Compressive Sensing Based Algorithms For Electronic Defence

Signals and Communication Technology
Amit Kumar Mishra
Ryno Strauss Verster
Compressive
Sensing Based
Algorithms
for Electronic
Defence
More information about this series at http://www.springer.com/series/4748
Amit Kumar Mishra Ryno Strauss Verster
•
Compressive Sensing Based

Algorithms for Electronic
Defence
123
Amit Kumar Mishra Ryno Strauss Verster
Department of Electrical Engineering Department of Electrical Engineering
University of Cape Town University of Cape Town
Rondebosch, Cape Town Rondebosch, Cape Town
South Africa South Africa
ISSN 1860-4862 ISSN 1860-4870 (electronic)

ISBN 978-3-319-46698-9 ISBN 978-3-319-46700-9 (eBook)
DOI 10.1007/978-3-319-46700-9
Library of Congress Control Number: 2016956629
© Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To our loving families and colleagues.
Contents
Part I Electronic Defence and Compressive Sensing

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Motivation and Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Outline and Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Electronic Defence Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Electronic Defence Overview . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Electronic Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.3 Electronic Attack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.4 Electronic Protect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Electronic Support Communication Applications . . . . . . . . . . . . . 13
2.2.1 Communication Electronic Support—CES . . . . . . . . . . . . 13
2.2.2 Communication Intelligence—COMINT . . . . . . . . . . . . . . 16
2.2.3 Signal Processing Techniques . . . . . . . . . . . . . . . . . . . . . . 16
2.2.4 Signal Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.5 Signal Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.6 Emitter Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Direction of Arrival Methods Used for Electronic
Support Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 DOA Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.1 Phase Interferometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.2 DOA Estimation Algorithms . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 Existing Compressive Based Direction-of-Arrival Methods . . . . . 31
3 Compressive Sensing: Acquisition and Recovery . . . . . . . . . . . . . . . . 33
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Compressive Sensing Formulation . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.1 Compressible Signal and Sensing Matrices . . . . . . . . . . . . 35
3.2.2 Implementation for 1-Dimensional Signals . . . . . . . . . . . . 37
vii
viii Contents
3.3 Compressive Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 39

3.3.1 Random Demodulator (RD) Analog-to-Information
Sampler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.2 Non-uniform Sampler (NUS) . . . . . . . . . . . . . . . . . . . . . . 42
3.3.3 Compressive Multiplexing (CMUX) Sampler . . . . . . . . . . 44
3.3.4 Summary of CS Sampling Methods . . . . . . . . . . . . . . . . . 44
3.4 CS Recovery Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.4.1 Convex Optimization-Based Algorithms . . . . . . . . . . . . . . 47
3.4.2 Fast ‘1 -Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4.3 Greedy Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4.4 Summary of CS Recovery Algorithms . . . . . . . . . . . . . . . 58
Part II Simulations of Compressive Sensing Used for Electronic

Support Applications
4 Design of CS Based DOA Estimation for Modulated Shift-Keying
Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.1.1 Shift-Keying CS Sensing Matrix Development . . . . . . . . . 66
4.1.2 CS Recovery Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.1.3 Implementing CS DOA Estimation . . . . . . . . . . . . . . . . . . 72
5 CS Based Shift-Keying Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.1 Simulation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.1.1 General Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.1.2 Performance Indicators . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.1.3 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2 Simulation 1.1.1—Phase CS Recovery for 2FSK . . . . . . . . . . . . . 80
5.3 Simulation 1.1.2—Phase CS Recovery for 2PSK . . . . . . . . . . . . . 81
5.4 Simulation 1.2.1—CS Recovery of BIE for 2FSK . . . . . . . . . . . . 82
5.5 Simulation 1.2.2—CS Recovery of BIE for 2PSK . . . . . . . . . . . . 83
5.6 Assessment of System Parameters for Shift-Keying CS
Recovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.7 Demodulation Capability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.8 Simulation 1.3.1—Computational Performance
of CS Recovery for 2FSK Signal . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.9 Simulation 1.3.2—Computational Performance of CS Recovery
for 2PSK Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.10 Computational Performance of CS Recovery . . . . . . . . . . . . . . . . 89
6 Modulation Specific CS DOA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1 Chapter Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1.1 General Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.1.2 Performance Indicators . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.1.3 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.2 Simulation 2.1—CS DOA for 2FSK Signals . . . . . . . . . . . . . . . . 96
Contents ix
6.3 Simulation 2.2—CS DOA for 2PSK Signals . . . ............. 98

6.4 Assessment of CS DOA Estimation Algorithm
for Shift Keying Modulated Signals . . . . . . . . . . . . . . . . . . . . . . . 100
6.4.1 Performance Assessment for ULA10 . . . . . . . . . . . . . . . . . 100
6.4.2 Performance Assessment for ULA3 . . . . . . . . . . . . . . . . . . 102
6.4.3 Comparison of ULA10 and ULA3 . . . . . . . . . . . . . . . . . . . 102
7 CS Based Spectrum Sensing for ES . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.2 Selective Spectrum Weighted CS Approach . . . . . . . . . . . . . . . . . 107
7.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.3.1 Case 1: Wide Band Spectrum Recovery . . . . . . . . . . . . . . 109
7.3.2 Case 2: Selective Spectrum Sensing . . . . . . . . . . . . . . . . . 111
Part III Concluding Statements and Appendices

8 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.1 CS Based DOA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.1.1 Accurate Phase Recovery Achievable Using Orthogonal
Matching Pursuit (OMP) . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.1.2 Demodulation Capability via CS Recovery for 2FSK
Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.1.3 Equivalent Computational Performance of CS DOA
as for Conventional DOA . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.1.4 Higher Accuracy of DOA Estimates for CS DOA
Than Conventional DOA for Low SNRs . . . . . . . . . . . . . 118
8.1.5 Reduction of Memory Required Using CS DOA
Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.1.6 Scalability of CS DOA Estimation for Electronic
Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.2 CS Based Spectrum Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.3 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
9 Appendix: Some Useful Theoretical Background . . . . . . . . . . . . . . . . 121
9.1 Electromagnetic Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
9.2 Receiver Components: Background . . . . . . . . . . . . . . . . . . . . . . . 122
9.2.1 Antennas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
9.2.2 RF Front-End Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
9.2.3 Radio Frequency Propagation Operations . . . . . . . . . . . . . 131
9.3 Typical ED System Configuration . . . . . . . . . . . . . . . . . . . . . . . . 138
9.4 Electronic Support Receiver Systems . . . . . . . . . . . . . . . . . . . . . . 141
9.4.1 ES Receiver Characteristics . . . . . . . . . . . . . . . . . . . . . . . . 142
9.4.2 Sensitivity Conversion. . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.4.3 Types of ES Receivers . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
x Contents
9.5 Compressive Sensing Mathematical Fundamentals . . . . . . . . . . . . 158

9.5.1 Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
9.5.2 Sets, Bases and Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
9.5.3 Matrix Construction for CS Sensing . . . . . . . . . . . . . . . . . 162
9.5.4 Null Space Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
9.5.5 Restricted Isometry Property . . . . . . . . . . . . . . . . . . . . . . . 164
9.5.6 Compliant RIP Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 165
9.5.7 Incoherence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.6 Sampling Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
9.6.1 Sampling Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
9.6.2 Conventional Sampling Mixing-Acquisition . . . . . . . . . . . 169
9.6.3 Under Sampling Techniques . . . . . . . . . . . . . . . . . . . . . . . 170
9.6.4 Direct Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
9.7 Wideband CS Sampling Techniques . . . . . . . . . . . . . . . . . . . . . . . 172
9.7.1 Multi-rate Asynchronous Sub-nyquist Sampling
(MASS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
9.7.2 Xampling-ADC and Modulated Wideband Converter
(MWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Part I
Electronic Defence and Compressive
Sensing
Chapter 1
Introduction
In this book we describe applications of Compressive Sensing (CS) based algorithms

for Electronic Support (ES) tasks such as direction-of-arrival (DOA) estimation of
modulated communication signals and spectrum sensing. We also provide the reader
with an overview of Electronic Defence and Compressive Sensing. The objective of
applying CS to ES tasks1 is to reduce the computation and memory requirements
and improve system performance.
The majority of modern ES systems are developed to function as receivers that
perform detection, classification, DOA estimation, and identification of radio fre-
quency (RF) signal simultaneously, these are performed in an environment rife with
high noise, interference, and frequency agile signal. Even in low intensive conflict
scenario ES systems have a requirement for computational performance [42]. As a
consequence, to match the computational load, modern ES receivers use field pro-
grammable gate array (FPGA) technologies [147] and digital signal processing (DSP)
cores for processing.
The adoption of FPGAs and DSP cores as part of the ES hardware has lead
to improved processing power. However, the problem of dealing with the mem-
ory requirements has not necessarily improved. In fact, modern DOA estimation
techniques require additional memory than conventional techniques. CS methods
provide an alternative solution, as it deals with the memory requirement by means
of sub-Nyquist sampling, resulting in a reduction of input data required.
CS research has rapidly developed, since its inception [28, 44] for real world
deployment on modern DSP platforms. In most cases, depending on the CS technique,
1 Electronic support is a sub-category of electronic defence (ED). ES equipment is developed to

collect, intercept, identify and locate enemy signals [3] in order to execute a specific task relative to
the threat level that received signal holds. Received signals can also be used for situational awareness
[1] or in other words, determine the type and location of enemy weapons or electronic capabilities.
ES systems need to optimize data throughput whilst gathering a considerable amount of real time
data. The objectives of ES systems are to determine the type of electromagnetic emitters present,
and where they can be located.
© Springer International Publishing AG 2017 3

A.K. Mishra and R.S. Verster, Compressive Sensing Based Algorithms
for Electronic Defence, Signals and Communication Technology,
DOI 10.1007/978-3-319-46700-9_1
4 1 Introduction
deployment is possible without requiring additional hardware. In other words, CS can

be implemented, in principal, by simply altering the signal processing software used
for ES tasks, with only a slight increase in cost on hardware. However, a marginal
cost on computation is required [173].
In the remainder of this section we outline the problem statement, scope, and
limitation of CS methods applied to DOA estimation techniques with specific appli-
cation to modulated communication signals. We then discuss our design objectives,
and lastly, present a summary of each of the chapters in this monograph.
1.1 Motivation and Problem Statement
The abundance of communication signal encountered in ES scenario presents a chal-

lenging environment for ES operations [42]. Moreover, modern communication sig-
nal comprises of several modulation techniques, which include conventional analog
type modulation (FM, AM, SSB, etc.) as well as digital modulation (FSK, MSK,
QPSK, BPSK, etc.). This abundance of communication signal creates a greater need
to optimize data throughput, whilst gathering considerable amounts of real time data
which results in higher memory requirements.
Most modern ES system processing is computed on digital processing platforms
(i.e. FPGAs and DSP cores), as opposed to prior analog sub-systems [175]. Even with
the computational core development, which has been tracking Moore’s law [155],
cost effective DSP platforms have not been realized. Hence, there is the incentive in
the ES domain to develop low-cost systems that are readily replaceable [34].
CS methods tackle the problem of memory load by sub-Nyquist sampling resulting
in optimized data throughput. Moreover, newly developed CS techniques can be
implemented using existing signal processing platforms, by adapting the software
and fine-tuning the method of signal acquisition. Thus, the unique attributes CS
methods hold, make it an ideal candidate to reduce the amount of data required.
We aim to show specific electronic defence (ED2 ) applications that CS tech-
niques could be applied to and produce comparable computational performance by
CS dependant methods, as opposed to conventional Nyquist dependent methods.
However, it shall be noted that we do not aim at showing the efficacy of CS for all
ED systems and hence it remains an open question whether CS techniques can be
developed for the complete ES/ED system.
2 Electronic defence and the term electronic warfare are interchangeably used in the open literature
and refer to the same objective—that is to protect and ensure the use of the electromagnetic spectrum
for friendly and hostile scenarios for the purpose of tactical military tasks which enable the safety of
assets (i.e. equipment and people) in the field. In this body of work we will use the term electronic
defence.
1.2 Outline and Contribution 5
1.2 Outline and Contribution
Throughout the remainder of this work we develop, design, test, and report on findings
of implementing CS based methods for ES tasks. We detail existing CS methods used
for ES type applications in the open literature and provide insights on several unique
techniques of using CS based recovery to determine the DOA of communication
signals and CS based spectrum sensing.
The framework of the book is detailed below in terms of its content, per chapter.
Chapter 2: Electronic Defence Systems
In this chapter we discuss and review conventional and modern Electronic Defence
systems with particular focus on Electronic Support Tasks such as detection and
estimation methods, and their associated computation and memory requirements.
Chapter 3: Compressive Sensing Acquisition and Recovery
In Chap. 3 we review the general CS framework. Thereafter the CS acquisition
scheme in the current literature is reviewed to determine the best method for CS
based recovery and acquisition to be used for ES tasks.
Chapter 4: Design of Modulation Specific Compressive Sensing Based Direction-of-
Arrival
Based on the literature reviewed, we describe a CS based approach to recover
the phase of digitally modulated input signal where the carrier frequency is known.
Thereafter, we outline how such phase estimates can be used to determine the direc-
tion of arrival of digitally modulated signal which is incident on a uniform linear
array (ULA).
Chapter 5: CS Based Shift-Key Modulation
In this chapter we show, by means of simulation, the capability of retrieving CS
phase for amplitude shift keying (ASK), phase shift keying (PSK), and frequency
shift keying (FSK) digitally modulated input signal for low SNR situations. Based on
the findings from the simulations, retrieval performance is discussed to form coherent
arguments and concluding remarks.
Chapter 6: Modulation Specific CS-DOA
In Chap. 6 we show, by means of simulation, the capability of accurate CS based
DOA estimation given ASK, PSK, and FSK input signals for low SNR environments.
Based on the findings from the simulations, various performance parameters relating
to computation and memory requirements are discussed to form coherent arguments
and concluding statements.
Chapter 7: Spectrum Sensing for ES
In this chapter we show the benefit of the recently proposed CS schemes on
reducing the load on acquisition for ED spectrum monitoring. Further, we describe
a modified CS scheme, which we denote as selective spectrum sensing, to further
improve signal estimation performance for spectrum sensing. The proposed scheme
leverages on a-priori knowledge of the frequency bands of interest and is shown to
perform efficiently under severe signal to noise ratio (SNR) conditions.
6 1 Introduction
Chapter 8: Concluding Remarks

In this chapter we conclude the monograph with some discussion on the future
directions CS based EW system design may take.
Chapter 9: Appendix on Some Useful Theoretical Background
In the appendix we give some useful theoretical background. In case the reader is
new to the domain it is strongly suggested that she goes through this chapter before
starting the monograph.
Chapter 2
Electronic Defence Systems
2.1 Introduction
Electronic Defence (ED) is defined as the art and science of preserving the use of
the Electro-Magnetic (EM) spectrum for friendly use while denying its use to the
enemy [157]. The inherent value that the EM spectrum has as a natural resource,
which is utilized and/or abused, is a pivotal reason why ED has such a vested interest
in preserving its use. ED protects the spectrum by utilizing task specific passive ED
receiver systems and ED transmitter systems (i.e. Jammers). ED receiver systems
(falling within the sub-category of electronic support (ES)) are designed to detect,
monitor and locate EM radiation sources (Friend or Foe), whereas ED countermea-
sures are designed to reduce the effectiveness of threats (i.e. Enemy EM radiation
sources).1
Concerning communication operations, the EM spectrum that is utilized commer-
cially and for exclusive military use, occupies large sections/bands of the Electro-
magnetic spectrum (i.e. FM, AM, GSM, UMTS, LTE, WiFi, WiMax, etc.). As is the
case with supportive measures, monitoring/sensing certain bands of interest is a typi-
cal technique to ensure that the spectrum usage is preserved. As a consequence other
operative tasks become available, allowing for tracking capability and intercepting
communications for intelligence gathering.
In this chapter we develop a critical literature review, detailing the context of ES in
the domain of ED and its application as well as implementation for communication
systems. We deal with a wide range of topics starting from ED passive communica-
tion systems (namely Electronic Support (ES) receivers), operations for detection,
communication techniques, RF propagation theory to direction finding.
1 Both ED receiver and transmitter systems are designed by sourcing from multi-disciplinary fields
such as radar, communications, digital signal processing, antenna theory, radio frequency systems,
high performance computing and computer networks to preserve the EM spectrum with high effec-
tiveness for the user.

DOI 10.1007/978-3-319-46700-9_2
8 2 Electronic Defence Systems
Fig. 2.1 Illustrates the electromagnetic spectrum usage in terms of frequency and wavelength,
sourced and adapted from [99]
2.1.1 Electronic Defence Overview
ED systems sense and monitor the EM spectrum. The frequencies that comprise the
EM spectrum range from Alternating Current (AC) to Gamma rays, which from a
governing and/or controlling entity’s (i.e. ED systems) point of view, is an enormous
band to sense and determine the usage thereof.
ED domain utilizes full extent of this usable electromagnetic (EM) spectrum
namely the radio frequency, infra-red, optical and ultraviolet spectrum [2]. Usability
of the EM spectrum is detailed in Fig. 2.1.
Classically, ED has been divided into following three domains.
• Electronic Support Measures (ESM),
• Electromagnetic countermeasures (ECM).
• Electromagnetic counter-countermeasures (ECCM).
However, in recent years these subdivisions were renamed and redefined under the
guidance of NATO [2]—now widely accepted in many countries, but not all.
Under the previous definition2 the ED subdivisions were understood as follows
(see Fig. 2.2 for detail).
• ESM—Receiver systems, mainly used for intercept purposes in ED.
• ECM—Jamming, chaff, flares used for the sole purpose to counter systems such
as radars, military communication and weapon systems.
• ECCM—Design or operational measures taken to counter radar and communica-
tion systems against the effectiveness of ECM.
Under the newly redefined view of NATO, ED subsystems/divisions are now
defined as Electronic Support (ES), Electronic Attack (EA) and Electronic Protect
(EP) subsystems. These three divisions are detailed below. In Sects. 2.1.2–2.1.4 we
shall describe how these divisions include classical definitions which can be corre-
lated to Fig. 2.2.
2 Definitions sourced and modified from [1, 2].

2.1 Introduction 9
Fig. 2.2 Illustrates the

classical definition
subdivision of the ED
domains for operational tasks
• ES—This is the same as the classical definition of ESM.

• EA—This is same as the previous ECM including jamming, chaff, and flares. Also
including anti-radiation weapons and direct energy weapons.
• EP—This is identical to the classical ECCM definition.
For clarity, it is important to distinguish ESM (or ES) from signal intelligence
(SIGINT) which contains two streams of intelligent systems, namely communica-
tion intelligence (COMINT) and electronic intelligence (ELINT) [2]. Differentiation
between these types of signal has become increasingly vague—as signal complexity
develops—for the purpose of transmissions received [1].
Purpose of the respective subdivisions:
• COMINT—The operational tasks involve receiving communication signals for
the purpose of extracting intelligence from the data/information carried by the
signals of interest.
• ELINT—These operations are interested in non-communication signals (i.e. radar
signals) to determine the type of electromagnetic system in use by an enemy,
in order to develop a counter measure. ELINT systems typically collect a large
amount of data over an extended period of time.
• ES/ESM—The modus operandi of ES is to collect, intercept, identify and locate
enemy signals [3] in order to execute a specific task relative to the threat level
that the received signal holds. The signal can also be employed for situational
awareness [1]. In other words, the signal can be used to determine the types and
locations of enemy weapons or electronic capabilities. ES systems typically need

to optimize data throughput whilst gathering a considerable amount of real time
data. The main objective of such systems is to determine which emitter types are
present and where they can be located.
The majority of the objectives, and hence the operational tasks for this monograph,
will be focused within the domain of ES. However, due to the nature of exclusively
dealing with communication signals, certain aspects of COMINT will be included
for a holistic approach of the topic.
2.1.2 Electronic Support
In combat or passive scenarios where assets are in the field, it is a high operational
priority to gather as much information about the physical environment and commu-
nications in the immediate vicinity in order to asses the threat level. This information
gathering ensures the safety of people and equipment [3]. ES undertakes this task
via electronic interception of communication and other RF signals (i.e. Radar). A
typical topology of this is shown in Fig. 2.3.
The objective of ES is to provide other electronic defence (ED) systems with
accurate combat information in order to alert and react appropriately to threats.
We refer the reader to the appendix for further information on ED theory and the
application of ES systems therein.
Fig. 2.3 The operation of a possible Electronic Support deployed in the field. This is a typical
application of an intercept layout of an ES system, intercepting communication from an adversary’s
transmissions from a communication node. Adopted from [3] and modified by the authors
2.1 Introduction 11
ES systems deliver the capability to search, intercept, identify and locate inten-
tional and unintentional sources of electromagnetic (EM) radiation [157]. These tasks
involve real-time signal acquisition and processing of combat or friendly information
gathered to generate intelligence and pass onto sub-systems. Intercepting communi-
cation signals comprise of several steps; namely receiving the signal, identifying the
type of signal, and finally locating the source of the radiation which is done by DOA
estimation methods. Information gained from a signal to infer the location of the
emitter source, namely DOA estimation, is regarded as a pivotal task for ES systems
[126].
2.1.3 Electronic Attack
Electronic Attack (EA) has at its core, the objective to restrict enemy signals access in
using the EM spectrum for communication, information exchange and/or other illicit
activities (i.e. infrared and optical detection and tracking). This denial of informa-
tion can be categorized into two schemes; firstly, in terms of information protection
(protecting your own communication link via deception or encryption), and secondly
as information attack (denying the user the use of their own communication link)
[3]. For most of the situations in ED, the attack of an adversary’s communication
systems to deny information exchange is one of the most important task in ensuring
a successful control strategy for information dominance.
Typically the denial of an adversary in exchanging information via RF commu-
nication is done via jamming. In brief, a communication jammer emits an excess
(large amounts) of RF energy in the RF link of the enemy [3, 157]. This in no way
reflects the entire scope of EA techniques and technologies. The sophistication of
such jammer technology in literature and industry [3, 34, 157] serves as a remainder
of how involved this aspect of ED is.
However, as far as the scope of the current work is concerned, more diverse EA
systems will not be critically analysed herein. Further investigation of EA systems
are left up to the reader (Fig. 2.4).
2.1.4 Electronic Protect
EP Systems are designed to restrict penetration and susceptibility of friendly systems

being interfered with by enemy ES and EA systems [1, 157]. This is an involved
process. A range of approaches must be adopted in achieving the intended task
of protecting friendly forces. EP systems have a focal objective to protect friendly
information, usually telecommunication, from being manipulated by an adversary.
Fig. 2.4 Illustrates the operation of a possible Electronic Attack deployed in the field. This is
a typical application of an EA system, whereby a Jammer is used to reduce the efficacy of an
adversary’s transmissions from a communication node. Adopted from [3] and modified by author
Some of the major kinds of EP measures are detailed below.

• Emission Control (EMCON) is the intelligent control that coordinates friendly
transmissions for a period of time. In particular, it limits access of certain trans-
mission sources at critical junctures as to not risk exposure for detection by enemy
ES and EA systems [3].
• Low Probability Intercept LPI (spread spectrum) signals is the use of deliberate
spreading of the transmitted data over frequency, and hence bandwidth. This is
intentional, so as to prevent any attempt of adversary ES and EA systems to
intercept, locate and jam signals [2].
• Screen Jamming is a clever, yet simple, technique to introduce RF energy between
friendly communication networks and the SIGINT systems of enemies to impede
the interception of transmission [3]. This is done by user specific jammer systems,
much the same as the EA jammer systems, but with a different objective.
• Encryption is the classical technique to ensure the fidelity of data transmission, in
cases where transmissions are intercepted and extracted for intelligence gains.
All these techniques are utilized to protect friendly communications being inter-
cepted or susceptible to adversary electronic tactics.
The discussion of EP serves as part of the overview for the ED domain, whereas
our focus will be restricted to ES systems and hence little more will be detailed on
EP herein this work.
2.2 Electronic Support Communication Applications 13
2.2 Electronic Support Communication Applications
The application of our work incorporates both ES and COMINT (a sub section of
SIGINT) tasks, as detailed in Sect. 2.1.2, with a focus on implementing the new
signal processing techniques using compressive sensing. As a result, the question of
which equipment platforms such techniques can be implemented on (i.e. digital or
analog), becomes a focal point.
In fact the typical equipment platforms that are used to perform ES and COMINT
tasks are communication electronic support (CES) and communication intelligence
(COMINT) equipments. The distinction between the two systems varies in terms of
both equipment architecture and purpose.
In this section we highlight the equipments needed for both systems, their appli-
cations and purpose. We also include common signal processing techniques used as
part of the processor unit. Then, we review how certain techniques such as emitter
identification, feature extraction, and classification are typically implemented for
communication ES purposes. Lastly, we discuss the implementation of spectrum
monitoring and direction finding techniques that are pertinent to our work.
2.2.1 Communication Electronic Support—CES
Communication Electronic Support systems provide immediate emitter signal infor-

mation to troops and other operators in order to make informed decisions in the
battlefield [139]. The functions that are tasked to CES equipment include:
• Search,
• Interception,
• Classification,
• Identification, and
• Direction Finding.
As pointed out earlier, the large frequency coverage as well as the congestion of the
RF spectrum requires a receiver system that can cope with the demand of complete
coverage of the intended spectrum. In the case of CES equipment, this is either
implemented by a wide-band channelized receiver or with a number of wide-band
Super-Heterodyne (SH) channelized receivers, coupled with the number of receiving
antennas, that rapidly scan over the RF bands [107]. These bands are known as the
instantaneous bandwidth (IBW) corresponding to the SH receiver channel width,
typically 40 MHz wide [139]. This receiver architecture is the most commonly used
one in practice for CES equipment [107] and serve as motivation for its review. We
refer the reader to the work done in [1, 8, 139] for further study on the working details
of different available architectures.
CES equipment typically comprise of the following sub systems (sourced from
[8, 107]). See Fig. 2.5 for the CES system block diagram.
Fig. 2.5 System block diagram for a typical CES system and equipment requirement. Sourced
from [107] and modified by the authors
1. Antenna array, composed of different sub-band antennas typically covering the

VHF, UHF and SHF bands
• These arrays typically form a sub-band circular array (UCA) useful for DF
methods [156], especially correlative methods. Further discussion of these
methods are detailed in Sect. 2.4.
• The UCAs are made up of an odd number of vertical poles to reduce the phase
ambiguities.
• An additional antenna used as a GPS antenna and receiver which provides
information about its global position, necessary for ground deployment, as
well as synchronizing a CES receiver with other systems in a CES sensor
network.
2. Antenna front-end (AFE) and sub-band array switching matrix
• Provides pre-selector filtering and low noise amplification of the received RF
signal before being passed to the RF receiver
• The switch matrix selects between the different sub-array of antennas in order
to perform DF measurement as the number of UCA channels are limited.
3. RF receiver
• The CES monitoring equipment exploits the SH architecture and channel
nature to do stepwise sweeping over the entire frequency spectrum. The IBW
is usually 40 MHz (as previously suggested).
• The sweeping speed is determined by the tuning time, typically in the order
of 10 s of microseconds.
Fig. 2.6 A two channel DRx shown performing detection and differential phase measurement using
two antenna channels from the IBW equipment. Sourced from [107]
• Adjusting the IBW tuning speed can be used to either increase/reduce the
frequency resolution for a specific purpose, such as detecting and locating
frequency agile signals, better known as frequency hopping (FH) signals.
• The required instantaneous dynamic range (IDR) is usually IDR > 60 dB, with
a signal-to-noise (SNR) ratio between 8 and 10 dB.
4. Processing Unit
• The IF signal from each channel of the RF receiver is then converted via a
12–14 bit ADC (providing the needed 60 dB IDR). All the channel data then
gets processed via FPGAs, which provide the DSP capability such that detec-
tion using a filter bank implementation and phase measurement algorithms
can be accomplished.
• The processing of these channels is shown in terms of processing blocks
in Fig. 2.6, which shows two channels from the UCA. The processing tasks,
shown as system blocks (i.e. Windowing, FFT, Detection etc.), are executed by
the FPGA. The blocks are merely system descriptions describing the process-
ing steps.
• Windowing applied to the data stream reduces the sidelobe response of fre-
quencies for Fast Fourier Transform (FFT) detection.
5. Axillary Units i.e. Human Machine Interface (HMI), computer, databases,
libraries and storage
• After preliminary detection, direction of arrival (DOA) and clustering are
performed. The auxiliary units use this information to perform classification
and feature extraction.
• Classification and feature extraction techniques are discussed in Sect. 2.2.4.
2.2.2 Communication Intelligence—COMINT
Communication Intelligence systems are used to analyse signal over an extended

period of time giving sufficient processing time to provide accurate intelligence
about signal characteristics and content [139]. Signal intelligence includes emit-
ter location, signal structure and the level of electronic counter counter measures
(ECCM) employed by an enemy to evade detection. These characteristics are then
used to compile and support mission control plans and provide insight to generate the
appropriate waveforms used in jamming enemy signals [40]. In this respect COMINT
is vital for jamming and protecting RF supremacy. The main functions ascribed to
COMINT are as follows.
• Signal Acquisition and feature extraction;

• Signal Classification;
• Signal demodulation;
• Voice signal demodulation, decryption and listening;
• Signal Recording; and
• Decoding, transmission standard recognition and speech recognition.
Furthermore, use for COMINT equipment extends to civilian application for spec-
trum monitoring, whereby surveillance of the spectrum use is monitored to deter-
mine if users are broadcasting within the legal specified bands [12]. Any RF emitter
broadcasting in a defined civilian area must comply with the regulatory body licens-
ing agreement that defines the broadcasting standards within a specific geographical
location. The Independent Communications Authority of South Africa (ICASA) is an
example of such a body. However, as far as DF tasks are concerned, COMINT equip-
ment typically does not include omnidirectional antennas and sub-systems needed to
perform DF. The DF (360◦ ) antennas are substituted for high gain directional anten-
nas in order to improve sensitivity and provide higher SNR for signal parameter
estimation [1].
It is worth mentioning that the architecture for COMINT equipment share simi-
larities with CES equipment including similar AFE, UCA, RF receiver, DRx (Dual
Receiver) channels, and processing units. However, as mentioned before, the DF
antennas are replaced by high-gain directional antennas and as a consequence the
DOA processing steps are omitted and additional number of processors, recording
systems (i.e. storage devices), and software tools are added.
2.2.3 Signal Processing Techniques
Any intelligent receiver system requiring classification and/or identification of signal

are highly dependent on some form of signal processing unit, and/or HMI (i.e. com-
puter) [68, 167]. Moreover, signal processing techniques—specific to transforms—
play a pivotal role in generating a representative base that allows different digital
methods to classify, identify, and extract features from RF signals in ES communi-

cation systems [42]. Some transforms that are integral to ES equipment include the
Fast Fourier Transform (FFT), Wavelet Transform (WT), Walsh Hadamard Trans-
form (WHT), and the Hilbert Transform (HT). A few of these tasks are described
below.
1. Filtering Methods based on cyclostationary signal properties [63] using an FFT
accumulation technique. A lower level implementation that does not involve fea-
ture detection, as cyclostationary techniques do, involve FIR filters that are quite
commonly seen as part of digital receiver systems to perform digital filtering of
signals.
2. Time-Frequency Analysis which is either implemented via a Wavelet Transform
or else, in some systems, via the Wigner-Ville distribution coupled with a quadra-
ture mirror filter (QMF) method. Time-frequency analysis of signals, especially
for ES systems, are vital to determine if a signal is frequency agile.
3. Signal Detection using a strip spectral correlation analysis (SSCA) [138] that
comprises of FFT blocks to achieve detection. Detection is ubiquitous with signal
processing of RF signals within ED. However, elementary detection techniques
such as threshold detection simply do not suffice for some kinds of signal. Hence
is the need for more probabilistic detection methods such as the SSCA and cyclo-
stationary signal processing that provide improved detection performance with
more deterministic parameters [7, 138].
2.2.3.1 Transforms
A transform, for the discrete case, is the process where an input signal is mapped from
one domain represented as real discrete values to another vector space. This mapping
process is the basis on which any transform is based. Transforms are ubiquitous in
signal processing. Herein we have selected the most frequently used and prominent
transforms used as part of the ES signal processing block. However many more
transforms exist and can be applied to ED, but such a discussion merits a study on
its own (Tables 2.1, 2.2 and Fig. 2.7).3
2.2.4 Signal Classification
Classification techniques in the equipment mentioned earlier serve to recognize spe-

cific characteristics about a received communication signal [139]. These are known
as classifiers of the system, whereby the output parameters from the receiver unit
of a CES or COMINT system (i.e. DF, frequency carrier and detection) are used as
3 Seethe following literature on the advances [165] and implementation of different transforms for
the uses in ES [32, 73, 110].
18
Table 2.1 A table detailing the expressions for the transforms and the associated complexity
Transforms Definition Fast algorithm Complexity
Name Complexity
Discrete Cosine Transform (DCT) DCT-II: N −1 π (FCT) O(N log2 (N )) O(N )
X k = n=0 xn cos (n + 1/2)k
N −1 N
Discrete Fourier Transform (DFT) X k = n=0 xn e− j2π kn/N (FFT) O(N log2 (N )) O(N )
X (m, w) =
Discrete Short Time Fourier Transform (WFFT) O(N log2 (N )) O(N )
∞ − jwn
(DSTFT) n=−∞ x[n]w[n − m]e
1
Walsh Hadamard Transform (WHT) (Hm )k,n = m/2 (−1) j k j n j where (FWHT) O(N log2 (N )) O(N )
2
k = i<m 0 ki 2i and n = i<m 0 n i 2i

Wavelet Transform (WT) Wψ , f (a, b) = (FWT) O(N ) − O(N log2 (N )) per scale O(N ) per scale

1 ∞ x −b
√ ψ f (x)d x
|a| −∞ a
2 Electronic Defence Systems
Table 2.2 A table detailing the advantages and disadvantages associated with different signal
processing transforms used to facilitate electronic support processing tasks
Transforms Application Advantage Disadvantages
(FCT) • Spectral methods • Efficient • Single basis to
• Lossy compression representation of represent signals
(i.e. MP3, Image signals • No phase
processing) • Smaller data length information
• Multiple DCT
variants (i.e.
I-VII-DCT)
(FFT) • Spectral methods • Magnitude & Phase • Assumption of
(i.e. Filter banks) information periodicity causes
• Demodulation • Efficient spectral leakage
• Broad application to representation of • No time information
RF detection, frequencies (exclusive to
identification, and frequency domain)
classification • Only two
representative basis
(sin & cos)
(WFFT) • Non-periodic signal • Provides • Computationally
analysis time-frequency intensive (Let time
• Sidelobe reduction information slots be M then
• Major reduction in complexity is
spectral leakage O(M N log2 (N ))
• Equal resolution for
time and frequency
transformations
(FWHT) • Frequency signal • Less • Reduction in true
processing computationally frequency
• Digital signal expensive than FFT, as representation
detection/estimation only 2 discrete states • Representative basis
• Used when signals exist (i.e. 2 bits) for leads to different
are choppy addition and frequency
subtractions interpretation
• Reduction in Bit
depth needed
(FWT) • Time-Frequency • Varying resolution • Prior knowledge of
signal processing for time-frequency the signal must
analysis transformations generally be known
• Signal detection, • Numerous wavelet • Computationally
estimation, and basis (i.e. Haar, more expensive
classification Daubechies, Coifman,
Symmlet) [37]
• Provides adequate
signal approximations
that have sharp
discontinuities
Fig. 2.7 Illustrates how signal processing in the context of ES maintains a hierarchy, in that the
lower level signal processing techniques all add up and aid the upper, more complex levels, in order
to accomplish the ultimate task of classification and identification of emitters
input to another system processor to classify the signal. Signal classification of the
signals include one or more of the following.
• Signal type recognition (i.e. Analog or Digital),
• Analog modulation recognition (i.e. AM, FM, PM),
• Digital Modulation recognition (i.e. FSK, PSK, ASK), and
• Type of multiplexing recognition (i.e. FDM, PPM).
Classifiers above are vital for emitter identification tasks.
In practice there are two major perspectives on classification for communication
signals that have been adopted in military applications [139]. These are based on
two algorithmic approaches, namely pattern recognition processing and decision
theoretic approach, sourced from [107, 139]:
• Pattern Recognition Processing—uses signal feature knowledge processed by

an artificial neural network type of algorithm [121]. Two major works that have
contributed to classification processing using neural networks include:
1. the Nandi-Azzouz classifier [121], that proved to be successful in distinguishing
between 13 analog and digital modulation types, namely AM, FM, FSK2-4,
ASK2-4, DSB, LSB, VSB, USB, PSK2-4, and combined (amplitude and phase)
modulation, and
2. the Assaleh-Farrell-Mammone classifier used to discriminate between differ-
ent digitally modulated signals [10] (The modulation types include CW, BPSK,
QPSK, BFSK, and QFSK.).
• Decision Theoretic Approach—depends on likelihood or probability to deter-

mine the modulation of a signal. Two of these decision-based methods include:
Fig. 2.8 Functional system block diagram describing the qLLR classifier used in the work by Kim
and Polydoros, taken from [135]
Fig. 2.9 System block describing the processing steps involved for emitter sorting process. Com-
piled by author, but sourced from [107]
1. the Sills classifier [161], which can discriminate between the three different
types of PSK (BPSK, QPSK, PSK-6) and three types of Quadrature Ampli-
tude Modulation (QAM) signals (i.e. 24 QAM, 25 QAM, 26 QAM) (The classifier
implements a maximum-likelihood (ML) algorithm for coherent classification
and validate the findings using a noncoherent ML version of the algorithm.);
and
2. the Kim-Polydoros classifier [141] which is an efficient means to discriminate
between modulation, relying on a quasi-log-likelihood ratio (qLLR) rule to base
decisions (Fig. 2.8).
2.2.5 Signal Feature Extraction
Signal feature extraction is a critical step, used after a signal has been classified, to
extract and catalogue the features that are assigned to the signal in order to identify
the emitter type [2], see Fig. 2.9. The process for feature extraction is similar to the
process used for emitter deinterleaving4 and sorting of radar warning receiver (RWR)
systems within ES [113].
Although the tasks are similar, to deinterleave and to sort the features based on
classifiers, the fundamental difference depends upon the emission description of the
signals [156]. In communication scenarios, emission descriptor words (EDW) that are
4 “Deinterleaving is a kind of clustering analysis, which clusters inter-weaved pulses—intercepted
by a scout or by other means—into distinct groups belonging to respective emitters, according to

the pulses’ features.” [88].
Fig. 2.10 System block diagram describing the processing steps involved to identify a RF emitter
assigned to communication signals are different in nature to that of radar signals (i.e.
RWR systems), as they are primarily pulsed [88]. Nevertheless, the process remains
similar, although the features assigned to signals differ. Typical features associated
with communication signals [139] which serve as an input to the clustering analysis
process (i.e. deinterleaving) are:
• Signal classification
• Frequency of operation
• RF bandwidth
• Modulation type
• Power Levels.
The features assigned to a particular signal are transformed into a digital word,
which is then passed onto the clustering analysis process based on a knowledge-
based algorithm which is mostly a histogram analysis method [88] (see Fig. 2.10 for
the sorting process). A key component of the feature extraction is to sort the EDW
into a subspace based on the parameters of the feature in order to generate a track file
of the emitters that are catalogued. This track file is then used as an input to identify
the emitter. The process of generating the track file is discussed in the following
section. Furthermore Fig. 2.10 adeptly contextualizes how the processing of an RF
communication signal, by means of detection, classification, and feature extraction,
enables an emitter to be identified. We refer the reader to the following literature [13,
145, 163] for further discussion.
2.2.6 Emitter Identification
Once the features of the signal emitter has been sorted and the accompanying track
file generated as depicted in Fig. 2.10, the following process takes over in identifying
2.3 Direction of Arrival Methods Used for Electronic Support Tasks 23
Fig. 2.11 System block diagram describing the flow process for emitter identification. Compiled
by author, but sourced from [107]
the most likely emitter based on a history of emitters as well as the database that
stores relevant parameters that comprise the details of a specific emitter [107].
Every parameter that is defined in the track file based on the EDW from the
signal features are used as a comparison to the measured values of an incoming
track. As depicted in Fig. 2.11 the process of identifying an emitter is based on
scoring the specific parameters based on the correlation from past emitters in the
database. Consider an emitter signal that has been received with a specific frequency
operation (FO) modulation type and signal power. The input signal is assigned with a
score based on the correlation measured of each of the values against emitters in the
database [108]. Then, the sum total score is calculated and if the total score exceeds
a selected threshold, the received signal can be identified with a probability relative
to the score, which is known as the confidence level [109].
2.3 Direction of Arrival Methods Used for Electronic

Support Tasks
DOA systems determine accurate estimates, within probabilistic bounds, of the direc-
tion from where a signal of interest (SOI) is received—also known as the line of
bearing (LOB). Once an accurate estimate of the signal DOA is determined a sec-
ond DOA must be acquired from a different geographical location, which can either
be done from a second receiver or a mobile intercept receiver. When two estimates
from varying locations are acquired the location of a communication signal can be
accurately determined, which is known as direction finding (DF).
It is important to distinguish between the signal processing task which constitute
DOA estimation methods and the operational task of DF. As explained, DF is the
operational objective in ES to use DOA estimates from multiple receivers to find the
direction of an EM source with high accuracy. In other words, DOA forms a part of
DF.
Main requirements of a DOA system are:
• High accuracy i.e. resolution less than 1◦ ;
• High sensitivity;
• Real time data capturing and processing;
• Short minimum requirement signal duration;
• Immunity to field distortion and polarization errors.
Both azimuth and elevation characteristics can be considered to provide a three-
dimensional estimate of a signal DOA. However, as elevation DOA is common
in air-to-ground scenarios, we do not continue the 3D discussion herein,5 we only
consider the case for azimuth DOA methods as we are interested in ground-to-ground
scenarios.
In this section we review conventional and modern DOA techniques used in ES,
and discuss why only phase interferometry is considered for our ES application,
especially communication DOA estimation. We then discuss modern DOA algo-
rithms used for ES systems and their associated performance to estimate the DOA
of a signal. Thereafter, we review the scalability of existing DOA estimation using
compressive sensing in the open literature.
2.4 DOA Methods
The two main DOA methods used in ES are amplitude comparison and phase interfer-
ometry [37]. Table 2.3 details the comparison of the two DOA measurement methods
and their associated benefits and drawbacks.
Amplitude comparison methods, as the name suggests, compare the amplitude
of a measured signal from multiple antennas [126] or in some instances a single rotat-
ing antenna [107], in order to determine the DOA of a signal. Although amplitude
comparison DOA is widely used, with adequate directional resolution and bandwidth
coverage, it is generally designed for pulsed transmissions. That is why they are pre-
dominantly used for radar warning receivers. Subsequently, the use of amplitude
comparison DOA in ES equipment is almost non-existent, except for some elec-
tronic intelligent (ELINT) equipment where rotating antennas are used [107]. As a
consequence, we do not develop amplitude comparison methods further in this work.
5 We refer the reader to the following literature for further reading on the subject of DOA and
elevation direction finding techniques [61, 21, 107, 126].
2.4 DOA Methods 25
Table 2.3 DOA measurement method comparison, sourced from [37]

Amplitude comparison Phase interferometer
Sensor configuration Typically 4–6 equispaced 2 or more RHC or LHC spirals
antenna elements for 360◦ in fixed array
coverage
θ 2 C λ
DF accuracy DFACC ≈ BW24S d B (Gaussian DFACC = 2π d cos θ θ
Shape)
DF accuracy improvement Decrease antenna BW; Increase spacing of outer
Decrease amplitude mistrack; antennas; Decrease phase
Increase squint angle mistrack
Typical DF accuracy 3◦ –10◦ rms 0.1◦ –3◦ rms
Sensitivity to High sensitivity; Mistrack of Relatively insensitive;
multipath/Reflections several dB can cause large DF Interferometer can be made to
errors tolerate large phase errors
Platform constraints Locate in reflection free area Reflection free area; Real
estate for array; Prefers flat
radome
Applicable receivers Crystal video; Channelizer; Superheterodyne
Acousto-optic; Compressive
CdB = Amplitude monoipulse ratio in dB; S = Squint angle in degrees; θBW = Antenna beamwidth
in degrees
Phase interferometry on the other hand, is widely used in ES equipment [37].

Phase interferometry is dependent on using multiple antennas, typically an array of
uniform spacing. The received signal’s phase between the individual receiving anten-
nas, is correlated to provide an estimate of the DOA based on the phase difference
[176]. Due to the breadth of implementation of this DOA method for ES, a review
of phase interferometry follows.
2.4.1 Phase Interferometry
Phase interferometry is considered as one of the best suited technique for communi-
cation signal DOA estimation [126]. If a scenario demands higher accuracy, in the
order of 0.1◦ –1◦ , the antenna spacing distance d = λ/2 (referred to as the baseline
width) can be decreased as it is relatively insensitive to phase errors [139]. More-
over, one can reduce phase mistrack by increasing the spacing of the outer antennas
in the array. Phase interferometry is relatively responsive, but requires complex RF
circuitry and processing when compared to other methods.
DOA phase interferometry system consists of the following components:
• an array of equidistant antennas which take various configuration—linear, circular
or lattice;
• front end channelized receiver circuitry, including mixers and analog-to-digital

converters (ADC); and
• digital signal processing back-end.
A typical configuration of this architecture is shown in Fig. 2.13.
Phase interferometry works on the following principle. Consider a plane wave
incident on a linear antenna array at an angle Φ, as shown in Fig. 2.12. Then based
on the geometry of the configuration the time difference of the signal being received
at every antenna can be expressed as phase difference. The phase difference can then
be be expressed in terms of the angle of arrival as shown in Eq. 2.1. Following the
notation in Fig. 2.12 gives:
d Δs
Φ=ω = 2π f ( ) = 2π(d sin θ )/λ, (2.1)
dt c
where
Φ phase of the signal,
ω angular frequency of the signal,
s vector component of the antenna separation distance,
f frequency of the signal,
θ angle of arrival,
d the antenna separation,
λ the wavelength of the RF signal,
c speed of light (3 × 108 ).
Fig. 2.12 An n length linear antenna array showing the 2 dimensional dynamics of phase inter-
ferometry, using phase representation of the time difference to solve the angle θ of an incident RF
wave. (Compiled by the authors.)
2.4 DOA Methods 27
Fig. 2.13 5-channel DF antenna processing architecture, taken from [144]
The phase information of the signal received at every branch of the receiver is
used by estimation algorithms to resolve the DOA. Before we review estimation
algorithms it is important to mention system considerations of phase interferometry
systems deployed in the field. Some of the system considerations restrict accurate
DOA estimation, and an awareness of them provide insight to which estimation
algorithm to choose.
A typical phase interferometric system using a uniform circular array (UCA)
dipole antennas is shown in Fig. 2.14. The processing back-end is shown in Fig. 2.13
implementing an FPGA back-end processing unit to perform correlative interfero-
metric estimation. This architecture and processing implementation in most cases is
considered as the standard approach for DOA tasks on modern ES equipments [176].
2.4.1.1 System Considerations
To deploy phase interferometry techniques in the field, there are several considera-
tions that have to be made, namely antenna spacing, coning error, system noise, and
signal-to-noise ratio (SNR).
Phase interferometric systems require minimal phase ambiguities as they distort
the accuracy of the field of view. When the antenna spacing is less than λ/2 the
field of view is 180◦ wide, which results from solving θ = 2 sin−1 (π/2d) [37].
Therefore, the spacing must match the highest frequency (i.e. smallest wavelength)
Fig. 2.14 Phase differences shown for two different incident waves for a 5-channel DF antenna
system, taken from [144]
of the received signal in the bandwidth of interest [149]. The restriction on the
antenna array for higher frequency cases (i.e. UHF/VHF communication) results in
ES DOA systems adopting multiple antenna arrays for various bandwidths of interest,
or resolving phase ambiguities with correlative algorithms; the latter solution being
more computationally expensive than the former.
When an EM source is elevated, for example an air-to-ground scenario, the ele-
vation of the incoming signal in relation to the receiver on the ground introduces
discrepancies in azimuth estimation for 2-Dimensional DOA task, which is known
as the coning error function [37]; adequately named because of the shape the locus
points create, which share the same phase delay.
Coning error adds to phase ambiguities when a signal is incident on the receiver
array at an elevated position, which in some cases can be large. Coning error can
be calculated by equating the ideal 2-D case φ = 2π(d sin θ )/λ with the 3D case
φ = 2π(d sin θ cos ϕ)/λ (with azimuth θ and elevation ϕ) which gives:
θ ∗ = sin−1 (sin θ cos ϕ). (2.2)
For cases when the emitter location is either on the horizon and/or restricted to
ground-to-ground application, the coning error is almost negligible [139] as elevation
increases. Fortunately for our application these effects are negligible.
The noise effects due to sensitivity and thermal noise contribute to the accuracy
in determining the DOA. The relationship for standard deviation of phase θφ relative
to noise is given as
1
σφ = √ , (2.3)
2S N R
which is then used to determine the common expression for angle accuracy using
interferometric techniques:
2.4 DOA Methods 29
c σφ c
σθ = = √ . (2.4)
2π di f cos θ 2π di f cos θ 2S N R
The restriction in the width between antenna elements in the array di , in order to
mitigate incorrect DOA estimation, requires a higher SNR of the system to process
and estimate the DOA of a SOI accurately. In some cases the required SNR could be
up to, and above 50 dB, which is unrealistic as interception systems operate in low
sensitivity environments [8].
Given such a high SNR demand, for certain tasks phase interferometry cannot
be used for ES. However, to overcome SNR demand, phase interferometric methods
in ES systems make use of circular harmonic (base-lengths are di = 2i−1 d1 ) and
non-harmonic (prime number multiples of base-length) antenna array with wider
baselines, which result in a lower SNR level requirement [37].
There are a variety of DOA estimation algorithms in the literature that are capable
of performing accurate DOA estimation given phase interferometric data. Several
estimation algorithms have seen successful implementation for ES systems, namely
the correlative interferometer algorithm (most widely used method) [15], multiple
signal classification (MUSIC) algorithm [158] and the estimation of signal parame-
ters via rotational invariance techniques (ESPRIT) algorithm [153].
2.4.2 DOA Estimation Algorithms
2.4.2.1 Correlative Interferometer [15]
The correlative interferometric method is based on a two step process. Firstly, the
respective phase differences between the antenna’s, respective to a primary antenna
(e.g. θ1 as shown in Fig. 2.14), are measured according to a known predefined bearing.
Then, based on a phase-history acquired during the system calibration of known
transmitter angles, the method performs a correlation between the different phase
measurements of the n-channel antennas and the stored phase history. The best
corresponding phase set is chosen for the phase of the received signal which results
in the correlative interferometric estimate of a incoming signal DOA.
The reliance on calibration history of some parameters make it susceptible to ele-
vation ambiguities and lower SNR, as well as lower resolution compared to MUSIC
and ESPRIT.
2.4.2.2 MUSIC [158] and Root-MUSIC [148]
The method of MUSIC as it applies to DOA was first formalized in [158] with beam-
forming [159] and maximum likelihood [191] DOA methods as seminal components
preceding its development. The algorithm is based on a probabilistic spectral search
method over all the angles in the subspace, using eigen decomposition methods to
Fig. 2.15 Histogram of

MUSIC and ESPRIT
results-random IO-element
linear array, source
correlation 90 %, small array
aperture (Δ = λ/4), taken
from [153]
resolve the DOA estimate. The search technique is computationally demanding and
therefore can be very expensive for some real-world applications. Developments
such as the alternative root-MUSIC algorithm [148] has shown to reduce the com-
putational complexity and improve estimation accuracy [137].
The conventional MUSIC algorithm, although computationally expensive works
for any antenna array configuration and multiple simultaneous RF signals. But it
remains vital for the algorithm to have knowledge—in terms of the spatial model—
of the positions of the antennas relative to one another. Furthermore, it is sensitive
to position, gain errors, and phase. Therefore careful consideration must be applied
for calibration.
2.4.2.3 ESPRIT [153]
ESPRIT is another estimation algorithm used for DOA estimation closely following
the MUSIC DOA method. The algorithm is based on a similar correlation matrix
generation and steering vector method as in MUSIC. The main difference is that
by using a non-singular matrix subject to the eigenvector noise subspace, a single
execution approach can be taken to determine the DOA instead of a search method.
This is sometimes referred to as a “one shot” approach.
Based on this single step process the computation for this algorithm is significantly
less as compared to MUSIC. However, due to the constraints imposed on the signal
model and matrix rank, the amount of antennas needed for ESPRIT is double that of
MUSIC which increases system cost. Furthermore, the use of total least square instead
of previous least square (LS) ESPRIT method reduces the error when SNR is low as
well as reducing error. The resolution of ESPRIT is reduced as compared to MUSIC.
Figure 2.15 shows the difference in DOA estimations for these two algorithms.
2.4 DOA Methods 31
Fig. 2.16 In a showing the actual and estimated DOA using both BCS snapshot methods as well
as the b error in terms of RMSE of the estimates with varying number of snapshots compared to
other DOA estimation methods. Courtesy [31]
In summary, where sub-space DOA estimation algorithms are considered for

our Compressive Sensing (CS) approach, MUSIC is the preferred DOA estimation
algorithm due to the lower requirement on the number of antennas compared to
ESPRIT, and higher resolution outcomes for DOA estimation.
2.5 Existing Compressive Based Direction-of-Arrival

Methods
In many respects compressive sensing (CS) based techniques used for DOA methods
are still in their infancy, which is interesting as CS theory is based on seminal works
from beamforming and super resolution [62] techniques.
In the open literature there are several CS methods used for DOA estimations,
which apply CS algorithms at various points in the processing chain. Majority of
the literature do not include CS sampling techniques for DOA estimation, rather,
only focus on applying CS recovery techniques. Such methods include the following
major work.
Bayesian CS based DOA estimation [31] develops a single and multi snapshot
approach using Bayesian compressive sensing (BCS) to estimate the DOA of a
narrowband signal. Rather than relying on compressive sampling, BCS determines
the estimates for DOA based on Nyquist-sampled voltage outputs directly from the
antenna elements. It was shown that by adopting this method, computation of the
covariance matrix for voltage outputs is not needed (as is the case with MUSIC [158]
and ESPRIT [153]). Also, robust and accurate estimates were determined without
the need for a-priori knowledge of the number of incident angles. However, the mag-
nitude estimates were somewhat degraded due to estimation error, but no such effect
to boresight-direction estimates were observed (see Fig. 2.16).
Fig. 2.17 DOA estimation of five simulated aeroplanes crossing the observation area. Left DOA
estimates from beamforming methods. Right amplitude distribution estimates by means of CS
methods. The true target position is highlighted by violet ‘x’ marker, whereas the estimates are
shown in green intensity points. Sourced from [52]
CS based radar DOA estimation [55] is another attempt to investigate the application
of CS for DOA estimation, specifically for radar (Fig. 2.17).
Even though this approach is successful in theory and has been shown to have
favourable results, the application value for implementation lacks system benefits
in terms of computational performance or sample reduction. In fact, it adds more
complexity to the system and requires additional processing time.
In summary, based on the open literature, there have been minimal investigations
as to how CS acquisition and recovery can be developed and applied to DOA esti-
mation tasks in ES resulting in optimized memory and computational use. This puts
the work of this monograph well in context.
Chapter 3
Compressive Sensing: Acquisition
and Recovery
3.1 Introduction
The pioneering steps taken toward digitization of signals can be attributed to the
theoretical work done by Kotelnikov, Nyquist, Shannon and Whittaker on sampling
continuous-time band-limited signals [87, 128, 160, 187]. Their work resolved the
issue of consistently recovering band-limited signals conditioned on the rate at which
the input signal is sampled, which later became known as the Nyquist Rate. This rate
empirically proved that a continuous band-limited signal can be accurately repre-
sented in the digital domain, if sampled at twice the highest frequency present [128].
The Nyquist rate remains the current convention for digital acquisition by means
of an analog-to-digital converter (ADC), whereas CS approaches the task of acquisi-
tion in a completely different way. Instead of restricting digital acquisition to twice
the highest frequency of the signal, CS acquires it by means of random sampling.
Thereafter, the randomized-sampled signal is used to recover the original signal by
means of linear optimization algorithms using sub-space modelling, which results in
acquisition at a lower rate than the Nyquist criteria under certain conditions. Hence,
CS is called a sub-Nyquist acquisition technique [53].
In this section we provide an overview of CS techniques as presented in the current
literature with respect to RF signal recovery and acquisition techniques. We discuss
the method of acquiring RF communication signal by CS techniques. Then, we shall
discuss the current CS acquisition schemes developed for RF signal acquisition and
propose the best suited scheme for DOA estimation. Thereafter, we shall review new
CS recovery algorithms that reduce memory and computation for applications in
DOA ES tasks.
The mathematical formulation of CS is extensive but vital in order to apply the
theory correctly. Therefore, we refer the reader to our theoretical review of CS math-
ematical formulation in the appendix 9.5, which describes signal requirements in
mapping to appropriate subspaces and the criteria on signal which result in a high
probability of recovery.

DOI 10.1007/978-3-319-46700-9_3
34 3 Compressive Sensing: Acquisition and Recovery
3.2 Compressive Sensing Formulation
Compressive sensing (CS) attempts to address the task of compression as it relates

to acquisition, storage and communication by means of randomization to reduce
the input data set [18]. Since its seminal work in [28, 44] CS has been applied to
various fields with varying success, viz. in signal processing [28], radar imaging [16,
69, 169], telecommunications [72, 140], data compression [174], image processing
[101, 152], and optical sensor applications [52].
CS theory provides the necessary conditions that a signal—being sparse in some
basis i.e. DFT, DCT, WHT—can be exactly recovered using a reduced set of samples
dependent on a sensing matrix satisfying the restricted isometry property (RIP) [44]
(refer Sect. 9.5.1 for further detail).
Taking an input signal F(t) with finite length N , which we represent as a vector
X [n] ∈ (N , 1) after sampling with its coefficients K N in some sparse basis.
Then as CS theory dictates, we choose an independent and identically distributed (iid)
Gaussian sampling matrix, Φ ∈ [M, N ], which when applied to the input vector X [n]
provides an output vector Y ∈ (M, 1). Once the vector Y is acquired—which is the
CS sampled signal—CS recovery algorithms solve the condition in Eq. 3.5, allowing
for the original input vector to be recovered with high probability, represented as a
vector ŝ.
The notations regarding CS acquisition and recovery, with respect to Fig. 3.1, are
shown below.
Y = ΦX (3.1)
The Fourier transform coefficients, denoted as s, are given for X based on the basis Ψ
comprising of the DFT matrix. Subsequently, the sensing matrix (see Sect. 9.5.3) is
composed of the DFT matrix and the randomized acquisition method for sampling.
X = Ψ ×s (3.2)
then Y = ΦΨ s letting ΦΨ = A (3.3)
gives Y = A × s. (3.4)
Fig. 3.1 CS measurement takes place for a vector x that is K-sparse in some other orthonormal
basis Ψ and sensed via a randomized sub-Gaussian matrix Φ where the number of measurements
M N . (Sourced from [18].)
3.2 Compressive Sensing Formulation 35
A typical method to recover the signal coefficients, which represent the signal,
involves solving the mathematical program:
ŝ = argmin||v||1 s.t Y = Av. (3.5)
Here v represents the solving vector being minimized in order to represent the final
estimated vector ŝ by means of iterative optimization. The equation above is a convex
optimization problem that can be solved via linear programming algorithms [150]
and several other algorithms, reviewed later in Sect. 3.4.
3.2.1 Compressible Signal and Sensing Matrices
A discrete signal is said to be sparse in a domain, if there exists a basis and/or frame
(Ψ ) that produce coefficients (α) that mostly comprise of zero coefficients [53]. If
this condition holds true, the sparsity of the signal can be exploited to compress the
signal for other applications. If we have a-priori knowledge that a signal is sparse in
some domain such as the Fourier domain for RF signal; we can use the knowledge of
that basis or frame to recover the input signal with reduced number of measurements.
In mathematical terms we describe a signal x as K -sparse when it has at most
K non-zero values, denoted as ||x||0 ≤ K . It is common to refer to a signal as K -
sparse, when in fact the signal x is actually K -sparse in terms of the representative
coefficients produced as a product of the basis and/or frame Ψ , i.e. x = Ψ α with
||α||0 ≤ K .
Most signal that we will be dealing with, in the RF domain, are rarely entirely
sparse for all applications. Thus, a better definition of compressibility is adopted in
describing a signal’s sparsity. The definition of compressibility of a signal x, requires
the sorted magnitude coefficients α—derived from the basis (or frame) Ψ —to decay
at a rate similar to that of the power law decay [150]. Importantly, if this definition
holds for x, it is compressible in the basis Ψ . The power law decay rate can be
expressed as:
1
| αs |≤ K q , s = 1, 2, ..., N , (3.6)
s
where K is an arbitrary constant, s is the sorted index, and q is the given rate of
decay.
One such compressible basis (i.e. Ψ ), which will be used extensively throughout,
is the Fourier transform, mapping time domain signal to a frequency-dependent
subspace with magnitude and phase coefficients [131]. However, the rate of decay
is not the only criteria to guarantee successful recovery by CS methods. For a basis
to be used for CS recovery it must also comply with the following criteria:
• an orthonormal basis;
• obey the null space property (NSP);
• obey the restricted isometery property (RIP); and
• have a lower bound for coherence.

Sparse Basis 0 if i = j
Orthonormal μ i· μ j =
(Ψ ) 1 if i = j
| < φ i ,φ j |
Def: μ(Φ ) =max
1≤i< j≤N
||φ i ||2 ||φ j ||2
Incoherence

Bound: μ(Φ ) = (2 log N)/M
Def: ||Δ (Φ x + e) − x||2 ≤ C||e||
1
Sensing Theorem: ||x||2 ≤ ||Φ x||2
Restricted Isometery Property C
Matrix (Φ )

N
Bound: M ≥ CK log
K
Spark (Φ ) > 2K
Null Space Property
||hΓ c ||1
Def: ||hΓ ||2 ≤ C √
K
Fig. 3.2 Illustrates the condition of a given basis Φ and the conditional requirements (i.e. Coher-
ence, NSP, RIP) for use in CS recovery, as well as the related sparse basis Ψ that has to be orthonormal
The representation in Fig. 3.2 diagrammatically relates the relationship of the basis
to the respective properties. We refer the reader to Sect. 9.5.3 where we detail the
theory for CS basis criteria.
If a matrix operates on an input vector X ∈ R N , producing a suitable vector
Y ∈ R M that allows for an unambiguous recovery of the input signal X [43] via CS
recovery algorithm then it is a suitable basis. This, as has been described above, is
possible only if it complies with the CS basis criteria.
Existing transform bases such as the discrete Fourier transform (DFT) does not
comply with the CS basis criteria, which is a problem for our investigation as the
DFT is pivotal to our digital processing goals. However, when a DFT is operated on
by an iid Gaussian matrix it results in an overall matrix that does comply with the CS
basis criteria. This result is further discussed in a later section, but it is important to
note that other discrete transforms such as DCT, WHT, etc. use the same operation
with an iid Gaussian matrix to achieve CS basis compliance.
3.2 Compressive Sensing Formulation 37
3.2.2 Implementation for 1-Dimensional Signals
The CS acquisition and recovery steps for a 1-dimensional input signal, using a
conventional RF receiver as the sensing system, in order to implement CS techniques
are entirely different from 2-dimensional signals typical with imaging equipment.
For the 1-dimensional case the operation of the matrix multiplication of the DFT
and iid Gaussian matrix is implemented by randomly sampling the input signal at
a sub-Nyquist rate where the total number of samples M = O(K log(N /K )) [35]
must be taken in order to guarantee successful CS recovery, with:
• M = number of CS samples required;
• K = the total number of sparse coefficients represented in the sparse basis i.e.
DFT; and
• N = the total number of samples of the input signal.
Once the input signal has been randomly sampled, a DFT matrix is applied to the
input signal to complete the CS sampled vector used for CS recovery of the original
N length vector X from the M length vector Y .
Figure 3.3 illustrates the system block proposed in order to apply CS for this work
in achieving compressive sampling with y[n] denoting the CS signal (Fig. 3.4).
Fig. 3.3 A basic block diagram of a CS RF receiver channel used to compressively sample and
recover a time domain RF input signal. Compiled by the authors
Fig. 3.4 a Is shows the frequency plot of the input time-domain signal f(t). In b the input signal
f (t) (in blue) corresponding to Fig. 3.3 and the CS random sampled signal (red) with c as the
recovered CS output estimate of the frequency spectrum of input signal f (t)
Fig. 3.5 a Illustrates the visual representation of compressed sensing via a matrix operation with
Φ the randomized sensing matrix operating on the input signal x. In (b) the approximation error
for two recovery techniques using different p strategies, namely Least squares (i.e. 2 ) and Basis
Pursuit (BP) (i.e. 1 ) are shown. (Courtesy of [78].)
As an aid to Fig. 3.3, the process that gets applied to the input signal x[n] can be
summarized by the matrix multiplication shown in Fig. 3.5 with the approximation
error denoted by p nor m with 0 < p < ∞.
Based on the discussion of the previous section, several conclusive conditions
apply to practical implementation of 1-dimensional signal processing. Given, that a
finite discrete signal x is K −sparse in some orthonormal basis, the sensing √ matrix
satisfies the RIP of order 2K , and has a low coherence of order K = O( K ). Then
its exact recovery, by some arbitrary recovery algorithm, is made possible by taking
M = O(K log(N /K )) [35] measurements. These conditions are depicted in both
Figs. 3.1 and 3.2.
It is worth mentioning here that the arbitrary recovery algorithm is based on the
proof and guarantee of l1 norm minimization [45]. The proof and guarantee predicts
that solving the optimization problem of the form below, will yield a solution that
optimally matches the K-sparse input signal.
x̂ = argmin||s||1 subject to y = As, (3.7)
where A is the product of ΦΨ and s the K-sparse vector. Note that x = Ψ s. Thus,
given the CS measurements y, x can be exactly recovered either with noise-free
y = Φx measurements (Theorem 4.1 in [150]) or measurements subject to noise
y = Φx + e (Theorems 4.3–4.4 of in [150]), bounded or Gaussian.
3.3 Compressive Sampling 39
3.3 Compressive Sampling
Before reviewing the techniques for CS recovery it is pivotal to determine which CS

sampling technique in the open literature serves as the best method for DOA estima-
tion of modulated communication signals assuming a conventional ES RF receiver
channel. In order to determine such a sampling technique we discuss conventional
sampling technologies used in modern RF receivers to determine a baseline for com-
parison. Thereafter we review several CS sampling techniques and determine the
best algorithm which is able to reduce memory and computational requirements.
Analog-to-digital converter (ADC)1 development lags behind the conventional
processing core (i.e. CPU) development by a large margin (i.e. Moore’s law [155]),
with sampling rates not doubling every year but every 2–4 years [173]. Slower devel-
opment of ADC technology is a major motivation to develop techniques that trade
processing power for sampling speed, in order to recover wider bandwidths. A variety
of techniques exist that exploit processing power and Nyquist theory to increase use
of bandwidth, reduce sampling rates, and remove the need for mixing and filtering
stages to reduce system costs. Such techniques comprise conventional mixing-filter
Nyquist sampling, bandpass sampling, direct sampling, and compressive sampling.
Although there exists several sub-Nyquist sampling techniques, which are mature
in their development; they only remain effective for special cases in relation to
low noise environments (bandpass), and limitations on ADC technology (direct).
Refer Table 3.1 for a quantitative comparison between such sampling techniques. For
our application CS sampling methods have a distinct advantage as they are custom
developed for CS recovery algorithms, and thus used in our investigation. Moreover,
CS has the capability of using lower rated ADCs to recover higher frequency sparse
signals with ever more bit depth resulting in improved dynamic range [18]. Perhaps
this is the must potential sub-Nyquist technique to break the barrier (i.e. red line in
b) of Fig. 3.6) which serves as the rate of innovation for current ADC technology.
In the open literature CS sampling techniques can be categorised into two groups.
The first consists of compression by modifying the input signal before sampling in
order to use a low rate ADC. The second, samples at rates equivalent to the Nyquist
frequency and then applies several techniques for compression.
Although both methods are applicable in theory, for practical purposes discussing
techniques that align with the former method, not the latter, are relevant to this
work. After all, the purpose for CS RF signal acquisition is to reduce the number
of samples needed before processing, in order to reduce memory and computational
1 Sampling a signal, in order to represent analog information (i.e. electromagnetic RF) in a digital
form, is done by means of an analog-to-digital converter (ADC) [185]. Many variations of these elec-
trical components exists, all sharing the same principle for acquisition but with varying techniques.
Additionally, some exhibit benefits over others in terms of bit depth and/or sampling rate. The ADC
types that exist and are widely used, include flash, sigma-delta, successive-approximation, ramp-
compare, and pipeline [183]. Current ADCs are capable of a conversion rate of up to 3.6 G S P S and
a bit depth of 12 bits. However these ADCs, although fast, do come at a price that for conventional
use in RF systems is exorbitantly high—in the range of > $4 000 per ADC, as of 2013 [76].
Table 3.1 Comparative table of different sampling techniques available for use in receiver systems,
compiled by the authors
Features Conventional Bandpass Direct CS
Number of samples (Memory load) M M H L
Bandwidth L-M L-M H H
Sparsity of frequency in spectrum L L-M L H
Resolution M-H M-H M-H L-M
Complexity L M H M
Computational load L M H M
Cost L L-M H L-M
KEY: H = High M = Medium L = Low
Fig. 3.6 a Shows the sampling rate versus bit depth for different variations of ADCs, and b the
rate of innovation for ADC technologies for different manufacturers. (Sourced from [116, 133].)
requirements. Therefore the requirements of CS sampling techniques that can be

applied for our application for DOA estimation of modulating communication signals
are:
• bandlimited CS sampling technique;
• able to simulate in software;
• practical implementation capability;
• able to utilize conventional CS recovery techniques; and
• minimized memory and computational requirements.
We refer the reader to Sect. 9.7 which deals with wideband CS sampling techniques
used for spectrum sensing tasks—as wideband CS sampling falls outside the scope
for our application.
Fig. 3.7 In a a system block is shown for the random demodulator analog-to-information sampler
method and in b the corresponding outputs with respect to time and frequency spectrum is shown.
The input signal is effectively mixed with random noise which shifts the spectrum by a relative
amount from the original; this is filtered and the original signal inferred by the CS recovery by
determining the shift of the spectrum. Sourced from [85, 173]
3.3.1 Random Demodulator (RD) Analog-to-Information

Sampler
The RD analog-to-information sampler utilizes a wideband pseudo-random demod-

ulator (i.e. a ±1 generator clocked at the Nyquist frequency) to mix with the input
signal f (t), apply a sample and hold (S/H) circuit that acts as a discrete low pass
filter, and finally sample the output from the S/H with a low-rate ADC (see Fig. 3.7).
The RD sampler has the effect of imposing randomization on the signal which
mimics an iid Gaussian RIP compliant matrix multiplication with the signal [85],
such that the recovery via CS is viable. The sampled signal y[n] can then be used to
recover the input signal f [n] via CS recovery algorithms, provided that frequency
support for f (t) is sparse. In Fig. 3.7 the distinct time and frequency signal spectrum
is shown at different stages of the RD process chain.
Work done in [85] suggested the feasibility of the RD sampler, and developed
further in [173] wherein it was shown to be an efficient CS sampling method for ban-
dlimited sparse signals. Moreover, a practical prototype of an analog-to-digital con-
version method, referred to as the compressive analog-to-digital converter (CADC),
based on the RD method was realized. Their work resulted in practical compressive
sampling and recovery using a 160 MHz DSP board of RF frequency sparse signal
up to 900 KHz with low-rate ADC sampling at 100 KHz.
A more recent practical implementation of the RD method was developed in
[190], known as the random demodulator pre-integrator (RDPI), shown in Fig. 3.8.
By adopting an eight channel RD architecture with a sub-channel correlated pre-
integrator design, high frequency sparse signal recovery was proven; with an effective
dynamic range of 54 dB, bandwidth coverage of 2.2 GHz and a sampling reduction
by a factor of 12.5 using a 90 nm CMOS chip design (refer [190]). However, this
technique is primarily used for radar applications, conditioned to short pulse trans-
missions, and has low applicability at ED bandwidths.
Fig. 3.8 System block of the RDPI, sourced from [190]
The RD method and its implemented RDPI extension is an effective CS sampling

method in reducing the sample size during acquisition. However, the inherent reliance
of the method on a large matrix inversion places a computational load on processing.
For example, taking a wideband frequency sparse bandlimited signal, requiring a
sampling rate of f N Q = 10 GHz, would need a matrix of dimension 2.6 × 108 by
1 × 107 for CS recovery—a large memory and computational requirement not prac-
tical for real-time processing at the current technology.
The constraint of bandwidth on the input signal impedes the RD method as a real-
time technique for use in wide bandwidths and high RF frequency cases, but feasible
for lower frequency operations. For scenarios where only a single bandlimited signal
is of interest, a pre-mixing stage can be added before the ADC to down convert the
input signal to an intermediary frequency (IF). Digitizing the IF signal requires less
memory and computation, making it feasible for real-time CS recovery.
3.3.2 Non-uniform Sampler (NUS)
The NUS attempts to randomly sample a signal at the level of the ADC, applying
innovative techniques to control the flow of data by means of S/H circuits before
quantization. In [182] a seminal prototype of an NUS IC device was developed,
using commercially off-the shelf components (COTS) for quantization and recovery
of signal from 800 MHz to 2 GHz sub-Nyquist sampling, using a 14 bit 400 MHz
ADC.
NUS relies on selecting, at random, integer multiples of the underlying Nyquist
rate allowing for corrective calibration, comparatively different to the random unre-
lated Nyquist sampling technique used by [91]. Nonetheless, using a S/H hold cir-
cuit controlled by a pseudo-random bit-sequence (PRBS) clocked at the Nyquist
frequency (i.e. 4.4 GHz), allows the NUS architecture to select and hold samples at
random, conditioned on the PRBS. The specifications for the NUS sampler, are a
bandwidth of 2 GHz, an occupied spectrum of 100 MHz, 5.8W power consumption,
Fig. 3.9 BER of the decoded GSM signal as a function of input power. Circular markers indicate
the performance of the uniform ADC for each of two randomly generated signals (denoted (1) or
(2)) at each of three levels of clutter (20, 50 and 100 MHz). Square markers indicate the performance
of the NUS on the same signals. The solid and dashed lines correspond to separate trials. (Taken
from [182].)
and a sample resolution of 8.8 ENOB with 55 dB of SNDR.2 The samples “held” by
the S/H circuits are then digitized by the under-sampling ADC and controlled by the
ADC sampling speed. Of every 8192 Nyquist rate samples only 440 are taken. There-
after, the samples are reconstructed and recovered on a desktop personal computer
with GPU hardware using a block algorithm.
The recovery method employed in [182] is non trivial as it requires multiple
interpolations, stitching, and windowing functions being applied before recovery of
the original signal is made possible. To some extent this complexity could serve as a
deterrent for practical implementation. However, the experimental evidence in [182]
suggests that real-time implementation is possible for high frequency bandlimted
signal with conditions on the spectral support and effective instantaneous bandwidth
(EIBW) dependent on the clocking frequency.
The experimental data indicate similar results for the bit-error rate (BER) of a
GSM input signal (see Fig. 3.9) when using the sub-sampling NUS architecture ver-
sus a conventional 4.4 GHz ADC. Moreover, the NUS technique allows for higher
bandwidth recovery than the RD and RDPI with significant improvements on recov-
ery of sparse signals.
A similar approach, following the same logic as the NUS sampler, known as the
random-ADC (RADC), was developed in [91]. It used a multiplexer and demulti-
plexer stage controlled via a PBRS, similar to [182]. The RADC method showed
2 ENOB refers to the effective number of bits and the SNDR denotes the signal to noise ratio +
distortion ratio.
Fig. 3.10 Showing the implementation of the RADC. (Taken from [91].)
promise for application in sparsogram and time-frequency applications. However,

practical implementation of the system has not yet been realized for high frequency
RF scenarios. Figure 3.10 shows the RADC implementation.
3.3.3 Compressive Multiplexing (CMUX) Sampler
CMUX, by definition, is a multi-channel architecture conceptualized in [162]

whereby a single ADC is used to sample the output of multiple channels. CMUX
has a dual approach to sub-sampling. The first approach relies on RF mixers to down
convert each channel to an IF. It then applies a psuedo-random chipping sequence
(PRCS), equivalent to [146], to each channel that has the effect of randomizing the
IF signal. Secondly, a bandpass sampling approach is taken, by first digitizing the
signal according to the respective Nyquist zones (see [4]), which is then mixed with
the PRCS.
Development of the CMUX draws from the RD and RDPI design with significant
differences in terms of parallel architecture and recovery. The recovery method differs
with other typical CS recovery methods, in that a condition known as joint sparsity
[51] must be used relating the sensing matrix to each branch in the CMUX scheme
for accurate recovery. The parallel architecture imposes a larger computational load
on recovery when compared with other common CS techniques.
A conceptual practical prototype is suggested in [182] with the remainder of
their work proven by simulation for low RF frequency signal applications (i.e.
10−20 MHz). However, using this method for wideband applications is possible
but not yet practically implemented.
3.3.4 Summary of CS Sampling Methods
For practical application of a CS sampling method, NUS and RDPI are the most
developed schemes scalable to the current DSP architectures. The RDPI implemen-
tation is preferable over the NUS sampling scheme for our application, due to less
complex software and sampling requirements, and the use of conventional CS recov-
3.4 CS Recovery Algorithms 45
Table 3.2 Comparison of the different CS sampling techniques reviewed for bandlimited signal
acquisition, compiled by the authors
Features RD RDPI NUS RADC CMUX
Software complexity L L H M H
Conventional CS recovery Y Y N Y N
Sampling complexity L L-M H M H
Computational requirements L-M M M M H
Reduction of samples M H H H H
Hardware implementation N Y Y Y N
KEY: H = High M = Medium L = Low
Y = Yes N = No
ery methods (detailed in a later section). See Table 3.2 for comparison of CS sampling
methods.
For simulation purposes RD, RDPI and RADC are adequate candidates as they
have similar computational requirements, sampling and software complexity, and
can use conventional CS recovery techniques. CMUX and NUS are effective CS
sampling techniques for simulation as well. However, their software complexity and
reliance on non-conventional CS recovery methods do not meet the requirements for
our application which depend on conventional CS recovery.
3.4 CS Recovery Algorithms
Once CS based sampling occurs, producing a sub-sampled signal y, it becomes the

responsibility of the CS recovery algorithm to produce an estimate of the input signal
x. There are numerous algorithms developed specifically for CS recovery, and can
be categorised into one of three classes; convex-optimization techniques, greedy
methods, and combinatorial methods.3 For adoption into the CS camp of recovery
algorithms, all sparse recovery methods must optimize the following.
• Minimal number of measurements.
• Robustness to noise and mismatch errors.
• Speed of computation.
• Performance guarantee on recovery.
3 Combinatorial techniques —developed by the theoretical computer science community [38]—

utilize the count-min, count-median or Bayesian methods. Combinatorial algorithms assume that
the origin of a signal of interest comes from a probability distribution, which imposes a belief of
propagation on the recovery [20]; or modelled for data network [72] and probabilistic learning
applications [79]. Thus, the relevance to our work—with exception to Fourier sampling in [64]—is
minimal, and we leave it to the reader to explore further.
Fig. 3.11 A work breakdown structure of the algorithms to be discussed; categorized according to
the two algorithmic groups, convex optimization and greedy algorithms, respectively. (Compiled
by the author.)
The type of CS recovery algorithm to use is dictated by the input signal considered.
For example, where the sparsity of the signal is known with high probability—greedy
algorithms are preferred. However, when the sparsity is unknown but the signal is
still sparse—convex-optimization algorithms are preferred. We discuss the reason
for the constraint of sparsity on the CS algorithm in the section to follow, by detailing
the CS recovery algorithms which constitute each category. Also, the advantages and
disadvantages of each algorithms are reviewed.
The objective of reviewing current CS recovery algorithms are to determine which
CS algorithm can be applied to our task of CS DOA estimation for modulated signals.
Specifically, we want to investigate which CS recovery algorithm can optimize speed,
memory, and computational requirements can match conventional DOA estimation
performance.
A work breakdown structure is provided in Fig. 3.11 revealing the relevant algo-
rithms reviewed for this work.
3.4.1 Convex Optimization-Based Algorithms
Convex optimization based algorithms adopt the definition of a convex problem,

where it is the objective to minimize a convex function f (x), for an unknown sparse
vector x, over a convex subset where f : x → R N [23]. Thus, based on the guarantees
of Eq. (3.7) the minimized output will result in the K -sparse input vector x. Although,
convex optimization algorithms employed by CS methods are categorically solved
by linear programs or iterative approaches, both adopt a similar set-up for solving
the convex problem.
Conventional convex optimization techniques do not provide adequate results
[150] when applied to CS recovery on their own. Therefore, CS specific sparse
recovery algorithms have been developed.
Recently developed CS recovery algorithms, based on convex optimization, take
the following approach. To recover an estimate x̂ given the measurements via y = Φx
: Φ ∈ R M×N a cost function J (x) = ||x||1 and a noisy cost function4 F(Φx, y) =
1
||Φx − y||22 are used. Based on basis pursuit de-noising (BPDN) [33, 100] an
2
adequate estimate of the K -sparse input vector x can be determined by solving the
following expressions.
Noise free case:
min{J (x) : y = Φx} (3.8)
x
Noisy case:
min{J (x) : F(Φx, y) ≤ ε}. (3.9)
x
where choosing a penalty parameter γ resolves this into an unconstrained case with
min x {J (x)} + γ F(Φx, y) with γ > 0 and determined by statistical trial [105].
3.4.1.1 Linear Programming
Linear programming (LP) is used to tackle the noise free implementation of the
1 − minimization problem, where a standard interior-point method [127] can be
used to solve the linear program in time complexity of order O(N 3 ) [150]. Although
exact recovery is guaranteed, the exponential growth in computational requirement is
undesirable for larger signal. Moreover, for practical purposes dealing with measured
signal, this approach is impractical for cases which include noise. Therefore more
robust algorithms have been developed in literature to solve problems with noisy
measurements. These, more robust convex optimization methods, are detailed in the
sections to follow.
4F cost function penalizes the difference in terms of Euclidean distance between the Φx and y in
vector form [150].
3.4.2 Fast 1 -Algorithms
For the case when noisy CS measurements are considered for recovery, a different
approach based on BPDN is adopted taking the form shown in Eq. (3.9) based on
the work in [30, 33]. More importantly, the results in [30] restructures the convex
optimization as a second order cone program (SOCP) [23] that can be solved by the
interior point method. Subsequently, this result has aided the development of fast
1 − norm algorithms.
A collection of the most relevant and widely reviewed algorithms in the open
literature, fitting the criteria of convex optimization, follows.
3.4.2.1 L1 Magic [26]
1 −magic comes as an algorithm package5 used for CS research that comprise two
fundamental solving algorithms, both able to robustly recover noisy CS measure-
ments using the interior point method described in Chap. 11 of [23]. However, as
one of the first CS recovery algorithms, and with the advent of newer algorithms,
these algorithms are comparatively slower, yet robust and accurate for CS recov-
ery. Nonetheless, they provide a solid theoretical introduction for other algorithms
wherein similar techniques are used, and therefore, mentioned herein.
The first, is a primal-dual algorithm for linear programming based on basis
pursuit and Newton’s iterative algorithm. This can, briefly, be described in terms of
a standard-form LP as in [26]:
minc0 , z subject to A0 z = b, (3.10)

z
f i (z) ≤ 0, (3.11)
where the search vector z ∈ R N , b ∈ R M , A0 is a M × N matrix with f i , i = 1, .., m

as a linear function
f i (z) = ci , z + di , (3.12)
for ci ∈ R N , di ∈ R. At the optimal point of the LP, there exists a dual vector
v ∈ R M , λ∗ ∈ R M , λ∗ ≥ 0 such that the Karush-Kuhn-Tucker (KKT) conditions
∗
are satisfied.
(KKT) c0 + A0T v ∗ + λi∗ ci = 0 (3.13)
i
Thus, in summary the primal dual algorithm finds the optimal z ∗ based on the optimal
dual vectors v ∗ and λ∗ by solving the above-mentioned non-linear equations. A
solution can be found by following the pseudo-code.
5 Distributed as open source code, written in Matlab and it can be accessed at [27].
Algorithm 1 Primal-Dual interior-point method [23]

Require: x that satisfies f 1 (x) < 0, .., f m (x) < 0, λ 0, μ > 1, ε f eas > 0, ε > 0,
repeat
1. Determine t, t ← μm/η̂
2. Compute primal dual search direction Δy pd
3. Line search and update
until ||r pi ||2 ≤ ε f eas , ||rdual ||2 ≤ ε f eas and η̂ ≤ ε
The primal, dual and central residuals provide a condition for the proximity of
(z, v, λ) which satisfies (3.13) in light of the slackness condition6 :

rdual = c0 + A0T v + λi ci (3.14)
i
rcent = −Λ f − (1/τ ) (3.15)
r pri = A0 z − b, (3.16)
where Λ is the diagonal matrix where Λii = λi , and f = ( f 1 (z)... f m (z))T .

The second approach, known as the log-barrier algorithm for SOCPs, is simi-
larly based on the work in [23] and, although SOCPs are more involved than LP, the
algorithm implementation is less involved than the primal-dual method mentioned
above. However, it is still founded on solving iterative Newton steps, much like the
LP problem. This can be written as follows (based on work in [26]).
min z c0 , z subject to A0 z = b, (3.17)

f i (z) ≤ 0, i = 1, ..., m (3.18)
where f i denotes the constraint (i.e. cost) taking the form of a second-order conic
1
f i (z) = (||Az||22 − (c1 , z + di )2 ). (3.19)
2
The log-barrier method modifies Eq. (3.17) into logarithmic form constituting a series
of linearly constrained programs, which can be expressed as:
1
min z c0 , z + − log(− f i z) subject to A0 z = b. (3.20)
τk i
It is important to mention that at log-barrier iteration k, by minimizing the series of

quadratic approximations generated by means of Newtons method, the algorithmic
approach begins to solve the system of equations. The quadratic approximations are
6λ f = 0, which converges subject to λik f i (z k ) = −1/τ k where the parameter τ k is increased

i i
progressively in accordance to the Newton iterations [26].
initialized around a point z based on Eq. (3.20) and expressed as
1
f 0 (z + Δz) ≈ z + gz , Δz + Hz Δz, Δz := q(z + Δz), (3.21)
2
1 1
where gz is the gradient given as gz = c0 + ∇ 2 f i (z), and Hz the Hessian
τ i
− f i (z)
matrix (see [26]).
Based on the above expressions the log-barrier algorithm pseudo code, to follow,
can be used as an outline to solve the SOCPs for CS noisy measurements.
3.4.2.2 1 -Least Squares (L1-LS) Regularization Algorithm [84]
L1-LS undertakes a different approach for solving CS measurements. This is also

known as a Gradient Projection method [189] and preceded by the Gradient Pro-
jection Sparse Reconstruction method in [59] on which the L1-LS is founded.
Consequently, the novelty of this approach stems from defining the convex opti-
Algorithm 2 Log-barrier algorithm for SOCPs [26]

Require: Feasible starting point z 0 , tolerance η, parameter μ and initial τ 1 with k = 1
repeat
if m/τ k < η then
α ← zk
else
τ k+1 ← μτ k , k ← k + 1
end if
until Solved (3.20) using Newton’s method subject to (3.21), using z k−1 as initial point. return
zk ← α
mization problem as a 1 -regularized least squares (LS) problem. By substituting the

sum of absolute values for the sum of squares used in Tikhonov regularization [65]
gives a general form,
min ||Ax − y||22 + λ||x||1 (3.22)
where λ is given as the Lagrange regularization parameter. Then, based on statistical

linear regression methods, a custom interior point-method similar to that of Eq. (3.20)
results, known as the Truncated Newton Interior-Point Methods (TNIPM). However,
the performance advantage depends on a preconditioned conjugate gradient (PCG)
step for initialization, whereas in [33] a LSQR [132] is used. The TNIPM therefore
transforms the 1 −regularized LS problem into a quadratic convex problem to be
solved, based on the following constraints:

n
min ||Ax − y||22 + λ u i s.t. u i ≤ xi ≤ u i , i = 1, ..., n. (3.23)
i=1
where TNIPM aims to solve a custom interior-point problem, given in [84] with
respect to the log-barrier method for bound constraints (3.23).

n
n
Φ(x, u) = − log(u i + xi ) − log(u i − xi ) (3.24)
i=1 i=1

n
φt (x, u) = t||Ax − y||22 + t λu i + Φ(x, u) (3.25)
i=1
An important result from [84] is based on defining a Lagrangian dual that places a
bound on an arbitrary x that produces a suboptimal x by constructing a dual feasible
point ν, so that G(ν) is the lower bound for the optimal value in Eq. (3.23) [84]. The
point ν is given as:
ν = 2s(Ax − y) (3.26)
s = min{λ/[|2((A Ax)i − 2yi )|]} , i = 1, ..., m
T
(3.27)
and the duality gap, known as G(ν), is used to determine the distance between the
objective value of x and the gap, which is denoted by η.
η = ||Ax − y||22 + λ||x||1 − G(ν) (3.28)
The Newton solution system is given as

Δx
H = −g (3.29)
Δu
with the Hessian H = ∇ 2 φt (x, u) and the gradient at a given iteration denoted as
g = ∇φt (x, u). All this, results in constructing the TNIPM Algorithm 3 as detailed
below.
Algorithm 3 Truncated Newton IPM for 1 −regularized LSPs. Courtesy of [84]

Require: Relative tolerance εr el > 0
Initialize : t := 1/λ, x := 0, u := 1
repeat
1. Compute the search direction (Δx, Δu) as an approximate solution to the Newton system
(3.29)
2. Compute the step size s by backtracking line search as in (3.26)
3. Update the iterate by (x, u) := (x, u) + s(Δx, Δu)
4. Construct dual feasible point ν from (3.26)
5. Evaluate the gap η from Eq. (3.28)
6. Update t
until η/G(ν) ≤ εr el
Finally, the development of the 1 −regularized LS algorithm7 proved to be an

improvement over the previous convex optimization algorithms (i.e. 1 −magic), in
that the recovery time could be reduced due to the initializing PCG step, prior to
solving the optimal estimate for x. This was shown in [84] by means of the problem
of MRI image recovery being investigated using a host of recovery algorithms.
3.4.2.3 L1-Homotopy [9]
One of the more recent algorithms, that have shown promising results for Fre-
quency Modulated Continuous Wave (FMCW) application using CS (see [9]), is
the 1 −homotopy algorithm.8 This method exploits the homotopy transformation
of the objective function (see Eq. (3.30)) from a 2 constraint to the 1 function.
Put differently, this method starts with an initial solution and finds a homotopy path
to the final solution. The progression along the homotopy path is governed by the
homotopy parameter, which corresponds to the two endpoints of the path given as
ε ∈ [0, 1) [9].
Given a CS measurement vector y = A x̃ + e, the 1 −homotopy algorithm solves
the 1 −norm minimization, by including a homotopy parameter and recasts the
1 −norm minimization as the following optimization problem:
1
min ||W x||1 + ||Ax − y||22 + (1 − ε)u T x (3.30)
x 2
where, by changing ε from 0 to 1 u can be defined as:
u = −W ẑ − A T (A x̂ − y), (3.31)
with W a diagonal matrix that has as its diagonals the positive weights w, and the
warm-start vector x̂ chosen arbitrarily given the corresponding matrix AΓTˆ AΓˆ . It
is important to realize that as ε changes from 0 to 1, the optimization problem in
Eq. (3.30) transforms into the standard 1 −norm, and therefore the solution follows
a piece-wise linear homotopy path towards the solution of
1
min ||W x||1 + ||Ax − y||22 . (3.32)
x 2
For optimal conditions the sub-differential of its objective function must be set to
zero [9]. The results in [9] and the homotopy optimization definition above leads to
the 1 − homotopy algorithm as shown in Algorithm 4.
The computational costs associated with this approach are significantly less, in
terms of time and iterative operations, than other state-of-the-art 1 solvers, namely
SpaRSA[188] and YALL1[9, 102]. Most of the cost relates to the update matrix
7 L1_LS source code (written in Matlab© ) and can be accessed in [129].

8 −homotopy code can be accessed at [129].
1
Algorithm 4 1 − homotopy algorithm. Courtesy of [9]

Require: A, y, W, x̂, and u (optional: inverse or decomposition factors of A Tˆ AΓˆ )
Γ
Ensure: x ∗
Initialize : t := 1/λ, x := 0, u := 1
repeat
Compute ∂ x (28) in [9]
Compute p and d (27b) of [9]
Compute δ ∗ = min(δ + , δ − ) (29) of [9]
if ε + δ ∗ > 1 then
δ∗ ← 1 − ε
x ∗ ← +δ ∗ ∂ x
break
end if
x ∗ ← +δ ∗ ∂ x
ε ← ε + δ∗
if δ ∗ = δ − then
Γ ← Γ /γ −
else
Γ ← Γ ∪ γ+
end if
until ε = 1
inverse operation and the update matrix factorization for A; with the complexity cost
in the order of M S + 2S 2 and M N + M S + 3S 2 + O(N ), respectively [9].
3.4.2.4 Fixed Point Continuation (FPC) Methods [70]
Another approach that falls within the domain of fast 1 −algorithms, is the fixed-
point continuation method, which deviates from the previous notion of the inte-
rior point method by applying a shrinkage method (a method applied to wavelet-
based denoising [150]). The FPC solves the 1 -minimization problem by defining
a convex-differentiable function H and employs an iterative shrinkage procedure.
Consequently, the 1 -minimization problem then takes the form of
min μ||x||1 + H (x), (3.33)

x
where the coefficient of x for the (k + 1)th time step is denoted as
xik+1 = shrink((x k − τ ∇ H (x k ))i , μτ ) (3.34)
with τ > 0 the step-length for gradient descent and μ is defined by the user
[70]. Moreover, specifying the typical convex cost function according to the resid-
ual squared norm gives H (x) = ||y − Φx||22 and its corresponding gradient
∇ H (x) = 2Φ T (y − Φx). Based on the selection of cost function, the program
is run until it converges to a fixed point and thus yields the sparse estimate vec-
tor x̂. The generic algorithm of FPC form is given below, in Algorithm 5, with the
respective penalty parameter used in Eq. (3.34).
Algorithm 5 Fixed point continuation algorithm. Courtesy of [70] & [150]

Require: CS matrix Φ, measurement y, parameter sequence μn
Ensure: Signal estimate x̂
Initialize :x̂0 = 0, r = y, k = 0
while alting criterion false do
1. k ←k+1
2. x ← x̂ − τ Φ T r {take a gradient step}
3. x̂ ← shrink(x, μk τ ) {perform soft thresholding}
4. r ← y − Φ x̂ {update measurement residual}
end while
return x̂ ← x̂
In previous studies it has been shown that FPC is a favourable candidate as opposed
to other recovery techniques based in the same category as 1 −minimization. It was
shown that algorithms based on FPC methods such as SpaRSA9 [188] and Fast
Iterative Shrinkage-Thresholding Algorithm (FISTA.10 ) [22] are able to out perform
algorithms such as 1 -LS in terms of computational time taken for recovery [98].
Moreover, in terms of direct comparison between the two best placed algorithms
using FPC techniques, FISTA outperforms SpaRSA by a computational factor of 4,
for Fourier based signal recovery [189] (a problem which is of importance for this
monograph).
3.4.2.5 Augmented Lagrange Methods [102]
Augmented Lagrange Multiplier (ALM)11 methods form a subclass of algorithms

included as part of convex programming. Instead of employing a cost function to
determine the optimal solution, concurrent estimates of the optimal solution and
Lagrangian multipliers are calculated at every step, to adjust for the solution out-
come [189]. By doing this the following 1 -minimization problems is recast as an
augmented Lagrangian function:
μ
L μ (x, λ) = ||x||1 + (λ, b − Ax) + ||b − Ax||22 (3.35)
2
9 Software package for SpaRSA algorithm can be accessed in [129].

10 Software package for FISTA algorithm can be accessed in [129].
11 Software package for ALM algorithm can be accessed in [129].
with μ > 0 used as a penalty parameter and λ given as the Lagrangian multiplier
vector. By increasing the penalty parameter μ the function can be written as the norm
of the residual, as in [23]. Therefore, the optimal solution x̂ can be expressed as:
x̂ = arg min L μ (x, λ̂). (3.36)

x
For an optimal solution to be reached efficiently, the chosen λ̂ must closely approx-
imate λ. Otherwise the iterative process can be exhaustive. Therefore an approach,
as discussed in [102], is adopted to compute an approximate estimate for both λ̂ and
x̂ for use in minimizing Eq. (3.35) rapidly. This process is expressed as:
xk+1 = arg min L μk (x, λk ) (3.37)

x
λk+1 = λk + μk (b − Axk+1 ) where λk+1 ← λ̂. (3.38)
It is important to note that this only becomes computationally feasible if the above
iterative process is less expensive iteratively than the minimization tasks of Eq. (3.35).
The algorithm known as YALL1,12 described in [102], is one of the seminal
works that applies ALM methods for use in CS recovery; and has shown success
over predecessor algorithms such as the 1 −LS.
3.4.2.6 Summary of CS Convex Optimization Algorithms
In summary, the 1 −homotopy algorithm provide the best performance for K -sparse
input signal when minimal CS measurements are available. However, when more CS
measurements are taken the FPC algorithm, namely FISTA, provide improved recov-
ery performance for computational time than 1 −homotopy. Therefore, depending
on the amount of CS measurements acquired, either of the two convex optimiza-
tion algorithms can produce optimized computational time. Work done in [22] sup-
port the previous statement—wherein several above-mentioned algorithms (1 -LS,
1 -Homotopy, SpaRSA, FISTA, ALM) were compared for computational recovery
time required for Fourier based CS measurements (see Table 3.3).
3.4.3 Greedy Algorithms
Greedy algorithms takes an entirely different approach to the problem of sparse recov-
ery via random CS measurements, by solving the non-convex program, expressed
as:
min{|ζ | : y = φi xi }, (3.39)
ζ
i∈ζ
12 Software package for YALL1 algorithm can be accessed in [129].

Table 3.3 Average run time for recovery for different 1 -fast algorithms. Courtesy of [189]
Corruption 0% 20 % 40 % 60 % 80 %
L1-LS 19.48 18.44 17.47 16.99 14.37
Homotopy 0.33 2.01 4.99 12.26 20.68
SpaRSA 6.64 10.86 16.45 22.66 23.23
FISTA 8.78 8.77 8.77 8.88 8.66
ALM 18.91 18.85 18.91 12.21 11.21
where ζ is given as the subset of indices i = 1, ..., N , and φi is the i th column of the
sensing matrix Φ. Thus, based on Eq. (3.39) the recovery technique applies a sparse
approximation to the actual signal, which is solved by greedily selecting columns
from Φ and forming a better fit approximation iteratively [150].
As discussed earlier in this chapter, for a recovery algorithm to be used in CS cer-
tain objectives need to be met namely. speed, robustness, performance guarantee, and
minimal measurements. Greedy algorithms, with the aid of a user defined estimate
for sparsity, significantly increase the recovery speed compared to 1 -minimization
algorithms at the cost of performance guarantee. Here the estimated output signal
approximates the input signal with less accuracy than its 1 -minimization algorithms.
However, greedy algorithms are still robust with regards to noisy measurements, and
require similar number of measurements.
In following section we detail the computational requirements, constraints, speed
and guarantee on performance of several Greedy algorithms. Thereafter we discuss
the constrains of applying Greedy algorithms for DOA estimation and comment on
the best algorithm to be used. Lastly, we compare Greedy algorithms with convex
optimization algorithms for use in our application which is CS DOA estimation.
3.4.3.1 Matching Pursuit [103]
Matching pursuit (MP) algorithm, first shown as a feasible solution to the sparsity
approximation problem in [103], is arguably the foundation on which most greedy
algorithms are based (in the field of CS) [105]. MP uses a given sampling matrix
Φ ∈ R M×N (otherwise referred to as a dictionary or basis) to construct a coefficient
index λk and residual r ∈ R M , where r is an iterative portion of the approximated
measurement, and λk selected from the basis [150].
The algorithm selects a vector, indexed by λk , from the basis that has a high
correlation with the residual r expressed similar to that in [103] as:
rk , φλ φλ
λk = arg max . (3.40)
λ ||φλ ||2
For each iteration of the algorithm, the following update for the approximation is
given.
rk , φλ φλ
rk = rk−1 − (3.41)
||φλ ||2
x̂λk = x̂λk−1 + rk , φλ φλ (3.42)
The update is repeated until a threshold is met, dependent on the norm r being
sufficiently smaller than a specified quantity (i.e. ||r ||2 < ε). The MP algorithm
is provided in pseudo-code below in Algorithm 6. We refer the reader to [103] for
further implementation details.
Algorithm 6 Matching Pursuit Algorithm. Courtesy of [103] & [150]

Require: CS matrix Φ, measurement y, stopping criteria ε
Ensure: Signal estimate x̂
Initialize :x̂0 = 0, r = y, i = 0
while alting criterion false do
1. i ←i +1
2. b ← Φ T r {form residual signal estimate}
3. x̂i ← x̂i − 1 + T (1) {update largest magnitude coefficient}
4. r ← r − Φ x̂i {update measurement residual}
end while
return x̂ ← x̂i
Based on the insight of [105] the MP algorithm cannot provide any guarantee on
recovery error for the estimate, and the requirement on iterations needed can become
computationally expensive. The complexity time of MP, given as O(M N T ) and
T denotes the number of MP iterations. Nonetheless, the approach of MP, given a
high sparsity signal, does provide favourable recovery time and can provide accurate
approximations for the signal x.
3.4.3.2 OMP [135]
Orthogonal Matching Pursuit (OMP) is an improvement on MP, due to the compu-

tational linear relationship on iterations T , that in special instances, has the potential
to grow excessively large. The proof of implementation using CS measurements and
sparse recovery was shown in [171] and founded on work in [135].
The novelty of the approach taken by OMP depends on modifying the residual r
in such a way as to remove unwanted portions of the residual to produce an improved
replica, at every iteration [171]. The residual modification is done by projecting r
onto a orthogonal subspace, based on the span of the sensing matrix Φ. The process
can be expressed according to the following sets of operations (based on [135]),
xk = arg min ||y − ΦΩ x||2 (3.43)

x
β̂t = ΦΩ xt (3.44)
rt = y − β̂t . (3.45)
The steps for r are repeated until the process converges. The pseudo-code steps are
shown in Algorithm 7.
In [171] it was shown that, convergence based on this approach for sparse recov-
ery, takes the form of O(M N K ) for computational complexity time. Consequently,
it proves that OMP is faster than MP and independent of the iteration. However,
guarantees on recovery are weaker than most convex optimization techniques, and
the robustness against noise is not clear. Moreover, the impact of large or small
noise additions can result in ambiguous recovery. Yet, if the sparsity K is small then
robustness can be insured.
3.4.3.3 CoSaMP
The adaptation of matching pursuit algorithms for CS based recovery (i.e. OMP,
MP etc.) forms a crucial base on which compressive sampling matching pursuit
(CoSaMP) is built, especially the precursory work based on regularized-OMP [124].
Interestingly, the approach of CoSaMP operates on an assumption that the RIP of a
given sensing matrix Φ of every subset K columns are roughly orthonormal [150].
This assumption results in a strong convergence with the added benefit of adding
and removing unwanted indices of the atoms chosen from the sensing basis. See
Algorithm 8 for further detail on implementation, represented by means of pseudo-
code. Currently, CoSaMP is arguably one of the fastest greedy algorithms for sparse
signal recovery using CS measurements [150]. It has been shown to require, under
specific sparsity conditions, a time complexity of O(M N ) for which it converges
independent of the sparsity level of the input signal. However, a priori knowledge
of the sparsity or at least a high probability of sparsity, denoted as K , of the sig-
nal is required. Otherwise, convergence and recovery guarantees are increasingly
ambiguous.
3.4.4 Summary of CS Recovery Algorithms
Having considered both classes of CS algorithms used for 1-dimensional signal

recovery, for the ideal case where sparsity levels can be accurately approximated
before signal acquisition, greedy algorithms out perform convex optimization algo-
rithms in terms of recovery time and computational load.
If a scenario exists where the sparsity of a signal is known for a desired sensing
basis, CoSaMP would be preferred over other greedy algorithms as it requires the
lowest complexity O(M N ) and is complete with the guarantee for recovery.
Algorithm 7 Orthogonal Matching Pursuit Algorithm. Courtesy of [135] & [171]

Require: CS matrix Φ, measurement y, the sparsity level m of the ideal signal
Ensure: Signal estimate ŝ in Rd , a set Λm ∈ Rm×d , residual rm = v − am
Initialize :r0 = v, Λ = , t = 1
repeat
1. Find index λt that solves
λk ← arg max |rt−1 , ϕ j |
j=1,..,d
If maximum occurs, break.
2. Augment the index set and the matrix of chosen atoms:
Λt = Λt−1 ∪ {λt }
Φt = [Φt−1 , ϕ j ]
3. Solve least square problem to obtain new signal estimate
xt ← arg min ||v − Φt x||2

x
4. at ← Φt xt , r ← v − at {update measurement residual}

5. t ← t + 1
until t ≥ m
j
return ŝ ← xt
If a sparse input signal is considered with unknown sparsity, convex optimization

are able to provide sufficient guarantees for accurate recovery,
Algorithm 8 Compressive Sampling Matching Pursuit Algorithm. Courtesy of [123].

Require: CS matrix Φ, measurement u, the sparsity level s of the ideal signal
Ensure: Signal estimate â of the target signal
Initialize :a0 ← 0 { Trivial initial approximation}
v ← u {Current samples = input samples }
k←0
repeat
1. k ←k+1
2. y ← Φ T v { Form signal proxy}
3. Ω ← supp(y2s ) {Identify large components}
4. T ← Ω ∪ supp(ak−1 ) {Merge supports}
5. b|T ← ΩT† u {Signal estimation by least-squares}
6. b|T c ← 0
7. ak ← bs {Prune to obtain next approximation}
8. v ← u − Φak {Update current samples}
until Halting criterion true
j
return ŝ ← xt
whereas greedy algorithms cannot. Thus, for cases where input signal spar-
sity is unknown, convex optimization algorithms are preferred, namely FISTA and
1 −homotopy (see Sect. 3.4.2.6). It is, however, important to mention that convex
optimization algorithms have a higher requirement on memory, computation, and
time complexity than greedy algorithms. Therefore, real-time application of convex
optimization algorithms are preferably processed off-line.
Part LXII
Simulations of Compressive Sensing Used
for Electronic Support Applications
Chapter 4
Design of CS Based DOA Estimation
for Modulated Shift-Keying Signal
4.1 Overview
In light of the data acquisition and recovery benefits that CS provides, as discussed
in the previous chapters, it remains a non-trivial task to adopt CS techniques in the
signal processing chain—from sampling to recovery—for DOA estimation. In order
to match the conventional DOA estimation performance there are several problems
that need to be considered when applying CS techniques. These problems include
phase recovery, choice of the sensing matrix, and choice of the recovery algorithm.
These problems are discussed below.
1. Phase recovery: From an RF signal processing perspective phase recovery seems
trivial, as one simply calculates the arctan of the complex components from the
DFT of an input time domain signal is expressed as

I m(X k )
∠X k = tan−1
Re(X k )
N −1 −2π j (kn/N )
where X k = n=0 xn e is the complex representation of the Fourier
transform of xn .
This results in a phase spectrum which, in general, is non-sparse when multi-
band signals are considered, with the exception of highly sparse frequency input
signals (see Fig. 4.1). Nevertheless, the correlation between magnitude and phase
coefficients remains non-trivial to determine under most circumstances for CS
phase recovery.1 In other words, the allocation of phase coefficients to their
1 Tothe best of our knowledge, based on the open literature, minimal avenues in terms of CS based
phase recovery have been investigated, with the exception of phase retrieval, which selectively
uses CS methods as part of recovering phase information given the Fourier magnitude data of
a signal [57]. Phase retrieval, traditionally, has a long-standing application in image processing
problems, particularly optics and x-ray crystallography [130]. Regardless of the success of CS as a

DOI 10.1007/978-3-319-46700-9_4
64 4 Design of CS Based DOA Estimation …
(a)
(b)
Fig. 4.1 A simple illustration of a the frequency and phase spectrum of a trivial input signal of a
frequency sparsity K = 1 determined using conventional methods. In b is shown the CS recovered
magnitude and phase spectrum using the l1_ls algorithm. Observe that the phase spectrum is non-
sparse and ambiguous for the case of CS phase recovery
corresponding frequency magnitude coefficients is difficult to determine, unless

the signal has frequency sparsity of k = 1.
Although CS recovery is well established for frequency-magnitude recovery of
sparse DFT based signals, due to the non sparse nature of the signal phase, most
CS recovered estimates of the phase spectrum are either poor or highly ambiguous
(see Fig. 4.1). Thus, an entirely different approach must be considered to enable
CS phase recovery by taking the signal structure into account, as DOA estimation
techniques require accurate phase information of the input signal.
(Footnote 1 continued)
candidate for phase retrieval [57, 104, 120, 130], and the similarities it shares with CS phase recovery
(due to only using Fourier magnitude data), its use in phase recovery is not trivial where sparsity
is conditional. This leaves the problem of phase recovery, by means of CS techniques as on open
problem according to the current literature.
4.1 Overview 65
2. Sensing matrix—For conventional RF communication signal processing the input

signal will be transformed by a DFT to the Fourier domain for further processing.
CS techniques employ the same technique by using a DFT sensing matrix during
the recovery process for similar RF communication signal. However, in order
to match conventional processing performance for large input signal requires
a large DFT sensing matrix for recovery which is computationally expensive
and therefore not viable in achieving similar performance (see Sect. 3.4). Thus,
an alternative CS sensing matrix and/or method must be found or developed to
match conventional processing performance.
3. Recovery algorithm usage—As mentioned in Sect. 3.4.4, the only viable algo-
rithms to use for CS recovery which match conventional DFT processing time
complexity are greedy algorithms. However, greedy algorithms require a-priori
information of the signal sparsity in some basis (i.e. sensing matrix) to guarantee
non-ambiguous recovery, which limits its application for communication signal
DOA estimation. The constraint on prior knowledge of sparsity for a communica-
tion signal can be addressed by limiting DOA estimation to scenarios for narrow
band signal and using a sensing basis where the input signal is highly sparse, in
the order of K = 1.
To address all three problems mentioned above, we propose using a generated
shift keying modulated sensing matrix (for amplitude—ASK, frequency—FSK and
phase—PSK) as the sensing matrix instead of the conventional DFT sensing matrix
used in similar RF communication CS recovery techniques.
By using a sensing matrix that consists of digitally modulated linear combinations
of finite length input vectors results in an orthonormal basis which complies to the
CS sensing matrix criteria (see Sect. 3.2.1); with sparsity of an input signal known
a-priori allowing the use of greedy algorithms, and recovery where sparsity is K = 1
allowing unambiguous recovery of phase.
Shift-keying modulated signals form the foundation for most digitally modulated
communication signal and has the potential to be scaled to more complex digital
modulated signal applications if our proposed method is successful. Thus, for simu-
lation purposes we only consider the case of 2-ary digital modulated signal. Higher
order digitally modulated signal is deemed out of scope for simulation purposes
within this work.
To summarize, our objectives as it relates to the problems outlined earlier include:
• Producing accurate phase recovered estimates of shift keying modulated signal for
DOA estimation;
• Frequency specific, narrowband application for digitally modulated communica-
tion signal in order to produce highly sparse input vectors for CS recovery;
• Optimize CS recovery by using greedy algorithms to match conventional process-
ing time complexity; and
• Reduce memory and computational requirements for DOA estimation whilst
matching processing performance with conventional DOA estimation techniques.
In the section to follow we detail the construction of the shift keying modulation
specific CS sensing matrices for our proposed method. We then detail the required
system parameters for use in CS based DOA estimation. Thereafter, we discuss our
simulation based implementation using an altered conventional DOA acquisition and
processing architecture.
4.1.1 Shift-Keying CS Sensing Matrix Development
To simulate a 2-ary shift keying modulated signal in discrete form, a finite length
input bit sequence is applied to a specific shift-keying expression—either frequency,
amplitude, or phase—to produce a discrete output vector. If this process is followed
iteratively spanning all m linear combinations of the binary bit stream a[n], one can
generate a modulation specific matrix.
The goal is to generate a square subset matrix, which replaces the DFT matrix
typically used for frequency CS recovery (see Sect. 9.5.3). Thus, we adopt the same
CS criteria for the shift-keying modulation specific subset matrices to ensure guar-
antees for CS recovery as well as phase recovery capability. These criteria are listed
below.
• Matrix must be square—M × M;
• Composed of complex sinusoidal elements; and
• The subset must be orthonormal.
We then denote this matrix as ψ2−ar y [m, n], with m = n where the generation
of ψ2−ar y [m, n] is done by calculating each successive row, based on the complex
forms of Eqs. 4.2, 4.5 and 4.9, for all binary combinations spanning 0 → 2 Nb − 1
and each row of the matrix corresponding to a specific binary sequence. In other
words, Ψ2−ar y can be defined as a linear independent orthonormal subset sampled
from complex sinusoids, determined by the form of shift keying modulation scheme
(i.e. 2ASK, 2FSK, 2PSK) and spanning the binary combinations of n b . Thus, we
have the following vector and sensing matrix notation for each shift-keying type.
2-Amplitude-Shift Keying (2ASK)
The discrete vector notation for 2ASK typically involves assigning two amplitude
values to a carrier sinusoidal waveform at a fixed frequency f c . In practice the ampli-
tude values are represented as voltages υ and span a limited range β, where each
voltage value can be denoted by
β
υj = j −β with j = 0, 1, ...L − 1. (4.1)
L −1
For our case the steps considered is limited to 2. For the discrete case, we have
the following expression for generating a 2ASK signal.

A 2π k
Sask [n] = (1 + vi [n]) cos[ n] (4.2)
2 N
4.1 Overview 67
where
Sask [n] = amplitude-shift keying output signal

vi [n] = amplitude related to binary input stream a[n], with
vi [n] taking on values of (−1; 1)
A
= carrier frequency sinusoidal amplitude
2
2π kc
= carrier angular frequency
N
kc = f c N
To generate an ASK output signal given an input bit stream a[i] of length Nb ,
where i represents the bit at a given sample time Tb , one can express the output
in terms of the input bit stream sampling rate. In other words, we want to obtain
samples of the bit stream a[i] for each given sample point of the digitally generated
2ASK signal, also known as upsampling [131]. This results in a re-sampled bit stream
written in discrete form as
a[n] = a[i/L] = a(i Tb /L), i = 0, L , 2L , ...., (4.3)
with L = N /Nb , which represents the samples per bit of the 2ASK output signal
reducing the period for sampling, denoted above as Tb . As a result, the number of
samples per bit are restricted to an integer multiple of the bit stream length, expressed
2 Nb
as L = (refer Fig. 4.2).
Nb
The matrix expansion of Eq. 4.2 for all linear combinations of 2ASK modulated
signal given a finite length bit sequence results in the sensing matrix, denoted herein
as Ψ2−AS K [m, n].
Given an input binary selection matrix A[m, n] where m = n and the rows of A
corresponds to the bit stream a[n] for a given binary range (20 )base2 − (2(Nb −1) )base2 ,
we can describe the 2ASK subset matrix as follows.
Ψ2 AS K [m, n] = (1 + A[m, n])(βe− j (2πkc n/N ) ) (4.4)
where {0 ≤ m ≤ 2 Nb m ∈ K ; 0 ≤ n ≤ 2 Nb n ∈ K }
and β is the amplitude of the carrier frequency
kc = f c N - where f c denotes the carrier frequency
2-Frequency Shift Keying (2FSK)

The discrete vector notation for 2FSK involves transmission of binary information
(i.e. a bit stream) in terms of discrete sinusoidal frequencies, which can be represented
as:
Binary Input Signal a[i]

2
1
0
−1
−2
0 1 2 3 4 5 6 7 8
Binary input Signal a[n]

2
1
0
−1
−2
0 50 100 150 200 250
2ASK modulation output signal

2
1
0
−1
−2
0 50 100 150 200 250
Fig. 4.2 A simple illustration for generating a 2ASK signal given a binary input stream. The input
bit stream a[i] is of length Nb = 10 whereas S AS K [n] is of length 2 Nb
S F S K [n] = A cos[2π(kc + k0 )n/N ], 0 ≤ n ≤ N − 1, for 1 (4.5)

S F S K [n] = −A cos[2π(kc + k1 )n/N ], 0 ≤ n ≤ N − 1, for 0 (4.6)
kc = f c N . (4.7)

2E b
A is the amplitude expressed in terms of A = with E b defined as the energy
Tb
per bit2 and f c denotes the carrier frequency.
The frequencies chosen to represent the binary information (i.e. k0 and k1 ) are
integer multiples of the sampling frequency 1/N (i.e. k0 = q/N and k1 = 2q/N
with q ∈ Z ). This ensures the output 2FSK signal is synchronized in phase and
periodic.
An example of this approach is shown in Fig. 4.3. For further reference to gener-
ation of higher order FSK signals, we refer the reader to [131].
2 Itmust be noted that for ED application purpose the signal amplitude is less relevant, as the final
joint matrix will be normalized for CS recovery. The amplitude is more relevant for specification
relating to Bit Error Rate.
4.1 Overview 69

2
1
0
−1
−2
0 1 2 3 4 5 6 7 8

2
1
0
−1
−2
0 50 100 150 200 250
2FSK modulation output signal

2
1
0
−1
−2
0 50 100 150 200 250
Fig. 4.3 A simple illustration for generating a 2FSK signal given a binary input stream. The input
bit stream a[i] is of length Nb = 10 whereas S F S K [n] is of length 2 Nb . (Simulated by the authors.)
The matrix expansion of Eq. 4.5 for all linear combinations of 2FSK modulated
signal given a finite length bit sequence, results in the sensing matrix denoted herein
as Ψ2−F S K [m, n].
We assume the same input binary selection matrix as for Ψ2−AS K [m, n] which
results in a 2FSK subset matrix and expressed as follows.
Ψ2F S K [m, n] = (2 −(A[m, n]−1))(βe− j (2π(kc +k0 )n/N ) )+((A[m, n]+1)−1)(βe− j (2π(kc +k1 )n/N ) )
(4.8)
2-Phase Shift Keying (2PSK)


2
1
0
−1
−2
0 1 2 3 4 5 6 7 8
2
1
0
−1
−2
0 50 100 150 200 250
2PSK modulation output signal
2
1
0
−1
−2
0 50 100 150 200 250
Fig. 4.4 A simple illustration for generating a 2PSK signal given a binary input stream. The input
bit stream a[i] is of length Nb = 10 whereas S P S K [n] is of length 2 Nb . (Simulated by the authors.)
The discrete vector notation for 2PSK involves transmission of binary information
using two phase states, where we use 0 and π respectively,3 which can be represented
as follows.

2π kc n
S2P S K [n] = A cos + qπ , q = 0, 1 (4.9)
N
kc = f c N , (4.10)
where q corresponds to the bit stream a[n] of length Nb . For the carrier frequency
we denote f c = m/T, m ∈ Z which is an integer multiple of the sampling time.
This ensures the output signal is synchronized and is void of phase discontinuities
other than 0 − π .
An example of this approach is shown in Fig. 4.4. For further reference on the
generation of higher order PSK signals, we refer the reader to [131].
3 Phase modulation states can be chosen as −π/2 and π/2. The choice of phase states can be
interchanged.
4.1 Overview 71
The matrix expansion of Eq. 4.9 for all linear combinations of 2PSK modulated
signal given a finite length bit sequence, results in the sensing matrix denoted herein
as Ψ2−P S K [m, n].
Given an input binary selection matrix A[m, n] where m = n and the rows of A
corresponds to the bit stream a[n] for a given binary range (20 )base2 − (2(Nb −1) )base2 ,
we can describe the 2PSK subset matrix as follows.
Ψ2P S K [m, n] = (2 − (A[m, n] + 1))(βe− j (2π kc n/N ) ) + ((A[m, n] + 1) − 1)(βe− j (2π(kc +k1 )n/N )(π ) )
(4.11)
When tested against the criteria for CS recovery—as outlined in Sect. 3.2.1—
requiring a sensing matrix to be orthonormal, we observe that matrix Ψ2 AS K does
not comply with the CS recovery criteria. Specifically, the columns of Ψ2 AS K are not
linearly independent from one another creating a non-orthonormal basis. Therefore,
we cannot pursuit the CS-DOA estimation for 2ASK digital modulation scheme,
further in this work. However, Ψ2F S K and Ψ2P S K conforms to the CS recovery criteria
and is used for DOA estimation simulations.
4.1.2 CS Recovery Method
The notation regarding CS acquisition and recovery for the general case, given a
shift-keying modulated signal as the input signal X , can be expressed as the sampled
CS signal
Y = Φ X, (4.12)
where Φ represents the randomized sampling matrix in the form as the notational
expansion of the sampling method described in Sect. 3.3. Furthermore, we can rep-
resent the shift-keying input vector X in terms of Ψ2ar y and the coefficient output
vector with sparsity K = 1, denoted as s, which corresponds to a single binary vector
from a column of Ψ2ar y ; expressed as,
X = Ψ2ar y × s (4.13)
then Y = ΦΨ2ar y s letting ΦΨ = A (4.14)
gives Y = A × s. (4.15)
The CS recovery mathematical program which must then be solved to determine

the coefficients of the shift-keying input vector is
Fig. 4.5 A simplified illustration of the CS modulation specific recovery method utilized for our
approach of CS based DOA and phase recovery
ŝ = argmin||v||1 s.t Y = Av. (4.16)
Here v is the update solving vector being minimized to represent the final estimated
vector ŝ.
Solving the mathematical program in Eq. 4.16 results in a CS recovered vector
output ŝ with sparsity K = 1 which contains the complex valued phase, magnitude
and binary index estimates of Ψ2ar y . The binary index corresponds to the row index
of the sensing matrix Ψ2ar y which we denote as the Binary Index Estimate (BIE).
To illustrate this approach, let us assume a bit stream of length 3, then we have
Nb = 3 and the sample length N = 23 = 8. We then choose the bit stream a[i] =
[1 1 0], and based on the CS recovery described above we should receive a vector
output for ŝ[n] = [0 0 0 0 a + ib 0 0 0 ] where the non-zero element of ŝ constitutes
the BIE, and the complex value element corresponding to magnitude and phase
components of the sampled signal. A simplified diagram of this method is illustrated
in Fig. 4.5.
The BIE serves as an elementary means of demodulation. Knowing the input vec-
tor sparsity a-priori allows optimal use of greedy algorithms for CS recovery simu-
lations. However, in order for successful CS recovery to occur using either Ψ2F S K
or Ψ2P S K , requires the carrier frequency and modulation type to be be identified
or known in advance so as to select and/or generate the corresponding shift-keying
sensing matrix. This places a limitation on application and scalability.
4.1.3 Implementing CS DOA Estimation
Our proposal to reduce cost on memory and computation, involves replacing the
conventional receiver-comparator architecture shown in Fig. 4.6 with that of a sub-
Nyquist CS acquisition receiver block (using the simulated CS acquisition RDPI
[190]) coupled with the CS recovery block, conditioned on the modulation specific
4.1 Overview 73
Fig. 4.6 A simplified block diagram of our approach in applying CS methods for modulation
specific DOA estimation. It should be noted that d ≤ λ/2 and input signals are still down converted
to baseband (i.e. IF) as for a conventional digital receiver. (Illustration compiled by the authors.)
sensing matrix. This results in CS resolved magnitude and phase estimates of respec-
tive signal of each channel. In conjunction with the steering vector this provides the
DOA estimates via sub-space algorithms.
The design of our proposed CS based DOA simulation is two-fold: first designing
accurate CS phase estimates for shift-keying modulated received signals; and second
investigating performance of sub-space DOA estimation algorithms given recovered
CS data. As shown in the previous section, phase recovery by CS means for a RF
signal is non-trivial and conventional CS recovery fails to produce non-ambiguous
results. Moreover, the DOA estimation algorithms that are able to integrate to the
system architecture proposed in Fig. 4.6 are limited.
We therefore adopt an approach which has the following assumptions to enable
modulation specific CS-DOA estimation:
1. Carrier frequency can be identified or is known in advance.
2. The signal is narrowband digitally modulated shift keyed.
3. DOA estimates are frequency specific.
4. DOA is dependant on digital modulation type.
Ultimately, our method must use signal data via an N -channel antenna array
receiver through a DOA estimator algorithm which outputs a bearing estimation for
an incoming shift-keyed modulating signal. We assume a system architecture4 as
shown in Fig. 4.6 and simulate the CS recovered scalar quantities per channel for
phase, magnitude, and BIE which are processed via sub-space DOA algorithms for
DOA estimation.
4 Thesystem blocks are simulated-only in MATLAB and have not been taken further for actual
implementation; this is deemed out of scope for this work.
Our implementation method follows on from the results in the previous section,
reliant on retrieving accurate phase estimates.5 On this basis our CS-DOA proceeds,
where the following steps describe the CS-DOA process corresponding to the system
blocks denoted in Fig. 4.6.
4.1.3.1 CS-DOA Steps
1. Input signal incident on a uniform linear array (ULA) antenna which is down
converted to base band (i.e. IF signal) for all channels via the N-channel receiver.
2. The IF signal is then randomly sampled by means of a simulated RDPI sampler.
3. The CS sampled signal is recovered via both greedy algorithms, viz. OMP and
CoSaMP. These algorithms are used for performance comparison purposes, their
time complexity being equal when an input signal has sparsity of 1.
4. The CS resolved signal output i.e. BIE, magnitude and phase estimates, are
recorded.
5. The CS phase and magnitude estimates become the input parameters for the sub-
space DOA estimation MUSIC algorithms to calculate the DOA estimates.
For the sub-space DOA estimation algorithm, namely MUSIC, we assume a signal
model with M signals incident on a ULA, given Gaussian noise associated with the
signals, which can be expressed in matrix form as:
x = SA + w (4.17)
A = [α1 , α2 , . . . ., α M ]
T
(4.18)
S = [s(φ1 ), s(φ1 ), . . . ., s(φ M )], (4.19)
where α is the input signals incident on the ULA of length N which results in A as
a N × M matrix. Also, S is the steering vector of size N × M. The goal is for the
sub-space algorithms (i.e. MUSIC) to produce orthogonal solutions for the steering
vector bearings (refer to Sect. 2.4 for further detail) which become the input signal
bearing estimates of the incidence angle.
Our method deviates from conventional estimation scheme, in describing the input
signal A in terms of CS recovered phase and magnitude estimates instead of time
domain input vectors.
5 Phase recovery for sub-space DOA estimation is much more critical for accurate bearing estimation
as oppose to magnitude estimates, (See Sect. 2.4).

Chapter 5
CS Based Shift-Keying Modulation
5.1 Simulation Outline
Shift-keying specific CS DOA estimation requires accuracy guarantees of CS recov-

ered phase and magnitude estimates before being used as input information for the
sub-space DOA estimation algorithms. Otherwise, if accuracy of phase and magni-
tude estimates cannot be guaranteed, the use of CS sampled and recovered informa-
tion for DOA estimation will fail. Only when the accuracy guarantees are determined
for shift-keying specific CS recovery, can we proceed to utilize CS recovered phase,
magnitude, and BIE estimates as scalar inputs to sub-space DOA estimation algo-
rithms.
In this section the performance and accuracy of the modulation specific shift-
keying CS recovery method (see Sect. 4.1.2) is investigated by means of simulation to
estimate the signal parameters (i.e. phase, and BIE) of a CS sampled modulated signal
(i.e. 2FSK and 2PSK) in a narrow bandwidth. The objectives of the investigation are
to determine the following:
1. Accuracy of phase recovery;
2. Accuracy of Binary Index Estimate;
3. Estimation performance in high noise environments (i.e. low SNR);
4. Most suitable CS recovery algorithm strategy;
5. Sampling compression ratio requirements for adequate CS recovery; and
6. Computational performance of CS recovery.
The input signal variables considered for simulations are SNR and the compres-
sion ratio (CR) of sampling. These variables are varied over a pre-defined range
typical for an ES receiver in a low SNR environment.
The performance indicators used for testing simulation outcomes are mean
squared error (MSE), Cramer Rao lower bound (CRLB), and probability of detec-
tion (PD), which provide a statistical base for CS recovery estimate performance
analysis. In addition, we test the time of computation for CS recovery estimates and

DOI 10.1007/978-3-319-46700-9_5
76 5 CS Based Shift-Keying Modulation
derive the number of floating point operations per second (FLOPS) for the respective
modulated signals considered.1
5.1.1 General Simulation Setup
The CS recovery block assumes that the respective basis matrix Ψ2F S K or Ψ2P S K for
recovery are generated given the carrier frequency of the input signals. The input
signal comes from a bit stream a[n] of length M = 256 (derived from the bit length
N = log2 M = 8 equal to a byte.), which can be denoted as a vector k of the sensing
basis Ψ [k, m] + w corrupted with Gaussian noise, denoted as w ∼ N (0, σ 2 ).
The bit sequences are considered as input signals with varied noise to represent
different SNR conditions. Also, for CS recovery the compression ratio (CR) for signal
sampling is varied over a selected range to determine its effect on the performance
of recovery. CR is ratio of CS sub-sampled vector length to that of the equivalent
Nyquist sampled vector length for the same input signal.
The output for all possible combinations of SNR and CR are measured against
performance indicators, namely MSE for phase and magnitude recovery, and proba-
bility of detection (PD) for the BIE. In addition, we use the Cramer Rao lower bound
(CRLB) as the ideal minimum variance of unbiased estimator (MVUE) to compare
with the MSE of our simulations. It can be noted here that the CRLB dictates the
lower bound of the best performing estimation achievable given an estimator MSE
[136].
5.1.2 Performance Indicators
5.1.2.1 MSE of Phase Estimates
Mean Squared Error (MSE) provides insights into estimation accuracy by measuring
the error of the actual signal parameter Y (i.e. phase) and the estimated parameter
Ŷ . MSE is calculated for each signal input length a[n], with iteration over the subset
of variables SNR and CR—their variable range is denoted as i, j respectively. Thus,
we can denote the MSE as follows:
1 2
N
M S E[i, j] = Y [n] − Ŷ [n] . (5.1)
N n=1
It is worthwhile noting that we can write MSE in terms of a variance of the estimator
distribution with a bias factor, as follows:
1 All simulations herein are simulated using MATLAB .

5.1 Simulation Outline 77
Fig. 5.1 A simplified diagram of the possible cases for various MVU estimators compared to the
CRLB, with the estimation value denoted as θ rather than Ŷ as described in Eq. 5.2. (Sourced from
[82].)
M S E(Ŷ ) = V ar (Ŷ ) + (Bias(Ŷ , Y ))2 . (5.2)
Therefore we can express the MSE of the phase estimates in terms of variance, com-
parable with the CRLB, where the respective CRLB for each estimate are derived2
as follows:
2
Phase : V ar (Ŷ ) ≤ . (5.3)
SN R × N
N represents the number of samples in the set and σ , the standard deviation of the
estimation.
If we assign the estimates for each respective compression ratio as a potential
MVUE, mapped over the range for SNRs, this provides a means of comparing the
estimation error with the CRLB for the same range, as a statistical bench mark test
for the estimation performance [83]. Figure 5.1 shows the various cases for a possible
MVUE and more interestingly, which comparison yields the most adequate result
for an estimator. Note that the closer the MVU estimate tends toward the CRLB the
more efficient it becomes.
5.1.2.2 Probability of Detection (PD) of BIE
Determining the probability of detection (PD) is trivial, in that the condition of

detection or correct recovery is measured as a summed boolean statement of the
correct corresponding BIE with the actual binary index of the input signal a[n]. This
can be expressed as

N
α = 1 if Y = Ŷ
P D[i, j] = α where (5.4)
n=1
α = 0 if Y = Ŷ .
2 For
the derivation of both CRLBs we refer the reader to [83] wherein the details are provided,
complete with worked examples.
5.1.2.3 Computational Performance
The computational performance is measured in terms of FLOPS which in theory,

given a computational platform, can be calculated according to the following expres-
sion.
CPU frequency FLOPS
No. Flops = Cores × × (5.5)
No. Cores Cycle
FLOPS are a standard measurement of computation in showing how a similar simula-

tion or algorithm will fare, given a different computational platform, say for example
an embedded system or DSP core.
However, from experimental knowledge the calculated number of FLOPS are typ-
ically much greater than the measured FLOP count during simulation [47]. Therefore
we conduct a bench mark test to determine the number of FLOPS during the simula-
tion and measure the time of computation for CS recovery as well,3 which provides
another measure for FLOPS and computational performance of CS recovery.
Simulations were conducted on a 4 core, 2.9 GHz Intel i7 platform.
5.1.3 Simulation Parameters
Figure 5.2 illustrates the typical output for a single iteration of the simulations to
follow, where magnitude, phase and BIE estimates (shown in red) are measured
against the actual signal parameters (in blue).
The simulation is based on takinga 2FSK signal,
with the carrier frequency known
a-priori, and varying the phase from − π2 : π2 corresponding to a 180 degree azimuth
range of a ULA receiver. The compression ratio is 15 % of the Nyquist sampling rate,
the SNR 5 dB, and the input vector length of 256 for this particular iteration of the
simulation. Several iterations of this simulation are carried out for 2FSK and 2PSK
signal according to the following signal variable parameters.
1. SNR [−5 dB : 20 dB]
2. CR of 3–36 %
3. The signal bit length M = log2 N = 8
4. Bit sequence varied in the range [0 : 28 ]
Only OMP and CoSaMP greedy algorithms are considered for CS recovery, due
to their computational efficiency properties when the input signal considered has the
sparsity of 1.
It should be noted that the phase estimates for Fig. 5.2 correspond to the actual
phase values accurately, and the BIE as well. However, the magnitude estimates do
not share the same accuracy levels of phase and BIE. Nonetheless, DOA estimation
3 The
bench mark tests for FLOPS/second are conducted using an open-source software, known as
Xbench.
5.1 Simulation Outline 79
Fig. 5.2 a Illustrates the phase recovery estimates, in red, with the actual phase shown in blue. In
b magnitude recovery estimates are shown in red, with the actual magnitude shown in blue. The
BIE estimate corresponds to the non-zero element in b with relation to the Ψ [m, n] sensing matrix
used for CS recovery. The input signal is a 2FSK vector of length 256 with a binary stream input
length 28 and modulated to equal a bit sequence equalling 28 = 1
algorithms are not dependent on magnitude information for accurate performance,

but require accuracy for phase for optimal performance. Therefore, we focus on
phase estimation further in this work.
Once all the various iterations of the simulations are complete for both 2FSK and
2PSK signal, according to the signal variable ranges, the calculated MSEs, PDs, and
FLOPs are analysed in line with the performance parameters. Thereafter, the results
of the analysis are used to determine the feasibility of using shift keying CS recovered
estimates for DOA estimation algorithms, which is covered in the next chapter.
5.2 Simulation 1.1.1—Phase CS Recovery for 2FSK
In this set of simulations CS phase recovery of 2FSK signal inputs were simulated
and resultant graphs drawn for the parameters SNR and CR.
Figure 5.3 depicts the detail of estimation MSE and how those figures compare to
the corresponding CRLB. Figure 5.4 serves as an aid to illustrate the relationship of
estimation error mapped over the variable range, providing an indicative character
of the estimates.
From Fig. 5.3a we observe CoSaMP producing MVU estimates for the various
CR values that are unbiased but not efficient. In (b) OMP produces improved MVU
estimates for the set of CR value that are unbiased and efficient for CR = 23 %, but
for lower values the respective MVUs are non efficient. For higher SNR values both
OMP and CoSaMP MSE tend towards the CRLB making them efficient regardless
of the CR.
(a) (b)
MSE of Phase recovered estimate for aFSK modulating input signal MSE of Phase recovered estimate for aFSK modulating input signal
and the CRLB vs SNR environment (low to medium) with varied CS sampling and the CRLB vs SNR environment (low to medium) with varied CS sampling
compression ratio as a percentage of original signal sample length. compression ratio as a percentage of original signal sample length.
CoSaMP greedy CS algorithm used for recovery OMP greedy CS algorithm used for recovery
0.7
1.2 CR−23 CR−23
CR−20 CR−20
CR−9 0.6 CR−9
1 CR−6 CR−6
CR−3 CR−3
CRLB 0.5 CRLB
0.8
0.4
MSE
MSE
0.6
0.3
0.4
0.2
0.2 0.1
−5 0 5 10 15 20 −5 0 5 10 15 20
Signal to Noise Ratio (SNR) in dB Signal to Noise Ratio (SNR) in dB
Fig. 5.3 The phase recovery Mean Squared Error (MSE) using greedy algorithms CoSaMP in a
and OMP in b, given an input signal that has type 2-Ary FSK modulation. The CRLB representing
the ideal MVUE and serves as the benchmark for recovery estimates. MSE scale of graphs are not
similar
5.2 Simulation 1.1.1—Phase CS Recovery for 2FSK 81
(a) (b)
MSE of Phase recovered estimate for aFSK modulating input signal MSE of Phase recovered estimate for aFSK modulating input signal
with varied SNR environment (low to medium) CS sampling compression with varied SNR environment (low to medium) CS sampling compression
ratio as a percentage of original signal sample length. ratio as a percentage of original signal sample length.
0.7
1.2
0.6
MSE of estimated phase

1 0.7
1.2
0.6 0.5
1
0.5
0.8 0.8
0.4
0.6 0.4
0.3
0.4 0.6 0.2
0.2 0.1 0.3
0.4
−5 −5 0.2
0 0
5 5
5 10 0.2 5 10 0.1
15 10 15
SNR in dB 10 SNR in dB
20 20
15 25 CR in % 15 25 CR in %
30 20 30
20
Fig. 5.4 Phase recovery MSE using greedy algorithms CoSaMP in a and OMP in b, given an input
signal that has type 2-Ary FSK modulation. The 3D graph shows the relationship of varying SNR
(y-axis) and CR (x-axis) on the accuracy of CS based phase recovery
The OMP algorithm outperforms the CoSaMP algorithm, in terms of MSE. As the
scales are not similar for graphs Fig. 5.3a, b, graphical illustrations on first inspection
might be misleading. However, notice that for all CR values, with SNR greater than
−8 dB, the MSE OMP is lower than CoSaMP.
For phase estimation of 2-Ary FSK using CS recovery, OMP serves as a more
efficient unbiased algorithm for DOA estimation. For higher SNR environments both
CS recovery methods work equally well.
5.3 Simulation 1.1.2—Phase CS Recovery for 2PSK
In this set of simulations CS phase recovery of 2PSK signal inputs were simulated
and resultant graphs drawn for the parameters SNR and CR.
Figure 5.5 depicts the estimation values compared to the corresponding CRLB,
whereas Fig. 5.6 serves as an aid in illustrating the relationship of the estimation error
mapped over the variable range of SNR, which provides an indicative character of
the estimates.
In Fig. 5.5a we see CoSaMP producing MVU estimates for the various CR values
that are unbiased but not efficient. In (b) OMP produces an improved MVU estimates,
for the set of CR values, which are unbiased but not efficient.
The OMP estimates are lower than the corresponding CoSaMP values in terms of
MSE. As the scales are not similar for graphs in Fig. 5.5a, b, graphical illustration on
first inspection might be misleading. However, notice that for CR lower than 16 %,
OMP MSE is lower than CoSaMP MSE, where CR is at its highest regardless of
SNR.
(a) (b)
MSE of Phase recovered estimate for aPSK modulating input signal MSE of Phase recovered estimate for aPSK modulating input signal
and the CRLB vs SNR environment (low to medium) with varied CS sampling and the CRLB vs SNR environment (low to medium) with varied CS sampling
compression ratio as a percentage of original signal sample length. compression ratio as a percentage of original signal sample length.
1.8
CR−23 1.8 CR−23
CR−20 CR−20
1.6
CR−9 1.6 CR−9
CR−4 CR−4
1.4 CRLB 1.4 CRLB
1.2 1.2
1
MSE
MSE
1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
−10 −5 0 5 10 15 20 −10 −5 0 5 10 15 20
Fig. 5.5 Phase recovery MSE using greedy algorithms CoSaMP in a and OMP in b, given an input
signal that has type 2-Ary PSK modulation. The CRLB representing the ideal MVUE and serves
as the benchmark for recovery estimates
(a) (b)
MSE of Phase recovered estimate for aPSK modulating input signal MSE of Phase recovered estimate for aPSK modulating input signal
with varied SNR environment (low to medium) CS sampling compression with varied SNR environment (low to medium) CS sampling compression
ratio as a percentage of original signal sample length. ratio as a percentage of original signal sample length.
1.8
1.8
1.6 1.6
1.4 1.4
1.5 1.5
1.2 1.2
1 1
1 1
0.5 0.5
0.8 0.8
0.6 0.6
−10 −10
−5 −5
0.4 0.4
0 5 0 5
5 10 5 10
15 0.2 15 0.2
SNR in dB 10 20 SNR in dB 10 20
15 25 15 25 CR in %
30
CR in % 30
20 20
Fig. 5.6 Illustrates the phase recovery MSE using greedy algorithms CoSaMP in a and OMP in b,
given an input signal that has type 2-Ary PSK modulation. The 3D graph shows the relationship of
varying SNR (y-axis) and CR (x-axis) on the accuracy of CS recovery for phase
For phase estimation of 2-Ary PSK using CS recovery, OMP serves as a more
efficient unbiased algorithm for DOA estimation deployment.
5.4 Simulation 1.2.1—CS Recovery of BIE for 2FSK
The BIE performance was simulated using CS recovery for all combinations of 2FSK
signal parameters, for a finite length input vector. The resultant graphs were drawn
for different SNR and CR, according to the simulation parameters (see Sect. 5.1.3).
Figure 5.7 detail the results obtained.
5.4 Simulation 1.2.1—CS Recovery of BIE for 2FSK 83
(a) (b)
Probability of detection of correct BIE estimate for aFSK modulating input signal Probability of detection of correct BIE estimate for aFSK modulating input signal
with varied SNR environment (low to medium). CS sampling compression ratio as a with varied SNR environment (low to medium). CS sampling compression ratio as a
percentage of original signal sample length. percentage of original signal sample length. OMP greedy CS algorithm used for recovery
CoSaMP greedy CS algorithm used for recovery
1
1
0.9
0.9
1
Probability of detection
1
0.9 0.8
0.9 0.8
0.8
0.8
0.7 0.7
0.7 0.7
0.6 0.6
0.6 0.6
0.5 0.5
0.4 0.5 0.4
0.5
0.3 0.3
0.2 0.4 0.2 0.4
15 0.3 15
10 10 30 0.3
5 30 5
0 25 0 25
0.2
−5 20 Signal to Noise Ratio −5 20
0.2
−10 15 Compression −10 15 Compression
SNR in dB (SNR) in dB
Percent Percent
Fig. 5.7 Binary index estimate (BIE) recovery probability of detection (PD) using greedy algo-
rithms CoSaMP in a and OMP in b, given an input signal that has type 2-Ary FSK modulation. The
3D graph shows the relationship of varying SNR (y-axis) and compression ratio (x-axis) on the PD
by means of CS recovery
In Fig. 5.7a we observe that CoSaMP estimates of BIEs for FSK input signals are
accurately estimated for CR ≥ 17% attaining a probability of detection (PD) of 1 for
the entire range of SNR. Virtually all BIEs will be correctly estimated when the SNR
and CR values do not exceed a CR ≤ 17 % and SNR ≤ 0 dB. Part b) in the same figure
shows that OMP estimates of the BIE, given the same input signal, have marginally
improved PD for CR ≥ 15. Similarly, BIE will be correctly estimated if the CR and
SNR values do not exceed a CR ≤ 15 % and SNR of ≤ 0 dB. As mentioned in the
previous chapter, if the BIE of an input vector can be correlated to the correct sensing
matrix row index, it would serve as a rudimentary form of demodulation using CS
recovery. Therefore, given that BIE for 2FSK can be achieved with high probability,
the use of shift-keying CS sensing matrix can provide additional signal information
without additional computation. Use of both OMP and CoSaMP for CS recovery
makes simultaneous demodulation capability of 2FSK plausible.
Notice that for both (a) and (b) (in Fig. 5.7) the PD remains high for SNR ≥ 5 dB
with CR tending to 10 %. Nevertheless, OMP serves as the superior BIE CS estimator
as it results in a wider operating range of SNR and CR whilst maintaining a PD of 1.
5.5 Simulation 1.2.2—CS Recovery of BIE for 2PSK
The BIE performance was simulated using CS recovery for all combinations of 2PSK
signal-parameters, given a finite length input vector, and resultant graphs were drawn
for given variable parameters SNR and CR, according to the simulation parameters
(see Sect. 5.1.3). Figure 5.8 details the results obtained.
In Fig. 5.8a, b we observe that neither CoSaMP nor OMP estimate BIEs for 2PSK
input signal accurately enough; PD of 1 is never attained for the entire SNR and CR
(a) (b)
Probability of detection of correct BIE estimate for aPSK Probability of detection of correct BIE estimate for aPSK
modulating input signal with varied SNR environment (low to medium). modulating input signal with varied SNR environment (low to medium).
CS sampling compression ratio as a percentage of original signal CS sampling compression ratio as a percentage of original signal
sample length. CoSaMP greedy CS algorithm used for recovery sample length. OMP greedy CS algorithm used for recovery
0.9
0.65
0.6
0.8
0.9
0.55
0.8 0.6
0.7 0.5
0.7 0.5
0.45
0.6 0.4
0.6 0.4
0.5 0.3 0.35
0.4 0.5 0.2 0.3
0.3
0.25
15 0.4
15 0.2
10 35 10 30
5 30 5
0.15
25 0 25
0 0.3
SNR in dB −5 20 SNR in dB −5 20
15 −10 15 CR in %
−10 CR in %
Fig. 5.8 Binary Index Estimate (BIE) recovery Probability of Detection using CoSaMP (a) and
OMP (b) greedy algorithm given an input signal that has type 2-Ary PSK modulation. The 3D graph
shows the relationship of varying SNR (y-axis) and Compression ratio (x-axis) on the probability
of detection by means of CS recovery
range. The low PD of the BIE estimates can be attributed to the phase modulated
structure of the sensing matrix vectors and row vector time delay similarities. The
sudden changes of phase between modulation segments creates non-linearity, which
introduces ambiguities for the greedy CS recovery algorithms as the sparsity is equal
to 1. Additionally, some row vectors of the sensing matrix can be represented by
other row vectors by adding a time shift to the original row vector, which is then
confused with the correct row vector (i.e. BIE) during CS recovery—reducing the
PD of the BIE estimates.
Even though a higher PD is achieved by CoSaMP than OMP, the condition for
demodulation was a PD of 1 for all BIEs. Therefore, our approach of shift-keying
sensing matrix CS recovery does not merit further development for BIE estimation
for 2PSK signal, as PD is too low for unambiguous BIE estimation.
5.6 Assessment of System Parameters for Shift-Keying CS

Recovery
Findings of CS recovery simulations for phase and BIE estimation performance are
summarized in Table 5.1 and regularly referred to for this discussion. The summarized
values provide insight for optimal system parameter choice for shift-keying specific
CS-recovery, where low-SNR environments are concerned.
For both 2FSK and 2PSK input signal we have shown that the phase estimation
performance of shift-keying sensing matrices using OMP and CoSaMP provides
sufficient accuracy, in terms of the MVUE criteria which is dependent on the CR
range. However, the more efficient algorithm to use, of the two greedy algorithms, is
Table 5.1 The summarized finding from the previous simulations and the operational range for modulation specific CS recovery for low to medium SNR and
compression ratios as small as 3 % of Nyquist sampling
Modulation CS Phase BIE Mod-CS operational range
type algorithm
Sim No. CR values MSE range Sim No. PD value 1
with reliable
MVUE (%)
CR SNR CR (for SNR (dB) Best Algo
BIE)
2FSK OMP 1.1.1 >3 0.01–0.42 1.2.1 >15 % All >15 % >−5 *
CoSaMP >6 0.02–0.6 >17 % All >17 % >−5
2PSK OMP 1.1.2 >4 0.01–0.2 1.2.2 NA NA – >−5 *
CoSaMP >9 0.03–0.16 NA NA – >−5
5.6 Assessment of System Parameters for Shift-Keying CS Recovery
85
OMP as it provides superior performance for phase estimates in terms of MSE, and
a lower requirement for CR.
Where 2FSK input signals are considered for DOA estimation using shift-keying
sensing matrices for CS recovery, the simulation outcomes dictate the following
operational variable range for accurate phase estimates:
1. CR ≥ 3 %,
2. SNR ≥ −5dB
For the case where we need BIE estimates along with 2FSK phase estimates the
operational range for CR will have to be increased to 15 %, the SNR range remaining
the same.
Where 2PSK input signals are considered for DOA estimation using shift-keying
sensing matrices for CS recovery, the simulation outcomes dictate the following
operational variable range for accurate phase estimates:
1. CR ≥ 4 %,
2. SNR ≥ −5dB.
It is important to note that phase estimates are required for DOA estimation, not
BIE estimates. BIE estimates are a like an extra parameter that can be estimated
using shift-keying matrices, which provides a means of demodulation as the BIE
represents the demodulated input signal in terms of its indexing.
Based on the simulations and the summary, as shown in Table 5.1, accurate phase
estimates, given shift-keying modulated 2FSK & 2PSK signal, are realizable using
CS techniques with the estimation performance margin (i.e. MSE and MVUE criteria)
dependent on the CR and SNR. CS phase estimates for such signal are possible for low
SNR environments, typical for ES receivers. However, for accurate BIE estimation
the system performance deviates, requiring an increase in the CR from 3 % to 15 %
and the guarantee of BIE estimates is only possible for 2FSK signals.
5.7 Demodulation Capability
Based on the values for BIEs from Table 5.1, we can assert that for CR ≥ 15 %
using OMP and only for 2FSK, BIE can be achieved. In addition we observe that the
computational performance requires 5 MFLOP for recovery (see Fig. 5.9).
The estimated BIE can be used as a rudimentary form of demodulation as it
corresponds to the correct modulation sequence of the row vector in the sensing
matrix. Therefore, to determine if the BIE can be used as such, it needs to be compared
with a conventional yet similar demodulation scheme. For this reason, we assume
a GSM input signal which uses GMSK for digital modulation equivalent to 2FSK
modulation.
In a typical GMSK demodulation process the input signal is demodulated via a
demodulator block done in real-time, which then immediately disqualifies our CS
recovery demodulation as competitive. Nevertheless, we know that GSM uses time
5.7 Demodulation Capability 87
Measure of computational performance in terms of FLOPS of aFSK

modulating input signal for a range of Compression ratio as
a percentage of original signal sample length.
OMP & CoSaMP CS algorithm used plotted for comparison.
7
5 x 10
OMP
4.5 CoSaMP
3.5
FLOPS for recovery
2.5
1.5
0.5
0
0 5 10 15 20 25 30 35
Compression Percent of original sampling
Fig. 5.9 Shows the computational performance in terms of FLOPS as it relates to compression
ratio for CS recovery. The input signal is of type FSK
division multiple access (TDMA) allowing a multiple channel transmission at the

same carrier frequency at 271 Kbits/s. Therefore, to demodulate the same input
signal using CS means, results in a FLOP count that can be implemented on a DSP
core. Considering that CS-recovery results in 5 MFLOPS for a bit length of 256, we
develop the following expression to determine the FLOP count required:
CS-FLOPS No. × transmission speed (k B/s)

Required Flops = (5.6)
CS bit length
5 × 10 × 271 × 103
6
= (5.7)
256
= 5.29 × 109 . (5.8)
At present similar FFT based deployment on currently available FPGA platform

yields 400 GFLOPS [6], whereas similar DSP deployment can achieve over 200
GFLOPS [75]. Therefore demodulation for modulation specific CS is achievable
using a FPGA or DSP platform, which only requires 5.29 GFLOPS where a GSM
signal is concerned. Thus, demodulation using the BIE of the CS recovered esti-
mate can be achieved using digital processing platforms at a comparable cost as
conventional demodulation algorithms.
5.8 Simulation 1.3.1—Computational Performance of CS

Recovery for 2FSK Signal
To determine the computational performance for 2FSK signal, simulations were

undertaken to find the number of FLOPS required given a 2FSK signal with varying
CR. Resultant graphs were obtained based on the simulation outcomes for different
values of CR. SNR is not considered for these simulations as computational perfor-
mance is independent of SNR. Figure 5.9 details the results obtained. From Fig. 5.9 it
is observed that the OMP algorithm (in blue), for this application, is computationally
more efficient than CoSaMP as it requires less FLOPS. As expected, the FLOPS
required increase with increasing CR. This is due to the increase in sample length
requiring more FLOPS for CS recovery. The operational range of FLOPS required
by OMP, given a 2FSK input signal of length N = 256, varies from 5 MegaFLOPS
to 11 MegaFLOPS and is dependent on the CR.
We therefore observe that OMP provides improved computational performance
for 2FSK signals. Thus, the use of OMP for CS based DOA estimation is further
motivated and preferred for deployment in terms of computational performance and
accuracy (see Simulation 1.1.1).
5.9 Simulation 1.3.2—Computational Performance of CS

Recovery for 2PSK Signal
In order to determine the computational performance for 2PSK signal, simulations

were undertaken to find the number of FLOPS required for CS recovery using OMP
and CoSaMP, given a 2PSK signal with varying CR. Resultant graphs were obtained
based on the simulations for different CR. SNR is not considered for these simulations
as computational performance is independent of SNR. Figure 5.10 details the results
obtained.
From Fig. 5.10 it is observed that the OMP algorithm (in blue), for this application,
is computationally more efficient than CoSaMP. The computational operational range
of FLOPS required by OMP, given a 2PSK input signal of length N = 256, varies from
5 MegaFLOPS to 10 MegaFLOPS which is dependent on CR.
We therefore observe that OMP provides improved computational performance
for consideration in our CS recovery approach for 2PSK signals. Thus, the use of
OMP for CS based DOA estimation is further motivated and preferred for deployment
in terms of computational performance and accuracy (see Simulation 1.1.2).
5.10 Computational Performance of CS Recovery 89
Measure of computational performance in terms of FLOPS of aPSK

modulating input signal for a range of Compression ratio as
a percentage of original signal sample length.
7
OMP & CoSaMP CS algorithm used plotted for comparison.
4.5 x 10
OMP
4 CoSaMP
3.5
FLOPS for recovery
2.5
1.5
0.5
00 5 10 15 20 25 30 35
Compression Percent of original sampling
Fig. 5.10 Shows the computational performance in terms of FLOPS as it relates to compression
ratio for CS recovery. The input signal is of type PSK
5.10 Computational Performance of CS Recovery
As CS-recovery is the only computational block that is additional to the conven-

tional Nyquist scheme, we are interested in the comparison of this to the conven-
tional schemes and whether the reduction in samples required (i.e. compression
ratio/percentage) provides an adequate return on computational time to make our
shift-keying specific CS method a competitive approach for deployment.
For both the OMP and CoSaMP algorithms the computational load (i.e. time
complexity) grows linearly at O(M N K ) for OMP and O(M N ) for CoSaMP (refer to
Sect. 3.4). However, it should be noted that due to the condition of sparsity K = 1 the
computational loads of both OMP and CoSaMP are equal, in terms of time complexity
O(M N ) where M is the size of the input vector and N the size of the square sensing
matrix. Therefore, computational performance has to be determined by means of
testing with regards to the specific application, which is depicted in Simulations
1.3.1–1.3.2 yielding a lower computational load for OMP—with a calculated range
between 5 and 11 MFLOPS.
To evaluate the computational load of a similar Nyquist scheme, we consider a
conventional FFT algorithm to produce phase estimates for comparison, which has
an associated time complexity of O(N log2 N ) [164]. Therefore, our approach for
CS-recovery would have to match a similar order of computation, which leads to the
the condition of similar time complexity expression where:
Compression percentage required to match similar

FFT simulation computational load via CS means.
60
Compression Percentage
50
Compression percentage
40
30
20
10
0
0 100 200 300 400 500 600 700 800 900 1000
Sample Length
Fig. 5.11 Compression ration required to match the equivalent FFT implementation via CS means
for phase recovery
M N = N log2 N (5.9)
M = log2 N . (5.10)
Thus, the CS sensing matrix size would have to comply with the above expression
in order to match the conventional FFT time complexity, where the compression
ratio, denoted as a percentile, is C R = 100 × M/N . The condition where the
conventional FFT approach matches our CS recovery approach follows the curve
shown in Fig. 5.11.
Therefore, the criteria of compression ratio (expressed as a percentage) required
to match the time complexity of the conventional FFT approach, given binary bit
lengths of an input signal for both 2FSK and 2PSK, is as follows.
• 256 3 %
• 512 1.7 %
• 1024 1 %
If comparative computational time using our CS recovery technique is the goal,
for sample lengths larger than 256, compression ratio smaller than 3 % is required.
The only scenario where the constraint on CR can be matched is for 2FSK phase
estimation using OMP for recovery (see Simulation 1.1.1). For scenarios where
higher CRs are required, the cost of computational load (i.e. processing time) must
be weighed against the reduction in sample size.
5.10 Computational Performance of CS Recovery 91
In summary the use of CS-recovery using shift-keying sensing matrices in con-

junction with the OMP algorithm, for phase estimates, provide similar computa-
tional performance as compared to the conventional FFT scheme when small sample
lengths are considered. For sample lengths ≤ 256—with our current phase modu-
lation CS scheme—comparative computational performance are realizable. Where
larger sample lengths are concerned Fig. 5.11 should be referred to for computational
performance comparison.
Chapter 6
Modulation Specific CS DOA
6.1 Chapter Outline
Based on the results of the foregoing chapter, wherein accurate phase estimations
were achieved using shift-keying sensing matrices for 2FSK and 2PSK signal, we fur-
ther our investigation by using CS recovered phase estimates for DOA estimation—
referred to as CS DOA in this chapter.
In this chapter the performance of the proposed CS DOA method is investigated
by means of simulations. The aim is to accurately estimate the DOA of modulated
shift-keying signals in a narrow bandwidth given CS phase estimates. CS sensing
matrices, developed in the previous chapter, are used to estimate the phase and thus
the DOA of SOIs by means of sub-space algorithms (i.e. MUSIC). These simulations
are structured to address the following tasks. The aims are to investigate
• the accuracy of CS DOA estimation compared to conventional Nyquist sampled
DOA estimation using similar sub-space algorithms (i.e. MUSIC);
• the estimation performance in high noise environments (i.e. low SNR);
• the compression ratio required for adequate CS DOA estimations; and
• the scalability of CS DOA estimation for ES.
Simulations are based on the method described in Sect. 4.1.3, taking the CS recov-
ered output data generated from simulation 1.1.1 (for 2FKS signals) and simulation
1.1.2 (for 2PSK signals) for N number of channels, assuming an ULA antenna as
inputs. Then sub-space DOA estimation algorithm MUSIC is used to determine the
CS DOA estimates. For comparison, we simulate a conventional Nyquist sampled
2FSK and 2PSK signal, via the same ULA antenna, and use MUSIC to determine
the DOA estimates. Similar SNR and CR values are considered for these simulation
as in Chap. 5.
All simulations herein are performed using MATLAB .1
1 The Phased Array System Toolbox is utilized for DOA estimation and CS outputs verified by means
of a custom written DOA MUSIC estimation algorithm.

DOI 10.1007/978-3-319-46700-9_6
94 6 Modulation Specific CS DOA
6.1.1 General Simulation Setup
The ULA antenna simulated comprise N elements spaced length d ≤ λ/2 apart,
where λ corresponds to the carrier frequency of the modulating signal input. Two
scenarios are considered of the ULA antenna, where N = 3 (the minimum number
required to estimate bearing) and N = 10 which represent a more typical deployment
of a conventional ULA.
A digital modulated shift-keying signal (i.e. 2FSK and 2PSK) is generated for
a specific bearing ranging from −75 : 75, which is then received at the simulated
ULA. A corresponding delay on each channel dependent on the bearing of incidence
is also imposed.
Each channel resolves the input signal by means of CS, as per Simulations 1.1.1
& 1.1.2, which provide the CS phase and magnitude recovery estimates for each
channel. Thereafter CS recovered estimates become the input parameters to DOA
estimation algorithm MUSIC. The DOA estimators use the CS DOA method. The
simulation of the MUSIC based CS DOA estimation follows on as per Sect. 4.1.3.
The shift keying input signal, given an incident angle, can be denoted as X [n] =
S(φ)a[n] + w[n] where the input modulating signal a[n] of length M (derived from
the bit length N = log2 M = 8 equal to a byte) can be denoted as a vector k
of the sensing basis Ψ [k, m]. S(φ) is the steering vector which is used in MUSIC
to determine the optimal DOA estimate and w ∼ N (0, σ 2 ) white Gaussian noise
corrupting the input signal. SNR is varied by changing the w and the CR by changing
M.
The output DOA estimates for the range of SNR (0 dB − 20 dB) and the range of
CR (1 % − 36 %) values are simulated and the results measured for the performance
indicators, viz. the MSE of the CS DOA estimates as compared to the actual DOA.
Figure 6.1 illustrates a single iteration of a simulation where SNR and CR are fixed,
which produce the CS DOA estimates (in black) and conventional DOA estimates
(in blue) compared to actual DOA (in red) of a 2FSK modulated input signal incident
from a bearing ranging [−80 : 80].
6.1.2 Performance Indicators
Mean Squared Error (MSE) is used to measure the difference between the actual
signal parameter Y (i.e. DOA incident signal) and the estimated Ŷ (using CS-DOA).
MSE is calculated for the set of signal angles N = [−80 : 80] of the input shift
keying modulated signal a[n], and with iterations over the subset of variables SNR
and CR with their variable range denoted as i and j respectively. Thus we can denote
the MSE as follows.
1 2
N
M S E[i, j] = Y [n] − Ŷ [n] (6.1)
N n=1
6.1 Chapter Outline 95
CS−DOA MUSIC, normal MUSIC and the true DOA estimates

for several FSK modulating input signal with SNR 1.1dB and
a compression of 3.9% of original samples.
3.5
2.5
Magnitude in dB
CS MUSIC
1.5 Norm−MUSIC
ACTUAL
0.5
−0.5
−80 −60 −40 −20 0 20 40 60 80
Bearing in degrees
Fig. 6.1 This figure illustrates the estimated CS DOA using our CS method versus the normal
Nyquist sampled DOA and the actual DOA for specified SNR and CR
MSE can also be written in terms of variance of the estimator distribution coupled
with a bias factor, as follows.
M S E(Ŷ ) = V ar (Ŷ ) + (Bias(Ŷ , Y ))2 (6.2)
6.1.3 Simulation Parameters
The following simulation parameters are chosen for the set of CS DOA estimation
simulations. The motivation for the choices are detailed as well.
1. Azimuth ranged in the limit [−80 : 80]. The range excludes the full 180◦ field of
view because phase ambiguities are high for bearings higher than ±80.
2. Elevation kept at 0 because our scope only considers azimuth for this investigation.
3. Distance between antenna elements is kept at d ≤ λ/2 which is required to
mitigate phase ambiguities.
4. SNR ∈ [0 dB : 20 dB] which is the typical SNR range for ES receivers.
5. CR is varied in the range of 1–36 % that of Nyquist sampling i.e. (ratio of: [0.02–
0.3]). This is based on results CS recovery operational range from previous chapter
(see Sect. 5.6).
6. Antenna ULA is of 3 and 10 elements. (Refer to Sect. 6.1.1.)
7. MUSIC sub-space DOA algorithm is considered. Although other sub-space DOA

algorithms are available, such as ESPRIT, MUSIC has a lower requirement on
the number of antennas and is more efficient and accurate.
8. The signal bit length M = log2 256 = 8. (Refer to Sect. 5.6.)
9. OMP algorithm is considered for CS recovery. (Refer to Sect. 5.6.)
6.2 Simulation 2.1—CS DOA for 2FSK Signals
CS DOA estimation of 2FSK signal was simulated using MUSIC estimation algo-
rithm and the resultant graphs are drawn for given variable parameters SNR and CR.
Separate simulations were done where the antenna elements in the ULA are 10 and
3. Figure 6.2 details the results obtained.
In Fig. 6.2a where 10 elements are simulated for the ULA it can be seen that for
SNR ≤ 17 dB the CS DOA method provides lower MSE values than the normal
DOA estimation for all CR values. This translates into lower variances between the
estimate and the actual DOA value using CS DOA for estimation, than the normal
DOA estimation.
Where low SNR values are considered (i.e. ≤ 5 dB) the MSE of the CS DOA
method ranges between 0.00132 : 0.0025, dependent on the CR considered the normal
DOA MSE, although not conspicuous in the graphs, varies from 0.026 to 0.0437.
The method of determining these ranges are shown in Fig. 6.3.
Taking the minimum and maximum MSE at SNR values 1.1 dB and 5 dB for both
normal and CS DOA in (a), we can describe the overall MSE improvement in terms
of a factor expressed as
MSE of the estimated DOAs ranging from [−75;75] for aFSK modulating
(a) input signal with varied SNR environment (low to medium). (b) MSE of the estimated DOAs ranging from [−75;75] for aFSK modulating
input signal with varied SNR environment (low to medium).
Using the MUSIC algorithm with conventional Using the MUSIC algorithm with conventional
−3 sampling via a ULA antennas of sampling via a ULA antennas of
x 10 10 3
CR−21 CR−21
0.35
MSE of estimated DOAs
3 CR−13 CR−13
CR−9 CR−9
CR−1 0.3 CR−3
2.5 Norm DOA Norm DOA
0.25
2
0.2
1.5
0.15
1 0.1
0.5 0.05
0 0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Fig. 6.2 Mean Squared Error (MSE) of the normal DOA method and the CS DOA method, com-
pared to the true DOA for each compression ratio denoted as a separate line. The OMP greedy
algorithm is used for CS recovery and the MUSIC algorithm for DOA estimation in both cases. The
MSE is measured for an input signal of type 2FSK modulation incident on an ULA antenna where
in a there are 10 antennas and b only 3. The DOA of the incident signal ranges from −75◦ : 75◦
6.2 Simulation 2.1—CS DOA for 2FSK Signals 97
(a)
(b)
Fig. 6.3 A simplified illustration with a showing how to determine the CS DOA MSE range for
low SNR, and b the MSE minimum and maximum factor as well as the conventional DOA MSE
range. The same convention holds for calculating similar values in other simulations in this section
MSE Norm DOA at 1.1 dB − MSE CS DOA at 1.1 dB

I Fmax = , (6.3)
MSE Norm DOA at 1.1 dB − MSE Norm DOA at 5 dB
MSE Norm DOA at 5 dB − MSE CS DOA at 5 dB
I Fmin = , (6.4)
MSE Norm DOA at 1.1 dB − MSE Norm DOA at 5 dB
where the minimum improvement factor (I Fmin ) is 14 and maximum improvement

factor (I Fmax ) is 16.2 compared with the conventional DOA scheme. The improve-
ment factors depend on the CR chosen for CS recovery, and as expected the MSE
improves when the CR is increased.
When 3 elements are simulated for the ULA, as shown in (b) of the same figure, a
different result is observed. For all SNR values the CS DOA method provides lower
MSE values than the normal DOA method. For low SNR values are considered (i.e.
≤ 5 dB) the MSE of the CS DOA method varies between 0.09 and 0.22 dependent
on the CR whereas the normal DOA MSE varies in the range of 0.294–0.47.
Again, taking the average of the minimum and maximum MSE at SNR 1.1 dB and
5 dB for both normal and CS DOA in (b), we obtain the overall MSE improvement
factors (IF) of
I Fmin = 2.45
and
I Fmax = 3.1
(as compared with the conventional DOA scheme). The improvement factor depends
on the CR chosen for the CS recovery. However, the relationship of MSE to CR is not
proportional as in (a). As CR decreases from 36 to 3 % the MSE improves, with CR
= 9 % providing the lowest MSE. The non-proportional relationship that results, as
opposed to the linear relationship in (a), can be attributed to the the effect of reduced
ULA elements.
In summary, the CS based algorithm results in improved DOA estimation accuracy
described in terms of a I Fmin and I Fmax . The improved CS DOA estimation accuracy
holds true when ULA elements are 3 or 10.
6.3 Simulation 2.2—CS DOA for 2PSK Signals
CS-DOA estimation of 2PSK signal was simulated using MUSIC estimation algo-
rithm and resultant graphs were drawn for different SNR and CR. Separate simu-
lations were done where the antenna elements in the ULA are 10 and 3. Figure 6.4
details the results obtained.
In Fig. 6.4a where the 10 elements are simulated in the ULA we observe that for
SNR ≤ 16 dB, the CS DOA method provides lower MSE values. This means when
SNR is higher than 16 dB the conventional DOA scheme provides lower MSE values,
and thus improved DOA estimates.
Where low SNR values are considered (i.e. ≤ 5 dB) the MSE of the CS DOA
method ranges between 0.00132 and 0.0026, dependent on the CR, whereas the nor-
mal DOA MSE, although not visible on the graph, varies from 0.0156 to 0.0268.
Following the convention for minimum and maximum improvement factor as detailed
in Eq. 6.3 for MSE at 1.1 dB and 5 dB for both normal and CS DOA result in improve-
ment factors of
I Fmin = 9.1
6.3 Simulation 2.2—CS DOA for 2PSK Signals 99
MSE of the estimated DOAs ranging from [−75;75] for aPSK modulating MSE of the estimated DOAs ranging from [−75;75] for aPSK modulating
(a) input signal with varied SNR environment (low to medium). (b) input signal with varied SNR environment (low to medium).
Using the MUSIC algorithm with conventional Using the MUSIC algorithm with conventional
−3 sampling via a ULA antennas of sampling via a ULA antennas of
x 10 10 3
3
CR−21 CR−21

CR−13 0.25 CR−13
2.5 CR−9 CR−9
CR−1 CR−3
Norm DOA
0.2 Norm DOA
2
0.15
1.5
1 0.1
0.5 0.05
0 0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Fig. 6.4 Mean Squared Error (MSE) of the normal DOA method and the CS-DOA method, com-
pared to the true DOA for each compression ratio. MSE is measured for an input signal of 2PSK
modulation incident on an ULA antenna where in a there are 10 antenna and in b 3. DOA of the
incident signal ranges from −75◦ : 75◦
and
I Fmax = 13
as compared with the conventional DOA scheme. The improvement factor depends
on the CR chosen for CS recovery and as expected the MSE improves when the CR
is increased.
When 3 elements are simulated for the ULA, as shown in b) of the same figure,
for all SNR values ≤ 16 dB the CS DOA method provides lower MSE values. Where
low SNR values are considered (i.e. ≤ 5 dB) the MSE of the CS-DOA method varies
between 0.121 and 0.281, whereas the normal DOA MSE varies between 0.56 and
0.937.
The minimum and maximum MSE at 1.1 dB and 5 dB for both normal and CS
DOA result in MSE improvement factors of
I Fmin = 3.1
and
I Fmax = 4.9
as compared with the conventional DOA scheme. The improvement factor depends
on the CR chosen for CS recovery. However, the relationship between MSE and CR
is not that of direct proportionality. As CR decreases from 21 to 3 %, CR = 9 %
provides the lowest MSE. The non proportional relationship that results, as opposed
to the linear relationship in a), can be attributed to the reduction of ULA elements.
In summary, using DOA estimation algorithm MUSIC for CS results in improved
DOA estimation accuracy.
6.4 Assessment of CS DOA Estimation Algorithm for Shift

Keying Modulated Signals
Findings of the previous CS DOA simulations are summarized in Table 6.1 and will
regularly be referred to for the remainder of this discussion. The summarized values
provide insight as to the optimal system parameters for CS DOA where low SNR
environments are concerned and ULA equal to 10 and 3, which we denote as ULA10
and ULA3 , respectively.
6.4.1 Performance Assessment for U L A10
Considering Simulations 2.1–2.2 where the elements of the ULA are 10, CS DOA
provides improved DOA estimates (in terms of MSE) as compared to the conventional
scheme. This is true for both 2FSK and 2PSK input signals, at a compression ratio
as low as 1 % and medium to low SNR environments (i.e. 0 dB ≤ SNR ≤ 15 dB).
In reality the improvement factors I Fmin & I Fmax , although large for some simula-
tion outcomes, do not result in large improvements on accuracy for DOA estimation.
For example, consider Simulation 2.1 where a 2FSK signal is the input and MUSIC
estimation algorithm is used for CS DOA estimation which yields the largest improve-
ment factor for all the simulations which is I Fmax = 16.2. Using I Fmax applied to
the CS DOA MSE at SNR 1.1 dB results in an MSE of 16.2 × 0.0025 = 0.0405 for
normal DOA estimation. Then the maximum deviation from the actual DOA can be
determined by calculating the statistical 95 % percentile of the deviation based on
both the MSEs. The 95 % of deviation for CS DOA estimation for Simulation 2.1
results in √
σ × 1.64 = 0.0025 × 1.64 = ±0.082◦ ,
whereas the normal DOA is

√
σ × 1.64 = 0.0405 × 1.64 = ±0.33◦ .
Therefore, although the CS DOA method provides improved DOA estimates than that
of the conventional DOA method, it only translates to a DOA accuracy improvement
of 0.33◦ − 0.082◦ = 0.248◦ .
The real value of our CS DOA approach, when large antennas are used, is the
reduction in the number of samples required. For example, consider an input signal
of length 1000 for ULA10 which requires 10 separate channels. If a digital receiver is
considered and no beamforming is done this results in 1000 × 10 = 10000 samples.
However, if CS DOA estimation method is implemented for the same scenario the
total number of samples required will reduce to 10000×1 % = 100 samples, which is
a significant reduction in memory with a small improvement for estimation accuracy
of DOAs.
Table 6.1 Summary of the findings from Simulations 2.1–2.2 and the operational range for CS-DOA for low to medium SNR, compression ratios ranging from
1 to 36 % of Nyquist sampling, and ULA elements of 10 and 3 for 2FSK and 2PSK CS recovered input signals
Modulation DOA Sim no. ULA = 10; 1 % ≤ CR ≤ 36 % ULA = 3; 3 % ≤ CR ≤ 21 %
type algorithm
MSE factor for SNR <5dB MSE range SNR range MSE factor for SNR <5dB MSE range SNR range where
CS-DOA where MSE CS-DOA MSE(CS-DOA) <
(CS-DOA) MSE (Norm-DOA)
< MSE (dB)
(Norm-
DOA)
(dB)
IFmin IFmax IFmin IFmax
2FSK MUSIC 2.1 14 16.2 0.001–0.0025 <17 2.45 3.1 0.09–0.35 <30
2PSK MUSIC 2.2 9.1 13 0.0013–0.0026 <16 3.1 4.9 0.12–0.28 <16
6.4 Assessment of CS DOA Estimation Algorithm for Shift Keying Modulated Signals
101
6.4.2 Performance Assessment for U L A3
Considering Simulations 2.1–2.2 where the elements of the ULA is 3, CS-DOA

provides improved DOA estimates (in terms of MSE) as compared to the conventional
DOA estimation scheme. The improved DOA estimates are true for both 2FSK and
2PSK input signal, at a compression ratio of as low as 3 % and for medium to low
SNR environments (i.e. 0 dB ≤ SNR ≤ 15 dB).
When ULA elements are restricted to ULA3 , MSE values increase compared to the
case where ULA10 for both CS DOA and conventional DOA estimates. Nonetheless,
the increased in MSE for ULA3 does not result in large deviations in terms of the
95 % percentile of deviation. As before, lets consider Simulation 2.1 where a 2FSK
signal is an input. MUSIC estimation algorithm is used for CS DOA estimation which
yields the largest improvement factor for all the simulations, i.e. I Fmax = 3.1 where
ULA3 (see Table 6.1). Using I Fmax applied to the CS DOA MSE at SNR 1.1 dB
results in an MSE of 3.1 × 0.35 = 1.0805 for normal DOA estimation. Thus the
95 % of deviation for CS DOA estimation is
√
σ × 1.64 = 0.35 × 1.64 = ±0.97◦ ,
whereas the normal DOA is

√
σ × 1.64 = 1.0805 × 1.64 = ±1.7◦ .
Therefore the CS DOA provides an improved DOA estimate, at a compression

ratio of 3 %, than the conventional DOA method. However, in reality the difference
in degrees for the conventional DOA and CS DOA estimate is less than 0.73◦ . Thus,
the deployment of either approaches will depend on the type of ES application and
the allowable tolerances for DOA deviation. For applications where large distances
are involved the tolerance for DOA deviation might be more crucial.
Again, the real value of the CS DOA approach is the reduction of samples required
to achieve equivalent, and in some cases, improved estimates. Consider an input
signal of length 1000 for the ULA3 which requires 3 separate channels. If a digital
receiver is considered and no beamforming is done, this implies 1000 × 3 = 3000
samples. However, if CS DOA is implemented instead, the total memory required
will reduce to 3000 × 3 % = 90 samples in total, which is a significant reduction in
memory.
6.4.3 Comparison of U L A10 and U L A3
In terms of sample reduction we can express the comparison of both the ULA cases
as per the following expression.
6.4 Assessment of CS DOA Estimation Algorithm for Shift Keying Modulated Signals 103
ULA10 = B × N × C R = 10 × N × 1 %, (6.5)
ULA3 = B × N × 3C R = 3 × N × 3(1 %), (6.6)
ULA3 = 0.9 × ULA10 , (6.7)
where B denotes the number of branches, N the number of samples.

Thus an interesting case can be made. Regardless of the number of antennas
used in the ULA for our CS DOA method, a similar reduction in memory is attained.
However, in terms of system cost, ULA3 provides an overall monetary cost reduction
due to less sampling branches.
Performance in terms of MSE and DOA accuracy is higher for the case where
ULA is greater. Therefore the deployment of either ULA10 or ULA3 has to then be
justified in terms of the performance criteria of the system. Considering ULA of 10
yields a maximum statistical 95 % percentile deviation of 0.082◦ , whereas for U L A3
the deviation is 0.97◦ .
For most DOA applications the operational performance where ULA is 3 (i.e.
ULA3 ), it is sufficient to determine accurate DOA estimation. Nevertheless, this will
depend on the performance criteria required for deployment. Moreover, it should be
noted that for ULA3 a similar DOA estimation accuracy is achieved compared to
the conventional DOA with ULA of 10, with CS DOA of ULA3 having maximum
95 % percentile deviation of ±0.97◦ and the the conventional DOA using ULA10 a
maximum with 95 % percentile deviation of ±0.33◦ .
In summary, for both the ULA cases the CS DOA estimation method, considering
2FSK and 2PSK CS recovered input signals, outperforms the conventional DOA esti-
mation method in terms of DOA accuracy. However, in some cases the improvement
on actual DOA estimation accuracy is negligible. More importantly, both the ULA
cases result in a large reduction of memory and according to Eq. 6.7 the memory
reductions for U L A10 and U L A3 are similar.
Chapter 7
CS Based Spectrum Sensing for ES
One of the major challenges of electronic defence (ED) systems is to sense a wide
spectral band in real time. In this chapter we show the use of compressive sensing
(CS) schemes to which have the potential of reducing the load on acquisition for
ED spectrum monitoring. We also propose a modified CS scheme, which we denote
as selective spectrum sensing, to further improve signal estimation for spectrum
sensing. The proposed scheme is shown to perform efficiently under severe signal
to noise ratio (SNR) conditions by leveraging a-priori knowledge of the frequency
bands of interest.
Most wideband sensing scenarios, from an electronic support (ES) perspective,
require high performance analog and digital systems to perform timely tasks such as
detection and identification accurately [3]. In a competitive technological field such
as ES, there is a continuous need to improve system management and performance
in order to reduce the risk of hardware and software bottlenecking which result in
system stagnation. Improved system performance is typically achieved by reducing
acquisition time and memory load whilst improving computational performance.
Our focus relates to the former—by reducing acquisition load using CS techniques.
Recent developments in making radio communication systems more intelligent,
flexible, and efficient for RF spectrum usage in a cognitive way, has led to the incep-
tion of the field of Cognitive Radio (CR) [93]. One of the major system functional
blocks of any CR system is the spectrum monitoring block. The functional block
of CR is a mutual framework common to ES sensing schemes. The noteworthy-
similarities are in sampling and detection requirements for a wide-band signal, which
allows a unique opportunity to exploit current advances made in the domain of CR
for use in ES spectrum detection and awareness.
Lately there have been attempts at using compressive sensing (CS) principles
[170] for more efficient spectrum sensing in CR [11]. These works can be categorised
into two types. The first type consist of attempts to reduce ADC rates of individual
spectrum sensors using CS [119], and the second type consists of the use of distributed
sensing employing multiple sensors to have an overall reduction in ADC and data

DOI 10.1007/978-3-319-46700-9_7
106 7 CS Based Spectrum Sensing for ES
rates [168]. The use of CS based algorithms in the domain of electronic defence (ED)
is sparse in the open literature.
In this chapter we address the current restrictions of using greedy algorithms in
resolving a wide band spectrum and then utilize a recent model based approach to
achieve selective spectrum sensing.Our work leverages a development made in the
work [54] which deals with frequency scenarios where the spectrum is block sparse,
and also, the model-based CoSaMP algorithm developed in [19] which allows for
signal recovery based on modelling the spectrum as segments. The main feature
of our work shows that by weighting the bands that are of interest in the wide
band spectrum, during the recovery process, these bands are selectively recovered
with better accuracy using the modified model based algorithm developed in [19].
It is an elegant solution to favour certain frequency bands for different modes of
operation and can be thought of as a varied form of discrete filtering. An added
benefit includes an increase of the spectral recovery in high SNR environments with
lower computational requirements than other CS methods.
7.1 Problem Statement
When using CS as a sensing technique to perform spectrum sensing, most perfor-

mance and optimization improvements are related to the recovery algorithms. As
has been discussed in the previous chapters iterative methods, although more accu-
rate [105], demands more computational time whereas greedy algorithms such as
OMP [172] or CoSaMP [122] tend to be preferred for faster recovery of the signal.
Subsequently these recovery algorithms require, as an input, the a-priori estimate
of the active frequencies within the spectrum. From a detection perspective, this is
not desirable but a necessity to obtain adequate recovery. This constraint causes the
spectrum estimation to degrade significantly if the estimate and true value do not
comply within a reasonable margin [172].
Fortunately when dealing with a wide frequency spectrum, such as in an ES
scenario spanning 30 GHz, we have the benefit of knowing that most communication
and radar signals operate within a specified operational bandwidth with unique signal
characteristics which are typically, if not always, known. This a-priori knowledge
of frequency band occupation provides an opportunity to model the signal offline,
and supply the recovery algorithm with a pre-weighted sensing matrix to recover
bands of interest. The CS recovery algorithm shown in [19] is used to investigate and
formulate this method.
Rather than estimating the number of frequencies present, the estimated band-
widths of the signal are given as an input during the recovery step. This approach
allows for more accurate estimation and faster computational time on signal recovery.
Signal that can be modelled as part of this recovery technique[19] are referred to as
block sparse signals. With most recovery techniques, recovering a signal within a low
SNR environment i.e. ≤5 dB, remains a high priority task [71] as recovery algorithms
are typically susceptible to high noise environments. Herein lies the inception of our
7.1 Problem Statement 107
approach to selectively recover specific bands of interest during the recovery step, as
detailed in Sect. 3.2. This is possible by, as discussed, using the a-priori knowledge
of the bandwidth of the signal of interest; and weighting the sparse basis (i.e. IDFT)
appropriately before recovery. This could potentially improve the spectral estimation
error of the signal in a high SNR environment. Details of the scheme are explained
in the following sections.
7.2 Selective Spectrum Weighted CS Approach
For a typical RF system the prioritising of bands would traditionally be achieved

by means of analog mixing, filtering, differential amplification, quantisation and
a wide band STFT-filter bank. For a detailed discussion on standard techniques
associated with wide band RF receiver systems the readers are referred to [97]. It
can be argued that our approach of selectively favouring/biasing a specific band, is
similar to bandpass filtering which could achieve the same task by adding a notch
filter in the RF receiver chain. However, this isn’t entirely similar, in that the selection
or biasing of a bandwidth of interest is done as part of the CS recovery step. This is
where the novelty in our approach stems.
Implementation of this method as part of the digital signal processing (DSP) back-
end, adds to the adept reconfigurability capability of a system. Subsequently, using
the CS-based scheme, the biased recovery of a sub-band of interest can be configured
in real-time without changes being made to either RF front end, analog or digital
filters.
We model the input signal as block sparse, similar to the model in both [54] and
[19], where the Fourier transform of the input signal can be modelled as follows.
X[k] = (X 1 , X 2 , X 3 , . . . , X N )

X [k] if k within (B1, B2, . . . , Bn )
s[k] =
0 if k is not within (B1, B2, . . . , Bn )
Where:
N
X [k] = k=0 x[n]e−2πkn/N
s[k] = approximation of the signal spectrum
x[n] = time domain signal ∼ x(t)
Bn = bandwidths of the block sparse signal
Although in practice the signal s[k] is identical to the spectrum of X [k], for the sake of
clarity, during the discussion it is appropriate to work with s[k]. Moreover, s[k] now
accurately resembles the spectrum in Fig. 7.1 which helps our cause. The proposed
Fig. 7.1 Showing a representative frequency spectrum that is block sparse, denoted by B1–B4.
Note that any frequencies outside of these bands are not of interest to the user
selective spectrum sensing scheme works in the following way. During the recovery
stage, as mentioned in Sect. 3.2, the sub-nyquist signal is represented as Y = As. The
CS based recovery algorithm needs the sensing matrix A = Φ F −1 as an input. We
show that weighting/modifying the columns of the matrix F −1 appropriately before
CS recovery, provides a viable means of favouring certain bands in the spectrum of a
signal thereby achieving selective spectrum sensing. This is illustrated by means of
a diagram in Fig. 7.2. This operation only modifies the matrix F −1 as the sampling
matrix Φ remains unchanged within our sensing scheme. The weighting of the matrix
F 1 is illustrated by matrix operations, as seen in Eq. (7.1) where diag creates a
diagonalized matrix from a vector.
F̂ = [(F −1 )T ∗ diag(δ(n))]T (7.1)
This gives a modified matrix denoted as F̂ −1 with δ(n) = (α1 , α−1 , . . . , αn ) equating
to the weighting vector that favours certain columns.
For our purposes δn can only occupy two values which are either a minimum
or maximum value, dependant on the interest in a specified bandwidth. These two
values are denoted as α1 → min and α2 → max, which are pre-set. This set of
extreme values was found to be working optimally when the elements differ by a
factor ≥100. More than two values can be assigned to δ. The result of assigning
multiple values to δ on CS recovery will be addressed as part of our future work.
Imagine we have an N x1 coefficient vector s[k] of a frequency spectrum that is
block sparse. For the sake of simplicity we take N to be 8 which in matrix operation
will lead to a F̂ −1 weighted DFT matrix subject to Eq. (7.1). In the recovery step, in
light of Eq. (3.1), this becomes a condition of
Y = Φ F̂ −1 s.
7.3 Simulation Results 109
Fig. 7.2 The system block diagram of how selective spectrum sensing is implemented. The quan-
tized signal y[k] sampled by compressive means is recovered via a biased matrix, depicted by A.
The recovery algorithm needs both the A and the expected bandwidth of the inverse Fourier matrix
F −1 and is biased by the input parameters α which take on either a high value or low value, similar
to a binary process. This combined with the bandwidth of interest as an input allows the construction
of the biased Fourier matrix
7.3 Simulation Results
In our simulation we explore two key aspects. The first one involves the recovery of
a signal spanning a wide band i.e. 0–20 GHz and the second is to show that we can
selectively sense frequency bands of interest (i.e. GSM signal uplink band) within
the wide band with improved mean square error (MSE) in a high noise environment.
In both the simulations we used four CS recovery algorithms, viz. Basis Pursuit
(BP), Orthogonal Matching Pursuit (OMP), Compressive Sampling Matching Pursuit
(CoSaMP) and Model Based CoSaMP (MB-CoSaMP). Those can be categorized as
either iterative or greedy (see Sect. 3.2). In these simulations we show that using the
MB-CoSaMP algorithm, provides improved spectral estimation capability compared
to the other algorithms.
7.3.1 Case 1: Wide Band Spectrum Recovery
An input signal with a wide band support was considered throughout the simulation,
consisting of M = 9 randomly located non-overlapping carrier frequencies f n with
bandwidths of Bn = 0 50 MHz varying with different magnitudes. The received
signal x(t) is sampled via a random sampling scheme which is sub-Nyquist. Signal
generation scheme closely follows that of [168].

M
x(t) = E n .Bn .sinc(Bn (t − Δ)). cos(2π f n (t − Δ)) + w(t)
n=1
The signal is modelled in such a way as to generate a signal in the frequency domain
that replicates the convolution of M delta functions convolved with a rectangular
function. The signal model contains a Sinc function which is defined as sinc(t) =
sin(πt)
πt
. Δ denotes time delay of the signal which also introduces phase shift and E n
includs different receive powers. The term w(t) represents additive white Gaussian
noise (AWGN) to simulate instrumentation and channel noise. The magnitude of
the received power E n remains the same throughout the sampling period of Ts .
Since the signal has a bandwidth of 20 GHz the sampling time was chosen to be
2 µs. As is the case with the conventional Nyquist sampling scheme this equates
to 2 W Ts = 80 000 number of samples (i.e. M). However, by using CS recovery
techniques i.e. MB-CoSaMP [19] the number of samples (i.e. N ) can be reduced to
well below 8000 samples with adequate probability of recovery.
Figure. 7.3 illustrates the estimation performance of the generated signal power
spectral density (PSD) as well as the CS recovered PSD, using the recovery algo-
rithm in [19]. The CS recovered signal, represented in red, shows that the spectral
information of the input signal is well recovered for almost all active sub-bands in a
high noise environment (i.e. SNR–2 dB). This claim is substantiated by the normal-
ized Mean Squared Error (nMSE) which was found to be lower than 8.9 × 10−2 .
This is within acceptable range for most electronic defence applications. However,
Fig. 7.3 Figure showing the power spectral density (PSD) of the CS recovered signal and the
original signal
Table 7.1 nMSE for different CS recovery algorithms varying the compression ratio’s for spectral
estimation of the wide band input signal
Algorithm Compression Ratio (M/N)
0.5 0.2 0.1
BP 0.0980 0.4195 0.4425
OMP 0.0949 0.1045 0.253
CoSaMP 0.0960 0.1186 0.251
MB-CoSaMP 0.0204 0.0458 0.0891
the spectral magnitudes of the recovered signal are offset by a small amount intro-
ducing error in the magnitude of the estimation, but not in the frequency estimation.
Performance degradation of the CS recovery spectral estimation, in terms of fre-
quency and magnitude, is observed with lower SNR environment i.e. ≤0 dB. The
Comparative nMSE for the spectral estimation of all four CS recovery algorithms
are shown in Table 7.1, highlighting the difference in error for different compression
ratios. The compression ratio is given as M/N with M being the number of samples
needed for CS recovery and N being to the number of samples needed as required
by the Nyquist criteria for signal acquisition. For all the compression ratios the MB-
CoSaMP recovery algorithm results in the lowest nMSE which motivates its use in
selective spectrum sensing and holds the most promise for better recovery of a wide
spectrum.
7.3.2 Case 2: Selective Spectrum Sensing
For the selective spectrum sensing case, we consider the same wide band signal, with
an added GSM 900 uplink band that is at a carrier frequency of 898.5 MHz with a
bandwidth of 25 MHz. We condition the matrix F̂ as in Eq. 7.1 according to the GSM
band of interest as well as adjusting the recovery constant for the block sparsity in
the MB-CoSaMP algorithm to the bandwidth of the signal of interest.
As mentioned in Sect. 3.2 there are numerous CS algorithms that can recover a
sparsely populated spectrum. This being the case, four algorithms were chosen on the
basis of the criteria defined in this Chapter and Sect. 3.2, namley BP, OMP, CoSaMP
and model-based(MB) CoSaMP. We have shown that out of these four the best
performing one for wide band application is the model-based CoSaMP. So we focus
on improving the error in spectral recovery using this algorithm. The other algorithms,
serve to provide comparative results and a measure for system performance statistics.
The graphs detailed below, follow the same structure, whereby a pair of graphs are
shown for either a high, medium or low SNR environment; with the first graph
illustrating the nMSE of all the recovery algorithms using normal recovery and the
second, with our approach of selective spectrum sensing applied to the recovery
process.
(a)
NON SELECTIVE: Normalized MSE vs CS compression ratio undersampling with SNR of 18.4934 dB
10 −3
BP
OMP
CoSaMP
MB−CoSaMP
10 −4
nMSE
10 −5
10 −6
0.1 0.12 0.14 0.16 0.18 0.2
Compression Ratio (M/N)
(b)
SELECTIVE: Normalized MSE vs CS compression ratio undersampling with SNR of 18.4934 dB
10 −3
BP
OMP
CoSaMP
MB−CoSaMP
10 −4
nMSE
10 −5
10 −6
0.1 0.12 0.14 0.16 0.18 0.2
Fig. 7.4 Normalized mean squared error (nMSE) for sampling compression ratio less than 0.2. For
non selective sensing a and selective sensing b as described in Sect. 7.2 for a high SNR environment,
i.e. 18 dB
In these simulations we only consider the nMSE related to the recovered signal
of the selected GSM band of interest that forms part of the wide band spectrum i.e.
0–20 GHz, such as in Sect. 7.3. In other words, any error of spectral recovery from
other frequencies do not contribute to the final nMSE values associated with the
recovered CS signals. Comparing the pairs of plots as shown in the figures above we
observe that regardless of the SNR environment, when Selective Spectrum Sensing
is applied the normalized mean squared error (nMSE), conditioned to the sampling
(a) NON SELCTIVE: Normalized MSE vs CS compression ratio undersampling with SNR of 9.2894 dB
10 −3
BP
OMP
CoSaMP
MB−CoSaMP
10 −4
nMSE
10 −5
10 −6
0.1 0.12 0.14 0.16 0.18 0.2
(b) SELECTIVENormalized MSE vs CS compression ratio undersampling with SNR of 9.2894 dB

10 −3
BP
OMP
CoSaMP
MB−CoSaMP
10 −4
nMSE
10 −5
10 −6
0.1 0.12 0.14 0.16 0.18 0.2
Fig. 7.5 Normalized mean squared error (nMSE) for sampling compression ratio less than 0.2.
For non selective sensing a and selective sensing b as described in Sect. 7.2 for a medium SNR
environment i.e. 9 dB
compression ratio improves. This results in more accurate estimation as denoted in

Sect. 7.2. Also, by applying selective spectrum sensing in the recovery stage provides
robustness in higher SNR environments (Figs. 7.4b, 7.5b and 7.6b).
(a)
−3
NON SELECTIVE: Normalized MSE vs CS compression ratio undersampling with SNR of −9.8086 dB
10
BP
OMP
CoSaMP
MB−CoSaMP
−4
10
nMSE
−5
10
−6
10
0.1 0.12 0.14 0.16 0.18 0.2
(b)
−3
SELECTIVE: Normalized MSE vs CS compression ratio undersampling with SNR of −9.8086 dB
10
BP
OMP
CoSaMP
MB−CoSaMP
−4
10
nMSE
−5
10
−6
10
0.1 0.12 0.14 0.16 0.18 0.2
Fig. 7.6 Normalized mean squared error (nMSE) for sampling compression ratio less than 0.2. For
non selective sensing a and selective sensing b as described in Sect. 7.2 for a low SNR environment
i.e. −9 dB
Part CXVI
Concluding Statements and Appendices
Chapter 8
Concluding Remarks
In this work we have covered the framework for electronic defence (ED) operations
and system requirements of current electronic support (ES) receivers, with specific
focus on how to use existing receiver types/architectures to aid in a theoretical imple-
mentation of compressive sensing (CS) techniques. More specifically, we identified
two main areas within the ES framework where CS can be implemented, namely
communication based direction of arrival (DOA) estimation and spectrum sensing,
and investigated the efficacy of both CS methods.
The efficacy of both the CS methods of implementation are summarized herein
separately, followed by a brief discussion on scalability in the ED domain.
8.1 CS Based DOA
8.1.1 Accurate Phase Recovery Achievable Using Orthogonal

Matching Pursuit (OMP)
Within the scope of greedy CS algorithms utilized for CS recovery, OMP yields an
improved MVUE phase recovery performance for both 2FSK and 2PSK input signal
for low SNR ranging 0–5 dB and CR as low as 3 %. We therefore assert that OMP is
the optimal CS greedy algorithm to use for our CS DOA estimation method.
8.1.2 Demodulation Capability via CS Recovery

for 2FSK Signals
Demodulation is possible using shift keying sensing matrix for CS recovery of 2FSK
input signals, but not 2PSK. The demodulation corresponds to the probability of

DOI 10.1007/978-3-319-46700-9_8
118 8 Concluding Remarks
detecting (PD) the correct binary index estimate (BIE), which requires a CR ≥ 15 %
and OMP for successful CS recovery estimates. Demodulation using BIE can be
performed in a low SNR environment (i.e. S N R ≥ −5 dB).
8.1.3 Equivalent Computational Performance of CS DOA

as for Conventional DOA
The use of CS-recovery on our uniform linear array (ULA) CS DOA architec-
ture yields equivalent computational performance compared to a conventional FFT
scheme for both 2FSK and 2PSK signals. The computational performance, in terms
of time complexity, only remains equivalent if the sample size of input signals remain
lower than 1024 samples, requiring CS sampling to have a CR = 1 % to the number
of samples required by Nyquist sampling.
For larger sample lengths of 2PSK and 2FSK signals than 1024 samples, the
conventional FFT based DOA estimation scheme requires less computation, resulting
in faster processing. Thus for larger sample lengths the CS DOA estimation does not
yield similar computational performance to conventional DOA estimation methods.
8.1.4 Higher Accuracy of DOA Estimates for CS DOA

Than Conventional DOA for Low SNRs
For low SNR values (i.e. SNR ≤ 5 dB) and CR ≥ 1 % the CS DOA ULA10 pro-
vides the most accurate DOA estimation over the conventional DOA method by an
improvement factor of I Fmax = 16.2 for 2FSK signals and I Fmax = 13 for 2PSK
signals. For higher SNR values the CS DOA estimation and conventional DOA esti-
mation scheme tend towards equivalent accuracy performance.
8.1.5 Reduction of Memory Required Using CS DOA

Estimation
For all the various parameters considered to determine an operational range for CS
DOA, a large memory reduction for adequate system operation can be obtained
through out. The memory reduction for the best case using CS DOA ULA10 only
requires 1 % of the total number of samples required by a conventional Nyquist
sampled DOA estimation method.
8.1 CS Based DOA 119
The large reduction of memory required by CS DOA does not result in degradation
to DOA accuracy, in fact the accuracy is equivalent to conventional DOA for high
SNR and improved for low SNR scenarios.
8.1.6 Scalability of CS DOA Estimation for Electronic

Support
Both 2PSK and 2FSK are sufficiently representative as modulation types for appli-
cation in our CS DOA method. However, direct application of 2FSK and 2PSK
modulated signals in ES systems are rare, in current communications systems. The
intent in choosing these digital modulation types were that they form the fundamen-
tal building blocks for more complex digital modulation schemes. Thus, if shown
to perform sufficiently well for stand-alone CS application, it would merit further
development for higher ordered of N -ary modulation.
The applications where 2FSK is currently used for electronic communication
in ES comprise GSM, Bluetooth 1 and FMCW (Frequency Modulated Continuous
Wave Radar), whereas for 2PSK the applications comprise wireless LAN standard
(IEEE 802.11b.1999 basic rate) and Bluetooth 2.
For practical RF application of our CS DOA method for ES tasks can be achieved
for GSM application as discussed in Sect. 6.4. Moreover, GSM uses GMSK for
modulation which can be described as a spectrally efficient and coherent form of
2FSK which would require minimal development for CS DOA. However, for practical
applications synchronization, multiple access and other signal characteristics will
have to be considered for full operational deployment. If the TDMA bursts can be
synchronized across the RF carriers for the entire GSM operational bandwidth of
25 MHz, it would allow our CS DOA estimation method to determine the direction
of 992 channels spanning a DOA of 180◦ using a ULA.
8.2 CS Based Spectrum Sensing
CS selective spectrum sensing as a wide spectrum sensing method can improve

spectral error recovery in severe signal to noise ratio (SNR) conditions as well as
improve computational costs. It is important to highlight that this can only be achieved
based on the a priori knowledge of the spectrum, which in most scenarios for ED
systems are available. And as motivated earlier, the model based CoSaMP recovery
algorithm, for our implementation, is well suited for spectrum sensing applications
and result in an improved recovery of the wide band spectrum as compared to other
CS algorithms.
120 8 Concluding Remarks
8.3 Final Remarks
Both implementations of CS for ES tasks have yielded practical and system perfor-
mance benefits for ES receiver systems which only realize for specific cases and are
subject to special conditions. However, our modest addition to the literature, by suc-
cessfully using CS methods for ED tasks shows the scalability of current CS theory,
and we are confident that further investigation of CS, as a new signal processing
method, can aid hardware and software performance for ED systems.
Chapter 9
Appendix: Some Useful Theoretical
Background
9.1 Electromagnetic Waves
The operation of using electromagnetic radiating fields to transmit encoded energy

(i.e. data) over a distance, for our purpose, can be expressed1 in the following way,
consider a source EM radiating element (i.e. antenna) located at an origin point in
space. A signal generated from such a point using an isotropic omni directional
antenna, radiates into free space as an ever-expanding sphere [39] which then is able
to interact with another resonating elements (i.e. receiver antenna). This expansion
of the field is called EM propagation and for illustration purposes detailed in Fig. 9.1.
At a significant distance from the transmit antenna the spherical wave-front starts
to approximate a planar wave which is known as the far-field distance and calculated
2D 2
in meters by R f f = m, where D is the antenna aperture and λ is the wavelength.
λ
From this far-field distance, the radiated electrical field of an arbitrary field can
be expressed in the following way
e− jk0 r
Ē(r, θ, φ) = [θ̂ Fθ (θ, φ) + φ̂ Fφ (θ, φ)] V/m. (9.1)
r
Ē → The electrical field

θ̂ , φ̂ → Are unit vectors according to the spherical coordinate system
r → The radial distance from the origin
k0 = 2π/λ → The free space propagation constant
λ = c/ f = 3 × 108 / f → The wavelength, dependant on the frequency (Hz)
Fθ (θ, φ), Fφ (θ, φ) → Are the pattern functions
1 Theoretical work was modified and sourced largely from the following sources, [3, 143, 181]. For
further detail, we refer the reader to the literature listed.

DOI 10.1007/978-3-319-46700-9_9
122 9 Appendix: Some Useful Theoretical Background
Fig. 9.1 Far field distance and the propagation dynamics needed from both, transmit and receive
systems. (Modified by the authors from [143])
In other words, considering the transverse electromagnetic wave equation in (9.1),

the electrical field propagates in the radial direction with a phase function of e− jk0 r ,
an amplitude of 1/r , and polarized in either φ or θ directions [143]. Also, according to
Maxwell’s equations, we know that any propagating electrical field has an associated
magnetic field that can be formalized using (9.1) where H and E denote the respective
magnetic and electrical fields.
Eθ
Hφ = , (9.2)
η0
Eφ
Hθ = , (9.3)
η0
where η0 = 377 and is also referred to as the wave impedance of free space. Further-
more, the poynting vector [80] which stipulates the directivity of the electromagnetic
field is given by the cross product of the electric and magnetic field vectors, as shown
below.
S̄ = Ē × H̄ W/m2 (9.4)
9.2 Receiver Components: Background
9.2.1 Antennas
In all ES receiver systems the purpose, namely the multiple tasks that a designed
receiver system needs to execute, determines which antenna will be used [1] as part
of the RF front end system. This desired propagation intent, in part, is the reason for a
single or multiple antenna implementation of various ES systems. The requirements
stipulating the type of antenna to be used are determined by numerous variables
9.2 Receiver Components: Background 123
Table 9.1 Typically used antenna performance parameters. Taken from [1]
Term Description
Gain The increase in signal strength (commonly
stated in dB) as the signal is converted by the
antenna from EM radiation to a voltage signal
4π Ae f f
(G = )
λ2
Frequency The coverage or range of frequency over which
the antenna can receive or transmit signals,
whilst providing the required parametric
performance
Bandwidth The frequency range of the antenna in units of
frequency. Often stated in terms of percentage
bandwidth [100 % × (maximum frequency –
minimum frequency)/average frequency]
Polarization and the orientation of the E and H waves
transmitted/received. Mainly vertical,
horizontal, or right- or left-hand circular
Beamwidth The angular coverage of the antenna, usually in
degrees, depicted by spatial radiation pattern
plots in terms of degree (deg) and decibels (dB)
related to azimuth and elevation
Efficiency The percentage of signal power transmitted or
received compared to the theoretical power
from the proportion of a sphere covered by the
antenna’s beam
known as antenna performance parameters, detailed in Table 9.1, which are consulted
when selecting an antenna. Performance requirements of antennas are extensive and
continually developing in the antenna design literature [14, 95, 181].
Most antennas, regardless of the application, can be categorized according to the
directivity of receiving or transmitting signal [37], therefore an antenna is either
defined as omnidirectional or directional. Omnidirectional antenna have equal gain
in a spherical/donut radiation pattern allowing for equal receiving and/or transmit-
ting signal-strength from all angles [2]. Similarly directional antenna, as the name
suggests, directs EM propagation in a specific direction for a required bearing, based
on the design of the antenna. This directionality allows for higher gain along the
bore-sight direction2 (See Fig. 9.2.).
Varying types of antennas have been developed as part of the proliferation of
application and technology. Figure 9.3 details typical antenna used for ED and ES
systems. Although numerous types of antenna exist, as suggested earlier, the applica-
tion dictates the choice of the antenna. From an application perspective with respect
to ES tasks such as direction finding (DF) and interception, RF receivers mostly
make use of omnidirectional antenna with high gains.
2 Bore-sight describes the direction at which the antenna is pointing.

Fig. 9.2 Common radiation pattern, in azimuth (Horizontal) and elevation (Vertical) planes.
Sourced from [24]
When given a specific operating frequency range for a ES system, namely the
VHF-UHF bands for our application, the choice of antenna becomes simple. The
only antenna that meets this requirement are dipole, whip, loop, biconical or swastika
antennas [1, 156] with the dipole, whip, and loop having narrow bandwidth coverage
and the biconical and Swastika having large bandwidths (refer Fig. 9.3). The former
antennas are more suited to direction finding tasks of narrow band signals and the
latter more suited to wideband spectrum sensing for ES application.
Diople antennas are preferred for DF-ES systems [60] due to their compact size,
omnidirectional nature, narrow bandwidth and relatively high gain (refer Fig. 9.4).
Hybrid omnidirectional wide bandwidth antennas (like biconical) are used as wide-
band antennas for tasks such as spectrum sensing/monitoring. However, when a sys-
tem requires more gain or directivity, a range of log periodic dipole arrays (LPDAs)
are sometimes added [142].
9.2.2 RF Front-End Systems
In radio frequency circuitry the term ascribed to the analog components between
the antenna and digital baseband systems (intermediate frequency - IF) are known
as the RF front end [39]. The RF front end is standard with the first stage of most
RF receiver systems, with exception to direct sampling systems that do not down-
convert the signal [5] to IF. When the receiver chain of an ES RF front end system is
designed, the typical system blocks used in the process can be accurately described
by components common to the super-heterodyne architecture (SHA) (as shown in
Fig. 9.5).
Numerous advances have been made in the domain of RF design and processing
in the recent years, for example, the advent of software defined radio (SDR) and
Fig. 9.3 Different types of antenna and their respective characteristics in EW applications. (Taken
from [1].)
Fig. 9.4 A typical tactical

direction finding antenna that
spans a frequency range of
20 MHz–3.6 GHz. Taken
from [142]
software radio (SWR) [177]. These systems still depend on a modified version of a
super-heterodyne architecture to acquire/quantize RF signals. In addition most SWR
applications are still not realizable due to limitations on ADCs. This limitation shows
how crucial RF front end systems are and suggests that RF front end systems are
inseparable for most current technologies (Table 9.2).
The super-heterodyne receiver (SHR) architecture, as seen in Fig. 9.5, is a common
model of the front end system. It describes all the individual system blocks that form
part of virtually all RF receiver front end systems. Different hybrid forms of RF
front end systems, stemming from SHR architecture, exist due to the proliferation of
microwave and circuit advances [157]. Such hybrid systems implement multi-stage
mixing, filter, amplification stages [80], channelizers, and filter banks [58, 97, 125,
178]. Most of these modifications are done in an attempt to improve RF reception
and processing of signal and nowadays manufactured as standard on-chip packages
[157].
9.2.2.1 Filter
Filter stages applicable for ES in RF front end systems perform the task of rejection
of unwanted frequencies bands, attenuating undesired mixing frequency artefacts
(see Sect. 9.2.2.3) and setting the IF bandwidth of the receiver [157]. Filters can be
described as two-port networks used to maintain and control the frequency response
in RF systems by only allowing transmission of frequencies within the passband
[143]. As Fig. 9.6 shows the characteristics that define the filter frequency response
can be categorized into three figures of merit, namely the passband, transition-band,
and stop-band.
BPF - Band pass filter IF - Intermediate Frequency

LNA - Low Noise Amplifier VGA - Variable Gain Amplifier
Mixer - Using a tunable local oscillator (LO) ADC - Analog-to-Digital Converter
Fig. 9.5 Illustrates the typical RF front end sub-systems shown in block diagram form which
constitute a theoretical RF receiver
Table 9.2 Describing the most important system blocks that comprise the RF front end
System block Description
Antenna Form the crucial step of converting EM energy
into electrical voltage
Amplifiers Apply a gain (dB) to low strength received
signal artefacts to the required power levels to
be processed for system tasks (i.e.
identifications, transformation, detection)
Filters Perform the necessary filtering by biasing
certain bandwidths of interest
Mixers Due to the high frequency that RF signals
propagate at, it is necessary to convert down
convert the signals to manageable frequencies
to digitize for digital processing. This is done
by mixing stages
Fig. 9.6 A typical frequency response for a bandpass filter, shown as part of the design process,
indicating the respective bands. The ripple effects in (a) are denoted by δ p & δs . Taken from [74]
Fig. 9.7 Shows the process system blocks used for filter design by the insertion loss method. (Taken
from [143].)
Filters comprise of discrete resistive, inductive and capacitive networks that oper-
ate to form the desired frequency response needed by the filter specifications. The
use of new material, novel filter design, and manufacturing processes are vast and
detailed within the open literature. However, our concern involve operational para-
meters of filters that need to comply with the requirements of RF receiver in ES
systems. These parameters involve the attenuation, ripple strength, phase character-
istics and transition band roll off [157]. Furthermore the design process commonly
used in designing such filters are known as the insertion loss method detailed in [143,
151] and shown by means of systems blocks in Fig. 9.7.
9.2.2.2 Amplifiers
Amplifiers in an RF front end, performs the pivotal task of applying gain to the signal
of interest (SOI) before it reaches the digital domain for further signal processing.
Both the design and implementation advances of these systems have come a long
way from the initial beginnings in the mid 1900s [92].
Currently, most amplifiers use three-terminal solid-state devices which include;
silicon or silicon-germanium bibolar junction transistors (BJT), field effect transis-
tors (FET), complimentary metal oxide semiconductors (CMOS) and high electron
mobility transistors (HEMTs) [66, 80, 181]. These devices have resulted in improved
gains, dynamic range, and bandwidth performance with amplifiers operating at fre-
quencies up to 100 GHz [157].
Both the RF and IF stage amplifiers sometimes comprise of multiple amplification
stages [143] which work in unison to improve gain linearity across the frequency
band of interest [181]. The RF amplifiers (i.e. low noise amplifiers- LNAs) increase
the power of weak received signal after the filter stage, as mentioned earlier in
Sect. 9.2.3, in order to increase dynamic range of the transmitted signal relative to
noise which compensates for loss due to the signal propagation. As is the case for
most interception tasks, the signal to noise ratio (SNR) is relatively low (e.g. <–5
dB) [156] due to noise levels overpowering the signal content. Different amplifier
designs are used to improve the SNR to result in a signal level that is sufficient for
operations like detection, identification and classification to be performed with high
confidence levels.
In the design stage of an amplifier there are certain operational parameters that
need to be consulted to ensure the required performance of an amplifier refer to
Table 9.3.
Table 9.3 Describing the important system blocks that comprise the RF front end
Term Description
Noise figure The measure of degradation due to noise
effects by RF components on the actual signal
noise. Mostly caused by device thermal noise
and measured in dB using the input and output
SNR ratio
Gain The measure in dB of the amplification added
to the input signal. Typically measured by the
input and output power figures
Bandwidth The frequency band at which the amplifier is
operational, centered at the operational
frequency. Typically, amplifiers have a limited
bandwidth wherein they are able to provide the
designed gain. This bandwidth is determined
by the stability of the system, which is
calculated using stability circles and smith
charts, as in [143]
Typical the amplifier stage3 in an ES receiver need to have a gain ≤60 dB and a
bandwidth >60 MHz [156] to deal with sensitivity and range requirements, which
are detailed in Sect. 9.4.1.
9.2.2.3 Mixing
Mixers are used in receiver system as frequency converters, modelled as a 3-port-

device and intentionally taking advantage of non-linear characteristics of diodes or
transistors to enable frequency conversion [181]. Even though it might not seem
obvious initially, frequency conversion is necessary due to the sampling rate limi-
tation of ADC technology not matching the RF frequencies. In fact, to sample RF
signal directly without the use of mixing would require double the sampling rate to
match the highest frequency present in the received signal [128]. However, recent
advances in techniques such as direct sampling have been shown to be effective for
frequencies up to 2–5 GHz [167], removing the need for mixers.
Traditionally mixer technologies depended on diodes to implement signal mix-
ing, however nowadays mixers are predominantly dependant on more effective and
reliable solid state devices such as field effect transistors (FETs) [66].
A mixer, as shown in terms of the symbol and functional blocks in Fig. 9.8, receives
the RF signal ( f R F ) and convolves a local oscillating frequency ( f L O ) signal which
3 Most amplifiers are manufactured as on-chip integrated circuits (IC’s) [77]. This places the empha-
sis on the task of choice of components rather than design. The design process, serves to arrive at
the required system specifications.
Fig. 9.8 The final outputs in terms of frequency plots of up conversion and down conversion
implemented by a mixer. Sourced from [143]
produces an output signal ( f I F ) usually known as the intermediate frequency (IF)

signal which comprise the sum and difference of the input frequencies.
Mixers can be utilized for up conversion and down conversion of a signal depen-
dant on the desired tasks. Following the notation as specified before, the determi-
nation from first principles follows. Consider the case when the received signal is
down-converted. The input RF signal is
x R F (t) = A cos(2π f R F t), (9.5)
and it is modulated with the local oscillator at a specific frequency using the mixer
x L O (t) = B cos(2π f L O t). (9.6)
The output signal produced is the intermediate frequency output in terms of the sum
and the difference of the respective input signals (in frequency domain). Conversely,
the same holds true for frequency up conversion, however as it is predominantly used
for transmitting purposes it is not used in the receiver chain.
Time Domain The time domain mathematical representation of down conversion

mixing shown below.
x I F (t) = C cos(2π f R F t) × cos(2π f L O t) (9.7)

C
= [cos(2π( f R F + f L O )t) + cos(2π( f R F − f L O )t)] (9.8)
2
Frequency Domain The frequency domain mathematical representation of down

conversion mixing shown below.
X I F (ω) = K π δ[(ω − ω R F ) + (ω + ω R F )] ∗ [π δ[(ω − ω I F ) + (ω + ω I F )]] (9.9)
The effect of down conversion in the frequency domain is shown in Fig. 9.8 and
detailed mathematically in Eq. (9.9). This frequency domain mixing can be rep-
resented, in terms of frequency, as the sum and difference of the inputs, f I F =
f R F ± f L O . Here the sum is better known as the upper side band (USB) and the
difference, the lower side band (LSB). When down conversion is applied it is impor-
tant to note that the spectrum of the LSB and USB are conserved whereas in up
conversion the LSB is inverted [157].
For consideration, in an ES receiver, the desired IF output would be determined
by the difference f I F = f R F − f L O which can be extracted using an appropriate low
pass filter (LPF).
The derivation denoted above holds true for an ideal case. In real systems mixers
will generate more artefacts due to non-linearity associated with individual com-
ponents (e.g. voltage controlled oscillators VCOs), inducing unwanted harmonics
and their effects [143]. Some of the effects on signal degradation involve image
frequency, conversion loss, increase in noise figure, intermodulation distortion, and
isolation.
All these factors account for system loss, frequency drift and spectral anomalies
that can adversely affect the effectiveness of a receiver if not accounted for properly.
Hence, mitigation of mixer effects form a crucial system consideration with respect
to later signal processing stages.
9.2.3 Radio Frequency Propagation Operations
RF communication, for the most part, used for ES activities are restricted to fre-
quency bands lower than 300 GHz, Table 9.4 indicates the current designation of
bands as defined for ED purposes. It is important to note that most tactical commu-
nication operations primarily take place in the HF, VHF, and UHF and nowadays
as communication bandwidths increase, in the SHF bands as well. Figure 9.9 details
the typical communication link, tactical communications bands, and propagations
modes used for ES communications.
It is well known that for different frequency bands, used for communication, differ-
ent propagation properties apply [143]. The higher frequency bands (i.e. >100 MHz)
rely on clear line of sight between the transmitter and receiver, whereas lower com-
munication bands (i.e. HF) can leverage on propagation phenomena such as sur-
face waves, reflected waves, and ducting [139] which do not rely on line of site
(but do introduce reception complexities). Although higher frequency bands have a
restriction of line-of-site for communication, they allow for higher bandwidth based
communications and hence higher data transfer rates.
Table 9.4 RF band designation as defined in the ED domain [156]. These bands are defined
differently depending on the domain
Frequency bands Wavelength Name Designation
3–30 KHz 100–10 km Very low frequency VLF
30–300 KHz 10–1 km Low frequency LF
0.3-3 MHz 1 km–100 m Medium frequency MF
3–30 MHz 100–10 m High frequency HF
30–300 MHz 10–1 m Very high frequency VHF
0.3–3 GHz 1–0.1 m Ultra high frequency UHF
3–30 GHz 0.1–0.01 m Super high frequency SHF
30–300 GHz 0.01–0.001 m Extra high frequency EHF
Fig. 9.9 Illustrating the

range of ED tactical
communication conducted in
the HF, VHF and UHF
bands. Sourced from [2]
VHF and UHF bands have the advantage of being more predictable, and hence can
be described more accurately by analytical expressions [2] which can be modelled
to analyse the effects of propagation. Such propagation effects, for communication
purposes above 100 MHz, include:
• Reflection (Due to ground and/or large objects) —– Two-way propagation model,
• Diffraction (Due to edges and corners of EM conductive environment) —– Knife-
Edge propagation model,
• Scattering (Due to foliage or small objects) —– Free Space propagation model,
and
• Attenuation (Due to atmospheric events, i.e. different forms of precipitation) —–
Experimental Environment model.
Our focus, concerns communication that takes place in the higher frequency bands,
namely VHF/UHF communications. Understanding the propagation phenomenon
(i.e. power requirements, losses measured, and modes of propagation) in these bands
are utmost. The review of lower communication bands (i.e. HF) propagation dynam-
ics are excluded in this work due to the majority of communication signals of interest,
for ES, operate at higher communication bands.
9.2.3.1 VHF/UHF Modes of Propagation
Propagation theory involves the means of modelling the environment through which
a communication link is established. A communication link is set up between a
transmitter (XMTR) and a receiver (RCVR), assuming a line-of-sight (LOS) link
under good weather condition, as shown in Fig. 9.10. The signal strength, shown
in dBm (typical notation for ED applications), leaves the transmitter at a specific
dBm level which gets either amplified by the antenna or simply propagates from the
antenna at unity gain (0 dBm), known as the emitted radiated power (ERP).
The EM signal propagates through the channel where it attenuates due to spreading
losses and atmospheric losses associated with the link. Although this is a simple
evaluation of the link losses, it adequately explains the general losses associated
with propagation, other more complex forms of attenuation are considered later.
Once the signal is received at the receiver antenna it is once again amplified and then
processed at a dBm level proportional to the distance from the transmitter.
Free Space Propagation
Is the case where spreading loss is the only propagation loss considered in the model
and reflection paths are minimal [2]. This loss usually applies to high altitude com-
munication, high frequencies, and narrow beam-width antennas [80]. The typical
equation associated with free space propagation determining the link losses, also
known as the link equation is shown below in both normal and log form as well.
L = (4π )2 d 2 /λ2 , (9.10)
L = 32.44 + 20 log( f ) + 20 log(d). (9.11)
Fig. 9.10 Illustrating the

propagation losses involved
over a communication
link....include db figures
Where
L is the loss to the system in dB

d is the distance between the transmitter and receiver
λ is the wavelength, calculated from the frequency λ = c/ f
32.44 in Eq. 9.11 equates to 20 log(4π ) which in realistic terms equates to a loss proportional
to 1/R 2 when considering free space propagation losses.
Two-Way Propagation
Is the case where significant reflective objects are in the vicinity of the communication
link, usually ground reflections accounting for most of these losses [1]. When this is
the case the two-way propagation model is commonly used to model the losses/gains
correctly by using Eq. 9.12. The losses associated with this model occur when both the
transmitter and the receiver are closer to earth’s surface, which exclude most cases of
air-to-air and air-to-ground links [156]. Moreover, the generic term describing most
of these reflections, in the communication domain, are known as multipath [143].
The losses are detailed below in both forms, namely normal (9.12) and logarithimic
(9.13). Overall propagation losses in approximate terms equate to a loss proportional
to 1/R 4 when dealing with ground reflection losses (i.e. multipath) [143].
L = (d)4 / h 2t h r2 , (9.12)
L = 120 + 40 log(d) − 20 log(h t ) − 20 log(h r ), (9.13)
where,
L is the loss to the system in dB

d is the distance between the transmitter and receiver
λ is the wavelength, calculated from the frequency λ = c/ f
ht is the height of the transmitter
hr is the height of the receiver
120 equates to a common dB amount associated to the transmit power (Pt ), transmit
antenna gain (G t ) and the receive antenna gain (G r ) - log(Pt G t G r ).
In a typical ED scenario it is preferable to determine which propagation model

must be used for the reception/interception of signals, which is commonly done by
Fig. 9.11 Illustrates the propagation losses associated with the free space model, note that the
height of the transmitter and receiver must be a significant height from the earth’s surface for this
model to be used [65]
calculating the Fresnel zone.4 If the communication link is within this zone, free space
propagation losses are considered. If it is outside this boundary, two-way propagation
losses are used [1].
ES propagation systems are typically designed for dynamic scenarios involving
a vast number of established and potential interception links. Thus, the propagation
models used are either free space, two-way, or knife edge propagation with the latter
model involving a special case of defraction not pertinent to our focus for commu-
nication link propagation, and consequently only mentioned herein (Figs. 9.11 and
9.12).
9.2.3.2 Attenuation/Propagation Loss
Attenuation, as defined by [181] in terms of RF design principles, is the propagation

losses experienced by a signal in the propagation path which cause signal power
to decrease. Causes that contribute to such losses are attributed to atmospheric phe-
nomenon, precipitation and man made structures [2, 143] with the former, commonly
playing a major role in the design for ED systems utilizing higher frequencies for
example, air-to-ground systems (satellite communication) and radar warning systems
(RWR) [178].
Shown in Fig. 9.13 is the typical atmospheric attenuation to propagation per km for
different frequencies, where operational frequencies below 10 GHz the propagation
losses are negligible.
Due to our focus on VHF-UHF communication bands, the effect of man made
structures add more of a significant propagation loss than atmospheric attenuation,
which is known as fading.
4 The Fresnel zone provides a means of calculating when reflected EM waves will arrive at the
receiver, either in phase or out of phase, which consequently affects the loss or gain of the signal.
The first Fresnel zone is where path-length phase shifts by 0–180◦ and in the second Fresnel zone
◦ nλd1 d2
path-length phase shifts 180–360 which can be calculated using the equation Fn = .
d1 + d2
Fig. 9.12 Illustrating the propagation losses involved for the two-way model. Where
h2 h2 e− jk0 Rd
PR = |v|2 /z 0 = Pt G t G r t 4r and v 2ck0 h t h r . Typically Ht and Hr are multiples
Rd Rd2
of the propagation wavelength
Fig. 9.13 Shows the propagation losses per km in dB for different frequency bands (horizontal
polarization), sourced from [151]
9.2.3.3 Fading
Most mobile and land communication scenario aggregate around populated areas
involving manmade structures (i.e. buildings, houses, cars etc.). This proximity to
man made structures allow for multiple scattering, reflection, and diffraction to take
place between the communication transmitter and receiver, causing fading [67].
Fading, is defined [36] as the phenomenon of small-scale variation to the mag-
nitude and phase of the transmitted signal due to a line-of-sight (LOS) not being
established. The result of fading causes the propagation of EM waves to rely solely
on reflection and defraction to arrive at the intended receiver point for communica-
tion.
The most widely used and accurate model in describing the statistical basis for
radio signal propagation with no LOS, is the Rayleigh fading model [67]. This model
allows for a possible ES system to compensate for the losses where mobile radio links,
mobile phones, and tactical VHF bands exist. The Rayleigh fading model is used for
ES systems for the following situations (sourced largely from [67]).
• Urban defence communication monitoring and direction finding as well as RCIED

(Radio controlled improvised explosive devices)
• Tactical air to ground links i.e. where a airborne platform (helicopter) is at a low
altitude.
• The receiver antenna is embedded in radio clutter,5 caused by close proximity to
the ground.
The statistical model given in Eqs. (9.14) and (9.15) describe the fading distribu-
tion involved where, of the two, the CDF is more important as it relates the likelihood
of a given value to be exceeded (Fig. 9.14).
The probability density function (PDF)
x x2
P(x) = e(− ) (9.14)
σ2 2σ 2
The cumulative density function (CDF)
x2
F(x) = 1 − e(− ) (9.15)
2σ 2
As mentioned, the variation of a propagating signal (see Fig. 9.15) is quantified
in terms of its standard deviation σ . In other words, using the Rayleigh model we
can relate the probability of exceeding a needed signal strength value to the standard
deviation of that signal strength in terms of dB. For our purposes—in ES—the typical
value is 10 dB.
For a given scenario where a minimum of –80 dBm signal strength is required by
a receiver (sensitivity level—see Sect. 9.4.1.2) and the propagation model indicates
a short sector of –70 dBm, creating a 10 dBm margin and an availability of 0.9
probability at the reception point.
Considering the effects of attenuation and fading within the context of an ES
receiver it should be noted, even though attenuation affects propagation losses, fading
5 Radio clutter is a term attributed to structures that influence radio propagation creating spurious
scattering signal.
Fig. 9.14 The Rayleigh probability density function f (x) and cumulative density function F(x)
with σ = 1. (Sourced and adapted from [65].)
Fig. 9.15 Received power

of a typical receiver versus
distance from the transmitter,
showing the small-scale
fading effects and the large
scale variations, sourced
from [143]
is more prevalent for interception techniques in ES due to its application in the

communication bands which operate at less than 2 GHz, where there is less than
0.02 dB/km propagation losses. Multipath and Rayleigh fading cause variations in
phase and magnitude that have the means of deteriorating tasks for ES receivers such
as detection and direction finding.
Typical fading effects on ES receiver tasks can be summarized as follows:
• The effects of fading causes an increase in bit error rate for modulation schemes,
• The phase ambiguities caused by multipath, scattering, reflection, and defraction
due to no line-of-sight reduce the ability of an ES system to determine the direction
of a RF transmission source [149].
9.3 Typical ED System Configuration
A typical ED system does not exist as a stand alone system which can be standardized
across all implementations of ED system application. Instead, an ED system’s objec-
tive determines what is necessary and what will be deemed as redundant. In other
9.3 Typical ED System Configuration 139
Fig. 9.16 Illustrates the differentiation, in terms of blocks, of operational functions within the
implementation of a typical ED system. The blue block indicates the receiver components, while
the red indicates the transmitter components. (Sourced from [3] and modified by the authors.)
words, if the only objective is to intercept a communication transmission, the receiver

block of the ED system will be included and a transmitter block removed, due to its
redundancy. However, from an overview perspective, the fundamental system blocks
are illustrated here to describe the entire ED system chain not for a specific use only.
An ED system chain is illustrated in Fig. 9.16 and will be used throughout this dis-
cussion to draw attention to specific components and their purpose in ED, especially
ES intercept systems for communication signal purposes. The blue dotted block
indicates the systems pertaining to this body of work. A discussion of the individual
system blocks are detailed below.
1. System Control: Deployment of such systems need a central hub that ensures
that all sub-systems are coherent and synchronized to perform operational tasks
[157]. Moreover, operations of control systems are typically performed using one
and/or several computers. These computers are either stand alone or distributed
which communicate via a network.
2. Antenna: From the theory of Electromagnetism6 we empirically know that an
antenna is an electrically conductive resonant material that extracts and facilitates
propagation of electromagnetic energy through an unbounded medium [81] by
converting EM energy to electrical signals that can be processed and interpreted.
For the purposes of ED antennas enable transmission of signal via a propagation
medium (i.e. air, free space), which enables tasks such as interception, direction
finding and jamming (using high gain antennas) to be achieved. See Sect. 9.2.1
for further detail.
3. Signal Distribution: Regardless of the size of an ED system, it is imperative to
have a splitting element to allow signal distribution to several receivers. Signal
splitters are placed between the antenna and receiver system, with a typical
impedance of 50 , requiring closely calibrated impedance matching [181].
This matching is done to reduce distortions between split signal channels and
ensure different receivers receive the same signal with respect to gain, magnitude,
and phase at the respective receiver systems.
4. Search Receiver: This system block, although generic and somewhat common,
is crucial to spectral intelligence gathering in the ED receiver chain, specifically
used to search the RF spectrum and characterize and classify sources of EM-RF
energy [48]. Systems that form a crucial part of this block and add to the reception
of the signal include LNAs, Analog Filters, and Mixers. Further discussion of
these receivers and subsystems are detailed in Sect. 9.4.
5. Set-On Receiver: These systems are used, sometimes in conjunction with search
receiver output data, for long-term analysis of the signal which includes measur-
ing parameters of signal for analytical use to the operator. In fact, these receivers
often comprise a channelized filter bank, using the search receiver RF front end,
in order to pre-select frequency bands of interest that will be passed on to the
operator.
6. Signal Processing: Realistically and within a modern context, this is where
most of the computationally intensive tasks take place. As covered in detail in
Sect. 2.2.3, the first task for ED use is to extract usable information of frequency
and bandwidth, energy of the signal, modulation type, and the baud rate of digital
communication signal. Then secondly, the information is used to detect, iden-
tify and classify signal accordingly. Techniques used within this block includes
high speed analog-to-digital (ADC) conversion, digital-filtering, DSP-blocks,
and signal transformation (i.e. Fast Fourier Transform—FFT, Walsh Hadamard
Transform—WHT etc.).
7. Direction Finding (DF) Signal Processing: Direction finding systems operate
on the principle that every electromagnetic wave propagates from a radiating
source in a specific direction through a medium which can be received at another
point [60]. Using multiple antennas, with the correct orientation, a DF receiver
6 Foundational work was done by J.C Maxwell [111]. The field equations relating to this work can
be found in the appendix.
9.3 Typical ED System Configuration 141
system could, by means of estimation algorithms [176], estimate the angle-of-

arrival of that signal.
8. Excitepr: An excitepr is simply an RF signal generator that has the capability of
modulating a generated signal. These systems are generally used for purposes in
EA, as detailed earlier, where objective of the RF system is to jam communication
systems. Jamming is typically achieved by generating and emitting random noise
over an RF channel—effective against most communication signal—or tone
jamming which is effective against modulation types such as frequency shift
keying (FSK) [3].
9. Power Amplifier: Although amplifiers are used as part of the receiver system for
ES purpose (the blue block in Fig. 9.16) when it comes to EA tasks these power
amplifiers convert the relatively weak signals (typically 0 dBm) into signal with
greater energy for transmission. For jamming EA tasks the signal converted up
and transmitted via the antenna can reach an order of several kilowatts (typi-
cally 1 kW). Issues with this system include, conversion efficiencies, frequency
coverage, spectral purity, and leakage [3].
10. Filters: Filters are used in most tasks relating to ED. For ES systems filters
operate on the input signal to define the bandwidth of interest or used to scan
a large bandwidth with a bandpass filter for interception tasks. For EA tasks
filters are placed before the antenna as a precautionary step in order to reduce
the likelihood of RF damage to friendly receiver systems due to the spurious
nature of power amplifiers [1].
11. Communications: This subsystem can be disseminated into two main operative
tasks, which are to send external data to the receiver and to deliver a report of
the data received to a operator [3]. Typically these tasks do not impose on the
performance of the overall system, however it does play an important role in
directing (i.e. command and control) the objectives of an ED system from a
remote location.
9.4 Electronic Support Receiver Systems
Reception and processing of communication signals in context of ED tasks are pre-

dominantly handled by means of an ES receiver [178]. However, these receivers are
able to support other types of signal reception and processing besides communication
signals.
The current demand from ES receiver systems to be effective for intercepting com-
munications, requires a level of sophistication beyond a single conventional system
[42]. Modern day ES receivers need to incorporate many closely integrated sub-
systems which perform different tasks and have high data rate throughput [178]. A
typical ES receiver architecture—shown in Fig. 9.17—highlights the number of sub-
systems (i.e. DF tasks, spectrum monitoring, frequency measurement) that comprise
an ES receiver.
Fig. 9.17 Illustrating a typical ES receiver architecture. Taken from [178]
The performance of such a receiver is an important consideration when developing

techniques to be implemented using ES receiver architecture. Although the current
literature on ES receiver component/system design, implementation, and use are
comprehensive,7 for our purpose a review of the operating parameters will be the
focus.
9.4.1 ES Receiver Characteristics
When characterising a receiver to operate within an environment, the dynamic range,

sensitivity levels and signal-to-noise ratio (SNR) values are key parameters to be
considered when designing for signal processing technique to be implemented for
detection capability of desired signals [178]. These parameters provide insight into
operational capabilities and provide the means to design/select receivers to inter-
cept/receive weak signals with improved probability of detection (Fig. 9.18).
9.4.1.1 Dynamic Range
In simple terms, dynamic range refers to the received input signal range, also
explained as the difference between the strongest and the weakest amplitude of a
signal that can be processed in real time (instantaneously) [107]. Yet, in practice, the
dynamic range remains a parameter that needs to be measured rather than estimated
based on theoretical calculations.
7 The literature that encompass ES receiver design and technical specification, are detailed in much
greater depth in the following works [39, 48, 178, 181].
9.4 Electronic Support Receiver Systems 143
Fig. 9.18 In a the dynamic range with regards to amplitude in frequency domain is depicted,
whereas b denotes the 1 dB compression point. (Taken from [174].)
The dynamic range—in an ES receiver—is predominantly determined by the oper-

ation of the amplifiers in the system, with exception to some components like ADC
dynamic range. However, taking note of this dependence the dynamic range is com-
monly defined with respect to the output power levels. The lower limit of the dynamic
range is specified by the sensitivity levels (the smallest that can be processed) and
the upper limit by the 1—dB compression point [178] (the point where the ampli-
fiers start following non-linear characteristics). Typical values of ES receivers range
between 50–90 dB depending on the type of receiver [37].8
9.4.1.2 Sensitivity Levels
Receiver sensitivity is defined as the minimum signal power that is required at the
receiver input to detect and/or process a desired signal [178]. This requirement pro-
vides insight into a receiver’s ability to distinguish a signal of interest from accom-
panying noise under weak signal conditions. However, the required sensitivity is a
loose term that does not apply to generic RF scenarios, for example when dealing
with modulated signals larger SNR values are required with regard to the carrier
frequency.
In the domain of ED it is seen as good practice to define where in the system chain
the receiver sensitivity is defined. These sensitivity levels are illustrated by Fig. 9.19
with two different definitions of sensitivities, with the correct value ascribed by the
sensitivity at the input of the receiver after the losses associated with cables, coupling,
and amplifiers have been included. However, there does exist special instances when
sensitivity can be defined differently, in terms of electric intensity (µV /m) instead
of log (dB) due to a complex relationship between the antenna and receiver [139].
Direction finding tasks are one of these exceptions which typically requires a conver-
sion from µV /m to dB, expressed in terms of P = signal strength (dB); E = electric
intensity (µV /m); F = frequency (M H z) as below.
8 Refer Sect. 9.4.3 for further detail on receiver types.

P = −77 + 20 log(E) − 20 log(F) (9.16)
The components that determine receiver sensitivity, which has been eluded to by
means of system loss, are attributed to thermal noise, noise figure, and signal to noise
ratio and defined as follows:
1. kTB—is defined as the thermal noise power level of an ideal receiver [156]
which is typically specified at 290 ◦ K with a constant receiver bandwidth
(i.e. M H z) nominally denoted in dBm/M H z, where a common value kTB =
−114 dBm/MHz and the following values hold:
• k—Boltzmann’s constant (1.38 × 10−23 J/◦ K)
• T —operating temperature, in ◦ K
• B—the effective bandwidth of the receiver.
Consider a receiver with an effective bandwidth of 10 MHz, the correct thermal
noise would correspond to kTB = –114 dBm/MHz + 10 dB = –104 dBm
2. Noise figure (NF)—defined by [1] as the ratio of the noise per kTB of actual
noise that would have to be added to an ideal, noiseless receiver in order to pro-
duce the actual noise that is present at the output. In other words, it is the thermal
noise that the receiver adds to the received signal with regards to the receiver input.
Each component in the receiver system comes with its own specified noise figure
as determined by the manufacturer. However, when determining a receiver system
parameter specifications, it is imperative to include and model the NF of the entire
system, since the sensitivity levels rely on the whole system (i.e. antenna, lossy
cable, amplifiers and distribution network), not simply the receiver sensitivity
level. When determining the entire NF of the system the value calculated is can
be expressed accordingly,
N F = L 1 + N p + D. (9.17)
Fig. 9.19 a Showing the

different definitions
associated with sensitivity
and b the losses before the
sensitivity of the ideal
receiver is considered. Taken
from [1]
Where D is determined from the graph in Fig. 9.20 and L 1 encompass all the
pre-amplifier losses and N p is the pre-amplifier noise figure. In non dB form the
noise figure can also be expressed as

Sin /Nin
F0 = . (9.18)
Sout /Nout
Typical values for receiver system noise figures are between 8 and 10 dB [2].
3. Signal to noise ratio (SNR)—SNR is the most important value to consider, more
so than kTB and/or NF, when determining and defining the sensitivity. In ED
sensitivity signal processing scenario the SNR is defined as

Sout put
S N R = 10 log , (9.19)
Nout put min
which is indicative of the minimum SNR that a receiver can still operate while
performing detection with high probability. Subsequently, in such a case, the sen-
sitivity (minimum detectable signal) can be calculated, in standard form according
to,
So
Smin = Sensitivit y = kT B F0 . (9.20)
No min
Take for example a scenario with the given parameter for calculating sensitivity,
then the system sensitivity can be expressed in dB form in the following way,
Fig. 9.20 Chart used to

determine the system
degradation, labelled as D in
Eq. 9.17. The intersection of
where the calculated value of
the sum of the preamp NF
(N p ), the preamp Gain (G p )
and the loss before the
receiver (l2 ), with the
receiver NF (N R ) determines
the degradation value and
enables system noise figure
to be calculated. Taken from
[1]
Fig. 9.21 Showing the spectrum of SNR values that are considered for ED scenario’s and typical
values associated with categorizing signals [151]
kT B + N F + r equir ed S N R (9.21)
= (−114 d Bm + 10 dB) + 10 dB + 20 dB = −74 dBm. (9.22)
It is important to note that the minimum SNR required by a receiver is highly

dependant on the task that needs to be performed. In the case where information
is modulated on to the received signal higher SNR values are required. This form
of SNR is known as the pre-detection SNR, called the RF SNR or the carrier-to-
noise ratio (CNR) [107]. In most cases the SNR of the signal is relatively higher
than the RF SNR.
Where frequency modulation (FM) is considered the receiver sensitivity is deter-

mined by the received power levels as well as the modulation characteristics.
Therefore, when determining the SNR of frequency-modulated signal, there
are two things to consider. First, the RF SNR which tends to be approxi-
mately 12 dB [1] in most ED systems and second, if the RF SNR is above this
threshold a FM improvement factor holds. This improvement factor equates to
I FF M (in dB) = 5 + 20 log10 β where β is the modulation index. Hence the SNR
for FM signals can be expressed as,
S N R = R F S N R + I FF M . (9.23)
With the above mentioned sensitivity parameters, it is prudent to know the com-
mon operating sensitivities that are associated with most ES receivers. These sensi-
tivity values are typically dependant on the type of receiver used (see Sect. 9.4.3),
which vary from –50 to –90 dBm with the ideal sensitivity preferably higher than
70 dB for tasks such as direction finding [37] (Fig. 9.21).
9.4.2 Sensitivity Conversion
The conversion from dBm to µV /m can be expressed as (taken from [1]):
E = 10(P+77+20 log[F])/20 (9.24)

and conversely as:
P = −77 + 20 log(E) − 20 log(F) (9.25)
These equations are based on power transmitted and antenna gain at the point of
reception, which are:
E2 A
P= (9.26)
Z0
Gc2
A= (9.27)
4π F 2
Where:
P = signal strength (in W) E = Electric intensity (in V/m)

A = Antenna area (in m 2 ) Z 0 = impedance of free space (120π ())
G = Antenna gain (= 1 for isotropic antenna) c = speed of light (3 × 108 )
F = Frequency (in Hz)
9.4.2.1 Effective Range
Considering the definition of sensitivity, it is imperative to determine the effective

range of a receiver in order to gauge at what physical distances a receiver is able to
operate for interception tasks
Various parameters are considered in specifying ES receivers to calculate the effec-
tive range within a certain margin, which include propagation loss (L s ), sensitivity
(equivalent to power received PR , but referred to as S R ), transmit power (PT ), trans-
mission gain (G T ), receive gain (G R ), and spreading loss (L s ).9 In short the effective
range can be determined by solving the following expression (written here in dB
form),
S R = PT + G T − L s + G R , (9.28)
where : L s = −32.4 − 20 log(F) − 20 log(d). (9.29)
As a means to describe the effect receiver parameters in real world operations,

consider two scenarios that are pertinent to this body of work. Given that an ES
receiver has two different sensitivity levels, namely –75 and –80 dBm.
9 Spreading loss can be determined from the nomograph in Fig. 9.22 or calculated using Eq. 9.11
equating to L s above.
Fig. 9.22 Illustrates the use

of a nomograph to determine
the propagation loss for a
frequency of 1000 MHz for a
given distance. Courtesy of
[2]
Scenario 1: Determine the effective range of intercepting a GSM signal from a

downlink tower transmission at 1800 MHz, and
Scenario 2: Determine the effective range of intercepting a GSM signal from the
uplink transmission of a handset 900 MHz.
Listed below are the effective ranges according to the above expression in Eq. 9.28.
Calculation steps scenario 1:

Description Symbol Value Reference
Power transmitted PT 15 W [154]
Gain transmitted PT 10 dB [37]
Frequency of transmission F 1800 MHz [143]
Gain received (omnidirectional antenna) GR 0 dBi [142]
Sensitivity LS −75; −80 dBm [37]
Effective range (distance) d in km
Sensitivity = −75 dBm:
20 log d = PT + G T − 32.4 − 20 log(F) + G R − L S (9.30)

20 log d = 10 log(15000) + 10 − 32.4 − 20 log(1800) + 0 − (−75) (9.31)
⎛ ⎞
29.25
⎝ ⎠
d = 10 20 = 29.02 km (9.32)

20 log d = 10 log(15000) + 10 − 32.4 − 20 log(1800) + 0 − (−80) (9.34)
⎛ ⎞
34.25
⎝ ⎠
d = 10 20 = 51.61 km (9.35)
Calculation steps scenario 2:

Description Symbol Value Reference
Power transmitted PT 2W [154]
Gain transmitted GT 0 dB [37]
Frequency of transmission F 900 MHz [143]
Gain received (omnidirectional antenna) GR 0 dBi [142]
Sensitivity LS −75; –80 dBm [37]
Effective range (distance) d in km

20 log d = 10 log(2000) + 10 − 32.4 − 20 log(900) + 0 − (−75) (9.37)
⎛ ⎞
29.25
⎝ ⎠
d = 10 20 = 6.7 km (9.38)

20 log d = 10 log(2000) + 0 − 32.4 − 20 log(900) + 0 − (−80) (9.40)
⎛ ⎞
34.25
⎝ ⎠
d = 10 20 = 12 km (9.41)
Table 9.5 A summary of the typical range of values associated with the important system character-
istics of ES receivers. This is valuable when considering signal processing tasks to be implemented
on such architectures. (Sourced from [116] and modified by the authors.)
Characteristic Typical range
Maximum instantaneous analysis bandwidth 0.05–2
(GHz)
RF range (GHz) 0.01–60
Dynamic range (dB) 40–90
Sensitivity (dBm) –70 – –90
Frequency resolution (MHz) 0.5–500
Minimum power (W) 60–200
These values provide an approximate range wherein the ES receiver has the poten-
tial to intercept such signals, as well as insights into the relationship that sensitivity
parameters hold on the capability of interception.
9.4.3 Types of ES Receivers
By way of introduction, with exception to receiver characteristics, the capability of

any receiver is highly dependant on the type of receiver, especially in the case of ES
receivers for electronic defence purposes [37]. ES receivers, by design, must make
system trade-offs due to several conflicting requirements as there are too many sce-
narios to execute all ES tasks simultaneously with optimal performance when using
a single receiver architecture. Hence, the need for specialized equipment, technolo-
gies, and techniques which result in the existence of a wide variety of ES receivers
today (Table 9.5).
The review of different ES receiver types are of paramount importance. Each of
the receiver’s capabilities, provide an insight, as to the approach of implementing
new/novel DSP techniques (i.e. Compressive Sensing etc.) on different receiver plat-
forms. The implementation parameters are summarized as follows:
• Receiver flexibility (Reconfigurability),

• Computational performance,
• Frequency resolution,
• Data rate,
• Ease of integration,
• Cost involved with each receiver.
All these parameters are considered in an attempt to determine the most appropri-
ate receiver type to utilize and/or consider for the purposes of this body of work. As
Table 9.6 Typical characteristics associated with different types of ES receivers. Taken from [1]
Receiver type General characteristics
Wideband crystal video Wideband instantaneous coverage; low
sensitivity and no selectivity; mainly for pulsed
signals
Tuned RF crystal video Similar to crystal video, however. Provides
frequency isolation and better sensitivity
IFM Coverage, sensitivity, and selectivity likened to
crystal video; measures frequency of received
signals
Superheterodyne wideband Most common type of receiver; good
selectivity and sensitivity
Superheterodyne narrowband Good selectivity and sensitivity; dedicated to
one signal
Channelized Combines selectivity and sensitivity with
wideband coverage
Microscan/Compressive Provides frequency isolation; measures
frequency; does not demodulate
Digital High flexibility; can deal with signals with
unknown parameters
a consequence such determination is leveraged by the characterization of the various

receivers in Tables 9.6 and 9.7.
In this section, a brief comparison of the various receiver types—that are typically
used for ES purposes—are reviewed. Below is a summarized table of the general
characteristics of some common ES receivers and Table 9.6 provides a summary of
the qualitative capabilities of each receiver type. Moreover, for the purpose of the
reader, a comprehensive table of the qualitative capabilities are detailed at the end of
this section in Table 9.7.
9.4.3.1 Crystal Video Receiver
This receiver type provides a simple yet cost effective means of doing instantaneous
detection whilst using inexpensive techniques. The receiver consists of a bandpass
filter and pre-amplifier circuit, then a crystal (diode) detector followed by a video
amplifier that has, as an output, the video band signal.
The diode detector circuit operates at low enough power, which is in the square
law10 region [2]. Benefits associated with this receiver are the simplicity in the tech-
nology which incurs less expense, instantaneous detection, and high probability of
interception (POI) in wide frequency range [178]. Whereas the drawbacks include,
no frequency resolution, poor sensitivity, and poor simultaneous signal performance
10 The output is a function is dependant on the input power rather than the signal voltage.
152
Table 9.7 Showing the summarized details of the qualitative capabilities of the various receivers
Receiver type Receiver qualitative capabilities
Measures Selectivity Sensitivity Dynamic Multiple Frequency Demodulation POI Cost
frequency range signals coverage
Crystal video N P P G N G G Y L
TRF Y M P G Y G P Y L/M
IFM Y P P M N G G N M
Superheterodyne Y G G G Y G P Y M/H
wideband
Superheterodyne Y G G G Y P P Y M/H
narrowband
Channelized Y G G G Y G G Y H
Microscan Y G G G Y G G N M/H
Digital Y G G G Y G M Y H
KEY G = Good M = Moderate P = Poor H = High L = Low Y = Yes N = No
9 Appendix: Some Useful Theoretical Background
Fig. 9.23 Showing the typical topology of a crystal video receiver
[37]. These receiver are typically used in radar warning receivers (RWRs), conse-
quently, ruling such a receiver type inadequate for our purposes. See Fig. 9.23.
9.4.3.2 Tuned RF Receiver
Tuned RF receiver type share a similar architecture as the wideband crustal video
receiver, mentioned previously, however in the earlier days of RF this receiver utilized
a YIG filter and oscillator to isolate the signal of interest at a specific frequency. Thus
increasing sensitivity but still suffering from slow response time and poor POI [178].
Due to advances in receiver technology this type of receiver has largely been replaced
by Superheterodyne receivers [1]. Tuned RF receivers nowadays are optionally used
for RWR and frequency measurements in hybrid scenarios [37].
9.4.3.3 Instantaneous Frequency Measurement (IFM) Receiver
As the name suggests, an IFM receiver measures the frequency of a received signal.
The received signal is split into two signal paths by means of a delay line, see
Fig. 9.24. One of the signal paths have a constant delay time τ which produces a
frequency dependant phase difference θ [178] whereas the other signal path remains
unchanged.
An IFM receiver takes advantage of this relationship, by measuring the phase differ-
ence between the two signals, whereby the frequency can then be inferred by using
the expression θ = 2π f 0 τ . Lastly, this frequency inference is digitized and passed
on to produce a direct digital frequency reading.
The preamplified IFM receiver exhibits the same sensitivity as the crystal video
receiver but less dynamic range. Benefits of an IFM receiver are that they are relatively
simple and compact with improved frequency resolution and high instantaneous POI.
Associative disadvantages include insufficient sensitivity for some ED scenarios and
therefore cannot be used in dense signal environments [37, 139, 178]. However some
Fig. 9.24 A diagram of a common IFM receiver. Sourced from [37]
Fig. 9.25 A diagram of a narrow band superheterodyne receiver layout. Sourced from [116]
scenarios/environment exist where IFM receiver are used in shipboard ES, Jammer
power management, and SIGINT equipment [37].
9.4.3.4 Superheterodyne: Wideband and Narrowband
Superheterodyne receivers are one of the most versatile receivers in use today [143].
The name super (i.e. higher) and heterodyne (i.e. linear shifts) join to describe the
operation of this receiver. It linearly shifts the received signal to an intermediate fre-
quency (IF) by means of a fixed and/or tuneable oscillator, this techniques is known
as mixing, see Sect. 9.2.2.3. Furthermore the isolation of other frequency aliasing is
done by means of bandpass filters, see Fig. 9.6, where the block diagram illustrates
the construction of such a receiver.
These receivers are utilize either a wideband bandpass filter to enable surveillance
and search tasks of a wide bandwidth of the RF spectrum, or a narrowband band-
pass filter that only isolates a small portion of the spectrum to be analysed with
higher frequency resolution and sensitivity. The usefulness of this approach, stems
from enabling the flexibility and control of the local oscillator which is varied in a
sawtooth-like fashion [178] to provide scanning ability of the frequency spectrum at
a cost to response time (Figs. 9.24, 9.25 and 9.26).
Fig. 9.26 A diagram of a wide band superheterodyne receiver layout. Sourced from [116]
Wideband superheterodyne
Advantages Improved response time to threats and probability of interception
Disadvantages Higher probability of spurious signal generation and less
sensitivity than its narrowband counterpart
Narrowband superheterodyne
Advantages High sensitivity, improved frequency resolution and no
interference of simultaneous signals
Disadvantages Slow response time, inadequate POI and suffers with signals that
are frequency agile
Both types of superheterodyne receivers are used in SIGINT equipment, Air (i.e.
Tactical air warning), shipboard ES, and the analysis system of a hybrid receiver
systems.
9.4.3.5 Channelized Receiver
Channelized receivers are widely considered as one of the ideal receiver types in
use for ES tasks. The technique behind its wide range adoption is the large number
of contiguous bandpass filters [178] for each channel, and as the name suggests,
channelizes (i.e. divides) the RF bandwidth by means of a power divider/multiplexer
into respective subbands whereby the signal from each channel is amplified, filtered,
and digitized further by means of a fixed tuned receiver (FTR), see Fig. 9.27. These
FTRs were classically comprised of surface acoustic wave (SAW) devices [107],
however in recent years rather use narrowband superheterodyne receivers and other
miniaturization technologies [1]. Typical implementation of a Channelized receiver
utilizes 10 to 20 TRF channels to cover 10 to 20 % of the RF range, coupled with
computer controlled switchable frequency translators allowing 100 % coverage of
the entire needed EW RF spectrum [2].
Fig. 9.27 A diagram of a channelized receiver layout
Advantages Wider RF bandwidth coverage, nearly instantaneous frequency read-

ings as well as moderate frequency resolution due to the limitation
on the number of realistic channels, improved sensitivity, and higher
selectivity [37]
Disadvantages With higher performance system, comes at a cost and channelized
receivers are no different. These receivers require higher complexity,
cost, and provide lower reliability with limited sensitivity
Usage/application SIGINT equipment and jammer power management
9.4.3.6 Microscan/Compressive Receiver
A Microscan receiver or Compressive receiver for ED tasks are essentially the same
receiver, with emphasis on describing the basic operation of the receiver in different
ways. Microscan receiver refers to the receivers ability to fast-sweep its local oscilla-
tor, subsequently mixing (see Sect. 9.2.2.3) the RF input signal to produce a chirped
frequency modulated (FM) signal [178]. Whereas Compressive receiver refers to the
compression of the output FM signal by the dispersive delay line (DDL) implemented
via SAW filters which are an integral component of this receiver architecture11 [178].
The concept of the compressive receiver was seminal in White’s work [186]
which allowed for a wide-band receiver to achieve fine-frequency resolution with
an output of narrow pulses that held a linear relationship between their position in
time and frequency of the RF input signal. The receiver operates by taking an input
RF signal—with frequency f 0 —which is mixed with a linear sweeping LO signal,
changing at a rate inversely proportional to the frequency-time gradient of the DDL
and then passed through a video detector, see Fig. 9.28 for illustration of the process.
The DDL must match the bandwidth of the sweeping LO which is defined by the
11 Adetailed discussion of the mathematical work needed to design such a receiver is beyond the
scope of this work. We refer the reader to the work done in [178, 186] for further detail on this topic.
Fig. 9.28 A diagram of a microscan/compressive receiver. Courtesy of [49]
signals that are within the spectrum that are of interest.
Advantages Near instantaneous processing of frequency, good resolution,

dynamic range, and improved simultaneous signal capability
Disadvantages Implementation of such a system is more complex. It is limited
by bandwidth with no pulse modulation information. Moreover, the
alignment of the saw filters are critical, if not manufactured correctly
can make a system ineffective
Usage/Application Typically used for SIGINT scenarios due to the large RF spectrum
needed to be surveyed. Sometimes used for applications needing
fine frequency analysis over a wide frequency range [37]
9.4.3.7 Digital Receivers
A digital receiver in some respects approximate an ideal receiver, where a received

signal is directly sampled at twice its highest frequency (i.e. Nyquist criteria) and the
digitized signal is made available to a dedicated DSP and/or processor unit where
digital filtering, demodulation, and other techniques are applied by means of com-
putation.
This ideal receiver, unfortunately, is not yet fully realizable for the entire RF spec-
trum that EW occupies, using current technology. The success of such a system is
dependant on the acquisition rates of current ADC’s (i.e. fastest commercially avail-
able ADC is 2.7 − 3.6 G S P S [76]), data throughput, processing power, and memory
requirements.
In current digital receivers this barrier of digitizing signals is circumvented by
applying a technique known as bandpass sampling [179]. We refer the reader to
Sect. 9.6.3.1 for a detailed discussion on the topic.
Functionally, as the name suggests, bandpass sampling samples at a lower rate (which
is attainable with current technologies) but within a specific bandwidth which allows
higher frequencies to alias or down convert to a zero IF where it is then digitized.
The performance and accuracy of quantizing the RF signal and removing spurious
signals are aided by filters and novel signal techniques in order to cover a RF range
of up to 20 GHz. An example of such a system can be found in [167].
Advantages Good simultaneous signal capability, good POI, improved selectivity

and dynamic range
Disadvantages A system such as this is highly complex and very expensive. It has
lower resolution than conventional receivers
Usage/application Land, sea, air ES, and SIGINT scenarios
9.5 Compressive Sensing Mathematical Fundamentals
Herein we discuss the fundamental building blocks—relating to 1-D time and fre-
quency domain signals - necessary for implementing CS theory. The notation fol-
lowed throughout this section is derived from work12 in [44, 53, 150]. Moreover,
in this section we will focus on the mathematical concepts that are associated with
CS theory and used for explanatory purposes rather than a literature comparison of
CS techniques. The review of implementing CS on digital systems are detailed in
Sect. 3.4.
9.5.1 Vector Space
As a prerequisite for CS theory, a signal must be sparse in order to be exactly recovered

[50], thus to recover such signals an understanding of vector spaces, bases, and frames
are required as a fundamental building block for recovery.
12 We advise the reader to the aforementioned literature for further comprehensive study of CS
and to provide insight relating CS to other signal processing applications, particularly in image
processing.
9.5 Compressive Sensing Mathematical Fundamentals 159
For signal processing purposes we model input signals as vectors (i.e. of length
N ) that exist in a discrete, finite domain known as a vector space, a N-dimensional
Euclidean space denoted by R N . A function frequently used for vector spaces in CS
is the p norm defined for p ∈ [1, ∞) and given as:
N 1/ p

||x|| p = | xi | p
, p ∈ [1, ∞). (9.42)
i=1
For application to signal approximation, a common technique is the norm as it mea-

sures signal strength and/or signal error which is minimized to obtain the best estimate
of the true signal in a particular vector space (Fig. 9.29).
As an example, consider a vector x ∈ R 2 that exists in an affine space that needs to
be approximated by a point x̂, with x̂ ∈ A where A is an arbitrary subspace; then
the following definition holds, by taking the approximation error using an p norm,
Fig. 9.29 A diagram of a typical digital receiver, however, not representing the ideal receiver but
a current realization of given todays technology limits. Courtesy of [176]
Fig. 9.30 Illustrates the different approximations of p norms, p = 1, 2, ∞ and the quasi-norm
1
with p = , approximating a point in R2 by means x̂, for the p norm onto A. Where A is a denotes
2
a low dimension subspace that is approximated by expanding the norm conditioned to x. Sourced
from [53]
with the objective to minimize ||x − x̂|| p subject to the vector space A for different
norms, the approximation of x̂ ∈ A can diverge to x with a measurable error based
on the p norm chosen.
In Fig. 9.30 this approximation is shown for different p norms. An observation
of the figure suggests that the approximations for different norms give rise to varied
approximation errors, with p > 1 error more evenly distributed or spread, whereas
p = 1 and the quasi-norm case error tends to be unevenly distributed or sparse [18].
The approximation error is the fundamental basis for sparse signals and their use for
CS recovery as it applies to higher dimensional signals and affine vector spaces.
9.5.2 Sets, Bases and Frames
Based on the fundamentals from linear algebra [112] sampled signals can represented
as a discrete vector in a finite-dimensional vector space V , where V ∈ R N and
K = {1, 2, . . . , N }, which is the vector space that comprise bases that span V .
Take for example a set Ψ = {ψ1 , ψ2 , . . . , ψ K } that is linearly independent and spans
V . Then Ψ can be defined as a basis for the vector space V .
Bases are vital in describing signals of similar origin or application which can
be represented by a linear combination of the vectors of the same basis, with varied
coefficients for each signal [53]. In some cases a basis is referred to as a dictionary
in CS literature.
In discrete mathematical terms, a signal can be decomposed into a linear set of
coefficients (ai ) and a basis (ψi ), such that the discrete signal x ∈ R N can be
expressed as:

x[n] = ai ψ̃i , (9.43)
i∈K
where ψ̃ is the dual basis or in matrix terms the inverse used to construct the original
signal. Sometimes it is useful, especially for signal reconstruction, to use a dual basis
in generating an orthonormal basis where a set of linearly independent vectors =
{µ1 , . . . , µ N , } that span V would constitute an orthonormal basis, if the following
condition holds:
0 if i = j
µi · µ j = (9.44)
1 if i = j.
Given a basis that is not orthonormal, using the set of vectors that comprise the
basis, an orthonormal basis can be generated following the Gram-Schmidt method
[112]. A crucial motivation for using orthonormal bases relates to the properties
associated with it, that is, its dual is equal to the hermitian adjoint (i.e. transpose)
such that Ψ = Ψ̃ T . Furthermore, it is useful to define the frame of a basis as it
sometimes provides a more developed representation of a signal due to the inherent
redundancies.
A frame is defined as a set of vectors (Ψi )i=1
n
in Rd where d < n, can be represented
as a matrix Ψ ∈ R , such that for all the vectors x ∈ Rd
d×n
A||x||22 ≤ ||Ψ T x||22 ≤ B||x||22 , (9.45)
where 0 < A ≤ B < ∞.13 In particular, a frame extends the definition of a basis
to include sets that are possibly linearly dependent, giving rise to infinitely many
coefficients α for an input signal x and frame Ψ such that x = αΨ . In the case where
Ψ is a d × N matrix, the values of A and B correspond to the eigenvalues of Ψ Ψ T
[150].
The infinitely many coefficients, attributed to the inclusive linear dependency of
the frame, provides a choice for coefficient vector when the dual frame is considered,
as the frame operated on by the signal is responsible in determining the coefficient
vector. Importantly, any dual frame Ψ̃ that satisfies
Ψ Ψ̃ T = Ψ̃ Ψ = I, (9.46)
is considered an alternative dual frame, with Ψ̃ = (Ψ Ψ̃ T )−1 Ψ enabling the dual

frame to be well-defined due to Ψ Ψ T being invertible based on the inequality that
A < 0 requires Ψ to have linear independent rows. Therefore, a means to obtain a
feasible coefficient vector β of a signal x, is by applying the following operation
β = Ψ T (Ψ Ψ̃ T )−1 x. (9.47)
13 The condition imposed on A > 0 implies that the rows of Ψ must be linearly independent. When
A is chosen to be the largest and B the least valued of the possible inequalities, it is known as the
optimal frame bound. When A = B the frame is known as A-tight. Finally, if A = B = 1 then Ψ is a
Parseval frame.
This coefficient vector in many CS recovery steps forms the first step of the algorithm,
see Sect. 3.4.
9.5.3 Matrix Construction for CS Sensing
A comprehensive study of constructing CS basis compliant matrices exists in the

open literature [25, 43]. Nonetheless, we provide an overview of the criteria and
the associated proofs required for compliance of possible matrices that match, and
subsequently can be used as a sensing matrix for CS. A large majority of the material
for this section is sourced from [105, 150] and we refer the reader to these for further
insight into CS.
First, let us assume a sensing matrix , with size M × N , operates on a K-sparse
input vector x, with size N × 1 (see Fig. 3.5), which produces the output vector y
with size M × 1 as
y = x. (9.48)
If we desire to recover all sparse signals from x using the measurements y via the
matrix , as seen above, it can quickly be deduced that any pair of different vectors
x, x̂ ∈ K = {x : ||x||0 ≤ K } must result in x = x̂. Otherwise, it is impossible
to distinguish between x and x̂ based on the measurement vector y as there will be
infinitely many solutions [18].
represents all x ∈ K only if the Null space of Ψ contains no
More formally,
vectors in 2K . The Null space of is defined as N () = {z : z = 0}. One
of the typical ways to characterise this property, and serves as the guarantee for
unambiguous recovery, is by means of the spark [46].
Spark
Definition 9.1 The spark of a given matrix is the smallest number of columns of
that are linearly dependent [105].
This definition yields the following guarantee based on Corollary 1 of [46].

Theorem 9.1 For any vector y ∈ R M , there exists at most one signal x ∈ K such
that y = x if and only if spark () > 2K [150].
Thus, the guarantee holds, based on the spark for the recovery of exactly sparse
signals [18], but does not extend to approximately sparse signals, which is dealt with
by the null space property (NSP).
9.5.4 Null Space Property
The null space property (NSP) can be considered as a condition that places an even
higher restriction on the null space of —denoted as N ()—to distinguish between
approximately sparse signals [46]. In other words, the N () must be kept free of
any vectors that are too compressible as well as from those that are sparse [150].
Subsequently, this enables the NSP to express, empirically, that the null space of
should be spread, and not be congruent on a small subset of indices [166].
The result of NSP provides a guarantee that a matrix of order 2K is adequate in
establishing exact recovery, subject to the condition in Eq. 9.49. In order to define
the operation of the NSP conditional check, we adopt the following notation based
on [150]. Let Γ be a subset of indices (Γ ⊂ {1, 2, . . . , N }) and its correspondent
Γ c = {1, 2, . . . , N }/Γ . When referring to a vector xΓ it designates the length N
vector by setting the values of x indexed by Γ c to zero. Applying the same logic, for
a matrix c with size M × N , results in a matrix with columns of indexed by Γ c
to zero.
Definition 9.2 (Definition 3.2 of [150]) A matrix satisfies the NSP of order K if
there exists a constant C > 0 such that
||h Γ c ||1
||h Γ ||2 ≤ C √ (9.49)
K
holds for all h ∈ N () and for all Γ such that |Γ | ≤ K .
Theorem 9.2 [35] Let : R N → R M denote a sensing matrix and Δ : R N → R M

denote an arbitrary recovery algorithm. If the pair (, Δ) satisfies 9.50 then
satisfies the NSP of order 2K.
Therefore, if a given matrix meets the criteria of NSP, then the only K-sparse vector
in its null space is h = 0. Although the recovery of the CS measurement vectors will
be detailed later, let us consider the example where we let the recovery algorithm be
denoted by Δ, then using the NSP inequality of 9.49, a guarantee can be established
such that,
σk (x)1
||Δ(x) − x||2 ≤ C √ (9.50)
K
for all x, where x is given as

σk (x) p = min||x − x̂|| p where x ∈ . (9.51)
K
Although the means of recovery will be covered in Sect. 3.4, the above equation
provides a guarantee for exact recovery for all K-sparse vectors, and suggests a high
likelihood for non-sparse signal recovery by some other K-sparse vector is possible
[105].
9.5.5 Restricted Isometry Property
The restricted isometery property (RIP) is a vital development that extends beyond
NSP conditions, allowing a guarantee for recovery where measurements are cor-
rupted by some form of error or noise. The guarantee, formalized by [29], places a
more strict condition on matrix , namely the isometry condition, which is ubiqui-
tous with compressive sensing.
Although the proof of the RIP is somewhat involved (see [17] for proof), in
simple terms if the RIP condition of order 2K holds for a matrix, say , then based
on Eq. 9.52 φ preserves the Euclidean distance between any pair of K-sparse vectors.
Definition 9.3 (Definition 3.3 of [150]) A matrix satisfies the restricted isometry
property of order K if there exists a δk ∈ (0, 1) such that
α||x||22 ≤ ||x||22 ≤ β||x||22 ) with α = (1 − δk ) and β = (1 + δk ) (9.52)

holds for all x ∈ K = {x : ||x||0 ≤ K },
√
where 0 < α ≤ β < ∞ and δ = (β−α)/(β+α) with multiplied with 2/(β + α)
provides a matrix that still satisfies the RIP, however, with a scaled matrix version
with symmetry about one in accordance with Eq. 9.52.
Stability is of critical concern for recovery of noisy measured sparse signals when
dealing with the RIP condition, especially what lower bound must be chosen. In other
words, if stable recovery of a signal is needed the RIP of the matrix must meet the
lower bound of Eq. 9.53. This lower bound criteria is denoted below, as the condition
for a matrix known as C-stable [105].
Definition 9.4 (Definition 3.4 of [150]) Let : R N → R M denote a sensing matrix
and Δ : R N → R M denote an arbitrary
recovery algorithm. We say that the pair
(Δ, ) is C-stable if for any x ∈ K and any e ∈ R M we have that
||Δ(x + e)||2 ≤ C||e||. (9.53)
The above definition dictates that if a small amount of noise is added to the measure-
ment of x, the result on the recovery of the signal is not to be unpredictable [150].
Moreover, Theorem 3.3 of [150] shows that by letting C → 1 forces to satisfy
the RIP (9.52) lower bound with δ K = 1 − 1/C 2 → 0. Hence, if we wish to reduce
the influence of noise on the signal recovery, must be adjusted to satisfy the RIP
lower bound with a smaller constant.
The measurement bounds of a potential sensing matrix (i.e. ), based on [17]
by the Theorem 3.4 in [150], has to result in a lower bound for the number of
measurements needed to achieve the RIP, with a high confidence level in terms of
(N, M, and K) by ignoring the impact of δ temporarily. This is given as,
Theorem 9.3 (Theorem 3.4 of [150]) Let be an M × N matrix that satisfies the
1
RIP of order 2 K with constant δ ∈ 0, . Then
2

N
M ≥ C K log (9.54)
K
√
where C = 1/2 log( 24 + 1) ≈ 0.28.
Based on the Johnson-Linderstrauss lemma [41] and related to the RIP, results in a
different bound on measurement for when δ is significantly lower. Namely, that
c0 log( p) 16c0 K
for small δ subject to M ≥ = an outcome for measurements
ε 2 c1 δ 2
K
of a RIP matrix of order K is proportional to 2 producing measurement bound of
δ
K log(N /K ) instead.
Finally, as shown in [105], a convenient result of the RIP condition surfaces, that
if a matrix satisfies the RIP then it also satisfies the NSP (see [105]). This relieves
the dependence on multiple checks of a matrix, providing a more efficient way to
test for compliance of a sensing matrix for use in CS recovery.
9.5.6 Compliant RIP Matrices
A pivotal technique is needed when generating a matrix that is potentially RIP com-
pliant; it is one thing to test for the RIP of a matrix and another problem entirely to
generate one. We follow the approach described in [17] whereby choosing entries φi j
from a probability distribution with two strict conditions on the distribution, ensures
that a matrix is RIP complaint. Firstly, the distribution must be norm-preserving
which leads to a variance of 1/M conditioned to
1
E(φi2j ) = . (9.55)
M
Secondly, the distribution must be sub-Gaussian14 stipulating that there exists a con-
stant c > 0 such that
E(eφi j t ) ≤ e−c t /2 for all t ∈ R.
2 2
(9.56)
The theorem that supports this result is given below.

Theorem 9.4 (Theorem 3.6 in [150]) Fix δ ∈ (0, 1). Let be an M × N random
matrix whose entries φi j are independent identically distributed (i.i.d) with φi j drawn
14 Sub-Gaussian refers to the decay rate of the tail of the distribution that must be similar to that of a
√
Gaussian distribution. Distributions that fit this definition is the Gaussian, Bernoulli with ±1/ M
[150].
according to a strictly sub-Gaussian distribution with c2 = 1/M. If

N
M ≥ k1 K log , (9.57)
K
then satisfies the RIP of order K with the prescribed δ with probability exceeding
1 − 2e−k2 M , where k1 is arbitrary and k2 = δ2 /2k ∗ − log(42e/δ)/k1 .
Although a sub-Gaussian random matrix can be chosen as a sensing matrix for
noisy measurements, in practice most signals that we are interested in, are not nat-
urally sparse but in some other basis Ψ . With this being the case, we then desire
that the product of the two matrices Ψ satisfy the RIP. Fortunately, based on the
findings in [105], if Ψ is an orthonormal basis it can be shown that its product with
the sub-Gaussian matrix is also Gaussian distributed. Provided that M is suffi-
ciently large, then Ψ will satisfy the RIP with a high confidence level. This result
is crucial, as it provides the means to construct numerous RIP compliant matrices
using a conventional transform basis.
9.5.7 Incoherence
The coherence of a given matrix, say given in Eq. 9.58 provides an easier check
for a unique recovery of a sparse signal, closely related to the spark, NSP and RIP.
As shown by Theorem 2 of [56] (see below) and work in [150] together with the
Welch Bound, provides an upper bound on the sparsity of a signal that guarantees
√ a
unique estimate on recovery given that the coherence is of order K = O( M).
Theorem 9.5 (Theorem 2 in [56]) The eigenvalues of an N × N matrix M with
entries m i j , 1 ≤ i. j ≤ N , lie in the union ofN discs di = di (ci , ri ), 1 ≤ i ≤ N ,
centred at ci = m ii and with the radius ri = j=i |m i j |.
Additionally, the coherence of a matrix is given as
| φi , φ j |
μ() = max (9.58)
1≤i< j≤N ||φi ||2 ||φ j ||2
with μ() the largest absolute inner product15 between any two columns φi , φ j
of . In [105] it is given √
that if M N then the lower bound for coherence is
approximately μ() ≥ 1/ M, also known as the Welch bound.
In essence, by means of [105], if has a low coherence μ() and spectral norm
||||2 , and if K = O(μ−2 () log N ), then using CS measurement y = x the input
signal x can be recovered with a high confidence level [150]. Moreover, if the Welch
bound is used in the definition of K , it gives K = O(M log N ) which results in a
15 The inner product is denoted here on out as β, α , with ℵ and denoting any variable.
linear dependence of sparsity and measurement, much like the condition imposed by
RIP [17].
9.6 Sampling Techniques
9.6.1 Sampling Theory
The system block used to mimic an ideal ADC, shown in Fig. 9.31, is the ideal
continuous-to-discrete-time (C/D) converter which will be used for explanatory pur-
poses. Given a analog continuous bandlimited input signal, the C/D gives as an output
the discrete time signal. Consider an input xc (t), sampled periodically with period
1
Ts = at the sampling frequency f s gives a discrete-time signal x[n] as an output,
fs
where x[n] = xc (nT ). see Fig. 9.31. Another way to look at the sampling, process,
is by means of mixing (see Appendix 9.2.2.3) xc (t) to xs (t) by means of a periodic
impulse response train.
∞
h(t) = δ(t − nT ) (9.59)
n=−∞
Mixed with the input signal xc (t) gives:

∞

xs (t) = xc (t) × h(t) = xc (t) δ(t − nT ) (9.60)
n=−∞
∞

= xc (nT )δ(t − nT ) (9.61)
n=−∞
Consequently, the Fourier transform of the continuous-time signals, xc (t) ⇒ X c ( jω)

: h(t) ⇒ H ( jω) is representative of a convolution in the frequency domain resulting
in:
∞
1 1
X s ( jω) = X c ( jω) ∗ H ( jω) = X c ( j (ω − kωs )) (9.62)
2π T k=−∞
Which in reality consists of copies of X c ( jω) shifted by integer multiples of the

sampling frequency f s and then superimposed [131]. As shown in Fig. 9.31, if the
bandlimited input signal is limited to half of the sampling frequency, then the original
signal can be recovered by means of a low pass filter (LPF). Otherwise, aliasing
will occur and the signal will be degraded significantly. As mentioned earlier, this
frequency is known as the Nyquist frequency, which is expressed as.
fs > ωN (9.63)
Fig. 9.31 An illustration of sampling a bandlimited signal, using an impulse train superimposed
and consequently filtered via a LPF to produce a non-ailised signal
The Nyquist frequency is given as ω N , and to recover the entire bandlimited signal
the frequency Nyquist rate is equal to 2ω N .
9.6 Sampling Techniques 169
Fig. 9.32 Illustrates the effect mixing has on a bandlimited signal, where down-conversion and
up-conversion are shown, respectively. Taken from [106]
9.6.2 Conventional Sampling Mixing-Acquisition
Conventional sampling scheme, utilizing an ADC for quantization, have a limited

bandwidth that can be exactly sampled without aliasing taking place due to limitations
on the ADC sampling rate as it related to the Nyquist criteria [128]. Two common
scenarios arise in any RF sampling system where either a limited bandwidth must
be sampled at a carrier frequency that exceed the limitations for current ADCs, or
the bandwidth itself is too wide requiring a different approach entirely.
For the first case, the typical approach makes use of a mixer and filter stage
(which can be done in the analog or digital domain) to down convert the bandwidth
to an appropriate sampling rate matched to the capability of the ADC, known as
the intermediate frequency (IF). Figure 9.32 shows how this applies in the frequency
domain. Although sampling at IF is the typically technique used for RF purposes (i.e.
SH receiver) [39], the inherent frequency ambiguities caused by mixer deviations and
drift are a source of concern, as errors occur and reduce the correct representation
of the acquired signal. Sampling at IF method is well established, cost effective,
and due to the reduced complexity of implementation is a preferred technique [39].
Moreover, for application in RF communications, this is typically the first method
for consideration in any system design.
For the second case, where an extensive bandwidth is required to be sampled,
most RF systems either use a sweeping technique as detailed in Sect. 2.2.2, employ-
ing a variable VCO to sweep across a wide bandwidth with a fixed sampling band.
Otherwise, a bank of samplers are used with each branch specific to a predeter-
mined sampled bandwidth that covers the entire spectrum of interest, as in a digital
receivers..
9.6.3 Under Sampling Techniques
9.6.3.1 Bandpass Sampling
It has long been known that discrete sampling of a bandlimited analog signal x(t)—
by taking advantage of aliasing—can result in high frequency signals being folded
down in frequency to an IF that can be recovered in the digital domain [179]. The
feasibility of this approach depends on a few factors that allow for coherent sampling
of the desired bandlimited signal at a specific carrier frequency. The method relies
on the assumption that modulated signals, which are unconverted for transmission at
a carrier frequency, still maintain the information embedded in the signal at a lower
oscillating frequency due to modulation. This lower information related modulation
is within a realistic bandwidth that can be under-sampled at a rate relative to the
modulation, instead of the Nyquist frequency dependent on the highest frequency in
the bandwidth. The sampling is conditioned on number of critical variables, namely
the carrier frequency ( f c ), upper frequency of the band ( f u ), lower frequency of the
band ( fl ), the sampling frequency (fs) and the bandwidth is given as B = f u − fl . In
order to sample the signal at zero IF without ambiguities, two crucial criteria must
hold. But first, it is necessary to define the band position which is measured from
the bandwidth origin to the lower band edge ( fl ) and is usually a fraction of the
bandwidth.
The criterion to ensure uniform sampling, with the sampling frequency at f s =
2B, dictates that the lower band edge must be an integer multiple of bandwidth B
i.e. fl = c( f u − fl ) with c = 0, ±1, ±2 . . . ± N , known as integer band positioning
[179]. Another case where uniform sampling will result for band sampling, when the
(2c + 1)
lower band edge is a half-integer multiple, conditioned to fl = ( f u − fl ),
2
is known as the half-integer positioning. Moreover, for uniform sampling one must
ensure that the sampling frequency f s obeys
2 fu 2 fl
≤ fs ≤ (9.64)
k k−1
so that aliasing of positive and negative edges do not overlap. Where k is some integer
multiple conditioned by 1 ≤ k ≤ f u /B.
The simplest and most effective sampling is to configure the band to be inte-
ger positioned [184]. In the sampling process the degradation of the signal due to
noise aliasing is unavoidable. However, by conditioning the sampling frequency to
quadrature sampling gives the least distortion ratio of the signal (see [179]). We refer
the reader to [184, 179] for further practical related issues when implementing this
method.
Bandpass sampling precedes other methods that resemble any sub-sampling or
sub-Nyquist qualities, allowing a system to operate with a lower ADC sampling
rate and still achieve adequate quantization of signals [4, 134]. Although bandpass
sampling has been used for a variety of applications (i.e. radar, communications,
9.6 Sampling Techniques 171
astronomy etc.) the application still depends on the signal being bandlimited at a
high operating frequency, low noise environment to limit noise distortion, as well
as minimal adjacent spectral occupation (due to the “folding” of the bands) [184].
Bandpassing imposes an inherent limit for multi-band signals and wideband signals,
hence for circumstances such as spectrum monitoring/sensing bandpass sampling is
not preferred.
9.6.3.2 Periodic Non-uniform Sampling (PNS)
PNS follows the same approach of bandpass sampling, by under-sampling an input

signal x(t). However, instead of the under-sampling rate being conditioned to a
single bandwidth, PNS adopts a sampling rate relative to the frequency support in
the spectrum for multiband signals that tend to be spectrally sparse [116].
As suggested by [90], a lower bound for the sampling rate can be achieved for
an input multiband signal with known frequency support relative to the number of
active bands in the spectrum N , with individuals bandwidths, B. The lower bound
results from the product of these values i.e. B × N .
In particular, the PNS allows for under-sampling with suitable recovery of the
signal by approximating this bound (B N ) without the need for complex circuitry
[116]. For practical implementation, PNS simply needs a set of time-delay elements
prior to the ADC, realizable by a time-interleaving ADC16 .
PNS consists of m under-sampling branches with related time-shifts y[n] = x(nTs +
φi ), 1 ≤ m ≤ where m/Ts is less than the Nyquist frequency f N Q with a final
processing step to recover the input signal x(t). In [86] it was demonstrated that a
bandpass signal could be exactly recovered using a PNS scheme at a sampling rate
of 2B samples/s, followed by [94] extending the approach to multiband signals.
Although PNS is a candidate for sub-Nyquist sampling, the practicality of the
approach is hampered by a reliance on Nyquist sampling track and hold circuits for
each branch in the interleaving ADC [89]. Also, due to imperfect time-delay produc-
tion in the separate sampling branches, requires a compensation for time skewing
causing frequency mismatch and harmonics [89] for correlative recovery. Combined
with a necessity for a priori knowledge of the multiband signal spectral support
impedes the effectiveness of PNS for generic and dynamic frequency scenarios.
9.6.4 Direct Sampling
In theory, as with ideal software radios [177], direct sampling would be capable
of sampling rates that could sample all communication signals (i.e. <30 GHz) in
16 A time interleaving ADC uses m ADCs with a sample rate equal to 1/m of the overall sampling
rate f N Q in parallel with one another. Each branch has a time delay imposed and the selection of
each branch is controlled by a MUX conditioned to the system clock.
Fig. 9.33 Illustrates the difference between traditional a and direct b sampling schemes. Taken
from [4]
order to remove the need of RF analog components (see Fig. 9.33). However, based
on current technologies the upper bound sampling rates restricts the acquisition to
a bandwidth of 1.75 GHz using direct sampling which has an associated monetary
cost.
Nevertheless, just as it is with most technological trends, current direct sampling
systems combine fast ADCs, bandpass sampling techniques, novel digital filtering,
and a bank of samplers to enable a large bandwidth to be sampled instantaneously,
mimicking the ideal case when an ADC will be available to directly sample a wide
bandwidth directly. Such a system can be found in [167] capable of sampling up to
20 GHz instantaneously.
Although the practical implementation of direct sampling have a large associate
monetary cost,17 it does carry benefits that comprise no frequency drift due to mixing,
digital processing where lossless processing is capable instead of analog processing
(i.e. filter, mixing, amplification), and instantaneous coverage of a wide bandwidth.
9.7 Wideband CS Sampling Techniques
9.7.1 Multi-rate Asynchronous Sub-nyquist Sampling

(MASS)
The MASS scheme is a unique approach to sampling, based on a multi channel

sampling scheme with sub-Nyquist ADCs that are prime multiples less than twice that
17 A system as shown in [167] will typically be in excess of R 2 million.

9.7 Wideband CS Sampling Techniques 173
of the Nyquist frequency ( f N Y Q ). First proposed by [168] with the intended purpose
for wideband spectrum monitoring used in cognitive radio (CR). In Fig. 9.34 the
proposed architecture depicts the system block diagram of MASS using a wideband
filter and parallel prime numbered sampling rate ADC channels with their respective
FFT block.
The subsequent frequency magnitude outputs from each branch is fed into a CS
recovery block utilizing a joint sparsity recovery technique [51], similar to that of
CMUX, in order to recover the spectrum magnitude plots. The reason behind the
prime numbered ADC branches stem from bandpass sampling theory [4] and the
effect of Nyquist folding of high frequency signals to baseband. However, instead
of a single ADC subject to bandpass sampling conditions (see Sect. 9.6.3.1), MASS
uses parallel prime numbered sampling rates as a bases to circumvent the effects of
bandpass aliasing, allowing for multiple band signal recovery over multiple wideband
signals (see [168] for further detail).
No further work, in the open literature, has yet been done as far as practical
implementation of this method, however the theory of MASS remains as one of the
most promising CS dependant wideband sampling techniques. The simulation results
in [168] suggests an application up to 20 GHz with detection capability in medium
to low SNR environments (see Fig. 9.35). Furthermore, as a sampling technique,
MASS exhibits one of the highest compression ratios as compared to similar multi-
channel sampling schemes [114, 162, 190] relative to the mean squared error (MSE)
in recovery of the spectrum.
Even though the implementation complexity of the system is relatively uncom-
plicated [168], the recovery for MASS is non-trivial and has a large requirement
for computation and memory load, but less than other CS based sampling schemes
for similar signal inputs. As a spectrum sensing technique MASS does not retain
any phase information after FFT processing, which is a drawback, but in reality CS
recovery phase information is mostly lost due to the inherent non-sparsity of phase.
Fig. 9.34 System block depicting the implementation of the MASS sampling scheme. Courtesy of
[168]
Fig. 9.35 a Illustrates the effect of sparsity and the compression ratio on the detection performance
of MASS, with received signals exposed to separate AWGN channels with average SNR = 10 dB.
b Illustrates the comparison between the proposed system and the existing approaches when the
compression ratio varies. Courtesy of [168]
9.7.2 Xampling-ADC and Modulated Wideband Converter

(MWC)
The application of Xampling technique relates to wideband scenarios with the built-
in assumption of a multiband input signal. In other words, the input signal constitutes
a finite number of bandlimited signals that are adequately spaced in frequency and
don not overlap, as shown in Fig. 9.36b.
The practical implemented modulated wideband converter (MWC) [115], which
follows the principles of Xampling [54, 117], leverage analog pre-processing tech-
niques via demodulation of the input signal to reduce sampling rates. However, for
MWC a multi-channel approach is adopted where all individual branches are modu-
lated down to baseband, sampled by a low-rate ADC, and then provide a final subset
of digital outputs yi [n], see the illustration of the MWC in Fig. 9.36.
The outputs from the MWC serves as the digital signal used by the continuous-to-
finite (CTF) block as shown in Fig. 9.37, used to infer the locations of the respective
bands and occupancy, and must be updated every time the band structure changes.
The output of the CTF is used for purposes of recovering the spectrum during the
final recovery stage.
A Hardware efficient realization of the MWC was implemented in [118] for a
multiband signal B N , with N = 6 bandlimited signals and adequate spacing, each
having a bandwidth B = 19 MHz. The set up follows the scenario of receiving three
concurrent transmissions with the Nyquist rate stated as f N Y Q = 2.075 GHz.
Adequate results were taken, which proved the feasibility of such a sampling scheme
for sub-Nyquist sampling using CS recovery techniques. However, the MWC flexi-
bility in handling non-multiband signals and frequency sparse signals were hampered
as the processing block struggle to group such signals in the contiguous block [182].
9.7 Wideband CS Sampling Techniques 175
Fig. 9.36 a Shows the system block diagram of the modulated wideband converter, and b the
spectrum of the output digital signals summed into a union subspace that is resolved via the CTF
and used by CS methods as inputs. Courtesy of [116, 118]
Fig. 9.37 a Showing the system block diagram of the CTF and b the hardware MWC boards
developed in [118]. Courtesy of [116]
The cost due to multiple ADCs and multichannel analog RF front end serves as
an impediment for adoption, as well as the limitation to finite multiband signals (i.e.
6–10). In realistic terms, a MWC cannot adequately recover a wideband spectrum
for use in spectrum monitoring applications (i.e. CR & Electronic Support) [182].
Furthermore, the case could be made that the MWC implementation of Xampling
can be likened to a filter-bank [58] design for wideband applications as the RF front
ends are similar in design.
References
1. Adamy, D.L.: EW 101 A First Course in Electronic Warfare. Artech House (2000). ISBN
9781580531696
2. Adamy, D.L.: EW 102 a Second Course in Electronic Warfare. Artech House (2004). ISBN
9781580536867
3. Adamy, D.L.: Introduction to Electronic Warfare Modeling and Simulation. Artech House
(2008). ISBN 1580534953
4. Akos, D.M., Stockmaster, M., Tsui, J.B.Y., Caschera, J.: Direct bandpass sampling of multiple
distinct RF signals. IEEE Trans. Commun. 47(7), 983–988 (1999a)
5. Alonso, M.T., López-Dekker, P., Mallorqui, J.J.: A novel strategy for radar imaging based on
compressive sensing. IEEE Trans. Geosci. Remote Sens. 48(12), 4285–4295 (2010)
6. Antoniadis, I., Glossiotis, G.: Cyclostationary analysis of rolling-element bearing vibration
signals. J. Sound Vib. 248(5), 829–845 (2001)
7. Ardoino, R., Megna, A.: LPI radar detection: SNR performances for a dual channel cross-
correlation based ESM receiver. In: Radar Conference, 2009. EuRAD 2009. European, pp.
113–116. IEEE (2009)
8. Asif, M.S., Romberg, J.: Sparse recovery of streaming signals using l1-homotopy (2013).
arXiv:1306.3331
9. Assaleh, K., Farrell, K., Mammone, R.J.: A new method of modulation classification for
digitally modulated signals. In: Military Communications Conference, 1992. MILCOM’92,
Conference Record. Communications-Fusing Command, Control and Intelligence., IEEE, pp.
712–716. IEEE (1992)
10. Axell, E., Leus, G., Larsson, E.G., Poor, H.V.: Spectrum sensing for cognitive radio: State-
of-the-art and recent advances. IEEE Signal Process. Mag. 29(3), 101–116 (2012a)
11. Axell, E., Leus, G., Larsson, E.G., Poor, H.V.: Spectrum sensing for cognitive radio: state-of-
the-art and recent advances. IEEE Signal Process. Mag. 29(3), 101–116 (2012b)
12. Azzouz, E.: Automatic Modulation Recognition of Communication Signals. Springer (1996)
13. Balanis, C.A.: Antenna Theory: Analysis and Design. Wiley (2012)
14. Balogh, L., Kollar, I.: Angle of arrival estimation based on interferometer principle. In: 2003
IEEE International Symposium on Intelligent Signal Processing, pp. 219–223. IEEE (2003)
15. Baraniuk, R.G., Cevher, V., Duarte, M.F., Hegde, C.: Model-based compressive sensing. IEEE
Trans. Inf. Theor. 56, 1982–2001 (2010)
16. Baraniuk, R., Steeghs, P.: Compressive radar imaging. In: 2007 IEEE Radar Conference, pp.
128–133. IEEE (2007)

DOI 10.1007/978-3-319-46700-9
178 References
17. Baraniuk, R.G.: Compressive sensing (lecture notes). IEEE Signal Process. Mag. 24(4), 118–
121 (2007)
18. Baraniuk, R., Davenport, M., DeVore, R., Wakin, M.: A simple proof of the restricted isometry
property for random matrices. Constr. Approx. 28(3), 253–263 (2008)
19. Baron, D., Sarvotham, S., Baraniuk, R.G.: Bayesian compressive sensing via belief propaga-
tion. IEEE Trans. Signal Process. 58(1), 269–280 (2010)
20. Barrick, D.E., Lipa, B.J.: Radar angle determination with music direction finding, 23 Nov
1999. US Patent 5,990,834
21. Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse
problems. SIAM J. Imaging Sci. 2(1), 183–202 (2009)
22. Boyd, S., Vandenberghe, L.: Convex optimization, October, 2013. http://www.stanford.edu/
~boyd/cvxbook/bv_cvxbook.pdf
23. Broto, D.: Antenna basic theory (2013). http://denmasbroto.com/article-13-antenna-basic-
theory.html
24. Calderbank, R., Howard, S., Jafarpour, S.: Construction of a large class of deterministic sensing
matrices that satisfy a statistical isometry property. IEEE J. Sel. Topics Signal Process. 4(2),
358–374 (2010)
25. Candes, E.J., Romberg, J.K., Tao, T.: Stable signal recovery from incomplete and inaccurate
measurements. Commun. Pure Appl. Math. 59(8), 1207–1223 (2006)
26. Candes, E., Romberg, J.: 1 −magic, Feb 2013. http://users.ece.gatech.edu/~justin/l1magic/#
code
27. Candes, E., Romberg, J.: l1-magic: recovery of sparse signals via convex programming (2005).
www.acm.caltech.edu/l1magic/downloads/l1magic.pdf
28. Candès, E.J.,: Compressive sampling. In: Proceedings oh the International Congress of Math-
ematicians: Madrid, August 22–30, 2006: Invited Lectures, pp. 1433–1452 (2006)
29. Carlin, M., Rocca, P., Oliveri, G., Viani, F., Massa, A.: Directions-of-arrival estimation through
bayesian compressive sensing strategies. IEEE Trans. Antennas Propag. 61(7), 3828–3838
(2013)
30. Chan, Y.T., Plews, J.W., Ho, K.C.: Symbol rate estimation by the wavelet transform. In: Pro-
ceedings of 1997 IEEE International Symposium on Circuits and Systems, 1997, ISCAS’97,
vol. 1, pp. 177–180. IEEE (1997)
31. Chen, S.S., Donoho, D.L., Saunders, M.A.: Atomic decomposition by basis pursuit. SIAM J.
Sci. Comput. 20(1), 33–61 (1998)
32. Cloete, C.: International ew trends and strategic influences, May, 2013. http://www.
aardvarkaoc.co.za/Proceedings/201305
33. Cohen, A., Dahmen, W., DeVore, R.: Compressed sensing and best k-term approximation. J.
Am. Math. Soc. 22(1), 211–231 (2009)
34. Collin, R.E.: Antennas and Radiowave Propagation, vol. 108. McGraw-Hill, New York (1985)
35. Cormode, G., Muthukrishnan, S.: Combinatorial algorithms for compressed sensing. In: Struc-
tural Information and Communication Complexity, pp. 280–294. Springer (2006)
36. Cripps, S.C.: RF Power Amplifiers for Wireless Communications, (Artech House Microwave
Library). Artech House, Inc. (2006)
37. Croisier, A., Esteban, D., Galand, C.: The diffraction of light by high frequency sound waves:
Part 1. In: Proceedings of the International Symposium on Information Circuis and Systems
(1976)
38. Dasgupta, Sanjoy: Gupta, Anupam: An elementary proof of the johnson-lindenstrauss lemma,
99–006. International Computer Science Institute, Technical report (1999)
39. Davenport, M., Duarte, M., Eldar, Y., Kutyniok, G.: Theory and Application. Compressed
sensing, Cambridge University Press (2012)
40. De Martino, A.: Introduction to EW systems. Artech House, Norwood (2012)
41. De Vore, R.A.: Deterministic constructions of compressed sensing matrices. J. Complex.
23(4), 918–925 (2007)
42. Donoho, D.L.: Compressed sensing. IEEE Trans. Inf. Theor. 52(4), 1289–1306 (2006a)
References 179
43. Donoho, D.L.: For most large underdetermined systems of linear equations the minimal
l1-norm solution is also the sparsest solution. Commun. Pure Appl. Math. 59(6), 797–829
(2006b)
44. Donoho, D.L., Elad, M.: Optimally sparse representation in general (nonorthogonal) dictio-
naries via l1 minimization. Proc. Natl. Acad. Sci. 100(5), 2197–2202 (2003)
45. Drentea, C.: Modern Communications Receiver Design and Technology. Artech House
(2010a)
46. Duarte, M.F., Baraniuk, R.G.: Spectral compressive sensing. Appl. Comput, Harmonic Anal
(2012)
47. Duarte, M.F., Davenport, M.A., Takhar, D., Laska, J.N., Sun, T., Kelly, K.F., Baraniuk, R.G.:
Single pixel imaging via compressive sampling. IEEE Signal Process. Mag. 25(2), 83–91
(2008)
48. Duarte, M.F., Sarvotham, S., Baron, D., Wakin, M.B., Baraniuk, R.G.: Distributed compressed
sensing of jointly sparse signals. In: Asilomar Conference on Signals, System, Computation,
pp. 1537–1541 (2005)
49. Eldar, Y.C., Kutyniok, G.: Compressed Sensing: Theory and Applications. Cambridge Uni-
versity Press (2012)
50. Eldar, Y.C., Mishali, M.: Block sparsity and sampling over a union of subspaces. In: 2009
16th International Conference on Digital Signal Processing, pp. 1–8. IEEE (2009)
51. Emmanuel J Candes and Terence Tao. Decoding by linear programming. IEEE Trans. Inf.
Theor. 51(12), 4203–4215 (2005)
52. Ender, J.H.G.: On compressive sensing applied to radar. Signal Process. 90(5), 1402–1414
(2010)
53. Falk, I.S.: Iterative einschließung der kleinsten (größten) eigenwerte eines hermiteschen
matrizenpaares, ii. Acta Mech. 49(1–2), 111–131 (1983)
54. Fannjiang, A., Liao, W.: Compressed sensing phase retrieval. In: 2011 Conference Record of
the Forty Fifth Asilomar Conference on Signals, Systems and Computers (ASILOMAR), pp.
735–738. IEEE (2011)
55. Farhang-Boroujeny, B.: Filter bank spectrum sensing for cognitive radios. IEEE Trans. Signal
Process. 56(5), 1801–1811 (2008)
56. Fernandez, M.R.: High performance computing—nodes, sockets and flops (2014). http://en.
community.dell.com/techcenter/high-performance-computing/w/wiki/2329
57. Figueiredo, M.A.T., Nowak, R.D., Wright, S.J.: Gradient projection for sparse reconstruction:
application to compressed sensing and other inverse problems. IEEE J. Sel. Top. Signal
Process. 1(4), 586–597 (2007)
58. Finne, M.: Methods for Direction-Finding of Direct-Sequence Spread-Spectrum Signals.
DIANE Publishing (1996)
59. Fuller, K.L.: To see and not be seen. In: IEE Proceedings F (Radar and Signal Processing),
pp. 1–9. IET (1990)
60. Gabriel, W.F.: Spectral analysis and adaptive array superresolution techniques. Proc. IEEE
68(6), 654–666 (1980)
61. Gardner, W.A.: The spectral correlation theory of cyclostationary time-series. Signal Process.
11(1), 13–36 (1986)
62. Gilbert, A.C., Strauss, M.J., Tropp, J.A., Vershynin, R.: One sketch for all: fast algorithms
for compressed sensing. In: Proceedings of the Thirty-Ninth Annual ACM Symposium on
Theory of Computing, pp. 237–246. ACM (2007)
63. Golub, G.H., Hansen, P.C., O’Leary, D.P.: Tikhonov regularization and total least squares.
SIAM J. Matrix Anal. Appl. 21(1), 185–194 (1999)
64. Gonzalez, G.: Microwave Transistor Amplifiers: Analysis and Design, vol. 2. Prentice Hall,
New Jersey (1997)
65. Graham, A.: Communications, Radar and Electronic Warfare. Wiley.com (2011)
66. Grout, I.: Digital Systems Design with FPGAs and CPLDs. Newnes (2011)
67. Gurbuz, A.C., McClellan, J.H., Scott, W.R.: A compressive sensing data acquisition and
imaging method for stepped frequency GPRs. IEEE Trans. Signal Process. 57(7), 2640–2650
(2009)
180 References
68. Hale, E.T., Yin, W., Zhang, Y.: A fixed-point continuation method for l1-regularized mini-
mization with applications to compressed sensing. CAAM TR07-07, Rice University (2007)
69. Haupt, J., Bajwa, W.U., Rabbat, M., Nowak, R.: Compressed sensing for networked data.
IEEE Signal Process. Mag. 25(2), 92–101 (2008)
70. Haupt, J., Castro, R., Nowak, R., Fudge, G., Yeh, A.: Compressive sampling for signal
classification. In: Fortieth Asilomar Conference on Signals, Systems and Computers, 2006.
ACSSC’06, pp. 1430–1434. IEEE (2006)
71. Hong, L., Ho, K.C.: Identification of digital modulation types using the wavelet transform.
In: Military Communications Conference Proceedings, 1999. MILCOM 1999. IEEE, vol. 1,
pp. 427–431. IEEE (1999)
72. Jacques, Laurent: Vandergheynst, Pierre: Compressed sensing: when sparsity meets sampling.
Chapter 23, 507–527 (2010)
73. Ji, S., Xue, Y., Carin, L.: Bayesian compressive sensing. IEEE Trans. Signal Process. 56(6),
2346–2356 (2008)
74. Johnk, C.T.A.: Engineering Electromagnetic Fields and Waves, 667 p. John Wiley, New York
(1975)
75. Johnson, R.C., Jasik, H.: Antenna Engineering Handbook, 1356 p. McGraw-Hill Book Com-
pany, New York (1984). No individual items are abstracted in this volume 1, 1984
76. Jorge, O.: Estimation, detection, and identification (2014). http://isr.ist.utl.pt/~pjcro/courses/
def0910/docs/EFI3CRLB.pdf
77. Kay, S.M.: Fundamentals of Statistical Signal Processing, Volume III: Practical Algorithm
Development, vol. 3. Pearson Education (2013)
78. Kim, S., Koh, K., Lustig, M., Boyd, S., Gorinevsky, D.: An interior-point method for large-
scale 1 -regularized least squares. IEEE J. Sel. Top. Signal Process. 1(4), 606–617 (2007)
79. Kirolos, S., Laska, J., Wakin, M., Duarte, M., Baron, D., Ragheb, T., Massoud, Y., Baraniuk,
R.: Analog-to-information conversion via random demodulation. In: 2006 IEEE Dallas/CAS
Workshop on Design, Applications, Integration and Software, pp. 71–74. IEEE (2006)
80. Kohlenberg, A.: Exact interpolation of band-limited functions. J. Appl. Phys. 24(12), 1432–
1436 (1953)
81. Kotelnikov, V.: On the carrying capacity of the ether and wire in telecommunications. Izd.
Red. Upr, Svyazi RKKA (1993)
82. Kozminchuk, B., Elsaesser, D.: Advanced communication esm system design considerations
(2013). http://www.dtic.mil/dtic/tr/fulltext/u2/a259014.pdf
83. Kuang, Y., Shi, Q., Chen, Q., Yun, L., Long, K.: A simple way to deinterleave repetitive
pulse sequences. In: 7th WSEAS International Conference on Mathematical Methods and
Computational Techniques in Electrical Engineering, Sofia, pp. 218–222 (2005)
84. Kurosawa, N., Kobayashi, H., Maruyama, K., Sugawara, H., Kobayashi, K.: Explicit analy-
sis of channel mismatch effects in time-interleaved adc systems. IEEE Trans. Circ. Syst. I
Fundam. Theor. Appl. 48(3), 261–271 (2001)
85. Landau, H.J.: Necessary density conditions for sampling and interpolation of certain entire
functions. Acta Math. 117(1), 37–52 (1967)
86. Laska, J., Kirolos, S., Massoud, Y., Baraniuk, R., Gilbert, A., Iwen, M., Strauss, M.: Random
sampling for analog-to-information conversion of wideband signals. In: 2006 IEEE Dal-
las/CAS Workshop on Design, Applications, Integration and Software. pp. 119–122. IEEE
(2006)
87. Levy, R., Cohn, S.B.: A history of microwave filter research, design, and development. IEEE
Trans. Microw. Theor. Tech. 32(9), 1055–1067 (1984)
88. Liang, Y.-C., Chen, K.-C., Li, G.Y., Mahonen, P.: Cognitive radio networking and communi-
cations: an overview. IEEE Trans. Veh. Technol. 60, 3386–3407 (2011)
89. Lin, Y.-P., Vaidyanathan, P.P.: Periodically nonuniform sampling of bandpass signals. IEEE
Trans. Circ. Syst. II: Anal. Digit. Signal Process. 45(3), 340–351 (1998)
90. Liu, D., Pfeiffer, U., Grzyb, J., Gaucher, B.: Advanced Millimeter-Wave Technologies: Anten-
nas. Wiley, Packaging and Circuits (2009)
References 181
91. Looney, M.: Advanced digital post-processing techniques enhance performance in time-
interleaved adc systems, October, 2013. http://www.analog.com/library/analogdialogue/
archives/37-08/post_processing.html
92. Lopez-Risueno, G., Grajal, J., Sanz-Osorio, A.: Digital channelized receiver based on time-
frequency analysis for signal interception. IEEE Trans. Aerosp. Electron. Syst. 41(3), 879–898
(2005)
93. Loris, I.: On the performance of algorithms for the minimization of 1 -penalized functionals.
Inverse Probl. 25(3), 035008 (2009)
94. Louis, E.K.: Electromagnetic spectrum, September, 2013. http://www.unwittingvictim.com/
Electromagnetic-Spectrum-3H.PNG
95. Lu, W., Vaswani, N.: Modified basis pursuit denoising (modified-bpdn) for noisy compressive
sensing with partially known support. In: 2010 IEEE International Conference on Acoustics
Speech and Signal Processing (ICASSP), pp. 3926–3929. IEEE (2010)
96. Lustig, M., Donoho, D., Pauly, J.M.: Sparse mri: the application of compressed sensing for
rapid mr imaging. Magn. Reson. Med. 58(6), 1182–1195 (2007)
97. Ma, S.: Alternating direction method of multipliers for sparse principal component analysis.
J. Oper. Res. Soc. China 1–22 (2011)
98. Mallat, S.G., Zhang, Z.: Matching pursuits with time-frequency dictionaries. IEEE Trans. Sig.
Process. 41(12), 3397–3415 (1993)
99. Marchesini, S.: Ab initio compressive phase retrieval (2008). arXiv preprint arXiv:0809.2006
100. Marki, F., Marki, C.: Taking measure of microwave mixers (2013). http://mwrf.com/
components/taking-measure-microwave-mixers
101. Matuszewski, J., Paradowski, L.: The knowledge based approach for emitter identification.
In: 12th International Conference on Microwaves and Radar, 1998. MIKON’98, vol. 3, pp.
810–814. IEEE (1998)
102. Matuszewski, J.: Specific emitter identification. In: Radar Symposium, 2008 International,
pp. 1–4. IEEE (2008)
103. Maurer, D.E., Chamlou, R., Genovese, K.O.: Signal processing algorithms for electronic
combat receiver applications. Johns Hopkins APL Tech. Digest 18(1), 69 (1997)
104. Maxwell, J.C.: A dynamical theory of the electromagnetic field. Philos. Trans. R. Soc. Lond.
155, 459–512 (1865)
105. Meyer, C.: Matrix analysis and applied linear algebra book and solutions manual. SIAM, 2
(2000)
106. Milojević, D.J., Popović, B.M.: Improved algorithm for the deinterleaving of radar pulses. In:
IEE Proceedings F (Radar and Signal Processing), vol. 139, pp. 98–104. IET (1992)
107. Mishali, M., Eldar, Y.C., Dounaevsky, O., Shoshan, E.: Xampling: analog to digital at sub-
nyquist rates. IET Circ. Dev. Syst. 5(1), 8–20 (2011)
108. Mishali, M., Eldar, Y.C., Dounaevsky, O., Shoshan, E.: Xampling: analog to digital at sub-
nyquist rates. IET Circ. Dev. Syst. 5(1), 8–20 (2011)
109. Mishali, M., Eldar, Y.C., Elron, A.: Xampling–part 1: Practice. Arxiv preprint (2009)
110. Mishali, M., Eldar, Y.C.: Blind multiband signal reconstruction: Compressed sensing for
analog signals. IEEE Trans. Sig. Process. 57(3), 993–1009 (2009)
111. Mishali, M., Eldar, Y.C.: Sub-nyquist sampling. IEEE Sig. Process. Mag. 28(6), 98–124 (2011)
112. Mishali, M., Eldar, Y.C.: Wideband spectrum sensing at sub-nyquist rates. IEEE Sig. Process.
Mag. 28(4), 102–135 (2011)
113. Moravec, M., Romberg, J., Baraniuk, R.: Compressed sensing phase retrieval. SPIE, Wavelets
XII 6701, 91 (2007)
114. Nandi, A.K., Azzouz, E.E.: Algorithms for automatic modulation recognition of communi-
cation signals. IEEE Trans. Commun. 46(4), 431–436 (1998)
115. National Instruments. Filter specifications (2013). http://zone.ni.com/reference/en-XX/help/
371325F-01/lvdfdtconcepts/dfd_filter_spec/
116. Naval Air Systems Command. Electronic warfare and radar systems engineering handbook,
January, 2013. http://www.microwaves101.com/encyclopedia/navy
182 References
117. Needell, D., Tropp, J.A.: Cosamp: iterative signal recovery from incomplete and inaccurate
samples. Appl. Comput. Harmon. Anal. 26(3), 301–321 (2009)
118. Needell, D., Tropp, J.A.: Cosamp: iterative signal recovery from incomplete and inaccurate
samples. Commun. ACM 53(12), 93–100 (2010)
119. Needell, D., Vershynin, R.: Signal recovery from incomplete and inaccurate measurements
via regularized orthogonal matching pursuit. IEEE J. Sel. Top. Sig. Process. 4(2), 310–316
(2010)
120. Neri, F.: Introduction to Electronic Defense Systems. SciTech Publishing (2005)
121. Neri, F.: Introduction to Electronic Defense Systems. Artech House, Norwood (1991)
122. Nesterov, Y., Nemirovskii, A.S., Ye, Y.: Interior-point polynomial algorithms in convex pro-
gramming. SIAM 13 (1994)
123. Nyquist, H.: Certain topics in telegraph transmission theory. Trans. AIEE 47, 617–644 (1928)
124. Ohlsson, H., Yang, A.Y., Dong, R., Sastry, S.S.: Compressive phase retrieval from squared
output measurements via semidefinite programming (2011). arXiv preprint arXiv:1111.6323
125. Oppenheim, A.V., Schafer, R.W., Buck, J.R., et al.: Discrete-Time Signal Processing, vol. 5.
Prentice Hall, Upper Saddle River (1999)
126. Paige, C.C., Saunders, M.A.: Lsqr: an algorithm for sparse linear equations and sparse least
squares. ACM Trans. Math. Softw. (TOMS) 8(1), 43–71 (1982)
127. Parisi, C.: Sensors smart sensors and sensor control electronics (2013). http://www.eeweb.
com/blog/carmen_parisi/P10
128. Patel, M., Darwazeh, I., O’Reilly, J.J.: Bandpass sampling for software radio receivers, and
the effect of oversampling on aperture jitter. In: IEEE 55th Vehicular Technology Conference,
2002. VTC Spring 2002, vol. 4, pp. 1901–1905. IEEE (2002)
129. Pati, Y.C., Rezaiifar, R., Krishnaprasad, P.S.: Orthogonal matching pursuit: recursive function
approximation with applications to wavelet decomposition. In: 1993 Conference Record of
The Twenty-Seventh Asilomar Conference on Signals, Systems and Computers, 1993, pp.
40–44. IEEE (1993)
130. Peleg, S., Porat, B.: The cramer-rao lower bound for signals with constant amplitude and
polynomial phase. IEEE Trans. Sig. Process. 39(3), 749–752 (1991)
131. Pesavento, M., Gershman, A.B., Haardt, M.: Unitary root-music with a real-valued eigende-
composition: a theoretical and experimental performance study. IEEE Trans. Sig. Process.
48(5), 1306–1314 (2000)
132. Po, K., Takada, J.-I.: Signal detection based on cyclic spectrum estimation for cognitive radio
in ieee 802.22 wran system. The Institute of Electronics, Information and Communication
Engineers, Technical Reportof IEICE (2007)
133. Poisel, R.A.: Introduction to communication electronic warfare systems. Artech House (2008).
ISBN 1580533442
134. Polo, Y.L., Wang, Y., Pandharipande, A., Leus, G.: Compressive wide-band spectrum sens-
ing. In: IEEE International Conference on Acoustics, Speech and Signal Processing, 2009.
ICASSP 2009, pp. 2337–2340. IEEE (2009)
135. Polydoros, A., Kim, K.: On the detection and classification of quadrature digital modulations
in broad-band noise. IEEE Trans. Commun. 38(8), 1199–1211 (1990)
136. Poynting defence. Tactical direction finding antenna (2013). http://www.poyntingdefence.
com/index.php?q=catalogue|productinfo,39,Tactical-Direction-Finding-Antenna
137. Pozar, D.M.: Microwave and RF Design of Wireless Systems. Wiley (2000)
138. Quinn, D.: An fpga implementation feasibility study of the correlative interferometer direction
finding algorithm—part 2: in depth. Technical report, NUTAQ (2013). http://nutaq.com/en/
blog/fpga
139. Quiroga, R.Q., Nadasdy, Z., Ben-Shaul, Y.: Unsupervised spike detection and sorting with
wavelets and superparamagnetic clustering. Neural Comput. 16(8), 1661–1687 (2004)
140. Ragheb, T., Laska, J.N., Nejati, H., Kirolos, S., Baraniuk, R.G., Massoud, Y.: A prototype
hardware for random demodulation based compressive analog-to-digital conversion. In: 51st
Midwest Symposium on Circuits and Systems, 2008. MWSCAS 2008, pp. 37–40. IEEE
(2008)
References 183
141. Rahman, Arifur: FPGA Based Design and Applications. Springer (2008)
142. Ren, Q.S., Willis, A.J.: Fast root music algorithm. Electron. Lett. 33(6), 450–451 (1997)
143. Rice, D.W.: Hf direction finding by wave front testing in a fading signal environment. Radio
Sci. 17(4), 827–836 (1982). ISSN 1944-799X. doi:10.1029/RS017i004p00827
144. Richard, B., Mark, D., Marco, D., Chinmay, H., Jason, L., Mona, S., Wotao, Y.: An Introduction
to Compressive Sensing (2012). http://cnx.org/content/col11133/1.5/
145. Rohde, U.L., Rudolph, M.: RF/microwave circuit design for wireless applications. Wiley.com
(2013)
146. Romberg, J.: Imaging via compressive sampling. IEEE Signal Process. Mag. 25(2), 14–20
(2008)
147. Roy, R., Kailath, T.: Esprit-estimation of signal parameters via rotational invariance tech-
niques. IEEE Trans. Acoust. Speech Signal Process. 37(7), 984–995 (1989)
148. Saeid, S.H.: Study of the cell towers radiation levels in residential areas. In: Communication
Systems, p. 87 (2013)
149. Schaller, R.R.: Moore’s law: past, present and future. IEEE Spect. 34(6), 52–59 (1997)
150. Schleher, D.C.: Electronic Warfare in the Information Age, 1st edn. Artech House Inc, Nor-
wood, MA, USA (1999). ISBN 0890065268
151. Schleher, D.C.: Introduction to Electronic Warfare. Artech House, Norwood, MA (1994)
152. Schmidt, R.: Multiple emitter location and signal parameter estimation. IEEE Trans. Antennas
Propag. 34(3), 276–280 (1986)
153. Shan, T.-J., Kailath, T.: Adaptive beamforming for coherent signals and interference. IEEE
Trans. Acoust. Speech Signal Process. 33(3), 527–536 (1985)
154. Shannon, C.: Communication in the presence of noise. Proc. Inst. Radio Eng. 37(1), 108211:21
(1949)
155. Sills, J.A.: Maximum-likelihood modulation classification for psk/qam. In: Military Commu-
nications Conference Proceedings, 1999. MILCOM 1999. IEEE, vol. 1, pp. 217–220. IEEE
(1999)
156. Slavinsky, J.P., Laska, J.N., Davenport, M.A., Baraniuk, R.G.: The compressive multiplexer
for multi-channel compressive sensing. In: 2011 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), pp. 3980–3983. IEEE (2011)
157. Soliman, S.S., Hsue, S.-Z.: Signal classification using statistical moments. IEEE Trans. Com-
mun. 40(5), 908–916 (1992)
158. Stearns, S.D., Hush, D.R.: Digital Signal Processing with Examples in MATLAB. CRC Press
(2012)
159. Stephens, J.P.: Advances in signal processing technology for electronic warfare. IEEE Aerosp.
Electron. Syst. Mag. 11(11), 31–38 (1996)
160. Stojnic, M., Xu, W., Hassibi, B.: Compressed sensing-probabilistic analysis of a null-space
characterization. In: IEEE International Conference on Acoustics, Speech and Signal Process-
ing, 2008. ICASSP 2008, pp. 3377–3380. IEEE (2008)
161. Strmbck, P., Schultz, H.: Direct sampled wide band digital receivers in rwr/esm applications,
October 2013. http://www.aardvarkaoc.co.za/index_files/Page5847.htm
162. Sun, H., Nallanathan, A., Wang, C.-X., Chen, Y.: Wideband spectrum sensing for cognitive
radio networks: a survey. IEEE Wirel. Commun. 20, 74–81 (2013)
163. Texas Instruments. Data converters, October 2013a. http://www.ti.com/lsds/ti/data-
converters/high-speed-adc-greater-than-1gsps-rf-sampling.page
164. Texas Instruments. Digital signal processors (2014). http://www.ti.com/lsds/ti/dsp/overview.
page
165. Texas Instruments. Low noise amplifiers, October 2013b. http://www.ti.com/lsds/ti/
amplifiers-linear
166. Tian, Z., Giannakis, G.B.: Compressed sensing for wideband cognitive radios. In: IEEE Inter-
national Conference on Acoustics, Speech and Signal Processing, 2007. ICASSP 2007, vol.
4, pp. IV–1357. IEEE (2007)
167. Tropp, J.A., Gilbert, A.C.: Signal recovery from random measurements via orthogonal match-
ing pursuit. IEEE Trans. Inf. Theor. 53(12), 4655–4666 (2007a)
184 References
168. Tropp, J.A., Laska, J.N., Duarte, M.F., Romberg, J.K., Baraniuk, R.G.: Beyond nyquist: effi-
cient sampling of sparse band limited signals. IEEE Trans. Inf. Theor. 56(1), 520–544 (2010)
169. Tsaig, Y., Donoho, D.L.: Extensions of compressed sensing. Signal Process. 86(3), 549–571
(2006)
170. Tsui, J.: Digital Techniques for Wideband Receivers. SciTech Publishing (2004)
171. Tuncer, T.E., Friedlander, B.: Classical and Modern Direction-of-Arrival Estimation. Access
Online via Elsevier (2009)
172. Tuttlebee, W.H.W.: Software defined radio: origins, drivers, and international perspectives.
John Wiley (2002)
173. University of Berkley: Software (2013). http://www.eecs.berkeley.edu/~yang/software/
softwarepage.html
174. Vaccaro, D.D.: Electronic Warfare Receiving Systems. Artech House (1993)
175. Vaughan, R.G., Scott, N.L., White, D.R.: The theory of bandpass sampling. IEEE Trans.
Signal Process. 39(9), 1973–1984 (1991)
176. Venosa, E.: Eeweb electrical engineering home, how to specify an adc for a digital commu-
nications receiver, February, 2013. http://www.eeweb.com/blog/elettra_venosa
177. Vizmuller, P.: Radio Frequency Design Guide. Artech House (1995)
178. Wakin, M., Becker, S., Nakamura, E., Grant, M., Sovero, E., Ching, D., Yoo, J., Romberg, J.,
Emami-Neyestanak, A., Candes, E.: A nonuniform sampler for wideband spectrally-sparse
environments. IEEE J. Emerg. Sel. Top. Circ. Syst. 2(3), 516–529 (2012)
179. Walden, R.H.: Analog-to-digital converter survey and analysis. IEEE J. Sel. Areas Commun.
17(4), 539–550 (1999)
180. Waters, W.M., Jarrett, B.R.: Bandpass signal sampling and coherent detection. IEEE Trans.
Aerosp. Electron. Syst. 1(6), 731–736 (1982)
181. Wepman, J.A.: Analog-to-digital converters and their applications in radio receivers. IEEE
Commun. Mag. 33(5), 39–45 (1995)
182. White Paper Altera. Radar provessing: Fpgas or gpus (2014). https://www.altera.com/en_US/
pdfs/literature/wp/wp-01197-radar-fpga-or-gpu.pdf
183. White, W.D.: Frequency domain conversion, September 27 1960. US Patent 2,954,465
184. Whittaker, E.: On the function which are represented by the expansions of interpolation theory.
Proc. Roy. Soc. Edinburgh, Sect. A 35, 1818211;194 (1915)
185. Wright, S.J., Nowak, R.D., Figueiredo, M.A.T.: Sparse reconstruction by separable approxi-
mation. IEEE Trans. Signal Process. 57(7), 2479–2493 (2009)
186. Yang, A.Y., Sastry, S.S., Ganesh, A., Ma, Y.: Fast 1 -minimization algorithms and an appli-
cation in robust face recognition: a review. In: 2010 17th IEEE International Conference on
Image Processing (ICIP), pp. 1849–1852. IEEE (2010)
187. Yoo, J., Becker, S., Monge, M., Loh, M., Candes, E., Emami-Neyestanak, A.: Design and
implementation of a fully integrated compressed-sensing signal acquisition system. In: 2012
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp.
5325–5328. IEEE (2012)
188. Ziskind, I., Wax, M.: Maximum likelihood localization of multiple sources by alternating
projection. IEEE Trans. Acoust. Speech Signal Process. 36(10), 1553–1560 (1988)

Compressive Sensing Based Algorithms For Electronic Defence

Uploaded by

Copyright:

Available Formats

Compressive Sensing Based Algorithms For Electronic Defence

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Compressive Sensing Based Algorithms For Electronic Defence

Uploaded by

Copyright:

Available Formats

What are the main components of electronic defence systems?

What are the main components of electronic defence systems?

What are some applications of electronic support?

What are some applications of electronic support?

Signals and Communication Technology

Compressive Sensing Based

ISSN 1860-4862 ISSN 1860-4870 (electronic)

© Springer International Publishing AG 2017

Printed on acid-free paper

This Springer imprint is published by Springer Nature

Part I Electronic Defence and Compressive Sensing

3.3 Compressive Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 39

Part II Simulations of Compressive Sensing Used for Electronic

6.3 Simulation 2.2—CS DOA for 2PSK Signals . . . ............. 98

Part III Concluding Statements and Appendices

9.5 Compressive Sensing Mathematical Fundamentals . . . . . . . . . . . . 158

In this book we describe applications of Compressive Sensing (CS) based algorithms

1 Electronic support is a sub-category of electronic defence (ED). ES equipment is developed to

© Springer International Publishing AG 2017 3

deployment is possible without requiring additional hardware. In other words, CS can

1.1 Motivation and Problem Statement

The abundance of communication signal encountered in ES scenario presents a chal-

1.2 Outline and Contribution

Chapter 8: Concluding Remarks

© Springer International Publishing AG 2017 7

2.1.1 Electronic Defence Overview

2 Definitions sourced and modified from [1, 2].

Fig. 2.2 Illustrates the

• ES—This is the same as the classical definition of ESM.

locations of enemy weapons or electronic capabilities. ES systems typically need

2.1.2 Electronic Support

2.1.3 Electronic Attack

2.1.4 Electronic Protect

EP Systems are designed to restrict penetration and susceptibility of friendly systems

Some of the major kinds of EP measures are detailed below.

2.2 Electronic Support Communication Applications

2.2.1 Communication Electronic Support—CES

Communication Electronic Support systems provide immediate emitter signal infor-

1. Antenna array, composed of different sub-band antennas typically covering the

2.2.2 Communication Intelligence—COMINT

Communication Intelligence systems are used to analyse signal over an extended

• Signal Acquisition and feature extraction;

2.2.3 Signal Processing Techniques

Any intelligent receiver system requiring classification and/or identification of signal

methods to classify, identify, and extract features from RF signals in ES communi-

2.2.4 Signal Classification

Classification techniques in the equipment mentioned earlier serve to recognize spe-

• Pattern Recognition Processing—uses signal feature knowledge processed by

• Decision Theoretic Approach—depends on likelihood or probability to deter-

2.2.5 Signal Feature Extraction

4 “Deinterleaving is a kind of clustering analysis, which clusters inter-weaved pulses—intercepted

by a scout or by other means—into distinct groups belonging to respective emitters, according to

2.2.6 Emitter Identification

2.3 Direction of Arrival Methods Used for Electronic

2.4 DOA Methods

Table 2.3 DOA measurement method comparison, sourced from [37]

Phase interferometry on the other hand, is widely used in ES equipment [37].

2.4.1 Phase Interferometry

• front end channelized receiver circuitry, including mixers and analog-to-digital

minc0 , z subject to A0 z = b, (3.10)

min z c0 , z subject to A0 z = b, (3.17)

mization problem as a 1 -regularized least squares (LS) problem. By substituting the

Algorithm 3 Truncated Newton IPM for 1 −regularized LSPs. Courtesy of [84]

Finally, the development of the 1 −regularized LS algorithm7 proved to be an

Algorithm 4 1 − homotopy algorithm. Courtesy of [9]