Digital Communication Systems

Engineering with
Software-Defined Radio
 To my parents, Lingzhen and Xuexun
DP

To my wife, Jen, my parents, Ted and Barbara,

and my sisters Laura and Joanne
AMW
Digital Communication Systems
Engineering with
Software-Defined Radio

Di Pu
Alexander M. Wyglinski

artechhouse.com
Library of Congress Cataloging-in-Publication Data
A catalog record for this book is available from the U.S. Library of Congress.

British Library Cataloguing in Publication Data

A catalog record for this book is available from the British Library.

ISBN-13: 978-1-60807-525-6

Cover design by Adam Renvoize

© 2013 Artech House

Additional material may be found at http://www.sdr.wpi.edu/. All rights

reserved. Printed and bound in the United States of America. No part of this
book may be reproduced or utilized in any form or by any means, electronic
or mechanical, including photocopying, recording, or by any information
storage and retrieval system, without permission in writing from the publisher.
All terms mentioned in this book that are known to be trademarks or service
marks have been appropriately capitalized. Artech House cannot attest to the
accuracy of this information. Use of a term in this book should not be regarded
as affecting the validity of any trademark or service mark.

10 9 8 7 6 5 4 3 2 1
 To my parents, Lingzhen and Xuexun
DP

To my wife, Jen, my parents, Ted and Barbara,

and my sisters Laura and Joanne
AMW
Contents

Preface xiii

 Chapter 1 
What Is an SDR? 1
1.1  Historical Perspective 1
1.2 Microelectronics Evolution and its Impact on Communications
Technology 2
1.2.1  SDR Definition 3
1.3  Anatomy of an SDR 5
1.3.1  Design Considerations 6
1.4  Build it and they Will Come 7
1.4.1  Hardware Platforms 8
1.4.2  SDR Software Architecture 10
1.5  Chapter Summary 13
1.6  Additional Readings 13
References 13

 Chapter 2 
Signals and Systems Overview 15
2.1  Signals and Systems 15
2.1.1  Introduction to Signals 15
2.1.2  Introduction to Systems 16
2.2  Fourier Transform 18
2.2.1  Introduction and Historical Perspective 19
2.2.2  Definition 20
2.2.3  Properties 20
2.3  Sampling Theory 23
2.3.1  Uniform Sampling 24
2.3.2  Frequency Domain Representation of Uniform Sampling 24
2.3.3  Nyquist Sampling Theorem 26
2.3.4  Sampling Rate Conversion 27
2.4  Pulse Shaping 30
2.4.1  Eye Diagrams 32
2.4.2  Nyquist Pulse Shaping Theory 32
2.4.3  Two Nyquist Pulses 34
2.5  Filtering 39
2.5.1  Ideal Filter 39

vii
viii Contents

2.5.2  Z-Transform 39
2.5.3  Digital Filtering 42
2.6  Chapter Summary 47
2.7  Problems 47
References 50

 Chapter 3 
Probability Review 53
3.1  Fundamental Concepts 53
3.1.1  Set Theory 53
3.1.2  Partitions 54
3.1.3  Functions 56
3.1.4  Axioms and Properties of Probability 57
3.1.5  Conditional Probability 57
3.1.6  Law of Total Probability and Bayes’ Rule 58
3.1.7  Independence 59
3.2  Random Variables 59
3.2.1  Discrete Random Variables 60
3.2.2  Continuous Random Variables 65
3.2.3  Cumulative Distribution Functions 69
3.2.4  Central Limit Theorem 70
3.2.5  The Bivariate Normal 71
3.3  Random Processes 72
3.3.1  Statistical Characteristics of Random Processes 74
3.3.2  Stationarity 76
3.3.3  Gaussian Processes 77
3.3.4  Power Spectral Density and LTI Systems 78
3.4  Chapter Summary 79
3.5  Additional Readings 80
3.6  Problems 80
References 88

 Chapter 4 
Digital Transmission Fundamentals 89
4.1  What is Digital Transmission? 89
4.1.1  Source Encoding 91
4.1.2  Channel Encoding 92
4.2  Digital Modulation 94
4.2.1  Power Efficiency 94
4.2.2  Pulse Amplitude Modulation 95
4.2.3  Quadrature Amplitude Modulation 98
4.2.4  Phase Shift Keying 99
4.2.5  Power Efficiency Summary 104
4.3  Probability of Bit Error 105
4.3.1  Error Bounding 107
Contents ix

4.4  Signal Space Concept 108

4.5  Gram-Schmidt Orthogonalization 110
4.6  Optimal Detection 113
4.6.1  Signal Vector Framework 114
4.6.2  Decision Rules 116
4.6.3  Maximum Likelihood Detection in an AWGN Channel 117
4.7  Basic Receiver Realizations 119
4.7.1  Matched Filter Realization 120
4.7.2  Correlator Realization 122
4.8  Chapter Summary 124
4.9  Additional Readings 125
4.10  Problems 125
References 130

 Chapter 5 
Basic SDR Implementation of a Transmitter and a Receiver 131
5.1  Software Implementation 131
5.1.1  Repetition Coding 132
5.1.2  Interleaving 134
5.1.3  BER Calculator 135
5.1.4  Receiver Implementation over an Ideal Channel 136
5.2  USRP Hardware Implementation 137
5.2.1  Frequency Offset Compensation 138
5.2.2 Finding Wireless Signals: Observing IEEE 802.11 WiFi
Networks 140
5.2.3  USRP In-phase/Quadrature Representation 141
5.3  Open-Ended Design Project: Automatic Frequency Offset Compensator 145
5.3.1  Introduction 145
5.3.2  Objective 146
5.3.3  Theoretical Background 147
5.4  Chapter Summary 149
5.5  Problems 149
References 152

 Chapter 6 
Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver 153
6.1  Software Implementation 153
6.1.1  Observation Vector Construction 153
6.1.2  Maximum-Likelihood Decoder Implementation 157
6.1.3  Correlator Realization of a Receiver in Simulink 159
6.2  USRP Hardware Implementation 162
6.2.1  Differential Binary Phase-Shift Keying 163
6.2.2  Differential Quadrature Phase-Shift Keying 166
6.2.3  Accelerate the Simulink Model that Uses USRP Blocks 166
6.3  Open-Ended Design Project: Frame Synchronization 167
6.3.1  Frame Synchronization 167
x Contents

6.3.2  Barker Code 168

6.3.3  Simulink Models 168
6.3.4  Hints for Implementation 172
6.3.5  Hints for Debugging 172
6.4  Chapter Summary 172
6.5  Problems 173
Reference 175

 Chapter 7 
Multicarrier Modulation and Duplex Communications 177
7.1  Theoretical Preparation 177
7.1.1  Single Carrier Transmission 177
7.1.2  Multicarrier Transmission 181
7.1.3  Dispersive Channel Environment 183
7.1.4  OFDM with Cyclic Prefix 185
7.1.5  Frequency Domain Equalization 186
7.1.6  Bit and Power Allocation 187
7.2  Software Implementation 189
7.2.1  MATLAB Design of Multicarrier Transmission 189
7.2.2  Simulink Design of OFDM 192
7.3  USRP Hardware Implementation 194
7.3.1  Eye Diagram 194
7.3.2  Matched Filter Observation 195
7.4  Open-Ended Design Project: Duplex Communication 197
7.4.1  Duplex Communication 197
7.4.2  Half-Duplex 198
7.4.3  Time-Division Duplexing 198
7.4.4  Useful Suggestions 199
7.4.5  Evaluation and Expected Outcomes 200
7.5  Chapter Summary 201
7.6  Problems 201
References 204

 Chapter 8 
Spectrum Sensing Techniques 207
8.1  Theoretical Preparation 207
8.1.1  Power Spectral Density 207
8.1.2  Practical Issues of Collecting Spectral Data 209
8.1.3  Hypothesis Testing 214
8.1.4  Spectral Detectors and Classifiers 218
8.2  Software Implementation 222
8.2.1  Constructing Energy Detector 222
8.2.2  Observing Cyclostationary Detector 226
8.3  USRP Hardware Experimentation 227
8.4  Open-Ended Design Project: CSMA/CA 230
8.4.1  Carrier Sense Multiple Access 230
Contents xi

8.4.2  Collision Avoidance 231

8.4.3  Implementation Approach 231
8.4.4  Useful Suggestions 232
8.4.5  Evaluation and Expected Outcomes 233
8.5  Chapter Summary 234
8.6  Problems 234
References 236

 Chapter 9 
Applications of Software-Defined Radio 239
9.1  Cognitive Radio and Intelligent Wireless Adaptation 239
9.1.1  Wireless Device Parameters 241
9.2  Vehicular Communication Networks 242
9.2.1  VDSA Overview 243
9.2.2  Transmitter Design 244
9.2.3  Receiver Design 245
9.2.4  VDSA Test-Bed Implementation 245
9.3  Satellite Communications 246
9.4  Chapter Summary 250
References 250

 Appendix A 
Getting Started with MATLAB and Simulink 253
A.1  MATLAB Introduction 253
A.2  Edit and Run a Program in Matlab 253
A.3  Useful MATLAB Tools 254
A.3.1  Code Analysis and M-Lint Messages 254
A.3.2  Debugger 255
A.3.3  Profiler 256
A.4  Simulink Introduction 257
A.5  Getting Started in Simulink 257
A.5.1  Start a Simulink Session 257
A.5.2  Start a Simulink Model 257
A.5.3  Simulink Model Settings 258
A.6  Build a Simulink Model 259
A.6.1  Obtain the Blocks 260
A.6.2  Set the Parameters 260
A.6.3  Connect the Blocks 262
A.7  Run Simulink Simulations 264
References 266

 Appendix B 
Universal Hardware Driver (UHD) 267
B.1  Setting Up Your Hardware 267
B.2  Installing UHD-Based USRP I/O Blocks 267
B.3  Burning the Firmware to an SD Card 268
xii Contents

B.4  Configure the Ethernet Card 268

B.5  Modify the Iptables 269
B.6  Each Time You Use 269
B.7  Problems with Unicode 269
References 270

 Appendix C 
Data Flow on USRP 271
C.1  Receive Path 271
C.1.1  Situation 1 272
C.1.2  Situation 2 274
C.2  Transmit Path 274
References 276

 Appendix D 
Quick Reference Sheet 277
D.1  LINUX 277
D.1.1  Helpful Commands 277
D.1.2  Modify the Iptables 277
D.2  MATLAB 278
D.2.1  How to Start MATLAB 278
D.2.2  The MATLAB Environment 278
D.2.3  Obtaining Help 278
D.2.4  Variables in MATLAB 278
D.2.5  Vectors and Matrices in MATLAB 279
D.3  USRP2 Hardware 279
D.3.1  XCVR2450 Daughtercard 279
D.3.2  Sampling 280
D.3.3  Clocking 281
D.3.4  DDC and DUC 281
D.4  Differential Phase-Shift Keying (DPSK) 282
Reference 282

 Appendix E 
Trigonometric Identities 283
About the Author 285
Index 287
Preface

The communications sector has witnessed a significant paradigm shift over the
past several decades with respect to how wireless transceivers and algorithms are
implemented for a wide range of applications. Until the 1980s, almost all wireless
communication systems were based on relatively static, hard-wired platforms using
technologies such as application-specific integrated circuits (ASICs). Nonetheless,
with numerous advances being made in the areas of digital processing technol-
ogy, digital-to-analog converters (DACs), analog-to-digital converters (ADCs), and
computer architectures, communication engineers started rethinking how wireless
transceivers could be implemented entirely in the digital domain, resulting in both
digital communications and digital signal processing algorithms being constructed
completely in programmable digital logic and software modules. Subsequently, this
rethinking by the communications sector has given rise to software-defined radio
(SDR) technology, where all the baseband processing of the wireless transceiver
is entirely performed in the digital domain (e.g., programmable logic, software
source code). Nevertheless, despite the fact that SDR technology has been around
for nearly three decades, undergraduate and graduate courses focusing on digital
communications have been mostly taught using educational methodologies and
techniques that were either based on transmission technologies from half a century
ago or entirely lacked any physical hardware context whatsoever. Consequently,
there exists a significant need by the wireless community for an educational ap-
proach that leverages SDR technology in order to provide individuals with the latest
insights on digital communication systems; this book was primarily written in order
to satisfy this need.
The objective of this book is to provide the reader with a “hands-on” educa-
tional approach employing software-defined radio (SDR) experimentation to fa-
cilitate the learning of digital communication principles and the concepts behind
wireless data transmission, and provide an introduction to wireless access tech-
niques. Based on four years of educational experiences by the authors with respect
to the instruction of a digital communications course for senior undergraduate and
junior graduate students at Worcester Polytechnic Institute (WPI) based on SDR
experimentation, it is expected that enabling individuals to prototype and evaluate
actual digital communication systems capable of performing “over-the-air” wireless
data transmission and reception will provide them with a first-hand understanding
of the design tradeoffs and issues associated with these systems, as well as provide
them with a sense of actual “real world” operational behavior. Given that the learn-
ing process associated with using an actual SDR platform for prototyping digital
communication systems is often lengthy, tedious, and complicated, it is difficult to
employ this tool within a relatively short period of time, such as a typical under-
graduate course (e.g., 7–14 weeks). Consequently, this book contains a collection
of preassembled Simulink experiments along with several software examples to
xiii
xiv Preface

enable the reader to successfully implement these designs over a short period of
time, as well as simultaneously synthesize several key concepts in digital commu-
nication system theory. Furthermore, in order to provide the reader with a solid
grasp of the theoretical fundamentals needed to understand and implement digital
communication systems using SDR technology, three overview chapters near the
beginning of this book provides a quick summary of essential concepts in signals
and systems, probability theory, and basic digital transmission.
The intended readers for this book are senior undergraduate and entry-level
graduate students enrolled in courses in digital communication systems, industry prac­
titioners in the telecommunications sector attempting to master software-defined
radio technology, and wireless communication researchers (government, industry,
academia) seeking to verify designs and ideas using SDR technology. It is assumed
that the reader possesses an electrical and computer engineering background with a
concentration in signals, systems, and information. This book, along with its corre-
sponding collection of software examples and 28 sets of lecture slides (which can be
viewed at http://www.artechhouse.com and at http://www.sdr.wpi.edu/), is designed
to introduce the interested reader to concepts in digital communications, wireless
data transmission, and wireless access via a structured, “hands-on” approach.
In order to efficiently utilize this book for learning digital communications with
SDR technology, it is recommended that the interested individual reads this book
in a sequential manner. Starting with an overview of SDR technology in Chapter 1,
the reader will then be provide a quick refresher of signals and systems, probability
theory, and basic digital communications in Chapter 2, Chapter 3, and Chapter 4.
Following this review of the fundamentals, the next four chapters, namely, Chapters
5 through 8, focus on performing basic wireless data transmission, designing differ-
ent types of wireless receivers, introducing the concept of multicarrier modulation,
and devising spectrum sensing techniques. Finally, Chapter 9 provides an overview
of the various advanced applications where SDR technology has made a signifi-
cant impact. A collection of appendixes at the end of this book provides the reader
with instant access to information in order to help them quickly overcome the rela-
tively steep learning curve associated with SDR implementations. In addition to the
hands-on experiments, end-of-chapter problems also provide the readers with an
opportunity to strengthen their theoretical understanding of the concepts covered in
this book. Note that the 28 sets of lecture slides are closely related to the sequence
of topics presented in this book, where each set of slides is intended for a lecture of
approximately 60–90 minutes in duration. As a result, this book and its associated
materials is well-suited for a senior-level undergraduate course on digital communi-
cations taught over a 7-week academic term or a 14-week academic semester.
This book was made possible by the extensive support of numerous individu-
als and organizations throughout the duration of this project. First, we are deeply
indebted to our collaborators and sponsors at the MathWorks in Natick, MA, for
their constant support. In particular, we would like to sincerely thank Don Orofino,
Kate Fiore, Mike McLernon, and Ethem Sozer for their endless guidance and assis-
tance throughout the entire project. Second, we would like to sincerely thank Dan-
iel Cullen for his assistance in getting this educational approach “off the ground”
in 2010 with the first offering of the SDR-based digital communications course on
which this book is based. Third, we would like to express our deepest gratitude
Preface xv

to the WPI ECE department, especially its department head, Fred Looft, for being
a unwavering supporter in making the course that this book is based on become a
reality and a permanent addition to the WPI ECE curriculum. Finally, we would like
to thank our families for their support and encouragement.

Di Pu
Worcester Polytechnic Institute, USA

Alexander M. Wyglinski
Worcester Polytechnic Institute, USA
 Chapter 1 1
2
What Is an SDR? 3
4
5
6
Modern society as we know it today is becoming increasingly dependent on the reliable 7
and seamless wireless exchange of information. For instance, a rapidly growing num- 8
ber of individuals are using smartphones in order to access websites, send and receive 9
emails, view streaming multimedia content, and engage in social networking activities 10
anywhere on the planet at any time. Furthermore, numerous drivers on the road today 11
are extensively relying upon global positioning system (GPS) devices and other wire- 12
less navigation systems in order to travel in a new neighborhood or city without the 13
need for a conventional map. Finally, in many hospitals and medical centers across the 14
nation and around the world, the health of a patient is now being continuously moni- 15
tored using an array of medical sensors attached to him/her that send vital information 16
wirelessly to a computer workstation for analysis by a medical expert. 17
The enabling technology for supporting any form of wireless information ex- 18
change is the digital transceiver, which is found in every cell phone, WiFi-enabled 19
laptop, BlueTooth device, and other wireless appliances designed to transmit and 20
receive digital data. Digital transceivers are capable of performing a variety of base- 21
band operations, such as modulation, source coding, forward error correction, 22
and equalization. Although digital transceivers were initially implemented using 23
integrated circuits and other forms of nonprogrammable electronics, the advent 24
of programmable baseband functionality for these digital transceivers is the result 25
of numerous advancements in microprocessor technology over the past several de- 26
cades. Consequently, there exists a plethora of digital transceiver solutions with a 27
wide range of capabilities and features. In this chapter, we will provide an introduc- 28
tion to software-defined radio (SDR) hardware platforms and software architec- 29
tures, as well as study how it has evolved over the past several decades to become a 30
transformative communications technology. 31
32
33
1.1  HISTORICAL PERSPECTIVE 34
35
The term “software-defined radio” was first coined by Joseph Mitola [1], although 36
SDR technology was available since the 1970s and the first demonstrated SDR proto­ 37
type was presented in 1988 [2]. However, the key milestone for the advancement 38
of SDR technology took place in the early 1990s with the first publicly funded SDR 39
development initiative called SpeakEasy I/II by the U.S. military [3]. The SpeakEasy 40
project was a significant leap forward in SDR technology since it used a collection 41
of programmable microprocessors for implementing more than 10 military com- 42
munication standards, with transmission carrier frequencies ranging from 2 MHz 43
to 2 GHz, which at that time was a major advancement in communication systems 44
engineering. Additionally, the SpeakEasy implementation allowed for software 45
46
1 47

ART_Wyglinski_Ch01.indd 1 Manila Typesetting Company 12/06/2012

 2 What Is an SDR?

1 upgrades of new functional blocks, such as modulation schemes and coding schemes.
2 The first generation of the SpeakEasy system initially used a Texas Instruments
3 TMS320C40 processor (40 MHz). However, the SpeakEasy II platform was the first
4 SDR platform to involve field programmable gate array (FPGA) modules for imple-
5 menting digital baseband functionality. Note that given the microprocessor technol-
6 ogy at the time, the physical size of the SpeakEasy prototypes were large enough to
7 fit in the back of a truck.
8
9
10 Q   Why were the first SDR platforms physically very large in size?
11
12
13 Following this initial SDR project, research and development activities in this
14 area continued to advance the current state-of-the-art in SDR technology, with one
15 of the outcomes being the Joint Tactical Radio System (JTRS), which is a next-
16 generation voice-and-data radio used by the U.S. military and employs the software
17 communications architecture (SCA) [4]. Initially developed to support avionics [5]
18 and dynamic spectrum access [6] applications, JTRS is scheduled to become the
19 standard communications platform for the U.S. Army by 2012.
20
21
22 1.2 MICROELECTRONICS EVOLUTION AND ITS IMPACT ON
23 COMMUNICATIONS TECHNOLOGY
24
25 The microelectronic industry has rapidly evolved over the past six decades, result­ing
26 in numerous advances in microprocessor systems that have enabled many of the ap-
27 plications we take for granted every day. The rate at which this evolution has prog­
28 ressed over time has been characterized by the well-known Moore’s Law, which
29 defines the long-term trend of the number of transistors that can be accommodated
30 on an integrated circuit. In particular, Moore’s Law dictates that the number of
31 transistors per integrated circuit approximately doubles every two years, which
32 subsequently affects the performance of microprocessor systems such as processing
33 speed and memory. One area that the microelectronics industry has significantly
34 influenced over the past half century is the digital communication systems sector,
35 where microprocessor systems have been increasingly employed in the implementa-
36 tion of digital transceivers, yielding more versatile, powerful, and portable com-
37 munication system platforms capable of performing a growing number of advance
38 operations and functions. With the latest advances in microelectronics and micro-
39 processor systems, this has given rise to software-defined radio (SDR) technology,
40 where baseband radio functionality can be entirely implemented in digital logic and
41 software, as illustrated in Figure 1.2.
42 There exists several different types of microprocessor systems for SDR imple-
43 mentations. For instance, one popular choice is the general purpose microprocessor,
44 which is often used in SDR implementations and prototypes due to its high level of
45 flexibility with respect to reconfigurability, as well as due to its ease of implementation
46 regarding new designs. On the other hand, general purpose microprocessors are not
47 specialized for mathematical computations, and they can be potentially power inef-

ART_Wyglinski_Ch01.indd 2 Manila Typesetting Company 12/06/2012

1.2  Microelectronics Evolution and its Impact on Communications Technology 3

ficient. Another type of microprocessor system, called a digital signal processor (DSP), 1
is specialized for performing mathematical computations, implementation of new dig- 2
ital communication modules can be performed with relative ease, and the processor | 3
is relatively power efficient (e.g., DSPs are used in cellular telephones). On the other 4
hand, DSPs are not well suited for computationally intensive processes and can be 5
rather slow. Alternatively, field programmable gate arrays (FPGAs) are computation- 6
ally powerful, but power inefficient, and it is neither flexible nor easy to implement 7
new modules. Similarly, graphics processing units (GPUs) are extremely powerful 8
computationally but difficult to use and it is implement new modules as well. 9
10
11
1.2.1  SDR Definition
12
Given these microprocessor hardware options, let us now proceed with formu- 13
lating a definition for an SDR platform. An SDR is a class of reconfigurable/ 14
reprogrammable radios whose physical layer characteristics can be significantly modi- 15
fied via software changes. It is capable of implementing different functions at different 16
times on the same platform, it defines in software various baseband radio features, 17
(e.g., modulation, error correction coding), and it possesses some level of software 18
control over RF front-end operations, (e.g., transmission carrier frequency). Since all 19
of the baseband radio functionality is implemented using software, this implies that 20
potential design options and radio modules available to the SDR platform can be 21
readily stored in memory and called upon when needed for a specific application, such 22
as a specific modulation, error correction coding, or other functional block needed 23
to ensure reliable communications. Note that due to the programmable nature of the 24
SDR platform and its associated baseband radio modules, these functional blocks can 25
potentially be changed in real-time and the operating parameters of functional blocks 26
can be adjusted either by a human operator or an automated process. 27
Although definitions may vary to a certain degree regarding what constitutes 28
an SDR platform, several key characteristics that generally define an SDR can be 29
summarized by the following list [7]: 30
31
• Multifunctionality: Possessing the ability to support multiple types of radio 32
functions using the same digital communication system platform. 33
• Global mobility: Transparent operation with different communication net- 34
works located in different parts of the world (i.e., not confined to just one 35
standard). 36
• Compactness and power efficiency: Many communication standards can be 37
supported with just one SDR platform. 38
• Ease of manufacturing: Baseband functions are a software problem, not a 39
hardware problem. 40
• Ease of upgrading: Firmware updates can be performed on the SDR platform 41
to enable functionality with the latest communication standards. 42
43
One of the most commonly leveraged SDR characteristic is that of multifunctionality, 44
where the SDR platform employs a particular set of baseband radio modules based on 45
how well the communication system will perform as a result of that configuration. To 46
illustrate how multifunctionality might work on an SDR platform, suppose we employ 47

ART_Wyglinski_Ch01.indd 3 Manila Typesetting Company 12/06/2012

 4 What Is an SDR?

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19 (a)
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35 (b)
36 Figure 1.1  An illustration of multifunctionality in an FPGA-based SDR platform. (a) Swapping dif­
37 fe­rent modulation schemes on the same FPGA. (b) Digital communication transmitter chain on an
FPGA using the multiprocessor system-on-chip (MPSoC) concept.
38
39
40 an FPGA-based implementation consisting of a baseband radio architecture shown in
41 Figure 1.1(a). Now, let us assume that the transmission environment has changed and
42 the SDR platform has determined that a change in the modulation scheme is needed
43 to continue guaranteeing sufficient transmission performance. Consequently, using the
44 multiprocessor system-on-chip (MPSoC) concept, the SDR platform quickly swaps out
45 the BPSK modulation software module and replaces it with a QPSK modulation soft-
46 ware module, allowing for the continued operation of the SDR platform with minimal
47 interruption. The way that the MPSoC concept works is shown in Figure 1.1(b), where

ART_Wyglinski_Ch01.indd 4 Manila Typesetting Company 12/06/2012

1.3  Anatomy of an SDR 5

a microblaze on the FPGA acts as a mini-microprocessor that makes decisions on what 1

parts of the FPGA to activate, thus enabling these specific baseband radio modules. 2
3
4
1.3  ANATOMY OF AN SDR 5
6
An SDR system is a complex device that performs several complicated tasks simul- 7
taneously in order to enable the seamless transmission and reception of data. In 8
general, digital communications systems consists of an interdependent sequence of 9
operations responsible for taking some type of information, whether it is human 10
speech, music, or video images, and transmit it over the air to an awaiting receiver 11
for processing and decoding into a reconstructed version of the original informa- 12
tion signal. If the original information is analog, it must first be digitized using 13
techniques such as quantization in order for us to obtain a binary representation of 14
this information. Once in a binary format, the transmitter digitally processes this 15
information and converts it into an electromagnetic sinusoidal waveform that is 16
uniquely defined by its physical characteristics, such as its signal amplitude, carrier 17
frequency, and phase. At the other end of the communications link, the receiver 18
is tasked with correctly identifying the physical characteristics of the intercepted 19
electromagnetic waveform transmitted across a potentially noisy and distortion- 20
filled channel, and ultimately returning the intercepted signal back into the correct 21
binary representation. The basic building block of a digital communication system 22
is shown in Figure 1.2. 23
We see in Figure 1.2 that the input to the transmitter and output of the receiver 24
originate from or are fed into a digital source and digital sink, respectively. These 25
two blocks represent the source and destination of the digital information to be 26
communicated between the transmitter and receiver. Once the binary information 27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
Figure 1.2  An illustration describing some of the important components that constitute a modern
digital communications system. Note that for a SDR-based implementation, those components indi- 46
cated as “programmable” can be realized in either programmable logic or software. 47

ART_Wyglinski_Ch01.indd 5 Manila Typesetting Company 12/06/2012

 6 What Is an SDR?

1 is introduced to the transmitter, the first task performed is to remove all redundant/
2 repeating binary patterns from the information in order to increase the efficiency
3 of the transmission. This is accomplished using the source encoder block, which is
4 designed to strip out all redundancy from the information. Note that at the receiver,
5 the source decoder reintroduces the redundancy in order to return the binary infor-
6 mation back to its original form. Once the redundancy has been removed from the
7 binary information at the transmitter, a channel encoder is employed to introduce
8 a controlled amount of redundancy to the information stream in order to protect
9 it from potential errors introduced during the transmission process across a noisy
10 channel. A channel decoder is used to remove this controlled redundancy and re-
11 turn the binary information back to its original form. The next step at the trans-
12 mitter is to convert the binary information into unique electromagnetic waveform
13 properties such as amplitude, carrier frequency, and phase. This is accomplished
14 using a mapping process called modulation. Similarly, at the receiver the demodula-
15 tion process converts the electromagnetic waveform back into its respective binary
16 representation. Finally, the discrete samples outputted by the modulation block
17 are resampled and converted into a baseband analog waveform using a digital-to-
18 analog (D/A) converter before being processed by the radio frequency (RF) front-
19 end of the communication system and upconverted to an RF carrier frequency.
20 At the receiver, the reverse operation is performed, where the intercepted analog
21 signal is downconverted by the RF front-end to a baseband frequency before being
22 sampled and processed by an analog-to-digital (A/D) converter.
23
24
1.3.1  Design Considerations
25
26 Given the complexity of an SDR platform and its respective components, as de-
27 scribed in the previous section as well as in Figure 1.2, it is important to understand
28 the limitations of a specific SDR platform and how various design decisions may
29 impact the performance of the resulting prototype. For instance, it is very desirable
30 to have real-time baseband processing for spectrum sensing and agile transmission
31 operations with high computational throughput and low latency. However, if the
32 microprocessor being employed by the SDR platform is not sufficiently powerful
33 enough in order to support the computational operations of the digital communi-
34 cation system, one needs to reconsider either the overall transceiver design or the
35 requirements for low latency and high throughput. Otherwise, the SDR implemen-
36 tation will fail to operate properly, yielding transmission errors and poor com-
37 munication performance. An example of when the microprocessor is not capable
38 of handling the computational needs of the SDR platform is shown in Figure 1.3,
39 where the transmission being produced by the SDR platform is occurring in bursts,
40 yielding several periods of transmitted signal interspersed with periods of no trans-
41 missions. Note that such a bursty transmission would be extremely difficult to han-
42 dle at the receiver due to the intermittent nature of the signal being intercepted.
43 Other design considerations to think about when devising digital communica-
44 tion systems based on an SDR platform include the following:
45
46 • The integration of the physical and network layers via a real-time protocol
47 implementation on an embedded processor. Note that most communication

ART_Wyglinski_Ch01.indd 6 Manila Typesetting Company 12/06/2012

1.4  Build it and they will Come 7

1
2
3
4
5
6
7
8
9
10
11
12
Figure 1.3  An example of sampling rate errors when the SDR platform is unable to keep up with
the transmission of the data. Note that the bursts can result in significant difficulty at the receiver 13
with respect to the interception and decoding of the received signal. 14
15
systems are divided into logically separated layers in order to more readily 16
facilitate the design of the communication system. However, it is imperative 17
that each layer is properly designed due to the strong interdependence be- 18
tween all the layers. 19
• Ensuring that a sufficiently wide bandwidth radio front-end exists with agil- 20
ity over multiple subchannels and scalable number of antennas for spatial 21
processing. Given how many of the advanced communication system designs 22
involve the use of multiple antennas and wideband transmissions, it is im- 23
portant to know what the SDR hardware is capable of doing with respect to 24
these physical attributes. 25
• Many networks employing digital communication systems possess a central- 26
ize architecture for controlling the operations of the overall network (e.g., 27
control channel implementation). Knowing your radio network architecture 28
is important since it will dictate what sort of operations are essential for one 29
digital transceiver to communicate with another. 30
• The ability to perform controlled experiments in different environments (e.g., 31
shadowing and multipath, indoor and outdoor environments) is important 32
for the sake of demonstrating the reliability of a particular SDR implementa- 33
tion. In other words, if an experiment involving an SDR prototype system is 34
conducted twice in a row in the exact same environment and using the exact 35
same operating parameters, it is expected that the resulting output and per- 36
formance should be the same. Consequently, being able to perform controlled 37
experiments provides the SDR designer with a “sanity check” capability. 38
• Reconfigurability and fast prototyping through a software design flow for 39
algorithm and protocol description. 40
41
42
1.4  BUILD IT AND THEY WILL COME 43
44
Given our brief overview of SDR technology and its definition, as well as a survey 45
of various microprocessor design options and constraints available when imple- 46
menting an SDR platform, we will now focus our attention on several well-known 47

ART_Wyglinski_Ch01.indd 7 Manila Typesetting Company 12/06/2012

 8 What Is an SDR?

1 SDR hardware implementations published in the open literature. Note that when
2 designing a complete SDR system from scratch, it is very important to have both
3 a hardware platform that is both sufficiently programmable and computationally
4 powerful, as well as a software architecture that can allow a communication system
5 designer to implement a wide range of different transceiver realizations. In this sec-
6 tion, we will first study some of the well-known SDR hardware platforms before
7 talking about some of the available SDR software architectures.
8
9
1.4.1  Hardware Platforms
10
11 Exploration into advanced wireless communication and networking techniques re-
12 quire highly flexible hardware platforms. As a result, SDR is very well suited due
13 to its rapidly reconfigurable attributes, which allows for controlled yet realistic
14 experimentation. Thus, the use of real-time test bed operations enables a large set
15 of experiments for various receiver settings, transmission scenarios, and network
16 configurations. Furthermore, SDR hardware provides an excellent alternative for
17 comprehensive evaluation of communication systems operating within a networked
18 environment, whereas Monte Carlo simulations can be computationally exhaustive
19 and are only as accurate as the devised computer model. In this section, we will
20 study several well-known SDR hardware platforms used by the wireless community
21 for research and experimentation.
22 One of the most well-known of all SDR hardware platforms is the Universal
23 Software Radio Peripheral (USRP) concept that was introduced by Matt Ettus,
24 founder and president of Ettus Research LLC, which is considered to be a relatively
25 inexpensive hardware for enabling SDR design and development [8]. All the base-
26 band digital communication algorithms and digital signal processing are conducted
27 on a computer workstation “host,” where the USRP platform acts as a radio pe-
28 ripheral allowing for over-the-air transmissions and the libusrp library file de-
29 fines the interface between the USRP platform and the host computer workstation.
30 Note that the USRP design is open source, which allows for user customization and
31 fabrication. Furthermore, USRP platform design is modular in terms of the sup-
32 ported RF front-ends, referred to as daughtercards. We will now talk about two
33 types of USRP platforms: the USRP1 and USRP2.
34 The Universal Software Radio Peripheral-Version 1 (USRP1) was designed and
35 manufactured by Ettus Research LLC for a variety of different communities interested
36 in an inexpensive SDR platform. The USRP1 consists of a USB interface between host
37 computer workstation and USRP1 platform, which resulted in a data bottleneck due
38 to the low data rates supported by the USB connection. The USRP1 supports up
39 to two RF transceiver daughtercards, possesses an Altera Cyclone EP1C12Q240C8
40 FPGA for performing sampling and filtering, contains four high-speed analog-to-
41 digital converters, each capable of 64 MS/s at a resolution of 12 bits, with an 85 dB
42 SFDR (AD9862), and contains four high-speed digital-to-analog converters, each ca-
43 pable of 128 MS/s at a resolution of 14 bits, with 83 dB SFDR (AD9862).
44 Following the success of the USRP1, Ettus Research LLC officially released the
45 Universal Software Radio Peripheral-Version 2 (USRP2) platform in September
46 2008, as shown in Figure 1.4(a). The USRP2 platform provides a more capable
47

ART_Wyglinski_Ch01.indd 8 Manila Typesetting Company 12/06/2012

1.4  Build it and they will Come 9

1
2
3
4
5
6
7
8
9
10
11
12
13
14
(a) (b)
15
Figure 1.4  Examples of software-defined radio platforms. (a) Front view of a Universal Software
Radio Peripheral–Version 2 (USRP2) software-defined radio platform by Ettus Research LLC. (b) Front
16
view of a Kansas University Agile Radio (KUAR) software-defined radio platform. 17
18
19
SDR device for enabling digital communication system design and implementation. 20
The USRP2 features include a gigabit Ethernet interface between host computer 21
workstation and USRP2 platform, supports only one RF transceiver daughtercard, 22
possesses a Xilinx Spartan 3-2000 FPGA for performing sampling and filtering, 23
contains two 100 MS/s, 14 bits, analog-to-digital converters (LTC2284), with a 24
72.4 dB SNR and 85 dB SFDR for signals at the Nyquist frequency, contains two 25
400 MS/s, 16 bits, digital-to-analog converters (AD9777), with a 160 MS/s without 26
interpolation, and up to 400 MS/s with 8x interpolation, and is MIMO-capable 27
for supporting the processing of digital communication system designs employing 28
multiple antennas. 29
The radio frequency (RF) front-ends are usually very difficult to design and 30
are often limited to a narrow range of transmission carrier frequencies. This is due 31
to the fact that the properties of the RF circuit and its components change across 32
different frequencies and that the RF filters are constrained in the sweep frequency 33
range. Consequently, in order to support a wide range of transmission carrier fre- 34
quencies, both the USRP1 and USRP2 platforms can use an assortment of modular 35
RF daughtercards, such as the following: 36
37
• BasicTX: A transmitter that supports carrier frequencies within 1–250 MHz; 38
• BasicRX: A receiver that supports carrier frequencies within 1–250 MHz; 39
• RFX900: A transceiver that supports carrier frequencies within 800–1000 40
MHz with a 200+mW output; 41
• RFX2400: A transceiver that supports carrier frequencies within 2.3–2.9 GHz 42
with a 20+mW output; 43
• XCVR2450: A transceiver that supports carrier frequencies within two 44
bands, namely, 2.4–2.5 GHz with an output of 100+mW and 4.9–5.85 GHz 45
with an output of 50+mW. 46
47

ART_Wyglinski_Ch01.indd 9 Manila Typesetting Company 12/06/2012

 10 What Is an SDR?

1 Another well-known SDR hardware platform was the Kansas University Agile
2 Radio (KUAR), which is a small form factor SDR platform containing a Xilinx
3 Virtex-II Pro FPGA board and a PCI Express 1.4 GHz Pentium-M microprocessor,
4 as shown in Figure 1.4(b) [9, 10]. For its size and capability, the KUAR was one
5 of the leading SDR implementations in its day, incorporating substantial computa-
6 tional resources as well as wideband frequency operations. Around the same time
7 period, the Berkeley BEE2 was designed as a powerful reconfigurable computing en-
8 gine with five Xilinx Virtex-II Pro FPGAs on a custom-built emulation board [11].
9 The Berkeley Wireless Research Center (BWRC) cognitive radio test-bed hardware
10 architecture consists of the BEE2, several reconfigurable 2.4-GHz radio modems,
11 and fiber link interfaces for connections between the BEE2 and the radios modems.
12 The software architecture consists of Simulink-based design flow and BEE2 specific
13 operating system, which provides an integrated environment for implementation
14 and simple data acquisition during experiments.
15 With respect to compact SDR platforms, Motorola developed and built a 10-MHz,
16 4-GHz CMOS-based small form factor cognitive radio platform prototype [12].
17 Fundamentally flexible, with a low-power transceiver radio frequency integrated cir-
18 cuit (RFIC) at the core of this experimental platform, this prototype can receive and
19 transmit signals of many wireless protocols, both standard and experimental. Car-
20 rier frequencies from 10 MHz to 4 GHz with channel bandwidths from 8 kHz to
21 20 MHz were supported. Similarly, the Maynooth Adaptable Radio System (MARS) is
22 a custom-built small form factor SDR platform [13]. The MARS platform had the
23 original objectives of being a personal computer connected radio front-end where all
24 the signal processing is implemented on the computer’s general-purpose processor.
25 The MARS platform was designed to deliver performance equivalent to that of a
26 future base station and the wireless communication standards in the 1700-MHz to
27 2450-MHz frequency range. Furthermore, the communication standards GSM1800,
28 PCS1900, IEEE 802.11b/g, and UMTS (TDD and FDD) are also supported.
29 Rice University Wireless Open Access Research Platform (WARP) radios in-
30 clude a Xilinx Virtex-II Pro FPGA board as well as a MAX2829 transceiver [14],
31 while the Lyrtech Small Form Factor SDR is developed by a company from the Ca-
32 nadian Province of Québec that leverages industrial collaborations between Texas
33 Instruments and Xilinx in order to produce these high-performance SDR platforms
34 that consist of an array of different microprocessor technology [15]. Finally, Epiq
35 Solutions recently released the MatchStiq SDR platform, which is a powerful yet
36 very compact form factor SDR platform capable of being deployed in the field to
37 perform a variety of wireless experiments, including their inclusion onboard ve-
38 hicles such as automobiles and unmanned aerial vehicles [16].
39
40
1.4.2  SDR Software Architecture
41
42 Given the programmable attributes of an SDR platform, it is vitally important to
43 also develop an efficient and reliable software architecture that would operate on
44 these platforms in order to perform the various data transmission functions we
45 expect from a wireless communications system. In this section, we will review some
46 of the SDR software architectures currently available for use with a wide variety of
47 SDR hardware platforms.

ART_Wyglinski_Ch01.indd 10 Manila Typesetting Company 12/06/2012

1.4  Build it and they will Come 11

One of the first Simulink interfaces to the USRP2 platform was implemented as 1
part of an MS thesis at WPI and generously sponsored by the MathWorks [17]. In 2
this research project, the focus was on creating a Simulink blockset capable of com- 3
municating with the USRP2 libraries, which can then allow for communications with 4
the USRP2 platform itself. The resulting blocksets from this thesis research are shown 5
in Figure 1.5. By creating a Simulink interface to this SDR hardware, it is expected 6
that the existing signal processing libraries provided by the MathWorks can be exten- 7
sively leveraged in order to create actual digital communications systems capable of 8
performing over-the-air data transmission with other USRP2 platforms. 9
The architecture of the Simulink transmit and receiver blocks are shown in 10
Figure 1.6. In Figure 1.6(b), we observe how the Simulink transmitter block calls 11
the functions implemented by the S-function at different parts of the simulation. 12
While the simulation is initializing, it calls on mdlStart such that a data handler 13
object and first-in first-out (FIFO) register are created and a USRP2 object is instan- 14
tiated. Once the USRP2 object has been created, the operating parameters set in the 15
mask are passed down to the USRP2 hardware while the model is in the process 16
of initializing. Furthermore, the data handler object loads data into the FIFO and, 17
while the simulation is running, Simulink is repeatedly calling mdlOutputs such 18
that a frame of data is read from the FIFO, converted to a Simulink data type, and 19
sent to the output port to be received in the simulation. Note that when the simula- 20
tion has finished, the FIFO and data handler are deallocated. The transmitter shown 21
in Figure 1.6(a) possesses a similar mode of operation. 22
From the MS thesis and the development of the first Simulink prototype block- 23
set interface with the USRP2 SDR platform, the MathWorks built upon the lessons 24
learned from this experience and ultimately created the SDRu blockset, which can 25
be downloaded from the MathWorks website and installed with MATLAB R2011a 26
or later along with the Communications Toolbox [18]. After several years of de- 27
velopment, the SDRu blocks are at the core of numerous SDR implementations 28
using Simulink and the USRP2, as well as educational activities such as those to be 29
discussed later in this book. 30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
Figure 1.5  The initial prototype Simulink transmitter and receiver interfaces for the USRP2 platform [17]. 47

ART_Wyglinski_Ch01.indd 11 Manila Typesetting Company 12/06/2012

 12 What Is an SDR?

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19 (a)
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39 (b)
40 Figure 1.6  Architecture of the initial prototype interfaces to the USRP2 platform [17]. (a) Initial
41 prototype Simulink transmitter interface. (b) Initial prototype Simulink receiver interface.
42
43 Other SDR software architectures include the popular open-source GNU Radio
44 software [19], which is a community based effort for devising an SDR software
45 architecture capable of interfacing with any SDR hardware platform, especially the
46 USRP family of products, and enabling them to reliably and seamlessly communicate
47 with other SDR platforms as well as conventional wireless systems. Given the large

ART_Wyglinski_Ch01.indd 12 Manila Typesetting Company 12/06/2012

1.6  Additonal Readings 13

open-source community supporting the GNU Radio software architecture, several 1

community members have posted their customized solutions that were not incorpo- 2
rated into GNU Radio directly on the Comprehensive GNU Radio Archive Network 3
(CGRAN) website for download by the rest of the community [20]. Finally, another 4
SDR software interface is the Implementing Radio in Software (IRIS) project [21], 5
which is led by the researchers at the Centre for Telecommunications Value-Chain 6
Research (CTVR) at Trinity College Dublin and used by many researchers across 7
Europe and around the world. 8
9
10
1.5  CHAPTER SUMMARY 11
12
In this chapter, we studied how the development of digital communication systems 13
and SDR technology is closely linked to the evolution of microprocessor technology. 14
Furthermore, we obtained some insight on the various microprocessor technologies 15
that are currently available to be implemented as part of an SDR hardware proto­ 16
type. Finally, we concluded this chapter with a literature survey of several SDR 17
hardware platforms and software architectures, and getting a better understanding 18
of the design decisions involved in implementing this systems. By providing some 19
insights on how this technology functions, it is expected that when we eventually 20
begin implementing SDR prototype systems we will know the limitations and capa- 21
bilities of our SDR platform. 22
23
24
1.6  ADDITIONAL READINGS 25
26
Although this chapter gave a brief introduction to the expanding area of SDR tech- 27
nology, several books available in the open literature can provide a more detailed 28
viewpoint of this topic. For instance, the book by Reed extensively covers many of 29
the issues associated with the software architecture of an SDR platform [7], while 30
many of the design considerations and approaches used to construct SDR hardware 31
prototype and their RF front-ends are covered in the book by Kensington [22]. For 32
a decent overview of some of the SDR prototype platforms covered in this chapter, 33
namely, the BEE2, the MARS project, and the Motorola SDR platform, check out 34
Chapter 19 in [23]. Finally, for a clean-slate viewpoint about SDR technology and 35
wireless systems in general, the interested reader is encourage to check out the con- 36
ference publication by Farrell et al. [24]. 37
38
39
References 40
41
[1] Mitola III, J. “Software Radios: Survey, Critical Evaluation and Future Directions,” IEEE
42
Aerospace and Electronic Systems Magazine, April 1993.
43
[2] Hoeher, P. and H. Lang, “Coded-8PSK Modem for Fixed and Mobile Satellite Services
based on DSP,” Proceedings of the First International Workshop on Digital Signal Process- 44
ing Techniques Applied to Space Communications, Noordwijk, The Netherlands, 1988. 45
[3] Lackey, R. J., and D. W. Upmal, “Speakeasy: The Military Software Radio,” IEEE Com- 46
munications Magazine, May 1995. 47

ART_Wyglinski_Ch01.indd 13 Manila Typesetting Company 12/06/2012

 14 What Is an SDR?

1 [4] White, B. E., “Tactical Data Links, Air Traffic Management, and Software Programmable Ra-
2 dios,” Proceedings of the 18th Digital Avionics Systems Conference, St. Louis, MO, 1999.
3 [5] Eyermann, P. A., “Joint Tactical Radio Systems—a Solution to Avionics Modernization,”
Proceedings of the 18th Digital Avionics Systems Conference, St. Louis, MO, 1999.
4
[6] Bergstrom, C. S. Chuprun and D. Torrieri, “Adaptive Spectrum Exploitation Using Emerg-
5
ing Software Defined Radios,” Proceedings of the IEEE Radio and Wireless Conference.
6 Denver, CO, 1999.
7 [7] Reed, J. H., Software Radio: A Modern Approach to Radio Engineering, Prentice Hall
8 PTR, 2002.
9 [8] Ettus Research LLC, “USRP Networked Series”. https://www.ettus.com/product/category/
10 USRP_ Networked_Series.
11 [9] Minden, G. J., J. B. Evans, L. Searl, D. DePardo, V. R. Petty et al., “KUAR: A Flexible
12 Software-Defined Radio Development Platform,” Proceedings of the IEEE International
Symposium on New Frontiers in Dynamic Spectrum Access Networks, Dublin, Ireland, 2007.
13
[10] Minden, G. J., J. B. Evans, and L. Searl, D. DePardo, V. R. Petty et al., “Cognitive Radio for
14
Dynamic Spectrum Access—An Agile Radio for Wireless Innovation,” IEEE Communica-
15 tions Magazine, May, 2007.
16 [11] Chang Chen, J. Wawrzynek., and R W. Brodersen., “BEE2: A High-End Reconfigurable
17 Computing System,” IEEE Design and Test of Computers Magazine, March/April, 2005.
18 [12] Shi Q., D. Taubenheim, S. Kyperountas, P. Gorday, N. Correal et al., “Link Maintenance
19 Protocol for Cognitive Radio System with OFDM PHY,” Proceedings of the IEEE Inter-
20 national Symposium on New Frontiers in Dynamic Spectrum Access Networks, Dublin,
21 Ireland, 2007.
22 [13] Farrell, R., M. Sanchez, and G. Corley, “Software-Defined Radio Demonstrators: An Example
and Future Trends” International Journal of Digital Multimedia Broadcasting, 2009.
23
[14] Rice University WARP, WARP: Wireless Open-Access Research Platform, http://warp.rice.
24
edu/.
25
[15] Lyrtech RD Incorporated, Lyrtech RD Processing Systems: Software-Defined Radios,
26 http://lyrtechrd.com/en/ products/families/+processing-systems+software-defined-radio.
27 [16] Epiq Solutions, MatchStiq: Handheld Reconfigurable RF Transceiver, http://epiqsolutions.
28 com/matchstiq/.
29 [17] Leferman M. J., “Rapid Prototyping Interface for Software Defined Radio Experimenta-
30 tion,” Masters Thesis of Worcester Polytechnic Institute, Worcester, MA, 2010.
31 [18] The MathWorks, USRP Hardware Support from MATLAB and Simulink, http://www.
32 mathworks.com/discovery/sdr/usrp.html.
33 [19] GNU Radio, Welcome to GNU Radio!, http://gnuradio.org/.
34 [20] Nychis, The Comprehensive GNU Radio Archive Network. http://www.cgran.org/.
35 [21] Sutton P. D., J. Lotze, H. Lahlou, S. A. Fahmy, and K. Nolan et al., “Iris: An Architecture for
Cognitive Radio Networking Testbeds” IEEE Communications Magazine, September, 2010.
36
[22] Kensington P., RF and Baseband Techniques for Software Defined Radio, Norward, MA:
37
Artech House, 2005.
38
[23] Cabric, D., D. Taubenheim, G. Cafaro, R. Farrell, “Cognitive Radio Platforms and Test-
39 beds” Cognitive Radio Communications and Networks: Principles and Practice, Academic
40 Press, 2009.
41 [24] Farrell, R., R. Villing, A. M. Wyglinski, C. R. Anderson, J. H. Reed et al., “Rationale for
42 a Clean Slate Radio Architecture,” Proceedings of the SDR Forum Technical Conference,
43 Washington, DC, 2008.
44
45
46
47

ART_Wyglinski_Ch01.indd 14 Manila Typesetting Company 12/06/2012

 Chapter 2

Signals and Systems Overview

This chapter covers the theory of linear time-invariant (LTI) signals and systems,
which forms the fundamental basis for the latter chapters. Sections 2.3 to 2.5 intro-
duce several more advanced topics, such as sampling, pulse shaping, and filtering.
However, this chapter does not discuss these topics at length, preferring to direct
the reader to more comprehensive books on these subjects.

2.1  Signals and Systems

2.1.1 Introduction to Signals
A signal is defined as any physical quantity that varies with time, space, or any other
independent variable or variables. Mathematically, we describe a signal as a func-
tion of one or more independent variables [1]. For example, a cosine wave signal
is defined as:
s(t) = Acos(2π ft) (2.1)
where the signal amplitude A and carrier frequency f are constants. Then, this signal
is a function of t. In other words, this signal varies with time. We can classify signals
based on different criteria.
In terms of distribution, signals can be classified as deterministic signals and
random signals. For example, the cosine wave signal defined in (2.1) is a determin-
istic signal because for each variable t, there is one and only one corresponding s(t).
However, the temperature of a location at each hour of a day is a random variable.
Even if we know the temperature for past hours, we cannot predict what the tem-
perature would be for the next hour.
With respect to the value of independent variable t, signals can be classified as
either a continuous time signal or a discrete time signal. If t can pick up any value
from –¥ to ¥, the signal is a continuous time signal. Otherwise, if t can only pick
up discrete values from –¥ to ¥, the signal is a discrete time signal. In this case,
we usually use n instead of t to denote the time variable, and the resulting signal is
written as s[n] instead of s(t). Therefore, a discrete time cosine wave signal can be
expressed as:
s[n] = A cos(2π fn), n = 1,…, N (2.2)

where n are the discrete values we can pick up for time variable. Figure 2.1(a)
shows an example of continuous time triangular wave, and Figure 2.1(b) shows an
example of discrete time triangular wave.
In this chapter, we will discuss converting a signal from continuous-time do-
main to discrete-time domain using sampling, as well as from discrete-time domain
15
16 Signals and Systems Overview

(a)

(b)
Figure 2.1  Use triangular wave as an example to illustrate continuous time signal and discrete time
signal. (a) Continuous time signal. (b) Discrete time signal.

to continuous-time domain, using reconstruction. These processes are referenced to

as analog-to-digital and digital-to-analog conversions.
Signals can be classified as either periodic signals or aperiodic signals. For a
constant time T0, if the signal s(t) satisfies:

s(t) = s(t + kT0 ), k = - N ,…,1,…, N (2.3)

then s(t) is a periodic signal of period T0; otherwise s(t) is an aperiodic signal. Simi-
larly, for discrete time signal s[n], if it satisfies:

s[n] = s[n + kT0 ], k = - N ,…,1,…, N (2.4)

then s[n] is a periodic signal of period T0; otherwise s[n] is an aperiodic signal. Fig-
ure 2.2(a) shows an example of periodic triangular wave, and Figure 2.2(b) shows
an example of aperiodic triangular wave.

2.1.2 Introduction to Systems
A system may be defined as a physical device that performs an operation on a signal
[1]. We usually denote a system using letter S. For example, assume the input sig-
nal to a system is a continuous-time signal x(t), after some operations of system S,
the output signal is a continuous-time signal y(t), as shown in Figure 2.3, then this
continuous-time system can be defined as:
2.1  Signals and Systems 17

(a)

(b)
Figure 2.2  Use triangular wave as an example to illustrate periodic signal and aperiodic signal.
(a) Periodic triangular wave. (b) Aperiodic triangular wave.

y(t) = S[ x(t)] (2.5)

where the output signal at time t depends on the input signal x(t), as well as the
system S.
Similarly, we can define a discrete-time system as:

y[n] = S[ x[n]] (2.6)

For most situations, the LTI systems are most frequently studied and used. This
type of system has two most important properties, linearity and time-invariance,
which are specified as follows:

• Linearity: Linearity property includes two aspects, namely, superposition

property and homogeneity. Superposition property means when there are
several inputs to the system, the resulting output equals to the summa-
tion of each independent output. Homogeneity means if the input signal
is multiplied by a constant a, the resulting output should also be multi-
plied by a constant a. Therefore, linearity property can be expressed as:

Figure 2.3  A continuous system S, with input signal x(t) and output signal y(t).
18 Signals and Systems Overview

If
S[ x1(t)] = y1(t), S[ x2 (t)] = y2 (t) (2.7)
then

S[a1x1(t) + a2 x2 (t)] = a1S[ x1(t)] + a2 S[ x2 (t)] (2.8)

• Time-invariant: Time-invariant can be expressed as:
If

S[ x(t)] = y(t) (2.9)

then

S[ x(t - t0 )] = y(t - t0 ) (2.10)

It means for a time-invariant system, when the input signal is delayed by a
time period of t0, the output signal will also be delayed by the same amount
of time t0.

There are several other important properties of systems. Since this chapter fo-
cuses on continuous-time systems, the following properties are discussed in the
continuous-time context:

• Causal: A system is causal if for t < t0, the input signal is 0, then the corre-
sponding output signal for t < t0 is also 0. In other words, the output signal
cannot exist before the input signal. Otherwise, the system is called a non­
casual system.
• Memoryless: A system is memoryless if the output value at any time t de-
pends and only depends on the input signal value at that same time t.

In order to analyze a linear time-invariant system, we usually need to build a

mathematical model for the system. There are two frequently used models, namely,
input-output model and state variable model. Then, two methods can be used to
solve these models. They are time domain analysis and transform domain analysis.
The latter leads us to the next section, which introduces one of the most important
transforms, the Fourier transform.

2.2  Fourier Transform

Electrical and computer engineers often deal with physical phenomena that are
structured as some form of energy transmission via wave emission. Whether it is
sound waves, wireless signals, or ultrasound pulses, it is very important to pos-
sess a collection of tools that can analyze and study these forms of transmission.
Throughout history, numerous scientists and engineers have studied and developed
mathematical tools that can be used to characterize these wave emissions.
One form of mathematical tools developed during the 18th century is Fourier
analysis, which can be used to analyze signals and systems in frequency domain.
2.2  Fourier Transform 19

From a human perspective, we are often more used to analyzing signals and systems
in the time domain. However, the perspectives in the frequency domain can provide
us with more insight on the various features of a particular signal and/or system, as
well as its manipulation by its combination with other signals and processing using
other tools.
Fourier analysis family mainly includes Fourier transform (FT), Fourier series
(FS), discrete time Fourier transform (DTFT), and discrete Fourier transform (DFT).
Each has its own applications. In this section, we will focus on Fourier transform,
but we will also briefly cover the other three.

2.2.1 Introduction and Historical Perspective

Fourier analysis is named after Jean Baptiste Joseph Fourier (March 21, 1768–
May 16, 1830). Fourier was a French mathematician and physicist best known for
initiating the investigation of the Fourier series and their applications to problems
of heat transfer and vibrations. Fourier’s Law is also named in his honor. Fourier is
also generally credited with the discovery of the greenhouse effect [2].
Besides Fourier, many other scientists and mathematicians have also contributed to
this work. For example, Daniel Bernoulli and Leonhard Euler introduced trigonometric
repr esentations of functions, and Lagrange gave the Fourier series solution to the wave

Figure 2.4  Jean Baptiste Joseph Fourier, 1768–1830 (from [3]).

20 Signals and Systems Overview

equation. The subsequent development of the field is known as harmonic analysis and
is also an early instance of representation theory. The first fast Fourier transform (FFT)
algorithm for the DFT was discovered around 1805 by Carl Friedrich Gauss when in-
terpolating measurements of the orbit of the asteroids Juno and Pallas [4].

2.2.2 Definition
As mentioned at the beginning of this section, each tool in the Fourier family has its
own application, which is listed in Table 2.1. These applications are classified based
on the properties of input signals, as shown in Figure 2.5.
The Fourier transform, in essence, decomposes or separates a continuous, ape-
riodic waveform or function, as shown in Figure 2.5(a), into sinusoids of different
frequency that sum to the original waveform. It identifies or distinguishes the dif-
ferent frequency sinusoids and their respective amplitudes [5].
The Fourier transform of x(t) is defined as [1]:
¥
X(ω) = ò x(t)e - jωt dt (2.11)
-¥

where t is the time variable in seconds across the time domain, and W is the fre-
quency variable in radian per second across frequency domain.
Applying the similar transform to X(w) yields the inverse Fourier transform [1]:

¥
x(t) = ò X(ω)e j 2πωt dω
-¥ (2.12)

where we write x(t) as a weighted sum of complex exponentials.

The previous Fourier transform pair can be denoted as [6]:
F
x(t) ↔ X(ω) (2.13)
F
where the left hand side of the symbol ↔ is before Fourier transform, while the
F
right hand side of the symbol ↔ is after Fourier transform. The most frequently
used Fourier transform pairs are listed in Table 2.2 for your quick reference.
The mathematical definitions of the other three Fourier tools are listed in Table
2.3, but we will not introduce them in detail. Please note that for Fourier series, x(t)
is a periodic signal with fundamental period Tp = l/F0, and {ck} are the coefficients.

2.2.3  Properties
Several properties of Fourier transform are useful when studying signals and sys-
tems in the Fourier transform domain:

Table 2.1  The Applications of Fourier Family Tools

Name Continuous/Discrete Periodic/Aperiodic
Fourier transform Continuous Aperiodic
Fourier series Continuous Periodic
Discrete time Fourier transform Discrete Aperiodic
Discrete Fourier transform Discrete Periodic
2.2  Fourier Transform 21

(a)

(b)

(c)

(d)
Figure 2.5  Use triangular wave as an example to illustrate four different types of signals. (a) Con-
tinuous, aperiodic signal. (b) Continuous, periodic signal. (c) Discrete, aperiodic signal. (d) Discrete,
periodic signal.
22 Signals and Systems Overview

Table 2.2  Fourier Transform Table: Selected Pairs [6]

x(t) X(wj
u(t) 1
πδ(ω) +
jω
d(t) 1
1 2pd(w)
e jw0t 2pd(w — w0)
cos(w0t) p[d(w – w0) + d(w + w0)]
sin(w0t) p/j[d(w – w0) – d(w + w0)]
cos(w0t)u(t) jω
π /2[δ (ω - ω 0 ) + δ (ω + ω 0 )] +
ω02 - ω 2
sin(w0t)u(t) jω
π /2 j[δ(ω - ω 0 ) + δ (ω + ω 0 )] +
ω02 - ω 2
e−atu(t), a > 0 1
a + jw
te−atu(t), a > 0 æ 1 ö
2

çè a + jw ÷ø
e−a|t|u(t), a > 0 2a
a + w2
2

• Linearity: Since Fourier transform is a linear transform, it has linearity. Ac-

cording to Section 2.1.2, for any constant an(n = 1,..., N), if the time-domain
N N
signals x(t) = å n =1 an xn (t), then its Fourier transform is å n =1 an Xn (ω).
• Symmetry: If a signal x(t) is symmetric in time domain, then its Fourier trans-
form X(w) is also symmetric in frequency domain.
• Shifting property: Shifting property includes time-shifting property and
­frequency-shifting property, namely:

–  Time-shifting property: If a signal x(t) is shifted by t0 on time domain, then

its Fourier transform X(w) will be multiplied by e–jwt0 on frequency domain.
Suppose the time signal is x(t) and its Fourier transform signal is X(w).
–  Frequency-shifting property: If a signal X(w) is shifted by w0 on frequency
domain, then its inverse Fourier transform x(t) will be multiplied by ejw0t
on time domain.

Table 2.3  The Mathematical Definitions of the Fourier Tools [1]

Name Direct Transform Inverse Transform
Fourier series 1 ¥
x(t) = å k = -¥ cke j 2π kF0t
Tp ò
ck = x(t)e - j 2π kF0t dt

Discrete time Fourier transform ¥

X(ω) = å -¥ x[n]e - jω n 1
2π ò
x[n] = X(ω)e jω n dω

Discrete Fourier transform 1 N -1

ck = å x[n]e - j 2π kn/N
N n =0
N -1
x[n] = å k = 0 cke j 2π kn/N
2.3  Sampling Theory 23

Table 2.4  Fourier Transform Properties, based on [6]

Property Time Signal Fourier Transform Signal
Linearity N N
å n =1 an xn (t) å n =1 an Xn (ω)
Symmetry x(–t) X(–w)
Time shift x(t – t0) C(w)e–jwt0
Frequency shift x(t)ejw0t C(w – w0)
Scaling x(at) 1 æωö
X
α èç α ø÷
Derivative dn (jw)nC(w)
x(t)
dt n
Integration ∞ X(ω)
∫ −∞ x(τ )dτ jω
+ π X(0)δ (ω)

Time convolution x(t) h(t) C(w)H(w)

Frequency convolution C(t) h(t) 1
X(ω) * H(ω )
2π

Suppose the time signal is x(t) and its Fourier transform signal is X(w).

• Scaling property: If a signal x(t) is scaled by a factor of a on time domain,

then its Fourier transform X(w) will be scaled by a factor of 1/a on frequency
domain.
• Derivative and integration: If a time-domain signal x(t) is taking derivative
or integration operation, then its Fourier transform X(w) will be multiplied
or divided by jw.
• Convolution theorem: Convolution theorem includes time-domain convolu-
tion and frequency-domain convolution, namely:

–  Time-domain convolution: If two signals x(t) and h(t) are convoluted on

time domain, then their Fourier transform X(w) and H(w) will be multi-
plied on frequency domain.
–  Frequency-domain convolution: If two signals X(w) and H(w) are convo-
luted on frequency domain, then their inverse Fourier transform x(t) and
h(t) will be multiplied on time domain.

The Fourier transform properties introduced here are summarized in Table 2.4
for your quick reference.

2.3  Sampling Theory

As mentioned in Section 2.1.1, a continuous-time signal can be converted to a discrete-

time signal using sampling. At the same time, sampling is the most important operation
in analog-to-digital conversion, as shown in Figure 2.6, where an analog input
signal xa(t) is converted to a digital output signal. Sampling is the conversion of a
continuous-time signal into a discrete-time signal obtained by taking “samples” of
the continuous-time signal at discrete-time instants [1]. Section 2.3.1 will introduce
a frequently used sampling method, namely, uniform sampling.
24 Signals and Systems Overview

Figure 2.6  Basic parts of an analog-to-digital converter (ADC) [1]. Sampling takes place in the
sampler block.

Similarly, a discrete-time signal can also be converted to a continuous-time signal

using reconstruction. However, reconstruction is not always successful. Sometimes,
the reconstructed signal is not the same as the original signal. Since for a given sam-
pled signal, we can fit infinite continuous-time signals through the samples. However,
if the sampling satisfies certain criterion, the signal can be reconstructed perfectly.
This criterion, called Nyquist sampling theorem, will be introduced in Section 2.3.3.

2.3.1 Uniform Sampling
Concerning sampling interval, there are infinite ways to do sampling. However, if we
specify the sampling interval as a constant number Ts, we get the most widely used sam-
pling, called uniform sampling, or periodic sampling. Using this method, we are taking
“samples” of the continuous-time signal every Ts seconds, which can be defined as:

x[n] = x(nTs ), - ¥ < n < ¥ (2.14)

where x(t) is the input continuous-time signal, x[n] is the output discrete-time sig-
nal, Ts is the sampling period, and fs = 1/Ts is the sampling frequency.
An equivalent model for the uniform sampling operation is shown in
­Figure 2.7(a), where the continuous-time signal x(t) is multiplied by an impulse
train p(t) to form the sampled signal xs(t), which can be defined as:
xs (t) = x(t)p(t) (2.15)

where the signal p(t) is referred to as the sampling function.

The sampling function is assumed to be a series of narrow pulses, which is either
zero or one. Thus, xs(t) = x(t) when p(t) = 1, and xs(t) = 0 when p(t) = 0. Since p(t) = 1
only at time instants Ts, the resulting xs(t) = x(nTs) = x[n], which proves that this is in-
deed an equivalent model for the uniform sampling operation. This model will help us
to obtain the frequency domain representation of uniform sampling in Section 2.3.2.

2.3.2  Frequency Domain Representation of Uniform Sampling

Since it is easier for us to observe and derive the Nyquist sampling theorem in
frequency domain, in this section, we will try to represent the uniform sampling
process in frequency domain.
2.3  Sampling Theory 25

(a)

(b)
Figure 2.7  An equivalent model for the uniform sampling operation. (a) A continuous-time signal
x(t) is multiplied by a periodic pulse p(t) to form the sampled signal xs(t). (b) A periodic pulse p(t).

According to Figure 2.7(b), we can define the sampling function p(t) as:
¥
p(t) = å δ (t - kTs ), k = 0,1, 2,… (2.16)
k =-¥

where at time instants kTs, we have p(t) = 1. According to [7], p(t) is a Dirac comb
constructed from Dirac delta functions.
Substitution of (2.16) into (2.15) gives:

¥
(2.17)
xs (t) = x(t)p(t) = x(t) å δ (t - kTs )
k =-¥

In order to learn the sampling process in frequency domain, we need to take the
Fourier transform of xs(t). According to frequency-domain convolution property in
26 Signals and Systems Overview

Section 2.2.3, multiplication in time domain will lead to convolution in frequency

domain. Therefore, multiplication of x(t) and p(t) will yield the convolution of
X(w) and P(w):
1
Xs (ω) = X(ω ) * P(ω ) (2.18)
2π
where X(w) is the Fourier transform of x(t), and P(w) is the Fourier transform of p(t).
The Fourier transform of a Dirac comb is also a Dirac comb [7], namely:
¥ ¥
2π æ 2π ö 2π
P(ω) =
Ts
å δ çèω - k Ts ÷ø = Ts å δ (ω - kωs ) (2.19)
k =-¥ k =-¥
where ws = 2p fs is the sampling frequency.
Substitute (2.19) into (2.18). Since convolution with delta function pulses rep-
licas at pulse location we have:
¥
1
Xs (ω) =
2π Ts
å X(ω - kω s ) (2.20)
k =-¥

Equation (2.20) tells us that the uniform sampling creates images of the Fourier
transform of the input signal, and images are periodic with sampling frequency fs.

2.3.3 Nyquist Sampling Theorem

Based on (2.20), we draw the spectrum of original signal x(t) and the sampled signal
xs(t) on frequency domain, as shown in Figure 2.8. Assume the bandwidth of the
original signal is [–fh, fh], as shown in Figure 2.8(a). For now, we do not pay at-
tention to the signal amplitude, so we use A and As to represent them. Assume the
sampling frequency is fs, then the sampled signal will have replicas at location kfs.
In order to reconstruct the original signal from the sampled signal, we will apply
a lowpass filter on the sampled signal, trying to extract the n = 0 term from Xs(f ),
as shown in Figure 2.8(b). Therefore, accomplishing reconstruction without error
requires that the portion of the spectrum of Xs(f ) at f = ±fs does not overlap with
the portion of the spectrum at f = 0. In other words, this requires that fs – fh > fh or
fs > 2fh, which leads to the Nyquist sampling theorem.
Nyquist sampling theorem applies for the bandlimited signal, which is a signal
x(t) that has no spectral components beyond a frequency B Hz [8]. That is:
X(ω) = 0, ω > 2π B (2.21)

The Nyquist sampling theorem states that a real signal, x(t), which is bandlim-
ited to B Hz can be reconstructed without error from samples taken uniformly at
a rate R > 2B samples per second. This minimum sampling frequency, Fs = 2B Hz,
is called the Nyquist rate or the Nyquist frequency. The corresponding sampling
interval, T = 1/2B, is called the Nyquist interval [1]. A signal bandlimited to B Hz
which is sampled at less than the Nyquist frequency of 2B, (i.e., which was sampled
at an interval T > 1/2B is said to be undersampled).
When a signal is undersampled, its spectrum has overlapping tails, where Xs(f )
no longer has complete information about the spectrum and it is no longer possible
2.3  Sampling Theory 27

(a)

(b)

(c)
Figure 2.8  The spectrum of original signal x(t) and the sampled signal xs (t) on frequency domain.
(a) The spectrum of original signal x(t), with bandwidth –fh to fh, and amplitude A. (b) The spectrum
of sampled signal xs(t), which satisfies Nyquist sampling theorem. (c) The spectrum of sampled signal
xs(t), which does not satisfies Nyquist sampling theorem and has aliasing.

to recover x(t) from the sampled signal. In this case, the tailing spectrum does not
go to zero, but is folded back onto the apparent spectrum. This inversion of the tail
is called spectral folding or aliasing, as shown in Figure 2.8(c) [5].

2.3.4  Sampling Rate Conversion

In real-world applications, sometimes, we would like to lower sampling rate be-
cause it reduces storage and computation requirements. Sometimes, however, we
prefer a higher sampling rate, because it preserves fidelity. Sampling rate conversion
is a general term for the process of changing the time interval between the adjacent
elements in a sequence consisting of samples of a continuous-time function [5].
The process of lowering the sampling rate is called decimation, which is achieved
by ignoring all but every Dth sample. In time domain, it can be defined as:

y[n] = x[nD] (2.22)

28 Signals and Systems Overview

where x[n] is the original signal, y[n] is the decimated signal, and D is the decima-
tion rate.
According to (2.22), the sampling rates of the original signal and the decimated
signal can be expressed as:
F
Fy = x (2.23)
D

where Fx is the sampling rates of the original signal, and Fy is the sampling rates of
the decimated signal.
Since the frequency variables in radians, wx and wy, can be related to sampling
rate, Fx and Fy, by:
2π F (2.24)
ω x = 2π FTx =
Fx
and
2π F
ω y = 2π FTy = (2.25)
Fy
it follows that wx and wy are related by:
ω y = Dω x (2.26)

which means that the frequency range of wx is stretched into the corresponding
frequency range of wy by a factor of D.
In order to avoid aliasing of the decimated sequence y[n], it is required that
0 £½wy½£ p. Based on (2.26), it implies that the spectrum of the original sequence
π
should satisfy 0 £ ω x £ . Therefore, in reality, decimation is usually a two-step
D
process, consisting of a lowpass anti-aliasing filter and a downsampler, as shown
in Figure 2.9. The lowpass anti-aliasing filter is used to constrain the bandwidth of
π
the input signal to the downsampler x[n] to be 0 £ ω x £ .
D
In frequency domain, the spectrum of the decimated signal, y[n], can be ex-
pressed as [1]:
1 D-1 æ ω y - 2π k ö æ ω y - 2π k ö
Y (ω y ) = å
D k =0
HD ç
è D
S
ø÷ èç D ÷ø (2.27)

where S(w) is the spectrum of the input signal s[n], and HD(w) is the frequency
response of the lowpass filter hD[n]. With a properly designed filter HD(w), the
aliasing is eliminated, and consequently, all but the first k = 0 term in (2.27) vanish
[1]. Hence, (2.27) becomes:
1 æ ωy ö æ ωy ö 1 æ ωy ö
Y (ω y ) = HD ç S = S (2.28)
D è D ø÷ èç D ø÷ D èç D ø÷

Figure 2.9  The structure of decimation, consisting of a lowpass anti-aliasing filter and a downsampler.
2.3  Sampling Theory 29

for 0 ≤ ω y ≤ π . The spectra for the sequence x[n] and y[n] are illustrated in
π
­Figure 2.10, where the frequency range of the intermediate signal is 0 £ ω x £ ,
D
and the frequency range of the decimated signal is 0 £ ω y £ π .
The process of increasing the sampling rate is called interpolation, which can
be accomplished by interpolating I – 1 new samples between successive values of
signal. In time domain, it can be defined as:

ì x[n/I ] n = 0, ± I, ± 2I,…
y[n] = í (2.29)
î0 otherwise

where x[n] is the original signal, y[n] is the interpolated signal, and I is the inter-
polation rate.
According to (2.29), the sampling rates of the original signal and the inter­
polated signal can be expressed as:

Fy = IFx (2.30)

where Fx is the sampling rates of the original signal, and Fy is the sampling rates
of the interpolated signal. Since (2.24) and (2.25) also hold here, it follows that wx
and wy are related by:

(a)

(b)
Figure 2.10  The spectra for the sequence x[n]and y[n] where the frequency range of wx is stretched
into the corresponding frequency range of wy by a factor of D. (a) Spectrum of the intermediate
sequence. (b) Spectrum of the decimated sequence.
30 Signals and Systems Overview

ωx (2.31)
ωy =
I
which means that the frequency range of wx is compressed into the corresponding
frequency range of wy by a factor of I. Therefore, after the interpolation, there will
be I replicas of the spectrum of x[n], where each replica occupies a bandwidth of
p /I. Since only the frequency components of y[n] in the range 0 £ |wy| £ p /I are
unique, (i.e., all the other replicas are the same as this one), the images of Y(w)
above wy = p/I should be rejected by passing it through a lowpass filter with the
following frequency response:
ì π
ïC 0 £ ω y £ (2.32)
HI (ω y ) = í I
ïî0 otherwise
where C is a scale factor.
Therefore, in reality, interpolation is also a two-step process, consisting of an
upsampler and a lowpass filter, as shown in Figure 2.11. The spectrum of the output
signal z[n] is:
ì π
ïCX(ω z I) 0 £ ω z £ (2.33)
Z(ω z ) = í I
ïî0 otherwise
where X(w) is the spectrum of the output signal x[n].
The spectra for the sequence x[n], y[n], and z[n] are illustrated in Figure 2.12,
where the frequency range of the original signal is 0 £ |wx| £ p, and the frequency
π
range of the decimated signal is 0 £ ω z £ .
I

2.4  Pulse Shaping

In Section 2.3, we have shown that an analog signal can be converted to a digital
signal using an analog-to-digital converter (ADC), as illustrated in Figure 2.6. These
digital signals consisting of “1” and “0” values have significant advantages in the
aspect of signal processing. However, they cannot be directly transmitted in a real-
world situation. Instead, they must be first converted into an analog signal in order
to be transmitted. This conversion is done by the “transmit” or “pulse-shaping”
filter, which changes each symbol in the digital message into a suitable analog pulse
[8]. This process is shown in Figure 2.13, where the impulse modulator is denoted
by p(t), and the impulse response of the transmit filter is denoted by hT(t).
Similar to the sampling function in Section 2.3.2, the impulse modulator can
be defined as:
∞
p(t) = ∑ δ (t − kT ) (2.34)
k= −∞

Figure 2.11  The structure of interpolation, consisting of an upsampler and a lowpass filter.
2.4  Pulse Shaping 31

(a)

(b)

(c)
Figure 2.12  The spectra for the sequence x[n], y[n], and z[n], where the frequency range of wx
is compressed into the corresponding frequency range of wy by a factor of I. (a) Spectrum of the
original sequence. (b) Spectrum of the intermediate sequence. (c) Spectrum of the interpolated
sequence.

where p(t) = 1 for t = kT, and p(t) = 0 for all the other time instants. Therefore, the
analog pulse train after the impulse modulator is:
¥ t = kT
ìs(kT )
sa (t) = s[n]p(t) = å s(kT )δ (t - kT ) = í
t ¹ kT
(2.35)
k =-¥ î0
32 Signals and Systems Overview

Figure 2.13  On the transmitter side, the impulse modulator and transmit filter convert the digital
symbols into analog pulses for the sake of transmission.

where each digital symbol s[n] initiates an analog pulse that is scaled by the value
of the symbol.
If we do not have a transmit filter in our system, these pulses sa(t) will be trans-
mitted through the channel right away. In theory, on the receiver side, we should
be able to get the same pulse train, although with some delay. However, in reality,
we actually cannot get it due to the interference between adjacent symbols. There
are two situations when adjacent symbols may interfere with each other: when the
pulse shape is wider than a single symbol interval T, and when there is a nonunity
channel that “smears” nearby pulses, causing them to overlap. Both of these situ-
ations are called intersymbol interference (ISI) [8]. In order to solve this problem,
pulse shaping filters are introduced to bandlimit the transmit waveform.

2.4.1 Eye Diagrams
In telecommunication, an eye diagram, also known as an eye pattern, is an oscil-
loscope display in which a digital data signal from a receiver is repetitively sampled
and applied to the vertical input, while the data rate is used to trigger the horizontal
sweep [8]. It is so called because, for several types of coding, the pattern looks like
a series of eyes between a pair of rails.
Several system performance measures can be derived by analyzing the display,
especially the extent of the ISI. As the eye closes, the ISI increases; as the eye opens,
the ISI decreases. Besides, if the signals are too long, too short, poorly synchronized
with the system clock, too high, too low, too noisy, or too slow to change, or have
too much undershoot or overshoot, this can be observed from the eye diagram. For
example, Figure 2.14 shows a typical eye pattern for the noisy QPSK signal.

2.4.2 Nyquist Pulse Shaping Theory

In a communication system, there are normally two pulse shaping filters. One on
the transmitter side, and the other on the receiver side, as shown in Figure 2.15,
where we use hT (t) and hR(t) to denote the impulse response of the transmit filter
and receive filter. For simplicity, when considering the Nyquist pulse shaping theory,
we usually use the concept of equivalent channel, which includes not only the chan-
nel itself, but also the two pulse shaping filters. Therefore, the impulse response of
the equivalent channel is:
2.4  Pulse Shaping 33

Figure 2.14  A typical eye pattern for the QPSK signal. The width of the opening indicates the time
over which sampling for detection might be performed. The optimum sampling time corresponds to
the maximum eye opening, yielding the greatest protection against noise. If there were no filtering
in the system, then the system, would look like a box rather than an eye.

h(t) = hT (t) ∗hC (t)∗hR(t) (2.36)

Now, let us consider under which condition we can assure that there is no in-
tersymbol interference between symbols. The input to the equivalent channel, sa(t),
has been defined in (2.35). As mentioned before, each analog pulse is scaled by the
value of the symbol, so we can express sa(t) in another way:

Figure 2.15  The equivalent channel of a communication system, which consists of the transmit
filter, the channel, and the receive filter.
34 Signals and Systems Overview

sa (t) = ∑ akδ (t − kT ) (2.37)

where ak is the value of the kth symbol. It yields the output to the equivalent chan-
nel, y(t), which is:
y(t) = sa (t)* h(t) = å ak [δ (t - kT )* h(t)] = å ak h(t - kT ) (2.38)

Therefore, given a specific time instant, for example, t = mT, where m is a constant,
the input sa(t) is:
(2.39)
a ∑
s (mT ) = a d (mT − kT ) = a
k m

Consequently, the output y(t) becomes:

y(mT ) = å ak h(mT - kT ) = a0 h(mT ) + a1h(mT - T ) + . . . + amh(mT - mT ) (2.40)

Since we do not want the interference from the other symbols, we would like the
output to contain only the am term, which is:
y(mT ) = am h(mT - mT ) (2.41)

Moreover, it means at a time instant t = mT, we need to have:

 C k = m (2.42)
h(mt − kT ) = 
 0 k ≠ m
where C is some nonzero constant.
If we generalize (2.42) to any time instant t, we can get the Nyquist pulse shap-
ing theory as follows.
The condition that one pulse does not interfere with other pulses at subsequent
T-spaced sample instants is formalized by saying that h(t) is a Nyquist pulse if it
satisfies:

C k = 0 (2.43)
h(t) = h(kT ) = 
0 k≠ 0
for all integers k.

2.4.3 Two Nyquist Pulses

In this section, we will introduce two important Nyquist pulses, namely, sine pulse
and raised cosine pulse. When considering (2.43), the most straightforward pulse we
can think of is a rectangular pulse with time-width less than T, or any pulse shape
that is less than T wide. But the bandwidth of the rectangular pulse (and other nar-
row pulse shapes) may be too wide. Narrow pulse shapes do not utilize the spectrum
efficiently, but wide pulse shapes may cause ISI, so what is needed is a signal that
is wide in time (and narrow in frequency) that also fulfills the Nyquist condition in
(2.43) [8].

2.4.3.1  Sinc Pulse

One possible solution is the sine pulse. In mathematics, the sine function is defined as:
2.4  Pulse Shaping 35

sin(π x)
sinc(x) = (2.44)
πx
As shown in Figure 2.16, when variable x takes an integer value k, the value of the
sine function will be:
ì1 k = 0 (2.45)
sinc(k) = í
î0 k ¹ 0
In other words, zero crossings of the sine function occur at nonzero integer values.
Another important property of sinc function is that sine pulse is the inverse
Fourier transform of a rectangular signal, as shown in Figure 2.17(a). Suppose the
rectangular signal is defined as [9]:
ì 1
ïT |ω | <
H(ω) = í 2T (2.46)
ïî 0 otherwise
Take the inverse Fourier transform of the rectangular signal, we can obtain the sine
signal as:

ætö (2.47)
h(t) = sinc ç ÷
èTø

Make t = kT in (2.47), and we will have:

æ kT ö
h(t) = h(kT ) = sinc ç = sinc (k) (2.48)
è T ÷ø

1.2

0.8

0.6
sinc(x)

0.4

0.2

4
0 5 10
x
Figure 2.16  The plot of sine function as defined in (2.44). The x-axis is x, and the y-axis is sinc(x).
36 Signals and Systems Overview

Since k is an integer here, according to (2.45), we can continue writing (2.48) as:
æ kT ö ì1 k=0
h(t) = h(kT ) = sinc ç = sinc (k) = í (2.49)
è T ÷ø î0 k¹0
Comparing (2.49) with (2.43), we can easily find that if we make t = kT, the sine
pulse exactly satisfies Nyquist pulse shaping theory in Section 2.4.2. In other words,
1
by choosing sampling time at kT (sampling frequency equals ), our sampling in-
T
stants are located at the equally spaced zero crossings, as shown in Figure 2.17(b),
so there will be no intersymbol interference.
Recall the Nyquist sampling theorem in Section 2.3.3 states that a real sig-
nal, x(t), which is bandlimited to B Hz can be reconstructed without error using a

(a)

1.2

0.8

0.6
sinc(k)

0.4

0.2

...... ......
0

4
0 5 10
k

(b)
Figure 2.17  The sine pulse on time domain and its Fourier transform, rectangular pulse, on fre-
quency domain. (a) The rectangular pulse on frequency domain, defined in (2.46). (b) The sine pulse
t
defined in (2.49). The x-axis is k, where k = , and the y-axis is sinc(k).
T
2.4  Pulse Shaping 37

minimum sampling frequency of Fs = 2B Hz. In this case, the minimum sampling

1
frequency is Fs = Hz. Therefore, the corresponding minimum bandwidth is:
T
F 1 (2.50)
B= s =
2 2T
which is exactly the bandwidth of the rectangular signal defined in (2.46). Based on
the previous discussion, this choice of sine pulse h(t) yields the minimum bandwidth
B = Bmin = 1 , so it is called Nyquist-I pulse [10].
2T  
2.4.3.2  Raised Cosine Pulse
Nyquist-I pulse is a type of very attractive pulse, because it is wide in time and nar-
row in frequency, which means it has the advantages of both spectrum efficiency
and no ISI. However, Nyquist-I pulse is not practical, as it has ISI sensitivity due
to timing errors. Take sine pulse, for instance; for large t, the sine pulse defined in
(2.47) has the following approximation:

1 (2.51)
h(t) ~
t
so it is obvious that timing error can cause large ISI.
Given that Nyquist-I pulse is not practical, we need to introduce Nyquist-II
pulse, which has a larger bandwidth B > Bmin, but with less ISI sensitivity. Since this
type of pulse is more practical, it is much more widely used in pratice.
The raised cosine pulse is one of the most important types of Nyquist-II pulses,
which has the frequency transfer function defined as:

ì 1- β
ïT 0 £ |f | £
2T
ï
ïT æ æ πT æ 1- β ööö 1-β 1+ β
HRC (f ) = í ç 1 + cos ç çè | f | - ÷ £ |f | £ (2.52)
ï2 è è β 2T ø ÷ø ÷ø 2T 2T
ï 1+ β
ï0 |f | ³
î 2T
β
where b is the rolloff factor, which takes a value from 0 to 1, and is the excess
bandwidth. 2T
The spectrum of raised cosine pulse is shown in Figure 2.18. In general, it has
the bandwidth b ³ 1/(2T). When b = 0, the bandwidth b = 1/(2T), and there is no
excess bandwidth, which is actually a rectangular pulse. On the other end, when
b = 1, it reaches the maximum bandwidth B = l/T.
By taking the inverse Fourier transform of HRC(f ), we can obtain the impulse
response of raised cosine pulse, defined as:

æ βt ö
cos ç π ÷
è Tø æ πt ö
hRC (t) = sinc ç ÷ (2.53)
æ βt ö
2 èTø
1- ç2 ÷
è Tø
38 Signals and Systems Overview

12

10 =0
= 0.5

8 =1
Amplitude

5 5 0 0.5 1 1.5
Normalized Frequency
Figure 2.18  Spectrum of a raised cosine pulse, defined in (2.52), which varies by the rolloff factor b.
The x-axis is the normalized frequency f0. The actual frequency can be obtained by f0/T.

Nyquist-II pulses do not have an ISI sensitivity because their peak distortion,
the tail of hRC(t), converges quickly, which can be expressed as:
¥
1
Dp = å |hRC ( e ¢ + (n - k))| ~
n3
(2.54)
n =-¥

Therefore, when timing error occurs, the distortion will not accumulate to infinity
in a similar fashion related to Nyquist-I pulses [10].
Actually, in many practical communications systems, root raised cosine filters are
used [11]. If we consider the communication channel as an ideal channel, and we as-
sume that the transmit filter and the receive filter are identical, we can use root raised
cosine filters for both of them, and their net response must equal to HRC(f ) defined
in (2.52). Since the impulse response of the equivalent channel can be expressed as:

h(t) = hT (t) * hC (t) * hR (t) (2.55)

where hC(t) is the impulse response of the communication channel, and hT(t) and
hR(t) are the impulse responses of the transmit filter and the receive filter, it means
on frequency domain, we have:
HRC (f ) = HT (f )HR (f ) (2.56)

Therefore, the frequency response of root raised cosine filter must satisfy:

|HT ( f )| = |HR(f )| = |HRC ( f )| (2.57)

2.5  Filtering 39

2.5  Filtering

When doing signal processing, we usually need to get rid of the noise and extract
the useful signal. This process is called filtering, and the device employed is called
a filter, which discriminates, according to some attribute of the objects applied at
its input, what passes through it [1]. Typical filter is a frequency-selective circuit,
namely, if noise and useful signal have different frequency distributions, by apply-
ing this circuit, noise will be attenuated or even eliminated, while the useful signal
is retained.
Filters can be classified from different aspects. For example, according to its
frequency response, filters can be classified as lowpass, highpass, bandpass and
bandstop. According to the signal it deals with, filters can be classified as analog
filters and digital filters. Specifically, analog filters deal with continuous–time sig-
nals, while digital filter deal with discrete-time signals. This section will focus on
digital filters.

2.5.1 Ideal Filter
As mentioned before, filters are usually classified according to their frequency-
­domain characteristics as lowpass, highpass, bandpass, and bandstop or band-
eliminated filters [1]. The ideal magnitude response characteristics of these types of
filters are shown in Figure 2.19.
According to Figure 2.19, the magnitude response characteristics of an ideal
filter can be generalized as follows: In passband, the magnitude is a constant. While
in stopband, the magnitude falls to zero. However, in reality, this type of ideal filter
cannot be achieved, so a practical filter is actually the optimum approximation of
an ideal filter.

2.5.2  Z-Transform
In Section 2.2, we have introduced the Fourier transform, which deals with
continuous-­time signals on the frequency domain. Since we are focusing on digital
filter design in this section, where discrete-time signals are involved, we need to
introduce a new type of transform, namely, the z-transform.
The z-transform of a discrete-time signal x[n] is defined as the power series:

X(z) =
¥
å x[n]z - n (2.58)
n =-¥

where z is a complex variable [1].

The z-transform is used to analyze discrete-time systems. Its continuous-time
counterpart is the Laplace transform, defined as follows:
¥
X(s) = ò x(t)e - st dt
-¥ (2.59)
where t is the time variable in seconds across the time domain, and s = s + jw is
a complex variable. When evaluated along the jw axis (i.e., s = 0), the Laplace
transform reduces to the Fourier transform defined in (2.11). Thus, the Laplace
40 Signals and Systems Overview

(a) (b)

(c) (d)
Figure 2.19  Ideal magnitude response characteristics of four types of filters on the frequency
range [0,2p]. (a) Lowpass filter. (b) Highpass filter. (c) Bandpass filter, where (wCl, wC2) is pass band.
(d) Bandstop filter, where (wCl, wC2) is stop band.

transform generalizes the Fourier transform from the real line (the frequency axis
jw) to the entire complex plane.
According to Section 2.3.2, we know that if a continuous-time signal x(t) is
uniformly sampled, its sampling signal xs(t) can be expressed as:
¥
xs (t) = å x(nT )δ (t - nT ) (2.60)
n =-¥

where T is the sampling interval. If we take the Laplace transform of both sides,
we will get:
¥ ¥ é ¥ ù
Xs (s) = ò
-¥
xs (t)e - st dt = ò
-¥
ê å x(nT )δ (t - nT )ú e - st dt (2.61)
ëê n =-¥ ûú
Since integration and summation are both linear operators, we can exchange their or-
der. Then, based on the sampling property of the delta function, we can further get:
¥ ¥
x(nT ) éê ò δ (t - nT )e - st dt ùú = å x(nT )e - snT
¥
Xs (s) = å ë -¥ û n =-¥
(2.62)
n =-¥

1
Let z = esT, or s = ln z; then (2.62) becomes:
T
¥
X(z) = å x(nT )z - n (2.63)
n =-¥
2.5  Filtering 41

Since T is the sampling interval, x(nT) = x[n]. The previous equation can be further
written as:
¥
X(z) = å x[n]z - n (2.64)
n =-¥

which is exactly the definition of z-transform in (2.58). Therefore, the z-transform

and the Laplace transform can be connected by:

z = e sT (2.65)
or
1
s= ln(z) (2.66)
T

According to (2.58), we know that z-transform is the series of z–1. Actually, the
z-transform has no meaning unless the series converge. Given a limitary-amplitude
sequence x[n], the set of all the z values that makes its z-transform converge is called
region of convergence (ROC). Based on the theory of series, these z-values must
satisfy:
¥

å |x[n]z -n
| < ¥ (2.67)
n =-¥

The frequently used z-transform pairs and their region of convergence are listed in
Table 2.5.
When discussing a linear time-invariant system, the z-transform of its system
impulse response can be expressed as the ratio of two polynomials:

bm z m + bm -1z m -1 + … + b1z + b0 B(z) (2.68)

H(z) = =
an z n + an -1z n -1 + … + a1z + a0 A(z)

where the roots of A(z) = 0 are called the poles of the system, and the roots of
B(z) = 0 are called the zeros of the system. It is possible that the system can have
multiple poles and zeros.
If we factorize the numerator B(z) and denominator A(z), (2.68) can be written
as:
(z - z1)(z - z2 )�(z - zm ) (2.69)
H(z) = C
(z - p1)(z - p2 )�(z - pn )
where C is a constant, {pk} are all the poles, and {zk} are all the zeros. It will help us
draw the pole-zero plot of H(z).
Suppose we have a linear time-invariant system whose system impulse response
is defined as:
h[n] = n 2 an u[n] (2.70)
According to Table 2.5, its z-transform is as follows:
42 Signals and Systems Overview

Table 2.5  Z-Transform Table: Selected Pairs [12]

x[n] X(z) Region of Convergence
d[n] 1 all z
n
a u[n] z ½z½>½a½
z-a

nanu[n] az ½z½>½a½> 0
(z - a)2

n2anu[n] az(z + a) ½z½> a > 0

(z - a)3

æ 1 1ö az bz æ 1 1ö
çè n + n ÷ø u[n] + z > max ç , ÷
a b az - 1 bz - 1 è a bø

anu[n] sin[w0n] az sin ω0 ½z½> a > 0

z 2 - 2az cos ω0 + a2

anu[n] cos[w0n] z(z - a cos ω0 ) ½z½> a > 0

z 2 - 2az cos ω0 + a2

eanu[n] z ½z½> e –a
z - ea

eanu[n] sin[w0n] ze a sin ω0 ½z½> e –a

2 2
z e a - 2ze a cos ω0 + 1

eanu[n] cos[w0n] ze a (ze a cos ω0 ) ½z½> e –a

2 2
z e a - 2ze a cos ω0 + 1

z(z + a) (2.71)
H(z) = a
(z - a)3

Comparing (2.71) with (2.69), we easily get that this system has two zeros, z1 = 0
and z2 = –a, and three poles, p1 = p2 = p3 = a. Therefore, its pole-zero plot is shown
in Figure 2.20.
There are several properties of the z-transform that are useful when studying
signals and systems in the z-transform domain. Since these properties are very simi-
lar to those of the Fourier transform introduced in Section 2.2.3, we list them in
Table 2.6 without further discussion.

2.5.3 Digital Filtering
In order to perform digital filtering, input and output of a digital system must both
be discrete-time series. If the input series is x[n], the impulse response of the filter is
h[n], then the output series y[n] will be:

y[n] = x[n] * h[n] (2.72)

2.5  Filtering 43

Figure 2.20  Pole-zero plot of the system defined in (2.71). Poles are denoted using crossings, and
zeros are denoted using circles. The region of convergence of this system is the region outside the
circle z = | a |.

According to the time convolution property in Table 2.6, on the frequency

­domain, (2.72) is equivalent to:
Y (z) = X(z)H(z) (2.73)

where X(z) and Y(z) are the z-transforms of the input and output series, x[n] and
y[n], and H(z) is the z-transform of h[n].

Table 2.6  Z-Transform Properties, Based on [12]

Property Time Signal Z-Transform Signal
Linearity N N
å m =1 am xm (t) å m =1 am Xm (z)
Symmetry x[–n] X(z–1)
Shifting x[n – m] z –mX(z)
Scaling n
a x[n] æ aö
Xç ÷
è zø
Derivative nx[n] d λ(z)
-z
dz
Integration n z
åm = -¥ x[m] z -1
X(z)

Time convolution x[n] *h[n] X(z)H(z)

Frequency convolution x[n]h[n] 1 æ z ö dv
2π j ò
X(v)H ç ÷
è vø v

Suppose the time signal is x[n] and its z-transform signal is X(z).
44 Signals and Systems Overview

As mentioned in Section 2.5.1, the ideal filters are not achievable in practice.
Consequently, we limit our attention to the class of linear time-invariant systems
specified by the difference equation [1]:

N M
y[n] = - å ak y[n - k] + å bk x[n - k] (2.74)
k =1 k=0

where y[n] is the current filter output, the y[n – k] are previous filter outputs, the x[n – k]
are current or previous filter inputs. This system has the following frequency response:

M
å bk e - z
H(z) = k =0 (2.75)
N
1 + å ak e - z
k =1

where the {ak} are the filter’s feedback coefficients corresponding to the poles of the
filter, the {bk} are the filter’s feed forward coefficients corresponding to the zeros of
the filter, and N is the filter’s order.
The basic digital filter design problem is to approximate any of the ideal fre-
quency response characteristics with a system that has the frequency response (2.75)
by properly selecting the coefficients {ak} and {bk} [1].
There are two basic types of digital filters, finite impulse response (FIR) and infi-
nite impulse response (IIR) filters. When excited by an unit sample d[n], the impulse
response h[n] of a system may last a finite duration, as shown in Figure 2.21(a), or
forever even before the input is applied, as shown in Figure 2.21(b). In the former
case, the system is finite impulse response, and, in the latter case, the system is infi-
nite impulse response.
A FIR filter of length M with input x[n] and output y[n] can be described by the
difference equation [1]:
M -1
y[n] = b0 x[n] + b1x[n - 1] + � + bM -1x[n - M + 1] = å bk x[n - k] (2.76)
k =0

where the filter has no feedback coefficients {ak}, so H(z) has only zeros.
IIR filter has been defined in (2.74), which has one or more nonzero feedback
coefficients {ak}. Therefore, once the filter has been excited with an impulse, there
is always an output.

2.5.3.1  Case Study: Cascaded Integrator-Comb Filter

Since cascaded integrator-comb (CIC) filter plays an important role in the univer-
sal software radio peripheral (USRP) hardware, this section introduces this special
filter. CIC filters, invented by Eugene B. Hogenauer, are a class of FIR filters used
in multirate processing. The CIC filter finds applications in interpolation and deci-
mation. Unlike most FIR filters, it has a decimator or interpolator built into the
architecture [13].
2.5  Filtering 45

(a)

(b)
Figure 2.21  The impulse responses of an FIR filter and an IIR filter. (a) The impulse response of an
FIR filter. (b) The impulse response of an IIR filter.

A CIC filter consists of one or more integrator and comb filter pairs. In the case
of a decimating CIC, the input signal is fed through one or more cascaded integra-
tors, then a down-sampler, followed by one or more comb sections whose number
equals to the number of integrators. An interpolating CIC is simply the reverse of
this architecture, with the down-sampler replaced with a up-sampler, as shown in
Figure 2.22.
To illustrate a simple form of a comb filter, consider a moving average FIR filter
described by the difference equation:

1 M (2.77)
y[n] = å x[n - k]
M + 1 k=0

The system function of this FIR filter is:

1 M -k 1 [1 - z -(M +1) ]
H(z) = å
M + 1 k=0
z =
M + 1 (1 - z -1)
(2.78)

Suppose we replace z by zL, where L is a positive integer; then the resulting

comb filter has the system function:

1 [1 - z - L(M +1) ] (2.79)

HL (z) =
M + 1 (1 - z - L )
46 Signals and Systems Overview

Figure 2.22  The structure of an interpolating cascaded integrator-comb filter [14], with input sig-
nal x[n] and output signal y[n]. This filter consists of a comb and integrator filter pair, and an up-
sampler with interpolation ratio L.

where L is decimation or interpolation ratio, and M is number of samples per stage,

usually 1 or 2. This filter has zeros on the unit circle at:

zk = e j 2π k / L(M +1) (2.80)

for all integer value of k except k = 0, L, 2L,..., ML, as shown in Figure 2.23.
The common form of the CIC filter usually consists of several stages, and the
system function for the composite CIC filter is:
N
æ 1 1 - z - L(M +1) ö
H(z) = HL (z)N = ç ÷ (2.81)
è M + 1 1 - z -L ø
where N is number of stages in the composite filter.
Characteristics of CIC filters include linear phase response and utilizing only
delay and addition and subtraction. In other words, it requires no multiplication
operations. Therefore, it is suitable for hardware implementation.

Figure 2.23  The zeros of a CIC filter defined in (2.79), where all the zeros are on the unit circle.
2.7  PROBLEMS 47

2.6  Chapter Summary

Software defined radio experimentation requires both strong computer skills and
extensive knowledge of digital signal processing. This chapter lists some useful top-
ics including sampling, pulse shaping, and filtering. The purpose of this chapter is
to help the readers get prepared for the experimentation chapters, especially for
the USRP hardware. For example, on the USRP transmit path, there are analog-
to-digital converters and digital down-converters, where the theory of sampling is
needed. On the USRP receive path, besides digital-to-analog converters and digital
up-converters, there are CIC filters, where filtering knowledge is useful. In addition,
when you build a real communication system, you will need a transmit filter and
receive filter, which requires expertise on pulse shaping.

2.7  PROBLEMS

1. [Periodic Signal] Suppose we have a discrete signal

4π
π j n
5
x [ n] = cos( n) + e (2.82)
2

(a) Is it a periodic signal?

(b) If yes, what is its period?

2. [LTI System] Suppose we have a discrete system

y [ n] = nx[ n] + x[ n− 1]. (2.83)

Is it an LTI system? Please justify.

3. [Casual System] Please determine whether the following systems belong to

casual systems or not and explain.

(a) y[n] = x[n] − x[n + 1]

(b) y [ n] = ∑
n
x[ k]
k = −∞
(c) h(t) = e−2t u(t)

4. [Fourier Transform] Assume the Fourier transform of a signal f(t) is F(jw).

(a) What is the Fourier transform of f(2t − 3)?

(b) Suppose we have another signal p(t) = cos(t) and fs(t) = f(t)p(t), what is
the Fourier transform of fs(t)?

5. [Nyquist Sampling Theorem] Suppose we have a signal x(t) whose frequency

response is

1 ω < 2 rad/s
X ( jω ) =  (2.84)
0 otherwise
48 Signals and Systems Overview

and assume we have another signal f(t) defined as

f (t) = x( t)cos(2 t). (2.85)

What is the Nyquist sampling frequency of f(t)?

6. [Sampling Rate Conversion] Let us cascade the decimator and the interpo-
lator introduced in Section 2.3.4 to perform a sampling rate conversion by
factor 3/4, as shown in Figure 2.24(a), where the input signal is i[n] and the
output signal is o[n]. Suppose the spectrum of the input signal is shown in
Figure 2.24(b).

(a) Design the system by specifying D, I, HD (w) and HI (w).

(b) Plot the spectrum of all the signals in this system, namely x[n], y[n], z[n]
and o[n].

7. [Sampling Rate Conversion] There is a cascaded system shown in Figure

n
æ 1ö
2.25, where the input x[n] = ç ÷ u[n]. Assume the upsampler and down-
è 3ø
sampler are ideal (with ideal low-pass filters) and the overall system response
F(ejw) is LTI. Determine this overall system response, namely, find

Y ( e jω )
F (e jω ) = . (2.86)
X ( e jω )

(a) A sampling rate conversion system.

(b) The spectrum of input signal.

Figure 2.24  Sampling rate conversion problem.

Figure 2.25  A cascaded system, with input signal x [n] and output signal y [n].
2.7  PROBLEMS 49

8. [Z Transform] Given a discrete signal

x [ n] = 3δ ( n − 2) + 2δ ( n − k), (2.87)
what is its z transform? What is its region of convergence?
9. [Z Transform] Given a z transform

5z
X (z) = 2
, (2.88)
z − 3z + 2
what is its inverse z transform?
10. [Z Transform] Let y[n] = x[n] * h[n] be the output of a discrete-time sys-
n
æ 1ö
tem, where * is the convolution operator. Let h[ n ] = çè ÷ø u[n ] and
3
( )
ì 1 n n ³ 0,
ï 2
x [ n] = í (2.89)
( )
n
ï 1 n < 0.
î 4
Find Y (z) and specify its region of convergence.
11. [Stable
 System] Suppose a linear time-invariant system is specified by the
following difference equation

y [ n] + 0.2 y[ n − 1] − 0.24 y[ n − 2] = x[ n] + x [ n − 1] (2.90)

(a) What is its frequency response H(z)?
(b) What is its zeros and poles?
(c) Is it a stable system?

12. [Ideal Filter] Suppose we have a system shown in Figure 2.26(a), where the
input signals are
sin(2t )
f (t) = , (2.91)
2π t
and

s(t ) = cos(1000t ). (2.92)

The frequency response of the bandpass filter is shown in Figure 2.26(b).

(a) What is output signal y(t)?
(b) Please draw the frequency response of y(t).

13. [FIR Filter] Suppose we have a three order real-coefficient FIR filter, whose
frequency response is
3
H (e jw) = ∑ h[n]e − jwn (2.93)
n =0
æ jπ ö
Given H(e ) = 2, H ç e 2 ÷ 7 - j 3 and H(ejp) = 0, what is H (z)?
j0
è ø
14. [FIR Filter] Design a second-order FIR filter whose frequency response has
π
a zero magnitude at w = . Specify the design using time-domain difference
2
equation.
50 Signals and Systems Overview

(a) Block diagram of the system.

- -

(b) Frequency response of the bandpass filter.

Figure 2.26  Ideal filter problem.

15. [Linear Filter] Draw a direct form realization of the linear filter with the fol-
lowing impulse response:

h[n] = {−2,5,8,3, −3}. (2.94)

References

[1]   Proakis, and J. Dimitris, G. Manolakis, Digital Signal Processing:Principles, Algorithms,

and Applications, 3rd edition, Prentice Hall, 1996.
[2]   Chang, Academic Genealogy of Mathematicians, World Scientific, 2011.
[3]   Joseph Fourier, http://en.wikipedia.org/wiki/Joseph_Fourier.
[4]   Fulop, S. A., Speech Spectrum Analysis. New York: Springer-Verlag, LC, 2011.
[5]   Cavicchi, T. J., Digital Signal Processing, John Wiley Sons, 2000.
[6]   Elali, T. S., and M. A. Karim, Continuous Signals and Systems with MATLAB, CRC
Press, 2001.
[7]   James, J. F., A Student’s Guide to Fourier Transforms: With Applications in Physics and
Engineering, 3rd edition, Cambridge University Press, 2011.
[8]   Johnson, C. R. and W. A. Sethares, Telecommunications Breakdown: Concepts of Com-
munication Transmitted via Software-Defined Radio, Prentice Hall, 2004.
[9]   Johnson C. R., W. A. Sethares, and A. G. Klein, Software Receiver Design: Build Your Own
Digital Communications System in Five Easy Steps, Cambridge University Press, 2011.
[10] Wyglinski, A. M., Nekovee and Y. T. Hou, Cognitive Radio Communications and Net-
works: Principles and Practice, Academic Press, 2009.
[11] Barry J., E. A. Lee, and D. G. Messerschmitt, Digital Communication, 3rd edition, Kluwer
Academic Press, 2004.
[12] Alan Jeffrey and Daniel Zwillinger, Table of Integrals, Series, and Products, 7th edition,
Academic Press, 2007.
2.7  PROBLEMS 51

[13] Hogenauer, E. B., “An Economical Class of Digital Filters for Decimation and Interpola-
tion” IEEE Transactions on Acoustics, Speech and Signal Processing, Vol. 29, No. 2, 1981,
pp. 155–162.
[14] Lyons, G., Streamlining Digital Signal Processing: A Tricks of the Trade Guidebook,
Wiley-IEEE Press, 2007.
 ­­C h a p t e r 3

Probability Review

We live in a world that possesses a significant degree of randomness. Whether it is the

final scores for a sporting event, the precise arrival time of a friend, or the next time that
your cellular phone rings, there exists a wide range of events occurring throughout our
daily lives that have some level of uncertainty associated with them. However, despite
the fact that these and many other phenomena appear “random” at first glance, it is
possible to extract a quantitative description of the behavior for these occurrences and
precisely characterize them using mathematical models. These models are referred to as
random variables for static events and random processes for time-varying events. In this
chapter, we will present the fundamental mathematical tools needed to understand, cre-
ate, and analyze these types of models within the context of wireless data transmission.

3.1  Fundamental Concepts

To accurately quantify the behavior of a random variable or a random process, it is

necessary to understand an assortment of fundamental mathematical tools that are
needed in order to develop these models. One of these tools is the concept of rela-
tive frequency, which is used to provide a basic metric regarding the rate at which a
specific outcome resulting from a random phenomenon occurs. Suppose we observe
the outcome of M trials of a random phenomenon, such as the tossing of a fair, two-
sided coin. Labeling each outcome as O1, O2, . . ., Om, what we are interested in is
the number of times a “heads” occur out of the total number of M coin tosses.
Defining Nn (Oi) as the number of trials for which the Oi outcome occurred over a
total of n trials, we can specify the relative frequency Nn (Oi) / n. Furthermore, in order
to obtain the probability measure or probability of an event having an outcome Oi, one
would take the limit of the relative frequency as it approaches infinity, namely:
Nn (Oi )
Pi = lim (3.1)
n ®¥ n
Throughout the rest of this chapter and the this book, we will be using the prob-
ability measure to characterize the likelihood that a specific outcome will occur.

3.1.1 Set Theory
When dealing with the outcomes of random events and phenomena, one useful
mathematical tool is set theory. Suppose we define the sample space as the set of
all possible, distinct, indecomposable measurements that could be observed. The
sample space is usually denoted as W and can either be continuous (e.g., W = [0, ¥))
or discrete (e.g., W = {0, 1, 2, . . .}). An outcome is an element of a point located within
the sample space W, while an event is a collection of outcomes.
53
54 Probability Review

The idea of sample spaces, outcomes, and events can be described using set
theory. For example, suppose we are given A and B as a collection of points, and w
is a point in W; then:

1. w Î W: “w is an element of W.”
2. A Ì B: “A is a subset of B.”
3. A = B: “A is an equal set to B.”
4. A Ì B and A ¹ B: “A is a proper subset of B.”

Graphically representing these relationships can be readily accomplished using

Venn diagrams. An example of a Venn diagram for the case of A Ì B is shown in
Figure 3.1. Furthermore, a collection of commonly used set identities are summa-
rized in Table 3.1, which can be generalized into:
æ ¥ ö ¥ æ ¥ ö ¥
B Ç çç ∪ An ÷÷ = ∪ (B Ç An ) and B È çç ∩ An ÷÷ = ∩ (B È An ) (3.2)
ç ÷ ç ÷
è n =1 ø n =1 è n =1 ø n =1

and
C C
 ∞  ∞  ∞  ∞




∩ 
An 

= ∪ AC
n and 


∪ A n 

= ∩ AnC (3.3)
 n =1  n =1  n =1  n =1

Finally, a summary of possible set operations and their associated Venn diagram
representations are shown in Table 3.2.

3.1.2 Partitions
Partitioning of a sample space is a very powerful tool in set theory since it allows
an individual to “divide and conquer” all the possible events into mutually exclu-
sive, complementary subsets. Suppose we have a family of nonempty sets Bn that
are pairwise disjoint and form partitions of the entire sample space W. as shown in
Figure 3.2(a), where the sets satisfy:

Bn ∩ Bm = ∅, ∀ n, m and n ≠ m,
N
∪ Bn = Ω (3.4)
n =1

Figure 3.1  Illustration of a Venn diagram representing A Ì B.

3.1  Fundamental Concepts 55

Table 3.1  Commonly Used Set Identities

Identity Name Definition
Commutativity A È B = B È A and A Ç B = B Ç A
Associativity A È (B È C) = (A È B) È C and A Ç (B Ç C) = (A Ç B) Ç C
Distributivity A Ç (B È C) = (A Ç B) È (A Ç C) and A È (B Ç C) = (A È B) Ç (A È C)
DeMorgan’s Law (A Ç B)c = Ac È Bc and (A È B)c = Ac Ç Bc

Consequently, we can carve up some set A using these partitions, such that:

A = A∩ Ω
 N 
= A ∩  ∪ Bn 
 
 n =1 

N
= ∪ (A ∩ Bn) (3.5)
n =1

This partitioning of A is shown in Figure 3.2(b), where A is carved up into A Ç

B1, A Ç B2, and A Ç B3. As we will see later in this chapter, the ability to partition
sets into several disjoint subsets will be very helpful in developing additional math-
ematical tools.

Table 3.2  Collection of Fundamental Set Operations

Operation Definition Venn Diagram
Complement A := {w Î W: w Ï A}
c

A
Ac

Empty set or null set Wc = Æ

W
Union A È B := {w Î W: w Î A or w Î B}
B A
W
Intersection A Ç B := {w Î W: w Î A and w Î B}
B A
W

Set difference B \ A := B Ç Ac
B
A
W

Disjoint set AÇB=f

or mutually exclusive B
A
W
56 Probability Review

(a)

(b)
Figure 3.2  An illustration of the partitioning of a set A into several disjoint subsets based on the
disjoint sets {Bn}. (a) Sample space W partitioned into disjoint sets {Bn}. (b) Set A partitioned by dis-
joint sets {Bn}.

3.1.3  Functions
By definition, functions are used to transform or map various values into another
quantitative representation. This mapping process can be readily represented in
Figure 3.3, where the values of x located within the domain X are mapped by a
function f to a collection of values y contained within a specific range, which forms
part of a potentially larger co-domain Y.
Notice how the range contains all possible values of f(x), where {f(x): x Î X}.
Furthermore, the range is a subset of the co-domain Y, and under specific circum-
stances the range and co-domain are one and in the same set of values. Whenever
dealing with functions, there is a collection of associated terminology describing the
mapping operations by functions such as f, such as:

• Onto: A function is “onto” if its range equals its co-domain such that for
every y ÎY, f(x) = y has a solution.
• One-to-one: The condition f(x1) = f(x2) implies x1 = x2.
• Invertible: For every y Î Y, there is a unique x Î X with f(x) = y, which means
that the function is both onto and one-to-one, and that for every y Î Y and
f(x) = y has a unique solution.

Range

Domain Co-domain
Figure 3.3  An illustration of how a function maps values from the domain to a range contained
within a co-domain.
3.1  Fundamental Concepts 57

• Inverse image: If we have f : X ® Y and if B Ì Y, then the inverse image of B

is f –1(B) := {x Î X : f(x) Î B}.

3.1.4 Axioms and Properties of Probability

Using the mathematical tools that we have developed thus far, it is now time to
establish a fundamental set of rules for probability theory. Given a nonempty set W
called the sample space, and a function P defined on the subset of W, we say P is a
probability measure if the following four axioms are satisfied:

1. The empty set f is called the impossible event. Moreover, P(f) = 0.

2. Probabilities are nonnegative (i.e., P(A) ³ 0, "A).
3. If A1,A2,. . . are events that are mutually exclusive or pairwise disjoint (i.e.,
 ∞  ∞
An Ç Am = f for n ¹ m, then P  ∪ An  = ∑ P(An ).
 n= 1  n= 1
4. The sure event or certain event occurs when finding the probability meas-
ure of the entire sample space W (i.e., P(W) = 1). Furthermore, the almost-
sure event occurs when P(A) = 1 where A ¹ W.

Several useful properties based on several manipulations of the Axioms of Prob-

ability are presented in Table 3.3.

Q Using the Axioms of Probability, verify that the properties presented

in Table 3.3 are indeed true.

3.1.5  Conditional Probability

We are often interested in determining whether two events, say A and B, are related
in the sense that knowledge of one influences the occurrence of the other. Conse-

Table 3.3  Several Common Probability Properties

Property Definition
Finite disjoint union æ N ö N
P çç ∪ An ÷÷ =
ç ÷
å P(An )
è n =1 ø n =1

Probability of a complement P(Ac) = 1 - P(A)

Monotonicity A Ì B ® P(A) £ P(B)
Inclusion-exclusion P(AÈB) = P(A) + P(B) - P(AÇB)
Limit properties æ ¥ ö æ N ö
P çç ∪ An ÷÷ = limN ®¥ P çç ∪ An ÷÷
ç ÷ ç ÷
è n =1 ø è n =1 ø

æ ¥ ö æ N ö
P çç ∩ An ÷÷ = limN ®¥ P çç ∩ An ÷÷
ç ÷ ç ÷
è n =1 ø è n =1 ø
58 Probability Review

quently, we employ the conditional probability mathematical tool in order to deter-

mine the relationship between two events, where this tool is defined as:
P(A ∩ B)
P(A| B) = , for P(B) > 0 (3.6)
P(B)
This definition can be interpreted using the concept of relative frequency that was
discussed earlier in this chapter. Suppose that an experiment is performed n times,
and that the event B occurs nB times while the event A occurs nA times and the event
A Ç B occurs nAÇB times. Then, the ratio of the number of times that A and B occur
given the number of times B occurs is defined as:

nA ∩ B nA ∩ B/n nA ∩ B/n P(A ∩ B)

= ∴ lim = (3.7)
nB n B /n n →∞ n B/n P( B)
In other words, the number of times that B occurs (i.e., nB), will normalize the
number of times that A Ç B occurs. As we will see next, the concept of conditional
probability will be very useful in developing some of the key concepts for prob-
ability theory.

3.1.6 Law of Total Probability and Bayes’ Rule

Recall from Section 3.1.2 how we could partition a sample space into a collection of
mutually exclusive events B1, B2, . . ., Bn. Let us proceed with partitioning the event
A using this collection, such that:
A = A ∩ (B1 ∪ B2 ∪ B3 ∪ � ∪ Bn ),
= (A ∩ B1) ∪ (A ∩ B2 ) ∪ � ∪ (A ∩ Bn ) (3.8)

Now taking the definition for the conditional probability and isolating for P (AÇBi),
we get the following:

P(A ∩ Bi ) = P(A|Bi )P(Bi ) (3.9)

which becomes the definition for the Law of Total Probability, namely:

P(A) = P(A|B1) P(B1) + P(A|B2 )P(B2 ) + � + P(A | Bn )P(Bn ) (3.10)

From this expression, suppose that B1,B2, . . ., Bn constitute partitioned components

of the sample space W and that we have defined an event A. We can now define the
probability of event Bi using the following expression:

P(A ∩ Bi ) P( A ∩ Bi )
P(Bi | A) = =
P(A) n (3.11)
∑ P(A|Bk )P(Bk)
k =1

which is referred to as Bayes’ Rule. With the Law of Total Probability and Bayes’
Rule now defined, it is possible to solve complex probability models within the
context of a set theory framework using these tools and their variants.
3.2  Random Variables 59

3.1.7  Independence
There are several occasions where a special relationship exists between two or more
events. For instance, if knowledge of an occurrence of an event B does not alter the
probability of some other event A, we can say that A is independent of B; that is,

P ( A ∩ B)
P ( A) = P(A |B) = (3.12)
P ( B)

By definition, two events A and B are considered to be independent when:

P(A ∩ B) = P(A)P(B) (3.13)

In general, when we have events Ai, i = 1, 2, . . ., we say that they are mutually in-
dependent when there are a finite number of events and satisfy:
 
P  ∩ Ai  = ∏ P(Ai ) (3.14)
 i∈I  i∈I

Note that if (3.14) holds for all subsets of I containing exactly two events but not
necessarily three or more, we say that the Ai are pairwise independent.

Q Prove the definition for independence, where P(A Ç B) = P(A)P(B).

3.2  Random Variables

In science and engineering, measurements and observations are expressed as nu-

merical quantities. These tend to have a degree of uncertainty, and we refer to this
uncertainty variability as a random variable. A random variable is a number as-
signed to every outcome of an experiment. This number could be the gain in a game
of chance, the voltage of a random source, the cost of a random component, or any
other numerical quantity that is of interest in the performance of the experiment.
A random variable is a mapping function whose domain is a sample space and
whose range is some set of real numbers:

X = X (ω ) (3.15)
where X is the random variable, and w is the outcome of an experiment found in the
sample space W. An example of the mapping process relating some experiment outcome
w by the random variable X to some real number value is shown in Figure 3.4.
Note that the sample space W is the domain of the function X(W), while Wx is
the range and the co-domain is .
Given this mapping function, let us now develop a notation for defining the
probability of an event or an outcome. First, we will let upper case letters denote
random variables, while lower case letters will denote the possible values of the
60 Probability Review

Figure 3.4  An example of a random variable X mapping an experimental outcome w to a real

number value.

random variables. Second, let us define the following expression for the probability
of a specific event, as well as its shorthand representation on the right-hand side:

P({w ∈ Ω : X(w ) < x}) → P(X < x) (3.16)

where P(.) denotes the probability measure, w Î W denotes the outcomes are ele-
ments of the sample space W, and X(w) denotes the mapping function that produces
an output value less than a deterministic quantity x. On the right-hand side, the
shorthand version is interpreted as “the probability that the random variable X will
produce an output that is less than the deterministic quantity x.”
Finally, some additional shorthand expressions and their equivalent full math-
ematical representations include the scenario where if B is a set of real numbers,
then the probability that X produces an output belonging to that set is given by:

P(X ∈ B) = P({X ∈ B}) = P({ω ∈ Ω : X(ω) ∈ B}) (3.17)

Another expression is when B is an interval such that B = (a, b], the probability that
the output of the random variable X is given by:

P(a < X ≤ b) = P({ω ∈ Ω : a < X(ω) ≤ b}) (3.18)

3.2.1 Discrete Random Variables

The first type of random variable we will be studying is discrete random variables.
Suppose that X is a discrete random variable, in which case there exists distinct real
numbers xi such that:

∑ P(X = xi ) = 1 (3.19)
i

Using the Law of Total Probability that we have studied in Section 3.1.6, we know that:

P(X ∈ B) = ∑ P(X = xi ) (3.20)

i: xi ∈B

where the set B is composed of a collection of values xi. A specific form of discrete
random variable is die integer-valued random variable, where its output values are
the integers xi = i; that is,
3.2  Random Variables 61

P(X ∈ B) = ∑ P(X = i) (3.21)

i∈B

Since each output of the random variable X possesses a different probability of

occurrence, we usually define the probability of a specific discrete output xi being
generated by X as:

pX (xi ) = P(X = xi ) (3.22)

where pX (xi) is referred to as the probability mass function (PMF). Note that the
values of the PMF are constrained by:

0 ≤ pX (xi ) ≤ 1 and ∑ pX (xi ) = 1 (3.23)

i

Several frequently used PMFs are specified in Table 3.4, including uniform, Poisson,
and Bernoulli random variables.
When dealing with multiple discrete random variables X 1 ,...,X n , the general-
ized definition for all of these random variables to be independent is given as:

 n  n
p ∩ {X j ∈ Bj }  = ∏ P(X j ∈ Bj ),∀ choices of sets B1 , …, Bn (3.24)
 j= 1  j= 1

Note that if X 1 , . . ., X n are independent, then so are any subset of them. Further-
more, if X 1 , . . ., X n , . . . is an infinite sequence of random variables, then they are
independent if the generalized expression holds for every finite n = 1, 2, . . .. Finally,
if for every B the probability P(Xj Î B) does not depend on j, men we say that all
the Xj are identically distributed. In the case when all the Xj are also independent
with each other, we refer to these random variables as independent and identically
distributed (i.i.d.).

Table 3.4  Several Frequently Used Probability Mass Functions (PMFs)

Random Variable PMF Definition Graphical Representation
PX ( x)
ïì ,
1 k = 1, . . ., n
Uniform pX (k) = í n
ïî 0, otherwise
1/n

0 1 2 3 n -1 n x
k -λ
λ e PX ( x)
Poisson pX (k) = k!
, k = 0, 1, 2, . . .

0 1 2 3 x

ì p, k =1 PX ( x)
ï
Bernoulli pX (k) = í1 - p, k=0 P
ï0, otherwise 1- p
î

0 1 2 3 x
62 Probability Review

3.2.1.1  Expectation
One way to characterize a random variable is by its expected value or mean, which
can be quantitatively determined using the expression:

mX = E[X ] = ∑ xi P(X = xi ) (3.25)

i

where the sum of each value is weighted by its probability of occurrence. Conse-
quently, using the definition of the PMF, we can rewrite (3.25) as:

E[X ] = ∑ xi pX (xi ) (3.26)

i

Suppose we have a random variable X and a real-valued function g(x) that maps
it to a new random variable Z (i.e., Z = g(X)). Recall that X is actually the mapping
of points from a sample space W to a collection of real numbers. Therefore, we are
performing a mapping of a mapping when solving for the random variable Z (i.e.,
Z(w) = g(X(w))). Thus, in order to compute the expectation of the random variable
Z, namely E[Z], we can employ the following expression:

E[Z] = E[ g(X)] = ∑ g(xi )pX (xi ) (3.27)

i

which is referred to the expectation of a function of a random variable. One of the

most important aspects of (3.27) is that the expectation of Z can be solved using
the PMF for X rather than having to determine the PMF of Z. Table 3.5 presents
several useful properties associated with the expectation operator.

E[ X r ]
Derive the Markov Inequality, P(X ³ a) £ , using the definition
Q for the expectation.
ar

3.2.1.2  Conditional Probability and Conditional Expectation

For random variables, we define the conditional probability as:

P(X ∈ B|Y ∈ C) = P({X ∈ B}|{Y ∈ C })

P({ X ∈ B} ∩ {Y ∈ C}) P(X ∈ B,Y ∈ C) (3.28)

= =
P({Y ∈ C}) P(Y ∈ C)

Furthermore, we can define the conditional PMF as:

P(X = xi , Y = yj ) pX Y(xi , yj )
p X |Y (x i | y j) = P (X = x i |Y = y j ) = =
P(Y = yj ) pY ( yj )

P( X = xi , Y = yj ) pX Y(x i , yj ) (3.29)
pY | X (yj | xi ) = P ( Y = yj | X = xi) = =
P ( X = xi ) pX ( x i)
3.2  Random Variables 63

Table 3.5  Several Useful Properties of Expectation

Name Definition
Linearity If a and b are deterministic constants and X and Y are random variables, then E[aX + bY] =
E[aX] + E[bY] = aE[X] + bE[Y].
Moment The nth moment (n ³ 1) of a real-valued random variable X is defined as the expected value
of X raised to the nth power (i.e., Momentn (X) = E[Xn]).
Mean The mean is the first moment of X ( i.e., Moment1 (X) = E[X] = m).
Central The second moment of X with its mean subtracted is its variance (i.e., Moment2(X − m) =
Moment E[(X − m)2] = var(X) = s 2).
Correlation The correlation between two random variables X and Y is defined to be equal to E[XY].
Covariance The covariance between X and Y is defined as cov(X, Y)
Correlation
coefficient
The correlation coefficient of two random variables X and Y is given as ρ XY = E ê σ X
ë (
é X - µX Y - µY
σY )( )ùúû.
E[ X ]
Markov If X is a nonnegative random variable, then for any a > 0 we have P(X ≥ α ) ≤ a
.
inequality
Chebyshev E[Y 2 ]
For any random variable Y and any a > 0, we have P(|Y |≥ α ) ≤
a2
inequality
Cauchy- The Cauchy-Schwarx Inequality states that | E[XY ] | £ E[X 2]E[Y 2 ] which becomes an equal-
Schwarz ity if and only if X and Y are linearly related.
inequality
Independence The random variables are independent if and only if E[h(X)k(Y)] = E[h(X)]E[k(Y)] for all
functions h(x) and k(y).

which can then be expressed as:

p X Y (x i , y j ) = p X | Y(x i | yj ) pY( y j ) = pY | X( yj | x i ) pX( x i) (3.30)

If Y is an arbitrary random variable, and we take A = {Y Î C}, where C Ì ,

then we define the Law of Total Probability as:

P(Y ∈ C) = ∑ P(Y ∈ C| X = xi )P(X = xi ) (3.31)

i

If Y is a discrete random variable, taking on distinct values yi and setting C = {yi}

yields:
P(Y = yj ) = ∑ P(Y = yj | X = x i )P( X = xi)
i

= ∑ pY | X ( y j) pX( x i) (3.32)
i

Often we encounter the situation where some random variable Z is a function of

X and some other discrete random variable Y, say, Z = g(X, Y). However, we are inter-
ested in determining P(Z = z). We know that by the Law of Total Probability that:

P( Z = z) = ∑ P(Z = z |X = xi )P(X = xi )
i

= ∑ P(g (X , Y ) = z | X = xi )P( X = xi ) (3.33)

i
64 Probability Review

Looking at the resulting equation more closely, we notice that since we already have
the entire expression conditioned on the fact that the random variable X produces
a deterministic output xi, namely, X = xi, we can substitute the random variable X
with the deterministic output xi such that:

P(g (X , Y ) = z | X = xi ) = P(g(x i , Y ) = z |X = xi ) (3.34)

This operation is referred to as the Substitution Law, and it can be used to help
solve expressions and operations involving multiple random variables.
Similarly, the conditional expectation is given as:

E [ g(Y )| X = xi ] = ∑ g ( yj ) pY | X( yj | xi ) (3.35)
j

the Substitution Law for Conditional Expectation is given as:

E [ g(X , Y )| X = xi ] = E[ g( xi ,Y )| X = xi ] (3.36)

and the Law of Total Probability for Expectation is defined as:

E [ g(X ,Y )] = ∑ E [ g( X, Y )| X = xi ] pX ( xi) (3.37)

i

Derive the resulting expressions for the Law of Total Probability for
Q Expectation and the Substitution Law.

3.2.1.3  A Communications Example: Binary Channel Receiver Design

In digital communications and wireless data transmission, we can view the opera-
tions of receivers as attempting to “guess” what was sent for a specific transmitter.
When guessing the value of a transmission whose values can be modeled as a ran-
dom variable X, and once it has been observed at the receiver whose values can be
modeled by a random variable Y, we need some sort of decision rule to figure out
which value of X was originally transmitted. This is illustrated in Figure 3.5, where
X can output values of either 0 or 1 while Y is observed at the receiver. Notice how
the observation Y can be accidentally interpreted as a value other than the one
transmitted (i.e., X ¹ Y).
Although there exist several approaches for deciding on what values were trans-
mitted based on the observation of the intercepted received signal, no decision rule
possesses a smaller probability of error than the maximum a posteriori (MAP) rule,
where having observed Y = j, the MAP rule says to decide on X = 1 if:

P( X = 1| Y = j) ≥ P( X = 0|Y = j ) (3.38)

and to decide X = 0 otherwise. In other words, the MAP rule decides X = 1 if

the posterior probability of X = 1 given the observation Y = j is greater than the
3.2  Random Variables 65

Figure 3.5  An example of a binary channel where transmission values generated by the random
variable X are being observed at the receiver as the random variable Y.

posterior probability of X = 0 given the observation Y = j. Furthermore, we can

observe that:
P(X = i, Y = j ) P (Y = j |X = i)P( X = i)
P(X = i |Y = j ) = = (3.39)
P(Y = j) P(Y = j)
which we can then use to rewrite (3.38) as:
P(Y = j | X = 1)P( X = 1) P(Y = j |X = 0)P( X = 0)
≥
P(Y = j) P (Y = j)
P(Y = j | X = 1)P(X = 1) ≥ P(Y = j |X = 0)P(X = 0) (3.40)

If X = 0 and X = 1 are equally likely to occur, we can simplify this expression such
that it yields the maximum likelihood (ML) rule:

P(Y = j |X = 1) ≥ P(Y = j | X = 0) (3.41)

Furthermore, in the general case, we can rearrange (3.40) such that it yields the
likelihood ratio, namely:

P(Y = j |X = 1) P( X = 0) (3.42)
≥
P(Y = j |X = 0) P(X = 1)
where the right-hand side is referred to as the “threshold” since it does not depend on j.

3.2.2  Continuous Random Variables

So far we have looked at random variables that assume only a discrete set of values.
We will now focus our attention on random variables that can produce a continuum
of output values, which we will refer to as continuous random variables. One of the
primary reasons continuous random variables are important to study is that they
can readily model numerous random phenomena that occur in the real world.
However, before moving forward, it must be noted that some of the language
used to describe discrete random variables will be slightly different when talking
about continuous random variables. For example, instead of referring to the prob-
ability mass function of a random variable, we will call them probability density
functions due to the fact that the possible outputs of a continuous random variable
can assume any infinitesimal value rather than some discrete quantity.
66 Probability Review

3.2.2.1  Densities and Probabilities

As we have seen earlier in this chapter, it is possible to mathematically compute the
probability of an event occurring when employing discrete random variables, namely:
b− 1
P(a ≤ X < b) = ∑ pX (i) (3.43)
i= a

where X is the discrete random variable and both a and b are boundaries of a subset
belonging to the sample space W. However, suppose now that X is a continuous
random variable that can take on the entire continuum of values within the interval
[a, b). In order to compute the same probability in this case. we can start by real-
izing this scenario as the summation of an infinite number of points in [a, b) with
the space between samples equal to Dx.
Given an infinite number of samples, our Dx becomes so tiny that Dx ultimately
converges to dx and our summation becomes an integral expression. Therefore, we
can rewrite (3.43) in terms of continuous random variables, yielding:
b
P(a ≤ X < b) = ∫ f (t)dt (3.44)
a
where f(t) is called the probability density function or PDF. Note that the PDF is the
continuous version of the PMF that we discussed previously in this chapter. More-
over, generally, we can express the probability of a continuous random variable
using the PDF by the expression:
+∞
P(X ∈ B) = ∫ f (t)dt = ∫ IB(t)f (t)dt (3.45)
B −∞

where Ib(t) is an indicator function that produces an output of unity whenever a

value of t belongs to the set B and produces an output of zero otherwise. Note that
+¥

ò f (t)dt = 1, f (t) is nonnegative, f(t) approximately provides the probability at a

-¥
point t, and:
P(a ≤ X ≤ b) = P(a < X ≤ b) = P(a ≤ X < b) = P(a < X < b) (3.46)
where the endpoints do not affect the resulting probability measure. Finally, a sum-
mary of several commonly used PDFs are presented in Table 3.6, including uniform,
Gaussian, and exponential random variables.

3.2.2.2  Univariate Expectation

Similar to the expectation of a single discrete random variable, the expectation for
a continuous random variable X with PDF f (x) can be computed using the follow-
ing expression:
+∞
E [ g(X)] = ∫ g( x) f ( x) dx (3.47)
−∞

where g(.) is some real function that is applied to the random variable X.
3.2  Random Variables 67

Table 3.6  Several Frequently Used Probability Density Functions

Random Variable PMF Definition Graphical Representation
f ( x)
1/ (b - a)
 1 , a≤ x≤ b
Uniform f (x) =  (b− a)
 0, otherwise
x
0 a b
f ( x)

 λ e ,− λx
x≥0 l
Exponential f (x) = 
 0, x< 0
x
0
f ( x)

f (x) = λ e− λ | x |
Laplace 2 l /2

x
0

f ( x)

f (x) = λ/ π ,λ >0 1 / pl
Cauchy λ 2 + x2
1 / 2pl
x
-l 0 l

f ( x)
2
Gaussian f (x) = 1 e−0.5((x− µ )/σ )
2π σ2

x
m

3.2.2.3  Conditional Probability and Conditional Expectation

Many random variables of practical interest are not independent, where it is often
the case that the outcome of one event many depend or be influenced by the result
of another event. Consequently, it is sometimes necessary to compute the condi-
tional probability and conditional expectation, especially in circumstances where
we have to deal with problems involving more than one random variable.
Unlike the conditional probability for a discrete random variable, the condi-
tional probability for a continuous random variable needs to defined in an alter-
native manner since the probability of a single exact point occurring is zero (i.e.,
P(X = x) = 0). As a result, if we employ the definition for the conditional probability
for a discrete random variable from (3.28), namely:

P(Y ∈ C , X = x)
P(Y ∈ C | X = x) = (3.48)
P( X = x)
we observe that the occurrence of P(X = x) = 0 would yield a “divide-by-zero” sce-
nario. Consequently, we need to determine another definition for the conditional
68 Probability Review

probability (and conditional expectation) that would work within the continuous
random variable framework.
It can be shown that in order to calculate the conditional probability, one must
employ a conditional density [1], which can be defined as:

fX Y (x, y)
fY | X (y |x) = (3.49)
f X ( x)

where fx(x) > 0. Thus, leveraging the conditional density, we can now compute the
conditional probability without concern for a “divide-by-zero” scenario by solving
for the following expression:

P(Y ∈ C | X = x) = ∫ fY | X (y | x) dy (3.50)
C

Furthermore, we can define the Law of Total Probability as the following:

+∞
P(Y ∈ C ) = ∫ P(Y ∈ C | X = x)fX (x) dx (3.51)
−∞

where we weigh all the conditional probabilities by the PDF of X before integrating
them together to form the overall probability of the event Y Î C. Finally, just as in
the discrete random variable case, we can employ a form of Substitution Law for
continuous random variables when dealing with conditional probability, which is
defined by:

P((X , Y ) ∈ A| X = x) = P((x, Y ) ∈ A| X = x) (3.52)

Note that if X and Y are independent, then the joint density factors, yielding the
following expression for the conditional density:

fX Y (x, y)
fY | X (y| x) =
f X ( x)
fX (x) fY ( y)
=
f X ( x)

= fY (y) (3.53)

which implies that when the two random variables are independent, we do not need
to condition one event on the other.
Similarly, the conditional expectation when dealing with continuous random
variables is defined as the following expression employing the conditional density,
namely:
3.2  Random Variables 69

+∞
E [ g(Y )| X = x] = ∫ g ( y)fY | X (y| x) dy (3.54)
−∞

Furthermore, the Law of Total Probability for a conditional expectation is given as:

+∞
E [ g(X , Y )] = ∫ E [ g(X ,Y )| X = x]fX ( x) dx (3.55)
−∞

and the Substitution Law for a conditional expectation is defined as:

E [ g (X , Y )| X = x] = E [g(x , Y )| X = x] (3.56)

3.2.3  Cumulative Distribution Functions

For both PDFs and PMFs, it is sometimes important to visualize the cumulative distri-
bution function (CDF), especially since it provides another perspective on how the ran-
dom variable behaves probabilistically. Furthermore, the CDF can sometimes be used
to solve problems that would otherwise be difficult to access via some other definition.
Mathematically speaking, we can define the CDF by the following expression:

x
FX (x) = P(X ≤ x) = ∫ f (t)dt (3.57)
−∞

which describes the probability that the outcome of an experiment described by the
random variable X is less than or equal to the dummy variable x.
As an example, suppose that we want to calculate the probability of P(a £ X < b)
using the PDF shown in Figure 3.6(a). One approach to quickly evaluating this
probability is to leverage the tail probabilities of this distribution, namely, P(X < a)
(shown in Figure 3.6(b)) and P(X < b) (shown in Figure 3.6(c)). Notice how the tail
probabilities are actually the CDFs of X based on (3.57), where FX(a) = P(X < a)
and FX(b) = P(X < b).
Consequently, given that we are only interested in the region of the PDF where these
two tail probabilities do not intersect, we can compute the following probability:

P(a ≤ X < b) = P(X < b) − P(X < a) = FX (b) − FX (a) (3.58)

where all we really need are the values for the CDF of X at x = a and x = b.
Several fundamental characteristics of the CDF include the fact that FX (x) is
bounded between zero and one, and that FX(x) is a nondecreasing function (i.e.,
FX(xi) £ FX(x2) if x1 £ x2). Furthermore, the PDF is the derivative of the CDF in
terms of the dummy variable x, which we can define as:

d
fX (x) = FX (x) (3.59)
dx
70 Probability Review

(a)

(b)

(c)
Figure 3.6  An example of how the CDF can be used to obtain the tail probabilities P(X < a) and
P(X < b) in order to quickly calculate P (a < X < b). (a) The region of the PDF of the random variable X
that needs to be integrated in order to yield P (a £ X < b). (b) The region of the PDF of the random
variable X that needs to be integrated in order to yield P (X < a). (c) The region of the PDF of the
random variable X that needs to be integrated in order to yield P (X < b).

The Q function is a convenient way to express right-tail probabili-

ties for Gaussian random variables, P(X > x). Mathematically, this is
equivalent to finding the complementary CDF of X, namely [2]:

Q(x) = 1 − FX (x) = 1 − P(X ≤ x)

i ∞
1 2
= P(X > x) = ∫ e− t /2
dt
2π
x

where FX(x) is the CDF of X.

3.2.4  Central Limit Theorem

There exists several instances where some sort of phenomenon is modeled as the sum-
mation of a collection of independently and identically distributed random variables.
3.2  Random Variables 71

In such cases, we can employ a useful theorem that is capable of helping us define the
overall behavior of the summed random variables. Referred to as the central limit the-
orem (CLT), this theorem states the conditions under which the mean of a sufficiently
large number of independent random variables, each with finite mean and variance,
can be approximated by a Normal (i.e., Gaussian) distribution. Note that the CLT
requires the random variables to be identically distributed. Since real-world quantities
are often the balanced sum of many unobserved random events, this theorem provides
a partial explanation for the prevalence of the normal probability distribution.
Suppose we have X1,X2,X3,..., Xn independent and identically distributed ran-
dom variables with a common mean m and common variance s 2. Let us define the
sum of these random variables by the expression:
n
1  Xi − µ 

Yn =
n
∑  σ 
 (3.60)
i= 1

where the random variable Xi is normalized by the standard deviation s and its
mean shifted to zero by m. Consequently, if we take the limit of n®¥, then by the
CLT the CDF of Yn becomes:
−∞
1 2
lim FYn (y) = ∫ e− t /2
dt (3.61)
n→∞ 2π y

where the integrand is the definition for the standard normal PDF. The reason why
this result for the CLT is so important is that if we take a collection of independent
and identically distributed random variables possessing any distribution and sum
them all together, we will obtain a resulting distribution that is Gaussian.

Given that a binomial random variable is formed by summing n in-

dependently and identically distributed Bernoulli random variables,
Q show for a “large” value of n that the binomial PDF begins to appear
similar to a Gaussian PDF by using die hist command in MAILAB.
Explain this result using the CLT.

3.2.5 The Bivariate Normal

As mentioned previously, there are several random phenomena occurring frequently
in the real world that can be characterize by a Gaussian (i.e., Normal) distribution.
For example, a univariate Gaussian distribution possesses a PDF that is defined by:

1 2
/ 2σ 2
fX ( x) = e- (x- µ) (3.62)
2πσ 2
where m is the mean of the Gaussian random variable X, and s 2 is the variance of X. In
the case that m = 0 and s 2 = 1, we refer to X as a standard Normal random variable.
Although the univariate Normal distribution occurs rather frequently in the
real world, there are several instances where a bivariate Normal distribution could
also be defined to characterize a specific distribution involving two Gaussian
72 Probability Review

random variables possessing some degree of correlation between each other. An illus-
tration of an example of a bivariate Normal distribution is shown in Figure 3.7. To
visualize the impact of correlation between two Gaussian random variables forming
a bivariate Normal distribution, let us examine the examples shown in Figure 3.8
and Figure 3.9, where the zero-mean Gaussian random variables X and Y produce
10,000 coordinate points in two-dimensional space. Since these random variables
are zero-mean, the density of the resulting scatter plot is mostly concentrated around
the origin at coordinates (0, 0). We can see in Figure 3.8 that since X and Y are
uncorrelated, the resulting density of coordinate outputs are circularly symmetrical
around the means of the two random variables. On the other hand, if X and Y are
correlated, the resulting density of output coordinates are heavily skewed along an
axis within the two-dimensional plane, as shown in Figure 3.9.
Mathematically speaking, the general definition for a bivariate Gaussian den-
sity with parameters mX, mY, σ X2 , 2, and correlation coefficient r is given by:
σY

æ æ
( ) ( ) ( ) + ( ) ö÷ø ö÷ø
y - µY 2
2 y - µY
x - µX x - µX
exp ç -1 2 ç σX
- 2ρ σX σY σY
è 2(1- ρ ) è
fXY (x, y) =
2πσ Xσ Y 1 - ρ 2 (3.63)

where the correlation coefficient is defined as:

é æ x - µ X ö æ y - µY ö ù
ρ = E êç ÷ç ÷ ú (3.64)
ëè σ X ø è σ Y ø û

3.3  Random Processes

As we have seen thus far, a random variable is simply a rule for assigning to every
possible outcome w from a sample space W a number X(w). Given this description,

Figure 3.7  Density of a bivariate Normal distribution with no correlation (i.e., r = 0).
3.3  Random Processes 73

Figure 3.8  Scatter plot of a bivariate Normal distribution with no correlation (i.e., r = 0).

we will now provide a definition for a random process, which is also referred to as a
stochastic process. A random process is a family of time-domain functions depending
on the parameters t and w (i.e., we can define a random process by the function).

X(t) = X(t , ω ) (3.65)

where the left-hand side is the shortened representation of a random process that
implicitly assumes the existence of a mapping of an outcome w from the sample
space W to a real-valued number. Note that the domain of w is W while the domain
of t is either  for continuous-time random processes or Z for discrete-time random
processes.

Figure 3.9  Scatter plot of a bivariate Normal distribution with a substantial amount of correlation
between Gaussian random variables X and Y.
74 Probability Review

Given this additional degree of freedom, namely, the time parameter t, we should
ask ourselves how we should interpret the behavior of a random process. Referring
to the random process illustration shown in Figure 3.10, we see that a random pro-
cess X(t,w) consists of a collection of deterministic time-domain functions, each of
which corresponds to an outcome wi from the sample space W. The “randomness”
associated with the random process is derived from the fact that at every time in-
stant, the real-valued output of the random process is generated by the completely
random selection of one of the time-domain functions at that specific time instant.
Furthermore, notice how the selection of the same deterministic time-domain func-
tion corresponding to the same outcome at two different time instants, say t0 and
t1, could yield a real-valued output that is different. In general, we can interpret the
random process X(t,w) as follows:

1. A family or ensemble of functions X(t,w) when t and w are variables;

2. A single time function when t is a variable and w is fixed;
3. A random variable when t is fixed and w is a variable;
4. A number when t and w are fixed.

3.3.1 Statistical Characteristics of Random Processes

Suppose we have a random process that is noncountable infinite for each time
instant t. Given this situation, we can define its first-order distribution F(x, t) and
first-order density f(x, t) as:

F(x, t) = P(X(t) £ x) (3.66)

and
∂F(x, t)
f (x, t) = (3.67)
∂x
When studying a random process at two time instants, say, t1 and t2, this is simi-
lar to studying two random variables, namely, X(t1,w) and X(t2,w). Therefore, the

Figure 3.10  An illustration of a random process X(t,w).

3.3  Random Processes 75

second-order distribution F(x1, x2; t1, t2) and second-order density f(x1, x2; t1, t2)
are defined as:

F(x1 , x2 ; t1 , t2 ) = P(X(t1) ≤ x1 , X(t2 ) ≤ x2 ) (3.68)

and
∂ 2F(x1 , x2 ; t1 , t2 )
f (x1 , x2 ; t1 , t2 ) = (3.69)
∂x1∂x2
Finally, if we extend these definitions to the nth-order distribution F(x1, . . ., xn;
t1,...,tn) and nth-order density f( x1, . . .,xn; t1, . . .,tn), we get:

F(x1 , …, xn ; t1 , …, tn ) = P(X(t1) ≤ x1 , …, X(tn ) ≤ xn ) (3.70)

and
¶n F(x1 , …, xn ; t1 , …, tn )
f (x1 , …, xn ; t1 , …, tn ) = (3.71)
¶x1 …¶xn
For determining the statistical properties of a random process, knowledge from
the function F(x1, . . .,xn; t1, . . .,tn) is required. However, in most applications only
certain averages are actually needed. For instance, one of the mostly commonly used
statistical characterizations for a random process is the mean function, where the
mean function mx(t) of a random process X(t,w) is the expectation of the random
process at a specific time instant t. This can be mathematically represented by:

µ X (t ) = E [ X(t, ω )] (3.72)
Another useful statistical characterization tool for a random process X(t,w) is
the autocorrelation function RXX(t1,t2), where we evaluate the amount of correla-
tion that the random process X(t,w) possesses at two different time instants t1 and
t2. We can define this mathematically by the expression:
RXX (t1 , t2 ) = E [X(t1 , ω )X∗ (t2 , ω)]
+∞ +∞
= ∫ ∫ x1x∗2 f (x1 , x2 ; t1 , t2 )dx1dx2 (3.73)
−∞ −∞

Note that the value of the diagonal for RXX(t1,t2) is the average power of X(t,w),
namely:

E [|X (t , ω)|2 ] = RX X(t, t) (3.74)

Several other useful properties and observations about the autocorrelation function
include:

1. Since RXX(t1,t2) = E[X(t1,w)X*(t2,w)], then RXX(t2,t1) = E[X(t2,w)X*(t1, w)] =

R*XX(t1,t2).
2. We have RXX(t,t) = E[X(t,w)|2 ] ³ 0.
3. A random process for which E[X(t,w)|2 ] < ¥ for all t is called a second-order
process.
76 Probability Review

4. For RXX(t,t) = E[X(t,w)|2 ] ³ 0 and given time instants t1 and t2, we have the
following inequality:

| RXX (t1 , t2 ) |≤ E [| X(t1 , ω ) | 2 ]E [| X(t2 , ω ) | 2 ]

5. A normalized process occurs when X(t, ω)/ CXX (t, t).

An extension of the definition for the autocorrelation function is the autocova-

riance function CXX(t1,t2 of X(t,w), which is the covariance of the random process
X(t,w) at time instants t1 and t2. Mathematically, we can represent the autocovari-
ance function by the expression:

CXX (t1 , t2 ) = RXX (t1 , t2 ) − µ X (t1)µ ∗X (t2 ) (3.75)

Note that for t1 = t2, the autocovariance function produces the variance of X (t,w).
Furthermore, we can sometimes represent the autocovariance function of a random
process X(t,w) using a normalize metric called the correlation coefficient, which we
define as:
CXX (t1 , t2 )
ρXX (t1 , t2 ) =
CXX (t1 , t1)CXX (t2 , t2 ) (3.76)

3.3.2 Stationarity
A stationary process is a random process that exhibits the same behavior at any
two time instants (i.e., the random process is time invariant). Two common forms
of stationary processes are strict-sense stationary (SSS) random processes and wide-
sense stationary (WSS) random processes. In this section, we will define these two
types of random processes.

3.3.2.1  Strict-Sense Stationarity

A random process is SSS whenever its statistical properties are invariant to a shift
of the origin (i.e., the random process X(t,w) and X(t + c,w) both possess the same
statistics for any time shift c). Therefore, the nth-order density of an SSS random
process would be equal to, by definition, the following expression:

f (x1 , …, xn ; t1 , …, tn ) = f (x1 , …, xn ; t1 + c, …, t n + c) (3.77)

for any time shift c.
It follows that f(x; t) = f(x; t + c) for any time shift c, which means that the first-
order density of X(t,w) is independent of the time t, namely:

f (x; t) = f (x) (3.78)

Furthermore, f(x1, x2; t1, t2) = f(x1, x2; t1 + c, t2 + c) is independent of c for any value
of c. Consequently, this means that the density becomes:

f (x1 , x2 ; t1 , t2 ) = f (x1 , x2 ; τ ), where τ = t1 − t2 (3.79)

3.3  Random Processes 77

Thus, the joint density of the random process at time instants t and t + t is indepen-
dent of t and is equal to f (x1, x2;t).

3.3.2.2  Wide-Sense Stationarity

Although SSS random processes can yield mathematically tractable solutions based
on their useful time-invariant property, the occurrence of SSS random processes in
the real world is not very frequent. As a result, another form of stationarity that
does occur rather frequently in the real world will now be discussed.
A random process X (t,w) is considered to be WSS whenever both of the follow-
ing properties are true:

• The mean function mX(t) does not depend on time t (i.e., mX(t) = E[X(t,w)] =
mX).
• The autocorrelation function RXX(t + t,t) depends only on the relative differ-
ence between t and t + t (i.e., RXX(t + t,t) = Rxx(t)).
Several observations about WSS random processes include the following:
• The average power of a WSS random process is independent of time since
E[|X(t,w)|2] = RXX(0).
• The autocovariance function of a WSS random process is equal to CXX(t) =
RXX (t) – |mX|2.
• The correlation coefficient of a WSS random process is given by rXX(t) =
CXX(t)/CXX(0).
• Two random processes X(t,w) and Y(t,w) are jointly WSS if each is WSS and
their cross-correlation depends on t = t1 – t2.
• If the random process X(t,w) is WSS and uncorrelated, then CXX(t) = qd(t),
where q is some multiplicative constant.

3.3.2.3  Cyclostationarity
There exists another form of stationarity characteristic that often occurs in wire-
less data transmission. A cyclostationary random process Y(t) is defined by a mean
function mY(t) that is periodic across time t as well as an autocorrelation function
Ryy(t + q,q) that is periodic across q for a fixed value of t. Consequently, a cyclosta-
tionary random process Y(t) with period T0 can be described mathematically by:
T0
1
RYY (τ ) =
T0 ∫ RXX ( τ + θ , θ )dθ (3.80)
0

3.3.3  Gaussian Processes

In probability theory and statistics, a Gaussian process is a stochastic process whose
realizations consist of random values associated with every point in a range of times
(or of space) such that each such random variable has a Normal distribution. More-
over, every finite collection of those random variables has a multivariate Normal
distribution.
78 Probability Review

Gaussian processes are important in statistical modeling because of properties in-

herited from the Normal distribution. For example, if a random process is modeled
as a Gaussian process, the distributions of various derived quantities can be obtained
explicitly. Such quantities include the average value of the process over a range of times,
and the error in estimating the average using sample values at a small set of times.
Given the following expression:

T
y = ò g(t)X(t) dt (3.81)
0

we can say that X(t) is a Gaussian process if:

• E(y2) is finite (i.e., does not blow up).

• Y is Gaussian-distributed for every g(t).

Note that the random variable Y has a Gaussian distribution, where its PDF is
defined as:

− (y− µY )2
1 2σ Y 2
fY (y) = e
2πσY2 (3.82)

where mY is the mean and σ Y2 is the variance. Such processes are important because
they closely match the behavior of numerous physical phenomena, such as additive
white Gaussian noise (AWGN).

Why is an uncorrelated random process referred to as “white” such as

Q in the case of additive white Gaussian noise?

3.3.4 Power Spectral Density and LTI Systems

To analyze a signal in the frequency domain, the power spectral density (PSD),
SXX(f ), is often used to characterize the signal, which is obtained by taking the
Fourier transform of the autocorrelation RXX(t ) of the WSS random process X(t).
The PSD and the autocorrelation of a function, RXX(t ), are mathematically related
by the Einstein-Wiener-Khinchin (EWK) relations, namely:

¥
SXX (f ) = ò RXX (τ )e - j 2π f τ dτ (3.83)
-¥

¥
RXX (f ) = ò SXX (τ )e + j 2π f τ df (3.84)
-¥

Using the EWK relations, we can derive several general properties of the PSD of
a stationary process, namely:
3.4  Chapter Summary 79

¥
• SXX (0) = ò RXX (τ )dτ
-¥
∞
• E{X (t)} = ∫ SXX (f )df
2
−∞
• Sxx(f )>0 for all f.
• Sxx(−f ) = Sx(f ).
• The power spectral density, appropriately normalized, has the properties usu-
ally associated with a probability density function:

SXX (f )
px (f ) = ∞
∫ −∞ SXX (f )df (3.85)

A very powerful consequence of the EWK relations is its usefulness when attempt-
ing to determine the autocorrelation function or PSD of a WSS random process that
is the output of a linear time-invariant (LTI) system whose input is also a WSS ran-
dom process. Specifically, suppose we denote H(f ) as the frequency response of an
LTI system h(t). We can then relate the power spectral density of input and output
random processes by the following equation:

SY Y (f ) = |H (f )|2SX X(f ) (3.86)

where SXX(f ) is the PSD of input random process and SYY (f ) is the PSD of output
random process. This very useful relationship is illustrated in Figure 3.11.

3.4  Chapter Summary

In this chapter, a brief introduction to some of the key mathematical tools for ana-
lyzing and modeling random variables and random processes has been presented.
Of particular importance, the reader should understand how to mathematically
manipulate Gaussian random variables, Gaussian random processes, and bivariate
Normal distributions since they frequently occur in digital communications and
wireless data transmission applications. Furthermore, understanding how station-
arity works and how to apply the EWK relations to situations involving random
processes being filtered by LTI systems is vitally important, especially when dealing
with the processing and treatment of received wireless signals by the communica-
tion system receiver.

Figure 3.11  An example of how the an LTI system h(t) can transform the PSD between the WSS
random process input X(t) and the WSS random process output Y(t).
80 Probability Review

3.5 ADDITIONAL READINGS

Although this chapter attempts to provide the reader with an introduction to some
of the key mathematical tools needed to analyze and model the random variables
and random processes that frequently occur in many digital communication systems,
the treatment of these mathematical tools is by no means rigorous nor thorough.
Consequently, the interested reader is encouraged to see some of the available text-
books that address this broad area in greater detail. For instance, the gold standard
for any textbook on the subject of probability and random processes is by Papoulis
and Pillai [1]. On the other hand, those individuals seeking to understand probabil-
ity and random processes theory within the context of communication networks
would potentially find the textbook by Leon-Garcia to be highly relevant [3].
For individuals who are interested in activities involving physical layer digital
communication systems and digital signal processing, such as Wiener filtering, the
textbook by Gubner would be a decent option given that many of the examples and
problems included in this publication are related to many of the classic problems
in communication systems [4]. Regarding textbooks that possess numerous solved
examples and explanations, the books by Hsu [5] and Krishnan [6] would serve as
suitable reading material. Finally, for those individuals who are interested in study-
ing advanced topics and treatments of random processes, the textbook by Grimmett
and Stirzaker would be a suitable publication [7].

3.6 Problems

1. [Probability of Event] A die is tossed twice and the number of dots facing up in
each toss is counted and noted in the order of occurrence.

(a) Find the probabilities of the elementary events.

(b) Find the probabilities of events A, B, C, A Ç Bc, and A Ç C, where set A cor-
responds to the event “number of dots in first toss is not less than the number
of dots in second toss,” set B corresponds to the event “number of dots in
first toss is 6,” and set C corresponds to the event “number of dots in dice
differs by 2.”

2. [Probability of Set] A number x is selected at random in the interval [-3,2]. Let

the events, A = {x < 0}, B = {| x – 0.5| < 0.5} and C = { x > 0.75}.

(a) Find the probabilities of A, B, A Ç B, and A Ç C.

(b) Find the probabilities of A È B, A È C, and A È B È C first by directly evalu-
ating the sets and then their probabilities, and second, by using appropriate
axioms or corollaries.

HINT: Assume that the probability of any subinterval is proportional to its length.

3. [Probability of Event] An asymmetric communications channel shown in Figure

3.12 operates as follows. The input can take one of the two discrete values,
3.6  Problems 81

Figure 3.12  An asymmetric communications channel with input values (“0,” “1”) and output
values (“0,” “1,” “2”).

namely a value of “0” with a probability a, and a value of “1” with a probability
1 − a. The output can take one of the three discrete values: “0,” “1,” and “2.”
Use the notation,

Ti = {i is transmitted}, i = 0, 1 with P[T0] = a, P[T1] = 1 − a, and

Rj = {j is received}, j = 0, 1, 2.

The conditional probabilities for the occurrence of the output symbols are also
shown in the figure. For example, P[R0|T0] = p1 (read as: “probability that a ‘0’
is received given that a ‘0’ was transmitted equals p1”).

(a) Find the probability that the output is “0” (i.e., find P[R0]).
(b) Find the probability that the input was “0” given that the output is “0” (i.e.,
find P[T0|R0]).
(c) Find the probability that the input was “1” given that the output is “2” (i.e.,
find P[T1|R2]).

4. [MATLAB TOOL] Referring to the asymmetric channel of Figure 3.12:

(a) Generate a 1 × 1, 000, 000 vector containing “0” and “1” with probabilities
a and 1 − a (NOTE: Use in this problem a = 0.1, 0.5, and 0.8). Show as a
function of vector index, n, the accumulation of the probabilities of “0” and
“1” as n increases from 1 to 1, 000, 000. Do these values correspond to the
theoretical (expected) outcome?
(b) Take the vector generated in part (a) and feed it to the asymmetric channel of
Figure 3.12. Plot the accumulated values of P[R0], P[T0|R0], and P[T1|R2] as
n increases from 1 to 1,000,000. Does the asymptotic behavior of these three
scenarios correspond to the theoretical results in question 3?
82 Probability Review

 ne way of writing a MATLAB script is explained as follows. Notice that the follow-
O
ing steps need to be implemented cumulatively as n increases from 1 to 1, 000, 000.

• Generate the transmit symbol vector using randsrc() function. The rand-
src() function allows you to generate the input vector such that each entry
of the alphabet occurs with the desired probability.
• Select the indices associated with zeros (ones) in the transmit symbol vector
using find() function.
• Randomize the indices associated with zeros (ones) in the transmit symbol
vector using randperm() function.
• Select the first 100 × p1% (100 × q1%) elements from the vector containing
randomized zero (one) locations and store them as zeros, select the next 100 ×
p2% (100 × q2%) elements from the same vector containing randomized
zero (one) locations and store them as ones, and so on. (You may have to use
round() function to select an integer number of zeros (ones).)
• Count the appropriate values, and store P[R0] as a function of n, P[T0|R0] as
a function of n and P[T1|R2] as a function of n.
• Substitute the numerical values in your result from question 3. Verify that the
plots of P[R0] versus n P[T0|R0] versus n and P[T1|R2] versus n settle at these
numerically obtained values as n increases.

5. [Probability Mass Function] Consider an information source that indefinitely

produces binary pairs that we designate as SX = {1, 2, 3, 4}. Find and plot the
probability mass function in the following cases:

(a) pk = p­1/k for all k in SX.

(b) pk = pk–1/2 for k = 2, 3, 4.
(c) pk = pk–1/2k for k = 2, 3, 4.
(d) Can the random variables in parts (a), (b), and (c) be extended to take on
values in the set {1, 2, 3,...}? If yes, specify the probability mass function of
the resulting random variables. If no, explain why not.

6. [Bernoulli Random Numbers] Suppose you have just been hired by the Acme
Randomness Company, Inc. and your first assignment is to deal with their
RNG3000 random number generator, as shown in Figure 3.13. This random
number generator operates as follows: Three Bernoulli random number genera-
tors are continuously producing sequences of “1” and “0” values, where the ith
Bernoulli random generator has a probability of pi of producing a “1” and a
probability of 1 − pi of producing a “0.” Although these three random number
generators are simultaneously producing binary values, only one is actually being
selected by a switch for a duration of N samples, and the switch randomly selects
a random number generator after every N samples. Note that the value for N is
selected at random from a range [1000, 10000] only once when the RNG3000
starts up. Finally, the output of the switch is then passed through a binary sym-
metric channel. The binary symmetric channel possesses a probability of p for
the case when the input samples remains the same at the output of the channel,
and a probability 1 – p when the sample value is flipped to the other binary
3.6  Problems 83

value. Note that the RNG3000 is capable of producing L = 10 000 000 random
samples in a single operation.

Given that p = 0.995, p1 = 0.9, p2 = 0.5, p3 = 0.1, and that we are not informed
about the value of N generated at the start of the process, design a device that
can be connected to the output of the binary symmetric channel capable of de-
termining the approximate value of N based solely on the continuous stream of
output samples from the binary symmetric channel, as well as determine which
Bernoulli random number generator is being selected by the switch for each seg-
ment of N samples. Note that you will need to implement Figure 3.13 in the first
place such that you can test your proposed design.

Several hints to get you in the right direction:

• Store in a single vector all the L samples produced at the output of the
RNG3000 that you have implemented for a fixed value of N (note that N is
generated only once via a uniform random number generator).
• To determine the size of N, apply a window of size N starting with N = 1000
and slide it across the vector of length L, recording the statistics of “0” and
“1” values per window. Based on the statistics, and iteratively repeating this
procedure for incrementally increasing window sizes, one can then determine
based on the statistics the appropriate size of N.
• Once N has been determined, one can use the statistics per window in order
to ascertain which one of the three Bernoulli random number generators
were employed.

7. [Random Variable] A random variable X has the following probability density

function (PDF):

ïìc (1 - x 2 ) -1 £ x £ 1
f X ( x) = í
ïî 0 elsewhere (3.87)

(a) Find c and plot the PDF.

(b) Find P[X = 0], P[0 < X < 0.5], and P[|X − 0.5| < 0.25].

Figure 3.13  Schematic of the Acme Corporation’s RNG3000 random number generator.
84 Probability Review

8. [Zero-mean Additive Noise] In this problem, we need to extract a secret message

for a noisy transmission possessing several unknown and random characteristics.
The noisy transmission is saved in the file secretmsg.mat.

 he objective of this problem is nullify the impact of the zero-mean additive

T
noise and attempt to locate all the actual transmission samples that have been
broadcast in bursts of various lengths rather than in a single continuous sig-
nal. The transmitter and the channel have the following properties:

• A secret message, contained within the intercepted signal, which consists of

26 lowercase letters of the alphabet. Each letter has been translated into a
5-digit binary word based its order in the alphabet (i.e., we have decimal rep-
resentation “a”=“0,” “b” = “1,” “c” = “2,”. . ., “z”=“25” that is translated
into binary words with the least significant bit located at the left-hand side;
for example, “3” = “1 1 0 0 0.” Note that since it takes 5 bits to represent
these 26 letters, and since there are 25 = 32 possible binary words, the re-
maining 3226 = 6 binary words are equal to a blank space.
• Once the stream of binary digits have been formed from the secret message,
all binary digits equal to “1” are replaced with a “+1” and all binary digits
equal to “0” are replaced with a “−1,” thus forming a sample train of “+1”
and “−1” values.
• Once this transmission stream of “+1” and “−1” values have been created,
a sample-and-hold technique is employed per sample, repeating it R = 1000
times.
• The transmitter does not send out the output of the sample-and-hold opera-
tion continuously but rather in bursts, where the duration of each burst in
terms of letters can be characterized by a discrete uniform random variable
distributed across [5, 20].
• The interburst duration in terms of the number of samples between the end
of the last burst and the start of the successive burst is characterized by an
exponential random variable of parameter l.
• The noise added to each sample of the transmitted signal in the channel can be
characterized by a zero-mean Gaussian random variable with variance s  2.
• Since noise is introduced into the channel, the original secret message prior
to binary encoding is repeated N = 3 times such that it is possible to recover
the message even in the presence of some corrupted bits.
• At the receiver, an integrate-and-dump approach can be used across each
of the R samples in order to determine whether a“+1”or a“−1” has been
transmitted. Note that this procedure simply adds up all the R samples cor-
responding to a specific binary digit and decides which value is more likely
after normalizing by R.

Once your receiver has been constructed, please answer the following questions:

(a) What is the approximate value of the parameter l of the exponential random
variable describing the interburst time?
3.6  Problems 85

(b) What is the approximate value of the variance s2 of the Gaussian noise added
to the transmitted message?
(c) Concatenating all the transmission bursts together, what is the secret message?

9. [Random Variable] A random variable X has the following cumulative distribu-

tion function (CDF):
ì0 for x < 0
ï
FX (x) = í 1 -2 x
ïî1 - 4 e for x ³ 0
(3.88)
(a) Plot the CDF and identify the type of random variable.
(b) Find P[X £ 2], P[X = 0], P[X < 0], P[2 < X < 6], and P[X > 10].
(c) Find the mean and variance of X.

10. [Random Variable] A random variable X has the following probability density
function (PDF):
ìïc (1 - x2 ) -1£ x £ 1
fX (x) = í
ïî0 elsewhere (3.89)

(a) Find c and plot the PDF.

(b) Plot the CDF of X.
(c) Find P[X = 0], P[0 < X < 0.5], and P[|X – 0.5| < 0.25].

11. [RLC Circuit] In this problem, suppose the Acme Corporation is focusing their
efforts on the design of an RLC circuit similar to the one shown in Figure 3.14.
The voltage transfer function between the source and the capacitor is:
1
H(ω ) = 2 (3.90)
(1 − ω LC) + jω RC
S uppose that the resonant frequency of the circuit, w0, is the value of w that
maximizes |H(w)|2.

(a) Solve for the expression of the resonant frequency in terms of R, L, and C.
(b) Given that these circuits are mass produced, then the actual values of R, L,
and C in a particular device may vary somewhat from their design specifica-

Figure 3.14  Schematic of the Acme Corporation’s RLC-1975 circuit.

86 Probability Review

tions, and, hence, so does the resonant frequency. Assuming L = 1, C = 1,

and R ∼ uniform[0, 2], plot the probability density function of the resonant
frequency w0 by testing N = 100,000,000 components.
(c) Assess the accuracy of your answer in part (b) by deriving and comparing with
the theoretical probability density function for the resonant frequency w0.
(d) Assuming L = 1, C = 1, and R ∼ uniform [0, 2], plot the probability den-
sity function of the resonant peak |H(w 0)|2 by testing N = 100,000,000
­components.
(e) Assess the accuracy of your answer in part (d) by deriving and comparing with
the theoretical probability density function for the resonant peak |H(w 0)|2.

12. [Bivariate Normal] In this problem, you have just been assigned another project
by the Acme Randomness Company, with respect to the characterization of some
test data that was recently collected during a measurement campaign. This test
data is saved in the file testdata.mat.
An initial analysis by the company has shown that this data is a concate­
nation of five bivariate Normal vector pairs, (X, Y), where the length for each
pair of vectors is derived from a discrete uniform random variable that produces
values between 100, 000 and 1, 000, 000. Note that all of these bivariate Normal
vectors are zero-mean and unit variance.
Furthermore, each of these vector pairs possess a different positive correla-
tion factor rXY between them, where rXY is defined as:

ìæ X - mX ö æ Y - mY ö ü
ρXY = E íç ÷ç ÷ ý (3.91)
îè σ X ø è σ Y ø þ

where mX and mY are the mean values of X and Y, while sX and sY are the stan-
dard deviation values of X and Y.
Your task is to determine the approximate lengths of these five vectors and to
find the approximate correlation factor pXY for each pair of vectors.

Several helpful suggestions:

• Use the plotting command scatter in order to show the correlation be-
tween X and Y. Please include these scatter plots in your solutions and make
sure to explain them thoroughly.
• Employ a sliding window approach in order to test the correlation of a spe-
cific segment. Changes in the correlation could be indicative of a transition
from one vector to another.

13. [Autocorrelation] Prove the following two properties of the autocorrelation

function Rxx(t) of a random process X(t):

(a) If X(t) contains a DC component equal to A, then Rxx(t) will contain a con-
stant component equal to A2.
(b) If X(t) contains a sinusoidal component, then Rxx(t) will also contain a sinu-
soidal component of the same frequency.
3.6  Problems 87

14. [Cross-correlation] Consider a pair of stationary processes X(t) and Y(t). Show
that the cross-correlations RXY(t) and RYX(t) of these processes have the following
properties:

(a) RXY(t) = RYX(–t)

1
(b) RXY (t ) ≤ [RX (0) + RY (0)]
2
15. [Advanced Topic] A sequence of L random variables {Xn , n ³ 1} is said to be a
discrete martingale with respect to another sequence {Yn , n ³ 1} if Xn is a func-
tion of Y1,Y2, . . .,Yn and if:

E [Xn+ 1 | Yn , …, Y1 ] = Xn , n ≥ 1 (3.92)

Moreover, if E[Xn+1|Yn, . . .,Y1] ³ Xn for n ³ 1, then Xn is said to be a discrete

submartingale. On the other hand, if E[Xn+1|Yn, . . .,Y1] £ Xn for n ³ 1, then Xn,
is said to be a discrete supermartingale.

Suppose we let {Zi} be some sort of Bernoulli process defined as:

 1, for success in the nth trial

Zn =  (3.93)
 −1, for failure in the nth trial
and
P(Zn = 1) = p
P(Zn = -1) = 1 - p = q (3.94)

where mz = p – q and σ Z2 = 4 pq. Let us define the following random walk process:

n n− 1
Wn = ∑ Zi = ∑ Zi + Zn = Wn−1 + Zn (3.95)
i= 1 i= 1

2
The mean of Wn is mw = n(p – q), while its variance is σ W = 4npq.
Prove that Yn = Wn – n(p – q) is a discrete martingale with respect to {Zk, 1 £ k £ n}.

16. [Advanced Topic] We define a random process {Y(t), t ³ 0} as a continuous mar-

tingale with respect to the random process {X(s), 0 £ s £ t} when:

E [Y (t)] < ¥,
E [Y (t) | X (s), 0 £ s £ t ] = Y (s) (3.96)

Moreover, when E[Y(t)|X(s), 0 £ s £ t] ³ Y(s), then {Y(t), t ³ 0} is a continuous

submartingale with respect to {X(s),0 £ s £ t}. Finally, when E[Y(t)|X(s),0 £ s £ t] £
Y(s), then {Y(t), t ³ 0} is a continuous supermartingale with respect to {X(s),
0 £ s £ t}.
Suppose that a Wiener process {W (t), t ³ 0} possesses a mean of zero and a
variance equal to s 2t. Moreover, let us define the random process:
88 Probability Review

2
(σ 2t / 2)
Y (t) = ekW (t)-k , t > 0 (3.97)
 here k is an arbitrary constant. Show that Y(t) is a martingale with respect to
w
{W(t), t ³ 0}.

17. [Advanced Topic] Suppose that we define a random process as Y(t) = W 2(t), where
W(t) is a Wiener process with variance s 2t. Show that Y(t) is a submartingale.
18. [Advanced Topic] Suppose that we have a Wiener process X(t). Using the follow-
ing arbitrary linear combination:
n
∑ ai X(ti ) = a1X(t1) + a2X(t2 ) + … + an X(tn ) (3.98)
i= 1

where 0 £ t1 £ . . .£ tn and ai are real constants, show that it can also be a Gauss-
ian process. [15 marks]

19. [Advanced Topic] Suppose we define the term continuous in probability for a
random process {X(t), t Î T} when for every Î > 0 and t Î T:

lim P(|X ( t + h) - X (t)| > ε) = 0 (3.99)

h ®0

Show that a Wiener process X(t) is continuous in probability.

20. [Advanced Topic] Suppose we define X(w) as the Fourier transform of a random
process X(t). If X(w) is nothing more than white noise with zero mean and an
autocorrelation function that is equal to q(w1)d(w1 – w2), then show that X(t) is
wide-sense stationary with a power spectral density equal to q(w)/ 2p.

References

[1] Papoulis A., and S. Unnikrishna Pillai, Probability, Random Variables and Stochastic Pro-
cesses, 4th Edition, McGraw-Hill Higher Education, 2002.
[2] Abramowitz, M., and I. Stegun, Handbook of Mathematical Functions: with Formulas,
Graphs, and Mathematical Tables, Dover Publications, 1965.
[3] Leon-Garcia, A., Probability, Statistics, and Random Processes for Electrical Engineering,
3rd Edition., Prentice Hall, 2008.
[4] Gubner J. A., Probability and Random Processes for Electrical and Computer Engineers,
Cambridge University Press, 2006.
[5] Hsu H., Schaum’s Outline of Probability, Random Variables, and Random Processes,
McGraw-Hill, 1996.
[6] Krishnan V., Probability and Random Processes, Wiley-Interscience, 2006.
[7] Grimmett G., and Stirzaker D., Probability and Random Processes, 3rd Edition, Oxford
University Press, 2001.
 ­­C h a p t e r 4

Digital Transmission Fundamentals

In this chapter, we will provide an overview of several key fundamental concepts

employed in the transmission of digital data. Starting with an understanding of how
binary data can be used to manipulate the physical characteristics of an electromag-
netic waveforms, we will then look at the basic anatomy of a digital communication
system before concentrating our attention on several different approaches for mod-
ulating electromagnetic waveforms using binary data. Once we have established
how a digital data transmission process operates, we will then explore one of the
key analytical tools for assessing the quantitative performance of such systems: the
bit error rate. Finally, this chapter will conclude with an introduction to the design
of digital receivers via a signal vector space perspective.

4.1  What Is Digital Transmission?

A digital transceiver is a system composed of a collection of both digital and analog

processes that work in concert with each other in order to handle the treatment and
manipulation of binary information. The purpose of these processes is to achieve
data transmission and reception across some sort of medium, whether it is a twisted
pair of copper wire, a fiber-optic cable, or a wireless environment. At the core of
any digital transceiver system is the binary digit or bit, which for the purposes of
this book is considered the fundamental unit of information used by a digital com-
munication system.
Therefore, a digital transceiver is essentially responsible for the translation be-
tween a stream of digital data represented by bits and electromagnetic waveforms
possessing physical characteristics that uniquely represent those bits. Since electro-
magnetic waveforms are usually described by sine waves and cosine waves, several
physical characteristics of electromagnetic waveforms commonly used to represent
digital data per time interval T include the amplitude, phase, and carrier frequency
of the waveform, as shown in Figure 4.1. Notice how different combinations of
bits represent different amplitude levels or different phase values or different carrier
frequency values, where each value uniquely represents a particular binary pattern.
Note that in some advanced mapping regimes, binary patterns can potentially be
represented by two or more physical quantities.
However, there is much more going on in a digital transceiver than just a map-
ping between bits and waveforms, as shown in Figure 4.2. In this illustration of the
basic anatomy of a digital transceiver, we observe that there are several functional
blocks that constitute a communication system. For instance, the mapping between
bits and electromagnetic waveform characteristics is represented by the modulation
and demodulation blocks. Additionally, there are the source encoding and source

89
90 Digital Transmission Fundamentals

Figure 4.1  Possible mappings of binary information to EM wave properties.

decoding blocks that handle the removal of redundant information from the binary
data, channel encoding and channel decoding blocks that introduce a controlled
amount of redundant information to protect the transmission for potential errors,
and the radio frequency front-end or RF front-end blocks that handle the conversa-
tion of baseband waveforms to higher carrier frequencies.
One may ask the question: Why do we need all these blocks in our digital commu-
nication system? Notice in Figure 4.2 the presence of a channel between the transmitter
and the receiver of the digital transmission system. The main reason that the design of a
digital communication system tends to be challenging, and that so many blocks are in-
volved in its implementation, is due to this channel. If the channel was an ideal medium,
where the electromagnetic waveforms from the transmitter are clearly sent to the receiver

Figure 4.2  Generic representation of a digital communication transceiver.

4.1  What Is Digital Transmission? 91

without any sort of distortion or disturbances, then the design of digital communication
systems would be trivial. However, in reality a channel introduces a variety of random
impairments to a digital transmission that can potentially affect the correct reception of
waveforms intercepted at the receiver. For instance, a channel may introduce some form
of noise that can obfuscate some of the waveform characteristics. Furthermore, in many
real-world scenarios many of these nonideal effects introduced by the channel are time-
varying and thus difficult to deal with, especially if they vary rapidly in time.
Thus, under real-world conditions, the primary goal of any digital communica-
tion system is to transmit a binary message m̂ (t) and have the reconstructed version
of this binary message m̂(t) at the output of the receiver be equal to each other. In
other words, our goal is to have P(m̂(t) ≠ m(t)) as small as needed for a particular
application. The metric for quantitatively assessing the error performance of a digi-
tal communication system is referred to as the probability of error or bit error rate
(BER), which we define as Pe = P(m̂(t) ≠ m(t)). Note that several data transmission
applications possess different Pe requirements due in part to the data transmission
rate. For instance, for digital voice transmission, a BER of Pe ~ 10−3 is considered
acceptable, while for an average data transmission application a BER of Pe ~ 10−5 −
10−6 is deemed sufficient. On the other hand, for a very high data rate application
such as those that employ fiber-optic cables, a BER of Pe ~ 10−9 is needed since any
more errors would flood a receiver given that the data rates can be extremely high.
To help mitigate errors that may occur due to the impairments introduced by
the channel, we will briefly study how source encoding and channel encoding works
before proceeding with an introduction to modulation.

4.1.1 Source Encoding
One of the goals of any communication system is to efficiently and reliably com-
municate information across a medium from a transmitter to a receiver. As a re-
sult, it would be ideal if all the redundant information from a transmission could
be removed in order to minimize the amount of information that needs to be sent
across the channel, which would ultimately result in a decrease in the amount of
time, computational resources, and power being expended on the transmission.
Consequently, source encoding is a mechanism designed to remove redundant in-
formation in order to facilitate more efficient communications.
The way source encoding operates is by taking a sequence of source symbols
u and mapping them to a corresponding sequence of source encoded symbols
v, vi ∈ v as close to random as possible, and the components of v are uncorrelated,
(i.e., unrelated). Thus, by performing this source encoding operation we hope to
achieve some level of redundancy minimization in vi ∈ v, thus limiting the amount
of wasted radio resources employed in the transmission of otherwise “predictable
symbols” in u. In other words, a source encoder removes redundant information
from the source symbols in order to realize efficient transmission. Note that in order
to perform source encoding, the source symbols need to be digital.
A single analog television channel occupies 6 MHz of frequency bandwidth.
i On the other hand, up to eight digitally encoded television channels can fit
within the same frequency bandwidth of 6 MHz.
92 Digital Transmission Fundamentals

4.1.2  Channel Encoding

To protect a digital transmission from the possibility of its information being corrupted,
it is necessary to introduce some level of controlled redundancy in order to reverse
the effects of data corruption. Consequently, channel encoding is designed to correct
for channel transmission errors by introducing controlled redundancy into the data
transmission. As oppose to the redundancy that is removed during the source encoding
process, which is random in nature, the redundancy introduced by a channel encoding
is specifically designed to combat the effects of bit errors in the transmission, (i.e., the
redundancy possesses a specific structure known to both the transmitter and receiver).
In general, channel encoding operates as follows: Each vector of a source en-
coded output of length K, namely, vl where l = 1,2,..., 2K, is assigned a unique
codeword such that the vector vl = (101010 . . .) is assigned a unique codeword
cl Î of length N, where  is a codebook. During this process, the channel encoder
has introduced N − K = r controlled number of bits to the channel encoding process.
The code rate of a communications system is equal to the ratio of the number of
information bits to the size of the codeword, (i.e., the code rate is equal to k/N).
When designing a codebook for a channel encoder, it is necessary to quanti-
tatively assess how well or how poorly the codewords will perform in a situation
involving data corruption. Consequently, the Hamming distance is often used to
determine the effectiveness of a set of codewords contained within a codebook
by evaluating the relative difference between any two codewords. Specifically, the
Hamming distance dH(ci , cj) between any two codewords, say, ci and cj, is equal to
the number of components in which ci and cj are different. When determining the
effectiveness of a codebook design, we often are looking for the minimum Ham-
ming distances between codewords, that is,

dH , min = min dH (ci , c j ) (4.1)

ci ,c j ∈C , i ≠ j

since our goal is to maximize the minimum Hamming distance in a given codebook
to ensure that the probability of accidentally choosing a codeword other than the
correct codeword is kept to a minimum. For example, suppose we have a codebook
consisting of {101,010}. We can readily calculate the minimum Hamming distance
to be equal to dH,min = 3, which is the bestpossible result. On the other hand, a code-
book consisting of {111, 101} possesses a minimum Hamming distance of dH,min = 1,
which is relatively poor in comparison to the previous codebook example.
In the event that a codeword is corrupted during transmission, decoding spheres
(also known as Hamming spheres) can be employed in order to make decisions on
the received information, as shown in Figure 4.3, where codewords that are cor-

Figure 4.3  An example of decoding spheres.

4.1  What Is Digital Transmission? 93

rupted during transmission are mapped to the nearest eligible codeword. Note that
when designing a codebook, the decoding spheres should not overlap in order to en-
able the best possible decoding performance at the receiver, (i.e.,→ dH.min = 2t + 1).

4.1.2.1  Shannon’s Channel Coding Theorem

In digital communications, it is sometimes necessary to determine the upper limit of
the data rate for a specific digital transceiver design. Consequently, in 1949 Claude
Shannon published his seminar paper that addressed this topic, “Communication
in the Presence of Noise” [1]. In this paper, he defined

A rate 1/3 repetition code with no source encoding would look like:
1 → 111 = c1 (1st codeword)

0 → 000 = c2 (2nd codeword)

Q
∴ C = {000, 111}

 hat are the Hamming distances for the codeword pairs dH (111,000)
W
and dH (111,101)?

a quantitative expression that described the limit on the data rate, or capacity, of a
digital transceiver in order to achieve error-free transmission.
Suppose one considers a channel with capacity C and we transmit data at a
fixed code rate of K/N, which is equal to Rc (a constant). Consequently, if we in-
crease N, then we must increase K in order to keep Rc equal to a constant. What
Shannon states is that there exists a code such that for Rc = K/N < C and as N → ∞,
we have the probability of error Pe → 0. Conversely, for Rc = K/N ≥ C, Shannon
indicated that no such code exists. Hence, C is the limit in rate for reliable com-
munications, (i.e., C is the absolute limit that you cannot go any faster than this
amount without causing errors).
So why is the result so important? First, the concept of reliability in digital com-
munications is usually expressed as the probability of bit error, which is measured
at the output of the receiver. As a result, it would be convenient to know what this
capacity is given the transmission bandwidth, B, the received signal-to-noise ratio
(SNR) using mathematical tools rather than empirical measurements. Thus, Shannon
derived the information capacity of the channel, which turned out to be equal to:

C = B log 2 (1 + SNR) [b/s] (4.2)

where this information capacity tells us the achievable data rate. Note that Shannon
provided us with only the theoretical limit for the achievable capacity of a data trans-
mission, but he does not tell us how to build a transceiver to achieve this limit.
Second, the information capacity of the channel is useful since this expression
provides us with a bound on the achievable data rate given bandwidth B and re-
ceived SNR, employed in the ratio η = R/C, where R is the signaling rate and C is
94 Digital Transmission Fundamentals

the channel capacity. Thus, as η → 1, the system becomes more efficient. Therefore,
the capacity expression provides us with a basis for tradeoff analysis between B and
SNR, and it can be used for comparing the noise performance of one modulated
scheme versus another.

4.2  Digital Modulation

In analog modulation schemes, the analog message signal modulates a continu-

ous wave prior to transmission across a medium. Conversely, digital modulation
involves having a digital message signal modulating a continuous waveform. As we
have seen earlier in this chapter, this can be accomplished by uniquely manipulating
the amplitude and phase information of a signal during each symbol period T based
on a specific pattern of bits. However, most digital modulation techniques possess
an intermediary step that we will focus on in this section, where collections of b
bits forming a binary message mb are mapped to a symbol, which is then used to
define the physical characteristics of a continuous waveform in terms of amplitude
and phase. In particular, for each of the possible 2b values of mb, we need a unique
signal si(t),1 ≤ i ≤ 2b that can then be used to modulate the continuous waveform,
as shown in Figure 4.4.
In this section, we will study several different families of approaches for mapping
binary data into symbols that can then be used to modulate continuous waveforms.
These modulation scheme families are defined by which physical characteristic or
combination of characteristics are manipulated by the mapping process in order to
uniquely represent a specific bit pattern. However, there exist various tradeoffs be-
tween these different families, including how efficiently a bit is mapped to a symbol
in terms of the transmit power expended. Consequently, we will first explore how
we assess this tradeoff before studying three major families of digital modulation
schemes and how they compare to each other.

4.2.1  Power Efficiency

In order to assess the effectiveness of mapping a bit to a symbol in terms of the
transmit power expended per symbol, we can employ the power efficiency metric.
Suppose we define the energy of a symbol s(t) as:
T

∫s
2
Es = (t)dt (4.3)
0

Figure 4.4  Modulation process of equivalent binary data.

4.2  Digital Modulation 95

where T is the period of the symbol. Then, for a modulation scheme consisting of M
symbols, we can define the average symbol energy via the following weighted average:
T T
Es = P(s1(t)) ⋅ ∫0 s1 (t)dt + � + P(sM (t)) ⋅ ∫0 sM (t)dt
2 2
(4.4)

where P(si(t)) is the probability that the symbol si(t) occurs. Furthermore, if we
would like to calculate the average energy per bit, we can approximate this using Es
and dividing this quantity by b = log2 (M) bits per symbol, yielding:
Es = Es
Eb = (4.5)
b log 2 (M)
To quantitatively assess the similarity between two symbols in terms of their
physical characteristics, we define the Euclidean distance as:
T
dij2 = ∫ (si (t) − s j (t))2 dt = E∆sij (4.6)
0

where ∆sij (t) = si(t) – sj(t). Since we are often interested in the worst-case scenario
when assessing the performance of a modulation scheme, we usually compute the
minimum Euclidean distance, namely:
T
∫
2
dmin = min (si (t) − s j (t))2 dt (4.7)
si (t ), s j (t ),i ≠ j 0

Thus, the power efficiency of a signal set used for modulation is given by the
expression:
d2
e p = min (4.8)
Eb

4.2.2  Pulse Amplitude Modulation

Of the various physical characteristics of a signal waveform that can be manipu-
lated in order to convey digital information, the most obvious choice is the signal
amplitude level. Leveraging this physical characteristic, pulse amplitude

Suppose we would like to find the p given the following waveforms:

s1(t ) = A × [u(t ) - u(t - T )] = s(t )

s2 (t ) = -A × [u(t ) - u(t - T )] = -s(t )

Q where u(t) is the unit step function. Compute the following:
• The minimum Euclidean distance dmin 2
.
• The average bit energy Esb.
• The power efficiency eP.

modulation­ (PAM) is a digital modulation scheme where the message information

is encoded in the amplitude of a series of signal pulses. Furthermore, demodulation
of a PAM transmission is performed by detecting the amplitude level of the carrier
at every symbol period.
96 Digital Transmission Fundamentals

The most basic form of PAM is binary PAM (B-PAM), where the individual
binary digits are mapped to a waveform s(t) possessing two amplitude levels ac-
cording to the following modulation rule:

• “1” → s1(t);
• “0” → s2(t).

where s1(t) is the waveform s(t) possessing one unique amplitude level, while s2(t) is
also based on the waveform s(t) but possesses another unique amplitude level. Note
that the waveform s(t) is defined across a time period T and is zero otherwise. Since
the duration of the symbols is equivalent to the duration of the bits, the bit rate for
a B-PAM transmission is defined as Rb = 1/T bits per second.
The energy of the waveform s(t) is defined as:

T
Es = ∫ s2 (t)dt [ Joules] (4.9)
0

Suppose we define s(t) as a rectangular waveform, namely:

s(t) = A ⋅ [u(t) − u(t − T )] (4.10)

where u(t) is the unit step function and A is the signal amplitude. Furthermore, sup-
pose that the bit “1” is defined by the amplitude A while the bit “0” is defined by
the amplitude − A. We can subsequently write our modulation rule to be equal to:

• “1” → s(t);
• “0” → s(t).
2
2 A
Therefore, the symbol energy is given by Es = E-s = A T = . From this result, we
Rb
can define the energy per bit for a B-PAM transmission as:
T T
Eb = P(1) × ò s12(t)dt + P(0) × ò s22(t)dt (4.11)
0 0

where P(l) is the probability that the bit is a “1,” and P(0) is the probability that the
bit is a “0.” Thus, if we define s1(t) = s(t) and s2(t) = –s(t), then the average energy
per bit is equal to:
T
Eb = Es {P(1) + P(0)} = Es = ∫ s2 (t)dt = A2T (4.12)
0

Calculating the minimum Euclidean distance, we get:

T T
2
dmin = ∫ (s(t) − (−s(t)))2 dt = ∫ (2s(t))2 dt = 4A2T (4.13)
0 0

which is then plugged into (4.8) in order to yield a power efficiency result for a
B-PAM transmission of:
d2 4A2T
e p = min = 2 = 4 (4.14)
Eb AT
4.2  Digital Modulation 97

As we will observe throughout the rest of this section, a power efficiency result of
4 is the best possible result that you can obtain for any digital modulation scheme
when all possible binary sequences are each mapped to a unique symbol.
Suppose we now generalize the B-PAM results obtained for the average bit en-
ergy, the minimum Euclidean distance, and the power efficiency and apply them to
the case when we try mapping binary sequences to one of M possible unique signal
amplitude levels, referred to as M-ary pulse amplitude modulation (M-PAM). First,
let us express the M-PAM waveform as:

si (t) = Ai ⋅ p(t), for i = 1, 2, …, M/ 2 (4.15)

where Ai = A(2i – 1), p(t) = u(t) – u(t – T), and u(t) is the unit step function. Graphi-
cally speaking, the M-PAM modulation scheme looks like the signal constellation
shown in Figure 4.5.
In order to compute the power efficiency of M-PAM, p,m-pam, we select the
2
dmin pair s1(t) = A × p(t) and s2(t) = –A × p(t) since this pair of signal waveforms are
the closest to each other in terms of shape. Thus, using this selection of waveforms,
we can solve for the difference between them:

∆s(t) = 2 A ⋅ p(t) (4.16)

which yields a minimum Euclidean distance of:

2
dmin = 4A2T (4.17)

In order to calculate the average symbol energy, Es, we can simplify the mathemat-
ics by exploiting the symmetry of signal constellation, which yields:

2 2 M/ 2
Es = A T ∑ (2i − 1)2
M i =1

(M 2 − 1)
= A2T which is simplified via tables (4.18)
3

Es A2T (22b − 1)
→ Eb = =
log 2 (M) 3b

Finally, solving for the power efficiency yields:

12b
e p, M− PAM = 2b (4.19)
2 −1

Figure 4.5  M-PAM Signal constellation.

98 Digital Transmission Fundamentals

4.2.3  Quadrature Amplitude Modulation

Similar to PAM, quadrature amplitude modulation (QAM) implies some sort of
amplitude modulation. However, QAM modulation is a two-dimensional signal
modulation scheme as opposed to PAM modulation. The two dimensions of the
QAM modulation, namely, the in-phase and quadrature components, are orthogo-
nal to each other, which implies that one can essentially double the transmission
data rate “for free.” Furthermore, rectangular QAM can be thought of as two or-
thogonal PAM signals being transmitted simultaneously.
Mathematically, if a rectangular QAM signal constellation consists of M unique
waveforms, this could potentially be represented as M-PAM transmissions operat-
ing simultaneously in orthogonal dimensions. Note that QAM signal constellations
could also take the form of nested circles (called circular QAM), or any other geo-
metric pattern that involves orthogonal modulators. Rectangular QAM is a popular
modulation scheme due to its relatively simple receiver structure, as shown in Fig-
ure 4.6, where each dimension employs a M-ary PAM detector.
In order to determine the power efficiency of M-QAM, let us first define the
mathematical representation of a signal waveform belonging to this form of modu-
lation:

sij (t) = Ai ⋅ cos(wct) + Bj ⋅ sin(wct) (4.20)

where wc is the carrier frequency, and Ai and Bj are the in-phase and quadrature am-
plitude levels. Notice how the cosine and sine functions are used to modulate these
amplitude levels in orthogonal dimensions. Visualizing M-QAM as a signal constel-
lation, the signal waveforms will assume positions in both the real and imaginary
axes, as shown in Figure 4.7.
To compute the power efficiency of M-QAM, ep ,1M-QAM, we first need to
calculate the minimum Euclidean distance, which becomes:
T

∫ ∆s
2 2
dmin = (t)dt = 2 A2T (4.21)
0
where we have selected the following signal waveforms without loss of generality:
s1(t) = A ⋅ cos(wct) + A ⋅ sin(w ct)
s2 (t) = 3A ⋅ cos(wct) + A ⋅ sin(wct) (4.22)

Figure 4.6  M-QAM receiver structure.

4.2  Digital Modulation 99

Figure 4.7  M-QAM signal constellation.

In order to derive the average symbol energy, Es, we leverage the expression from
M-ary PAM by replacing M with M such that:
2 M −1
Es = A T (4.23)
3
which can then be used to solve:
Es 2b − 1
Eb = = A2T (4.24)
log 2 (M) 3b
Thus, the power efficiency is equal to:
3!b
e p, M− QAM = b
(4.25)
2 −1

4.2.4  Phase Shift Keying

Phase shift keying (PSK) is a digital modulation scheme that conveys data by chang-
ing, or modulating, the phase of a reference signal, (i.e., the carrier wave). Any digi-
tal modulation scheme uses a finite number of distinct signals to represent digital
data. PSK uses a finite number of phases, each assigned a unique pattern of binary
digits. Usually, each phase encodes an equal number of bits. Each pattern of bits
forms the symbol that is represented by the particular phase. The demodulator,
100 Digital Transmission Fundamentals

which is designed specifically for the symbol-set used by the modulator, determines
the phase of the received signal and maps it back to the symbol it represents, thus
recovering the original data. This requires the receiver to be able to compare the
phase of the received signal to a reference signal; such a system is termed coherent.
Phase shift keying (PSK) characterizes symbols by their phase. Mathematically,
a PSK signal waveform is represented by:

 π 
si (t) = A cos  2π fct + (2i − 1)  , for i = 1,..., log 2 m (4.26)
 m
π
where A is the amplitude, fc is carrier frequency, and (2i − 1) is the phase offset of
m
each symbol. PSK presents an interesting set of tradeoffs with PAM and QAM. In
amplitude modulation schemes, channel equalization is an important part of decod-
ing the correct symbols. In PSK schemes, the phase of the received signal is much
more important than the amplitude information.
There are several types of PSK modulation schemes based on the number of
M possible phase values a particular PSK waveform can be assigned. One of the
most popular and most robust is binary PSK, or B-PSK, modulation, whose signal
constellation is illustrated in Figure 4.8. In general, the modulation rule for B-PSK
modulation is the following:
“1” → s1(t) = A ⋅ cos(ωc t + θ )

“0”→ s2(t) = −A ⋅ cos(ωc t + θ )

= A ⋅ cos(ωc (t) + θ + π )
= −s1(t ) (4.27)

In other words, the two signal waveforms that constitute a B-PSK modulation
scheme are separated in phase by θ.
In order to derive the power efficiency of a B-PSK modulation scheme, ep,bpsk,
2
we first need to compute the minimum Euclidean distance dmin by employing the
definition and solving for:

Figure 4.8  BPSK signal constellation.

4.2  Digital Modulation 101

∫
2
dmin = (s1(t) − s2 (t))2 dt
0
T
= 4A2 ∫ cos2 (ω ct + θ )dt
0
T
4A2T 4A2
=
2
+
2 ∫ cos(2ω ct + 2θ )dt
0

= 2 A2T (4.28)

Notice in (4.28) how the second term disappeared from the final re-
sult. This is due to the fact that the second term possessed a carrier
frequency that was twice that of the original signal. Since the carrier
signal is a periodic sinusoidal waveform, integrating such as signal
possessing as high frequency would result in the positive portion of
i the integration canceling out the negative portion of the integration,
yielding an answer than is close to zero. Consequently, we refer to this
term as a double frequency term. Also note that many communication
systems filter their received signals, which means the probability of
filtering out the double frequency term is also quite high.
2
Note that another way for computing dmin is to use the concept of correlation,
which describes the amount of similarity between two different signal waveform. In
this case, we can express the minimum Euclidean distance as:
T

2
dmin = ∫ (s2 (t) − s1(t))2 dt = Es1 + Es2 − 2 r12 (4.29)
0
where the symbol energy for symbol i, ESi, and the correlation between symbols 1
and 2, r12, are given by:
T T

Esi = ∫ si2 (t )dt and ρ12 = ∫ s1(t )s2 (t )dt

0 0
Employing the definition for Eb, we can now solve for the average bit energy
of the B-PSK modulation scheme by first solving for the symbol energies of the two
signal waveforms and then averaging them; that is,
T T

∫ s12 (t)dt ∫
2
Es1 = =A cos2 (ω ct + θ )dt
0 0
2 2 T
AT A
=
2
+
2 ∫ cos(2ω ct + 2θ)dt
0

A2T
=
2
A2T
Es2 =
2
A2T
Eb = P(0) ⋅ Es2 + P(1) ⋅ Es1 = (4.30)
2
102 Digital Transmission Fundamentals

Note that since the number of bits represented by a single symbol is equal to one,
both the bit energy and symbol energy are equivalent.
Finally, applying the definition for the power efficiency, we get the following
expression:
2
dmin
ε p,BPSK = = 4 (4.31)
Eb
This is suppose to be the largest possible value for p for a modulation scheme
employing all possible signal representations, (i.e., M = 2b waveforms). Notice when
using the correlation approach to calculate the minimum Euclidean distance, in or­
2
der to get a large ε p, we need to maximize dmin, which means we want ρ12 < 0. Thus,
to achieve this outcome, we need the following situation:

Es1 = Es2 = E = A2T/2 (4.32)

2
which means dmin = 2(E − ρ12 ) and consequently ρ12 = −E.

Show that for the following signal waveforms:

s1(t ) = A ⋅ cos(wc t + q )
Q
s2 (t ) = 0
the power efficiency is equal to εp = 2.

Show that for the following signal waveforms:

s1(t ) = A ⋅ cos(wc t + q)
Q s2 (t ) = A ⋅ sin(wc t + q)

the power efficiency is equal to ε p = 2.

So far we have studied a PSK modulation scheme that consist of only just one of
two waveforms. We will now expand our PSK signal constellation repertoire to in-
clude four distinct waveforms per modulation scheme. In quadrature PSK (QPSK)
modulation, a signal waveform possesses the following representation:

si (t ) = ± A ⋅ cos(wc t + q) ± A ⋅ sin(wc t + q) (4.33)

where each signal waveform possesses the same amplitude but one of four possible
phase values. This kind of phase modulation is illustrated by the signal constella-
tion diagram shown in Figure 4.9, where each waveform is located at a different
phase value.
In order to derive the power efficiency of QPSK, εp,qpsk, we first need to solve
for the minimum Euclidean distance, dmin 2 , which is equal to:

T

2
dmin = ∫ ∆s 2 (t)dt = 2 A2 T (4.34)
0
4.2  Digital Modulation 103

Figure 4.9  QPSK signal constellation.

Next, we would like to find Eb , which requires us to average over all the signal
waveforms. Consequently, this is equal to:

(Es1 + Es2 + Es3+ Es4) /4 A 2 T

Eb = =
log 2 (M ) 2 (4.35)

where the symbol energy of all four symbols is equal to Es1 = Es2 = Es3 = Es4 = A2T.
Finally, solving for the power efficiency using (4.8), we get:
2
dmin
e p ,QPSK = = 4 (4.36)
Eb
which is the same as BPSK but with 2 bits per symbol, making this a fantastic result!
Finally, let us study the general case when a PSK modulation scheme has a
choice of M possible phase values, where the distance of a signal constellation point
to the origin is always a constant and the signal constellation consists of M equally
spaced points on a circle. Referred to as M-ary PSK (M-PSK), a signal waveform
can be mathematically represented as:

 2p i 
si (t ) = A ⋅ cos  w c t +  , for i = 0,1, 2, …, M − 1 (4.37)
 M 
Note that there are several advantages and disadvantages with this modulation
scheme. For instance, as M increases the spacing between signal constellation points
decreases, thus resulting in a decrease in error robustness. Conversely, having the
information encoded in the phase results in constant envelope modulation, which
is good for nonlinear power amplifiers and makes the transmission robust to am-
plitude distortion channels.
Regarding the derivation of the power efficiency for an M-PSK modulation
scheme, εp,m-psk, suppose we define two adjacent M-PSK signal waveforms as
s1(t) = A × cos(ωct) and s2(t) = A ⋅ cos(ωct + 2p/M). Calculating the minimum Euclid­
ean distance using:

2
dmin = Es1 + Es2 − 2ρ12 (4.38)
104 Digital Transmission Fundamentals

where we define the symbol energy as:

T
A2T
∫ si (t)dt =
2
Esi = , for i = 1, 2 (4.39)
0
2

and the correlation between the two signal waveforms as:

T
A2T  2p 
ρ12 = ∫ s1(t )s2 (t )dt = cos  (4.40)
2  M 
0
this yields:
 2p  
2
dmin = A 2 T 1 − cos   (4.41)
  M 
Es = Es where E s = A 2 T / 2. Using
The average bit energy Eb is equal to Eb =
log 2 (M ) b
the definition for the power efficiency from (4.8), we see that:
  2π  2 π 
ε p,M-PSK = 2b  1 − cos   = 4b sin  b  (4.42)
  M  2 

4.2.5  Power Efficiency Summary

After examining the power efficiency performance of several different modulation
schemes, it is important to assess the tradeoffs between the different schemes such
that we can make the appropriate design decisions in the future when implementing
a digital communication system. To determine how much power efficiency we are
losing relative to ep,qpsk, which possesses the best possible result, we use the follow-
ing expression:
 e p ,QPSK 
dSNR = 10 ⋅ log10   (4.43)
 e p ,other 
Using this expression, we created a table of dSNR values for the modulation schemes
studied in this chapter, as shown in Table 4.1.
From Table 4.1, notice how the two-dimensional modulation schemes perform
better than the one-dimensional modulation schemes. Furthermore, notice how all
of the modulation schemes studied are linear modulation schemes, which means
they possess a similar level of receiver complexity. Given these insights on the power
efficiency performance of these modulation schemes, we now turn our attention to
the robustness of a modulation technique in the presence of noise.

Table 4.1  dSNR Values of Various Modulation Schemes

M b M-ASK M-PSK M-QAM
2 1 0 0 0
4 2 4 0 0
8 3 8.45 3.5 (??)
16 4 13.27 8.17 4.0
32 5 18.34 13.41 (??)
64 6 24.4 18.4 8.45
4.3  Probability of Bit Error 105

4.3  Probability of bit error

One of the most commonly used quantitative metrics for measuring the performance of
a digital communication system is the probability of bit error or BER, which is the prob-
ability that a bit transmitted will be decoded incorrectly. This metric is very important
when assessing whether the design of a digital communication system meets the specific
error robustness requirements of the application to be support (e.g., voice, multimedia,
data). Furthermore, having a metric that quantifies error performance is helpful when
comparing one digital communication design with another. Consequently, in this sec-
tion we will provide a mathematical introduction to the concept of BER.
Suppose that a signal si(t), i = 1, 2, was transmitted across an AWGN channel
with noise signal n(t), and that a receiver intercepts the signal r(t). The objective of
the receiver is to determine whether either s1(t) or s2(t) was sent by the transmitter.
Given that the transmission of either s1(t) or s2(t) is a purely random event, the only
information that the receiver has about what was sent by the transmitter is the ob-
served intercepted signal r(t), which contains either signal in addition to some noise
introduced by the AWGN channel.
Given this situation, we employ the concept of hypothesis testing [2] in order to
set up a framework by which the receiver can decide on whether s1(t) or s2(t) was
sent based on the observation of the intercepted signal r(t). Thus, let us employ the
following hypothesis testing framework:
H1 : r(t) = s1(t) + n(t), 0 ≤ t ≤ T

H0 : r(t) = s2 (t) + n(t), 0 ≤ t ≤ T

where 0 and 1 are hypothesis 0 and hypothesis 1.

Leveraging this framework, we next want to establish a decision rule at the
receiver such that it can select which waveform was sent based on the intercept
signal. Suppose we assume that s1(t) was transmitted. In general, we can determine
the level of correlation between two signals x(t) and y(t) over the time interval 0 ≤
t ≤ T using the expression:
T

∫ x(t)y(t)dt
0

Consequently, our decision rule on whether s1(t) or s2(t) was transmitted given that
we observe r(t) is defined as:
T T

∫ r(t)s1(t)dt ≥ ∫ r(t)s2 (t)dt (4.44)

0 0

where we assume that s1(t) was transmitted. Recall that correlation tells us how
similar one waveform is to another waveform. Therefore, if the receiver knows the
appearance of s1(t) and s2(t), we can then determine which of these two waveforms
is more correlated to r(t). Since s1(t) was assumed to be transmitted, ideally the
received signal r(t) should be more correlated to s1(t) than s2(t).
On the other hand, what happens if some distortion, interference, and/or noise
is introduced in the transmission channel such that the transmitted signal waveforms
106 Digital Transmission Fundamentals

are corrupted? In the situation where a transmitted signal waveform is sufficiently cor-
rupted such that it appears to be more correlated to another possible signal waveform,
the receiver could potentially select an incorrect waveform, thus yielding an error
event. In other words, assuming s1(t) was transmitted, an error event occurs when:
T T

∫ r(t)s1(t)dt ≤ ∫ r(t)s2 (t)dt (4.45)

0 0
Since r(t) = s1(t) + n(t), we can substitute this into the error event in order to obtain
the decision rule:
T T T T

∫ s12 (t)dt + ∫ n(t)s1(t)dt ≤ ∫ s1(t)s2 (t)dt + ∫ n(t)s2 (t)dt

0 0 0 0
T
Es1 − ρ12 ≤ ∫ n(t)(s2 (t) − s1(t))dt
0

Es1 − ρ12 ≤ z

From this expression, we observe that both Es1 and ρ12 are deterministic quan-
tities. On the other hand, z is based on the noise introduced by the transmission
channel, and thus it is a random quantity that requires some characterization. Since
n(t) is a Gaussian random variable, then z is also a Gaussian random variable. This
is due to the fact that the process of integration is equivalent to a summation across
an infinite number of samples, and since we are summing up Gaussian random vari-
ables, the result in is also a Gaussian random variable. With z ~ N(0, s 2), we now
need to calculate the variance of z, s 2, which can be solved as follows:
T
N0
σ 2 = E{z 2 } =
2 ∫ (s1(t) − s2 (t))2 dt
0
N
= 0 (Es1 + Es2 − 2ρ12 ) → Assume Es1 = Es2 = E
2
= N0 (E − ρ12 )
T T
where E = Ei = ∫ si2 (t)dt and ρ12 = ∫ s1(t)s2 (t)dt. Note that we are assuming that the
0 0
channel is introducing zero-mean noise, which means the sum of these noise contribu-
tions, (i.e., z, will also be zero-mean).
With both deterministic and random quantities characterized, we can now pro-
ceed with the derivation for the probability of bit error. The probability of an error
occurring given that a “1” was transmitted, (i.e., P(e|1)), is equal to:
 E − ρ12  Since z ∼ N(0, σ 2 )
P(z ≥ E − ρ12 ) = Q   →
 σ  and E − ρ12 is constant
 
 (E − ρ12 )2 
= Q  

→ Use σ 2 = N0 ( E −ρ 12 )
 σ2 


E − ρ12 
= Q 
 N0 
4.3  Probability of Bit Error 107

where the Q-function is defined as:

∞
1 2
Q(x) = ∫ e− t /2
dt (4.46)
2π x
The next step is to optimize the probability of bit error by minimizing P(e|1),
which can be achieved by optimizing the correlation term ρ12. Intuitively, the best
choice we can make is when s2(t) = –s1(t), which gives us ρ12 = −E. Consequently,
this result yields:

2Eb 
P(e |1) = Q  (4.47)
 N0 
2
Note that when Es1 ≠ Es2, we can then use dmin = Es1 + Es2 − 2 ρ12, which yields the
following expression for the probability of bit error:
 2 
 dmin 
Pe = Q  (4.48)


2N0 

Show that the total probability of bit error is equal to:


E − ρ12 
Q Pe = P(e|1)P(1) + P(e| 0)P(0) = Q 
N0 
(4.49)


When dealing with a large number of signal waveforms that form a modulation
scheme, the resulting probability of error, Pe, is expressed as a sum of pairwise er-
ror probabilities (i.e., the probability of one received symbol being another, specific
received symbol). The pairwise error probability of si(t) being decoded when sj(t)
was transmitted is given as:
 dij2 
Q  (4.50)
 2 N0 
where N0 is the variance of the noise. An important note here is that we are assum-
ing the noise is AWGN. since Q functions apply specifically to Gaussian random
variables. Therefore, the complete expression for Pe can be expressed as:
æ d2 ö æ d12j ö æ dMj
2 ö
Q ç min ÷ £ Pe £ Q ç ÷ + …+ Qç ÷, i ¹ j (4.51)
è 2 N0 ø èç 2N0 ø÷ èç 2N0 ø÷

where the second half of the relationship is the summation of every single pairwise
error probability.

4.3.1  Error Bounding

Computing each pairwise error probability is not always practical. It is possible to
create an upper and lower bound on Pe by computing only the pairwise errors of
points that are within one degree of the point of interest. Consider the behavior of the
Q function Q(.). As the input to Q(.) increases, the resulting output of the Q function
108 Digital Transmission Fundamentals

approaches zero. You will find that computing the pairwise error probability of points
farther away yields negligible contributions to the total Pe, but can save a significant
amount of time as well as cycles. Thus, an accurate estimate of P(e) can be computed
from the following bounds. These upper and lower bounds can be expressed as:
æ d2 ö æ dij2 ö
Q ç min ÷ £ P (e ) £ å Q ç ÷ (4.52)
è 2 N0 ø i ÎI èç 2N0 ÷ø

where I is the set of all signal waveforms within the signal constellation that are
immediately adjacent to the signal waveform j. In order to accurately assess the per-
formance of a communications system, it must be simulated until a certain number
of symbol errors are confirmed [3]. In most cases, 100 errors will give a 95 percent
confidence interval, which should be employed later on in this book in order to char-
acterize the bit error rate of any digital communication system under evaluation.

4.4 Signal Space Concept

Until this point we have studied digital communication systems from a signal wave-
form perspective. Leveraging this perspective, we have developed mathematical
tools for analyzing the power efficiency and BER of different modulation schemes.
However, there are several instances where the use of a signal waveform framework
can be tedious or somewhat cumbersome. In this section, we will introduce another
perspective on how to characterize and analyze modulation scheme using a differ-
ent mathematics representation: signal vectors.
Suppose we define φj(t) as an orthnormal set of functions over the time interval
[0, T] such that:
T
1 i= j

∫ φi (t)φj (t)dt =  0 otherwise
0

Given that si(t) is the ith signal waveform, we would like to represent this wave-
form as a sum of several orthonormal functions; that is,
N
si (t) = ∑ sikφk (t) (4.53)
k= 1

which can be equivalently represented by the vector:

si = (si1 , si 2 , si 3 , …siN ) (4.54)

where the elements of the vector si define the amplitude scaling of each orthonor-
mal function used to represent the waveform. An illustration of a signal waveform
represented by three orthonormal functions is shown in Figure 4.10. Consequently,
given this relationship between the signal waveform and the orthonormal functions,
where the former can be represented as the weighted sum of the latter, we can read-
ily describe the signal waveform as a simple vector, which we will see next possesses
the advantage of enabling us to employ relatively straightforward mathematical
operations based on linear algebra.
4.4  Signal Space Concept 109

Figure 4.10  Sample vector representation of Si (t) in three dimensional space using basis functions
φ1(t), φ2(t), and φ3(t).

In order to find the vector elements, sil, we need to solve the expression:
T N T
ò si (t)φl (t)dt = å sik ò φk (t)φl (t)dt = sil (4.55)
0 k =1 0
which is essentially a dot product or projection of the signal waveform si(t) on the
orthonormal function fi(t). At the same time, if we perform the vector dot product
between the signal waveforms si(t) and sj(t), we get a correlation operation that is
equal to:
T
∫ si (t)s j (t)dt = si ⋅ s j = ρij (4.56)
0
while the energy of a signal si(t) is equal to:
T
Esi = ∫ si2 (t)dt = si ⋅ si =||si||2 (4.57)
0

All of these mathematical operations will be employed when determining the

power efficiency of a modulation scheme or deriving the optimal decision rule for a
receiver.
Suppose we would like to compute the power efficiency for a modulation
scheme using a signal vector approach rather than a signal waveform approach.
The first step would be to calculate the minimum Euclidean distance, which can be
solved using the following:
110 Digital Transmission Fundamentals

T T

∫ ∫
2
dmin = ∆sij2 (t)dt = (si (t) − s j (t))2 dt
0 0

= || si − s j ||2 = (si − s j ) ⋅ (si − s j )

= Esi + Esj − 2 ρij

where the correlation term between signal waveforms Si(t) and Sj(t) is given by:
T
ρij = ∫ si (t)s j (t)dt = si ⋅ s j (4.58)
0
In order to solve for the power efficiency, we choose a set of orthonormal basis
functions fi(t), i = 1,2,...,k, where k is the dimension of the signal vector space.
Given this set of functions, we can now represent the vector si, i = 1,2,..., M where
si = (si1, si2,…sik) and:
T
sij = ∫ si (t)φ j (t)dt (4.59)
0

Consequently, using the vector representations for the signals and the orthonor-
mal functions, we can calculate the minimum Euclidean distance:
2
dmin = min || si − s j || 2 (4.60)
i≠ j

the average symbol and bit energy values:

1 M
Es = ∑
M i= 1
|| si || 2

Eb = Es / log2(M) (4.61)
and the power efficiency:

2
ε p = dmin / E b (4.62)

4.5 Gram-Schmidt Orthogonalization

In mathematics, particularly linear algebra and numerical analysis, the Gram-

Schmidt orthogonalization process is a method for creating an orthonormal set
of functions in an inner product space such as the Euclidean space n. The Gram-
Schmidt orthogonalization process takes a finite set of signal waveforms {s1(t),...,
sM(t)} and generates from it an orthogonal set of functions {f1(t),..., fi(t)} that spans
the space n. Note that an orthonormal function possesses the following property:
T
1 i=j
∫ φi (t) φj (t)dt = 0 otherwise
0
Furthermore, it is possible to represent a signal waveform si(t) as the weighted
sum of these orthonormal basis functions; that is,
4.5  Gram-Schmidt Orthogonalization 111

N
si (t) = ∑ sikφk (t) (4.63)
k= 1

However, what we need now is an approach to generate the set of orthonormal

basis functions {fj(t)}.
To derive a set of orthogonal basis functions {f1(t),..., fi(t)} from set of signal
waveforms denoted by {s1(t),..., sm(t)}, let us first start with s1(t) and normalize it:
s1(t)
φ1(t) =
Es1
where Es1 is the energy of the signal s1(t). This normalized version of the signal wave-
form s1(t) will serve as our first orthonormal basis function from which we will con-
struct the rest of our orthonormal basis function set. In other words, we are effectively
“bootstrapping” a set of orthonormal basis functions based on the existing signal wave-
forms of the modulation scheme to be used. Note that we can represent s1(t) as:
s1(t) = Es1 φ1(t) = s11φ1(t)

where the coefficient s11 = Es1 and the orthonormal function f1(t) satisfy the unit
energy constraint as required.
Next, we would like to create the second orthonormal function, f2(t).In order to
accomplish this task, we use the signal waveform S2(t) as a starting point. However,
S2(t) may contain elements of f1(t), and thus we need to remove this component
before normalizing it in order to transform this waveform into f2(t). To achieve
this, we first determine how much of f1(t) is contained within s2(t) by taking the
dot product between these two functions and determining how much s2(t) projects
onto f1(t); that is,
T
s21 = ∫ s2 (t)φ1(t)dt
0
To help in getting the basis function f2(t), we define the intermediate function:

g2 (t) = s2 (t) − s21φ1(t)

which is orthogonal to f1(t) over the interval 0 £ t £ T by virtual of the fact that we
have removed the f1(t) component from s2(t). Finally, normalizing g2(t) yields the
basis function f2(t):
g2 (t)
φ2 (t) =
T

∫ g22 (t)dt (4.64)

0

which can be expanded to:

s2 (t) − s21φ1(t)
φ2 (t) =
2
Es2 − s21 (4.65)
where Es2 is the energy of the signal s2(t). A quick sanity check clearly shows that
the orthonormal basis function f2(t) satisfies the constraint:
112 Digital Transmission Fundamentals

T T

ò φ2 (t)dt = 1 ò φ1(t)φ2 (t)dt = 0

2
and
0 0

In general, we can define the following functions that can be employed in an

iterative procedure for generating a set of orthonormal basis functions:
i− 1
gi (t) = si (t) − ∑ sijφ j (t)
j= 1
T
sij = ∫ si (t)φ j (t)dt , j = 1, 2, …, i − 1
0
gi (t)
φi (t) = , i = 1, 2, …, N
T

∫ gi2 (t)dt (4.66)

0
We will now work out an example that deals with the Gram-Schmidt orthogonal-
ization process.

4.5.1 An Example
Suppose we want to perform the Gram-Schmidt orthogonalization procedure of the sig-
nals shown in Figure 4.11 in the order s3(t), s1(t), s4(t), s2(t) and obtain a set of orthonor-
mal functions {fm(t)}. Note that the order in which the signal waveforms are employed
to generate the orthonormal basis functions is very important, since each ordering of
signal waveforms can yield a potentially different set of orthonormal basis functions.

Figure 4.11  Example signal waveforms.

4.6  Optimal Detection 113

Starting with s3(t), we get:

s3(t) s3(t)
φ1(t) = = (4.67)
Es3 3
Then, leveraging the result for f1 (t), we then derive the orthonormal basis function
f2(t) using S1(t):
 1 / 3, 0 ≤ t< 2
=  2 / 3,
2
g 2(t) = s1(t) − s12φ1(t) = s1(t) − s (t) 2≤t<3
3 3  0,
 t≥3
 1/ 6 , 0≤ t< 2
g2 (t) 
∴ φ 2 (t) = = 2 / 6 , 2 ≤ t <3
T  0,
t≥3
∫ g22 (t)dt  (4.68)
0

We subsequently repeat this operation for s4(t):

2
g3(t) = s4 (t) − ∑ s4 jφ j (t) = 0
j= 1

∴ φ3(t) = 0 (4.69)
but we notice the resulting f3(t) is equal to zero. This implies that the signal wave-
form s4(t) can be entirely characterized by only f1(t) and f2(t). Finally, for s2(t), we
get the following:
3
g 4 (t) = s2 (t) − ∑ s2 jφ j (t) = 0
j= 1

g 4 (t) s2 (t)
∴ φ4 (t) = =
T 2
∫ g 42 (t)dt (4.70)
0

Consequently, with the orthonormal basis functions {f1(t), f2(t), f4(t)} defined, we
can now express the four signal waveforms as:

• s1 = (2 / 3 , 6 / 3, 0);
• s2 = (0, 0, 2);
• s3 = ( 3 , 0, 0);
• s4 = (−1/ 3 ,−4/ 6 , 0).

4.6 Optimal Detection

Detection theory, or signal detection theory, is used in order to discern between signal
and noise [2]. Using this theory, we can explain how changing the decision threshold
will affect the ability to discern between two or more scenarios, often exposing how
adapted the system is to the task, purpose, or goal at which it is aimed.
114 Digital Transmission Fundamentals

4.6.1 Signal Vector Framework

Let us assume a simple digital transceiver model as shown in Figure 4.12. As men-
tioned previously, the receiver only observes the corrupted version of si(t) by the
noise signal n(t), namely, r(t). The noise signal n(t) usually represents the culmina-
tion of all noise sources into a single variable. Therefore, our detection problem in
this situation can be summarized as: given r(t) for 0 £ t £ T, determine which si(t),
i = 1, 2,..., M, is present in the intercepted signal r(t).
Suppose we decompose the waveforms si(t), n(t), and r(t) into a collection of
weights applied to a set of orthonormal basis functions, namely:

N N N
si (t) = å sikφk (t), r(t) = å rkφk (t), n(t) = å nkφk (t)
k =1 k =1 k =1

Given that all of these signal waveforms use the same orthonormal basis functions,
we can rewrite the waveform model expression r(t) = si(t) + n(t) into:
N N N
å rkφk (t) = å sikφk (t) + å nkφk (t)
k =1 k =1 k =1

r = si + n

Since r consists of a combination of the deterministic waveform si and probabi-

listic signal n, our attention now turns to mathematically characterizing n. Since the
noise signal n(t) is assumed to be a Gaussian random variable, we need to determine
how the characteristics of this random variable translates into a signal vector repre-
sentation. We know that the noise vector element nk is equal to:

T
nk = ò n(t)φk (t)dt (4.71)
0

which is the projection of the noise signal waveform on the orthonormal basis
function fk(t). Since the noise signal n(t) is a Gaussian random variable and the
integration process is a linear operation, this means that nk is a Gaussian random
variable as well. Thus, the noise signal vector n is a Gaussian vector. Let us now
proceed with determining the statistical characteristics of n in order to employ this
knowledge in signal waveform detection.
First, we would like to calculate the mean of these vector elements. Thus, by
applying the definition for the expectation, this yields:

Figure 4.12  Simple digital transceiver model.

4.6  Optimal Detection 115

 T 
E{nk } = E  ∫ n(t)φk (t)dt 
 0 
T
= ∫ E{n(t)}φk (t)dt
0

=0 (4.72)
since since E{n(t)} = 0, which ultimately means that the mean of the noise signal
vector is E{n} = 0.
The next step is to calculate the variance of these vector elements. Suppose we
let (nnT)kl = nknl be equal to the (k, l)th element of nnT. Therefore, in order to deter-
mine E{nknl}, where nk and nl are defined by:
T T
nk = ∫ n(t )φ k (t )dt , nl = ∫ n( ρ)φl (ρ)dρ
0 0

we can apply the definition for E{nknl}, which yields:

ïìæ ö æT ö ïü
T
E{nk nl } = E íç ò n(t)φk (t)dt ÷ ç ò n(ρ)φl (ρ)dρ ÷ ý
ïîè 0 øè0 ø ïþ

ìïT T üï
= E í ò ò n(t)n(ρ)φk (t)φl (t)dtdρ ý
îï 0 0 þï
Solving E{nknl} yields:
TT
E{nk nl } = ò ò E{n(t)n(ρ)}φk (t)φl (t)dtd ρ
0 0
TT
N0
= òò 2
δ (t - ρ)φk (t)φl (t)dtd ρ
0 0
T
N0
=
2 ò φk (t)φl (t)dt
0

N0
= δ (k - l),
2 (4.73)
where the integration of the product of the two orthonormal functions fk(t) and
fl(t) yields a delta function since only when k = l do these two functions project onto
each other. As a result, the matrix equivalent of this outcome is equal to:
N0
E{nnT } =I N × N (4.74)
2
Given the vector representation of the Gaussian random variable obtained in (4.74),
we need to define the joint probability density function of this representation in order
to characterize the individual elements of this vector. Leveraging the assumption that
the noise elements are independent to each other, we can express the joint probability
density function as the product of the individual probability density functions for each
element, yielding:
116 Digital Transmission Fundamentals

N
1 2 2
p(n) = p(n1 , n2 ,… , nN ) =
(2 πσ 2) N / 2
∏ e − n / 2σ
i

i=1

= p(n1)p(n2)…p(nN)
1 − n2 / 2σ 2
where p(ni ) = e i is the probability density function for the vector element
σ 2π N 2 N0
ni. Since we know that E{n kn l } = 0 δ (k − l) we can then solve E{nk } = = σ 2.
2 2
Additionally, we know that the dot product of a vector can be written as the sum-
mation of the squared elements, namely:
N
∑ ni2 = ||n||2
i=1 (4.75)
which can then be used to yield the following expression for the joint probability
density function:
1 2 2
p(n) = p(n1 , n2 ,… , nN ) = 2 N /2
e − || n || / 2σ
(2π σ ) (4.76)

4.6.2  Decision Rules

With the formulation of the joint probability density function derived in (4.76), we
can now define a rule for the receiver that can be used to determine which signal
waveform is being intercepted given the presence of some noise introduced by the
channel. Suppose we define the following criterion for the receiver as:

Minimize P(error) → P(m

ˆ i ≠ mi)
Maximize P(correct) → P(m
ˆ i = mi) (4.77)

where the probability of error is P(e) = P(error), the probability of correct reception
is P(c) = P(correct), and P(e) = 1 – P(c) is the complementary relationship between
these two probabilities. Then, using the law of total probability, the overall prob-
ability of correct detection is equal to:

P (c ) = ∫ P(c| r = ρ) p(ρ)dρ
V (4.78)
where P(c|r = ³) ³ 0 and p(ρ) ³ 0. Therefore, we observe that when P(c) attains a
maximum value, this occurs when P(c|r = ρ) also possesses a maximum value.
In order to maximize P(c|r = ρ), we use the following decision rule at the
receiver:

P(sk | ρ ) ≥ P ( si | ρ), for i = 1, 2 , … , M and i ≠ k (4.79)

for i = 1,2,..., M and i ¹ k. Note that for this decision rule we are assuming that sk
is present in ρ such that:

ρ = sk + n → m
ˆ = m k (4.80)
4.6  Optimal Detection 117

Employing a mixed form of Bayes Rule that is composed of probability density

functions and probabilities, namely:
p(ρ |s i )P(s i )
P(s i |r = ρ ) =
p( ρ ) (4.81)
we would like to manipulate this decision rule into a formulation that can be
­employed by a receiver. Specifically, recall how we wanted to maximize P(c|r = ρ)
earlier in this section. By employing the mixed Bayes Rule formulation, the optimal
detector can be rewritten such that it is equal to:
p(ρ|s i )P(s i )
max P(s i |r = ρ) = max
si si p(ρ) (4.82)
for i = 1,2,..., M. Since p(ρ) does not depend on si, we can simplify the optimal de-
tector expression such that:
max p(ρ|s i )P(s i )
si (4.83)
for i = 1,2,..., M.
Based on our result in (4.83), two types of detectors can be derived based on
this expression. The first type of detector is referred to as a maximum a posteriori
(MAP) detector, which can be expressed as:

P(s i |r = ρ) = max p (ρ| s i )P(s i ) (4.84)

si
1
for i = 1,2,..., M. However, in the event that P(s i ) =
, which implies that P(si) does not
M
depend on si, we can omit the P(si) term from the optimal detector expression, yielding
the second type of detector, referred to as a maximum likelihood (ML) detector:

P(s i |r = ρ) = max p(ρ| s i ) (4.85)

si

for i = 1,2,..., M. In the next section, we will mathematically derive the maximum
likelihood detector given the optimal decision rule for data transmissions being
performed across AWGN channels.

4.6.3 Maximum Likelihood Detection in an AWGN Channel

Maximum likelihood detection is a popular statistical method employed for fitting a sta-
tistical model to data and for identifying model parameters. In general, for a fixed set of
data and underlying probability model, a maximum likelihood approach selects values
of the model parameters that produce the distribution that are most likely to have re-
sulted in the observed data (i.e., the parameters that maximize the likelihood function).
Suppose that a data transmission is operating across an AWGN channel prior
to interception by the receiver. Recall that the transmission model for this scenario
is given by:

r = si + n (4.86)
where si is the ith signal waveform sent by the transmitter, n is the noise introduced
to the data transmission by the AWGN channel, and r is the intercepted signal
118 Digital Transmission Fundamentals

waveform by the receiver. Given that si is a deterministic quantity, and n is a ran-

dom entity that has just been characterized by a joint probability density function,
what is needed now is a characterization of r, which can be derived from the char-
acterization of n coupled with the deterministic properties of si.
Suppose we consider the conditional probability of a single element of the re-
ceived vector r = ρ, say the kth element, given that the signal waveform si was as-
sumed to be transmitted:
1 2 2
p(ρk |s ik) = e -(ρk -s ik) / 2 σ
2πσ 2 (4.87)
where the kth element of the noise vector is equal to nk = rk – sik. Since we assume
that the AWGN vector elements are uncorrelated, (i.e., independent), we can re-
write this conditional probability expression as:
N
p(ρ|s i ) = Õ p(ρk |s ik), for i = 1, 2,…, M
k =1 (4.88)
Consequently, this product of multiple elemental probability density functions will
ultimately yield the following expression:
1 2 2
p(ρ|s i ) = 2 N /2
e -||ρ -s i || / 2σ
(2πσ ) (4.89)
Notice how we now have a formulation for the conditional probability that is
entirely represented in terms of si,ρ, and their respective elements. Leveraging this
expression, we can proceed with mathematically determining the maximum likeli-
hood detector.
Since we would like to solve for max p(ρ | s i ), suppose we take the expression
si
for p(ρ|si), apply it to the detector, and take the natural logarithm of the resulting
expression. Performing these operations would yield the following:
N æ 1 ö || ρ - s i ||2
ln(p(ρ|s i )) = ln ç ÷- (4.90)
2 è 2πσ 2 ø 2σ 2
Note that the natural logarithm was employed in order to get rid of the exponential base
in the expression, thus yielding a linear expression for the optimal decision rule. Fur-
thermore, since natural logarithms are monotonic functions (i.e., if x2 ³ x1 then ln(x2) ³
ln(x1)), the decision rule would still remain valid when the inequality is employed.
Solving for this linear decision rule, and given the monotonic behavior of the
natural logarithm, we can derive the following:
æ N æ 1 ö ||ρ - s i ||2 ö
maxln(p(ρ|s i )) = max ç ln ç -
si si è 2 è 2πσ 2 ø÷ 2σ 2 ø
÷

æ ||ρ - s i ||2 ö
= max ç - ÷
si è 2σ 2 ø

= max æçè -||ρ - s i ||2 ö÷ø

si

= min||ρ -s i || (4.91)
si
4.7  Basic Receiver Realizations 119

Since we are interested in the choice of si that yields the maximum value for the
decision rule, we can rewrite this decision rule as:

s k = arg min ||ρ - s i || ® m

ˆ =m
si (4.92)
Note that one of the advantages of employing a vector representation for these
decision rules is that the entire scenario can be interpreted in terms of distance. Spe-
cifically, the term ||ρ – si|| actually represents the distance between the heads of two
vectors, ρ and si, whose tails are located at the origin. Thus, a maximum likelihood
detector is the equivalent of a minimum distance detector.

Referring to the signal constellation diagram shown in Figure 4.13,

implement a simple QPSK transceiver operating in an AWGN channel
and implement a maximum likelihood detector. Does this decision
Q rule match the decision regions in Figure 4.13? Does this decision rule
adequately declare si as the transmitted signal based on which quadrant ρ
appears in? What is the impact of the AWGN channel for different values
for the variance given that the noise is zero-mean?

4.7 Basic Receiver Realizations

The fundamental challenge of digital communications is recovering what was trans-

mitted after it has passed through a channel and been corrupted by noise. The first
receiver structure we will examine is based on filtering the received signal with a
static filter that maximizes the SNR of the channel, which will subsequently mini-
mize the bit error rate. However, one of the disadvantages of a matched filtering
approach is that it requires a priori knowledge of all the possible signal waveforms
sent by the transmitter.

Figure 4.13  Decision regions for QPSK signal constellation.

120 Digital Transmission Fundamentals

4.7.1 Matched Filter Realization

When designing a receiver, we are interested in a decision rule where the receiver
yields the correct result more often than any other decision rule that can be em-
ployed by the receiver. In other words, we are interested in detecting a pulse trans-
mitted over a channel corrupted by noise.
Suppose we employ the following transmission model:

x (t) = g(t) + w(t), 0 ≤ t ≤ T (4.93)

where g(t) is a pulse signal, w(t) is a white noise process with mean m = 0 and power
spectral density equal to, N 0 and x(t) is the observed received signal. Assuming
2
the receiver knows all the possible waveforms of g(t) produced by the transmitter,
the objective of the receiver is to detect the pulse signal g(t) in an optimum manner
based on an observed received signal x(t). Note that the signal g(t) may represent a
“1” or a “0” in a digital communication system.
In order to enable the receiver to successfully detect the pulse signal g(t) in an
optimal manner given the observed received signal x(t), let us filter x(t) such that
the effects of the noise are minimized in some statistical sense so the probability of
correct detection is enhanced. Suppose we filter x(t) using h(t) such that the output
of this process yields:

y (t) = g0 (t ) + n(t ) (4.94)

where n(t) is the result of the noise signal w(t) filtered by h(t) and g0(t) is the filtered
version of g(t) by h(t). The transmission model and filtering operation by h(t) is il-
lustrated in Figure 4.14.
Let us rewrite this filtering operation in the frequency domain, where the time-
domain convolution operations become frequency-domain products. Thus, taking
the inverse Fourier transform of H(f)G(f), which is equivalent to a convolution of
h(t) and g(t), we get the following expression for the filtered version of g(t):
∞ j 2π ft
g 0 (t) = ∫ −∞ H (f )G(f )e df (4.95)

where the inverse Fourier transform returns the filtering operation back to the time
domain.
Let us now calculate the instantaneous power of the filtered signal g0(t), which
is given as:
¥
|g 0 (t )|2 = |ò H (f )G(f )e j 2π ft df |2
-¥ (4.96)

Figure 4.14  Filtering process for detecting g(t).

4.7  Basic Receiver Realizations 121

In order to determine a quantitative metric that would indicate when we have

achieve the goal of maximizing g0(t) relative to n(t), let us employ the peak pulse
SNR, which is defined as:
|g (T )|2
η= 0 2 (4.97)
E{n (t)}
where |g0(T)|2 is the instantaneous power of the output signal at sampling instant T,
and E{n2(t)} is the average power of the output noise. Thus, goal of this matched
filter realization is to maximize g0(t) with respect to n(t) using the peak pulse SNR
metric, which can be achieved by designing a filter h(t) that can yield the largest
possible value for η.
In order to design h(t), we need to mathematically solve for h(t), which consists
of evaluating the expression:
¥ 2

ò
2 j 2π ft
|g 0 (t)| = H (f )G(f )e df (4.98)
-¥
which is the magnitude squared of the inverse Fourier transform of H(f )G(f ) =
N
F{h(t)*g(t)}. Since w(t) is a white Gaussian process with power spectral density 0 ,
2
we know that by the EWK Theorem that the power spectral density of the filtered
N0
noise signal n(t) is equal to SN (f ) = | H (f ) | 2. Therefore, applying the definition
2
for η and including these expressions will yield:
¥
|ò H (f )G(f )e j 2π fTdf |2
-¥
η=
N0 ¥
2 ò-¥
| H (f )|2df (4.99)

From this resulting expression, we see that we need to solve for frequency
­response H(f ) such that it yields the largest possible value for the peak pulse SNR η.
In order to obtain a closed-form solution let us employ Schwarz’s Inequality. Sup-
pose that we have two complex functions, say, φ1(x) and φ2(x), such that:
¥ ¥
ò |φ1(x )|2 dx < ¥ and ò |φ2 ( x)|2 dx < ¥ (4.100)
-¥ -¥
Then, by Schwarz’s Inequality we can rewrite the following integral expression as
an inequality:
2
¥ æ¥ ö æ¥ ö
ò ò ò
2 2
φ 1 (x )φ 2 ( x) dx £ ç |φ 1 ( x)| dx ÷ × ç |φ 1 ( x)| dx ÷ (4.101)
-¥ è -¥ ø è -¥ ø

with this expression becoming an equality when φ 1(x ) = K ⋅ φ∗2 ( x).

Therefore, leveraging Schwarz’s Inequality in our expression for the peak pulse
SNR, it can be shown that the numerator of (4.99) can be rewritten as:
2
¥ æ¥ ö æ¥ ö
ò
j 2π ft
H (f )G(f )e df £ ç ò |H (f )|2 df ÷ × ç ò |G(f )|2 df ÷ (4.102)
-¥ è -¥ ø è -¥ ø

which then yields the following inequality for η:

122 Digital Transmission Fundamentals

¥
2
η£
N0 ò |G(f )|2 df (4.103)
-¥
Thus, in order to make this expression an equality, the optimal value for H(f) should
be equal to:

H opt (f ) = K ⋅ G∗ (f )e− j 2π fT (4.104)

whose time-domain representation can be mathematically determined using the
­inverse Fourier transform:
¥
hopt (t ) = K × ò G* (f )e- j2π fTe - j2 π ftdf = K × g(T - t) (4.105)
-¥
Notice that when we are performing a matched filtering operation, we are con-
volving the time-flipped and time-shifted version of the transmitted pulse with the
transmitted pulse itself in order to maximize the SNR.

The reason that we call these filters matched filters is because when we
convolve the time-flipped and time-shifted version of the transmitted
pulse signal with itself, the process is SNR maximizing. Consequently,
i if a receiver intercepts some unknown noise-corrupted signal, it can
readily identify which one was sent by a transmitter by matching this
intercepted signal to all the available signal waveforms known at the
receiver using an implementation illustrated in Figure 4.15.

Referring to Figure 4.16, suppose we have a signal g(t). Show that

Q h(t) and g0(t) are the corresponding matched filter and filtered output
signals.

4.7.2  Correlator Realization

Recall that a matched filter realization assumes some sort of knowledge regarding
the transmitted data. However, if the receiver possesses this information about the

Figure 4.15  Schematic of matched filter realization of receiver structure.

4.7  Basic Receiver Realizations 123

(a)

(b)

(c)
Figure 4.16  An example of the time-domain input and output signal waveforms processed by a
matched filter. (a) Time-domain representation of the input signal to the matched filter. (b) Time-
domain impulse response of the matched filter. (c) Time-domain representation of the output signal
of the matched filter.

transmitter and its signal characteristics, it is also possible to employ a more sta-
tistical approach for determining which signal waveforms have been sent, even in
the presence of a noisy, corruption-inducing channel. Specifically, we can employ
the concept of correlation such that we only need to assume knowledge about the
waveforms themselves.1
Suppose we start with the decision rule derived at the beginning of this section
and expand it such that:

min ||ρ - s i ||2 = min(ρ - s i ) × (ρ - s i )

si si

= ρ × ρ - 2 ρ × s i + s i × s i (4.106)
Since ρ ⋅ ρ is common to all the decision metrics for different values of the signal
waveforms si , we can conveniently omit it from the expression, thus yielding:

min (−2 ρ ⋅ si + si ⋅ si ) = max (2ρ ⋅ si − si ⋅ si ) (4.107)

si si

where ρ ⋅ si and si ⋅ si are defined by:

1.  or a matched filtering implementation, knowledge of both the transmission signal waveforms and
F
the statistical characteristics of the noise introduced by the channel is needed by the receiver.
124 Digital Transmission Fundamentals

Figure 4.17  Correlator realization of a receiver structure assuming perfect synchronization.

T T
ρ ⋅ si = ∫ ρ (t )s i (t )dt si ⋅s i = ∫ si2 (t )dt = E si
0 0

We can observe that the waveform representation of ρ ⋅ si is equal to the correla-

tion of r(t) = p(t) with respect to si(t). Thus, when sk(t) is present in r(t), the optimal
detector is equal to:
 T
Esi 
s k = arg max 
i 
∫ ρ(t )s i (t )dt −
2 
 (4.108)
0

Based on this result, we can design a receiver structure that leverages correla-
tion in order to decide which signal waveform was sent by the transmitter based on
the observed intercepted signal at the receiver. A schematic of a correlation-based
implementation is shown in Figure 4.17. Given r(t) = si(t) + n(t) and we observe only
r(t) = p(t) at the input to the receiver, we first correlate r(t) with si (t) across all i.
Next, we normalize the correlation result by the corresponding signal energy Esi
in order to facilitate a fair comparison. Note that if all energy values are the same
for each possible signal waveform, we can dispense with the energy normalization
process since this will have no impact on the decision making. Finally, the resulting
decision values for each of the branches are compared against each other, and the
branch with the largest resulting value is selected.

4.8  CHAPTER SUMMARY

A deep and thorough understanding of digital communication theory is vitally es-

sential when designing and evaluating software-defined radio implementations. In
this chapter, an overview of several useful topics, including several different types
of modulation schemes, the derivation of the probability of error, Gram-Schmidt
orthogonalization, formulation of the optimal decision rule, and two receiver struc-
tures were studied in order to provide the reader with the fundamentals needed in
order to master these versatile yet complex systems. In the subsequent chapters,
the fundamental knowledge obtained in this chapter as well as the two previous
4.10  Problems 125

chapters will be leveraged extensively when implementing digital communication

systems and networks based on software-defined radio technology.

4.9 ADDITIONAL READINGS

Given the introductory nature of this chapter with respect to the topic of digital
communications, the interested reader is definitely encouraged to explore the
numerous books that provide a substantially more detailed and advanced treat-
ment of this topic. For instance, the latest edition of the seminal digital commu-
nications textbook by Proakis and Salehi [4] provides a rigorous, mathematical
treatment of many of the concepts covered in this chapter, in addition to many
other topics not presented, such as spread spectrum technologies, equalization,
and RAKE receiver implementations. To complement this mathematically terse
document, the authors also published a companion textbook that treats digital
communications from a more applied perspective, including examples in MATLAB
and Simulink [5].
As for introductory textbooks on digital communications, Sklar wrote an excel-
lent document that provides the reader with a balance of mathematical rigor, detailed
explanations, and several well-crafted examples [6]. The textbook by Couch is also
in the same category as Sklar, but it treats both analog and digital communications
[7], which is well suited for individuals who do not possess a background in the
communications field. Rice wrote his introductory textbook on digital communica-
tions from a discrete-time perspective, which is suited for an individual possessing
a background in discrete-time signal and systems [8]. The textbook also provides
numerous end-of-chapter problems as well as MATLAB examples available online,
providing the reader with many opportunities to practice the theoretical concepts
covered within this text.
The classic digital communications book by Barry, Messerschmitt, and Lee [9]
is an excellent reference textbook for those individuals who possess some under-
standing about digital communications but need convenient and immediate access
to detailed information. Similarly, the textbooks by Madhow [10] and Pursley [11]
both provide readers with a more advanced treatment of digital communication
theory. Finally, the book by Hsu [12] is an excellent reference textbook that consists
mostly of a popular collection of solved problems.

4.10  Problems

1. Power efficiency: Find and compare the power efficiency for the following
three binary signal sets. Give the relative performance in decibels. Assume
the s1(t) and s2(t) are equally likely.
(a) s1(t) = Bsin(ω0t + φ) and s2(t) = Bsin(ω0t – φ) for 0 ≤ t ≤ T and where
2
Eb ≤ A2T . Find the best (B,φ).
A2T
(b) s1(t) = Asin(ω0t + θ) and s2(t) = Bsin(ω0t) for 0 ≤ t ≤ T and where Eb ≤ 02
and A0 is known. Find the best (A, B, θ).
126 Digital Transmission Fundamentals

(c) s1(t) = Asin(2 – pf0t + Dt) and s2(t) = Acos(2 – pf0t – Dt – θ) for 0 £ t £ T.

Find the best peak frequency deviation, D, in terms of T.
2. Decision boundaries: Three different eight-point constellations are proposed
as shown in Figure 4.18. Draw the appropriate decision boundaries for each
technique in two dimensions, and express the probability of symbol error
Ps in terms of the peak energy-to-noise density ratio. Repeat for an average
energy normalization. Which of the demodulators would be easier to imple-
ment? Justify.
3. Power efficiency: The signal constellation for a communications system with
16 equiprobable symbols is shown in Figure 4.19. The channel is AWGN
N0
with noise power spectral density of . Compute and compare the power
2
efficiency of this system with a 16-level PAM system with adjacent amplitude
differences of 2A.
4. Matched filtering: Consider the set of M = 4 waveforms shown in Figure 4.20.
(a) Apply the Gram-Schmidt orthogonalization procedure to obtain a set of
K basis functions that spans the set of signals.
HINT: Use the order s3(t), s0(t), s2(t), and s1(t).
(b) What is the dimension of the signal space that contains these four signals?
(c) Plot the signal constellation corresponding to the signal set.

Figure 4.18  Three eight-point signal constellations (8-PSK, “7-around-1,” and “Box”).
4.10  Problems 127

Im

3A

A Re

Figure 4.19  16-QAM signal constellation.

(d) Suppose that the symbol period is T = 1; sketch the optimal matched

filter designs h0(t), h1(t), h2(t), and h3(t) for each of these signals.
(e) Sketch the matched filter realization when the receiver intercepts a
signal r(t).
NOTE: The received signal r(t) can be composed of s0(t), s1(t), s2(t), and
s3(t) signal waveforms that have been transmitted through an additive
white Gaussian noise (AWGN) channel.
(f) Sketch the output of the matched filters h0(t), h1(t), h2(t), and h3(t) when
the input is equal to h0(t).

Figure 4.20  Collection of M = 4 waveforms employed in a digital transceiver.

128 Digital Transmission Fundamentals

5. Error analysis: A ternary communication system transmits one of three

equiprobable signals s(t), 0, or −s(t) every T seconds. The received signal is r(t) =
s(t) + z(t), r(t) = z(t), or r(t)= −s(t) + z(t), where z(t) is white Gaussian noise with
E[z(t)] = 0 and Rz(τ) = E[z(t)z*(t)] = 2N0δ(t – τ).
The optimum receiver computes the correlation metric:
T 
U = Re  ∫ r (t)s∗ (t)dt  (4.109)
 0 
a nd compares U with a threshold A and a threshold −A. If U > A, the deci-
sion is made that s(t) was sent. If U < −A, the decision is made in favor of
−s(t). If −A <U < A, the decision is made in favor of 0.
1 T
(a) Suppose we define E = ∫ s(t)s∗ (t)dt and N = Re  ∫ z(t)s∗ (t)dt , where
T

2 0 
 0 
N is a Gaussian random variable. Solve for the mean and variance of N.
(b) Determine the three conditional probabilities of error: Pe given that s(t)
was sent, Pe given that −s(t) was sent, and Pe given that 0 was sent.
HINT: Solve in terms of E, N, and A, as well as express the answer
using the Q-function:
¥
1
Q( x) =
2π ò e-u/2 du (4.110)
x

(c) Determine the average probability of error P e as a function of the thresh-

old A, assuming that the three symbols are equally probable a priori.
(d)  Determine the value of A that minimizes P e.
HINT: Use Leibnitz Rule:

d  ∞ 
  = − df g( f ( x))

dx  ∫ g (a)da
 dx
(4.111)
f (x)

6. Maximum a posteriori detection: Suppose our transmitter is operating in a

zero-mean additive white Gaussian noise (AWGN) channel of variance σ n2
such that the received signal vector is r = si + n, for i = 0,1, 2, 3.
(a) What is the mathematical expression for the conditional probability den-
sity function (CPDF) p(r|si)?
HINT: Use the mathematical expression for a zero-mean Gaussian ran-
dom variable.
(b) What is the general expression for the maximum a posteriori (MAP)
detector?
HINT: Recall hybrid presentation of Bayes Rule with respect to the ­usage
of both PDFs and probabilities.
(c) Using the answer from part (a), give the correlator realization of this
receiver structure based on part (b) when the probabilities of the trans-
mitted signals are P(s0) = P(s1) = 2P(s2) = 3P(s3). Draw the receiver
structure.
HINT: Use log-likelihood functions with the expression from part (a).
4.10  Problems 129

Figure 4.21  Four signaling waveforms employed in binary pulse amplitude modulation scheme.

7. Modulation and power efficiency: Suppose we are given the four waveforms
shown in Figure 4.21 as candidates for a binary pulse amplitude modulation
(PAM) modulation scheme. Let us consider me signal pairs P1 : (s1(t),s2(t)),
P2: (s1(t),s3(t)), and P3 : (s1(t), s4(t)).
2
(a) Determine the minimum Euclidean distance, dmin , for signal pairs P1,
P2, and P3.
(b) Calculate the average energy per bit, Eb, for signal pairs P1, P2, and P3.
Note that the pair of signals represent either a binary 1 or a binary 0, each
of which has an equal probability of occurring in the transmission.
(c) Of the three signal waveform pairs, which possesses the best power
­efficiency ∈p and explain in your own words why this is the case.
8. Modulation and power efficiency:
(a) Calculate the average symbol energy, E s, for the tertiary signal constella-
tion shown in Figure 4.22. Assume all three symbols are equiprobable.
2
(b) Calculate the minimum Euclidean distance, dmin , for the tertiary signal
constellation shown in Figure 4.22.
(c) Suppose that each signal constellation point in Figure 4.22 requires b = 2 bits
in order to achieve unique representation. Find the power efficiency, ∈p.
9. Modulation and power efficiency:
(a)  Given the four waveforms in Figure 4.23, compute the Euclidean dis-
tance between all possible pairs of waveforms:
2
dij2 = ∫ (si (t ) − s j (t ))2 dt , i ≠ j (4.112) 2
dmin
0

Figure 4.22  Tertiary signal constellation.

130 Digital Transmission Fundamentals

Figure 4.23  Four potential waveforms available for modulation.

(b) Suppose that you are designing a modulator that maps b = 1 bit into
a waveform and that you are given the responsibility of choosing two
waveforms from the available four waveforms shown in Figure 4.23.
To achieve the best possible result when operating in an additive noise
channel, which two waveforms would you choose? Justify your answer.
HINT: Use you results in part (d).

References

[1] Shannon, C. E., “Communication in the Presence of Noise” Proceedings of the Institute of

Radio Engineers, Vol. 37, No. 1, January, 1949. p. 1021.
[2]  Poor, An Introduction to Signal Detection and Estimation, Springer, 2010.
[3] Wyglinski, A. M., “Physical Layer Loading Algorithms for Indoor Wireless Multicarrier
Systems”. PhD Thesis of McGill University. Montreal, Canada, 2004.
[4]  Proakis, J. and M. Salehi, Digital Communications, 5th Edition, McGraw-Hill, 2007.
[5] Proakis, J. G., M. Salehi, and G. Bauch, Contemporary Communication Systems Using
MATLAB, 2nd Edition. Brooks/Cole, 2003.
[6] Sklar, B., Digital Communications: Fundamentals and Applications, 2nd Edition, Prentice
Hall PTR, 2001.
[7]  Couch, W., Digital and Analog Communication Systems, 7th Edition, Prentice Hall, 2006.
[8] Rice, M., Digital Communications: A Discrete-Time Approach, 3rd Edition, Pearson/
Prentice Hall, 2009.
[9] Barry, J. R., D. G. Messerschmitt, and E. A. Lee, Digital Communication, 3rd Edition,
Springer, 2003.
[10] Madhow, U., Fundamentals of Digital Communication, Cambridge University Press,
2008.
[11]  Pursley, M. B., Introduction to Digital Communications, Prentice Hall, 2004.
[12]  Hsu, H. P., Schaum’s Outline of Theory and Problems of Analog and Digital Communica-
tions, 2nd Edition, McGraw-Hill, 2002.
 Chapter 5

Basic SDR Implementation of a

Transmitter and a Receiver

This chapter will serve as an introduction to understanding how communication

algorithms operate under real-world conditions using software-defined radio (SDR)
technology. This chapter will also provide the foundation for subsequent digital
communication algorithms and their implementation in SDR, which will be dis-
cussed later in this book. In this chapter, we will get our first hands-on experience
with SDR by implementing a bare bones communication system in Simulink. For
information about the theoretical background for the experiments in this chapter,
please refer to Section 4.1 through Section 4.3, which cover the concepts of encod-
ing, modulation, and probability of bit error. For information about MATLAB and
Simulink products, please refer to Appendix A.

5.1  Software Implementation

Before implementing any designs in SDR, it is imperative to first understand any

design or implementation using a computer simulation package in order to better
characterize the behavior of the implementation, as well as provide a set of bench-
mark results that can be used when constructing these prototypes in actual hard-
ware. Furthermore, many communications engineers and researchers approach the
implementation of a communication system design via a methodological sequence
of computer simulation followed by hardware implementation.
In order to get things started, let us first open the txbasic.mdl file,1 as shown
in the Figure 5.1, which contains a very basic digital transmitter model in Simulink.
Note that there is no channel in this model, and thus there are no impairments
intro­duced to the data stream being produced by this transmitter. Random “1”
and “0” values are generated by the transmitter, then sent using 36 bit long frames,
which are processed by an encoding scheme to make them more robust to noise.
Following the encoding, this bit stream is then modulated to a 16 QAM symbol
format prior to transmission.
As mentioned before, although we can see all of the blocks of this system, it
is not entirely complete. Specifically, the user-defined functions are left empty, and
throughout this chapter we will proceed to fill in the remaining blocks in order to
complete this system as part of the experiment defined in this chapter. Before we
start writing code, we should examine the assigned parameters of the defined blocks
in the mask parameters and attempt to understand why we have selected those
values. For example, Figure 5.2 shows the mask parameters of the Bernoulli Binary
Generator block, which can be accessed by double-clicking the block.

1. The experiment files used in this book are available at [1]. 131
132 Basic SDR Implementation of a Transmitter and a Receiver

Figure 5.1  A very basic digital transmitter model in Simulink.

When working in Simulink, the format of the data is extremely important. For
example, if a [36 ´ 1] signal is sent, to an input port expecting a [37 ´ 1] signal,
as highlighted in Figure 5.2, not only will the model not run, the software will
produce numerous error messages, as shown in Figure 5.3. Understanding how
each block affects the data format will greatly simplify the design process of a
communication system when we use Simulink.

5.1.1 Repetition Coding
One of the key building blocks of any communication system is the forward error
correction (FEC), where redundant data is added to the transmitted stream to make

Figure 5.2  Mask parameters of the Bernoulli Binary Generator block, where the samples per frame
is 36.
5.1  Software Implementation 133

Figure 5.3  Error messages from Simulink. By clicking on the error message, a detailed description
of the source of this error will be provided in the lower window to help with the debugging
process.

it more robust to channel errors. There are many types of FEC techniques, such
as the repetition coding approach, where each transmitted bit is repeated multiple
times. For example, Problem 1 in Section 5.5 introduces a repetition code with rep-
etition factor R = 5 in a binary pulse amplitude modulation (BPAM) transmission
system. In this section, we will explore together one technique for combating the
introduction of errors to data transmissions by implementing a simple repetition
coder (repetition factor R = 4). So what does it mean by a repetition coder with
repetition factor R = 4? A simple definition is that if a “0” symbol is to be transmit-
ted, this “0” symbol will be repeated four times by the repetition coder, such that
the output would be “0000”.
Let us first start by double clicking on the repetition coder block, which will
result in a MATLAB function block editor opening, in which we can write cus-
tomized MATLAB code. As mentioned previously, setting break points is a great
way for understanding and debugging M-files. For more information about break
points and how they can be used to debug and evaluate the code, please refer to
Appendix A.

The repmat function in MATLAB can be used to realize a simplistic

repetition coding scheme. For example, to repeat a vector u for 4
i times, the following expression can obtain this result:
y=repmat(u,4,1);

What are the trade-offs to consider when choosing between a high

Q or a low repetition factor?
134 Basic SDR Implementation of a Transmitter and a Receiver

5.1.2  Interleaving
A repetition code is one of several useful tools for a communication systems engi-
neer in order to enhance a robust data transmission. However, it is sometimes not
enough, since it does not address the issue when a large quantity of data is cor-
rupted in contiguous blocks. For instance, if a transmitter sends the data stream
“101101”, a repetition coder with a repetition factor of 4 will yield:

111100001111111100001111,
where each input bit is repeated four times. While this encoding scheme may appear
robust to error, it is still possible during a data transmission that a significant noise
burst occurs over many consecutive bits, corrupting numerous binary digits in the
transmission, and yields the following outcome:

111100 ----------------------- 1100001111,

where some of the original data is completely irretrievable.

Why is it that even with repetition coding, our data transmission

Q can still be severely affected? What could be done to make it even
more robust?

Interleaving is an approach where binary data is reordered such that the correla-
tion existing between the individual bits within a specific sequence is significantly
reduced. Since errors usually occur across a consecutive series of bits, interleaving
a bit sequence prior to transmission and de-interleaving the intercepted sequence at
the receiver allows for the dispersion of bit errors across the entire sequence, thus
minimizing its impact on the transmitted message. A simple interleaver will mix up
the repeated bits to make the redundancy in the data even more robust to error. It
reorders the duplicated bits amongst each other to ensure that at least one redun-
dant copy of each will arrive even if a series of bits are lost. For example, if we use
an interleaving step of 4, it means we reorder the vector by index [1, 5, 9, ..., 2,
6, 10, ...]. As a result, running “111100001111111100001111” through such an
interleaver will yield the following output:

101101101101101101101101.
The interleaving step can be any of the factoring numbers of the data length. How-
ever, different mixing algorithms will change the effectiveness of the interleaver.

The reshape function in MATLAB can be used to realize the inter-

i leaving.

Once we have implemented the interleaver, let us combine the repetition coder
and the interleaver into a single FEC sub-system. Although the simple interleav-
ing technique introduced above is sufficient for our implementation, there are
5.1  Software Implementation 135

various other forms of interleaving, that we will investigate in the next two sub-
sections.

5.1.2.1  Block Interleaving

The first approach to interleaving is to employ a block interleaver, as shown in
­Figure 5.4. Block interleaving is one method for reordering a bit sequence, where
N × M bits fill an N column by M row matrix on a column basis, and then each
resulting row is concatenated with each other in serial and outputted from the
interleave block. At the transmitter side, the block interleaver is loaded column
by column with N codewords, each of length M bits. These N codewords are then
transmitted row by row until the interleaver is emptied. Then the interleaver is
loaded again and the cycle repeats. The main drawback of block interleavers is the
delay introduced with each column-by-column fill of the interleaver [2].

5.1.2.2  Convolutional Interleaving

Another approach to interleaving is to employ a convolutional interleaver [3], as
shown in Figure 5.5. At the transmitter, the bit sequence is shifted into a bank of
N registers, each possessing an increasing amount of buffer memory. The bits in
the bank of registers are then recombined via a commutator and transmitted across
the channel. At the receiver, the reverse process is performed in order to recover the
original sequence. Compared with block interleavers, convolutional interleavers
reduce memory requirements by about one-half [4]. However, the delay problem
associated with the initial fill still exists.

5.1.3  BER Calculator

After examing several techniques for protecting the reliability of a data trans-
mission subject to interference and noise, we now need an approach to measure

Figure 5.4  Schematic of a block interleaver.

136 Basic SDR Implementation of a Transmitter and a Receiver

Figure 5.5  Schematic of a convolutional block interleaver (from [3]).

how well these techniques perform quantitatively. Referring back to Chapter 4,

we saw that bit error rate (BER) is a commonly used metric for the evaluation
and comparison of digital communication systems. One straightforward way of
calculating the BER is to count the number of received bits of information and
then determine which ones are received in error. In other words, the ratio of bit
errors to the total number of bits received can provide us with an approximate
BER calculation. Note that the more bits that are received, the more accurate this
metric becomes.
In this model, the alignment block is accounting for the delay of the signal as it
propagates through the other blocks. As the model is configured, the BER should
be zero. Check this by connecting both the input ports of the alignment block to
the initial bit sequence.

5.1.4 Receiver Implementation over an Ideal Channel

The next step is to implement the receiver for the transmitter we have just con-
structed. We can now remove the AWGN Channel block for this step and assume
this is an ideal channel. Therefore, there is no noise and therefore no impairments to
account for in the data transmission. The receiver design is simply the inverse of the
transmitter design, as shown in Figure 5.6. To test out our implementation, let us
feed the decoded signal into the BER calculator. Since we are using an ideal channel,
we should expect that the resulting BER is equal to zero because no interference nor
distortion is present during transmission.

1. Note the two persistent variables - what does persistent mean?

2. Once you believe you have achieved success, intentionally invert
Q one of the bits. Does your BER calculated behave as expected?
3. What effect might different signal constellations have on the bit
error rate?
5.2  USRP Hardware Implementation 137

Figure 5.6  Approximate appearance of final Simulink transceiver model. This model implements
a digital communication system using 16-QAM and repetition coding. We can simulate the com-
munication channel by adding the AWGN Channel block.

1. When configuring the user functions, please remember that for-

matting is important. We may need to set the expected port sizes
in the model explorer.
2. By default, the output of a user-defined function is sample­
i based. Certain blocks may expect frame based inputs. We can
use the To frame block to rectify this. A double-lined connec-
tion indicates the signal is frame based; a single line connection
indicates the signal is sample based.

5.2  USRP Hardware Implementation

In this section, we will deal with several fundamental but important concepts in
digital communication system designs via SDR using the ethernet-based USRP plat-
forms.2 First is frequency offset compensation. Due to the imperfection of devices,
frequency offset exists in every communication system, which prevents the ideal
signal reception. Therefore, users need to find out the amount of frequency offset
and try to compensate for it. The other important concept is in-phase/quadrature
(I/Q) representation of the signals. Suppose­we have a signal s(t) with the phase f(t),
we can represent this signal using its I and Q component, namely,
sI (t) = s(t)cos(φ(t)) (5.1)
and
sQ (t) = s(t)sin(φ(t)) (5.2)

2. Simulink and MATLAB support most of the Ethernet-based USRP platforms, uncluding USRP2, USRP
N210 rev2 and rev 4.
138 Basic SDR Implementation of a Transmitter and a Receiver

Since the default data type of USRP is I/Q, understanding how I and Q data
flows in USRP will greatly help us with the SDR design and implementation. In
addition, we will be familiarized with some of the useful Simulink tools for digital
communication system.

5.2.1 Frequency Offset Compensation

Most of the signal processing techniques introduced so far happen in baseband,
where the center frequency is around 0 Hz. However, in order to enable radio
transmission or broadcasting, baseband signals need to be modulated to a higher
frequency band, which is usually referred to as a radio frequency (RF) band. The
carrier wave used in this modulation process is provided by a circuit called oscilla-
tor circuit, where oscillator is the main component. Figure 5.7 shows a photograph
of a local oscillator on an SDR PCB. However, oscillator crystals are not perfect,
since they do not oscillate at exactly the specified frequency. To measure this imper-
fection, oscillators and other frequency control devices specify their peak variation
in units of parts per million (ppm). When operating on a certain frequency, it will
lead to a peak frequency variation, which can be calculated using the following
expression:

f × ppm (5.3)
∆f =
106
where ppm is the peak variation (expressed as +/-), f is the center frequency (in Hz),
and Df is the peak frequency variation (in Hz). Due to imperfections and tolerances
in the manufacturing process, oscillator crystals typically have inaccuracies of about
20 or 50 parts per million (ppm), which are guaranteed by the manufacturer.

Figure 5.7  A 10.7 Mhz local oscillator on the SDR PCB (from [5]), which is used to generate the
carrier wave in order to modulate signals from baseband to RF band.
5.2  USRP Hardware Implementation 139

Inaccuracies in the oscillator crystal frequency cause errors in the RF carrier

frequency. Note that even relatively small differences in frequency between the
transmitter and the receiver can prevent the receiver from correctly decoding the
transmission. It is also worth noting that oscillator circuits provide the clock source
for the digital electronics. In particular, the clocks for the analog-to-digital and
digital-to-analog converters affect their sampling rates. Therefore, accurate oscilla-
tors are essential for both the RF and digital systems.
The GNU Radio FAQ [6] states that the “USRP2 reference clock stability” is
“about 20 ppm unless you lock to an external reference.” Note that the USRP2
has an input connector for an external frequency reference, which can provide a
highly precise timing signal for USRP2. Suppose that we select a carrier frequency
of 2.45 GHz for our USRP2. An offset of 20 ppm may sound small, but it is quite
significant at high frequencies. According to (5.3), the peak frequency variation in
this scenario can be:

2.45 ´ 109 ´ 20
Df = = 49, 000 Hz (5.4)
106

In other words, when using a carrier frequency of 2.45 GHz, the worst case fre-
quency offset could be as much as about 50 kHz. Note that if there are two USRP2s
trying to communicate with each other, where one has an offset of -50kHz and the
other has an offset of +50kHz, the overall difference in frequency will be 100kHz.
It is necessary to compensate for this frequency offset in order to achieve success-
ful digital communication between the USRPs. For example, we have measured some
of the USRP2s to have frequency offsets of as much as 45kHz (45kHz/2.45GHz =
18ppm). With frequency offset compensation, the USRP2 works fine, but without
it the digital communication system will not work at all.
In order to devise a software-based approach for frequency offset compensator,
let us perform the following steps to measure and compensate this phenomenon.
First, arbitrarily select a carrier somewhere within either of the supported bands of
the XCVR2450 daughtercard (2.4 to 2.5 GHz and 4.9 to 5.9 GHz). For example,
suppose we choose 2.45 GHz. Note that it is not recommended to choose this ex-
act frequency since other individuals using wireless equipment including software-
defined radio might be transmitting all at the same time and frequency.
Then, let us run siggen.mdl on one USRP and observeFFT.mdl on
­another USRP, as shown in Figure 5.8 and Figure 5.9. By setting the Center fre-
quency para­meter in both SDRu Transmitter and SDRu Receiver blocks, we can
use the carrier frequency selected in the previous step, as shown in Figure 5.10(a)
and Figure 5.10(b). Please note that these two blocks can be obtained by typing
sdrulib in the command window.
As long as these two models are running, we can now use the FFT plot from
spectrum scope to measure the frequency offset between the received signal and
the desired carrier. In order to have a better observation from the scope, we can
right click on the plot from the scope, and choose “Autoscale”. An example of the
output from the FFT scope is shown in Figure 5.11(a). Once the value of the fre-
quency offset is obtained, we can proceed to stop the models, and adjust the carrier
frequency of the transmitter by the amount of this offset. For example, by visually
140 Basic SDR Implementation of a Transmitter and a Receiver

Figure 5.8  The structure of siggen.mdl. In this model, since the SDRU Transmitter block requires
the complex input, the Real-Imag to Complex block converts real and imaginary inputs to a complex-
valued output signal.

inspection, we determine the value at which the peak occurred is Df Hz, as shown
in Figure 5.11(b). Then, we can set the center frequency of the transmitter to be
2.45 GHz+Df Hz. By iteratively adjusting the carrier frequency and observing the
result, we should be able to determine the carrier frequency that we will use on the
transmitter in order to observe the correct carrier frequency on the receiver.
Keep in mind that both the transmitter and the receiver have frequency offsets. For
example, 2.45 GHz on the receiver is somewhere within 2.45 GHz +/- 20ppm and is
not exactly 2.45 GHz. When tuning the transmitter to eliminate the frequency offset
on the receiver, we are essentially compensating for the offset between them. It is the
relative offset between the transmitter and receiver that matters. Note that we can
adjust either the transmitter or the receiver (or both) to compensate for the offset. The
basic idea is to get the two devices synchronized in carrier frequency with each other.
If there are more than two USRPs communicating with each other, it is important to
make sure that each pair of transmitters and receivers are all tuned to each other.

5.2.2 Finding Wireless Signals: Observing IEEE 802.11 WiFi Networks

Let us now use the knowledge that we have so far picked up in this chapter to look
at a commercially available network. Specifically, we can use USRP to observe the
activities of electromagnetic spectrum of commercial wireless local area networks
(WLANs) within the geographical vicinity. Most high population density areas em-

Figure 5.9  The structure of observeFFT.mdl. In this model, FFT Display is an enabled subsystem. We
use the Data Len parameter to qualify the execution of this part. Specifically, when Data Len contains
a zero value, there is no data, so FFT Display cannot be enabled.
5.2  USRP Hardware Implementation 141

(a) (b)
Figure 5.10  Set Center frequency of the SDRu Transmitter block and the SDRu Receiver block. (a)
Set Center frequency of the SDRu Transmitter block to be 2.45 GHz; (b) Set Center frequency of the
SDRu Receiver block to be 2.45 GHz.

ploy numerous wireless communication networks for a variety of applications, such

as the wireless network at school, company, or home. Consequently, we can use the
USRP experimentation platform to plot their magnitude spectrum.
IEEE 802.11 [7] is one type of WLAN standard that possesses a list of car-
rier frequencies for a collection of Wi-Fi channels. For example, The IEEE 802.11
standard defines Channel 1 of the 2.4 GHz band to be centered at 2.412 GHz.
Now, specify the Center frequency parameter of SDRu Receiver with this carrier
frequency, and use the Spectrum Scope to plot their magnitude spectrums. Since the
XCVR2450 daughtercard supports the 5 GHz band as well, use the Spectrum Scope
to plot spectrum in the 5 GHz band.

1. It is recommended to turn on “Autoscale” (right click the scope

display) and adjust the amplitude, since the peaks of these wire-
less signals could be rather low.
2. It might be useful to change the “Decimation” parameter of
i SDRu Receiver to adjust the frequency resolution for a better
graph of the spectrum.
3. In order to capture the full bandwidth of a Wi-Fi signal, we may
need to decrease the downsampling of the received USRP data.

5.2.3  USRP In-phase/Quadrature Representation

Understanding how a signal is represented can greatly enhance our ability to
analyze and design baseband digital communication systems. As mentioned in
142 Basic SDR Implementation of a Transmitter and a Receiver

(a)

(b)
Figure 5.11  Determine the offset Df based on the output from the FFT scope. (a) An example of
the output from the FFT scope; (b) by visual inspection, we can determine the offset Df, the value at
which peak occurred.

Section 4.2.3, a bandpass signal can be represented by the sum of its in-phase (I)
and quadrature (Q) components. Moreover, the data input to the USRP board is
complex, which includes I and Q. Since I and Q play an important role in digital
communications, we are going to observe the I and Q data on transmitter and
receiver sides.
5.2  USRP Hardware Implementation 143

5.2.3.1  Observing In-phase and Quadrature Data

For the transmitter side, let us use siggen.mdl from Section 5.2.1. For the re-
ceiver side, let us use the observeiq.mdl model, where the main subsystem of
this model is shown in Figure 5.12. The model consists of two Vector Scope blocks
that are attached to the real and imaginary output. With these two scopes, we can
observe the real and imaginary parts of a complex number separately.
By changing the two blocks attached to the Real-Image to Complex block, we
can specify the sources of siggen.mdl using six sets of values defined in Table 5.1.
For each source input, let us record the corresponding output in observeiq.mdl
by saving the plots of Vector Scope. For example, Figure 5.13 shows the sample
plots from the two scopes when both I and Q inputs are sine waves.

Q Is the output the same as the input you have specified?

5.2.3.2  Mathematical Derivation

The behavior of the in-phase and quadrature data is closely related to the struc-
ture of USRP board, especially the digital up converter (DUC) and digital down
converter (DDC), which are used by the field-programmable gate array (FPGA) to
convert between baseband and intermediate frequency (IF).
At the transmitter path, a baseband I/Q complex signal is sent to the USRP
board. The digital up converter will interpolate the signal, upconvert it to the IF
band, and finally send it through the DAC. The structure of DUC is shown in
­Figure 5.14, which is named as complex multiplier. In this structure, the numerically
controlled oscillator (NCO) is a digital signal generator which creates a ­synchronous
(i.e., clocked), discrete-time, discrete-valued representation of a waveform, usually
sinusoidal. NCOs are often used in conjunction with a digital-to-analog converter
(DAC) at the output to create a direct digital synthesizer (DDS) [8].

Figure 5.12  The structure of main subsystem of observeiq.mdl. Since the SDRu Receiver block pro-
duces the complex output, the Complex to Real-Image block converts the complex-valued signal to
real and imaginary components.
144 Basic SDR Implementation of a Transmitter and a Receiver

Table 5.1  The input sources of siggen.mdl

Set Index Real Imaginary
1 0 0
2 0 1
3 1 0
4 0 Sinc
5 Sinc 0
6 Sinc Sinc

(a)

(b)
Figure 5.13  Sample plots from the two scopes when both I and Q inputs are sine waves. (a) Output
from Real Scope; (b) output from Imag Scope.
5.3  Open-Ended Design Project: Automatic Frequency Offset Compensator 145

Figure 5.14  The structure of digital up converter on transmitter path, where a baseband I/Q com-
plex signal Iin (t) and Qin (t) is upconverted to the IF band by a carrier frequency of wc.

At the receiver path, the standard FPGA configuration includes digital down
converters (DDC) implemented with a 4-stage cascaded integrator-comb (CIC), as
shown in Figure 5.15. First, the DDC down-converts the received signal from the
IF band to the base band. Next, it decimates the signal so that the data rate can be
adapted by the Gigabit Ethernet and is reasonable for the processing capability of
computers. The complex input signal (IF) is multiplied by the constant frequency
(usually the IF) exponential signal. The resulting signal is also complex, and cen-
tered at 0. Then the signal is decimated with a factor M. The decimator can be
treated as a low pass filter followed by a down sampler. In this structure, the coor-
dinate rotation digital computer (CORDIC) [9], also known as the digit-by-digit
method and Volder’s algorithm, is a simple and efficient algorithm to calculate
hyperbolic and trigonometric functions. It is commonly used when no hardware
multiplier is available as the only operations it requires are addition, subtraction,
bitshift and table lookup.
Given this insight on the structure of DUC and DDC, we can now perform the
following mathematical derivations:

5.3 Open-Ended Design Project: Automatic Frequency

Offset Compensator

5.3.1  Introduction
Starting from this chapter, there will be several open-ended design projects involv-
ing the implementation of an advanced digital transmission/reception system using
SDR. These projects are related to what we have learned and done in each chapter,
but require additional thinking and investigation. For each design project, a poten-
tial solution will be provided. However, there is no single solution for these projects,
so please do not feel constrained to this solution. You are highly encouraged to
propose your own solution and try it out. Each reader is expected to come up with
his/her own innovative design.

The trigonometric identities are needed when expressing the I and

i Q data. Please refer to Appendix E.
146 Basic SDR Implementation of a Transmitter and a Receiver

Figure 5.15  The structure of digital down converter on receiver path, where the input RF signal
I’(t) and Q’(t) is first down-converted to the IF band by a carrier frequency of wc, and then decimated
with a factor M.

1. In Figure 5.14, suppose the input I/Q complex signal is Iin(t)

and Qin(t), the carrier frequency is wc, what is the output of the
DUC, denoted as I¢(t) and Q¢(t)?
2. In Figure 5.15, suppose the input I/Q complex signal is I¢(t) and
Q¢(t), which has been expressed in the previous step. Also assume
that the carrier frequency has a little offset Dw compared to the
DUC, so it is wc + Dw. What is the output after CORDIC and
before CIC? Please express them as I¢¢ (t) and Q¢¢ (t).
Q 3. In Figure 5.15, for simplicity, the rest part after CORDIC, in-
cluding CIC and 1/2 Band, can be viewed as a low pass filter.
I¢¢ (t) and Q¢¢ (t) will go through the low pass filters, where the
double-frequency component like 2wc will disappear. What is
the output of the low pass filters, denoted as Iout(t) and Qout(t)?
4. Given Iout(t) and Qout(t), can you accurately recover Iin(t) and
Qin(t)? If yes, mathematically give the relationship between Iin(t),
Qin(t) and Iout(t), Qout(t). If no, give the reasons.

5.3.2 Objective
Although every communications system will need to compensate for frequency offsets
introduced by nonideal RF crystals, it is difficult and often impractical to compensate
for this manually. Full communications systems often include automatic frequency
offset compensation that can correct this nonideality without end-user intervention.
This feature is essential for devices that must communicate with other devices with
unknown or changing frequency offsets. The objective of this project is to design and
implement a software-defined radio (SDR) communication­ system capable of auto-
matically calculating the frequency offset between two USRP ­platforms.
In Section 5.2.1, we have already found the frequency offset of two USRPs
manually, namely, by comparing the FFT of the transmitted and received signals. In
this project, let us discover this offset in an automatic way. In other words, in the
receiver model, the only thing we need to do is to hit the “start simulation” button,
which will result in this model providing us with the value of the frequency offset.
5.3  Open-Ended Design Project: Automatic Frequency Offset Compensator 147

5.3.3 Theoretical Background
Before applying any method to USRP boards, it is highly recommended to test
the method with Simulink-only model without the USRP transmitter and receiver
blocks and see whether it works. This is used to examine the approach taken under
ideal conditions before actually testing it in the lab setting.

Although we are required to find the frequency offset of two USRP

i boards, we are actually trying to find the frequency offset of a received
signal compared to a transmitted signal.

In hardware implementation, the frequency offset is introduced by the real-world

communication channel. While in the Simulink-only model, we can introduce the
frequency offset by Phase/Frequency Offset block, as shown in Figure 5.16. Let us
specify a frequency offset in this block, and see whether the Simulink model can
give us the number we have specified.
The solution we provide here is based on the theory that raising the received
signal to the power of M removes the modulation from the signal and creates a
tone at M ´ Df, where Df is the frequency offset between the transmitted signal and
the received signal. Specifically, if we take the square of a transmitted signal, the FFT of
the received signal will be shifted double of the frequency offset. Assume the original
transmitted signal is x(t), and the original received signal is y(t), whose FFT is Y(f ).
If the transmitted signal is squared, which is x(t)2, then the FFT of the received
signal becomes Y(f - 2Df ). Before the square operation, the maximum magnitude of
Y(f ) is located at 0 Hz, then after the the square operation, the maximum magnitude
will move to 2Df Hz. Therefore, if we can find the frequency which corresponds to
the maximum magnitude of the new FFT, this frequency, when divided by 2, repre-
sents the frequency location of the offset. This is the estimated offset frequency. In
Simulink, we can easily find the magnitude of FFT using the Magnitude FFT block.
However, unlike Section 5.2.1, we cannot read the frequency value directly
from the FFT plot. Instead, we should employ an automatic way of finding it out.

Figure 5.16  The frequency offset of the Simulink model can be introduced by filling out the Phase
offset and Frequency offset parameters of the Phase/Frequency Offset block.
148 Basic SDR Implementation of a Transmitter and a Receiver

­ herefore, next to Magnitude FFT, we need another important block, namely,

T
Maximum. This block returns the value and index of the maximum elements of
the input signal coming out from Magnitude FFT block. Indices are the locations of
maximums, which should be equal to double of the offset, 2Df, in this system.
Another key to the correct frequency representation is the conversion of the
FFT indices found in the previous step into actual frequency units.

Suppose the FFT index found in the previous step is idx, the frame
size used in this model is Fsize the FFT length of Magnitude FFT
block is N, then the actual frequency offset in Hz can be calculated
by:

idx × Fsize
∆f =
i N×2

More generally, if we take the M-th order of the transmitted signal,

the actual frequency offset in Hz can be calculated by:

idx × Fsize
∆f =
N× M

The techniques discussed above all happen on the receiver side. While on the
transmitter side, we might need the following blocks in the model:

1. Random Integer Generator: Generate random uniformly distributed inte-

gers in the ranges of 0 and 1. These integers are served as the digital input to
the communication system.
2. Baseband Modulator: Modulate the input signal. In order to transmit the
digital input through the channel, modulation is required to convert the
signal from digital domain to analog domain.
3. Raised Cosine Transmit Filter: Upsample and filter the input signal using a
square root raised cosine FIR filter. As mentioned in Section 2.4, the purpose
of using transmit filter is to eliminate the intersymbol interference.

But of course, we are not constrained to these blocks. Instead, we can use all the
blocks available in Simulink. If necessary, we can also use MATLAB to implement
our method.

1. When you are done with your design, compare your results here
from what you have obtained in Section 5.2.1. Are they identical
to each other? If not. which one do you think is more accurate?
Why?
Q 2. Using siggen.mdl, you can generate various frequency offsets
over a wide frequency band. Record these offset values, as well
as the results provided by your design, on which frequency band
do they have a good match of each other? Can you explain the
reason?
5.4  Chapter Summary 149

5.4  CHAPTER SUMMARY

This chapter serves as a starting point for the SDR experimentation. In this chap-
ter, we first implement a bare bones communication system in Simulink, where we
apply two error-correction techniques, namely, repetition coding and interleaving.
Then, several experiments with USRP hardware make us familiar with this plat-
form, which highlights the frequency offset and I/Q data of USRP. In the end, the
open-ended design project helps to find out the frequency offset between two USRP
boards.

5.5 Problems

1. Repetition Coding Suppose that a repetition code with repetition factor R = 5

is used in a binary pulse amplitude modulation (BPAM) transmission system,
where a binary digit “0” is mapped to an amplitude value of “-1” while a binary
digit “1” is mapped to an amplitude value of “+1”. Note that the bit period Tb is
equal to the symbol period TS. The BPAM transmission is sent across an AWGN
channel, where the noise possesses a mean of m = 0 and a standard deviation of
s = 0.1. At the receiver, a limiter is employed to decode the BPAM transmission
intercepted from the channel into the appropriate binary digits, where the de-
coding threshold received amplitude value is equal to 0.

(a) What is the probability that a single received symbol accidentally gets
­decoded incorrectly?
(b) What is the probability that repetition decoder yields an output that is
incorrect?
(c) What would be the impact on the error performance of the repetition decoder
when the noise standard deviation of the AWGN channel is increased to
s = 1?

2. Binary Partial Response Signaling A binary pulse amplitude modulation (BPAM)

partial-response signaling (PRS) system shown in Figure 5.17 generates its
pulse-shaped transmission output by exciting a linear time invariant (LTI) pulse-
­shaping filter g(t) by an impulse train sequence consisting of amplitude values
defined by:

Bn = I n + I n -1 (5.5)

at a rate 1/T symbols per second. The sequence {In} consists of binary digits se-
lected independently from the alphabet {−1, +1} with equal probability. Hence,
the filtered signal has the form:
∞
1
v (t) = ∑ Bn g(t − nT), T=
2W
(5.6)
n =−∞
150 Basic SDR Implementation of a Transmitter and a Receiver

Figure 5.17  Schematic of a Binary PAM partial-response signaling system.

(a) Sketch the signal space diagram for v(t) and determine the probability of
occurrence of each symbol.
(b) Determine the autocorrelation of the sequence {Bn}.
(c) Solve for the power spectral density (PSD) of the sequence {Bn}.
(d) Given that g(t) = u(t) − u(t − T), where u(t) is a unit step function, solve for
the PSD of the output signal v(t).

HINT: The following Fourier Transform Pair may be useful:

æt ö
P ç ÷ « Tsinc (Tf ) (5.7)
èT ø

3. Quadrature Partial Response Signaling A quadrature partial-response signaling

(QPRS) system is shown in Figure 5.18, where the in-phase and quadrature com-
ponents of the transmitted signal each consist of the BPAM PRS implementation
discussed in the previous problem. The transmission output of the QPRS system
is equal to s(t) = Re{v(t)ej2p(fc)t}, where v(t) = vc(t) + jvs(t), Bn = In + In−1, and Cn =
Jn + Jn−1. Note that the sequences {Bn} and {Cn} are independent, and that the
binary outputs of In and Jn are equiprobable. Assuming that g(t) = u(t) − u(t - T),
where u(t) is a unit step function:

(a) Sketch the signal space diagram for the QPRS signal and determine the
probability of occurrence of each symbol.
(b) Determine the autocorrelation of vc(t), vs(t), and v(t).
(c) Solve for the power spectral density (PSD) of vc(t), vs(t), and v(t).

4. Block Interleaving Refer to Section 5.1.2.1 for the definition of block interleav-
ing, and answer the following questions:

(a) Show how a sequence of bits labeled b1 through b20 can be block in-
terleaved when N = 5 and M = 4. What is the resulting interleaved
­sequence?
(b) If the interleaved sequence experiences an error burst lasting five bit
periods, show how the error burst is dispersed in the de-interleaved
­sequence.
(c) What is the amount of end-to-end delay caused by the block interleaving
and de-interleaving processes?
5.4  Chapter Summary 151

Figure 5.18  Schematic of a Quadrature partial-response signaling system.

5. Convolutional Interleaving Refer to Section 5.1.2.2 for the definition and block
diagram of convolutional interleaving, and answer the following ­questions:

(a) Show how a sequence of bits labeled b1 through b30 can be interleaved
when N = 4. What is the resulting interleaved sequence?
(b) If the interleaved sequence experiences an error burst lasting five bit
periods, show how the error burst is dispersed in the de-interleaved
­sequence.
(c) What is the amount of end-to-end delay caused by the interleaving and
de-interleaving processes?

6. Frequency Offset Compensation In Section 5.2.1, you conduct several experi-

ments on two USRPs with frequency offset, and then try to compensate the
offset. As a theoretical background, in this problem, we will study the frequency
offset between transmitter and receiver on frequency domain, and see how this
causes the problem to modulation and demodulation.
The modulation and demodulation process is shown in Figure 5.19.
On transmitter, assume x(t) is the baseband signal we are going to transmit.
The carrier is a cosine wave with frequency fc, so the modulated signal is x(t)

Figure 5.19  The modulation and demodulation process.

152 Basic SDR Implementation of a Transmitter and a Receiver

Figure 5.20  The spectrum of transmitted signal x(t), and two carrier waves.

cos(2pfct). On frequency domain, the spectrum of x(t) is X(f ), with bandwidth

W, as shown in Figure 5.20. The spectrum of the carrier cosine wave is also
shown in Figure 5.20, as two delta function located at fc and −fc.
Assume we have an ideal channel, so there is no distortion to the signal dur-
ing transmission. On receiver, in order to demodulate the signal, we multiply
the received signal by cos(2pfrt), where fr is receiver carrier frequency, which is
slightly larger than fc. The spectrum of this cosine wave is also shown in Figure
5.20, as two delta function located at fr and −fr .
In the end, we use a rectangular shape low pass filter with bandwidth W to
cut high frequencies and reconstruct the original signal.

(a) Sketch the spectrum of signal after modulation.

(b) Sketch the spectrum of signal after demodulation.
(c) Sketch the spectrum of signal after lowpass filter.
(d) Compare the spectrum of signal after lowpass filter with the spectrum
of original signal X(f ). Is this signal successfully recovered or not?
Please explain.

References

[1] SDR Experimentation Using Simulink, http://dot.ece.wpi.edu/?page_id=239.

[2] Jacobsmeyer, J. M., Introduction to Error-Control Coding, www.pericle.com/papers/Error_­
Control_ Tutorial.pdf.
[3] Forney, G. D., “Burst-Correcting Codes for the Classic Bursty Channel,” in the IEEE
Transactions on Communications. COM-19: 772-781, 1971.
[4] Sklar, B., Digital Communications Fundamentals and Applications, Prentice Hall, 1988.
[5] ASSA Radio Astronomy Group. Radio Astronomy Projects. http://www.radio-assa.org.au/
book/export/ html/14.
[6] USRP2 FAQ. http://gnuradio.org/redmine/wiki/gnuradio/USRP2GenFAQ.
[7] IEEE 802.11 Working Group, IEEE 802.11: The Working Group Setting the Standards for
Wireless LANs. http://www. ieee802.org/11/.
[8] Kroupa, V. F., Direct Digital Frequency Synthesizers, Wiley-IEEE Press, 1998.
[9] Vankka, J., Digital Synthesizers and Transmitters for Software Radio, Springer, 2005.
 Chapter 6

Receiver Structure and Waveform

Synthesis of a Transmitter
and a Receiver

In this chapter, we will introduce several techniques for designing and implementing
two different receiver structures commonly used in the creation of a digital com-
munication system, as well as assess their performance observed during over-the-air
transmission. At the same time, we will show how to construct a series of ortho-
normal basis functions that can be combined in order to produce a wide range of
signal waveforms that could be employed by a digital transmitter and receiver. In
the experimental portion of this chapter, we will implement two different modula-
tion structures and then observe their performance. Finally, to wrap up this chapter,
we will develop a frame synchronization design for the open-ended design project.

6.1  SOFTWARE IMPLEMENTATION

Recall from Sections 4.7.1 and 4.7.2 the design of a maximum-likelihood receiver
structure based on a matched filter realization, as well as a correlator-based ap-
proach. Using the theory introduced in Section 4.6, we will now proceed to build
our own implementation for the matched filter and correlator-based receivers. Spe-
cifically, we will implement a two-step approach, where the first step creates the
observation vector from the received signal waveforms, followed by the decision-
making process in the second step. For the purpose of this experiment, we will use
PAM transmission and an AWGN channel to validate the design presented in this
chapter, as shown in Figure 6.1.

6.1.1  Observation Vector Construction

Given a waveform, we would like to use an observation vector for the receiver, so
the first step of the implementation is to convert the received waveform x(t) to a
vector representation X, which is shown in Figure 6.2. Referring to Figure 6.2, this
part of the experiment will construct an observation vector X, which will later be
used as the input to a correlator-based detector. This section employs the ortho-
normal functions {fm(t)} specified in problem 1 of Section 6.5 to generate a bank of
correlators operating on the received signal x(t).

153
154 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

Figure 6.1  The received signal x(t) is generated by a PAM transmission and an AWGN channel.

Given the presence of noise in the transmission channel, it is neces-

sary to collect as much information about the received signal as
i possible in order to form an accurate decision. With more informa-
tion of the received signals, the effect of noise is more likely to be
dismissed.

Let us download and open the correlator.m file. The code has been partially
implemented, and we will finish the rest of the code with the following steps and
plot the results.
For the first part of the code, let us first start off this design by defining and
implementing the four equiprobable transmitted signals specified in problem 1 of
Section 6.5, s1(t), s2(t), s3(t), s4(t), as shown in Figure 6.3. Each of them corresponds
to one of the four transmitted symbols mi, i = 1, 2, 3, 4. Since these four transmitted
symbols are of equal probability, these four transmitted signals are also equiprobable.
In this part, {si(t)} are given in the form of signals in time domain. As specified by
correlator.m, the duration of each signal is 3s, and the number of symbols is 100, de-
fined by the number variable, so we should set the transmission time to 300s in order
to ensure that 100 symbols are transmitted and received in each simulation loop.

Plot a randomly generated stream of these waveforms s(t) in

Q MATLAB. What do you observe? For example, a sample plot of s(t)
generated by 10 transmitted symbols mi is shown in Figure 6.4.

Figure 6.2  A bank of correlators operating on the received signal x(t) to produce the observation
vector X that will be used for a correlator-based detector.
6.1  Software Implementation 155

2 2

s1(t) 1 1

s2(t)
0 0

2
0 1 2 3 0 1 2 3
Time, t /s Time, t /s

2 2

1 1
s3(t)

s4(t)
0 0

1 1

2 2
0 1 2 3 0 1 2 3
Time, t /s Time, t /s
Figure 6.3  The waveform of four equiprobable transmitted signals, s1(t), s2(t), s3(t), and s4(t). Each
of them lasts for 3s, and corresponds to one of the transmitted symbols.

1.5

0.5
s(t)

5
0 5 10 15 20 25 30
Time, t /s
Figure 6.4  A sample plot of s(t) generated by 10 transmitted symbols mi.
156 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

1. Although the design presented here is for continuous time

signals, we need to employ a discrete time approximation
of L = 1000 samples per second in order for this to work in
MATLAB. In other word, the sampling rate used is 1000 sam-
ples per ­second.
i 2. The set of signals {si(t)} can be generated by concatenating scaled
ones and zeros matrices together.
3. For every 3s, we will need to choose one of the four signals
randomly. One approach for doing this task efficiently is to use
switch and case in MATLAB.

Next, let us create an AWGN channel that will be used to add zero-mean white
Gaussian noise of variance to the transmitted signal s(t), which can be expressed
as:

x (t ) = s(t ) + n(t ) (6.1)

where s(t) is the transmitted signal, x(t) is the received signal, and n(t) is the white
Gaussian noise that will be generated in this step in order to model the distortion
introduced by the transmission channel.

Plot the time domain representations of the input signals s(t) and
output signals x(t) for the channel for several different noise vari-
Q ances. Explain how the noise could potentially impair the success-
ful decoding of the intercepted signal at the receiver.

We can use randn in MATLAB to define the Gaussian noise. For instance,
i to create 100 zero–mean Gaussian values with standard devi­ation
2, we can use the following command: 2 . * randn (100,1) ;

Once the received signal waveforms have been implemented in MATLAB, let us
proceed to define the orthonormal functions {fm(t)} derived in problem 1 of Section
6.5. This step is similar to the step that {si(t)} are defined, just make sure that the
vector for {si(t)} and {fm(t)} has the same length.
In the end, let us implement an “integrate-and-dump” block in order to obtain
the observation vector X, defined as:

X = (x1,1, x1,2 , x1,3),(x2,1, x2,2 , x2,3),…,(x100,1, x100,2 , x100,3) (6.2)

where {xn,m} is generated by:
T
xn , m = ∫0 xn (t)fm (t)dt, n = 1, 2, 3,..., 100, m = 1, 2, 3 (6.3)
This block integrates a series of values, uses the result, and restarts integrating, as
shown in Figure 6.5. Basically, the integrate-and-dump block does a resampling
operation on the filtered output.
6.1  Software Implementation 157

Sampling

Dump

Figure 6.5  The operations done by an integrate-and-dump block. Notice how the integrator accu-
mulates the input waveform across the time period T before it resets (i.e., dumps, in order to repeat
the process again).

Since we are employing discrete time approximation for continuous

i time signals integration could be realized by sum in MATLAB.

When you have done all these steps test your implementation with
a transmission of duration 300s, randomly consisting of one of the
four equiprobable signals of duration T = 3s each with zero-mean
Q white Gaussian noise of variance 0.5 added. Assume perfect syn-
chronization between the received signal and the communications
system. Plot each of the elements of the observation vector X using
the stem command in MATLAB.

6.1.2  Maximum-Likelihood Decoder Implementation

As mentioned at the beginning of Section 6.1, we are implementing a two-step ap-
proach, where the first step creates the observation vector from the received signal
and the second step makes the decision. The implementation of the correlator in
Section 6.1.1 has created the observation vector X, which is the transformation of
the received signal into vector space. So now, in this section, we will proceed to the
second step, which is finding the best match by comparing the observation vector to
the signal space representation of the transmitted waveforms. This step is known as
the maximum-likelihood (ML) decoder, as shown in Figure 6.6. This decoder will
operate on the observation vector X from Section 6.1.1 to produce an estimate mn
of the nth transmitted symbol mn in a way that would minimize the average prob-
ability of symbol error.
According to Figure 6.6, the maximum-likelihood algorithm works as follows:
Given the observation vector X at the input to the decoder as well as the vector
representations of the transmitted signals {s}, we first correlate X with si across
all i using a element-by-element product and an accumulator. Next, we normalize
the correlation result by the corresponding signal energy Ei in order to facilitate
a fair comparison. Finally, the resulting decision values for each of the branches
are compared against each other, and the branch with the largest resulting value is
selected.
158 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

Figure 6.6  A maximum-likelihood (ML) decoder system operating on the observation vector X
from Section 6.1.1 to produce an estimate m of the transmitted symbol mi to minimize the average
probability of symbol error. Notice how the element-by-element product of the vectors X and si fol-
lowed by an accumulator is equal to the dot product, XTsi.

In order to get started with this stage of the implementation, let us download
and open the decoder.m file. Although this code is incomplete, it will serve as a
starting point for the rest of this experiment.
For the first part of the code, let us define the four vector representations of the
signals {si(t)}, which are denoted as sv in the code and have been derived in problem
1 of Section 6.5. Each signal vector svi should be a 1 ´ 3 vector.
As long as we have the svi vector, we can calculate the energy of each signal us-
ing the following expression:
3
Ei = || svi ||2 = < svi , svi > = å svij2 , i = 1, 2, 3, 4 (6.4)
j =1
where <svi, svi> is the inner product of svi and svij is the component of each signal
vector.
Given the energy of each signal, let us proceed to implement the accumulator
and subtraction as shown in Figure 6.6. For accumulator, it adds up all the element-
by-element product between X and si.

i Accumulator can be realized by sum in MATLAB.

Plot the output of the accumulator for each of the branches. Pro-
Q vide your observations and explain.

For subtraction, we normalize the accumulated result by the corresponding signal

energy Ei calculated in (6.4). The subtraction is necessary due to the theory of the
maximum-likelihood decoder, presented in Section 4.6.3. The decoder decision rule
minimizes the probability of error, which is expressed in (8.11).
6.1  Software Implementation 159

Let us rewrite the decision rule as follows:

sk = arg min || x j - si ||
si

= arg min(x j - si )(x j - si ) (6.5)

si

= arg min(x j × x j - 2x j × si + si × si )
si

where xj . xj is common to all decision metrics for different si, so we can omit it,
thus yielding:
min(-2x j × si + si × si ) = max(2x j × si - si × si )
si si

= max 2
si
(ò0
T
x j (t )si (t )dt - 0.5ò
T
s 2 (t)dt
0 i
) (6.6)

Therefore, it is necessary to subtract half of the signal energy in each branch. Note
that if all energy values are the same for each possible signal waveform, we can
dispense with the energy normalization process since this will have no impact on
the decision making. In the end, select the largest branch and decode it to produce
the estimate m̂.

i We can use max in MATLAB to select the largest branch.

Remove the energy subtraction from each of the branches and

repeat your experiment.
Q 1. Do you get the same results as before?
5. What would be required for the results with and without
­energy subtraction to be equivalent?

There are several factors that can impact the performance of the entire system, such
as limits of integration, energy subtraction, and SNR of the channel. Energy sub-
traction has been discussed before. As for the other two factors, if we integrate over
a shorter duration than the symbol length, we are correlating the received symbol
with a smaller portion of the transmitted symbol. If we increase the SNR value, it
implies that the transmit power is also increased.

6.1.3  Correlator Realization of a Receiver in Simulink

In Section 6.1.1 and 6.1.2, we have implemented a matched filter-based receiver
structure using MATLAB. In this part of the experiment, we will implement another
type of the receiver structure, namely, the correlator realization, in Simulink. The
receiver structure has been shown in Figure 4.17, which employs the same transmit-
ted signal si(t) as in Section 6.1.1. Consequently, we will be constructing a transmit-
ter that is capable of randomly generating s1(t), s2(t), s3(t), and s4(t), devising an
160 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

additive white Gaussian noise (AWGN) channel, and a four-branch receiver using
correlation in each branch to help determine which symbol was transmitted every
T = 1 second. Let us assume perfect synchronization in this system.

Given these factors, try out the following changes to the system
and compare them with the original results:
1. In Section 6.1.1, do not integrate the entire period T; integrate
until 0.75T. Plot both the estimate m and the actual transmit-
ted symbol mi and compare it with the original plot.
2. In Section 6.1.2, do not subtract the energy of si(t) from each
branch. Plot both the estimate m and the actual transmitted
symbol mi and compare it with the original plot.
Q 5. Combine your implementation from Sections 6.1.1 and 6.1.2.
Change the parameters of AWGN channel in Section 6.1.1
to make the signal-to-noise ratio (SNR) range from 10−5 to
10−1. Compare the m in Section 6.1.2 with the input x in Sec-
tion 6.1.1 to get the bit error rate. Plot a BER curve. NOTE:
In order to get an accurate BER curve, you may need a large
number of transmitted signals, especially when the SNR value
is high.

In order to prepare for our implementation, let us download and open the inte-
grate_and_dump.mdl file, which is shown in Figure 6.7. In this Simulink model,
we can find three important components that will be used in our own implemen-
tation. Note that the Integrate and Dump block accumulates samples according
to the signal possessing the most samples per unit time. In this case, the Gaussian

Figure 6.7  A simple Simulink model of a digital transceiver employing an integrate-and-dump

block at the receiver.
6.1  Software Implementation 161

Noise Generator is producing 100 samples per 1 second duration. We can run the
experiment for different pulse durations T (the default is T = 1 second).

Explain the end-to-end operation of this basic Simulink model,

from the binary PAM transmitter all the way across to the output
Q of the integrate-and-dump block. Use time domain plots to help illus-
trate your explanation.

Now let us create a new Simulink model that will contain our design for the four-
branch correlator-based receiver. First, implement a transmitter that can generate
each of the four waveforms shown in Figure 6.17. The Multiport Switch block can
be used to determine which waveform is transmitted for each symbol period of
time.

Plot the output of this transmitter to show the randomly generated

Q sequence of waveforms s1(t), s2(t), s3(t), and s4(t).

Next, pass this signal stream through an AWGN channel in order to corrupt
the signal with noise. Vary the amount of noise introduced to the signal stream and
observe the shape of the received signal r(t). Note that other channel effects such as
fading can also corrupt the signal. However, we are only focusing on the AWGN
in this exercise.
Given the corrupted intercepted signal r(t), we will conduct the following steps
to create the receiver: First, for each branch, we will repeat si(t) indefinitely such
that it aligns with the corresponding intercepted signal r(t). One approach is to use
the repeating sequence block in the basic Simulink sources blockset. Then, multiply
the repeating si(t) waveforms in each of the branches with the intercepted signal
r(t). Take the output of this multiplication and input it into the integrate-and-dump
block.

How much noise must be added to make the signal stream

Q ­unrecognizable?

Q What did you set as the number of samples to be integrated?

As discussed before, due to the unique symbol energies, for proper operation, the
system must compensate for the differences in each branch. Therefore, take the
output of the integrate-and-dump block and subtract from it the energy of the cor-
responding waveform si(t). Since the receiver knows about all the waveforms ahead
162 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

of time, simply calculate manually the symbol energy and use the constant source
block and the addition block to subtract off the energy from the integrated signal
stream. The relevant signal energy is scaled and subtracted from each branch, thus
producing a measure of the difference between the received signal and each of the
waveforms for each sample instant.

What do you notice about the relative energy levels for each of the
Q branches?

Finally, implement the decision-making block that selects the maximum value at
each time instant T = 1 seconds. The ML decoder just chooses the symbol with
the highest correlation as the estimate of the transmitted symbol, which is then
converted to binary and compared to a binary version of the transmitted symbol.
Note that there is a delay of one sample interval in the decoder output. In order to
calculate the error rate, the binary input also needs to be delayed by one sample
interval to align with the decoded symbol.

Plot the output of the decision-making process, indicating which

Q of the four waveforms were selected. Why is there a delay in the
decoded signal relative to the originally transmitted signal?

As long as we have created the transmitter and the receiver, let us calculate the
bit error rate of this receiver design for a range of noise variance values. Plot
the BER curve of this model down to 10-3. A MATLAB script can be created to
set various values for the SNR, which are used to control the Simulink model
to determine the BER parameters. The BER block can be used to terminate the
simulations for each iteration. The termination condition is set to terminate after
either a certain number of errors are generated or a certain number of symbols
have been transmitted.

Suppose you removed the energy subtraction for all the branches.
Q Does the BER performance of the system change?

6.2  USRP HARDWARE IMPLEMENTATION

In the previous section, we studied how to design actual digital receivers in software
based on the ML concept. However, there is significant value in evaluating these
designs in actual hardware as well in order to provide insights that only real-world
implementations can provide. Therefore, we will start off with the design of two
different modulation schemes employed in a basic transmitter and receiver imple-
mentation. From this experiment, we will get our first experience of how the simu-
lation world differs from actual hardware world.
6.2  USRP Hardware Implementation 163

6.2.1  Differential Binary Phase-Shift Keying

Differential phase shift keying is a noncoherent form of phase shift keying that
avoids the need for a coherent reference signal at the receiver. Noncoherent receiv-
ers are relatively easy and cheap to build and hence are widely used in wireless
communications.

6.2.1.1  Transmitter
The transmitter has a basic structure, which is made up of four blocks, as shown
in Figure 6.8.
The purpose and special parameters of each block is as follows: Signal From
Workspace block is the source of this model. In this block, we use a valid MATLAB
expression ([1 0]’) to specify the Signal parameter, which generates a repeated ‘10.’
We use this signal since it is easy to observe. In the Sample time parameter, there is a
variable named BitRate. This variable is specified by callback functions, which will
be introduced shortly. We set Samples per frame to be 179, because the input frame
size of the SDRu Transmitter block is set as 358, and Raised Cosine Transmit Filter
has an upsampling factor of 2.
In a DBPSK system, the input binary sequence is differentially modulated us-
ing a DBPSK modulator. In Simulink, this operation is conducted by the DBPSK
Modulator Baseband block. The output of this block is a baseband representation
of the modulated signal.
The Raised Cosine Transmit Filter block upsamples and filters the input signal
using a square root raised cosine FIR filter. The icon of the block shows the impulse
response of the filter. Similar to the Signal From Workspace block, the variable
named oversampling is defined in callback functions.
The SDRu Transmitter block is the sink of this model, which can be obtained
by typing sdrulib in the command window. Please note that the input frame size to
this block does not have to be 358. It can be any number you would like.

6.2.1.2  Receiver
The receiver behaves in an inverse manner relative to the transmitter, but there are
more things going on, so it is more complicated than the transmitter. The receiver
model is made up of three parts, as shown in Figure 6.9.

Figure 6.8  The structure of DBPSK transmitter, where a binary source is modulated using DBPSK
and filtered using raised cosine pulse shaping filter.
164 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

Figure 6.9  The structure of a DBPSK receiver, where the transmitted signal is received by the SDRu
Receiver block and fed into the DBPSK Receiver subsystem. Frequency offset between two USRP2s
should be compensated to ensure signal reception.

The first part is the frequency offset compensator. We should use the frequency
offset found in Section 5.2.1 in this part. The second part is the SDRu Receiver
block and the third part is an enabled subsystem, where all the main operations of
this model reside, as shown in Figure 6.10.
The purpose and special parameters of each block inside the enabled subsystem
is as follows: Since the output of the SDRu Receiver block is sample based, we use
the Frame Conversion block to convert it to framebased. This block does not make
any changes to the input signal other than the sampling mode.
In order to form a pulse shaping filter with the Raised Cosine Transmit Filter,
the Raised Cosine Receive Filter block filters the input signal using a square root
raised cosine FIR filter. The icon of the Raised Cosine Receive Filter block shows
the impulse response of the filter.
One of the most significant advantages of a DPSK modulation scheme is that we
do not need to worry about the carrier recovery, which estimates and compensates
for frequency and phase differences between a received signal’s carrier wave and the
receiver’s local oscillator for the purpose of coherent demodulation. However, we

Figure 6.10  The structure of DBPSK Receiver subsystem, where a symbol is filtered using raised
cosine pulse shaping filter and demodulated using DBPSK. The timing recovery is conducted by the
Mueller-Muller Timing Recovery block.
6.2  USRP Hardware Implementation 165

still need to implement some form of timing recovery. The purpose of the timing re-
covery is to obtain symbol synchronization. Two quantities must be determined by
the receiver to achieve symbol synchronization. The first is the sampling frequency,
and the other is sampling phase. The Mueller-Muller Timing Recovery block recov-
ers the symbol timing phase of the input signal using the Mueller-Muller method.
This block implements a decision-directed, data-aided feedback method that re-
quires prior recovery of the carrier phase.
Corresponding to the DBPSK Modulator Baseband on the transmitter side, the
DBPSK Demodulator Baseband block demodulates a signal that was modulated
using the differential binary phase shift keying method. The input is a baseband
representation of the modulated signal.
In the end, To Workspace and Display blocks serve as the sink of this model.
Using these two blocks, we can observe the received data in two ways. We can ei-
ther observe a portion of data directly through Display, or save a larger number of
data to workspace through To Workspace.

6.2.1.3  Using Callback Functions

Callback functions are a powerful way of customizing the Simulink model. A call-
back is a function that executes when you perform various actions on your model,
such as clicking on a block or starting a simulation. We can use callbacks to execute
a MATLAB script or other MATLAB commands. We can use block, port, or model
parameters to specify callback functions.

Figure 6.11  Create model callback functions using the Callbacks pane of the model’s Model Prop-
erties dialog box. In our DBPSK example, several parameters are defined in InitFcn.
166 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

Common tasks you can achieve by using callback functions

­include
• Loading variables into the MATLAB workspace automatically
i when you open your Simulink model;
• Executing a MATLAB script by double-clicking on a block;
• Executing a series of commands before starting a simulation;
• Executing commands when a block diagram is closed.

You can create model callback functions interactively or programmatically.

Use the Callbacks pane of the Model Properties dialog box to create model call-
backs interactively, as shown in Figure 6.11. You can access it by right clicking
on the blank space of your model. In Callbacks pane, the main callbacks you can
define include the following: PreLoadFcn, PostLoadFcn, StartFcn, StopFcn, and
CloseFcn.
To create a callback programmatically, use the set_param command to assign a
MATLAB expression that implements the function of the model parameter corre-
sponding to the callback. Read the online documentation [1] for more information
about creating a callback programmatically.

6.2.2  Differential Quadrature Phase-Shift Keying

With the theory of DBPSK presented in the previous section, let us now construct a
DQPSK system, which employs many of the same design features as DBPSK. First,
let us create a DQPSK transmitter model called DQPSKTx.mdl, which should con-
tain the DQPSK Modulator Baseband block and the Raised Cosine Transmit Filter
block. Just as in the DBPSK implementation, we will need to select several operating
parameters for this model, such as Signal, Sample time, and Samples per frame. Next,
construct a DQPSK receiver named DQPSKRx.mdl. Use the DQPSK Demodulator
Baseband block and the Raised Cosine Receive Filter block. It is vitally important
that we compensate the frequency offset between the transmitter and the receiver.
Then, run these two models and plot the data you have saved to workspace.

Compare the performance between DBPSK and DQPSK. Which

Q one is better? Justify your answer.

6.2.3  Accelerate the Simulink Model that Uses USRP Blocks

In Section 6.2.1, when we plot the received data from DBPSK, we may find some
constant zeros even if we have compensated the frequency offset. At the same time,
by observing Overrun of SDRu Receiver block, we can find that there are packets
being dropped during the USRP2 transmission to the host. This problem is mainly
due to the Simulink model not being fast enough to keep up with the processing
6.3  Open-Ended Design Project: Frame Synchronization 167

speed of the USRP2 board, such that the model cannot be executed in real time.
Here is a collection of performance improvements we can make in our Simulink
models to approach, if not achieve, real time:

1. In the Simulation ® Configuration Parameters ® Data Import/Export dia-

log, turn off all logging.
2. Make sure that the model is single rate. If the model requires resampling,
then choose rational coefficients that will keep the model single rate.
3. Do not add any Buffer blocks to the model.
4. You should try to run with Rapid Accelerator instead of Normal mode. Be
aware that some scopes do not plot data when run in Rapid Accelerator
mode, but scopes inevitably slow down a model in any case. (See Point 6.)
5. Try to avoid feedback loops. Typically, such loops imply scalar processing,
which will slow down the model considerably.
6. Perhaps the most obvious trick of all is not to use scopes unless absolutely neces-
sary. To visualize your data, send it to a workspace variable and postprocess it.
7. If you are using Accelerator or Rapid Accelerator, set the Simulation ®
Configuration Parameters ® Optimization ® Compiler optimization level
to “Optimizations on (faster runs).”
8. If the model has lots of Constant blocks, it could help slightly if Simulation ®
Configuration Parameters ® Optimization ® Inline parameters is checked
on. This will cause the sample time of those Constant blocks to truly become
inf such that Simulink will get the values once and only once during a run.
5. If the model generates code, the Solver setting should be Fixed-step/discrete.
The tasking mode should be Single Tasking.

Applying several or all of these approaches, we should be able to improve the real-
time processing ability of our Simulink models with USRP blocks. Besides, as of
MATLAB 2012b, Simulink will have a so-called Performance Advisor, which can
automatically improve a model’s performance.

6.3  OPEN-ENDED DESIGN PROJECT: FRAME SYNCHRONIZATION

In the previous experiments in this chapter, have sent digital information consisting
of ones and zeros. However, in real-world situations, we usually need to send mes-
sages that are made of packets or frames of data. Therefore, knowing where each
frame starts is a crucial step for performing any format of data reception. In this
section, we will develop frame synchronization mechanism in an open-ended man-
ner using the principle of correlation.

6.3.1  Frame Synchronization

Frame synchronization is the process in the telecommunications transmission system
to align the digital channel (time slot) at the receiving end with the corresponding
time slot at the transmission end as it occurs. For example, you transmit a packet
that contains numerous frames. At the receiver side, if you want to know where a
specific frame actually starts, then you need to implement frame synchronization.
168 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

Frame synchronization in packet-based protocols is often realized using a pre-

amble in the transmitted signal. This preamble is a known signal chosen to be easily
detected and is transmitted immediately before the user’s data. When the preamble
is detected, the receiver now knows where the user data begins and can ensure that
it is handled correctly. Thus, frame synchronization involves the following steps: In
the first step, the transmitter injects a fixed-length symbol pattern, called a marker,
into the beginning of each frame to form a marker and frame pair, which is known
as a packet. Packets are then converted from symbols into a waveform and trans-
mitted through the channel. The receiver detects the arrival of packets by searching
for the marker, removes the markers from the data stream, and recovers the trans-
mitted messages. Marker detection is the most important step for frame synchroni-
zation. In this open-ended design project, a 13-bit Barker code is proposed to form
preambles. This code is chosen to have an extremely peaky result when correlated
with itself, enabling easy detection.
In this project, we will first introduce two Simulink models that implement
frame synchronization. In these models, Barker code is employed as the marker.
Then, we are going to build on the second model and incorporate the USRP2 into
it. In the end, we will transmit “Hello world” from end to end using two USRP2s.

6.3.2  Barker Code

Barker codes are commonly used for frame synchronization in digital communica-
tion systems. A Barker code is a sequence of N values of +1 and –1:

aj for j = 1, 2,..., N (6.7)

such that:
N -v
å a j a j +v £1 (6.8)
j =1
for all 1 £ v < N.
Barker codes have a length of at most 13 bits and possess low correlation side-
lobes. A correlation sidelobe is the correlation of a codeword with a time-shifted
version of itself. An example of autocorrelation function of Barker-7 code is shown
in Figure 6.12. It is obvious from this figure that the sidelobes are low.

6.3.3  Simulink Models

In this section, we will introduce two Simulink models that implement frame syn-
chronization using Barker code as the marker. The first model illustrates the basic
idea, and the second model employs the idea.

6.3.3.1  Basic Model

To get started, let us download and open the mFindFrameSync.mdl file, as shown
in Figure 6.13. There are several key blocks in this model:
6.3  Open-Ended Design Project: Frame Synchronization 169

6
Largest correlation value
5

Low sidelobes
2

0 2 4 6 8 10 12 14
Figure 6.12  Autocorrelation function of Barker-7 code, which has a low sidelobes.

At the beginning, in the Barker Code Generator block, let us specify the length
of the Barker code, and this block will give the corresponding codeword, which is
either 1 or –1. However, the Bernoulli binary is either 1 or 0. In order to attach the
Barker code to the frame, we need to change its format using the Bipolar to Uni-
polar Converter block. Then, the Barker code is attached to the beginning of each
frame using Matrix Concatenate to form a marker and frame pair, which is known
as a packet. In the middle, the Delay block is used to simulate a real channel, which
incurs a delay of 5. In the end, the Error Rate Calculation block uses the delay value
we have found to calculate the computation delay and verifies whether we have
found it correctly. If the value is correct, the error rate should be 0.

Figure 6.13  Simulink model that realizes frame synchronization using Barker code.
170 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

(a)

(b)

Figure 6.14  Two key subsystems in Simulink model mFindFrameSync.mdl. (a) Compute Delay
subsystem, which calculates the delay of the channel by detecting the peak of the correlation of
the received data and the Barker code. (b) Consecutive Delay Comparison subsystem, which
checks whether the calculated delay remains the same for several iterations.

In addition, there are two key subsystems in this model: Compute Delay (Align
Signals ® Enabled Delay Computation ® Compute Delay), as shown in Fig-
ure 6.14(a), calculates the delay of the channel by detecting the peak of the correla-
tion of the received data and the Barker code. Consecutive Delay Comparison (Align
Signals ® Consecutive Delay Comparison), as shown in Figure 6.14(b), checks
whether the calculated delay remains the same for numDelayCalcs ­iterations.
If it is, the Enabled Delay Computation subsystem is disabled, such that we
don’t need to calculate the delay again and again. This subsystem guarantees the
result of this model and also improves the efficiency.

6.3.3.2  Application of the Basic Idea

According to the basic idea, we can construct a more advanced digital communication
system. Let us download and open the mFindFrameStart.mdl file, as shown in
Figure 6.15. Previously, what we have done is transmit binary bits. However, using this
model, we can transmit a sentence, such as “Hello world.” Therefore, in addition to
this Simulink model, we need another m-file to conduct the converter between charac-
ters and binary bits, which has been realized in the charToBitsAndBack.m file.
The variable “Delay” is a most important value in this model, as it provides the
information concerning where the start of the frame is, such that the receiver could
know where should be the starting point of decoding.
6.3  Open-Ended Design Project: Frame Synchronization 171

Figure 6.15  Simulink model that realizes frame transmission using Barker code.

Compared to the basic model, there are three different blocks: First, in the
Signal From Workspace block, let us specify a source from the workspace in or-
der to use it as our test message. For example, suppose we would like to transmit
the test message “Hello world” to the awaiting receiver. Since “Hello world” is
in ASCII character format, in order to be transmitted, we need to convert it to
the format that the computer can process, which is the binary format. We run
charToBitsAndBack.m and get the corresponding bit streams. We save this bit
stream in a variable called “sBit” so that it can be used as the signal source. In this
block, “samples per frame” should be the same as the length of “sBit.” Although
the length of bit stream for “Hello world” is 77, 10 zeros are added to the end of
this bit stream, so the length of “sBit” becomes 87.
Next, in MATLAB Function block, we use a MATLAB Function to pick up the
useful information out of a frame. In this example, delay+13+1 is the first bit in the
stream and delay+13+77 is the last bit. The number 13 corresponds to the length of
the Barker code, and 77 is the length of bit stream for “Hello world.” Therefore, if
we want to transmit some other sentences, or we want to use a Barker code of some
other length, these values need to be modified.
At last, in To Workspace block, we save the useful bits to workspace,
and these bits can be converted to ASCII characters using the second half of
charToBitsAndBack.m. The converted results are illustrated in Figure 6.16,
which shows that we have successfully received the transmitted messages.

Figure 6.16  The converted results at the receiver side.

172 Receiver Structure and Waveform Synthesis of a Transmitter and a Receiver

Figure 6.17  Four-signal waveforms for the Gran-Schmidt Orthogonalization procedure.

6.3.3.3  Incorporating USRPs

When incorporating USRPs into the second model, it is suggested to use DBPSK to
ensure the best performance of bit transmission.

6.3.4  Hints for Implementation

According to the input frame size of the SDRu Transmitter block, we will need to
buffer N frames of data, which makes it less “real time,” but that is OK. Note that
there is always a tradeoff between buffer size and latency. It is always suggested to
try working with the raw data offline in MATLAB before trying out real-time data
retrieval or the Simulink model that do this automatically.

6.3.5  Hints for Debugging

When you have done the complete implementation, you can use the following steps for
debugging. First, collect incoming data from the SDRu Receiver block and save it in an
mat file. Then, correlate the Barker sequence with stored incoming data in MATLAB
and find the mode of the largest correlation value every 179 samples. You should visual-
ize it, not just expect it to be “13.” Next, take the mean of the mode, which can be off
by ±1. In the end, feed incoming data back into your Simulink receiver model without
USRP2 and fix the delay. Once verified, implement the system entirely in Simulink.

6.4  CHAPTER SUMMARY

This chapter covers two basic receiver structures commonly used in the creation of a
digital communication system and assesses their performance observed during over-
the-air transmission. It also shows how to construct a series of orthonormal basis
6.5  Problems 173

s(t)

T/4 3T/4 T t

–1

Figure 6.18  Pulse representing binary "1."

functions that can be combined to produce a wide range of signal waveforms. In the
hardware experimental part of this chapter, two different modulation schemes are
implemented using USRP harware. Finally, a frame synchronization approach is de-
signed for the open-ended design project, which enables the packet transmission.

6.5  PROBLEMS

1. [Gram-Schmit
 Orthogonalization] Carry out the Gram-Schmidt orthogonal-
ization procedure of the signals in Figure 6.16 in the order s3(t), s1(t), s4(t),
s2(t) and thus obtain a set of orthonormal functions {fn(t)}. Then, determine
the vector representation of the signals {sn(t)} by using the orthonormal func-
tions {fm(t)}. Also, determine the signal energies.

2. [Matched
 Filter Realization] Consider the signal: s(t ) = A tcos(ω t ) for 0 £ t £ T
T c

(a) Determine the impulse response of the matched filter for the signal.
(b) Determine the output of the matched filter at t = T.
(c) Suppose the signal s(t) is passed through a correlator that correlates the
input s(t) with s(t). Determine the value of the correlator output at t =
T. Compare your result with that in part (b).
3. [Maximum
 Likelihood Detection] A certain digital baseband modulation
scheme uses the pulse shown in Figure 6.18 to represent binary symbol “1”
and the negative of this pulse to represent binary “0.” Derive the formula

s1(t) s2(t)
1 1

T/2 3T/2 5T/2 3T t T 2T 3T t

–1 –1

Figure 6.19  A pair of orthogonal signals.

 Figure 6.20  A 16-QAM constellation.

for the probability of error incurred by the maximum likelihood detection

procedure applied to this form of signaling over an AWGN channel.

4. [Receiver
 Design] A pair of orthogonal signals s1(t) and s2(t) over the obser­
vation interval 0 < t < 3T are shown in Figure 6.19. The received signal is
defined by
x(t) = sk(t)+ w(t), 0 < t < 3T, k =1, 2, (6.10)
 here w(t) is additive white Gaussian noise of zero mean and power spectral
w
density N0/2.
(a) Design a receiver that decides in favor of signals s1(t) and s2(t), assum-
ing that these two signals are equiprobable.
(b) Calculate the average probability of symbol error incurred by this re-
ceiver for E/N0 = 4, where E is the signal energy.
5. [Correlator
 Realization] A binary digital communication system employs
the signals

s0(t) = 0, 0 £ t £ T
s1(t) = A, 0 £ t £ T (6.1)
f or the transmission of information The demodulator is implemented as a
bank of correlators followed by samplers to sample the output of the cor-
relators at t + T.

ART_Wyglinski_Ch06.indd 174 Manila Typesetting Company 12/22/2012

6.5  PROBLEMS 175

(a) Determine the optimum detector for an AWGN channel and the opti-
mum threshold, assuming that the signals are equiprobable.
(b) Determine the probability of error as a function of the SNR. How does
the above signaling scheme compare with antipodal signaling?
6. [Error Bound] The signal constellation for a communication system with 16
equiprobable symbols is shown in Figure 6.20. The channel is AWGN with
noise power spectral density of N0/2.

(a) Using the union bound, find a bound in terms of A and N0 on the error
probability for this channel.
(b) Determine the average SNR per bit for this channel.
(c) Express the bound found in part (a) in terms of the average SNR per bit.
(d) Compare the power efficiency of this system with a 16-level PAM system.

Reference

[1]  The MathWorks, Using Callback Functions, http://www.mathworks.com/help/toolbox/

simulink/ug/f4-122589.html.
 Chapter 7

Multicarrier Modulation and Duplex

Communications

In Section 4.2, we studied several single-carrier modulation schemes, where input

binary bits are modulated by a carrier signal with the center frequency fc. Instead
of having one center frequency, multi-carrier modulation (MCM) multiplexes serial
input data into several parallel streams and transmit them over independent subcar-
riers. These subcarriers can be individually modulated and manipulated, allowing
for optimization with respect to the channel. This is especially true for a wireline
communications system where the channel is well defined. This chapter will intro-
duce fundamental principles of multicarrier modulation and show why MCM is an
integral part of modern communications systems. Following this, we will look at
the hardware implementation of a duplex communications system, which will pro-
vide us with the fundamentals of the media access control protocols. Note that for
hardware implementation and open-ended design projects, we will focus on single
carrier, due to inherent challenges of implementing multicarrier on USRP. This will
be left as an exercise to the interest reader.

7.1  Theoretical Preparation

Why do digital communication systems use multicarrier modulation? What are the
disadvantages of transmitting at a high data rate using a single carrier transmission?
What happens if part of the channel is severely attenuated? MCM addresses these
issues via a “divide-and-conquer” approach. More specifically, MCM transmits
data in parallel frequency bands simultaneously rather than in a single large chan-
nel. Since dividing the transmission into small bands allows for individual treatment
of these subcarriers, a multicarrier transmission scheme becomes robust to fast-
fading channels and narrowband interference.
Note that there are several important distinctions between multicarrier trans-
mission in a wired communications environment versus a wireless communications
environment. Without loss of generality, we will be focusing in this book chapter
on a wireless communications environment, but is important to note that certain
assumptions made here are not necessarily optimal for a wireline implementation.

7.1.1  Single Carrier Transmission

Most modern communication systems, including all digital subscriber line (xDSL)
modems [1–3], as well as most commercial communication standards, including
WiFi [4, 5], WiMAX [6, 7] and LTE [8, 9], use digital signaling techniques to

177
178 Multicarrier Modulation and Duplex Communications

transmit large quantities of information over a short period of time while being
robust to transmission noise. The fundamental unit of information in these digital
communication systems is the bit, which consists of only two values (“on” and
“off,” “1” and “0,” and so). However, these systems usually map, or modulate,
groups of bits into symbols prior to transmission. At the receiver, each symbol is
compared to a known set of symbols and is demodulated into the group of bits cor-
responding to the closest matching symbol. In the next subsection, we will look at
the modulation scheme that will be employed in this chapter.

7.1.1.1  Modulation and Demodulation of QAM Symbols

Quadrature amplitude modulation (QAM) is the technique of transmitting data
on two quadrature carriers (i.e., they are 90° out of phase with each other, mak-
ing them orthogonal). In what follows, we will consider a digital implementation
of the modulation, so that the carrier signals are cos(wkn) and sin(wkn), with car-
rier frequency wk. The amplitude of each of these two carriers is specified by the
sequence of input bits and is changed every 2N samples, which is called the QAM
symbol period. The amplitudes can take on only a finite number of possible values.
Figure 7.1 represents a discrete-time model of the QAM modulator, where the se-
quence of input bits d[m] is used to determine the amplitudes of the two carriers.
For each symbol period, D bits of input are taken from the input bit stream d[m]
and used to select one of the 2D combinations of amplitudes for the two carriers.
Calling these amplitudes a[ℓ] and b[ℓ], respectively, where ℓ represents the symbol
time index, these amplitudes are kept constant for the duration of a symbol period
by the upsampling and rectangular window filtering shown in Figure 7.1, yielding
the piecewise constant signals a¢ [n] and b¢ [n]. One can then write the expression of
the modulated signal as

s[n] = a′ [n]cos(ω k n) + b′ [n]sin(ωkn) (7.1)

where wk = 2pk/2N is the carrier frequency and 2N is the period of the symbol.
Note that we have to limit the possible carrier frequencies to integer multiples of
2p/2N in order to keep orthogonality between the sinusoidal and cosinusoidal car-
riers in a digital implementation (see the following derivations).
In this chapter, we will be dealing with rectangular QAM signal constellations,
such as those shown in Figures 7.2(a), 7.2(b), and 7.2(c) for 4-QAM, 16-QAM, and
64-QAM, respectively. This basically amounts to saying that a[ℓ], b[ℓ] Î{±(2k – 1)E,

Figure 7.1  Rectangular QAM modulator (multirate model of a digital implementation), where the
carrier signals are cos(wkn) and sin(wkn), and their amplitudes are determined by input bits d[m].
7.1  Theoretical Preparation 179

00 10

Quadrature
1

01 11

1 3
In phase
(a)

1100 0100 0110 1110

3

1000 0000 0010 1010

Quadrature

1001 0001 0011 1011

1101 0101 0111 1111

1 3
In phase
(b)

110000 111000 011000 010000 010010 011010 111010 110010

7
110100 111100 011100 010100 010110 011110 111110 110110
5
100100 101100 001100 000100 000110 001110 101110 100110
3
Quadrature

100000 101000 001000 000000 000010 001010 101010 100010

1
100001 101001 001001 000001 000011 001011 101011 100011

100101 101101 001101 000101 000111 001111 101111 100111

110101 111101 011101 010101 010111 011111 111111 110111

110001 111001 011001 010001 010011 011011 111011 110011

1 3 5 7
In phase
(c)
Figure 7.2  Three types of QAM signal constellations. The in-phase values indicate the amplitude of the
carrier cos(wkn), while the quadrature values indicate the amplitude of the carrier sin(wkn). (a) 4-QAM signal
constellation. (b) Rectangular 16-QAM signal constellation. (c) Rectangular 64-QAM signal constellation.
180 Multicarrier Modulation and Duplex Communications

k = 1,..., 2D/2–1}, where E is some positive constant that scales the energy of the sent
signals and D/2 is the number of bits used to represent the amplitude level of one of
the carriers during a symbol.
One of the advantages of QAM signaling is the fact that demodulation is rela-
tively simple to perform. From Figure 7.3, the received signal, r[n], is split into
two streams and each multiplied by carriers, cos(wkn) and sin(wkn), followed by
a summation block (implemented here through filtering by a rectangular window
followed by downsampling). This process produces estimates of the in-phase and
quadrature amplitudes, â[l] and b̂[l], namely,
2 �N + 2 N − 1
aˆ[� ] = ∑ r[ n]cos(ω k n)
n = 2 �N

2 �N + 2 N − 1 2 � N +2 N −1
= ∑ a′ [n]cos(ω k n)cos(ωkn) + ∑ b′ [n]sin( ω kn)cos(ω kn )
n = 2 �N n = 2� N

2 �N + 2 N − 1 2 � N + 2 N −1
 2π kn   2π kn  2π kn 2π kn
= ∑ a′ [n]cos 
2 N
 cos 
2N
 + ∑ b′[n]sin  2N  cos  2N 
n = 2 �N     n =2 � N    
2 �N + 2 N − 1
a′ [n]
= ∑ 2 (7.2)
n = 2 �N

and

2 �N + 2 N -1
bˆ[� ] = å r[n]sin(ωk n)
n = 2 �N

2 �N + 2 N -1 2 �N + 2 N -1
= å a¢[n]cos(ωk n)sin(ωk n) + å b¢[n]sin(ωk n)sin(ω k n)
n = 2 �N n = 2 �N

2 �N + 2 N -1 2 �N + 2 N -1
æ 2πkn ö æ 2π kn ö æ 2π kn ö æ 2π kn ö
= å a¢[n]cos ç
è 2 N
÷ sin ç
ø è 2 N
÷+ å
ø
b¢[n ]sin ç
è 2 N
÷ sin ç
ø è 2N ø
÷
n = 2 �N n = 2 �N

2 �N + 2 N -1
b¢[n]
= å 2
(7.3)
n = 2 �N

Figure 7.3  Rectangular QAM demodulator, where the received signal, r [n], is split into two streams
and each multiplied by carriers, followed by a summation block.
7.1  Theoretical Preparation 181

where, due to the orthogonality of the two carriers, the cross terms vanish, leaving
the desired amplitude (after some trigonometric manipulation). The bits estimated
from â[ℓ] and b̂[ℓ] bits are then multiplexed together, forming the reconstructed
version of d[m], dˆ[m].
We have so far dealt with modulation and demodulation in an ideal setting.
Thus, one should expect that d[m] = d[m] with probability of 1. However, in the
next subsection we will examine a physical phenomenon that distorts the transmit-
ted signal, resulting in transmission errors.

7.1.2  Multicarrier Transmission

As we have seen thus far, signals can be transmitted on a single carrier frequency.
For instance, (7.1) modulates to the carrier frequency wk. However, the transmitted
signal may only use up a small portion of the total available bandwidth. To increase
bandwidth efficiency and throughput, it is possible to send additional QAM signals
in other portions of the unused bandwidth simultaneously. An example of this is
shown in Figure 7.4, where we have taken the QAM modulator of Figure 7.1 and
put several of them in parallel, each with a different carrier frequency. The data for

a�k−1[n]
ak−1[�]
↑ 2N u[n] − u[n − 2N ]
cos ωk−1 n
dk−1 [m]
+
+
sin ωk−1 n
↑ 2N u[n] − u[n − 2N ]
bk−1 [�]
b�k−1[n]

ak [�] a�k [n]

↑ 2N u[n] − u[n − 2N ]
cos ωk n
dk [m] +
+ + s[n]
+
sin ωk n +
bk [�] b�k [n]
↑ 2N u[n] − u[n − 2N ]

ak+1[�] a�k+1[n]
↑ 2N u[n] − u[n − 2N ]
cos ωk+1n
dk+1[m]
+
+
sin ωk+1n
↑ 2N u[n] − u[n − 2N ]
bk+1[�]
b�k+1[n]

Figure 7.4  Transmitter of an orthogonally multiplexed QAM system, where several QAM modula-
tors are put in parallel, each with a different carrier frequency.
182 Multicarrier Modulation and Duplex Communications

each modulator came from portions of a high-speed bit stream. This is principle
behind multicarrier modulation.
Orthogonal frequency division multiplexing (OFDM) is an efficient type of
multicarrier modulation that employs the discrete Fourier transform (DFT) and in-
verse DFT (IDFT) to modulate and demodulate the data streams. Since the carriers
used in Figure 7.4 are a sinusoidal function of 2pkn/2N, it should come as no sur-
prise that a 2N-point DFT or IDFT can carry out the same modulation, since it also
contains summations of terms of the form e±2pkn/2N. The setup of an OFDM system
is presented in Figure 7.5. A high-speed digital input, d[m], is demultiplexed into
N subcarriers using a commutator. The data on each subcarrier is then modulated
into an M-QAM symbol, which maps a group of log2(M) bits at a time. Unlike the
representation of (7.1), for subcarrier k we will rearrange ak[ℓ] and bk[ℓ] into real
and imaginary components such that the output of the “modulator” block is pk[ℓ] =
ak[ℓ] + jbk[ℓ]. In order for the output of the IDFT block to be real, given N subcarri-
ers we must use a 2N-point IDFT, where terminals k = 0 and k = N are “don’t care”
inputs. For the subcarriers 1 £ k £ N – 1, the inputs are pk[ℓ] = ak[ℓ] + jbk[ℓ], while
for the subcarriers N + 1 £ k £ 2N – 1, the inputs are pk[ℓ] = a2N–k[ℓ] + jb2N–k[ℓ].
The IDFT is then performed, yielding
1 2 N −1
s[2�N + n] = ∑ pk [�]e j(2π nk/ 2N)
2 N k =0
(7.4)

where this time 2N consecutive samples of s[n] constitute an OFDM symbol, which
is a sum of N different QAM symbols.
This results in the data being modulated on several subchannels. This is achieved
by multiplying each data stream by a sin(Nx)/sin(x), several of which are shown in
Figure 7.6.
The subcarriers are then multiplexed together using a commutator, forming the
signal s[n], and transmitted to the receiver. Once at the receiver, the signal is de-
multiplexed into 2N subcarriers of data, ŝ[n], using a commutator and a 2N-point
DFT, defined as

[ ] []
[ ]
[ ]
Mod IDFT

Channel

_
^[ ] ^[] []

^[ ]
Demod Equalizer DFT ^[ ]

Figure 7.5  Overall schematic of an orthogonal frequency division multiplexing system, where the
DFT and IDFT are employed to modulate and demodulate the data streams.
7.1  Theoretical Preparation 183

Frequency Content
6

0
0 2 4 6
Frequency (rad)
Figure 7.6  Characteristics of orthogonal frequency division multiplexing: frequency response of
OFDM subcarriers.

2 N −1
p k[� ] = ∑ sˆ[2�N + n]e − j(2π nk/ 2 N ) (7.5)
n =0

and applied to the inputs, yielding the estimates of pk[ℓ], pk[ℓ]. The output of the
equalizer, pk[ℓ], then passes through a demodulator, and the result multiplexed to-
gether using a commutator, yielding the reconstructed high-speed bit stream, dˆ [m].

7.1.3  Dispersive Channel Environment

Until now, we have considered only an OFDM system operating under ideal con-
ditions (i.e., the transmitter is connected directly to the receiver without any in-
troduced distortion). Now we will cover channel distortion due to dispersive
propagation and how it is compensated for in OFDM system.
From our high school physics courses, we learned about wave propagation and
how they combine constructively and destructively. The exact same principles hold
in high-speed data transmission.
For example, in a wireless communication system, the transmitter emanates
radiation in all directions (unless the antenna is “directional,” in which case, the en-
ergy is focused at a particular azimuth). In an open environment, like a barren farm
field, the energy would continue to propagate until some of it reaches the receiver
antenna. As for the rest of the energy, it continues on until it dissipates.
In an indoor environment, as depicted in Figure 7.7(c), the situation is differ-
ent. The line-of-sight component (if it exists), p1, arrives at the receiver antenna
first, just like in the open field case. However, the rest of the energy does not simply
dissipate. Rather, the energy is reflected by the walls and other objects in the room.
Some of these reflections, such as p2 and p3, will make their way to the receiver
antenna, although not with the same phase or amplitude. All these received compo-
nents are functions of several parameters, including their overall distance between
the transmitter and receiver antennas as well as the number of reflections. At the
receiver, these components are just copies of the same transmitted signal, but with
different arrival times, amplitudes, and phases. Therefore, one can view the channel
as an impulse response that is being convolved with the transmitted signal. In the
184 Multicarrier Modulation and Duplex Communications

1
6
0.8 4

Magnitude (db)
Amplitude

2
0.6
0
0.4

0.2

0
0 1 2 3 4 0 1 2 3
Time Frequency (rad/sec)
(a) (b)

p2
Rx

p1

Tx

p3

(c)
Figure 7.7  Example of a channel response due to dispersive propagation. Notice the three distinctive
propagate paths, p1, p2 and p3, that start at the transmitter Tx and are intercepted at the receiver Rx.
(a) Impulse response. (b) Frequency response. (c) The process by which dispersive propagation
arises.

open field case, the channel impulse response (CIR) would be a delta, since no other
copies would be received by the receiver antenna. On the other hand, an indoor
environment would have several copies intercepted at the receiver antenna, and
thus its CIR would be similar to the example in Figure 7.7(a). The corresponding
frequency response of the example CIR is shown in Figure 7.7(b).
In an xDSL environment, the same principles can be applied to the wireline
environment. The transmitted signal is sent across a network of telephone wires,
with numerous junctions, bridging taps, and connections to other customer appli-
ances (e.g., telephones, xDSL modems). If the impedances are not matched well in
the network, reflections occur and will reach the devices connected to the network,
including the desired receiver.
With the introduction of the CIR, new problems arise in our implementation
that need to be addressed. In Section 7.1.4, we will look at how to undo the smear-
ing effect the CIR has on the transmitted signal. In Section 7.1.5, we will look at
one technique employed extensively in OFDM that inverts the CIR effects in the
frequency domain.
7.1  Theoretical Preparation 185

7.1.4 OFDM with Cyclic Prefix

The CIR can be modeled as a finite impulse response filter that is convolved with a
sampled version of the transmitted signal. As a result, the CIR smears past samples
onto current samples, which are smeared onto future samples. The effect of this
smearing causes distortion of the transmitted signal, increasing the aggregate BER
of the system and resulting in a loss in performance.
Although equalizers can be designed to undo the effects of the channel, there is
a tradeoff between complexity and distortion minimization that is associated with
the choice of an equalizer. In particular, the distortion due to the smearing of a pre-
vious OFDM symbol onto a successive symbol is a difficult problem. One simple
solution is to put a few “dummy” samples between the symbols in order to cap-
ture the intersymbol smearing effect. The most popular choice for these K dummy
samples are the last K samples of the current OFDM symbol. The dummy samples
in this case are known as a cyclic prefix, as shown in Figure 7.8(a).
Therefore, when the OFDM symbols with cyclic prefixes are passed through the
channel, the smearing from the previous symbols are captured by the cyclic prefixes,
as shown in Figure 7.8(b). As a result, the symbols only experience smearing of
samples from within their own symbol.
At the receiver, the cyclic prefix is removed, as shown in Figure 7.8(c), and the
OFDM symbols proceed with demodulation and equalization. }

}
}

CP SymbolM- 1 CP SymbolM CP SymbolM+1

n
(a)

h[n] h[n] h[n]

CP SymbolM- 1 CP SymbolM CP SymbolM+1

(b)

CP CP CP

SymbolM- 1 SymbolM SymbolM+1

(c)
Figure 7.8  The process of adding, smearing capturing, and removal of a cyclic prefix. (a) Add cyclic
prefix to an OFDM symbol. (b) Smear channel h(n) from the previous symbol into the cyclic prefix.
(c) Remove the cyclic prefix.
186 Multicarrier Modulation and Duplex Communications

Despite the usefulness of the cyclic prefix, there are several disadvantages. First,
the length of the cyclic prefix must be sufficient to capture the effects of the CIR. If
not, the cyclic prefix fail to prevent distortion introduced from other symbols. The
second disadvantage is the amount of overhead introduced by the cyclic prefix. By
adding more samples to buffer the symbols, we must send more information across
the channel to the receiver. This means to get the same throughput as a system with-
out the cyclic prefix, we must transmit at a higher data rate.

7.1.5  Frequency Domain Equalization

Once the cyclic prefix has taken care of the intersymbol interference, the received
signal has the DFT applied to it. However, smearing between samples still exists and
must be compensated for. This is achieved by multiplying the subcarriers with the
inverse of the channel frequency response.
The interset of the cyclic prefix resides in the fact that it actually transforms
the linear convolution between the transmitted signal s[n] and the channel impulse
response h[n] into a symbol-by-symbol circular convolution. To clearly see this,
let us take a closer look at a given OFDM symbol with a cyclic prefix, the symbol
starting at time n = 0. Denoting by s[0],..., s[2N – 1] the 2N samples of output of
the transmitter IDFT for the first OFDM symbol, the addition of the cyclic prefix
gives rise to a new signal, namely,
 s[n + 2N − K ] 0 ≤ n ≤ K −1
s [n] = 
 s[n − K ] K ≤ n ≤ 2N − 1
Denoting now by r [ n] the result of the convolution of the signal s [n] with the
length-L channel impulse response h[n], one has that
L− 1
r [ n] = ∑ h[ k] s[ n − k]
k= 0

 n− K L −1
 ∑ h[k ]s[ n − K − k] + ∑ s[n − k + 2 N − K] K ≤ n ≤ K + L −1
 k= 0 k = n −K+1
=
L− 1

 ∑ h[k ]s[ n − K − k] K + L ≤ n ≤ 2N − 1
 k= 0
From this equation, it is rapidly seen that, after removal of the cyclic prefix, the
received sequence r[ n] = r [ n + K] is
2N − 1
r[n] = ∑ h[k]s[((n − k))2 N ] = h[n ] 2N s[n ] (7.6)
k= 0
Thus, the received samples, after removal of the cyclic prefix, are just made up
of the circular convolution of the sent signal (i.e., 2N samples per symbol) with the
channel impulse response h[n]. If one now looks at (7.6) in the frequency domain,
it comes that

R[k] = H[k] ⋅ S[k] ,

7.1  Theoretical Preparation 187

where capital letters represent 2N-point DFTs of the corresponding sequences. Re-
ferring back to Figure 7.5, the 2N-point DFT R[k] of the received samples is already
computed and is denoted by p k[�].
Referring to Figures 7.6 and 7.7(b), if we consider the multiplication of the
corresponding frequency samples together, we notice that each of the subcarriers
experiences a different channel “gain” H[k]. Therefore, what must be done is to
multiply each subcarrier with a gain that is an inverse to the channel frequency
response acting on that subcarrier. This is the principle behind per tone equaliza-
tion. By knowing what the channel frequency gains are at the different subcarriers,
one can use them to reverse the distortion caused by the channel by dividing the
subcarriers with them. For instance, if the system has 64 subcarriers centered at
frequencies wk = 2pk/64, k = 0, . . . , 63, then one would take the CIR h[n] and take
its 64-point FFT, resulting with the frequency response H[k], k = 0, . . . , 63. Then,
to reverse the effect of the channel on each subcarrier, simply take the inverse of the
channel frequency response point corresponding to that subcarrier,

1
W [k] = (7.7)
H[k]

and multiply the subcarrier with it.

7.1.6  Bit and Power Allocation

Resource allocation is an important aspect of OFDM design, since it determines the
bit error rate (BER) performance of the OFDM system. Given a fixed bit rate and
a transmitted power constraint, the BER can be minimized by properly allocating
the bit and power levels over the subcarriers. In this section, we will introduce the
theory and technique concerning the bit and power allocation.

7.1.6.1  Bit Allocation

Most OFDM systems use the same signal constellation across all subcarriers, as
shown in Figure 7.9(a), where the commutator allocates bit groupings of the same
size to each subcarrier. However, their overall error probability is dominated by the
subcarriers with the worst performance. To improve performance, adaptive bit al-
location can be employed, where the signal constellation size distribution across the
subcarriers varies according to the measured signal-to-noise ratio (SNR) values, as
shown in Figure 7.9(b), where the commutator allocates bit groupings of different
sizes. In extreme situations, some subcarriers can be “turned off” or nulled if the
subcarrier SNR values are poor.
Assuming that both the transmitter and receiver possess knowledge of the sub-
carrier SNR levels, one is able to determine the subcarrier BER values. Since the
subcarrier BER values are dependent on the choice of modulation used for a given
SNR value, we can vary the modulation scheme used in each subcarrier in order
to change the subcarrier BER value. Larger signal constellations (e.g., 64-QAM)
require larger SNR values in order to reach the same BER values as smaller constel-
lations (e.g., 4-QAM), which have smaller SNR values.
188 Multicarrier Modulation and Duplex Communications

N 0=4 N 0=1

N 1=4 N 1=3

N 2=4 N 2=6

N 3=4 N 3=4

M- 1 M- 1
N Total = Ni N M-3 =4 N Total = Ni N M-3 =7
i= 0 i= 0
N M-2 =4 N M-2 =2

N M-1 =4 N M-1 =3

(a) (b)
Figure 7.9  Comparison of variable and constant rate commutators with equivalent total rate.
(a) Constant-rate commutator. (b) Variable-rate commutator.

We will simply allocate the number of bits bi on subcarrier i according to

æ γ ö
bi = log 2 ç 1 + i ÷ (7.8)
è G ø
where gi is the SNR of subcarrier i (not in dB). Of course, (7.8) gives rise to nonin-
teger numbers of bits. Round the resulting bis appropriately.

7.1.6.2  Power Distribution of the Transmitted Signal

Channel responses are not always linear. In particular, certain frequencies can be far
more attenuated than others. It can be shown that the optimum power distribution
for the subcarriers should be a constant K. Determining what this level should be
is done through “water pouring,” or choosing the power at a frequency P(f ) such
that all P(f1) = P(fn).
P(f ) is given by

ì N (f )
ïλ - , f eF
P (f ) = í | H (F) |2 (7.9)
ï 0, f eF ¢
î

where N(f ) is the PDF of Gaussian noise, H(f ) is the transfer function representing
a linear channel, and l is the value for which:

òF P(f )df = P (7.10)

or the area under the curve in Figure 7.10. This water filling is done so that the
probability of bit error is the same for all subcarriers.
7.2  Software Implementation 189

Figure 7.10  An illustration of the water filling principle. Notice how power is allocated to each
subcarrier, such that the resulting power would be a constant K.

7.2  Software Implementation

Using the theory that has been introduced in Section 7.1, we will now proceed
to build our own implementation for one type of the MCM transmission system,
namely, an OFDM system. Specifically, we will design and prototype an OFDM
communication system using MATLAB, followed by its implementation in Simu-
link, both of which consist of a transmitter, a channel, and a receiver. With these
implementations, we can observe the BER performance and the transmission spec-
trum of the OFDM system.

7.2.1  MATLAB Design of Multicarrier Transmission

In this section, we will design and simulate an OFDM communication system using
MATLAB. When the system is completed, as shown in Figure 7.5, it should include
the following main components:

• A transmitter, which includes commutator, QAM modulator, IFFT, and cyclic

prefix add modules.
• A channel, which simulates the noise and the channel impulse response of an
actual transmission environment.
• A receiver, which includes cyclic prefix remove, FFT, QAM demodulator, and
commutator modules.

Since the complete OFDM system is a rather complicated structure, let us break
down the whole system into several smaller pieces and start the implementation
from these basic functions. For each function, after its implementation, we will
evaluate it in order to make sure that it is functioning properly. This is a very useful
and widely employed strategy for engineers when developing and testing a large
scale, complicated system.
First of all, let us implement the rectangular QAM modulator as shown in Fig-
ure 7.1, given an arbitrary carrier frequency wk. The M-QAM modulator modu-
lates every b = log2 M binary bits into one of the M complex numbers. We can
get a better idea of this modulation process by looking at the constellation plot,
as shown in Figure 7.2. Corresponding to the QAM modulator, let us implement
the rectangular QAM demodulator as shown in Figure 7.3, given an arbitrary
190 Multicarrier Modulation and Duplex Communications

carrier frequency wk. The M-QAM demodulator converts the complex numbers
to binary bits.

1. Test the cascade of the QAM modulator and demodulator to

see if the input is completely recovered at the output. Do you
see any discrepancies?
2. Plot the frequency magnitude response of each single carrier
Q transmitter after the channel and add them up in one plot.
What does it look like?
3. Implement the constellations of 4-QAM, 16-QAM, and
64-QAM. Can you get the same results as shown in Figure 7.2?

When it comes to the channel, AWGN is generated to simulate the noise in the
real transmission environment. Let us implement an AWGN generator that accepts
as inputs the desired noise variance s 2, the noise mean m, and the number of ran-
dom points R.

1. Verify that AWGN generator works by creating a histogram

of its output and compare against the Gaussian PDF as de-
scribed by (7.11) for m = 3, s 2 = 2.
( x - µn )2
1 -
2σ n2
fX (x) = e
2πσ n2 (7.11)
Q
where mn and σ n2 are the mean and variance.
2. Verify that rectangular QAM modulator/demodulator works
by introducing the appropriate amount of noise and compar-
ing the resulting BER with the BER in Figure 7.11 at 10–3.
What values of SNR did you employ?

Next, let us create several copies of rectangular QAM modulator at transmitter,
as well as several copies of rectangular QAM demodulator at receiver. With the help
of the commutators, we can implement Figure 7.4 and its corresponding receiver.
At the transmitter, according to the OFDM algorithm, the high-speed digital input

1
10
Probability of Bit Error

2
10

3
10

4
10

10
0 5 10 15 20 25
Signal to Noise Ratio (dB)
Figure 7.11  Probability of bit error curves for 4-QAM, rectangular 16-QAM, and 64-QAM.
7.2  Software Implementation 191

x[m] is demultiplexed into N subcarriers dk[m] using a commutator, in order to

adjust the size of the signal to fit in the M-QAM modulator. While at the receiver,
the commutator essentially performs parallel to serial conversion to regain the bit
stream.

Verify that Figure 7.4 and its corresponding receiver work under
Q ideal conditions. What carrier frequencies did you not use in the
implementation.

Then, let us implement Figure 7.5 using IDFT and DFT blocks. In MATLAB,
these two blocks can be realized using ifft and fft functions. Note that the size
of the input/output should be a variable. At the transmitter, in order to get a real-
valued time series s[n], according to the IDFT properties, a 2N-point IDFT should
be implemented. However, the output of the M-QAM modulator are Appoint com-
plex numbers. Therefore, before going through the IDFT block, we need to rear-
range and expand the N-point complex numbers to 2N-point as follows:

X * [2N - k] = X[k], for k = 1, 2,..., N - 1 (7.12)

where X[k] are the 2N-point complex numbers and X*[k] are the 2N-point rear-
ranged complex numbers. Note that X[0] and X[N] values can be assigned with any
real numbers without affecting the IDFT output, which is also 2N-point. Since the
FFT works well with sequences of length 2n due to its structure and implementa-
tion, we zero pad the data in case we do not have this initial data format. At the
receiver, the DFT block reverses the IDFT operation performed at the transmitter.
Thus, the output of the DFT block are 2N-point complex numbers, N of which can
be used to recover the output of M-QAM modulator on transmitting side.
Besides the AWGN, a channel filter should be employed here to simulate the
CIR effect existing in a real-world communication channel, which can be modeled
as a finite impulse response filter.

Implement the channel filter, given as an input the channel impulse re-
sponse, and add the noise generator. Test your implementation with the
following parameters, whose impulse response are shown in Figure 7.12:

h1 = [1 0 0 0 0 0 0 0]
Q h2 = [1 0.1 0.0001 0 0 0]

h3 = [1 0.3 0.4 0.32 0.2 0 .1 0.05 0 .06 0 .02 0 .009]
(7.13)
with m = 0 and s 2 = 0.001.

The last component we need to take into account is the cyclic prefix. Let us
implement the cyclic prefix add and remove blocks of the OFDM system, as shown
in Figure 7.8, with the length of the cyclic prefix an input variable. At the transmit-
ter, cyclic prefix is added to eliminate the intercarrier interference (ICI) caused by
192 Multicarrier Modulation and Duplex Communications

Impulse Response of h
1
1
Amplitude
0.5

0
0 1 2 3 4 5 6 7 8 9
Time
Impulse Response of h2
1
Amplitude

0.5

0
0 1 2 3 4 5 6 7 8 9
Time
Impulse Response of h
3
1
Amplitude

0.5

0
0 1 2 3 4 5 6 7 8 9
Time
Figure 7.12  The impulse responses of three channels, which will result in three different ICI
­effects.

multipath propagation effects and the frequency-selective nature of the communi-

cations channel that is generally attributed to this real-world phenomenon. This
selectivity can be represented by a channel impulse response, which will cause dis-
tortion of the transmitted signal and increase the aggregate BER of the system. For
example, (7.13) represents three different channel impulse responses, which will
result in three different ICI effects. At the receiver, cyclic prefix section is removed
from the received data to recover the valid message. The cyclic prefix is no longer
necessary here since it was added at the transmitter to avoid ICI and has thus served
its purpose.
With the system completely implemented, we can now study how it performs in
a variety of conditions. The performance measure commonly used in digital trans-
mission systems is the BER. Given an SNR value and channel conditions, the BER
is the fraction of bits that are errors to the total number of bits transmitted. The
only question is “how many bits need to be transmitted in order to obtain a reliable
BER?” The rule-of-thumb answer is that one continues transmitting bits until 100
errors have been reached [10].

7.2.2  Simulink Design of OFDM

In the previous section, we implemented an OFDM system in MATLAB and ob-
served the BER performance. Since many of the MATLAB functions have their cor-
responding blocks in Simulink, we can try out the OFDM design in Simulink to see
whether it simplifies our design and implementation. For this purpose, let us down­
7.2  Software Implementation 193

load and open the ofdm.mdl file. The example OFDM transmitter and receiver has
already been implemented and provided, as shown in Figure 7.13. We will finish the
rest of the model with the following steps and experiment with the parameters.

1. With your implementation, plot the BER curves for the three
channel realizations in (7.13) between BER values of 10–2 and
10–4. In particular, simulate your system when it uses N = 16,
32, and 64 subcarriers and employs a cyclic prefix of length
N/4. The high speed input x[m] is a uniformly distributed.
Q Plot the BER results.
2. Employ both constant-rate commutators, which either use
BPSK, 4-QAM, 16-QAM, and 64-QAM, and the variable-rate
commutator. Plot the BER results. Which approach attains the
best performance? Please justify.

In order to finish the model, let us start off this design by getting the AWGN
block from Simulink Library Browser and connecting the transmitter and receiver
via AWGN block. Then, connect Spectrum Scope after the AWGN channel and
connect the transmitter and receiver to the inputs of Error Rate Calculation block.
By doing this, we have completed this model. Now, let us play around this model
by changing some parameters.

1. Change the SNR value of the AWGN channel and observe the
effect in the frequency domain via Spectrum Scope.
Q 2. Open the OFDM Transmitter and OFDM Receiver subsys-
tems, change the number of subcarriers, and observe the effect
in the frequency domain via Spectrum Scope.

Figure 7.13  OFDM transmitter and receiver implementation in Simulink.

194 Multicarrier Modulation and Duplex Communications

7.3  USRP Hardware Implementation

A matched filter is a theoretical framework that should not be mistaken for the
name of a specific type of filter. It is a filtering process designed to take a received
signal and minimize the effect of noise present in it. Hence, it maximizes the SNR
of the filtered signal. It happens that an optimum filter does not exist for each signal
shape transmitted, and it is a function only of the transmitted pulse shape. Due to
its direct relationship to the transmitted pulse shape, it is called a matched filter.
One commonly used matched filter design employed by many communication sys-
tems is the square root raised cosine filter.
In this section, we use two models to illustrates a typical setup in which a
transmitter uses a square root raised cosine filter to perform pulse shaping and the
corresponding receiver uses a square root raised cosine filter as a matched filter to
the transmitted signal. At the receiver, we can plot an eye diagram from the filtered
received signal, from which we can learn the effect of matched filter in a communi-
cation system.

7.3.1 Eye Diagram
In telecommunication, an eye diagram, also known as an eye pattern, is an oscil-
loscope display in which a digital data signal from a receiver is repetitively sampled
and applied to the vertical input, while the data rate is used to trigger the horizontal
sweep. It is so called because, for several types of coding, the pattern looks like a
series of eyes between a pair of rails.
Several system performance measures can be derived by analyzing the display.
If the signals are too long, too short, poorly synchronized with the system clock,
too high, too low, too noisy, too slow to change, or have too much undershoot or
overshoot, this can be observed from the eye diagram. For example, Figure 7.14
shows a typical eye pattern, where the x-axis represents timing and the y-axis repre-
sents amplitude. We can also obtain the information concerning time variation and
amount of distortion from the eye diagram. In general, the most frequent usage of
the eye pattern is for qualitatively assessing the extent of the ISI. As the eye closes,
the ISI increases; as the eye opens, the ISI decreases.
Best sampling instant

Distortion Time
amount variation

“The eye”
Figure 7.14  A typical eye pattern, where the x-axis represents timing and the y-axis represents
amplitude.
7.3  USRP Hardware Implementation 195

Figure 7.15  A typical eye pattern for the QPSK signal. The width of the opening indicates the time
over which sampling for detection might be performed. The optimum sampling time corresponds to
the maximum eye opening, yielding the greatest protection against noise. If there were no filtering
in the system, then the system would look like a box rather than an eye.

7.3.1.1  Discrete-Time Eye Diagram Scope

The Discrete-Time Eye Diagram Scope block displays multiple traces of a modu-
lated signal to produce an eye diagram. We can use this block to reveal the modula-
tion characteristics of the signal, such as pulse shaping or channel distortions. This
is a most important block that will be used in this section. For example, Figure
7.15 shows a typical eye pattern for the noisy QPSK signal. Generally speaking, an
open eye pattern corresponds to minimal signal distortion. Distortion of the signal
waveform due to intersymbol interference and noise appears as a closure of the eye
pattern, which results in a significant probability that the received data might be
incorrectly decoded.

7.3.2  Matched Filter Observation

In this section, we will use the DBPSK transmitter model on the transmitter side. On the
receiver side, let us download and open the rcfilter_receiver.mdl file, as shown
in Figure 7.16. The top level of this model looks similar to the DBPSK receiver model.
However, its enabled subsystem contains the information that we are interested in.
For the first enabled subsystem, let us go into the matched filter subsystem, as
shown in Figure 7.17, where there is a Raised Cosine Receive Filter and a Discrete-Time
Eye Diagram Scope. The Raised Cosine Receive Filter is employed to be the matched
196 Multicarrier Modulation and Duplex Communications

Figure 7.16  A model that illustrates a typical receiver setup, in which the receiver uses a square
root raised cosine filter as a matched filter corresponding to the square root raised cosine filter on
the transmitter.

filter for the Raised Cosine Transmit Filter on the transmitter. By double clicking the
Discrete-Time Eye Diagram Scope, we can observe the eye diagram of our system.
Perform the following tasks and plot your observations.

1. Compare Raised Cosine Receive Filter and Raised Cosine

Transmit Filter on the transmitter side, and see whether they
match each other.
2. Change the parameters (group delay, rolloff factor, and so
on) of the Raised Cosine Receive Filter to make it match the
Raised Cosine Transmit Filter. How does the eye diagram
appear? a
Q 3. Change the parameters of the Raised Cosine Receive Filter
to make it mismatch the Raised Cosine Transmit Filter. How
does the eye diagram appear?
4. Delete the Raised Cosine Receive Filter block, as shown in
Figure 7.18. How does the eye diagram appear?
a To get better observation, you need to use “Autoscale” of eye
diagram.

Figure 7.17  The subsystem of the receiver model, in which the receiver plots an eye diagram from
the filtered received signal. Based on this diagram, we can learn the effect of matched filter in a com-
munication system.
7.4  Open-Ended Design Project: Duplex Communication 197

Figure 7.18  The subsystem of the receiver model without a Raised Cosine Receive Filter.

Please note that when the eye diagram has two widely opened “eyes,” it in-
dicates the appropriate instants at which to sample the filtered signal before de-
modulating. It also indicates the absence of intersymbol interference at the sampling
instants of the received waveform. A large SNR in the channel will produce a low-
noise eye diagram. We can also construct a Simulink-only model using AWGN
Channel block and then change the SNR parameter in the AWGN Channel block
to see how the eyes in the diagram change.

7.4 Open-Ended Design Project: Duplex Communication

In this section, we will design a duplex single carrier communication system based
on what we have accomplished thus far in this book. A duplex communication
system is a system composed of two connected parties or devices that can commu-
nicate with one another in both directions. Duplex systems are often employed in
many communications networks, either to allow for a communication “two-way
street” between two connected parties or to provide a “reverse path” for the moni-
toring and remote adjustment of equipment in the field.
Systems that do not need the duplex capability use instead simplex communica-
tion. These include broadcast systems, where one station transmits and the others
just “listen.” Several examples of communication systems employing simplex com-
munications include television broadcasting and FM radio transmissions.

7.4.1  Duplex Communication

The models that we have been used so far in this book are operated in the simplex
communication manner, such as the DBPSK transmitter and receiver models studied
in the previous experiments. Since we are using the XCVR2450 transceiver daugh-
terboard in our experiments described in this book, our USRP2 can only be either a
transmitter or a receiver. However, we cannot employ both transmitter and receiver
at the same time due to the design of this specific daughterboard. As described in
[11], the XCVR2450 uses a common local oscillator (LO) for both receive and
transmit, as shown in Figure 7.19. A change of one LO setting will be reflected
in the other LO setting. Therefore, the XCVR2450 does not support full-duplex
mode; attempting to operate in full-duplex will result in transmit-only operation.
198 Multicarrier Modulation and Duplex Communications

Figure 7.19  A photograph of XCVR2450 RF transceiver daughterboard (from [12]), where a com-
mon local oscillator is used for both receive and transmit.

Consequently, in this open-ended design project, we will implement a system that is

capable of doing half-duplex communication.

7.4.2 Half-Duplex
A half-duplex (HDX) system provides communication in both directions, but only
one direction at a time (not simultaneously). Typically, once a party begins receiv-
ing a signal, it must wait for the transmitter to stop transmitting before replying.
An example of a half-duplex system is a two-party system such as a “walkie-talkie”
style two-way radio, wherein one must use “over” or another previously designated
command to indicate the end of transmission and ensure that only one party trans-
mits at a time, because both parties transmit and receive on the same frequency. A
good analogy for a half-duplex system would be a one-lane road with traffic con-
trollers at each end. Traffic can flow in both directions, but only one direction at a
time, regulated by the traffic controllers. There are several different ways to control
the traffic in a half-duplex communication system, and we suggest you to use time-
division duplexing.

7.4.3  Time-Division Duplexing

Time-division duplexing (TDD) refers to a transmission scheme that allows an
asymmetric flow for uplink and downlink transmission that is more suited to data
transmission. In a time-division duplex system, a common carrier is shared be-
tween the uplink and downlink, the resource being switched in time. Users are al-
located one or more timeslots for uplink and downlink transmission, as shown in
Figure 7.20.
For radio systems that aren’t moving quickly, an advantage of TDD is that
the uplink and downlink radio paths are likely to be very similar. This means that
techniques such as beamforming work well with TDD systems. Examples of time-
division duplexing systems are as follows:
7.4  Open-Ended Design Project: Duplex Communication 199

Mes
sage

ge
ssa
Me

Mes
sage

ge
ssa
Me

Time Time
Figure 7.20  A half-duplex communication system using time-division duplexing. A common carrier
is shared between station A and station B, the resource being switched in time.

• UMTS 3G supplementary air interfaces TD-CDMA for indoor mobile tele-

communications [13];
• DECT wireless telephony [14];
• Half-duplex packet mode networks based on carrier sense multiple access
[15]; for example, two-wire or hubbed Ethernet, wireless local area net-
works, and Bluetooth can be considered TDD systems;
• IEEE 802.16 WiMAX [16].

7.4.4  Useful Suggestions

Suppose we have a radio station A and a radio station B; let us perform the fol-
lowing tasks in time sequence: station A transmits “Hello world” to station B, sta-
tion B transmits “Hello world” to station A, station A transmits “Hello world” to
station B, station B transmits “Hello world” to station A, and so on, as shown in
Figure 7.20.
At the beginning, let us try something simple to see whether it works. For ex-
ample, we can transmit “1010”… instead of “Hello world,” such that frame syn-
chronization is not necessary. We can test one cycle only, (i.e., station A transmits
“Hello world” to station B and station B transmits “Hello world” to station A).
Then, in order to make it an automatic communication system, we need to
write a MATLAB code for station A and another MATLAB code for station B, such
that all we need to do is to click the “Run” button on two stations at the beginning
of simulation. There are two useful MATLAB functions in this step. A MATLAB
function called sim allows us to run our Simulink model from MATLAB code.
Another MATLAB function called pause allows us to specify the interval between
two actions in our MATLAB code. This function is necessary due to the different
computation ability of PCs. We want both of them get ready before continuing to
the next step. For example, a very simple piece of code is as follows:
200 Multicarrier Modulation and Duplex Communications

sim(‘tx’); pause(10); sim(‘rx’);

According to this code, the radio station first runs a Simulink model called “tx,”
then it pauses for 10-s and in the end it runs a Simulink model called “rx.”
At last, in order to specify the simulation time of our model, we can go to
­Simulation ® Configuration Parameters ® Solver tab, and change Stop time.

7.4.5 Evaluation and Expected Outcomes

With individual models and MATLAB scripts for both stations completed, we can
proceed to the testing process. Let us first start the two scripts at the same time,
which makes station A initiate transmission and station B begin listening. After a
certain period of time, station B will evaluate the received message and send a mes-
sage back if it has received the correct message. In this way, the two stations will
relay the message back and forth to each other, demonstrating a TDD half-duplex
communication system. The most important task in this process it to find out the
pause time for each station between the transmitter mode and the receiver mode.
An example of the status output at station B is displayed in Figure 7.21. According
to the figure, we can see that station B received the correct message for four times
before it received a wrong message. When the wrong message is received, the script
stops executing, and the communication process ends.

Figure 7.21  An example of the status output at station B, where it shows the switch between the
transmitter mode and the receiver mode as well as the received messages.
7.5  Chapter Summary 201

7.5 CHAPTER SUMMARY

This chapter focus on the principle and implementation of multicarrier modulation.

In this chapter, the theory of multicarrier modulation is first introduced. Then, we
implement the multicarrier transmission in both MATLAB and Simulink. In the
end, the open-ended design project realizes a duplex communication system.

7.6  PROBLEMS

1. [Orthogonal Carriers] In Section 7.1.1.1, prove that the carriers cos(wkn)

and sin(wkn) are orthogonal over the symbol period.1 Are there any values
of k that are exceptions? Why?
2. [Carrier Frequency] Prove that (7.2) and (7.3) are true. What are the excep-
tions for the carrier frequencies for perfect recovery under ideal conditions?
Why is k constrained to be below N?
3. [QAM Demodulator] Draw a schematic for an orthogonally multiplexed
QAM demodulator as in Figure 7.4. Show that the subcarriers of the or-
thogonally multiplexed QAM system are orthogonal.
4. [2N-Point IDFT] In Section 7.1.2, why are k = 0 and k = N “don’t care”
inputs?
5. [OFDM Configuration] Show that the OFDM configuration and the orthog-
onally multiplexed QAM system are identical. Which of the two implemen-
tations is faster if the OFDM system employs FFTs? Prove it by comparing
the number of operations.
6. [Cyclic Prefix Remove] Explain (7.6) in detail, and prove that r[n] = h[n]
2N s[n].
7. [Bit Allocation] Based on Figure 7.7(b), explain clearly how different subcar-
riers experience different SNR values and thus different BERs. Explain how
bit allocation can be used advantageously.
8. [Subcarrier SNR] Find the expression for the subcarrier SNR, given the chan-
nel impulse ­response, the subcarrier transmit power levels, and the noise
variance s 2.
9. [Multicarrier Transceiver] Suppose we are given a generic multicarrier
transceiver as shown in Figure 7.22, where we have selected N = 4 for this
problem. The frequency (magnitude) response of the synthesis filter g(0)(n) is
shown in Figure 7.23(a), and the frequency responses of the synthesis filters
g(1)(n), g(2)(n) and g(3)(n) are simply frequency shifted versions of g(0)(n) cen-
tered at p/2, p, and 3p/2. Furthermore, the corresponding analysis filters pos-
sess the same frequency (magnitude) responses. Finally, we have arbitrarily
defined the signal spectra for d(0)(n), d(1)(n), d(2)(n), and d(3)(n) in terms of the
spectral shapes shown in Figures 7.23(b)–7.23(e).

(a) Initially assuming an ideal transmission channel between the transmit-

ter and the receiver, sketch across the frequency range [0, 2p) the signal
P–1
1. Two real sequences h[n] and g[n] are said to be orthogonal over a period 0 to P if ∑ n=0 h[n]g[n] = 0.
202 Multicarrier Modulation and Duplex Communications

(a)

(b)
Figure 7.22  Schematics of generic multicarrier modulation employing synthesis and analysis filter-
banks. (a) Generic multicarrier transmitter. (b) Generic multicarrier receiver.

spectra for the following: (a) upsampled signal y(0)(n), (b) output of the
synthesis filter g(0)(n), (c) composite signal s(n), (d) output of the analy-
sis filter f (1)(n), (e) downsampled signal dˆ(1)(n).
NOTE: Please make sure to properly indicate values of the x-axis.
(b) Suppose now that a frequency selective channel exists between the
transmitter and receiver, as shown in Figure 7.23(f). Sketch the result-
ing signal spectra at the input to the multicarrier receiver, namely, r(n).
If each of the subcarrier equalizers w(0)(n), w(1)(n), w(2)(n), and w(3)(n)
consist of only one coefficient, what should be its values and why?
HINT: Refer to Figure 7.23(f) and the frequency response values at the
center of each subcarrier.
(c) Suppose that the average frequency attenuation per subcarrier is equal to:

C = {C i }8i =1 = {0.5, 0.25, 0.05, 0.75}

the noise spectral density is equal to N0 = 10–6, the transmit power level
for each subcarrier is Pi = 1 × 10–3, and the SNR gap is equal to G = 20.
We understand that the number of bits that can be allocated to subcar-
rier i is equal to:
æ γ ö
bi = log 2 ç 1 + i ÷, i = 1, …, N (7.14)
è G ø
where the subcarrier signal-to-noise ratio (SNR) is equal to:
Pi | Ci |2
γi = (7.15)
N0
7.5  Chapter Summary 203

(a)

(b)

(c)

(d)

(e)

(f)

Figure 7.23  Sketches of the synthesis filter g(0)(n), all four signal spectra, and the channel response.
(a) Frequency response for synthesis filter g(0)(n). (b) Signal spectrum for d (0)(n). (c) Signal spectrum
for d (2)(n). (d) Signal spectrum for d (2)(n). (e) Signal spectrum for d (3)(n). (f) Frequency response of
the wireless channel.

What is the bit allocation across all these subcarriers rounded to

the nearest integer? What is the average quantization error between the
optimal (real) bi values and the rounded (integer) values across all the
subcarriers?
(d) Finally, suppose now that one of two possible modulation schemes can
be employed per subcarrier, where modulation scheme 1 (i.e., MS1)
204 Multicarrier Modulation and Duplex Communications

transmits two bits per symbol epoch while modulation scheme 2 (i.e.,
MS2) transmits four bits per symbol epoch. Using the information pro-
vided in part (c), the probability of bit error for each subcarrier when
employing MSI is PeMS1 = {7 × 10–7, 2 × 10–6, 1 × 10–5, 5 × 10–8} while for
MS2 it is Pe,MS2 = {9 × 10–6, 2 × 10–5, 1 × 10–3, 1 × 10–6}. In order to maxi-
mize the throughput given an average probability of bit error constraint
of 1 × 10–5, what modulation scheme should be used for each subcarrier?
What is the maximum throughput? Please justify your answer.
HINT: When computing the average probability of bit error, remember
to weigh each subcarrier quality by the number of bits being supported
and divide by the total bits to be transmitted.
10. [Bit and Power Allocation]: Suppose we have an orthogonal frequency-­
division multiplexing (OFDM) communication system consisting of N =
8 subcarriers that is transmitting over a frequency selective fading chan-
nel possessing an average frequency attenuation per subcarrier equal
to:

C = {C i }8i = 1 = {0.5, 1.0, 0.1, 0.25, 0.05, 0.75, 0.6, 0.2}

and a noise spectral density of N0 = 10−8. The number of bits that can be
allocated to subcarrier i is equal to:
æ γ ö
bi = log 2 ç 1 + i ÷ , i = 1, …, N (7.16)
è Gø
where the subcarrier signal-to-noise ratio (SNR) is equal to:
Pi | Ci |2
γi = (7.17)
N0
and Pi is the transmit power for subcarrier i.

(a) Suppose that Pi = 1 × 10−5 for all subcarriers, and the SNR gap is equal
to G = 20. What is the bit allocation across all these subcarriers rounded
to the nearest integer? What is the average quantization error between
the optimal (real) bi values and the rounded (integer) values across all
the subcarriers?
(b) Suppose bi = 3 bits across all subcarriers given an SNR gap equal to G = 20.
What should be the approximate subcarrier power values Pi?

References

[1] Cioffi, J. M., “Asymmetric Digital Subscriber Lines,” C34 in Communications Handbook,

CRC Press in Cooperation with IEEE Press, 1997.
[2] Golden, P., H. Dedieu, and K. Jacobsen, Fundamentals Of DSL Technology, Auerbach
Publications, 2004.
[3] Golden, P., H. Dedieu, and K. Jacobsen, Implementation And Applications of DSL Tech­
nology, Auerbach Publications, 2007.
[4] Hanzo, L. L., Y. Akhtman, L. Wang, and M. Jiang, MIMO-OFDM for LTE, WiFi and
7.5  Chapter Summary 205

WiMAX: Coherent versus Noncoherent and Cooperative Turbo Transceivers, Wiley-IEEE

Press, 2010.
[5] Lee, B. G., and S. Choi. Broadband Wireless Access and Local Networks: Mobile WiMax
and WiFi, Artech House, 2008.
[6] Pareek, D., WiMAX: Taking Wireless to the Max, Auerbach Publications, 2006.
[7] Nuaymi, L., WiMAX: Technology for Broadband Wireless Access, Wiley, 2007.
[8] Sesia, S., I. Toufik, and M. Baker, Lte - The Umts Long Term Evolution: From Theory to
Practice, 2nd edition, Wiley, 2011.
[9] Cox, C., An Introduction to LTE: LTE, LTE-Advanced, SAE and 4G Mobile Communica-
tions, 2nd edition, Wiley, 2012.
[10] Jeruchim, M. C., P. Balaban, and K. S. Shanmugan. Simulation of Communication Sys-
tems: Modeling, Methodology and Techniques, 2nd edition, Springer, 2000.
[11] http://files.ettus.com/uhd_docs/manual/html/dboards.html#xcvr-2450.
[12] http://www.olifantasia.com/drupal2/en/node/12.
[13] Muratore, F., UMTS Mobile Communications for the Future, Wiley, 2001.
[14] Phillips, J. A., and G. M. Namee. Personal Wireless Communication with DECT and
PWT, Artech House, 1998.
[15] Spurgeon, C., Ethernet: The Definitive Guide, O’Reilly Media, 2000.
[16] Radha G. S. V., K. Rao, and G. Radhamani, WiMAX: A Wireless Technology Revolution,
Auerbach Publications, 2007.
 Chapter 8

Spectrum Sensing Techniques

With advanced digital communication systems such as cognitive radio [1, 2] being
used in a growing number of wireless applications, the topic of spectrum sensing has
become increasingly important. This chapter will introduce the concept, the funda-
mental principles, and the practical applications of spectrum sensing. In the experi-
mental part of this chapter, two popular spectral detectors will be implemented and
their performance will be observed by applying them to several Simulink generated
signals. This is followed by a USRP hardware implementation of a wideband spectrum
sensing technique. Finally, the open-ended design project will focus on a commonly
used medium access control (MAC) scheme referred to as carrier sense multiple access
with collision avoidance (CSMA/CA), which heavily relies on spectrum sensing.

8.1  Theoretical Preparation

In recent years, a large portion of the assigned spectrum has been observed to be sparsely
and sporadically utilized during several spectrum measurement campaigns [3, 4], as il-
lustrated in Figure 8.1. In particular, spectrum occupancy by licensed transmissions are
often concentrated across specific frequency ranges while a significant amount of the
spectrum remains either underutilized or completely unoccupied. Therefore, dynamic
spectrum access [6, 7] has been proposed for increasing spectrum efficiency via the
real-time adjustment of radio resources using a combination of local spectrum sensing,
probing, and autonomous establishment of local wireless connectivity among cognitive
radio (CR) nodes and networks. During this process, spectrum sensing is employed for
the purpose of identifying unoccupied licensed spectrum (i.e., spectral “white spaces”).
Once these white spaces have been identified, secondary users (SU) opportunistically
utilize these spectral white spaces by wirelessly operating across them while simultane-
ously not causing harmful interference to the primary users (PU).1 Currently, there ex-
ists several techniques for spectrum sensing. This chapter will emphasize two of them,
namely, the energy detection and cyclostationary feature detection.

8.1.1  Power Spectral Density

Although the concept of power spectral density has been introduced in Section 3.3.4,
we will briefly revisit it in this section since it is a crucial concept in spectrum sensing.
To analyze a signal in the frequency domain, the power spectral density (PSD), SXX ( f ),

1. Primary users are licensed users who are assigned certain channels, and secondary users are unlicensed users who
are allowed to use the channels assigned to a primary user only when they do not cause any harmful interference
to the primary user [8].

207
 208

SPECTRAL WHITESPACES

Power (in dBm)

0.088 0.35 0.65 0.95 1.25 1.55 1.85 2.15 2.45 2.686
Frequency (in GHz) 9
x 10
Figure 8.1  A power spectral density snapshot of wireless spectrum ranging from 88 MHz to 2686 MHz measured on July 11, 2008, in Worcester, MA, at coordi-
nates 42 °16′8′′N, 71 °48′14′′W (from [5]).
Spectrum Sensing Techniques
8.1  Theoretical Preparation 209

is often used to characterize the signal, which is obtained by taking the Fourier
transform of the autocorrelation RXX (t) of the WSS random process X(t).
The PSD and the autocorrelation of a function are mathematically related by the
­Einstein-Wiener-Khinchin (EWK) relations [9], namely:
∞ − j 2π f τ
SXX (f ) = ∫ −∞ RXX (τ )e dτ (8.1)

¥
RXX (f ) = ò SXX (τ )e+ j 2π fτ df (8.2)
-¥

A very powerful consequence of the EWK relations is its usefulness when at-
tempting to determine the autocorrelation function or PSD of a WSS random pro-
cess that is the output of a linear time-invariant (LTI) system whose input is also
a WSS random process. Specifically, suppose we denote H(f ) as the frequency re-
sponse of an LTI system h(t). We can then relate the power spectral density of input
and output random processes with the following equation:

SYY (f ) = |H(f )|2SXX(f ) (8.3)

where SXX (f ) is the PSD of input random process and SYY (f ) is the PSD of output
random process. As we will see later, the experiments in Section 8.2 of this chapter
are all based on this very useful relationship, as illustrated in Figure 8.2.

8.1.2  Practical Issues of Collecting Spectral Data

Although spectrum sensing is a rather intuitive process, its implementation pos-
sesses several engineering tradeoffs that can potentially impact the performance
of the overall operation the results obtained. One of the key considerations when
designing a spectrum sensing system is how to collect and store spectrum measure-
ment data via a spectrum sweep. A spectrum analyzer can be employed to provide
a nearly instantaneous snapshot of radio frequency bandwidth via the sampling of
intercepted signals at a specified rate. For example, Figure 8.3(a) shows that Agilent
CSA-N1996A spectrum analyzer is used to take spectrum measurements, and Fig-
ure 8.3(b) shows the PSD of a pulse-shaped QPSK signal collected by the spectrum
analyzer.
There are several practical issues when parameterizing a spectrum sweep, such
as the sweep time and sweep resolution (i.e., the speed and detail at which the
spectrum sweeps are obtained). Higher resolution sweeps result in longer sweep
time, but provide a more accurate measurement of the spectrum. The sweeps used

Figure 8.2  An example of how the an LTI system h(t) can transform the PSD between the WSS
random process input X(t) and the WSS random process output Y(t).
210 Spectrum Sensing Techniques

(a)

(b)
Figure 8.3  A spectrum analyzer is employed to provide a snapshot of radio frequency bandwidth.
(a) Using Agilent CSA-N1996A spectrum analyzer to take spectrum measurements (from [10]).
(b)  Power spectral density of a pulse shaped QPSK signal, collected by an Agilent Technologies
spectrum analyzer [11].

later in this chapter are the average of thousands of spectrum sweeps across a single
bandwidth. Note that for the same measurement equipment, the speed at which
spectrum measurements can be obtained varies from seconds to hours and days
depending on the choice of several sweep parameters. However, the selection of
these sweep parameters is heavily dependent on what signals are being observed,
what sort of characteristics are being sought after in the spectrum measurements,
and how the spectrum measurement information will be post-processed afterward.
8.1  Theoretical Preparation 211

Table 8.1  Five Parameters Used to Define Spectrum Measurement Processes

Parameter Related Equation Main Impacted Aspect
Sweep resolution (δ) B1 Accuracy
δ=
NFFT
Window size (W) B Time
N step =
W
Dwell time (tdwell) t sweep Time
t dwell =
N sweep
Averaging sweeps limN ®¥ {E[n(t )]} ® 0 Accuracy
Sweep time (tsweep) t sweep = t dwell ´ N sweep Time

As a result, Table 8.1 summarizes several parameters used to define spectrum mea-
surement processes and the aspects of measurement they mainly affect. Then in the
following subsection, we will understand how the choice of these parameters can
directly impact the outcome of the measurements obtained.

8.1.2.1  Sweep Resolution

In order to provide a snapshot of the radio frequency bandwidth, a spectrum analyzer
samples the intercepted signals at a specified rate. Considering a unit time frame, each
sample in this frame displays a degree of freedom. Given a larger number of samples
in this frame, more information of the intercepted signals can be represented. Thus, for
an FFT-based spectrum analyzer, given the bandwidth of one single sweep, the sweep
resolution is decided by the number of FFT points, which can be expressed as:
B1
δ= (8.4)
NFFT
where B1 is the bandwidth of one single sweep and NFFT is the number of FFT points.
Based on (8.4), we observe that a larger number of FFT points will result in a higher
sweep resolution. For example, in Figure 8.4, there are two FFT plots of the same sine
wave. Figure 8.4(a) is generated using a 32-point FFT, and Figure 8.4(b) is a 512-point
FFT. Comparing these two plots, we find that 512-point FFT provides a higher resolu-
tion as well as a more accurate description of the spectrum than the 32-point FFT.
However, in order to handle the additional FFT samples, a spectrum analyzer
that possesses a stronger computational processing ability is required, as well as a
longer sensing time. Thus, there is an upper bound on the sweep resolution due to
the limits of computational processing ability. On the other hand, a smaller FFT
size may incur time-domain aliasing due to the undersampling in the frequency
domain. Therefore, the input spectrum data is usually zero-padded to be consistent
with the larger FFT size that provides the required resolution.

8.1.2.2  Window Size

The bandwidth of the spectrum we would like to analyze is usually larger than the
bandwidth that one single sweep can support, so the concept of “window” needs
to be introduced here. For each single sweep, the spectrum analyzer sweeps the size
212 Spectrum Sensing Techniques

N = 32
20

18

16

14

12

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a)

N = 512
20

18

16

14

12

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

(b)
Figure 8.4  Two FFT plots of the same sine wave. (a) A 32-point FFT plot. (b) A 512-point FFT plot.
8.1  Theoretical Preparation 213

of a window, and then some post-processing approaches are employed to integrate

these FFT results from the windows.
The bandwidth of the spectrum we would like to analyze can be calculated by:

B = fstop − fstart (8.5)

where fstop and fstart are two input parameters to the spectrum analyzer, which
specify the stop frequency and the start frequency. The total number of the sweeps
to cover the whole bandwidth is referred to as frequency steps, Nstep, defined as:
B
Nstep = (8.6)
W
where B is the total bandwidth and W is the window size. For example, in Figure
8.5, the whole bandwidth is four times the window size, so the number of frequency
steps is four.
Equation (8.6) implies that the larger the window size, the fewer frequency
steps are required. However, according to (8.4), for a fixed FFT size the sweep reso-
lution would become lower. In addition, the larger window size implies a higher
sampling frequency that the processing computer must be able to keep up with.
Thus, although a larger window size can effectively reduce the number of sweeps, it
may also incur some other issues that need to be taken into account.

8.1.2.3  Dwell Time

Given the window size, the dwell time is the amount of time that the spectrum
analyzer spends in each “window” or subband. This is one of the specifications
provided by most commercial spectrum analyzers, since it is more likely to generate
an accurate spectrum sensing result if enough time is spent in each subband. The
dwell time can be adjusted by changing the number of samples per FFT point input
parameter to the spectrum. If this input value is set as N, then N samples will be
collected in order to calculate the FFT for each frequency point. It usually holds
that the larger the dwell time, the larger the overall sweep time will be. These two
factors can approximately be related by:
tsweep
tdwell = (8.7)
Nsweep
where tdwell is the dwell time, tsweep is the overall sweep time, and Nsweep is the total
number of the sweeps.

Figure 8.5  A simple example showing the relation between the total bandwidth B and the window
size W.
214 Spectrum Sensing Techniques

8.1.2.4  Effect of Averaging Sweeps

When analyzing electro-magnetic spectrum, we usually sweep the spectrum several
times and then take the average of these values. This is due to the fact that the aver-
aging process can smooth out the spectrum. Since we usually assume the noise n(t)
to be an additive white Gaussian process with zero mean, as the number of samples
approaches infinity, we will have:

lim {E[n(t)]} → 0 (8.8)

N →∞

where E [ ·] is the expectation operator and N is the number of sweeps. For ex-
ample, Figure 8.6(a) shows a single sine wave whose amplitude is 1 with the added
white Gaussian noise. Figure 8.6(b) shows the result by accumulating 1000 sine
waves in Figure 8.6(a) and taking their average. Comparing these two plots, we
find that the effect of additive white Gaussian noise can be greatly eliminated by
the averaging process.
In real-world applications, as the number of sweeps used in the averaging pro-
cess increases, the noise level will be better captured. This can be used to estimate
the noise statistics and would be very useful when determining the energy threshold
for the SDR implementation of an energy detector.

8.1.2.5  Sweep Time

Given that the dwell time is the time spent in each subband, we define the sweep
time as the time spent in the entire bandwidth of interest. It is determined by sev-
eral factors, including the dwell time, window size, and the number of sweeps. For
the first factor, increasing the dwell time usually increases the sweep time. For the
second factor, increasing the window size decreases the number of frequency steps
according to (8.6), and hence decreases the sweep time. For a fixed window size,
although increasing the FFT size will result in a finer frequency resolution, the num-
ber of the frequency steps remains the same. Therefore, without accounting for the
computational processing ability, the sweep time would be the same. However, in
a real-world situation, with a larger FFT size, the resulting higher frequency resolu-
tion increases the computational processing time and hence the total sweep time.
For the last factor, it is obvious that a larger number of the sweeps will result in a
longer sweep time.
In conclusion, when collecting spectral data, we should consider all of the issues
introduced here, in order to generate an accurate result, while at the same time be
time efficient.

8.1.3 Hypothesis Testing
As mentioned at the beginning of Section 8.1, in dynamic spectrum access net-
works, spectrum sensing is employed for the purpose of identifying unoccupied
licensed spectrum, which is equivalent to detecting the frequency locations of the
primary signals. Therefore, spectrum sensing can be interpreted as a signal detec-
tion problem. Most signal detection problems can be formulated in the framework
8.1  Theoretical Preparation 215

1
x(t)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

t
(a)

1.5

0.5
x(t)

5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t

(b)
Figure 8.6  The sine wave with and without the averaging process. (a) A sine wave whose amplitude is
1 with the added white Gaussian noise. (b) Accumulating 1000 sine waves and taking their average.

of an M-ary hypothesis test, where we have an observation (possibly a vector or

function) upon which we wish to decide among M possible statistical situations
describing the observations [12]. According to this criterion, the spectrum sens-
ing performs a binary hypothesis testing in order to decide whether or not there
216 Spectrum Sensing Techniques

are primary signals in a particular channel. The two hypotheses are denoted as
follows:
H0 : no primary signals

H1 : primary signals exist (8.9)

where 0 is usually referred to as null hypothesis and 1 is usually called alterna-
tive hypothesis.
For a null hypothesis, since there are no primary signals present, the received
signal is just the noise in the RF environment. On the other hand, for the alterna-
tive hypothesis, the received signal would be the superposition of the noise and the
primary signals. Thus, the two hypotheses in (8.9) can be represented as:

H0 : x[k] = n[k]

H1 : x[k] = s[k] + n[k] (8.10)

for k = 1, ..., N, where N is the number of received signals, x [k] is the received
signal, n[k] is the noise in the RF environment, and s[k] is the primary signal. Con-
sequently, the spectrum sensing can be considered a detection problem, as based
on the observation x, we need to decide among two possible statistical situations
describing the observation, which can be expressed as:

 1 x ∈ Γ1
δ (x) = 
 0 x ∈ Γ 1c (8.11)

When the observation x falls inside the region G1, we will choose 1. However,
if the observation falls outside the region G1, we will choose 0. Therefore, (8.11)
is known as decision rule, which is a function that maps an observation to an ap-
propriate hypothesis [12]. In the context of spectrum sensing, different spectral
detectors and classifiers are actualy the implementations of different decision rules.
In Section 8.1.4, two decision rules will be introduced.
Regardless of the precise signal model or detector used, sensing errors are inevi-
table due to additive noise, limited observations, and the inherent randomness of
the observed data [13]. In testing 0 versus 1 in (8.9), there are two types of errors
that can be made, namely, 0 can be falsely rejected or 1 can be falsely rejected
[12]. In the first hypothesis, there are actually no primary signals in the channel, but
the testing detects an occupied channel, so this type of error is called a false alarm
or type I error. In the second hypothesis, there actually exist primary signals in the
channel, but the testing detects only a vacant channel. Thus, we refer to this type of
error as a missed detection or type II error. Consequently, a false alarm may lead to
a potentially wasted opportunity for the SU to transmit, while a missed detection
could potentially lead to a collision with the PU [13].
Given these two types of errors, the performance of a detector can be character-
ized by two parameters, namely, the probability of false alarm (PF), and the prob-
ability of missed detection (PM) [14], which correspond to type I and type II errors,
respectively, and thus can be defined as:
8.1  Theoretical Preparation 217

PF = P{ Decide H 1 | H 0 } (8.12)
and

PM = P{ Decide H 0 | H1 } (8.13)
Note that based on PM, another frequently used parameter is the probability of
detection, which can be derived as follows:

PD = 1 - PM = P{ Decide H 1 | H1 } (8.14)
which characterizes the detector’s ability to identify the primary signals in the chan-
nel, so PD is usually referred to as the power of the detector.
As for detectors, we would like their probability of false alarm as low as pos-
sible, and, at the same time, their probability of detection as high as possible. How-
ever, in real-world situation, this is not achievable, because these two parameters are
constraining each other. To show their relationship, a plot called receiver operating
characteristic (ROC) is usually employed [15], as shown in Figure 8.7, where its
x-axis is the probability of false alarm and its y-axis is the probability of detection.
From this plot, we observe that as PD increases, the PF is also increasing. An opti-
mal point that reaches the highest PD and the lowest PF does not exist. Therefore,
the detection problem is also a tradeoff, which depends on how the type I and type
II errors should be balanced.

0.9

0.8
Probability of Detection PD

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of False Alarm PF

Figure 8.7  A typical receiver operating characteristic (ROC), where the x-axis is the probability of
false alarm (PF) and the y-axis is the probability of detection (PD).
218 Spectrum Sensing Techniques

8.1.4  Spectral Detectors and Classifiers

The spectral detectors and classifiers are implementations of the different decision
rules. For example, the energy detector in Section 8.1.4.1 assumes that the received
signal will have more energy when the channel is occupied than when it is vacant;
thus, the two decision regions in (8.11) are divided by an energy threshold. The
cyclostationary feature detector in Section 8.1.4.2 is not only a detector, but also
a classifier that recognizes the cyclic features of different modulation schemes in
advance, so the decision regions are divided according to the cyclic features of the
modulation schemes.

8.1.4.1  Energy Detector

Energy detection uses the energy spectra of the received signal in order to identify
the frequency locations of the transmitted signal. This detection approach relies on
only the energy present in the channel, and no phase information is required. Since
the energy of a signal x (t) is defined as:

¥
E=ò | x (t ) |2 dt (8.15)
-¥

a decision statistic for energy detector can be:

N
D= å (x[k ])2 (8.16)
k =1

where x[k] is the received signal and N is the total number of the received signals.
The underlying assumption is that with the presence of a signal in the channel, there
would be significantly more energy than if there was no signal present. Therefore,
energy detection involves the application of a threshold T in the frequency domain,
which is used to decide whether a transmission is present at a specific fre­quency, as
shown in Figure 8.8. With any portion of the frequency band where the energy
­exceeds the threshold is considered to be occupied by a transmission, with the deci-
sion rule is as follows:

ì1 D >T
δ (D) = í (8.17)
î0 D <T

where D is the decision statistic calculated by (8.16) and T is the predefined energy
threshold. Therefore, in Figure 8.8, any portion of the frequency band where the
energy exceeds –100 dBm is considered an occupied channel.
Since different transmitters employ different signal power levels and trans-
mission ranges, one of the major concerns of energy detection is the selection of
an appropriate threshold. A threshold that works for one transmission may not
be appropriate for another. Figure 8.9 shows two typical detection errors caused
by inappropriate energy detection threshold. In Figure 8.9(a), the threshold is
too low, so some noise is considered primary signals, resulting in a type I error,
8.1  Theoretical Preparation 219

–50

–60

Energy spectral density (in dBm) –70

–80

–90

–100

–110

–120

–130
600 620 640 660 680 700 720 740
Frequency (in MHz)
Figure 8.8  Energy detection threshold level, denoted as T. Any portion of the frequency band
where the energy exceeds the threshold is considered occupied by a transmission.

false alarm. In Figure 8.9(b), the threshold is too high, so some primary signals
are ignored, incurring the type II error, missed detection.

8.1.4.2  Cyclostationary Feature Detector

Modulation recognition and signal classification have been subject’s of considerable
research for more than over two decades. Classification schemes can generally be
separated into one of two broad categories: likelihood-based (LB) approaches and
feature-based (FB) approaches [16]. Cyclostationary feature detection is an FB tech-
nique based on the fact that communications signals are not accurately described
as a stationary process, but rather more appropriately modeled as a cyclostationary
process [17].
A cyclostationary signal possesses statistics that vary periodically with time. By
definition, a signal x ( t ) is wide-sense cyclostationary if its mean and autocorrela-
tion are periodic [18], namely:

M x (t ) = M x (t + T0 ) (8.18)
and
Rx (t ,τ ) = Rx (t + T0 , τ) (8.19)
where Mx(t) is the mean value of the signal x ( t ) and R(t, t) is the autocorrela-
tion function of the signal x ( t ) . These periodicities occur for signals possessing
well-­defined characteristics due to several processes such as sampling, scanning,
modulating, multiplexing, and coding, which can be exploited to determine the
modulation scheme of the unknown signal [17].
The periodic nature of the signal allows it to be expressed as a Fourier series
[18, 19]:
220 Spectrum Sensing Techniques

–50

–60

Energy spectral density (in dBm) –70

–80

–90

–100

–110

–120

600 620 640 660 680 700 720 740

Frequency (in MHz)
(a)
–50

–60
Energy spectral density (in dBm)

–70

–80

–90

–100

–110

–120

–130
600 620 640 660 680 700 720 740
Frequency (in MHz)
(b)
Figure 8.9  Inappropriate energy detection threshold levels. (a) Low energy detection threshold
level yields type I error, false alarm. (b) High energy detection threshold level yields type II error,
missed detection.

ì æ τö æ τ öü
Rx (t ,τ ) = E íx ç t + ÷ x * ç t - ÷ ý = å Rxα (τ )e j 2παt (8.20)
î è 2 ø è 2 ø þ {α }

where E{.} is the expectation operator, {a} is the set of Fourier components, and
Rαx (τ ) is the cyclic autocorrelation function (CAF) given by:

1 T /2
Rxα (τ ) = lim ò-T / 2 Rx (t ,τ )e
- j 2παt
(8.21)
T ®¥ T

Alternatively, in the case when Rx (t, t) is periodic in t with period T0, (8.21)
can be expressed as:
8.1  Theoretical Preparation 221

1 T0 / 2 - j 2πα t (8.22)
Rxα (τ ) = ò-T / 2 Rx (t , τ )e
T0 0

Consequently, the Fourier transform of the CAF, which is referred to as the

spectral correlation function (SCF), is given by:

¥
S xα (f ) = ò Rαx (τ )e - j 2π f τ dτ (8.23)
-¥

which can be shown to be equivalent, assuming cycloergodicity, to the following

expression [18]:

α 1 Dt / 2 1 æ αö * æ αö
SX (f ) = lim lim
T ®¥ Dt ®¥ Dt
ò-Dt / 2 T XT çè t , f + 2 ÷ø XT çè t , f - 2 ÷ø dt (8.24)

where XT (t, f ) is the time varying Fourier transform defined as:

t +T / 2
XT (t , f ) = ò x (u)e j 2π fudu (8.25)
t -T / 2

A significant advantage of the SCF is its lack of sensitivity to additive white

noise, since modulated information is a cyclostationary process while noise is not.
In other words, the spectral components of white noise are uncorrelated, so it does
not contribute to the resulting SCF for any value of a ¹ 0. This property is especially
useful when the noise power exceeds the signal power, which would make the signal
undetectable when using a simple energy detector. As a result, cyclic detectors can
successfully operate even in low SNR environments.
To derive a normalized version of the SCF, the spectral coherence function
(SOF) is given by:
α
α SX (f )
CX (f) =
0
[S X (f + α / 2) × SX0 ( f − α / 2)]1 / 2 (8.26)

Therefore, the three important functions derived in cyclostationary feature de-

tector can be summarized as in Table 8.2.
The magnitude of the SOF varies from 0 to 1 and represents the strength of
the second order periodicity within the signal. The SOF contains the spectral fea-
tures of interest. These features are nonzero frequency components of the signal
at various cyclic frequencies. All modulation schemes contain a range of spectral
components at different cyclic frequencies, thus distinguishing them from other
modulation schemes, (i.e., the spectral components form a spectral fingerprint for
the specific modulation scheme). The SOFs of two typical modulation schemes,
QPSK and 4PAM, are shown in Figure 8.10(a) and Figure 8.10(b). Notice how the
SOF for each modulation scheme generates a highly distinct spectral image. These
distinctions allow signals to be classified from cyclic analysis. In general, the plot of
SOF is a three-dimensional figure, where the x-axis represents the cyclic frequency
222 Spectrum Sensing Techniques

Table 8.2  Three Important Functions Derived in Cyclostationary Feature Detector

Function Name Equation Characteristic
Cyclic autocorrelation 1 T /2 Fourier series of autocorrelation
  function (CAF) Rxα (τ ) = limT ®¥ ò-T / 2 Rx (t , τ )e
- j 2 πα t
T
Spectral correlation ¥ Fourier transform of the CAF
S xa (f ) = ò Rxα (τ )e - j 2π f τ dτ
  function (SCF) -¥
Spectral coherence α
α
SX (f ) Normalized version of the SCF
  function (SOF) CX (f ) = 1/ 2
 SX
0 0 
 (f + α / 2) × SX (f − α / 2)

a, the y-axis represents the spectral frequency f, and the z-axis represents the cor-
responding magnitude of the SOF for each (a, f ) pair. Note that when a does not
equal zero, the SOF values are approximately zero.
Cyclostationary feature detectors can be implemented via FFTs. Knowledge
of the noise variance is not required to set the detection threshold. Hence, the
detector does not suffer from the “SNR wall” problem of the energy detector.
However, the performance of the detector degrades in the presence of timing
and frequency jitters (which smear out the spectral lines) and RF nonlineari-
ties (which induce spurious peaks). Representative papers that consider the ap-
proach are [20–22].

8.2  Software Implementation

In Section 8.1.4, two forms of spectral detectors and classifiers were presented. In
this section, we will implement these detectors using software and apply them to
some Simulink-generated signals in order to make detection. From these software
implementations, we will have a better understanding of the characteristics of these
two detectors, as well as their specific applications.

8.2.1 Constructing Energy Detector

In the first portion of the experiment, we will generate several signals and construct
the first type of spectral detector, namely, the energy detector, according to the
theory presented in Section 8.1.4.1. Following this, we will explore one important
aspect of this detector, namely, the selection of an appropriate threshold. Note that
this portion of the chapter is entirely performed in Simulink.

8.2.1.1  Signal Generation

To get started, the first step of this experiment is to generate several signals belong-
ing to different families of modulation schemes. Later, we will apply the spectral
detectors to these signals and observe how certain modulation schemes can vary
spectrally while others possess similar characteristics.
First, let us download and open the datagen.mdl file, as shown in
­Figure 8.11. In this Simulink model, the Random Integer Generator blocks gen-
erate ­M-ary integers depending on the modulation scheme in the given branch.
8.2  Software Implementation 223

(a)

(b)
Figure 8.10  Distinctive cyclic features of different modulation schemes. (a) SOF of QPSK signal in
an AWGN channel at 10 dB SNR. (b) SOF of 4PAM signal in an AWGN channel at 10 dB SNR.
224 Spectrum Sensing Techniques

Then, the modulated signals are pulse shaped for over-the-air transmission. In the
end, vectors of each transmission are saved to the MATLAB workspace by Signal
To Workspace blocks.
Once the model has been downloaded and opened in Simulink, let us run the
model once and check the MATLAB workspace in order to see whether the sig-
nals have been saved. If yes, let us attach the Spectrum Scope block to the AWGN
Channel block to observe the output of each channel. Please note that in order to
get a satisfactory frequency resolution, the FFT length of the Spectrum Scope block
should be set properly.

8.2.1.2  Energy Detector Construction

Given these signals in the workspace, we will now proceed to construct a simple
energy detector in Simulink that will analyze the signals in the workspace and
determine whether or not a signal is present based on the decision statistic and
a threshold. However, the decision statistic defined in (8.16) is quite inflexible,
particularly in the case of narrowband signals and sinewaves. An alternative
approach could be devised by using a periodogram to estimate the spectrum via
squared magnitude of the FFT, as depicted in Figure 8.12 [27]. Correspondingly,
the following steps will assist in producing the decision statistic of the inter-
cepted signal:

Vary the SNR value of the AWGN Channel block and observe the
Q output spectrum of each channel. At what point do the signals be-
come unobservable due to noise?

Figure 8.11  Signal generation model, where signals of three different modulation schemes are
saved to the MATLAB workspace.
8.2  Software Implementation 225

1.  Prefiltering of the intercepted signal extracts the frequency band of interest.
2. Analog-to-digital conversion (ADC) transforms the filtered intercepted sig-
nal into discrete time samples.
3. Fast Fourier transform (FFT) provides the frequency representation of the
signal.
4. Square-law device yields the square of the magnitude of the frequency re-
sponse from the FFT output.
5. Average N samples of the square of the FFT magnitude can be done by a
moving average filter of length N, which windows and time-averages the last
N samples of input. The window size N should be adjusted depending upon
the performance of the system.2

Please note that in real-world applications, the first two blocks are usually im-
plemented in the hardware, which depends on the software-defined radio equip-
ment itself and is not under our control. Therefore, for the implementation in this
section, only the last three blocks need to be implemented, and the signals from
Section 8.2.1.1 will serve as the discrete-time signals.

8.2.1.3  Energy Threshold Selection and Hypothesis Testing

To perform the actual energy detection, we need to apply a threshold to the deci-
sion statistic obtained in Section 8.2.1.2 in order to decide whether a transmission
is present at a specific frequency. In other words, any portion of the frequency
band where the energy exceeds the threshold is considered occupied by a transmis-
sion, since energy detection is a binary decision-making process consisting of two
hypotheses: 0 (idle) or 1 (occupied). However, one of the primary challenges
of energy detection is the selection of an appropriate threshold. This is due to the
fact that the threshold may work for one transmission but may not be sufficient for
another, since the transmitters might be employing different signal power levels or
the transmission ranges may vary altogether.
Let us now conduct a series of trials across a range of SNR values to determine
the threshold. For each SNR value, let us conduct 10 trials, where each trial picks a

Figure 8.12  Schematic of an energy detector implementation employing prefiltering and a square-
law device to produce the decision statistic of the intercepted signals.

2. You can perform a linear averaging, which averages the N samples in a nonweighted manner. Or, you can perform
an exponential averaging, which averages the N samples in a weighted manner, and the most recent samples are
given more weighting than the older ones.
226 Spectrum Sensing Techniques

different threshold. Using this threshold, we compute the probability of false alarm
(the number of times we incorrectly say that there is a signal present) and the prob-
ability of missed detection (the number of times we did not find the signal). After
10 trials, we select the threshold that performs the best concerning both metrics
and set it as the threshold value that will be employed in the energy detection for
this specific SNR.

Which modulation scheme is most often detected? Which modula-

Q tion scheme is most often not detected?

8.2.2 Observing Cyclostationary Detector

In the second portion of the experiment, we will apply the second type of spectral
detector, namely, the cyclostationary detector, according to the theory presented in
Section 8.1.4.2. Since the theory of the cyclostationary detector is relatively com-
plicated, we will employ some ready-made M-files already containing this imple-
mentation of a cyclostationary detector and observe the results.3 Furthermore, we
will continue to use the Simulink model in Figure 8.11 to generate signals and
examine their spectra. First, let us open the datagen.mdl file and reset the para­
meters of each channel to the following values: SNR = 100 dB, β = 0.5. Run the
datagen.mdl model and save the received signals in your workspace, as this will
be used later.
All modulation schemes contain a range of spectral components at different
cyclic frequencies, thus distinguishing them from other modulation schemes. Recall
that the spectral coherence function (SOF) defined in (8.26) is a measure of the
correlation between cyclic frequencies, and cyclostationary detection relies on the
fact that the SOF for each modulation scheme generates a highly distinct spectral
image. These distinctions allow signals to be classified based on a cyclic analysis.
Therefore, in this subsection, we will focus on observing the spectral coherence
functions generated by different modulation schemes. First, let us download and
open the cyclic.m file, which generates the SOF, and use the following lines of
code to plot the SOF:

[ SCF Cx ] = cyclic ( qpsk );

Cxplot = surf ( abs ( Cx ) );
set ( Cxplot , ‘ edgecolor ’ , ‘ interp ’);

Please note that in this code, Cx is the SOF generated from the received QPSK
signal, and SCF is the corresponding spectral correlation function. The SOF is
plotted by the surf function, and the specifications of the plot is listed by the set
­function.

3. The interested reader can check out [16–19] for more information about cyclostationary detectors.
8.3  USRP Hardware Experimentation 227

1. Generate an AWGN vector and plot its SOF. Does this plot sup-
port the conclusion that cyclostationary detectors are robust to
Gaussian noise? Why?
2. Why might cyclostationary detectors be susceptible to frequency
Q selective fading?
3. Adjust the roll-off factors of the pulse shaping filters and re-
examine the coherence functions. What effect does the excess
bandwidth have on the effectiveness of the detector?

8.3  USRP Hardware Experimentation

So far we have used Spectrum Scope frequently in our Simulink experiments, as it

displays the mean-squared spectrum or power spectral density of the input signal.
However, the bandwidth that we can observe using this scope is limited. For exam-
ple, in Section 5.2.1, Spectrum Scope is used to observe the FFT of the received sig-
nal, as shown in Figure 8.13, where the observed frequency range is from –100 kHz
to 100 kHz. In other words, the bandwidth is 200 kHz.
If we want to observed the FFT over a wider frequency range, for example,
4.9 GHz to 5.9 GHz, which is one of the frequency ranges that the XCVR2450
daughterboard supports, can we still use the Spectrum Scope? The answer is yes.
However, we need to divide the wideband channel into several narrowband sub-
channels in order to use the Spectrum Scope on each subchannel, and then apply
a “sliding window” approach to integrate the FFT results. In this section, we will
implement a wideband spectrum sensing technique using two types of sliding win-
dow. This can be accomplished by the following steps.
To get started with this implementation, let us take the Simulink design of
energy detection from Section 8.2.1.2 and incorporate into it the SDRu Receiver
block. By doing this, we can use USRP2 to collect the spectrum data.
Now, let us proceed to develop a sliding window approach upon the energy
detection implementation from the previous step such that it can cover a wider
frequency range, for instance, 200 MHz. Two types of windows will be employed
here, namely,

• A rectangular window, as shown in Figure 8.14(a), which starts at f1 and

ends at f2 sharply.
• A tapered window, such as a Hamming window [23], as shown in Figure
8.14(b), which also starts at f1 and ends at f2, but with gradual transitions
on both ends.

The purpose here is to sweep individual segments of a much wider frequency

range using these two types of windows and then post-process them into a single
frequency sweep result, as shown in Figure 8.15. When performing spectral analysis
on finite length of data, a key function of these windows is to minimize the dis-
continuities of truncated spectrum, thus reducing spectral leakage [24], since the
228 Spectrum Sensing Techniques

Figure 8.13  Using Spectrum Scope to observe the FFT of the received signal, where the observed
frequency range is from −100 kHz to 100 kHz and the bandwidth is 200 kHz.

amount of spectral leakage depends on the amplitude of the discontinuity. As the

discontinuity becomes larger, spectral leakage increases, and vice versa. Compar-
ing Figure 8.15(a) and Figure 8.15(b), we can find that the tappered windows can
effectively reduce the amplitude of the discontinuities at the boundaries of each
period, and the resulting spectrum gradually tapers to zero at the ends. Minimizing
the discontinuities along the transition edges is one of the advantages of tapered
windows over the rectangular windows.
As long as these two types of windows have been implemented, let us apply
them for 5.1 GHz to 5.3 GHz frequency range and repeat the spectrum sens-
ing for 25 times, respectively. In the end, average the results for both window
types.

Q What do you observe?

Then, let us transmit a sinusoidal signal in that frequency range and perform
the test again.

Q Can you see the sinusoidal signal on your results?

8.3  USRP Hardware Experimentation 229

(a)

(b)
Figure 8.14  Two types of windows, which will be used to divide the wideband channel into narrow­
band subchannels. (a) Rectangular window. (b) Hamming window.

(a)

(b)
Figure 8.15  A single frequency sweep result from two types of windows. (a) A single frequency
sweep result using a rectangular window. (b) A single frequency sweep result using a Hamming
window. Note there are overlaps in transition areas.
230 Spectrum Sensing Techniques

Let us summarize our observations by answering the following

questions:
  1. How long does it take to sweep 200 MHz of spectrum?
  2. I s there a relationship between sweep time, FFT size, and fre-
quency resolution?
  3. What do you notice about the quality of your spectrum sweep
Q when you average more sweeps?
  4. What is the impact of the individual window size when sweep-
ing a large frequency band?
  5. Is it possible to obtain the frequency response of a wideband
spectral range using a single FFT and no windowing? If yes,
is there a tradeoff with frequency resolution? If no, please ex-
plain the reason.

8.4 Open-Ended Design Project: CSMA/CA

In the previous experiments found within this book, we were dealing with a wireless
network consisting of only two nodes. However, in many real-world situations, we
encounter multiple nodes communicating with each other in a distributed fashion,
in which these nodes share a common channel or medium as well as employ a me-
dium access control (MAC) protocol that enables everyone access to this medium.
In this section, we will design and implement an SDR communication system ca-
pable of achieving medium access control using a simple random backoff process
for collision avoidance and a spectrum sensing procedure that has been designed in
the earlier parts of this chapter.

8.4.1 Carrier Sense Multiple Access

Carrier sense multiple access (CSMA) is a probabilistic media access control
protocol in which a node verifies the absence of other traffic before transmit-
ting on a shared transmission medium, such as an electrical bus, or a band of
the electromagnetic spectrum [25]. CSMA is composed of two logically defined
components:

• Carrier sense describes the fact that a transmitter uses feedback from a re-
ceiver that detects a carrier wave before trying to send. That is, it tries to
detect the presence of an encoded signal from another station before attempt-
ing to transmit. If a carrier is sensed, the station waits for the transmission in
progress to finish before initiating its own transmission.
• Multiple access describes the fact that multiple stations send and receive on
the medium. Transmissions by one node are generally received by all other
stations using the medium.
8.4  Open-Ended Design Project: CSMA/CA 231

8.4.2 Collision Avoidance
Carrier sense multiple access with collision avoidance (CSMA/CA) is a variant of
CSMA. Collision avoidance is used to improve CSMA performance by not allow-
ing wireless transmission of a node if another node is transmitting, thus reducing
the probability of collision due to the use of a random binary exponential backoff
time.
In this protocol, a carrier sensing scheme is used, namely, a node wishing to
transmit data has to first listen to the channel for a predetermined amount of time
to determine whether or not another node is transmitting on the channel within
the wireless range. If the channel is sensed “idle,” then the node is permitted to
begin the transmission process. If the channel is sensed as “busy,” the node defers
its transmission for a random period of time, namely the backoff time. A simple
example of CSMA/CA is shown in Figure 8.16, where node 1 wants to transmit,
but the channel is sensed as “busy,” so node 1 defers its transmission for five
slots.

8.4.3 Implementation Approach
Although medium access control is usually employed in multiple node communica-
tion systems, as a starting point, we will implement a two-node (radio 1 and radio 2)
communication system in this section. These two radios are both transceivers, shar-
ing a common communication channel and using CSMA/CA protocol. Since they are
transceivers, they can switch between transmit and receive modes. We can define the
duration of each mode. Please note since this system is also a half-duplex communica-
tion system, we can build it upon the system obtained in Section 7.4. However, unlike
Section 7.4, where two MATLAB scripts are used to control the switch between dif-
ferent modes, in this section, CSMA/CA is employed to switch the modes. Therefore,
our system should perform the following three stages, as shown in Figure 8.17:

Figure 8.16  An example of a 3-node communication system employing CSMA/CA protocol (based
on [26]), where node 1 employs a backoff time of five time slots.
232 Spectrum Sensing Techniques

• Stage 1:
– Radio 1 is in transmit mode. It switches between transmitting a random num-
ber of “Hello world” frames and listening for a codeword from radio 2.
– Radio 2 does the spectrum sensing using energy detection. If the channel
is sensed as “busy,” radio 2 enters a random backoff time. If the channel
is sensed “idle,” radio 2 begins to transmit a codeword to radio 1 and the
system enters stage 2.
• Stage 2:
– Radio 2 is in the transmit mode. It repeats sending a random number of
“change” to radio 1.
– Radio 1 is in the receive mode, and it does the spectrum sensing using en-
ergy detection. If the channel is sensed as “busy,” radio 1 keeps receiving
the incoming message. If the channel is sensed “idle,” radio 1 begins to
decode the message it has received, and the system enters stage 3.
• Stage 3:
– If radio 1 can get at least one “Change” in its decoded message, radio 1
begins to broadcast “Goodbye,” and their communication ends.
– If radio 1 does not get any “Change” in its decoded message, the whole
system returns to the origin and starts from stage 1.

8.4.4  Useful Suggestions

Since energy detection is used to determine whether the channel is idle or busy,
we should select a reasonable threshold for our detector. Besides the two radios,
we need to implement an exponential random number generator, which is used
to generate the backoff time for CSMA/CA protocol while the system is run-
ning. The number of “Hello world” frames and the number of “change” code-
words can also be exponential random variables. However, they don’t need to

Figure 8.17  Three stages in CSMA/CA protocol implementation.

8.4  Open-Ended Design Project: CSMA/CA 233

Figure 8.18  An example of a successful implementation at radio 2, which demonstrates the three
stages introduced in Section 8.4.3.

be generated­while the system is running. Instead, they can be defined before the
system is running.

8.4.5 Evaluation and Expected Outcomes

With individual models for both radios completed, we can proceed to the testing
process. Let us start the two models at the same time, which makes radio 1 initiate­
transmission and radio 2 begin sensing. As soon as radio 2 detects an occupied
channel, it will evaluate the received message and send a “Change” message back if
it has received the correct message “Hello world.” Since radio 1 is in sensing mode
right now, once it detects an occupied channel, it will decode the received message
and send a “Goodbye” message back if a “Change” message is received. In this way,
the two radios will relay the message back and forth to each other, demonstrating
a half-duplex communication system. The most important task in this process it to
find out the proper sensing time, transmission time, and back-off interval for both
radios.
An example of a successful implementation at radio 2 is displayed in Figure
8.18, where the three stages introduced in Section 8.4.3 are clearly demonstrated.
According to the figure, we can see that once “Goodbye” is received, radio 2 enters
234 Spectrum Sensing Techniques

stage 1 and keeps sensing. When “Hello world” is received, radio 2 continues to
stage 2 and begins transmitting. After a predetermined transmission time, radio 2
proceeds to stage 3, in the listening mode.

8.5 CHAPTER SUMMARY

This chapter focuses on the principle and implementation of two spectrum sensing
techniques, namely, the energy detection and the cyclostationary feature detection.
In this chapter, the theory of spectrum sensing and the practical issues of collecting
spectral data are first introduced, followed by the implementation of the energy
detector and the observation of the cyclostationary feature detector in MATLAB.
Using USRP, a wideband spectrum sensor is implemented in the hardware experi-
ment. In the end, the open-ended design project realizes the CSMA/CA protocol,
which heavily relies on the spectrum sensing techniques and enables the multiple
access schemes.

8.6  PROBLEMS

1. [Power Spectral Density] Calculate and plot the power spectral density of a
100 MHz sinusoidal tone. How would you expect this to look on a spectrum
analyzer?
2. [Power Spectral Density] The power spectral density of a narrowband noise
signal, n(t), is shown in Figure 8.19. The carrier frequency is 10 Hz.
(a) Find the power spectral densities of the in-phase and quadra-
ture components of n(t).
(b)  Find their cross-spectral densities.
3. [Power Spectral Density] The input to a time-invariant linear system with
impulse response h(t) is a white Gaussian noise process X(t) with two-
No
sided spectral density . The output is the random process Y(t). The fil-
2
ter’s response is given in terms of pT(t), the rectangular pulse of duration T,
sin(2π t)
by h(t) = * pT (t).
T
(a) Find the cross-correlation function RXY (t), −¥ < t < ¥.
(b) Find RY Y (0).

1 S nn (f )

–12 –10 –8 8 10 12 f
Figure 8.19  Power spectral density of a narrowband noise signal, n(t).
8.6  PROBLEMS 235

4. [Hypothesis Testing] Given a random variable X ~ N (μ, s2), where μ and s2

1 n
are unknown. x1, x2, . . . , xn are n samples of X. Let us define x = å xi,
n n i =1
Q = å (xi - x)2. Suppose we have the following hypothesis testing:
i =1
H0 : µ = µ0
(8.27)
H1 : µ ¹ µ0
(a)  What should be the decision statistic D?
(b)  What is the region to reject H0?
5. [HypothesisTesting] Given a random variable X ~ N(μ, s2), where μ is un-
known. x1, x2, . . . , xn are n samples of X, whose variance is S 2. Suppose we
have the following hypothesis testing:
H0 : σ 2 ³ 16
(8.28)
H1 : σ 2 < 16
(a)  What should be the decision statistic D?
(b)  What is the region to reject H0?
6. [Hypothesis Testing] Given two random variables X ∼ N(µ1 , σ 12 ) and
Y ∼ N(µ2 , σ 22 ), where σ 12 and σ 22 are unknown. x1 , x2 ,..., xn1 are n1 samples
of X, y1 , y2 ,..., yn2 are n2 samples of Y. All these samples are independent.
Suppose we have the following hypothesis testing:
H0 : µ1 = µ 2
(8.29)
H1 : µ1 ¹ µ 2
(a)  What should be the decision statistic D?
(b)  What is the distribution of D?
7. [Hypothesis Testing] A factory manufactures one type of 68 mm nail. In real
production, its length satisfies the Gaussian distribution of N(μ, 3.62). Sup-
pose we have the following hypothesis testing:
H0 : µ = 68
(8.30)
H1 : µ ¹ 68
Assume x is the mean of n samples and the testing will be conducted according
to the following criteria: When x - 68 > 1, reject H0; otherwise, accept H0.
(a) When n = 36, what is the probability of type I error?
(b) When n = 64, what is the probability of type I error?
(c) When n = 64 and H0 does not hold (assume μ = 70), what is
the probability of type II error?
8. [Hypothesis Testing] Consider the problem of detecting between two real
2
Gaussian random variables with means μi and variance σ i , i = 0, 1. Show that
the PFA is given by
æ τ - µ0 ö
P(Z > τ | H 0) = Q ç ÷ (8.31)
è σ0 ø
1 ¥ -t 2 / 2
where Q(τ ) :=
2π òτ e dt is the standard Gaussian tail function.
 9. [Cyclostationary Process] If x(t) is wide-sense cyclostationary, its mean and
τ τ
autocorrelation are periodic. Express Rx çæ t - , t + ö÷ as a Fourier series.
è 2 2ø
If x(t) is cyclostationary, what does that imply about frequency components
at m/T0, where m is an integer?
10. [Cyclostationary Process] Given the cyclic autocorrelation function in
(8.21) and using the Einstein-Wiener-Khintchine (EWK) relation, derive
the spectral correlation function of a wide sense cyclostationary process.

References

[1]  M itola III, J., and G. Q. Maguire, “Cognitive Radio: Making Software Radios More Per-
sonal,” IEEE Personal Communications, Vol. 6, No. 4, Aug. 1999, pp. 13–18.
[2]  Haykin, S.,”Cognitive Radio: Brain-Empowered Wireless Communications,” IEEE Jour-
nal on Selected Areas in Communications, Vol. 23, No. 2, Feb. 2005, pp. 201–220.
[3]  McHenry, M. A., P. A. Tenhula, D. McCloskey, D. A. Roberson and C. S. Hood, “Chicago
Spectrum Occupancy Measurements Analysis and a Long-Term Studies Proposal,” Proceed-
ings of Workshop on Technology and Policy for Accessing Spectrum, Boston, MA, 2006.
[4]  Pagadarai, S., and A. M. Wyglinski, “A Quantitative Assessment of Wireless Spectrum
Measurements for Dynamic Spectrum Access,” Proceedings of the IEEE International
Conference on Cognitive Radio Oriented Wireless Networks and Communications,
­Hannover, Germany, 2009.
[5]  Pagadarai, S., R. Rajbanshi, and A. M. Wyglinski, “Agile Transmission Techniques,” Cog-
nitive Radio Communications and Networks: Principles and Practice, Elsevier, 2009.
[6]  Akyildiz, I. F., W. Y. Lee, M. C. Vuran, and S. Mohanty, “NeXt Generation/Dynamic
Spectrum Access/Cognitive Radio Wireless Networks: A Survey,” Computer Networks,
Vol. 50, No. 13, Sept 2006, pp. 2127–2159.
[7]  Zhao, Q., and B. M. Sadler, “A Survey of Dynamic Spectrum Access,” IEEE Signal Pro-
cessing, Vol. 24, No. 3, May 2007, pp. 79–89.
[8]  Peng Ning, Y. L., and H. Dai, “Authenticating Primary Users’ Signals in Cognitive Radio
Networks via Integrated Cryptographic and Wireless Link Signatures,” IEEE Symposium
on Security and Privacy, Oakland, CA, May 2010.
[9]  Hanssen, A., and L. Scharf, “A Theory of Polyspectra for Nonstationary Stochastic Processes,”
IEEE Transactions on Signal Processing, Vol. 51, No. 5, May 2003, pp. 1243–1252.
[10]  WPI Wireless Innovation Lab, Welcome to SQUIRRELWeb, http://www. spectrum.wpi.
edu/.
[11]  Leferman, M. J., “Rapid Prototyping Interface for Software Defined Radio Experimenta-
tion,” Master of Thesis Worcester Polytechnic Institute, Worcester, MA, 2010.
[12]  Poor, H. V., An Introduction to Signal Detection and Estimation, Springer, 2010.
[13]  Zhao, Q., and A Swami, “Spectrum Sensing and Identification,” Cognitive Radio Com-
munications and Networks: Principles and Practice, Elsevier, 2009.
[14]  Kay, S. M., “Statistical Decision Theory I,” Fundamentals of Statistical Signal Processing,
Volume 2: Detection Theory, Prentice Hall, 1998.
[15]  Shanmugan, K. S., and A. M. Breipohl, “Signal Detection,” Random Signals: Detection,
Estimation and Data Analysis, Wiley, 1988.
[16]  Like, E. C., Non-Cooperative Modulation Recognition Via Exploitation of Cyclic Statis-
tics, Master’s Thesis, 2007.
[17]  Like, E. C., V. D. Chakravarthy, P. Ratazzi, and Z. Wu, “Signal Classification in Fading
Channels Using Cyclic Spectral Analysis,” EURASIP Journal on Wireless Communications
and Networking, 2009.
8.6  PROBLEMS 237

[18]  G ardner, W. A., W. A. Brown, and C.-K. Chen, “Spectral Correlation of Modulated
Signals—Part II: Digital Modulation,” IEEE Transactions on Communications, Vol. 35,
No. 6, 1987, pp. 595–601.
[19]  Gardner, W. A., Cyclostationarity in Communications and Signal Processing, IEEE Press,
Piscataway, NJ, 1993.
[20]  Cabric, D., S. M. Mishra, and R. W. Brodersen, “Implementation Issues in Spectrum Sens-
ing for Cognitive Radios,” Proceedings of the Asilomar Conference on Signals, Systems
and Computers, 2004.
[21]  Kim, K., I. A. Akbar, K. K. Bae, J.-S. Um, and C. M. Spooner, “Cyclostationary Approaches
to Signal Detection and Classification in Cognitive Radio,” Proceedings of IEEE Sympo-
sium on New Frontiers in Dynamic Spectrum Access Networks, Dublin, Ireland, 2007.
[22]  Sutton, P. D., K. E. Nolan, and L. E. Doyle, “Cyclostationary Signatures in Practical Cogni-
tive Radio Applications,” IEEE Journal on Selected Areas in Communications, 2008.
[23]  Hamming, R. W., “Windows,” Digital Filters, 3rd edition, Dover Publications, 1997.
[24]  Harris, F. J., “On the Use of Windows for Harmonic Analysis with the Discrete Fourier
Transform,” Proceedings of IEEE, Vol. 66, No. 1, Jan. 1978, pp. 51–83.
[25]  Tanenbaum, A. S., “The Medium Access Control Sublayer,” Computer Networks, 4th edi-
tion, Prentice Hall, 2002.
[26]  http://www.hpl.hp.com/personal/Jean_Tourrilhes/Linux/Linux.Wireless.mac.html.
[27]  Cabric D., A. Tkachenko, and R. W. Brodersen, “Experimental Study of Spectrum Sensing
Based on Energy Detection and Network Cooperation,” Proceedings of the First Interna-
tional Workshop on Technology and Policy for Accessing Spectrum, Boston, MA, 2006.
 Chapter 9

Applications of Software-Defined Radio

Given the latest advances in SDR technology with respect to computational horse-
power, energy efficiency, and form factor compactness, communication system de-
signers are devising innovative wireless transmission implementations capable of
enabling a wide range of wireless data transmission applications deemed either im-
possible or prohibitively costly only several years ago. Until this chapter, the focus
of this book was primarily on providing a solid theoretical foundation with respect
to digital communications and related areas, as well as synthesizing those concepts
in a series of practical, hands-on exercises that help provide a deeper understanding
about how digital transceivers actually operate in the real world.
In this chapter, we provide a context for the theory and practice covered thus
far in this book by exploring three state-of-the-art applications where SDR technol-
ogy has made significant, almost revolutionary, contributions. Specifically, we first
start in Section 9.1 with an exploration of wireless systems capable of performing
autonomous decision making, referred to as cognitive radio. Then, we focus in Sec-
tion 9.2 on the use of SDR technology in enabling next-generation wireless com-
munications in vehicular networks. Finally, we study how SDR technology can be
adapted for satellite communication systems in Section 9.3 before concluding this
chapter.

9.1  Cognitive Radio and Intelligent Wireless Adaptation

Cognitive radio [1] is a communication systems paradigm that focuses on employ-

ing highly agile, environmentally aware, intelligent wireless platforms in order
to autonomously choose and fine-tune device operating parameters based on the
prevailing radio and network environmental conditions. In many situations, SDR
technology is primarily responsible for making cognitive radio a reality due to its
operational agility resulting from a digital baseband that is entirely implemented in
software and/or programmable logic.
In order to enable a substantial level of intelligent decision making in these
cognitive radio systems, machine learning techniques have been proposed to either
partially or entirely automate the (re)configuration process (see [2–6] and refer-
ences therein). However, the solutions produced by these techniques often depend
on explicit inputs by the human operators, which require some knowledge of the
wireless device and the target networking behavior (e.g., high data rates, high error
robustness, and low network latency [7]). Moreover, the target behavior specified
by the human operators may not match the actual target behavior.
Most modern wireless networking devices support multiple applications, such
as streaming multimedia content via the Internet, voice-over-IP telephony, online

239
240 Applications of Software-Defined Radio

distance learning, and electronic file downloading/sharing. Since these applications

possess different device configurations in order to achieve the desired user experi-
ence, many wireless networking devices have some degree of flexibility when choos-
ing a configuration. However, the configuration selection process is still mostly
conducted by a human operator or via an exhaustive search approach when the
number of parameters representing the configuration is small. Nevertheless, for
scenarios possessing a large device configuration solution space, researchers are
studying the use of machine learning techniques in adaptation algorithms for wire-
less networking devices [4–6].
One major issue is the accuracy of decisions made by the adaptation algo-
rithm, which is based on the quality and quantity of inputs to the device. Thus,
with more information available to the algorithm, it becomes capable of making
more informed decisions that achieve the desired user experience more precisely.
Referring to Figure 9.1, three types of parameters employed by the adaptation
algorithm are:

1. Device configurations: Essentially the “knobs” of the wireless networking

device, these are a collection of parameters that can be altered to change the
current operating state of the device. Note that several potential configu-
rations may not be possible to implement and are thus disallowed by the
adaptation algorithm.
2. Environmental parameters: Represented as the “dials” in Figure 9.1, these
parameters represent the information about the current status of the device,
as well as its sensed wireless environment using external sensors.
3.  Target networking experience: This is usually represented by quantitative
“metric” that approximately describes the average human user’s experience
when operating the wireless networking device. The goal of the adaptation
algorithm is to achieve the best-possible value for a given metric.

Figure 9.1  Wireless networking device employing “knobs” and “dials” in order to perform adaptation.
9.1  Cognitive Radio and Intelligent Wireless Adaptation 241

Note that a combination of a device configuration and a set of environmental

parameters may yield a high value for a specific metric given one wireless network-
ing application, while resulting in a low value for another different application. The
following subsection describes these types of parameters in more detail.

9.1.1  Wireless Device Parameters

The definition of an optimal decision is a combination of device configuration
and environmental parameters that achieve the target networking experience
better than any other combination available. Defining a proper list of parameters
constituting a device configuration to be employed in an adaptation algorithm
is of prime importance. A well-constructed list consists of common wireless
parameters that each possess a large impact on the target networking behavior.
Table 9.1 shows a list of nine transmission parameters commonly used in wireless
networking devices.
Environmental parameters inform the system of the surrounding environment
characteristics. These characteristics include internal information about the device
operating state and external information representing the wireless channel environ-
ment. Both types of information can be used to assist the cognitive radio in making
decisions. These variables along with the target networking experience are used as
inputs to the algorithm. Table 9.2 shows a list of six environmental parameters that
can affect the operational state of a cognitive radio device.
The purpose of a machine learning-based adaptation algorithm is to autono-
mously improve the wireless networking experience for the human operator. How-
ever, improving this experience can mean several different things. Much research is
focused on improving the accommodation of many wireless users within the same
network. Other important aspects include providing error-free communications,
high data rates, limiting interference between users, and even the actual power
consumption of the wireless networking device, which is extremely important in
mobile applications. As shown in Table 9.3, we have defined five common target
networking experiences that guide the adaptation algorithm to a specific optimal
solution for the cognitive radio system.
The target experiences presented in Table 9.2 represent the means for guid-
ing the state of the adaptation process for the wireless networking device. The

Table 9.1  Several Common Wireless Networking Device Configuration Parameters (from [1, 5])
Parameter Description
Transmit power Raw transmission power
Modulation type Type of modulation
Modulation index Number of symbols for a given modulation scheme
Carrier frequency Center frequency of the carrier
Bandwidth Bandwidth of transmission signal
Channel coding rate Specific rate of coding
Frame size Size of transmission frame
Time-division duplexing Percentage of transmit time
Symbol rate Number of symbols per second
242 Applications of Software-Defined Radio

Table 9.2  Several Common Wireless Networking Environmental Parameters (from [1, 5])
Parameter Description
Signal power Signal power as seen by the receiver
Noise power Noise power density for a given channel
Delay spread Variance of the path delays and their amplitudes for a channel
Battery life Estimated energy left in batteries
Power consumption Power consumption of current configuration
Spectrum information Spectrum occupancy information

Table 9.3  Several Common Wireless Networking Target Experiences (from [1, 5])
Objective Description
Minimize bit error rate Improve overall BER of the transmission environment
Maximize data throughput Increase overall data throughput transmitted by radio
Minimize power consumption Decrease amount of power consumed by system
Minimize interference Reduce the radio interference contributions
Maximize spectral efficiency Maximize the efficient use of the frequency spectrum

algorithm makes use of these experiences through relationships that describe

how modifying the device parameters achieve these objectives. To facilitate the
decision-­making process, each target networking experience must be represented
by a mathematical relationship that relates a device configuration and environ-
mental parameters to a scalar value that describes how well this set achieves the
specific goal [1, 5]. These functions will provide a way for the adaptation algo­
rithm to rank combinations of configurations and environmental parameters, ul-
timately leading to a final decision.
Note that it is possible to specify several target networking experiences si-
multaneously, with the final score being represented by a numerical value. In this
case, the individual scores of the target experiences are weighted according to
their importance in the specific application and summed together, forming the
final overall score [5].

9.2 Vehicular Communication Networks

All forms of wireless communications require access to radio frequency (RF) spec-
trum. However, given the finite amount of RF spectrum and the rapidly growing
number of wireless applications and end users, many sectors are experiencing the
effects of spectrum scarcity due to the inefficient utilization of this natural re-
source. One solution that is capable of accommodating this growing demand for
RF spectrum while providing greater wireless access is dynamic spectrum ­access
(DSA). As opposed to the conventional spectrum allocation paradigm, which fo-
cuses on static licensed spectrum assignments where the license holders possess
exclusive transmission rights, the DSA paradigm enables unlicensed devices to
temporarily “borrow” unused licensed spectrum while ensuring the rights of the
incumbent license holders are respected. One technology capable of achieving
9.2  Vehicular Communication Networks 243

DSA is cognitive radio, which often employs autonomous and flexible communi-
cations techniques implemented via an SDR programmable wireless communica-
tions platform.1
There has been much research and development over the past several decades
in order to facilitate ubiquitous, reliable, and efficient wireless data connectivity
between individual vehicles as well as with roadside information infrastructure [8].
The objective of these research efforts into vehicular networking is to further en-
hance the overall driving experience while on the road [9–12]. In particular, wire-
less connectivity can be used to enable a greater level of vehicular safety such that
the driver’s situational awareness is enhanced [13]. Given the potential benefits of
vehicular networking, it is expected that commercial vehicle manufacturers will
introduce this technology in the near future [14–16].
Despite the advantages of employing wireless systems in vehicular applica-
tions, there is increasing concern that the rapid growth of this sector coupled with
significant bandwidth requirements of the vehicular wireless transmissions will
result in a spectrum scarcity issue. One remedy for this issue is to employ SDR
technology that is specifically designed to perform DSA networking across several
vehicles as well as with associated roadside infrastructure, which is referred to
as vehicular dynamic spectrum access (VDSA) [15, 17, 18]. Given the potential
of this solution, the feasibility of VDSA has been extensively studied for several
candidate frequency bands [17], networking protocols enabling VDSA commu-
nications have been assessed [14–16], the theoretical capacity of VDSA networks
have been derived [18], and experimental prototype VDSA platforms and test-
beds have been constructed [19, 20].

9.2.1 VDSA Overview
VDSA operates by establishing a wireless data transmission link between any two
vehicular communication devices across unoccupied portions of licensed spectrum.
Thus, a vehicular wireless device must be capable of identifying all the unoccupied
frequency bands via spectrum sensing prior to any data transmission. An example
of VDSA is shown in Figure 9.2 where “car A” communicates with “car B” by
forming a secondary (unlicensed) user (SU) transmission link LAB across some un-
occupied spectrum that is located between two primary (licensed) user (PU) signals.
Suppose that all vehicular communications is decentralized in this scenario, and
that all information is shared between vehicles using multihop relaying. Conse-
quently, wireless data transmission is subsequently performed between “car B” and
“car C” via the transmission link LBC , which is also located within a region of
unoccupied spectrum, in order to relay the information to all vehicles within the
geographical vicinity.

1. Given the growing number of wireless interfaces being employing on a single vehicle, it is also anticipated
that automobile manufacturers will eventually integrate all the communication systems on a vehicle into a
single SDR platform. Such an approach would facilitate remote updating of wireless standards and protocols
on vehicles without the need for an expensive upgrade recall, as well as simplify the communications
infrastructure on the vehicle itself.
244 Applications of Software-Defined Radio

Frequency

Figure 9.2  An illustration of the vehicular dynamic spectrum access (VDSA) concept (from [20]). In this
scenario, the vehicular secondary (unlicensed) users (SUs) temporarily access unoccupied spectrum
from primary (licensed) users (PUs) for performing wireless data transmission while on the road.

9.2.2 Transmitter Design
An architecture for a VDSA transmitter is shown in Figure 9.3, where the primary
concern when transmitting information within a VDSA networking architecture is
the potential for interference with existing primary and other secondary user signals
located within the spectral and geographical vicinity. As a result, several factors
must be considered when designing the VDSA transmitter, namely [20]:

• Finding primary user – unoccupied frequencies: The identification of unoccu-

pied frequency bands for performing VDSA will be conducted using a com-
bination of spectrum database searching and active spectrum sensing. The
former is expected to provide a rough assessment of all spectrum opportuni-
ties within a geographical vicinity, while the latter fine tunes the spectrum
situational awareness of the VDSA network by indicating the instantaneous
availability of a specific frequency band.
• Avoiding secondary users: The mantra of any DSA paradigm and their vari-
ants is that the secondary users can never interfere with any of the incumbent
primary (licensed) transmissions co-existing within a specific geographical and
frequency vicinity. Since there might not be any centralized coordination with
respect to the spectrum access between primary and secondary users, the latter

Figure 9.3  An illustration of a VDSA transmitter design showing how it searches for and avoids both
PUs and SUs, as well as determine the most optimal location for the next SU carrier frequency, prior
to secondary transmission [20].
9.2  Vehicular Communication Networks 245

may need to employ some sort of contention-based physical layer medium ac-
cess control (MAC) in order to check out the availability of a frequency band
to avoid interference with the primary user or other secondary users.
• Message generation and protocol: Assuming that the VDSA network is op-
erating in a decentralized mode, it is necessary that the header information
of each frame shared within the network contains information needed to
facilitate the formation and continuation of any VDSA transmission links.
Examples of various information to be included in the header are the unique
identification number of the source vehicular wireless platform, time stamp
information, number of hops the message can be relayed before it is consid-
ered out of date, and identification number of the destination vehicle.

9.2.3  Receiver Design

The receiver design also heavily relies on the occupancy behavior of the wireless
spectrum. For instance, referring to Figure 9.3, the power spectral density (PSD)
is derived from the spectrum sensing measurements and all peaks exceeding three
standard deviations of the mean PSD are identified, (i.e., three-sigma outliers). The
center frequency for each of the three-sigma outlier peaks is located, and the inter-
cepted message is then demodulated and decoded. Consequently, the authors in [20]
proposed two important steps that a VDSA receiver should possess when determin-
ing how to access the wireless spectrum in a secondary manner:

• Spectrum sensing and statistical analysis: Using energy detection, which is a

simple form of discriminating between noise and signals across a frequency
band via the use of an energy detection threshold [16], the transmitter and
receiver of a VDSA wireless link attempt to locate each other across a fre-
quency band in order to lock onto each other to perform data transmission.
Given that energy detection is a relatively simple spectrum sensing technique,
the value of the energy detection threshold used to differentiate between a
signal and spurious noise was selected such that only three-sigma outliers
[21] of the energy detector were identified as potential signals.
• Center frequency detection: Once the frequency locations of the VDSA trans-
missions have been identified, the surrounding noise is filtered out and the
process of fine tuning the center frequency commences. This can be accom-
plished by integrating the possible signal PSD over signal bandwidth, which
removes the impact of the noise present in the channel, and then finding the
frequency value that partitions the overall signal energy by half assuming a
symmetric PSD. Once the center frequency of the potential signal has been
determined, the receiver attempts to demodulate and decode the signal at
that center frequency.

9.2.4 VDSA Test-Bed Implementation

Experimentation of an actual VDSA network employing vehicles operating under high-
way conditions was conducted at Worcester Polytechnic Institute (WPI) on March 12,
2011 at 12:50 PM. The experiment consists of five vehicles, as shown in Figure 9.4,
246 Applications of Software-Defined Radio

Figure 9.4  A photograph of the actual VDSA networking experiment conducted on March 12,
2011 at 12:50 PM [20]. Picture taken from car D. The figure shows the VDSA network implemented
on the highway I-190 north of Worcester, MA. As indicated on the figure, the leading car is perform-
ing spectrum sensing and the rest of the cars form the proposed VDSA networking architecture.

where the leading car was equipped with a spectrum analyzer in order to record all
the VDSA transmissions occurring during the experiment, while each of the remaining
cars was equipped with a USRP2 SDR platform with its antenna anchored to the ex-
terior of each vehicle. The experiment was initiated in Worcester, MA along Interstate
I-190 north to Leominster, MA followed by heading east along Route 2 until the in-
tersection with Interstate I-495, where the proposed VDSA network proceeded south
until the intersection between Interstate I-495 and Interstate I-290, where the vehicles
finally headed back west on Interstate I-290 to Worcester, MA.
During the experiment, car 1 of the VDSA network was continuously broadcast-
ing hazard messages, while the remaining three cars constantly performed spectrum
sensing over a 100-MHz-wide frequency band from 2.4 GHz to 2.5 GHz and per-
formed multihop relaying in the event that a message was successfully intercepted,
as shown in Figure 9.5. The receiver would lock on to the frequency and start to
decode the message as soon as transmission was detected using the three-sigma rule.
If the message was successfully decoded and recorded by the USRP2, it switched to
transmit mode and broadcasted the message to the rest of the cars. In the event that
no message was decoded, or the decoded message did not correspond to the desired
network protocol, the USRP2 ignored the transmission and started another itera-
tion of spectrum sensing. Each radio was connected to a laptop that would record
and time stamp every received message for performance analysis.

9.3 Satellite Communications

In the past, communication systems often required specialized hardware in order

to achieve some level of functionality. Thus, to keep production costs low, these
9.3  Satellite Communications 247

Figure 9.5  Three-dimensional plot showing a sample of the experimental results obtained from
the VDSA test-bed involving the secondary user transmission in the neighborhood of licensed user
signals. Notice the multihop transmission of “car 1”®“car 2”®“car 3”®“car 4,” where the transmis-
sion frequencies were dynamically chosen based on availability.

systems contained only the hardware necessary to perform the tasks that they were
designed for, which made these systems difficult to modify and upgrade. However,
with the advent of various digital processing and computing technologies, commu-
nication systems have evolved to the point where many of the encoders, modula-
tors, filters, and decoders used by these communication systems can be implemented
either partially or entirely using software and/or programmable logic [22]. Those
communications systems that implement all of their baseband functionality in soft-
ware are referred to as SDR [23].
While some SDR platforms employ general-purpose processors or even digital
signal processors (DSPs), many use a technology called field-programmable gate ar-
rays (FPGA). An FPGA consists of reconfigurable logic elements and a switch matrix
to route signals between them. These devices can be configured to support simple
logic operations, such as addition, or more complex systems, such as digital filters.
In addition, some FPGAs can support dynamic reconfiguration, where a system is
capable of swapping components as needed, without any reprogramming. Conse-
quently, the firmware on an FPGA can be updated remotely in order to install new
blocks and remove unused blocks. Although relatively difficult to use with respect
to the design and prototype of a communication system, one of the primary advan-
tages of an FPGA is that it is very computationally powerful, making it a suitable
choice for any SDR implementation involving complex operations. Commercially
available examples of FPGA-based SDR platforms include Lyrtech’s small form
factor (SFF) SDR development platform and the Wireless Open-Access Research
Platform (WARP) [24, 25]. Although the USRP family of SDR platforms available
from Ettus Research LLC possess an FPGA, these have been mostly used to perform
sampling and filtering operations on incoming and outgoing data streams.
248 Applications of Software-Defined Radio

One wireless application where FPGA-based SDR platforms are beginning

to emerge is in the area of satellite communications, where SDR systems are em-
ployed on satellites in order to enable advanced communications with terrestrial
ground stations, as well as with other satellites in orbit around the Earth [22,
26, 27]. In particular, one type of satellite platform that is beginning to witness
a substantial increase in the utilization of SDR technology is CubeSat systems. A
CubeSat is a small form-factor satellite conforming to the CubeSat specifications
published by California Polytechnic State University [28]. A 1U form factor Cube-
Sat possesses a length, width, and height of 10 cm on each side and has a total
weight that cannot exceed 1.33 kilograms. In addition to this 1U configuration, the
CubeSat specifications allow for 2U (10 × 10 × 20 cm) and 3U (10 × 10 × 30 cm)
configurations.
One the main design constraints for implementing a satellite system is the physi-
cal size of the platform itself. Given the significant costs involved in deploying a
satellite system in orbit, each module that forms part of the satellite system must use
the smallest amount of space possible. Consequently, off-the-shelf solutions for SDR
systems that can be employed on satellite systems simply do not exist, and there cur-
rently exists significant research and development efforts by government laboratories,
industry, and academia in order to create a family of SDR platforms capable of being
used on satellite systems. This design requirement is further compounded by the rela-
tively small form factor associated with CubeSat systems, which also include a suite
of sensors and other instruments necessary for the functionality of the satellite. An
example of an FPGA-based SDR platform for a CubeSat is shown in Figure 9.6.
As a result, there exists several desired characteristics for a CubeSat FPGA-
based SDR platform that is sought after by the community, namely [30]:

1. Reconfigurability. The system is easily reconfigured to support any number

of encoding, modulation, and other signal processing schemes.
2. Small size. The system fits the dimensions for a 1U CubeSat (10 × 10 × 10
cm) and meets all other CubeSat standard requirements [28].
3. Plug and play. The system uses the Space Plug-and-Play Avionics (SPA) pro-
tocol for plug and play capabilities.
4. Open source. The system uses open source hardware and software so that
others can adapt it to fit their project goals and so that they can create de-
rivative systems as needed.

One such CubeSat FPGA-based SDR platform that attempts to meet these
characteristics was presented in [30–32], with the hardware platform shown in
Figure 9.6 and the SDR firmware architecture shown in Figure 9.7. The actual
SDR communications hardware is based on the redesign of a USRP1 SDR platform
by Ettus Research LLC, which publishes their open source SDR hardware designs
and schematics online. At the center of the SDR architecture is the Micro­Blaze
softcore microprocessor, which acts as a traffic mediator, guiding data ­between
the USB2 port to the various other devices in the system. The SPI and I2C control-
lers provide the users with a means of sending configuration information to the
USRP and to the RF daughtercards. One of the RS232 controllers connects to an
serial port on the board, while the other connects to the AT90, which controls the
9.3  Satellite Communications 249

Figure 9.6  A photograph of the COSMIAC CubeSat SDR system. The bottom board is the FPGA
board while the one on top is the RF daughtercard [29].

ASIM for the device. The other components simply control the peripherals that
give them their names.
Several advantages of this implementation of a CubeSat FPGA-based SDR plat-
form include the following [30]:

• It employs well-tested components from Xilinx including the DDR2 memory

controller, the SPI and I2C controllers, and the MicroBlaze microprocessor.
• The Embedded Development Kit (EDK) software from Xilinx provides users
with a relatively simple way to design MicroBlaze-based systems for custom
hardware.
• Using a MicroBlaze processor to allows one to write much of the control
logic in the C programming language.
• This system should be compatible with GNU Radio with few modifications
since the MicroBlaze and other components can emulate the FX2 USB2 con-
troller with proper configuration.

As for disadvantages, this system presented in [30] employs components that only
work with the Xilinx toolset, so the system loses its platform independence. Fur-
thermore, the USB2 controller that is at the core of the USRP1 implementation is
currently priced at $14,000, which can be prohibitively expensive.
250 Applications of Software-Defined Radio

Figure 9.7  Block diagram of the proposed CubeSat SDR firmware. The MicroBlaze processor acts
as the central control for all of the various communications blocks [30].

9.4  Chapter Summary

In this chapter, we briefly explored three applications where SDR technology is em-
ployed in order to facilitate wireless data transmission in challenging environments.
Whether it is intelligently and adaptively seeking out transmission opportunities
across time, frequency, and space; enabling the transfer of information between ve-
hicles while driving along a stretch of highway; or facilitating the exchange of data
between a satellite in orbit and a ground station back on Earth, SDR is increasingly
becoming the core technology that is enabling these and many other applications
in today’s society.

References

[1] Wyglinski, A. M., M. Nekovee, and T. Hou, Cognitive Radio Communications and Net-
works: Principles and Practice, Academic Press, 2009, http://www.elsevierdirect.com/
ISBN/9780123747150/ Cognitive-Radi-Communications-and-Networks.
[2] Barker, B., A. Agah, and A. M. Wyglinski, “Mission-Oriented Communications Properties
for Software Defined Radio Configuration,” Cognitive Radio Networks, CRC Press, 2008.
[3]  Newman, T., A. M. Wyglinski, J. B. Evans, “Cognitive Radio Implementation for Efficient
Wireless Communication,” Encyclopedia of Wireless and Mobile Communications, CRC
Press, 2007.
[4] Newman, T., R. Rajbanshi, A. M. Wyglinski, J. B. Evans, and G. J. Minden, “Population
Adaptation for Genetic Algorithm-Based Cognitive Radios,” ACM/Springer Mobile Net-
works and Applications, 2008.
9.4  Chapter Summary 251

[5] Newman, T. R., B. A. Barker, A. M. Wyglinski, A. Agah, and J. B. Evans et al., “Cognitive

Engine Implementation for Wireless Multicarrier Transceivers,” Wiley Journal on Wireless
Communications and Mobile Computing, Vol. 7, No. 9, Nov. 2007, pp. 1129–1142.
[6]  Rieser, C. J., “Biologically Inspired Cognitive Radio Engine Model Utilizing Distributed
Genetic Algorithms for Secure and Robust Wireless Communications and Networking”
PhD Thesis, Virginia Polytechnic Institute and State University, Blacksburg, VA, 2005.
[7]  Troxel, G. D., E. Blossom, S. Boswell, M. Brodsky, A. Caro, I. Castineyra et al., “Adap-
tive Dynamic Radio Open-Source Intelligent Team (ADROIT): Cognitively-Controlled
Collaboration Among SDR Nodes,” Proceedings of the IEEE Communications Society
Conference on Sensors, Mesh and Ad Hoc Communications and Networks (SECON)—
Workshop on Networking Technologies for Software-Defined Radio Networks, Reston,
VA, Aug. 2006.
[8]  Intelligent Transportation Society of America, IntelliDrive Task Force, http://intellidrive
­taskforce.itsa. wikispaces.net/.
[9]  Willke, T. L., P. Tientrakool, and N. F. Maxemchuk, “A Survey of Inter-Vehicle Communica-
tion Protocols and Their Applications,” IEEE Communications Surveys and Tutorials, Vol.  11,
2009.
[10]  Khaled, Y., M. Tsukada, J. Santa, and T. Ernst, “On the Design of Efficient Vehicular Appli-
cations,” Proceedings of the IEEE Vehicular Technology Conference, Dresden, ­Germany,
Apr. 2009.
[11]  Papadimitratos, P., A. de La Fortelle, K. Evenssen, and R. Bringnolol, “Vehicular Commu-
nication Systems: Enabling Technologies, Applications, and Future Outlook on Intelligent
Transportation,” IEEE Communications Magazine, Vol. 47, 2009, pp. 84–95.
[12]  Chennikara-Varghese, J., W. Chen, T. Hikita, and R. Onishi, “Local Peer Groups and
Vehicle-to-Infrastructure Communications,” IEEE Global Communication Conference—
Workshop, Nov. 2007.
[13]  Chen, W., and S. Cai, “Ad Hoc Peer-to-Peer Network Architecture for Vehicle Safety Com-
munications,” IEEE Commun. Magazine, Apr. 2005.
[14]  Tsukamoto, K., S. Matsuoka, O. Altintas, M. Tsuru, and Y. Oie, “Distributed Channel Co-
ordination in Cognitive Wireless Vehicle-to-Vehicle Communications,” Proceedings of the
Wireless Access in Vehicular Environments (WAVE) Conference, Dearborn, MI, Dec. 2008.
[15]  Tsukamoto, K., Y. Omori, O. Altintas, M. Tsuru, and Y. Oie, “On Spatially-Aware Channel
Selection in Dynamic Spectrum Access Multi-Hop Inter-Vehicle Communications,” Proceed-
ings of the IEEE Vehicular Technology Conference—Fall, Anchorage, AK, Sep. 2009.
[16]  Lacatus, C., R. Vuyyuru, O. Altintas, D. Borota, and Ivan Seskar, “Evaluation of Energy-
Based Spectrum Sensing Algorithm for Vehicular Networks,” Proceedings of the SDR Fo-
rum Technical Conference, Washington DC, Dec. 2009.
[17]  Pagadarai, S., A. M. Wyglinski, and R. Vuyyuru, “Characterization of Vacant UHF TV
Channels for Vehicular Dynamic Spectrum Access,” Proc. IEEE Vehicul. Networking
Conf., Oct. 2009.
[18]  Chen, S., A. M. Wyglinski, R. Vuyyuru, and O. Altintas, “Feasibility Analysis of Vehicular
Dynamic Spectrum Access via Queueing Theory Model,” Proc. IEEE Vehicul. Networking
Conf., Dec. 2010.
[19]  Altintas, O., M. Nishibori, R. Vuyyuru, Y. Fujii, and K. Nishida et al., “Implementation
and Evaluation of Distributed Control and Data Channel Coordination Algorithms for
V2V Dynamic Spectrum Access,” Proceedings of the SDR Forum Technical Conference,
Washington DC, Nov. 2010.
[20]  Rzayev, T., Y. Shi, A. Vafeiadis, S. Pagadarai, and A. M. Wyglinski, “Implementation of a
Vehicular Networking Architecture Supporting Dynamic Spectrum Access,” Proceedings of
the IEEE Vehicular Networking Conference, Amsterdam, The Netherlands, Nov. 2011.
[21]  Pukelsheim, F., “The Three Sigma Rule,” The American Statistician, Vol. 48, No. 2, May
1994, pp. 88–91.
252 Applications of Software-Defined Radio

[22] Wireless Innovation Forum,“The Three Sigma Rule,” What is Software Defined Radio,
http://data.memberclicks.com/site/sdf/SoftwareDefinedRadio.pdf.
[23] Youngblood, G., “A Software-Defined Radio for the Masses, Part 1,” QEX, Jul/Aug:
13–21, 2002, http://www.arrl.org/files/file/Technology/tis/info/pdf/020708qex013.pdf.
[24] Mouser—Lyrtech SFF SDR Development Platform, April 2011, http://www.mouser.com/
Lyrtech/.
[25]  Wireless Open-Access Research Platform, April 2011, http://warp.rice.edu/index.php.
[26]  Vulcan Wireless SDR Products, April 2011, http://www.vulcanwireless.com/products/.
[27] Lynaugh, K., and M. Davis, Software Defined Radios for Small Satellites, Nov. 2010,
http://www.cosmiac. org/ReSpace2010/.
[28] California Polytechnic State University, CubeSat Design Specification Rev. 12, 2009, http://
www.cubesat.org/images/developers/cds_rev12.pdf.
[29] COSMIAC CubeSat SDR Photograph, April 2011, http://www.cosmiac.org/images/
CCRBwSDR.JPG.
[30] Olivieri, S., J. Aarestad, L. H. Pollard, A. M. Wyglinski, C. Kief et al.,“Modular FPGA-
Based Software Defined Radio for CubeSats,” Proceedings on the IEEE International Con-
ference on Communications, Ottawa, Canada, June 2012.
[31] Olivieri, S., A. M. Wyglinski, L. H. Pollard, J. Aarestad and C. Kief, “Responsive Satellite
Communications via FPGA-Based Software-Defined Radio for SPA-U Compatible Plat-
forms,” Presented at the ReSpace/MAPLD 2010 Conference in Albuquerque, NM, http://
www.cosmiac.org/ReSpace2010/. Nov. 2010.
[32] Olivieri, S. J., “Modular FPGA-Based Software Defined Radio for CubeSats,” Master’s
Thesis, Worcester Polytechnic Institute, Worcester, MA, May 2011.
 Appendix A

Getting Started with MATLAB and Simulink

You will be using MATLAB and Simulink for the experiments, as well as for the
open-ended design projects in this book. This appendix serves as a brief refresher of
MATLAB, since you should have used it before. However, if you don’t have exten-
sive experience with Simulink, then this appendix shows you how to get started with
the tool. Please note the MATLAB portion of this appendix is mainly based on the
MATLAB Documentation presented in [1] and the Simulink portion is based on the
Simulink Basics Tutorial presented in [2], but here we extract the most important and
fundamental concepts such that you can quickly get started by reading this appendix. For
more information about these two products, you are encouraged to refer to [1] and [2].

A.1  MATLAB INTRODUCTION

MATLAB is widely used in all areas of applied mathematics, in education and research
at universities, and in industry. MATLAB stands for MATrix LABoratory and the soft-
ware is built up around vectors and matrices. Consequently, this makes the software
particularly useful for solving problems in linear algebra, but also for solving algebraic
and differential equations as well as numerical integration. MATLAB possesses a col-
lection of graphical tools capable of producing advanced GUI and data plots in both 2D
and 3D. MATLAB also has several toolboxes useful for performing communications
signal processing, image processing, optimization, and other specialized operations.

A.2  EDIT AND RUN A PROGRAM IN MATLAB

When writing programs, you will need to do this in a separate window, called the edi-
tor. To open the editor, go to the “File” menu and choose either the “New...Script” (if
you want to create a new program) or “Open” (to open an existing document) option.
In the editor, you can now type in your code in much the same way that you would use a
text editor or a word processor. There are menus for editing the text that are also similar
to any word processor. While typing your code in the editor, no commands will be per-
formed. Consequently, in order to run a program, you will need to do the following:

1. Save your code as <filename>.m, where <filename> is anything you wish to

name the file.
2. Go to the command window. If necessary, change directories to the directory
containing your file. This can be accomplished using the cd command com-
mon to UNIX and DOS. Or, alternatively, you can select the button labeled
“...” in the Current Directory window, and change directories graphically.
You can view your current directory by typing the pwd command.
253
254 Getting Started with MATLAB and Simulink

3. In order to run the program, type the name of the file containing your pro-
gram at the prompt. When typing the filename in the command window, do
not include “.m”. By pressing enter, MATLAB will run your program and
perform all the commands given in your file.

In case your code has errors, MATLAB will complain when you try to run the
program in the command window. When this happens, try to interpret the error
message and make the necessary changes to your code in the editor. The error that
is reported by MATLAB is hyperlinked to the line in the file that caused the prob-
lem. Using your mouse, you can jump directly to the line in your program that has
caused the error. After you have made the changes, make sure you save your file
before trying to run the program again in the command window.

A.3  Useful Matlab Tools

This section introduces general techniques for finding errors, as well as using automatic
code analysis functions in order to detect possible areas for improvement within the
MATLAB code. In particular, the MATLAB debugger features located within the Edi-
tor, as well as equivalent Command Window debugging functions, will be covered.

Debugging is the process by which you isolate and fix problems with your code.
Debugging helps to correct two kinds of errors:

• Syntax errors: For example, misspelling a function name or omitting a

parenthesis.
• Run-time errors: These errors are usually algorithmic in nature. For example,
you might modify the wrong variable or code a calculation incorrectly. Run-
time errors are usually apparent when an M-file produces unexpected results.
Run-time errors are difficult to track down because the function’s local work-
space is lost when the error forces a return to the MATLAB base workspace.

A.3.1  Code Analysis and M-Lint Messages

MATLAB can check your code for problems and recommend modifications to maxi-
mize the performance and maintainability through messages, sometimes referred to as
M-Lint messages. The term lint is the name given to a similar tool used with other pro-
gramming languages such as C. In MATLAB, the M-Lint tool displays a message for
each line of an M-file it determines possesses the potential to be improved. For example,
a common M-Lint message is that a variable is defined but never used in the M-file.

You can check for coding problems using three different ways, all of which report
the same messages:

• Continuously check code in the Editor while you work: View M-Lint mes-
sages and modify your file based on the messages. The messages update au-
tomatically and continuously so you can see if your changes addressed the
A.3  Useful Matlab Tools 255

issues noted in the messages. Some messages offer extended information,

automatic code correction, or both.
• Run a report for an existing MATLAB code file: From a file in the Editor,
select Tools > Code Analyzer > Show Code Analyzer Report.
• Run a report for all files in a folder: In the Current Folder browser, click the
Actions button, then select Reports > Code Analyzer Report.
For each message, review the message and the associated code in order to make
changes to the code itself based on the message via the following process:
• Click the line number to open the M-file in the Editor/Debugger at that line.
• Review the M-Lint message in the report and change the code in the M-file,
based on the message.
• Note that in some cases, you should not make any changes based on the M-
Lint messages because the M-Lint messages do not apply to that specific situ-
ation. M-Lint does not provide perfect information about every situation.
• Save the M-file. Consider saving the file to a different name if you made
significant changes that might introduce errors. Then you can refer to the
original file as you resolve problems with the updated file.
• If you are not sure what a message means or what to change in the code as a
result, use the Help browser to look for related topics.

You can also get M-Lint messages using the mlint function. For more information
about this function, you can type help mlint in the Command Window. Read the
online documentation [3] for more information about this tool.

A.3.2  Debugger
The MATLAB Editor, graphical debugger, and MATLAB debugging functions are
useful for correcting run-time problems. They enable access to function workspaces
and examine or change the values they contain. You can also set and clear break-
points, which are indicators that temporarily halt execution in a file. While stopped
at a breakpoint, you can change the workspace contexts, view the function call
stack, and execute the lines in a file one by one.
There are two important techniques in debugging: one is the breakpoint while the
other is the step. Setting breakpoints to pause the execution of a function enables you to
examine values where you think the problem might be located. While debugging, you
can also step through an M-file, pausing at points where you want to examine values.
There are three basic types of breakpoints that you can set in the M-files, namely:

• A standard breakpoint, which stops at a specified line in an M-file.

• A conditional breakpoint, which stops at a specified line in an M-file only
under specified conditions.
• An error breakpoint that stops in any M-file when it produces the specified
type of warning, error, or NaN or infinite value.

You cannot set breakpoints while MATLAB is busy (e.g., running an M-file) unless
that M-file is paused at a breakpoint. While the program is paused, you can view
the value of any variable currently in the workspace, thus allowing you to examine
256 Getting Started with MATLAB and Simulink

values when you want to see whether a line of code has produced the expected re-
sult or not. If the result is as expected, continue running or step to the next line. If
the result is not as expected, then that line, or a previous line, contains an error.
While debugging, you can change the value of a variable in the current workspace
to see if the new value produces expected results. While the program is paused, assign
a new value to the variable in the Command Window, Workspace browser, or Array
Editor. Then continue running or stepping through the program. If the new value does
not produce the expected results, the program has a different or another problem.
Besides using the Editor, which is a graphical user interface, you can also debug
MATLAB files by using debugging functions from the Command Window, or you
can use both methods interchangeably. Read the online documentation [4] for more
information about this tool.

A.3.3  Profiler
Profiling is a way to measure the amount of time a program spends on performing
various functions. Using the MATLAB Profiler, you can identify which functions
in your code consume the most time. You can then determine why you are calling
them and look for ways to minimize their use. It is often helpful to decide whether
the number of times a particular function is called is reasonable. Since programs of-
ten have several layers, your code may not explicitly call the most time-consuming
functions. Rather, functions within your code might be calling other time-consuming­
functions that can be several layers down into the code. In this case, it is important
to determine which of your functions are responsible for such calls.

Profiling helps to uncover performance problems that you can solve by:

• Avoiding unnecessary computation.

• Changing your algorithm to avoid costly functions.
• Avoiding recomputation by storing results for future use.

When you reach the point where most of the time is spent on calls to a small num-
ber of built-in functions, you have probably optimized the code as much as you can
expect. You can use any of the following methods to open the Profiler:

• Select Desktop ® Profiler from the MATLAB desktop.

• Select Tools ® Open Profiler from the menu in the MATLAB Editor/Debugger.
• Select one or more statements in the Command History window, right-click
to view the context menu, and choose Profile Code.
• Enter the following function in the Command Window: profile viewer.

To profile an M-file or a line of code, follow these steps after you open the Profiler,
as shown in Figure A.1:

1. In the Run this code field in the Profiler, type the statement you want to run.
2. Click Start Profiling (or press Enter after typing the statement).
3. When profiling is complete, the Profile Summary report appears in the Pro-
filer window. Read the online documentation [5] for more information about
this tool.
A.5  Getting Started In Simulink 257

Figure A.1  The profiler window.

A.4 SIMULINK INTRODUCTION

Simulink (Simulation and Link) is an extension of MATLAB created by MathWorks

Inc. It works with MATLAB to offer modeling, simulation, and analysis of dynami-
cal systems within a graphical user interface (GUI) environment. The construction
of a model is simplified with click-and-drag mouse operations. Simulink includes
a comprehensive block library of toolboxes for both linear and nonlinear analy-
ses. Models are hierarchical, which allow using both top-down and bottom-up
approaches. As Simulink is an integral part of MATLAB, it is easy to switch back
and forth during the analysis process and thus, the user may take full advantage of
features offered in both environments. This section presents the basic features of
Simulink and is focused on Communications System Toolbox.

A.5  Getting Started In Simulink

A.5.1 Start a Simulink Session

To start a Simulink session, you need to bring up the MATLAB program first. Then,
from MATLAB command window, enter:
>> simulink
Alternatively, you may click on the Simulink icon located on the toolbar, as
shown in Figure A.2.
The library browser window of Simulink, as shown in Figure A.3, will pop up
presenting the libraries for model construction. To see the content of the toolbox,
click on the “+” sign at the beginning of each toolbox. For example, Figure A.3
shows the content of Communications System Toolbox.

A.5.2 Start a Simulink Model

To start a model, click on the new file icon as shown in Figure A.3. Alternatively,
you may use the keyboard shortcut CTRL+N. A new window will appear on the
screen. You will be constructing and simulating your model in this window. A
screenshot of a typical working model window is shown in Figure A.4.
258 Getting Started with MATLAB and Simulink

Figure A.2  Start a Simulink session by clicking on the Simulink icon, which is circled.

A.5.3 Simulink Model Settings

In this book, most of the simulations and implementations are designed for com-
munication systems, so a proper Simulink setting for this purpose is very helpful.
commstartup changes the default Simulink model settings to values more ap-
propriate for the simulation of communication systems. The changes apply to new
models that you create later in the MATLAB session, but not to previously created
models. To install the communications-related model settings each time you start
MATLAB, invoke commstartup from your startup.m file. To be more specific,
the settings in commstartup cause models to:

• Use the variable-step discrete solver in single-tasking mode.

• Use starting and ending times of 0 and Inf, respectively.

Figure A.3  Simulink library browser. To create a new model, click on the new file icon, which is circled.
A.6  Build A Simulink Model 259

Figure A.4  A screenshot of a typical Simulink model window.

• Avoid producing a warning or error message for inherited sample times in

source blocks.
• Set the Simulink Boolean logic signals parameter to Off.
• Avoid saving output or time information to the workspace.
• Produce an error upon detecting an algebraic loop.
• Inline parameters if you use the Model Reference feature of Simulink.

If your communications model does not work well with these default settings, you
can change each of the individual settings as the model requires.

A.6  Build A Simulink Model

To demonstrate how a system is represented using Simulink, we will build the block
diagram for a simple model consisting of a random number input and two scope
displays, which is shown in Figure A.5.
This model consists of four blocks: Random Number, Abs, and two Scopes.
The Random Number is a source block from which a random signal is generated.
This signal is transferred through a line in the direction indicated by the arrow to

Figure A.5  A simple Simulink model consisting of four blocks.

260 Getting Started with MATLAB and Simulink

the Abs Math Block. The Abs block modifies its input signal (calculates its absolute
value) and outputs a new signal through a line to the Scope block. The Scope is a
sink block used to display a signal like an oscilloscope.

Building a Simulink model is usually accomplished through the following steps:

1. Get the necessary blocks from the Library Browser and place them in the
model window.
2. Modify the parameters of the blocks according to the system that we are
modeling.
3. Connect all the blocks with lines to complete the Simulink model.

Each of these steps will be explained in detail using our example system. Once a
system is built, simulations can be run to analyze its behavior.

A.6.1  Obtain the Blocks

Each of the blocks we will use in our example model will be gathered from the
Simulink Library Browser. To place the Random Number block into the model
window, follow these steps:

1. Click on the “+” in front of “Simulink”, since “Sources” is a subfolder be-

neath the “Simulink” folder.
2. Scroll down until you see the “Sources” folder. Click on “Sources”, the right
window will display the various source blocks available for us to use and
Random Number is one of them, as shown in Figure A.6.
3. To insert the Random Number block into your model window, click on it in
the Library Browser and drag the block into your window.

The same method can be used to place the Abs and Scope blocks in the model
window. The Abs block can be found in the “Math Operations” subfolder and the
Scope block is located in the “Sink” subfolder. Arrange the four blocks in the model
window by selecting and dragging an individual block to a new location so that
they look similar to Figure A.7.

A.6.2 Set the Parameters

Simulink allows us to modify the default settings of the blocks in our model so that
they can accurately reflect the characteristics of the system we are analyzing. For
example, we can modify the Random Number block by double-clicking on it. Do-
ing so will cause the following window to appear, as shown in Figure A.8. Note that
Simulink gives a brief explanation of the block’s function in the top portion of this
window. In this case, it shows “Output a normally (Gaussian) distributed random
signal. Output is repeatable for a given seed.”
This window allows us to adjust the mean, variance, and seed of the random
input. The sample time value indicates the time interval between successive readings
of the signal.
A.6  Build A Simulink Model 261

Figure A.6  Obtain Random Number block from Simulink Library Browser.

Let us assume that our system’s random input has the following Gaussian dis-
tribution: Mean = 0, Variance = 1, Seed = 0. Enter these values into the appropriate
fields and leave the “Sample time” set to 0.1. Then click “OK” to accept them and
exit the window.
The Abs block simply returns the absolute value for the input and the Scope
block plots its input signal as a function of time, so there are no system parameters

Figure A.7  Obtain all the necessary blocks for the model and arrange them in the model window.
262 Getting Started with MATLAB and Simulink

Figure A.8  By double-clicking on the Random Number block, we can open the block and set the pa-
rameters. The “Variance” parameter is highlighted because its value will be changed in Section A.7.

that we can change for these two blocks. We will look at the Scope block in more
detail after we have run our simulation.

A.6.3  Connect the Blocks

In order to accurately reflect the system we are modeling, the Simulink blocks must
be properly connected. In our example system, the signal output by the Random
Number block is transmitted to the Abs block. The Abs block outputs its absolute
value to the Scope block, which graphs the signal as a function of time. Thus, we
need to draw lines from the output of the Random Number block to the input of the
Abs block, and from the output of the Abs block to the input of the Scope block.
Lines are drawn by dragging the mouse from where a signal starts (output terminal
of a block) to where it ends (input terminal of another block). When drawing lines, it
is important to make sure that the signal reaches each of its intended terminals. Simu-
link will turn the mouse pointer into a crosshair when it is close enough to an output
terminal to begin drawing a line, and the pointer will change into a double crosshair
when it is close enough to snap to an input terminal. A signal is properly connected if
its arrowhead is filled in, as shown in Figure A.9(a). If the arrowhead is open, as shown
in Figure A.9(b), it means that the signal is not connected to both blocks. To fix an open
signal, you can treat the open arrowhead as an output terminal and continue drawing
the line to an input terminal in the same manner as explained before.
When drawing lines, you do not need to worry about the path you follow. The
lines will route themselves automatically. Once blocks are connected, they can be
repositioned for a neater appearance. This is done by clicking on and dragging each
block to its desired location (signals will stay properly connected and will reroute
A.6  Build A Simulink Model 263

(a)

(b)
Figure A.9  Two blocks can be connected by drawing a line between them. (a) Two blocks are
properly connected, where an arrowhead is filled in. (b) Two blocks are not connected, where the
arrowhead is open.

themselves). After drawing in the lines and repositioning the blocks, the example
system model should look like Figure A.5.
In some models, it will be necessary to branch a signal so that it is transmitted
to two or more different input terminals. For example, in this Simulink model, the
output of the Random Number block is transmitted to two different input termi-
nals, one is Abs block, the other is Scope 1 block. This can be done by first placing
the mouse cursor at the location where the signal is to branch. Then, using either
the CTRL key in conjunction with the left mouse button or just the right mouse
button, drag the new line to its intended destination. The routing of lines and the
location of branches can be changed by dragging them to their desired new posi-
tion. To delete an incorrectly drawn line, simply click on it to select it, and hit the
DELETE key.

Figure A.10  Start running a simulation by clicking on the Start option of the Simulation menu,
which is circled.
264 Getting Started with MATLAB and Simulink

A.7  Run Simulink Simulations

Now that our model has been constructed, we are ready to simulate the system. To do
this, go to the “Simulation” menu and click on “Start”, as shown in Figure A.10, or
just click on the “Start/Pause Simulation” button in the model window ­toolbar (looks
like the “Play” button on a VCR) if you are using Windows operating system. Since
our example is a relatively simple model, its simulation runs almost instantaneously.
With more complicated systems, however, you will be able to see the progress of the
simulation by observing its running time in the lower box of the model window.
Double click the Scope and the Scope 1 block to view the random signal and its abso-
lute value for the simulation as a function of time. Once the Scope window appears, click
the “Autoscale” button in its toolbar (looks like a pair of binoculars) to scale the graph

(a)

(b)
Figure A.11  The scope displays of the Simulink model. (a) The random signal displayed by Scope 1.
(b) The absolute value of the random signal displayed by Scope.
A.7  Run Simulink Simulations 265

to better fit the window. Having done this, you should be able to see the results shown in
Figure A.11. It is very obvious that corresponding to each point in Figure A.11(a), Figure
A.11(b) returns its absolute value, which satisfies the theory of this model.
As mentioned in Section A.6.2, we can easily change the parameters of the
block by double clicking the block. How would this change affect the output ob-
served by the scopes? Let us change the “Variance” of the Random Number block
to be 2 and keep the other parameters as they were. Then, re-run the simulation and
view the scopes. Please note the scope graphs will not change unless the simulation
is re-run, even though the “Variance” has been modified. The new scope graphs
should now look like Figure A.12.
Comparing Figure A.12 with Figure A.11, we notice that the only difference
between this output and the one from our original system is the amplitude of the

(a)

(b)
Figure A.12  The scope displays of the Simulink model with variance = 2. (a) The random signal
displayed by Scope 1. (b) The absolute value of the random signal displayed by Scope.
266 Getting Started with MATLAB and Simulink

random signal. The shapes are identical due to the same seed value. However,
since the variance is doubled, the new amplitude is about 2 times as the original
amplitude.

References

[1]  The MathWorks, MATLAB Documentation, http://www.mathworks.com/help/techdoc/.

[2] University of Michigan, Simulink Basics Tutorial, http://www.engin.umich.edu/group/ctm/
working/mac/ simulink_basics/.
[3]  The MathWorks, Avoid Mistakes While Editing Code, http://www.mathworks.com/help/
techdoc/matlab_ env/brqxeeu-151.html.
[4]  The MathWorks, Debugging Process and Features, http://www.mathworks.com/help/techdoc/
matlab_env/ brqxeeu-175.html.
[5]  The MathWorks, Profiling for Improving Performance, http://www.mathworks.com/help/
techdoc/matlab_ env/f9-17018.html.
 Appendix B

Universal Hardware Driver (UHD)

The UHD is the “Universal Software Radio Peripheral” hardware driver. It works
on all major platforms (Linux, Windows, and Mac), and can be built with GCC,
Clang, and MSVC compilers [1].
The goal of the UHD is to provide a host driver and API for current and future
Ettus Research products. Users will be able to use the UHD driver standalone or
with 3rd party applications, such as Simulink. However, there are a few things you
need to do before you can actually use your ethernet-based USRP hardware with
the MATLAB R2011b1 software via UHD interface, which will be introduced in
the following sections. In this appendix, we assume that the employed operating
system is Ubuntu [2].

B.1  SETTING UP YOUR HARDWARE

Connect the ethernet-based USRP board to your computer using a gigabit Ethernet
cable. For optimal results, attach the USRP directly to the computer, on a private
network (i.e., one without any other computers, routers, or switches). Connecting
the USRP board directly to your computer requires the following:

• A version of USRP firmware which supports UHD. It will be introduced in

Section B.2.
• A gigabit Ethernet card, which supports flow control and has a static IP ad-
dress. For more information, follow these guidelines below.

B.2  INSTALLING UHD-BASED USRP I/O BLOCKS

With the latest UHD interface support and Communications System Toolbox, you
can use ethernet-based USRP with MATLAB and Simulink. Starting MATLAB
2012b, USRP support package can be downloaded and installed using “target-
installer” command in MATLAB. Before MATLAB 2012b, go to [3], download the
support package. You need to fill out a form and submit it before you can save the
package in your local computer. After you get the package, install it following
the steps 1–4 in README.txt.

1. MATLAB and Simulink supports USRP radios since version R2011b.

267
268 Universal Hardware Driver (UHD)

B.3  BURNING THE FIRMWARE TO AN SD CARD

MATLAB R2011b support for ethernet-based USRP devices has been tested using
UHD binary download release 003.002.003. For versions other than R2012b, the
user should check the MathWorks documentation for the support package to find
out which version of UHD is supported. You can download the necessary firmware
images from [4], and burn them to an SD card following the instructions below.
Please note you need to be a root user in order to perform the following steps on
Ubuntu operating system.

1. Insert the SD memory card into the USB burner.

2. From the terminal, change to the directory containing the USRP SD Card
Burner GUI using ‘cd’ command.
3. Launch the USRP SD Card Burner GUI by typing ‘sudo python usrp2_card_
burner_gui.py’ at the terminal if you are using this card with a USRP2 radio.
usrp2_card_burner_gui.py can be downloaded from [5]. Note: Do not save
directly from the link on this page. Go into the link, and choose “Down-
load” to save.
4. Select the drive for the memory card device. For example, if the memory card
device is on the sdd drive, select /dev/sdd.
5. Click the “Firmware Image” button and navigate to the firmware image file.
6. Click the “FPGA Image” button and navigate to the image file.
7. Click the “Burn SD Card” button.
8. If a prompt indicating that burning the firmware appears, confirm that it
states the verification information passed and click “OK.”
9. Insert the SD memory card into the USRP radio and power on.

B.4  CONFIGURE THE ETHERNET CARD

In Ubuntu operating system, in order to configure the Ethernet card for your USRP
hardware, you need to perform the following tasks:

1. Select System ® Preferences ® Network Connections

2. Select the network card to which the USRP will be attached. If you are not
sure which network card the USRP is attached to, you can use ifconfig
command to find it out.
3. Click the “Edit” button
4. Go to the “IPv4 Settings” tab
5. Select “Manual” for “Method”, and click the “Add” button
6. Use the following addresses:
• Address: 192.168.10.1
• Netmask: 255.255.255.0
• Leave Gateway blank

Since USRP radio has an IP address of 192.168.10.2, you need to select a number
192.168.10.X where X is anything but 2.
B.7  Problems With Unicode 269

B.5  MODIFY THE IPTABLES

Iptables is a firewall, installed by default on all official Ubuntu distributions (Ubuntu,

Kubuntu, Xubuntu). When you install Ubuntu, iptables is there. However, in order
to use the SDRu Transmitter and Receiver blocks, you need to modify the Iptables
to allow all traffic. You can access the Iptables on your computer by typing the fol-
lowing commands on the terminal:

• sudo bash
• cd /root/scripts/
• emacs iptables.sh

Now, you should have the Iptables open in front of you. You only need to modify
two lines concerning the default firewall policies:

• IPTABLES_COM -P INPUT ACCEPT

• IPTABLES_COM -P FORWARD ACCEPT

When this modification is done, you should save the Iptables and execute it by typ-
ing the following command on the terminal:

• ./iptables.sh

B.6  EACH TIME YOU USE

In future MATLAB sessions, you must run setupsdru to use Communications

with USRP. setupsdru is under the commusrp directory.
You can automate this step by including the commands prompted by the “Add
SDRu” shortcut to your startup.m file. If you prefer to run it manually you can use the
shortcut labelled “Add SDRu”, that has been provided on the MATLAB shortcut bar.
The installer will run setupsdru when you first install Communications with USRP.

B.7  Problems With Unicode

A user may get the following error when he runs findsdru (after running setupsdru):
*??? The conversion from a local code page string to unicode changes the num-
ber of characters. This is not supported. Error in Þ usrp_uhd_mapi
This error occurs due to incompatible localization between the USRP hardware
and your host computer. It can be fixed on Linux by going to a command shell and
executing:

• export LANG=C, or
• export LANG=en_US.ISO8859-1

These commands reset the locale to be compatible with the usrp_uhd_mapi code.
270 Universal Hardware Driver (UHD)

References

[1]  http://ettus-apps.sourcerepo.com/redmine/ettus/projects/uhd/wiki/
[2]  http://www.ubuntu.com/
[3]  http://www.mathworks.com/programs/usrp/
[4]  http://files.ettus.com/binaries/uhd_stable/releases/uhd_003.002.003-release/images-only/
[5]  https://ettus-apps.sourcerepo.com/redmine/ettus/projects/uhd/repository/revisions/master/
show/host/utils
 Appendix C

Data Flow on USRP

This appendix introduces how data flows on the USRP board. After reading this ap-
pendix, you should have a better insight on the USRP board. To make things clearer,
the data flow is introduced in two parts, receive (RX) path and transmit (TX) path,
as shown in Figure C.1.

C.1  Receive Path

At the RX path, we have four ADCs, and four DDCs. Generally speaking, on each
daughterboard, the two analog input signals are sent to two separate ADCs. The
digitized samples are then sent to the FPGA for processing. Upon entering the
FPGA, the digitized signals are routed by a multiplexer (MUX) to the appropriate
digital down-converter (DDC). There are four DDCs. Each of them has two inputs
(I and Q). The MUX is like a router or a circuit switcher. It determines which ADC
(or constant zero) is connected to each DDC input. We can control the MUX using
usrp.set_mux () method in Python. This allows for having multiple channels
selected out of the same ADC sample stream.
The standard FPGA configuration includes digital down-converters imple-
mented with four stages cascaded integrator-comb (CIC) filters. CIC filters are
very-high-performance filters using only adds and delays. For spectral shaping and
out-of-band signals rejection, there are also 31 tap halfband filters cascaded with
the CIC filters to form complete DDC stage. The standard FPGA configuration
implements two complete digital down-converters. Also there is an image with four
DDCs but without halfband filters. This allows one, two, or four separate RX
channels. It’s possible to specify the firmware and fpga files that are to be used by
loading the corresponding rbf files. By default, an rbf is loaded from /usr/local/
share/usrp/rev{2,4}. The one used unless you specify the fpga_filename con-
structor argument when instantiating a usrp source or sink is std_2rxhb_2tx.rbf.
Table C.1 lists the current three different rbf files and their usage.
Now let’s take a closer look at the digital down-converter. What does it do?
First, it down-converts the signal from the IF band to the base band. Second, it deci-
mates the signal so that the data rate can be adapted by the USB 2.0 and is reason-
able for the computers’ computing capability. Figure C.2 shows the block diagram
of the DDC. The complex input signal (IF) is multiplied by the constant frequency
(usually also IF) exponential signal. The resulting signal is also complex and entered
at 0. Then we decimate the signal with a factor M.
Note that when there are multiple channels (up to four), the channels are inter-
leaved. For example, with four channels, the sequence sent over the USB would be “I0
Q0 I1 Q1 I2 Q2 I3 Q3 I0 Q0 I1 Q1,” and so on. The USRP can operate in full-duplex

271
272 Data Flow on USRP

Figure C.1  USRP receive and transmit paths [1].

mode. When in this mode, the transmit and receive sides are completely independent of
one another. The only consideration is that the combined data rate over the bus must be
32 MBps or less. Finally, the I/Q complex signal enters the computer via the USB.
To illustrate the effect of half-band filter in Rx path, two different situations
are discussed. Given the bandwidth of each Rx channel is f, if we set the center
frequency at 0Hz, then, at frequency domain, it is [–f/2, f/2].

C.1.1  Situation 1
Suppose there are four users (u1, u2, u3 and u4), and each of them uses one of the
Rx channels, as shown in Figure C.3. In frequency domain, the bandwidth they
are occupying are all from -f/2 to f/2. In this situation, the FPGA rbf file should be
std_4rx_01x.rbf, which means the FPGA firmware contains four Rx paths
without halfbands and zero tx paths.

Table C.1  Three rbf Files and Their Usage

rbf File Name FPGA Description
multi_2rxhb_2tx.rbf There are two or more USRPs
std_2rxhb_2tx.rbf Contains two Rx paths with halfband filters and two tx paths
std_4rx_0tx.rbf Contains four Rx paths without halfbands and zero tx paths
C.1  Receive Path 273

Figure C.2  The structure of digital down converter on receiver path, where the input RF signal I’(t)
and Q’(t) is first down-converted to the IF band by a carrier frequency of wc , and then decimated
with a factor M.

In analog-to-digital converter (ADC), continuous signals are converted to dis-

crete digital numbers. A common and useful way to do this is through the use of a
periodic impulse train signal [2], p(t), multiplied by the continuous time signal x(t).
This is known as impulse train sampling, as introduced in Section 2.3.2.
Then, the low-pass filter takes out the signal around center frequency and sup-
presses the rest replicas. If the decimation value of DDC is M/2, after entering the

u1
u1
f
–f/2 f/2 f
f –Mf/4 Mf/4

u2

u2 f
f –Mf/4 Mf/4

f
–f/2 f/2

u3
u3

f f
f –Mf/4 Mf/4
–f/2 f/2

u4

f
f –Mf/4 Mf/4

u4

f
–f/2 f/2

Figure C.3  4 Rx channels (labeled 1, 2, 3, and 4), starting from the inputs A1, A2, B1, and B2, pass-
ing through one of the ADCs and one of the DDCs, then reaching the USB in the end.
274 Data Flow on USRP

u1
u1

f f
f
–f/2 f/2 –Mf/4 Mf/4 –Mf/2 Mf/2
f

f
u2

u2 f
f –Mf/2 Mf/2
–Mf/4 Mf/4
f f
–f/2 f/2

Figure C.4  Two Rx channels (labeled 1 and 2), starting from the inputs A1 and B1, passing through
one of the ADCs and one of the DDCs with half-band filter, then reaching the USB in the end.

DDC, the bandwidth of each signal will be expanded to [-M f/4, M f/4]. The sum
of all the bandwidth is 2Mf.

C.1.2  Situation 2
Suppose there are two users (u1 and u2), and each of them uses one of the Rx
channels, as shown in Figure C.4. In frequency domain, the bandwidth they are
occupying are both from –f/2 to f/2. In this situation, the FPGA rbf file should be
std_2rxhb_2tx.rbf, which means FPGA firmware contains two Rx paths with
half-bands and two tx paths. If the decimation value of DDC is M/2, after entering
the DDC, the bandwidth of each signal will be expanded to [–M f/4, M f/4]. Then,
after entering the half-band filter, the bandwidth of each signal will be further ex-
panded to [–M f/2, M f/2]. The sum of all the bandwidth is still 2Mf.

C.2 Transmit Path

At the TX path, the story is pretty much the same, except that it happens in reverse.
We need to send a baseband I/Q complex signal to the USRP board. The digital up-
C.2  Transmit Path 275

Figure C.5  The structure of digital up-converter on transmitter path, where a baseband I/Q com-
plex signal Iin (t) and Qin (t) is upconverted to the IF band by a carrier frequency of wc . I’(t) and Q’(t)
will be sent to the DAC.

converter (DUC) will interpolate the signal, up-convert it to the IF band, and finally
send it through the DAC, as shown in Figure C.5.
More specifically, first, interleaved data sent from the host computer is pushed
into the transmit FIFO on the USRP. Each complex sample is 32 bits long (16 bits
for in-phase, and 16 bits for quadrature). This data is de-interleaved and sent to the
input of an interpolation stage. Assuming, the user has specified an interpolation
factor of L, this interpolation stage will interpolate the input data by a factor of L/4
using CIC filters.
The output of the interpolation stage is sent to the demultiplexer (DEMUX).
The DEMUX is less complicated than the receiver MUX. Here, the in-phase and
quadrature output of each CIC interpolator is sent to in-phase and quadrature in-
puts of one of the DAC chips on the motherboard. The user specifies which DAC
chip receives the output of each CIC interpolator [3].
Inside of the AD9862 CODEC chip, the complex-valued signal is interpolated
by a factor of four using half-band filter interpolators. This completes the re-
quested factor of L interpolation. After the half-band interpolators, the complex-
valued signal is sent to a DUC. Note, at this point the signal is not necessarily
modulated to a carrier frequency. The daughterboard might further up-convert
the signal.
The in-phase and quadrature output of the DUC are sent as 14-bit samples to
individual digital-to-analog converters, which convert them to analog signals at a
rate of 128 mega-samples per second. These analog signals are then sent from the
AD9862 to either daughterboard interface J667 or J669, which represent slots TxA
and TxB, respectively.
Different from the DDCs, the DUCs on the transmit side are actually contained
in the AD9862 CODEC chips, not in the FPGA. The only transmit signal processing
blocks in the FPGA are the interpolators. The interpolator outputs can be routed to
any of the 4 CODEC inputs [1].
As we have mentioned before, the multiple RX channels (1, 2, or 4) must all
be the same data rate (i.e., same decimation ratio). The same applies to the 1, 2, or
4 TX channels, which each must be at the same data rate (which may be different
from the RX rate) [4].
276 Data Flow on USRP

References

[1]  Patton, Lee K., A Gnn Radio Based Software-Defined Radar, Master’s Thesis, 2007.
[2]  Proakis, John G., and Dimitris G. Manolakis, Digital Signal Processing Principles, Algo-
rithms, and Applications, 3rd ed., Upper Saddle River, NJ, 1996.
[3]  Shen, Dawei, Gnn Radio Tutorials, http://www.nd.edu/˜dshen/GNU/, [Online]. 2005.
[4]  Abbas Hamza, Firas, The USRP under 1.5X Magnifying Lens!, gnuradio.org/redmine/
attachments/129/USRP_ Documentation.pdf, 2008.
 Appendix D

Quick Reference Sheet

D.1  LINUX

D.1.1  Helpful Commands

sudo - Run commands with administrative privileges.
Note: If you use “sudo” to run a single command as root, you might see error
messages. To avoid these error messages, simply use “sudo -i” or “sudo su” to log
in to run multiple commands as root. Or, use “sudo -H” to force sudo to use root’s
home directory instead.

ifconfig - Configure network interfaces.

• Use the “-a” option to list all interfaces.

• Use “up”/“down” to enable/disable interfaces.

D.1.2  Modify the Iptables

Iptables is a firewall, installed by default on all official Ubuntu distributions (Ubuntu,
Kubuntu, Xubuntu). When you install Ubuntu, iptables is there. However, in order
to use the SDRu Transmitter and Receiver Blocks, you need to modify the Iptables
to allow all traffic.
You can access the Iptables on your computer by typing the following com-
mands on the terminal:

• sudo bash
• cd /root/scripts/
• emacs iptables.sh

Now, you should have the Iptables open in front of you. You only need to
modify two lines concerning the default firewall policies:

• IPTABLES_COM -P INPUT ACCEPT

• IPTABLES-COM -P FORWARD ACCEPT

When this modification is done, you should save the Iptables and execute it by
typing the following command on the terminal:

• ./iptables.sh

277
278 Quick Reference Sheet

D.2  MATLAB

D.2.1  How to Start MATLAB

• Windows: Choose the submenu “Programs” from the “Start” menu. From the
“Programs” menu, open the “MATLAB” submenu. From the “MATLAB”­
submenu, choose “MATLAB.”
• Linux: At the prompt, type matlab.
• Mac: Double-click on the icon for MATLAB.

You can quit MATLAB by typing exit in the command window.

D.2.2  The MATLAB Environment

The MATLAB environment consists of menus, buttons, and a writing area similar
to an ordinary word processor.

D.2.2.1  Command Window

The writing area that you will see when you start MATLAB, is called the command
window. In this window, you give the commands to MATLAB. Once you have
typed the command you wish MATLAB to perform, press <enter>. If you want to
interrupt a command that MATLAB is running, type <ctrl> + <c>. The command
window is also useful if you just want to use MATLAB as a scientific calculator or
as a graphing tool.

D.2.2.2  Command History Window

The commands you type in the command window are stored by MATLAB and can
be viewed in the command history window. To repeat a command you have already
used, you can simply double-click on the command in the history window, or use
the <up arrow> at the command prompt to iterate through the commands you have
used until you reach the command you desire to repeat.

D.2.3  Obtaining Help

To obtain help on any of the MATLAB commands, you simply need to type help
<command> at the command prompt. For example, to obtain help on the “setups-
dru” function, we type at the command prompt: help setupsdru.
You may also get help about commands using the “Help Desk,” which can be
accessed by selecting the MATLAB Help option under the Help menu.

D.2.4  Variables in MATLAB

You can define your own variables in MATLAB. Suppose you need to use the value
of 2cos(3) repeatedly. Instead of typing 2*cos(3) over and over again, you can de-
note this variable as x by typing the following: x=2*cos(3).
D.3  USRP2 Hardware 279

If you do not want to have the result of a calculation printed out to the com-
mand window, you can put a semi-colon at the end of the command; MATLAB still
performs the command in the “background.”

• To know what variables have been declared, type whos. Alternatively, you
can view the values by opening the workspace window. This is done by se-
lecting the Workspace option from the View menu.
• To erase all variables from the MATLAB memory, type clear.
• To erase a specific variable, say x, type clear x.
• To clear two specific variables, say x and y, type clear x y, that is separate
the different variables with a space. Variables can also be cleared by selecting
them in the Workspace window and selecting the delete option.

D.2.5  Vectors and Matrices in MATLAB

D.2.5.1  Vectors
Create a vector by putting the elements within [] brackets. For example, x=[1 2
3 4 5 6 7 8 9]. We can also create this vector by typing x=1:9.
The vector (1 1.1 1.2 1.3 1.4 1.5) can be created by typing x=[1 1.1 1.2
1.3 1.4 1.5] or by typing x=1:0.1:1.5. The default step size is 1, as in the
previous example. So if the actual step size is not 1, you need to specify it in your
definition, as in this example.

D.2.5.2  Matrices
Matrices can be created according to the following example. The matrix A is cre-
ated by typing A=[1 2; 3 4; 5 6], (i.e., rows are separated with semi-colons).

D.2.5.3  Additional Information

• In MATLAB, the first index of a vector or matrix starts at 1, not 0 as is com-

mon with other programming languages.
• If the matrices (or vectors, which are special cases of matrices) are of the
same dimensions, then matrix addition, matrix subtraction, and scalar multi­
plication works just like we are used to.
• If you want to apply an operation such as squaring each element in a matrix,
you have to use a dot (.) before the operation you wish to apply. The dot al-
lows you to do operations elementwise. All built-in functions such as sin,
cos, exp, and so on automatically act elementwise on a matrix.

D.3  USRP2 Hardware

D.3.1  XCVR2450 Daughtercard

• The XCVR2450 daughtercard supports two frequency bands: 2.4 to 2.5 GHz
and 4.9 to 5.9 GHz.
280 Quick Reference Sheet

• There are two cables inside the USRP2 that connect the RF1 and RF2 SMA
connectors on the front of the USRP2 case to connectors J2 and J1 on the
daughtercard.
• The XCVR2450 always uses J1 for transmit and J2 for receive. This behavior
appears to hard-coded into the FPGA firmware and is not software-controllable
at present time.

D.3.2  Sampling

D.3.2.1  Decimation
The sampling frequency of the USRP2 is:

100
fs = MHz (D.1)
D
where D is the decimation, whose value can be any multiple of 4 from 4 through
512. (Nonmultiples of 4 yield different types of cascaded integrator-comb (CIC)
filters, but this probably isn’t what you want.)
The decimation parameter affects the bandwidth of the received signal, since it
affects the samping rate. Decimation essentially involves lowpass-filtering and then
downsamping by the decimation factor. Recall that downsampling by a factor of X
is equivalent to sampling the original signal at a factor of fs/X instead of fs. There-
fore, larger decimation factors result in smaller sampling rates and thus smaller
bandwidths. (After all, the bandwidth depends on the Nyquist frequency, or fs/2.)
The frequency resolution of the FFT is determined by the sampling frequency
(fs) and the number of points (N) in the FFT:

fs
frequency resolution = (D.2)
N

For example, an FFT with 1024 points and a sampling frequency of 128 kHz
has a resolution of
128 kHz
frequency resolution = = 125 Hz (D.3)
1024 points

In other words, each bin of the FFT is 125 Hz wide. (Spectrum analyzers com-
monly refer to this value as resolution bandwidth.)
Note there is the tradeoff between resolution in time and resolution in frequency.
FFTs with more points have finer-grain frequency resolution, but require more time to
collect the samples (since there are more points). FFTs with fewer points have coarser
resolution in frequency. Also note that larger FFTs are more computationally expensive,
which is important to keep in mind because you will be running the FFTs in real time.
Larger decimation values result in a smaller sampling rate; thus, if you keep
your FFT at 1024 points, larger decimation values will yield finer-grain fre-
quency resolution because (BW)/(N points) will be a smaller resolution band-
width value.
D.3  USRP2 Hardware 281

The punchline: You should use a high decimation value (e.g., 400 or 512) and
a sufficient number of FFT points (e.g., 1024) when taking fine measurements, such
as your oscillator frequency offset error measurements. This will afford the best
possible measurement precision.
Finally, here are a few additional useful links:

• http://en.wikipedia.org/wiki/Decimation_(signal_
processing)
• http://zone.ni.com/devzone/cda/tut/p/id/3983

D.3.2.2  Interpolation
The interpolation parameter is used for transmitting in DAC whereas the decima-
tion parameter is used for receiving in ADC. Interpolation has the same range of
values (any multiple of four from 4 to 512) as decimation.
According to [1]: “The FPGA talks to the DAC at 100 MS/s just like it talks
to the ADC at 100 MS/s. The interpolation from 100 MS/s to 400 MS/s happens
inside the DAC chip itself. Unless you are doing something fancy, you can think of
the DAC as operating at 100 MS/s.”

D.3.3  Clocking
The clock rate of the Spartan-3 FPGA in the USRP2 is 100 MHz.

D.3.4  DDC and DUC

The USRP2 FPGA uses a digital down-converter (DDC) and a digital up-converter
(DUC) to convert between baseband and intermediate frequency (IF).

D.3.4.1  Digital Down-Converter

Our standard FPGA configuration includes digital down-converters implemented
with cascaded integrator-comb (CIC) filters. CIC filters are very-high-performance
filters using only adds and delays. The FPGA implements four digital down con-
verters, which allows one, two, or four separate RX channels. At the RX path, we
have four ADCs and four DDCs. Each DDC has two inputs, I and Q. Each of the
four ADCs can be routed to either of I or the Q input of any of the four DDCs.
This allows for having multiple channels selected out of the same ADC sample
stream.
The DDC basically does two things. First, it down-converts the signal from
the IF band to the base band. Second, it decimates the signal so that the data rate
can be adapted by the USB 2.0 and is reasonable for the computers’ computing
capability.
The DDCs in USRP2 are very similar to the USRP1 with a four-stage CIC but
with two HBF filters. The CIC can decimate in the range [1...128]. The high rate
HBF has seven taps and it decimates by two. The low-rate HBF has 31 taps and
also decimates by two.
282 Quick Reference Sheet

D.3.4.2  Digital Up-Converter

The digital up-converter interpolates the signal, up-converts, it to the IF band, and
finally sends it through the DAC. There is a four-stage CIC and two HBF filters each
working in interpolation mode.

D.3.4.3  More Information

For more information on these converters, see the following pages:

• http://gnuradio.org/redmine/wiki/gnuradio/USRP2GenFAQ
• http://gnuradio.org/redmine/wiki/gnuradio/UsrpFAQDDC

D.4  DIFFERENTIAL PHASE-SHIFT KEYING (DPSK)

Differential phase shift keying (DPSK) is a form of phase modulation that conveys data
by changing the phase of the carrier wave. For example, a differentially encoded BPSK
(DBPSK) system is capable of transmitting a binary “1” by adding 180° to the current
phase, while a binary “0” can be transmitted by adding 0° to the current phase.
Since a physical channel is present between the transmitter and receiver within a
communication system, this channel will often introduce an unknown phase shift to
the PSK signal. Under these circumstances, the differential schemes can yield a bet-
ter error rate than the ordinary schemes, which rely on precise phase information.
Referring to both BPSK and QPSK modulation schemes, there is phase ambiguity
whenever the constellation is rotated by some amount due to the presence of phase
distortion within the communications channel through which the signal passes. This
problem can be resolved by using the data to change rather than set the phase.

Reference

[1] http://gnuradio.org/redmine/wiki/gnuradio/USRP2GenFAQ.
 Appendix E

Trigonometric Identities

exp(± jθ ) = cos(θ ) ± j sin(θ)

1
cos(θ ) = [exp(jθ ) + exp(− jθ)]
2
1
sin(θ ) = [exp(jθ ) − exp(− jθ)]
2j

sin2 (θ ) + cos2 (θ) = 1

cos2 (θ ) − sin2 (θ ) = cos(2 θ)

1
cos2 (θ ) = [1 + cos(2θ)]
2
1
sin2 (θ ) = [1 − cos(2θ)]
2

2sin(θ )cos(θ) = sin(2θ)

sin(α ± β ) = sin(α)cos(β ) ± cos(α)sin( β)

cos(α ± β ) = cos(α )cos( β ) ∓ sin(α)sin( β)

tan(α ) ± tan( β )
tan(α ± β ) =
1 ∓ tan(α )tan( β)
1
sin(α )sin(β) = [cos(α − β ) − cos(α + β )]
2
1
cos(α )cos(β) = [cos(α − β ) + cos(α + β )]
2
1
sin(α )cos(β ) = [sin(α − β ) + sin(α + β )]
2

283
About the Authors

Di Pu was born in Suzhou, China. She received her B.S. degree (with distinction)
from Nanjing University of Science and Technology (NJUST), Nanjing, China, in
2007 and M.S. degree from Worcester Polytechnic Institute (WPI), Worcester, MA,
in 2009. Since March 2008, she has been a member of the Wireless Innovation
Laboratory, where she is conducting research into cognitive radio system imple-
mentations. Di Pu is a recipient of the 2007 Institute Fellowship for pursuing Ph.D.
studies at WPI in electrical engineering.
Dr. Alexander M. Wyglinski is an associate professor of electrical and computer
engineering at Worcester Polytechnic Institute (WPI) and director of the Wireless Innova-
tion Laboratory (WI Lab). He received his Ph.D. degree from McGill University in 2005,
his M.S.(Eng.) degree from Queens University at Kingston in 2000, and his B.Eng. degree
from McGill University in 1999, all in electrical engineering. Prior to joining WPI,
Dr. Wyglinski was at the Information and Telecommunication Technology Center (ITTC)
of the University of Kansas from 2005 to 2007 as an assistant research professor.
Dr. Wyglinski is very actively involved in the wireless communications research
community, especially in the fields of cognitive radio systems and dynamic spectrum
access networks. He currently serves on the editorial boards of both IEEE Communi-
cations magazine and IEEE Communications Surveys and Tutorials; he was a tutorial
co-chair for the 2008 IEEE Symposia on New Frontiers in Dynamic Spectrum Access
Networks (IEEE DySPAN 2008); he is currently the student travel grants chair for
the 2010 IEEE Symposia on New Frontiers in Dynamic Spectrum Access Networks
(IEEE DySPAN 2010); he is the Communications Techniques and Technologies track
co-chair of the 2009 IEEE Military Communications Conference; he was the Wire-
less Mobile Communications track chair of the 2008 IEEE Military Communications
Conference; he has been a guest co-editor for the IEEE Communications magazine
with respect to two feature topics on cognitive radio (May 2007, April 2008); he was a
technical program committee co-chair of the Second International Conference on Cog-
nitive Radio Oriented Wireless Networks and Communications (CrownCom 2007);
he was a track chair for both the 64th and 66th IEEE Vehicular Technology Confer-
ence (VTC); and he is currently a technical program committee member on numerous
IEEE and other international conferences in wireless communications and networks.
Dr. Wyglinski is a senior member of the IEEE, IEEE Communications Society,
IEEE Signal Processing Society, IEEE Vehicular Technology Society, IEEE Women
in Engineering, Eta Kappa Nu, and Sigma Xi. Dr. Wyglinski’s current research in-
terests are in the areas of wireless communications, wireless networks, cognitive
radios, software-defined radios, transceiver optimization algorithms, dynamic spec-
trum access networks, spectrum sensing techniques, machine learning techniques
for communication systems, signal processing techniques for digital communica-
tions, hybrid fiber-wireless networking, and both multihop and ad hoc networks.

285
Index
A Cyclic autocorrelation function (CAF), 220
Accelerate the Simulink model, 166–167 Cyclic prefix, 185–186
Analog-to-digital conversion, 23 Cyclostationary feature detection, 219–222
Aperiodic, 16 Cyclostationary process, 219
Applications of software-defined radio, 239 Cyclostationary random process, 77
ASCII, 171
Autocorrelation function, 75 D
Autocovariance function, 76 Decimation, 27–28
Averaging sweeps, 214 Decision region, 119
AWGN channel, 117–118 Decision rule, 116
Decision statistic, 218
B Detection theory, 113
Barker code, 168–169 Differential phase shift keying (DPSK),
Bayes’ rule, 58 163–165
Bit allocation, 187–188 Differential quadrature phase-shift keying
Bit error rate (BER), 105 (DQPSK), 166
Bivariate normal, 71–72 Digital down converter (DDC), 146
Block interleaving, 135 Digital modulation, 94
Digital transceiver, 1
C Digital transmission, 89–91
Digital up converter (DUC), 145
Callback functions, 165
Dirac comb, 25
Carrier sense multiple access (CSMA), 230
Discrete-time, 15
Cascaded integrator-comb (CIC) filter, 44–46
Discrete Fourier transform (DFT), 182
Causal, 18
Dispersive channel environment, 183–184
Central limit theorem (CLT), 71
Duplex communication, 197
Channel encoding, 92–93
Dwell time, 213
Channel filter, 191
Dynamic spectrum access (DSA), 207
Cognitive radio, 207, 239
Collecting spectral data, 209–214
E
Collision avoidance (CA), 231
Commutator, 182 Einstein-Wiener-Khinchin (EWK) relations, 79
Conditional expectation, 62–64 Energy detection, 218–219
Conditional probability, 57–58 Energy subtraction, 158–159
Continuous-time, 15 Energy threshold, 218
Convolutional interleaving, 135 Error bounding, 107–108
Correlator realization, 122–124, 159–162 Euclidean distance, 97
Cumulative distribution function (CDF), 69 Expected value/mean, 62
Eye diagram, 32

287
288 Index

F Matched filter, 194

False alarm/Type I error, 216 Matched filter realization, 120–122
FFT plot, 142 Maximum a posteriori (MAP) detector, 117
Field programmable gate array (FPGA), 3 Maximum likelihood decoder, 157–159
Filtering, 39 Maximum likelihood (ML) detector, 117
Finite impulse response (FIR), 44 Medium access control (MAC), 230
Fourier transform, 18–23 Memoryless, 18
Frame synchronization, 167–168 Missed detection/Type II error, 216
Frequency domain equalization, 186–187 Multicarrier modulation (MCM), 177
Frequency offset, 138 Multicarrier transmission, 181–183
Frequency offset compensation, 138–140
Function, 56–57 N
Nyquist pulse, 34
G Nyquist pulse shaping theory, 32
Gaussian process, 77–78 Nyquist sampling theorem, 26–27
Gram-Schmidt orthogonalization, 110–112
O
H Observation vector, 153–154
Half-duplex (HDX), 198 Orthogonal frequency division multiplexing
Hypothesis testing, 105, 214–217 (OFDM), 182–183
Orthonormal basis functions, 108
I Oscillator, 138–139

IEEE 802.11, 140

P
In-phase and quadrature, 98
Independence, 59 Partition, 54–56
Infinite impulse response (IIR), 44 Parts per million (ppm), 139
Integrate-and-dump, 156 Peak frequency variation, 139
Intercarrier interference (ICI), 191–192 Per tone equalization, 187
Interleaving, 134 Periodic, 16
Interpolation, 29–30 Phase shift keying (PSK), 99–104
Intersymbol interference (ISI), 32 Poles and zeros, 41
Inverse discrete Fourier transform (IDFT), 182 Power allocation, 188
Power efficiency, 94–95
K Power of the detector, 217
Power spectral density (PSD), 78
Kansas University Agile Radio (KUAR), 10
Probability density function (PDF), 66
Probability mass function (PMF), 61
L
Probability of false alarm, 216–217
Laplace transform, 39 Probability of missed detection, 216–217
Linear time-invariant (LTI), 17 Pulse amplitude modulation (PAM), 95–97
Low-pass, high-pass, band-pass and band-stop, Pulse shaping, 30
40
Q
M
Quadrature amplitude modulation (QAM),
Machine learning, 239 98–99
Mask parameter, 131–132
Index 289

R T
Raised cosine pulse, 37–38 Tapered window, 229
Random process, 72 Test-bed implementation, 245–246
Random variable, 59–60 Time division duplexing (TDD), 198
Receiver design, 245 Total probability, 58
Receiver operating characteristic (ROC), 217 Transmitter design, 244–245
Receiver structure, 153
Rectangular QAM, 178 U
Rectangular window, 229 Uniform sampling, 24
Region of convergence (ROC), 41 Universal Software Radio Peripheral (USRP),
Relative frequency, 53 8–9
Repetition coding, 132–133 University Wireless Open Access Research
Platform (WARP), 10
S
Sample space, 54 V
Sampling, 23 Vector representation, 108–109
Sampling function, 24 Vehicular communication network, 242
Satellite communication, 247 Venn diagram, 54
SDRu Receiver block, 141
SDRu Transmitter block, 141 W
Set theory, 53–54
Water filling principle, 189
Shannon’s channel coding theorem, 93–94
Waveform synthesis, 153
Signal, 15
Wide-sense stationary (WSS), 77
Signal constellation, 179
WiFi network, 141
Signal waveform, 108
Window size, 211
Simplex communication, 197
Wireless local area network (WLAN), 140–141
Sinc pulse, 34–37
Single carrier transmission, 177–181
X
Sliding window approach, 227
Software-defined radio (SDR), 1 XCVR2450 daughtercard, 139
Source encoding, 91
SpeakEasy, 1–2 Z
Spectral coherence function (SOF), 221 Z transform, 39
Spectral correlation function (SCF), 221
Spectrum analyzer, 210
Spectrum sensing, 207
Square root raised cosine filter, 194–196
Stationary process, 76
Strict-sense stationary (SSS), 76–77
Substitution law, 64
Sweep resolution, 211
Sweep time, 214
System, 16