and Electronic Systems, no. 4, pp. 446–451, 1980.
[46] R. T. Behrens and L. L. Scharf, “Signal processing applications of oblique projection
operators,” IEEE Transactions on Signal Processing, vol. 42, no. 6, pp. 1413–1424, 1994.
[47] L. L. Scharf and B. Friedlander, “Matched subspace detectors,” IEEE Transactions on Signal
Processing, vol. 42, no. 8, pp. 2146–2157, 1994.
[48] J. R. Guerci, J. S. Goldstein, and I. S. Reed, “Optimal and adaptive reduced-rank STAP,”
IEEE Transactions on Aerospace and Electronic Systems, vol. 36, no. 2, pp. 647–663, 2000.
[49] J. Liu, H. Li, and B. Himed, “Threshold setting for adaptive matched filter and adaptive
coherence estimator,” IEEE Signal Processing Letters, vol. 22, no. 1, pp. 11–15, 2015.
6 Fast and Robust Sparsity-Based
STAP Methods for Nonhomogeneous
Clutter
Xiaopeng Yang, Yuze Sun, Xuchen Wu, Teng Long, and Tanpan K. Sarkar
6.1 Introduction
Space–time adaptive processing (STAP) can effectively suppress clutter and achieve
better moving target detection, as compared with traditional nonadaptive methods [1].
However, in a practical nonhomogeneous environment, the clutter characteristics would
change fast and dynamically, making it difficult to collect a sufficient number of inde-
pendent and identically distributed (i.i.d.) training samples for effective estimation of the
clutter covariance matrix. Meanwhile, the computation of high-dimensional covariance
matrix inversion requires a prohibitive computational complexity [2–4]. Therefore it is
important to develop STAP methods for robust clutter suppression and fast computation
capabilities.
In the past three decades, many suboptimal STAP algorithms have been developed
[4–8]. The reduced-dimension STAP algorithms, such as the pulse repetition interval
(PRI)-staggered algorithm [4] the extended factored algorithm (EFA) [5,6], and the
joint-domain localized (JDL) algorithm [7] can reduce the requirement of training
samples. However, the covariance matrix inversion in computing the adaptive weight
vector is still a time-consuming task for clutter suppression. On the other hand, the
reduced-rank STAP algorithms [9–13], such as the principle components (PCs) [9–11],
conjugate gradient (CG) [12], and the projection approximation subspace tracking
(PAST) [13,14], can also reduce the requirement of training samples. However, because
the performance of reduced-rank STAP algorithms is sensitive to the variation of clutter
environment, an inappropriate rank selection may lead to a severe clutter suppression
performance loss. In the past decade, some sparsity-based STAP methods have been
investigated that significantly reduce the requirement of training samples [15–25]. It has
been explored that the intrinsic sparsity of the clutter spatial-temporal power spectrum
and the space–time adaptive weight vectors can be well utilized for clutter suppression
[20,23]. However, the conventional sparse recovery methods cannot obtain desirable
sparse recovery accuracy and convergence speed in a practical environment. Therefore,
it is necessary to study on the fast and robust sparsity-based STAP methods for the
practical complex clutter environment.
In this chapter, some sparsity-based STAP methods are developed by exploiting the
intrinsic sparsity of the clutter spatial-temporal power spectrum and the space–time
adaptive weight vectors. Firstly, the signal model of airborne phased array radar is
introduced, and then the intrinsic sparsity of STAP is analyzed according to the clutter
165
166 Yang, Sun, Wu, Long, and Sarkar
spatial-temporal power spectrum and the space–time adaptive weight vectors. Secondly,
based on the sparsity of the clutter spatial-temporal power spectrum, a robust and fast
iterative sparse recovery method for STAP is proposed, which can not only alleviate the
effect of noise and dictionary mismatch but also reduce the computational complexity
by recursive inverse matrix calculation. Afterwards, based on the sparsity of space–time
adaptive weight vectors, a fast STAP method based on PAST with sparse constraint is
proposed, which can provide a more robust and stable estimation of clutter subspace
for a small set of training samples. Based on the simulated and the actual airborne
phased array radar data, it is verified that the proposed methods can provide better per-
formance with small training sample support in a practical complex nonhomogeneous
environment.
The side-looking antenna array configuration of an airborne phased array radar is con-
sidered and the corresponding geometry is shown in Figure 6.1. The N-element uniform
linear antenna array aligns with the velocity direction of the platform, and the interele-
ment spacing is dA . During each coherent processing interval (CPI), M identical pulses
are transmitted with a pulse repetition frequency (PRF) of fr . The operation wavelength
is λ, the height of platform is h with a velocity denoted by va , and L fast time samples
are collected to cover the detection region in each pulse repetition interval (PRI). Each
CPI data of the received signal is stored as an N × M × L data-cube, as shown in
Figure 6.2. Each slice corresponding to a fast-time sample of the data cube is an N × M
matrix, which is stacked as an N M × 1 vector according to the channel order.
It is known that radar target detection is a binary hypothesis problem, where hypoth-
esis H1 corresponds to target presence and hypothesis H0 corresponds to target absence
when x is composed of clutter xc , the received target echo xt , and the zero-mean com-
plex Gaussian, spatially and temporally white noise xn .
According to the geometry shown in Figure 6.1, the target can be model as a strong
scatter point, whose azimuth angle is θt and the elevation angle is ϕt , the relative
velocity is vt . Therefore the target echo xt ∈ CN M×1 can be given as
where b(ωc,i ) = [1, exp(j 2πωc,i ),. . ., exp(j (M − 1)2πωc,i )]T ∈ CM×1 is the tem-
poral steering vectors and a(ϑc,i ) = [1, exp(j 2πϑc,i ),. . ., exp(j (N − 1)2πϑc,i )]T ∈
CN×1 is the spatial steering vectors, respectively. Thus, based on Melvin’s model in [4],
the space–time clutter snapshot xc ∈ CN M×1 can be expressed as
Nc
xc = ξ̃i v(ωc,i ,ϑc,i ), (6.5)
i=1
where ξ̃i denotes the random complex amplitude corresponding to the ith clutter patch.
The conventional STAP weight vector w ∈ CN M×1 is derived with the minimum
noise variance (MNV) principle [1], which is shown as the following constrained power
minimization problem
4 42
4 4
min J (w) = E 4wH x4 s.t. wH v(ωt ,ϑt ) = 1, (6.6)
w 2
where v (ωt ,ϑt ) is the steer vector of the target. Then by using the method of Lagrange
multipliers, the optimal adaptive weighting vector is obtained as
It is well known and analyzed in some reference that the space–time covariance matrix
R ∈ CN M×N M is usually estimated from K i.i.d. training samples around the lth range
cell under test (CUT) [1,2,4], i.e.,
1
K−1
R̃ = E xx H ≈ xk xkH = R̃c + R̃n . (6.8)
K −1
k=1,k!=l
In (6.8), R̃c denotes the clutter covariance matrix, the noise covariance matrix is
R̃n = σ2 I, the noise power is σ2 and I is the N M × N M identity matrix.
R = Uc c Uc + σ2 Un UH
n, (6.9)
where Uc denotes the clutter subspace spanned by the principal eigenvectors, c = diag
(ς1,. . .,ς P ) consists of the P principal eigenvalues of R, and Un denotes the noise
subspace. The clutter covariance matrix can be denoted by only P principal eigenvalues
instead of Nc clutter space–time steering vector. The clutter is sparse with respect to
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 169
0.012
pi/2
pi/3
0.01 pi/6
Degree of correlation
0.008
0.006
0.004
0.002
0
0 0.1 0.2 0.3 0.4 0.5 0.6
Spatial angle/rad
Figure 6.3 Correlations between space–time steering vectors with spatial angle π/2, π/3, and π/6.
the system degrees of freedom (DoFs). In order to further demonstrate the sparsity of
clutter, the space–time correlation coefficient is given as
4 4
$ %4v(ωp,ϑp )H v(ωq ,ϑq )4
cor v(ωp,ϑp ),v(ωq ,ϑq ) = 4 4 4 4, (6.10)
4v(ωp,ϑp )4 4v(ωq ,ϑq )4
which describes the degree of correlation between different space–time steering vectors,
and the correlation results corresponding to three different spatial angles are shown in
Figure 6.3. It is found that the space–time steering vector is highly correlated to the
vectors that are spatially adjacent. It means that the clutter has a high correlation with
the spatial angle, and the space–time steering vector near the clutter ridge can replace all
the vectors in the spatial-temporal plane to approximate the clutter [27]. Therefore the
received space–time data shows high sparsity in the angle-Doppler domain [19,20]. As
shown in Figure 6.4, the major components of clutter spectrum are distributed near the
ridge. Thus, compared with the whole spatial-temporal plane, the complex amplitude of
received spectrum in most area is rather small, so the received space–time data is sparse
with respect to the whole angle-Doppler domain.
Based on this property, the homogeneous clutter can be well estimated by the sparse
recovery approaches. The spatial-temporal plane is discretized into a grid with Ns spatial
bins and N$ d Doppler
% bins, where each grid point is associated with a space–time steering
vector v fd,i ,fs,j . Therefore the space–time snapshot of the lth range cell can be
given as
xl = xc + xn = ϒ + xn, (6.11)
170 Yang, Sun, Wu, Long, and Sarkar
0.5 0
0.4 –5
0.3
–10
0.2
Normalized Doppler
–15
0.1
0 –20
–0.1
–25
–0.2
–30
–0.3
–0.4 –35
–0.5 –40
–0.5 –0.25 0 0.25 0.5
Normalized angle
Figure 6.5 The procedure of clutter suppression based on sparsity of spatial-temporal spectrum.
where ϒ = γ̃1,1, γ̃1,2,. . ., γ̃Ns ,Nd ∈ CNs Nd ×1 is the complex amplitude of the spectral
distribution. The space–time overcomplete dictionary matrix ∈ CN M×Ns Nd is given
as the collection of all space–time steering vectors, i.e.,
$ % $ % $ %
= v fd,1,fs,1 ,. . .,v fd,i ,fs,j ,. . .,v fd,Nd ,fs,Ns . (6.12)
Therefore, when the complex amplitude of the spectral distribution and the corre-
sponding space–time steering vectors are estimated effectively based on the property of
sparse distribution, the space–time covariance matrix R can be well reconstructed for
clutter suppression. As Ns Nd is much bigger than the system DoFs N × M, the space–
time dictionary is overcomplete and highly correlated, so that (6.11) is underdeter-
mined. However, based on the theory of sparse recovery [28,29], the ill-posed equation
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 171
can be solved effectively with limited number of training sample. The sparse recovery
of clutter spectrum can be solved by the following L1 -norm optimization
where L1 -norm guarantees the sparsity of complex amplitude ϒ, and the L2 -norm
restrains the estimation error within ε. Equation (6.13) can also be given as
1
P Nd
Ns
$ % $ %
R= γ̃p,i,j 2 vp fd,i ,fs,j vH fd,i ,fs,j + σ2 I, (6.15)
P
p=1 i=1 j =1
$ %
where P is the number of training data, vp fd,i ,fs,j is the space–time steering vector
in the overcomplete dictionary corresponding to the recovery result of each training data.
Figure 6.6 The relationship between the full space–time dimension and clutter subspace.
172 Yang, Sun, Wu, Long, and Sarkar
recursive least square (RLS)-type algorithms applied to system identification are studied
in [30,31], which results in a performance improvement for sparse systems, meanwhile,
several novel STAP methods with sparse constraint have recently been proposed in
[23,24]. Among these methods, the concept of the L1 -norm sample matrix inversion
(SMI) method is reviewed in this section. By employing the sparse constraint to the
MNV cost function, the problem is described as the following constrained optimization
problem
4 4
4 H 42
min J1 (w) = E 4w xl 4 + 2κ (w) s.t. wH vt (ωt ,ϑt ) = 1, (6.16)
w 2
where (w) is a term to characterize the sparsity of the weight vector, and κ is a
positive scalar that provides a trade-off between the sparsity and the output power. A
larger value of κ implies that more components will be shrunk to zero. Because of the
convexity of the L1 -norm constraint, it is common practice to approximate the sparse
constraint as (w) = w1 [24]. Therefore the adaptive weight vector w deduced by
Yang et al. [23] is
−1
R̂ + κ vt (ωt ,ϑt )
w= −1 , (6.17)
t (ωt ,ϑ t ) R̂ + κ
vH vt (ωt ,ϑt )
where = diag |w11|+ε , |w21|+ε ,. . ., |wNM1 |+ε and ε is a small positive constant and wi ,
i = 1,2,. . .,NM are the entries of the adaptive weighting vector w. Compared with the
adaptive weight vector obtained in (6.7), it is obvious that the sparse constraint yields
an additional term κ in the inversion of the estimated clutter covariance matrix R̂.
However, the adaptive weight vector in (6.17) is not a closed-form solution, since is
a function of w. Some iterative methods, such as L1 -norm RLS and L1 -norm recursive
SMI methods, can be used to compute the adaptive weight vector [23,24].
6.4.1 Robust and Fast Iterative Sparse Recovery Methods for STAP
In the past decade, by exploiting the intrinsic sparsity of the clutter in the angle-Doppler
domain, some sparse-based STAP methods have been proposed to achieve the sparse
recovery of the clutter spatial-temporal spectrum with limited training data [17–22].
A global matched filter (GMF) [17] is firstly applied to the STAP, which demonstrates
that both target and clutter can be identified based on a single snapshot without prior
estimation of the clutter covariance matrix. Then, by assuming the knowledge of the
clutter ridge in the spatial-temporal plane, targets are estimated as the sparse solution in
the angle-Doppler domain outside this clutter ridge [18]. In these methods, the sparse
recovery of the clutter spatial-temporal spectrum is usually formulated as a regularized
optimization problem, which can be solved by convex optimization [19]. The compu-
tational complexity of such an operation will increase beyond our capacity when the
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 173
Nd
Ns
$ %
xl = xc + xn = γ̃i,j v fd,i ,fs,j + xn = ϒ l + xn, (6.18)
i=1 j =1
A. Method Formulation
a. Sparse Recovery Processing
The effect of additive noise is not considered in the basic sparse recovery method [19].
However, the additive noise is inevitable in practical environments, which will increase
the recovery error. Therefore, the regularized processing is employed to reduce the effect
of additive noise in the sparse recovery. In this method, the fast approximation FOCUSS
method is utilized for clutter sparse recovery. The sparse recovery problem in (6.14) can
be converted as
N
s Nd
4 4
min J (ϒ l ) = 4 γl,i 4 s.t. xl − ϒ l 2 ≤ ε, (6.19)
ϒl p
i=1
174 Yang, Sun, Wu, Long, and Sarkar
where α denotes the Lagrange multipliers that match the noise level. By solving the
gradient of ϒ l , we can get
8 $ %
∇L (ϒ l ) = |p| (ϒ l ) ϒ l + α H ϒ l − H xl , (6.21)
9 $ p−2 p−2 %
where (ϒ l ) = diag γl,1 ,. . ., γl,Ns Nd , then the appropriate Hessian matrix
can be obtained by
8
∇ 2 L (ϒ l ) = |p| (ϒ l ) + αH . (6.22)
$ %−2 9 $ (k) %
where δ=1/α. By defining W (k) = ϒl and (k) = W (k) , where
& p p '
(k) 1− /2 1− /2
W (k) = diag γl,1 ,. . ., γl,Ns Nd
(k)
is the diagonal weighting matrix at the
kth iteration, the solution at kth iterative can be derived as
(k+1) $ %H 0 $ %H 1−1
ϒl = W (k) (k) δI + (k) (k) xl . (6.25)
It is easily found from (6.25) that the matrix inversion is still needed to calculate
complex amplitude, which will significantly influence the convergence of iteration.
Although the adaptive subspace selection [19] can be applied to reduce the dimension
of complex amplitude in the iterations, the direct inverse calculation still cannot be
avoided. However, based on the mathematical analysis, it can be proved that the
$ %H
matrix T=δI+(k) (k) is a Hermitian matrix, therefore the matrix inversion can be
calculated recursively based on the block Hermitian matrix property [33,34].
The low-resolution estimation based on the Fourier spectrum is employed as the ini-
(0)
tial value of ϒ l , i.e., ϒ l = H xl , and then the calculation can be executed iteratively
as (6.25). During the iterations, the prominent components in are gradually reinforced,
while the remaining small components are suppressed until they become close to zero.
(k)
Finally, when the absolute difference of ϒ l is smaller than the convergence threshold,
the sparse recovery result is obtained. From (6.25), it can be found that when the noise
level is reduced to 0, i.e., δ → 0, the proposed method will degenerate to the FOCUSS
method. Because the regularized processing is applied in the iteration, the proposed
method can effectively improve the recovery performance under noise.
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 175
the iteration will stop, and then the sparse recovery is achieved. Then the reconstruction
of space–time covariance matrix R and adaptive weighting vector can be calculated
correspondingly.
B. Method Verification
In this section, simulated data and two actual measured airborne phased array radar
data sets (MCARM [34] and one other data) are used to verify the clutter suppression
performance of the proposed methods, and compared with SMI, EFA [3], STAP using
CVX [19], and FOCUSS [19] methods.
a. Simulated Data
The simulation parameters are listed in Table 6.1, where ρs and ρ d are set to be 4 in the
simulations, and all the results are averaged over 500 Monte Carlo runs.
The output SINRs versus the number of snapshots based on SMI, EFA, STAP using
CVX, FOCUSS, and the proposed method are investigated and the results are shown
in Figure 6.7. It is found that the sparsity-based STAP methods can obtain desirable
176 Yang, Sun, Wu, Long, and Sarkar
24
22
20
Output Power [dB]
18
16
14
OPT
SMI
12 EFA
FOCUSS
STAP using CVX
Proposed Method
10
124 8 16 32 48 64 72 96 128
Range Cells
SINR performance with small training sample support, which exhibits much faster
convergence than conventional STAP methods. Meanwhile, the proposed method can
provide better performance than other sparsity-based methods with the same number of
training samples, because the calibration of space–time overcomplete dictionary and the
regularization in the sparse recovery processing are applied in the proposed method.
The output SINRs with four training samples are investigated correspondingly and
the results are shown in Figure 6.8. It is found that the SMI and EFA methods cannot
obtain desirable SINR with minimal training sample support, owing to the insufficient
estimation of the clutter covariance matrix, so that these two methods could not provide
desirable target detection performance in a practical clutter environments. However,
the sparsity-based STAP methods could provide much better performance in the instance
of good estimation of clutter distribution, so that the clutter covariance matrix can be
well reconstructed. Moreover, the proposed method can obtain better sparse recovery
performance than conventional sparsity-based STAP methods, because of the calibration
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 177
25
15
5
SINR [dB]
−5
−15
−25 OPT
SMI
EFA
−35 FOCUSS
STAP using CVX
Proposed Method
−45
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Normalized Doppler
Figure 6.8 Output SINRs of SMI, EFA, STAP using CVX, FOCUSS, and our proposed method.
b. MCARM Data
The MCARM data [34] is used to verify the STAP methods in this section. The array was
an L-band phased array antenna using 22 elements arranged as a 2 × 11 configuration.
The PRF of the radar is 1,984 Hz, 128 pulses are contained in one CPI, the platform
velocity is 100 m/s, and the height of the platform is 3,078 m. We note that ρ s and ρ d
are both set to 6 while 12 pulses and 8 elements data of MCARM are used, the target
is located at the 299th range cell with −0.15 Doppler frequency, and 4 range cell data
around the 299th range cell are selected as the training samples.
The output SINRs are investigated and the results are shown in Figure 6.10. It is
similar with the previous simulated results that SMI and EFA methods cannot obtain
desirable SINR with minimal training sample support. Moreover, the proposed method
178 Yang, Sun, Wu, Long, and Sarkar
40
30
20
Output Power [dB]
10
−10
−20
−30 SMI
EFA
−40 FOCUSS
STAP using CVX
Proposed Method
−50
1 50 100 150
Range Cells
Figure 6.9 Range detections of SMI, EFA, STAP using CVX, FOCUSS, and our proposed
method.
25
10
SINR [dB]
−10
SMI
EFA
FOCUSS
−20 STAP using CVX
Proposed Method
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Normalized Doppler
Figure 6.10 Output SINRs of SMI, STAP using CVX, FOCUSS, and our proposed method based
on MCARM data.
can obtain better SINR performance in both main-lobe and side-lobe regions than in
conventional sparsity-based STAP methods.
The range detections are also investigated and the results are shown in Figure 6.11.
It is similar to the previous simulated results, in that the SMI and EFA methods cannot
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 179
−10
Output Power [dB]
−20
−30
SMI
EFA
−40 FOCUSS
STAP using CVX
Proposed Method
250 270 290 310 330 350
Range Cells
Figure 6.11 Range detections of SMI, EFA, STAP using CVX, FOCUSS, and our proposed
method based on MCARM data.
detect the target effectively, while the sparsity-based STAP methods can provide desir-
able detection. The proposed method can also obtain a larger difference of the output
power between the tested range cell and adjacent range cells than the FOCUSS method
and STAP using CVX, so that the target can be detected correctly.
25
10
Output SINR [dB]
−10
SMI
EFA
FOCUSS
−20
STAP using CVX
Proposed Method
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Normalized Doppler
Figure 6.12 Output SINRs of SMI, STAP using CVX, FOCUSS, and our proposed method based
on actual measured airborne radar data.
85
75
Output Power [dB]
65
55
45 SMI
EFA
FOCUSS
STAP using CVX
35 Proposed Method
200 220 240 260 280
Range Cells
Figure 6.13 Range detections of SMI, EFA, STAP using CVX, FOCUSS, and our proposed
method based on actual measured airborne radar data.
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 181
The PAST method [35] is derived by minimizing the mean squares error between the
space–time received data and its projection on the clutter subspace, which is denoted by
54 42 6
J2 (Ŵc ) = E 4xl − Ŵc ŴH 4
c xl 2
$ % $ % $ H % (6.32)
c R̂Ŵc + tr Ŵc R̂Ŵc · Ŵc Ŵc ,
= tr R̂ − 2tr ŴH H
where the N M × P matrix Ŵc spans the clutter subspace as Uc , and tr(R̂) denotes the
trace of the covariance matrix R̂. The cost function J2 (Ŵc ) obtains its global minimum
only when Us contains the principal eigenvectors of R̂. Otherwise, all the stationary
points of J2 (Ŵc ) are saddle points. Therefore, similar to the L1 -norm SMI method, by
imposing the sparse constraint to J2 (Ŵc ), the cost function of the proposed method is
described as
54 42 6 $ %
J3 (Ŵc ) = E 4xl − Ŵc ŴH 4
c xl 2 + 2κ Ŵc
$ % $ %
= tr R̂ − 2tr ŴH (6.33)
c R̂Ŵc
$ H % $ %
c Ŵc + 2κ Ŵc .
+ tr Ŵc R̂Ŵc · ŴH
The cost function in (6.33) can be represented in the following exponentially weighted
form
k 4 42 $ %
4 4
J3 (Ŵc (k)) = ρ K−i 4xi − Ŵc (k)ŴcH (k)xi 4 + 2κ Ŵc (k)
2
i=1
$ % $ % (6.34)
= tr R̂(k) − 2tr ŴcH (k)R̂(k)Ŵc (k)
$ % $ %
+ tr ŴcH (k)R̂(t)Ŵc (k) · ŴcH (k)Ŵc (k) + 2κ Ŵc (k) ,
where ρ is a forgetting factor to ensure that data in the distant past are downweighted,
which can provides a trade-off between the tracking capability and evaluated error
when the system operates in a nonstationary environment. In addition [11,37], R̂(k)
= ki=1 ρ k−i xi xHi = ρ R̂(k − 1) + xk xk denotes the exponentially weighted covariance
H
matrix. As J3 (Ŵc (k)) is a fourth-order function of the elements of Ŵc (k), iterative
processing is thus necessary to minimize J3 (Ŵc (k)). The core idea of the PAST method
is to employ ŴH c (i − 1)xi to approximate the projection Ŵc (k)xi for 1 ≤ i ≤ k.
H
k 4 42 & '
4 4
J4 (Ŵc (k)) = ρ k−i 4xi − Ŵc (k)ŴcH (k)xi 4 + 2κ Ŵc (k)
2
i=1
(6.35)
k 4 42 & '
4 4
= ρ k−i 4xi − Ŵc (k)yi 4 + 2κ Ŵc (k) ,
2
i=1
where yi = ŴH c (i − 1)xi . The projection approximation changes the error performance
surface of J3 (Ŵc ). For stationary or slowly time-varying signals, the difference between
ŴHc (i − 1)xi and Ŵc (k)xi is small, and as the number of iterations increases, the
H
effect of the past input data gradually becomes insignificant. As a result, J4 (Ŵc (k))
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 183
can effectively approximate J3 (Ŵc ) and a desirable estimation of the clutter subspace
can be obtained by minimizing J4 (Ŵc (k)).
Note that J4 (Ŵc (k)) is a second-order function of the estimated clutter subspace
$ Ŵc%(k). Therefore,
approximation $ %based on the reference in [23], the sparse constraint is
given as Ŵc (k) = sign Ŵc (k) , the sub-gradient of (6.35) with respect to Ŵc (k) is
given as
$ %
∇JŴc (k) = −rxy + Ŵc (k)Ryy + κsign Ŵc (k) , (6.36)
where
sign (·) is a component-wise sign function, which is defined as sign (x)
x/|x| for x != 0
= [22,23], and rxy (k) and Ryy (k) are given as
0 for x = 0
k
rxy (k) = ρ k−i xi yH
i = ρrxy (k − 1) + xk yk
H
i=1
(6.37)
k
Ryy (k) = ρ k−i yi yH
i = ρRyy (k − 1) + yk yk .
H
i=1
By equating the above gradient terms to zero, the clutter subspace can be estimated as
$ $ %%
Ŵc (k) = rxy (k) − κsign Ŵc (k) R−1 yy (k). (6.38)
It is obvious that using (6.38) directly is computationally consuming, but the computa-
tional complexity can be effectively reduced by employing the RLS method. Let
$ %
T (k) = rxy (k) − κsign Ŵc (k) . (6.39)
Then, substituting (6.37) into the above equation, T (k) can be described by the
following recursive equation
0 $ % $ %1
T (k) = ρT (k − 1) + xk yHk − κsign Ŵc (k) − ρκsign Ŵc (k − 1) . (6.40)
Because the instantaneous error of the weight vector changes slowly in each time step,
the sign of the weights does not change rapidly. As such, T (k) can be approximated by
$ % $ %
T (k) ≈ ρT (k − 1) + xk yH k + κ ρ − 1 sign Ŵc (k − 1) . (6.41)
h(k)
where h(k) = P (k − 1) yk , g(k) = . By substituting (6.41) and (6.44) into
ρ+yH
k h(k)
(6.42), the iterative relationship between Ŵc (k) and Ŵc (k − 1) can be obtained as
Ŵc (k) = Ŵc (k − 1) + xk − Ŵc (k − 1)yk gH (k)
1 $ % (6.45)
+κ 1− sign Ŵc (k − 1) P (k − 1) − g (k) hH (k) .
ρ
After estimating the clutter subspace Ŵc (k) from the proposed method, the space–
time adaptive weight vector can be obtained as
$ %
I − Ŵc ŴHc vt (ωt ,ϑ t )
wSC−P AST = $ % . (6.46)
t (ωt ,ϑ t ) I − Ŵc Ŵc vt (ωt ,ϑ t )
vH H
Compared with the solution of the conventional PAST method [35], because of the
sparse constraint, there is an additional term in (6.38). It has been proved in [38] that
the mean squared error (MSE) of L1 -norm constraint RLS method is lower than that
of RLS, which means that the L1 -norm constraint RLS method represents a signifi-
cant performance improvement over the conventional RLS method. As in the proposed
method, the RLS method is employed for estimating the clutter subspace, thus a more
robust solution of the estimated clutter subspace can be obtained because of the sparse
constraint.
The following three important observations are in order. First, the initial value of P(k)
must be a Hermitian and positive-definite matrix, and the initial value of Ŵc (k) should
be composed of P orthogonal vectors. Therefore, without loss of generality, P(0) is set
to the P × P identity matrix and the columns of Ŵc (0) are set to the P leading unit
vectors of the N M × N M identity matrix. Second, in order to maintain the Hermitian
symmetry, an operation Tri {·} denoted by the operator is employed in the processing,
which indicates that only the upper (or lower) triangular part of P(k) is reserved, and its
Hermitian transposed version is copied to the opposite triangular part. Third, as PAST
may not converge to an orthonormal basis [35], an operation denoted by the operator
Orth {·} is employed in each iteration to guarantee the orthogonality between the vectors
in Ŵc (k).
B. Method Verification
In this section, simulated data and two actual measured airborne phased array radar
data sets (MCARM [34] and one other data) are used to verify the clutter suppression
performance of the proposed L1 -norm PAST method. The performance is compared
with those obtained from the SMI, EVD, PAST, L1 -norm CCG [23,24], L1 -norm mod-
ified conjugate gradient (L1 -norm MCG) [23,24], and shrinkage operator STAP [11]
methods.
a. Simulated Data
The parameters of the simulated data are listed in Table 6.2. In the simulations, the
SMI method is implemented with a loading factor of 10 dB above the noise power, the
forgetting factor ρ is set to 0.97, and zero vectors are used to initialize the adaptive
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 185
70 70
60 60
Improvement Factor[dB]
Improvement Factor[dB]
50 50
40 40
30 30
OPT OPT
SMI L1−norm CCG
20 20
EVD L1−norm MCG
PAST Shrinkage Operator STAP
L1−norm PAST L1−norm PAST
10 10
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
Number of Samples Number of Samples
Figure 6.14 IF performance of EVD, PAST, L1-norm PAST, L1-norm CCG, L1-norm MCG, and
shrinkage operator STAP methods versus the number of training samples.
weight vector in the L1 -norm CCG and L1 -norm MCG methods. A scaled N M × N M
identity matrix εI is used as the initialization matrix for the PAST, L1 -norm PAST,
L1 -norm CCG, and L1 -norm MCG methods, respectively, where ε = 1 × 10−3 is used.
All results are averaged over 500 Monte Carlo runs.
The improvement factor (IF) performance of the proposed method with respect to
the number of training samples is given and compared with the EVD, PAST, L1 -norm
PAST, L1 -norm CCG, L1 -norm MCG, and shrinkage operator STAP methods. It is
assumed that the selected rank is 15, and the value of sparse constraint parameter is set to
κ = 1000. The results are shown in Figure 6.14. It is clear that the SMI method obtains
the worst convergence performance, whereas the proposed L1 -norm PAST converges
with the lowest number of training samples and outperforms the conventional PAST
method. On the other hand, the EVD, L1 -norm CCG and shrinkage operator STAP
methods achieve comparable performance. In addition, compared with the L1 -norm
PAST, the L1 -norm CCG methods and the shrinkage operator STAP method, the L1 -
norm MCG method suffers from significant performance loss when the training sample
support is small.
The IF performance of different methods with respect to the target Doppler frequency
using a small training samples support is shown in Figure 6.15. It is assumed that the
selected rank is 15, the number of training samples is 40, and the sparse constraint
186 Yang, Sun, Wu, Long, and Sarkar
70 70
60 60
50 50
Improvement Factor [dB]
30 30
20 20
OPT OPT
SMI L1−norm CCG
10 EVD 10 L1−norm MCG
PAST Shrinkage Operator STAP
L1−norm PAST L1−norm PAST
0 0
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Normalized Doppler Normalized Doppler
Figure 6.15 IF performance of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG, L1-norm
MCG, and shrinkage operator STAP methods versus the target Doppler frequency.
50 50
40 40
Improvement Factor [dB]
Improvement Factor [dB]
30 30
SMI
L1−norm CCG
EVD
20 20 L1−norm MCG
PAST Shrinkage Operator STAP
L1−norm PAST L1−norm PAST
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4
Normalized Doppler Normalized Doppler
Figure 6.16 IF performance of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG, L1-norm
MCG, and shrinkage operator STAP methods based on 200 training samples of MCARM data.
parameter is κ = 1000. Similar to Figure 6.14, the SMI method yields the worst IF
performance, while the EVD, the L1 -norm CCG, and shrinkage operator STAP methods
perform the best. The proposed L1 -norm PAST obtains better IF performance than the
conventional PAST method. On the other hand, the L1 -norm MCG method suffers from
some performance loss for the underlying small training sample support case when
compared with the L1 -norm PAST and L1 -norm CCG methods.
b. MCARM Data
In the verifications, 12 pulses and 8 elements data of the MCARM are applied, and the
target is located at the 299th range cell with a normalized Doppler frequency of −0.15.
The rank is selected as 19, the SMI method is implemented with a loading factor of 10
dB above the noise power, the value of the sparse constraint parameter is chosen to be
κ = 1000, and the initialization matrix for the PAST, L1 -norm PAST, L1 -norm CCG,
and L1 -norm MCG methods are chosen to be 10−3 I.
First, 201 training samples from range cells number 200 to number 400 are used,
and the IF results and range detection are shown in Figures 6.16 and 6.17. It is
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 187
5 5
L1−norm CCG L1−norm CCG
L1−norm MCG L1−norm MCG
0 0
Shrinkage Operator STAP Shrinkage Operator STAP
L1−norm PAST L1−norm PAST
−5 −5
−10 −10
Output Power [dB]
−20 −20
−25 −25
−30 −30
−35 −35
−40 −40
280 290 300 310 320 330 340 350 360 280 290 300 310 320 330 340 350 360
Range Cell Range Cell
Figure 6.17 Range detection of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG,
L1-norm MCG, and shrinkage operator STAP methods based on 200 training samples of
MCARM data.
40 40
Improvement Factor [dB]
Improvement Factor [dB]
30 30
20 20
10 10
SMI L1−norm CCG
EVD L1−norm MCG
PAST Shrinkage Operator STAP
L1−norm PAST L1−norm PAST
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Normalized Doppler Normalized Doppler
Figure 6.18 IF performances of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG,
L1-norm MCG, and shrinkage operator STAP methods based on 20 training samples of
MCARM data.
seen that, when a sufficient number of training samples are available, all the STAP
methods achieve acceptable IF performance and the target can be detected correctly.
Next, we only use 21 training samples from range cells number 260 to 280, and the
yielding range detection and IF results are depicted in Figures 6.18 and 6.19. It is
evident that, as the number of training samples becomes smaller, the SMI and L1 -
norm MCG methods no longer perform properly, whereas the EVD, PAST, L1 -norm
PAST, L1 -norm CCG, and shrinkage operator STAP methods still provide satisfactory
performance.
5 5
0 0
−5 −5
−10 −10
Output Power [dB]
−20 −20
−25 −25
Figure 6.19 Range detection of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG, L1-norm
MCG, and shrinkage operator STAP methods based on 20 training samples of MCARM data.
60 60
50 50
Improvement Factor [dB]
Improvement Factor [dB]
40 40
30 30
20 20
SMI L1−norm CCG
EVD L1−norm MCG
10 PAST 10 Shrinkage Operator STAP
L1−norm PAST L1−norm PAST
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Normalized Doppler Normalized Doppler
Figure 6.20 IF performances of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG, L1-norm
MCG, and shrinkage operator STAP methods based on 300 training samples of actual measured
airborne radar data.
SMI method is implemented with a loading factor of 10 dB above the noise power, the
value of the sparse constraint parameter is selected as κ = 1000, and 10−3 I is used
as the initialization matrix for the PAST, L1 -norm PAST, L1 -norm CCG, and L1 -norm
MCG methods.
We first use 301 training samples from range cells number 120 to 400, and the
yielding range detection and IF results are shown in Figures 6.20 and 6.21. Similar
to the results based on the MCARM data, as the number of training samples is sufficient
in this case, all the STAP methods achieve acceptable IF performance and the target can
be detected correctly. We then reduce the number of training samples to 21 collected
between range cells number 120 and 140, and the corresponding range detection and IF
results are shown in Figures 6.22 and 6.23. Again, similar to the MCARM data case,
as the number of the training samples is insufficient in this case, the SMI and the L1 -
norm MCG methods do not offer satisfactory performance, whereas the EVD, PAST,
L1 -norm PAST, L1 -norm CCG, and shrinkage operator STAP methods still function
well.
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 189
0 0
SMI L1−norm CCG
EVD L1−norm MCG
PAST Shrinkage Operator STAP
−10 L1−norm PAST −10 L1−norm PAST
−20 −20
Output Power [dB]
−40 −40
−50 −50
150 180 210 240 270 300 150 180 210 240 270 300
Range Cells Range Cells
Figure 6.21 Range detection of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG, L1-norm
MCG, and shrinkage operator STAP methods based on 300 training samples of actual measured
airborne radar data.
60 60
50 50
Improvement Factor [dB]
Improvement Factor [dB]
40 40
30 30
20 20
L1−norm CCG
SMI
EVD L1−norm MCG
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Normalized Doppler Normalized Doppler
Figure 6.22 IF performances of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG, L1-norm
MCG, and shrinkage operator STAP methods based on 20 training samples of actual measured
airborne radar data.
0 0
−10
−10
−20
Output Power [dB]
Output Power [dB]
−20
−30
−30
−40
−40
−50 SMI L1−norm CCG
EVD L1−norm MCG
PAST Shrinkage Operator STAP
−60 −50
L1−norm PAST L1−norm PAST
150 180 210 240 270 300 150 180 210 240 270 300
Range Cells Range Cells
Figure 6.23 Range detection of SMI, EVD, PAST, L1-norm PAST, L1-norm CCG, L1-norm
MCG, and shrinkage operator STAP methods based on 20 training samples of actual measured
airborne radar data.
190 Yang, Sun, Wu, Long, and Sarkar
6.5 Conclusions
In this chapter, fast and robust sparsity-based STAP methods for practical environments
have been developed by exploiting the intrinsic sparsity ofthe clutter spatial-temporal
power spectrum and the space–time adaptive weight vectors. Firstly, the signal model of
received space–time data for airborne phased array radar is introduced, and the intrinsic
sparsity of STAP is analyzed according to the clutter spatial-temporal power spectrum
and the space–time adaptive weight vectors. Secondly, based on the sparsity of clutter
spatial-temporal power spectrum, a robust and fast iterative sparse recovery method for
STAP is introduced, which can not only alleviate the effect of noise and dictionary
mismatch, but also reduce computational complexity by the recursive inverse matrix
calculation. Afterwards, based on the sparsity of space–time adaptive weight vectors, a
fast STAP method based on PAST with sparse constraint is introduced, which provides
a more robust and stable estimation of clutter subspace when there is only a small set of
training samples is available. Based on the simulated and the actual airborne phased
array radar data, it is verified that the proposed methods can provide better perfor-
mance with minimal training sample support in practical complex nonhomogeneous
environments.
References
[1] J. R. Guerci, Space-Time Adaptive Processing for Radar. Artech House, 2003.
[2] J. Ward, “Space-time adaptive processing for airborne radar,” Technical Report 1015, MIT
Lincoln Laboratory, Dec. 1994.
[3] W. Zhang, Z. He, and J. Li, “A method for finding best channels in beam-space post-Doppler
reduced-dimension STAP,” IEEE Trans. Aerosp. Electron. Syst., vol. 50, pp. 254–264, 2013.
[4] W. L. Melvin, “Space-time adaptive radar performance in heterogeneous clutter,” IEEE
Trans. on Aerospace and Electronic Systems, vol. 36, pp. 621–633, 2000.
[5] Y. L. Wang, Y. N. Peng, and Z. Bao, “Space-time adaptive processing for airborne radar with
various array orientation,” IET Radar, Sonar Navigation, vol. 144, pp. 330–340, 1997.
[6] W. Zhang, Z. He, and J. Li, “A method for finding best channels in beam-space post-Doppler
reduced-dimension STAP,” IEEE Trans. on Aerospace and Electronic Systems, vol. 50, pp.
254–264, 2013.
[7] H. Wang and L. Cai, “On adaptive spatial-temporal processing for airborne surveillance
radar systems,” IEEE Trans. on Aerospace and Electronic Systems, vol. 30, pp. 660–670,
1994.
[8] A. K. Shackelford, K. Gerlach, B. D. Blunt, “Partially adaptive STAP using the FRACTA
algorithm,” IEEE Trans. on Aerospace and Electronic Systems, vol. 45, pp. 58–69, 2009.
[9] T. Long, Y. Liu, X. Yang, and Y. Sun, “Improved eigenanalysis canceler based on data-
independent clutter subspace estimation for space-time adaptive processing,” Science China:
Information Sciences, vol. 56, no. 10, pp. 1–10, 2013.
[10] R. Fa and R. C. de Lamare, “Reduced-rank STAP algorithms using joint iterative optimiza-
tion of filters,” IEEE Trans. on Aerospace and Electronic Systems, vol. 47, pp. 1668–1684,
2011.
Fast and Robust Sparsity-Based STAP Methods for Nonhomogeneous Clutter 191
[11] S. Sen, “Low-rank matrix decomposition and spatio-temporal sparse recovery for STAP
radar,” IEEE Journal of Selected Topics in Signal Processing, vol. 9, no. 8, pp. 1510–1523,
2015.
[12] Y. L. Wang, W. J. Liu, and W. C. Xie, “Reduced-rank space-time adaptive detector for
airborne radar,” Science China: Information Sciences, vol. 57, no. 8, pp. 1–11, 2014.
[13] W. J. Liu, W. C. Xie, R. F. Li, Z. T. Wang, and Y. L. Wang, “Adaptive detectors in the Krylov
subspace,” Science China: Information Sciences, vol. 57, no. 10, pp. 1–11, 2014.
[14] P. Parker, A. L. Swindlehurst, “Space-time autoregressive filtering for matched subspace
STAP,” IEEE Trans. on Aerospace and Electronic Systems, vol. 39, no. 2, pp. 510–520,
2003.
[15] M. A. Herman and T. Strohmer, “High-resolution radar via compressed sensing,” IEEE
Trans. on Signal Processing, vol. 57, no. 6, pp. 2275–2284, 2009.
[16] K. R. Varshney, M. Cetin, J. W. Fisher, and A. S. Willsky, “Sparse representation in
structured dictionaries with application to synthetic apertureradar,” IEEE Trans. on Signal
Processing, vol. 56, no. 8, pp. 3548–3561, 2008.
[17] S. Maria and J. J. Fuchs, “Application of the global matched filter toSTAP data an efficient
algorithmic approach,” in Proc. IEEE Int. Conf. Acoust. Speech Signal Process. Toulouse,
France, May 2006, pp. 14–19.
[18] I. W. Selesnick, S. U. Pillai, K. Y. Li, and B. Himed, “Angle-Doppler processing using sparse
regularization,” in Proc. IEEE ICASSP, Dallas, TX, Mar. 2010, pp. 2750–2753.
[19] K. Sun, H. Zhang, G. Li, H. Meng, and X. Wang, “A novel STAP algorithm using sparse
recovery technique,” in Proc. IGARSS, Cape Town, South Africa, July 2009, pp. 336–339.
[20] Z. Yang, X. Li, H. Wang, and L. Nie, “Sparsity-based space-time adaptive processing using
complex-valued homotopy technique for airborne radar,” IET Signal Processing, vol. 8, no.
5, pp. 552–564, 2014.
[21] X. Yang, Y. Sun, T. Zeng, and T. Long, “Robust and fast iterative sparse recovery method
for space-time adaptive processing,” Science China: Information Sciences, vol. 59, no. 6,
pp. 1–13, 2016.
[22] Q. Wu, Y. D. Zhang, M. G. Amin, and B. Himed, “Space-time adaptive processing and
motion parameter estimation in multi-static passive radar exploiting Bayesian compressive
sensing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 2, pp. 944–
957, 2016.
[23] Z. Yang, R. C. de Lamare, and X. Li, “L1-regularized STAP algorithms with a generalized
sidelobe canceller architecture for airborne radar,” IEEE Trans. on Signal Processing, vol.
60, no. 2, pp. 674–686, 2012.
[24] Z. Yang, R. C. de Lamare, and X. Li, “Sparsity-aware space-time adaptive processing
algorithms with L1-norm regularisation for airborne radar,” IET Signal Processing, vol. 6,
no. 5, pp. 413–423, 2012.
[25] X. Yang, Y. Sun, T. Zeng, and T. Long, “Fast STAP method based on PAST with sparse
constraint for airborne phased array radar,” IEEE Trans. On Signal Processing, vol. 64, no.
17, pp. 4550–4561, 2016.
[26] A. Haimovich, “The eigencanceller: Adaptive radar by eigenanalysis methods,” IEEE Trans.
on Aerospace and Electronic Systems, vol. 32, no. 2, pp. 532–542, 1996.
[27] M. E. Tipping, “Sparse Bayesian shrinkage and selection learning andthe relevance vector
machine,” Journal of Machine Learning Research, vol. 1, no. 9, pp. 211–244, 2001.
192 Yang, Sun, Wu, Long, and Sarkar
A radar system emits probing signals and records the reflections. Estimating the relative
angles, delays, and Doppler shifts from the received signals allows to determine the
locations and velocities of objects. However, due to practical constraints, the prob-
ing signals have finite bandwidth B, the received signals are observed over a finite
time interval of length T only, and a radar typically has only one or a few transmit
and receive antennas. These constraints fundamentally limit the resolution up to which
objects can be distinguished. Specifically, a radar cannot distinguish objects with delay
and Doppler shifts much closer than 1/B and 1/T , respectively, and a radar system with
NT transmit and NR receive antennas cannot distinguish objects with angels closer than
1/(NT NR ). As a consequence, the delay, Doppler, and angular resolution of standard
radars is proportional to 1/B and 1/T , and 1/(NT NR ). In this chapter, we show that the
continuous angle-delay-Doppler triplets and the corresponding attenuation factors can
be resolved at much finer resolution, using ideas from compressive sensing. Specifically,
provided the angle-delay-Doppler triplets are separated either by factors proportional to
1/(NT NR −1) in angle, 1/B in delay, or 1/T in Doppler direction, they can be recovered
at a significantly smaller scale or higher resolution.
7.1 Introduction
193
194 Heckel
the relative distances and velocities of objects with a single measurement. However, the
angular resolution of a MIMO radar is 1/(NT NR ), where NT and NR are the number of
transmit and receive antennas, and is thus again limited by a physical constraint, namely
the number of antennas (see Section 7.6 for a detailed argument on the resolution).
Even though objects that are simultaneously much closer than 1/(NT NR ) in angle,
1/B in delay, and 1/T in Doppler direction, are impossible to distinguish for real-world
radar systems in general, it is still possible to determine the locations of the objects with
a much higher degree of accuracy than the resolution limit of (1/(NT NR ),1/B,1/T ). In
this chapter, we discuss signal recovery techniques for building super-resolution radar
systems that can achieve localization accuracy below the resolution limit. In more detail,
we study the problem of recovering the continuous delays and Doppler shifts in a SISO
radar system, and the problem of recovering the angles, delays, and Doppler shifts in a
MIMO radar system, in both cases from the responses to known and suitably selected
probing signals. As we see later, those problems – termed the super-resolution radar
and super-resolution MIMO radar problems – amount to recovering a signal that is
sparse in a continuous dictionary from linear measurements, and can thus be viewed
as a generalization of the traditional compressive sensing problem.
If the objects may be assumed to lie on a sufficiently coarse grid, compressed
sensing-based [3] approaches provably recover the delay-Doppler pairs for SISO radar
system [4–6], and the angle-delay-Doppler triplets for MIMO radar systems [7,8].
However, to establish those results, the aforementioned papers assume that angles,
delays, and Doppler shifts lie on a sufficiently coarse grid, specifically a grid with
spacing 1/(NT NR ),1/B, and 1/T , in angle, delay, and Doppler direction, respectively
(see Section 7.2.3). Since NT ,NR ,B, and T are physical problem parameters, they
can in general not be made (arbitrarily) large in order to make the grid finer. In fact,
the coarseness of the grid is required for the measurement matrix to be incoherent,
therefore the aforementioned results cannot straightforwardly be extended to a grid
with significantly finer spacing. In some special cases, however, off-the-grid recovery
is possible with standard spectral estimation techniques. For example, for a single
input antenna and either known and constant delays (see Section 7.2.4), or known and
constant Doppler shifts, the super-resolution radar problem reduces to a standard 2D
line spectral estimation problem [8, sec. 5]. For these special cases, the object locations
can be recovered – off the grid – with standard spectral estimation techniques such as
Prony’s method, MUSIC, and ESPRIT [9]. In general, however, the super-resolution
radar problems cannot be reduced to the classical line spectral estimation problem.
Therefore, traditional spectral estimation techniques are not directly applicable.
Recently, an alternative, convex optimization based approach to solve the classical
line spectral estimation has been proposed that is much more generally applicable than
traditional line spectral estimation techniques. Specifically, the paper [10] shows that
the frequency parameters, which are the unknowns in the line spectral estimation prob-
lem, can be perfectly recovered by solving a convex total-variation norm minimization
program, provided they are sufficiently separated. Related convex programs have been
studied for compressive sensing off the grid [11], denoising [12], signal recovery from
short-time Fourier measurements [13], and the SISO and MIMO super-resolution radar
Super-Resolution Radar Imaging via Convex Optimization 195
problems [14,15], and the generalized line-spectral estimation problem [16]. The focus
of this chapter is on explaining how this convex optimization based approach enables
high resolution in radar. In particular we discuss the results in [14,15], showing that
a convex program recovers the continuous angles, delays, and Doppler shifts perfectly,
provided that they are sufficiently separated. Furthermore, we show that a simple convex
1 -minimization program recovers the angles and delay-Doppler shifts on an arbitrarily
fine grid, again provided they are sufficiently separated. Finally, we provide numerical
results demonstrating robustness to noise.
The remainder of this chapter is organized as follows. In the first part we consider
the SISO radar model. In more detail, Section 7.2 contains the radar model and formal
problem statement, in Sections 7.3 and 7.3.2 we present the convex optimization based
recovery approach and corresponding performance guarantees, and in Section 7.4 we
show that 1 -minimization recovers the locations on an arbitrarily fine grid. In Section
7.4.2, we provide numerical results demonstrating that the approach is robust to noise
and in Section 7.5.2 we outline the proof of the main technical statements. In the second
part, Section 7.6, we explain how the results for SISO radar can be extended to the
MIMO case. We conclude in Section 7.7 with a discussion on challenges in applying
those ideas in practice, and current and future research directions.
A radar system with a single transmit and single receive antenna is typically modeled
as a linear system. The response y recorded at the receive antenna, to a probing signal,
x, emitted at the transmit antenna, is a weighted superposition of delayed and Doppler-
shifted versions of the probing signal x:
y(t) = s(τ,ν)x(t − τ)ei2πν t dνd τ. (7.1)
Here, s denotes the spreading function, which describes the scene being sensed, and τ
and ν are the delays and Doppler shifts. Often, the moving objects are modeled by point
scatterers. Mathematically, this means that the spreading function specializes to
S
s(τ,ν) = bj δ(τ − τ̄j )δ(ν − ν̄j ).
j =1
S
y(t) = bj x(t − τ̄j )ei2πν̄j t . (7.2)
j =1
Thus, the received signal is a superposition of the reflections of the probing signal by
the point scatterers. The relative distances and velocities of the S-many objects can be
trivially obtained from the delay-Doppler pairs ( τ̄j , ν̄j ). In order to locate the objects,
196 Heckel
we therefore need to estimate the delay-Doppler pairs and the corresponding attenuation
factors bj from a single input–output measurement, i.e., from the response y to a known
and suitably selected probing signal x. As we will see later, the particular choice of the
probing signal is crucial for good localization performance.
This is not a restrictive assumption as this region can have area BT " 1, which is
typically very large. In fact, for many applications, it is reasonable to assume that the
delay-Doppler pairs lie in a region of area significantly smaller than one [18–20], an
assumption often referred to as the linear system being “underspread”. We do not make
or require this assumption here.
By the 2W T -Theorem [21], band and approximate time limitation of the response
y implies that y is essentially characterized by on the order of BT coefficients. We
therefore sample y in the interval [−T /2,T /2] at rate 1/B, so as to collect L := BT
samples, denoted by yp := y(p/B) (for simplicity we assume in the following that
L = BT is an odd integer). As detailed in [14, sec. 5], those samples are given by
S
L−1
yp = bj [Fνj Tτj x]p, p = −N,. . .,N, N := , (7.3)
2
j =1
Super-Resolution Radar Imaging via Convex Optimization 197
where
2 3
1
N N
pk
−i2π k
[Tτ x]p := x e L e−i2πkτ ei2π L (7.4)
L
k=−N =−N
and
[Fν x]p := xp ei2πpν .
Here, we defined the time-shifts τj := τ̄j /T and frequency shifts νj := ν̄j /B. To
avoid ambiguity, from here onwards we refer to (τ̄j , ν̄j ) as a delay-Doppler pair and
to (τj ,νj ) as a time–frequency shift. From (τ̄j , ν̄j ) ∈ [−T /2,T /2] × [−B/2,B/2] we
have (τj ,νj ) ∈ [−1/2,1/2]2 . Since Tτ x and Fν x are 1-periodic in τ and ν, we assume
in the remainder of the chapter without loss of generality that (τj ,νj ) ∈ [0,1]2 . The
operators Tτ and Fν have an interesting interpretation as fractional time and frequency
shift operators in CL . In fact, if the parameters τ and ν lie on a (1/L,1/L) grid, the
operators Fν and Tτ reduce to the “natural” time frequency shift operators in CL , i.e.,
The definition of a time shift in (7.4) as taking the Fourier transform, modulating the
frequency, and taking the inverse Fourier transform is a very natural definition of a
continuous time-shift τj ∈ [0,1] of a discrete vector x = [x0,. . .,xL−1 ]T .
Finally, note that to obtain the input–output relation (7.3) (see [14, sec. 5]) from (7.2),
a periodic sinc function is approximated with a finite sum of sinc functions (this is where
partial periodization of x becomes relevant). Thus, if we take the probing signal to be
essentially time-limited, then equality in (7.3) does not hold exactly. However, in [14,
sec. 5] it is shown that for a random probing
√ signal, as considered in this chapter, the
incurred relative 2 -error decays as 1/ L and is therefore negligible for large L. It
is confirmed numerically in the same paper that the approximation error made in this
process is negligible. Moreover, if we took x to be T -periodic, then the input–output
relation (7.3) becomes exact, but at the cost of the probing signal x not being time-
limited.
L-periodic sequence x . In order to see this, and to better understand the super-resolution
radar problem, it is instructive to consider two special cases.
S (νj L)p
yp = bj xp−τj L ei2π L , p = −N,. . .,N .
j =1
y = Gx b.
2
Here, the vector y contains as entries the samples yp , Gx ∈ CL×L is the Gabor matrix
with window x, defined by
kp
[Gx ]p,(k,) := xp− ei2π L , k,,p = −N,. . .,N, (7.5)
2
and b ∈ CL is a sparse vector with the j -th nonzero entry given by bj and indexed by
(τj L,νj L).
Thus, recovery of the triplets {(bj ,τj ,νj )}Sj=1 amounts to recovering the S-sparse
vector b from the measurement vector y. A – by now standard – recovery approach
is to solve a convex 1 -norm-minimization program. From [22, thm. 5.1] we know
that, provided the x are i.i.d. sub-Gaussian random variables, and provided that
S ≤ cL/(log L)4 for a sufficiently small numerical constant c, with high probability, all
S-sparse vectors b can be recovered from y via 1 -minimization. Note that the result
[22, thm. 5.1] only applies to the Gabor matrix Gx and therefore does not apply to the
super-resolution problem where the “columns” Fν Tτ x are highly correlated. In fact, the
two problems are conceptually very different: [22, thm. 5.1] shows that the columns of
the Gabor matrix Gx are nearly orthogonal, while the “columns” Fν Tτ x are extremely
correlated for two pairs of time–frequency shifts that are close.
S
yp = xp bj ei2πpνj , p = −N,. . .,N . (7.6)
j =1
Note that the samples {yp } above are samples of a mixture of S complex sinusoids, and
estimation of the coefficients {(bj ,νj )} corresponds to determining the magnitudes and
the frequency components of these sinusoids. Estimating the coefficients {(bj ,νj )} from
the samples {yp } is known as a line spectral estimation problem and can be solved using
classical approaches, such as Prony’s method [23, ch. 2], as well as convex programming
based approaches [10]. An analogous situation arises when there are only time shifts
(νj = 0 for all j ) as taking the discrete Fourier transform of yp yields a relation exactly
equal to (7.6).
We next present a convex optimization based recovery algorithm. Even though the
corresponding convex program can be solved in polynomial time, standard solvers are
currently computationally very expensive, limiting the practical applicability. However,
in Section 7.4 we will discuss a very closely related convex program that has a signif-
icantly better computational efficiency at the cost of making a small griding error that
is due to a discretization step. Since the results and intuition for both approaches are
nearly the same, we start by discussing the continuous case here.
Here, FH ∈ C L2 ×L2 is the (inverse) 2D discrete Fourier transform matrix with the entry
qk+r
in the (k,)-th row and (r,q)-th column given by [FH ](k,),(r,q) := L12 ei2π L , and the
entries of the vector f are given by [f(r)](r,q) := e−i2π(r τ+q ν ) , k,,q,r = −N,. . .,N
(here, and in the following we use for convenience a two- or three-dimensional index
2
to refer to entries of vectors and matrices). Moreover, Gx ∈ CL×L is the Gabor matrix
defined in (7.5).
The significance of the representation in (7.7) is that recovery of the unknowns
{(bj ,rj )} from z is a 2D line spectral estimation problem that can be solved with
standard spectral estimation techniques such as Prony’s method [9]. Therefore, we only
2
need to recover z ∈ CL from y ∈ CL . To do so, we use that z is a sparse linear
combination of atoms in the set A := {f(r),r ∈ [0,1]3 }. A regularizer that promotes
such a sparse linear combination is the atomic norm induced by these signals [24],
defined as
200 Heckel
zA := inf |bk | : z = bk f(rk ) .
bk ∈C,rk ∈[0,1]2
k k
We estimate z by solving the basis pursuit type atomic norm minimization problem
problem
4 4
AN(y) : minimize 4z̃4A subject to y = Az̃. (7.8)
z̃
We remark that the rk may be obtained more directly from a solution to the dual of (7.8)
[14, sec. 6]; see also [12, sec. 3.1], [10, sec. 4], [11, sec. 2.2] for details on this approach
as it is applied to related problems.
Since computation of the atomic norm involves taking the infimum over infinitely
many parameters, finding a solution to AN(y) may appear to be daunting. For the
one-dimensional case (i.e., only time or frequency shifts), the atomic norm can be
characterized in terms of linear matrix inequalities [11, prop. 2.1]. This characterization
is based on the Vandermonde decomposition lemma for Toeplitz matrices, and allows
us to formulate the atomic norm minimization program as a semidefinite program that
can be solved in polynomial time. While this lemma generalizes to higher dimensions
[25, thm. 1], it fundamentally comes with a rank constraint that appears to prohibit an
straightforward characterization of the atomic norm in terms of linear matrix inequal-
ities. Nevertheless, based on [25, thm. 1], one can obtain a semidefinite programming
relaxation of AN(y), which can be solved in polynomial time. Similarly, a solution of
the dual of AN(y) can be found with a semidefinite programming relaxation. Since the
computational complexity of the corresponding semidefinite programs is quite large,
we will not dive into the details of those semidefinite programming relaxations. As
mentioned before, instead, we show in Section 7.4 that the parameters {rj } can be
recovered on an arbitrarily fine grid via 1 -minimization. While this leads to a gridding
error, the grid may be chosen sufficiently fine for the gridding error to be negligible
compared to the error induced by additive noise, and in practice, there is always some
additive noise.
problematic and can be done in practice. Theorem 7.6 shows that, with high probability,
the triplets {(bj ,τj ,νj )} can be recovered perfectly from the samples by solving a con-
vex program, provided that the number of time–frequency shifts is sufficiently smaller
than the number of measurement, and provided that the following minimum separation
condition holds:
Note that the time–frequency shifts must not be separated in both time and frequency,
for example the minimum separation condition can hold even when τj = τj # for some
j != j # . The main result on recovery via atomic norm minimization from the paper [14]
is stated next.
theorem 7.2 Assume that the samples of the probing signal x ∈ CL are i.i.d.
N (0,1/L) random variables. Consider a signal where the sign of the attenuation
factors {bj }Sj=1 is i.i.d. uniform on {−1,1} or the complex unit disc, and suppose
that the time–frequency shifts {(τj ,νj )}Sj=1 obey the minimum separation condition.
Furthermore, choose δ > 0 and assume that the number of nonzero attenuation factors,
S, and the number of measurements, L, obey
L
S≤c 3
,
log (L/δ)
S
where c is a numerical constant. Then, with probability at least 1− δ, z = j =1 bj f(rj )
is the unique minimizer of AN(y), y = Gx FH z.
This result is essentially optimal in terms of the allowed sparsity level, as the number
S of unknowns can be linear – up to a logarithmic-factor – in the number of obser-
vations L. Even when we are given the time–frequency shifts (τj ,νj ), we can only
hope to recover the corresponding attenuation factors bj by solving the linear system of
equations in (7.3), provided that S ≤ L.
Since the complex-valued coefficients bj in the radar model describe the attenuation
factors, it is natural to assume that the phases of different bj are independent from
each other and are uniformly distributed on the unit circle of the complex plane. Indeed,
standard models for wireless communication channels and radar [26], assume the coeffi-
cients {bj } to be complex Gaussian distributed. Nevertheless, we believe that the random
sign assumption is not necessary for Theorem 7.2 to hold. In Section 7.7, we discuss
a closely related problem, which does not require the random sign assumption, and
thus provides a basis for the claim of the random sign assumption not being necessary.
Finally, we would like to point out that Theorem 7.2 continues to hold for sub-Gaussian
sequences x , for example random signs.
202 Heckel
is large, then in the presence of even a tiny amount of noise, stable recovery of the
attenuation factors and time–frequency shifts even from
S
z= bj f(rj ),
j =1
where we set, with a slight abuse of notation, [f(νj )] = ei2πj νj , is not possible (see [27,
thm. 1.1] and [10, sec. 1.7]). Condition (7.9) allows us to have 0.4 S # time–frequency
#
shifts in an interval of length 2S
L , which is optimal up to the constant 0.4.
This argument illustrates that for stable recovery, it is relevant whether a number
of frequencies, say S many, cluster together in a small interval of size smaller than
S/L. To illustrate this point further, consider again the simpler problem of line
S
spectral estimation, i.e., recovery or the frequencies from z = j =1 bj f(νj ), with
[f(νj )] = ei2πj νj . Consider the following (Vandermonde) matrix parameterized by S
and :
V = [f(0),f((1 − )/L),. . .,fL ((2S − 1)(1 − )/L)].
We next state a theorem, which provides a lower bound on the condition number of V.
The lower bound implies that there are signals with S many frequencies in an interval
smaller than S/L, that are indistinguishable even under a tiny amount of additive noise.
theorem 7.3 [28, thm. 1.3] Fix some ∈ (0,1), let K = 1− 1
L, and let
S = O(log(L/(1 − ))). Then the matrix V has condition number at least e O(S) .
The theorem implies that there exists a vector b with unit norm that obeys
Vb2 ≤ e−O(S) . As a consequence,
⎛ ⎞2
N
⎝ bj ei2πνj + bj ei2πνj ⎠ = Vb22 ≤ e−O(S),
=−N j odd j even
which means that there are two sets of S many point sources each, with separation
2(1−)
L , but telling them apart requires an exponentially small additive error. To obtain
intuition on the constants involved in the statement, we plot the condition number of V
for different values of S and in Figure 7.1.
While those two arguments show that some form of separation between the time–
frequency shifts is necessary, the exact form of separation required in (7.9) may not
be necessary for stable recovery and less restrictive conditions may suffice. Indeed, in
Super-Resolution Radar Imaging via Convex Optimization 203
1 S=2
S=4
S=8
1/κ
0.5 S = 16
S = 32
0
0.6 0.8 1
1−
Figure 7.1 Inverse of the condition number κ of the matrix V with entries
[V]pq = e−i2πpq(1−)/L , L = 200, and q = 1,. . .,S, for different values of the number of
sources S as a function of the separation between frequencies of (1 − )/L.
the simpler problem of line spectral estimation, Donoho [27] showed that stable super-
resolution is possible via an exhaustive search algorithm, even when condition (7.9) is
#
violated locally, as long as every interval of the ν-axis of length 2SL contains less than
S # /2 frequencies and S # is small (in practice, think of S # 10).
of area L = BT " 1. The ability to handle a potentially large region in which delay-
Doppler pairs can lie in is important in radar applications, since we might need to resolve
objects with large relative distances and relative velocities.
A practical approach to estimating the triplets {(bj ,τj ,νj )} from the received signal y in
the input–output relation (7.3) is to suppose that the time–frequency shifts lie on a fine
grid, and solve the problem on that grid. In general this leads to a gridding error, which,
however, is minimized as the grid grows finer [31]. We next discuss the corresponding
(discrete) sparse signal recovery problem.
Suppose the time–frequency shifts lie on a fine grid with spacing (1/K,1/K), where
2
K is an integer obeying K ≥ L. Let b ∈ CK be the signal with each entry bm,n
corresponding to one of the grid points, with nonzeros equal to the attenuation factors
bj for the time–frequency shifts (τj ,νj ),j = 1,. . .,S. See Figure 7.2 for an illustration.
With this assumption, the input–output relation (7.3) becomes:
K−1
yp = bm,n [Fm/K Tn/K x]p, p = −N,. . .,N .
m,n=0
y = Rb,
Super-Resolution Radar Imaging via Convex Optimization 205
1
K
1
L
ν
1
L
Figure 7.2 Time frequency shifts that lie on a grid: (1/L,1/L) is the “natural” grid, and
(1/K,1/K) is the fine grid. Each dot corresponds to a potential nonzero, and the larger dots
correspond to the actual nonzeros {bj }.
2
where, as before, y is the vector containing as entries the values yp , and R ∈ CL×K ,
is the matrix with (m,n)-th column given by Fm/K Tn/K x. The matrix R contains as
columns “fractional” time–frequency shifts of the sequence x . If K = L, R contains as
columns only “whole” time–frequency shifts of x and R is equal to the Gabor matrix Gx ,
defined by (7.5). In this sense, K = L is the natural grid (see Section 7.2.3) and the ratio
SRF := K/L can be interpreted as a super-resolution factor. The super-resolution factor
determines by how much the (1/K,1/K) grid is finer than the original (1/L,1/L) grid.
A standard approach to the recovery of the sparse signal b from the underdetermined
linear system of equations y = Rb is to solve the following convex program:
4 4
4 4
L1(y) : minimize 4b̃4 subject to y = Rb̃. (7.11)
b̃ 1
The following theorem is the main result from [14] for recovery on the fine grid.
theorem 7.4 Assume that the samples of the probing signal x are i.i.d. N (0,1/L)
random variables, L = 2N + 1, and that L = 2N + 1 ≥ 1024. Consider a signal b
supported on S ⊆ {0,. . .,K − 1}2 , and suppose that it satisfies the minimum separation
condition
1 2.38
min max(|m − m# |,|n − n# |) ≥ .
(m,n),(m#,n# )∈S : (m,n)!=(m#,n# ) K N
Moreover, suppose that the nonzeros of b are i.i.d. uniform on {−1,1} or the complex
unit disk. Choose δ > 0, let y = Rb be the measurement corresponding to b, and
suppose that the number of nonzeros of b is sufficiently smaller than the number of
samples L
L
S≤c 3
,
log (L/δ)
where c is a numerical constant. Then, with probability at least 1 − δ, b is the unique
minimizer of L1(y), y = Rb.
206 Heckel
Note that Theorem 7.4 does not impose any restriction on K, in particular it can be
arbitrarily large. The proof of Theorem 7.4, discussed in Section 7.5.2, is closely linked
to that of Theorem 7.2.
The parameter δ is chosen on the order of the noise variance or the expected grid-
ding error.
The paper [14] considered a synthetic problem with L = 201, where each problem
instance is generated by drawing√S = 10 time–frequency shifts (τj ,νj ) uniformly at
random from the interval [0,2/ 201]2 . This amounts to drawing the corresponding
delay-Doppler pairs ( τ̄j , ν̄j ) from the interval [0,2] × [0,2]. The attenuation factors bj
corresponding to the time–frequency shifts were drawn uniformly at random from the
complex unit disc, independently across j . Measurements were then obtained according
to the input–output relation (7.3).
Figure 7.6 depicts the average resolution error versus the super-resolution factor
. = K/L. The resolution error is defined as the average over j = 1,. . .,S of
SRF
L ( τ̂j − τj )2 + (ν̂j − νj )2 , where the ( τ̂j , ν̂j ) are the estimates of the time–frequency
shifts extracted from a solution of L1-ERR, obtained with the SPGL1 solver [33].
Note that the resolution attained at SRF = 1 corresponds to the resolution attained
by matched filtering and by the compressive sensing radar architecture [4] that was
discussed in Section 7.2.3.
Super-Resolution Radar Imaging via Convex Optimization 207
0.4
2 -resolution error
0.3 Standard/CS Radar
0.2
10 dB SNR
0.1 30 dB SNR
Noiseless
0
1 5 10 15
Super-resolution factor
1
T
1
B
Figure 7.3 Super-resolution radar uniformly provides better resolution error than standard
or CS radar. The plot show the resolution error for the recovery of S = 10 time–frequency shifts
from the observation y with and without additive Gaussian noise n of a certain signal-to-noise
. SNR := y2 /n2 . The resolution error is defined as the average over
ratio
L ( τ̂j − τj )2 + (ν̂j − νj )2 , where (τj ,νj ) are the original time–frequency shifts, and the
( τ̂j , ν̂j ) are the time–frequency shifts on the grid, obtained by solving L1-ERR, for different
super-resolution factors. The different super-resolution factors are illustrated below the plot.
As mentioned before, there are two error sources that were incurred by this approach.
The first is the gridding error obtained by assuming that the points lie on a fine grid
with grid constant (1/K,1/K), which decays in K. The second is the additive noise
error, which is constant. The figure shows that for SRF larger than 1, the resolution is
significantly improved using the super-resolution radar approach. Moreover we see that
for small super-resolution factors SRF, the gridding error dominates, while for large
values of SRF, the additive noise error dominates. In this experiment, the gridding error
approximately decays as 1/SRF. The experiment demonstrates that in practice solving
the super-resolution radar problem on a fine grid is essentially as good as solving it
on the continuum – provided the super-resolution factor is chosen sufficiently large, so
that the gridding error becomes negligible relative to the error due to additive noise.
In this section, we discuss the proofs of theorems 7.2 and 7.4 from [14], which are
closely linked. Specifically, both theorems 7.2 and 7.4 follow from the existence of
certain dual certificates, and the certificate for proving theorem 7.4 is obtained directly
from the certificate constructed to prove theorem 7.2.
208 Heckel
Moreover, suppose that the vectors {Af(rj )}Sj=1 are linearly independent. Then z is the
unique minimizer of AN(y).
have that
4 #4
4z 4 − zA = |bj# | −
i
|bj | > Re Q∗ (r#j )bj# − Re Q∗ (rj )bj
A
j j j j
qH Af(r#j )bj# − Re
ii
= Re qH Af(rj )bj
j j
iii
= 0.
Here, inequality (i) follows from the dual polynomial interpolating the sign pattern,
and from |Q(r#j )| < 1 for at least one r#j , which in turn follows from {r#j } !⊆ {rj } and
|Q(r)| < 1 for all r ∈/ {rj }, by assumption. Inequality (ii) follows from the definition of
the dual polynomial, and inequality (iii) follows from Az# = Az, by assumption. This
contradicts that z# is an optimal solution.
We now turn to the construction of a dual certificate obeying the conditions of Propo-
sition 1, which concludes the proof of Theorem 7.2. The construction of the dual certifi-
cate Q is inspired by the construction of related certificates in [10,11]. First, recall that
the entries of f(r),r = [τ,ν], are given by [f(r)](k,p) = ei2π(kτ+pν ) . From
< =
Q(r) = q,Af(r) = AH q,f(r) ,
Here, F is the squared Fejér kernel, which is a particular trigonometric polynomial with
coefficients gj , i.e.,
N
F (t) = gj ei2πtj .
j =−N
Shifted versions of the polynomial F (i.e., F (t −t0 ) for some t0 ∈ R) and the derivatives
of the polynomial F are also one-dimensional trigonometric polynomials of degree N .
Therefore Ḡ, its partial derivatives, and shifted versions thereof are two-dimensional
trigonometric polynomials of the form q̄,f(r). The construction of Q̄ is concluded by
showing that the coefficients ᾱj , ᾱ 1j , ᾱ 2j , ᾱ 3j , can be chosen such that Q̄ reaches global
maxima at the parameters {rj }.
The construction of Q in the paper [14] for proving Theorems 1 follows a similar
program. Specifically, the polynomial Q is constructed > such that
? it interpolates the signs
sign(bj ) at rj with the functions Gn (r,rj ) = Agn (rj ),Af(r) . Here, gn (r),n = (n1,n2 )
is the vector with (v,k)-th coefficient given by
gk gp (i2πk)n1 (i2πp)n2 e−i2π(τk+ν p),
where the {gj } are the coefficients of the squared Fejér kernel F , just defined. With this
definition, we have
E Gn (r,rj ) = Ḡn (r − rj ).
This follows from
0 1 0 1
E AH A = FH E GH
x Gx F = I,
210 Heckel
where we used E GH x Gx = I, shown at the end of this section. Moreover, Gn (r,rk )
concentrates around Ḡn (r − rk ).
Now, Q is constructed by interpolating the signs sign(bj ) at rj with G(0,0) (r,rj ),
j = 1,. . .,S, and slightly adopting this interpolation near the points {rj } with linear
combinations of the functions G(1,0) (r,rj ) and G(0,1) (r,rj ), in order to ensure that local
maxima of Q are achieved exactly at the rj . Specifically, we set
S
Q(r) = α j G(0,0) (r,rj ) + α1j G(1,0) (r,rj ) + α2j G(0,1) (r,rj ). (7.15)
j =1
N
kp k# p
∗
x Gx ](k,),(k #,# ) =
[GH xp− xp−# e−i2π L ei2π L .
p=−N
#
Noting that E [x ] = 0, we conclude that E [GH x Gx ](k,),(k #,# ) = 0 for != . For
# ∗
= , using the fact that E[xp− xp− ] = 1/L, we arrive at
x Gx ](k,),(k #,# ) ] =
E[[GH e L .
L
p=−N
The
latter is equal to 1 for k = k# and 0 otherwise. This concludes the proof of
E Gx Gx = I.
H
then b is the unique minimizer of L1(y). Here, vS its the vector consisting of the entries
of v indexed by S, and likewise vS c consists of the entries not indexed by S, i.e., the
entries indexed by the complement of S.
Theorem 7.4 now follows directly from the existence of the polynomial Q that was
constructed in the previous section. To see this, define v as [v](m,n) = Q([m/K,n/K])
and note that v satisfies (7.16), since Q([m/K,n/K]) = sign(b(m,n) ) for (m,n) ∈ S and
|Q([m/K,n/K])| < 1 for (m,n) ∈ / S.
θ
object
dR
Figure 7.4 Principle of MISO radar: × and correspond to transmit and receive antennas. The
transmit antenna sends a probing signal, and the reflections of the object are received by three
receive antennas. Estimating the relative delays of the probing signal allows to estimate the angle
of the object relative to the antenna array, in addition to distance and velocity.
212 Heckel
r = 0,j = 0 ×
θ
× Object 1
dT
d
×
Figure 7.5 Principle of MIMO radar: × and correspond to transmit and receive antennas.
Throughout, we assume the spacing of the NT transmit and NR receive antennas to be dT = 2fc
c
and dR = cN T
2fc , where fc is the carrier frequency.
receive antennas can give a much larger angular resolution from far fewer antennas.
Specifically, by arranging NT transmit and NR receive antennas in a particular way
(see Figure 7.5), a MIMO radar can obtain the same resolution obtained by a MISO
(or SIMO) radar with NT NR uniformly spaced receive (or transmit) antennas. This is
often called a MIMO virtual array. In this section, we discuss a MIMO radar model,
and show that the fundamental limit for resolving the angle-delay-Doppler triplets is
(1/(NT NR ),1/B,1/T ). We furthermore show that this limit can be overcome in the
sense that triplets can be resolved on a much finer grid, provided they are sufficiently
separated.
receive antennas [34,35]. As explained in Section 7.6.2, the (baseband) signal yr (t) at
continuous time t received by antenna r = 0,. . .,NR − 1, consists of the superposition
of the reflections from the objects of the transmitted probing signals xj (t),j = 0,. . .,
NT − 1, and is given by
S T −1
N
yr (t) = bk ei2πrNT βk ei2πj βk xj (t − τ̄k )ei2πν̄k t . (7.17)
k=1 j =0
S T −1
N
i2πrNT β k
[yr ]p = bk e ei2πj βk [Fνk Tτk xj ]p . (7.18)
k=1 j =0
Here, yr contains the samples of the received signal yr taken at rate 1/B (i.e.,
[yr ]p = yr (p/B) in the interval p/B ∈ [−T /2,T /2]), and xj is the vector containing
the samples of the probing signal ([xj ]p := xj (p/B)).
We have reduced the problem of identifying the locations of the objects under the con-
straints that the probing signals xj are band-limited and the responses yr are observed
over a finite time interval only, to the estimation of the parameters bk ∈ C, (βk ,τk ,νk ) ∈
[0,1]3,k = 1,. . .,S, from the samples [yr ]p,r = 0,. . .,NR − 1,p = −N,. . .,N , in the
input–output relation (7.18). We call this the super-resolution MIMO radar problem.
2d sin(θ)(dT j + dR r) 2(dT j + dR r)
τ̃ := + = τ̄ − β .
c c c
For the second equality, we defined the angle parameter β := − sin(θ)/2 and the delay
τ̄ := 2d
c . Taking the Doppler shift into account, the reflection of the j -th probing signal
received by the r-th receive antenna is given by
Here, b̃ ∈ C is the attenuation factor associated with the object, and ν̄ := 2v c fc is the
Doppler shift, which is a function of the relative velocity, v, of the object. By choosing
the antenna spacing as dT = 2fc c and dR = cN T
2fc , the reflection of the j -th probing signal
received by the r-th receive antenna in (7.19) becomes
j +rNT
i2π(fc +ν̄ )β
b̃xj (t − τ̃)ei2π(fc +ν̄ )(t−τ̄) e fc ≈ b̃xj (t − τ̄)ei2π(fc +ν̄ )(t−τ̄) ei2πβ(j +rNT ) .
Here, the approximation follows by the Doppler shift ν̄ being much smaller than the
carrier frequency fc , therefore fcf+c ν̄ ≈ 1, and τ̃ ≈ τ̄. If follows that the reflection of the
j -th probing signal received by the r-th receive antenna, after demodulation, is
where we defined b = b̃e−i2πν̄ τ̄ . Next, consider S objects with parameters (bk ,βk ,
τ̄k , ν̄k ). Since, for S objects, the (demodulated) signal yr received by antenna r consists
of the superposition of the reflections of the probing signals xj ,j = 0,. . .,NT − 1,
transmitted by the transmit antennas, we obtain the input-output relation in (7.17) simply
by summing over the reflections given by bk xj (t − τ̄k )ei2πν̄k t ei2πβk (j +rNT ) .
has entries
Similarly as before, we use for convenience a three dimensional index to refer to entries
2
of a vector. Finally, the matrix A ∈ CNR L×NR NT L is defined as follows. The expression
T −1
N
i2πrNT β
wr,p := e ei2πj β [Fν Tτ xj ]p,
j =0
Super-Resolution Radar Imaging via Convex Optimization 215
with
1
N
k
ap,k,j = [xj ] ei2π(−p) L .
L
=−N
Let fp,j ∈ CL be the vector with kth entry [fp,j ]k = ap,k,j , k = −N,. . .,N ,
2
and let Aj ∈ CL×L be the block-diagonal matrix with fTp,j on its pth diagonal,
p = −N,. . .,N . With this notation, A is defined as the block-diagonal matrix with the
2
matrix [A0,. . .,ANT −1 ] ∈ CL×NT L on its diagonal, for all NR blocks on the diagonal.
With this notation, (7.3) becomes (7.20).
Similarly as for the SISO radar problem, recovery of the unknowns bk ,rk = [βk ,τk ,νk ]
from the measurement z = Sk=1 bk f(rk ) is a 3D line spectral estimation problem that
can be solved with standard spectral estimation techniques. In order to recover the
vector z from the measurement y, we use that z is a sparse linear combination of atoms
in the set A := {f(r),r ∈ [0,1]3 }, and estimate z by solving the basis pursuit type atomic
norm minimization problem
4 4
AN(y) : minimize 4z̃4A subject to y = Az̃,
z̃
where
zA := inf |bk | : z = bk f(rk ) .
bk ∈C,rk ∈[0,1]3
k k
To summarize, as for the SISO radar problem, we estimate the parameters bk ,rk from y
by:
i. solving AN(y) in order to obtain z,
ii. estimating the rk from z by solving the corresponding 3D-line spectral estimation
problem, and
iii. solving the linear system of equations y = S−1k=0 bk Af(rk ) for the bk .
10 5 5
|β j − βj # | ≥ or |τj − τj # | ≥ or |νj − νj # | ≥ . (7.21)
NT NR − 1 N N
As before, |τj − τj # | is the wraparound distance on the unit circle.
Note that the triplets must not be separated in angle, time, and frequency simulta-
neously; for the MIMO minimum separation condition to be satisfied, it is sufficient if
they are separated in at least one of those dimensions.
theorem 7.6 Assume that the samples of the probing signals xj ,j = 0,. . .,NT − 1,
are i.i.d. zero-mean Gaussian random variables with variance 1/(NT L), and let
L = 2N + 1 ≥ 1024 and NT NR ≥ 1024. Consider a signal where the signs of
the attenuation factors {bj }Sj=1 are i.i.d. uniform on {−1,1} or the complex unit disc,
and suppose that the triplets {(βj ,τj ,νj )}Sj=1 obey the MIMO minimum separation
condition. Furthermore, choose δ > 0 and assume that the number of nonzero
attenuation factors, S, obeys
min(L,NT NR )
S≤c , (7.22)
log3 (L/δ)
S
where c is a numerical constant. Then, with probability at least 1 − δ, z = k=1 bk f(rk )
is the unique minimizer of AN(y), y = Az.
Theorem 7.2 guarantees that, with high probability, the attenuation factors and loca-
tion parameters can be recovered perfectly from the observation y by solving a convex
program (recall that the parameters bk ,rk can be obtained from z), provided that the loca-
tions rk = [β k ,τk ,νk ] are sufficiently separated in either angle, time, or frequency, and
provided that the total number of objects satisfies condition (7.22). Note that, translated
to the physical parameters τ̄k , ν̄k , the MIMO minimum separation condition becomes:
For all k,k # : k != k # ,
10 10.01 10.01
|βk − βk # | ≥ or | τ̄k − τ̄k # | ≥ or |ν̄k − ν̄k # | ≥ .
NT NR − 1 B T
Theorem 7.2 is essentially optimal in the number of objects that can be located, since
S can be linear – up to a log-factor – in min(L,NT NR ), and S ≤ min(L,NT NR ) is a
necessary condition to uniquely recover the attenuation factors bk even if the locations
rk are known. To see this, note that for the linear system of equations (7.20) to have a
unique solution, the vectors Af(rk ) must be linearly independent. If βk = 0, for all k, or
if τk = 0 and νk = 0, for all k, the vectors Af(rk ),rk = (βk ,τk ,νk ),k = 0,. . .,S − 1 can
only be linearly independent provided that S ≤ L and S ≤ NT NR , respectively. This is
seen from
⎡ T −1 i2πj β ⎤
ei2π0β N j =0 e F ν T τ xj
⎢ ⎥
Af(r) = ⎢
⎣
..
.
⎥.
⎦
T −1 i2πj β
ei2πNT (NR −1)β N j =0 e F T x
ν τ j
We finally note that Theorem 7.2 is proven by constructing a dual certificate in a similar
manner to our certificate construction for the SISO in Section 7.7.
Super-Resolution Radar Imaging via Convex Optimization 217
y = Rb,
where R ∈ CNR L×K1 K2 K3 is the matrix with (n1,n2,n3 )-th column given by
Note that the nonzeros of b and its indices correspond to the attenuation factors bk and
the locations rk on the grid. A standard approach to the recovery of the sparse signal b
from the underdetermined linear system of equations y = Rb is to solve the following
convex program:
4 4
4 4
L1(y) : minimize 4b̃4 subject to y = Rb̃. (7.23)
b̃ 1
What follows is the main result for recovery on the fine grid.
Note that Theorem 7.7 does not impose any restriction on K1,K2,K3 , in particular
they can be arbitrarily large. The proof of Theorem 7.7 is closely linked to that of
Theorem 7.6; specifically, similarly to the SISO case, the existence of a certain dual
certificate guarantees that b is the unique minimizer of L1(y). The dual certificate is
obtained directly from the dual certificate for the continuous case, which, as mentioned
218 Heckel
before, is constructed for the MIMO case in a similar way as the certificate for the SISO
case has been constructed in Section 7.7.
with δ chosen on the order of the noise variance. There are two error sources incurred
by this approach: the gridding error obtained by assuming the points lie on a grid
with spacing (1/K1,1/K2,1/K3 ), which decreases in SRF and becomes negligible, and
the additive noise error, which is constant. The results of the simulations, depicted in
Figure 7.6, show that the object resolution of the super-resolution approach is signifi-
cantly better than that of the compressed sensing-based approach [7,8] corresponding to
recovery on the coarse grid, i.e., SRF = 1. Moreover, the results show that the approach
SNR = 5dB
0.6 SNR = 10dB
Resolution error
SNR = 20dB
0.4 Noiseless
0.2
1 2 3 4 5 6
SRF
Figure 7.6 Resolution error for the recovery of S = 5 objects from the samples y with and
without additive Gaussian noise n of a certain signal-to-noise ratio SNR = y22 /n22 , for
varying super-resolution factors (SRFs). The resolution error is defined as the average over
(NT2 NR2 (β̂ k − βk )2 + L2 ( τ̂k − τk )2 + L2 (ν̂k − νk )2 )1/2 , k = 1,. . .,S, where (β̂k , τ̂k , ν̂k ) are the
locations obtained by solving L1-ERR.
Super-Resolution Radar Imaging via Convex Optimization 219
0.15 IAA
Resolution error L1-ERR
0.1
0.05
0
30 20 10
SNR in dB
Figure 7.7 Resolution error (smaller is better) of L1-ERR and IAA applied to y + n,
where n ∈ CNR L is additive Gaussian noise, such that the signal-to-noise ratio is
SNR := y22 /n22 . As before, the resolution error is defined as
(NT2 NR2 (β̂ k − βk )2 + L2 ( τ̂k − τk )2 + L2 (ν̂k − νk )2 )1/2 , where (β̂ k , τ̂k , ν̂k ) are the
locations obtained by solving L1-ERR.
is robust to noise and that even under noise, the localization accuracy is significantly
improved over a standard approach to radar.
We next compare our approach to the iterative adaptive approach (IAA) [36], pro-
posed for MIMO radar in [37]. IAA is based on weighted least squares and has been
proposed in the array processing literature. IAA can work well even with only one
snapshot only and can therefore be directly applied to the MIMO super-resolution prob-
lem. However, to the best of our knowledge, no analytical performance guarantees are
available in the literature that attest IAA similar performance than the 1 -minimization-
based approach. We compare the IAA algorithm [36, table II, entitled “The IAA-APES
Algorithm”] to L1-ERR, for a problem with parameters NT = 3,NR = 3, and L = 41,
as before, but with SRF = 3 and (βk ,τk ,νk ) = (k/(NR Nt ),k/L,k/L),k = 1,. . .,S, so
that the location parameters lie on the fine grid and are separated. As before, we draw
the corresponding attenuation factors bk i.i.d. uniformly at random from the complex
unit disc. Our results, depicted in Figure 7.7, show that L1-ERR performs better in this
experiment than IAA, in particular for small signal-to-noise ratios.
In this section we discuss a class of signal recovery problems that are closely related
to the SISO and MIMO radar problem, corresponding results, and open theoretical
research problems, and comment on computational challenges in applying the methods
discussed here in practical radar systems.
The SISO and MIMO radar problems discussed in this chapter are versions of a
more general problem, namely that of recovering a signal that is sparse in a continu-
ously indexed dictionary, with the index corresponding to the locations and velocities of
objects. In contrast, traditional compressive sensing research has focused on the recov-
ery of signals that are sparse in discretely indexed dictionaries via convex programs [3]
220 Heckel
amongst other methods. As discussed in this chapter in the context of radar, signals
that are sparse in continuously indexed dictionaries can be recovered via a convex
program either by solving an atomic norm minimization problem, or by discretizing
the continuous parameter space. However, the discretization step induces a gridding
error. While in practice – provided the grid is chosen sufficiently fine – the gridding
error is negligible, fine discretization leads to dictionaries with extremely correlated,
i.e., coherent, columns, and the theory of compressive sensing, and many practical
algorithms, rely on the dictionary to be incoherent and therefore does not apply to
fine grids. The primary difficulty with recovering signals in such dictionaries is that
the elements of the dictionaries are very close to each other – which is the case both
for continuously indexed dictionaries as well as for finely discretized signals. Stable
recovery of signals that are sparse in such dictionaries requires excluding signals that
are supported on elements of the dictionary that are very close to each other. Such signals
are excluded here by imposing the minimum separation condition.
More specifically, the SISO and MIMO radar problems belong to a class of signal
recovery problems where the goal is to recover unknown coefficients {bj } and location
parameters {rj } from the measurement
S
y= bj Af(rj ).
j =1
provably recovers the coefficients {bj } and location parameters {rj } perfectly, provided
that the minimum separation condition holds (see Section 7.3.2). While standard spec-
tral estimation techniques such as Prony’s method, MUSIC, and ESPRIT [9] also prov-
ably succeed for the noiseless case, even without requiring a separation condition, an
advantage of the convex program is that it does not require knowledge of S, and perhaps
more importantly is provably robust [12,40,42].
Specifically, Tang et al. [42] show that the atomic norm regularized least squares
estimate enables near-optimal denoising of z from a noisy measurement y = z + e,
S
where z is a signal of the form z = j =1 f(νj ), and e is zero-mean Gaussian noise
with variance σ I. Specifically, provided the minimum separation condition holds, the
2
with high probability. This result is essentially optimal; to see this, note that even if we
knew the location parameters {νj } exactly, the best bound we could achieve would be
σ2 S [43], only by a logarithmic factor short of the result. In addition, the paper [42]
shows that the corresponding estimator localizes the frequencies up to a certain (small)
error, provided that the number of samples is sufficiently large.
Next, suppose that A is a M × L Gaussian random matrix with i.i.d.
N (0,1/M) + iN (0,1/M) entries, with M typically much smaller than L. Assume we
are given a noisy measurement y = Az + e, with z a signal of the form z = Sj=1 f(νj )
and e is noise. To estimate the signal z from such a measurement one can use an atomic
norm optimization problem of the form
1
ẑ := arg min y − Az̄22 subject to z̄A ≤ τ, (7.24)
z̄ 2
with τ a tuning parameter. Theorem 4 in the paper [16] shows that, as long as
M ≥ cS log(L),
with c a fixed numerical constant, the minimizer of (7.24) with τ = zA obeys
4 4
4ẑ − z42 ≤ cS log(L)σ2,
2
with high probability. Again, this results is essentially optimal both with respect to the
number of measurements M required relative to the number of unknowns S, as well as
with respect to the best bound we can achieve for estimation under noise. Moreover,
there results do not make any assumptions on the coefficients bj . Unfortunately, the
corresponding proof strategy does not carry over to the case where A is a structured
random matrix, as in the SISO and MIMO radar problems considered here. However,
numerical simulations suggest that for a number of structured random matrices, includ-
ing the matrices parametrizing the SISO and MIMO radar problems, the performance of
the nuclear norm minimization program, as well as 1 -norm regularized least squares,
is similar.
222 Heckel
References
[1] D. W. Bliss and K. W. Forsythe, “Multiple-input multiple-output (MIMO) radar and
imaging,” in Asilomar Conf. on Signals, Syst. and Comput., 2003, pp. 54–59.
[2] J. Li and P. Stoica, “MIMO radar with colocated antennas,” IEEE Signal Process. Mag.,
vol. 24, no. 5, pp. 106–114, 2007.
[3] E. J. Candès, 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. 489–509, 2006.
[4] M. Herman and T. Strohmer, “High-resolution radar via compressed sensing,” IEEE Trans.
Signal Process., vol. 57, no. 6, pp. 2275–2284, 2009.
[5] R. Baraniuk and P. Steeghs, “Compressive radar imaging,” in IEEE Radar Conf., 2007,
pp. 128–133.
[6] R. Heckel and H. Bölcskei, “Identification of sparse linear operators,” IEEE Trans. Inf.
Theory, vol. 59, no. 12, pp. 7985–8000, 2013.
[7] D. Dorsch and H. Rauhut, “Refined analysis of sparse MIMO radar,” J. Fourier Anal. Appl.,
vol. 23, 2017.
[8] T. Strohmer and H. Wang, “Adventures in compressive sensing based MIMO radar,” in
Excursions in Harm. Anal., ser. Appl. Num. Harm. Anal., 2015, pp. 285–326.
[9] P. Stoica and R. L. Moses, Spectral Analysis of Signals. Pearson Prentice Hall, 2005.
Super-Resolution Radar Imaging via Convex Optimization 223
[31] G. Tang, B. N. Bhaskar, and B. Recht, “Sparse recovery over continuous dictionaries-just
discretize,” in Asilomar Conf. on Signals, Syst. and Comput., Pacific Grove, CA, 2013,
pp. 1043–1047.
[32] S. R. Becker, E. J. Candès, and M. C. Grant, “Templates for convex cone problems with
applications to sparse signal recovery,” Math. Prog. Comp., vol. 3, no. 3, pp. 165–218, 2011.
[33] E. van den Berg and M. P. Friedlander, “Probing the pareto frontier for basis pursuit
solutions,” SIAM J. Sci. Comput., vol. 31, no. 2, pp. 890–912, 2008.
[34] B. Friedlander, “On the relationship between MIMO and SIMO radars,” IEEE Trans. Signal
Process., vol. 57, no. 1, pp. 394–398, Jan. 2009.
[35] T. Strohmer and B. Friedlander, “Analysis of sparse MIMO radar,” Appl. Comput. Harm.
Anal., vol. 37, no. 3, pp. 361–388, 2014.
[36] T. Yardibi, J. Li, P. Stoica, M. Xue, and A. B. Baggeroer, “Source localization and sensing:
A nonparametric iterative adaptive approach based on weighted least squares,” IEEE Trans.
Aerosp. Electron. Syst., vol. 46, no. 1, pp. 425–443, 2010.
[37] W. Roberts, P. Stoica, J. Li, T. Yardibi, and F. Sadjadi, “Iterative adaptive approaches to
MIMO radar imaging,” IEEE J. Sel. Topics Signal Process, vol. 4, no. 1, pp. 5–20, 2010.
[38] B. Bernstein and C. Fernandez-Granda, “Deconvolution of point sources: A sampling
theorem and robustness guarantees,” Commun. Pur. Appl. Math., to appear, 2017.
[39] N. Antipa, G. Kuo, R. Heckel et al., “Diffusercam: Lensless single-exposure 3D imaging,”
Optica, vol. 5, no. 1, p. 19, 2018.
[40] E. J. Candès and C. Fernandez-Granda, “Super-resolution from noisy data,” J. Fourier Anal.
Appl., vol. 19, no. 6, pp. 1229–1254, 2014.
[41] C. Fernandez-Granda, “Super-resolution of point sources via convex programming,” Infor-
mation and Inference, vol. 5, pp. 251–303, 2016.
[42] G. Tang, B. N. Bhaskar, and B. Recht, “Near minimax line spectral estimation,” IEEE Trans.
Inf. Theory, vol. 61, no. 1, pp. 499–512, 2015.
[43] F. Bunea, A. Tsybakov, and M. Wegkamp, “Sparsity oracle inequalities for the Lasso,”
Electron. J. Stat., vol. 1, pp. 169–194, 2007.
[44] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear
inverse problems,” SIAM Journal on Imaging Sciences, vol. 2, no. 1, p. 183202, 2009.
[45] N. Boyd, G. Schiebinger, and B. Recht, “The alternating descent conditional gradient
method for sparse inverse problems,” SIAM J. Optimiz., vol. 27, no. 2, pp. 616–639, 2017.
[46] N. Rao, P. Shah, and S. Wright, “Forward-backward greedy algorithms for atomic norm
regularization,” IEEE Trans. Signal Process., vol. 63, no. 21, pp. 5798–5811, 2015.
8 Adaptive Beamforming via
Sparsity-Based Reconstruction
of Covariance Matrix
Yujie Gu, Nathan A. Goodman, and Yimin D. Zhang
8.1 Introduction
Adaptive beamforming is an effective spatial filtering technique that adjusts the beam-
forming weight vector to increase the strength of the signal of interest while suppress-
ing interference and noise. As a ubiquitous task in array signal processing, adaptive
beamforming has been widely used in radar, sonar, wireless communications, radio
astronomy, seismology, speech processing, medical imaging, and many other areas (see,
for example, [1–5] and the references therein). Unlike conventional data-independent
beamformers (e.g., fixed or switched beamformers), adaptive beamformers depend on
the array received data and hence are expected to provide better capabilities for inter-
ference suppression and signal enhancement. Nevertheless, it is also well known that
adaptive beamformers are extremely sensitive to model mismatch, especially when the
training samples used for the calculation of the beamforming weight are contaminated
by the desired signal. In practice, such model mismatch commonly occurs. For example,
the data covariance matrix cannot be accurately estimated due to the limited number of
training samples, and the steering vector of the desired signal may also be imprecise
or even unknown due to look direction error, imperfect calibration, and other effects.
Whenever model mismatches exist, classical adaptive beamformers (e.g., Capon beam-
former [6]) will suffer severe performance degradation. To this end, adaptive beam-
former design with robustness against model mismatch has been an intensive research
225
226 Gu, Goodman, and Zhang
topic in the past decades, and various robust adaptive beamforming techniques have
been proposed (see, for example, [4,7,8] and the references therein). Based on the
principle of adaptive beamforming, these robust adaptive beamformers can be classified
into two major categories.
In the first category, robust adaptive beamforming techniques process the sample
covariance matrix, because the exact interference-plus-noise covariance matrix is usu-
ally unavailable in practical applications. The sample covariance matrix is a maximum
likelihood estimate of the data covariance matrix, and thus leads to the optimal output of
the resulting adaptive beamformer when the sample size tends to infinity. Unfortunately,
the sample size is often limited in practice, thus resulting in significant performance
degradation, especially when the desired signal is present in the training samples [9,10].
The most popular robust beamforming technique in this category is the diagonal loading
technique [9,11–13], which adds a scaled identity matrix to the sample covariance
matrix to reduce the conditional number. A major problem with diagonal loading is that
there is no clear rule to choose the optimal diagonal loading factor in different scenarios.
In order to choose the diagonal loading factor adaptively, rather than in an ad hoc way,
several user parameter–free adaptive beamforming algorithms were proposed (see, for
example, [14] and the references therein). The shrinkage estimation approach [15] in the
sense of minimizing mean squared error (MSE) can automatically compute the diagonal
loading levels without the need to specify any user parameters. However, this approach
leads to an estimate of the statistical covariance matrix of the array received data rather
than the required interference-plus-noise covariance matrix. In such a case, the perfor-
mance degradation becomes severe with the increase of the desired signal power, even
when the desired signal steering vector is exactly known. The eigenspace decomposition
technique [16,17] is another popular approach for robust adaptive beamforming that is
applicable to an arbitrary steering vector mismatch case. The key idea of this technique
is to use the projection of the presumed steering vector onto the sample signal-plus-
interference subspace. This approach requires the knowledge of the dimension of the
signal-plus-interference subspace. It is known that this approach suffers severe perfor-
mance degradation from the subspace swap1 when the signal-to-noise ratio (SNR) is
low [18,19]. It also suffers from the signal self-nulling problem, especially at high SNR
levels. A sparsity-based iterative adaptive approach (IAA) [20] can iteratively update
the spatial power estimates in the whole observation field and subsequently update
the covariance matrix used for adaptive beamformer design. Although it does improve
the power estimate, the IAA beamforming algorithm is not robust against direction-of-
arrival (DOA) mismatch because its weight is simply that of the scanning grid point
corresponding to the assumed DOA of the desired signal.
In the second category, robust adaptive beamforming techniques process the pre-
sumed desired signal steering vector because the exact knowledge of the steering vector
is not easy to obtain in practice. In practical situations, steering vector mismatch can
1 A subspace swap occurs when the measured data is better approximated by some components of the noise
subspace than by some components of the signal subspace, i.e., there is a switch of vectors between the
estimated signal and noise subspaces.
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 227
easily occur due to look-direction errors [21,22] or imperfect array calibration and dis-
torted antenna shape [23]. Besides these, other common causes leading to steering vector
mismatches include array manifold mismodeling because of source wavefront distor-
tions that result from environmental inhomogeneities [24,25], near-far problem [26],
source spreading and local scattering [27–30], as well as other effects [10]. In this
category, the linear constrained minimum variance (LCMV) beamformer [31] is most
commonly used. It provides robustness against uncertainty in the signal look direction
by broadening the main lobe of the beampattern. However, the additional imposed
constraints reduce the degrees of freedom (DOFs) of the resulting adaptive beamformer.
More importantly, the LCMV beamformer becomes less robust when any other types
of steering vector mismatch beyond the look-direction errors become dominant. To
improve the robustness of adaptive beamformers against arbitrary unknown steering
vector mismatches, the worst-case performance optimization-based technique [32–34]
makes explicit use of an uncertainty set of the signal steering vector. This method
requires that the upper bound of the norm of the mismatch vector is a priori unknown.
Moreover, the worst operating conditions may not always occur. Hence, this adaptive
beamforming technique is also an ad hoc approach and will suffer from performance
degradation whenever the upper bound of the norm of the mismatch vector is either
overestimated or underestimated. Another representative technique in this category is to
estimate the desired signal steering vector by maximizing the beamformer output, under
the constraint that the convergence of the steering vector estimate to any interference
steering vector or their combinations is prohibited [35,36]. However, the imposed norm
constraint on the steering vector is too strict to be satisfied, particularly when there exist
local scattering encountered in, e.g., mobile communications and indoor speech signal
processing. In such cases, gain perturbations in different sensors cannot be ignored, and
then the norm constraint no longer holds.
As mentioned previously, these two categories of adaptive beamforming techniques
were developed almost independently in the past four decades. Obviously, these adaptive
beamforming techniques are not optimal, because they respectively assume that either
the desired signal steering vector or the interference-plus-noise covariance matrix is
exactly known. Since the pioneering work of Vorobyov et al. [37], adaptive beamform-
ing has been required to be jointly robust against covariance matrix uncertainty and
steering vector mismatch [38–42]. It is worth noting that, in [40], the interference-plus-
noise covariance matrix is reconstructed by integrating the outer products of interfer-
ence steering vectors weighted by the Capon spatial spectrum over a region separated
from the desired signal direction, thus removing the desired signal component from the
covariance matrix used for adaptive beamformer design. The reconstructed interference-
plus-noise covariance matrix is then used to correct the presumed signal steering vector
in order to maximize the beamformer output power under the only constraint that the
corrected steering vector does not converge to any interference steering vector or their
combinations. Based on the reconstructed interference-plus-noise covariance matrix
and the estimated desired signal steering vector, the resulting adaptive beamformer
provides a near-optimal output performance with a fast convergence rate. However,
the computational complexity of covariance matrix reconstruction is high due to the
228 Gu, Goodman, and Zhang
integral operation. In addition, there is a certain level of performance loss when the
number of training samples is small, because the source power obtained from the Capon
spatial spectrum is underestimated and, as a result, the estimated interference-plus-noise
covariance matrix is inaccurate.
In this chapter, we will elaborate adaptive beamforming via sparsity-based recon-
struction of the interference-plus-noise covariance matrix [43]. By exploiting the spar-
sity of sources in the observed spatial domain, the interference covariance matrix is
reconstructed as a linear combination of the outer products of the interference steering
vectors weighted by their individual power, which can be estimated from a sparsity-
constrained covariance matrix fitting problem. As such, the proposed technique provides
a signal-free interference-plus-noise covariance matrix to enable robust adaptive beam-
former design that avoids the signal self-nulling problem. It requires low computational
complexity, as there is no matrix inversion or eigen-decomposition involved in the
sparsity-constrained covariance matrix fitting problem. Hence, the proposed adaptive
beamforming technique is suitable for an arbitrary number of training samples [44].
When the number of training samples is larger than the number of array sensors, the
formulated sparsity-constrained covariance matrix fitting problem can be effectively
solved by using the known array structure, i.e., estimate the directions of sources and
their power in turn. The proposed adaptive beamformer is compared to existing state-
of-the-art adaptive beamformers in terms of computational complexity, output signal-
to-interference-plus-noise ratio (SINR) performance, and convergence rate. Numerical
simulations clearly demonstrate the near-optimal output performance and faster conver-
gence rate of the proposed adaptive beamforming algorithm exploiting the sparsity of
sources in the spatial domain.
In this section, we first build the narrowband array signal model, then briefly review
adaptive beamforming criteria and classical adaptive beamformers.
as well as source direction, e.g., a s a(θs ), where θs is the direction of the desired
signal impinging on the array. For example, the ideal steering vector of a uniform linear
array (ULA) has the form of
0 2π 2π
1T
a(θ) = 1,e−j λ d sin θ,. . .,e−j λ (M−1)d sin θ , (8.3)
√
where θ is the DOA of the source, j = −1 is the imaginary unit, λ is the wavelength
of the narrowband signal, and d = λ/2 is the inter-element spacing of the array.
Similarly, the steering vector of the interference has the similar form with a different
source direction. In contrast, there is no such form for the additive noise because noise
does not have a fixed direction.
In practical array applications including radar, however, the data covariance matrix
cannot be accurately estimated due to the limited training samples, and the signal steer-
ing vector may not be precisely known because of the imperfect knowledge of the
source location, propagation environment and/or array calibration. In such cases, the
MPDR beamformer suffers severe performance degradation, which becomes obvious
with the increase of input signal power. Hence, the MPDR beamformer underperforms
the MVDR beamformer in practical applications.
1
K
R̂ = x(k)x H (k), (8.12)
K
k=1
where K is the number of snapshots (i.e., training samples). The resulting adaptive
beamformer,
−1
R̂ ā s
wSMI = −1
, (8.13)
ā H
s R̂ ā s
is called the sample matrix inversion (SMI) beamformer [48], where ā s = a(θs ) is
the presumed signal steering vector. Whenever there exists a desired signal in the array
received signal x(k), the SMI beamformer is in essence an MPDR beamformer (8.10)
rather than an MVDR beamformer (8.8). As K → ∞, R̂ will converge to R, and
the corresponding output SINR will approach the optimal value under stationary and
ergodic assumptions. However, when K is small, the large gap between R̂ and R is
known to dramatically affect the output performance of the SMI beamformer, especially
when there is a desired signal in the training samples [9,10].
In order to reduce the sensitivity of the SMI beamformer to model mismatches,
many different beamforming algorithms have been developed in the past decades and
successfully applied in a wide range of areas (see, for example, [3,4,7,14] and the
references therein). In the following, several classical adaptive beamforming algorithms
are briefly reviewed.
The performance of the DL-SMI beamformer depends on the diagonal loading factor
ξ. It is usually chosen in an ad hoc way, typically about ten times the noise power, i.e.,
ξ = 10σn2 , where the noise power σn2 is assumed to be known [9]. Obviously it is not
optimal, nor is the instantaneous noise power easy to know. In order to adaptively choose
the loading factor, several user parameter–free approaches have been proposed for adap-
tive beamforming [14]. However, this method leads to an estimate of the data covariance
matrix rather than that of the interference-plus-noise covariance matrix. Regardless of
the value of the chosen diagonal loading factor, the performance loss of the DL-SMI
beamformer is inevitable, and this degradation becomes more severe with the increase
of the desired signal power [49]. The main reason is that the desired signal component is
always active in any kind of diagonal loading beamformer and its effect becomes more
pronounced with the increase of input SNR [40].
Worst-Case Beamforming
The worst-case performance optimization-based adaptive beamforming [32] guarantees
a distortionless response for all possible steering vectors in a predetermined set. The
worst-case adaptive beamforming problem can be formulated as
where es = a s − ā s denotes the mismatch vector between the actual signal steer-
ing vector a s and the presumed signal steering vector ā s , and ε is the upper bound
of the norm of the mismatch vector es . Here, · 2 denotes the 2 -norm, also called
the Euclidean norm. Because the constraint condition is nonlinear and nonconvex, the
worst-case adaptive beamforming problem (8.16) is a semi-infinite nonconvex quadratic
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 233
program, and is NP-hard2 . By using the special structure of the objective function and
the constraints, the nonconvex optimization problem can be reformulated as a second-
order cone programming (SOCP) problem
min wH R̂w s.t. wH ā s ≥ ε w2 + 1,
w
$ %
Im wH ā s = 0, (8.17)
which is convex and can be efficiently solved in polynomial time using the well-
established interior point methods. Here, Im ( · ) denotes the imaginary part of a
complex number.
The worst-case beamformer is robust to arbitrary unknown signal steering vector
mismatch with an upper-bounded norm. However, in practical applications, neither the
mismatch vector nor its upper bound is a priori known. Either overestimation or under-
estimation of the upper bound of the norm of the steering vector mismatch will degrade
the performance of the worst-case beamformer. In addition, the worst-case beamformer
also suffers the signal self-nulling problem because it uses the sample covariance matrix
R̂ rather than the interference-plus-noise covariance matrix R i+n .
source locations in the observed field (or the number of scanning points), which is
usually much larger than the true number of sources. Then, the IAA algorithm repeats
the following iterative process
R̄ = A(θ)diag(p̂)AH (θ)
for l = 1,. . .,L
−1
R̄ a(θl )
wl = −1
a H (θl )R̄ a(θl )
p̂l = wH
l R̄wl
end for (8.18)
to converge, where A(θ) = [a(θ1 ),a(θ2 ),. . .,a(θL )] ∈ CM×L is the array steering
matrix, p̂ = [p̂1, p̂2,. . ., p̂L ]T ∈ RL
+ is the estimated spatial spectrum. Here, R+ denotes
L
2 In optimization theory, NP-hard problems represent a class of extremely difficult problems that cannot be
solved in polynomial time.
234 Gu, Goodman, and Zhang
As such, the IAA algorithm can achieve the signal waveform (and hence signal power)
estimation by way of sparse signal representation. It performs well when there is no
model mismatch. However, when there is a slight model mismatch on the signal steering
vector, e.g., signal-look direction mismatch, performance degradation would occur, and
the degradation becomes severe with the increase of the input SNR.
Q
R i+n = σi2q a(θiq )a H (θiq ) + σn2 I, (8.19)
q=1
where Q is the number of interferers, a(θiq ) is the steering vector of the q-th interfer-
ence impinging from the DOA θiq , and σi2q is the corresponding interference power.
Hence, in order to have an accurate estimate of the interference-plus-noise covariance
matrix R i+n , we need to know the steering vectors of all interferers via DOAs and their
individual power, together with the noise power. When these pieces of information are
unavailable, the interference-plus-noise covariance matrix can be reconstructed as [40]
R̂ i+n = pCapon (θ)a(θ)a H (θ)d θ, (8.20)
¯
1
pCapon (θ) = −1
(8.21)
a H (θ)R̂ a(θ)
is the Capon spatial spectrum estimator, and ¯ is the complement sector of . Here,
is a known or estimated angular sector in which the desired signal is located. Hence, the
covariance matrix estimator R̂ i+n collects all interference and noise in the out-of-sector
¯ which effectively excludes the desired signal component.
,
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 235
can dramatically improve the performance regardless of the desired signal power
(see [40] and accompanying simulations). Nevertheless, the estimation accuracy of
R̂ i+n in (8.20) is poor because the Capon estimator (8.21) underestimates the true
power, especially when the number of snapshots is limited. On the other hand, the
computational complexity is high because the covariance matrix reconstruction process
introduces the unnecessary integral operation, where the number of interferers is
actually countable.
p ≥ 0,
σn2 > 0, (8.23)
where p ∈ RL + is the spatial spectrum distribution on the sample grids of the observed
spatial domain (e.g., {θ1,θ2,. . .,θ } ¯ ∈ RL×L
L ∈ ∪ ), P = diag(p) + is the corre-
sponding diagonal matrix, A = a(θ1 ),a(θ2 ),. . .,a(θL ) ∈ C M×L is the array manifold
matrix, and · F and · 0 , respectively, denote the Frobenius norm of a matrix
and the 0 “norm” of a vector. Note that, although it does not satisfy the positive
homogeneity, the 0 “norm”, which counts the number of nonzero elements in a vector,
is an ideal measure of sparsity. According to the sparse observation, the number of
potential sources is much larger than the true number of sources, i.e., L " Q + 1. The
idea behind (8.23) is intuitive in the sense that it tries to find the sparsest spatial spectrum
distribution p and the noise power σn2 such that the difference between the resulting
covariance matrix AP AH + σn2 I and the sample covariance matrix R̂ is minimized.
236 Gu, Goodman, and Zhang
However, the true number of sources is a priori unknown. Even if known, it is under-
stood that (8.23) is a difficult combinatorial optimization problem due to the nonconvex
0 “norm” constraint, and is intractable even for moderately sized problems. In the past
decades, many approximation methods have been proposed to solve this nonconvex
optimization problem, such as greedy approximations [51,52] and lp (p ≤ 1) convex
relaxations [53,54]. When the solution p is sufficiently sparse, the 0 “norm” can be
approximately replaced by the 1 -norm. By introducing the 1 -norm convex relaxation,
the nonconvex optimization problem (8.23) can be formulated as a convex one:
4 4
K
4 4
min 4R̂ − AP AH − σn2 I 4 s.t. p1 ≤ σs2 + σi2k + σn2 + δ,
p,σn2 F
k=1
p ≥ 0,
σn2 > 0, (8.24)
K
where the 1 -norm of p equals the power sum of all sources (i.e., σs2 + k=1 σi2k + σn2 ),
and a small number δ > 0 is added to the power constraint in order to allow a space for
the optimization algorithm to search for p. However, in practical applications, the true
number of sources is not easy to know, not to mention their power.
Alternatively, the convex optimization problem in (8.24) can be reformulated as a
basis pursuit denoising (BPDN) problem [55] as
4 4
4 4
min 4R̂ − AP AH − σn2 I 4 + γp1 s.t. p ≥ 0,
p,σn2 F
[3,53] and the references therein). In general, the DOAs are estimated either from a
spectral search algorithm (see, for example, [50,57]) or from a search-free polynomial
rooting algorithm (e.g., [58] and the references therein). For convenience, here we
simply use the classical Capon spatial spectrum p Capon (θ) in (8.21) to estimate the
DOAs of sources. The estimated DOAs provide the support of the sparse vector defined
in the proposed sparsity-constrained covariance matrix fitting problem.
Let p denote the set of directions corresponding to the peaks of p Capon on the entire
observed spatial domain (i.e., ∪ ), ¯ for which the cardinality is usually greater than
the true number of sources because of the spurious peaks (i.e., |p | = p0 > Q + 1).
Here, | · | denotes the cardinality of a set. In order to minimize the 0 “norm” p0 to
find the sparsest solution of (8.23), a common method is to remove the spurious peaks by
setting a threshold, such as the noise power, which can be approximately$ estimated
% as the
minimum eigenvalue of the sample covariance matrix (i.e., σ̂n2 = λmin R̂ ) [59], where
λ min ( · ) denotes the minimum eigenvalue of a matrix. In theory, there are M − Q − 1
eigenvalues that equal the actual noise power σn2 . However, in practical applications with
a limited number of snapshots, the minimum eigenvalue of the sample covariance matrix
is always smaller than the noise power. Hence, if the value of a peak in the Capon spatial
spectrum pCapon is lower than the threshold, it will be regarded as a spurious peak and
its corresponding direction will be removed from the set p . After removing all the
spurious peaks, the residual set is denoted as ˜ p = { θ̃p,1,. . ., θ̃ } with cardinality
p, Q̃
˜
|p | = Q̃ ≤ |p |. In such a case, p0 = Q̃ ≥ Q + 1.
After finding the DOA support θ̃p = [ θ̃p,1,. . ., θ̃p, Q̃ ]T , the sparsity-constrained
covariance matrix fitting problem (8.23) degenerates into an inequality-constrained
least-squares problem:
4 4
min 4R̂ − A( θ̃ p )P ( θ̃p )AH ( θ̃p ) − σn2 I 4F s.t. p( θ̃ p ) > 0,
p(θ̃ p ),σn2
where P ( θ̃p ) = diag(p( θ̃ p )) is a diagonal matrix with the power distribution p( θ̃p ) ∈
Q̃
R++ on the DOA support θ̃p , and A( θ̃p ) = a(θ̃p,1 ),. . .,a( θ̃p, Q̃ ) ∈ CM×Q̃ is the cor-
Q̃
responding array manifold matrix. Here, R++ denotes the set of Q̃-dimensional vectors
of positive real numbers. The strict inequality constraint enforced here indicates that
the signal power on the found DOA support θ̃p are always positive. The optimization
problem (8.26) is convex and can be solved using highly efficient interior point methods.
It is noted that covariance matrix reconstruction-based adaptive beamformers are
not very sensitive to the estimation error in the noise power. Therefore, for the sake
of simplicity, the optimization variable of noise power σn2 is taken to be the mini-
mum eigenvalue of R̂, which leads to a simplified inequality-constrained least-squares
problem:
4 4
min 4R̂ − λmin (R̂)I − A( θ̃p )P ( θ̃p )AH ( θ̃ p )4F s.t. p( θ̃p ) > 0. (8.27)
p(θ̃ p )
238 Gu, Goodman, and Zhang
where vec ( · ) denotes the vectorization operator, and denotes the Khatri–Rao prod-
uct. Without the inequality constraint, the closed-form solution to (8.27) is given by
−1
p( θ̃p ) = GH G GH r, (8.29)
where G = A( θ̃p ) A( θ̃ p ) = a(θ̃p,1 ) ⊗ a( θ̃p,1 ),. . .,a( θ̃p, Q̃ ) ⊗ a( θ̃p, Q̃ )
= vec(a( θ̃p,1 )a H ( θ̃p,1 )),. . .,vec(a( θ̃p, Q̃ )a H ( θ̃p, Q̃ )) ∈ CM ×Q̃ is obtained by stack-
2
2
ing the outer products of the sources steering vectors, and r = vec(R̂−λ min (R̂)I ) ∈ CM
is vectorized from the sample covariance matrix subtracted by an estimated noise
covariance matrix. Here, ⊗ denotes the Kronecker product. Then, the estimated spatial
spectrum of (8.23) is Q̃-sparse and is expressed as
p( θ̃p ) θ∈˜ p,
p(θ) = ˜ p. (8.30)
0 θ∈
/
Namely, only Q̃ entries of the estimated spatial spectrum p(θ) are nonzero and all
other L − Q̃ entries are zero. An example of spatial spectrum comparison between the
proposed sparse spectrum (8.30) and the Capon spectrum (8.21) is illustrated in Figure
8.2, where three sources impinge from DOAs of −50◦ , −20◦ , and 5◦ with the SNR of
30 dB, 30 dB, and 20 dB, respectively. It is clear that the proposed method achieves a
more accurate estimate of the signal power.
30
Capon spectrum
25 Sparse spectrum
20
15
Power (dB)
10
–5
–10
–15
–90 –60 –30 0 30 60 90
θ( )
However, when some sources are very weak, there may be negative entries in
p(θ̃p ) (8.29), which is obtained by discarding the inequality constraint in (8.27). With-
out loss of generality, assume that the q̃-th entry of p( θ̃ p ) is negative, i.e., p( θ̃p, q̃ ) < 0.
In such a case, the inequality constraint in (8.26) will not be satisfied, and the closed-
form solution in (8.29) should be modified. A simple method is to force p( θ̃p, q̃ ) to be a
small positive value δ > 0 (for example, δ = 10−5 is used in our simulations), and the
power estimation of other sources, θ̄p = [θ̃p,1,. . ., θ̃p, q̃−1, θ̃p, q̃+1,. . ., θ̃p, Q̃ ]T ∈ RQ̃−1 ,
will be modified as
p̄(θ̄ p ) = [Ḡ Ḡ]−1 Ḡ r̄,
H H
(8.31)
where Ḡ = vec(a(θ̃p,1 )a H ( θ̃p,1 )),. . .,vec(a( θ̃p, q̃−1 )a H ( θ̃p, q̃−1 )),vec(a( θ̃p, q̃+1 )a H
( θ̃p, q̃+1 )),. . .,vec(a( θ̃p, Q̃ )a H ( θ̃p, Q̃ )) ∈ CM ×(Q̃−1) , and r̄ = vec(R̂ − λmin (R̂)I
2
2
−δa( θ̃p, q̃ )a H ( θ̃p, q̃ )) ∈ CM . In other words, we recalculate the source powers after
fixing the power of weak sources, thus resulting in a modified spatial spectrum as
⎧
⎪
⎨ p̄( θ̄p ) θ ∈ θ̄ p,
p(θ) = δ θ = θ̃p, q̃ , (8.32)
⎪
⎩ 0 θ∈/ θ̃p .
where a(θiq )a H (θiq ) is the outer product of the q-th interference steering vector a(θiq ).
Because there are at most Q̃ elements in the set ¯ ∩ ˜ p , the integral operation in (8.20)
is effectively simplified to be a summation operation (8.33) by using the sparse charac-
teristics of sources in the observed spatial domain. Note that there is no desired signal
component in the reconstructed interference-plus-noise covariance matrix.
Considering the possible look direction mismatch, the DOA of the desired signal can
be located by searching for the peak of pCapon in , i.e., θ̃s = arg maxθ∈ pCapon (θ),
and the corresponding steering steering vector is denoted as ã s = a( θ̃s ). When ∩ ˜p
is empty, which is common at low SNRs, we simply use the presumed signal steering
vector for adaptive beamformer design, i.e., ã s = ā s , even if there is a look direction
mismatch.
Substituting the reconstructed interference-plus-noise covariance matrix R̂ i+n and
the estimated signal steering vector ã s into the MVDR beamformer (8.8) together, we
can propose the adaptive beamformer as
−1
R̂ i+n ã s
w= −1
. (8.34)
ã H
s R̂ i+n ã s
Step 1: Estimate the DOAs of the sources by, e.g., searching for the peaks of the Capon
spatial spectrum p Capon (θ) (8.21);
Step 2: Solve the least-squares problem (8.27) to obtain the Q̃-sparse spatial spectrum
p(θ) (8.30) or (8.32);
Step 3: Reconstruct the interference-plus-noise covariance matrix R̂ i+n (8.33) and
estimate the signal steering vector ã s ;
Step 4: Compute the proposed adaptive beamformer w (8.34).
however that, if the spatial estimate of the sources in the entire region is desired, the
SMI beamformer has the complexity of O(LM 2 ) as well.
of interference sources. In the worst-case beamformer (8.16), the upper bound of the
mismatched vector is ad hoc chosen to be ε = 0.3M, as suggested in [32]. Without
loss of generality, the angular sector covering the direction of the desired signal in
the reconstruction-based beamformers is set to be = [θ̄s − 5◦, θ̄s + 5◦ ] (namely,
¯ = [−90◦, θ̄s − 5◦ ) ∪ ( θ̄s + 5◦,90◦ ]
[0◦,10◦ ]), and the corresponding out-of-sector is
◦ ◦ ◦ ◦
(namely, [−90 ,0 ) ∪ (10 ,90 ]). The sampling grid is uniform in ∪ ¯ with 0.1◦
increment between adjacent grid points. As a benchmark, the optimal SINR (8.5)
is also shown in all figures, which is calculated from the exact interference-plus-
noise covariance matrix and the actual desired signal steering vector. Considering
that the output performance of the interference-plus-noise covariance matrix (sparse)
reconstruction-based beamformers is very close to the optimal SINR regardless of
the input signal power [40,43], we also compare the output performance in terms of
deviation from the optimal SINR. For fair comparison, the actual steering vector of the
desired signal is normalized so that a22 = M(= 10) [32,33]. The CVX software [60]
is used to solve the related convex optimization problems.
60
Optimal SINR
50 SMI
DLSMI
40 Eigenspace
Worst-Case
30
Output SINR (dB)
IAA
Reconstruction
20
Sparsity
10
–10
–20
–30
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(a)
2
SMI
1.8 DLSMI
Deviations from optimal SINR (dB)
1.6 Eigenspace
Worst-Case
1.4 IAA
Reconstruction
1.2 Sparsity
1
0.8
0.6
0.4
0.2
0
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(b)
30
25
20
Output SINR (dB)
15
10 Optimal SINR
SMI
5 DLSMI
Eigenspace
0 Worst-Case
IAA
–5 Reconstruction
Sparsity
–10
10 20 30 40 50 60 70 80 90 100
Snapshots
(c)
Figure 8.3 First example: exactly known steering vectors. (a) output SINR versus input SNR;
(b) deviations from optimal SINR versus input SNR; (c) output SINR versus number of
snapshots.
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 243
20
10
0
Beampattern (dB)
–10
–20
–30
Optimal
–40 SMI
DL-SMI
–50 EIG
WorstCase
–60
–90 –60 –30 0 30 60 90
θ( )
20
10
0
Beampattern (dB)
–10
–20
–30
–40 Optimal
IAA
–50 Reconstruction
Sparsity
–60
–90 –60 –30 0 30 60 90
θ( )
In Figure 8.4, we compare the beampattern of the proposed beamformer with those
of the other tested beamformers for K = 30 and SNR = 20 dB, where the vertical solid
line denotes the direction of the desired signal and the vertical dashed lines denote the
directions of interference. It is evident that the beampattern of the proposed adaptive
beamformer almost exactly matches that of the optimal one.
60
Optimal SINR
50 SMI
DLSMI
40 Eigenspace
Worst-Case
30
Output SINR (dB)
IAA
Reconstruction
20
Sparsity
10
–10
–20
–30
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(a)
2
DLSMI
1.8 Eigenspace
Deviations from optimal SINR (dB)
1.6 Worst-Case
IAA
1.4 Reconstruction
Sparsity
1.2
0.8
0.6
0.4
0.2
0
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(b)
30
25
20
15
Output SINR (dB)
10
5 Optimal SINR
SMI
0
DLSMI
–5 Eigenspace
Worst-Case
–10 IAA
Reconstruction
–15
Sparsity
–20
10 20 30 40 50 60 70 80 90 100
Number of snapshots
(c)
Figure 8.5 Second example: fixed signal DOA mismatch. (a) output SINR versus input SNR;
(b) deviations from optimal SINR versus input SNR; (c) output SINR versus number of
snapshots.
246 Gu, Goodman, and Zhang
60
Optimal SINR
50 SMI
DLSMI
40 Eigenspace
Worst-Case
30
Output SINR (dB)
IAA
Reconstruction
20
Sparsity
10
–10
–20
–30
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(a)
2
DLSMI
1.8 Eigenspace
Deviations from optimal SINR (dB)
1.6 Worst-Case
IAA
1.4 Reconstruction
Sparsity
1.2
0.8
0.6
0.4
0.2
0
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(b)
30
25
20
Output SINR (dB)
15
10 Optimal SINR
SMI
5 DLSMI
Eigenspace
0 Worst-Case
IAA
–5 Reconstruction
Sparsity
–10
10 20 30 40 50 60 70 80 90 100
Number of snapshots
(c)
Figure 8.6 Third example: random sources look direction mismatch. (a) output SINR versus input
SNR; (b) deviations from optimal SINR versus input SNR; (c) output SINR versus number of
snapshots.
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 247
where a(θt ),t = 1,2,3,4, correspond to coherently scattered paths. The steering vector
of the t-th path, a(θt ), can be modeled as a plane wave from the direction of θt .
The DOAs of scattered paths follow independent normal distribution θt ∼ N ( θ̄s ,4◦ ),
t = 1,2,3,4, and the phases of scattered paths follow independent uniform distribution
ψt ∼ U [0,2π) ,t = 1,2,3,4. Note that the tested adaptive beamformers are imple-
mented in a block adaptive manner, which means that both θt and ψt ,t = 1,2,3,4
change from run to run but do not change from snapshot to snapshot. From Figure 8.7,
the output performance loss of the proposed beamformer is less than 0.7 dB, which is
much smaller than what is suffered by the other tested beamformers.
4
a s (k) = s0 (k)ā s + st (k)a(θt ), (8.37)
t=1
where st (k) ∼ N (0,1),t = 0,1,2,3,4, are independent and identically distributed (i.i.d.)
zero-mean complex Gaussian random variables that change from snapshot to snapshot,
θt ∼ N ( θ̄s ,4◦ ),t = 1,2,3,4, are the random DOAs changing from run to run while
remaining fixed from snapshot to snapshot. This corresponds to the case of incoherent
248 Gu, Goodman, and Zhang
60
Optimal SINR
50 SMI
DLSMI
40 Eigenspace
Worst-Case
30
Output SINR (dB)
IAA
Reconstruction
20
Sparsity
10
–10
–20
–30
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(a)
2
DLSMI
1.8 Eigenspace
Deviations from optimal SINR (dB)
1.6 Worst-Case
IAA
1.4 Reconstruction
Sparsity
1.2
0.8
0.6
0.4
0.2
0
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(b)
30
25
20
Output SINR (dB)
15
10
Optimal SINR
5 SMI
DLSMI
0
Eigenspace
–5 Worst-Case
IAA
–10 Reconstruction
Sparsity
–15
10 20 30 40 50 60 70 80 90 100
Number of snapshots
(c)
Figure 8.7 Fourth example: coherent local scattering. (a) output SINR versus input SNR;
(b) deviations from optimal SINR versus input SNR; (c) output SINR versus number of
snapshots.
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 249
60
Optimal SINR
50 SMI
DLSMI
40 Eigenspace
Worst-Case
30
Output SINR (dB)
IAA
Reconstruction
20
Sparsity
10
–10
–20
–30
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(a)
2
SMI
1.8 DLSMI
Deviations from optimal SINR (dB)
1.6 Eigenspace
Worst-Case
1.4 IAA
Reconstruction
1.2 Sparsity
1
0.8
0.6
0.4
0.2
0
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(b)
30
25
20
15
Output SINR (dB)
10
5 Optimal SINR
SMI
0
DLSMI
–5 Eigenspace
Worst-Case
–10 IAA
Reconstruction
–15
Sparsity
–20
10 20 30 40 50 60 70 80 90 100
Number of snapshots
(c)
Figure 8.8 Fifth example: wavefront distortion. (a) output SINR versus input SNR; (b) deviations
from optimal SINR versus input SNR; (c) output SINR versus number of snapshots.
250 Gu, Goodman, and Zhang
local scattering [30], where the rank of the signal covariance matrix R s is higher than
one. In the general-rank case, the output SINR should be rewritten as [10]
wH R s w
SINR = , (8.38)
wH R i+n w
w = P{R −1
i+n R s }, (8.39)
8.4.7 Discussion
From the extensive simulation results illustrated for different scenarios in this chapter,
it is clear that the proposed adaptive beamformer, based on interference-plus-noise
covariance matrix sparse reconstruction, consistently enjoys the best performance as
compared to other tested beamformers. More specifically, having benefitted from the
interference covariance matrix reconstruction, which excludes the desired signal com-
ponent, the output performance of the proposed beamformer is always close to or equal
to the optimal SINR regardless of the input SNR. In contrast, there is a slight output per-
formance degradation for the interference-plus-noise covariance matrix reconstruction-
based beamformer because the Capon spatial spectrum estimator underestimates the
power of interferers and thus decreases the estimation accuracy of the interference-plus-
noise covariance matrix. On the other hand, other tested beamformers use the signal-
contaminated covariance matrix, and thus degrade the output performance particularly
when the input SNR is high. Another significant advantage of the proposed adaptive
beamformer is that it achieves faster convergence, and only requires the number of
snapshots to be slightly higher than the number of array sensors in order to approach the
optimal SINR. In contrast, if an average performance loss of less than 3 dB is required
for the SMI beamformer, the number of signal-free snapshots must be at least twice the
number of array sensors [48]. However, in the underlying problem where the snapshots
are contaminated by the desired signal, the convergence becomes much slower, i.e., a
much higher number of snapshots are required [17].
Note that, although the ULA is adopted in our simulations, the idea of covariance
matrix sparse reconstruction proposed in this chapter for the adaptive beamformer
design can be generalized to arbitrary arrays, e.g., the coprime array [66,67]. Having
benefitted from the larger array aperture due to the sparse deployment, the coprime
array adaptive beamforming algorithm proposed in [67] achieves high robustness
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 251
60
Optimal SINR
50 SMI
DLSMI
40 Eigenspace
Worst-Case
30
Output SINR (dB)
IAA
Reconstruction
20
Sparsity
10
–10
–20
–30
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(a)
2
SMI
1.8 DLSMI
Deviations from optimal SINR (dB)
1.6 Eigenspace
Worst-Case
1.4 IAA
Reconstruction
1.2 Sparsity
1
0.8
0.6
0.4
0.2
0
–30 –20 –10 0 10 20 30 40 50
Input SNR (dB)
(b)
30
25
20
Output SINR (dB)
15
10 Optimal SINR
SMI
5 DLSMI
Eigenspace
0 Worst-Case
IAA
–5 Reconstruction
Sparsity
–10
10 20 30 40 50 60 70 80 90 100
Number of snapshots
(c)
Figure 8.9 Sixth example: incoherent local scattering. (a) output SINR versus input SNR;
(b) deviations from optimal SINR versus input SNR; (c) output SINR versus number of
snapshots.
252 Gu, Goodman, and Zhang
against model mismatches with a significant reduction in the number of antennas and
the associated radio frequency chains.
It is also worth noting that the extension of the proposed narrowband adaptive
beamforming technique to the broadband adaptive beamforming is straightforward [3].
For example, applying fast Fourier transform (FFT) to the broadband signal yields nar-
rowband components, which can then be independently processed using the proposed
(narrowband) covariance matrix sparse reconstruction-based adaptive beamforming
technique. Subsequently, the time-domain broadband beamformer output is obtained by
applying an inverse FFT to the output of the individual narrowband beamformers.
8.5 Conclusion
References
[1] B. D. Van Veen and K. M. Buckley, “Beamforming: A versatile approach to spatial filtering,”
IEEE ASSP Mag., vol. 5, no. 2, pp. 4–24, Apr. 1988.
[2] H. Krim and M. Viberg, “Two decades of array signal processing research: The parametric
approach,” IEEE Signal Process. Mag., vol. 13, no. 4, pp. 67–94, July 1996.
[3] H. L. Van Trees, Optimum Array Processing: Part IV of Detection, Estimation, and
Modulation Theory. John Wiley & Sons, 2002.
[4] J. Li and P. Stoica, Eds., Robust Adaptive Beamforming. John Wiley & Sons, 2005.
[5] W. Liu and S. Weiss, Wideband Beamforming: Concepts and Techniques. Wiley, 2010.
[6] J. Capon, “High-resolution frequency-wavenumber spectrum analysis,” Proc. IEEE, vol. 57,
no. 8, pp. 1408–1418, Aug. 1969.
[7] S. A. Vorobyov, “Principles of minimum variance robust adaptive beamforming design,”
Signal Process., vol. 93, no. 12, pp. 3264–3277, Dec. 2013.
[8] K. Yang, T. Ohira, Y. Zhang, and C.-Y. Chi, “Super-exponential blind adaptive beamform-
ing,” IEEE Trans. Signal Process., vol. 52, no. 6, pp. 1549–1563, June 2004.
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 253
[9] H. Cox, R. M. Zeskind, and M. H. Owen, “Robust adaptive beamforming,” IEEE Trans.
Acoust. Speech Signal Process., vol. 35, no. 10, pp. 1365–1376, Oct. 1987.
[10] A. B. Gershman, “Robust adaptive beamforming in sensor arrays,” Int. J. Electron.
Commun., vol. 53, no. 6, pp. 305–314, Dec. 1999.
[11] W. F. Gabriel, “Spectral analysis and adaptive array superresolution techniques,” Proc.
IEEE, vol. 68, no. 6, pp. 654–666, June 1980.
[12] Y. I. Abramovich, “Controlled method for adaptive optimization of filters using the criterion
of maximum SNR,” Radio Eng. Electron. Phys., vol. 26, no. 3, pp. 87–95, Mar. 1981.
[13] B. D. Carlson, “Covariance matrix estimation errors and diagonal loading in adaptive
arrays,” IEEE Trans. Aerosp. Electron. Syst., vol. 24, no. 4, pp. 397–401, July 1988.
[14] L. Du, T. Yardibi, J. Li, and P. Stoica, “Review of user parameter-free robust adaptive
beamforming algorithms,” Digital Signal Process., vol. 19, no. 4, pp. 567–582, July 2009.
[15] P. Stoica, J. Li, X. Zhu, and J. R. Guerci, “On using a priori knowledge in space-time
adaptive processing,” IEEE Trans. Signal Process., vol. 56, no. 6, pp. 2598–2602, June 2008.
[16] L. Chang and C.-C. Yeh, “Performance of DMI and eigenspace-based beamformers,” IEEE
Trans. Antennas Propagat., vol. 40, no. 11, pp. 1336–1347, Nov. 1992.
[17] D. D. Feldman and L. J. Griffiths, “A projection approach for robust adaptive beamforming,”
IEEE Trans. Signal Process., vol. 42, no. 4, pp. 867–876, Apr. 1994.
[18] J. K. Thomas, L. L. Scharf, and D. W. Tufts, “The probability of a subspace swap in the
SVD,” IEEE Trans. Signal Process., vol. 43, no. 3, pp. 730–736, Mar. 1995.
[19] M. Hawkes, A. Nehorai, and P. Stoica, “Performance breakdown of subspace-based
methods: Prediction and cure,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process.,
Salt Lake City, UT, May 2001, pp. 4005–4008.
[20] T. Yardibi, J. Li, P. Stoica, M. Xue, and A. B. Baggeroer, “Source localization and sensing:
A nonparametric iterative adaptive approach based on weighted least squares,” IEEE Trans.
Aerosp. Electron. Syst., vol. 46, no. 1, pp. 425–443, Jan. 2010.
[21] L. C. Godara, “The effect of phase-shift errors on the performance of an antenna-array
beamformer,” IEEE J. Ocean. Eng., vol. 10, no. 3, pp. 278–284, July 1985.
[22] J. W. Kim and C. K. Un, “An adaptive array robust to beam pointing error,” IEEE Trans.
Signal Process., vol. 40, no. 6, pp. 1582–1584, June 1992.
[23] N. K. Jablon, “Adaptive beamforming with the generalized sidelobe canceller in the presence
of array imperfections,” IEEE Trans. Antennas Propagat., vol. 34, no. 8, pp. 996–1012, Aug.
1986.
[24] A. B. Gershman, V. I. Turchin, and V. A. Zverev, “Experimental results of localization of
moving underwater signal by adaptive beamforming,” IEEE Trans. Signal Process., vol. 43,
no. 10, pp. 2249–2257, Oct. 1995.
[25] J. Ringelstein, A. B. Gershman, and J. F. Böhme, “Direction finding in random inhomoge-
neous media in the presence of multiplicative noise,” IEEE Signal Process. Lett., vol. 7, no.
10, pp. 269–272, Oct. 2000.
[26] Y. J. Hong, C.-C. Yeh, and D. R. Ucci, “The effect of a finite-distance signal source on a
far-field steering applebaum array – two dimensional array case,” IEEE Trans. Antennas
Propagat., vol. 36, no. 4, pp. 468–475, Apr. 1988.
[27] K. I. Pedersen, P. E. Mogensen, and B. H. Fleury, “A stochastic model of the temporal and
azimuthal dispersion seen at the base station in outdoor propagation environments,” IEEE
Trans. Veh. Technol., vol. 49, no. 2, pp. 437–447, Mar. 2000.
[28] J. Goldberg and H. Messer, “Inherent limitations in the localization of a coherently scattered
source,” IEEE Trans. Signal Process., vol. 46, no. 12, pp. 3441–3444, Dec. 1998.
254 Gu, Goodman, and Zhang
[29] D. Astely and B. Ottersten, “The effects of local scattering on direction of arrival estimation
with MUSIC,” IEEE Trans. Signal Process., vol. 47, no. 12, pp. 3220–3234, Dec. 1999.
[30] O. Besson and P. Stoica, “Decoupled estimation of DOA and angular spread for a spatially
distributed source,” IEEE Trans. Signal Process., vol. 48, no. 7, pp. 1872–1882, July 2000.
[31] O. L. Frost, “An algorithm for linearly constrained adaptive array processing,” Proc. IEEE,
vol. 60, no. 8, pp. 926–935, Aug. 1972.
[32] S. A. Vorobyov, A. B. Gershman, and Z.-Q. Luo, “Robust adaptive beamforming using
worst-case performance optimization: A solution to the signal mismatch problem,” IEEE
Trans. Signal Process., vol. 51, no. 2, pp. 313–324, Feb. 2003.
[33] J. Li, P. Stoica, and Z. Wang, “On robust Capon beamforming and diagonal loading,” IEEE
Trans. Signal Process., vol. 51, no. 7, pp. 1702–1715, July 2003.
[34] R. G. Lorenz and S. P. Boyd, “Robust minimum variance beamforming,” IEEE Trans. Signal
Process., vol. 53, no. 5, pp. 1684–1696, May 2005.
[35] A. Hassanien, S. A. Vorobyov, and K. M. Wong, “Robust adaptive beamforming using
sequential quadratic programming: An iterative solution to the mismatch problem,” IEEE
Signal Process. Lett., vol. 15, pp. 733–736, Nov. 2008.
[36] A. Khabbazibasmenj, S. A. Vorobyov, and A. Hassanien, “Robust adaptive beamforming
based on steering vector estimation with as little as possible prior information,” IEEE Trans.
Signal Process., vol. 60, no. 6, pp. 2974–2987, June 2012.
[37] S. A. Vorobyov, A. B. Gershman, Z.-Q. Luo, and N. Ma, “Adaptive beamforming with
joint robustness against mismatched signal steering vector and interference nonstationarity,”
IEEE Signal Process. Lett., vol. 11, no. 2, pp. 108–111, Feb. 2004.
[38] Y. J. Gu, W.-P. Zhu, and M. N. S. Swamy, “Adaptive beamforming with joint robustness
against covariance matrix uncertainty and signal steering vector mismatch,” Electronics
Lett., vol. 46, no. 1, pp. 86–88, Jan. 2010.
[39] Y. Gu and A. Leshem, “Robust adaptive beamforming based on jointly estimating covariance
matrix and steering vector,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process.,
Prague, Czech Republic, May 2011, pp. 2640–2643.
[40] Y. Gu and A. Leshem, “Robust adaptive beamforming based on interference covariance
matrix reconstruction and steering vector estimation,” IEEE Trans. Signal Process., vol. 60,
no. 7, pp. 3881–3885, July 2012.
[41] L. Huang, J. Zhang, X. Xu, and Z. Ye, “Robust adaptive beamforming with a novel
interference-plus-noise covariance matrix reconstruction method,” IEEE Trans. Signal
Process., vol. 63, no. 7, pp. 1643–1650, Apr. 2015.
[42] J. Yang, G. Liao, J. Li, Y. Lei, and X. Wang, “Robust beamforming with imprecise array
geometry using steering vector estimation and interference covariance matrix reconstruc-
tion,” Multidim. Syst. Signal Process., vol. 28, no. 2, pp 451–469, Apr. 2017.
[43] Y. Gu, N. A. Goodman, S. Hong, and Y. Li, “Robust adaptive beamforming based on
interference covariance matrix sparse reconstruction,” Signal Process., vol. 96, Part B,
pp. 375–381, Mar. 2014.
[44] Y. Gu and Y. D. Zhang, “Single-snapshot adaptive beamforming,” in Proc. IEEE Sensor
Array Multichannel Signal Process. Workshop, Sheffield, UK, July 2018.
[45] Y. C. Eldar, A. Nehorai, and P. S. La Rosa, “An expected least-squares beamforming
approach to signal estimation with steering vector uncertainties,” IEEE Signal Process. Lett.,
vol. 13, no. 5, pp. 288–291, May 2006.
Adaptive Beamforming via Sparsity-Based Reconstruction of Covariance Matrix 255
[46] K. Kumatani, T. Gehrig, U. Mayer et al., “Adaptive beamforming with a minimum mutual
information criterion,” IEEE Trans. Audio Speech Lang. Process., vol. 15, no. 8, pp. 2527–
2541, Nov. 2007.
[47] Y. Rong, Y. C. Eldar, and A. B. Gershman, “Performance tradeoffs among adaptive
beamforming criteria,” IEEE J. Sel. Top. Signal Process., vol. 1, no. 4, pp. 651–659, Dec.
2007.
[48] I. S. Reed, J. D. Mallett, and L. E. Brennan, “Rapid convergence rate in adaptive arrays,”
IEEE Trans. Aerosp. Electron. Syst., vol. 10, no. 6, pp. 853–863, Nov. 1974.
[49] L. Du, J. Li, and P. Stoica, “Fully automatic computation of diagonal loading levels for
robust adaptive beamforming,” IEEE Trans. Aerosp. Electron. Syst., vol. 46, no. 1, pp. 449–
458, Jan. 2010.
[50] R. O. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Trans.
Antennas Propag., vol. 34, no. 3, pp. 276–280, Mar. 1986.
[51] J. A. Tropp and A. C. Gilbert, “Signal recovery from random measurements via orthogonal
matching pursuit,” IEEE Trans. Inf. Theory, vol. 53, no. 12, pp. 4655–4666, Dec. 2007.
[52] A. J. Miller, Subset Selection in Regression. Chapman and Hall, 2002.
[53] D. Malioutov, M. Çetin, and A. S. Willsky, “A sparse signal reconstruction perspective
for source localization with sensor arrays,” IEEE Trans. Signal Process., vol. 53, no. 8,
pp. 3010–3022, Aug. 2005.
[54] Y. C. Eldar and G. Kutyniok, Eds., Compressed Sensing: Theory and Applications.
Cambridge University Press, 2012.
[55] S. Chen, D. Donoho, and M. Saunders, “Atomic decomposition by basis pursuit,” SIAM J.
Sci. Comput., vol. 20, no. 1, pp. 33–61, 1998.
[56] R. Tibshirani. “Regression shrinkage and selection via the lasso,” J. Roy. Stat. Soc. B, vol.
58, no. 1, pp. 267–288, 1996.
[57] R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters via rotational invariance
techniques,” IEEE Trans. Acoust. Speech Signal Process., vol. 37, no. 7, pp. 984–995, July
1989.
[58] A. B. Gershman, M. Rubsamen, and M. Pesavento, “One- and two-dimensional direction-
of-arrival estimation: An overview of search-free techniques,” Signal Process., vol. 90, no.
5, pp. 1338–1349, May 2010.
[59] K. Harmanci, J. Tabrikian, and J. L. Krolik, “Relationships between adaptive minimum
variance beamforming and optimal source localization,” IEEE Trans. Signal Process., vol.
48, no. 1, pp. 1–12, Jan. 2000.
[60] M. Grant, S. Boyd, and Y. Y. Ye, “CVX: MATLAB software for disciplined convex
programming,” Dec. 2017. Available: http://cvxr.com/cvx/
[61] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht, “Compressed sensing off the grid,” IEEE
Trans. Inf. Theory, vol. 59, no. 11, pp. 7465–7490, Nov. 2013.
[62] Y. Li and Y. Chi, “Off-the-grid line spectrum denoising and estimation with multiple
measurement vectors,” IEEE Trans. Signal Process., vol. 64, no. 5, pp. 1257–1269, Mar.
2016.
[63] C. Zhou, Y. Gu, X. Fan et al., “Direction-of-arrival estimation for coprime array via virtual
array interpolation,” IEEE Trans. Signal Process., vol. 66, no. 22, pp. 5956–5971, Nov.
2018.
[64] C. Zhou, Y. Gu, Z. Shi, and Y. D. Zhang, “Off-grid direction-of-arrival estimation using
coprime array interpolation,” IEEE Signal Process. Lett., vol. 25, no. 11, pp. 1710–1714,
Nov. 2018.
256 Gu, Goodman, and Zhang
9.1 Introduction
257
258 Aubry, Carotenuto, De Maio, and Govoni
• The first technique [22] leans on the iterative adaptive algorithm (IAA) [23–25],
which is a sequential procedure aimed at enhancing the spectrum estimate,
suitably reducing the leakage effect suffered by the conventional filter bank, and
includes a Bayesian information criterion (BIC)-based stage [26] to promote
block-sparsity in the recovery process;
• The second approach [27] retrieves the space–frequency profile as solution to a
regularized maximum likelihood (RML) estimation problem, where a term pro-
moting the block-sparsity of the 2-D profile is directly included in the objective.
This penalty function is intertwined with the “lq -norm” (0 < q ≤ 1) of the
vector and contains the space–frequency source energies pushing for a sparse
2-D profile estimate. Since this procedure extends the sparse learning via iterative
minimization (SLIM) algorithm that was developed in [28] to the block-sparsity
scenario, in the following, it is referred to as block SLIM (BSLIM).
At the analysis stage, some case studies are illustrated to assess the effectiveness of
the proposed 2-D spectrum sensing strategies. Specifically, the capabilities of the differ-
ent signal processing techniques to recover the actual space–frequency occupancy maps
are evaluated for both simulated and measured data (acquired via the software-defined
radio (SDR) device “RTL-SDR R820T2 RTL2832U 1PPM TCXO”). The results high-
light that both BSLIM and IAA may provide a reliable space–frequency electromagnetic
cognizance with significant performance improvements as compared with the classic
filter bank at the price of increased computational complexity.
To summarize, the main contributions of this chapter are:
• The introduction of a 2-D sensing signal model exploiting the data collected at the
end of each radar pulse repetition interval (PRI) to address the spectrum-sensing
task for a modern (potentially cognitive) radar equipped with a phased-array
antenna and several TX-RX modules. These data are already used in modern
radars to establish (with heuristic tools) the “less disturbed frequency” and
accomplish frequency diverse transmissions.
Spectrum Sensing for Cognitive Radar via Model Sparsity Exploitation 259
Let us consider a sensor that is equipped with M antennas1 that collects the signals trans-
mitted by K sources located in specific, but unknown, angular directions θ̄1,. . ., θ̄K ,
with given, but unknown, spectral extensions, i.e., bandwidths.2 Figure 9.1 provides a
pictorial representation of the considered sensing scenario, where the goal is to charac-
terize the space–frequency features of the radio environment surrounding the radar of
interest.
The baseband discrete-time signal at the output of the receiving array (obtained by
sampling the continuous-time signals according to the Nyquist rate for the sensing
bandwidth B) can be written as
K
y(n) = s( θ̄h )xh (n) + w(n), n = 1,. . .N, (9.1)
h=1
1 A large number of channels/antennas is generally assumed to be the sensing task performed by a modern
(possibly cognitive) radar.
2 Symbols presenting the overbar usually refer to the true parameters.
260 Aubry, Carotenuto, De Maio, and Govoni
Figure 9.1 Sensing scenario: a sensor equipped with multiple receive antennas is used to collect
signals emitted by multiple transmit sources. ©[2018] IEEE. Reprinted, with permission from
[“Two-dimensional spectrum sensing for cognitive radar,” 2018 IEEE Radar Conference
(RadarConf18), Oklahoma City, OK, USA, 2018.].
NF
x h = [xh (1),xh (2),. . .,xh (N )]T = s F (ωk )āh,k , (9.2)
k=1
signal samples emitted by the h-th source (during the observation time) evaluated at the
frequency ωk . Hence, (9.1) can be written as
K
NF
y = [y(1)H ,y(2)H ,. . .,y(N )H ]H = s̄( θ̄h,ωk )āh,k + w, (9.3)
h=1 k=1
where
where
• w1,. . .,wN1 ∈ CN M are the interference vectors affecting the different data-
windows, modeled as i.i.d. zero-mean circularly symmetric white Gaussian ran-
dom vectors, with covariance matrix σ2 I ;
• ām,k,h is proportional to the Fourier transform of the signal emitted by the m-th
source (during the h-th data-window), evaluated at the frequency ωk .
In order to proceed, let {θi }K 1
i=1 be a grid of angular locations that are assumed
fine enough such that the true positions of the existing sources lie on the grid, i.e.,
{ θ̄1,. . ., θ̄K } ⊆ {θi }K 1 3
i=1 . Under this mild assumption , the signal model (9.4) can be
recast as
K1 NF
yh = s̄(θi ,ωk )ai,k,h + wh, (9.5)
i=1 k=1
where s̄(θi ,ωk ), i = 1,. . .,K1 , k = 1,. . .,NF , define the overall dictionary and ai,k,h ,
i = 1,. . .,K1 , k = 1,. . .,NF , represent the overall space–frequency profile at the h-
th snapshot. Notice that the space–frequency profile in general depends on h, since
the signal transmitted by each emitter changes with the data-window, due to both data
modulation and channel condition state. However, if the space–frequency pair (i
,k
)
does not belong to the space–frequency support of the emitters at the h
-th snapshot (in
the following referred to as an inactive pair at the h
-th snapshot), then it is assumed that
ai
,k
,h = 0 for all the snapshots. This is tantamount to requiring that both the angular
location as well as the spectral support of the emitters are stationary along the overall
data acquisition time. We write this formally as
(i
,k
) is inactive at h
⇒ ai
,k
,h = 0, ∀h = 1,. . .,N1 . (9.6)
Based on (9.4) and (9.5), 2-D spectrum sensing is tantamount to recovering the overall
space–frequency profile. Precisely speaking, the space–frequency occupancy map can
be obtained from the estimated profile starting from the set of angle-frequency bins
whose energy is different from zero, i.e.,
N1
|ai
,k
,h |2 > 0.
h=1
3 In the presence of mismatches, some refinements can be accomplished, for instance resorting to the
RELAX algorithm [25,29,30].
262 Aubry, Carotenuto, De Maio, and Govoni
To shed light on the intrinsic structure of the considered signal model (9.5), let us
introduce the following compact vectorial representation
y h = H x h + wh, h = 1,. . .,N1, (9.7)
where
According to (9.9), the received signal for each data-window is the superposition of
the weighted angular-frequency components associated with the different sources. Such
configuration, usually referred to as the subspace signal model [31,32], is ever-present
in signal processing applications including direction of arrival evaluation and spectral
analysis where the functional form of the steering vectors is application-dependent.
Regardless of the actual steering vector structure, several algorithms have been proposed
in the open literature to estimate the unknown model parameters [24,33–35]. In the
following section, some advanced recovery strategy exploiting the block-sparsity of the
unknown matrix X induced by condition (9.6) are presented.
Before concluding this section, some interesting remarks about the importance of 2-D
spectrum awareness for cognitive radars are now provided. Precisely:
In this section, two adaptive signal processing techniques with the ability to recover
the space–frequency occupancy map via block-sparsity exploitation are described:
BIC-based IAA processing and BSLIM .
mum number of space–frequency sources able to appropriately fit the available data by
means of a smart successive interference cancellation approach.
In order to proceed, let us introduce the IAA technique. It is a sequential iterative
algorithm that tries to evaluate the unknown 2-D profile iteratively refining the estimates
provided by the conventional filter bank approach usually suffering the so-called leakage
effect [25]. Focusing on the ith steering vector, the idea is to consider all the other
contributions as interference. In this respect, assuming that the phases of xi,h , i ∈
{1,. . .,i − 1,i + 1,. . .,K1 NF }, h = 1,. . .,N1 , are i.i.d. random variables uniformly
distributed over [0,2π]5 , the average (over the N1 data-windows) covariance matrix
of the interference experienced by the source corresponding to the ith steering vector is
K
1 NF
Qi = l + σ I,
Pl hl hH 2
i = 1,. . .,K1 NF , (9.13)
l=1
l!=i
with Pl = N1 X l 2 ,
1 2 l = 1,. . .K1 NF , and
T
X k = Xk,1,Xk,2,. . .,Xk,N1 ∈ CN1 ,
is the k-th row of the matrix X. In other words, Pl is the average power (over the different
data-windows) of the l-th source. As shown in [24], the best linear unbiased estimator
of Xi (notice that the different data-windows are statistically uncorrelated) is given by
−1
hH
i Qi Y
X̂ i = −1
, i = 1,. . .,K1 NF (9.14)
hH
i Qi hi
Now, letting
K
1 NF
R= l + σ I,
Pl hl hH 2
(9.15)
l=1
it follows that
Qi = R − Pi hi hH
i . (9.16)
Therefore, applying the matrix inversion lemma [24,34] to (9.16)
hH −1
−1 i R
i Qi =
hH −1
, i = 1,. . .,K1 NF , (9.17)
1 − Pi hH
i R hi
consequently, the estimator of X i can be written as
hH −1
i R Y
X̂ i = −1
, i = 1,. . .,K1 NF (9.18)
hH
i R hi
5 This is in accordance with the key references [24,25,43], which lay the theoretical background of the
IAA approach. Nevertheless, at the analysis stage arbitrary signals, i.e., not necessarily characterized by
frequency bin weights with i.i.d. phases, are considered.
Spectrum Sensing for Cognitive Radar via Model Sparsity Exploitation 265
Now, since R depends on the unknowns X i 22 , i = 1,. . .,K1 NF , the IAA iteratively
replaces Pi with the estimates P̂i = N11 X̂ i 22 in (9.15) and initializes the process with
the filter bank outputs [22]. Otherwise stated, at each step the IAA predicts the matrix
R, R̂ say, using in (9.15) the source contributions X̂ that were estimated at the previous
step, according to (9.18). This procedure is iterated until a convergence condition is
reached, e.g., a maximum number of iterations p̄ is performed, or
K
1 NF (p)
(p−1)
X̂ i 2 − X̂ i ¯
2 ≤ ,
i=1
where p > 0 is the iteration step and ¯ is a maximum distance among two successive
profile estimates. The IAA process is summarized in Algorithm 1.
Finally, the per-iteration computational complexity of the IAA [24,25] is O(NF K1
N 2 M 2 ). Some computationally efficient implementations can be also conceived see
[43–46].
Let us now promote block sparsity in the recovery process jointly using the profile
estimate obtained by Algorithm 1 and the BIC framework. To this end, let us denote by
where N MN1 is the size of the observation data, i.e., the product between the
number of available snapshots, antennas, and data-windows, and the factor 4 in
the second term of (9.19) accounts for the number of unknowns for each source,
i.e., its complex valued amplitude, angle, and frequency.
Then, at step k, 1 ≤ k ≤ K̄, the space–frequency source with index
h̃
k = arg min BICk (h̃)
h̃!∈I (k)
BI C X̂ i if i ∈ I(k
)
X̂ i = . (9.20)
0 otherwise
It is worth pointing out that the procedure requires the storage of the two K̄-dimensional
real-valued vectors (ordered sets)
[h̃
1, h̃
2,. . ., h̃
K̄ ],
BIC1 (h̃
1 ),BIC2 (h̃
2 ),. . .,BICK̄ (h̃
K̄ ) .
where
2 0$ 1
K
1 NF
%q/2
f1 (X) = Xk 22 + −1 (9.22)
q
k=1
the k-th row of the matrix X and > 0 a smoothing factor making (9.22) differentiable.
In (9.21), σL2 and σU
2 are respectively a lower bound and an upper bound for the white
interference. σL can be evaluated characterizing the power level associated with the
2
isolated operation of the receiver components, whereas σU2 can be obtained via measure-
K N
1 F
2
2 Xk 2 − N1 NF ,
q
k=1
: 'q/2 ;
2 &4
K81 NF
4
fX (X) ∝ exp − 4Xk 42 + , (9.24)
q
k=1
which is equivalent P (it is enough to consider the negative logarithm of the objective
2 → ∞, a so called improper prior is assumed for σ 2 at the
in (9.25)). Notice that, as σU
recovery stage [54,55], tantamount to assuming that σ2 has equal probability over the
range [σL2 ,∞].
• Rq̄ be the set of selected active row indices (see equation (9.27) and the related
description for more details on its evaluation);
• h(q̄) = |Rq̄ |, namely the number of selected active rows;
q̄
• X̄ be the least-squares estimate of X associated with the selected active rows
Rq̄ .
Hence, denoting by Iq ⊆ ]0,1] the discrete set of the considered q̄ values, the overall
BSLI M q
In equation (9.26), N MN1 is the size of the observation data, i.e., the product between
the number of available snapshots, antennas, and data-windows, while the factor
(2N1 + 2) in the second term of (9.26) represents the number of unknowns for each
source, i.e., the N1 complex valued amplitudes, the angle, and the frequency.
q̄
As to the evaluation of Rq̄ , again a BIC-based strategy is employed. Let X̂ be the
q̄ q̄
profile recovered by Algorithm 2 as q = q̄, X̂o be the matrix obtained from X̂ sorting
q̄ q̄ q̄
its rows so that X̂o,1 22 ≥ X̂o,2 22 ≥ . . ., ≥ X̂ o,N1 22, namely, the per-row energy of
q̄ q̄ q̄ q̄ q̄
X̂o is arranged in decreasing order, and r o = [ro (1),ro (2),. . .,ro (K1 NF )]T ∈ NK1 NF
q̄ q̄
be the vector containing the corresponding ordered row indices, i.e., X̂o,i = X̂ r q̄ (i),
i = 1,. . .,K1 NF . Then
q̄ q̄
Rq̄ = ro (1),. . .,ro (kq̄
) , (9.27)
where
kq̄
= arg min BICq̄ (k),
k∈{1,... K̄}
with
& k '
q̄ k q̄
BICq̄ (k) = 2N MN1 log H o X̂ o − Y 22 + (2N1 + 2)k log (2N MN1 ) ,
(9.28)
k = 1,. . ., K̄.
In (9.28)
q̄ k q̄ q̄
• H o ∈ CMN,k is the matrix containing the first k columns of H o , with H o the
matrix obtained from H sorting its columns according to the permutation induced
q̄
by the vector r o ;
q̄ k q̄
• X̂ o is the matrix containing the first k rows of X̂ o ;
• K̄ is an upper bound to the actual number of space–frequency sources (result-
ing from some upper bounds on the number of sources K and their frequency
support);
• the second term of (9.28) represents the BIC-based penalty.
270 Aubry, Carotenuto, De Maio, and Govoni
In this section, the capability to perform 2-D spectrum sensing of both BSLIM and
BIC-based IAA processing algorithms is assessed on both simulated and measured data
scenarios.
90 1
80
70 0.9
60
0.8
50
40
0.7
30
20 0.6
q (deg)
10
0 0.5
–10
–20 0.4
–30
0.3
–40
–50
0.2
–60
–70 0.1
–80
–90 0
–2.5–2.3–2.1–1.9–1.7–1.5–1.3–1.1–0.9–0.7–0.5–0.3–0.1 0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5
f(Hz) ×108
Figure 9.2 Space–frequency profile of the analyzed scenario. The gray scale is proportional to the
energy in each space–frequency bin.
Spectrum Sensing for Cognitive Radar via Model Sparsity Exploitation 271
(a)
90 1
80
70 0.9
60
0.8
50
40
0.7
30
20 0.6
q (deg)
10
0 0.5
–10
–20 0.4
–30
0.3
–40
–50
0.2
–60
–70 0.1
–80
–90 0
–2.5–2.3–2.1–1.9–1.7–1.5–1.3–1.1–0.9–0.7–0.5–0.3–0.1 0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5
f(Hz) 8
×10
(b)
Figure 9.3 Space–frequency occupancy map, first trial: (a) BSLIM; (b) BIC-based IAA
processing. The detected bins correspond to the hotter pixels.
the considered angle direction, normalized to the maximum value. The plot clearly
highlights the space–frequency portions occupied by the K emitters.
Figures 9.3 and 9.4 show the space–frequency occupancy maps recovered via the
BLSIM technique with adaptive selection of q and the BIC-based IAA processing for
two independent trials. In Algorithm 1, the exit condition based on the maximum num-
ber of iterations p̄ = 4 is considered and the BIC-based IAA processing assumes
K̄ = 180. In Algorithm 2 δ = 10−1 , = 10−6 , σL2 = 1, and σU 2 = 10. Also, the
(a)
90 1
80
70 0.9
60
0.8
50
40
0.7
30
20 0.6
q (deg)
10
0 0.5
–10
–20 0.4
–30
0.3
–40
–50
0.2
–60
–70 0.1
–80
–90 0
–2.5–2.3–2.1–1.9–1.7–1.5–1.3–1.1–0.9–0.7–0.5–0.3–0.1 0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5
f(Hz) 8
×10
(b)
Figure 9.4 Space–frequency occupancy map, second trial : (a) BSLIM; (b) BIC-based IAA
processing. The detected bins correspond to the hotter pixels.
Table 9.1 Empirical false alarm rate, PFA , and empirical detection
probability, PD , for BSLIM and BIC-based IAA processing for the
sensing scenario of Figure 9.2.
Recovery Algorithm PF A PD
Figure 9.5 Empirical ROC for the filter bank approach (solid curve), IAA algorithm (dashed
curve), BSLIM q = 1 (dashed-dotted curve), and BSLIM q = 0.45 (dotted curve).
for both BSLIM and BIC-based IAA techniques, a moving average filtering (with 2
equal weights) along the frequency dimension is performed over the recovered profile
(containing in each space–frequency bin the estimated average energy) so as to handle
the intrinsic on-off behavior of the map. Hence, PF A is evaluated counting the detections
(over 20 independent trials, each composed of N1 = 10 independent data-windows6 ) in
the space–frequency bins where no emitters are present. Analogously, PD is obtained
counting the number of detections, over the same 20 independent trials, where space–
frequency sources are located. Inspection of the table confirms previous considerations
and highlight that BSLIM can provide, for the analyzed situations, better performance
than BIC-based IAA.
Finally, in Figure 9.5 the empirical receiver operating characteristic (ROC) curves of
the filter bank approach, IAA processing, and BLSIM (with q equal to either 0.45 or 1)
are reported, assuming the same simulation setup as in Figure 9.2, but with SNR = 5dB.
Specifically, for each approach, the threshold is set so as to ensure the desired empirical
6 This setup corresponds to 157,320 space–frequency bins, which are emitters free and 2,080 bins occupied
by RF sources, where some guard cells for the communication transmitters have been included (the 3-dB
bandwidth measure is considered for each emitter).
274 Aubry, Carotenuto, De Maio, and Govoni
• M = 1 receive antenna;
• N = 100 samples per snapshot;
• N1 = 10 snapshots;
• NF = 2N frequency bins.
As in Subsection 9.4.1, Algorithm 1 involves a maximum number of iterations given by
p̄ = 4 and the BIC-based IAA processing uses K̄ = 180. Additionally, in Algorithm 2,
δ = 10−1 , = 10−6 , σL2 = 0.004, and σU 2 = 10. The lower bound to the noise
power level is estimated via a measured data spectrum where no RF sources are present,
Spectrum Sensing for Cognitive Radar via Model Sparsity Exploitation 275
Figure 9.6 Experimental setup for measured data acquisition via SDR device.
and this value is also used to evaluate the interference covariance matrix in the IAA-
based procedures. The upper bound is higher than 10 times the power of the available
data. Moreover, the adaptive selection of q assumes K̄ = 180 and Iq = {0.01,0.12,
0.23,0.34,0.45,0.56,0.67,0.78,0.89,1}.
For comparison purposes also the behavior of the filter bank technique, Algorithm 1
without BIC-based stage, and Algorithm 2 for a fixed q is analyzed. To this end, it is
assumed that they perform a noncoherent energy integration of the N1 = 10 estimated
profiles, and the resulting value in each bin is compared with a threshold, i.e., the noise
power level times a multiplicative factor (empirically set).
In Figure 9.7, the results corresponding to the 1,000 collected samples are displayed.
In particular, Figure 9.7.a illustrates the square modulus of the acquired data Fourier
transform versus frequency, i.e., the data spectrum, while Figures 9.7(b), 9.7(c), 9.7(d),
9.7(e), and 9.7(f) illustrate the frequency occupancy maps obtained via the filter bank
approach, the IAA algorithm, the BIC-based IAA processing, the BSLIM algorithm
with adaptive selection of q, and Algorithm 2 with q = 1, respectively. Specifically,
the vertical yellow lines in these maps identify the frequency bins where RF sources
are detected.
Inspection of the maps highlights the superiority of the IAA-based strategies and
BSLIM techniques over the conventional filter bank approach. Indeed, the latter pro-
276 Aubry, Carotenuto, De Maio, and Govoni
(a) (b)
(c) (d)
(e) (f)
Figure 9.7 Results associated with the first set of 1,000 samples collected in the TETRA
bandwidth. (a) data spectrum. Frequency occupancy map: (b) filter bank approach; (c) IAA
algorithm; (d) BIC-based IAA processing; (e) BSLIM (q = 1); (e) BSLIM (adaptive q
selection). The detected frequency bins correspond to the hotter pixels.
vides an unreliable recovery of the frequency occupancy profile mainly due to the energy
spillover effect induced by the multiple spectral contributions. On the other hand, both
IAA-based algorithms and BSLIM strategies are able to accurately retrieve the actual
frequency bins occupied by RF sources. However, BLIM shares a lower computational
complexity than the counterpart. Notice that a possible missed detected emitter could
be present around the frequency −0.9 MHz. However, it can not be a priori claimed if it
is an emitter or a noise spike. Remarkably, BSLIM algorithm with adaptive selection of
q and the BIC-based IAA processing are totally adaptive and do not require any ad hoc
threshold selection, which is a valuable feature from a practical point of view.
Spectrum Sensing for Cognitive Radar via Model Sparsity Exploitation 277
(a) (b)
(c) (d)
(e) (f)
Figure 9.8 Results associated with the second set of 1,000 samples collected in the TETRA
bandwidth. (a) data spectrum. Frequency occupancy map: (b) filter bank approach; (c) IAA
algorithm; (d) BIC-based IAA processing; (e) BSLIM (q = 1); (e) BSLIM (adaptive q
selection). The detected frequency bins correspond to the hotter pixels.
Figure 9.8 reports the results of the described spectrum-sensing algorithms for
another dataset still composed of 1,000 samples, assuming the same setup as in
Figure 9.7. The preceding remarks hold true also in this case, namely the IAA-based
strategies and BSLIM techniques prove effective to recover the frequency occupancy
maps.
In the second experiment the focus is on a sensing bandwidth belonging to the spec-
trum occupied by global system for mobile (GSM) communications [58]. Specifically,
a center frequency f0 = 943 MHz, a one-sided bandwidth of B = 2.8 MHz, and a
sampling frequency of 2.8 complex Mega-samples per second are considered.
278 Aubry, Carotenuto, De Maio, and Govoni
(a) (b)
(c) (d)
(e) (f)
Figure 9.9 Results associated with the first set of 1,000 samples collected in the GSM bandwidth.
(a) data spectrum. Frequency occupancy map: (b) filter bank approach; (c) IAA algorithm;
(d) BIC-based IAA processing; (e) BSLIM (q = 1); (e) BSLIM (adaptive q selection).
The detected frequency bins correspond to the hotter pixels.
In this analysis, the parameters used by all the procedures are the same as in the
first experiment7 but for σL2 = 0.01 and the threshold used after noncoherent energy
integration in filter bank technique, Algorithm 1 without BIC-based stage, and Algo-
rithm 2 with q = 1. Indeed, a different multiplicative factor to the noise power level is
considered.
Figures 9.9 and 9.10 show the results associated with two different experiments,
each corresponding to a dataset of 1,000 samples. Like the previous analysis, both the
(a) (b)
(c) (d)
(e) (f)
Figure 9.10 Results associated with the second set of 1,000 samples collected in the GSM
bandwidth. (a) data spectrum. Frequency occupancy map: (b) filter bank approach; (c) IAA
algorithm; (d) BIC-based IAA processing; (e) BSLIM (q = 1); (e) BSLIM (adaptive q
selection). The detected frequency bins correspond to the hotter pixels.
acquired data spectrum and the obtained frequency occupancy maps are provided for
each dataset.
The obtained maps confirm the ability of the IAA-based and BSLIM strategies to
correctly detect and identify frequency bins occupied by RF sources. Finally, the con-
ventional filter bank approach exhibits performance close to the other algorithms for the
datasets in the GSM bandwidth, unlike the other analyses.
As final and concluding remark, the reported analysis has highlighted the effective-
ness of the introduced approaches to endow spectrum awareness. Moreover, accounting
for both computational complexity and recovery ability, BSLIM appears to be the best
choice.
280 Aubry, Carotenuto, De Maio, and Govoni
9.5 Conclusions
2-D spectrum sensing has been considered to gather real-time space–frequency scenario
awareness, assuming a sensor equipped with multiple receive channels. A formal
discrete-time sensing signal model has been developed and two adaptive signal
processing algorithms have been introduced for recovering the space–frequency
occupancy map via block-sparsity exploitation. The former employs the IAA technique
and incorporates a BIC-based stage to promote block-sparsity in the recovery process.
The latter applies the RML estimation paradigm to automatically push for block-sparsity
in the 2-D profile evaluation.
At the analysis stage both simulated and measured data scenarios are considered to
evaluate the capability of the proposed procedures to retrieve the actual space–frequency
occupancy maps. The reported results clearly show the effectiveness of the developed
tools and their superiority over the conventional filter bank approach. Precisely, both
BIC-based IAA and BSLIM with adaptive q selection are able to grant an accurate
and reliable space–frequency occupancy map recovery without requiring an ad hoc
threshold selection. Furthermore, better performance than the classic filter bank is
obtained at the price of a higher computational complexity. Finally, considering both
computational complexity and recovery ability, it is reasonable to use BSLIM for 2-D
spectrum sensing.
Future research might concern the use of constant false alarm rate (CFAR) strategies
to pick up the most relevant space–frequency contributions within the estimated profile,
as well as the analysis of different model order selection rules both in the IAA context
and for the adaptive selection of the parameter q.
Acknowledgment
The work of Drs. Augusto Aubry and Antonio De Maio was sponsored by the US Army
RDECOM International Technology Center – Atlantic in partnership with the US Army
Research Laboratory under USAITC-A Seedling Project W911NF-17-2-0134.
References
[1] M. Wicks, “Spectrum crowding and cognitive radar,” in 2010 2nd International Workshop
on Cognitive Information Processing, Italy, June 2010, pp. 452–457.
[2] H. Griffiths, L. Cohen, S. Watts et al., “Radar spectrum engineering and management:
Technical and regulatory issues,” Proceedings of the IEEE, vol. 103, no. 1, pp. 85–102,
2015.
[3] A. Farina, A. De Maio, and S. Haykin, Eds., The Impact of Cognition on Radar Technology.
Schitech Publishing, 2017.
[4] H. He, P. Stoica, and J. Li, “Waveform design with stopband and correlation constraints for
cognitive radar,” in 2010 2nd International Workshop on Cognitive Information Processing,
Elba, Italy, June 2010.
Spectrum Sensing for Cognitive Radar via Model Sparsity Exploitation 281
[21] M. Tang, G. Ding, Q. Wu, Z. Xue, and T. A. Tsiftsis, “A joint tensor completion and
prediction scheme for multi-dimensional spectrum map construction,” IEEE Access, vol. 59,
no. 4, pp. 8044–8052, 2016.
[22] A. Aubry, A. De Maio, and M. Govoni, “Two-dimensional spectrum sensing for cognitive
radar,” in IEEE Radar Conference, Oklahoma City, OK, April 2018.
[23] S. D. Blunt and K. Gerlach, “A novel pulse compression scheme based on minimum
mean-square error reiteration,” in Proceedings of the International Conference on Radar,
Adelaide, Australia, September 2003.
[24] W. Roberts, P. Stoica, J. Li, T. Yardibi, and F. A. Sadjadi, “Iterative adaptive approaches to
mimo radar imaging,” IEEE Journal of Selected Topics in Signal Processing, vol. 4, no. 1,
pp. 5–20, 2010.
[25] T. Yardibi, J. Li, P. Stoica, M. Xue, and A. B. Baggeroer, “Source localization and
sensing: A nonparametric iterative adaptive approach based on weighted least squares,”
IEEE Transactions on Aerospace and Electronic Systems, vol. 46, no. 1, pp. 425–443, 2010.
[26] P. Stoica and Y. Selen, “Model-order selection: A review of information criterion rules,”
IEEE Signal Processing Magazine, vol. 21, no. 4, pp. 36–47, 2004.
[27] A. Aubry, V. Carotenuto, A. De Maio, and M. Govoni, “Multi-snapshot spectrum sensing for
cognitive radar via block-sparsity exploitation,” IEEE Transactions on Signal Processing,
vol. 67, no. 6, pp. 1396–1406, 2019.
[28] X. Tan, W. Roberts, J. Li, and P. Stoica, “Sparse learning via iterative minimization with
application to mimo radar imaging,” IEEE Transactions on Signal Processing, vol. 59, no. 3,
pp. 1088–1101, 2011.
[29] J. Li and P. Stoica, “Efficient mixed-spectrum estimation with applications to target feature
extraction,” IEEE Transactions on Signal Processing, vol. 44, no. 2, pp. 281–295, 1996.
[30] J. Li, P. Stoica, and D. Zheng, “Angle and waveform estimation via relax,” IEEE Transac-
tions on Aerospace and Electronic Systems, vol. 33, no. 3, pp. 1077–1087, 1997.
[31] S. M. Kay, Ed., Fundamentals of Statistical Signal Processing, Volume II: Detection Theory.
Prentice Hall, 1998.
[32] L. L. Scharf, Ed., Statistical Signal Processing. Addison-Wesley Reading, 1991.
[33] M. A. Richards, J. A. Scheer, and W. A. Holm, Eds., Principles of Modern Radar: Basic
Principles. SciTech Publishing, 2010.
[34] P. Stoica and R. L. Moses, Eds., Spectral Analysis of Signals. Pearson Prentice Hall, 2006.
[35] S. D. Blunt, K. Gerlach, and T. Higgins, “Aspects of radar range super-resolution,” in IEEE
Radar Conference, Boston, MA, April 2007.
[36] J. R. Guerci, Ed., Cognitive Radar: The Knowledge-Aided Fully Adaptive Approach. Artech
House Inc., 2010.
[37] D. R. Fuhrmann and G. San Antonio, “Transmit beamforming for mimo radar systems using
signal cross-correlation,” Transactions Aerospace Electronic System, vol. 44, no. 1, pp. 171–
186, 2008.
[38] N. Shariati, D. Zachariah, and M. Bengtsson, “Minimum sidelobe beampattern design for
mimo radar systems: A robust approach,” in IEEE International Conference on Acoustics,
Speech, and Signal Processing, Florence, Italy, May 2014.
[39] A. Aubry, A. De Maio, and Y. Huang, “MIMO radar beampattern design via PSL/ISL
optimization,” IEEE Transactions on Signal Processing, vol. 64, no. 15, pp. 3955–3976,
2016.
[40] A. Konar and N. D. Sidiropoulos, “Hidden convexity in qcqp with toeplitz-hermitian
quadratics,” IEEE Signal Processing Letters, vol. 10, no. 22, pp. 1623–1627, 2015.
Spectrum Sensing for Cognitive Radar via Model Sparsity Exploitation 283
[41] A. Aubry, V. Carotenuto, and A. De Maio, “New results on generalized fractional program-
ming problems with Toeplitz quadratics,” IEEE Signal Processing Letters, vol. 10, no. 22,
pp. 1623–1627, 2015.
[42] O. Aldayel, V. Monga, and M. Rangaswamy, “Successive qcqp refinement for mimo radar
waveform design under practical constraints,” IEEE Transactions on Signal Processing,
vol. 14, no. 64, pp. 1623–1627, 2015.
[43] M. Xue, L. Xu, and J. Li, “IAA spectral estimation: Fast implementation using the Gohberg–
Semencul factorization,” IEEE Transactions on Signal Processing, vol. 59, no. 7, pp. 3251–
3261, 2011.
[44] G. O. Glentis and A. Jakobsson, “Efficient implementation of iterative adaptive approach
spectral estimation techniques,” IEEE Transactions on Signal Processing, vol. 59, no. 9,
pp. 4154–4167, 2011.
[45] G. O. Glentis and A. Jakobsson, “Superfast approximative implementation of the iaa spectral
estimate,” IEEE Transactions on Signal Processing, vol. 60, no. 1, pp. 472–478, 2012.
[46] W. Sun, H. C. So, Y. Chen, L. T. Huang, and L. Huang, “Approximate subspace-based
iterative adaptive approach for fast two-dimensional spectral estimation,” IEEE Transactions
on Signal Processing, vol. 62, no. 12, pp. 3220–3231, 2014.
[47] Y. C. Eldar and M. Mishali, “Robust recovery of signals from a structured union of
subspaces,” IEEE Transactions on Information Theory, vol. 55, no. 11, pp. 5302–5316,
2009.
[48] Y. C. Eldar, P. Kuppinger, and H. Bolcskei, “Block-sparse signals: Uncertainty relations and
efficient recovery,” IEEE Transactions on Signal Processing, vol. 58, no. 6, pp. 3042–3052,
2010.
[49] Y. C. Eldar and G. Kutyniok, Eds., Compressed Sensing: Theory and Applications.
Cambridge University Press, 2012.
[50] Y. Wang, J. Wang, and Z. Xu, “On recovery of block-sparse signals via mixed l2 / lq (0 <
q ≤ 1) norm minimization,” EURASIP Journal on Advances in Signal Processing, vol. 76,
pp. 1–17, 2013.
[51] M. Lai, Y. Xu, and W. Yin, “Improved iteratively reweighted least squares for unconstrained
smoothed lq minimization,” SIAM Journal on Numerical Analysis, vol. 51, pp. 927–957,
2013.
[52] M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence analysis of block succes-
sive minimization methods for nonsmooth optimization,” SIAM Journal on Optimization,
vol. 23, no. 2, pp. 1126–1153, 2013.
[53] A. Aubry, A. D. Maio, A. Zappone, M. Razaviyayn, and Z. Luo, “A new sequential
optimization procedure and its applications to resource allocation for wireless systems,”
IEEE Transactions on Signal Processing, vol. 66, no. 24, pp. 6518–6533, 2018.
[54] A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin, Eds., Bayesian Data Analysis. Boca
Raton, FL: Chapman & Hall/CRC, 2004.
[55] C. P. Robert and G. Casella, Eds., Monte Carlo Statistical Methods.Springer Science +
Business Media, 2004.
[56] R. W. Stewart, K. W. Barlee, D. S. Atkinson, and L. H. Crockett, Software Defined Radio
Using MATLAB & Simulink and the RTL-SDR. Strathclyde Academic Media, 2015.
[57] J. Dunlop, D. Girma, and J. Irvine, Digital Mobile Communications and the TETRA System.
John Wiley & Sons, 2013.
[58] M. Rahnema, “Overview of the GSM system and protocol architecture,” IEEE Communica-
tions Magazine, vol. 31, no. 4, pp. 92–100, 1993.
10 Cooperative Spectrum Sharing
between Sparse Sensing-Based
Radar and Communication Systems
∗ ∗∗
Bo Li and Athina P. Petropulu
10.1 Introduction
Radio spectrum has been essential to a wide array of technologies, including com-
munications, position and navigation systems, and radars. Until recently, radar and
communication systems jointly consumed most of the spectrum below 6 GHz, with
commercial and noncommercial (i.e., military radar) uses assigned on distinct bands.
However, recent studies have shown that large chunks of spectrum designated for
radar applications are underutilized [1], while there is spectrum congestion in bands
devoted to commercial wireless communications. Further, as the number of connected
devices grows, the spectrum congestion problem will worsen. The nonuniform spectrum
utilization is clearly illustrated in the data collected in downtown Berkeley, California,
as shown in Figure 10.1 [2].
To address the need for more efficient use of spectrum, US government agencies have
been examining the possibility of allowing wireless broadband systems to operate in
the 3500–3650 MHz band, previously used exclusively by high-powered shipborne, air-
borne, and ground-based radar systems operated by the Department of Defense [3,4]. On
examining the viability of coexistence of radar and communication systems in that band,
the National Telecommunications and Information Administration (NTIA) proposed the
use of exclusion zones [5], which would protect base stations from radar interference.
However, those zones cover large US metropolitan areas, and such an approach would
not solve the problem of improving spectral efficiency.
The main problem with different systems using the same spectrum is the interference
that one system exerts to the other. According to a 2006 NTIA report [6], levels of
interference to noise ratio between −9 dB and −2 dB, which are below the thermal noise
floor level of the radar receiver, can reduce the probability of target detection. Also, the
interference generated by the radar reduces the throughput of a communication system
operating nearby.
Spectrum-sharing methods are focused on enabling radar and communication
systems to share the spectrum efficiently by minimizing interference effects. The
∗ Bo Li is now with Aurora Innovation, Inc. This work was supported by NSF under Grant ECCS-1408437.
∗∗ This work was supported by NSF under Grant ECCS-1408437.
284
Cooperative Spectrum Sharing 285
the systems’ willingness to cooperate, with payback reduced interference for the radar
and higher throughput for the communication system. Of course, the amount of infor-
mation that can be shared and the privacy issues involved would have to be evaluated
in each case. Examples of radar systems that could be amenable to such cooperation
include radar for autonomous vehicles, weather monitoring, etc.
The remainder of this chapter is organized as follows. Section 10.2 provides some
background on MIMO-MC radars. Section 10.3 presents the coexistence model between
a MIMO-MC radar and a MIMIO communication system, and outlines all assumptions
made. Section 10.4 formulates the problem of spectrum sharing and provides efficient
algorithms for solving it. Finally, Section 10.5 presents simulation results demonstrating
the performance gains of cooperation, and Section 10.6 offers some concluding remarks.
MIMO radars transmit different waveforms from their transmit (TX) antennas. Their
receive (RX) antennas forward their measurements to a fusion center, where a data
matrix is populated with the information received by each RX antenna. This matrix is
then used in array processing methods to estimate target information. For a relatively
small number of targets as compared to the number of TX and RX antennas, the data
matrix is low rank [37], and thus can be reconstructed with provable accuracy (under
certain conditions) based on a small set of its entries via matrix completion (MC)
[38,39]. This observation is the basis of MIMO-MC radars. In MIMO-MC radars, each
RX antenna forwards to the fusion center a small number of pseudo-randomly sub-
Nyquist sampled measurements, along with their sampling times, and partially fills a
row of the data matrix. Under certain conditions [39], the full data matrix, corresponding
to Nyquist sampled data, can be stably recovered via MC techniques, with reconstruc-
tion error proportional to the noise level. The recovered matrix can be subsequently used
for target detection via standard array processing methods [37]. The sub-sampling at
the antennas avoids the need for high-rate analog-to-digital converters, and the reduced
number of samples translates into power and bandwidth savings over the antenna-fusion
center link. Because MIMO-MC radars recover all missing entries of the data matrix,
they do not suffer from signal-to-noise ratio (SNR) loss due to data sub-sampling.
MIMO-MC radars can achieve the high resolution of traditional MIMO radars with
significantly fewer samples and reduced hardware complexity. We should note that
compressive sensing (CS)–based MIMO radars [40–42] are another sample-reduction
approach in the literature. Compared to MIMO-CS radars, MIMO-MC radars avoid
basis mismatch problems.
Let us consider a collocated MIMO radar system using uniform linear arrays (ULA)
for both transmission and reception. Let Mt,R and Mr,R be the number of TX and
RX antennas, respectively, and dt and dr the inter-element spacing at the transmit and
receive arrays, respectively. The radar transmits L pulses, with pulse repetition interval
(PRI) TPRI and carrier wavelength λc . The radar operates in two phases; in the first phase
the TX antennas transmit waveforms and the RX antennas receive target returns, while
Cooperative Spectrum Sharing 287
in the second phase, the RX antennas forward their measurements to a fusion center.
The K far-field targets are with distinct angles {θk }, target reflection coefficients {βk },
and Doppler shifts {νk }, and are assumed to fall in the same range bin. In the absence of
clutter, the noisy data matrix at the fusion center can be formulated as [37,39,43]
Y R = V r V Tt P S + W R , (10.1)
where the m-th row of Y R ∈ CMr,R ×L contains the L fast-time raw samples for-
warded by the m-th antenna [44]; S = [s(1),. . .,s(L)], is the waveform matrix, with
s(l) = [s1 (l),. . .,sMt,R (l)]T being the l-th snapshot across the transmit antennas; the
transmit waveforms are assumed to be orthogonal, i.e., it holds that SS H = I Mt,R
[37]; W R denotes additive noise; and P ∈ CMt,R ×Mt,R denotes the transmit precoding
matrix. V t [vt (θ1 ),. . .,vt (θK )] and V r [vr (θ1 ),. . .,vr (θK )] respectively denote
the transmit and receive steering matrix and vr (θ) ∈ CMr,R is the receive steering vector
defined as
r T
vr (θ) e−j 2π0ϑ ,. . .,e−j 2π(Mr,R −1)ϑ ,
r
(10.2)
where ϑr = dr sin(θ)/λc denotes the spatial frequency with respect to the receive array.
vt (θ) ∈ CMt,R is the transmit steering vector and is respectively defined. Matrix is
defined as diag([β1 ej 2πν1 ,. . .,β K ej 2πνK ]). D V r V Tt is the target response
matrix. At the fusion center, Y R passes through matched filters, after which target
estimation is performed via standard array processing methods [45].
When K is smaller than Mr,R and L, the noise-free data matrix, M DP S, is
low rank and under certain conditions can be provably recovered based on a subset of
its entries. This observation gave rise to MIMO-MC radars [37,39,43], where each RX
antenna sub-samples the target returns and forwards the samples to the fusion center.
The partially filled data matrix at the fusion center can be mathematically expressed as
follows (see [37, scheme I]):
Y R = (M + W R ), (10.3)
where denotes the Hadamard product; is the sub-sampling matrix containing 0’s
and 1’s. The sub-sampling rate p equals 0 /(LMr,R ), where 0 denotes the num-
ber of 1’s in . When p equals 1, the matrix is filled with 1’s, and the MIMO-MC
radar is identical to the traditional MIMO radar. At the fusion center, the completion of
M can be achieved by the following nuclear norm minimization problem [38]
where ·∗ denotes the matrix nuclear norm; δ > 0 is a parameter determined by the
sampled entries of the noise matrix, i.e., W R . It is shown in [38] that the recovery
of M is stable against noise. The matrix recovery error is proportional to the noise level
δ, while the recovery conditions are given in terms of the coherence of M. [38]
288 Li and Petropulu
where U ∈ CMr,R ×K and V ∈ CL×K contain the left and right singular vectors of M;
U ·k denotes the k-th column vector of U ; · ∞ denotes the maximum entry of the
matrix. ; and μ(V ) is the coherence of subspace spanned by basis matrix V , defined as
: ;
L L
μ(V ) max V l· ∈ 1,
2
, (10.7)
K 1≤l≤L K
where V l· denotes the l-th row vector of V . The coherence of subspace U is similarly
defined.
According to [46], assuming that matrix M is sampled uniformly at random at m
points, there exist constants C and c, such that if
5 6
1/2
m ≥ C max μ21,μ0 μ1,μ0 n1/4 nK β log n (10.8)
for some β > 2, the minimizer of the nuclear norm problem is unique and equal to M
with probability at least 1 − cn−β .
The condition in (10.8) implies that the smaller the coherence parameters, the fewer
samples are needed for recovering the full matrix M. Upper bounds of the coherence
parameters can give us an idea of how small those parameters can be.
The coherence parameters of M for the case P = I Mt,R are given in the following
theorem [43].
theorem 10.1 (Coherence of M when P = I Mt,R ) Let the minimum spatial fre-
quency separation of the K targets be ξt and ξr with respect to the transmit and receive
2
arrays. On denoting the Fejér kernel by Fn (x) = n1 sin 2(πnx) , and for dt = dr = λc /2
sin (πx)
and
Mr,R Mt,R
K ≤ min , , (10.9)
FMr,R (ξr ) FMt,R (ξt )
it holds that
Mr,R
μ(U ) ≤
μr0 . (10.10)
Mr,R − (K − 1) FMr,R (ξr )
Further, if every snapshot of the waveforms S ·l ≡ s(l) satisfies the following equation:
Mt,R 0 π π1
|S T·l vt (θ)|2 = , ∀l ∈ N+
L ,θ ∈ − , , (10.11)
L 2 2
then μ(V ) is upper bounded by
Mt,R
μ(V ) ≤
μt0 . (10.12)
Mt,R − (K − 1) FMt,R (ξt )
Cooperative Spectrum Sharing 289
Consequently,
√ the matrix M has coherence parameters μ0 max{μr0,μt0 } and μ1
K μ0 .
The bounds in Theorem 10.1, along with the orthogonality property of the radar wave-
forms in (10.11), were used to design waveforms with good incoherence properties. The
work of [43] involves numerical optimization on the complex Stiefel manifold, which
has high computational complexity. However, radar waveforms need to be updated
frequently as security against adversaries, which makes the issue of computational cost
more severe. It was also observed in [43] via simulations that using a random unitary
matrix [47] as the waveform matrix results in performance very close to that of the opti-
mum waveform, indicating that a random unitary matrix might be a good approximation
of the optimal waveform.
The matrix completion performance degrades when the signal-to-
interference-plus-noise ratio (SINR) drops below 10 dB [37], which suggests that
along with “good” radar waveforms, a precoder, designed to mitigate interference
would be very important. In the following subsection, we consider a MIMO-MC radar
that uses a random unitary matrix as the waveform matrix S, and a nontrivial precoder
maxtrix P [35,36].
theorem 10.2 Consider the MIMO-MC radar presented in Section 10.2, using a
random unitary waveform matrix S, and with M as defined in (10.3). For any transmit
precoder P such that for K0 = Rank(M) it holds that K0 ≤ K, and for an arbitrary
transmit array geometry and target angles, the coherence of the right singular vector
subspace of M is bounded as
√
K0 + 2 3K0 ln L + 6 ln L
μ(V ) ≤ μ̃t0, (10.13)
K0
lemma 10.3 Let SN be a χ 2 random variable with N degrees of freedom. Then for
each t > 0
√
P SN − N ≥ t 2N + t 2 ≤ e−t /2 .
2
(10.14)
290 Li and Petropulu
It is clear that K0 is not larger than K. Recall that M has a compact singular-value
decomposition (SVD) given as
M = U V H , (10.15)
where U ∈ CMr,R ×K0 and V ∈ CL×K0 contain the left and right singular vectors of M;
∈ RK0 ×K0 is diagonal containing the singular values. Consider the QR decomposition
of V r and S T P T V t :
V r = Qr R r ,
(10.16)
S P V t = Qt R t ,
T T
∗
1 Qt = Qr U 1 1 (Qt V 1 ) ,
M = Qr U 1 1 V H T H
(10.17)
which is a valid SVD of M. The uniqueness of the singular values of a matrix indicates
that ≡ 1 . Therefore, we can choose U = Qr U 1 and V = Q∗t V 1 . We have
Mr,R
μ(U ) = sup (Qr )m· U 1 22
K0 m∈N+
M r,R
(10.18)
Mr,R K r
≤ sup (Qr )m· 22 = μ ,
K0 m∈N+ K0 0
M r,R
where μr0 is the upper bound on μ(U ) defined in Theorem 10.1. We also have
L L
μ(V ) = sup (Q∗t )l· V 1 22 = sup (Qt )l· 22 . (10.19)
K0 l∈N+ K0 l∈N+
L L
S T P T V t F = Qt (R 1 R 2 ). (10.20)
We can represent Qt as
Qt = S T P T V t F K0 R −1
1 , (10.21)
Cooperative Spectrum Sharing 291
It holds that
$ −1 %H $ H %−1
R −1
1 R1 = R1 R1
$ %−1
= RH H
1 Qt Qt R 1
$ H ∗ ∗ T T
%−1 (10.23)
= FH
K0 V t P S S P V t F K0
$ H ∗ T
%−1
= FH
K0 V t P P V t F K0 ,
where the last equality holds because SS H = I Mt,R . Consider the QR decomposition of
P T V t F K0 given by
P T V t F K0 = Qa R a , (10.24)
where Qa ∈ CMt,R ×K0 contains orthonormal columns, and R a ∈ CK0 ×K0 is upper
triangular and full rank. Substituting (10.23) and (10.24) into (10.22), we have
L $ %−1 H ∗
μ(V ) = sup s T R a R H
a Ra Ra s l
K0 l∈N+ l
L
(10.25)
L L
= sup s T s ∗ = sup s l 22,
K0 l∈N+ l l K0 l∈N+
L L
where s l QTa S ·l , and the second equality holds because R a is invertible. Based on
[49, theorem 3], if Mt,R = O(L/ ln L), the entries of S can be approximated by i.i.d.
Gaussian random variables with distribution CN (0,1/L). Since Qa has orthonormal
columns, s l ∈ CK0 ,∀l ∈ N+ L also contains i.i.d. Gaussian random variable with distri-
bution CN (0,1/L), and Ls l 22 is distributed according to χK2 . Based on Lemma 10.3
√ 0
and setting t = 6 ln L, it holds that
P Ls l 22 ≥ K0 + 2 3K0 ln L + 6 ln L ≤ L−3 . (10.26)
From the derivation, the bound on μ(V ) holds for any target angles, array geometry, and
precoding matrix P as long as P T V t F K0 has full column rank K0 . This completes the
proof of Theorem 10.2.
Based on Theorem 10.2, we have the following theorem for the coherence parameters
of M.
with probability 1 − L−2 , where μ̃t0 is defined in Theorem 10.2. μr0 is the upper bound
on μ(U ) given in Theorem 10.1.
The above result holds for any precoding matrix P such that the rank of M is K0 .
Proof The theorem can be proved by substituting the bounds on μ(U ) and μ(V ) in
Theorem 10.1 with the bounds derived in Theorem 10.2.
Some comments are in order. First, if K0 is O(ln L), the upper bound μ̃t0 > 1 is a
small constant O(1); this can be seen from the definition of μ̃t0 in (10.13) by choosing
K0 > ln L. Therefore, based on (10.13), the coherence parameter of M is close to 1,
which means that M has good coherence property. A similar bound was provided on
the coherence of the subspaces spanned by a random orthogonal basis in [46]. Second,
unlike the results in [43, theorem 2], the probabilistic bound on μ(V ) is independent
of the target angles and array geometry. Third, the results in Theorem 10.4 hold for
any random unitary matrix S. The radar waveform can be changed periodically, which
would be good for security reason, without affecting the matrix completion perfor-
mance. Fourth, the probabilistic bound on μ(V ) in Theorem 10.2 is independent of P .
This means that we can design P , for example for the purpose of transmit beamforming
and interference suppression, without affecting the incoherence property of M. This key
observation validates the feasibility of precoding for spectrum sharing between MIMO-
MC radar and communication systems, a topic that will be discussed next. Note that this
chapter focuses on the design of radar precoder in the spatial domain, not the waveform.
The radar precoder will not affect the waveform ambiguity properties in the time and
Doppler domains.
Cooperative Spectrum Sharing 293
… …
Collocated MIMO radar
… …
Communication TX Communication RX
Figure 10.2 A MIMO communication system sharing spectrum with a collocated MIMO
radar system.
Figure 10.3 Radar-communication coexistence signal model during one radar PRI.
294 Li and Petropulu
Figure 10.4 The spectrum-sharing architecture. The cooperation is coordinated by the control
center, a node with high computing power that also serves as the radar fusion center. The control
center collects information from radar and communication systems, computes jointly optimal
signaling schemes for both systems and sends each scheme back to the corresponding system.
Transmitted Signals
It is assumed that the two systems transmit narrowband waveforms with the same sym-
bol period. To evaluate the feasibility of radar and communication systems having the
same symbol period, let us consider an S-band search and acquisition radar with range
resolution equal to 300 m (a typical range resolution is between 100 m and 600 m
[51,52]). The corresponding radar sub-pulse duration is 2 μs. Communication symbol
Cooperative Spectrum Sharing 295
duration of 2 μs is quite typical in model cellular systems [53]. The transmitted signal is
narrowband if the channel coherence bandwidth is larger than the signal bandwidth [53–
55]. In a macro-cell, typical values for the channel coherence bandwidth are of the order
of 1 MHz [56,57], which is larger than the signal bandwidth of 0.5 MHz (or symbol
interval 2 μs). Thus, the narrowband assumption is typically valid.
If higher signal bandwidth is needed, OFDM signaling can be used for both radar
[15,17] and the communication system [56,57]. Our coexistence model can still be valid
on each OFDM carrier, over which the signal can be considered as narrowband.
Fading
We assume that H , G1 , and G2 are flat fading, which is valid when the transmitted
signals are narrowband. The flat-fading assumption is common practice in the radar-
communication system coexistence literature [22–26]. In addition, all channels are
assumed to be block fading over the radar PRI. For a radar with medium pulse repetition
frequency, the PRI is usually between 30 μs and 0.3 ms. The typical channel coherence
time for 2.5 GHz and 5.8 GHz carrier frequency ranges from 2 ms up to 200 ms [58].
The channel coherence time is much larger than the radar PRI. As for the moving
targets, the resulting Doppler shifts are usually assumed to be constant during one PRI
[44,59]. Therefore, channel block fading is a reasonable assumption.
Figure 10.5 TDM-based CSI estimation and feedback and reception of design results from the
control center.
Under the aforementioned assumptions on transmitted signals, fading, CSI, and the
radar mode of operation, let us consider a target scene at a particular range bin as in
Section 10.2 but with clutter. The baseband signal received by the radar receivers during
L symbol durations in one radar PRI can be expressed as
$ %
Y R = DP
F GH SI + CP S + G2 X2 + W R . (10.32)
F GH I FGHI
signal interference noise
Y C = FGHI
H X + G1 P S1 + W C . (10.33)
F GH I FGHI
signal interference noise
Y C and W C denote the received signal and additive noise at the communication RX
antennas, respectively. X [x(1),. . .,x(L)] is a matrix whose columns, x(i)’s are code-
words from the codebook of the communication system. W R/C contains i.i.d. random
entries distributed as CN (0,σR/C
2 ). i ,i ∈ {1,2} are diagonal matrices containing the
j α
random phase offset e il between the MIMO-MC radar and the communication system
at the l-th symbol. These phase offsets are time-varying and arise due to the random
phase jitters of the oscillators between the radar transmitter and the communication
receiver [31,71]. Note that the Doppler shift will not be an issue in the design considered
in this chapter. This is because the radar signal model in (10.32) is for the fast-time
samples received in one radar pulse, and during a pulse, the Doppler shift is usually
assumed to be constant [41,44,59,72]. In that case, the Doppler shift can be absorbed
into the target RCS, and does not affect our design.
The control center aims to protect the radar system and maximize the spectrum effi-
ciency. In the following, we present a joint design of the communication and radar
transmissions, which will be implemented at the control center, so that the interference
at the radar RX antennas is minimized, thus allowing for successful matrix completion,
while certain communication system requirements are met [36].
Cooperative Spectrum Sharing 297
In this section, we formulate the MIMO-MC radar and MIMO communication spectrum
sharing problem and present an algorithm to solve it [36].
For the communication system, the covariance of interference plus noise is given by
R Cin = G1 GH
1 + σC I
2
(10.34)
1
L
Cavg ({R xl },) C(R xl ,), (10.36)
L
l=1
Figure 10.6 The sub-sampling at the radar receiver modulates the interference channel from the
communication transmitter to the radar receiver G2 . As shown in the left figure, the null space of
G2 is typically empty; thus the communication system transmission would always introduce
interference to the radar. Due to the random sub-sampling, the null space of the modulated
interference channel G2l shown in the right figure becomes nonempty, and thus, it is possible for
the communication system to introduce zero EIP to the radar receiver if it transmits in the null
space of G2l .
tr (D t )
ESINR = L $ % .
tr (C t ) + l=1 tr G2l R xl G2l /(pLMr,R ) + σR
H 2
One can see that the joint design of the communication TX covariance matrices {R xl },
the radar precoder P (embedded in ), and the radar sub-sampling scheme is neces-
sary to maximize the ESINR. In Theorem 10.4, we prove that the radar precoder P can
be designed without affecting the incoherence property of M.
Cooperative Spectrum Sharing 299
At the control center, the spectrum sharing problem can be formulated as follows:
where V k v∗t (θk )vTt (θk ). The constraint of (10.39a) restricts the communication
rate to be at least C, in order to support reliable communication and avoid service
outage. The constraints of (10.39b) restrict the total communication and radar transmit
power to be no larger than PC and PR , respectively. The constraints of (10.39c) restrict
the power of the radar probing signal in directions of interest to be no smaller than
the power achieved by the uniform precoding matrix trM() t,R
I , i.e., vTt (θk )v∗t (θk ) ≥
ξvTt (θk ) trM()
t,R
I v∗t (θk ) = ξtr (). It holds that ξ ≥ 1, which is used to control the
beampattern at the target angles of interest. The purpose of this constraint is to ensure
fairness across the multiple targets. The constraint in (10.39d) imposes restrictions on
the radar sub-sampling matrix such that it corresponds to a fixed sub-sampling rate p
and has large spectral gap. In the matrix completion literature, is either a uniformly
random sub-sampling matrix [38], or the adjacency matrix of a regular bipartite graph
with large spectral gap [76]. The spectral gap of a matrix is defined as the difference
between the largest and the second largest singular values [76].
In order for the control center to formulate and solve the problem of (10.39) it needs:
(i) the communication and radar system CSI; estimation and feedback of CSI are
discussed in Section 10.3.
(ii) target angles, and clutter parameters {σβ2c } and {θkc }. Since the control center
k
integrates the radar fusion center functionality, the target angles obtained from
the previous tracking cycle will be available. In practice, the clutter parameters
could be estimated when the targets are absent [74]. If the parameters {σβ2c } are
k
not known, we can use a single value, σ02 , for all the targets. With such choice,
the objective treats all target directions equally. One possible choice for σ02 is the
smallest target RCS variance that could be detected by the radar. Note that the
solution of (P1 ) is independent on the specific value of σ02 .
(iii) all parameters in the constraints. Parameters like power budget and required com-
munication rate could be provided by the radar and communication systems.
Problem (P1 ) is nonconvex with respect to optimization variable triplet ({R xl },,).
In Subsection 10.4.1 we present an algorithm to find a local solution via alternating opti-
mization, while in Subsection 10.4.2, we discuss the feasibility and solution properties
of (P1 ) [36].
300 Li and Petropulu
Problem (PR ) is convex and involves multiple matrix variables, the joint optimization
with respect to which requires high computational complexity. The semidefinite matrix
variables {R xl } have LMt,C
2 real scalar variables, which would result in a complexity
solving Problem (PR ) can be implemented based on the Lagrangian dual decomposition
[77]. The Lagrangian of (PR ) can be written as
L
$ %
L({R xl },λ1,λ2 ) = tr G2l R xl GH
2l
l=1
L
$ %
+ λ1 tr (R xl ) − PC + λ2 C − Cavg ({R xl }) ,
l=1
where λ1 ≥ 0 and λ2 ≥ 0 are the dual variables associated with the transmit power and
the communication rate constraints, respectively. The dual problem of (PR ) is
The domain of the dual function, i.e., dom g, is λ1,λ2 ≥ 0 such that g(λ1,λ2 ) > −∞.
The problem is called dual feasible if (λ1,λ2 ) ∈ dom g. The dual function g(λ 1,λ2 )
can be obtained by solving L independent subproblems, each of which can be written
as follows
&& ' '
(PR -sub) min tr GH 2
l G2 + λ 1 I R xl
R xl 0 (10.41)
− λ2 log2 I + R −1
wl H R xl H H
.
Before giving the solution of (PR -sub), let us first state some observations.
Observation 1) The average capacity constraint should be active at the optimal point.
This means that the achieved capacity is always C and λ2 > 0. To show this, let us
Cooperative Spectrum Sharing 301
assume that the optimal point {R ∗xl } achieves Cavg ({R ∗xl }) > C. Then we can always
shrink {R ∗xl } until the average capacity reduces to C, which would also reduce the
objective. Thus, we end up with a contradiction. By complementary slackness, the
corresponding dual variable is positive, i.e., λ2 > 0.
$ % +
Observation 2) GH 2
l G2 + λ 1 I is positive definite for all l ∈ NL . This can be shown
by contradiction. Suppose that there exists l such that G2
l G2 + λ1 I is singular. Then
H
lemma 10.5 [63,64] For given feasible dual variables λ1,λ2 ≥ 0, the optimal solu-
tion of (PR -sub) is given by
−1/2 −1/2
R ∗xl (λ1,λ2 ) = l U l l U H
l l , (10.42)
−1/2 −1/2
where l GH 2
l G2 + λ 1 I ; U l is the right singular matrix of H̃ l R wl H l ;
+
l = diag(βl1,. . .,β lr ) with βli = (λ2 − 1/σli ) , r and σli ,i = 1,. . .,r, respectively,
2
being the rank and the positive singular vales of H̃ l . It also holds that
r & '+
H
log2 I + R −1
wl H R ∗
xl H = log(λ2 σli2 ) . (10.43)
i=1
Based on Lemma 10.5, the solution of (PR ) can be obtained by finding the optimal
dual variables λ∗1,λ∗2 . The cooperative spectrum sharing problem (PR ) can be solved via
the procedure outlined in Algorithm 1. The convergence of Algorithm 1 is guaranteed by
the convergence of the ellipsoid method [78]. The complexity of the dual decomposition
based algorithm is only linearly dependent on L.
l=1
Recall that the sampling matrix is required to have large spectral gap. However, it is
difficult to incorporate such conditions in the optimization problem (10.44). Based on
the fact that row and column permutation of the sampling matrix would not affect its
singular values and thus the spectral gap, a suboptimal approach is to search the best
sampling scheme by permuting rows and columns of an initial sampling matrix 0 , i.e.,
where ℘(0 ) denotes the set of matrices obtained by arbitrary row and/or column per-
mutations. 0 is generated with binary entries and pLMr,R ones, where x denotes
the largest integer smaller or equal to x. Therefore, the constraint on the number of 1’s
in can also be satisfied. One good candidate for 0 would be a uniformly random
sampling matrix, as such matrix exhibits large spectral gap with high probability [76].
Multiple trials with different 0 ’s can be used to further improve the choice of .
However, the search space is very large since the cardinality of ℘(0 ) is Mr,R ! L!,
where operator ! denotes the factorial. One can reduce the search space as follows [36]:
where ℘r (0 ) denotes the set of matrices obtained by arbitrary row permutations. The
search space in (10.46) equals Mr,R !, i.e., the cardinality of ℘r (0 ), which is greatly
reduced compared to that in (10.45). Furthermore, the following proposition shows that
such reduction of search space comes without any performance loss.
proposition 10.6 For any 0 , searching for an in ℘r (0 ) can achieve the same
EIP as searching in ℘(0 ).
Proof We can prove the proposition by showing that the EIP achieved by any 1 ∈
℘(0 ) can also be achieved by a certain 2 ∈ ℘r (0 ). For the pair (1,{R xl }), the
same EIP can be achieved by the pair (2,{R̃ xl }), where
The problem in (10.46) aims to find the best one-to-one match between the rows
of 0 and the rows of Q. Let us construct a cost matrix C r ∈ RMr,R ×Mr,R with
[C r ]ml 0m· (Ql· )T . The problem turns out to be a linear assignment problem with
cost matrix C c , which can be solved efficiently in polynomial time O(Mr,R
3 ) using the
where Al is defined as
T
∂C(R xl ,)
Al −
∂Re() = ¯
= + 2 −1
σC I ) − (G1 GH H −1 ¯.
1 + σC I + H R xl H ) ]G1 =
2
GH H
1 [(G1 G1
(10.48)
The sequential convex programming technique is applied to solve by repeatedly
solving the following approximate optimization problem
tr (D t )
(P ) max ,
0 tr (C t ) + ρ
(10.49)
s.t. tr () ≤ PR /L,tr (A) ≤ C̃,
tr (V k ) ≥ ξtr () ,∀k ∈ N+
K,
where
L
$ %
C̃ = ¯ + tr (A
C(R xl , ) ¯ l) − C ,
l=1
L
A= Al , (10.50)
l=1
1
L
$ %
ρ= 2
l G2 + σR .
tr R xl GH 2
pLMr,R
l=1
max ˜ t ),
tr (D
˜
0,φ>0
˜ t ) = 1 − φρ
s.t. tr (C (10.51)
tr () ˜
˜ ≤ φPR /L,tr (A) ≤ φ C̃,
$ %
˜ k − ξI ) ≥ 0,∀k ∈ N+ .
tr (V K
∗
The optimal solution of (10.51), denoted by (˜ ,φ∗ ), can be obtained by using any
standard interior point method–based SDP solver with a complexity of O((Mt,R 2 )3.5 ).
∗
˜ /φ∗ . In each alternating iteration with respect
The solution of (10.49) is given by
to , it is required to solve several iterations of SDP due to the sequential convex
programming.
It is easy to show that the objective function, i.e., ESINR, is nondecreasing during the
alternating iterations of {R xl }, and , and is upper bounded. According to the mono-
tone convergence theorem [81], the alternating optimization is guaranteed to converge.
The cooperative spectrum sharing algorithm in the presence of clutter maximizing the
effective radar SINR is summarized in Algorithm 2.
Feasibility
A necessary condition on C for the feasibility of (P1 ) with respect to {R xl } is C ≤
Cmax (PC ) where
1
L
Cmax (PC ) max log2 I + σC−2 H R xl H H ,
{R xl }0 L
l=1
(10.52)
L
s.t. tr (R xl ) ≤ PC .
l=1
Cooperative Spectrum Sharing 305
The above optimization problem is convex and has a closed-form solution based on
water-filling [54]. It can be shown that Cmax (PC ) is essentially the largest achievable
communication rate when there is no interference from radar transmitters to the com-
munication receivers. Note that C = Cmax (PC ) will generate a nonempty feasible set
for {R xl } only if G1 GH1 = 0 (omitting the trivial case = 0), i.e., the radar transmits
in the null space of the interference channel G1 to the communication receivers.
A necessary condition on ξ for the feasibility of (P1 ) with respect to is ξ ≤ ξmax ,
where
ξmax max ξ,
0,ξ≥0
(10.53)
s.t. tr (V k ) ≥ ξtr (),∀k ∈ N+
K.
Note that the above optimization problem is independent of tr (). Without loss of
generality, we assume that tr () = 1, based on which we have the following equivalent
SDP formulation
ξmax max ξ, s.t. tr () = 1,
0,ξ≥0
(10.54)
tr (V k ) ≥ ξ,∀k ∈ N+
K.
It is easy to check that ξmax ≥ 1, which can be achieved by set (,ξ) to be (I /Mt,R ,1).
The following proposition provides a sufficient condition for the feasibility of (P1 ).
proposition 10.7 If C,ξ,PC > 0,PR > 0 are chosen such that C < Cmax (PC ) and
ξ ≤ ξmax , then (P1 ) is feasible.
Proof If C < Cmax (PC ), the feasible set for {R xl } determined by constraints
in (10.39a) and (10.39b) F{Rxl } is nonempty as long as tr () is sufficiently small.
If ξ ≤ ξmax , the feasible set for determined by constraints in (10.39c) F1 is
nonempty and has no restriction on tr (). If ∈ F1 , then α ∈ F1,∀α > 0.
The overall feasible set for , F , is the intersection of feasible sets determined
by (10.39a), (10.39b), and (10.39c). F is nonempty as long as F1 and F{Rxl } are
nonempty because we can choose any ∈ F1 and scale it down to make (P1 ) feasible.
The claim is proven.
where γ is a real positive constant. The following proposition relates the optimal solu-
tions of problems (10.49) and (10.55).
306 Li and Petropulu
Proof Denote ∗1 and ∗2 the optimal solutions of (10.49) and (10.55), respectively. It
is clear that ∗1 is a feasible point of (10.55). This means that tr (∗2 ) ≤ tr (∗1 ) ≤ PR .
Therefore, ∗2 is a feasible point of (10.49). It holds that
tr (∗1 D t ) tr (∗2 D t )
SINRmax ≡ ≥ ≥ SINRmax . (10.56)
tr (∗1 C t ) + ρ tr (∗2 C t ) + ρ
It is only possible when all the equalities hold. In other words, ∗2 is optimal for (10.49).
This completes the proof.
In order to characterize the optimal solution of (10.55), we need the following key
lemma:
which could be shown by right multiplying (X + Y ) on both sides of the equality. Since
X, Y , and Z are Hermitian, we have
Since (X + Y )−1 is invertible, there exists a unique positive definite matrix V , such that
(X + Y )−1 = V 2 . Simple algebra manipulation shows that
V −1 ZV −1 = (V −1 X−1 V −1 )(V Y V )
(10.59)
= (V Y V )(V −1 X−1 V −1 ),
Based on Lemma 10.9, we prove the following result by following the approach
in [80]:
proposition 10.10 Suppose that (10.55) is feasible when γ is set to SINRmax . Then,
the following claims hold:
1) Any optimal solution of (10.55) has rank at most K.
2) All rank-K solutions ∗K of (10.55) have the same range space.
Cooperative Spectrum Sharing 307
3) Any solution ∗K − with rank less than K has range space such that Range (∗K − ) ⊂
Range (∗K ).
4) (10.49) and (10.51) always have solutions with rank at most K and with the same
range space properties as that for (10.55).
K
K
+ λ2 Dt + νk V k = I + λ1 A + λ2 γC t + νk ξI (10.60a)
k=1 k=1
= 0 (10.60b)
0, 0,λ1 ≥ 0,λ2 ≥ 0,{νk } ≥ 0 (10.60c)
tr (D t ) ≥ γtr (C t ) + γρ (10.60d)
tr (V k ) ≥ 0,∀k ∈ N+
K, (10.60e)
where 0,λ1 ≥ 0,λ2 ≥ 0, and {νk } ≥ 0 are dual variables. We can rewrite (10.60a)
as follows
K
Rank () + Rank λ2 D t + νk v∗t (θk )vTt (θk )
k=1
(10.61)
K
≥ Rank I + λ1 A + λ2 γC t + νk ξI .
k=1
2 ∗ ∗
Recall that D t = k σβ k vt (θk )vt (θk ). It holds that λ 2 D t +
T T
k νk vt (θk )vt (θk )
has rank at most K. Since A and C t are positive semidefinite, the matrix on right
hand side of (10.61) has full rank. Therefore, Rank () is not smaller than Mt,R − K.
From (10.60b) and (10.60d) we conclude that any optimal solution must have rank at
most K.
The second claim asserts that if there are multiple solutions with rank K, they all have
the same range space. This can be proved using contradiction. Suppose that ∗1 and ∗2
are rank-K solutions of (10.55) and Range (∗1 ) != Range (∗2 ). Based on convex theory,
any convex combination of ∗1 and ∗2 , saying ∗3 α∗1 + (1 − α)∗2,∀α ∈ (0,1), is
also a solution of (10.55). However, ∗3 is with rank at least K + 1, which contradicts
the fact that any solution must have rank at most K. The third claim could also be
proved using contradiction. Suppose that ∗1 and ∗2 are respectively rank-K solution
and solution with rank smaller than K, and Range (∗2 ) \ Range (∗1 ) is nonempty. Then
any convex combination of ∗1 and ∗2 , saying ∗3 α∗1 + (1 − α)∗2,∀α ∈ (0,1),
is also a solution of (10.55). However, ∗3 is again with rank at least K + 1, which
contradicts the fact that any solution must have rank at most K.
The last claim on the solutions of (10.49) and (10.51) follows from Proposition
10.8.
Proposition 10.10 indicates that the rank of the optimal precoding matrix will not be
larger than the number of the targets.
308 Li and Petropulu
where
tr (D t )
ESINR# = $ % (10.63)
2 /(pLMr,R ) + σR
tr (C t ) + tr
G2 R x GH 2
and
= L l=1
l is diagonal and with each entry equal to the number of 1’s in the
corresponding row of . Similar techniques in Algorithm 2 can be used to solve (P#1 ).
We can see that (P#1 ) has much lower complexity because there is only one matrix
variable for the communication transmission. However, the drawback of the constant-
rate communication is that R x cannot adapt to the variation of the effective interference
channel G2l . On the other hand, the adaptive communication transmission considered
in (P1 ) can fully exploit the channel diversity introduced by the radar sub-sampling
procedure.
Another consequence is that the ESINR# depends on only through
. Since
is searched among the row permutations of a uniformly random sampling matrix, the
number of 1’s in each row of is close to pL, or equivalently,
will be very close
to the scaled identity matrix pLI . To further reduce the complexity, the optimization
with respect to in (P#1 ) is omitted because all row permutations of will result in a
very similar ESINR# . From a different perspective, if the radar sub-sampling matrix
is not available for the radar and communication cooperation, we can safely replace
with pLI in the ESINR# . The discussion in this paragraph asserts that, for the case of
constant-rate communication transmission almost no performance degradation occurs
due to the absence of the knowledge of .
the block fading assumption. The spectrum sharing problem has the same form as (P#1 )
with the objective function being
tr (D t )
SINR = $ % . (10.64)
tr (C t ) + tr G2 R x GH2 /Mr,R + σR
2
• The radar effective SINR, i.e., the objective of the spectrum sharing problem;
• The matrix completion relative recovery error, defined as M − M / F /MF ,
/
where M is the completed data matrix at the radar fusion center;
• The radar transmit beampattern, i.e., the transmit power for different azimuth
angles vTt (θ)P v∗t (θ);
• The MUSIC pseudo-spectrum and the relative target RCS estimation RMSE
obtained using the least-squares estimation on the completed data matrix M. /
Monte Carlo simulations with 100 independent trials are carried out to get an average
performance.
Table 10.1 The radar ESINR, MC relative recovery errors, and the relative target
RCS estimation RMSE for MIMO-MC radar and communication spectrum sharing.
40
20
–20
–50 0 50
Azimuth Angle
0
–5
Spatial Spectrum in dB
–10
–15
–20
–25
–30
–50 0 50
Azimuth Angle
Figure 10.7 The radar transmit beampattern and the MUSIC spatial pseudo-spectrum for
MIMO-MC radar and communication spectrum sharing. The true positions of the targets and
clutters are labeled using solid and dashed vertical lines, respectively.
probability of missing the second target is increased. We should note that neither the
uniform nor the NSP precoding schemes have any capability of clutter mitigation. From
Figure 10.7, we observe that the jointly designed precoding scheme successfully focuses
the transmit power toward the three targets and nullifies the power toward the point
scatterers. The three targets can be accurately estimated from the pseudo-spectrum
obtained by the joint design. Meanwhile, the communication system can still achieve
312 Li and Petropulu
the required rate by aligning its transmission along a channel subspace that does not
interfere with the radar emissions. This significant advantage is enabled by the joint
design of radar and communication transmissions.
1
MC Relative Recovery Error
30
0.8
20 Uniform Precoding + Selfish Comm. Uniform Precoding + Selfish Comm.
ESINR in dB
0.2
–10
1
Relative RCS Estimation RMSE
0.8
Uniform Precoding + Selfish Comm.
NSP Precoding + Selfish Comm.
0.6
Uniform Precoding + Design R xl&W
Design P + Selfish Comm.
0.4 Jointly Designed Precoding
0.2
Figure 10.8 Comparison of spectrum sharing with different levels of cooperation between the
MIMO-MC radar and the communication system under different PR .
Cooperative Spectrum Sharing 313
the achieved ESINR, the MC relative recovery error, and the relative target RCS esti-
mation RMSE. The algorithms that use trivial uniform and NSP-based radar precoding
perform poorly because the point scatterers are not properly mitigated. The scheme that
designs P only could mitigate the scatterers but the interference from the communica-
tion transmission is not controlled. The joint design of P , R xl , and simultaneously
addresses the clutter and the mutual interference between the radar and the communi-
cation systems, and thus achieves the best performance amongst all the algorithms. The
performance gains come from high-level cooperation between the two systems.
RCS estimation RMSE for different values of σG 2 are shown in Figure 10.9. The value
2
of σG2 varies from 0.05 to 0.5, which effectively simulates different distances between
2
the communication transmitter and the radar receiver. It is clear that the adaptive
communication transmission outperforms the constant-rate counterpart under various
values of interference channel strength. As discussed in Section 10.4.3, the adaptive
Figure 10.9 Comparison of spectrum sharing with adaptive and constant-rate communication
transmissions under different levels of variance of the interference channel from the
communication transmitter to the radar receiver.
314 Li and Petropulu
communication transmission can fully exploit the channel diversity of G2l introduced
by the radar sub-sampling procedure. The price for the performance advantages is high
complexity. The average running times on a laptop with Intel Core i5 dual-core 2.4 GHz
CPU for the adaptive and constant-rate communication transmissions are respectively
15.6 and 4.8 seconds. The choice between these two transmission schemes can be made
depending on the available computing resources.
Figure 10.10 Comparison of spectrum sharing with traditional MIMO radars and MIMO-MC
radars with different sub-sampling rates p.
Cooperative Spectrum Sharing 315
interfering the radar system when the number of actively sampled radar RX antennas
is small, i.e., sub-sampling is small. In addition, the larger ESINR of the MIMO-MC
radar results in a larger output SINR than that of the traditional radar. Furthermore, the
MIMO-MC radar achieves better target RCS estimation accuracy than the traditional
radar if its sub-sampling rate is between 0.4 and 0.7. For p larger than 0.7, the target
RCS estimation accuracy achieved by the MIMO-MC radar is worse than that achieved
by the traditional radar because small ESINRs for p ≥ 0.7 introduce high distortion
in the completed data matrix. The results in Figure 10.10 could be used to help the
selection of radar sub-sampling rate p. For the best target RCS estimation accuracy,
p = 0.6 is the best choice, while for the biggest savings in terms of samples and
similar performance as traditional radars, p = 0.4 is the best choice. Since there is no
closed-form solution for the joint design problem, it is difficult to provide a theoretical
justification.
Based on these results, we conclude that MIMO-MC radars can coexist with commu-
nication systems and achieve better target RCS estimation than traditional radars, while
saving up to 60% of data samples. Such significant advantage is introduced by sparse
sensing (i.e., sub-sampling) in MIMO-MC radars, as discussed in Section 10.4.4.
10.6 Conclusions
In this chapter, we have considered the coexistence of a MIMO-MC radar and a wire-
less MIMO communication system by sharing a common carrier frequency. The radar
transmits random unitary waveforms, and both radar and communication systems use
precoders. The precoders and the radar sub-sampling scheme have been jointly designed
by the control center to maximize the radar SINR while meeting certain rate and power
constraints for the communication system. Random unitary waveforms can be easily
generated and updated for waveform security. We should note that the presented joint
design–based spectrum sharing method can also be applied to traditional MIMO radars,
which is a special case of MIMO-MC radars with 100% sub-sampling rate, i.e., p = 1.
The jointly designed spectrum sharing scheme has been evaluated via extensive sim-
ulations. Specifically, we have shown that cooperative design brings a significant perfor-
mance advantage as compared to noncooperative design. The jointly designed spectrum
sharing scheme successfully focuses the transmit power towards the targets and nulli-
fies the power towards the clutter. It achieves significant improvement in ESINR, MC
relative recovery error, and target RCS estimation accuracy. We have also compared
the performance and complexity of the adaptive and the constant-rate communication
transmission schemes for radar-communication spectrum sharing. Finally, we have pro-
vided simulation-based comparison of MIMO-MC radars and traditional MIMO radars
coexisting with communication systems. We have observed that the MIMO-MC radar
achieves better performance in terms of ESINR and output SNR. Our simulations sug-
gest that MIMO-MC radars can coexist with communication systems and achieve better
target RCS estimation than traditional radars, while saving up to 60% in data samples.
Of course these advantages come at increased computations for MC.
316 Li and Petropulu
We should note that the constraint requiring that the number of targets is smaller
than the number of radar antennas results in an inefficient usage of the MIMO radar
degrees of freedom. However, the high resolution of traditional MIMO radar is retained
by MIMO-MC radars with a great reduction of sample and hardware complexity. The
considered signal model is for narrow-band waveforms. Broadband MIMO systems
typically use OFDM waveforms [16]. In such case, the joint design still applies on indi-
vidual component carriers. This would substantially expand the application scenarios of
the results presented in this chapter.
References
[1] “Realizing the full potential of government-held spectrum to spur economic growth,” The
Presidents Council of Advisors on Science and Technology (PCAST), technical report, July
2012. [Online]. www.dtic.mil/dtic/tr/fulltext/u2/a565091.pdf.
[2] D. Cabric, I. D. O’Donnell, M. S. Chen, and R. W. Brodersen, “Spectrum sharing radios,”
IEEE Circuits and Systems Magazine, vol. 6, no. 2, pp. 30–45, 2006.
[3] “FCC proposes innovative small cell use in 3.5 GHz band,” Federal Communications
Commission (FCC), news release, December 2012. [Online]. apps.fcc.gov/edocs_public/
attachmatch/DOC-317911A1.pdf.
[4] G. Locke and L. E. Strickling, “An assessment of the near-term viability of accommodating
wireless broadband systems in the 1675–1710 MHz, 1755–1780 MHz, 3500–3650 MHz,
and 4200–4220 MHz, 4380–4400 MHz bands,” US Dept. of Commerce, the National
Telecommunications and Information Administration, technical report TR-13-490, 2012.
[5] E. Drocella, J. Richards, R. Sole, F. Najmy, A. Lundy, and P. McKenna, “3.5 GHz exclusion
zone analyses and methodology,” US Dept. of Commerce, the National Telecommunications
and Information Administration, technical report TR-15-517, 2015.
[6] F. H. Sanders, R. L. Sole, B. L. Bedford, D. Franc, and T. Pawlowitz, “Effects of RF
interference on radar receivers,” US Dept. of Commerce, the National Telecommunications
and Information Administration, technical report TR-06-444, 2006.
[7] A. Lackpour, M. Luddy, and J. Winters, “Overview of interference mitigation techniques
between WiMAX networks and ground based radar,” in 20th Annual Wireless and Optical
Communications Conference, April 2011, pp. 1–5.
[8] F. H. Sanders, R. L. Sole, J. E. Carroll, G. S. Secrest, and T. L. Allmon, “Analysis and res-
olution of RF interference to radars operating in the band 2700–2900 MHz from broadband
communication transmitters,” US Dept. of Commerce, the National Telecommunications
and Information Administration, technical report TR-13-490, 2012.
[9] M. R. Bell, N. Devroye, D. Erricolo, T. Koduri, S. Rao, and D. Tuninetti, “Results on
spectrum sharing between a radar and a communications system,” in 2014 International
Conference on Electromagnetics in Advanced Applications (ICEAA), 2014, pp. 826–829.
[10] Q. Zhao and B. M. Sadler, “A survey of dynamic spectrum access,” IEEE Signal Processing
Magazine, vol. 24, no. 3, pp. 79–89, May 2007.
[11] E. Hossain, D. Niyato, and Z. Han, Dynamic Spectrum Access and Management in Cognitive
Radio Networks. Cambridge University Press, 2009.
[12] L. S. Wang, J. P. McGeehan, C. Williams, and A. Doufexi, “Application of cooperative
sensing in radar-communications coexistence,” IET Communications, vol. 2, no. 6, pp. 856–
868, July 2008.
Cooperative Spectrum Sharing 317
[29] A. Hassanien, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Signaling strategies for dual-
function radar-communications: An overview,” IEEE Aerospace and Electronic Systems
Magazine, vol. 31, no. 10, pp. 36–45, October 2016.
[30] B. Li and A. P. Petropulu, “Spectrum sharing between matrix completion based MIMO
radars and a MIMO communication system,” in IEEE International Conference on Acous-
tics, Speech and Signal Processing, April 2015, pp. 2444–2448.
[31] B. Li, A. P. Petropulu, and W. Trappe, “Optimum co-design for spectrum sharing between
matrix completion based MIMO radars and a MIMO communication system,” IEEE
Transactions on Signal Processing, vol. 64, no. 17, pp. 4562–4575, September 2016.
[32] B. Li and A. P. Petropulu, “Radar precoding for spectrum sharing between matrix comple-
tion based MIMO radars and a MIMO communication system,” in IEEE Global Conference
on Signal and Information Processing, December 2015, pp. 737–741.
[33] B. Li, H. Kumar, and A. P. Petropulu, “A joint design approach for spectrum sharing between
radar and communication systems,” in IEEE International Conference on Acoustics, Speech
and Signal Processing, March 2016, pp. 3306–3310.
[34] B. Li and A. P. Petropulu, “MIMO radar and communication spectrum sharing with clutter
mitigation,” in IEEE Radar Conference, May 2016, pp. 1–6.
[35] B. Li and A. P. Petropulu, “Matrix completion based MIMO radars with clutter and
interference mitigation via transmit precoding,” in 2017 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), March 2017, pp. 3216–3220.
[36] B. Li and A. P. Petropulu, “Joint transmit designs for coexistence of MIMO wireless
communications and sparse sensing radars in clutter,” IEEE Transactions on Aerospace and
Electronic Systems, vol. 53, no. 6, pp. 2846–2864, December 2017.
[37] S. Sun, W. Bajwa, and A. P. Petropulu, “MIMO-MC radar: A MIMO radar approach based
on matrix completion,” IEEE Transactions on Aerospace and Electronic Systems, vol. 51,
no. 3, pp. 1839–1852, July 2015.
[38] E. J. Candès and Y. Plan, “Matrix completion with noise,” Proceedings of the IEEE, vol. 98,
no. 6, pp. 925–936, June 2010.
[39] D. S. Kalogerias and A. P. Petropulu, “Matrix completion in colocated MIMO radar: Recov-
erability, bounds and theoretical guarantees,” IEEE Transactions on Signal Processing,
vol. 62, no. 2, pp. 309–321, Jan 2014.
[40] C. Chen and P. P. Vaidyanathan, “Compressed sensing in MIMO radar,” in Asilomar
Conference on Signals, Systems and Computers, 2008, pp. 41–44.
[41] Y. Yu, A. P. Petropulu, and H. V. Poor, “MIMO radar using compressive sampling,” IEEE
Journal of Selected Topics in Signal Processing, vol. 4, no. 1, pp. 146–163, February 2010.
[42] M. G. Amin, Compressive Sensing for Urban Radar. CRC Press, 2014.
[43] S. Sun and A. P. Petropulu, “Waveform design for MIMO radars with matrix completion,”
IEEE Journal of Selected Topics in Signal Processing, vol. 9, no. 8, pp. 1400–1414,
December 2015.
[44] M. A. Richards, Fundamentals of Radar Signal Processing. McGraw-Hill, 2005.
[45] H. Krim and M. Viberg, “Two decades of array signal processing research: The parametric
approach,” IEEE Signal Processing Magazine, vol. 13, no. 4, pp. 67–94, 1996.
[46] E. J. Candès and B. Recht, “Exact matrix completion via convex optimization,” Foundations
of Computational Mathematics, vol. 9, no. 6, pp. 717–772, 2009.
[47] K. Zyczkowski and M. Kus, “Random unitary matrices,” Journal of Physics A: Mathemati-
cal and General, vol. 27, no. 12, p. 4235, 1994.
Cooperative Spectrum Sharing 319
[68] X. Hou and C. Yang, “How much feedback overhead is required for base station cooperative
transmission to outperform non-cooperative transmission?” in 2011 IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP), 2011, pp. 3416–3419.
[69] P. Stoica, J. Li, and Y. Xie, “On probing signal design for MIMO radar,” IEEE Transactions
on Signal Processing, vol. 55, no. 8, pp. 4151–4161, 2007.
[70] G. Cui, H. Li, and M. Rangaswamy, “MIMO radar waveform design with constant modulus
and similarity constraints,” IEEE Transactions on Signal Processing, vol. 62, no. 2, pp. 343–
353, 2014.
[71] R. Mudumbai, G. Barriac, and U. Madhow, “On the feasibility of distributed beamforming
in wireless networks,” IEEE Transactions on Wireless Communications, vol. 6, no. 5,
pp. 1754–1763, 2007.
[72] C. Chen and P. P. Vaidyanathan, “MIMO radar ambiguity properties and optimization using
frequency-hopping waveforms,” IEEE Transactions on Signal Processing, vol. 56, no. 12,
pp. 5926–5936, 2008.
[73] S. N. Diggavi and T. M. Cover, “The worst additive noise under a covariance constraint,”
IEEE Transactions on Information Theory, vol. 47, no. 7, pp. 3072–3081, November 2001.
[74] Z. Chen, H. Li, G. Cui, and M. Rangaswamy, “Adaptive transmit and receive beamforming
for interference mitigation,” IEEE Signal Processing Letters, vol. 21, no. 2, pp. 235–239,
February 2014.
[75] C. Chen and P. P. Vaidyanathan, MIMO Radar Spacetime Adaptive Processing and Signal
Design. John Wiley & Sons, 2008, pp. 235–281.
[76] S. Bhojanapalli and P. Jain, “Universal matrix completion,” in Proceedings of The 31st
International Conference on Machine Learning, 2014, pp. 1881–1889.
[77] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.
[78] R. G. Bland, D. Goldfarb, and M. J. Todd, “The ellipsoid method: A survey,” Operations
Research, vol. 29, no. 6, pp. 1039–1091, 1981.
[79] H. W. Kuhn, “The Hungarian method for the assignment problem,” Naval Research Logistics
Quarterly, vol. 2, no. 1-2, pp. 83–97, 1955.
[80] Q. Li and W.-K. Ma, “Optimal and robust transmit designs for MISO channel secrecy
by semidefinite programming,” IEEE Transactions on Signal Processing, vol. 59, no. 8,
pp. 3799–3812, 2011.
[81] J. Yeh, Real Analysis: Theory of Measure and Integration. World Scientific, 2006.
[82] “3GPP TS LTE evolved universal terrestrial radio access (E-UTRA) physical layer proce-
dures,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 36.213 V8.0,
2009.
[83] M. T. Kawser, B. Hamid, N. Hasan, M. S. Alam, and M. M. Rahman, “Downlink SNR to
CQI mapping for different multiple antenna techniques in LTE,” International Journal of
Information and Electronics Engineering, vol. 2, no. 5, p. 757, 2012.
11 Compressed Sensing Methods for
Radar Imaging in the Presence of
Phase Errors and Moving Objects
Ahmed Shaharyar Khwaja,∗ Naime Ozben Onhon, and Mujdat Cetin
Compressed sensing (CS) is a useful tool for processing sparse signals, i.e., signals
that have a few large values and many small or zero values. In [1], the authors state
that if an object has a sparse representation in a basis, a lesser number of nonadaptive
measurements contains enough information to reconstruct the object via basis pursuit.
A dictionary for sparse representation can be implicitly defined through a transform or
explicitly defined through a collection of signals that can represent an object or a signal
in a compressed form, e.g., a piece-wise smooth object can be represented by only a few
coefficients after taking a wavelet transform, or a narrow-band signal can be described
in terms of a few entries of a dictionary consisting of exponential signals with varying
frequencies. In the former case, CS reconstruction can be described as a dictionary-less
approach, whereas in the latter case, it can be described as a dictionary-based approach.
In [2], the authors explain that a sparse or piece-wise constant signal can be recovered
from a random number of its partial Fourier coefficients. Candès and Tao show in [3] that
a measured coded signal corrupted by errors can be recovered using linear programming,
provided the signal is sparse and there are certain constraints on the coding matrix.
References [4,5] provide a review of CS, as well as identify potential application areas
for data compression, channel coding, inverse problems, data acquisition, single-pixel
imaging, etc.
The purpose of radar imaging is to extract an underlying signal present in the elec-
tromagnetic data received by an antenna. The underlying signal can be the image of an
area or of an object being observed by the antenna. In many cases, these signals are
sparse, thus making them suitable for applications of CS. References [6] introduces a
new approach to radar imaging based on CS where a low-dimensional, nonadaptive,
linear projection is used to acquire an efficient representation of a compressible signal
directly, using just a few measurements. The use of this approach can result in the
design of new radar systems, where a matched filter is not needed at the receiver and
the analog-to-digital conversion bandwidth can be reduced. Reference [7] demonstrates
the advantage of CS for radar imaging by achieving higher resolution and artifact-free
∗ Ahmed Shaharyar Khwaja is supported by The Scientific and Technological Research Council of Turkey
(TUBITAK) - 2236 Co-Funded Brain circulation grant.
321
322 Khwaja, Onhon, and Cetin
images from randomly undersampled data. A new radar imaging system based on CS is
presented in [8].
In this chapter, we present CS and its applications in radar imaging, specifically
synthetic aperture radar (SAR) and inverse synthetic aperture radar (ISAR) imaging,
focusing on scenarios involving unknown motion. This chapter is organized as follows:
• We first provide the relevant mathematical expressions for CS and SAR, and then
formulate the problem of CS SAR imaging.
• Thereafter, we consider the case where there are unknown motion errors present
during the SAR acquisition process. The compensation of these errors blindly is
called autofocus. We formulate the problem, present a general autofocus-based
CS solution, and then review existing literature on this topic.
• In the next section, we formulate the problem of SAR moving target imaging,
discuss the types of existing CS-based solutions and present a survey of existing
methods in the literature.
• In the next section, we present CS ISAR imaging, followed by a literature review
of this topic.
We would like to mention that in this chapter, all equations that represent functions
with arguments show matrices or vectors obtained using each combination of values of
their arguments, e.g., S(k,y) = exp{−j 2ky} represents a function. The arguments of
this function are vectors k and y having lengths equal to Nt and Ny , respectively. This
function results in the creation of a two-dimensional (2D) matrix S whose number of
rows and columns are equal to Ny and Nt , respectively. The qth row of S is obtained
by using all entries in k and the qth entry of y, i.e., exp{−j 2kyq }. All other equations
not representing functions with arguments, e.g., z = α are obtained by following the
normal matrix multiplication rules.
Compressed sensing has been shown to be a useful tool in applications of radar imaging
[9,10]. Reference [11] demonstrates the advantage of CS for achieving higher resolution
compared to traditional processing techniques. In [12] and [13], the authors validate the
applications of CS with real data by showing that even after many missing samples in the
raw data, an image can still be reconstructed without any loss of resolution. Examples of
traditional reconstruction with full and limited data, and CS-based reconstruction using
limited data, are shown in Figure 11.1. These examples have been taken from [9]. It can
be observed that CS-based reconstruction can generate a high resolution image even
from limited data.
The concept of CS states that a sparse unknown signal can be recovered from incom-
plete sets of linear measurements by a specifically designed nonlinear recovery algo-
rithm, offering the possibility of signal compression, hardware simplification, as well
as emphasizing certain features in an image [14]. The CS theory mainly involves for-
mulating a system model first and, subsequently, the incorporation of the model into a
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 323
Figure 11.1 The reconstructions of the Slicy target from the MSTAR data set. (a) The reference
image reconstructed from high-bandwidth data. (b) The conventional image reconstructed from
limited-bandwidth data. (c) Image reconstructed from limited-bandwidth data using compressed
sensing. (Taken from [9] with permission.)
y = Aα + (11.1)
or
y = α + , (11.2)
where denotes noise and = A. If the number of measurements is less than the
number of unknown variables, the recovery of α from y is hampered by the fact that
there can be many solutions to this under-determined problem. However, if satisfies
a property known as the restricted isometric property (RIP), α can be recovered by
solving the following problem:
The RIP means that the columns of should not be very similar, as this would ensure
a sufficient number of linearly-independent measurements. The above problem is a
combinatorial optimization problem and is solved by either using a relaxed version
324 Khwaja, Onhon, and Cetin
of the problem that replaces the l0 -norm with an lp -norm, e.g., with p = 1, or using
greedy algorithms. The first term in (11.3) represents the closeness of the solution to the
observations, while the second term represents a priori sparse information. The term λ
is used to balance these two terms and is known as a hyper-parameter. Its choice can
influence the solution. Note that (11.3) can also be obtained by assuming a Laplacian
prior for α and Gaussian distribution for the noise, and subsequently solving for α using
maximum a posteriori estimation.
This pulse has a length Tp , a chirp-rate K, and t is the across-track time sampled at
a frequency ft , i.e., t = 0, f1t , f2t ,. . ., Nft −1
t
with Nt being the total number of across-
track time samples. The received pulse is a delayed version of the transmitted pulse and
this delay is given by the two-way distance between the antenna and the scatterer. The
total received raw data are given by an integration of the received pulse over the whole
scene. However, as a discretized scene is considered, the raw data are considered as a
sum of signals received from each of the scatterers observed by the antenna at different
positions in the azimuth direction, and can be described by the following equation after
demodulation:
N
S(t,y) = σ(rn,yn )Pn (t,y),N = nr × ny , (11.5)
n=1
where
N
(r,y) = σ(rn,yn )δ(r − rn,y − yn ), (11.8)
n=1
where δ(.,.) represent a 2D Dirac pulse. This reflectivity map represents the intensity
of
each scatterer in an observed
scene. The sensor-target distance is given as dn (y) =
rn + (y − yn ) , where rn = xn2 + H 2 is the slant-range, xn is the ground-range,
2 2
and H is the height of the SAR antenna above the flat scene. Different SAR processing
algorithms, e.g., the wavefront reconstruction algorithm [17], the omega-k algorithm
[18], range-Doppler algorithm [19], or the chirp scaling algorithm [20] can be used to
generate images from the raw data.
The expression for raw data in a one-dimensional (1D) wavenumber domain calcu-
lated using the principle of stationary phase (POSP) [19] is given as
.
S(k,y) = p(
k) σ(rn,yn ) exp −j 2k rn2 + (y − yn )2 , (11.9)
n
where p(
k) is the Fourier transform (FT) of p(t),
k is the range wavenumber, i.e.,
k = {−Nt /2×2πft /(Nt c),(−Nt /2+1)×2πft /(Nt c),. . .,(Nt /2−1)×2πft /(Nt c)},
k = kc +
k, and kc = 2πfc /c.
where σn = σ(rn,yn ),n = {1,2,. . .,N}, we can write the received data s as a column of
size Nt Ny × 1 as follows:
s = σ + , (11.13)
where a nonzero value of the reflectivity leads to the selection of raw data of a column
from . This selected column corresponds to the position of the nonzero value of the
reflectivity.
In case of data loss or undersampling in the azimuth and/or range, we can consider
different rows or columns to be missing in , thus we can rewrite (11.13) as
s = σ + , (11.14)
326 Khwaja, Onhon, and Cetin
Figure 11.3 SAR images of a vehicle from 360◦ aperture. (a) Standard Fourier image.
Compressed sensing-based solution using lp -norm instead of l0 -norm in (11.3), where
(b) p = 1 and (c) p = 0.8. (Taken from [10] with permission.)
Figure 11.4 Original resolution: 0.3 m: (a) Conventional image. (b) Compressed sensing
reconstruction. Original resolution: 0.6 m: (c) Conventional image. (d) Compressed sensing
reconstruction. (Taken from [14] with permission)
standard Nyquist sampling. The authors further demonstrate that the data obtained
by these patterns can be used for successful CS SAR reconstruction.
The conditions for CS reconstruction can be satisfied for the SAR imaging problem
when there are a few strong scatterers embedded in a weak background. In this case,
the reflectivity reconstruction can be carried out by solving the following minimization
problem:
σ̂ = argmins − σ22 + λσ0 . (11.15)
σ
This equation can be solved using different recovery methods: linear programming [2],
orthogonal matching pursuit (OMP) [24], iterative shrinkage/thresholding [25], modi-
fied quasi-Newton method [14], etc.
The reconstructed reflectivity vector σ̂ is then reshaped into a 2D form as follows:
ˆ
(r,y) = ξ −1 ( σ̂), (11.16)
which shows the estimated value of the reflectivity for different range and azimuth
positions. This estimation does not suffer from artifacts or reduction in resolution, which
may arise if the undersampled data are used for reconstruction using traditional SAR
imaging techniques. Examples of CS reconstruction are shown in Figure 11.4, taken
from [14].
328 Khwaja, Onhon, and Cetin
The presence of superscript n in ren (y) indicates that these errors are dependent on the
range position of a scatterer, whereas the absence of n in ye (y) means that these errors
are only dependent on the aperture positions. Please note that this model does not make
a narrow-beamwidth assumption, but considers the variation of motion errors within the
aperture for each point that is being imaged. Now the expression of raw data in the 1D
wavenumber domain can be written as
.
Sn (k,y) = p(
k)σ(rn,yn ) exp(−j 2k (rn2 + (y − yn )2 )
Sen (k,y), (11.18)
where
. .
Sen (k,y) = exp −j 2k n
(rn − re (y)) − (y − yn − ye (y)) + rn + (y − yn )
2 2 2 2
(11.19)
is the error term due to the residual motion errors.
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 329
Considering the case where the residual motion errors do not create any significant
major shift in the range position, we can write the motion errors as
. .
n n
se (kc,y) = exp −j 2kc (rn − re (y)) + (y − yn − ye (y)) − rn + (y − yn )
2 2 2 2
(11.20)
and the raw data can be written as
$ %
Sen (t,y) = σ(rn,yn )p t − 2den (y)/c
sen (kc,y), (11.21)
which differs from the raw data in case of no motion errors by the term
sen (kc,y).
where now the solution consists of estimating the reflectivity and motion errors for the
nth target.
Equation (11.24) can be solved using the method called sparsity-driven autofocus
(SDA) proposed in [33]. This method solves for joint SAR imaging and phase-error
correction. Phase errors are incorporated in the problem as model errors, and phase error
correction is performed during the image formation process. Example results from [33]
are shown in Figure 11.6. The proposed method handles the problem as an optimization
problem in which the cost function is composed of a data fidelity term that depends
on the phase errors and a regularization term, which is the l1 -norm of the reflectivity.
330 Khwaja, Onhon, and Cetin
The given cost function is minimized jointly with respect to the reflectivity and the
phase error using coordinate descent technique. The algorithm is an iterative two-step
algorithm, which cycles through steps of image formation and phase error estimation
and compensation.
1. In the first step, the cost function is minimized with respect to the reflectivity as
follows:
σ̂n = argmins̃n − A
n σ22 + λσ1 . (11.25)
σ,
n
2. In the second step, the estimated reflectivity is used to solve for the unknown
motion errors.
ˆ n = argmins̃n − A
n σ22 .
(11.26)
n
The above two steps are repeated until there is no significant change in the values of the
estimates. The solution from this method is dependant on an appropriate choice of the
variable λ. As the method does not require creating a dictionary of phase errors, it can
be seen as a dictionary-less approach.
Consider another case where the motion errors are not range- and azimuth position–
dependent, but only depend on the aperture position, i.e., the trajectory errors are given
only by ye (y). Given these motion phase errors as φe (yq ) for the qth aperture position,
the problem can be formulated as
s = A
σ + , (11.27)
where
= diag[
se (kc,y),
se (kc,y),. . .,
se (kc,y)] (11.28)
is a vector of size 1 × Ny . Figure 11.5 shows the system model for CS SAR autofocus.
The figure shows that in the presence of trajectory deviations, the received data have a
different form compared to the case of no trajectory deviations. This difference comes
from a “perturbing” term
. This term is a result of radar-target distance that is
different in the presence of trajectory deviations compared to the radar-target distance
in the absence of trajectory deviations.
In this case, the solution can be written as follows:
ˆ = argmins − A
σ22 + λσ1 .
σ̂,
(11.30)
σ,
This solution does not require making any sub-patches and can be obtained using the
SDA method. Besides the SDA method, other methods have also been proposed to solve
the CS SAR autofocus problem. A brief overview of some of these references is given
in the following paragraphs.
In [34], the authors propose a sparsity-based SAR autofocus (SBA) method that
corrects phase errors within the image reconstruction process like the SDA. Actually,
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 331
Azimuth
y
Ψ
den (Y) : radar-target distance
S = AΔΨΨσ A/D
t
Range
Figure 11.6 Images reconstructed by (a) conventional imaging method, (b) sparsity-driven
imaging method, (c) sparsity driven autofocus method. The reconstructed image obtained using
sparsity-driven autofocus method is focused. This can be seen by highly localized four points in
the center of the figure. The images obtained using conventional imaging and sparsity-driven
imaging methods are defocused, as seen by multiple artifacts apparent throughout the
corresponding figures. (Taken from [33] with permission)
both methods aim to solve approximately the same problem within a block relaxation-
based framework. Different from the SDA, the proposed method incorporates an addi-
tional surrogate parameter of the SAR reflectivity field into the optimization problem to
make the algorithm stable and to guarantee the convergence to an accumulation point
or a connected set of accumulation points. Experimentally, it is also shown that this
formulation results in a faster convergence.
Reference [35] proposes a sparsity-based SAR imaging algorithm, called perturbed
autofocus SAR (PA-SAR), which jointly solves for autofocus and off-grid target errors,
i.e., for scatterers that are not on the discrete grid. The proposed algorithm uses a vari-
ant of the OMP algorithm called a parameter perturbation-based orthogonal matching
pursuit (PPOMP) algorithm for efficient estimation of off-grid scatterer locations. The
location of a target, which is not exactly on the grid, is described using the closest grid
node coordinates with an additive unknown perturbation. Since off-grid targets with
strong reflectivities may adversely effect the reconstruction of neighboring targets with
weaker reflectivities, the off-grid oriented structure of PA-SAR provides an advantage
in terms of the resolvability of these targets.
332 Khwaja, Onhon, and Cetin
Reference [36] uses a different sparse reconstruction approach called the expectation
maximization-based matching pursuit (EMMP) algorithm [37] for solving the SAR
autofocus problem. The EMMP algorithm treats the compressive measurements as
incomplete data and constructs through iterative expectation and maximization (EM)
steps the complete data corresponding to a set of SAR data for each strong target. The
EM iterations provide more accurate and efficient estimation of the individual target
parameters as well as enable the estimation of unknown phases for each complete
data component. In conclusion, the proposed EMMP-based SAR imaging algorithm
is described as being more greedy, computationally less complex, and having lower
reconstruction errors compared to l1 -norm minimization.
In [38], the authors compare the results of SDA [33], SBA [34], [36] EMMP with a
parameter-free variant of EMMP called the OMP-based autofocus algorithm (AOMP),
in terms of mean-square error (MSE), entropy, target-to-background ratio (TBR), and
signal-to-noise ratio (SNR). Furthermore, these sparsity-based techniques are compared
to the well-known PGA [30]. Comparison results show that all of the sparsity-based
techniques form focused and sparse images with some slight qualitative variations.
The performance of these techniques depends highly on the selection of the hyper-
parameters. Moreover it is shown that, in terms of phase error estimation performance,
SBA, SDA, and AOMP work better compared to the PGA and EMMP. As expected,
considering the run time, the performance of the PGA is much faster than the sparsity-
based techniques.
Reference [39] presents a method called autofocusing iteratively re-weighted aug-
mented Lagrangian method (AIRWALM), which is based on an iteratively re-weighted
augmented Lagrangian method (IRWALM) [40] and a sparsity-driven autofocus (SDA)
method [33]. The proposed method optimizes over the reflectivities and phase errors
jointly to solve a constrained formulation of the sparsity driven autofocus problem
with an lp -norm, p ≤ 1 cost function. Instead of solving the unconstrained problem
in (11.30), the AIRWALM solves the following problem:
p
argmin σp
σ,
s.t. s − A
σ2 ≤ . (11.31)
The motivation for this problem formulation is that it is easier to determine the
error bound , rather than choosing a regularization parameter. Moreover, the use
of p-norms further enhances the sparsity, which may result in a better phase error
estimation. This formulation offers a reduced computation time compared to the SDA
algorithm.
In [41], the authors propose a novel signal processing algorithm for joint SAR image
formation and autofocus in a synthesis dictionary-based sparse representation frame-
work. The proposed algorithm can be applied broadly to scenes that exhibit sparsity
with respect to any dictionary. This is done by extending the SBA imaging framework
from [34] to joint SAR image formation and autofocus. Phase error vector is estimated
using a MAP estimator and compensated through an iterative algorithm to produce
focused images.
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 333
The work in [42] uses a parametric sparse representation to compensate for platform
motion errors in SAR that improves the imaging quality compared to other autofocus
methods. The imaging consists of estimating the reflectivities of the scatterers, followed
by calculating azimuth velocity errors and range acceleration errors.
Compressed sensing was initially applied to scenes that were assumed to contain only
static targets. However, an observed scene can have moving targets such as vehicles,
boats, etc. For this case assuming the absence of moving targets can lead to the appear-
ance of artifacts in the processing results, similar to the case when a scene was processed
without taking into account residual motion errors. Therefore, CS SAR processing has to
be modified to take into account moving targets. In the following sections, we describe
the moving target imaging problem, give details about CS SAR processing for moving
targets, and review existing literature on this topic.
• A shift in azimuth position due to the range velocity vnr creating an azimuth
vn
wavenumber shift of 2kc Vr .
• A blurring effect caused by the azimuth velocity vny .
These effects can be further seen in Fig. 11.7, taken from [43], where two moving
vehicles are shown. One of the vehicles moving in range direction is shifted from its
original position on the road, while the second vehicle moving in azimuth direction is
blurred.
334 Khwaja, Onhon, and Cetin
Figure 11.7 Effects of target motion on processed image. Image courtesy Artemis, Inc.
For multiple moving targets, the expression for raw data is given as a weighted sum
of data from individual targets as follows:
Nm
Sm (t,y) = σ(rn,yn )Pmn (t,y),Nm = nr × ny × nvr × nvy , (11.35)
n=1
where
2d n (y)
Pmn (t,y) = p t − m ,n = {1,2,. . .,Nm } (11.36)
c
and nvr and nvy are the total number of range and azimuth velocities considered. After
processing, each target will undergo a different amount of azimuth shift and defocusing
due to different velocities. These velocities should be taken into account during process-
ing to compensate for these effects. Therefore, the velocities should be estimated first
so that the estimated values could be used for focusing of the images.
as follows:
Nm
(r,y,vr ,vy ) = σ(rn,yn,vnr,vny )δ(r − rn,y − yn,vr − vnr,vy − vny ). (11.37)
n=1
Note that if range and/or azimuth acceleration is considered, the reflectivity will have
higher than four dimensions. This matrix shows the reflectivity of a target corresponding
to its range and azimuth positions, and having certain range and azimuth velocities. The
matrix is rearranged in a vector of size Nm × 1 as follows:
σm = [σm
1 2
,σm Nm T
,. . .σm ] , (11.38)
where σm n = σ(r ,y ,vn,vn ),n = {1,2,. . .,N }.
n n r y m
Traditionally a fixed dictionary–based approach was used for CS-based moving-target
imaging, where the dictionary would consist of data for all possible range and azimuth
velocities for all the points in a considered scene. Stojanovic and Karl [44] use a fixed
dictionary–based approach for moving target imaging and showed that CS can be used
to estimate velocities and positions of moving objects considering a single scatterer
in each observation pixel for mono- and multistatic SAR configurations. The authors
considered a high signal-to-clutter ratio (SCR), i.e., a scenario in which the moving
targets’ amplitude are much higher than the stationary targets. In [45], a fixed dictionary-
based CS processing method is used to focus targets moving in range direction only,
considering a low SCR, where a clutter cancellation filter was used to increase the SCR.
A fixed dictionary is given as
Nm
m = [pm
1 2
,pm ,. . .pm ], (11.39)
n is the response of the nth moving target given as a N N × 1 vector:
where pm t y
n (y)
4πfc dm 2d n (y) 2
n
pm = exp −j + j πK t − m ,n = {1,2,. . .,Nm }. (11.40)
c c
Using the undersampling operator A, the data corresponding to moving targets can
be expressed as
sm = Am σm + . (11.41)
The dictionary becomes very large if a large number of velocity parameters are
considered. The OMP algorithm is typically used as recovery algorithm for motion
parameter estimation due to low computational complexity and ease of implementation
for reconstruction of reflectivities. Moreover, the process of selecting basis vectors one
by one can be useful in cases where correlation exists amongst the basis vectors. The
algorithm solves the following problem:
σ̂m = argminsm − Am σm 22 + λσm 0 . (11.42)
σm
The steps involved in the OMP algorithm consist of iteratively: 1) correlating the
received data with each column of the sub-sampled dictionary, 2) finding the column
336 Khwaja, Onhon, and Cetin
with the maximum correlation, 3) estimating the reflectivity using the selected column
via a least-squares solution, and 4) removing the contribution of the estimation from
the received data. The solution shows the amplitude for each combination of motion
parameter values.
The reconstructed reflectivity vector σ̂ can be rearranged as a four-dimensional matrix
as follows:
ˆ m (r,y,vr ,vy ) = f −1 ( σ̂m ). (11.43)
The four-dimensional matrix can be further divided into 2D reflectivity maps showing
estimated reflectivity for range and azimuth positions corresponding to different values
of velocities. Example results from [45] are shown in Figure 11.8.
In [46], the authors apply the CS SAR imaging approach to estimate the target reflec-
tivities and motion parameters of targets having rotational and vibrational motions based
on fixed dictionaries. In [47], a fixed dictionary–based approach is used for indication
of human motion with through-the-wall imaging.
Different approaches were also suggested in the literature to reduce the dictionary
size and enable the applications of CS SAR moving target imaging to realistic scenes.
In [48], a fixed dictionary–based CS imaging approach is used to focus data moving in
azimuth direction, where the range velocity is estimated before this step using a Radon
transform approach. Using such an approach can help in decreasing the dictionary size
as the fixed dictionary would consist of only azimuth velocities. The fixed dictionary–
based method is also applied to motion parameter estimation and focusing of the pro-
cessed images in [49], where the relative localization of the defocused targets provides
an opportunity to generate the locations in the dictionary only around a limited region
of the scene. Results obtained using this method are shown in Figure 11.9.
References [49] and [50] examine the CS imaging performance in the presence of
dictionary mismatch, i.e., when the discrete parameters used to generate the dictionary
elements are different from the parameters in the data. In this case, the data are assumed
#
to be generated from a dictionary m , and hence the reconstruction is carried out as
follows:
# #
σ̂m = argmins − Am σm 22 + λσ0 . (11.44)
σ
The reconstructed reflectivity in this case is different from the actual reflectivity and this
difference depends on the correlation between the actual and the assumed dictionaries.
The reconstructed reflectivity will become
# #
σ̂m = m m σ. (11.45)
These references also showed that to deal with this problem an oversampled dictio-
nary can be created, and the oversampling requirement is more severe for range velocity,
i.e., the dictionary has to be created such that the range velocity spacings are very
small. This can be seen in Figure 11.10, where MSE versus dictionary mismatch in
positions and velocities can be seen. It can be noted that a small mismatch causes a
sudden increase of MSE to a maximum value. The reason is that a small mismatch in
range velocity causes the position of the reconstructed reflectivity to be different from
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 337
Scene containing Multiple Moving Points in a Pixel Reconstructed Scene (Velocity = 3 m/s)
4
2.5
3 2
Amplitude
Amplitude
1.5
2
1
1
0.5
0 0
40 40
30 15 30 15
20 10 20 10
10 5 10 5
0 0 0 0
Azimuth Bins Range Bins Azimuth Bins Range Bins
3 2.5
2.5 2
Amplitude
Amplitude
2
1.5
1.5
1
1
0.5 0.5
0 0
40 40
30 15 30 15
20 10 20 10
10 5 10 5
0 0 0 0
Azimuth Bins Range Bins Azimuth Bins Range Bins
2.5
2
Amplitude
1.5
0.5
0
40
30 15
20 10
10 5
0 0
Azimuth Bins Range Bins
Figure 11.8 Simulated and reconstructed scene. The reconstructed scenes show estimated
reflectivities for different values of velocities. (Taken from [45] with permission)
the actual position, as dictionary elements with different combinations of range velocity
and azimuth positions can be highly correlated.
The technique in [51] aims to recover moving targets up to a membership in an
equivalence class. These equivalence motion classes involve different combinations of
starting positions and velocities that correspond to the same SAR-to-target range history.
338 Khwaja, Onhon, and Cetin
0.6
Azimuth bins
400 250 0.6
0.5
500 300 0.5
Road
0.4 350
0.4
600
400
0.3 0.3
700 450
0.2 0.2
800 500
0.1 0.1
550
900
0 600 0
2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
Range bins Range bins
(a) Difference of cross-polarized channels. (b) Focused result at velocity 7.9 m/s.
150 0.35
200
0.3
Azimuth bins
250
0.25
300
Road
350 0.2
400 0.15
450
0.1
500
0.05
550
600 0
2 4 6 8 10 12 14 16 18 20
Range bins
Figure 11.9 Comparison of input image and images with retrieved motion parameters. (Taken
from [49] with permission)
0.007
0.006
MSE
0.005
0.004
0.003
0.002
0.001
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Mismatch
Figure 11.10 Mean-square error versus dictionary mismatch. (Taken from [49] with permission)
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 339
s = A(
m )σm + , (11.46)
where
m is a matrix of size Nt Ny × Nm containing the phase errors due to motion
and represents element-by-element multiplication. The matrix is given as
m = [
ψ1m,
ψ 2m,. . .,
ψN m
m ], (11.47)
ψ nm = [exp{−j kc φen (y1 )},. . ., exp{−j kc φen (y1 )}, exp{−j kc φen (y2 )},. . .,
exp{−j kc φen (y2 )}. . . exp{−j kc φen (yNy )},. . ., exp{−j kc φen (yNy )}]T .
(11.48)
σ̂m = argmins − A(
m )σm 22 + λ1 σm 1 + λ2 β − 11,s.t.|β(k)| = 1∀k,
σm,β
(11.49)
where
β T = [β1 T ,β2 T ,. . .,βNy T ] (11.50)
and
βn = [exp(−j kc φe1 (yn ), exp(−j kc φe2 (yn ),. . ., exp(−j kc φeNm (yn ))]T . (11.51)
The extra term is used for facilitating the solution. As the term β represents the phase
errors due to the moving points, and there are only a few moving points in a scene,the
340 Khwaja, Onhon, and Cetin
50 50
100 100
150 150
200 200
250 250
300 300
350 350
400 400
450 450
500 500
50 100 150 200 250 300 350 400 450 500 50 100 150 200 250 300 350 400 450 500
(a) (b)
50 50
100 100
150 150
200 200
250 250
300 300
350 350
400 400
450 450
500 500
50 100 150 200 250 300 350 400 450 500 50 100 150 200 250 300 350 400 450 500
(c) (d)
Figure 11.11 (a) Original scene. (b) Defocused image obtained using traditional processing. (c)
Image reconstructed by sparsity-driven imaging assuming a stationary scene. (d) Image
reconstructed by the sparsity-driven moving object imaging approach. (Taken from [59] with
permission)
term β−1 becomes sparse, and is hence represented as β−11 . An example of imaging
carried out using this approach is shown in Figure 11.11.
In [55], the authors extend the applicability of sparsity-driven moving target–
focusing methods to very low signal-to-clutter ratio environments. First, SAR raw
data are divided into subapertures in the azimuth direction. Subsequently, low-rank and
sparse decomposition is applied using the multiple subapertures data to accomplish the
separation of moving targets from the stationary SAR background.
In [56], the authors use a variant of fractional FT as a basis to estimate the Doppler
rate from undersampled data, and hence the azimuth velocity. They use Radon transform
to estimate range velocity. In [57], moving targets in a processed imaged are refocused
by first extracting a ROI sub-image and then following a two-step iterative procedure.
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 341
A sparse image is first obtained using an initial phase compensation parameter. Then
the phase compensation parameter is estimated.
In [58], the smeared images of moving targets are removed from a processed image
to obtain a clear SAR image from multi-channel SAR images. The presented approach
uses a fixed-dictionary CS-based method to separate SAR data from moving targets and
noise. The fixed dictionary is based on frequency shifts for different range velocities
and angle between flight path and a moving object that has dependency on both range
and azimuth velocity of the moving target.
Inverse synthetic aperture radar imaging consists of a moving target observed by a static
antenna. The target’s motion creates a synthetic aperture, which can be used to obtain
a high resolution image. However, a target can have an unknown maneuvering motion,
which is similar to the case of SAR imaging in the presence of residual motion errors.
This unknown maneuvering motion needs to be estimated and compensated to obtain
a high-resolution artifact-free image. The moving target in a static background can be
considered as a sparse scene, therefore, CS can be applied for ISAR imaging. In the
following, we formulate the problem of ISAR imaging and CS application for ISAR
imaging, followed by a literature review of existing CS ISAR imaging approaches.
where τ is the along-track time sampled at a frequency fτ , i.e., τ = 0, f1τ , f2τ ,. . ., Nfτ −1
τ
,
with Nτ being the total number of along-track time samples. The slant-range distance
to the rotation center of the observed moving point is represented by rn ; xn and yn
denote the range and azimuth distance of the point from the rotation center, respectively.
The rotational motion is responsible for creating a high-resolution in the along-track
direction. Note that it is normally assumed that the whole object is rotating with the
same rotation angle, i.e.,
θn (τ) =
θ0 (τ)∀n.
Generally, ISAR imaging problems consider range-compressed and range-aligned
ISAR data, where any movement across range cells due to translational motion is
assumed to be eliminated. Therefore, the data are aligned along the azimuth direction
for each range bin. Such data, corresponding to a range bin and a point n, can be
written as
sn (τ) = σn pn (τ). (11.53)
342 Khwaja, Onhon, and Cetin
The reflectivity of the point n is given by σn , and pn (τ) describes the range-aligned
signal based on the rotational motion of the target as follows:
τ $ %
pn (τ) = rect exp −j kc dan (τ) , (11.54)
Tτ
& '
where rect Tττ is a limit on the pulse size according to pulse-width Tτ arising from
the total target observation time in azimuth direction, and dan (τ) = xn cos(
θn (τ)) +
yn sin(
θn (τ)) is the radar-target distance after range-alignment. The total range-
compressed and aligned raw data for a total number of Nm moving points in a single
range bin can then be written as a sum of received data from each point as follows:
Nm
s(τ) = σn pn (τ). (11.55)
n=1
The purpose of ISAR imaging is to obtain a focused image for each range bin from
the raw data. For ISAR, if the rotational motion is not properly accounted for in the
imaging process, it will result in blurring in the processed image in the along-track
direction. Therefore, the rotational rate should be estimated and compensated to get a
focused image.
as βn = k2c yn ω̇n , and γn = k6c yn ω̈n is defined as the rotational acceleration rate phase
term. Then, the pulse pn (τ) for a point n can be redefined as
& '
τ
pn (τ,α n,βn,γn ) = rect exp −j αn τ − j βn τ 2 − j γn τ 3 . (11.57)
Tτ
As in Section 11.4, a reflectivity vector σm
of size Nm × 1 is defined such that its
entries have the form σmn = σ(α ,β ,γ ), n = {1,2,. . .,N }, where N = N × N ×
n n n
m
m α β
Nγ . The total number of considered points for rotational rate, rotational acceleration,
and rotational acceleration rate are Nα,Nβ and Nγ , respectively.
The dictionary is defined as
Nm
= [pm
1 2 3
m ,pm
,pm ],
,. . .,pm (11.58)
n is the response of the nth rotating target given as
where pm
& '
n τ
pm = rect exp −j αn τ − j βn τ 2 − j γn τ 3 ,n = {1,2,. . .,Nm
}. (11.59)
Tτ
The dictionary m has a size Nτ × Nm . Making use of the dictionary definition, the
received raw data for one range bin can be written as
σm
s = Am + ε. (11.60)
In Figure 11.12, the model based on (11.60) is illustrated. Equation 11.60 can be
solved for each range bin using the fixed dictionary–based approach as follows:
= argmins − Am
σ̂m σm
2 + λσm
2
0 . (11.61)
σm
The reconstructed reflectivity σ̂m for each range bin is then converted to a 1D form,
which shows the reflectivity estimate versus the Doppler pixel position.
In [61] and [62], it is further shown that to avoid the degradation of performance
caused by dictionary mismatch, the fixed dictionary should be sufficiently upsampled
in rotational acceleration and rotational acceleration rate. Figure 11.13 gives one such
example, where the degradation of performance with respect to mismatch in rotational
0.9
1 0.8
0.8 0.7
0.6 0.6
MSE
0.4 0.5
0.4
0.2
0.3
0
5 0.2
0.95 5
0.7 0.95 0.1
0.45 0.7
0.45
0.2 0.2
γ mismatch β mismatch
Figure 11.13 MSE versus β mismatch and γ mismatch. The circled region shows variation of
MSE with fixed β mismatch and varying γ mismatch, and vice versa. (Taken from [61] with
permission)
acceleration and rotational acceleration rate can be seen. The performance degrada-
tion is evaluated in terms of MSE between actual and reconstructed reflectivity in the
presence of mismatch. Therefore, to avoid any performance degradation, the dictionary
should be constructed with very fine spacing of rotational acceleration and the rotational
acceleration rate.
A parametric dictionary or a dictionary-less approach can also be used to avoid
the high upsampling required for a fixed dictionary–based approach. In the parametric
dictionary approach, the motion parameters of the moving targets are estimated as part
of the solution directly. In this case, the reconstruction problem can be defined as
σ̂m
= argmin s − Am
(α,β,γ)σ2 + λσm
2
0, (11.62)
σm
,α,β,γ
where the term m (α,β,γ) now shows that dictionary elements are generated based on
the motion model given in (11.56), and the parameters are found such that the resulting
dictionary elements are well matched to the received raw data. One possible approach
to solve this problem is as part of the OMP algorithm, where the dictionary elements
are iteratively generated such that they are best matched to received data. This can be
expressed as
α̂ k , β̂k , γ̂k = argmax| < s,(Am
(α k ,β k ,γk )) > |, (11.63)
α k ,β k ,γk
where k is the iteration number and | < .,. > | represents correlation. An example of
results obtained using a parametric dictionary-based approach is shown in Figure 11.14.
This figure was taken from [61].
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 345
200 200
150 150
Azimuth
Azimuth
100 100
50 50
10 20 30 40 50 60 10 20 30 40 50 60
Range Range
(a) (b)
Comparison of reference image and reconstructed image Comparison of reference image and reconstructed image
250 250
Reference image Reference image
Reconstructed image Reconstructed image
200 200
150 150
Azimuth
Azimuth
100 100
50 50
10 20 30 40 50 60 10 20 30 40 50 60
Range Range
(c) (d)
Figure 11.14 (a) Reference image. Note that the reference and reconstructed images are
overlapped for comparing the reconstruction performance: a successful reconstruction will result
in the overlap looking like the reference image, whereas an unsuccessful reconstruction will
result in the overlap having points around the reference image. (b) Overlap of reference and
reconstructed images using a parametric dictionary approach with 50% downsampled data.
(c) Overlap of reference image and reconstructed image using a technique based on integrated
high-order matched phase transform with 50% data; partial reconstruction can be seen as the
technique was proposed for imaging with fully sampled data. (d) Overlap of reference image
and reconstructed image using a dictionary without a rotational acceleration rate phase term
with 50% data. The resulting reconstruction using this dictionary fails to generate
a discernible image. (Taken from [61] with permission)
where it is assumed that the target is moving with a uniform rotational velocity, or
only such a small number of continuous samples in azimuth direction are acquired
that the target can be considered to have uniform motion during the acquisition. Such
an approximation can be seen as a small-angle approximation, or a uniform rotation
346 Khwaja, Onhon, and Cetin
n τ
pm
= rect exp (−j α n τ) ,n = {1,2,. . .,Nα }. (11.65)
Tτ
where ζn = k2c xn ωn2,n = {1,2,. . .,Nα × Nζ }. The authors use a parametric dictionary
for reconstruction. Such a model also allows determination of the actual azimuth posi-
tion of the moving targets. In [74], a similar motion model is used, and the phase error
after translational motion compensation is considered. A two-step imaging process is
used, where reflectivity is estimated first using compressed sensing reconstruction and
starting with a phase error estimate, then the phase errors are estimated using entropy
minimization method.
In [68], the authors deal with translational motion phase error compensation and
a target undergoing motion according to the model given in (11.70). The proposed
imaging approach involves a two-step process: 1) A coarse error compensation is
achieved using the minimum entropy criterion; 2) a sparsity-driven optimization using
a 2D Fourier-based dictionary is carried out, where the residual phase errors are treated
as model error and removed to achieve a fine correction.
In other approaches, the rotation motion is assumed to consist of nonuniform rotation
up to a second-order term as follows:
yn
θn (τ) ≈ yn ωn τ + ω̇n τ 2 . (11.68)
2
The dictionary is composed of the following terms:
n τ $ %
pm = rect exp − j α n τ − j βn τ 2 ,n = {1,2,. . .,Nα × Nβ }. (11.69)
Tτ
In [62], a dictionary is presented that can deal with both first-order rotational velocity
and second-order rotational acceleration phase terms. An analysis is also carried out to
quantify the effects of spacing of different parameters in the dictionary on the imaging
performance. In [75], the authors consider rotational velocity and uniform acceleration.
The work in [76] consider uniform and nonuniform rotational motion in the presence
of unavailable or corrupted data and demonstrate the application of quadratic time
frequency–based representations for CS imaging when rotational acceleration is present
in the data.
In [77], the authors consider rotational velocity and uniform acceleration. Most of the
autofocus approaches lose their efficacy in subaperture cases, because the vacant aper-
tures break the FT relationship between the range-compressed data and the RD image.
Eigenvector-based autofocus is an exception. Rotational acceleration is compensated by
searching, then the residual motion errors are calculated by eigenvector-based autofocus.
348 Khwaja, Onhon, and Cetin
The authors in [78] consider uniformly accelerated rotation targets. The maneuvering
signal model is formulated as chirp code and represented using a chirp-Fourier basis.
Then sparse representation is applied to realize range-Doppler imaging from the sparse
apertures, where the superposition of chirp parameters is acquired using the modi-
fied discrete chirp Fourier transform (MDCFT). After preprocessing involving sample
selection, rotation center determination, and noise reduction, the chirp parameters are
used to estimate the parameters of rotational motion using the weighted least square
(WLS) method. Finally, a high-resolution scaled-ISAR image is achieved by rescaling
the acquired RD image using the estimated rotational velocity.
In [79], the authors propose a weighted eigenvector-based phase correction method to
correct for unknown phase errors, followed by using a partial Fourier matrix to achieve
CS ISAR imaging. In [80], it is assumed that the translational motion compensation has
been successfully accomplished and only rotational motion is considered. Rotational
acceleration is considered and a dictionary based on scaled nonuniform Fourier trans-
form is used. The reflectivity is modeled as a complex Gaussian distribution.
In [81], the authors further extend the technique in [68] to the maneuvering target
model undergoing uniform rotational acceleration. In this case, the rotational motion is
approximated as
yn xn
θn (τ) ≈ yn ωn τ + ω˙n τ 2 − ωn2 τ 2, (11.70)
2 2
and the dictionary is composed of
n τ $ %
pm = rect exp − j α n τ − j ηn τ 2 ,n = {1,2,. . .,Nα × Nη }, (11.71)
Tτ
where ηn = k2c (yn ω˙n − xn ωn2 ),n = {1,2,. . .,Nα × Nη }.
Some other techniques do not assume any particular form of motion errors, and
can deal with random phase errors. In [82], the authors propose an iterative two-step
method: in the first step, an estimate of the reflectivity is obtained using an expansion-
compression variance component–based method. In the second step, the phase errors are
estimated making use of maximum likelihood method, similar to the method proposed in
[33] for SAR autofocus. The whole process is repeated until no significant improvement
can be seen. The method can deal with any type of motion errors.
In [83], the authors propose a novel sparse Bayesian ISAR imaging algorithm with a
newly proposed logarithmic Laplacian prior, which is achieved by putting a logarithm
on the exponent of the Laplacian prior. Compared to the Gaussian scale mixture and
Laplacian priors, the proposed logarithmic Laplacian prior has a narrower main lobe
and higher tail values, and performs better on sparseness representation. Then, the
logarithmic Laplacian prior-based ISAR image is reconstructed by MAP estimation,
and arbitrary phase errors are estimated based on the minimum entropy criterion during
the iterative process of sparse signal recovery. In [84], the authors impose continuity in
range by imposing a Bayesian model, and propose an autofocus algorithm to compen-
sate for random phase errors.
In [85], the authors assume a target only having a range velocity and micro-Doppler
motion. They use a joint sparsity model to separate micro-Doppler from the main target
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 349
11.6 Conclusions
References
[1] D. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1289–1306,
2006.
[2] E. Candès and T. Tao, “Decoding by linear programming,” IEEE Trans. Inf. Theory, vol. 51,
no. 12, pp. 4203–4215, 2005.
[3] E. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruc-
tion from highly incomplete frequency information,” IEEE Trans. Inf. Theory, vol. 52, no. 2,
pp. 489–590, 2006.
[4] E. Candès and M. Wakin, “An introduction to compressive sampling,” IEEE Signal Process.
Mag., vol. 25, no. 2, pp. 21–30, 2008.
[5] R. Baraniuk, “Compressed sensing,” IEEE Signal Process. Mag., vol. 24, no. 4, pp. 14–20,
2007.
[6] R. Baraniuk and P. Steeghs, “Compressive radar imaging,” in Proc. IEEE Radar Conference,
2007, pp. 128–133.
350 Khwaja, Onhon, and Cetin
[7] V. Patel, G. Easley, D. Healy, and R. Chellappa, “Compressed synthetic aperture radar,”
IEEE J. Sel. Topics Signal Process., vol. 4, no. 2, pp. 244–254, 2010.
[8] K. Aberman and Y. C. Eldar, “Sub-nyquist SAR via fourier domain range-doppler process-
ing,” IEEE Trans. Geosci. Remote Sens., vol. 55, no. 11, pp. 6228–6244, 2017.
[9] M. Cetin, I. Stojanovic, N. Onhon et al. “Sparsity-driven synthetic aperture radar imaging:
Reconstruction, autofocusing, moving targets, and compressed sensing,” IEEE Signal
Process. Mag., vol. 31, no. 4, pp. 27–40, 2014.
[10] L. Potter, E. Ertin, J. Parker, and M. Cetin, “Sparsity and compressed sensing in radar
imaging,” Proc. IEEE, vol. 98, no. 6, pp. 1006–1020, 2010.
[11] M. Herman and T. Strohmer, “High-resolution radar via compressed sensing,” IEEE Trans.
Signal Process., vol. 57, no. 6, pp. 2275–2284, 2009.
[12] M. T. Alonso, P. Lopez-Dekker, and J. J. Mallorqui, “A novel strategy for radar imaging
based on compressed sensing,” IEEE Trans. Geosci. Remote Sensing, vol. 48, no. 12,
pp. 4285–4295, 2010.
[13] X. Dong and Y. Zhang, “A novel compressive sensing algorithm for SAR imaging,” IEEE J.
Select. Topics Appl. Earth Observ. Remote Sensing, vol. 7, no. 2, pp. 708–729, 2013.
[14] M. Cetin and W. C. Karl, “Feature-enhanced synthetic aperture radar image formation based
on nonquadratic regularization,” IEEE Trans. Image Processing, vol. 10, no. 4, pp. 623–631,
2001.
[15] D. L. Donoho and M. Elad, “Optimally sparse representation in general (nonorthogonal)
dictionaries via l1 minimization,” Proceedings of the National Academy of Sciences,
vol. 100, no. 5, pp. 2197–2202, 2003.
[16] D. M. Malioutov, M. Cetin, and A. S. Willsky, “Optimal sparse representations in general
overcomplete bases,” Proc. IEEE International Conference on Acoustics, Speech, and Signal
Processing, pp. 793–796, 2004.
[17] M. Soumekh, Synthetic Aperture Radar Signal Processing. Wiley, 1999.
[18] C. Cafforio, C. Prati, and F. Rocca, “SAR data focusing using seismic migration techniques,”
IEEE Trans. Aerosp. Electron. Syst., vol. 27, no. 2, pp. 194–207, 1999.
[19] I. Cumming and F. Wong, Digital Processing of Synthetic Aperture Radar Data. Artech
House, 2005.
[20] R. K. Raney, H. Runge, R. Bamler, I. Cumming, and F. Wong, “Precision SAR processing
using chirp scaling,” IEEE Trans. Geosci. Remote Sens., vol. 32, no. 4, pp. 786–799, 1994.
[21] R. Baraniuk, V. Cevher, M. Duarte, and C. Hegde, “Model-based compressive sensing,”
IEEE Trans. Inf. Theory, vol. 56, pp. 1982–2001, 2010.
[22] L. Applebaum, S. Howard, S. Searle, and R. Calderbank, “Chirp sensing codes: Determin-
istic compressed sensing measurements for fast recovery,” Appl. Comput. Harmon. Anal.,
vol. 26, pp. 283–290, 2009.
[23] I. Stojanovic, M. Cetin, and W. C. Karl, “Compressed sensing of monostatic and multistatic
SAR,” IEEE Geosci. Remote Sens. Lett., vol. 10, no. 6, pp. 1444–1448, 2013.
[24] J. Tropp and A. Gilbert, “Signal recovery from random measurements via orthogonal
matching pursuit,” IEEE Trans. Inf. Theory, vol. 53, no. 12, pp. 4655–4666, 2007.
[25] M. Elad and M. Zibulevsky, “Iterative shrinkage algorithms and their acceleration for l1l2
signal and image processing applications,” IEEE Signal Process. Mag., vol. 27, no. 3,
pp. 78–88, 2010.
[26] D. P. Belcher and N. C. Rogers, “Theory and simulation of ionospheric effects on synthetic
aperture radar,” IET Radar, Sonar & Navigation, vol. 5, no. 5, pp. 541–551, 2009.
[27] G. Franceschetti and R. Lanari, Synthetic Aperture Radar Processing. CRC Press, 1999.
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 351
[45] A. S. Khwaja and J. Ma, “Applications of compressed sensing for SAR moving target
velocity estimation and image compression,” IEEE Trans. Instrum. Meas., vol. 60, no. 8,
pp. 2848–2860, 2011.
[46] S. Zhu, A. M. Djafari, H. Wang et al., “Parameter estimation for SAR micromotion target
based on sparse signal representation,” EURASIP J. Adv. Sig. Proc., vol. 2012, 2012.
[47] F. Ahmad and M. Amin, “Through-the-wall human motion indication using sparsity-driven
change detection,” IEEE Trans. Geosci. Remote Sens., vol. 51, no. 2, pp. 881–890, 2013.
[48] Q. Wu, M. Xing, C. Qiu, B. Liu, Z. Bao, and T.-S. Yeo, “Motion parameter estimation in
the SAR system with low PRF sampling,” IEEE Geosci. Remote Sens. Lett., vol. 7, no. 3,
pp. 450–454, 2010.
[49] A. S. Khwaja and X. P. Zhang, “Motion parameter estimation and focusing from SAR
images based on sparse reconstruction,” IEEE Geosci. Remote Sens. Lett., vol. 11, no. 8,
pp. 1350–1354, 2011.
[50] A. S. Khwaja, M. Naeem, and A. Anpalagan, “Analysis of moving object imaging from
compressively sensed SAR data in the presence of dictionary mismatch,” Int. J. Antenn.
Propag., vol. 2013, 2013.
[51] J. Gunther, J. Hunsaker, H. Anderson, and T. Moon, “Sparse reconstruction of equivalence
classes of moving targets using single-channel synthetic aperture radar,” in Proc. IEEE
International Conference on Acoustic, Speech and Signal Processing, 2014.
[52] L. Prunte, “Gmti from multichannel SAR images using compressed sensing under off-grid
conditions,” in Proc. International Radar Symposium, 2013.
[53] C. Ekanadham, D. Tranchina, and E. P. Simoncelli, “Recovery of sparse translation-invariant
signals with continuous basis pursuit,” IEEE Trans. on Signal Processing, vol. 59, no. 10,
pp. 4735–4744, 2011.
[54] N. O. Onhon and M. Cetin, “SAR moving target imaging in a sparsity-driven framework,”
in Proc. SPIE Optics+Photonics, Wavelets and Sparsity XIV, 2011, pp. 8138–8139.
[55] M. Yasin, M. Cetin, and A. S. Khwaja, “SAR imaging of moving targets by subaperture
based low-rank and sparse decomposition,” in Proc. IEEE Signal Processing and Commu-
nications Applications Conference, 2017.
[56] X. Zhang, G. Liao, S. Zhu, D.Yang, and W. Du, “Efficient compressed sensing method for
moving-target imaging by exploiting the geometry information of the defocused results,”
IEEE Geosci. Remote Sens. Lett., vol. 12, no. 3, pp. 517–521, 2015.
[57] Y. Chen, Q. Zhang, G. Li, and J. Sun, “Refocusing of moving targets in SAR images via
parametric sparse representation,” Remote Sensing, vol. 9, no. 8, pp. 1–15, 2017.
[58] L. Prunte, “Compressed sensing for removing moving target artifacts and reducing noise in
SAR images,” in Proc. European Conference on Synthetic Aperture Radar, 2016.
[59] N. O. Onhon and M. Cetin, “SAR moving object imaging using sparsity imposing priors,”
EURASIP J. Adv. Sig. Proc., vol. 2017, 2017.
[60] V. Chen and H. Ling, Time-Frequency Transforms for Radar Imaging and Signal Analysis.
Artech House, 2002.
[61] A. S. Khwaja and M. Cetin, “Compressed sensing ISAR reconstruction considering highly
maneuvering motion,” Electronics, vol. 6, no. 1, 2017.
[62] A. S. Khwaja and X. P. Zhang, “Compressed sensing ISAR reconstruction in the presence of
rotational acceleration,” IEEE J. Sel. Top. Appl. Earth Observ., vol. 7, no. 7, pp. 2957–2970,
2014.
[63] Z. Liu, X. Wei, and X. Li, “Decoupled ISAR imaging using rsfw based on twice compressed
sensing,” IEEE Trans. Aerosp. Electron. Syst., vol. 50, no. 4, pp. 3195–3211, 2014.
Compressed Sensing, Radar Imaging, Phase Errors, and Moving Objects 353
[81] G. Xu, L. Yang, G. Bi, and M. Xing, “Maneuvering target imaging and scaling by using
sparse inverse synthetic aperture,” Signal Process., vol. 137, pp. 149–159, 2017.
[82] W. Su, Y. Qin, H. Wang, and Q. Yang, “Joint ISAR imaging and phase error correction based
on sparse bayesian learning,” Int. J. Sig. Proc. Sys., vol. 4, no. 6, pp. 487–493, 2016.
[83] S. Zhang, Y. Liu, X. Li, and G. Bi, “Logarithmic Laplacian prior based Bayesian inverse
synthetic aperture radar imaging,” Sensors, vol. 16, no. 5, 2016.
[84] L. Zhao, L. Wang, G. Bi et al. “Structured sparsity-driven autofocus algorithm for high-
resolution radar imagery,” Signal Process., vol. 125, pp. 376–388, 2016.
[85] L. Sun, X. Lu, and W. Chen, “Joint sparsity-based ISAR imaging for micromotion targets,”
IEEE Geosci. Remote Sens. Lett., vol. 13, no. 11, pp. 1734–1738, 2016.
[86] Q. Hou, Y. Liu, and Z. Chen, “Reducing micro-Doppler effect in compressed sensing ISAR
imaging for aircraft using limited pulses,” Electron. Lett., vol. 51, no. 12, pp. 937–939, 2015.
Index
355
356 Index