Signal Design and Detection

Marvin K. Simon=Sami M. Hinedi= William C. Lindsey

 Information Information
Channel Sink
Source

Information Signal Baseband Signal

Source Generator Channel Detector ™~

Information Signal IF
Modulator Demodulator
Source Generator Channel

Information Signal IF IF
Source Generator Transmitter Receiver

RF RF RF
Information Signal Modulator
Transmitter Channel Receiver
Source Generator Transmitter
(Media)

Information Signal \F Transmit Physical Receive

Modulator Transmit Channel Antenna
Source Generator Antenna

6g
|
| | ; Channel Layer 6 |
|
|
| | Channel Layer 5 |
|
|
| | Channel Layer 4

|
| Channel Layer 3

Channel Layer 2

Channel Layer 1

Systems Engineer Perspective

Figure 1.5 Information Transportation System (ITS) pyramid; uncoded communication system functional architecture

OPAMP TECHNICAL BOOKS

1033 North Sycamore Avenue
Los Angeles, CA 90038
Tel. (213) 464-4322— (800) 468-4322 —
Monday-Saturday 8:00-5:30 !
 Information
Sink

Signal information
Detector b | Sink

| Demodulator
Signal
Detector Information
Sink

IF Signal Information
Demodulator
Receiver Detector Sink
 -» Digitized by the Internet Archive
in 2022 with funding from
Kahle/Austin Foundation

https ://archive.org/details/digitalcommunica0000simo
Digital Communication Techniques
Signal Design and Detection

Marvin K. Simon
Jet Propulsion Laboratory

Sami M. Hinedi
Jet Propulsion Laboratory

William C. Lindsey
LinCom Corporation
University of Southern California

PTR Prentice Hall

Englewood Cliffs, New Jersey 07632
Library of Congress Cataloging-in-Publication Data
Simon, Marvin Kenneth,
Digital communication techniques / Marvin K. Simon, Sami M.
Hinedi, William C. Lindsey.
p. em.
Includes bibliographical references and index.
Contents: v. 1. Signal design and detection
ISBN 0-13-200610-3
1. Digital communciations. I. Hinedi, Sami M. II. Lindsey,
William C. _ IIL. Title.
TK5103.7.S52 1994
621.382—dc20 93-9053
CIP

Editorial/production supervision and interior design: Harriet Tellem

Cover design: Aren Graphics
Manufacturing buyer: Alexis Heydt
Acquisitions editor: Paul W. Becker
Editorial assistant: Maureen Diana

© 1995 by PTR Prentice Hall

Prentice-Hall, Inc.
A Paramount Communciations Company
Englewood Cliffs, New Jersey 07632

The publisher offers discounts on this book when ordered in bulk quantities:
Corporate Sales Department
PTR Prentice Hall
113 Sylvan Avenue
Englewood Cliffs, NJ 07632
Phone: 201-592-2863
FAX: 201-592-2249

All rights reserved. No part of this book may be reproduced, in any form or by any means,
without permission in writing from the publisher.

Printed in the United States of America

OROES SON uAaSo ie

ISBN: 0-135-200610-3

Prentice-Hall International (UK) Limited, London

Prentice-Hall of Australia Pty. Limited, Sydney
Prentice-Hall Canada Inc., Toronto
Prentice-Hall Hispanoamericana, S.A., Mexico
Prentice-Hall of India Private Limited, New Delhi
Prentice-Hall of Japan, Inc., Tokyo
Simon & Schuster Asia Pte. Ltd., Singapore
Editora Prentice-Hall do Brasil, Ltda., Rio de Janeiro
 Dedication

Marvin K. Simon would like to dedicate this book to his wife Anita in appreciation for her
consultation at every step of the way through the preparation of this manuscript. Indeed
this author could have never completed his portion of this book nor for that matter any
of his previous books without the support, encouragement, guidance, and sharing in the
many anxieties that accompany a project of this magnitude. This book is also dedicated to
his children Brette and Jeff who have made him very proud through their development and
accomplishments both as human beings and in their respective fields of interest.

Sami M. Hinedi would like to thank his father Mounir and his mother Hikmat for their
constant support, encouragement and guidance and for the various sacrifices they made
in order to provide their children with the best education one can hope for. He also would
like to thank his sister Fadwa, his brother Taha and last but not least, his dear uncle Abdul
Kader for their help and for being there when he needed them most. This book is dedicated
to all of them.

William C. Lindsey would like to dedicate this book to Dorothy, my wife, and to John, my
son, for the support system they have provided me throughout by professional career. Also,
to Corinne Leslie, my secretary for 27 years, for her loyal, continuous, and dedicated sup-
port; and to my sisters, Loretta and Jean for their lifelong inspiration and encouragement,
 reer,
aAO eer
695@ (Ee .

a oe Pret inves
7 Ne :
= EY

5 awl et

* pw Sea”
= \ irc bed ie

5
;
=
 Contents

Preface XV

Chapter 1 Introduction to Telecommunications 1

1.1. Digital Communication System Functional

Architecture 2
1.1.1 Communications Channel, 3
1.1.2 Uncoded Digital Communication System
Architecture, 7
1.1.3 Coded Digital Communication System Architecture, 7
1.1.4 ISO-OSI Network Architecture, 11
1.2 Telecommunication Networks and Services 14
1.2.1 Deep Space Network (DSN), 14
1.2.2 Tracking and Data Relay Satellite System (TDRSS), 14
1.2.3 Global Positioning System (GPS), 15
1.2.4 Integrated Services Digital Network (ISDN), 16
1.2.5 Military Satellite Communications, 16
1.2.6 Mobile and Personal Communications, 17
1.3. Performance Criterion and Link Budgets 27
1.3.1 Relation between Power and Bandwidth, 27
1.3.2 Link Budgets, 32
1.3.3 Signal-to-Noise Ratio as a Performance Criterion, 40
1.4 Shannon’s Capacity Theorem 42
1.5 Digital Communication with Subcarriers and Data
Formatting 43
Problems 47
References 53

Chapter 2 Power Spectral Density of Digital Modulations 56

2.1 Power Spectral Density of a Synchronous Data Pulse

Stream 56
2.1.1 Power Spectral Density of a Synchronous Data Pulse
Stream Generated by a Binary, Zero Mean, WSS =
Sequence, 57
vi Contents

2.1.2 Power Spectral Density of a Synchronous Data

Pulse Stream Generated by a Binary, Zero Mean,
Cyclostationary Sequence, 59
2.1.3 Power Spectral Density of a Synchronous Data
Pulse Stream Generated by a Binary, Nonzero Mean,
Cyclostationary Sequence, 60
22 Power Spectral Density of a Generalized M-ary Markov
Source 63
2.2.1 NRZ Baseband Signaling, 65
2.2.2 RZ Baseband Signaling, 65
2.2.3 Biphase (Manchester) Baseband Signaling, 66
2.2.4 Delay Modulation or Miller Coding, 66
Pa) Power Spectral Density for Data Pulse Streams with
Data Asymmetry 68
2.3.1 NRZ Data, 68
2.3.2 Manchester Data, 69
2.4 Power Spectral Density of Convolutionally Coded
Modulations 72
2.4.1 Convolutional Encoder Model, 73
2.4.2 Encoder Output Spectrum for Independent Binary
Symbol Input, 75
2.4.3 Encoder Output Spectrum for First-Order Markov
Input, 80
be) Comparison of the Spectral Occupancy of Various I-Q
Modulation Techniques 83
2.6 Power Spectral Density of Continuous Phase
Modulation (CPM) 88
Problems 92
References 96

Chapter 3 Scalar and Vector Communications Over the

Memoryless Channel 98
3:1 Scalar Communications Over the Memoryless
Channel 100
3.1.1 Scalar Communications Over the Discrete Memoryless
Channel, 102
3.1.2 Scalar Communications Over the Continuous
Memoryless Channel, 107
3.1.3 Optimum Scalar Receivers, 108
3.1.4. Error Probability Performance, 113
meP4 Vector Communications Over the Continuous
Memoryless Channel 115
3.2.1 A Posteriori Probability Computer, 116
3.2.2 Optimum Vector Correlation and Minimum Distance
Receivers, 117
Contents vii

3.2.3 Decision Rules, Functions, Regions, and

Boundaries, 126
3.2.4 Error Probability Performance, 128
3.2.5 Effects of Translation and Rotation of the Signal
Vectors on Message Error Probability, 131
3.2.6 Translation to Minimize the Required Signal
Energy, 133
3.2.7. Union Bound on M-ary Vector Receiver
Performance, 135
3.2.8 Error Probability Performance of Two-Dimensional,
M-ary Vector Receivers—A New Approach, 137
3A Other Detection Criteria 141
3A.1 The Bayes Criterion, 141
3A.2 The Min-Max Criterion, 143
3A.3 The Neyman-Pearson Criterion, 145
3B The Gaussian Integral Q(x) 147
3B.1_ Table of Q(x), 147
3B.2 Series Approximation for Q(x), 147
3C Gram-Schmidt Orthogonalization Procedure 154
3D Connecting the Linear Vector Space to the Physical
Waveform Space 157
z)3 The Pairwise Probability of Error (PPE): A Tool
for System Performance Evaluation and Signal
Design 160
Problems 162
References 176

Chapter 4 Coherent Communications with Waveforms 178

4.1 Optimum Waveform Receivers in White Gaussian

Noise 179
4.1.1 Binary Cross-Correlation Receivers, 180
4.1.2 Matched Filter Receivers, 181
4.1.3 M-ary Waveform Receivers, 184
4.1.4 Error Probability Performance, 186
4.1.5 Limiting Error Probability for Block Codes as
M > ~, 221
4.1.6 Matrix Generation of Binary Signal Sets, 223
4.1.7. Coherent Detection of Differentially Encoded M-PSK
Signals, 225
4.2 Optimum Waveform Receivers in Colored Gaussian
Noise 235
4.2.1 Time-Sampling Approach, 235
4.2.2 Karhunen-Loéve (K-L) Expansion Approach, 237
4.2.3 Whitening Approach, 246
Wipf Cire tent Seat ct AnDrOach. 259
viii Contents

4.3 Inphase and Quadrature Modulation and

Demodulation 261
4.3.1 Real and Complex Signal Models, 261
4.3.2 Effect of Imperfect Carrier Synchronization, 267
4.3.3 Effect of Imperfect Bit Synchronization, 268
4.3.4 Effect of Data Asymmetry, 271
4.3.5 Suboptimum Detection, 277
4A Derivation of the Symbol Error Probability for
Polyphase Signals 285
4B Bandpass Random Process Representation 287
Problems 292
References 308

Chapter 5 Noncoherent Communication with Waveforms 309

5.1 Noncoherent Receivers in Random Phase

Channels 310
5.1.1 Optimum M-FSK Receivers, 316
5.1.2 Suboptimum M-FSK Receivers, 319
5.2 Performance of Noncoherent Receivers in Random
Phase Channels 321
5.2.1 Performance of M-ary Orthogonal Signals, 322
5.2.2 Performance of M-ary Nonorthogonal Signals, 327
5.2.3 Performance of M-ary Equicorrelated Signals, 333
5.2.4 Performance of Binary Nonorthogonal Signals, 334
5.2.5 Effect of Frequency Error on Orthogonal M-FSK
Detection, 338
5.2.6 Effect of Timing Error on Orthogonal M-FSK
Detection, 343
5.2.7 Effect of Timing and Frequency Errors on Orthogonal
M-FSK Detection, 349
5.2.8 A Bound on the Performance of Orthogonal M-FSK
Detection in the Presence of Frequency Error, 354
5.2.9 Performance of Two Suboptimum Receivers, 361
5.3. Noncoherent Receivers in Random Amplitude and
Phase Channels 368
5.3.1 Optimum Receivers in Rayleigh Channels, 370
5.3.2 Optimum Receivers in Rician Channels, 372
5.4 Performance of Noncoherent Receivers in Random
Amplitude and Phase Channels 375
5.4.1 Performance in Rayleigh Channels, 375
5.4.2 Performance in Rician Channels, 376
5A Useful Probability Density Functions 385
SA.1 Central Chi-Square Density, 385
SA.2_ Noncentral Chi-Square Density, 386
_9A.3 Rayleigh Density, 387
Contents

5A.4_ Rician Density, 388

SA.5 On the Joint Density of Amplitude and Phase, 389
5B Derivation of Equation (5.43)392
aC Generalized M-ary Symbol Error Probability
Bound 395
Problems 397
References 404

Chapter6 Partially Coherent Communication with Waveforms 406

6.1 Optimum Partially Coherent Receivers Based on One
Symbol Observation 408
6.2 Performance of Partially Coherent Receivers Based on
One Symbol Observation 409
6.2.1 Performance of Binary Antipodal Signals, 410
6.2.2 Performance of Binary Orthogonal Signals, 412
6.2.3 Performance of Binary Nonorthogonal Signals, 413
6.2.4 Performance of M-ary Orthogonal Signals, 415
6.3 Optimum Partially Coherent Receivers Based on
Multiple Symbol Observation 419
6.3.1 Maximum-Likelihood Block-by-Block Detection of
M-PSK Signals, 419
6.4 Performance of Partially Coherent Receivers Based on
Multiple Symbol Observation 422
6.4.1 Evaluation of the Pairwise Error Probability, 424
6.4.2 Special Cases, 425
6.4.3 Performance Sensitivity to Mismatch, 430
Problems 431
References 437

Chapter 7 Differentially Coherent Communication with

Waveforms 438

a Classical M-DPSK 439

7.1.1 Symbol Error Probability Performance, 441
7.1.2 Bit Error Probability Performance, 446
7.1.3 Effect of Noise Correlation and Power Imbalance, 447
7.1.4 Effect of Imperfect Timing, 449
7.1.5 Double Error Rate Performance, 452
7.1.6 Error Probability Performance Comparison of DPSK
and Noncoherent FSK, 456
hee, Optimum Waveform Receivers Based on a Larger Than
Two Symbol Observation Interval 460
7.2.1 Maximum-Likelihood Differential Detection of
M-PSK—Block-by-Block Detection, 461
7.2.2 Bit Error Probability Performance, 466
7.2.3 Implementation Algorithms, 474
x Contents

7.3. DPSK—Noncoherent Detection of Overlapped Binary

Orthogonal Modulation 477
7.3.1 Generalization to Overlapped M-ary Orthogonal
Modulation, 478
7.4 Suboptimum Waveform Receivers 484
7.4.1 Wideband IF Filter, Integrate-and-Dump Detector, 485
7.4.2 Narrowband IF Filter, Sampling Detector, 487
7.4.3 Multiple Decision Feedback Symbol-by-Symbol
Detector, 488
7.5 Symmetric DPSK 490
7A Derivation of the Optimum Receiver for Classical
(2-Symbol Observation) M-DPSK 493
7B Probability Distribution of the Phase Angle between
Two Vectors Perturbed by Additive Gaussian
Noise 497
Problems 498
References 505

Chapter 8 Double Differentially Coherent Communication with

Waveforms 508

8.1 Definition of Higher Order Phase Difference Modulation

(PDM) 509
8.2 Structures Motivated by the Maximum-Likelihood (ML)
Approach 511
8.3. Autocorrelation Demodulation (ACD) of Binary PDM-1
Signals 514
8.4 I-Q Demodulation of Binary PDM-1 Signals 518
8.5 Comparison of ACD and I-Q Demodulation of PDM-1
Signals 522
8.6 Serial (Absolutely Frequency Invariant) Structures 525
8.6.1 Autocorrelation Demodulation (ACD) of PDM-2
Signals, 525
8.6.2 I-Q Demodulation of PDM-2 Signals, 533
8.6.3 Multiple Symbol Differential Detection of PDM-2
Signals, 538
8.6.4 Further Enhancements, 542
8.7 Parallel (Relatively Frequency Invariant) Structures—
Multiphase ACD 546
Problems 550
References 553

Chapter9 Communication over Bandlimited Channels 555

9.1 Definition and Characterization of a Bandlimited

Channel 557
9.2 Optimum Pulse Shape Design for Digital Signaling
through Bandlimited AWGN Channels 558
Contents xi

9.2.1 Nyquist Criterion for Zero ISI, 561

9.2.2 Apportionment of Filtering between Transmitter and
Receiver, 564
9.2.3 The Eye Pattern of a Bandlimited Communication
System, 567
9.2.4 Systems with Controlled ISI—Partial Response, 568
9.3. Optimum Demodulation of Digital Signals in the
Presence of ISI and AWGN 578
9.3.1 Evaluation of the MLSE—The Viterbi Algorithm, 582
9.3.2 Error Probability Performance, 591
9.4 Equalization Techniques 600
9.4.1 Zero-Forcing Linear Equalizer, 602
9.4.2 _Mean-Squared Error Linear Equalizer, 603
9.4.3 Decision-Feedback Equalization, 605
9.5 Further Discussion 611
Problems 613
References 619

Chapter 10 Demodulation and Detection of Other Digital

Modulations 621

10.1 I-Q Modulations 622

10.1.1 Unbalanced QPSK (UQPSK), 622
10.1.2 Quadrature Amplitude Modulation (QAM), 627
10.1.3 Quadrature Biorthogonal Modulation, 641
10.2 Continuous Phase Modulation (CPM) 651
10.2.1 Continuous Phase Frequency Modulation
(CPFM), 651
10.2.2 Continuous Phase FSK (CPFSK), 653
10.2.3 Minimum-Shift-Keying (MSK), 675
10.2.4 Sinusoidal Frequency-Shift-Keying (SFSK), 686
10.2.5. Continuous Phase FPM and QFPM, 688
10.3 Further Discussion 695
Problems 699
References 708

Chapter 11 Coded Digital Communications 711

11.1 Information Transfer Using Bit-by-Bit Binary

Signaling 712
11.2 Coded Digital Communication System
Architectures 713
11.2.1 Coding-Decoding System Architecture (Level 3), 714
11.2.2 Coded-Modulation and Demodulation-Decoding
System Architecture (Level 5), 718
11.2.3 Quantizing the Demodulator/Detector Outputs, 732
11.2.4 Creating the DMC for Channel Decoding, 735
xii Contents

11.3 Interleaving-Deinterleaving to Mitigate Channel

Memory 736
11.3.1 Periodic Block Interleavers-Deinterleavers, 737
11.3.2 The Helical Interleaver, 738
11.3.3. The Periodic Convolutional Interleaver-
Deinterleaver, 739
11.3.4 Other Interleavers-Deinterleavers, 741
11.4 Synchronization of I-D Systems 741
11.5 Coded System Architecture for Channels with
Memory 742
11.5.1 Coding and Modulation Matching for Concatenated
Coding Systems, 744
11.6 Classification of Error Correction and Detection
Codes 744
Problems 748
References 750

Chapter 12 Block-Coded Digital Communications 752

12.1 Block-Coded Digital Communication System

Architecture 753
12.1.1 Coding from the Baseband and IF Channel Model
Perspectives, 755
12.1.2 Fundamental Parameters of a Block-Coded
Communication System, 757
12.2 Performance of Block-Coded Communication
Systems 757
12.2.1 Transmission of Orthogonal Modulation Vectors, 760
12.3 Special Types of Binary Block Codes 761
12.3.1 Orthogonal Binary Codes, 762
12.3.2 Biorthogonal Binary Codes, 763
12.3.3 Transorthogonal Binary Codes, 763
12.3.4 The Bandwidth Issue for Binary Coded BPSK
Modulation, 764
12.3.5 Channel Throughput and Time-Bandwidth
Requirements, 764
12.4 The Ensemble of Block-Coded Modulation
Systems 765
12.4.1 Performance of the Ensemble of Binary Coded
Systems Using BPSK Modulation, 766
12.4.2 MEP Performance Comparison of Orthogonal Codes
with the Ensemble of Binary Codes Using BPSK, 770
12.4.3 Performance of the Ensemble of Q-ary Codes Using
Q-ary ASK and QASK, 771
12.4.4 Comparison of the Ensemble Performance of Q-ary
Codes with Energy Constrained Codes, 777
Contents xiii

12.4.5 Q-ary Coded M-PSK, 778

12.4.6 Further Discussion Regarding Ro and R*, 779
12.4.7 The Bhattacharyya Bound and Distance for Isometric
Coded Modulation, 780
£25 Shannon’s Channel Coding Theorem 781
12.5.1 Shannon Limits and Information Transfer
Physics, 781
12.5.2 The Channel Capacity, Matched Filter, and Spread
Spectrum Communications Interconnect, 784
12.5.3 Spread Spectrum Communications, 785
12.5.4 The Union Bound for Binary Orthogonal Codes
Written in Terms of the Shannon Capacity Limit, 786
12.6 Gallager’s Upper Bound on the Ensemble Error
Probability 786
12.6.1 Block versus Convolutional Coding for the Very Noisy
Channel, 788
127 Demodulator Design Using the Ro Criterion 789
12.7.1 Design of Hard-Decision Binary Coded BPSK
Demodulators, 790
12.7.2 Design of Soft-Decision Coded BPSK
Demodulators, 791
12.7.3, Choice of the Number of Quantization Levels, 794
12.7.4 Coded DPSK and DDPSK Modems with Hard
Decisions, 795
12.7.5 Binary Coded DPSK with Infinite Quantization, 796
12.7.6 Optimization of Codes and Modem Design Using the
Ro Criterion, 797
12.8 Linear Block Codes 799
12.8.1 The Union Bound Expressed in Terms of the Hamming
Distance for Binary Codes, 800
12.8.2 The Plotkin Bound for Binary Codes, 802
12.8.3 The Role of the Minimum Hamming Distance in
Specifying Good Error Control Codes, 802
12.8.4 Error Probability Performance for Error-Correcting
Block Codes, 803
12.8.5 Hamming Codes and Their Error Probability
Performance, 804
12.8.6 Golay Codes and Their Error Probability
Performance, 805
12.8.7 Cyclic Codes, 805
12.8.8 BCH Codes and Their Error Probability
Performance, 806
12.8.9 Reed-Solomon Codes and Their Error Probability
Performance, 809
Problems 820
References 824
xiv Contents

Chapter 13 Convolutional-Coded Digital Communications 826

131 Convolutional Encoders 826

1
ey Representation of Convolutional Codes 827
13.2.1 Polynomial Representation, 827
13.2.2 The State Diagram, 832
13.2.3 The Tree Diagram, 832
13.2.4 The Trellis Diagram, 833
13 Decoding Convolutional Codes 836
13.3.1 Maximum-Likelihood Decoding of Coded BPSK
Modulation, 838
13.3.2 _Maximum-Likelihood Decoding Using the Viterbi
Algorithm, 842
13.3.3 Hard-Decision and Soft-Decision Viterbi
Decoding, 843
13.3.4 Path Memory and Synchronization Requirements, 847
13.4 Probability of Decoding Error for Convolutional
Codes 850
13.4.1 Distance Properties of Convolutional Codes, 850
13.4.2. The Transfer Function Approach, 852
1335 Error Probability Performance of Convolutional Coded
BPSK Modulation 854
13.5.1 Hard- and Soft-Decision Decoded Error
Probability, 854
13.5.2 Bit Error Probability, 855
13.6 Error Probability Performance of Convolutional Codes
Using the Viterbi Algorithm 856
137 Concatenated Reed-Solomon and Convolutional
Codes 856
13.8 Other Convolutional Decoding Techniques 860
13.8.1 Sequential Decoding, 861
13.8.2 Threshold Decoding, 863
13.9 Punctured Convolutional Codes 864
Problems 874
References 876

Index 879
Tg

Preface
—_—————[— js SSS

Future applications of Digital Communication Techniques to architecting and implement-

ing global information transportation and computing systems have never been brighter. This
outlook is driven by social, economical, political, and technological reasons. From a techni-
cal perspective, it is recognized by most communication engineers that communications is
required to accomplish computing while computers are required to accomplish communica-
tions. From a technology viewpoint, this technical perspective is rapidly being accomplished
using emerging digital microelectronic technologies (DSP and VLSI) to implement digital
communication systems.
Digital Communication Techniques are exciting and are of vital importance to all
societies. Countries have failed to be competitive simply because they did not succeed in
establishing good communication infrastructures. Consequently, one major purpose of this
textbook is to present, in a unique and innovative way, a functional architecture and a the-
ory for use in the design of uncoded and coded digital communication systems. The sys-
tem architecture is pyramidal and the theoretical development is unique in that it is pre-
sented, for the first time, from a Systems Engineering perspective for both bandlimited
and power limited communication channels. This perspective adopts the point of view that
coding and modulation are both components of the signal design problem and that de-
modulation and decoding are both components of the signal detection problem. Beginning
with Chapter 1, the subject matter progresses top down and systematically in a hierarchial
way. The geometric concepts, first introduced by Shannon and Kotelnikov, and later docu-
mented by Wozencraft and Jacobs in their book Principles of Communication Engineering,
are used in Chapter 3 to set the foundations for signal design and detection. Starting with
Chapter 4 and ending in Chapter 8, coherent, noncoherent, partially coherent, and dif-
ferentially coherent detection techniques are treated for numerous uncoded modulation
techniques, such as BPSK, QPSK, MPSK, and MFSK. In Chapter 10, these same detec-
tion techniques are applied to more advanced forms of uncoded modulation such as QAM,
CPFSK, MSK, QFPM, and CPQFPM. As opposed to the M-ary error probability criterion
used in designing uncoded systems, the Ro-criterion is introduced for use in optimizing the
design of coded systems. Since Ro is a function of the codec and modem choice, this cri-
terion leads to a combined codec-modem design that employs the most effective coding
and modulation technique. In fact, soft-decision demodulators can be systematically de-
signed using Ro. Chapters 11, 12, and 13 consider block, convolutional, and concatenated _ )
coding techniques from the systems perspective. In addition, the counterparts of maximum- °
likelihood (ML) decoding and ML decoding using the Viterbi algorithm are given. A va-
riety of interleaving-deinterleaving techniques (block, convolutional, helical, hybrid) are
presented. To understand the connection among the various coding techniques presented,
 xvi Preface
1

a Venn diagram for error correcting codes is constructed; emphasis is placed on present-
ing the communciations efficiency achieved using Hamming, Golay, Bose-Chaudhuri-
Hocquenghem (BCH), Reed Solomon (RS), and convolutional codes.
There are a vast number of textbooks on the market today that deal with the subject
of digital communications. In fact, in a quick survey of the textbook literature, we were
able to come up with at least ten books that include the words “digital communications” in
their titles. At least three bear these words as their entire title implying, somewhat, that they
are an all-inclusive treatment of the subject. Many of these books are quite broad in scope,
but also quite shallow in detail. Striking a proper balance between these two attributes,
yet maintaining a high level of readability, is no simple task. We believe that our book
accomplishes this important goal and sets it apart from all other digital communication texts
currently on the market. Several key features that distinguish our book from the others are
as follows:

e A top-down perspective of digital communication system design, using a pyramid

structure to describe the system functional architecture.
e A top-down presentation of the theory needed to perform uncoded and coded system
design.
Includes Ro criterion for use in the design of coded systems.
e Includes more recent developments in the field that have occurred over the last 20 or
30 years.
A universal appeal to graduate students as well as system architects and engineers.
Written by authors whose combined industrial and university experience exceeds 60
years.

There are many specific features that make this book unique and beneficial to its read-
ers. With the advent of today’s advances in the solid state microelectronic technologies, a
variety of novel digital communication systems are appearing on the market and are serv-
ing as motivation for the introduction of new telecommunication and information services.
Chapter 1 of this book provides the reader with examples of such services and top level
system architectures thereby indicating the highly complex nature of these systems. We
believe that Chapter 2, which discusses the computation of power spectral density of dig-
ital modulations, is the best treatment of spectrum efficiency evaluation found anywhere.
This computation is essential to assessing the bandwidth (spectral) occupancy requirement
of a digital modulation, yet it is ignored in many books. Another key feature of this book
is the organization and order of presentation of the material in Chapters 3 through 7. By
first describing coherent detection and then successively following with noncoherent, par-
tially coherent, and differentially coherent detection, the reader is provided with a logical
flow starting with the conceptually simplest technique and proceeding top down to the more
complex techniques. The discussion of double differentially coherent detection in Chapter 8
is unique to our book.! Here the reader will learn how to design differentially coherent com-
munication systems that are robust to frequency offsets due, for example, to Doppler and
_ oscillator instabilities. Chapter 9 treats the voluminous subject of bandlimited communica-
tions in a condensed and unified way. Included here are the important subjects of Nyquist

1. There are a few Russian-authored textbooks that discuss double differentially coherent detection, but as yet,
they have not been translated into English.
Preface XVii

and partial response signaling, maximum-likelihood detection in the presence of intersym-

bol interference, and equalization. To the authors’ knowledge, Chapter 10 of this book is the
most complete and up-to-date treatment of advanced modulation techniques. It guides the
reader through the most recent modulation techniques described in the literature and how
they compare in terms of such properties as modulation type, pulse shaping, continuity of
phase, variation of envelope, I and Q channel data rates, and parameter offsets. Still another
key feature is the proper identification of the important link (mapping function) between the
modulation and coding functions in coded digital communication systems. Most books that
discuss both modulation and coding treat these two functions as separate and independent
entities. In some books where modulation and coding are treated as combined, the treat-
ment is strictly limited to trellis coded modulation. This book is unique in that it presents a
general formulation for coded communication systems by properly defining the key parame-
ters and interfaces between the various modulation and coding functions. This discussion is
presented in Chapter 11 which includes many examples to clearly elucidate this often over-
looked but all important aspect of system design. Chapter 12 discusses the use of the Ro
criterion in the design of coded systems. Finally, Chapter 13 presents a compact yet author-
itative discussion of the design of convolutionally-coded communication systems, a subject
that, by itself, can occupy an entire textbook.
We recognize that a complete study of reliable and efficient communication of infor-
mation requires a full and detailed treatment of the two important disciplines: information
theory and communication theory. Since the main focus in this book is on the latter, we do
not treat the problem of efficient packaging of information (data compression of text, voice,
video, etc.) nor do we treat the important problem of designing the ultimate error control
coding-decoding technique which achieves the ultimate transmission speed (channel capac-
ity). The solution to these problems are best treated in separate books on information theory
and error control coding, and, indeed, there are such books available.
This book has been written for use as a textbook at universities involved in teaching
Communication Sciences. It has also been designed to accommodate certain needs of the
systems architect, systems engineer, the professor, and communication sciences researcher.
The lecture material has been organized and written in a form whereby theory and prac-
tice are continuously emphasized. Most of the problems suggested at the conclusion of each
chapter have evolved from teaching graduate level courses to students at the University of
Southern California and the California Institute of Technology. Through homework assign-
ments, most of the problems have been field tested, corrected, and enhanced over the years.
The architecture for this book is predicated upon two graduate level Communica-
tion Theory courses (EE564 and EE664) taught at the University of Southern California’s
Communication Sciences Institute over the past 25 years. In this sense, two semesters are
required to cover its contents. The organization and presentation of the material is largely
based upon the academic and course design work of Professors William C. Lindsey and
Robert A. Scholtz. The authors’ approach to presenting the solution to the problem of vector
communications in the presence of colored noise and the representation of bandpass random
processes are largely due to Professor Scholtz. The Digital Communication System architec-
ture presented in the pyramids of Chapter 1 was created by Professor Lindsey. In addition,
certain exercises given in and at the conclusion of Chapters 1, 3, 4, 5, and 6 were taken **
— < ae,

eet
with permission from problems created and used by Professors Robert M. Gagliardi, Vijay
Kumar, William C. Lindsey, Andreas Polydoros, Charles L. Weber, and Robert A. Scholtz
xviii Preface

all of the University of Southern California’s Communication Sciences Institute. The au-
thors wish to thank these individuals for permission to use these challenging problems as
homework exercises.
The authors are grateful for the help of many colleagues and students during the
preparation of this manuscript, and are particularly indebted to: Dr. Michael Mandel, for-
merly a graduate student at the California Institute of Technology; Dr. Debbie Van Alphen,
Professor of Electrical Engineering at California State University, Northridge; and Dr.
Jorge M.N. Pereira of GTE; for their comments and criticisms. Furthermore, we thank our
colleagues at the Jet Propulsion Laboratory, Pasadena, CA, namely: Dr. Tien Nguyen, Dr.
Haiping Tsou, Dr. Daniel Raphaeli, Mr. Biren Shah, Mr. Sampson Million, Mr. Alexander
Mileant, Mr. Mazen Shihabi, Dr. Ramin Sadr, Mr. Roland Bevan, and Dr. Victor Vilnrotter
for reading and reviewing various parts of the manuscript.

Marvin K. Simon
Sami M. Hinedi
William C. Lindsey

Pasadena, CA
 Chapter 1
Introduction
to Telecommunications

Today’s advances in solid state electronic technologies are impacting the design and imple-
mentation of various digital communication systems, and specifically, terrestrial and space-
related telecommunication networks and services. Many systems/networks that once were
considered impractical due to their complexities are now well within the reach of current
technology and are presently being designed, implemented, and operated. Many microelec-
tronic technologies are maturing so fast that what is considered impractical today may well
be implementable tomorrow. As a result, novel telecommunication and information services
are continuously being introduced into the market.
At the system component level, microprocessors are operating faster and faster, front-
end amplifiers are achieving gains in frequency regions that were once considered unattain-
able with noise figures that approach the theoretical limit, antenna surfaces are becoming
“smoother,” thus enabling higher frequency transmission and reception. Various other ad-
vanced technologies, especially wireless, are affecting and sometimes dictating the system
design architecture and strategy. At the systems level, satellites are moving into an era
of multiple beams and high-gain spot beams allowing for frequency reuse, supplementing
global communications coverage and even allowing the use of “very” small antenna aper-
tures. Cellular terrestrial and proposed space radio systems around the world are testing
digital systems to replace their analog FM equipment, while cordless telephones are moving
outside the boundaries of the home and may merge with cellular systems, enabling users to
carry phones with them. Mobile satellite experiments [1] are currently being carried out to
test the next generation system in which a mobile user, located in a car, on a plane or a ship,
can directly communicate with a satellite, bypassing the need for a relay station. Existing
telephone networks are being upgraded using fiber-optic technology [2] to allow for higher
data rates and better transmission quality enabling novel information services. These ad-
vances and others are resulting in improved system performances, higher reliability, easier
operability and maintainability. Consequently, systems life cycles are increased dramatically
and life cycle costs decreased. Global mobile and personal communication systems are the
wave of the future, and their applications are endless.
Telecommunication engineering and more specifically, coding, modulation, ‘demod-
ulation, decoding, detection, and synchronization techniques are the technical topics to be
covered in the chapters to follow. Fundamentals of signal detection, being coherent, non-
coherent, partially and differentially coherent, are the key methodologies to the design of
 2 Introduction to Telecommunications Chap. 1

optimum receivers. The choice of the modulation scheme from the “simple” binary phase
shift keying (BPSK) to the more “elaborate” quadrature amplitude modulations (QAM) is
essential in trading channel bandwidth and achievable bit error rate for a given and fixed
transmitted signal power. The synchronization function is fundamental in operating any
communication link. At the receiving site, various timing and phase references are needed
and are derived from the incoming noisy signal; these include carrier phase and frequency
estimation, possibly subcarrier phase and frequency, and definitely symbol (bit) timing to
recover the transmitted information bits. Synchronization precludes communication and in-
cludes these two functions: acquisition and tracking. Each of the carrier, subcarrier, and
symbol acquisition processes consists of first frequency, then phase acquisition, and the total
lock-up (or acquisition time) depends on the specific structures or algorithms employed. The
transition from acquisition to tracking function is nonuniquely defined and is typically said
to have occurred when the instantaneous phase error decreases and remains below a pre-
determined threshold. Automatic gain control (AGC) circuitry is essential in maintaining
reasonable received power levels in the receiver and in providing some protection against
“Jarge” intentional or accidental interferences. Finally, signal reference generators at both
the transmitter and receiver require some degree of time and frequency stability to maintain
a fixed reference throughout the system.
This chapter introduces the reader to the architecture of a digital communication
system and introduces the various terms and key parameters used throughout the book.
Section 1.2 presents practical “telecommunication networks” and discusses future networks.
The various elements and key functions of an end-to-end communication link are discussed
in Section 1.3. The key performance parameters such as signal-to-noise ratios, bandwidth,
and so on, are defined along with the various losses that need to be accounted for in a
link budget analysis. The “information” capacity of a communication channel as defined
by Shannon is discussed in Section 1.4 and in Chapter 11. Communication with subcarriers
and data formatting of various signals is the topic of Section 1.5.

1.1 DIGITAL COMMUNICATION SYSTEM FUNCTIONAL ARCHITECTURE

For the most part, the functional architecture of a digital communication system [3] has
evolved over the past 35 years; with few exceptions, the functional architecture matches
the physical structure. This evolution has been driven by two major forces, viz., the de-
velopment of communication and information theory (communication sciences) and the
development of communications, computer, and information technology. In presenting the
functional architecture of a digital communication system, the top-down approach will be
used. This approach takes advantage of the hierarchical nature of any system, in particular,
an information transportation system which transports information using a digital commu-
nication system. In addition, the top-down approach fits well with the presentation and de-
velopment of the digital communication techniques and theory presented in the chapters that
follow.
Our top-down approach begins with the simplest architectural level, Level 0: The Con-
ceptual Level, and proceeds downward until the bottom level, viz., the Physical Level, is
reached. Figure 1.1 demonstrates that the basic building blocks of an Information Trans-
portation System (ITS) in level one of the hierarchy are: (1) an information source to be
transported, (2) a communication channel or information pipe, and (3) an information user
or a sink. In order to connect the information source to the channel, a transducer is needed.
In level 2 of the hierarchy in Fig. 1.1, this transducer is identified as the communication
Sec. 1.1 Digital Communication System Functional Architecture 3

transmitter. Simultaneously, a transducer is also needed at the channel output of level one,
and this is accomplished in level 2 by adding a communication receiver to interface with the
information user. Thus, we conclude that the basic building blocks of a digital communi-
cation system are a transmitter, a physical channel (communications pipe), and a receiver.
These blocks will be further partitioned and discussed.
As we shall see, the system engineering and design of a digital communication sys-
tem best proceeds top-down, while implementation and system integration proceeds best
bottom-up, that is, top-down design driven by system requirements followed by bottom-up
integration of the system building blocks yields a system that meets the system user require-
ments. From the system requirements, system engineers generate design specifications and
from these specifications, implementations (software and hardware) are perceived. Execut-
ing this top-down design process, which is driven by the system requirements and bottom-up
integration, constitutes what is frequently referred to in practice as Systems Engineering and
Integration (SE&]I).
The interfaces between the functional blocks are control signals and are specified by
interface control documents (ICDs). In what follows, a further breakdown of the functional
architecture of the information transportation system will be presented in the form of a
hierarchy. All systems are hierarchical in nature—one engineer’s part is another’s system.
The transmitter, channel, and receiver can easily be identified at each level in the hierarchy.
Figure 1.2 serves to demonstrate the building blocks for the communications transmitter
and the communication receiver in level 2 of a digital communication system architecture
in Fig. 1.1. This figure shows that each system building block is a distinct architecture within
a larger architecture in the hierarchy. These building blocks will be modeled and elaborated
upon below, as well as in Chapters 3 through 13.

1.1.1 Communications Channel

We observe from Fig. 1.1 that communication system architecture centers or pivots around
the communications channel. The architecture of this channel can best be demonstrated
by introducing the channel hierarchy. The architecture of this channel hierarchy can be
conveniently represented using six levels; these levels are identified in Fig. 1.3. In this
chapter and Chapter 2, we are primarily concerned with modeling and analysis of levels
4, 5, and 6. In the remaining chapters, we will be incorporating channel levels 4, 5, and 6
as a part of the IF channel in level 3. In other words, the mathematical characterization of
the IF channel must serve to mimic the features of the RF channel, the Antenna-to-Antenna
Channel, and the Physical Channel. In a sense, the IF channel model will serve as a “virtual
channel” model that mimics the physics of the hardware and the physical channel at levels
4, 5, and 6. The concept of a virtual channel is key to understanding the modeling of the
communications process. A virtual channel is a set of resources necessary to complete the
communication link between information source and user. In this book, we elect to model
their effects, for the most part, below level 3 in the Information Transportation System (ITS)
hierarchy. This model will consider the effects of power (energy) reduction as a function
of the building blocks and source-user separation primarily below level 3 of the system
hierarchy. In the following section, we further identify the building blocks and features at
each channel level in the hierarchy demonstrated in Fig. 1.3. In essence, the building blocks
of the communication channel at any level are always defined by those building blocks in
the architectural levels below.
Figure 1.4 demonstrates the communication systems engineering education pyramid.
 UOITELUIO}U|
oneuodsues)
wa}shs

eo ee ee ee oe ae ee

4UIS

20/N0S
UOITELUJO}U|
UOITEWUJOJUY
UONTEWIOJU]
ug

UOyeTUNUTIUIO?7)
wiaysAS oInps}IYIV
eansiq
*] poylduig uoNeorunWIUIOD
wia}shs sinjoajtyore pojutidoy)
wor ([g¢]
 Joyiuusuel]

puegeseg-8pop fal-d du Buriwsues)

Joyejn ;
JOSS89014 JOAC9|18}U| Jopiwsuel] JBUBAUDD
| Jeyiusuel euuejuy

UOIVEUOJU|
MO|4

JOAIB0eY

-10po0eq
pueqeseg -UMOG dy Bulaisoay
JOSS89014 BARa|JA]UIBG JeyeAUOD JEAlIe0ey euuajuy

aanSIy
7] JoyIUIsuURLL
pue JaAtooa1 Surpyingsysojq payutday)
wou ([g¢]
 jauueyo
puegeseg
|eUUBYDO
jeoisAyd jeuueyy

CuUUsUY-0O}-eULA}UY
OWA.
= ejel0siq ssejAuowey
jouueyD

VANSI
=F" JouULYD sINJOOIIYOIe
SNSIOA UONROTUNUIUIOD
Wa}SKs ainqoaytyose poyutidoy)
Wor ([g¢]
Sec. 1.1 Digital Communication System Functional Architecture 7

This pyramid summarizes the relationships between the theory needed to perform analysis
and/or system simulation at various channel levels in the overall architecture. This figure
demonstrates that the lower the level to be worked, the larger the knowledge base required
from a system engineering viewpoint.

1.1.2 Uncoded Digital Communication System Architecture

In considering the more detailed breakdown of the architecture at the lower levels in the
hierarchy, it is convenient to break this into two components, viz., (1) the functional archi-
tecture of an uncoded digital communication system and (2) the functional architecture of
a coded digital communication system. Chapters 3 through 10 will be concerned with the
theory and techniques associated with the design and optimization of an Uncoded Digital
Communication System; the performance criterion will be “bit error probability.” Chap-
ters 10 through 13 will provide the necessary interface that will lead to presenting error
control coding techniques associated with the design and optimization of a Coded Digital
Communication System; the performance criteria will be the “Ro criterion” and bit error
probability.
Figure 1.5 demonstrates the Information Transportation System (ITS) Pyramid for the
architecture of an uncoded digital communications system. At each level in the architecture,
there exists a channel that serves to characterize and model the physical features associated
with the signal flow path below it. This channel hierarchy serves to define (one-to-one)
the elements associated with the channel layers illustrated in Fig. 1.3. The frequency up
converter present in practice between the IF and RF transmitter and the down converter
between the RF receiver and IF receiver (see Fig. 1.2) have been included respectively in
the RF transmitter and receiver building blocks. In this chapter, we are primarily concerned
with modeling and characterizing the power transfer and channel effects in channel layers 5
and 6.

1.1.3 Coded Digital Communication System Architecture

When information source and error control coding techniques are used to mitigate the dele-
terious effects of channel interference and memory on communication transmission, the
functional architecture changes. Figure 1.6 demonstrates the /7S Pyramid when source and
error control coding techniques are assumed. Comparing the pyramids in Figures 1.5 and
1.6, we see that a source encoder and a channel encoder are required on the transmit side
and a channel decoder and a source decoder are required on the receive side. In practice,
these elements are frequently referred to as the source codec and the channel codec. The
informed reader will note that the interleaver and deinterleaver functions demonstrated in
Fig. 1.2 have been omitted in Fig. 1.6; these functions, designed to alleviate channel mem-
ory effects, are assumed to be incorporated in the channel coder and decoder of Fig. 1.6,
respectively. The details of these functions are covered in Chapters 11 through 13 from a
system design and engineering perspective. As before, the channel at any level of the coded
system architecture serves to incorporate the features of the channel below it with regard to
any deleterious effects on signal (information) flow.
The ITS Pyramids serve to demonstrate the feature of the signal (information) flow
path (the communication channel) that constitute hardware versus those that constitute the
communication medium (that is, cable, waveguide, atmospheric, and so on.) In essence,
the ITS Pyramids serve to answer the question: “What functional elements constitute the
 ([g¢] wor pojutiday) pruresAd uonvonpe Bulisaursue suiajsks UONeOIUNWIWOD = pT dAINSI
uolediunwWwoD
seubewospaja|y
Aioay|
voyesiunwwog
Aoay,
Buipog
Buipog Koay

‘uonewojul

‘UOoITeWO}U}
WUOjU}
BR

‘uolyewsoju|
Buipoa
pue

pue

pue
|uonesiunwwog
Aioay pue
|seuBewose|3
Aioay

‘UOIeWJOJU|‘UO!TeWJO}U}
‘UONEIUNWWOD
‘BuUBlUY
‘uoNeduNWWOD
Buuajuy

uoneBedoig
soiskyd
aeauBewos9e/3
Aioay|
pue
 euvey)he]
g

BUUEUDv9A27
p

puueussafe)
|

sutaysXs Jsaursuy aAQdadsiag

ans
S*T
= UONULIOJUT uoneyodsuesy,
waysks(Sif) ‘prureshd papooun[eu8Ip
UONeoTUNUIUIOD
Waishs [euoNUNy oinjooytyose pajutiday)
wor ([¢]
uy) J0]09
uo yur pus (1oded
 (saded pus yoeq uo 10[09 uy)
([g¢] wos poyutidoy) ainjooy1yose [euONoUN] W9\sAs UONROTUNUIWIOD
[eusIp pepoo ‘prureshd (SJ) waiIsks uoyeyodsuvsy] UONeUIOJUT QT ans
aatpadsiag Joautsuq suiayshS
Sec. 1.1 Digital Communication System Functional Architecture 11

Source Information

ee
Ee ee
ane ieriveneend| Se
ite pyencrypll ln) unas
(Seer
(RC
eeoe
ey ae
Ret |
Pr pate 5 paces |\inih
{ eneMansmion oped 4)
ay he eae pps Pa ee
|
|
|
i a [REmitTranemit
Antenna
J} Chann
Moderel
| Trans
Antenna-to- _
| RF Channel
Antenna Coded
IF Channel | Channel! Channel
|
aE pReceNE ee
ence en

|
[IF
Receive |
[—Demodulate
| __ —
Deinterleave
| J
4
\
ieaarer
en
fa
espe
a = eee
User Information

Figure 1.7. Simplex communication channel architecture

communication channel from the theoretical perspective?” The digital communication sys-
tem building blocks elements which define the channel, the signal processing functions, the
building blocks and hardware/software interfaces are integrated in the communications pipe
illustrated in Fig. 1.7. It is suggested that the reader carefully examines Fig. 1.7 to under-
stand this important communications pipe.

1.1.4 ISO-OSI Network Architecture

In digital communication networks, data and voice can be sent from one point to another in
one of two fashions. A circuit connection can be set up in the network between the source
and the user. Then the message is transmitted, and when it finishes, the circuit connection
is deleted. Another method for communication is to break up the message into smaller
packets and transmit the packets through the network (possibly through different paths) to
the destination, and then reconstruct the message at the user site. The latter technique results
in packet-switched data networks while the former technique results in circuit-switched
networks.
Figure 1.8 demonstrates the seven layers that make up the ISO-OSI Network Archi-
tecture for packet-switched networks and shows how the architecture demonstrated by the
ITS Pyramids fits into the Physical Layer (Layer 1) of the ISO-OSI Architecture: Open Sys-
tem Interconnection (OSI) network architecture is an approach to interprocessor commu-
nications developed by the International Organization for Standardization (ISO) and is the
focus of that research effort. For details of this important and detailed standard, the reader
is referred to the basic material contained in [4—10]. The functions of each of the layers are
summarized in the following list.
[pour sNJOaITYOIe YIOMJOU [TSO-OS] g’T ans
amypa}yIY YIOMION

sake
eyeg
YIOMIONyur
Jake
4eAe7 yodsuel]

12
Sec. 1.1 Digital Communication System Functional Architecture 13

1. The Physical Layer: The architecture of this layer is presented in Figs. 1.5 and 1.6. It
is further detailed in Fig. 1.7. This layer is the physical communication medium that
supports the transmission of bits and is concerned with the modulation and demodu-
lation process and functions such as synchronization and framing.
2. The Data-Link Layer: This layer provides the basic link frame, including informa-
tion packet, start and end headers, and an error check field. It ensures that packets are
reliably transmitted through the network and it shields higher layers from any con-
cerns about the physical transmission medium. The service provided by the data link
layer is the reliable transfer of a packet over a single link in the network. This is ac-
complished through positive and negative acknowledgments for error checking in the
link.
3. The Network Layer: This layer provides the networkwide (or internetwork) services
required to transmit, route, and relay data between computers (or nodes) without
regard for the communications medium. This layer generates (at the source) and reads
(at the destination) the address information and control fields for purposes of flow
control. It is responsible for establishing, maintaining, and terminating the network
connection between two users and for transferring data along that connection.
4. The Transport Layer: This layer is responsible for providing data transfer between
two users at an agreed on quality of service (QOS). When a connection is established
between two users, the layer is responsible for selecting a particular class of service,
to monitor transmissions to ensure the appropriate service quality is maintained, and
to notify the users if it is not.
5. The Session Layer: This layer provides dialogue services for those functions that re-
quire the establishment of connection and synchronization between any two machines
prior to the exchange of information. This layer provides the “Are you there”—“Yes
I am” exchange prior to the exchange of application data. A primary concern of the
session layer is controlling when users can send and receive, based on whether they
can send concurrently or alternately.
6. The Presentation Layer: This layer provides a common representation of application
data that is communicated between application entities. Common representation refers
to the encoding of data and the order of bits and bytes exchanged between processors.
This may include character code translation, data conversion, or data compression and
expansion. For example, the exchange of data between a processor using American
Standard Code for Information Interchange (ASCII) encoded characters and a proc-
essor using Extended Binary Coded Decimal Interchange Code (EBCDIC) encoded
characters requires a data translation before the information can be utilized.
7. The Application Layer: This layer provides the user program with an interface to an
OSI system. In this case, the user is any computer program or application program that
requires interprocessor communications. The layer includes services used to establish
and terminate the connections between users and to monitor and manage the systems
being interconnected and the various resources they employ. A number of common
communication functions have been identified and grouped into so-called application
entities with standard services and protocols. The application entities are the key to
open systems.
 14 Introduction to Telecommunications Chap. 1

1.2 TELECOMMUNICATION NETWORKS AND SERVICES

Telecommunication networks are used extensively in the modern world to interconnect ge-
ographically separated nodes or “terminals” for the purpose of providing a specific service.
The elements of a network can be spread over a relatively small area, such as a building or
a university campus, or over cities or even continents. The elements or nodes themselves
can be stationary or dynamic depending on the application. The number of nodes typically
ranges from a few elements to a few thousand or more. Each node can consist of a simple
telephone or a computer terminal to a more elaborate “receiver” or signal processing cen-
ter. The network performs a function or provides a service from which its name is typically
derived. Basically, any interconnected set of elements that communicate with each other in
some manner qualify as a telecommunication network. A very simple network can consist
of a set of two personal computers that communicate with each other to share a common
data base. It is worthwhile pointing out at this time the technology forecast made in 1945
by Arthur C. Clarke [11] about using rocket stations to provide worldwide radio coverage.
It ranks among the most accurate and history-changing of all technology forecasts in mod-
ern times. The majority of current networks rely on satellites or “rocket stations” to com-
municate the information necessary to provide their respective services. Examples of such
networks are the National Aeronautics Space Administration’s (NASA’s) Deep Space Net-
work, the Telecommunication and Data Relay Satellite System, and the Global Positioning
System.

1.2.1 Deep Space Network (DSN)

The Deep Space Network (DSN) [12, 13] is the largest, most sensitive scientific telecom-
munications and radio navigation network in the world. Its principal responsibilities are to
support automated interplanetary spacecraft missions and radio and radar astronomy ob-
servations in the exploration of the solar system and the universe. The DSN also supports
high-Earth orbiters, lunar and Space Shuttle missions.
The DSN is managed, technically directed, and operated for the National Aeronautics
and Space Administration (NASA) by the Jet Propulsion Laboratory (JPL), California Insti-
tute of Technology, Pasadena, California. The DSN is a collection of three communication
complexes located in California (USA), Madrid (Spain), and Canberra (Australia) that sup-
port deep space missions. Each complex consists of basically three or more “large” antennas
that receive and transmit commands or data to the space probes. The location of the com-
plexes was chosen to provide global support and maximize coverage availability. Figure 1.9
depicts the DSN’s largest antenna, which is 70 meters in diameter. Each node of this net-
work consists of either a complete communication complex or a deep space spacecraft. The
network is nonstationary in the sense that the relative distance between some of the nodes is
changing as a function of time. The use of large antennas in addition to supercooled front-
end receivers has enabled the DSN to receive some of the weakest signals (—185 dBm) used
in communication systems.

1.2.2 Tracking and Data Relay Satellite System (TDRSS)

The Tracking and Data Relay Satellite System (TDRSS) [14] consists of a space network
(SN) and a ground network (GN) and provides two-way relay of communications between
a user spacecraft and the ground terminal located at White Sands, New Mexico. The system
has evolved in two phases, TDRSS-I and TDRSS-II. In TDRSS-I, the two operational space
 Sec. 1.2 Telecommunication Networks and Services 15

Figure 1.9 The Deep Space Network

(DSN) 70-meter antenna (Courtesy of Jet
Propulsion Laboratory)

segment elements, TDRS-East and TDRS-West, are shown in Fig. 1.10a and are located
respectively at about 41° and 171° west longitude in geosynchronous orbits. A spare satellite
(TDRS-Spare), located at 61° west longitude, can be repositioned at either location in case
of a satellite malfunction. The SN provides tracking and data acquisition with more than
85% coverage for orbit altitudes of 200 to 1,200 km, and 100% coverage for orbit altitudes
of 1,200 to 12,100 km. Return data rates of up to 300 Mb/sec from user spacecraft are
supported, as are a number of different services. In TDRSS-II depicted in Fig. 1.10b, the
Zone of Exclusion (ZOE) is removed using the Out-Of-View (OOV) TDRS-II. Return data
rates of up to 600 Mbits/sec can be supported. TDRSS supports the primary communication
links to the space shuttle, as illustrated in Fig. 1.11.

1.2.3 Global Positioning System (GPS)

The third network to be discussed is NAVSTAR’s Global Positioning System (GPS) [15,
16]. Its purpose is to provide three-dimensional position and velocity information as well
as time transfer to users anywhere in the world at any time. The position determination
is based on measuring the delay of radio frequency (RF) signals from four satellites of
a total constellation of 24. A minimum of four measurements are necessary to solve for
the four unknowns—three-dimensional position in addition to the user’s clock offset. The
24 satellites are shown in Fig. 1.12 and operate in 12-hour orbits at an altitude of 20,183
km. The constellation provides visibility of six to eleven satellites at five degrees or more
above the horizon to users located anywhere in the world at any time. The four satellites
offering the best geometry can be selected automatically by the receivers using ephemeris
 16 Introduction to Telecommunications Chap. 1

ANTENNA TDRS-WEST
TDRS-EAST TDRSS GROUND STEERING LONG. 171° WEST
LONG.41° WEST TERMINAL ANGLE
LAT. 32° NORTH (225° EW A,
LONG. 107° WEST 31° N/S)

Se SEs
S. SPACECRAFT
=

ZONE OF EXCLUSION
(FOR USER
SPACECRAFT HEIGHT
OF 1200 km OR LESS)

Figure 1.10a Tracking and Data Relay Satellite System-I (TDRSS-I)

information transmitted by the satellites. Accuracies on the order of 10 meters in position

are anticipated using commercial receivers and | meter or less using military receivers.
Comparable position location accuracies can also be achieved using the GLONASS system,
which has been designed and implemented by the former Soviet Union.
GPS has revolutionized position location and navigation in this century. For example,
all commercial jets are being equipped with GPS receivers to allow for more efficient navi-
gation and routing between two cities. GPS is also allowing for smarter highways and trans-
portation systems where cars get information on traffic and adaptive routing is employed to
minimize congestion [17].

1.2.4 Integrated Services Digital Network (ISDN)

The Integrated Services Digital Network (ISDN) and its broadband version have been un-
der development for several years and are currently undergoing field tests [18,19]. The three
fundamental principles embodied in ISDN and broadband ISDN are digitization with high
bandwidth, worldwide standards, and integrated functionality and services. ISDN will fi-
nally integrate voice and data services enabling novel applications limited only by the hu-
man imagination [20]; for example, personalized newspapers in which an individual, using
his home “terminal,” can retrieve the sports section from the Los Angeles Times, the food
section from the New York Times, the world news from the Christian Science Monitor, and
so on, and print all sections in the form of one “personal” newspaper. Another example is
the cost-effective home office in which an employee can communicate with his employers
through voice, facsimile, and video simultaneously, thus enabling him or her to hold meet-
ings with fellow employees also located at home or in different company locations.

1.2.5 Military Satellite Communications

In the military domain, the next generation satellite communications (MILSATCOM) sys-
tem, designed to serve both the nation’s strategic and tactical forces, is currently in the
 Sec. 1.2 Telecommunication Networks and Services 17

46°W
rr Spare TORS RS | Il 171°W
.
“<n
174°W

Nominal
In-View Relay (IVR) In-View
TDRS Il Non-Relay (IVNR)
TDRSII

7443°E

Out-of-View (OOV)
TDRS Il

Figure 1.10b Tracking and Data Relay Satellite System-II (TDRSS-ID)

deployment phase. The new system, called MILSTAR [21] and depicted in Fig. 1.13, will
provide a worldwide, highly jam-resistant, survivable and enduring MILSATCOM capa-
bility under conditions of both electronic warfare and physical attack. Through the use of
extremely high frequency (EHF) and other advanced techniques, the system will recover
very quickly from the propagation degradation caused by a high-altitude nuclear detona-
tion. MILSTAR satellite crosslinks will operate in the W-band (60 GHz) to prevent eaves-
dropping and interference from enemy ground stations. They will also enable message relay
around the world above the atmosphere without reliance on intermediate downlinks to vul-
nerabie ground stations. This capability would ensure, for instance, that a presidential jet
flying over the Pacific would be able to communicate with a tactical nuclear force in Europe
even if all ground stations had been destroyed.

1.2.6 Mobile and Personal Communications

Satellite systems for mobile and personal communications and their interconnection with
the Public Switched Telephone Network (PSTN) will no doubt be the next layer of telecom-
munication infrastructure developed for purposes of global communications and navigation.
Mobile satellite communication systems (MSCS) are capable of delivering a range of ser-
vices to a wide variety of terminal types. The architecture of a future MSCS is illustrated in
Fig. 1.14a with objectives depicted in Fig. 1.14b; namely, communication from one hand-
held terminal to another hand-held terminal anywhere in the world. Examples of mobile
satellite terminal platforms include land vehicles, aircraft, marine vessels, remote data col-
lection and control sites. Additionally, services can be provided to portable terminals, which
are currently the size of a briefcase and no doubt will be reduced to hand-held sizes for fu-
ture systems. Many future services will include position reporting to determine the position
 Introduction to Telecommunications Chap. 1
18

SHUTTLE

GROUND STATION
WHITE SANDS, N.M.

Figure 1.11 TDRSS/User spacecraft configuration

of the mobile user, with transmission of information via the mobile satellite communications
system. In the future, Mobile Satellite Services and Radio Determination Satellite Services
(RDSS) will provide the ultimate in Global Personal Communications when interoperated
with the Public Telephone Switched Network. Ultimately, mobile satellite systems should be
capable of providing personal communications services (PCS) such as voice and data (low-
rate) to very small terminals including hand-held units. Applications requiring much higher
data rates should be possible with somewhat larger mobile and portable terminals. Mobility
and high capacity are the major new requirements that these future systems must satisfy.
At least three approaches have been proposed to the Federal Communication Com-
mission (FCC). These include different network architectures consisting of satellite constel-
lations in low Earth orbits (LEO), medium Earth orbits (MEO), and geosynchronous orbits
(GEO). In addition, Europe is considering satellite constellations in highly eccentric (ellip-
tic) orbits (HEO). These various orbits are depicted in Fig. 1.15. Presently, the International
Maritime Satellite (Inmarsat) [22] Organization and the American Mobile Satellite Corpo-
ration (AMSC) systems utilize GEO. 7
Currently, there are several proposed systems: (1) Motorola’s proposed/ridium [23]
and Loral/Qualcomm’s proposed GlobalStar [24] consisting of a satellite constellation in
LEO, (2) TRW’s proposed Odyssey [25] consisting of a satellite constellation in MEO, and
Sec. 1.2 Telecommunication Networks and Services 19

Figure 1.12 The Global Positioning

System (GPS) orbit configuration

(3) Europe’s proposed Archimedes [26] consisting of satellites in HEO. It is clear that the
multinational development of global wireless communications will be a ubiquitous reality
in the twenty-first century. Table 1.1 summarizes the characteristics of the proposed systems
in terms of number of satellites, number of orbit planes, orbit altitude, and so on.
The final system/network architecture associated with this wireless world is difficult
to speculate on at this time; however, it seems safe to predict that before the year 2000, small
shirt-pocket size, low-weight, full-featured personal communication hand-held terminals
will be available with long talk time and high quality voice, data and position location
capabilities that can be used anywhere in the world—either at home, in the office, on
the street, or while on travel with either terrestrial and/or satellite facilities. This personal
communications network (PCN) will be implemented such that any of the network facilities
will be totally transparent to the end user. In the next few sections, we will give a brief
overview of several systems proposed as the infrastructure for future generation wireless,
mobile, personal communications.

Inmarsat. Inmarsat [22] is currently the world’s largest mobile satellite commu-
nications organization, backed by more than 60 national telecommunications enterprises,
as well as a worldwide network of equipment manufacturers, distributors, and service
providers. Established in 1979, Inmarsat has its headquarters in London and has 64 member
countries as of May 1992.
The worldwide services that are currently available through Inmarsat include: direct-
dial telephone, facsimile, telex, electronic mail, data transmission, flight deck voice and
data services, position reporting, and fleet management. In the near future, global paging,
briefcase telephones, and voice and data for the business traveler will be available. Inmarsat
currently uses a total of 10 satellites in its global system. While some of these satellites are
carrying commercial communications traffic, others are in orbit as backup spares to ensure
continuity of service. The satellites, as shown in Fig. 1.16, are arranged in geostationary
orbit to provide communications coverage of all areas of the globe with the exception
20 Introduction to Telecommunications Chap. 1

og ey
a

Figure 1.13 The MILSTAR system

of the extreme polar regions. To provide enough capacity to meet the huge expansion in
demand, Inmarsat has acquired Inmarsat-2 satellites for the 1990s. A new generation of
Inmarsat-3 satellites with dedicated spot beams is in production for launch in 1994—95, and
its characteristics are shown in Fig. 1.17.

Iridium. Motorola’s proposed Jridium [23] is a LEO satellite-based system de-

signed to provide global telecommunication services on a continuous basis with coverage of
all units from the North Pole to the South Pole. A system overview is depicted in Fig. 1.18.
Five types of basic service are presently planned, each with a separate terminal or family of
terminals:

|. Radio determination service and two-way messaging: All subscriber units, except
some pagers, will contain this capability, which includes automatic location reporting.
Units with GPS or GLONASS services will also be available, which will provide
greatly increased geolocation speed and accuracy.
NM Digital voice communications service: The system will provide duplex high-quality
4.8 kb/s voice communications.
3. Paging: An alphanumeric pager for instantaneous direct-satellite paging in any region
of the world is being designed in a package only slightly larger than today’s pagers.
4. Facsimile: Two types of mobile facsimile units are planned, a stand-alone facsimile
unit and a unit to be used with an /ridium telephone.
5. Data: A 2.4 kb/s modem is being developed for use within the Jridium network.
 Figurell4a Mobile saciluc communication
sysiem achaccme

The backbone of the Iridium system is the constellation of 77 small satellites in po-
jar Low-Earth-Orbit (LEO) at an altitude of 420 nautical miles (778 km). This relatively
icate
further
low altitude assures that 2 user will never have to commun than about 1.250
nautical miles (2,315 km) to a satellite. This allows the use of lightweight low-power radio-
telepho nes
with small low-profile antennas compared to those required toutilize geostation-
ary saicMines at ranges of more than 19,000 nautical miles (35.188 km). The constellation
comprises seven planes of 11 satellites, each plane equally spaced in longitude. The orbital
period of each saseMite is approximately 100 minutes.
The satellites are interconnected via microwave crosslinks, thus forming a global
network in space, with linkage to the ground via gateways. Each gateway interfaces with
the local Public Switched Telephone Network (PSTN), providing local area user/customer
recon keeping and interconnection capability between Iridium and non-Iridium users. The
decision to locate a gateway in any given area can be made by local governments and/or
servers wilely out of a desire tooffer local service, rather than to satisfy a functiona need
l
by the system for thegatew to ay
be a node in the networ to k
assure global coverage. Satellite
telemetry, tacking and control, and overall system control will be accompanied by one or
more centrally located System Control Facilities utilizing earth terminals and microwave
finks.
The radio links between users and satellites will operate somewhere in the 1.600 MHz
“1 1,700 Mz region of L-Band, while the gateway feeder links and satellite crosslinks
rat
will opein e ofthe 18 GHz to 30 GHz region of K-Band. L-Band is best
small portions
22 Introduction to Telecommunications Chap. 1

Communicate from one hand-held terminal to another hand-held terminal

anywhere in the world

Figure 1.14b Mobile Satellite Communication System (MSCS) objective

suited for direct satellite-to-portable-user links because it offers a good mix of relatively
short wavelengths and low propagation losses, and it includes the upper frequency limits
of economical commercial hardware using available technologies. Each satellite employs
an L-Band phased array antenna complex that forms 37 interlaced spot beams (cells). Each
beam operates independently. Each carrier of each beam has its downlink transmit power
electronically controlled to maintain link quality for each user, allowing for variations in
range and blockage/foliage. Further design details can be found in [23].

Odyssey. TRW’s proposed Odyssey [25] is a MEO satellite system consisting of

(as few as) 12 satellites in MEO for personal communications as depicted in Fig. 1.19.
Odyssey will provide high-quality, wireless communications from satellites worldwide. The
telecommunication services include voice, data, paging, radio, determination (position and
navigation), and messaging. It will use Code Division Multiple Access (CDMA) and Fre-
quency Division Multiple Access (FDMA) techniques to achieve compatibility with other
systems.
The medium altitude at which Odyssey satellites will operate is 10,240 km, and they
will be capable of communicating with hand-held transceivers. The system’s space segment
consists of 12 satellites. Each satellite has two 19-beam, steerable antennas that are used to
communicate with the system users. The satellite/user links are implemented in the RDSS
Band. The satellites also communicate with the Earth stations, in Ka-band via a high gain
steerable antenna. In addition to providing the user with RDSS functions, the system pro-
vides access to the public switched network by use of digital voice and spread spectrum
techniques. The smallest hand-held user transceiver is proposed to use 0.5 Wattts of RF
transmit power and is capable of carrying a digital voice signal at 4.8 kbps. The satellites
operate as simple frequency translating transponders.
Sec. 1.2 Telecommunication Networks and Services 23

LEO: LOW EARTH ORBIT

(Circular: 900 —1,400 km Altitude
Elliptic: 400 —3,000 km Altitude Range)

MEO: MEDIUM EARTH ORBIT

(Circular: 10,000 km Altitude)

GEO: GEOSTATIONARY ORBIT

(Circular: 35,800 km Altitude Over
Equator)

HEO: HIGHLY ECCENTRIC ORBIT

(E.G.: 1,000 -26,555 km)

The Lower the Orbit the Faster its Period of

Rotation Around the Earth

Figure 1.15 Classification of satellite orbits

Table 1.1 Characteristics of Proposed Systems

Odyssey GlobalStar Iridium

Number of satellites 12 48 66
Number of orbit planes 3 8 6
Orbit inclination (deg) 55) By) 87
Orbit altitude (km) 10354 1389 765
Intersatellite links no no yes
Repeater type bent-pipe bent-pipe processing
Frequency Bands (GHz)
¢ User to satellite 1.6 1.6 1.6
* Satellite to user Ds pie 1.6
Multiple-access mode CDMA CDMA TDMA

High-elevation angle operation, that is, greater than 30°, reduces the possibility of
signal blockage for a mobile user. The use of steerable antennas permits the coverage of
selected areas of the Earth for extended periods of time, as the satellite passes overhead.
This significantly reduces the amount of beam-to-beam switching that must be done and
simplifies the overall system design. For RDSS purposes, the system employs a combination
of ranging measurements and Doppler measurements to determine the location of a user.
For the user/satellite links, the 16.5 MHz frequency bands are divided into three sub-
bands of 5.5 MHz each. Each beam, of the 19 beam user/satellite antennas, transmits or
receives in only one of the three sub-bands. The antennas are designed to have similar spot
beam sizes and beam positions in both the 1.6 GHz and 2.5 GHz bands. For satellite/Earth
station link, each of the 19 received or transmitted sub-bands are frequency shifted to a dif-
ferent frequency requiring 16.5 MHz total bandwidth. This value includes data channels,
 24 Introduction to Telecommunications Chap. 1

160 140 120 100 80 60 40 20 O 20 40 60 80 100 120 140 160 180

160 140 120 100 80 60 40 20 0 20 40 60 80 100 120 140 160 180

Figure 1.16 Inmarsat worldwide mobile communications network

order-wire channel, and guard bands. Because of the relatively high values of rain atten-
uation in the 20 and 30 GHz frequency bands, diversity Earth station techniques will be
employed. Further detailed discussion of the design and performance can be found in [25].

GlobalStar. Loral/Qualcomm’s proposed GlobalStar system, with characteristics

shown in Fig. 1:20, is a LEO satellite-based mobile communications system providing radio
determination, mobile voice, and data communications services. The GlobalStar system is
interoperable with the current and future Public Switched Telephone Network (PSTN) and
Public Land Mobile Networks (PLMN), such as cellular telephone systems, Personal Com-
munications Networks (PCNs), or Personal Communications Services (PCS). The Global-
Star system also is designed to interface with private networks and provide mobile services
to private network users.
The GlobalStar space segment consists of a constellation of 48 LEO satellites in
circular orbits with 750 nautical miles (1389 km) altitude to provide 100% global coverage.
In this constellation, there are eight orbital planes and each plane has six satellites, which
are equally phased within the orbital plane. Each orbital plane has an inclination angle
of 52 degrees. Each satellite has a 7.5 degree phase shift to the satellite in the adjacent
orbital plane. Initially, a constellation of 24 operating satellites will be launched to provide
maximum U.S. coverage. As worldwide traffic demand develops, 24 (or more) additional
satellites will be launched to provide continuous global coverage. Further design details can
be found in [24].

Archimedes. Archimedes is a proposed mobile satellite system conceived by the

European Space Agency (ESA) to effectively serve the European market for Mobile Radio
Services (MRS). Archimedes |26] was conceived to effectively serve the growing demands
for Mobile Radio Services (MRS). The communications systems architecture is very similar
to that for any GEO satellite communications system, the major differences being due to
Sec. 1.2 Telecommunication Networks and Services 25

Marecs A&B 2
Launch weight 1,006kg
Height 2.56m
Solar array span 13.8m
Type 3-axis stabilized
Capacity Approx. 75 two-way
voice circuits
Launch vehicle Ariane
First launch 1982
Manufacturer British Aerospace/Marconi

Inmarsat-2
Launch weight 1,230-1,320kg
Height 3.36m
Solar array span 15.23m
Type 3-axis stabilized
Capacity 250 two-way voice circuits
Launch vehicles Ariane/Delta ‘“
First launch 1990
Manufacturer British Aerospace consortium

intelsat-V
Launch weight 1,970kg
Height 6.58m
Solar array span 15.59m
Type 3-axis stabilized
Mobile Capacity 50 two-way voice circuits
Launch vehicles Atlas Centaur/Ariane
First launch 1982
Manufacturer Ford Aerospace and
Communications Corp.

Marisat
Launch weight 655kg
Height 3.81m
Diameter 2.61m
Type Spin stabilized
Capacity 10 two-way voice circuits
Launch vehicle Thor Detta
First launch 1976
Manufacturer Hughes Aircraft Co.

inmarsat-3
Scheduled for launch In 1995
Launch weight c.1,900kg
Height c.2m
Solar array span c.21m
Type 3-axis stabilized
Capacity 2,000 two-way voice circuits ¢ sa,
Launch vehicles = Atlas Centaur/Ariane/Proton a iy
Manufacturer Martin Marietta Astro Space

Figure 1.17 Characteristics of satellites used by the Inmarsat organization (Courtesy of Inmarsat)

the fact that the Archimedes satellites move with respect to a point on the Earth’s surface.
These differences are: the requirement to regularly hand over from one satellite to the next,
differential path length, and differential Doppler shift. These factors add to the complexity
of both fixed and mobile terminals as circuitry has to be built in to mitigate these effects.
Services are provided in much the same way as for a GEO satellite system. Broadcast-
ers uplink their channels via high frequency fixed feeder links to the satellite as shown in
26 Introduction to Telecommunications Chap. 1

Subscriber ;
Calls Subscriber
Calls

L-Band
Control/Status
K-Band

Figure 1.18a /ridium system overview

Fig. 1.21. The satellite then transmits the broadcast signal at the lower mobile communica-
tions frequencies to enable reception by mobiles throughout the coverage area. Figure 1.21
illustrates the communications system architecture showing various potential user types.
The overall coverage of the satellite system is Europe, but to reduce RF power per chan-
nel, it is advantageous to use high-gain spot beams. The formation of these spot beams is
intrinsically linked to the choice of frequency due to the laws of antenna geometry.
A critical parameter in determining the cost per channel of the Archimedes system is
the capacity of the system, in terms of the number of channels the satellite communication
system can simultaneously support. Satellites are limited in the amount of power that they
can supply to their communications payloads, and therefore, the capacity of the system is
directly proportional to the power requirement of each channel onboard the satellite. The
first generation Archimedes system is intended to serve Europe. Europe has a relatively high
latitude and therefore, the elevation angle to a GEO satellite is relatively low. The propa-
gation environment is very important for mobile communications and can be significantly
improved if higher elevation angles are given. This drives the consideration of alternative
orbits, particularly highly eccentric (elliptic) orbits (HEO). The increased elevation angles
given for HEO can be converted into decreased RF power requirements by considering prop-
agation margins. This is the most severe propagation environment Archimedes is intended to
be served by terrestrial retransmission. These measurements show the minimum advantage
of HEO over GEO to be 7 dB.
The mobile terminals are intended to be very simple and hence have nonsteerable
antennas. This means the satellite must be within the antenna beam at all times. For HEO,
 Sec. 1.3 Performance Criterion and Link Budgets 27

@ Supports Hand-Held & Moblie Users (4800 BPS Digital

Voice and Other Services)

© Constallation of 77 Satellites in 7 Polar Planes Provides

Global Coverage

®@ Low Altitude Enables Hand-Held Users—at 900 km Min.

Elevation Angle to Satellite is 10 deg.

@ Each Satellite Has a 37-Cell Terrestrial Antenna Pattern

Fixed Relative to Satellite but Moves Quickly Accross Earth

® Satellites Networked by Using Space Crosslinks

@ System is Mirror Image of Cellular Telephone System

@® TDMA/FDMA Combination
77 Processing Satellites
@ L-Band for Both Up- & Down-Link from User
488 KG; 1429 W
®@ Complex Networking

Figure 1.18b /ridium characteristics

the satellite is always within a cone directly overhead and therefore a highly directional
antenna can be used, whereas for GEO, the satellite can be anywhere within a fairly large
toroid and therefore, a low-gain toroidal coverage antenna is needed. For further details of
the proposed system, see [26].
There are too many telecommunication networks proposed or in existence to enumer-
ate or discuss, but the examples described above are representative of the majority. In con-
clusion, a network is an interconnected set of “elements” or “terminals” that communicate
with each other to provide a particular service.

1.3 PERFORMANCE CRITERION AND LINK BUDGETS

Characterizing the performance of a digital communication system is a key step for compar-
ison of different systems and techniques. In this section, we first define power, bandwidth,
and their relation to power spectral density. Then, the various elements and design parame-
ters of a link are examined and unified in the link budget equation, followed by a discussion
of signal power-to-noise ratio as a performance criterion.

1.3.1 Relation between Power and Bandwidth

Both power and bandwidth play a key role in system design and analysis. Often, they
must be traded off against each other. The average power of a wide-sense stationary (WSS)
random process x(t) is given by
+00

P,=E {370 |= R,,(0) = / San (1.1)

—0o

where R,xx(t) is the autocorrelation function! of the random process and S;x(f) its re-
spective Power Spectral Density (PSD) [27] in units of Watts/Hz. By replacing statistical

1. The autocorrelation function R,,(t) and its corresponding power spectral density S,,.(f) are often denoted
using one subscript, such as R,(t) and S,(f).
28 Introduction to Telecommunications Chap. 1

Figure 1.19 Odyssey satellite constellation

averaging with time averaging, the above definition can be extended to include ergodic pro-
cesses. Although the power of a waveform can be defined precisely, the notion of waveform
bandwidth is difficult to define and is sometimes vague and misleading. The basic objective
for specifying bandwidth is to indicate the frequency range over which most of the wave-
form power is distributed. For random processes,” bandwidth can be derived from the PSD
Syx(f) which is the Fourier transform of the autocorrelation function R,,(t). The most
commonly used bandwidths (in Hertz) are the “3-dB bandwidth,” the “p% fractional power
bandwidth,” the “Gabor or root-mean-square bandwidth,” and the “equivalent noise band-
width,” defined as follows (one-sided definitions):*

1. 3-dB Bandwidth (B,_ qp): Most useful for unipodal PSDs, it is the width along the
positive frequency axis between the frequencies where S,,(f) has dropped to half its
maximum value, that is,

2. When discussing bandwidth of a filter H(f), replace the PSD S,,(f) with the magnitude squared of the filter
transfer function, that is, |H(f)|*, and all subsequent equations are valid.
3. We assume that the PSD is symmetric, that is, S..(f) = Sxx(—/f).
 Sec. 1.3 Performance Criterion and Link Budgets 29

* Relies technically on Qualcomm's Cellular CDMA

¢ 48 Bentpipe Satellite Architecture that maximizes use of

terrestrial infrastructures (mobile and fixed)

* 6 satellites in each of 8 52-deg. inclined orbits at 1,389 km;

Bol Mass = 263 kg; Power = 875 w

* 6 isoflux spot beams at L-& S Bands with frequency

reuse of 2

* Services include voice, data, & RDSS to hand-held & mobile

¢ CDMA with 1.25 MCPS spread spectrum

e Variable voice coding rate: 1.2 - 9.6 KBPS

¢ Minimum elevation angle: 5 deg.

¢ Rake reception for coherent combining of simultaneous

signals through 2 or 3 satellites Figure 1.20 GlobalStar overview

uy
Sixx (B3_ dB) == 5 Sxx max (1.2)

This definition has the advantage that it can be computed directly from the spectral
density function. It has the disadvantage that it may be misleading for spectral densi-
ties with slowly decreasing tails and it may become ambiguous and nonunique for
spectral densities with multiple peaks, that is, there may be several frequencies at
which the spectral density is half its maximum value.
. p% fractional Bandwidth (By): It is the frequency width along the positive frequency
axis within which p% of the one-sided spectrum lies, that 1s,

(Pe Sxx(f)df we
(1.3)
Spare 100
where f2 — fi;= Bp. This definition has the advantage of being unique. However, it
can be misleading for spectral densities with somehow slowly decaying tails. Also,
it has the disadvantage that it must be evaluated by integration of the power spectral
density.
. Gabor Bandwidth (B;ms): It is defined as the root-mean-square (rms) bandwidth of
the PSD, that is,
(oe) 1/2
Brms = {| PSec Dat
| (1.4)
0

This bandwidth is mostly used in radar systems and it also has the disadvantage of
being evaluated by integration.
. Equivalent Noise Bandwidth (B,,): It is the most common definition used in studies of
systems with stochastic inputs. It is defined as the width along the positive frequency
axis within which a flat spectral density gives a value of power equal to that of the
given power spectral density where the value of the flat spectral density corresponds
to the maximum of the given spectral density (see Fig. 1.22). In mathematical terms
 30 Introduction to Telecommunications Chap. —

Radio, Voice, and Data

Radio, Voice and Data

Radio
Station Voice and Data
Telephone
Radio Station

Radio Only
Satellite Control Centre

Figure 1.21 Archimedes communications architecture

jie ban. Can

Te,@)
)
B, = 2 (1.5)
Sxx Gi )max

Often, a two-sided noise bandwidth is also defined. In that case

+00 ‘
Jo Set Pat
Wa = 2B, = ~~. (1.6)
Sixx (f)max

An alternate interpretation of noise bandwidth is obtained by passing white noise n(t)

of two-sided spectral density level No/2 through a filter H(f). Then, the PSD of the
output process x(t) is

: No )
Sxx(f) = = |H(f) | (1:7)

with average power

Pee i SUC
(o.@)

/ SH) Pag
CO N in

ee)
(1.8a)
If we were to pass the white noise through another brick-wall filter with gain
Sec. 1.3 Performance Criterion and Link Budgets 31

Sxx (f)

ERR
Figure 1.22 The noise bandwidth

|H(f)|max and two-sided bandwidth W,,, then the output process y(t) has power
N
P= Was HDs max (1.8b)

For the two processes x(t) and y(t) to have identical powers, we require P, = Py
resulting in

ae
ae oc me 2B, (1.9)
he ae TG A Pe
In this context, W,, is the total (positive and negative) spectral width of an ideal (brick-
wall) filter whose amplitude characteristic is equal to the maximum of |H(f)| and
which yields the same power at its output as does H(f) when white noise is present
at the input. For any of these bandwidths, the fractional out-of-band power,. Pop, is
given by

Pipe foVine t SxAaf (1.10)

: Spar
Example 1.1
The transfer function of a single-pole Butterworth filter is given by

i i
A(f)= a+j2xf | Ap a
a

It is clear that

A (f )max = H(0) = 1

The 3-dB bandwidth is thus obtained by

2 1 1 a
| H(B3_ gp) "= 2 Se eaedh an
2n B 2
1+ (TR
iB )
32 Introduction to Telecommunications Chap. 1

On the other hand, the 90% bandwidth is the solution to

fo HAP af _
Ig? Hf) Pdf
which gives
Boo

and
_, (27 Boo 1 6.3la
tan Ij“
( F ))}=0.9=;
05 Boo=
90 an = 6.31B 3- dB

Finally, the noise bandwidth is evaluated using

5, Drape a
22]
Mi 2
| H(f) | df =atan eae ———
0 a 0 4

Therefore,
a= a
= ie eerily
Example 1.2
For an N-pole Butterworth filter, we have

Pek ON ee ry
1+ (24)
anf

and the 3-dB bandwidth is still given by a/27. However, the noise bandwidth becomes
o0 I
By =f) Fae
i aaa
a

and simplifies to

__@/2N)
7 sin(t/2N) 3-4B
As N gets larger, the noise bandwidth approaches the 3-dB bandwidth, as expected.

Fortunately, most of the waveform spectra encountered in typical communication sys-

tems have a single peak with relatively fast spectral roll-off. As such, all the preceding def-
initions yield approximately the same value and none of the above-mentioned ambiguities
occur. For a more elaborate treatment, Scholtz [28] discusses cases in which discrepancies
among the various definitions do occur and thus a precise bandwidth definition is essential.

1.3.2 Link Budgets

A key characteristic of a communication channel is the physical transmission media em-

ployed; this constitutes the physical channel (the media) in Figs. 1.5, 1.6 and 1.7. Depending
on the nature of the information and the propagation distance, a variety of transmission me-
dia can be employed ranging from a pair of wires to waveguides or even optical fibers. The
Sec. 1.3 Performance Criterion and Link Budgets 33

#
o) 3
TRANSMISSION
i
a DESIGNATION
a
5) APPLICATIONS
MEDIA > 9
= i
10° ULTRAVIOLET Ks
10
OPTICAL FIBERS z VISIBLE LIGHT OPTICAL COMMUNICATION
10°
é INFRARED 14]
es (1 micron)| a 10 1c

30-300 GHz
WAVEGUIDES, LINE-OF-SIGHT EXTREMELY HIGH fs 10!! RETR E HRe Gre ee
+ cm FREQUENCY (EHF) hed SYSTEMS
LINE-OF-SIGHT RELAYING, 3-30 GHz SATELLITE AND SPACE COMMUNICATION,
LINE-OF-SIGHT IONOSPHERE SUPERHIGH = 10!°| MICROWAVE RELAY, RADAR (AIRBORNE,
PENETRATION, WAVEGUIDES 10 em PREQUENCY (SHF) F APPROACH, SURVEILLANCE, AND WEATHER)
0.3-3 GHz TV (UHF), SPACE TELEMETRY
Peace ULTRAHIGH — 109 |RADAR, MILITARY SATELLITE
+ MELEREQUENCY (UHF) F3 |
COMMUNICATION
COAXIAL CABLES, SKYWAVE 30-300 MHz TV (VHF) AND FM, LAND
(IONOSPHERIC AND VERY HIGH — 108 |TRANSPORTATION (TAXIS, BUSES,
TROPOSPHERIC SCATTER) 10 ME FREQUENCY (VHF) RAILROADS), AIR TRAFFIC CONTROL
3-30 MHz , |BUSINESS, AMATEUR AND CITIZENS
HIGH FREQUENCY fy 10” | BAND, MILITARY COMMUNICATION,
COAXIAL CABLES, (HF) = MOBILE RADIO TELEPHONE
IONOSPHERIC REFLECTION 100 M Aas a
(SKY WAVE) MEDIUM 106 |AM BROADCASTING, AMATEUR,
FREQUENCY (MF) = MOBILE, PUBLIC SAFETY
eeee ee
30-300 kHz NAVIGATIONAL AIDS, RADIO BEACONS,
LOW FREQUENCY 10° |INDUSTRIAL (POWER LINE)
(LF) Ss COMMUNICATIONS
10 km
3-30 kH
VERY Low L104. |NAVIGATION, TELEPHONY, TELEGRAPHY,
WIRE PAIRS, SURFACE DUCTING FREQUENCY (VLF) FREQUENCY ANDEEE
EES
TIMINGEEEESTANDARDS
EEE ee eee
(GROUND WAVE) 100"km 0.3-3 kHz
VOICE FREQUENCY TELEPHONY, DATA TERMINALS
1000 km
(VF)
30-300 Hz
EXTREMELY LOW MACROWAVE SUBMARINE
yee
FREQUENCY (ELF) COMMUNICATION

Figure 1.23 Various frequency bands with typical uses and transmission media

transmitting antenna transforms the electric signal into an electromagnetic wave for prop-
agation while the receiving antenna performs the reciprocal process. Depending on its fre-
quency, the wave can take on various forms ranging from a star light, to a television broad-
cast or even a laser beam. The electromagnetic spectrum has been divided into different
bands which are assigned to serve specific types of communication as shown in Fig. 1.23.
As an example, television broadcast occupies either the 30-300 MHz range (VHF band) or
the 0.3-3 GHz range (UHF band).
In certain applications, such as terrestrial mobile communications, the radiated en-
ergy from one transmitting antenna may reach the receiving antenna by direct propagation
or by reflections by the ground or the sky as depicted in Fig. 1.24. This phenomenon is re-
ferred to as multipath, which indicates that the total received electromagnetic energy is the
weighted sum of “many” delayed versions of the transmitted signal where each component
is randomly attenuated and delayed according to the propagation path it took to the receiver.
Depending on the carrier frequency, sky reflections can result from ionospheric or tropo-
spheric scattering (reflection and refraction). The ionosphere is a collection of gases, atoms,
water droplets, pollutants, and so on, that absorb and reflect large quantities of radiated en-
ergy. It is located in the upper part of the atmosphere at an altitude of about 100-300 km
34 Introduction to Telecommunications Chap. 1

IONOSPHERE

IONOSPHERIC
“| WAVES

TROPOSPHERIC WAVE

100-300 km

TROPOSPHERE

REFLECTED WAVE
DIRECT WAVE
TRANSMITTER
ANTENNA

RECEIVING
ANTENNA

Figure 1.24 Multipath propagation in the troposhere and ionosphere

and is known to reflect signals in the VLF band. On the other hand, the troposphere is about
10 km from the earth’s surface and is known to reflect carriers in the SHF band.
The design of any telecommunication link makes use of the so-called communication
equation which contains all of the basic elements and design parameters inherent in the
communication system and makes possible the specification of the received signal power-
to-noise spectral density ratio (units of Hz) with a high degree of accuracy, viz.,
2
FE=PG, (=) LL pbalrs GEM)

Here, P, is the total received signal power (Watts), No the one-sided power spectral noise
density (Watts/Hz), P; the transmitted total power (Watts), G; the peak transmitting an-
tenna gain (unitless), 4 = c/f the carrier wavelength (meters) and c is the speed of light
(meters/sec), d the distance between transmitter and receiver (meters), Lg the pointing loss
(unitless), Lp the polarization loss (unitless), Lq the atmospheric loss (unitless), L, the an-
tenna surface tolerance loss (unitless), G, the peak receiving antenna gain (unitless), k Bolz-
mann’s constant (1.38 10-23 Watts/Hz-° K)and finally, 7° the receiver temperature (Kelvin,
°K). Each element of the “power budget equation” given in (1.11) is defined and discussed
here as to its effect on the system design.
A typical communication link is shown in Fig. 1.25 where various “virtual” channels
are identified. At the transmitter, each information bit or block of bits is transformed into
a unique baseband waveform by the baseband modulator. This signal is then upconverted
twice to IF and then again to RF. Following power amplification, the signal is fed to the
antenna for propagation. The physical channel is the propagation media between the trans-
mitting and the receiving antenna. It can be a waveguide or propagation through rain as
discussed earlier. Modeling the effects of the physical channel on the propagating electro-
magnetic field in different frequency bands and under various scenarios is a continuous and
Sec. 1.3 Performance Criterion and Link Budgets 35

ongoing task [29]. At the receiving antenna output, the RF signal undergoes filtering and
amplification using a low noise amplifier (LNA). The physics of these operations and their
equivalent set at the transmitter are referred to as the RF channel or subsystem. Some of
these operations are highly nonlinear and might introduce signal amplitude and phase dis-
tortion. Examples of such include Travelling Wave Tube (TWT) amplifiers [30] used prior
to transmission which can introduce severe AM/PM effects, depending on their operating
point. At the receiver, the output of the LNA is then downconverted to an appropriate IF for
further filtering and processing. Communication system engineers deal mainly with model-
ing the IF signal in both the transmitter and receiver, or commonly designated the JF channel
or subsystem. At the receiver, the signal model in the IF subsystem incorporates the various
effects of both the physical and the RF channel such as fading due to shadowing, cross-
polarization effects, AM/AM and AM/PM conversion in TWT amplifiers, amplitude and
phase scintillation due to plasmas, and so on. Typically, the IF signal is again downconverted
to baseband using inphase and quadrature references for further processing including bit de-
tection. In modern receivers, the latter stage is typically implemented digitally as long as the
data rates are well within the technology capabilities. The baseband signals or equivalently
the baseband channel or subsystem are equivalent to their IF counterparts and are dealt with
by communication system engineers interchangeably as the IF conversion to baseband con-
stitutes a mere frequency shift with no additional effects.
In this section, we concentrate on the link budget for the additive white Gaussian
noise (AWGN) channel. All effects in the physical channel are frequently modeled as power
degradation in the received signal. This assumption is valid for the majority of the commu-
nication links encountered in practice. Let P, denote the maximum power at the amplifier
output. Then, the actual transmitted power P; is equal to P,L. where L, denotes the circuit
coupling loss and is given by
4AZaZLr
Le (1.12)
Za Zi)?
Here, Z, and Z, denote the output impedance of the amplifier and the input impedance
of the antenna respectively. Only when the impedances are matched can maximum power
transfer occur (that is, P; = P,). The transmitted electromagnetic signal is characterized by
the Effective Isotropic Radiated Power (EIRP), defined by
EIRP = P,G;LoLy (PS)

where G; is the transmitting antenna gain in the receiver direction and Lo, L, the point-
ing and antenna surface tolerance losses, respectively. A rough surface can cause energy
scattering and loss of gain in the desired direction. The antenna surface tolerance loss is
approximately given by

eee ro), (1.14)

where o is the rms roughness of the antenna surface in wavelength dimensions [31]. For
a /A = 0.02, the degradation is about 0.27 dB.
Antenna gain is defined as the ratio of the capture area of the antenna to that of an
isotropic antenna which radiates the same power in all directions. The peak antenna gain G;
is given by
G;
= max 27(¢z,1) (1.15)
where g;(@-, $1) is the antenna gain as a function of both azimuth angle ¢, and elevation
36
4! TANNVHO ie

| =] peea ee Satl
ye
enaptn
eee
IWOISAHd TANNVHO |
| |
ONVd3Sva
| |
ONVsasve AWNoOIS|
OUL: ‘OURS ¥yOLVINGOW
ONvsasva
YOLVINGOW30

aansigq
sz°]T y poylfdus
[eIISIP UONROTUNUIWIOD
sks wo
Sec. 1.3 Performance Criterion and Link Budgets 37

Peak Gain, G

(a)

46 beamwidth
6 is the antenna 3 dB beamwidth -3 dB from
peak
Second First
sidelobe sidelobe

Peak Gain, G

Transmitting x)
antenna be

Antenna
mainlobe

Antenna (b)
Sidelobes

Figure 1.26 Antenna gain pattern

angle ¢ and G, is the maximum gain taken over all angles (see Fig. 1.26a). Assuming
perfect pointing, the gain of the transmitting antenna in the direction of the receiver would
be exactly G,. However, in more realistic scenarios, the antenna is not pointed perfectly
and as a result, the gain in the receiver direction is slightly less than the peak gain G,. That
degradation is accounted for in the pointing loss Lg. Another important parameter of an
antenna gain pattern is its “beamwidth,” which reflects the width of the gain in space as
illustrated in Fig. 1.26b. For a circular reflector of diameter d,, the maximum (peak) gain
G, and the half power beamwidth yp (radians) are given by
38 Introduction to Telecommunications Chap. 1

a\e
G; ~ pa (=) (1.16)

Xr
~ ela)
Me da./Pa
where fg is the antenna efficiency factor which is a function of the aperture field intensity
distribution over the dish. This pg is typically on the order of 55%. Note from (1.16) and
(1.17) that the larger the antenna or the higher the frequency, the higher is the gain and
the narrower the beamwidth. For a parabolic antenna with pointing error @-, the antenna
pointing loss Lg is approximately given by
Lg © 072 6(Ge/ 0)” (1.18)
which indicates an exponential decrease in gain as a function of pointing error $e. Typical
pointing errors are about 0.1° — 0.3° for geostationary satellites depending on the tracking
capabilities of the antenna. Spatial pointing is maintained in the course of a track by using
several techniques, one of which is conical scan in which the antenna forms small circles
around the target while measuring the received power to determine the “best” spatial coor-
dinates.
After propagating a distance d, the field flux density (fF D—in units of Watts/meter?)
at the receiver becomes
(ET RP)LE
F D(d) = (1.19)
4rd?
where L, is the atmospheric loss due to rainfall, ionized atoms in the ionosphere, and so
on. The atmospheric losses are caused by absorption and scattering of the electromagnetic
field by the various molecules in the atmosphere. The problem becomes more severe as
the wavelength of the carrier approaches the size of these molecules. For example, severe
absorption due to water vapor molecules occur at 22 GHz while oxygen absorption occurs
at about 60 GHz. Other effects include degradation due to rainfall which becomes severe
when the wavelength approaches the water drop size, which is dependent on the type of
rainfall. For example, at a rainfall rate of 10 mm/hr, the attenuation is about 0.04 dB/km at
6 GHz but increases to 0.25 dB/km at 11 GHz. That degradation is also a function of the
antenna elevation angle as the latter affects the path length of the propagation through the
rain [29].
Assuming a receiving antenna with area A, perpendicular to the direction of the
transmitter, the collected signal power becomes [32]
(EITRP)LgA;Lp
P, = FD(d)A;Lp = (1.20)
4nd?
where L, accounts for any polarization loss due to misalignment in the received field polar-
ization. Polarization is the orientation in space of the electromagnetic field and it is typically
linear or circular. In the first, the electromagnetic field is aligned in one planar direction,
whereas in the latter, it is radiated with two orthogonal components that when combined,
-appear to rotate on a circle or on an ellipse. In both cases, the receiving antenna must have
a matching orientation. Circular polarization has the advantage of not being affected by an
erroneous phase rotation in the electromagnetic field during propagation, unlike linear po-
larization, which suffers a degradation.
Sec. 1.3 Performance Criterion and Link Budgets 39

Using the reciprocity principle [32], the effective receiving antenna area can be ex-
pressed in terms of the antenna gain as

Ar = 7 8r($.,o})
Ae / /
(1.21)
where g,(¢,, ¢;) denotes the receiving antenna gain in the direction of the transmitter. Here
again, the receiving antenna is typically not pointed perfectly and the gain in the transmitter
direction might not be the peak gain G,. Using (1.21) in (1.20), we have

P, = P,G,;LsLoLpLal,G, (1222)

where L, denotes the space loss and is given by (A/4zrd)*. Both coupling as well as pointing
losses can also occur in the receiver and need to be accounted for in (1.22).
The parameter No in (1.11) represents the normalized noise power spectral density in
Watts per Hertz (W/Hz). This factor is usually referenced to the input stage of the receiver
since the signal-to-noise ratio is established at that point. It is given by

No =kT° (1.23)
where k is Boltzmann’s constant and T° the system noise temperature in Kelvin (°K). This
equivalent system noise temperature can be expressed as

jad Ped Be (1.24)

where 7, is the background noise temperature and 7, the temperature of the receiver
electronics. The latter is equivalently given in terms of the receiver noise figure F [33]

Ef Chey 290" (1:23)

The standard technique employed to measure the noise temperature of a system is to com-
pare, in a fixed bandwidth B, a measurement of system noise power P, with the noise figure
from a resistor at a known temperature. Since P, and B are related by P, = kT°B, the
unknown system noise temperature can be determined quite simply. The quantity B is the
single-sided noise bandwidth, and noise power is always measured with respect to it.
The link budget of (1.11) is often rewritten as

P, [EIRP G,
No
ee
k Jen 15]
ye EXel li |e
ce)
r

where the transmitter parameters are lumped in the first bracket, propagation parameters in
the second, and the receiver figure of merit in the third.
Example 1.3
Table 1.2 depicts a link budget for an RF communication link between Mars and Earth as
shown in Fig. 1.27. Two different link options are considered: a direct link and an indirect
link. The first option consists of communicating from Mars to a Mars relay satellite (MRS)
and then to Earth directly at a frequency of 8.4 GHz (X-band) with a data rate of 20 Mbits/sec
[Table 1.2(a)]. The second option is similar except that the link to Earth will be relayed through
a satellite (Earth relay satellite—ERS) and the carrier frequency is at 32.4 GHz (Ka-band) and
the data rate is reduced to 10 Mbits/sec [Table 1.2(b)]. Since Ka-band suffers large degradations
in bad weather, the ERS is essential in increasing link availability.
 Introduction to Telecommunications Chap. 1
40

EARTH RELAY
SATELLITE (ERS) MARS RELAY
SATELLITE (MRS)
Ka/W-BAND

X/Ka-BAND

Figure 1.27 |Mars-Earth RF communication link

1.3.3 Signal-to-Noise Ratio as a Performance Criterion

Any performance criterion should assess the capability of a communication system to pro-
duce at its output a desired signal, that is, it should relate the desired operation to the actual
operation in the presence of the channel disturbances.
One of the most widely used measures of performance, particularly in systems with
stochastic additive disturbances, is the signal-to-noise ratio (SNR). The scenario consists of
a desired waveform (baseband or modulated carrier) contaminated by an additive interfering
waveform. When this combined waveform appears at any point in the system, the SNR at
that point (typically at the output of the receiver) is defined as
SNR = Power in desired signal waveform
— : 2t,
Power in interfering waveform ( )

SNR is an indication of how much stronger the desired signal is than the interfering signal
at the point in question. It is common to express SNR in dB, that is,

(SNR)ag = 10logi9(SNR) = 10log,9(P) — 10 log;o(N) (1.28)

where P and W are the powers* of the desired signal s(t) and interference (noise) n(t),
respectively. Thus, SNR is useful as a performance criterion when the faithfulness of the
reconstructed waveshape is not the primary concern as in digital communication and radar
systems. In comparing two systems on the basis of SNR performance, one must compute
the SNR at the same point in both systems (typically, the receiver output where the final
reconstructed information waveform is to be extracted) when s(t) + n(t) is input to both.
The system with the higher SNR would be considered the more favorable.

4. From here on in the book, P denotes the received signal power and the subscript “r” will be omitted.
- Sec. 1.3 Performance Criterion and Link Budgets

Table 1.2(a) Mars Return Link Direct to Earth (20 Mbps at f = 8.4 GHz)

Frequency GHz 8.4

Data Rate Mbps 20.0

(1) Mars relay EIRP dBw 73.4

TX output power watts 200.0
Circuit losses dB 0.8
Antenna gain dBi SES
Diameter m 5.0
Efficiency % 70.0
Pointing loss dB 0.1
Pointing error degrees 0.04
(2) Space loss dB 282.4
Range AU 2S
(3) Clear sky loss dB 0.1
(4) Rain attenuation dB 0.0
(5) Polarization loss dB 0.2
(6) Ground G/T dB/K 60.7
Antenna gain dBi 73.6
Antenna diameter m 70.0
Antenna efficiency % 60.0
Pointing loss dB 0.1
Pointing error degrees 0.002
System noise temp K 19.2
Space background K 3.0
Atmospheric noise K 6.3
Maser noise temp K 10.0
Maser noise figure dB 0.15
(7) Boltzmann’s constant dBw/Hz-K —228.4
(8) P,/No dB-Hz 80.0
(9) Mars-to-MRS degradation dB 0.2
(10) Bandwidth dB-Hz 73.0
(11) E£,/Npo into demodulator dB 6.8
(12) Ground equipment degradation dB 1.5
(13) Required E,/No (with coding) dB 2.4
Required E;,/
No (without coding) dB 9.6
Coding gain dB Tks?
(14) Link margin dB 2.9

Often, we require that a system be designed to guarantee that the SNR performance
at a given point is above a specified level, often called threshold SNR. Then the ratio of the
actual SNR at that point to the threshold SNR is called the system or design margin at that
point, that is,

SNR actual
Design margin = (1.29)
SNR threshold
which is commonly expressed in decibels. Design margins are extremely important since
 42 Introduction to Telecommunications Chap. 1

Table 1.2(b) Mars Return Link through the Earth Relay Satellite
(10 Mbps at f = 32.4 GHz)

Frequency GHz 32.4

Data Rate Mbps. 10.0

(1) Mars-to-MRS degradation dB 0.2

(2) MRS EIRP dBw 81.0
TX output watts 90
Circuit losses dB 0.8
Antenna gain dBi 62.7
Diameter m 5.0
Efficiency % 65.0
Pointing loss 0.4
Pointing error degrees 0.04
(3) Space loss (2.5 AU) dB 294.2
(4) Polarization loss dB 0.2
(5) ERS G/T dB/K 62.2
Antenna gain dBi 76.0
Antenna diameter m 25.0
Antenna efficiency % 55.0
Pointing loss dB 0.2
Pointing error degrees 0.0025
System noise temp K 23.0
Space background temp K 30)
RX noise temp K 20.0
Circuit loss dB 0.1
LNA noise figure dB 0.25
(6) Boltzmann’s constant dBw/Hz-K —228.6
(7) P,/No at ERS dB-Hz TS
(8) ERS-to-Earth degradation dB 0.2
(9) P,/No at ground dB-Hz Tia
(10) Data rate dB-bps 70.0
(11) E,/No into demodulator dB Wolk
(12) Ground equipment degradation dB 1:5
(13) Required E;/No (with coding) dB 2.4
Required E,/No (without coding) dB 9.6
Coding gain dB The
(14) Link margin dB 30,

they indicate the amount by which a system can be tolerant to unexpected events or ad-
ditional degradation. The links of Example 1.3 have design margins of 2.9 and 3.2 dBs
respectively.

1.4 SHANNON'S CAPACITY THEOREM

It is important to note here the premise theorized by Shannon [34-36] (also see Chapter 12)
with regard to the efficiency of a communication system. Shannon shows that for the addi-
 Sec. 1.5 Digital Communication with Subcarriers and Data Formatting 43

tive white Gaussian noise channel, coding schemes exist for which it is possible to commu-
nicate data error free over the channel provided the system data rate Ry, (in bits per second)
is less than or equal to what he called the channel capacity C. If Rp > C, then error-free
transmission is not possible. The channel capacity C, in bits per second, is defined by
P

Ca Bt og2(1 1 ++ ——
NoB> if
(1.30)

where P represents the received signal data power in watts, B the one-sided channel band-
width in Hz, and No the one-sided PSD level of the noise. Now P/NoB can be written as
1% es Ppp Me PT,/No ns Ep
ht)
Nob eaNgBH o. 1By Ree > No
where 7) is the time duration of a data bit in seconds, that is, the inverse of the data rate,
6p = Rp/B is the ratio of the data rate to the one-sided channel noise bandwidth in bits/Hz
and E,/No = PT)/No is the energy per bit-to-noise spectral density ratio. Frequently, the
parameter E;,/No is referred to as the relative efficiency or communication efficiency of a
particular communication scheme.
Substituting (1.31) and B = R»/d, into (1.30) and assuming communication at chan-
nel capacity (R, = C), we get
E
5p = logy(1 + 5—) (1.32)
No
Equation (1.32) is plotted in Fig. 1.28 as a function of 1/5, and E,/No. Notice that as
the value of E,/No decreases, the transmission bandwidth necessary to achieve 100% effi-
ciency, operation at C, increases. It is therefore instructive to allow B to approach infinity in
(1.30) or equivalently let 5, go to zero in (1.32). Taking the limit gives (see also Chapter 12)
Ep _
In(2).=--1.6.0dB (1.33)
0
Consequently, coding schemes exist for which error-free transmission is possible in the
unconstrained bandwidth, additive, white Gaussian noise channel provided that E,/No >
—1.6 dB. Unfortunately, Shannon’s theorem does not specify the data encoding method to
accomplish that.
The relative efficiency of any practical communication system can be compared with
the most efficient on the basis of how closely it approaches the Shannon limit, (1.33), for a
given error probability. In the following chapters, we consider various modulation, demod-
ulation, coding, decoding, and data-detection methods that attempt to approach the limit
expressed by Shannon’s theorem. Typical conclusions taken from later studies of various
methods are compared with the Shannon limit in Fig. 1.28. The Shannon capacity is made
more rigorous in Chapter 12.

1.5 DIGITAL COMMUNICATION WITH SUBCARRIERS AND DATA FORMATTING

Let P denote the total power of an RF signal, whether at the receiver or at the transmitter,
then it is typically divided among the carrier and the data according to

Pg=(1 —Lm)P (1.34)

 Introduction to Telecommunications Chap. 1
44

NONCOHERENT
FSK (Pp = 10°)

COHERENT
PSk(Pe= 10P)
COHERENT
PSK (P2= 10%)
ne
w)2

SHANNON'S LIMIT
-1.6 dB (2n2)

1.0 10 100
grit BLS BANDWIDTH
6 %p\NFORMATION
BIT RATE
Figure 1.28 Relative efficiency versus the ratio of bandwidth to information bit rate

and

J
By I Lope (1.35)

where the parameter Lj, represents the transmitter modulation loss. Essentially, it specifies
the power division between the carrier and information sidebands resulting from a particular
modulation process. If the basic information carrier is a microwave, continuous-wave (CW)
sinusoidal voltage, the modulation can be impressed on the transmitted carrier by angle
modulation [phase (PM) or frequency (FM)], amplitude modulation (AM), or a simultane-
ous combination of the two. Analytically speaking, the instantaneous value of a modulated
RF carrier can be expressed by

s(t) = V¥2Pa(t) cos(@et + 0(t)) (1.36)

where a(t) characterizes the amplitude modulation, 6(t) characterizes the angle modulation,
Ww is the carrier radian frequency, and /2P denotes the peak carrier amplitude when a(t) is
unity. The modulation signals, the information to be transmitted, can be classified as digital,
analog or a combination of the two.
One typical telemetry modulation is to phase-shift-key the data onto a square-wave
subcarrier and then phase-modulate onto a sinusoidal carrier (PCM/PSK/PM) as depicted
in Fig. 1.29. However, an alternate scheme phase-modulates the data directly on the carrier.
In this latter case, the data being phase-shift keyed is called PCM/PM. In either case, the
modulated carrier can be represented mathematically by
 Sec. 1.5 Digital Communication with Subcarriers and Data Formatting 45

ANALOG
MODULATED
DATA DATA
SOURCES | ANALOGTO SYMBOLS SUBCARRIER | SUBCARRIER
DIGITAL CARRIER SurpuT
MULTIPLEXER PHASE PHASE
CONVERTER MODULATOR 8 MODULATOR

DIGITAL SOURCES

SUBCARRIER CARRIER
GENERATOR GENERATOR

POWER POWER
SPECTRUM SPECTRUM
CARRIER

BINARY WAVEFORM (DATA) MODULATED

en MIDEEANOS
ae SUBCARRIER

f f
SSS ean ) bye 0 fo-lse fo fotlec

A B Cc

Figure 1.29 Typical modulation spectrum involving subcarriers

K
s(t) = V2P sin(wet + YA; Dj(t)) (1.37)
a

where A; denotes modulation index (in degrees) for the ith “data” signals Was 2s poe
K the total number of data channels and D;(t) the normalized data sequence (in the
case of PCM/PM) or a normalized modulated square-wave subcarrier (in the case of
PCM/PSK/PM), that is,
dj(t), PCM/PM
POT |di(t)Sin(se,t), _PCM/PSK/PM ue
where d;(t) = +1 is the normalized data sequence, Sin(ws¢,t) = +1 the square-wave sub-
carrier, and Ws, the i‘h subcarrier frequency. Using trigometric identities with the fact that
Did hee aed. b= 15.2,2.2,,.h, (1.3/7) may be expanded for K = | as

s(t) = V2P cos Aq sin(wet) + V2P D,(t) sin Ay cos(@et) (1.39)

If 0° < A, < 90°, this phase-modulated carrier comprises a pilot tone (residual carrier) and
a Double-Sideband (DSB) modulated carrier. A system with A; < 90° is called a residual
carrier system. A system with A; = 90° is called a suppressed carrier system. A residual
carrier employs a phase-locked loop to track the pilot tone and provide a coherent carrier
phase reference for demodulating the DSB-modulated carrier. A Costas loop may be used
as a suppressed carrier receiver. For K = 2, (1.37) may be expanded as

s(t) = V2P{cos Aj cos A2 sin(w-t) + D)(t) sin Ay cos A2 cos(@et) (1.40)

+ Do(t) cos Ay sin Az cos(@ct) — D,(t) D2(t) sin A, sin A2 sin(@-t)}
46 Introduction to Telecommunications Chap. 1

When we have a single data channel, as in (1.39), the first term is the carrier, while
the second term is the data channel. Hence the modulation index A, has allocated the total
power P in the transmitted signal s(t) to the carrier and to the data channel, where the
carrier power and the data power are, respectively
P.= P cos”Ay (1.41)
P, = Psin*A, (1.42)
Comparing (1.41) to (1.35), it is clear that the modulation loss Lm is equal to cos A}.
When we have two data channels as in (1.40), the first term is the carrier component, the
second term is the modulated subcarrier D|(t), the third term is the modulated subcarrier
D»2(t) component, and the fourth term is the cross-modulation loss component. Hence, the
corresponding powers in the four components are allocated by both modulation indices Aj
and A> with

Lae cos” Ay cos” A? (1.43)

(i Sde sin? Ay cos” A? (1.44)

P42 = P cos” Aj sin* A2 (1.45)

Pm = P sin* A; sin* Ao (1.46)
Similar expressions may be carried out for the cases K > 3.
In some applications, the subcarrier is a sine-wave, in which case

Dj(t) = d;(t) sin(@s5¢;t) (1.47)

Using the identity

exp[jBsin(ot + ¢)1= S > Jn(B)expLin(ot + $)] (1.48)

we have from (1.37) for K = 1 (letting wsc = @s¢,)

s(t) = V2P sin(wet + Ajd)(t) sin @sct)

= V2P {cos(A] Sin @sct) SiN Wet + d,(t) sin(A, sin @sct) COS Wet}
[o.@)

= V2PJo(A1) sin@ct + V2P sin wet [>>

2zcan cx@nen)} (1.49)
n=1

CoO

+ V¥2Pd\(t) cos wet 2 2 Jon+1(A4) sin((2n + baxd|

n=0

where Jp (x) = (—1)"Jn(—x) and J,(x) is the n’” order Bessel function [37]. The first term
of (1.49) is the discrete carrier, the second term is other discrete components at @. + 2nas¢,
and the last term consists of suppressed carrier modulations centered at w. + (2n + l)ase.
The baseband data dj(t) can have different formats as illustrated in Fig 1.30. With
a Non-Return to Zero-Level (NRZ-L) data format, a logical “one” is represented by one
level and a logical “zero” by the other. With NRZ-M (Mark), a logical “one” is represented
by a change in level and a logical “zero” by no change. Two other formats, Biphase and
Miller, are also defined in Fig. 1.30. It is worthwhile noting that Bi-@-L is identical to NRZ-
L when the latter is modulated on a square-wave subcarrier with a cycle period equal to
 Problems 47

eae ube A ie DATA SEQUENCE

Ppl |
| | | | | NRZ-LEVEL (OR NRZ CHANGE)
NRZ-L CRS | "ONE" IS REPRESENTED BY ONE LEVEL
7 : "ZERO" IS REPRESENTED BY THE OTHER LEVEL
|_| | NAZ-MARK (DIFFERENTIAL ENCODING)
Cet | || || | "ONE* IS REPRESENTED BY A CHANGE IN LEVEL
ad Oe Peal | "ZERO" IS REPRESENTED BY NO CHANGE IN LEVEL
| NRZ-SPACE (DIFFERENTIAL ENCODING)
NRZ-S | | "ONE" 1S REPRESENTED BY NO CHANGE IN LEVEL
| — | | "ZERO" 1S REPRESENTED BY A CHANGE IN LEVEL

"ONE" IS REPRESENTED BY A HALF-SYMBOL WIDE PULSE

“ZERO” IS REPRESENTED BY A NO PULSE CONDITION
|
| BIL-PHASE-LEVEL (SPLIT PHASE OR MANCHESTER CODE)
| “ONE” IS REPRESENTED BY A HALF-SYMBOL WIDE PULSE OF
| ONE POLARITY
Bio-L "ZERO" IS REPRESENTED BY A HALF-SYMBOL WIDE PULSE OF
| THE OPPOSITE POLARITY
BI-PHASE-MARK
Bio-M |A TRANSITION OCCURS AT THE BEGINNING OF EVERY SYMBOL
| PERIOD. “ONE” IS REPRESENTED BY A SECOND TRANSITION
ONE-HALF SYMBOL PERIOD LATER
| "ZERO" IS REPRESENTED BY NO SECOND TRANSITION
|
BI-PHASE-SPACE
| A TRANSITION OCCURS AT THE BEGINNING OF EVERY SYMBOL
PERIOD. "ONE" 1S REPRESENTED BY NO SECOND TRANSITION
Bio-S | "ZERO" IS REPRESENTED BY A SECOND TRANSITION
ONE-HALF SYMBOL PERIOD LATER
|
eeDELAY MODULATION eee
(MILLER CODE)
"ONE" IS REPRESENTED BY A SIGNAL TRANSITION AT THE
| | | MIDPOINT OF THE SYMBOL PERIOD
an | "ZERO" IS REPRESENTED BY NO TRANSITION UNLESS IT IS
— | | : FOLLOWED BY ANOTHER ZERO. IN THE LATTER INSTANCE, A
| | TRANSITION IS PLACED AT THE END OF THE SYMBOL PERIOD
ae} el sVMBot
{peal |
INTERVAL!
! | |
| | OF THE FIRST SYMBOL :

t
0 27; 4 47, & 67, 4 87, 4107,
Tee Sle mile) Vlg V97s AT

Figure 1.30 Various binary PCM waveforms

the bit period, so that a transition occurs at midpoint. Thus, PCM/PM with Bi-f-L data
formatting is a special case of PCM/PSK/PM with NRZ-L data formatting.

PROBLEMS
1.1 Let Z, and Z> be two independent Gaussian random variables with moments

E{Z\}=A,, Var{Zi}=o7
48 Introduction to Telecommunications Chap. 1

E{Zp}= Ao, Var{Z2} = of

It is desired to find the probability density function of the product random variable Z = ZZ2.
One way of doing this is to represent Z in the following form:
Zi+Z Z1-Z
PEA ye alas
2
Ses 2Saath DS Fel
(a) Find the probability density functions (pdf’s) of Z4 and Z_.
(b) Are Z, and Z_ correlated? Are they dependent?
(c) Find the pdf’s of Z4 and Z?..
(d) Based on your results in (a), (b), and (c), find the characteristic function of Z and indicate
how you would find (do not attempt to carry out the evaluation) its pdf.
1.2 Let a time-invariant system be given by the following differential equation:
dy(t) ‘
Fe gee
Let the input to the system x(t) be a wide sense stationary (WSS) random process with mean
m, = | and covariance function K,(t) = 3 exp(—|t|). Compute the following quantities:
(a) The mean of y(t),
(b) The autocorrelation function of y(t) and its PSD,
(c) The cross-correlation and cross-power spectral density of the input x(t) and output y(r).
1.3 Consider the discrete-time system represented by the linear constant coefficient difference
equation [x (1) is the input and y(n) is the output]

3 1
y(n) — go" Saal aig ee 2) =x(n) + 2x(n — 1)
(a) Draw a block diagram representation of the system in terms of adders, delays, and coeffi-
cient multipliers.
(b) Let the input x(n) be a white WSS random sequence with PSD equal to «7. Compute the
PSD of the output y(n).
(c) Design a digital causal filter (specify its impulse response) which operates on y(n) (that is,
cascaded with the first filter) such that its output z(7) is white.
(d) Suppose that a third filter H(z) is cascaded at the output of the whitening filter. Is the output
of H(z) still white if
i, Be alaié
HZ) Falieeces |
re tel
What did the insertion of H(z) do?
1.4 Consider the system illustrated in Fig. P1.4. Assume that the input is a WSS random process
with autocorrelation R,(t) and corresponding PSD S,(f).
(a) Find the autocorrelation function of y(t) in terms of R,(t).

DELAY 5 sec

Figure P1.4
Problems 49

(b) Compute the PSD of y(t) in terms of S,(f).

(c) Compute the PSD of y(t) in terms of S,(f). If S,(f) is flat for all frequencies, that is,
Sx(f) = 1, sketch S,(f) and its‘autocorrelation R(t). What is the power of the signal
z(t)?
(d) Find the cross-correlation function R,,(t) and the corresponding cross-power spectral
density S,\(f).
1.5 A WSS random process x(t) has PSD S,(f) given by

le O=lf\s fa — a)

Sf) =} co? {EU —1+al}, fall a) <Ifl< fall +a)

0, otherwise

This is referred to as a raised cosine roll-off characteristic with excess bandwidth factor a
where 0 < a@ < i.
(a) Suppose that x(t) is sampled at a rate f, = 1/T; =2f,(1 + @) producing the sampled
function
co

x)= YP x(nT,)6(t — nTs)

n=—CO

Are the samples uncorrelated? (Explain.)

(b) Suppose now that x(t) is sampled at a rate f; = 1/7; = 2 f, (note that this corresponds to
the case of undersampling). Determine and sketch the PSD of the sampled process x*(t)
(show all scaling of both axes). Are the samples uncorrelated?
(c) Assume now that x*(t) of (b) is passed through the ideal filter

nif) =| If S fa
0, otherwise

producing X(t). Write a sampling expansion appropriate to the filter output x(t). Are the
samples of this expansion uncorrelated?
1.6 An autocorrelation function which frequently arises in practical problems is

R(t) =e"! cos Bt

(a) Calculate the corresponding PSD S,(f).

(b) For what value of f does S,(f) assume its maximum value?
(c) Show that S,(/) has a maximum at a frequency unedHe to zero ifa < V3.
(d) Compute the maximum value of S,(/).
(e) Show that the half-power point occurs at f = r/(27) (where r2=aez+ B?) and is always
to the right of the peak value.
lg! Consider a random process x(t) consisting of an ensemble of rectangular waves that alternate
between the values E and —E. Assume that the distribution of zero crossings is Poisson, that
is, the probability of n zero crossings (transitions) occurring in a duration of time T is given by
n

Pr(n) = Ory er
n!

where A is the average number of zero crossings per unit time. Also, assume that x(t) takes on
the values E and —E with equal a priori probability.
(a) Sketch a sample waveform of the random process x(f).
(b) Compute the autocorrelation function of x(t).
(c) Compute the PSD corresponding to the result in (b).
50 Introduction to Telecommunications Chap. 1

1.8 Consider a random data sequence with PSD given by

2)
sin(z Tf)
S(f) =F (
ee) Watts/Hz
(xT f)
(a) Find the rms bandwidth.
(b) Find the equivalent noise bandwidth.
(c) Find the 3-dB bandwidth.
(d) Find the p% fractional bandwidth.
1.9 Repeat Problem 1.8 if the PSD is given by
4 (af T/2)
sin’
S. = T-—_—_—

1.10 Repeat Problem 1.8 if the PSD is given by

m°T (cosa fT)*
Co ——
4[(r fT)? — 42]
1.11 (a) Find the 3-dB and equivalent noise bandwidth of the Butterworth spectra of order n if

Sn(@) = |Hn (@) Hn (—@)| = |Hn(@)|?

where
ae era 1 Om, Wr.
Ay, (@) & Tr, [io = eit [2itn=1)/2n}]

(b) What is the limiting equivalent noise bandwidth as n approaches infinity?

1.12 The impulse response of a filter corresponding to a raised-cosine spectrum is
ie pee cosamt/T
PW) = (at 1 — @at/TY
(a) Sketch the pulse shapes for a = 0.25, 0.5, and 1.
(b) Determine and sketch the raised cosine spectra for a = 0.25, 0.5, and 1.
1.13 Define the frequency range involved in: (a) HF, (b) UHF, (c) VHF, (d) EHF, (e) ELF, (f) L-band,
(g) S-band, (h) X-band, and (i) K-band.
1.14 (a) Determine the space loss at a distance of 22,000 miles (35,420 km), typical of a satellite in
synchronous orbit, when the carrier frequency is 150 MHz, 1.2 GHz, 2.3 GHz, 8.6 GHz,
and 32 GHz; that is, VHF, L-band, S-band, X-band, and Ka-band.
(b) Repeat (a) for a distance of 870 miles (1,400 km) which is typical of a satellite in a low
Earth orbit (LEO).
(c) Repeat (a) for a distance of 6,211 miles (10,000 km) which is typical of a satellite in a
medium Earth orbit (MEO).
1.15 In order to maintain the antenna surface tolerance loss below 0.3 dB, what is the required rms
roughness of the antenna at 1.2 GHz, 2 GHz, 8 GHz, and 32 GHz?
1.16 In the deep space network (DSN), parabolic antennas with 26, 34, and 70 m in diameter are
used.
(a) Compute the peak gain for each antenna at L-band (1.2 GHz), X-band (8.4 GHz), and Ka-
band (32 GHz), assuming an antenna efficiency of 45%.
(b) Compute the half-power beamwidths for each antenna in the various frequency bands.
1.17 A transmitting rectangular antenna of width w and length / produces the gain function

gi(@ied) 100 vue sin gy) /A] sin[(z/ sin Ely

[(rw sin $)/A][z1 sin $,)/A]
Problems 51

Figure P1.19

when operating with wavelength 4. Determine the half-power beamwidth in the ¢; = 0 and in
the @, = 0 plane.
1.18 A communication satellite is to transmit to Earth at L-band (1.2 GHz).
(a) Assuming a LEO (1,400 km altitude) satellite, compute the required diameter of a
parabolic transmitting antenna so that the beam covers a 1,000 km? area. Assume an an-
tenna efficiency of 40%. What is the resulting antenna gain?
(b) Repeat (a) for a MEO (10,000 km altitude) satellite and a desired ground footprint of
2,000,000 km? {roughly the size of continental U.S. (CONUS)].
In the Iridium network, a phone call might require routing through various satellites to get it to
its destination. Consider the 2-satellite system depicted in Fig. P1.19. Assume ideal frequency-
translating satellites, white noise, and equal bandwidths.
(a) Compute the received signal power at the receiver as a function of uplink power P,,, satel-
lite gains G; and G2, and path losses L; and L.
(b) Compute the received noise power at the receiver as a function of Py}, Pn2, Pn3, Gi, Go,
Lj, and Lp.
(c) Compute the signal-to-noise ratio at the receiver.
1.20 A satellite at 1,000 km altitude is to use multiple downlink spot beams to achieve frequency
re-use, while covering a distance on Earth of 1,200 km (use one-dimensional antenna patterns
here). The beams are to have center-to-center separation of 0.1 radians with the gain pattern
depicted in Fig. P1.20. A receiving station in any beam must have 50 dB isolation from any
other beam (the received carrier power at any ground station from any other beam after front
end filtering is to be less than 50 dB from the carrier power at peak gain in its own beam). The
downlink signal is transmitted with bandwidth B, but can be located at any carrier frequency
fe. fo + B, fe + 2B, etc. Neglecting dual polarization, determine the amount a frequency can
be reused with this system.
1.21 In Table 1.2(a), assume the frequency is moved to Ku-band (12 GHz) and that the antenna
diameter is reduced from 5 to 3 m.
(a) Compute P,/No in dB-Hz.
(b) What is the link margin?
(c) Assume that a 34-m antenna is used on the ground instead of a 70-m antenna, compute
P,/No and the link margin.
52 Introduction to Telecommunications Chap. 1

GAIN PATTERN

0.05 0.1 0.15 0.2 RADIANS

Figure P1.20

(d) With a 6 dB link margin and a 34-m antenna on the ground, what is the maximum data rate
that can be supported?
1.22 In Table 1.2(b), assume that the Mars Relay Satellite (MRS) transmitted power is reduced from
90 to 50 watts but the Earth Relay Satellite (ERS) antenna is increased to 30 m in diameter.
(a) Compute P,/No in dB-Hz at ERS.
(b) Compute P,/No in dB-Hz on the ground.
(c) What is the link margin with 10 Mbps?
(d) Given a desirable link margin of 6 dB, what is the maximum data rate (in bps) that can be
supported?
1.23 Consider the case in which 100 kbps is to be transmitted over a 3 kHz voice grade circuit. Is
it possible to achieve error-free transmission with a SNR of 10 dB? If not, what modifications
would you suggest?
1.24 Consider a voice grade telephone circuit with a bandwidth of 3 kHz. Assume the circuit can be
modeled as an additive white Gaussian noise (AWGN) channel.
(a) What is the capacity of such a circuit if the SNR is 30 dB?
(b) What is the minimum SNR required for a data rate of 4800 bps on such a voice grade
circuit?
1.25 (a) Find the average capacity in bps that would be required to transmit a high-resolution
black-and-white TV signal at the rate of 32 pictures per sec if each picture is made up
of 2 x 10° picture elements (pixels) and 16 different brightness levels. All picture elements
are assumed to be independent and all levels have equal likelihood of occurrence.
(b ~— For color TV, this system additionally provides for 64 different shades of color. How much
more system capacity is required for a color system compared to the black and white
system?
(c) Find the required capacity if 100 of the possible brightness-color combinations occur with
probability of 0.003 each, 300 of the combinations occur with probability 0.001 each, and
624 of the combinations occur with probability 0.00064 each.
1.26 Assume E,/No = 3.4 dB at T = 28°K and the data rate is Ry = 1/7, = 115,000 bps.
(a) Find the power-to-noise energy ratio, P/ No in dB-Hz, at T = 28°K.
 References 53

(b) For T = 0°K, 28°K, and 300°K, what is the noise power?
(c) What are the communication efficiencies at the temperatures given in (b)?
(d) Find the equivalent bit mass, say m,,, that satisfies Einstein’s result relating energy to mass,
that is, E = mc’, if T = 28°K.
(e) Find the equivalent energy associated with an electron and the ratio of the bit mass given
in (d) to that of an electron.
(f) If each spacecraft TV picture transmitted from Jupiter consists of 2,000,000 bits, find how
many electron masses make up one TV picture. (Assume T = 28°K.)
(g) What is the channel capacity if the single-sided bandwidth is B =200,000 Hz? (Assume
TE DEAS)
(h) If a picture is worth a thousand words, how many bits/word are there in a picture transmit-
ted from Jupiter?
(i) If each word contains 5 bits of information, how many bits are there in a “1,000-word”
picture?
1.27 Consider phase-shift keying with a carrier component which can be written as

x(t) = Asin (27fet + cos! (ad(t)))

Assume that d(t) is a +1-valued binary message signal. Show that the ratio of power in the
carrier and modulation components of x,(f) is

REFERENCES
1. NASA’s Mobile Satellite Experiment, Pasadena, CA: Jet Propulsion Laboratory, MSAT-X Quar-
terly, no. 1, October 1984.
2. Special Issue on Optical Communications, Proceedings of the IEEE, vol. 58, no. 10, October
1970.
3. Lindsey, W. C. and M. K. Simon, Telecommunication Systems Engineering, Englewood Cliffs,
N.J.: Prentice Hall, 1973. Reprinted by Dover Press, New York, NY, 1991.
4. International Organization for Standardization, /nformation Processing Systems—Open Systems
Interconnection—Basic Reference Model, ISO 7498, New York: American National Standards
Institute, October 1984.
5. Henshall, J. and S. Shaw, OSI Explained, End-to-End Computer Communication Standards, West
Sussex, England: Ellis-Horwood Limited, pp. 10-20, 1988.
6. Knightson, K. G., J. Larmouth and T. Knowles, Standards for Open Systems Interconnection,
New York: McGraw-Hill, pp. 12-19, 20-54, 259-301, 1988.
7. International Organization for Standardization, Information Processing Systems—Open Systems
Interconnection—Manufacturing Message Specification, ISO 9506, New York: American Na-
tional Standards Institute, 1989.
8. International Organization for Standardization, Information Processing Systems—Open Systems
Interconnection—Basic Reference Model, Part 4: Management Framework, ISO 7498-4, New
York: American National Standards Institute, 1988.
9. International Organization for Standardization, Information Processing Systems—Open Systems
Interconnection—Systems Management Overview, ISO 10164, New York: American National
Standards Institute, 1990.
10. International Organization for Standardization, Information Processing Systems—Open Systems
54 Introduction to Telecommunications Chap. 1

Interconnection—Structure of Management Information, ISO 10165, New York: American Na-

tional Standards Institute, 1990.
. Clarke, A. C., “Extra-Terrestrial Relays,” MSN & CT, August 1985, pp. 59-67, also in Wireless
World, 1945.
. Deep Space Network, Flight Projects Interface Design Handbook, Vol I: Existing DSN Capabili- .
ties, Pasadena, CA: Jet Propulsion Laboratory, Document 810-5, Rev. D, March 1990. .
. Deep Space Network, Flight Projects Interface Design Handbook, Vol II: Proposed DSN Capa-
bilities, Pasadena, CA: Jet Propulsion Laboratory, Document 810-5, Rev. D, March 1990.
. Space Network (SN) User’s Guide, National Aeronautics and Space Administration, Goddard
Space Flight Center, Revision 6, September 1988.
. Global Positioning System, Vol. I, Washington, D.C.: The Institute of Navigation, 1980.
. Global Positioning System, Vol. II, Washington, D.C.: The Institute of Navigation, 1984.
. Kirson, A. M.,“RF Data Communication Considerations in Advanced Driver Information Sys-
tems,” JEEE Transactions on Vehicular Technology, vol. 40, no. 1, February 1991, pp. 51-55.
. Décima, M., W. S. Gifford, R. Potter and A. A. Robrock, eds., “Special Issue on Integrated
Services Digital Network: Recommendations and Field Trials I?” JEEE Journal On Selected
Areas in Communications, vol. SAC-4, no. 3, May 1986.
. Décima, M., W. S. Gifford, R. Potter and A. A. Robrock, eds., “Special Issue on Integrated
Services Digital Network: Technology and Implementations II,’ JEEE Journal On Selected Areas
In Communications, vol. SAC-4, no. 8, November 1986.
. “ISDN: A Means Towards A Global Information Society,’ JEEE Communications Magazine, vol.
25, no. 12, December 1987.
. Ricci, F. and D. Schutzer, U.S. Military Communications, Maryland: Computer Science Press,
1986.
. Wood, P., “Mobile Satellite Services for Travellers,’ JEEE Communications Magazine, vol. 29,
no. 11, November 1991.
. Grubb, J. L., “The Traveller’s Dream Come True,” JEEE Communications Magazine, vol. 29, no.
11, November 1991.
. Kwan, R. K. and R. A. Wiedeman, “GlobalStar: Linking the World via Mobile Connections,”
International Symposium on Personal, Indoor and Mobile Communications, London, U.K.,
September 1991, pp. 318-23.
Px Rusch, R. J., P. Cress, M. Horstein, R. Huang and E. Wismell, “Odyssey, A Constellation For Per-
sonal Communications,” Fourteenth AIAA International Communication Satellite System Confer-
ence, Washington, D.C., March 1992.
26. Taylor, S. C. and W. C. Shurvinton, “The Archimedes Satellite System,’ American Institute of
Aeronautics and Astronautics, March 1992.
Dil Papoulis, A., Probability, Random Variables and Stochastic Processes, New York: McGraw-Hill,
1965.
28. Scholtz, R. A., “How Do You Define Bandwidth,” Proceedings of the International Communica-
tion Conference, Los Angeles, October 1972, pp. 281-88.
29: Ippolito, L., et al., Propagation Handbook for Satellite Systems Design, NASA Reference Publi-
cation 1082, December 1981; 1083, June 1983.
30. Strauss, R., J. Bretting and R. Metivier, “TWT for Communication Satellites,” Proceedings of the
IEEE, vol. 65, March 1977, pp. 387—400.
Sie Ruze, R., Effect of Surface Error Tolerance on Efficiency and Gain of Parabolic Antennas, Pro-
ceedings of the IEEE, vol. 54, April 1966.
3). Rusch, W. and P. Potter, Analysis of Reflector Antennas, New York: Academic Press, 1970.
References 55

33. Bennett, W., Electrical Noise, New York: McGraw-Hill, 1960.

34. Shannon, C. E. and W. Weaver, The Mathematical Theory Of Communication, Urbana, IL: Uni-
versity of Illinois Press, 1959.
35% Gallager, R. G., Information Theory and Reliable Communication, New York: John Wiley and
Sons, 1968.
36. Wozencraft, J. M. and I. M. Jacobs, Principles of Communication Engineering, New York: John
Wiley and Sons, 1965.
Sih Abromowitz, M. and I. Stegun, Handbook of Mathematical Functions, Washington, D.C.: Na-
tional Bureau of Standards, 1965.
38. Lindsey, W. C., “System Architecting and Engineering Digital Communication Systems,” Lin-
Com Corporation, Technical report 0850-100-09.01, LinCom Corporation, 5110 Goldleaf Circle,
Los Angeles, CA, 90056, September 1992.
 Chapter 2
Power Spectral Density
of Digital Modulations

In Chapter 1, several definitions of bandwidth were given as measures of the spectral occu-
pancy of a chosen signaling or modulation technique. All of these were in some way related
to the power spectral density (PSD) of the signaling scheme. Thus, the purpose of this chap-
ter is to characterize the PSD of various types of synchronous data pulse streams which
include as special cases the data formats introduced in Chapter 1.

2.1 POWER SPECTRAL DENSITY OF A SYNCHRONOUS DATA PULSE STREAM

In this section, we derive a general result for the power spectral density of a synchronous
data pulse stream where the underlying data sequence {a,} that generates the pulse stream
has known statistical properties and the transmitted waveform is a single known pulse shape,
p(t). We shall assume a data model that covers a wide class of applications, namely, the gen-
erating sequence is wide sense stationary (WSS) with known autocorrelation function R(m).
Such a model includes the commonly encountered cases of Markov and purely random data
sources.
In the former case, each data bit, for example, a, in the sequence takes on M possible
values (states) {aj;i = 1,2,..., M} with probabilities {pj;i = 1,2,..., M)} (often called
the stationary probabilities). Furthermore, the transitions between states occur with tran-
sition probabilities {pix; i,k = 1,2,..., M} where pj;x denotes the probability that a, is
transmitted in a given transmission interval following the occurrence of a; in the previous
transmission interval. These transition probabilities are conveniently arranged in an M x M
transition matrix P = {pix}. A special case of the above is a purely random source, that is,
one that emits a data bit in a given signaling interval independent of those emitted in pre-
vious signaling intervals. Such a source can be modeled as a degenerate case of the above
M-ary Markov source (that is, P~! does not exist) whose transition matrix P has identical
rows. When M = 2, and the two states a; and @2 are equal and opposite, we get the familiar
results for antipodal binary signaling.
Occasionally, one finds situations where the generating data sequence is not WSS
but, however, is cyclostationary, that is, its first two moments are periodic. One such case
occurs at the output of a convolutional encoder where the period (in code symbols) is equal
to the reciprocal of the code rate (more about this later on in the chapter). Thus, in order
to characterize the spectral properties of the output of such an encoder, one must develop
an expression for the power spectral density of a synchronous data pulse stream whose
Sec. 2.1 Power Spectral Density of a Synchronous Data Pulse Stream 57

generating sequence is cyclostationary. We begin with the derivation of the power spectral
density of a synchronous data pulse stream generated by a WSS sequence.

2.1.1 Power Spectral Density of a Synchronous Data Pulse Stream Generated

by a Binary, Zero Mean, WSS Sequence

Consider the binary (+1), zero mean, WSS sequence {a,} for which it is known that

Gn =0; Gndm
= Ra(m — n) (2.1)
and the overbar denotes statistical expectation. From this sequence, we form the data pulse
stream
oe)

m(t)= > anp(t —nT) (2.2)

n=—Co

where p(t) is the elementary signal (pulse shape) and is not necessarily restricted to be
time-limited to only a single signaling interval, that is, the pulse train m(t) can contain
overlapping pulses. Irrespective of the properties of the generating sequence {a,}, the pulse
stream m(t) is itself cyclostationary since the expected value of the product m(t)m(t + T)
is, in addition to being a function of tT, a periodic function of t. Thus, to compute the PSD
of m(t), namely, S,,(f), we must first average the statistical correlation function

R(t; T) A m(t)m(t + 7) (2.3)

over ft [the average is performed over the period of R,,(t; t)] and then take the Fourier
transform of the result. Thus,

Sm(f) & F ((Rm(t; T))} (2.4)

where <> denotes time average and F denotes Fourier transform. Substituting (2.2) into
(2.3) and making use of (2.1) and the definition of Fourier transform, we get!

Sm(f) = i!(r) |Ra(m —n)p(t —nT)pt +t — ne Pa (2.5)

m n
T

Since

p(t) = ( P(fyel"df (2.6)

ON

where P(f) is the Fourier transform of p(t), then substituting (2.6) in (2.5) gives

Oge
njP"( P*
p1"
Bee. i/Sy RaGn =
y z m n

(Qe)
x Ba Oe fra ded ydz

18

1. Unless otherwise noted, all summations and integrals range from —oo to oo.
58 Power Spectral Density of Digital Modulations Chap. 2

Recalling that

/e JT dz — §(x) (2.8)
T

then (2.7) simplifies to

P*(— fie ft te PAOD

Sm(f) — |Ss pe Ra(m — n)P(y)

(2.9)
ar
. eae) dy

Recalling that the time average is performed over the period of Rm(t; T) , that is, T, then
T/2
(e204) a al ; jon GA pee ee: Oe) (2.10)
= Erie mY fT
Substituting (2.10) in (2.9) and letting / = m — n gives

j
=D Rate —j2nflT
Sn(f) es [roverx eee
pSMOt fT]

y (Qe)
v, y e JOT PAT ty

Finally, from Poisson’s formula, we have that

Dian ws = -oe (xaS =) 2.1)

n

which when substituted in (2.11) yields

1 1 k in zk
Sn(f) = E~ Ra seer = dsP ( ee *)P*( a (2.13)
Since

sr ={o a (2.14)
and |P(f)|? is an even function of f, we get the desired result?

Sm(f) = Sp(f) Sa(f) (2.15)

where

Si) & aIPDP (2.16)

2. The result in (2.15) can also be obtained (see Problem 2.1) directly from the basic definition of the PSD of an
arbitrary random process m(t), namely, S,,(f) = yim E |i |Mo(f P| where Mo(f) is the Fourier transform of
0-00
a To-sec observation of an ensemble member of m/(t).
Sec. 2.1 Power Spectral Density of a Synchronous Data Pulse Stream 59

is the power spectral density of the pulse shape p(t) and

ee)

Sop i Raye et (2.17)

l=—co

is the power spectral density of the sequence, that is, the discrete Fourier transform of its
autocorrelation function. Note that if the data sequence is purely random, that is,

Zntin= Ram ~n) ={ to (2.18)

then, from (2.17)

Sa(f) = 1 (2.19)
and hence

Sm(f) = Spf) (2.20)

which is a commonly used result.

2.1.2 Power Spectral Density of a Synchronous Data Pulse Stream Generated

by a Binary, Zero Mean, Cyclostationary Sequence

Suppose now that we have a binary sequence {a,} that has the properties

Gn =0; Gndm = Ra(n;m

—n) (2.21)

Furthermore

kU =) — RAWAM ING Se (6D sally, Beran @22)

where N denotes the period of the autocorrelation function Ra(n; m — n). Then, following
the steps leading up to (2.9), we see that the analogous result is now

Swf) = f S2 Relniim =P Ore eherr er

ah (2.23)
x e f2n(y+f)nT (ef2xO+\ ay

Thus far, we have not made use of the cyclostationary property given in (2.22). If now we
use this property to evaluate the sum on n in (2.23), and as before let / = m — n, then it can
be easily shown that
N
bpR, (n;1) Te gad
ye Ra (n; 1) e Ja (yt fmt _
n=)
‘i (2.24)
x bpee

k
60 Power Spectral Density of Digital Modulations Chap. 2

or making use of (2.12)

N
JA (yt fynT a bsRa (n; 1) eae |
es I (n; l) e

f= ;
lt
(225)
1 k
ated 5 ae
. Ez d (v+s a)
Substituting (2.25) into (2.23) and performing the integration on y gives

1 a! sls
Srl) byRa(n; De cual
k
ie —j2xfITp (
(_ tr. Me
=z) \ip (—f){e
t= (pees

Since R,,(t; T) is now periodic in t with period NT, then

eee ae tes l2mkt/NTyy— sin wk a f° k=0 (2.27)

aN T NEY wk 0; k#0

Finally, substituting (2.27) into (2.26) gives the desired result

Sm(f) = Sp(f) Sa(f) (2.28)

where S)(f) is still given by (2.16) and

Sif)= Do E Daal stem ae (2.29)

[=—60 n=

is the power spectral density of the equivalent stationary data sequence, that is, the period-
icity of the correlation function caused by the cyclostationary behavior of the sequence {ay}
must be “averaged out” before taking the discrete Fourier transform.

2.1.3 Power Spectral Density of a Synchronous Data Pulse Stream Generated

by a Binary, Nonzero Mean, Cyclostationary Sequence

When the generating sequence {a,} has nonzero mean, then the spectrum of the correspond-
ing synchronous data pulse stream will have a discrete component in addition to the cus-
tomary continuous component. An example of a situation where this might occur is at the
output of a convolutional encoder whose input is random but not equiprobable binary data.
The procedure for handling this case is as follows:
Define the zero mean cyclostationary sequence {A,} by

An = Qn — Qn (2.30)

which has the properties

An=0; AnAm
= Ra(n; m —n) (2.31)
Then, using the results of the previous section, the continuous component of the power
spectrum (S;,(f)), for the synchronous data pulse stream generated by {ay} is given by

(Sm(P)c = Spf) Sa(f) (2.32)

Sec. 2.1 Power Spectral Density of a Synchronous Data Pulse Stream 61

where
ee) 1 N

Si(A= 0 i Dd Ratn; p rna (2.33)

l=—oo n=1

The discrete spectral component (Sin(f))q is found from

SuiDa=F{( oaeamPC —nT)p(t+t -nr)| (2.34)

Making use of the Fourier transform relation of (2.6) gives the alternate form

(Sn Pa = /ee
ane POP) (el ay as)
y n m

Since, for a cyclostationary sequence, a, and Gp» are both periodic with period N, then
analogous to (2.25) we have

— ,—j2aynT
N
OIE ||
1 y }
k
—————
DyGn’ a [Some NDE (> oP
(2.36)

ee ae owe
— ,—j2afnT -- Saal u k
n m=1 k

Substituting (2.36) into (2.35) and integrating on y yields after simplification

(Sm(Aa
Poe = ae F(A)NT
hee, Samer
c ema
k m=1
(2.37)
peaI | pik l
aa) p® |2.k\~< 5
—j2nnl/N ain
eae
. NT (<7) ( ai) n=1domme
Performing the time average over the period NT gives as before

jontktye/nT\ — Sinw@(K +1) fl; k=-l

( is Aika op \.Oi, keel (25°)
Finally, substituting (2.38) into (2.37) and recalling that IP(f)|* is an even function of f
gives the desired result
2 N |
1 k
(Sma =Typ aii (7) SD He

m=\
(2.39)

x “ Ane
®
—_,j ‘nk/|N
dank 5
(F k )
Pe, uae
NT

Note that when {a,} is a WSS sequence (that is, N = 1), then (2.39) reduces to

i) k\|? k (2.40)
(Sm( fa = 7? d IP(=) os be T
62 Power Spectral Density of Digital Modulations Chap. 2

Example 2.1
Consider a binary, zero mean Markov source characterized by

Pr {dn+1 # an} = Pt
, (2.41)
Pr{@n41 =4n}=1— pr

The autocorrelation function for such a source is given by

Ra (1) = (1 — 2p,)!" (2.42)

Substituting (2.42) into (2.17) gives
[e.@)

Cail +2 00 — 2p,)' cos

2x fT (2.43)
l=!

From [1; p. 84, Eq. (454)], we have that

n ee n+2 oS
ee wce cee a“ +a"™*cosn@ — a" rel cos[(n+ 1)0
1)6] (2.44)
reer 1 —2acosné + a?

which for a < 1 andn — o6 becomes

ee cos6 — a?
Sea cos k@ = : a 5 (2.45)
=e 1—2acos0+a

Furthermore

ee 1—a?
1+2) a‘ cosko = a 2.46
k=1
1 —2acos6 + a2 ( )

Applying (2.46) to (2.43) with a = 1 — 2p; and g = 27 fT immediately gives the desired result,
namely

LG 2p
Sa(f) =
Lb = 2p) = 2 ap cos eet
(2.47)
ne 4p;(1 — pr)
2 (1 —2p;) (1 — cos2afT) + 4p?
Furthermore, if the data stream generated by this sequence uses rectangular pulses, that is,

Ati Oestrssull
p= |0; otherwise eae.
then, using (2.16) and (2.48) we get
ey)
: sin’ (z fT) 4p:,01 — p
Sm(f) = ACT = Pe be | (2.49)
(wfT)° 20 — 2p) —cos2xfT) + 4p7

Example 2.2
Consider the sequence formed by interleaving N independent first order Markov sources with
respective transition probabilities p,,;n = 1,2,...,.N. Then, the resulting sequence is cyclo-
stationary with correlation function

leat)
: =
(1—2p,,)"/";
“
1=0,4N,+42N,...
d j
ann : all other integer/ ey)
 Sec. 2.2 Power Spectral Density of a Generalized M-ary Markov Source 63

The power spectral density 5z(f) of (2.29) is computed as (letting / = kN, k =0, +1, +2,...)

(= LOT (I~ 2p, een

ix Ne
m=l k
(2.51)
ee oo
N7|
iz oe f+250 (1 ~ 294)on2nfR
n=1 k=1

Noting the similarity between (2.43) and (2.51) (for fixed n), we can immediately write down
the result, namely

Sa(f) = 1 > 4p; r Pt,) (2 52)

N n=]
2(1 — 2p,,)( — cos 2x fNT) + 4p; ;

2.2 POWER SPECTRAL DENSITY OF A GENERALIZED M-ARY MARKOV SOURCE

Consider a random M-ary source which every T seconds emits an elementary signal from
the set {s;(t);7 = 1,2,..., M} with probability p; (the stationary probabilities). Note that,
in contrast to the data model in Section 2.1, here it is the waveshape that changes with the
data rather than merely the amplitude of a fixed waveshape (pulse). For a Markov source,
the sequence of waveforms is further characterized by the transition probability matrix
P = {pix} where now pj;x denotes the probability that signal s,(t) is transmitted in a given
transmission interval after the occurrence of the signal s;(t) in the previous transmission
interval.
From the above statistical description of the source, it can be shown [2], by an ap-
proach analogous to that taken in Section 2.1, that the power spectral density of a modula-
tion m(t) generated by this source is given by”

fi
Smf)= = Pil Si (A)?
i=l

¥
ns (*) ‘3(7 a =)
2
1 Co

iohous i)
(2.53)
n=—Oo ll

2 Bete abt 0h oe
a5 7 ke dns (f) SCP) pik le J2 i]
p= k=

where S;(/) is the Fourier transform of the i‘" elementary signal s;(t) and
co N
pilz) & > pizts —si(t) = si(t) — D- pest) (2.54)
n= k=]
n)
Also, S/(f) is the Fourier transform of s;(t). The quantity De is defined as the probability
that the elementary signal s,(t) is transmitted n signaling intervals after the occurrence of
s;(t). Hence, from the properties of Markov sequences, Py is the ik" element of the matrix
P”. Also, by definition, p\)’ = pix.
3. Note that this result is consistent with those in Section 2.1 when s;(t) = a; p(t).
64 Power Spectral Density of Digital Modulations Chap. 2

Notice that the second term (the discrete or line spectrum) of (2.53) vanishes when
M n =

oypis: (=) =0 | (2.55)

which implies that a necessary and sufficient condition for the absence of a line spectrum is
that
M

>. pisi(t) =0 (2.56)

f=)

Many special classes of signals exist for which the general PSD of (2.53) can be
simplified. The first class is generated by the source defined by the conditions:

(i) for each signal waveform s;(t) of the set of elementary signals, the negative —s;(t) is
also in the set;
(ii) the stationary probabilities on s;(t) and —s;(t) are equal, and
(iii) the transition probability p;x = p-s whenever s;(t) = +s,(t) and sx(t) = +5;(f).

Such a signaling source is said to be negative equally probable (NEP). Its overall spectrum
is characterized by the absence of a line spectrum and furthermore is independent of the
transition probabilities themselves. For this special case, the PSD reduces to the first term of
(2.53), namely
1 M

Smif) = = DPS AP (2.57)

=I

which is the weighted sum of the energy spectrum of the elementary signaling set.
Another class of signals of interest is those generated by a purely random source
whose properties were discussed in Section 2.1. Such a source, that is, one whose transition
matrix P has identical rows, has the property that P” = P for all n > 1. In this case, the PSD
of (2.53) simplifies to
M
1 2
Sm(f) =D Pill — Pid) Si (AYP
=

° (7 a =) (2.58)
2 M M
— FY pire Re [SHAS]
i=1

When M = 2 and p; = p, p2 = 1 — p, (2.58) reduces to

1
Sm(f) = = PC = p)|Si (f) — Sof)?
jpese ‘ Ala r (2.59)

+72 [ps (F)+0-n%(F)[ 8(F-F)

Sec. 2.2 Power Spectral Density of a Generalized M-ary Markov Source 65

Obviously, if s;(t) = —s2(t) = p(t) and p = 1/2, then the signaling format becomes NEP
and (2.59) reduces to (2.57), which is then identical to (2.16).
At this point we turn to an evaluation of the PSDs of the various PCM signaling
formats discussed in Chapter 1.

2.2.1 NRZ Baseband Signaling

Referring to Fig. 1.30 and assuming that the binary waveform levels are +A, the NRZ
signaling format falls into the class of signals whose spectrum is given by (2.59). Since the
elementary signal is a rectangular pulse of width 7, its Fourier transform is

p sin(z fT)
Si(f) = —S2(f) AT exp(—jxfT) (2.60)
Ce

Substituting (2.60) into (2.59) and letting E = A2T, we get

Sm(f) 1 > sin?

(a fT)
E = 7! py)
—(1 —2p)*6(f) 5(f) +4pd
+ 4p —
- p) —-
(@fTY (2.61)
2.61

This result could also be obtained from (2.32) and (2.40) by recognizing that

_ Fa 1— (1 —2p)*=4p(i — p); =
Gn =(1—2p); AnAm= |0: ( P) P( P) aie
(2.62)
P k\ _JAT; k=0
Pet ks k#0

When p = 1/2, the dc spike at the origin disappears and the NRZ signaling format falls into
the NEP class with

Suh) Cosine GrfT)

(2.63)
Be PAG ty
2.2.2 RZ Baseband Signaling

Here we have a situation where S|(f) =0 and S2(f) corresponds to the Fourier transform
of a half-symbol wide pulse, that is,

AT ja fT \ [sin (fT /2)

So(f) = hae (- 5 )are (2.64)

Since the source is again purely random, then substituting (2.64) into (2.59) gives

Seed Te
Se te tea P eile2h s(f= Ln3)
ie (2.65)
tig sin?
(1 fT/2)
Bea? | fT jay
66 Power Spectral Density of Digital Modulations Chap. 2

2.2.3 Biphase (Manchester) Baseband Signaling

Here the two elementary signals are defined by

oe As Os7=T/2
si(t) = ee a (2.66)

S2(t) =—s1(@)
Substituting the Fourier transform of (2.66) into (2.59) gives

Sm(f) 1 = oh We n
WP -Fa-2 Yo (=) a(t-F)
ea (2.67)
sin’
(2 f T /2)
ea 4p(1 - Clap) ——
GefT 2

For p = 1/2, the line spectrum disappears and

4
Sm(f) — sin'(afT/2) (2.68)
E (x fT/2)*
2.2.4 Delay Modulation or Miller Coding

As indicated by Hecht and Guida [3], the Miller coding scheme can be modeled as a
Markov source with four states whose stationary probabilities are all equal to 1/4 and whose
transition matrix is given by

Oia l/2> sO ietbio

ito Ae AY >:
5 ig Oey Pa) 0 eae
Bi2E Qi Pas MeO
Another property of the Miller code is that it satsifies the recursion relation

PV T= -7P'r i>0 (2.70)

where I is the signal correlation matrix whose ik" element is defined by

1 r
vik & BE, Ih si(t)sy(t)dt i,k =1,2,3,4 (2.71)

For the Miller code, the four elementary signals are defined by

si(t)=—sq(t) =A; O<t<T

2) =—n(y
Ay ovOp Tye (2.72)
=|Ai T2S¢eT

and E; = A?T; i=1,2,3,4. Substituting (2.72) into (2.71) and arranging the results in
Sec. 2.2 Power Spectral Density of a Generalized M-ary Markov Source 67

26
ey —-—-— NRZ
22 ------ BIPHASE
a DELAY MODULATION
1.8
WHERE:
1. 6 T = SIGNAL PULSE WIDTH
ee
S A= SIGNAL AMPLITUDE
E12
10%. E=A°T
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1.0 We 1.4 1.6 1.8 2.0
de

Figure 2.1 Power spectral densities: NRZ, biphase, and delay modulation (Reprinted from [15])

the form of a matrix, we get

1 0 0 -1
Sd (ae aeh tao
Our te) (2.73)
Alt OF 0, 1
Finally, using (2.69), (2.70), and (2.73) in the general PSD result of (2.53) yields the
result for the Miller code:
Sm(f) = ——_—_——
———
I _(23 — 2080 — 22 cos20 — 12 cos 30
E 262 (17 + 8 cos 86) (2.74)

+ 5cos
46 + 12cos 50 + 2cos60 — 8cos 76 + 2 cos 80)

where 0 Aa fT.
The power spectral densities of the NRZ, biphase, and delay modulation signaling
schemes as given by (2.63), (2.68), and (2.74), respectively, are plotted in Fig. 2.1.
Spectral properties of the Miller code that make it attractive for magnetic tape record-
ing as well as phase-shift-keyed signaling include:

1. The majority of the signaling energy lies in frequencies less than one-half of the data
fate, Rah
2. The spectrum is small in the vicinity of f =0. This spectral minimum facilitates
carrier tracking, which because of property | can also be more efficiently achieved
than Manchester coding. A reduced spectral density in the vicinity of f = 0 is an
important consideration in tape recording, primarily because of the poor de response
of the channel.
3. As a result of property 2, a lower magnetic tape recording speed (higher packing
density) can be used.
 68 Power Spectral Density of Digital Modulations Chap. 2

4, The Miller code is insensitive to the 180° phase ambiguity common to NRZ-L and
Manchester coding.
5. Bandwidth requirements are approximately one-half those needed by Manchester
coding.
Other well-known techniques for reducing spectral bandwidth rely on ternary rather than
binary alphabets [4]; however, this is achieved at the expense of an increase in error proba-
bility performance and a complication of implementation.

2.3 POWER SPECTRAL DENSITY FOR DATA PULSE STREAMS WITH DATA
ASYMMETRY
In many applications [5S—7], the data pulse stream experiences data asymmetry in that the
rising and falling transitions do not occur at the nominal data transition time instants. Such
asymmetry in the data waveform causes the presence of a line spectrum in the PSD as
well as distortion of the continuous component of the PSD. Both of these ultimately cause
degradation of the error probability of the receiver. In order to assess this degradation, it is
desirable to evaluate the PSD of NRZ and Manchester streams with data asymmetry.

2.3.1 NRZ Data

In this section, we use the data asymmetry model proposed in [5] which assumes that +1
NRZ symbols are elongated by AT/2 (relative to their nominal value of T sec) when a
negative-going data transition occurs and —1 symbols are shortened by the same amount
when a positive-going data transition occurs. When no data transition occurs, the symbols
maintain their nominal T-sec value. As such, we can use the generalized M-ary source
model of Section 2.2 where now M = 4 with
HO) Ay = [staAe
"10: — otherwise

—Ae =—f/2<7
<= T(k— A
200 =|6 / ote ( )/2
: otherwise
Ors)
AR) Totes 2
SOs |0; otherwise

By ey ae alot ea at
CS |0; otherwise
The stationary probabilities associated with these four elementary waveforms are, respec-
tively

Pi\=ppPr; p2=(1— p)pr;

(2.76)
p3=P(l— pr); pa=(1— pi — p;)
where p; is the transition probability which for a purely random source is related to the a
priori probability of a +1 NRZ symbol, p, by

Pt =2pC — p) (2.77)
Sec. 2.3 Power Spectral Density for Data Pulse Streams with Data Asymmetry 69

Taking the Fourier transform of the four elementary signals in (2.75) and substituting the
results in (2.58) gives after much simplification [6]

aD apd
S, 227s (x fT) i, 7,, sin® (1 f Tn)
(f) yee O12) + Ante. rl
Teepr APT Aste
sin (27 fT)
2 ee [Aa(p, Pr.) — As(p, Pr)] (2.78)

+A? [2p — (1 ~ mp) P3(f) + >ariLe Cin, p.)8(f - =)

where

Ai (pr) = pr — pr) (1 + 20 — py) —

A2(P. Pron) = (3p7 + Pu(l — px) LL + 20 — 2p)]) cos? fT)

A3(pr,) = Pr (1+p? — pr)cos* 1 fT + p? cos(2
fT n)

|
Aa(P, Pr.) = pr(1 — pr)(1 — 2p) Ecos(2xfTn) — p sin afm (2.79)
1
As(P, Pr) = PC =p CLs". p)

CQ, ph) = sin?(n7n) [cos*(nzn) +(1- 2p)” sin*(nz7)|

tS

Figure 2.2 is a plot of the continuous component of (2.78), (Sm(f))-, normalized by the
energy E = A°T, for various values of data asymmetry 7 in percent.

2.3.2 Manchester Data

In this section, we use the data asymmetry model proposed in [5] which assumes that for
a +1 data bit, the first half of the Manchester symbol is elongated by AT /4 (relative to its
nominal value of 7/2 sec) and for a —1 symbol, the first half of the Manchester symbol
is shortened by the same amount. When a data transition follows, the second half of the
Manchester symbol retains its nominal value (i.e., 7/2). When no data transition occurs,
the second half of the Manchester symbol adjusts itself so that the total Manchester symbol
maintains its nominal T-sec value. In view of the above asymmetry model, we can use the
generalized M-ary source model of Section 2.3 where now M = 4 with
 70 Power Spectral Density of Digital Modulations Chap. 2

(Sm(f))c/E

7 = Data Asymmetry

Figure 2.2 Normalized power spectra

versus data asymmetry—NRZ data
fT (Reprinted from [6] © IEEE)

nF —T/2<t<AT/4
- AT/4<t <(T/2)0+A/2)
: otherwise

—A; —T/2<t<-—AT/4
A; —AT/4<t <(T/2)0 — A/2)
(OR otherwise
(2.80)
A; —T/2<t<AT/4
—A; AT/4<t<T/2
0; otherwise

> SPO <te eats

sq4(t) i
= A; —AT/4<t<T/2
0; otherwise

The stationary probabilities associated with these four elementary waveforms are again
given by (2.76). Taking the Fourier transform of the four elementary signals in (2.80) and
substituting the results in (2.58) gives after much simplification
Sec. 2.3 Power Spectral Density for Data Pulse Streams with Data Asymmetry 71

where for p = p; = 1/2, the discrete component (S,,(f)),4 is given by [7]

9
(Snf))a = Arf)
DS (2.82)
ee=p a a) 1) ei
ti 1Gay) 2 Gnsayet
Pa] HAG
3 +7) 13 (¢ - =)
T

with

n=A/4
feeRvene sin? (man) [1 + 2h1(m, n)]?
4 (2.83)
H>(m, n) = sin? (2mzn)
H3(m,9) = 2 sin? (min) (cosmmn) [1 + 2h1(m, n)]

The parameter h;(m, 7) in (2.83) is defined as

(2.84)

Sp =—
ey
2
MTG YC) HOP ve
ery ailcre.
bspiectigi ete)
(#)
(2.85)

S
Sa
+ A°TC6(n)

where
 72 Power Spectral Density of Digital Modulations Chap. 2

ine {| afTU+n) ae rfEl=7) ]+eosarro|

es ee D D,
T ]
ey asta
Culm) = gcosfTn {2sin?eeZ e
=

C3= ;[— 4cos (“)]

Cia) = ee {sin(7)

(2.86)
+ sin (=) [1 - cosn/T cos nl}

ie 1
— =sin (=) sin (=) 58 8 [2 cos3z fTn + cos
2x fTn]

-- IEC!
Cs(n) = ;sin’peo”2]_ =sin? aoe
SaL2

=ant TUE] Va
cay eo ee |ee ea)
5
ee
eee
feces ed
|cosmer

For 7 = 0, (Sm(f))q = 0 and (Sm(f)), of (2.85) reduces to (2.68). Figure 2.3 is a plot of
(Sm(f)), of (2.85) normalized by the energy E = A°T for various values of data asymmetry
n in percent. Note that as 7 increases, the null at f = 0 disappears.

2.4 POWER SPECTRAL DENSITY OF CONVOLUTIONALLY CODED MODULATIONS

In this section; we apply the results of Section 2.1 to the case where the data sequence {ap}
corresponds to the output of a convolutional encoder (see Chapter 13 for a more detailed
discussion of convolutional codes). For most of the optimum performance convolutional
codes, a purely random input gives a purely random output with a bandwidth expansion
equal to the reciprocal of the code rate. As such, the PSD of the encoder output would
merely be a frequency scaled version of the input spectrum. There is, however, a class of
codes (to be defined shortly) for which the above is not a true statement, and indeed some
of the optimum performance codes fall into this category. In these instances, the bandwidth
expansion produced by the encoder can be considerably less than the reciprocal of the code
rate. A discussion of the implications of this statement will be given later on. Another
consideration is the case where the input to the code is not a purely random source. Along
these lines, we shall again consider two data models, namely, an unbalanced NRZ source
where the two signaling states of the binary waveform are not assumed to occur with equal
probability, that is, p and (1 — p), and a first-order Markov source which allows transition
probabilities, p;, between 0 and 1.
We begin by noting that the symbol sequence at the output of a convolutional encoder
is a cyclostationary process (not necessarily zero mean) with period equal to the number of
modulo-2 summers, n. As such the results of Section 2.1.3, in particular, are the ones that
apply.
 Sec. 2.4 Power Spectral Density of Convolutionally Coded Modulations 73

0.6

7 = Data Asymmetry

0.5

0.4

0.3

(f))o/E
(Sim

0.2

VY
0.1

Figure 2.3 Normalized power spectra

1.0 2.0 versus data asymmetry—Manchester data
(Reprinted from [7] © IEEE)

2.4.1 Convolutional Encoder Model

Consider a convolutional encoder with constraint length K and rate b/n, as shown in
Fig. 2.4. In the m" information interval, b information symbols amp+j; j =0,1,2,...,b—
1 enter the encoder and n channel symbols xmy+); v = 1,2,...,n exit the encoder. The
structure of such a convolutional encoder can also be defined by a generator (connection)
matrix, namely

bie MAW” oheioe Ail

SOT IPI OLNO cae MdGa)
G= ’ 5 (2.87)

Sei oe a e SRK

where g;,; is either one or zero depending, respectively, on whether or not the i th modulo
summer is connected to the 7° shift register stage. For mathematical convenience, we
shall assume that both the input symbols {a+ ;} and the output symbols {x ,+)} take on
values plus and minus one. This allows modulo-2 summation operations to be replaced by
algebraic products. Thus, the encoder of Fig. 2.4 has the input output relation
Kb
Xmn+v = I] erie] (2.88)
v=
74 Power Spectral Density of Digital Modulations Chap. 2

Figure 2.4 A general constraint length K, rate b/n convolutional encoder (Reprinted from [8], [16] © IEEE)

Clearly, the output sequence {Xmn+.} is cyclostationary with period n as previously men-
tioned.
If, analogous to (2.2), a modulation m(t) with pulse shape p(t) is formed from the
sequence {Xmn+y}, then the spectrum of this modulation is given by (2.32) and (2.34). In
terms of the present notation, the continuous component of the spectrum (S,,(f)),. would
be

(Sm( Pc = Sp(fSZ(P) (2.89)

where S,(f) is given by (2.16) and

] n (ee) Rec

S¥(f) = F » pesRx(v; Ke!” a (2.90)

where

Rx(V3 Kk) = (Xy — Xv) Arte — Xv¢u) A XvXv4n (2.91)

Letting

K=nlaag—vi ~ viel een (2.92)

we can rewrite (2.90) as

i ee
Sx(f) = — dd de Rx(v; nl +g — vje S27 ta-wWfT (2.93)

Equation (2.93) will be more useful later on since it shows that x, is the symbol generated
from the v'" interleaving tap output and Xni+q is the symbol generated from the g" inter-
leaving tap output.
From the definition of a cyclostationary sequence, we have that

XyXni+g = Xy—niXq (2.94)

Sec. 2.4 Power Spectral Density of Convolutionally Coded Modulations 75

Thus, using (2.94), we can rewrite (2.93) in the more compact form
1 n n CO

POD
S-(f). = — ies dd, B epRx(v;
nl + qq — (nl +q—v)fT]fT
v)cos[2m(nl+g —v) (2.95)
2.95

where ¢ is the Neumann factor defined by

ile =)
o) = |2 140 (2.96)

We shall define the “memory” of the cyclostationary process as the smallest integer /* such
that X,Xni+g =0 for] > I*.
If {x;} is not zero mean, then in addition to the continuous PSD component of (2.89),
we will also have a discrete spectral component (Sj(f))q given by

OS
oi
D
wn Lye. (4)
nT Ysvc0s
Ysveos( (=
n ).
=—00
(2.97)

+|
Sosesin(4)] fo(s-4)
2.4.2 Encoder Output Spectrum for Independent Binary Symbol Input

Consider the input to the encoder of Fig. 2.4 to be a sequence of independent binary symbols
{a;} that take on values +1 with probabilities p and (1 — p). Let m = 0 in (2.88) with no
loss in generality; then since the symbols are independent, the mean of x, is given by

ee (wea (aeons ee Oeat (2.98)

where
Kb
ay AD ani (2,99)
i=l
that is, the algebraic sum of the number of +1’s in the v'" row of G. Furthermore, it can be
shown that
(aye teqt2Bvq() Le (a) et: a= k= 1

Rx(v; nl +4 —v) = § Geta t2Fa(—) _ Gytea, —(K —1) <1 <0 (2.100)
0; otherwise

where
(K—lb
Bug) 2 Me 8v,i8q,i+lb (2.101)
il

that is, the cross-correlation of the v'" row* of G with the g'" row shifted by /b elements.
Finally, substituting (2.98) and (2.100) into (2.95) and (2.97) gives the desired results for

4. The cross-correlation is performed only on the parts of the v'h and shifted g'" rows whose elements overlap.
76 Power Spectral Density of Digital Modulations Chap. 2

the continuous and discrete components of the encoder output spectrum, namely

Sx(f) = aysss= El ae at] cos [2m (nl + q — v) fT]

v=lg=—Iel—0 (2.102)

2
sin af T
(Sm( fe = (=) Se (f)

where a unit rectangular pulse has been assumed for p(t) and

EE ESTE a m vk ?
(Sm(P)a == DU (=ra > (a cos (= )
k=—6S n
(2.103)
2D
Ley Man bee ae eS
[Se n( =) a(/ =)

Several interesting points can be deduced from (2.102). First, note from (2.99) and
(2.103) that for v = q and / = 0, the exponent a, + a — 2Byg(/) = 0 independent of the
code. Thus, for these terms
(G)%t%4—2B vq) = 1 (2.104)

for all @ (including zero). It is also possible that the above exponent can be zero for other
combinations of v, g, and /. The conditions on the elements of G under which this is
possible are derived in [8] and are given by

Sei =O Bal, 2.6.40 (2.105)

8v,i = 8q,i+lb; i — Ile 2 Cr) (K ae l)b

which are graphically illustrated in Fig. 2.5. This defines a class of convolutional codes
whose significance will become apparent shortly. In the meantime, it is sufficient to remark
that if the code is such that (2.105) can be satisfied for any values of v, g, and / (other than
v = q and/ = 0), then (2.104) is satisfied for all a (again including zero).
Returning to (2.102), we further note that independent of the code, the exponent
a, +g is never equal to zero. Thus, the term (a)*’t will be equal to one only when @
is equal to one and will vanish when a is equal to zero (that is, p = 1/2). This last statement
is crucial to the results that now follow.

The case of a purely random data input (p = 1/2). Consider the special case
of a purely random data input with rectangular pulses and p = 1/2 (a = 0). If the code is
such that (2.105) is not satisfied for any combination of v, gq, and / (other than v = g and
! =0), then the only nonzero terms in (2.102) will be the n terms correponding to v = q
and / = 0 for which (2.104) applies. Thus, Sz(f) = 1 and the continuous component of the
spectrum is just given by (2.63). Furthermore, the discrete component vanishes. Thus, the
conclusion to be reached is that for all codes that do not satisfy (2.105), a random NRZ data
input results in a random NRZ data output scaled (expanded) in frequency by the reciprocal
 ul
Sec. 2.4 Power Spectral Density of Convolutionally Coded Modulations 77

qth ROW OF G: Lb presen ssmash:

op AOI Nit yee
{a Gea a ie nen Cn a eee

1 Oe e © ee eo ee 1 0 Oeeeeed

(k-2)b Lb
Figure 2.5 An illustration of the constraints of Equation (2.105) (Reprinted from [8], [16] © IEEE)

of the code rate, that is, n/b. We define this class of codes as uncorrelated convolutional
codes.>
If, on the other hand, the code is such that (2.105) can be satisfied for at least one
combination of v, q, and / (other than v = q and / =0), then S;(f) of (2.102) will be
frequency dependent and the encoder output spectrum will not be merely a frequency scaled
version of the input. In fact, let {v, g,/} denote the set of combinations of v, q, and / (other
than v = q and/ = 0) for which (2.105) can be satisfied; then (2.102) simplifies to

S2(f)=14+->» Y= > e1'cos[2x (nl + q — v) fT] (2.106)

{v4.1}
Note that the discrete component still vanishes. We define this class of codes as correlated
convolutional codes and observe that for such codes the encoder output spectrum will differ
(in form) from the input spectrum.
An example of a correlated convolutional code is one whose generator matrix is
such that two or more rows are identical since it satisfies (2.105) with / = 0. Larsen [9]
tabulates short constraint length codes (up to and including K = 14) noncatastrophic codes
with maximal free distance for rates 1/2, 1/3, and 1/4. Included in this tabulation is a rate
1/3, K =3 code with generator matrix
eg) ban
c= l | (2.107)
| ey ae
This code was first found by Odenwalder [10]. The two combinations of v, q, and / that sat-
isfy (2.105) are v = 2, g =3,/ =0, and v = 3, g = 2, / = 0 in which case (2.106) becomes

Sg(f) = 1+ cos 2x fT (2.108)

5. Note that all rate 1/2, noncatastrophic convolutional codes are uncorrelated codes.
78 Power Spectral Density of Digital Modulations Chap. 2

CODE GENERATOR MATRIX

i}(@) 4
edie
Tee

ete
2.0Cq abe

1
. [| :

TI
=
1.6 a ice
B
1.4

We2
uy

ee :
eee
€ = 23
ay secee ai ss
WY =]

. BEEEEEEEEH a t

SESeSSseeeeeeieait :
as Sooo SeSseeecueaia za
| SeeeeeeeeeeeseeetCeseesare
ae SEEGER EERE EERE
SGSEESTEECCEEESS veeTEE
0.2 ee as a. Rea wa a SS

0.0 seeon me Zan = fa HHH SEE = 1

-4 -3 -2 -1 0 1 2 3 4

Figure 2.6 Spectrum for best rate 1/3, constraint length 3 convolutional code; dotted
curve is spectrum of uncoded NRZ (Reprinted from [8], [16] © IEEE)

Multiplying (2.108) by (2.63) gives the encoder output spectrum illustrated in Fig. 2.6.
Another interesting result may be gleaned from the example just presented. Since
codes whose generator matrix have at least two identical rows are a subclass of correlated
convolutional codes for which / = 0, then for these codes (2.99) simplifies still further to
1
S;(f) =1+—- cos [27(q — v) fT] (2.109)
: ‘ 2 »
where the set {v, g} now represents all pairs of integers corresponding to pairs of identical
rows in G. Since (2.109) only depends on the difference v — q, and small values of v — qg
correspond to less oscillatory behavior of S(f), then it is clearly desirable from a spectral
concentration standpoint to have the identical rows packed close together in G. For example,
if the bottom row of G in (2.107) is moved to the top, then v — g = 2 and

2
Sef) =1+ 5 cosdnfT (2.110)
which is illustrated in Fig. 2.7. .
Sec. 2.4 Power Spectral Density of Convolutionally Coded Modulations 79

CODE GENERATOR MATRIX

=
SreeFEE

(Sin
(Do
/E

Raeesa
GRRE
BABS
GARE
NZ|ABER
EO
RDIALGRD
eaeaes By|
|
SQRDIRERRISEE
LS

Figure 2.7 Spectrum for best rate 1/3, constraint length 3 convolutional code with
permuted code generator matrix (Reprinted from [8], [16] © IEEE)

The important point to realize from the above is that permuting the rows of the gen-
erator matrix of a convolutional code can be used to change the encoder output spectrum
without any effect on the bit error probability performance of the code.

The case of an unbalanced NRZ input (p + 1/2). Consider the optimum

rate 1/2, constraint length 3 code whose generator matrix is given by the first two rows
of (2.107). For this code a; = 2, a2 = 3 and the sets of exponents @, + ag and a, + a —
2Py,g); 1=0,1,2; v,qg=1, 2 are tabulated below:

l v q 2hAeg =) ay + Og @p.-- ag 2Bv,4 )

Sa
See)
SS NO
KS
RS
KS fF
Nn
Ae
DN Cy
SS
Ie
SS
80 Power Spectral Density of Digital Modulations Chap. 2

Lov oq W+q-v ata ay+a,—26,4(0)

2 1 ] 5 6)
2 Ms Be 6 2.
2 l ] 4 4 2
p l Ps 5 5 3
2 Z 1 3 5 3
2 2 2 + 6 4

Using these results in (2.102) gives the closed form result

i
Sef) = 1-5 (a4— a) + (@+a — 24) cos2nfT
a (@ —a°) [cos 4a fT + cos 8 fT] (2.111)

+ (a - a°)[2cos6zfT + cos 10x fT]

Since this code is an uncorrelated code, for a = 0 (2.111) reduces to Sx (f) = | as it should.
Figure 2.8 is an illustration of the corresponding continuous spectrum (Sm(f)), [ob-
tained by multiplying (2.111) by (2.63)] for several values of p = (1 — a)/2. We note that
the output encoder spectrum narrows as p is decreased despite the fact that the continuous
component of the input spectrum remains unaltered in shape as p is varied [see (2.61)].
The discrete component of the output spectrum is easily found from (2.103) to be

es ae k
(@-#)
(SnMa=Sse Yo (=) e(¢-35) (2.112)
|sed IO here

2.4.3 Encoder Output Spectrum for First-Order Markov Input

For the first-order Markov source, we first transform the information sequence {a,} into a
transition sequence {tj} where t; = ajaj_1; j = 1,2,... The sequence {tj} is a stationary
sequence of independent +1 random variables with transition probability p,. For any / and
n > 1, we have

n n

ai-n = if pts aL =a] I] +i (27153)

fei i=]

Defining

hyi = 8v,i ® Bvit1 B... B Bi, Kb (2.114)

where ® denotes modulo-2 summation, and

Kb
pu eh (2.115)
[=2

then after considerable algebraic manipulation, it can be shown [8] that

Sec. 2.4 Power Spectral Density of Convolutionally Coded Modulations 81

CODE GENERATOR MATRIX

101
()"s) aa

:
ale ais
Geom
ine
Bites

ABS. ne BA
1 vt -1
fT

Figure 2.8a Spectrum for best rate 1/2, constraint length 3 convolutional code; p = 0.1 (Reprinted from [16])

nl
|
Wl

Sx (DH=-n Soo (Lh = fu — hg + 2hy,thg1)

v lq =]

El ( tVe —2&y, (1)

(l —hg,1) (7) set

x cos [27 (nl +q —v) fT] (2.116)

cos (27 (nK +q —v) fT] —#’ cos [2m(n(K — 1) +q —v) fT]
1 —2t 7b cos2anfT +t 72b
where

Fa Ou
82 Power Spectral Density of Digital Modulations Chap. 2

3/%()s)

Figure 2.8b Spectrum for best rate 1/2, constraint length 3 convolutional code; p 0.3 (Reprinted from [16])

Equation (2.117) is used in (2.89) to get the continuous component of the encoder output
spectrum.
The discrete spectrum is obtained as

(Sm(P)a
(2.118)

Note that a discrete spectrum can potentially exist at the encoder output despite the fact
that the encoder has only a continuous spectrum at its input such as that in (2.49) with
Jp — (1 ay t) /2. Note that if the code is transparent, that is > each row of G has an odd
number of ones ’ then from (2.114)

(25119)

which when substituted in (2.118) results in (Sj, Fa = (),

As an example of the application of (2.116), again consider the optimum rate 1/2
constraint length 3 convolutional code for which the modified generator matrix H whose
elements are defined in (2.114) is given by
 Sec. 2.5 Spectral Occupancy of Various I-O Modulation Techniques 83

(Sin
()Q/E PCH
COOP
Be
fspibiay
Bie
Bo
Be
aS
be
SRE
Bh
ce a
aa
aes
|
am
an
Be)
ab
Bas
im

Ba
IZ

rely a (2.120)

The corresponding continuous component of the encoder output spectrum may be obtained
from the following:
. |
Se(f) =1- al (2121)
f Pie ry er i) cos4afh ert =i? LP cos 8arr
1 —2tcos4nfT +?
Figure 2.9 is an illustration of (Sn(f)),. [using (2.121)] for several values of transition
density p; = (1 —f) /2.

2.5 COMPARISON OF THE SPECTRAL OCCUPANCY OF VARIOUS I-O MODULATION

TECHNIQUES
Thus far, we have considered evaluation of the PSD only for baseband modulations. In a
large part of the remainder of this book, we shall address ourselves to a form of bandpass
modulation in which two lowpass modulations are impressed on inphase and quadrature
carriers. Such an inphase-quadrature (I-Q) signal has the generic form
84 Power Spectral Density of Digital Modulations Chap. 2

(Do
(Sin
/E
ee ||

:oH

Figure 2.9a Spectrum for best rate 1/2, constraint length 3 convolutional code;
first-order Markov source; p; = 0.1 (Reprinted from [8], [16] © IEEE)

s(t) = Aym](t) cos @,t — Agmo(t) SIN Wet

| (2.122)
=Re {3(nei'} 3(t)
= Aym)(t) + jAgmo(t)
where w, = 27 f, is the radian carrier frequency, m;(t) and mg(t) are baseband modula-
tions of the form in (2.2), that is,
ee)

mi(t)= >> aapi(t—nT;); mog(t)= )> bapo(t—nTg) (2.123)

n>=—OO n=—CO

and s(t) is the equivalent complex lowpass signal. Assuming that {a,} and {b,} are £1, zero
mean WSS sequences as in (2.1) but are independent of each other, then the PSD of s(t) is
related to the PSD’s of m;(t) and mg(t) by
A2
SHGp = =a [Sm(f — fe) + Smi(f + fe)]
(2.124)
Ag
ea eA [Smrate - i) ae Smo(f ot |

One popular carrier modulation to be discussed in Chapter 4 that falls in the class
defined by (2.122) is quadrature phase-shift-keying (QPSK) in which {ay} and {by} are
Sec. 2.5 Spectral Occupancy of Various I|-O Modulation Techniques 85

@
||
a
||
=
=
=
f
||
=
Ba
f
oe
eB
=
S
i
||
||
Ba
Si
sea]
a
||
el
||
S
(Qo
(Sm
/E ||
{isa}
is
B
2
past
B
5

SB
SB
a
I
a

ace

1-3}
|
eeoeEeele
Ee
ASN
BARAT
PARA
AAA
PALI]
Beso
eeeee
BAAO2
"! BOLE
BEARS
ORCI
eAewa
Bee
ABBon
GHG
USRRISSRAORESNO
4 CERES
PORRABOAs
RARE
RS
GRE
PRPs
RI|
BRA
Bes
BRAN

Figure 2.9b Spectrum for best rate 1/2, constraint length 3 convolutional code;
first-order Markov source; p; = 0.3 (Reprinted from [16])

random data sequences, Ay = Ag = A, T; = Tg = 2Ty (Tp is the bit time), and p;(t) =
po(t) = p(t) is a unit rectangular pulse of duration 27). Using the results of Section 2.1.1,
we find that the PSD of QPSK is given by
ee 2
PT; Ee mf —— fe)Tths ‘iSie2 ge Cit Fe)
ii 4
Ss(f) = (2.125)
Catf=foty “Ox + poy’
where P = A? is the power in s(t).
Another popular carrier modulation technique to be discussed in Chapter 10 is
minimum-shift-keying (MSK) [11] which has an I-Q representation in which {a,} and {by}
are random data sequences, Ay = Ag = A, T; = Tg = 2Thp, and

- COS 3755 —T, <t<Tp , sin ne Gat 215 7126

pit) = 0; otherwise ° 2 es 0; otherwise ( )
Again using the results of Section 2.1.1, the PSD of MSK is found to be

S(f\=—>
8PT, cos? 2n(f — fe)Ti
ete me eee
cos* 2m(f + fe)Tb
(2.127)
Pee toy yi) (l= 16(F + f-)°7,) it 4
Here P = A*/2. Note that, asymptotically, the PSD of MSK rolls off at a rate hpi
compared with QPSK which rolls off as f a:
86 Power Spectral Density of Digital Modulations Chap. 2

“s)
3/0)
z
FE
0.60
aie
Lai
Pt

Figure 2.9¢ Spectrum for best rate 1/2, constraint length 3 convolutional code;
first-order Markov source; p; = 0.7 (Reprinted from [16])

Even faster spectral roll -offs can be achieved by further smoothing of the pulse shape.
For example, so called sinusoidal frequency-shift-keying (SFSK) |12] (see Chapter 10) has
an I-Q representation in the form of (2.122) where {a,} and {b,} are random data sequences,
Aj = Ag=A,T; =Tg = 27}, and

sin 2t/Tp
2mut/Tp y= otherwise
<t<Tp

SS
a PS ke
5IN R|> Bao)a)aS
7)ee (2.128)
ee
RIS eS, VIah BN
[co le
Ale yee Ss

Q Q ~
— a
ll ee
rey
is aM —!
SS | ls a: te
Aik
hs ~~)
Sa
==!
evil
=) so
a
ke)eee
A aeSIN
ers
sof
S
D
oO
=| ss
reVIVi
ta

For SFSK, the PSD can be obtained in the form of an infinite series, in particular [12]
Sec. 2.5 Spectral Occupancy of Various I-Q Modulation Techniques 87

1.0

0.9

0.8

0.7

0.6

(Sm
(Do
/E

Figure 2.9d Spectrum for best rate 1/2, constraint length 3 convolutional code;
first-order Markov source; p; = 0.9 (Reprinted from [8], [16] © IEEE)

PI; | — I
Ss(f) = s |
E (3)Ao(f eS Te) ee SS J2n (;)Bon(f rc fe)
all

Pe
= I ee
6 a Jon-1 (;)Bon—-1(f ax ‘|

(2,129)
| wo | : .
oF E (3)Ao(f ale Sc) spe y J2n (3)Bon(f ap te)
hn

2
= ]
oF oe J2on-1 (;)Bon—-1(f = ra|
i

where J;,(x) is the n' order Bessel function of argument x and

 88 Power Spectral Density of Digital Modulations Chap. 2

na sin 27 f Tp
AS are,

Beka I’ 4% cos2xfT
Aot= 54 (i+ an) +3415 a5 )= Tae
1 2n 1 2n
A= 5a(f+Z) +54 (4-2 )3
‘ (2.130)
1 2n — I 1 2n — I
Arm-1(f) = 54 (r+ a )-34 (+- i )

1 I 1 1
Bon-1(f) = Spe (FFP a) ae 5 Aon I G= a)

It can be shown that (2.130) rolls off asymptotically as f~°.

For all three of the above modulations, the bandpass PSD can be expressed in the form
1
ENG oer GC ic)! (2.131)
where G(f) is the equivalent two-sided lowpass PSD, that is, the PSD of the complex
lowpass signal s(t) defined in (2.122). Figure 2.10 is a plot of G( f)/2PT) (in dB) versus
fT» for the three modulations. This figure clearly illustrates the trade-off between the width
of the main lobe and rate of roll-off of the spectrum. Figure 2.11 is a plot of fractional out-
of-band power, 7 (in dB), for the same three modulations versus B7; where the equivalent
lowpass bandwidth, B, is defined, as in Chapter 1, by

a) =BG(f)d
B

ey (2.132)
jes G(f)df

2.6 POWER SPECTRAL DENSITY OF CONTINUOUS PHASE MODULATION (CPM)

Not all digital modulations can be expressed in I-Q form. One very popular example of this
is continuous phase modulation (CPM) in which a carrier is frequency modulated with a
digital pulse train of the form in (2.2). Mathematically speaking, a CPM signal has the form®
CO

i(owsann Se wan?)
S(t) =FAlRe Hire ate

(2.133)
CO

= Acos (o+ 27h > ang(t — n)

n=—OO

6. See Chapter 10 for a more detailed discussion of CPM.

 Sec. 2.6 Power Spectral Density of Continuous Phase Modulation (CPM) 89

G(f)/2PT,

Figure 2.10 Equivalent lowpass power

spectral densities of various carrier
modulations (Reprinted from [17] © IEEE)

where h is the modulation index, and the phase shaping pulse, q(t), is given by
t
q(t) a r(t)dt (2.134)

with r(t) the frequency pulse, and T the signaling interval. Typically, h is chosen as a
rational number, that is, h = J/M in which case the data symbols {a,} take on values +1,
+3,...,2+(M — 1) and the number of states in CPM is finite. Furthermore, r(t) is assumed
to be of finite duration, that is, r(t) is nonzero only in the interval 0 < ¢ < LT where L is
the pulse length in symbols. The case L = 1 is referred to as full response CPM whereas
the more general case of L > | wherein the data symbols overlap is called partial response
CPM. Also r(t) is assumed normalized such that

ra |
/ r(t)dt = = (2.135)
5 4
The special case where L = | and the frequency pulse r(t) is rectangular, namely
Ae
raajwe
0;
Osis?
otherwise
(2.136)
is referred to as continuous phase frequency-shift-keying (CPFSK). The special case of
CPFSK for which h = 0.5 corresponds to MSK which, as we have already seen, has an
I-Q representation.
 90 Power Spectral Density of Digital Modulations Chap. 2

(dB)
POWER,
OUT-OF-BAND
FRACTIONAL
n

-60 : :
0.01105 040/.151120..25 30 35. 400452 60, Perec tt Peto
of various carrier modulations (Reprinted
Bly from [17] © IEEE)

To compute the power spectrum of a full response CPM signal with rational modu-
lation index, we make use of the signal representation due to Rimoldi [13] which can be
written in the form
s(t) = ARe Pal ae) = Acos (Wet + 2a fut + W(ts un, On)) (2.137)

where

fo ==—h ance
fu aT 2.138
(2.138)

and W(t; Un, On) is referred to as the physical tilted phase which, for M even and 0 < t < T,
is given by

W(t; Un, On) = 270hon + 4rthung(t) + w(t) (2.139)

Here uy is a translated and scaled version of ay, that is, un = (a, + M — 1)/2 so that up,
takes on only nonnegative values, namely, 0, 1, 2,..., M — 1 and
t
w(t) =mh(M — le — 2rh(M — 1)q(t) (2.140)

hence, the notion of tilted phase. Finally, o, denotes the state of the modulator in the n™
transmission interval and the transition between states is described by

On+1 = Rg [Un + On) (2.141)

Sec. 2.6 Power Spectral Density of Continuous Phase Modulation (CPM) 91

where Ry[X] denotes the remainder of X after division by M, and

ee l.c.m.(J, M)
(2.142)
J
is the number of modulator states with l.c.m. denoting “least common multiple.”
The state sequence {o,} behaves as a Markov sequence with S$ states and thus is
characterized by an S x S probability transition matrix P = {p;,}. It can be shown [14] that
all of the elements of P have value 1/S and thus the stationary probabilities, {p;}, also all
have value 1/S. Following a procedure analogous to that in Section 2.2, Biglieri and Visintin
[14] have computed the equivalent lowpass PSD, G(f). Omitting the details, we present the
following result:

OD aa, De ee
k=0
(2.143)
2 2 Pale,
ae (f—fu) = Dyel hk “Na 1A
2 ai(f)

where

BM) =F |Aci ir ana tet) (2.144)

Note the absence of a discrete spectrum.
For the special case of CPFSK, q(t) =t/2T and consequently, w(t) =0. Thus,
(2.144) becomes

ax(f) = Ao (fT — hk) eI S~ fT Ak) (2.145)

where
sina(f — fH)T
A T) = AT —————_ 2.146
ott) af — fut ane
Substituting (2.145) together with (2.146) into (2.143) gives

Wintec s.2
GN=z E d A(T — mn)
, | Met
ae El =j2e(f—fet rr
- Re |: 2. X e j2mhka.(
o( f ¢T — hk ) (2.147)

1 M-1
Gees ge Oe A PT Fe Ik)

M 1=0

Finally, letting h = 1/2 (ie., J = 1 and M =2), T = Tp, hate. (2? = A2/2, (2.147) reduces,
after much manipulation, to
32PT, cos’ 2nfTp
G(f)= 5 ae (2.148)
1 (1 — 16f27;)
which together with (2.131) yields (2.127).
 92 Power Spectral Density of Digital Modulations Chap. 2

PROBLEMS
2.1 Starting with the basic definition of the PSD for an arbitrary random process m(t), namely
|
Sm(f) = pies
lim E}—
|Ts ||Mo(f)I’
o( f)|

where Mo(f) is the Fourier transform of a 7o-sec observation of an ensemble member of

m(t), derive Eq. (2.15) corresponding to m(t) in the form of a binary pulse train as in (2.2).
For convenience, let 7) correspond to an integer number of bits over the interval -NT <t <
(N + 1)T, that is, 77 = (2QN + 1)T.
2.2 Consider a binary random pulse train, m(t), such as that characterized by (2.2) together with
(2.18). Suppose now that m(t) is periodically pulsed on (for Tz sec) and off (for T, — Ty sec)
as in Figure P2.2, thus producing the random process, z(t). Assume that Ty and T, are integer
multiples of T. Using the approach of Problem 2.1, find the PSD of z(t). How does it compare
to the PSD of the nonpulsed modulation as given by (2.15)?
2.3 Consider a deterministic +1 periodic waveform w(t) as shown in Figure P2.3a. Suppose now
that w(t) is periodically pulsed on (for Tz sec) and off (for T, — Tg sec) as shown in Figure
P2.3b, thus producing the random process, z(t). Assume that Ty and T, are integer multiples of
T. Using the approach of Problem 2.1, find the PSD of z(t).
2.4 For the purpose of frame synchronization, a deterministic +1 waveform (sync pattern) w(t) is
periodically inserted into a binary data pulse stream m(t), thus producing the waveform z(t) as
shown in Figure P2.4. Evaluate the PSD of the synchronized data stream z(t). Note that z(t)
can be expressed as the sum of two random processes, one in the form of z(t) in Problem 2.2
and one in the form of z(t) in Problem 2.3.
2.5 An FSK signal is characterized in the n' transmission interval nT =t <(n-- ll by,

Figure P2.2

Figure P2.3a
Problems 93

Aw
m(t) = Acos |
(0+ an) PSE a

where a, takes on values of +1 with equal probability and the 6,’s are uniformly distributed in
(—z, m) and are independent of each other.
(a) Find the time-averaged autocorrelation R,,(T) & (mime +>) where the statistical aver-
age is taken over the 6,,’s and the data sequence {ay}.
(b) Compute the PSD corresponding to your answer in part (a).
2.6 Repeat Problem 2.5 for the case of an M-FSK signal where now a, takes on values of +1,
+3,...,+(M — 1) with equal probability.
2.7 A binary-valued (+1) periodic sequence {d,} is defined by the property dp+x = dk,k =1,2,...
where p is the number of sequence positions in a single period.
Dp
(a) Show that the autocorrelation function of {d,,}, namely, Ry(k) & 2 YS dndn+k, is given by
n=l

A(k) — D(k)
= A(k) + Dik)
where A(k) is the number of sequence positions in which the two sequences agree in
polarity and D(k) is the number of sequence positions in which they disagree.

-sync “| data —-} sync + data aan sync

Figure P2.4
94 Power Spectral Density of Digital Modulations Chap. 2

(b) A pseudonoise (PN) sequence is a particular type of binary-valued periodic sequence that
has the following additional properties:
1. The periodp is equal to 2% — 1 (N integer).
2. Ina single period of the sequence, the number of —1’s exceeds the number of 1’s by
one. :
3. The product of a PN sequence and a cyclic shift of itself produces another cyclic shift of
the original PN sequence. Mathematically speaking, {dx} x {dk+m} = {dx+i} where m is
in general unequal to /.
Show that the autocorrelation function of a PN sequence is given by

ile k= 0
Rak) =| 1, REST
Vee be oscrea
together with Ra(k) = Ra(k +np), n= +1, +2,... That is, the autocorrelation of a PN
sequence is a periodic (with period p) two-level discrete function of the discrete shift
variable, k, possessing one value for k = 0 and another value for all other k 4 0 within
the period.
(c) Consider a binary pulse stream d(t) generated by the PN sequence defined in part (b)
where the pulse shape p(t) is a unit rectangular pulse of width T. Find the autocorrelation
function Ry(t) of the waveform d(r).
(d) Find the PSD of the PN waveform of part (c). Is it a continuous spectrum, a discrete
spectrum, or a combination of both?
2.8 Consider a binary random pulse train, m(t), such as that characterized by Eq. (2.2) together
with (2.18). After each N data bits, a parity check bit bit, b;, is now inserted into the sequence.
Thus, the sequence that generates m(t) would appear as ...d_y, d_N+1,---, Ga{nO = oad
cop Os QUSOIN AA AONEDy 0 an CPN GDo sou JUNE jth parity check bit, b;, is related to the previous
N data bits by

N
bj =- I] aiN+k
k=1
Compute the PSD of the resulting m(t). Does it have any discrete components?
2.9 Consider a binary random pulse train, m(t), such as that characterized by Eq. (2.2) together
with (2.18). This modulation is transmitted over a multipath channel that produces a single
delayed and attenuated reflection which adds to m(t). Thus, the received modulation r(t) can
be expressed as r(t) = m(t) + am (t — Ty) where a denotes the attenuation and T, denotes the
delay. Compute the PSD of r(t).
2.10 A digital pulse stream generated by a binary, zero mean, WSS sequence is impulse sampled at
arate 1/7*. The resulting modulation, m*(t), can be written as

[oe]

m*(t)=m(ti(t)=m(t) S* 3 —iT*
—19)
i=—0o

where m(t) takes the form of (2.2) together with (2.1) and the epoch fo is a uniformly dis-
tributed random variable in the interval (—7*/2, T*/2) which is independent of the data se-
quence.
(a) Find the time-averaged statistical autocorrelation function of m*(t), namely R(t) A
(m*Om*e AF 7))where the statistical average is taken both over the data sequence and
the random epoch fo. Hint: Use the Poisson sum relation of (2.12) to express the doubly
infinite train of impulse functions as a doubly infinite train of exponentials.
Problems 95

(b) Show that the PSD corresponding to Rm (t) of part (a) can be expressed as

1 ee 1
Sin® (f) a a ) iS;ee (s= iz)
ia
1=—CO

where Sj, (f) is the PSD of the unsampled modulation as in (2.15). That is, the PSD of
a uniform impulse sampled digital pulse stream is a doubly infinite sum of translations of
the PSD of the unsampled modulation where the frequency interval between translations is
equal to the sampling frequency.
2.11 Suppose now that in Problem 2.10 the sampling pulse train is nonideal in that i(t) is now given
by

i@t)= )) f@-iT*1)
where f(f) is a narrow pulse with unit area. Show that analogous to the result in part (b) of
Problem 2.7, the PSD is now given by

Sah)
4
* ew

(C4) eknie T* ‘sme faisT*

l —_— — L aie,

where F(f) is the Fourier transform of f(t).

2.12 A sinusoidal carrier is phase modulated with a binary data-modulated subcarrier thus forming
the signal

s(t) = V2P sin (@ct + Bym(t) sin (Wsct + Osc) + Oc)

where 0 < f,, < 2/2 is the modulation angle, w, and ws. are the carrier and subcarrier radian
frequencies, respectively, and 6,, 9s. are the carrier and subcarrier phases which are uniformly
distributed in (—7z, 2) and independent of each other.
(a) Using simple trigonometry, express s(t) in terms of its unmodulated and modulated com-
ponents at the carrier frequency, and even and odd harmonics of the subcarrier frequency
above and below the carrier frequency.
(b) Using the form of s(t) found in part (a), find its PSD. Identify the various components of
the PSD as being discrete or continuous and the average power of each.
2.13 Suppose that in Problem 2.12 a unit square-wave subcarrier were used instead of the sinusoidal
one. As such the signal is now of the form

s(t) = V2P sin (wet + Bmm(t)Sq (@sct + Osc) + Ac)

where Sg (Ws-t +s) is as illustrated in Figure P2.13. How do your answers to parts (a) and
(b) of Problem 2.10 change?

Sq(,,t
+ 9,.)
na-@0 se

Figure P2.13
 96 Power Spectral Density of Digital Modulations Chap. 2

2.14 Consider a sinusoidal carrier that is phase modulated by two binary data modulations. One
is directly modulated onto the carrier with a modulation angle of 2/2 radians. The other is
first multiplied by the first one and then applied to a sinusoidal subcarrier which is then phase
modulated onto the carrier with a modulation angle f,,. The resulting signal is of the form

s(t) =V/2P sin (ct a Smit) + Bmmy(t)m(t) sin (Wsct + Osc) + 6.)
(a) Using simple trigonometry, express s(t) in terms of its various components at the carrier
frequency, and even and odd harmonics of the subcarrier frequency above and below the
carrier frequency. Are there any discrete (unmodulated) components in this signal? If so, at
what frequencies?
(b) Using the form of s(t) found in part (a), find its PSD. Identify the various components of
the PSD as being discrete or continuous and the average power of each.
2.15 Consider a staggered quadriphase pseudonoise (PN) modulation where identical (except for
the 1/2-chip delay between them) PN waveforms (see Problem 2.7) modulate each of the
quadrature carriers. The mathematical form of such a modulation is

m(t) = V2P [d(t) cosw-t — d (t — T/2) sinawct]

where , is the radian carrier frequency, T is the chip time interval, and d(t) is a PN wave-
form that has a statistical autocorrelation function Rg(t) as determined in Problem 2.7. For
simplicity, assume that the code sequence period, p, is very much greater than unity.
(a) Find the statistical autocorrelation function R,,(t) of the cyclostationary process m(t)
defined above.
(b) Find the corresponding PSD S,,(f).
(c) Discuss how the PSD found in part (b) is different from that corresponding to a staggered
quadriphase PN modulation with different and independent PN sequences modulating the
quadrature carriers, that is,

m(t) = V2P [d)(t) cos @t — dz (t — T/2) sin wet]

where d}(t) and d2(t) are still assumed to have the same chip period 7.
2.16 Starting with (2.53) and the properties of the Miller Code as described by (2.69), (2.70) and
(2.73), derive its PSD, namely, (2.74).
2.17 Starting with the general form of the equivalent lowpass PSD for a CPFSK signal as given by
(2.147), derive the equivalent lowpass PSD for the special case of MSK as given by (2.148).

REFERENCES
1. Jolley, L. B. W., Summation of Series, New York: Dover Publications, 1961.
2. Titsworth, R. C. and L. R. Welch, “Power Spectra of Signals Modulated by Random and Pseu-
dorandom Sequences,” Pasadena, CA: Jet Propulsion Laboratory, Technical Report No. 32-140,
October 1961.
3. Hecht, M. and A. Guida, “Delay Modulation,” Proceedings of the IEEE, vol. 57, no. 7, July 1969,
pp. 1314-16.
4. Deffebach, H. L. and W. O. Frost, “A Survey of Digital Baseband Signaling Techniques,” NASA
No. N71-37703, June 1971.
5. Simon, M. K., K. Tu and B. H. Batson, “Effects of Data Asymmetry on Shuttle Ku-Band Com-
munications Link Performance,” JEEE Transactions on Communications, vol. COM-26, no. 11,
November 1978, pp. 1639-51.
References 97

6. Nguyen, T. M., “The Impact of NRZ Data Asymmetry on the Performance of a Space Telemetry
System,” JEEE Transactions on Electromagnetic Compatibility, vol. 33, no. 4, November 1991,
pp. 343-S0.
. Nguyen, T. M., “Space Telemetry Degradation due to Manchester Data Symmetry Induced Car-
rier Tracking Phase Error,’ JEEE Transactions on Electromagnetic €ompatibility, vol. 33, no. 3,
August 1991, pp. 262-68.
. Simon, M. K. and D. Divsalar, “Spectral Characteristics of Convolutionally Coded Digital Sig-
nals,” JEEE Transactions on Communications, vol. COM-28, no. 2, February 1980, pp. 173-86.
. Larsen, K. J., “Short Convolutional Codes with Maximal Free Distance for Rates 1/2, 1/3, and
1/4,” IEEE Transactions on Information Theory, vol. IT-19, no. 3, May 1973, pp. 371-72.
. Odenwalder, J. P., “Optimal Decoding of Convolutional Codes,” Ph.D. dissertation, University of
California, Los Angeles, 1970.
. Doelz, M. L. and E. H. Heald, “Minimum-Shift Data Communication System,’ U. S. Patent no.
2,977,417, March 28, 1961 (assigned to Collins Radio Company).
. Amoroso, F., “Pulse and Spectrum Manipulation in the MSK Format,” JEEE Transactions on
Communications, vol. COM-24, no. 3, March 1976, pp. 381-84.
. Rimoldi, B. E., “A Decomposition Approach to CPM,” JEEE Transactions on Information The-
ory, vol. 34, no. 3, March 1988, pp. 260-70.
. Biglieri, E. and M. Visintin, “A Simple Derivation of the Power Spectrum of Full-Response CPM
and Some of Its Properties,’ JEEE Transactions on Communications, vol. 38, no. 3, March 1990,
pp. 267-69.
. Lindsey, W. C. and M. K. Simon, Telecommunication Systems Engineering, Englewood Cliffs,
NJ: Prentice Hall, 1973. Reprinted by Dover Press, New York, NY, 1991.
. Simon, M. K. and D. Divsalar, “Spectral Characteristics of Convolutionally Coded Digital Sig-
nals,” Pasadena, CA: Jet Propulsion Laboratory, JPL Publication 79-93, August 1, 1979.
. Simon, M. K., “A Generalization of Minimum-Shift-Keying (MSK)-Type Signaling Based Upon
Input Data Symbol Pulse Shaping,” JEEE Transactions on Communieations, vol. COM-24, no. 8,
August 1976, pp. 845-56.
 Chapter 3
Scalar and Vector
Communications Over
the Memoryless Channel

In the previous chapters, a system level description of a communication link was presented
and the power spectral densities of various modulation techniques were computed. Modula-
tion and demodulation were discussed without much detail into their respective operations.
The various tasks performed in an IF demodulator can be grouped into two main functions,
namely, synchronization and detection. For the purpose of this chapter, we assume perfect
synchronization (or timing) and concentrate on the basics of detection and hypothesis test-
ing. The degradation due to imperfect synchronization on the detection performance will be
the topic of a future chapter.
Information about the world is acquired by observation and physical measurement.
All measurements are subject to error. All efforts to completely eliminate this error finally
reach a bound set by nature’s underlying chaos, which infects all observations of physical
phenomenon with an innate uncertainty (error). Acquiring information about any physical
system involves making decisions. Generally, we must decide which of a set of hypothe-
ses (statements) best describe the system insofar as observations permit one to judge. Such
hypotheses exist and, in the context of some theory, signify the state of the system. Such
a theory requires modeling the sources of uncertainty (error) that corrupt the observations.
The irreducible minimum component of error in making decisions is characterized by ana-
lyzing the decision rules that minimize some convenient measure of the average amount of
error.
The transportation of information requires a communication system. Such a system
corrupts the information by the introduction of random background noise. The need to fer-
ret weak signals out of such noise has called for the development of a Theory of Signal
Detection. This need arose during World War II in the context of determining the presence
or absence of an enemy emitter. At that time, it was realized that decisions made in the face
of uncertainty could be treated by the Theory of Signal Decisions, commonly known as Hy-
pothesis Testing. Such an elegant theory had been fully developed in the 1930s by Neyman
and Pearson [1,2], Wald [3] and others. The motivation of this theory came from “An Essay
Toward Solving a Problem in the Doctrine of Chances” which appeared in the Philosophical
Transactions of The Royal Society of London in 1763 [4]. Therein, the insightful Reverend
Thomas Bayes treated the problem of making decisions under conditions of uncertainty, and
Scalar and Vector Communications Over the Memoryless Channel 99

he recommended selecting that hypothesis for which the a posteriori probability is great-
est. This is called Bayes strategy, and the general formulation of hypothesis testing is often
called “Bayesian.”
In the context of communication of information, the abstract set of hypotheses
{Hm; m=0,...,M — 1} takes on the physical meaning of messages. For convenience,
we denote this message set as {mm, m=0,1,..., M — 1} such that hypothesis H,, has a
one-to-one correspondence with message m,,. We will apply the theory of hypothesis test-
ing to the problem of communication of information with minimum probability of error.
Our top-down approach to demonstrating the theory of signal design (message represen-
tation) and signal detection (message decision) is to apply Statistical Decision Theory to
solve this communication problem. As we shall see, it is rich in results of great practical
interest.
In the communication of information, the message-bearing signal is observed as val-
ues of the voltage r(t) appearing at the terminals of an antenna during a known interval
[0, 7] of observation. This voltage waveform is a random process described by probabil-
ity density functions that embody what is known about the message-bearing signals to be
detected and the statistical properties of the noise that corrupt them. Statistical decision the-
ory, as applied to the communication problem, shows how best to process that input voltage
r(t) in order to decide among several hypotheses (messages) about what r(t) contains. As
we shall see, it prescribes certain kinds of processing, manipulations, and appropriate log-
ical operations that end in decisions. The probabilities of error that these decisions incur
reveal the ultimate limits on the detectability (Probability of Message Error) of the message
embedded in r(f).
Using the top-down approach to decision making depicted in Fig. 3.1, we first con-
sider the scalar communication problem in which the observation is a discrete random vari-
able taking on values r = ro, 71,...,77—1. At first, this may seem to be a gross oversimpli-
fication of the problem; ironically enough, we shall see, in Chapters 4, 11 through 14, that
it is indeed the model that one must ultimately characterize in order to evaluate the perfor-
mance of an uncoded (quantized) and of a coded digital communication system. Next, we
model our observation as a continuous random variable taking on values r = peR (R is the
set of real numbers). A generalization of this model allows for our observation to be a ran-
dom vector taking on values r = pe R™ (R% is the set of N-tuples of real numbers). This
model allows for evaluation of the performance of most uncoded communication systems,
as will be discussed in future chapters. Moreover, the vector model offers great insight into
the detection problem wherein the transmitted messages are represented by waveforms. It
will also be shown that the waveform communication problem can be reduced to a corre-
sponding vector problem with the appropriate transformation, and hence the vector problem
serves as a platform from which the waveform problem can be approached. Figure 3.2 il-
lustrates the detection problem for all cases of interest. The information source outputs one
of M messages, which is transmitted using a scalar, a vector, or a waveform. The channel
is represented by the probabilistic transition mechanism which outputs an element of the
observation space. Finally, the decision rule maps the observed element into one of M hy-
potheses (messages), according to a certain prescribed criterion. A communication link from
a systems engineer perspective is shown in Fig. 3.3. This perspective has been discussed in
great detail in Chapter 1. For our purpose, it suffices to note that depending on the appli-
cation, an appropriate architectural level can be chosen to model the channel as a discrete
memoryless channel (DMC) or a continuous memoryless channel (CMC).
 100 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Top Down

{s,;} KNOWN SCALAR

n\S A RANDOM VARIABLE

CASE 2:
{s,;} KNOWN VECTORS
n lS A RANDOM VECTOR

CASE 3:
{s,(t)} KNOWN WAVEFORMS
n(t) IS A RANDOM PROCESS

Figure 3.1 Scalar, vector, and waveform communications (top-down approach)

PROBABILISTIC OBSERVATION
INFORMATION
TRANSITION SPACE
SOURCE DECISION
MECHANISM
RULE

Figure 3.2 Elements of the detection system

3.1 SCALAR COMMUNICATIONS OVER THE MEMORYLESS CHANNEL

In scalar communications, the transmitter transforms each message m into a distinct voltage
s for transmission. The channel is assumed to corrupt the transmitted voltage s by adding
a random variable n representing the noise. The receiver observes r = s + n and estimates
the transmitted message m. For an additive Gaussian noise channel, the random variable n
has a normal (or Gaussian) distribution. The channel can be modeled differently depending
on the application. A more general channel model is given by r = f(s,n) in which the
received voltage is a nonlinear two-dimensional function of both the signal and noise. An
example of a nonlinear channel is the case where the signal is amplified by a Travelling
Wave Tube (TWT) amplifier operating in the saturation mode. Another example is the
fading channel where the receiver observes r = as +n and a is a random variable that
models signal attenuation due to the fade. The attenuation itself can be time-varying and
frequency dependent.
As previously mentioned, the detection problem is best formulated as a hypothesis
testing problem in which the function of the receiver is to decide which of the M hypotheses
(messages) is most likely. Letting H; denote the i‘” hypothesis, then

mo — Ho:r=so+n
m, — M:r=s;
+n
GaD)

mMuM—1 — Hy-1:r=SsSm—-1+n
 OjU|WW UOITe
doy umogq

UOIJEWO}U|
HUIS

UOIeWJ

Oyu]gounos
UONeWOJU jeubis JeuBis
gounos Joyesauayg 10198}9Q
YUIS

ZUMOUS (ug {(a|'s)g}

UONeUWO}U

DANSI SWIa}SAG-¢"E
Ja9uIsUd datdadsiod

101
102 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

for the additive noise channel where s; is a known deterministic voltage representing the i’
message and n is in general a zero-mean noise with a known probability density function.
The function of the receiver is to operate a decision rule that maps the observed random
variable r into one of the M hypotheses. Typically, the optimization criterion will be the
minimization of the message error probability, but another criteria such as minimizing the
average cost, as in Bayesian detection, can be chosen and is discussed in more detail in
Appendix 3A in the case of vector communications.

3.1.1 Scalar Communications Over the Discrete Memoryless Channel

Let’s consider first the discrete memoryless channel (DMC) which corrupts the signal by
adding a discrete random variable n. In this case, the output r is constrained to assume one
of a set of J discrete values. We assume that the M a priori (before reception) message
probabilities P(m;) are known. These are the probabilities with which each message gets
transmitted over the channel. In this case, the channel is described by MJ conditional
(transition) probabilities p;,; = P(rj/m;) where M is the number of messages and J the
number of channel observable outputs, as depicted in Fig. 3.4. Often, these probabilities are
grouped into a single M by J matrix P given by

P0,0 PO0,1 a POV"

P1,0 PAY ae Plena
i : , ; : (822)

JOM MO) JOYS P99 OS

Examples of such channels are depicted in Fig. 3.5. In the binary channel (BC), M = J =2
and the two messages mo and m, experience in general different transition probabilities. On
the other hand, they experience identical transition probabilities in the binary symmetric
channel (BSC). In other channels, the message either is received correctly or is erased.
Such is the case in the binary erasure channel (BEC). Other examples include the noiseless
channel in which p;,; = 1 fori = j and 0 fori ¥ j, the useless channel in which pj, ;= p;
for alli and j, and finally, the M-ary symmetric channel defined by pj, ;= pj,i-
For any discrete memoryless channel (DMC), we would like to derive the optimum re-
ceiver that minimizes the probability of error. That is, upon observing the received symbol
r;, the receiver should decide which message m; was transmitted in such a way as to max-
imize the probability of correct decision, P(C). We therefore expect the optimum receiver
to outperform any other receiver after a “long” series of transmissions. Using the method of
conditional probabilities, the joint probability of transmitting m; and receiving rj is given
by
P(m;, rj) = P(mj)P(rj/mi) (3.3)

where P(mj) depends solely on the message source. The total probability of receiving r j AS
the sum of all joint probabilities P(m;, r;) over all possible transmitted messages, that is

M-\
P(rj)= » P(mi, rj) (3.4)
i=0

Using Bayes rule [5,6], the a posteriori (after reception) probability that m; was transmitted
Sec. 3.1 Scalar Communications Over the Memoryless Channel 103

M9
m4

DISCRETE
MEMORYLESS |
CHANNEL
(DMC)

U2
™M-1 4

M MESSAGE J OBSERVABLE
INPUTS OUTPUTS
(a)

ah PL t9/mo |

m4

ney

ba PL y.4/ my]
Figure 3.4 The discrete memoryless
(b) channel (DMC)

given that r; is received (observed) is

P(rj/mj)P (mj)
P(mj/rj) = P(r;)
oe
(

Before transmission, each message m; has an a priori probability P(m;) of being

transmitted. Therefore, if the receiver were to decide on which message was transmitted be-
fore making any observation, the message with the highest probability of transmission [that
is, highest a priori probability P(m;)] would be chosen. After observing r;, the probabil-
ity that m; was transmitted is altered from the a priori probability P(mm;) to the a posteriori
probability P(m;/r;). We expect that a “smart” receiver will exploit this change to make a
better decision on the transmitted message.
On the basis of this set of known a posteriori probabilities, the goal is to find out
how does the optimum receiver make a decision (that minimizes the probability of error)
regarding the transmitted message? That is, how does the receiver map the observed symbol
r; onto the message input set {m;} such that each possible received symbol is assigned one
and only one element (but not necessarily a distinct element) from the message set? Let
104 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

m9
Poo 1-p
Mo uy Mo uf)

P10 p
B

Poy P i,
m4 4 m4 "4
Py a 1-p i

(a) BINARY CHANNEL (b) BINARY SYMMETRIC (c) BINARY ERASURE

(BC) CHANNEL (BSC) CHANNEL (BEC)

Figure 3.5 Examples of discrete memoryless channels: (a) Binary channel (BC), (b)
Binary symmetric channel (BSC), (c) Binary erasure channel (BEC)

m(r;) denote the transmitted symbol attributed to observing r;; for example, the receiver
decides that m; was transmitted if r3 is received, that is, m(r3) =m. Then, the conditional
probability of being correct given rj is observed is just the probability that m(rj) was
actually transmitted, that is

P(C/rj) = P(m(rj)/rj) (3.6)

The conditional probability P(C/r;) is then maximized by choosing m(r j) to be that mem-
ber of the set {m;} with the largest a posteriori probability. This maximum a posteriori
(MAP) probability decision rule determines the optimum receiver. The unconditional prob-
ability of a correct decision, P(C), is the weighted average of the above conditional proba-
bilities, that is
J=1
F(Cy= a RCir PG) (3.7)
j=0
and the probability of error is therefore

P(E) =1—P(C) (3.8)

Since the probabilities P(r;) are positive and independent of the decision rule, the sum in
(3.7) is maximized if each of the P(C/rj;) terms is maximum. Since P(r;) is independent
of the transmitted message, then using (3.5), the decision rule simplifies to: set m(rj) = mx
if and only if (iff)

P(m
x) P(rj/my) = P(mj;)P(rj/m;) for alli (3.9)

and is referred to as the maximum a posteriori (MAP) decision rule. In the particular case
where the a priori probabilities are equal, the decision rule simplifies further to: set m(r p=
my, if and only if

P(rj/mg) > P(rj/mj) for alli (3.10)

This special case of the MAP decision rule is referred to as the maximum likelihood (ML)
decision rule. Thus, we conclude that the optimum receiver is a probability computer, in
particular, an a posteriori probability computer.
Sec. 3.1 Scalar Communications Over the Memoryless Channel 105

Figure 3.6 Channel of Example 3.1

Example 3.1
Consider the mathematical model of a discrete communication channel having 3 possible input
messages {a, b,c} and 3 possible output symbols {1, 2, 3} as shown in Fig. 3.6. The channel
model is completely described by the set of nine conditional probabilities

Pia) 016) P:2/a)—03" PB/a)—01

P(1/b) =0.1 P(2/b)=0.5 PG/b) =0.4
Pave) =O019PC/c)=01 PG/c)=08
which specify the probability of receiving each output symbol conditioned on each input mes-
sage. Assume that we know the set of the three a priori probabilities P(a), P(b), and P(c) with
which the input messages are transmitted

}X@)) = Vs 12) =O) IAC

1. Given that the received symbol is 1, what is the decision of the optimum receiver regarding
which message is transmitted?
Using Bayes rule, we have

_ P(1/a) PG)
P(a/l)= mak Veh W

where
P(1) = P(1/a)P(a) + P(\/b)P(b) + P(/c)P(c)

= (0.6)(0.3) + (0.1)(0.5) + (0.1)(0.2) = 0.25

Therefore
(0.6)(0.3)
P(a/1) = ———___—
(0.25) = 0.7
().72

Similarly
P(1/b)P(b) — (0.1)(0.5)
ese) 1X) (0.25)

P(1/c)P(c) (0.1)(0.2)
; aes = = 0.08
ey P(1) (0.25)
Since P(a/1) is greater than both P(b/1) and P(c/1), the optimum receiver will decide that
message a was transmitted given that the received symbol is 1.
2. Repeat (1) for the received symbol 2.
Similarly

P(2/a)P(a)
P(a/2)= P(2)
106 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

where
P(2) = P(2/a)P(a) + P(2/b)P(b) + P(2/c)P(c)

= (0.3)(0.3) + (0.5)(0.5) + (0.1)(0.2) = 0.36

Therefore :
(0.3)(0.3) _= 0.25
P(@j2)= 0.36

and
P(2/b)P(b) (0.5)(0.5)
= = = ()'6944 =.
Fey? P(2) (0.36)

P(2/c)P(c) (0.1)(0.2)
= = —__
— _=()'0555...;
Rte) P(2) (0.36)
If the received symbol is 2, the receiver decides that message b was transmitted.
3. Repeat (1) for the received symbol 3.

EGS)
_i= —
P@/a)P(a)
Pa)

where

PG) =PGC/a)P@) +P C/b) PO) + PC/oe) Fe)

= (0.1)(0.3) + (0.4)(0.5) + (0.8)(0.2) = 0.39

Therefore

P\(a/3) = Oe) = (0 ,09Ree

(0.39)

P(3/b)P(b) (0.4)(0.5)
P(b/3) PG)
= —___— 0.39) _ =
__= ——— 0.5128
0.5128...

Be P38 eR) af (0.8)(0.2) ¥

ACB = PQ) = 630 OMNO2=e.

The receiver will decide that message b is transmitted if the received symbol is 3. Note from
(2) that message b is also decided when symbol 2 is received. Therefore, message oe 9? is never
chosen by the MAP receiver. This is acceptable since each received symbol is assigned a unique
transmitted message, but not necessarily a distinct message. This problem illustrates that in
some DMCs, the MAP decision rule, which minimizes the average probability of error, does
not always decide on all messages in the transmitted set.
4. Given that the received symbol is 2 and the receiver decides that message b was sent, what
is the probability that the decision is correct, that is, that message b was actually transmitted?

0.25
P (correct decision/2) = P(b/2) = 0.36 = 0.6944...

5. What is the probability of error P(E) of the optimum receiver?

P (correct decision/1) = P(C/1) = P(a/1) = au =O

P(correct decision/2) = P(b/2) = 0.6944...

P(correct decision/3) = P(b/3) = 0.5128...

Sec. 3.1 Scalar Communications Over the Memoryless Channel 107

MESSAGE SCALAR SCALAR

SOURCE TRANSMITTER RECEIVER

(a)

Ss

{m;} — {si}

Figure 3.7 Continuous memoryless channel (CMC) model for scalar communications

Therefore, the probability of correct decision is

P(C) = P(C/1)P() + P(C/2)PQ) + P(C/3)P@)
= (0.72) (0.25) + (0.69) (0.36) + (0.51) (0.39) = 0.6273...
and the probability of error P(E) is thus given by

PCE) =1 = P(E) =) — 0:6273'= 013727.

3.1.2 Scalar Communications Over the Continuous Memoryless Channel

In most channels, the noise n is a continuous random variable with probability density
function (pdf) fy, (a).! As a result, the received random variable r is also of the continuous
type and the continuous memoryless channel (CMC) in the scalar case is modeled as shown
in Fig. 3.7. For a particular observed value of r, say r = ~, assume that the receiver sets
m(p) =m;. Then, the conditional probability of correct detection is just the probability that
m; was actually transmitted given that r = p is observed, that is

P(C/r = p) = P(m() = m;/r = p) = P(m;/r = p) G.01)

So the probability of being correct, P(C), is thus
[o,¢)

P(C) a P(C/r = p) fr(p)dp (3.12)

—0o

1. Strictly speaking, the subscript n in f,(a@) denotes the random variable while the argument is an independent
variable (in this case, w). Sometimes, we will denote the independent variable by the same symbol as the random
variable, i.e., fn(n).
108 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Select m(p) =m;

PROBABILITY iff
COMPUTER
max {Plm,|r = p]} = P[m|r = p]
P(my-1 pe p)

Figure 3.8 Optimum receiver for scalar communications

where f;(o) is the probability density function of the received voltage r. Here again,
P(C) is maximized when P(C/r = p) is maximized since f,(p) > 0. But, P(C/r = p) =
P(m;/r = p) is maximized when m; is actually transmitted. As in the discrete memory-
less channel, the optimum receiver maps p into m; for which the a posteriori probability
P(m;/r = p) is maximum. Therefore, the decision rule is: set m(o) = m; if and only if

P(mi/r = p) = max {P(me/r = p)} for all k (3.13)

Thus, in cases of both discrete and continuous noise, the optimum receiver is a “prob-
ability computer” as depicted in Fig. 3.8. The optimum receiver computes the a posteriori
probability, given that r = p, for all messages in the set and decides on the message with the
largest computed a posteriori probability.

3.1.3 Optimum Scalar Receivers

So far, we have shown that the optimum receiver is a probability computer that computes
all the a posteriori probabilities. Indeed, this is the case for any channel with arbitrary
signal and noise statistics. We would like to know if the receiver can be simplified for some
specific channels and how it can be mechanized in a relatively “simple” manner. We will
first consider the case of binary communications to understand the basic fundamentals, and
then generalize the results to M-ary communications.

Binary scalar receivers. In the binary case (M = 2), the optimum detector sets
m(p) = mo if and only if

P(mo/r
= p) > P(m/r =p) (3.14)
Using Bayes rule in mixed form, we obtain

fr (p/m) P(m;)
P(m;/r=p)= (3-05)
fr(p)
where f,(o/m;) is the conditional probability density of r given that m; was transmitted.
Inserting (3.15) into (3.14), we have the following decision rule: decide that mo was trans-
mitted if and only if

fr(p/mo)P(mo) _ fr(e/m)P(m))
fr(p) ‘i fr(e) ee
Cancelling out f;(o) from both sides, we have
mo
f-(e/mo)P(mo) = f-(e/m)P (m1) (3.17)
m\
Sec. 3.1 Scalar Communications Over the Memoryless Channel 109

f - (p/ my ) P(m4)

fr (p/mg ) Pim )

So | S{

Figure 3.9 Division of observation space in binary scalar communications

The above notation indicates that the receiver decides that message mg was transmit-
ted if the left-hand side of (3.17) is greater than its right-hand side. However, message m is
chosen if the reverse is true. In the case of equality, either message can be chosen randomly.
Figure 3.9 depicts a typical division of the observation space into two disjoint decision re-
gions, say Jp and J). If pel), then m(o) = mj, that is, m, is decided. On the other hand, if
pelo, then m(p) = mo. For p = p* (that is, p¢/p and pe/; simultaneously), both messages
are equally likely to have been transmitted and the receiver can choose either of the two
randomly. Sometimes, (3.17) is expressed as a ratio, that is

A,(p) 4
fr(p/mi) ZI POmo) _ (3.18)
f(p/mo) mo Pom)
This is referred to as the likelihood ratio and it measures which of the two messages is more
likely (in terms of larger a posteriori probability) to have been transmitted. Note that the
threshold 2* depends only on the a priori probabilities of the messages.
To further evaluate the rule, we need to make additional assumptions on the channel
and the statistics of the noise and the signal. Assuming that the signal is disturbed by
additive noise and that the latter is independent of the signal source, then if s; is transmitted,
r = :¢ is received if and only if the noise takes on the value n = p — sj, that is

fr(e/mi) = fnle — si/mi) = fnle — Si) (3.19)

where the latter equality is due to the independence of the noise from the signal. Therefore,
the decision rule of (3.17) simplifies to: set m(p) = mo if and only if

fn(e — 80)
P(mo) > fn(e — 51)
P(m1) (3.20)
Assuming further that the additive noise is Gaussian with probability density function (pdf)

exp(—x?/207)
fnr(x%) = (3.21)
V2102

where o? is the noise variance (and is assumed to be known at the receiver), (3.20) reduces
110 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

s4

Decide m (P) = my
iff
fi(p)
Aq.>a 2
Otherwise decide m,

So

Figure 3.10 Implementation of binary scalar receivers

to
eee,
exp [-| exp [-os) |
P(mo) > P(m\) (3.22)
V2102 V2n02
Note that when the noise has a nonzero mean, the latter can be added to the signals sg and
s, and (3.22) still holds for the modified signals. Taking the natural logarithm of both sides,
(3.22) can be simplified to yield

=(p — 50)? > —(p — 51)? + 20%Ine™

P(mo)
(3.23)
or equivalently

Aq £ p(so — 81) > a (3.24)

where the threshold a is given by

ga Sori vh a Bey (3.25)

This rule suggests the mechanization depicted in Fig. 3.10 in which the received
signal, p, is multiplied (correlated) with the two expected signals so and s; or the difference
signal so — s; to form Aq, and the result is compared to a threshold which is a function
of the signal powers ae en the noise power co”, and the a priori probabilities P(mo) and
P(m}). In the special case where the signals have equal power (that is, s9 = —s;) and equal
a priori probabilities, the threshold a becomes zero and the receiver reduces to multiplying
r = p with so and s; and choosing the larger of the two.
Example 3.2
Consider a binary communication link where message mo is transmitted by sending so = 0 and
message m, is transmitted by sending a random variable s with a known probability density
function. To be more specific, consider the following binary hypothesis testing problem:

Hj): mj—r=n
Ay: m,;—r=s-+n
Sec. 3.1 Scalar Communications Over the Memoryless Channel 111

where s and n are independent random variables with respective pdf’s (assume a > 0, b > 0
and a 4 b)
—aS
ON Cem eS ()
fs(8) Fa S<0O
and

be°" N>O
n(N = 4 <
Joh) fe N <0
Find the decision rule that minimizes the error probability if mo and m, are equally likely.
Using the likelihood ratio test of (3.18), the decision rule becomes

2) fr(p/m}) es
A(p)
fr(p/mo) in

as the a priori probabilities are equal. Since r is the sum of two independent random variables
when message mj, is transmitted, the conditional pdf f,(o/m},) is the convolution of the indi-
vidual pdf’s, that is

fr(p/m) = fs(p) * fn(P)

= /* geale-N be NaN
0
b
Se axb

The pdf of r, conditioned on mo, is simply

fr(p/mo) = fu(p) = be"?

Hence, the decision rule becomes

a ml
A(p)= (ise er ae
ap mo

Rearranging and then taking the natural logarithm of both sides yields

Baz Inf—], ab

M-ary scalar receivers. We have shown that the optimum M-ary receiver sets
m(p) =m; if and only if the a posteriori probability P(m;/r = p) is the largest. Using
Bayes rule in mixed form as given in (3.15) and cancelling f,-(p), the decision rule becomes:
set m(p) = m; if and only if

fr(p/mi)P (mi) = fr(e/mx)

P(mg) for all k (3.26)
Again assuming Gaussian noise with pdf given by (3.21), we get (ignoring the 1/¥V 2a
factor)
Pee Ne Le 2

P(mj;) exp "| > P(mx) exp eo] (3.27)

Simplifying, we have

207InP(m;) — p* + 2s; — s? > 207InP(m,) — p? + 2s, — sp (3.28)

112 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

4) bo

Select m (P)=m,
iff
max iGd= 95

Figure 3.11 Implementation of M-ary scalar cross-correlation receivers

So, the optimum receiver sets ™(o) = m; if and only if

ps; + bj = psx + by for all k (3.29)

where the “bias” by is given by

bp A= [207In Pom) = sz (3.30)

This rule suggests the functional diagram (mechanization) shown in Fig. 3.11, where
gi = ps; + b;. Note that the optimum decision rule is to multiply the received voltage with
each of the signals, add a bias to account for different signal powers and a priori probabili-
ties, and finally choose the largest. This receiver is linear since the decision rule is a linear
function of p and this is the result of the Gaussian assumption on the noise. For other noise
pdf’s, the receiver may indeed be a nonlinear function of p. In the special case where the a
priori probabilities are equal, the 7 In(1/M) term can be ignored and the bias reduces to
1 5
be = — 55h (3.31)
Adding (—p7/2) to both sides of (3.29) and multiplying by (—2) yields

(CN ei (3.32)
or equivalently, the receiver decides m(o) = m; if and only if

lo — s;| < |o — sx| for all k (3:33)

In other words, the receiver decides on the message m; whose voltage s; is “ closest” to the
received voltage p. This rule is referred to as minimum distance decoding and it will be used
again when considering vector communications.
Example 3.3
A system is used to transmit M equally likely messages using the voltages s; =iA, i =
0, 1,..., M — 1. The received voltage, when the i'" message is transmitted, is r = s; +n where
Sec. 3.1 Scalar Communications Over the Memoryless Channel 113

m 0
| l | l
m m m m
| lal Taal Bs M-1
| | l | |
ALS lo Shs ee | laa i
0 l A 2A | 3A | Og 6 | (m-1) A ;
| | | |
| | | |
1 1 i] i} i}

A 3A 5A 7A

2 2 22 2

Figure 3.12 Decision regions of Example 3.3

n is a zero-mean Gaussian noise with variance o*. Draw the decision regions of the optimum
receiver.
Since the noise is Gaussian, the optimum rule is minimum distance decoding as shown in Fig.
3.12. Hence, if r< A/2, decide mo. If A/2 <r < 3A/2, decide m, and so on.

3.1.4 Error Probability Performance

In this section, the error probability is used to assess the performance of the optimum
receivers for both binary and M-ary scalar signaling. In other fields such as radar, signal
acquisition, antenna pointing, and so on, other performance criteria are sometimes used in
detection and these are discussed in Appendix 3A.

Binary scalar receiver performance. We now evaluate the probability of mes-

sage error, P(E), for additive Gaussian noise as a function of the transmitter voltage levels
(so, 51) and the a priori probabilities [P (mg), P(m,)]. Without loss of generality, assume
that so > s; and that the threshold a given by (3.25) is positive. When m; is transmitted for
either i= 0 or 1, the receiver is correct if and only if r lies in region /;, depicted in Fig. 3.9.
The probability of this occurrence is

P(C/mj) = P(rel;/mi) =i) fr(p/mi)dp (3.34)

lj
and therefore

P(C)= 57P(mj)P(C/mi)= 5 P(m; | Sr(p/mi)dpe (3.35)

0) is

But. f,(0/m;) = fu(e — sj); hence

Pees 0° at
exp[— et f exp[—
exp] (eos
sy ae
m m fo ce
: a’ V 2107 : —oo ree
where a’ A a/(so — 51). The probability of error, P(E), is thus given by

P(E) =1— P(C) = P(mo) + P(m) — P(C) (3.37)

Using P(C) from (3.36), we have
(p—s0)" / fo-sy"
=a
Pe exp asa5 @ exp[— ]
Pe ey a! JV Ino2
Pn, — | .—— —oo Too
dp) 3.38)
114 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Substituting a = °° and B = >", (3.38) becomes

(ene) 2
P(E) = P(mo) (1- Paclueteyia)
ooo oe
; (3.39)
a —S] 2

« exp(—B*/2)
(m})
+ P(m 1 -|i: re
————d B

If we define the Gaussian probability integral as

© exp(—y*/2)
0
O@) (3.40)
56 /2

then using the property Q(—x) = 1 — Q(x), pore

reduces to
re
P(E) = P(mo) Q(— (3.41)
A table of the Gaussian integral is given in Appendix 3B along with an n'" order series
approximation. If the signals were equally probable, that is, P(mo) = P(m)) = 1/2, then
a’ =e)
= “OTS! and
aoe
my (3.42)
P(E)= Q(
where |x| denotes the absolute value of x. The nen value in (3.42) is necessary to handle
both cases sg > s; and s; > so. Hence, the probability of error depends only on the ratio of
the “distance” or “potential difference” between the signals to the noise standard deviation
o. If we were to choose the transmitted voltages so and s; with the additional constraint that
se aE se is fixed (that is, total power is constraint), then the optimum signal design [7] that
will minimize the probability of error is to set the voltages the furthest apart using

S| =—so ='s (3.43)

which results in

|s| Se
Pmin(E) = Q =] — Q =) (3.44)
oO

Thus, the error performance depends only on the signal-to-noise power ratio (SNR) given
by s*/o”. The larger this ratio, the smaller is the probability of message error.

M-ary scalar receiver performance. The optimum receiver in the M-ary case
divides the observation space (a line in the scalar communication problem) into disjoint
regions denoted Jo, 1), ..., —1 respectively, as shown in Fig. 3.13. If the received signal
pel, then m(p) = m; and the probability of being correct is given by
M-1
PC)= S> Pom) P(C/m) = Y* Pom) i}fr(o/midp
i=0 i=
(3.45)
Example 3.4
Compute the probability of error of the receiver in Example 3.3.
 Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 115

P(Mo) fn (Pp - So)

P(m4)fn (Pp- $4)

P(mm-1) fn (P= Sm)

So | $1 | eee | SM-1 Pp

— |
Ip ee ly cEgit FC (al emit IM-1

Figure 3.13 Partitioning of observation space in M-ary scalar communications

The probability of error is the weighted sum of the conditional probabilities of error, that is
, Mol
P(E) = — E P(E/m;)

But due to symmetry, we have

© exp(—r?/207) (=
P(E/mo)
NS = P(E
lS ==Uae|| ama re ad ie
———_—_——— dr = =)

and

P(E/m,) = P(E/m2) =---= P(E/my-2)

where
A/2 —(r — A)2/202 fore) Zo (cual 2 2
PiE/my)= | Lea | ene lex(orsbane2e ar =20(2)
—o0 JV2102 3A /2 V2n02 20

rera(3)0(8)+2(4?)0(8)=2(“)e(8)
Hence

3.2 VECTOR COMMUNICATIONS OVER THE CONTINUOUS MEMORYLESS

CHANNEL
In communication systems, each message m; is transmitted using a unique waveform 5; (tf).
We will later map the waveform communication problem into an equivalent vector problem
and show that the mapping does indeed preserve the inherent properties. Note that the signal
vectors are defined with respect to an orthonormal basis. Depending on the choice of the
latter, the dimensionality N of the problem can be reduced as discussed in Appendix 3C.
In this section, we will proceed with the vector problem and consider the vector
communication system model shown in Fig. 3.14. Here again, the source outputs one of M
messages which is mapped into its respective vector s; by the vector transmitter. We assume
116 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

VECTOR
A (p)
VECTOR VECTOR
SOURCE TRANSMITTER CHANNEL RECEIVER

Select m(p)=m;
iff
PROBABILITY
COMPUTER Nees {P(m,/ r=p)}=P(mj/ r=p)

Figure 3.15 Optimum receiver for vector communications

that the a priori probabilities {P(m;)}, the signal set {s;}, and the statistical characterization
of the channel {f;(p/s = s;)} are all known. The receiver observes the vector p and decides
which message m; was transmitted.

3.2.1 A Posteriori Probability Computer

Let P(C/r = p) denote the conditional probability of correct decision given that r= p
is observed and assume that the receiver sets m(p) = mx. Then the probability of correct
decision is the probability that m, was indeed transmitted, that is

P(C/r = p) = P(i(p) = mx/r = p) = P(m/t = p) (3.46)

The unconditional probability of correct detection is then

pcy= ff Pcir=eyflordp (3.47)

er
N— fold

where N is the dimensionality of the observation space. Since f-(@) > 0, then P(C) is
maximized when the conditional probability P(C/r = p) is indeed maximum. Hence, the
optimum receiver sets m(p) = mx if and only if

P(m,/r
= p) = P(m;/r
= p) for alli (3.48)

Here again, the optimum receiver is an a posteriori probability computer as depicted in Fig.
3.15. Using Bayes rule in mixed form, we have

Pomfr= o)= PDes= 8) P(m;j) fr =)

3.4%
Thus, the optimum receiver sets m(p) = mx when the decision function (rule)

P(m;) fr(e/S = Si) (3.50)

is maximum for i = k. This defines the maximum a posteriori probability vector re-
ceiver. Another related receiver is the maximum likelihood vector receiver which sets
m(p) = mx if and only if

Sr(p/S = Sx) = max{/r(p/s= $i)} (3.51)

Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 117

Note that the maximum likelihood receiver does not depend on the knowledge of the a priori
probabilities {P(m;)} and that if all messages are equally probable, then the maximum like-
lihood vector receiver is equivalent to the maximum a posteriori probability vector receiver.

3.2.2 Optimum Vector Correlation and Minimum Distance Receivers

When the channel transition probability densities f,(p/s = $;) are known, the maximum a
posteriori probability vector receiver can be simplified further. We will first consider the
optimum receivers in the presence of additive white Gaussian noise (AWGN) and then
investigate the receiver structures when the noise is colored.
When the signal s is disturbed by an additive N-dimensional noise vector n with zero
mean, equal variance and uncorrelated components, the noise vector is defined as “white.”
In the special case when it is Gaussian in addition to being white, the noise components
are independent and we define the channel to be additive white Gaussian noise (AWGN).
Letting n = [719, 145.0'.. nn—1]!, we have

fnj(0j) = Foams
I ox [-«2/20
Dae (3.52)
resulting in
N-1 1 N/2 eee
fn(@) = I fnj(@)j) = (=) exp |-s30 a (3.53)

where a! =[ag,a1,...,a@n_1]. For the AWGN channel, o = /No/2 where No = kT°

Joules.

Binary vector receivers in the AWGN channel. In the case of binary signal-
ing, the optimum receiver sets m(p) = mo if and only if

P(mo) fr(p/s = So) > P(m1) fr(p/s = 81) (3.54)

But since r=s; +n, f-(p/s =8;) = fn(p — Sj) as the signals and the noise are indepen-
dent. Using (3.53), the decision rule becomes (since a! @= |a|? and ignoring the 1/270?
factor)

P(mo) exp |e
lo =i] 2°
rs P(m}) exp
lo — eysil?
oe he (3.55)

Taking the natural logarithm gives

“Ss 2 mo or 2
By dena +InP(mo) = — Ieee + In P(m) (3.56)
202 m\ 20
or equivalently, the receiver sets m(p) = mo if and only if

P(m})
le —— sol’ $7 < le —§,|?
—Sil — 2071
o nf ee (3.57)

For equally probable messages, the rule becomes: set m(p) = mo if and only if

lo — Sol? <|p — sil? (3.58)

118 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

DECISION @?2
4 BOUNDARY
Ho

i (P)

© denotes vector inner product

(a) DECISION SPACE (b) RECEIVER MECHANIZATION

Figure 3.16 Decision space and receiver mechanization of Example 3.5

This rule decides on the message whose corresponding signal s is “closest” to g. This rule
leads to the notion of Minimum Distance Decoding encountered earlier. Coding theorists
work from this perspective.
Example 3.5
One of two equally likely messages is to be communicated over the AWGN channel. Assuming
that the transmitter uses the signal vectors

(i) (3)
draw the decision space and show a mechanization for the optimum receiver.
Since the messages are equally likely and the channel is AWGN, the optimum receiver employs
minimum distance decision and the observation space can be divided into two distinct regions
as shown in Fig. 3.16(a). Let on = (| £2), then the decision rule becomes

mo
i 2 > v3 a
2p So —|Sol” = 2p-s: —'|s3|
m|

or

mo
2(p1 + p2) —2 = 2(—2p; — 22) — 8
my

or

mo
pit+m= —1
m
Hence, the receiver can be mechanized as shown in Fig. 3.16(b). The vector operation p' so iS
referred to as cross-correlation in practice.

M-ary vector receivers in the AWGN channel. Simplifying (3.50) in AWGN

channels, the receiver in the M-ary case sets m(p) = mg if and only if

p's + By > p's; + B; for alli (3.59)

 Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 119

Select m @) =m,
iff

max {q,} = 4;

Figure 3.17 M-ary vector cross-correlation receiver for the AWGN channel

where

B, A (207 In Gi = sil?) (3.60)

Re
Nl

The vector cross-correlation receiver can be mechanized as shown in Fig. 3.17, where
Gip= p's; + B;. The received signal p is correlated with each possible signal s; (that is, form
the inner product p’s;) and a bias B; is added to compensate for different signal powers and
a priori probabilities. When all the signals have equal lengths and a priori probabilities, the
bias can be ignored.

M-ary vector receivers in the colored Gaussian noise channel. So far,

we have considered the case when the channel noise is Gaussian and white. In this section,
we consider the case when the noise is still Gaussian, but not necessarily white. In general,
the N-dimensional Gaussian pdf is given by

fn(a) & G(@; mn; Kn)

| ] oe: (3.61)
oo a ay ae PERT TT p {50 - mn)! Ky, lq@ ms)
Ory dete
where WN is the dimension of the vector n, My is the mean vector of n (that is, my = E{n}
where E(.) denotes statistical expectation), Ky is the NxN symmetric covariance matrix of
n given by

Ky = E {(n—my)(n ~ mq)"| (3.62)

and det Ky, denotes the determinant of the square matrix Ky. Note that if the Gaussian
vector n is white and each component has variance a”, then Ky = o7Z (Z denotes the NxN
‘identity matrix), det Ky = (a7), and (3.61) reduces to (3.53) as expected. From (3.50),
the MAP receiver sets m(o) = m; if P(m;) fr(p/s = $;) is maximum. Since fr(p/s = $;) =
120 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Kilso wei
DONS
—s Kos + In P(mg )
0

Select m (Pp) =m,

iff

max {q.}= 4,
a KAS a tin P(t)

Figure 3.18 Optimum cross-correlation receiver mechanization in the presence of colored Gaussian noise

fn(p — §;) and taking the natural logarithm, the decision rule in the binary case becomes

I T y-! ee I Ty-l
In P(mo) — 50 — So) Ky (@ — So) = In P(m) — 5 (0 —spiky (@—s1) -G-65)
my)
To arrive at the above equation, we assumed that the noise has zero-mean. The
nonzero mean case can be handled by incorporating the means into new modified vectors,
given by s; = s; + my. Expanding the products and simplifying, the rule reduces to
mo
pK,
so siee 8 (3.64)
mM
where
1
Tiss oa a P(m})
A 5 {80DK n 180—S,K,
a iG 1
Si} |
+1n P(mo) ;
(3.65)

Note that in arriving at (3.64), we used the fact that p’ K, 'so = s)kK, |p since both quanti-
ties are scalars and are the transpose of each other (recall that lis symmetric since Ky is
symmetric and hence K,! = (K,!)"). In the M-ary case, the decision rule can be general-
ized to give: select m(p) = m; if and only if
ae bas
Dp Kee sae 581 Ky Is; + In P(m;) (3.66)
is maximum and the resulting cross-correlation receiver mechanization is depicted in Fig.
3.18. In the case of AWGN channel, K,! = (1/07) and (3.66) reduces to (1/07)(p"s; +
B;) where B; 1s given by (3.60).
Let’s consider an alternative derivation of the optimum receiver in the presence of
colored Gaussian noise. Knowing how to perform signal detection in the presence of white
Gaussian noise, a natural extension would be to perform an operation on the received vector
p, which contains colored noise, such that the resulting observation is corrupted by white
noise only. This is referred to as a whitening operation and assuming that it is linear, the
Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 121

output noise vector, in addition to being white, would remain Gaussian, since it is a by-
product of a linear operation on a Gaussian vector. Following the whitening process, the
optimum detection rule (assuming equally likely messages) is minimum distance decoding
as we have seen before. However, the detection has to be performed on the modified set of
transmitted vectors at the output of the whitening operation. This two-stage approach to the
detection problem is very intuitive, but it is not clear if it is equivalent to the optimum MAP
receiver.
Let’s start by drawing on a key result from the field of linear algebra on the factor-
ization of matrices, namely that any nonnegative definite symmetric matrix of dimension
NXN has N associated orthogonal eigenvectors e;, i = 1,..., N [8,9]. As a result of this,
the covariance matrix of any random vector, being nonnegative definite symmetric, can be
decomposed as

Ka = Q2AQ7 (3.67)
where &2 is the matrix of orthonormal eigenvectors given by

N= [e; €2 ... en] (3.68)

and A is a diagonal matrix of nonnegative eigenvalues given by

Rie saan oO

pest Cearaong0) (3.69)

: 0 Se 0)
(ree “Lge is
The eigenvectors and eigenvalues are solutions to the following sets of linear equations:

Kenep= Gs 10r Lae 12. aN (3.70)

Factoring the matrix A as

A=AlPal/2 (3.71)
where
aa wae 0

ize eat pe aie® Pe: (3.72)

Oe Perce ok()
0 ae OS AN

then, (3.67) can be rewritten as

Ka = 2A1/2(A/4)T
QT = (QA4)(QA'/)F = QOT (3.73)
where the matrix Q is defined as

0=2A!/2 (3.74)
Now in order to whiten the zero-mean vector n with a covariance matrix Ky, suppose we
form the linear transformation
w=Pn (3.75)
122 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

DECISION
Uf WHITENING WHITE NOISE
| TRANSFORMATION SIGNAL
OBSERVATION NEW OBSERVATION DETECTION
IN COLORED IN WHITE NOISE
NOISE

Figure 3.19 Vector detection in colored noise

where P is an NxN matrix chosen such that the resulting vector w is white, that is, Ky = Z.
From (3.75), we have

pee ws {ww7} =f {Pnn’P? |= Ph ep! (3.76)

But, we have shown that K, can be factored as QO! ,hence

Kn = ROO P= (POCO) (3.77)

The matrix P could be chosen as P = Q™! because then Ky = Z and the noise has been
whitened. Therefore, in conclusion, the noise vector n can be whitened using the linear
transformation Q~! where Kn = QQ! and O= QA!/?.
Returning to the detection problem in the presence of colored noise, our approach
would be to whiten the observation first and then perform signal detection in the presence
of white noise, as shown in Fig. 3.19. Hence, the problem (in the binary case) reduces to
deciding between the two modified hypotheses

mo —> Ho: =Q7's +w

aes “i (3.78)
m, — H:y=Q ‘si+w

where the new observation y = Q-'p. From (3.58), the optimal decision rule for white
noise is

my
ly — O7'so/? z ly — Q7!s;/ (3.79)
0

or equivalently

y" (Q's ~ Q-'s0) moZ 5{1Q-tsil? — 12-Is6/?] (3.80)

my
>
<=

The left-hand side (LHS) of the above equation reduces to

i =i
LHS = (2"'p) Q7' (s; — s0) =p" (Q7) Q7' (si — 50)
re (3.81)
=p" (20°) (si — So) = p’ Ky! (S1 — so)

Similarly, the right-hand side (RHS) becomes

Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 123

RHS = :{(o-'s1)' (Q-'s:) - (2-'s9)" (o"'s)|

1 Ip T
=5 {st(Q"') O-'s; — sf (o"') O-'so| (3.82)
1
ae, {s7Ky 's1 ae Kn so

and hence, the decision rule reduces to

pKa
Tiel
80) a2 5Depa)
Gie §) SsKe $0)
Pre
(3.83)
mo
which is identical to (3.64) for equally likely signals. Hence, we have shown that the op-
timum MAP receiver, in the presence of colored noise, does indeed whiten the observa-
tion and then perform signal detection in the presence of white noise. Again, the cross-
correlation process is involved in signal detection.
Example 3.6
Consider transmitting the equally likely messages mo, m, using the signals

0 0
so= 10 and s;= 10

over the additive vector Gaussian noise channel with covariance matrix

m= (2 5)
4 2

Derive the optimum receiver structure and draw the decision space.
Using (3.64), we have

cates(3 4)(20)
4
=3 (2)
Ren eee ey NEEGS

and the threshold b = 0. Hence, the decision rule becomes

Sea

5 ue
q(-p1 +202) < 0
3 m\

In this case, the factor 5/3 can be dropped without changing the rule. Also, note that the first
component of the observation, 1, is being used in the detection process, even though the
transmitted signals have identical first components. This is because the noise is not white and
the observation of p; offers information on the noise component of 2. If the noise vector were
white, the decision rule would have been

and p; would not be used in the detection process. The decision spaces of both cases of white
and colored noises are shown in Fig. 3.20.

In all derivations so far, we have assumed that the covariance matrix Ky is nonsingular
(or equivalently, that the inverse does exist). What if K, | does not exist? In this case, the
noise vector lies in a space of dimension less than NV, even though it has N components
(similar to three vectors lying in a plane even though each vector can have three components
that are nonzero). Expanding the noise vector in terms of the eigenvectors of Kp, it can be
124 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Po DECISION LINE FOR

COLORED NOISE
s
DECISION LINE IF L
NOISE IS WHITE

Py

Figure 3.20 Decision space in Example 3.6

CHANNEL
1

CHANNEL
2
MULTIPLE
VECTOR VECTOR fi (p)
TRANSMITTER RECEIVER

Figure 3.21 Diversity communications

shown that n has at least one noise-free dimension or direction. Perfect detection is then
possible in this case, provided that the projections of the transmitted vectors in these noise-
free dimensions are distinct. This is referred to as singular detection. If the projections
are not distinct, then the detection should be performed in the reduced dimension space,
spanned by the noise vector [8].

M-ary multiple (diversity) vector receivers. In some applications, the signal

is received through various communications channels as depicted in Fig. 1.9 and the receiver
has to jointly process all the received vectors, in order to minimize the probability of error.
This is referred to as diversity reception and is depicted in Fig. 3.21. Such is the case in
multipath channels which create signal fading as a consequence of scattering and delay in
the physical media. In practice, messages are transmitted using waveforms and diversity
can be accomplished in time, frequency, spatially, or through electromagnetic field polarity.
In all cases, the total receiver input can be modeled by an augmented received vector r
given by

r? = (rT,es art rf) (3.84)

where L denotes the number of received vectors (or diversity channels) and

rf)(12,9. PN (3.85)
Given that r’ = p’ = (p,", py", ..., p,,") is observed, the MAP receiver sets m(p) = m;
Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 125

SCALAR
TRANSMITTER

MULTIPLE
SCALAR
RECEIVER

tg=Ny*No
No

Figure 3.22 Diversity channel of Example 3.7

if and only if

is maximum for j =/, or equivalently if

P(mj) fr(p/s = Si) (3.87)

is maximum. Using the fact that

Sr( py, Pr,---,Pp)= fry /r7 es r,(P|/Pr; tees PL) Sr Hee ry,(Pr; =. Dy) (3.88)

(3.87) simplifies to

P(mj) fry /r,....7,,5(P 1/P2> 1+; Pz,S=Si) fr Rites r,/s(P2, ++, Py_/S=S§;) (3.89)

If rj conditioned on r2,..., ry is statistically independent of s, the decision rule simplifies

to: set m(p) = m; if and only if

P(mi) fr, Fre83 ry /s(P2, +--+, Pp /S=Si) (3.90)

is maximum. In this case, the knowledge that r); = p, is ignored by the MAP receiver
as it does not enter into the maximization of (3.87) with respect to 7. Hence, r} = p, is
irrelevant to the decision process and the performance of the receiver is not degraded [10].
This reasoning can be applied iteratively to (3.90) to ignore all irrelevant observations and
simplify the MAP multivector receiver to the extent possible.
Example 3.7
In the multiscalar communication system depicted in Fig. 3.22, the transmitted signals (s; =
—so =) and the noise nj, 2 are all statistically independent random voltages. Assuming that
the messages are equally likely and that the noises are Gaussian with zero-mean and variance
o”, derive the structure of the optimum multiscalar receiver.
The decision rule is given by (since L = 2)

m(
Fersrr(P1, P2/Mo) = fri,r2(01, 22/1)
m|

Applying Bayes rule, we have

mo
frs/r:(02/P1,Mo)
fr,(P1/Mo) = fro/r) (02/01,
1) fr,(1/1)
my|
126 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

But

fro/r\ (02/1, Mo) = fny(r2 — 11 + 50)

Similarly
Sroir (p2/ Pi, M1) = fny(r2 =n +51)
fr,(e1/mo) = fa,(ri — So) and f,(p1/m1) = fn,(ri — 51)
Hence, the decision rule becomes

exp \-333
I
((o1 +)? 2 + (p2 = 1-9)53 )|a2 0 |—357l ((o1 — 5)?
\2
+ (p2 a1 +)’)2
Taking the logarithm and rearranging, we get (assuming s > 0)
my\
pi-=—2 = 0
2° mo

3.2.3 Decision Rules, Functions, Regions, and Boundaries

The optimum decision rule partitions the N-dimensional observation space into M disjoint
decision regions Ro, Rj, R2,..., Ry—1 corresponding to the M messages. The decision
regions are defined by the decision rule which sets m = mx if the decision function |r —
s;|2 — 207In P(m;) is minimum for i= k.
To gain geometric insight into this decision rule, consider first the case of equiprob-
able messages. In this case, the second term in the above decision function can be dropped
and the receiver decides m = m, if the reduced decision function |r — §; | is minimum for
i =k. Noting that |r — s;| is the Euclidean distance between r and s;, then the decision rule
has a very simple geometric interpretation, namely, the decision is made in favor of that
signal (message) that is closer to r. This is called minimum distance decoding which is a
familiar concept in coding theory.
When the messages are not equiprobable, then a particular message m; will be most
probable. In this case, a decision is made more often in favor of m; and the decision regions
will be biased in favor of m;. This point is illustrated by the appearance of the term In P(m;)
in the decision function.

Binary communication with equal a priori probabilities. To be more

specific, let us assume that M = N = 2. Consider a two-dimensional space and the two
equiprobable vectors s9 and s; shown in Fig. 3.23a. From this figure, it is clear that when
the decision is made in favor of that message corresponding to the signal vector that is clos-
est to r, the decision boundary that separates the two decision regions is the perpendicular
bisector of the line that joins the vectors sp and s;. This is true since any point on the bound-
ary is equidistant from so and s). If r falls on the boundary itself, one flips an unbiased coin
to decide whether to select mo or my.

Binary communication with unequal a priori probabilities. Consider

now the binary case with unequal a priori probabilities for the two messages. Let do =
|r — So| and d; = |r ~ s;| denote the Euclidean distance of r from so and s1, respec-
tively. From (3.57), we see that a decision is made in favor of mg if de — d; ea
20% In[P(mo)/P(m},)] and in favor of my, if ae ~ ae > c?, Along the decision boundary,
dé — d* =c? . We shall now show that this boundary is a straight line perpendicular to (but
 Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 127
9,

(b)
Figure 3.23 Decision regions and decision boundaries for binary communication

not the bisector of) the line that joins the vectors $9 and $s; at a distance jz from So (see Fig.
3.23b). Here, pp = (c* + d*)/2d = (07 /d) In{ P(mo)/
P(m )] + d/2 where d is the distance
between So and $1.
To prove the above, we make use of the geometry in Fig. 3.23c. It is clear from this
figure that ae =a? +? and d? = a* + (d — 1)”, so that dp — d? = 2du — d* =c?. This
is the desired result since along the decision boundary, d? — d; is constant and equal to c?.
Note that for equal a priori probabilities, c = 0, and along the boundary, dp = d).

M-ary communication with N = 2. The decision boundaries for the decision

regions are determined along similar lines to those discussed above for the binary case. In
particular, consider the case of M = 3 equiprobable messages with N = 2. The decision
regions are as illustrated in Fig. 3.24. The decision boundaries are constructed by taking the
vectors So, $;, and $2 in pairs and constructing the three perpendicular bisectors of the lines
joining the vector pairs. When the signals are not equiprobable, then the decision boundaries
will still be lines drawn perpendicular to the lines joining the vector pairs. However, they
will be shifted away from their midpoint positions in accordance with the ratios of a priori
probabilities as above.
128 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Figure 3.24 Decision regions and decision boundaries for ternary communication

M-ary communication in N-dimensions. Finally, for /@ signal vectors in N-

dimensional space, the decision regions will be N-dimensional polyhedra and the decision
boundaries will be (NV — 1)-dimensional hyperplanes. Proof of this fact is left as an analytic
geometry exercise for the reader.

3.2.4 Error Probability Performance

In this section, we derive the message error probability of the optimum MAP vector receiver
in AWGN channels. First, we consider the case of binary signaling and then generalize
to M-ary communications. The performance in the presence of colored noise can also be
derived following similar steps.

Binary vector receiver performance. In the binary case, the probability of

message error can be found using the theorem on total probability, that is

P(E) = P(mo) P(E/mo) + P(m1) P(E/m)) (3.91)

where P(E/m;) is the conditional probability of error given m;. The MAP decision rule
is given by (3.57) and assuming that mo is actually transmitted, the probability of correct
decision is then the probability that
P

|e — sol” < Je — si? — 20°In ( Ul (3.92)

P(mo) /
Expanding the squares and simplifying, one obtains

| IC
p™ (si — 80) < 5 (Isi? = [sol?) — 07 In ( on (3.93)
P(mo)
Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 129

or equivalently (since p = so + n when mo is transmitted)

1 P(m})
So +n)? +m)’ (Ss;
(So — so)
(Ss; — So) <=< =5 ([s1|7
(isi = Iso| |sol?)
) — o* 07In
In (|——
P(mo) (3.94)
Performing the inner products and simplifying, we have
1 P(m,)
n’ is(S1
(Sj— So)
So) < =|S;
2 —o~12
5 iar — So|“
50g n(5) (3.95)
19>

i inner product n’ (s; — so) is a zero-mean Gaussian random variable z(u), with Saran
oa? = E{z*(u)} == (S} — So)! Kn(S1 — So). Since Ky = 07Z, the variance reduces to o?
ae so|?. Using the transformation z(u)= o|s; — So|x(u), where x(u) is a nonmalited
(zero-mean, unit-variance) Gaussian random variable, (3.95) can be rewritten as
dio oO
x(u) < — — —In (Fe) (3.96)
20 dio P(mo)

where do = |S; — So| denotes the distance between vectors sp and s;. Using the Gaussian
tail integral, the probability of error, conditioned on mo, is then given by

P(E/mo) = Q oo= Fn (Fe (3.97)

20 dio \P(mo)
Similarly, it can be shown that

Aging? P(m})
P(E = — +— |] 3.98
Shee be dip ster She
Hence, the total probability of error becomes

P(E) =P(mo)Q
mr \26 dio
a) )
Ses \-P(mo)
(3.99)

+ P(m)Q
ics 2z in (Fe)
han ARG
When the signals are equally likely, the expression reduces to
d
P(E) = o(3) (3.100)
20

which is a function of the ratio of the distance between the signals to the noise standard
deviation. This is a fundamental result which suggests selecting the binary modulation
technique that maximizes the “distance” between the vectors that represent the messages
in the channel.

M-ary vector receiver performance. In the M-ary case, the receiver partitions
the observation space into M disjoint decision regions R; as depicted in Fig. 3.13 [since the
figure deals with scalar communications, R; (a multidimensional region) is replaced with /;
(a one-dimensional region)] and the conditional probability of correct decision given that m;
was transmitted is given by

P(C/mi) = P(peRifm) = f+ f fr(p/mi)dp (3.101)

Rj
——
N-—fold
130 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Figure 3.25 Signal set of Example 3.8

and
M-1
P(C)= S47 P(m;) P(C/mi) (3.102)
i=0

For some specific signal constellations, the multidimensional integral can be reduced to the
product of N one-dimensional integrals when the noise is white and Gaussian. However, in
the general case, the integral cannot be simplified any further.
Example 3.8
A ternary communication system employs the vector signal set illustrated in Fig. 3.25 to trans-
mit equally probable messages mo, mj, and m2 over the additive white Gaussian noise channel
with zero-mean and covariance matrix Ky = o°Z. Write a mathematical expression for the ex-
act probability of error P(E).
Noting that cos 60° = 1/2 and sin 60° = /3/2, the signals vectors are given by

= =e,
4(5) = Yuk
i=
elle
A( 7H) and n=
nah elle
a( /f, )

Since the noise is white and Gaussian, minimum distance decoding is the optimal rule, as shown
in Fig. 3.25. Since the signals are symmetric and equally probable

P(E) = 1 — P(C/mo)

But

P(C/mo) = P(pe Ro/mo)

Assuming that mo is transmitted, @ = so + n resulting in

fr(p/mo) = fn(e — So)

Hence

P(C/mo) =i! fn(p — So)dp

Ro
- 1
l Dee?
= i Imo2 XP {0 —A)> = > dprdp\
Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 131

or

P(C/mo)
0 =f i
Fin SN exp | 5: ((01 - A)’
) +0} d p2dp

and

P(E) = 1 — P(C/mpo)

The above integral cannot be simplified any further and needs to be evaluated numerically for
performance prediction purposes.

3.2.5 Effects of Translation and Rotation of the Signal Vectors on Message

Error Probability

To facilitate analysis and/or to minimize the average required transmitted energy, it is often
advantageous to translate and rotate the signal constellation {sg, s1,..., Sy—1}.
Consider the conditional probability of correct decision given that m; is transmitted,
that is, P(C/m;). This conditional probability is the probability that the received vector
r=s; +n lies in region R;, as depicted in Fig. 3.26a. This is also the probability that
the noise vector drawn from the point s; lies within R;. Since this probability does not
depend on the origin of the coordinate system, we can translate the coordinate system to
facilitate the analysis. Such a translation is equivalent to translating the signal point and
the corresponding decision region. Consequently, P(C/m;) = P(C/s;) for the translated
system shown in Fig. 3.26b is identical to that of the system in Fig. 3.26a. We conclude
that translations of the entire signal constellation and the corresponding decision regions do
not alter message error probability. This in turn suggests that “different” signal sets yield
equivalent performances.
In the case of spherically symmetric additive white Gaussian noise, we make another
important observation, namely, the rotation of the coordinate system does not affect the
message error probability because the pdf fn(@) possesses spherical symmetry. To prove
this, consider Fig. 3.26c which is generated from Fig. 3.26a by translation and rotation.
Notice that rotation of the coordinate system is equivalent to rotating the signal constellation
in the opposite sense. The probability that the noise vector n (drawn from s;) lies in R; is the
same in Fig. 3.26a as the probability that the noise vector (drawn from s;) belongs to R; in
Fig. 3.26c. This is because the probability P(C/m;) is characterized by the integral over
the region R; of the noise pdf fnp(@ — s;) in Fig. 3.26a. If one translates s; to the origin, the
probability in region Rj is not affected by any rotation of the region R; because symmetric
Gaussian noise (same variance in all directions) possesses spherical (circular) symmetry.

Example 3.9
One of sixteen equally likely messages is to be communicated over a vector channel which adds
a Statistically independent zero-mean Gaussian random variable with variance No/2 to each
transmitted vector component. Assuming that the transmitter uses the signal vectors shown in
Fig. 3.27(a), express the P(E) produced by an optimum receiver in terms of the Gaussian tail
integral.
Since the channel is additive white Gaussian, the decision rule is the minimum distance de-
coding rule as shown in Fig. 3.27(b). Dividing the observation space into three major types as
depicted in Fig. 3.27(c), there are 4 regions of type ©, 4 regions of type @, and 8 regions of
type @. Since translation and rotation does not affect P(E) and hence P(C), we have
132 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

(c) Translation and rotation of $; and R

Figure 3.26 Translation and rotation of signal vectors and decision regions

ee ares I Ee
= - =——exp |= (x +9)
y?)[axa
bs
4 J_aj2 J—a/2 ™No No :
d/2 lo) 1
1 1
ou >/ —— exp a (x 35 )|dxd
2 J—aj2 J—aj2 No No | :
1 ee ee i 1 A ;
A ae ND) ls (x ae )|
dxd
4 J—aj2 J—aj2 No No 3 ?
Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 133

(a) SIGNAL (b) DECISION REGIONS (c) REGION TYPES

Figure 3.27 Signals and decision regions of Example 3.9

I 1
P(C)
(C) = —-(1—
4! P) p)” eae p)( 2p) +7 2p) 2

where p = QO(d/./2No). Hence

“(1-2
3 ps
P(E)=1~ P(C)=1- (1-52)

3.2.6 Translation to Minimize the Required Signal Energy

Based upon our arguments presented in the previous section, there exists infinitely many
signal constellations which yield the same message error probability. From a system de-
sign prespective, it makes sense to choose the set of signal vectors that minimizes the mean
energy required to yield a specified message error probability. This suggests choosing the
signals which are closest to the origin. By translating and rotating the signal set, one can
always find an arrangement of signals which require minimum energy. For example, trans-
lating a set of orthogonal vectors to minimize the required energy yields the transorthogonal
(simplex) set.
Consider the mean energy

ESE (s;)= )> P(mi)lsi|? (3.103)

of the signal set. Translation of the signal set to minimize the required energy does not affect
the message error probability and is equivalent to substracting some vector a from each
134 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

vector Ss; to create a new vector s; = s; —a for all i. In other words, we wish to find the
vector a which minimizes the new mean energy
M-1
E'= )> P(m)lsil « C100)
i=0

One can easily show (see Problem 3.42) that the vector a specifies the center of gravity of
the M vectors So, $1, ..., Sy—1, that is
M-1
a= Dy P(m;)s; = E {s;} (3.105)
1=0

This result readily follows from the fact that the energy E’ is equivalent to the moment
of inertia about a point a of the “M particles of probability masses” P(mo), P(m}),.--,
P(my-1), respectively.
Example 3.10
Consider the equilikely binary orthogonal signal vectors s) = (W/E, 0) ands? = (0, VE). Cre-
ating the mean signal vector a and subtracting it from So and s; yields the new signal vectors
sy’ = (—/E/2, 0) and s,7 = (/E/2, 0) = —sp. The message error probability for both sig-
nal sets is Q(./E/No); however, the latter requires a smaller mean energy. This is one of the
reasons why BPSK (antipodal vectors) is selected in practice over BFSK (orthogonal vectors).
Example 3.11
Transorthogonal (simplex) signal vectors are the minimum energy equivalent set of equilikely
orthogonal signal vectors. This is illustrated in Fig. 3.28 for the case of two- and three-
dimensional signal space. Note that energy minimization effectively reduces the number of
dimensions (bandwidth) of the signal space by one. For two-dimensional orthogonal signal vec-
tors (BFSK), the corresponding transorthogonal signal vectors lie on the same straight line, that
is, the one-dimensional antipodal signals (BPSK) as shown in Fig. 3.28a. For three-dimensional
orthogonal vectors, the plane formed by the three signals is constructed where the center of
gravity of the signal vectors on this plane becomes the new origin. It is obvious that the equiv-
alent transorthogonal signal vectors are reduced to two dimensions as depicted in Fig. 3.28b.
Similarly for the four-dimensional case, the minimum energy set lies at the vertices of a tetra-
hedron (as shown in Fig. 4.21 in Chapter 4) reducing it to a three-dimensional space. It can
be shown (in Chapter 12) that for a given mean energy, the equilikely transorthogonal signal
vectors are optimum in the sense of minimum error probability in white Gaussian noise.
For the orthogonal signal vectors s; = VE@;, the mean signal vector a is given by

and

so that

1
B= 15)?=6 (1~ x)
M
Hence for the same message error probability, the mean energy of the transorthogonal set is
(1 — 1/M) times that of the orthogonal set. For large M, orthogonal and transorthogonal signal
sets have the same message error probability.
 Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 135

S> |\~#—§ \7E ee 0

(a) BFSK to BPSK (b) 3-FSK to 3-PSK

Figure 3.28 Equivalent signal vector sets

3.2.7 Union Bound on M-ary Vector Receiver Performance

For most signaling schemes, the computation of the exact probability of error is tedious and
an upper bound on the performance might be sufficient. In this section, we will derive such
a bound for vector communications over the additive white Gaussian noise channel. The
bound is general and can be applied to any set of vectors.
We know from the previous discussion that the optimum a posteriori probability re-
ceiver sets m(p) = mz, if and only if

p's +B > p's; +B; foralli (3.106)

where B; is a constant given by (3.60). When the signals are of equal length and are
equally probable, the constants can be ignored and the rule simplifies to evaluating the cross-
correlations

p's > p's; (3.107)

Hence, the conditional probability of error, given that m, was transmitted, is

P(E/m,) = P (07s, <p's; for some j 4 k/me) (3.108)

Let E; (j #k) denote the event that p's; > p's. Then, an error can result if only one
event E; occurs for j #k, or if any two events Ej, Em (j Ak, m #k) occur simultaneously,
136 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

or if any three events Ej, Em, En (j #k, mAk andn#k) occur jointly, ..., or if all
M — events Eo,..., Ex—1, Ex41,..-, Em—i occur together. Mathematically, this can be
denoted using a union of events as

P(E/m,) = P(Eo/mg) + +++ + P(Ex-1/mg) + P(Ek+i/mg) + +--+ P(Em-1/mx)

+ P(Eo, E,/mx) +--- + P(Em-_2, Eu-i/mx)
Eee
+ P(Eo, .. = Ex-1, Ex41,---, Em-1/mk) (3.109)
M-1
=P) |Ej/mi
j=0
xk
The signs in (3.109) depend on M since P(A + B) = P(A) + P(B) — P(AB) but P(A +
B+C)= P(A)+ P(B)+ P(C) — P(AB) — P(AC) — P(BC) + P(ABC). Since P(A +
B) < P(A) + P(B), applying this inductively to the union in (3.109), we obtain a bound on
P(E/m,), namely
M-1
P(E/my) < >) P(Ej/mx) (3.110)
jak
or, equivalently
M-1
P(E/my) < S~ P (07s, < p's;/mx) (3.111)
j=0
J#k

This bound is referred to as the Union bound and is very useful in communication prob-
lems. Consider the event that {p/sx < p's; /mx}. Assuming that the noise is white Gauss-
ian, the probability of this event is given in (3.100) as

dk j
P(E ;/m) =| — (3.112)
20
where dx; is the distance between the vectors sx and s;. Hence, the Union bound simplifies
to

M-\
dk
P(E/m) < 9) Q @ (3.113)
j=0
i#k
The probability of error P(E) can then be bounded using (3.113) with the appropriate a
priori probabilities. This is a key result that will be used in the chapters on coded communi-
cations.
Example 3.12
Assuming that a communication system uses M orthogonal signals to communicate over an
AWGN channel with Ky, = 027, derive a bound on P(E) when |s; |? 4 Dei m—a Ole eS oe M —-1.
The signal can be modeled by

7 =VEO) = 01M
Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 137

where

vad aay (ORE |

o; d; ae (ee xj

for all i and j in the set {0, 1,..., M — 1}. The distance between any two vectors is given by

d€d=VE+E=V2E

and is independent of which pair we choose due to the orthogonality. Using (3.113), we get

jh;
PREYS 1)0 (z

where we have assumed that 0? = No/2 =kT°/2 Joules. At room temperature, T° = 300K =
28°C and kT° = k300K = 1/40 eV (electron-volt).

3.2.8 Error Probability Performance of Two-Dimensional, M-ary Vector

Receivers—A New Approach

In Section 3.2.4 we described the traditional approach for evaluating the probability of
correct decision for N-dimensional, M-ary vector receivers wherein the observation space is
partitioned into M disjoint N-dimensional regions Rj; i = 0, 1,2,..., M — 1 surrounding
the signal vectors. The origin of coordinates for defining these decision regions is assumed
to be the same for each of them and is located in the N-dimensional space defined by the
received vector r. In particular, for the i signal transmitted, the conditional probability of
correct decision is described by the N-dimensional integral in (3.101) which, in the general
case, cannot be simplified any further. Even for the case of a two-dimensional signal vector
set and an AWGN channel (such as in Example 3.8), we saw that the expression for the
conditional probability of correct decision still took the form of a double integral. Thus, in
general, the average probability of correct decision, or equivalently, the average probability
of error, would be expressed as the sum of M such double integrals.
Recently, Craig [16] cleverly showed that the evaluation of average probability of
error for the two-dimensional, AWGN channel case could be considerably simplified by
choosing the origin of coordinates for each decision region in the two-dimensional space de-
fined by the corresponding signal vector. This shift of the coordinate system origin from the
received vector space to the signal vector space allows the integrand of the two-dimensional
integral describing the conditional probability of error to be independent of the transmitted
signal. In particular, this integrand is simply the pdf of the two-dimensional noise vector
n which is characterized by a two-dimensional zero mean Gaussian pdf with independent
components. When expressed in polar coordinates, that is, (r, @), this pdf can be integrated
on r to yield a closed form result, that is, a single exponential, regardless of the complexity
of the integration limits or their dependence on @, the latter being dependent on the shape
of the decision boundaries. Thus, at worst, the average probability of error would now be
expressible as the sum of M single integrals. In short, this new approach of computing av-
erage error probability has shifted the complexity of the computation from the integrand to
the integration limits.
To illustrate the above approach, we consider a generalization of Example 3.8 cor-
responding to an M-ary communication system with a vector signal set having M points
138 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Figure 3.29 The geometry for the decision

region corresponding to the signal point so

uniformly distributed around a circle of radius A. The geometry for the decision region as-
sociated with the signal point

—1
So=A ( ‘)

is illustrated in Fig. 3.29. An error is made if the noise vector n lies outside the wedge-
shaped decision region bounded by the semi-infinite lines OA and OC, as for example, the
point Z. In particular, the probability of an error given that signal So is transmitted is twice
the probability that Z lies above the boundary AOD, namely
(M—1)2/M re
/2KGS, imo) =2 f ao | fa (r,
9) dr (3.114)
0 R
where R is the distance from the signal point to the boundary point E (in general, a func-
tion of #), and fn (r, @) is the above-mentioned bivariate Gaussian pdf of the noise vector
expressed in polar coordinates, that is,

Te r2
far, 0) ae exp (-z) (Selilis))

which is clearly independent of 6. Substituting (3.115) into (3.114) and carrying out the
integration on r gives

1 (M—1)x/M AZ sin2 4
P (Elmo) = — | ex dé 3.116
mt Jo P 202 sin? (0+ a) ( )

where, from the law of sines

s LS
Asin 71
= WENT
(0 fg2) (3.117)
Sec. 3.2 Vector Communications Over the Continuous Memoryless Channel 139

Making the change of variables ® 4 [(M — 1)/M] — 0, the final form for P (E |mo) is

J fdM@-Dx/M A? sin? 4
P (Elmo) =— f ep) eo 3.118
Tt Jo : 202 sin? © ( )

Finally, since from the symmetry of the vector signal set, we see that P (E |m;) is indepen-
dent of 7, then the average error probability is also given by (3.118), that is,
1 p(@-Dx/M Avene ES
Pe@)=— f exp | d® (3.119)

For the ternary case (M = 3) as in Example 3.8, (3.119) becomes

P(E) = af a - | d® (3.120)
m Jo 20? sin? ®
which is a considerable simplification of the result previously obtained for this case in the
form of one minus a double integral. Furthermore, for small values of P (E), the direct nu-
merical evaluation of (3.120) will yield improved accuracy relative to the previous indirect
result where a number close to unity (the probability of correct decision) is first evaluated
and then subtracted from unity.
For the binary case (M = 2), (3.119) evaluates to
1 m/2 AZ
P(E) = - exp -ss| do .121)
Tw JO 202 sin* ©
However, for this case, the result as obtained from the traditional evaluation approach in
Section 3.2.4 is expressible as a Gaussian tail integral, namely
A
p(e)=0(=) (3.122)

Thus, we see that the Gaussian tail integral, which is normally defined as the integral over a
semi-infinite interval of a Gaussian pdf, can be expressed as the finite integral
1 m/2 x2
= — exp |-——~— |d® (Gall23))
24) =} | 75
which is useful for numerical evaluation purposes.
It is straightforward to generalize the above results to signal sets whose decision
region boundaries are polygons surrounding each of the signal points. A typical example
of such is described by the geometry in Fig. 3.30 where the nth signal point s,, is illustrated
inside its correct decision region described by the four-sided polygon ABCDA. If signal s,
(corresponding to message m,) is transmitted, then an error occurs if the tip of the noise
vector lies outside of this polygon. The region outside of this polygon is made up of the four
disjoint regions ©, @, @, and @. The probability that the tip of the noise vector falls in region
@ is described by an integral analogous to (3.116), namely

6 ew)
Pi (Elm)=— f | alba h | (3.124)
1 0 202 sin? (0 + W1)
140 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

\ Figure 3.30 The geometry for the decision

\ region corresponding to the signal point s,,

The probability that the tip of the noise vector falls in any of the other three regions is
given by an expression similar to (3.124). Since the four regions are disjoint and comprise
the entire error space, then the conditional probability of error given that message m, was
transmitted is
G; __ Apsin’
A? sin? wi
EACEMin;, = (CEMA) = dé Brl25
i he ey Chas aefe o|-
S52 Ga (0
ase
+ Wi)
The conditional probability of error for any other point in the signal constellation is
given by an expression analogous to (3.125) except that the number of terms in the sum
might be different depending on the number of sides in the polygon that defines its decision
region. Finally, the average symbol error probability would be obtained by averaging all
of these conditional error probabilities over the a priori probabilities of the message set.
As such, the average symbol error probability is expressible as a weighted sum of single
integrals with finite limits and an exponential integrand.
 Appendix 3A
Other Detection Criteria

So far, we have considered the probability of error detection as the only performance mea-
sure for the receivers. Depending on the application (for example, in radar system design),
other criteria [11] might be more suitable and these make use of differing amounts of in-
formation and problem specification. The two criteria for the decision rule that we have en-
countered are the maximum likelihood (ML) criterion and the maximum a posteriori (MAP)
criterion. In this appendix, we will consider three more criteria: the Bayes decision rule,
the minimum-maximum rule, and finally, the Neyman-Pearson criterion. Only the basics of
each rule are discussed assuming a binary hypothesis and the reader is referred to appropri-
ate references for a more detailed treatment.

3A.1 THE BAYES CRITERION

In both detection criteria considered so far, we have assumed that all errors and all correct
decisions are equally important and thus, are treated equally by maximizing the a posteriori
probabilities. In some applications, some errors are more costly than others and as a result,
some kind of weighting is necessary to reflect that. In Bayesian detection, each course of
action is assigned a cost which is a function of both the chosen and true hypothesis. These
costs are denoted by C;;, where in the binary case, we have

Coo = Cost of deciding hypothesis Hp when Ho is true

Cio = Cost of deciding hypothesis H; when Ho is true
Co; = Cost of deciding hypothesis Hp when H, is true
Ci; = Cost of deciding hypothesis H; when Hy, is true
The expected value of the cost, referred to as the risk R, is given by

R = CooP (Ho) P(Ho/Ho) + CioP (Ho) P(A1/Ho)

(3A.1)
+ Co, P(A) P(Ho/ M1) + Ci P(A) P(A / 1)
where P(H;) is the a priori probability of hypothesis H; and P(H;/H;) is the probability
of deciding on hypothesis H; given that hypothesis H; is true. In the binary case, the
observation region is divided into two distinct regions, Ro and Rj, where we decide Ho if
pe Ro and decide H, if pe Rj. Hence

P(Hi/H;)= |Rj felo/H))dp GA2)

However, note that since the received vector belongs to either Ro or to Ry (that is, Ro U Ry
= complete observation space), then

|.stoitae=1- f fr(p/Ho)dp (3A.3)

Ry Ro

and

|.seostvae=1- | fr(p/Hi)dp (3A.4)

R, Ro
142 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Hence, the risk can be expressed as

R= CyP (Ho) + einen

3A.5
+} {PUN)(Coi — Ci) fe(e/
Hi) — P(Ho)(C10
— Coo) fr(o/Ho)} de Soe
Ro

Since correct decisions are less costly than erroneous ones, we can safely assume that

Cor Gime ands Cig E00 (3A.6)

so that Co} — Cj; > 0 and Cio — Coo > 0. Bayes decision criterion selects the decision re-
gion Ro in order to minimize the risk. Since the first two terms of (3A.5) are independent of
Ro, they are not involved in the minimization. The integral, on the other hand, is minimized
by selecting Rg to contain all the observations p that satisfy

P(Ho)(C10 — Coo) fr(e/ Ho) > P(A1)(Co1 — Ci) fr(e/ A) (3A.7)

This results in the most negative possible value of the integral and as a result, the risk is at
its minimum. Since Ro and R are disjoint sets, all observed vectors p for which

P(Ho)(Cio — Coo) fr(e/ Ho) < P(A1)(Coi — Ci) fr(e/A1) (3A.8)

are included in R,. Hence, the Bayes decision rule is given by

Ho
P(Ho)(Ci10 — Coo) fr(@/Ho) = P(A1)(Co1 — Cit) fr(@/A1) (3A.9)
Ay
or equivalently

Ar(p) A
fe(p/ Hi) 21 P(Ho)(Cio=Coo) _ , (3A.10)
fe(o/Ho) py P(H)(Cor = Cui)
which is a likelihood ratio test. Note that when Co; = Cio = 1 and Coo = Cj; = 0, the Bayes
rule reduces to the MAP decision rule, as given in (3.18) for the scalar case. All results
derived herein can be generalized to M-ary decisions as given in [12-14].
Example 3A.1
A decisioning system must decide between hypothesis Ho and H;, where Ho is the hypothesis
that an observed random variable x(u) has a pdf

l x?
fx(x/ Ho) = ex (-5)
Vir, Wee
and H is the hypothesis that x(w) has a pdf

i F

0, otherwise

The a priori probabilities of Ho and H; are 3/4 and 1/4, respectively. The costs are given by

Co =0 Ci=1

Cor =2 Cy =0
Determine and describe the optimal Bayes decision procedure.
 Sec. 3A.2 The Min-Max Criterion 143

fe) 1.081 3

Figure 3A.1 Decision regions in Example 3A.1

Since f,(x/H)) = 0 for x < 0 and x > 3, then decide Hp in this region. For 0 < x < 3, use the
ratio in (3A.10) to get

v2 (1/3) 2! @/4y(1) forO

<x <3
exp(—x?/2) Fy, (1/4)(2)’
or

pen——.,
exp(x*/2) for0<x <3
i,0) 2/2
or
A, 9

~S PQ An( )=1.081, for0<x

<3
Ho \ 2/20 a ks
Hence, the decision regions are as shown in Fig. 3.A1.

3A.2 THE MIN-MAX CRITERION

In some applications, the a priori probabilities P(Hp) and P(H)) are not known, in which
case, the minimum-maximum (min-max) criterion can be employed. The average cost in
(3A.5) can be rewritten as
R=Cjo(1 — P(Aj)) + C11 P(A) (3A.11)

+ | {P(H1)(Co1 — Ci) fr(e/ M1) — ( — P(A1)) (C10 — Coo) fr(o/Ho)}

dp
Ro
Since P(H;) is unknown, the min-max rule maximizes the risk with respect to P(H) and
then minimizes that maximum with the choice of Ro. Hence, the rule is given by

min max R (3A.12)

Ro P(A)
In most applications, the minimization and maximization operations can be interchanged
resulting in
min max R= max minR (3A.13)
Ro P(A) P(H\) Ro
The minimization of R, with respect to Ro, is simply the Bayes rule associated with P(H)).
Hence, Ro contains the observed vectors p for which

P(A})(Coi — Ci) fr(e/Mi) < (l — P(A1))(Ci0 — Coo) fr(e/Ho) = (3A.14)

144 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

giving the decision rule

felp/Ho) Zz
$0 PCHi)(Cor = Cus) Gen
:
fr(p/Hi) yy, Cl — P(A1))(Ci0 — Coo)
associated with P(A), or equivalently

fr(e/A1) Ua
> lbey(19 poe— P(A1))(C10 a— Coo) (3A.15b)
fr(e/Ho) Hy = PH) (Cor — Cin)
This is followed by a maximization to find the least favorable P(H)}).
Example 3A.2
Assuming that the a priori probabilities are unknown, repeat Example 3A.1 for the min-max
rule using the pdf’s

fx(x/Ho) = exp(—x), x>0

and

fr(x/H\) = 2exp(—2x), x>0

Using (3A.15a), we get the likelihood ratio test

exp(—x) = 2( P(A) )
2 exp(—2x) i, 1 — P(AM;)
or

siPte (Luo
A
Mae e(€2)))
or

Ho =
rare gels (soe
a) By.
P(Ao)

The decision rule is valid as long as 7T* is positive (that is, P( Ho) < 4/5). In this case, the risk
is given by [using (3A.1)]

R = P(Ho) P(A\/ Ho) + 201 — P(Ao)) P(Ao/ A)

But
Siss

P(H\/Ho) = / exp(—z)dz
0
and
foe)

P(Ho/H\) = / 2 exp(—2z)dz
ti!

Substituting for 7* and solving the integrals, we get

9 P?(Ho)—8
P(Ho)
R= |BP Ho)=1) for P(Ho) = 4/5
2(1— P(Ao)) for P( Ho) > 4/5
Now, we need to solve for P( Ho) which maximizes the risk R. For P( Ho) < 4/5, the maximum
is obtained by setting dR/d P(Ho) = 0 and solving for P(Ho) to get P(Ho) = 2/3 (with corre-
sponding R = 1/2). For P(Ho) => 4/5, the maximum R is 2/5 achieved at P( Ho) = 4/5. Hence
 Sec. 3A.3 The Neyman-Pearson Criterion 145

P(Ho) = 2/3 maximizes the risk and results in 7* = In 2. Therefore, the min-max decision rule
reduces to

A
Rie In?)
Ay

3A.3 THE NEYMAN-PEARSON CRITERION

In radar, signal acquisition, and antenna pointing, a decision between one of two hypotheses
must be made. Here, hypothesis Ho denotes target (signal) absent while H; denotes target
(signal) present. The performance of the radar detector (signal detector) is characterized by
the probability of each of these four possible events:

1. decide target present when target is present (H)/H}),

2. decide target not present when target is present (Ho/ HA),
3. decide target present when target is not present (H\/ Hp),
4. decide target not present when target is not present (Ho/ Ho)

The first event is called “target detection” and the probability associated with this event
is called the probability of detection, Pp. The second event always occurs if event (1)
does not occur; therefore, the target is missed and the miss event occurs with probability
Py = 1 -— Pp. Event (3) is called a false alarm since a target presence is declared even
though it is not actually present; its probability is denoted by Pr4. Using the same reasoning
as in events (1) and (2), the probability of event (4) must be the complement of event (3),
or Pc = 1 — Pra, and is sometimes called the probability of correct dismissal. Since the
decision process just described is binary, only two probabilities are needed to fully specify
the radar (signal presence indicator) performance. Notice that the error probability is given
by P(E) = P(Ho) Pra + P(A1)(1 — Pp), when the prior probabilities are known. In terms
of the channel transition densities, we have

Pra & Probability of false alarm = P(H/ Ho) = i fr(e/Ho)dp (3A.16)

R|

Pc 4 Probability of correct dismissal = P(Ho/ Ho) = / fr(e/Ho)dp = 1 —-— PrA3A.17)

Ro

Pp & Probability of detection = P(H,/H\) = / fr(p/Hi)dp (3A.18)

R|

Py 4 Probability of miss = P(Ho/H1) =| fr(p/Hi)dp =1— Pp (3A.19)

Ro

The Neyman-Pearson criterion maximizes the probability of detection, Pp, under the con-
straint that the probability of false alarm, Py, is equal to B. The solution is obtained by
using Lagrange multipliers to reformulate the problem as an unconstrained one. Construct

T= Pp— A(Pra
— B)

2 / {fe(o/H1)
~Afe(p/Ho)} do+28 3A.20
ae
Ry
146 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Clearly, if Py4 = 6, then maximizing [ maximizes Pp. As before, the integral is maxi-
mized if R; is set to contain all observed vectors p, for which

fr(p/ Hi) > Afr(p/ Ho) (3A.21)

resulting in the likelihood ratio test
H
Ar(p) =
fr(@/ Ho) Ho

The parameter A is then selected such that the constraint

Lra= p= itfr(p/Ho)dp (3A.23)

is satisfied. Note the similarity of (3A.22) to the maximum likelihood criterion, in which the
threshold A is unity.
Example 3A.3
Consider the binary problem with the following conditional pdf’s
1
fx (x/ Ho) a Jin exp(—x"/2)
IU

1
iy) = exp(—(x — m)?/2)
iaG
We require that Pr, = B. Derive the Neyman-Pearson test.
From (3A.22), we have

exp(—(x — m)*/2) " x

exp(—x?/2) i,
or equivalently

Ay InXt m
nie rey
Ho mM
The false-alarm probability is therefore
[e,2)
1
Prea=P(Hi/Ho) = | =

When 6 = 0.1 andm = 1, we have

|
o(m +3)
Ind+-=)]=0.1

From the table in Appendix 3B, we find that

Q(1.282) = 0.1
and hence, Ind + 1/2 = 1.282 resulting in A = 2.18. The decision rule reduces to

A
We 1280
Ho
 Appendix 3B
The Gaussian Integral Q(x)
In this appendix, we present a numerical table and a series approximation for Q(x), which
are useful in decision-theory problems. The Gaussian tail integral is defined by
foe) 1 sie
Q(x) aftae ex lamers
(-5) dy=1-—
dice Q(-x
Q(-x) 3B:1
(3B.1)

We also define the error function erf(x) as

2 8

a)
ace) 8 =u") ddu
f exp (—u?) (3B.2 )

and the complementary error function erfc(x) as

erfc(x) A 1 — erf(x) (3B.3)

The Gaussian tail integral Q(x) is related to erfc(x) through

erfc(x) = 20 (v2x) (3B.4)

or, inversely

O(y= pee (=) (3B.5)

2 J2

3B.1 TABLE OF Q(x)

A numerical table of Q(x) is given in Table 3B.1.

3B.2 SERIES APPROXIMATION FOR Q(x)

In this section, we present an n'" order series approximation for Q(x), along with a bound
on the percentage error. For x > 0, Q(x) is approximately given by

Q(x) ~ Qn(x) = Vian

| on (-%)
3 {14Son
“ WOO ced
,(1))(5)...
Qk =1 Ane
Ke

where

(1)(3)G)...(2n
Pr waa
+ 1) >0 (3B.7)
% error <
x?

In order to derive that bound, rewriting Q(x) as

| 2
Om) = Te : me exp (-5) dz (3B.8)
148 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Table 3B.1 Q(x). Source: Abramowitz, M. and |. A. Stegun, Handbook of Mathematical

Functions, New York: Dover Publications, 1970.

x K=O) ca 1—O(%)
0.50000 00000 00000 . 0.84134 47460 68543
. 0.50797 83137 16902 0.84613 57696 27265
0.51595 34368 52831 ° 0.85083 00456 69019
0.52392 21826 54107 0.85542 77003 36091
eecess
aorflno 0.53188
ooo°o°o 13720 13988 ococo
Oolhno 0.85992 89099 11231
0.53982 78372 77029 0.86433 39390 53618
e 0.54775 84260 20584 Ha 0.86864 31189 57270
. 0.55567 00048 05907 0.87285 68494 37202
0.56355 94628 91433 0.87697 55969 48657
Onrrno 0.57142
Pree
eY 37159 00901 0.88099 98925 44800

0957925 97094 39103 0.88493 03297 78292

0.58706 44226 48215 0.88876 75625 52166
0.59483 48716 97796 e 0.89251 23029 25413
0.60256 81132 01761 0.89616 53188 78700
0.6102€ 12475 55797 e 0.89972 74320 45558

0.61791 14221 88953 0.90319 95154 14390

0.62551 58347 23320 0.90658 24910 06528
0.63307 17360 36028 0.90987 73275 35548
0.64057 64332 7 0.91308 50380 52915
oOOCCO°O
o@
ooo°od
ooOoCCoo 0.64802 72924 24163 0.91620 66775 84986

0.65542 17416 10324 0.91924 33407 66229

.ee
°° 0.66275 UCUED enlifepl ° 0.92219 61594 73454
0.67003 14463 39407 0.92506 63004 65673
0.67724 18897 49653 0.92785 49630 34106
0.68438 63034 83778 0.93056 33766 66669

0.69146 24612 74013 0.93319 27987 31142

0.69846 82124 53034 0.93574 45121 81064
0.70540 14837 84302 0.93821 98232 88188
° 0.71226 02811 50973 0.94062 00594 05207
0.71904 26911 01436 0.94294 65667 62246

0.72574 68822 49927 0.94520 07083 00442

0.73237 11065 31017 0.94738 38615 45748
0.73891 37003 07139 0.94949 74165 25897
0.74537 30853 28664 @see 0.95154 27737 33277
0.75174 77695 46430 0.95352 13421 36280

0.75803 63477 76927 0.95543 45372 41457

0.76423 75022 20749 0.95728 37792 08671
0.77035 00028 35210 0.95907 04910 21193
Le.
@ore 0.77637 27075 62401 0.96079 60967 12518
0.78230 45624 14267 0.96246 20196 51483

0.78814 46014 16604 0.96406 96808 87074

0.79389 19464 14187 0.96562 04975 54110
0.79954 58067 39551 0.96711 58813 40836
0.80510 54787 48192 0.96855 72370 19248
0.81057 03452 23288 0.96994 59610 38800
HSE
PPR
Pee
Pere
Pree
oe
Ree
eee OnrLno 0.97128
BOMOMOO
SOCSCHK
SO
BHSYYN
PEP
UVUrwuwNwM
WWYWW
HRP
NNNMNN
PEPER
RRROoclnoe
aOnrfno
Onrflno
OoclLnNo
Onolne
o
OcLno
Onlno 34401 83998
aOnrpbno
Onrhno
aOrflno
anrlno 0.81593
oOonrlnoeo
oOr
OnLnNno
Lno 98746 53241
so
aonmnowwan
QOnona
P>pHppAHHh
MNP
aa
NNN
Won
MM
WWWWW
0.82121 0.97257 10502 96163
36203 85629
0.82639 12196 0.97381 01550 59548
61376 0.97500 21048
0.83147 23925 33162 51780
0.83645 69406
0.97614 82356 58492
72308
0.84134 47460 0.97724 98680 51821
68543
Sec. 3B.2 Series Approximation for Q(x) 149

Table 3B.1 Q(x). Source: Abramowitz, M. and |. A. Stegun, Handbook of Mathematical

Functions, New York: Dover Publications, 1970.
a es a ee
x 1 = Ox) x POG)
ee a EE ee ee ee
2.00 0:97724 98680 51821 3.00 0.99865 01020
2.02 0.97830 83062 32353 3,05 0.99885 57932
2.04 0.97932 48371 33930 3.10 0.99903 23968
2.06 0.98030 07295 90623 3.15 0.99918 36477
2.08 0.98123 72335 65062 3.20 0.99931 28621

2.10 0.98213 55794 37184 3.25 0.99942 29750

ail72 0.98299 69773 52367 3.30 0.99951 65759
2.14 0.98382 26166 27834 3.35 0.99959 59422

2.20 0.98609 65524 86502 3,50 0.99976 73709

exe 0.98679 06161 92744 3.55 0.99980 73844
2.24 0.98745 45385 64054 3.60 0.99984 08914
2.26 0.98808 93745 81453 RIES 0.99986 88798
2.28 0.98869 61557 61447 3.70 0.99989 22003

2.30 0.98927 58899 78324

2.32 0.98982 95613 31281 ae eee areal
2.34 0.99035 81300 54642 385 0°99994 09411
2.36 0.99086 25324 69428 3°90 0°99995 19037
2.38 0.99134 36809 74484 395 0°59996 09244

2.40 0.99180 24640 75404

2.42 0.99223 97464 49447 4.00 Oroeed Beers
2.44 0.99265 63690 44652 4.05 oreueautes ae
2.46 0.99305 31492 11376 4.10 5 be ere
2.48 0.99343 08808 64453 er fenoae peeas

2.50 0.99379 03346 74224

2.52 0.99413 22582 84668 4.25 0.99998 93115
2.54 0.99445 73765 56918 4.30 0.99999 14601
2.56 0.99476 63918 36444 4.35 0.99999 31931
2.58 0.99505 99842 42230 Das posses aeane

2.60 0.99533 88119 76281

2.62 0.99560 35116 51879 4.50 0.99999 66023
2.64 0.99585 46986 38964 4.55 0.99999 73177
2.66 0.99609 29674 25147 4.60 0.99999 78875
2.68 0.99631 88919 90825 4,65 So ete
4.70
2.70 0.99653 30261 96960
2.72 0.99673 59041 84109 4.75 0.99999 89829
2.74 0.99692 80407 81350 4,80 0.99999 92067
2.76 0.99710 99319 23774 4.85 0.99999 93827
2.78 0.99728 20550 77299 4.90 0.99999 95208
4.95 0.99999 96289
2.80 0.99744 48696 69572
2.82 0.99759 88175 25811 5.00 0.99999 97133
2.84 0.99774 43233 08458
2.86 0.99788 17949 59596
2.88 0.99801 16241 45106

2.90 0.99813 41866 99616

2.92 0.99824 98430 71324
2.94 0.99835 89387 65843
2.96 0.99846 18047 88262
2.98 0.99855 87580 82660

3.00 0.99865 01019 68370

150 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

let’s integrate by parts using f udv = uv — { vdu where

and get

Le ail x ale Zz
O= aS {©exp (-5) =| oe (-5) ac} (3B.9)

Since is 5 exp (-5) dz > 0, then for x > 0, we have

1 ae
Q(x) < ——— ex (-5) (3B.10)
J 20x ae
Also integrating by parts the second term in (3B.9) with

1 3
= ee du = me

we obtain
| 22 1 x2 oO 3 22
; noe ar az =~ SAD ore ~| AOD ro dz (3B.11)

Thus

Ol) ett (x l Fe 8 a i
x) = ——j-exp(—-—]-—
ae p 5 = &XP -—5 #f —ox (—_—=) cf (3B.12)

Ignoring the third term which is always positive, we have for x > 0

«)> =l (1-5)
1
ex (-5)
x2
3B.
Rilke Se \ ape ee erie

= (-8)=o(-2 anaes
Combining (3B.10) and (3B.13), we have

Sane ay p 5) =1O(%) r= Reet exp (-5) (3B.14)

This is an upper and lower bound on Q(x) for x > 0. This bound can be improved by
rewriting (3B.12) as

1 oe "00 3 i
1
Q(x) = one {(1— =)exp (-5) +x] a exp (-5) a:| (3B.15)
Sec. 3B.2 Series Approximation for Q(x) 151

Integrating by parts the second term with

ee A SOG).

we have

| ae x2 Pr(G)G6) 22
Oa) =
Fa (int w)or(-F) fo SP ew (-5)ac}caro
Again, integrating the second term by parts with
~ B&O) _ (3)(5)(7)
u= eg yp —— es eee
4 ke
z 2
dv =zexp (-5) dz, v=—exp (-5)

we have
2

O(x)=pees
anepie b 2aa3)4G)
{1 ier =
(3B.17)
2 (ove) 2)
+X exp (=) / ae exp (-5) a:|

From above, we obtain after integrating by parts n times

1 xe 2 (1)(3)(5)...
(2k — 1)
Q(x) = c= t= le 2 a1) eR
p( ) )| xk n (3B.18)

where R,, denotes the residue after n integrations and is given by

2 lee) ns
R, = (—1)"*!x exp (5) [ Sea (-5) dz (3B.19)
zent2

To simplify Rn, we write it as

Ry = (-D" 1G)... n+ Dx | eae (3B.20)

lo,@) Sle 2 p)

In the integral, let t = (1/2)(z? — x”). Then dz = (2t + x*)~'/*dt and

he i
[ (2t + x2)ntl

3 (3B.21)
jouO Red) fe BENG a
=p OO ey : (1+ 5) exp(—t)dt

or

...(2n-1
Rn = fey ee |(x) (3B.22)
152 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

where

pel
D(x) =i ( + =) ; exp(—t)dt (3B.23)
0 x

Note that ®(x) < | since

=
i: (:+- =)
ary} exp(—t)dt < i exp(—t)dt = 1
~
0 x 0

Let

l ¥2 n
(1)(3)(S)...
2k — 1)
Qn(x) = exp (-5) fne NN ee ener (3B.24)
J 2X k=1

be the n'” order approximation to Q(x), then the percentage error is given by

exp (-5) Rn |
Q(x)
— Qn(x)
% error = |—————_—
OG) _| = SOs (3B.25 )

Since ®(x) < 1, we have

exp ae
% error < J2nx
(-5) OO
Q(x) x2(n+1)
rs (3B.26)

Using the lower bound in (3B.13), we obtain

% error < DCIS) 1 Cra) ee (3B.27)

x2nt2 (1Ba 4)

Example 3B.1
Assume that x = 3. Calculate a simple bound on the percentage error when Q(x) is approxi-
mated by the first n terms in the asymptotic series, for n = 2, 3. Repeat for x = 5 and n = 2, 3,
and 4.
(a) For x =3

1)3)(5
ned Ghtor Se oa eaa
36 (12 )
(1)(3)(5)(7)
n = 3, % error < = 0.018 = 1.8%
38(1- 4)
Sec. 3B.2 Series Approximation for Q(x) 153

(b) For x = 5
DDO
n= 2, % error’< eine:
56(1 seei)
(1)(3)(S)7)
n = 3, % error < = 0.0003 = 0.03%
s8(1- 4)
n= 4, % error<
(DAE) = 0.0001 = 0.01%
Hel = 4)
 Appendix 3C
Gram-Schmidt Orthogonalization Procedure
The Gram-Schmidt orthogonalization procedure [15] is a useful tool in constructing a suit-
able coordinate system to define a decision space with the least possible coordinates or
equivalently, with the smallest dimensionality. Given a set S = {so(t), s1(t),--.,8u—1()}
of M signals, we wish to find a corresponding set of orthogonal signals, such that all signals
in the original set S can be represented as a linear combination of the orthogonal signals,
such that the number of the orthogonal signals is minimized. The procedure is very simple,
and it consists of choosing the first orthogonal signal f(t) as the first signal, that is

Bi(t) = so(t) (3C.1)

The second orthogonal signal 62(t) is that component of the second signal s;(t) which is
linearly independent of 6; (t), that is

> Bi (t)si(t)dt
Bo(t) = s1(t) — (ene |Bilt) (3C.2)
Jo*°
By@at
The third orthogonal signal 63(t) is again that component of s2(t) which is linearly indepen-
dent of £;(t) and B2(f), that is
Ts d

B3(t) = so(t) — Ga
Ts 2 Bi(t)
(t)dt
: Jo Bi (C.3)
Jo* B2(t)s2(t)dt
SE
9 By(t)dt

In general, we have

— ( Jo’ Bi Osx-1at
Belt) = se-1) — )|| > —— |8) (3C.4)
j=l Jo’ By (t)dt
for allk =2,..., M. To get these signals normalized, we simply define [for B;(t) 4 0]
Bi(t) :
(O=——— t= 1M GC5)
Jo’ Beat
to obtain an orthonormal (orthogonal and normalized) set of signals. The same procedure is
applicable to vectors by noting the 1s si(t)sj(t)dt =s!;.
Given the orthonormal set of signals (or vectors), each signal s;(t) (or vector s;) can
then be written as a linear combination of the orthonormal functions (or vectors) as
N
si(t) =) sijj(2) (3C.6)
j=l
for 0 <t < T, and s;(t) = 0 elsewhere. Here
Ts
Sij= [ Sj (t)pj(t)dt (3C.7)
App. 3C Gram-Schmidt Orthogonalization Procedure 155

Note that the number of orthonormal elements N is always less than or equal to the number
of original signals M. If the set S consists of linearly independent signals, then M = N
and the number of orthonormal elements is identical to the number of signals. Once the
set of orthonormal functions {¢;(t)} are identified, each transmitted waveform given in
(3C.6) is completely determined by the vector transpose s! = (Filo HApc ooo Gon) or all =
0, 1,..., M — 1. In vector notation, each N-tuple denotes the vector

sj = > si; (3C.8)

where the inner product oo j = 4ij and where @ ; denotes the unit vector along the j'*-axis,
j =1,2,..., N. In this sense, we can visualize the M vectors {s;} as defining M points
in the N-dimensional Euclidean space, the signal space. The coefficients s;; in (3C.7) are
equivalently found from the vector s; by evaluating the projection s;; = s! o ; for alli and j.
Example 3C.1
Consider the following three vectors:

si = (1,1, 1),s} = (1,0, 1),sf= (0, 1,0)

Then

Bi =si = (1,1,1)
T
piPore
=s' |(SiS
Pre Jaf
ee=c.0.0 oe(1,1,
1)ae
=3(1,2 2,1)

eras - (Bi)ot(25)0
T T

I
= (07 1,0) = 3 (1 LL) alae. 1) = (0, 0, 0)

Note that B, can not be in the new set since by definition, the elements of a basis are a set of
nonzero vectors. Check orthogonality property of the new vectors:

pip, =*-2**
L—2+1
=o
Normalizing B, and B,, we have

fe
TM shied 1)
Cig a”
4i
Re Js
$2 |B>| Gen a

Expressing the original signals in terms of the new orthonormal vectors, we have
156 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

80 = (85 01), + (8)bo). = V30)

5 See, Harenee ele

fet ae
Fe
1 2,
Lael
So = So — 8S} = —@, - —
wales?
In terms of the new orthonormal base (@,, @), we can write

ik T ae? if 1 2
So = (J3, 0), s; = ( ’ Ve Ss = ( a aes )
2 v3 A/6 NS ans
We have therefore defined a new observation space @, and @, such that the signals can be
represented in two dimensions. Note that in terms of the standard base o, = (50:0); o3 =
(0, 1, 0), and o; = (0, 0, 1), the signals were represented using three coordinates even though
they lied in two dimensions (that is, a plane). As can be seen from this example, choosing
the appropriate base can reduce the dimensionality of the observation space and could greatly
simplify the detection problem and reduce the receiver’s operations.
 Appendix 3D
Connecting the Linear Vector Space to the Physical Waveform Space

The relationship between signals and vectors will be needed; this suggests the possibility
of a geometrical representation of signals and an application of analytic geometry. Consider
an N-dimensional vector space with unit vectors @,, @>,..., @y along the coordinate axis.
These vectors are called the basis vectors and form a complete set. The set of N basis
vectors is not unique. One can choose any N independent vectors; they may or may not
be mutually orthogonal. In our work, we shall assume the set of basis vectors to be mutually
orthogonal and each normalized to a unit length. Such a set is called an orthonormal set.
Once the N basis vectors are chosen, any vector r in this space can be specified by
N numbers r? = (r},72,..., rN), which represent the magnitude of components of r along
the N basis vectors, respectively. The vector r can be expressed as
N
ee aCe (3D.1)
n=

where, for normalized basis vectors, the projection of r onto @,, is the scalar product
Lage t @,
reat: (3D.2)

for all n. The vector r can be represented geometrically by a point (71, r2,...,7N) in the
coordinate system formed by the basis vectors.
We have an analogous situation for signals which evolve with the passage of time.
Analogous to the base vectors, suppose we have N waveforms ¢$1(t), $2(t),..., @n(t)
which form an orthonormal set of basis signals, N may be finite or infinite. Then , an ar-
bitrary signal r(t) in this N-dimensional Euclidean space can be expressed as
N
r(t}=) raga@), O<t<T (3D.3)
n=l

where, for orthonormal basis signals, the projections of r(t) onto the n!” basis signal is
7
tar i r(t)dn(t)dt (3D.4)
0
for all n. The set of basis signals {f,(t)} must form a complete set of signals for the space.
In practice, the basis signals are usually generated using a local oscillator to generate the
inphase and quadrature components 2/7 cos wct and /2/T sin wet.
Once the basic signals {#,(t)} are specified, one can also represent the signal r(t)
by an N-tuple (r1,7r2,..., 7). Alternatively, we may represent this signal geometrically
by a point (r1,7r2,...,7N) in an N-dimensional space. We can now associate a vector r
with signal r(t). Note that the basis signal $)(t) is represented by the corresponding basis
vector @, and ¢2(t) is represented by @, and so on. The vector r is called the projection
of r(t) onto the vector space. In the applications that follow, the linear vector space needed
by the optimum vector receiver is created from the physically observable waveform r(t) =
s(t) +n(t) by projecting the waveform onto that space. Such projections are equivalent to
filtering and sampling the received waveform, namely, the operation of cross-correlation.
158 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

The scalar product. In a certain signal space, let s(t) and r(t) be two signals
represented by their projection vectors s? = (51, 52,..., SN) and r? =(r,1r2,..., rn). If
{dn (t)} are the orthonormal basis signals, then
N
s(t) =) )snGn(t) (3D.5a)
n=1

r(t) = >> rnGn(t) (3D.5b)

n=1

Hence,
iD N

/ r(t)s(t)dt = eres (3D.6)

0 n=1

since

[‘ Pn(t)bm(t)dt dee
= (i a
eee (3D.7)
However, the right-hand side of the above equation is by definition the scalar product or
inner product of vectors r and s
if
r’s= i}r(t)s(t)dt (3D.8)
0

We conclude that the integral of the product of two signals is equal to the scalar product of
the corresponding vectors.

Energy of asignal. For an arbitrary signal s(t), the energy E is given by

T
E= i s(t)dt Joules (3D.9)
0

It follows from (3D.5Sa) that

E=s's=|s/? (3D.10)
where |s|? is the square of the length of the vector s and |s| is called the norm of s. Hence,
the signal energy is given by the square of the length of the corresponding vector.
Example 3D.1
Consider the two-dimensional signal vectors depicted in Fig. 3D.1. For Binary On-Off Keying
(OOK), we use the inphase component generated by a local oscillator to write

so = 0 > so(t) =0

s| = VE¢, > s(t) = V2E/T cos wet

for 0 <t <T. For Binary Frequency Shift Keying (BFSK), we use the inphase and quadrature
components of a local oscillator to write

So = VE¢, — so(t) = /2E/T cos wet

s) = VE$, > 9\(t) = J2E/T sinwet

 App. 3D Connecting the Linear Vector Space to the Physical Waveform Space 159

$4

So 34 So S} So

|si| = JE |so| = |s1| = VE |so| = |s1| = JE

(a) On-off keying (OOK) (b) Binary frequency shift keying (BFSK) (c) Binary phase shift keying (BPSK

Figure 3D.1 Signals in Example 3D.1

for 0 < t < T. For Binary Phase Shift Keying (BPSK), we have

sy = —S; = VE, > s(t) = —51(t) = V2E/T cos wet

for0
=f 7.
 Appendix 3E
The Pairwise Probability of Error (PPE): A Tool for System Performance
Evaluation and Signal Design

One of the most powerful analytical tools used in the development of communication theory
is that of linear vector spaces, needed in the representation of signals (codes). In vector com-
munications, the Pairwise Probability of Error (PPE) depends on the Euclidean distance (a
true mathematical distance measure induced by the inner product or cross-correlation) and
the signal space with this distance forms a Hilbert space. Analysis and synthesis of vector
communication system techniques can be reduced, in many practical cases, to consideration
of the geometrical features of this Hilbert space.
A basic building block in the development of M-ary vector communications is the
so-called PPE. It also serves as the basic building block in the development of the Union
bound, which is used to characterize the symbol error probability (system performance) and
is composed of M — | terms, each of which is a PPE. This bound is used to develop the
ensemble symbol error probability arising in the Shannon theory, which demonstrates that a
code exists that exhibits a zero asymptotic error probability in the presence of noise, under
some conditions.
The pairwise error event is defined as

Ej & [rite — sj < ir —s¢P?/mi| (3E.1)

where we have assumed that message m, (signal s,) was transmitted and s; is any other
signal (code) vector in the transmitted set. The geometry of this event is illustrated in Fig.
3.E1. With r = s; + n, then the event can be equivalently written as

Eje=[r:2n"(se— 8)) + |sx — 8j1 < 0/me|

ay

where dj, denotes the distance between the vectors. Defining the unit vector ej,A (s;—
sx)/|S; — S,| and introducing the random variable z(u) = ne jk» One can write the pairwise
error event as

Ej. = Eo > “A (3E.3)

The probability of this event is just one minus the value of the probability distribution
function of the random variable z(u) evaluated at d jk/2, that is

P(E jx) & Po(Sj, sx)

d;

=p (<a > “jms. (3E.4)

which says that the pairwise error event occurs if the projection z(u) of the noise vector
onto the unit difference signal vector e jx is greater than one-half the distance between signal
(code) vectors s; and s; as shown in Fig. 3E.1. It is easy to show that z(u) is a zero-mean
App. 3E The Pairwise Probability of Error 161

ae

PAIRWISE ERROR
DECISION BOUNDARY

PAIRWISE ERROR EVENT: [z(u) > d jy/ 2]

Pr [z(u) 2 djy/2 ] = Pr [r:|r-sj| s|r-s,]/m]

«© d ik 2
dik‘ /4No
= exp [-22\o Jaz = a| |<@
I J™No V ‘0

e Characterizes exact performance of a binary communication system

employing two equally likely messages

e Holds for two arbitrary vectors in any normed vector space

e Basic building block in developing:

* Shannon's random coding theorem

e Union bound (which characterizes performance for an arbitrary

set of signal vectors)

« The ensemble symbol error probability (Rg -criterion)

¢ Performance of systems employing sequence estimation

Figure 3E.1 Pairwise probability of error (PPE)

Gaussian random variable with variance 0* = No/2. Using this fact in (3E.4), we obtain
1 CO

PS); 5.) = = / exp(—z7/No)dz (3E.5)

VINO Jdjx/2
which is just the Gaussian integral
aie
P(s;,: Sk) = O(——)
J2No
(3E.6)
for all 7 #k. This PPE has numerous applications. For example, it characterizes (1) the
 162 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

exact error probability performance of all binary communication systems transmitting equal
energy, equal probable signals, (2) the symbol error probability performance of an M-ary
system is Union bounded by a linear combination of PPE’s, that is
M-1
P(E/my)< 2 Pris}. (3E.7)
ink
This suggests that the performance of an M-ary vector communication system is bounded
by a linear combination of the performances of M — | binary systems employing signal
vectors $; and Sx, that is
M-1M-1
P(E) < DD) Ply) Pa(Sj, St) (3E.8)
k=0 j=0
i#k
Finally, there is no loss in generality in allowing s; and sx to be any two arbitrary
encoded data sequences (vectors) so that (3E.6) can be used in evaluating the performance of
coding-decoding algorithms of the sequential type, for example, decoding of convolutional
codes using the Viterbi algorithm. This CORY, building block will be applied in Chapters 4,
12, and 13. If one identifies the vectors Ss) = (Sm, ---,Smn) With the coded message m of N
symbols, then the PPE that message m ’ (consisting of another coded message of N symbols)
represented by si SeNS ras! Tone 6 ae Sen ES USE

d2,
Pim->m)=0( 4 (3E.9)
V2No

where Ci A |Sm —Sm’|’ /|2 denotes the squared Euclidean distance between the two vectors
Sm and S,,’. This result will be used numerous times in later chapters.

PROBLEMS
3.1 An observation x is defined as follows:
Ah: x=mb+n
lava eNO

CLG b and n are two independent, zero-mean Gaussian random variables with variances op
and Cp respectively and m is a constant. Assuming equiprobable hypotheses,
(a) Find and sketch the optimum (MAP) receiver.
(b) Calculate the minimum probability of error for the receiver found in (a).
(c) Repeat (a) and (b) when the random variable b has mean value my.
(d) Using (c), find the error probability when 0, — 0 and discuss your result.
3.2 An observation x is characterized under two hypotheses according to the following pdf’s:

Ho: —fx(x/ Ho) = |/Fexet- cae

53 e8)

Hie!) ie cai
0,
i
we ())
The probabilities of the two hypotheses Ho and H; are 5/8 and 3/8, respectively.
Problems 163

(a) Find the minimum error probability test to choose between Ho and A).
(b) Find the minimum error probability that results from the test in (a).
3.3 Consider the following scalar communication problem:

Hi: fx(x/ Hi) = ce Os

0, x <0
Assuming equiprobable hypotheses,
(a) Find a likelihood ratio test that minimizes the probability of error.
(b) For mo = 0 and m, = 1, it is convention to refer to the probability of choosing H; when
indeed Hp is true as the false alarm probability, Pra. Similarly, the probability of choosing
Hj, when indeed H is true is referred to as the probability of detection, Pp. Evaluate Pr 4
and Pp numerically.
3.4 Consider the following scalar communication problem:

662 JeSop thi sO, Il

where so = 1, s} = —1 and

1 1 : o| ae
f—a)- fe 1 . o| (n +a)?
ee
fr(n) = 5
g 256 Digg 256

and the hypotheses are equiprobable.

(a) Assuming that a < 1, describe the optimum receiver that will yield the minimum probabil-
ity of error. Repeat for a > 1.
(b) Find an appropriate expression for the probability of error in each case.
3.5 Consider the scalar communication problem where one of two voltage levels so = 0 and s; = 1
volt is transmitted over an additive noise channel where the noise pdf is characterized by

SSS ee <n<c7
fn(n) _
m(1 +n?)

(a) Determine the likelihood ratio test that specifies the maximum-likelihood receiver.
(b) Evaluate the false alarm probability Pr4 = Pr{H is chosen/ Hp is chosen} and the prob-
ability of detection Pp = Pr{H, is chosen/H is chosen} if the threshold voltage is set to
0.5 volts. Also calculate the probability of error.
3.6 (a) Using just a single observation, what is the optimum receiver to choose between the hy-
potheses
“i . 5 r D)
Ho: the sample is zero-mean Gaussian, with variance 09
r é i : D) ) 9
H,: the sample is zero-mean Gaussian, with variance oy (oy > 99)

if the a priori probability of H; is twice that of Ho.

(b) In terms of the observation, what are the decision regions /o and /;?
(c) What is the probability of choosing H; when Ho is true?
(d) Determine the probability of error if a; /ont =e ando; = 1.
3.7 Based on N independent samples, design a test to choose between

LA at
Ate Sa len
Hy: r=-l+n”

where n is a zero-mean Gaussian random variable with variance 0”. Assume equal a priori
164 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

probabilities for each hypothesis. Show that the test statistic may be chosen to be

a ri
=|-—
i=l

Find the decision region for 7.

3.8 Given the Poisson distributed random variable K with parameter x, that is
k,-x
Prob {K =k} = —, Her—i()). ila poe

Assume that x takes on one of two known values, say

Jaye (PS

Joh =a sew

with equal probability.

(a) Develop the decision rule which minimizes the probability of error in announcing the value
of x given the observable k.
(b) Assume now that x = 4 if Hj is true and x = 2 if Hp is true, simplify the decision rule
found in (a).
3.9 A binary hypothesis test is made on the basis of the observed statistics:

Hy > fr(p/ Ho) ae

p exp (—p?/20?) 5 Oz2U
o2

2 2
p po +a ap
i felo/ ty =F exp (= Fo2 )0(%). p=0

where q@ and o are known and /(x) is the zero-order modified Bessel function of the first kind.
(a) Specify the likelihood ratio test if

]
Prob{ Ho} = Prob{ H,} = 5

(b) Write an expression for the probability of false alarm, Pra.

(c) Write an expression for the probability of detection, Pp.
(d) Write an expression for the probability of missed detection, Py.
(e) Write an expression for the probability of error, Pr.
(f) Table P3.9 lists approximate values for the Marcum Q-function for integer arguments. That
is, it lists Q(a, B) where @ is as shown in the left margin and £ is given in the top row.
More complete tables can be found in papers by Marcum. What is the threshold setting
which gives Pr, = 0.01 (assume o = | volt)?
(g) For this threshold setting, determine the sensitivity of the false alarm probability to in-
creases and decreases of o by +25% about one volt. (Plot results.)
(h) Repeat (g) for Pp as a function of a/o.
Note: Marcum Q-function!
lo) me! 2
Q(a, B) & i x exp (-* “ )Io(ax)dx
B

3.10 Consider the following binary decision problem representing the detection of a known signal s

1. See M. S. Roden,““Digital and Data Communication Systems,’ Englewood Cliffs, N.J.: Prentice Hall, 1982.
Problems 165

Table P3.9
ce ee a 8 oe a
B
a 1 2 3 4 5 6 7 8 9 10
A i A OE ea a 0
1 0.875 0.268 0.044 0.003 0 0 0 Oe 0 0
2 ONS" = 01603, 970213" (501034. 0,002 0 0 0 0 0
3) 0991-05885 0595 —0:195° 0.031 ~~0002 0 0 0 0
4 0.999 ©0986 0:875 0.575 «420.186 ~=—:0.029._-—«- -:0002 0 0 0
5 0.999 0.999 0.984 0.865 0.564 0.180 0.027 0.002 0 0
6 0.999 0.999 0.999 0.984 0.860 0.543 0.176 0.027 0.002 0
TE FOSO9 E9999 50,999 920.999. 0.981 <0:855.. 0523 0.173 0,026... 0.002
8 0.999 0.999 0.999 0.999 0.999 0.980 0.850 0.511 0.170 0.026
9 0.999 0.999 0.999 0.999 0.999 0.999 0.977 0.845 0.497 0.167
10 0.999 0.999 0.999 0.999 0.999 0.999 0.999 0.975. 0.840 0.482

in additive white noise n:

Message m9: r=n

Message mj: r=s+n

The pdf which characterizes the additive random noise voltage is the generalized pdf

fn(a) = [en(o, c)/2T (1/c)] exp {—n(@, c)|a|°}

where

ioe) Bo IP Gle)/ Tey

for |a| < oo. Here c is a known positive parameter controlling the rate of decay, I'(.) is the
Gamma function, and o7 is the noise variance. Note for c = 2, this density reduces to the
Gaussian density whereas for c = 1, it becomes the Laplace density.
(a) Find the optimum decision rule which minimizes the probability of error. Assume that
Prob{mo} = Prob{m,} = 1/2.
(b) Sketch the optimum detector suggested by the rule found in (a).
(c) Find an expression for the error probability.
(d) Repeat (b) for c = 1 and c =2.
3.11 Consider the following binary hypothesis testing problem:

let 0 Vesa)

Higa — sce

where s and n are independent random variables with

fs(S) =aexp{—aS}, S>0

fr(N) = bexp{—bN}, N>O

Find the decision region which minimizes the error probability if Ho and Hj are equally
probable.
166 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

r=s+n

Figure P3.15

3.12 The two equally likely hypotheses are

1
Hy: UIA) Ss OSE):

Hy: f-(R) === exp(—=R?)

(a) Find the optimum decision rule.
(b) Compute the decision regions.
3.13 Let the conditional densities for a binary detection problem be given by the Cauchy densities
1
fem) = Tyee 1 =O
If the two messages mo and m, are equally likely and if ap and a, are known constants with
a| > ao,
(a) Find the a posteriori probabilities for the messages, P(mo/r) and P(m,/r);
(b) State the decision rule that will result in minimum probability of error;
(c) Compute the resulting probability of error.
3.14 The likelihood ratio is defined as

as fr(e/ A)
A,(p)
Fr(e/ Ho)

If we replace the observation p in A,(p) by the random variable r, then A,(r) is itself a random
variable with special properties. Prove:
(a) E{A,(r)"/Hi} = E{Ap*!(r)/Ho}
(b) E{A,(r)/Ho} = 1
(c) E{A,(r)/Hi} — E{A,(r)/Ho} = Var{A;(r)/
Ho}
3.15 The random variable n in Fig. P3.15a is Gaussian, with zero-mean. If one of two equally
likely messages is transmitted, using the signals of Fig. P3.15b, an optimum receiver yields
PEs O00K
(a) Compute the variance o? of n.
Problems 167

-4 0 4

(a)
Figure P3.16

(b) Erasures are sometimes employed in receivers to allow them to avoid a decision when it
seems that a decision might contribute an error. Consider the decision rule of Fig. P3.15c:

ifr > $, decide 5)

if r< —$, decide so
it |/Alex g, decide an erasure

Compute d such that the erasure probability is 10-4. Compute all necessary probabilities
P(r;/m) to reduce the continuous channel to a discrete memoryless channel (DMC) with
2 inputs (M = 2) and 3 outputs (J = 3).
(c) In some applications, it is better to quantify the quality of the received symbol. This is typ-
ically accomplished by measuring the confidence level of the observed voltage. Consider
the decision rule depicted in Fig. P3.15d. In this case, the receiver distinguishes between
a soft “1” (denoted by 1—) and a strong “1” (denoted by 1+). The new decision rule is as
follows:

(ijp —§, decide a strong “0” (0+)

if —$ <r <0, decide a soft “0” (0-)
if 0<r < $, decide a soft “1” (1-)
i 7 g, decide a strong “1” (1+)

Compute d such that the probability of any soft symbol is 10~*. Compute all necessary
probabilities P(r; /m ;) to reduce the continuous channel to a discrete memoryless channel
(DMC) with 2 inputs (M = 2) and 4 outputs (J = 4).
3.16 The random variable n in Fig. P3.15a is Gaussian, with zero-mean. If one of two equally
likely messages is transmitted, using the signals of Fig. P3.15b, an optimum receiver yields
P(E) = 0001.
(a) What is the minimum attainable probability of error when the channel of Fig. P3.15a is
used with the three equilikely signals of Fig. P3.16a?
(b) How about with the four equilikely signals of Fig. P3.16b?
(c) How do the answers of (b) change if it is known that the noise has mean m.
3.17 A communication system employs two channels to transmit a voltage s;, i=0,1, as shown in
Fig. P3.17. The voltages sg and s; correspond, respectively, to the messages mo and m;. Two
received voltages r; and rz are available to the decision device, that is, the receiver makes its
decisions based on a joint observation of r; and r2. Assume that the additive noises n; and n2
are zero-mean Gaussian random variables with variances o; and oF, respectively. Also, assume
that s;, 11, and np are statistically independent. Letting so = VE with probability P(mg) and
s} = —VE with probability P(m)),
(a) Determine the structure of the optimum receiver that results in the minimum probability of
error.
168 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

ue)

DECISION
TRANSMITTER DEVICE

Figure P3.17

s;° a

Figure P3.19

(b) Calculate the average probability of error for the case P (mo) = P(m,) and o? =o.
3.18 (a) Sketch the optimum vector receiver for each of the signal sets illustrated in Fig. 3D.1.
(b) Sketch the decision regions and boundaries.
(c) Compute the minimum probability of error for each case.
3.19 (a) Given two equilikely messages, mo and mj, to be transmitted (using the signals illustrated
in Fig. P3.19) over a channel that adds white Gaussian noise, find an expression for the
minimum error probability.
(b) What can one conclude regarding the energy of the two signals from the result found
in (a)?
3.20 Consider the following N-dimensional decision problem: the received vector r = s; + n where

sais iy, 2) Pa yD, 354

= (ifn EN)
Assume that the noise vector is Gaussian with mean 0 and covariance matrix o?Z, where T is
the N x N identity matrix.
(a) Design a decision rule to choose between the four equally likely hypotheses.
(b) For the case N = 2, assume the received vector is r = (4, 0)’. Which hypothesis would the
receiver choose?
3.21 The two equally likely hypotheses are

1 xe xh
Hos. Fis %2/ Ho) = XD ae
2
Ne OC eG
20 or 209 2092

1 x2 x2 x? x?
Hictaas (Xiseos exp (i ee aeexe ee
41 0001 Ss 207 209 P 20 20; E

—COl< Xoo
Problems
169

Circle of radius VE. Radius

Figure P3.23

Find the optimum decision rule.

3.22 One of four equilikely messages is to be communicated over a vector channel which adds
to each transmitted vector component a statistically independent zero-mean Gaussian random
variable with variance No/2.
(a) Assuming that the transmitter uses the signal vectors shown in Fig. P3.22a, express the
probability of error in terms of the function Q(x) defined in (3.40).
(b) Assume that the optimum receiver is modified such that any received vector in the square
centered at the origin and of ‘diameter d is declared an erasure as shown in Fig. P3.22b.
Compute the probability of erasures as a function of d and E,/No. Compute all necessary
probabilities P(r; /m ;) to reduce the continuous channel to a discrete memoryless channel
(DMC) with 4 inputs (VM = 4) and 5 outputs (J = 5).
(c) Repeat (b) if the square is replaced by a circle of radius d as shown in Fig. P3.22c.
3.23 For the equilikely signals illustrated in Fig. P3.23, sketch the optimum decision boundaries in
an AWGN channel.
3.24 When the two signal vectors so and s; shown in Fig. P3.24a are transmitted over a channel
170 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

So S1

(a)
Figure P3.24

Venere
d
Ss Lea S
7
ie Go
N ” wo N N.
N

Figure P3.25

Figure P3.26

disturbed by additive white Gaussian noise, an optimum receiver will achieve P(E) = q for
equilikely messages. Assuming that the seven vectors indicated by x’s in Fig. P3.24b are used
as signals over the same channel, compute the probability of error achieved by an optimum
receiver in terms of g, 0, and /, assuming equilikely messages.
3.25 Consider the PCM (pulse code modulation) signaling system where each PCM word has a
length of 3 bits, that is, there are 27 = 8 possible symbols. The signal vectors are placed on
the vertices of a “3-dimensional hypercube” as shown in Fig. P3.25. Assume all signals are
equilikely and that the noise is an AWGN with PSD of No/2 watts/Hz. Find an expression for
the probability of symbol error.
3.26 Consider the 6-QAM constellation set depicted in Fig. P3.26. Assume all signal vectors are
equilikely to be transmitted and that the channel is disturbed by AWGN with PSD No/2.
(a) Find the boundaries for the decision regions.
(b) Find the probability of symbol error.
 Problems 171

Figure P3.27

135° 45°

AMPLITUDE = y2-
180° 0° ABSOLUTE

270° Figure P3.28

Sil Consider the 4 x 4 QAM signal constellation depicted in Fig. P3.27. Assume an AWGN chan-
nel and equilikely signals.
(a) Draw the optimum decision boundaries and illustrate the decision regions.
(b) Using an appropriate basis, write an expression for each of the 16 vector signals.
(c) Compute the average energy for this signal constellation.
(d) Draw the optimum detector.
(e) Write an expression for the probability of symbol error.
3.28 Consider the V.29 signal constellation depicted in Fig. P3.28, which transmits 3 bits/symbol.
Assume an AWGN channel and equilikely signals.
(a) Draw the optimum decision boundaries and illustrate the decision regions.
(b) Using an appropriate basis, write an expression for each of the eight vector signals.
(c) Compute the average energy for this signal constellation.
(d) Draw the optimum detector.
(e) Write an expression for the probability of symbol error.
3.29 Consider the signal constellation shown in Fig. P3.29. This signaling scheme is referred to as
8-QAM (quadrature amplitude modulation). Note that the energy of each signal is not equal.
(a) Find the average energy of the signaling set. Assume all signals are equilikely.
(b) Find an upper bound for the probability of symbol error.
3.30 Consider a 3-dimensional “biorthogonal” signaling scheme as shown in Fig. P3.30. Assume
all these signals are equilikely and that they are sent over an AWGN channel with PSD No/2.
Using the union bound, find an upper bound for the probability of error as a function of E;/No
where £, denotes signal energy.
172 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

Figure P3.29

Figure P3.30

3.31 Consider the V.29 signal constellation depicted in Fig. P3.31, which transmits 4 bits/symbol.
Assume an AWGN channel and equilikely signals.
(a) Draw the optimum decision boundaries and illustrate the decision regions.
(b) Using an appropriate basis, write an expression for each of the 16 vector signals.
(c) Compute the average energy for this signal constellation.
(d) Draw the optimum detector.
(e) Write an expression for the probability of symbol error.
3.32 Consider a 32-QAM signal set as shown in Fig. P3.32. Assume all signals are equilikeiy and
are transmitted over an AWGN channel with PSD No/2.
(a) Calculate the “average” energy of this set.
(b) Find an upper bound for the probability of symbol error.
3.33 Consider the problem in which you must decide which of two hypotheses is true, based on the
observation of x:

Hy: x=so+no

Ay Ve sq 27)

where sg and s; are known constants and no, n; are zero-mean independent Gaussian random
variables with respective variances OG and at:
Problems 173

Figure P3.31

Figure P3.32

(a) Derive the decision rule that minimizes the probability of error (assuming equilikely prob-
abilities) and write expressions to characterize P(error/ Ho) and P(error/H}).
(b) Suppose that the problem is reformulated as follows: decide which of two hypotheses is
true based on the observation of the vector x where
Ho: x=Sot+n

A: x=s,;4+n

where so = (so, 0)7, s; = (0, 51)’ and n is a zero-mean Gaussian random vector with
covariance matrix
oO2
bal cine)
0
Kn=( °
174 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

What is the receiver that minimizes the probability of error?

(c) Write expressions to characterize P(error/Ho) and P (error/ Hj ) for the receiver in (b).
(d) Are the two receivers equivalent in performance? Explain your answer.
3.34 Given a random vector x with zero-mean and covariance matrix

if W1/2- =1/2
r= (-172 1 =")
=i) ltt
(a) Let the random variable y be formed by the inner product y = h’ x, where h is constrained
to have unity magnitude. Find the h vector which maximizes the mean-squared value of y.
(b) Determine the mean and variance of the random variable y with h chosen as in (a).
(c) Let w = Gx. Find the matrix G such that the resulting vector w has a covariance matrix
equal to the identity matrix (that is, perform a whitening transformation on x).
(d) Express x as
n

x= yy J riwiei
i=l

where the w;’s are zero-mean, uncorrelated random variables with unit variance (that is,
find the A;’s and the e;’s).
3.35 Let x be a correlation matrix with zero determinant. Prove that the matrix ee has eigenval-
ues zero or one where €,, is the matrix whose columns are the eigenvectors corresponding to
nonzero eigenvalues of 7Rx.
3.36 Based on the observation of the random vector x, determine the detector that will minimize the
probability of error between the two equilikely hypotheses:

Ho: x=(1,1,0,0)7
+n
Hy x=(0,031,-1)
ton
where the covariance matrix of the Gaussian zero-mean noise vector is given by

al 0 0
a OFeEO
Kn =
ORO leo
On0 St l

Hint: Is the covariance matrix singular? Is perfect decisions possible? How about preprocessing
the observation x with a matrix 7 such that Hx has a nonsingular covariance matrix?
O55, You are asked to design an optimum decision rule to decide between the following two hy-
potheses:

Ho: x(n) =so(n)

+ w(n)

A}: x(n) =s\(n)

+ w(n)
where so(n) = cos(nz), s;(n) = cos((n + 1)z)) and pz(n) is a Gaussian zero-mean wide sense
stationary (WSS) sequence.
(a) Based on observing x(10) and x(11), design the optimum decision rule assuming {1(n) is
white with PSD S,,(f) = 0.
(b) Bound the probability of error of the receiver in (a).
(c) Derive an exact expression for the probability of error for the receiver in (a).
Problems 175

(d) Assuming that the noise s(n) is no longer white, but has PSD

Su(f) =
ga ee) with
0 < |p| < 1
1 — 2ocos(27f) + p?
derive the optimum decision rule. Check to see that when p = 0, the decision rule reduces
to that in (a).
(e) Bound the probability of error of the receiver in (d).
(f) Derive an exact expression of the receiver in (d). Check to see that the expression reduces
to that in (c) when p = 0.
3.38 A decisioning system must decide between the hypothesis Hp and H; where

1 2
Ho: frlp)= = exp(—>)
apI WES Shes)
Joie ok Setheee
fr(0) | 0, otherwise
The a priori probability of Ho is 3/4 and H, is 1/4. The cost of decisioning is

Co =0 Ci=1

Co=1 Cy =90
Determine and describe the optimal Bayes decision procedure for this problem.
3.39 Consider the hypotheses

ley Ra

Aye Sr= sean

where s and n are independent random variables with pdf’s

j(S\=exp(—S), S20

fr(N) = 10exp(—10N), N>=O

(a) Compute the conditional pdf’s f,(p/Ho) and f,(e/ 1).
(b) Find the likelihood ratio test, A,(p).
(c) If Ho has a priori probability 1/4 and H, a priori probability 3/4, Co, = Cio = 3 and
Coo = C1; = 0, find the threshold for a Bayes test.
(d) Show that the likelihood ratio test of (c) can be expressed as

Ho
pe y
A
Find the numerical value for y for the Bayes test of (c).
(e) Compute the risk for the Bayes test in (c).
(f) Compute the threshold for a Neyman-Pearson test with Pr, less than or equal to ime
Find Pp for this threshold.
(g) Reducing the Neyman-Pearson test of (f) to

find Pr, and Pp for an arbitrary y.

 176 Scalar and Vector Communications Over the Memoryless Channel Chap. 3

3.40 A decisioning system must decide between two hypotheses with pdf’s

fr (e/ Ho) = 2exp(—2p), p29

fi(o/M1) =1/2, -lsp <1

Assuming that the a priori probabilities are unknown, formulate a decision rule using the min-
max criterion. j
3.41 A decisioning system must decide between two hypotheses with pdf’s

filo Ho) =}4 lp,” jor=<1

Ip| >1
Ef.2a les, alela2
fr(e/ Hi) = rf \p| —)

Choosing Hy when H, is true costs three times as much as choosing H; when H, is true.
Correct choices cost nothing. Formulate a decision rule using the min-max criterion.
3.42 Show that the vector
M-1
a= P(mj)s;
i=0
minimizes the required mean energy
_ Mal
E’ = )° P(m)|s; — al?
i=0
without altering the message error probability.

REFERENCES
1. Neyman, J. and E. Pearson, “The Testing of Statistical Hypotheses in Relation to Probability A
Priori,’ Proc. Comb. Phil Soc., vol. 29, (4), pp. 492-510 (1933).
2. Neyman, J. and E. Pearson, “On The Problem of the Most Efficient Tests of Statistical Hypothe-
sis,” Phil. Trans. Roy Soc. (London), Ser. A, 231(9), pp. 289-337 (1933).
3. Wald, A., Statistical Decision Functions, New York: John Wiley and Sons, 1950.
4. Bayes, T., “An Essay Toward Solving a Problem in the Doctrine of Chances,” Phil. Trans. Roy.
Soc. (London) , pp. 370-418 (1763). Reprinted in Biometrika 45, pp. 293-315 (1958).
5. Papoulis, A., Probability, Random Variables and Stochastic Processes, New York: McGraw-Hill,
1965.
6. Larson, H. J. and B. O. Shubert, Probabilistic Models in Engineering Sciences, Vol. I: Random
Variables and Stochastic Processes, New York: John Wiley and Sons, 1979.
7. Weber, C. L., Elements of Detection and Signal Design, New York: Springer-Verlag, 1987.
8. Strang, G., Linear Algebra and Its Applications, 3rd ed., San Diego: Harcourt Brace Jovanovich,
1988.
9. Friedberg, S. H., A. J. Insel and L. E. Spence, Linear Algebra, Englewood Cliffs, N.J.: Prentice-
Hall, 1979.
10. Wozencraft, J. M. and I. M. Jacobs, Principles of Communication Engineering, New York: John
Wiley and Sons, 1965.
11. Blackwell, D. and M. A. Girshick, Theory of Ganies and Statistical Decisions, New York: Dover
Publications, 1954.
References 177

12: Van Trees, H. L., Detection, Estimation and Modulation Theory, Part 1, New York: John Wiley
and Sons, 1968.
13% Helstrom, C. W., Statistical Theory of Signal Detection, 2nd ed., Oxford: Pergamon Press, 1968.
14. Melsa, J. L. and D. L. Cohn, Decision and Estimation Theory, New York: McGraw-Hill, 1978.
ID: Friedberg, S. H., A. J. Insel and L. E. Spence, Linear Algebra, Englewood Cliffs, N.J.: Prentice-
Hall, 1979.
16. Craig, J. W., “A New, Simple and Exact Result for Calculating the Probability of Error for Two-
Dimensional Signal Constellations, ” IEEE MILCOM’91 Conference Record, Boston, MA, pp.
25.5.1-25.5.5.
 Chapter 4
Coherent Communications
with Waveforms

In the previous chapter, we investigated the problem of scalar and vector communications
and derived decision rules that minimize the probability of message error. Other perfor-
mance criteria were also considered to illustrate alternative parameter optimization. In the
presence of white Gaussian noise, the decision rules resulted in optimum receiver structures
that correlate the received vector with all possible transmitted vectors (with the addition of
the appropriate bias), and then decide on the message with the maximum correlation. In this
chapter, we consider the problem of waveform communications in which each message is
transmitted using a distinct voltage function. We start with a baseband signal model which
assumes perfect carrier phase coherency. We will first derive the optimum receiver structures
in the presence of white Gaussian noise and then evaluate the resulting performance for var-
ious sets of transmitted waveforms. The approach is then generalized to handle nonwhite
Gaussian noise, using various techniques, namely, the sampling approach, the Karhunen-
Loéve (K-L) expansion, a direct derivation using a sufficient statistic, and a “whitening”
operation on the observed waveform. Finally, inphase (I) and quadrature (Q) modulation and
demodulation are introduced to establish the link between the derived receiver structures and
communication signals. We will then relax our assumptions regarding the carrier phase co-
herency and the effects of imperfect carrier and symbol synchronization are investigated for
the simple model of a slow time-varying phase error process. Finally, the deleterious effects
of data asymmetry and suboptimum detection on system performance are treated.
Our approach to waveform receiver design [1] is to first approximate the waveforms
using finite samples to provide a quick but nevertheless correct derivation of the receiver
structures. The more rigorous approach, which does not require the finite sample approx-
imation, uses the K-L expansion. It will be shown that the results obtained are the same;
however, it is felt that time sampling is more appealing from a practical viewpoint. In the
current case, the transmitter maps each message m into a distinct voltage waveform s(t) for
transmission as shown in Fig. 4.1. It is assumed that each transmitted message lasts for T
seconds and that the system is perfectly synchronized in the sense that the receiver has full
knowledge of the start and end epoch of each signal. The receiver observes a sample func-
tion p(t), 0 <t <T, of the random process r(t), and maps the observation into one of the
messages in the transmitted set. Formulating the detection problem as a hypothesis testing
problem, then
 Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 179

MESSAGE |_™_,| waverorm | S00) a aae,

A

LOL) waverorm Le (t)

SOURCE [fm} [TRANSMITTER |F, (py RECEIVER
Figure 4.1 Waveform communications model

mo > Ho : r(t) =so(t) + n(t)

iy Hy : r(t) = sit) +n@)
(4.1)

my-1— Hy-i : r(t)=sm-i1(t) +n(t)

for the additive noise channel where s;(t) is a known deterministic time function, of duration
T seconds, used to transmit message m;,i =0,1,..., M — 1. It is also assumed that a full
statistical description of the additive noise process n(t) is available, that is, the joint pdfs
Fn(ty),n(t9),...,n(t) (M1, N2, ..., Nx) are known for any finite k and for any set {t), to, ..., ty}.

4.1 OPTIMUM WAVEFORM RECEIVERS IN WHITE GAUSSIAN NOISE

To gain insight into the waveform detection problem, we will first consider the case when
the channel noise is white and Gaussian. The received signal, assuming that message m; is
transmitted, is given by

rt)=s@Q)+n@), O<t<T (4.2)

where the Gaussian random process n(t) has a symmetric power spectral density No/2
Watts/Hz double-sided in the frequency range —B to B Hz and BT > 1. Now according
to the sampling theorem [2], n(t) is completely characterized by its samples taken every
1/2B seconds, that is,
CO

n(t)= ) > n(t,) sinc(2Bt — k) (4.3)

k=—oo

where sinc(x)A (sin x)/(mx) and t, 4 k/2B. Moreover, we shall have occasion not only
to represent time functions as samples but also reconvert from sums of samples to time
integrals. Thus, we note that

aen“(t)dt = —he n“ De (tk) (4.4)

represents the energy in the sample function n(t). Now the joint probability density function
of the vector n? = (n(t1), n(t2), ---,n(kT)) created from sampling a T-second strip of n(t)
is given by

fn(ty),--;n(kT)() = C exp |- INoB THT ma) (4.5)

as the 2BT samples are uncorrelated and Gaussian. Here, C represents a normalization con-
stant. Using (4.4) to convert from sums of samples to a time integral, we readily approximate
180 Coherent Communications with Waveforms Chap. 4

(4.5) via
ie n*(t)dt
|- i (4.6)
Fni(ty),--:n(kT)(M) ~ C exp

for the k’” time interval 7; = [(k — 1)T, kT]. Without loss in generality, we set k = 1 and
consider the time interval [0, 7]. Hence

_ fo Pat
2(t)dt
Fn(O),---.n(T)(M) = C exp N (4.7)
0

Assuming that message m; is transmitted in that interval, then r! = (r(t), r(t2),---,r(T))

and s! = (s;(t1), si(t2), ---, si(T)) and hence
fr(p/mi) = fn(p — si/mi) = fn(e — Si)
Z 2
t) — s;(t))“dt
= Corp | 200s (4.8)
No
This will be the basis from which the optimum receivers in white Gaussian noise will be
derived.

4.1.1 Binary Cross-Correlation Receivers

In the binary case, the receiver needs to decide between messages mg and m,. Using MAP
decision theory, the receiver sets m(p) = mo iff

P(mo/r
= p) > P(m/r=p) (4.9)
or equivalently iff
P(mo) fr(e/mo) s P(m1) fr(e/m1)
(4.10)
fr(p) fr(p)
Using (4.8) in (4.10) and assuming equilikely messages, the decision rule becomes: set
m(p(t)) = mo iff
ai T (p(t) — so(t))“dt
2
5 exp | 6T &Ot) =
— 510)
2
s,(t))*dt rey
No No
Taking natural logarithm and simplifying, the rule becomes: set m(o(t)) = mo iff
r T
i (p(t) — so(t))?dt < if(o(t) — si(t))?dt (4.12)
0
or equivalently, choose the message corresponding to the waveform s;(t) that is “closest” to
p(t) where distance is defined through

d (s(t), p(t)) & [ (o(t) — s(t))?dt (4.13)

This is referred to as minimum distance decoding and it was encountered earlier in vector
communications. Expanding (4.12), we have
T T iD T
i!so(t)dt ao p(t)so(t)dt <| si(t)dt -2f p(t)s\(t)dt (4.14)
. 0 0
 Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 181

DECISION

St . Figure 4.2 Binary waveform receiver

o(t)—s, (t) (equal energy, equiprobable signals)

DECISION

S ,;(t)

Figure 4.3 Alternative implementation of the binary waveform receiver (equal energy, equiprobable signals)

Letting E; denote the energy of the i’” signal, that is

if
E; ay s?(t)dt (4.15)
0
then the rule reduces to: set m(p(t)) = mo iff

L 1
aga [ P(t) (so(t) — s1(t)) dt > meee (4.16)
0
For equal energy signals, Eg = FE) and the test reduces to
MW mo
ag= | pte (sot) —s)) 2 0
0 my
(4.17)
Hence, the receiver correlates p(t) with the difference signal so(t) — s,(t) over (O, T) and
compares the outcome to zero as shown in Fig. 4.2. If the correlation is positive, the receiver
selects message m,. Otherwise, it selects mo. Alternatively, the receiver can be implemented
using two correlators as depicted in Fig. 4.3 and selects the signal corresponding to maxi-
mum correlation. Note that the decision rule operates linearly on the received signal and that
no other rule (linear or nonlinear) can produce a smaller error probability. The zero thresh-
old is due to the assumption made on the binary signal. If the signals are not equal energy
or equally probable, the threshold will be biased away from zero.

4.1.2 Matched Filter Receivers

To gain further insight into the structure of the optimum waveform receiver, let’s consider
the problem of detecting the presence of a known signal s(t) using the block diagram shown
182 Coherent Communications with Waveforms Chap. 4

r (t) y (t) y (T) DECISION

h(t) Z0
ea
Figure 4.4 Detection of the presence of a signal

in Fig. 4.4, where the received signal is given by

Oa sO ne) (4.18)
and s(t) is nonzero for te[0, 7] and n(t) is AWGN with two-sided PSD No/2 Watts/Hz. The
output y(t) of the linear time-invariant (LTI) filter h(t) is sampled at the end of the signal
duration, that is, at t = 7, and the resulting voltage compared to zero. The goal is to design
the filter h(t) such that the signal-to-noise ratio (SNR) of the sample y(7) is maximized,
where SNR is defined to be the ratio of the mean squared to the variance of the sample, that
is
E*{y(T
SNR = ne (4.19)
oO
y(T)
Using convolution, y(t) can be written as
[o.@)

y(t) =)! r(A)h(t — A)dar (4.20)

—0o

Since the noise is zero-mean, we have

CO

E{y(T)} =i S(A)A(T — A)dr (4.21)

—0o

Similarly, the variance becomes

> CO CO

Or) = (L / Rn(aAy — A2ACT — Ay)A(T — d2)dajar (4.22)

—oo J—00

Since the noise is white, R,(t) = (No/2)d(t) and (4.22) reduces to

No of
0) => / h? (Add (4.23)
—0Oo

Therefore, the SNR as given by (4.19) becomes

a -Less aval — rary’

ss
(No/2) [22 h2(a)da ee
4.24

or alternatively

(25, 5(7 — B)h(B)dB)”

= (4.25)
(No/2) [o> h2 (ada
Using Schwartz’s inequality [3], which states that
oxo) 2 ee) es)
(
/ hn (haat 2 i h*(t)dt / h3(t)dt (4.26)
—oo
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 183

with equality if and only if h(t) =ch,(t) (c being any constant), (4.25) can be upper
bounded by

SNR
PO
—
cal aie eo
nae de
: re) [Wada ae
or since E = PT = (e s*(t)dt, then
2 ® , 2E
SNR < — / s°(T —A)ddA = — =2(PG)(CNR) (4.28)
ING dies No
where PG denotes the processing gain of the matched filter and is given by PG=WT and
CNR = P/(NoW) where P is the signal power and W is the front-end one-sided bandwidth
of the receiver. Here, CNR denotes the signal-to-noise ratio in the front-end bandwidth W
where as E/No = PT /No denotes the signal-to-noise ratio in the data rate bandwidth 1/7.
The ratio of these two bandwidths is the processing gain PG of the matched filter. Exact
equality in (4.28) occurs only when (c any constant)

i CSC) (4.29)
Therefore, SNR is maximized when the filter h(t) is matched to the incoming signal s(t),
thus the terminology matched filter. Of course, such a filter with a finite impulse response is
not physically realizable. However, as a practical matter, the filter becomes both physically
realizable and essentially optimum whenever almost all of the energy in the signal s(t) is
located within 0 < t < T and —B < f < B where B is the one-sided baseband bandwidth
of the receiver and is related to the IF bandwidth of the receiver through W = 2B. When
implementing a matched filter, the output y(t) can be expressed as

y(t) = / r(A)A(t — A)daA

a CO
(4.30)
=i) r(A)s(T —t +A)da
—0o

When sampled at t = 7, then

(o@)

y(T) = ik r(A)s(A)da (4.31)

=o)

which is exactly the correlation between the incoming waveform r(t) and the signal s(t).
From Fig. 4.3, the optimum binary waveform receiver can alternatively be implemented us-
ing matched filters as shown in Fig. 4.5. In this case, two filters matched to the possible
incoming waveforms are used to filter the observed signal and their outputs sampled appro-
priately at t = T to provide maximum SNR. The resulting voltages are fed into a comparator
for selection of maximum value. Note that the signal component of the matched filter output
is given by
(oe)

E[y(t)] | s(A)s(T —t +A)dr (4.32)

—0o

Recalling that the time-autocorrelation function [4] of any real function g(r) is given by

Rec) 4 f g(t)g(t + t)dt (4.33)

184 Coherent Communications with Waveforms Chap. 4

Select
message Decision
corresponding
to largest
cross-correlation

Figure 4.5 Matched filter implementation of the binary waveform receiver of Figure 4.3

Hence

Ely@l=k =) (4.34)

and at the sampling instant t = T, the mean becomes

Ely(T)] = Rs) = E (4.35)

which is the maximum value of R;(t). Thus, the time-correlation operation can be equiva-
lently performed using a matched filter. Note that when the output of the matched filter is
sampled at t = T + 8 due toa timing error 6, then E{y(T + d}= Rs (5) and the signal-to-
noise ratio is degraded accordingly.

4.1.3 M-ary Waveform Receivers

In the case of M-ary signals, the optimum receiver sets m(p(t)) = m; iff

Pera ae T OD) D
t) — sj(t))“dt
No
é (4.36)
canek {Pemex eo rare
9)
t) — sg(t))“dt
No
Simplifying and taking the logarithm, the receiver reduces to maximizing

i Me fish
In P(m;) — — / s?(t)dt + p(t)s;(t)dt (4.37)
No Jo No Jo
or equivalently, the optimum receiver sets m(o(t)) = m; iff
i T
qi ah p(t)sj(t)dt + B; es {| P(t)s(t)dt + a,| (4.38)
0 0
where the bias term By, is given by
N E
BA = In P(m,) — a (4.39)
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 185

Set m (p(t) = m,
iff

q;La = max{q,}
ieeo

Figure 4.6 Structure of the M-ary waveform receiver

where E;, is the signal energy given in (4.15). In the case of equally likely messages, the
bias reduces to — E;/2. In addition, if the signals had equal energy, the bias term can be ig-
nored. Note that the bias term is equivalent to that given in (3.60) when 0” = No/2. Hence,
the receiver should correlate the observed signal with all possible transmitted waveforms
(add the appropriate bias for unequal energy or unequal a priori probabilities) and select
that message corresponding to maximum correlation as shown in Fig. 4.6. This implemen-
tation requires M correlators and hence, the receiver complexity increases linearly with the
number of transmitted signals. If the signals were designed such that (see Appendix 3C)
N
HO=) sy)(0) (4.40)
= i)

then
; ig T N
[ ewsmar= [oY syojtoat
0 a
: (4.41)
N V1 N
=Yisj f pwmajnar =) sur
‘a1 0 a
186 Coherent Communications with Waveforms Chap. 4

SIGNAL
WEIGHTING
MATRIX

202 406 i0N

A

E ioaeeeeriN Set m(p(t)) =m,

if m (p (t))
q; palets {q,}
SM-11 SM-12esM-IN |

Figure 4.7 Alternative implementation of the M-ary waveform receiver using N correlators

Hence, the correlation can be achieved by correlating p(t) with the N orthogonal functions
and then weighting the output with the signal coefficients to form the required signal cor-
relation. This processing is depicted in Fig. 4.7 and requires only N correlators for any
signal set. When the number of signals M is greater than the dimensionality N, the sec-
ond configuration requires fewer correlators and is more suitable for implementation. Since
the correlation operation can also be implemented using a filter and a sampler, the opti-
mum receiver can alternatively be implemented using either M or N matched filters (with
the appropriate weighting) as illustrated in Fig. 4.8. Note that when the signal waveforms
satisfy (4.40), we can map the waveform problem into an equivalent vector problem where
s! = (si1, Si2,+-+, Sin), i =0, 1,+++, M — 1 and use all the tools developed in Chapter 3.

4.1.4 Error Probability Performance

In this section, the performances of the various coherent receivers are derived for both
binary and M-ary signaling. First, binary signals are considered to gain insight into the
problem [5], and then the results are generalized to the M-ary case. Equiprobable signals are
considered as they reflect most communication systems. But the results can also incorporate
unequiprobable signals with the appropriate weights.

Binary waveform receiver performance. For equiprobable, equal energy sig-

nals, the optimum receiver computes Ag = ia p(t)(so(t) — s1(t))dt and compares the re-
sulting random variable to zero. Then, the conditional mean and variance of Ag become
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 187

S02 eeeSon

S11 $12 eoeoSin

m (p (1)

(b) N MATCHED FILTERS

Figure 4.8 Implementation of the M-ary waveform receiver using matched filters

T
EtAg/mi= | si(t) (so(t) — s1(t)) dt (4.42)

and
T 1
Ka — a {(f n(t)(so(t) — su()ar (| n(u)(so(u) — sitw)du) |
0)
(4.43)
N L.

= ee ( (so(t) — s1()ar
2 0
Because of the equal energy assumption, ig sp(t)dt = Us sa(t)dt = E and (4.43) reduces
to
r
OKa — Nf {ze
— / suesottnae} (4.44)
0

If we denote

Vij & dt (4.45)

0 ‘ ue
VEi JE;
as the correlation coefficient between signal s;(t) and s(t), then

Ago & E{Aq/mo] = E(1 — vio) (4.46)

188 Coherent Communications with Waveforms Chap. 4

Aq & E[Aq/m1] = E(y0 — Y) (4.47)

and

oi, =NoEU—yio) - (4.48)

Note that
/ 1/2
| =|[22
ae a Z eal ay of =1 (4.49)
ui Dido EB 0 \VEj
where the upper bound is obtained using Schwartz’s inequality of (4.26). Due to the equal
probable, equal energy assumption, the probability of error, conditioned on mo, P(E/mo),
is identical to P(E/m},). As a result, the unconditional probability of error, P(E), for
equilikely signals is given by

0 1 (Aq — Bqo)”
Pe) =f = Op d(Aq) (4.50)
Fee [ 2OK, 20Kg

which, using the substitution x = (Ag — Ago)/oaq, can be simplified to

P(E)=0 (=) (4.51)

Substituting (4.47) and (4.48) into (4.51), we have

E
P(E)=Q (\
Na ea vo) (4.52)

where E’/ No denotes the signal energy-to-noise ratio. Alternatively, (4.52) can be expressed
as
d
P(E)=@Q (=) (4.53)

where d;; denotes the distance between signals s;(¢) and s;(t) and is given by
ij
dj, A i (si (¢) <= sj(t))° dt
0 (4.54)
= Ej; + Ej —2/ EE jvij
For the two binary signals with equal energy, din = 2E(1 — yo) and since the most negative
value 19 can take is minus one, the corresponding P(E) is then

[2E
P(E ea (az) Antipodal (BPSK) signals (4.55)
)

which occurs when so(t) = —s}(t) and the pair is referred to as antipodal signals. Note that
any pair of antipodal signals will generate the same P(E) for a given energy E. Examples
of three pairs of antipodal signals are given in Fig. 4.9, namely: (a) pulse code modulation
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 189

So (t) n S
—_ = oo

VET

(a) PULSE CODE MODULATED (PCM) SIGNALS

So (t)

VE/T
VE/T
T
t t
ali

- JER Vert
(b) MANCHESTER OR BI-o SIGNALS

Sq (t)
S(t)
/J2E/T J2E/T
+
a ~

(c) BINARY PHASE SHIFT KEYED (BPSK) SIGNALS

Figure 4.9 Examples of antipodal signals: (a) Pulse code modulated (PCM) signals, (b)
Manchester or Bi-¢ signals, (c) Binary phase-shift-keyed (BPSK) signals

(PCM), (b) coded PCM typically referred to as Manchester or Bi-¢, and (c) binary phase-
shift-keyed (BPSK) signals. The various signals are given by
PCM:

so) = -s= |0SEIT Dele

otherwise
(4.56)
Bi-@:
A Ey OS had
sot) =—s1(()= 4 —/E/T T/2<t<T (4.57)
0 otherwise
BPSK:

so(t) = —sj(t) = J2E/T sin(o@et +0) O<t<T, f,.T = integer (4.58)

The latter is referred to as a phase shift keyed signal since so(t) = —s\(t) = /2E/T sin(wct
+6 + 7) and hence, so(t), s(t) are transmitted by assigning the phase of the signal to 0 or
+z, respectively. Since the phase is changing between two defined states, this scheme is
referred to as binary phase-shift-keying (BPSK). Later, we will consider M-PSK signals
where the phase is varied among M states, equally spaced around the unit circle.
190 Coherent Communications with Waveforms Chap. 4

Figure 4.10 Correlation coefficient of

PHA aoe ¢
1,0 coherent frequency shift signals as a function
3r of 22 fioT

For orthogonal signals, ie sy(t)so(t)dt = 0 (that is, yi9 = 0) and P(E) of (4.52)
reduces to

E :
P(E) =O (4 Orthogonal signals (4.59)
\ 0

which is a factor of two (or 3 dB) worse in E’/ No than the performance of antipodal signals.
Another signaling scheme is coherent frequency-shift-keying (FSK) where the messages are
transmitted using two distinct frequencies, that is

DB
Sit yaa) cp sin(@jt +0), O<t<T (4.60)

I2E 3
So(t) = ap eae) OST (4.61)

and wT, woT > 1. In this case, the correlation coefficient is given by

1 Ve
sin(@, — wo)T e sin 27 f\, oT
Y10 = =f si (t)so(t)dt = = 4.62
E Jo (@| — wo)T 2 fiol ( )

where the correlation is plotted in Fig. 4.10 as a function of the frequency separation f|.9 =
fi — fo. The minimum possible correlation is given by yj9 = —<4, resulting in

E 2
ECE) =O —j{1+— Coherent BFSK (4.63)
No 3

which is 2/(1 + ys) = 2.1 dB worse in SNR than antipodal signals. The corresponding fre-
quency separation is (f| — fo)T = f\,0T = 3/4. On the other hand, when f;,97 = 0.5, that
is , the frequency separation is equal to half the message rate, orthogonal binary signaling is
obtained and coherent FSK with this minimum frequency separation is sometimes referred
to as minimum shift keying (MSK).
In all cases so far, we have considered equiprobable, equal energy signals. For unequal
energy and unequiprobable signals, it can be shown (see Problem 4.10) that
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 191

d\0 LANG RP. On4y)

P(E) =P a
() inno( qo Voee 7)
(4.64)
: dio ] No P(m))
+ P(m\)Q ++ 4 In
20%) d10.% 2°." Po)
For the special case of equiprobable, unequal energy signals, (4.64) simplifies to

P(E) = dio
(E) o(=) (4.65)
4.65

as expected. An example of a scheme with unequal powers is “On-Off” keying where so(t)
is sent by turning the power off and s;(t) is sent by turning the power on, that is

Oana Cera t (4.66)

otherwise

and so(t) = 0 for 0 <t < T. In this case, yj9 = 0 since so(t) = 0 and de =F Hence

P(E)=0Q (Fz) On-Off keying (4.67)

Note that the maximum energy delivered is E, but the average energy is E/2 since both
signals are equally likely. Hence, (4.67) can be expressed in terms of the average energy
Eayg as

Eavg
P(E)=O ie (4.68)

which is identical to (4.59).

M-ary waveform receiver performance. The performance of an M-ary MAP

waveform receiver is first derived for any signal set of equally probable, equal energy
signals. Then, the results are applied to signals of great interest, namely orthogonal (_L),
biorthogonal (Bi-L), transorthogonal, L-orthogonal, multiple-phase-shift-keyed (M-PSK)
and finally, multiple amplitude modulated signals (M-AM).
Referring to Fig. 4.6, which depicts the structure of the optimum M-ary waveform
receiver in an AWGN channel, the output of each integrate-and-dump filter is denoted
by gq; and the receiver selects the message or signal corresponding to the largest q;. At
this point, we need to differentiate between symbols and bits as both will be used in this
section. In binary communications (M = 2), each message or signal is transmitting one
binary bit of information, that is, a binary “O” or a binary “1.” Hence, the signal duration
T and energy F are the bit duration and bit energy, denoted more appropriately by 7)
and Ep, respectively. In M-ary communications (with M = 2”), each signal is transmitting
n = log, M binary bits. We will thus refer to each message or signal as a symbol and to
the duration and energy of the signal as symbol duration 7; and symbol energy Ey. Since
in every 7; sec, logy M bits are transmitted, (7;/7p) = (Es/Ep) = log, M. We will also
differentiate between symbol and bit error probabilities, denoted respectively by P;(£) and
P,(E). The bit error probability P,(E) will be used throughout this section to compare
192 Coherent Communications with Waveforms Chap. 4

the performances of binary and M-ary communication systems. The probability of message
(symbol) error, P;(E), can in general be expressed as
M-1
P,(E) = )) Ps(E/mi)P (mi) (4.69)
i=0
where P,(E/m;) is the probability of message or symbol error assuming that m; is transmit-
ted. In order to evaluate P,(E/m;), the outputs of the integrate-and-dump filters have to be
jointly characterized through the M-dimensional pdf. Recalling that

qk = [™ pidsaltyat (4.70)
where p(t) = s;(t) + n(t) when m; is transmitted, the output of the je? integrator becomes

guj = Esai + [* n@se(ddt (4.71)

with conditional mean

Gk/i & Elgxsi] = Es vei (4.72)

Similarly, any pair of outputs g, and gj; will have conditional covariance (that is, subtract
the means)
Ts Ts N Ts Ni
aj=E| | nooa(tdt f n(u)s,(wau} -F{ sx (O)5(t)dt = > Esty (4.73)
Hence, the outputs go, q1,.-.,Gm—1, conditioned on m;, are Gaussian random variables
with nonzero means and nonzero correlations. Their joint M-dimensional pdf is given by
(3.61) as

fa(a/m:) = G (q; Gj; Kq)

1 l 2x: <li eting . (4.74)
= exp |-5 (a — q;) a. (a = 4:
(27) M/2idet Kgl? : a ae )
where q = (40, 91, ---,9m—1)/ ,G; denotes the conditional mean vector given by

qi = GO0/i, Ti, ---,@M—11) = Es(yois Nis --- Ym—1i)” (4.75)

and Kg denotes the covariance matrix given by

900 901 ia q0M-1

710 ql oa 1M—
Kq = 3 ; 2 4 is

QM-10 4M-11 °** YM-1M-1

(4.76)
yoo YOl ae YOM—1
7 NoEs Y10 Yul BARD YiM-1

reas eerie 7
YM-10 YM-11 °** YM-1M-1
Equation (4.74) assumes that Kgq is invertible. If Kg is singular, the dimensionality of the
observed vector can be reduced without any loss in generality and can result in singular
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 193

detection. The conditional probability of symbol error P;(E/m;) = 1 — Ps(C/mj;), where

P;(C/m;) is the probability that all g,’s, except k = i, are less than q;, that is

Ps(C/m;) = Prob {gi > (Go, 91,°*+s Gi-1, Git1s°**sQM_—1)/mi} (4.77)

co Gi Gi
=f if eas fq(a/mi)dqodq :- -dqi—1dgi+1 ---dqu—idqi
—CO vd —C —O0
ee
M-—| integrals

Note the order of integration where the last integral is over gj. Since P(C/m;) depends on
yji, J =0,1,...,M —1, the resulting probability of symbol error, P,(E), depends solely
on the symbol energy E,, the noise spectral level No, and the signal correlation matrix T'!
given by

¥00 Y01 ve YOM-1

v10 VAT MU) YOM
C= ; : : : (4.78)

YM-10 YM-11 ‘** YM-1M-1

If we normalize the g;’s by introducing the change of variables

Zi = (Gi — Gi)
/VSGi (4.79)
then (4.77) may be rewritten as

Co rzt,/2Es/NoU—v%1) Zita/2Es/No(l-yim-1)
Pe(C/m) = f if af fi(z/mj)dz (4.80)
—oo J—00 —0oO

where z! = (zo, z1, Z2,-*:,ZM—1) and f,(z/m;) = G(z; 0;T), independent of m;. Using a
similar procedure as in Chapter 3, a Union bound can be derived for waveform communica-
tions resulting in

M-1
ie
P(E/m) < o(5) (4.81)
h »» J2No
i#k

where d;; is the distance given by (4.54).

The bit error probability is what a design engineer needs in a link budget. In order to
compare binary and M-ary communications, we will therefore use the bit error probability
for both signaling schemes as a measure. In binary communications, each signal transmits
one binary bit and hence, the probability of signal error is identical to the probability of bit
error. In M-ary communications, let M = 2” so that each signal or symbol is transmitting
log, M =n binary bits. Let’s consider the transmission of a sequence of K bits where K is
very large. When employing M-ary signaling, we can write K = hn where h is the number
of symbols and n is the number of bits per symbol. The probability of bit error P,(E) can
be written as

; a. NoEs
1. The covariance matrix Kg is given by Kg = 21.
194 Coherent Communications with Waveforms Chap. 4

Average number (#) of erroneous bits

P(E) = Total # of bits
(Avg. # of erroneous bits per symbol)(# of symbols)
(4.82)
< Total # of bits
(Avg. # of erroneous bits per symbol)h
Fi hlog, M
which is independent of / and hence of K. But
M-1
(Avg. # of erroneous bits per symbol) = Dee (Avg. # of erroneous bits/m;) P(m;)

a) Mo!
(4.83)
= ‘Ss (Avg. # of erroneous bits/m;)
i=0
The last equality is due to the fact that we are transmitting a “large” number of symbols and
as a result, all messages occur equally likely. Hence
1 M-l
»(E)
P,(E) Taipei
= —— Avg. g # of erroneous bits/m; /m) (4.84)
4.84

Now let n;, ; denote the event that message i is transmitted but message j is decided (that is,
vector $; is closest to observed vector p) and (i, j) the number of bits that symbol i differs
from symbol /; then
M-1
(Avg. # of erroneous bits given symbol 7) = SC Pn, jG J) (4.85)
j=0
Hence
1 M-1M-1
P(E) == Ppp GD) (4.86)
MGR ME =o j=0
The probability of bit error, P,(£), depends on the symbol-to-bit mapping, otherwise
known as the code. The computation of P(nj;,;) is very difficult if not impossible for a
general M and signal constellation. It is therefore desirable to obtain an upper bound on
P,(E). Let §;,; denote the event that message i is transmitted but signal j is closer to the
received vector, ignoring all other signals. The probability of the event &; ; is referred to as
the pairwise probability of error (PPE) and it will be used extensively in future chapters
to bound the probability of bit error. Figure 4.11 illustrates the regions where no,1 and & |
occur for an example of four signals. In all cases, the region where the event &;,; is true is
greater than the corresponding region where nj; ; is true. As a result

P(,j;) = P(ni,j) for alli andj (4.87)

and hence, (4.86) can be upper bounded by
1 M-1M-1
Bp Mion M D 2
ee ODP
Gi, J)PE. j) CL (4.88)
4.88

Equations (4.86) and (4.88) are exact and upper bound expressions for the bit error probabil-
ity for any signaling format and any code (that is, symbol-to-bit mapping). Both expressions
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 195

No.1

IN no,» IS CLOSER TO IN€o 1,p 1S CLOSER TO s,,

s,;, THAN ANY OTHER THAN $9 BUT $5 MIGHT BE
SIGNAL THE CLOSEST TO p (CONSIDER (X))
“. S$; IS THE CLOSEST TO p

Figure 4.11 Example of 7o,; and &,|

can be further simplified by noting that (7, 7) = (j,i) for all i and 7. However, the events
ni,j and n;,; are not always identical and depend on the signal set of interest.
Example 4.1
Consider the four signals illustrated in Fig. 4.12(a) with the corresponding bit mapping. This set
corresponds to quadrature phase-shift-keyed (QPSK) signals, which will be discussed in more
detail when considering M-PSK signals. Using (4.88), we have (M = 4)

1 e) 5
P,(E)
»(E) <<———~
Dew 22! i, j)P(&.,j)
j)P (Ei,j

6 {(0, 1) P(&0,1) + (0, 2) P (0,2) + (0, 3) P (0,3) + (1, 0) P10)

+ (1, 2)P 1,2) + (1, 3)P (1,3) + (2, 0) P 2,0) + (2, IP (&2,1)
+ (2, 3)P(&2,3) + (3, 1) P(E3,1) + G, 2)PE3,2) + GB, 0)P(E3,0) }
But

(O,.1) = 0) 3) = (1,0) = (1, 2) = (21D = @, 3) = G6, 0) =G, 2) =1

and

(0, 2) = (2, 0) = (1,.3) = G, 1) =2

Also from Fig. 4.12(b) and (c), it is clear that

Es
P (0,1) = P (0,3) = P(&1,0) = P (1,2) = P (2,1) = P(&2,3) = P(&3,2) = P(§30) = Q (‘|

and

2E;
P (0,2) = P(&1,3) = P (2,0) = P(&3,1) = @ ( No
196 Coherent Communications with Waveforms Chap. 4

S|

So

$3

(b) (c)
Figure 4.12 Signals and bit mapping of Example 4.1

Hence

E, 2E,
nie) =0(/%) +0 |

and E,/No =2E,/No. Note that P,(E) can be made arbitrarily small by making E,/No larger
and larger. This is not the case in binary or bit-by-bit signaling where the probability of a bit
error occurring approaches unity for any E,/No when transmitting many bits. To illustrate
that, consider the probability of a bit error in a single transmission [that is, Q(,/2E,/No)] as-
suming antipodal signals. Then, the probability of a bit error in K transmissions is 1 — [1 —
Q(./2E,/No)|* and approaches unity as K — oo (for a fixed E,/No). Note that this compar-
ison is between block transmission as the block length increases and bit-by-bit transmission as
the number of bits grows. The key advantage is obtained by varying the block length in one case
versus a fixed block (or bit) in the other.

Orthogonal Signals. Orthogonal signals are characterized by the fact that any two
distinct signals have zero correlation, that is

n= {9 ee
OMe
(4.89)
Hence, their signal correlation matrix [9 is given by
PO mes
On uk. siOe cee
To=]|. ae . | =Im (M x M identity matrix) (4.90)
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 197

In this case, the integrate-and-dump filter outputs go, g1, ---, gy—1 are Gaussian, i.i.d., with

Okey
qk/i = Bathe (4.91)

and variance o” = NoE;/2. From (4.77), we have

M-1
ie.) Gi 1 qe
P;(E/m;)
5 ( / i) =1— ae if ———
pea) exp
o( |-—£
#) }da

(4.92)
x 1 exp _ Gi ra Eeye ae

V 2102 ee
Since P;(E/m;) is independent of hypothesis i, the unconditional probability of symbol
error P;(E) becomes (using Appendix 3B)

pEy=1—
sh) = [- enn
ae NG (pp eed A i ale
ee 4.93
(4.93)

and is depicted in Fig. 4.13 as a function of E;/No for M = 2, 4, 8, and 16. The performance
degrades for increasing M because even though each M-ary system is transmitting at the
same symbol rate R, (symbols/sec), the corresponding bit rate 72, (bits/sec) is changing
and is given by
|
Rie a log, M (bits/sec) (4.94)
AY

Hence, as M increases, the bit rate also increases. In order to compare all M-ary systems at
the same bit rate, we need to assess P;(E£) as a function of bit SNR, Ey/No, given by
Ep = Ee 1
(4.95)
No Nolog,M
Figure 4.14 depicts P;(E) versus E,/No when M = 2”. It is clear that the bit rate adjust-
ment leads to an improvement in performance as M is increased.
To convert from symbol to bit errors, note that with orthogonal signals, the incorrectly
decided symbol is equally likely to be any of the remaining M — | symbols since each
symbol (equivalent signal vector) is equidistant from all others. Consider Fig. 4.15, which
depicts the bits for eight symbols (that is, M = 2? = 8). In any column of bits, there are
exactly 2”—! ones and 2"! zeros (four of each in Fig. 4.15). Thus, if a symbol in incorrectly
decided, then for any given bit position (or column), there are 2"! out of a possible 2” — 1
ways in which that bit can be in error. Thus, the probability of bit error, P,(£), is given by
n—|
P(E) = Ps(E) (4.96)
2" — |
and is plotted in Fig. 4.16 as a function of E,/No for several values of n. Using the union
bound of (4.81), P;(£) can be upper bounded by

P(E) =(M —1)Q (i) (4.97)

which is easily computable.

 198 Coherent Communications with Waveforms Chap. 4

iO.

10:7

Le
a”

10°

104

Tore Figure 4.13 Symbol error probability

a 0 # 8 12 16 db for M orthogonal signals as a function of
Es/No, 08 symbol SNR (Reprinted from [1])

Example 4.2
A typical set of signals that may be generated in practice for M-ary transmissions is the set of
M nonoverlapping sinusoidal pulses

a) = ;
V2PM sinact, iT,
FF oe< G+dT
i
, otherwise

This set is orthogonal and is useful in Pulse Position Modulation (PPM) systems. Another
orthogonal set of signals consists of sinusoidal pulses of different frequencies

s(t) = V2P sin(@et + =), esi SI

where the frequency separation is 1/27 Hz to ensure orthogonality. Such a set is referred to as
multiple frequency-shift-keyed (M-FSK) signals.

Biorthogonal Signals. Biorthogonal (Bi-L) signals can be constructed from or-

thogonal signals by augmenting the set with the negative of the signals. Let’s start with 4
orthogonal signals, denoted by so(t), si(t),---,8
"Sm _j(t). Adding their negatives to the set
we have SM (t)= —so(t), Sm, \() = yi oes) Mea) = —Sm_ (0). Hence, the signal
correlation ‘coefficients are given by
 Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 199

n i}
We
i]

n i
n=
n

n
n CORN
Ss)
Gn)
(GO):
NI)
00
mee}
Sh—-
nm
©

Figure 4.14 Symbol error probability for

M orthogonal signals (M = 2”) as a function
of bit SNR (Reprinted from [12])

SIGNAL PPIN

Mo _ 0 0 0— CORRECT SYMBOL
m, 00
mM, 010

Ms O17 | EQUALLY LIKELY M-1

m, 100 INCORRECT SYMBOLS
m 14001
5
Me V0
mz 1 |1

4 OUT OF 7 WAYS
THIS BIT CAN BE Figure 4.15 Symbol-to-bit mapping for
IN ERROR orthogonal signals (M = 8)

0 i#j,\i-sjl4#4
Wij 1 | i. (4.98)

-1 ji-jl=%
and as a result
Im -—Im
Ip = ( 2 2 ) (4.99)
200 Coherent Communications with Waveforms Chap. 4

| et

n=1
n=2

10"! n=3
n=4

n=5

n=6

10? ew
n= 8

n=9
x n=10
n= 15
192 n=20

104

-5
1¢ i 15 5 Figure 4.16 Bit error probability for M
orthogonal signals (M = 2”) as a function of
E., /No, dB bit SNR (Reprinted from [12])

where Zm is the u x a identity matrix. Since the signals and their negatives are in the
2
signal set, the receiver can be simplified and implemented using ¥ correlators as shown
in Fig. 4.17. In the decision rule, the signal is first chosen to correspond to the maximum
correlation in absolute value. Then, the sign of the signal is chosen to correspond to the
sign of the correlation. The joint pdf of go, g1,°--,4 M_\ is the multivariate Gaussian pdf of
(4.74) with an identity covariance matrix corresponding to the set of u orthogonal signals.
As a result, the probability of correct symbol detection, when m; is transmitted, is the
probability that |g;| > |qx| for all k Ai andi =0, 1,---, 4 — | and has the correct sign
(assumed positive without any loss in generality), that is

P;(C/m;) = Prob {4 > 0, |gil > (gol, --+, lgi-il, l@i+il, +: lay _,D/mil
fe i Ch

~ Jo —$
pert (4.100)
“_1 integrals
x fq(a/mi)dqodqi ---dqi-1dqi+1-: dqu_\dqi

Since the qj’s, i =0,1,---, 4 — 1 are i.id., then (with change of variables)
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 201

DECISION
DEVICE

. SELECTs; (t)
WHERE
a= Gn
2. SIGN OF s| (t)
CORRESPONDS
TO SGN (qi)

(4.101)

ogg? ; Es
p(E)=1— f ae —erfe (q+ [7 dq (4.102)

The symbol error probability, P;(£), is plotted in Fig. 4.18 as a function of bit SNR, E,/No.
Since for any choice of signals, one of the remaining M — 1 is antipodal (distance = 2./ E;)
 202 Coherent Communications with Waveforms Chap. 4

fel

pee
(PSK)
P(E)
iT]

iT]

Wu
now

er
eh
la)
e)
eb
Sle!
el
SiS
= Moe
DY
=|
FW
ON
ON
+0ono

a 15 20
Figure 4.18 Symbol error probability
for M biorthogonal signals (M = 2”) as a
ae be function of bit SNR (Reprinted from [12])

while the rest are orthogonal (distance = /2F;), P;(E£) can be bounded using (4.81) to
produce

P,(E) < (M —2)Q BS +Q abs (4.103)

s(E) S( No No ;

In order to compute the bit error probability, there are two types of symbol errors that
can occur—when the negative —s;(t) of the transmitted signal s;(t) is selected or when any
one of the M — 2 remaining orthogonal signals is decided. We will refer to the former event
as errors of the first type and to the latter event as errors of the second type. The probability
of symbol error of the first type is given by

Pyi(E) = Prob {qi<0, |gil > (Idols *++5 lgi-tls laiils+s la _D/mi}
=f ak
iL abst
iM (4.104)
wal
Honneeats

x fq(a/mi)dqodq ---dqi-\dqi+1-: dqu_\dqi

 Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 203

IGNA BIT MAPPING

pegtt) 000
S,(t) 001
S,(t) 010
S,(t) 011
s(t) =-So(t) ea
S_(t)=-s;(t) al 0
S,(t)=-so(t) 101
S5(t)=-s3(t) 100

3 OUT OF 6 WAYS THIS BIT CAN BE

IN ERROR (ASSUMING A SYMBOL Figure 4.19 Complementary bit encoding
ERROR OF THE SECOND TYPE) for M = 8 biorthogonal signals

Substituting for the joint pdf and simplifying, P,;(E) becomes

[fee EB; ie
Pa(E) = | ed —~1+erfc(q + Re dq (4.105)

The probability of error of the second type, P;2(E), is greater than P,,(E) since the or-
thogonal signals are “closer” to the transmitted signal than its negative version. Since both
error events are disjoint and form all possibilities of symbol errors, then P;;(E) + P,2(E) =
P;(E) and as a result

Poo(E) = Ps(E) — Psi (E) (4.106)

In order to proceed ahead and compute the probability of bit error, we need to assume a
symbol to bit mapping. Since the distance between a signal and its negative is the maximum
it can be, we will encode a symbol and its negative with complementary bit patterns as
shown in Fig. 4.19. This will ensure that the event when all bits are decided in error occurs
with the least possible probability, that is, a symbol error of the first type. Hence, if a
symbol error of the first type is made, all bits are in error and the conditional bit error
probability is exactly one. For a symbol error probability of the second type, in any bit
position (or column) of the remaining 2” — 2 possible incorrect symbols, there are an equal
number, 2”~! — 1, of ones and zeros. Thus, the probability of a bit being in error is one-
half given that a symbol error of the second type occurred. Thus, the bit error probability for
biorthogonal signals is given by

l
+E) Sete Lene) (4.107)

and is plotted in Fig. 4.20 versus bit SNR, Ep/No.

Transorthogonal Signals. Transorthogonal signals are equally correlated signals

with the most negative correlation that can be achieved with M signals. Recall that in bi-
nary waveform communications, antipodal signals, with minus one signal correlation, offer
the best performance. The motivation for transorthogonal signals is that their performance
would be superior compared with orthogonal signals, because they are constructed with the
 204 Coherent Communications with Waveforms Chap. 4

P(E)

Figure 4.20 Bit error probability for M

20 biorthogonal signals (M = 2”) as a function
Ey INo dB of bit SNR (Reprinted from [12])

largest achievable negative correlation [6]. The value of the signal correlation is limited as
follows:
M-1M-1
Ts M-\ Ts M—l Ts
| OD s(nyar = f a soars [ De DE silts) (hat (4.108)
0 i=0 0 j=0 ee | ay
= ME, +M(M — lEsy
where y is the correlation between any pair of distinct signals. Since Io Saran S (t))?dt >
0 must be true for any signal set, then

ME,+M(M —- 1)Esy =0 (4.109)

or

es
> — =
(4.110)
This is the maximum negative correlation achievable with M equally correlated signals. If
we choose a set with y = —1/(M — 1), the signal set is called a simplex signal set with
corresponding signal correlation matrix
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 205

Mal ie mah anat

ph ee ee (4.111)

en ie eT
One method to construct simplex signals is to start with M orthogonal signals, s;(t),
i =0, 1,---, M —1, each with energy E,, and translate each signal in the set by the average
of the signals, that is, let

, Mel
a(t) A 2 si(t) (4.112)

Then, form the new signals

50) =s0) =a@) 1=0,1,-:-,M—1 (4.113)

The signal energy of the new set is given by

Ts Ts Ts Ts
E\ = / (s}(t))?dt = i s?(t)dt + if a*(t)dt —2 / si(t)a(t)dt
0 0 0 0
, . ; (4.114)
=E;+ yn Es) roy we =E, (1= x)

and the new signal normalized correlation coefficients become

1 Ts 1 ie 1 M-1 1 M-1
Vii = ey if Oe / /
Oi ! (si) — dXSKS) — 7 »Dsi(t))dt
- a (4.115)
1 Es Es ME, ] : . .
=—— |0 +, |= ——-,._ independent
of jandj
E. M M M M-1

Hence, the new signal set achieves the minimum correlation possible for M signals and the
correlation is independent of the signal pair. The simplex signals for M = 2,3 and 4 are
depicted in Fig. 4.21. When M = 2, the simplex set reduces to two antipodal signals as
expected. Since the simplex set is obtained from an orthogonal set by translating the signals,
the symbol error rate remains the same as it is insensitive to signal translation or rotation in
AWGN channels. Using (4.93) and realizing that the signal energy is scaled by the factor
dad - a) then

M-\ 2
Fo sk Es M e?

where we let E, denote the energy of the simplex signal (that is, replaced E{ by E;). Since
all the signals are equidistant from each other, a symbol error is equally likely to be in favor
of any incorrect symbol and as a result, the bit error probability is related to the symbol error
probability in the same way it was for orthogonal signals. The Union bound for simplex
206 Coherent Communications with Waveforms Chap. 4

a) M=2

(b) M =3:
(EQUILATERAL TRIANGLE)

®o

92

c) M=4
(REGULAR TETRAHEDRON)

Figure 4.21 Simplex signals M = 2, 3, and 4 (Reprinted from [1])

signals is identical to that of orthogonal signals of (4.97) except for the adjustment in energy,

Multiple Phase-Shift-Keyed (M-PSK) Signals. This signal set corresponds to M

equally spaced vectors around the unit circle. The elements of the covariance matrix that
define this set of signals are given by

l=]

so that
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 207

1 cos(4%7) +++ cos (Cay)

20 : :
ioe cos( =
Gi? I 4 :
(4.119)
: os cos(4%
7)
cos (24-5) ASE ee 1
For M = 2, the signal set reduces to antipodal signals, while for M = 4, we obtain a
biorthogonal signal set. Notice that this matrix is singular since the sum of elements in every
column is zero; hence, the matrix has zero determinant.
For multiple phase-shift-keying (M-PSK), a convenient set of signals that may be
generated is the polyphase signal set

SO=V2P sin (Og +6)), Ost <7, (4.120)

where , is the carrier frequency and 6; = 27i/M fori =0,1,---, M — 1. Notice that for
M = 2, we get antipodal signals or binary-phase-shift-keying (BPSK), the case M = 4 pro-
duces quadriphase-shift-keying (QPSK), and M = 8 yields octaphase-shift-keying (8PSK).
The three signal sets are depicted in Fig. 4.22, along with the decision regions of an opti-
mum receiver (which employs minimum distance decoding).
Since the matrix [y—psx is singular, the dimensionality of the integration in (4.77)
must be reduced in order to evaluate P,(E). A simpler method is to transform the problem
into polar coordinates by representing the i‘” integrator output in terms of its amplitude and
phase. The optimum receiver computes for all i the quantities
Ts Ts
Gn = / r(t)s;({)dt = V3 | r(t) sin(@ct + 6;)dt
0 0
t Ts (4.121)
LIAS co6 | r(t)sin@-tdt + sin a | rn cosextdt|
0 0
= Acos[6; — n]
where defining
Ts
ne a r()||v2PP sin we t|dr (4.122)
0
Ts
A. af ri) V2P cos wet|dt (4.123)
0
we have

AA,/A2+ A2 (4.124)

and

n tan! ae (4.125)
As
On the basis of this representation of g;, it is important to note that the receiver of Fig.
4.6 requires M correlators, whereas an equivalent receiver (see Fig. 4.23) requires only two
 208 Coherent Communications with Waveforms Chap. 4

[°2

(a) ANTIPODAL (b) QUADRIPHASE

SIGNALS (M = 2) SIGNALS (M = 4)

(c) OCTOPHASE SIGNALS (M = 8) Figure 4.22 The signal vectors for various
polyphase signal sets: (a) Antipodal signals
THE DOTTED LINES REPRESENT DECISION (M = 2), (b) Quadriphase signals (M = 4),
THRESHOLDS IN THE RECEIVER (c) Octaphase signals (M = 8)

correlators. Since for any 7, gj represents a sample of a Gaussian process, the joint pdf of A
and n, given that s;(t) was transmitted, is

A A?—2E,A cos(n—6;) +E?

f(A, n) = |ZE,No °XP [|- Se ee |, USA S00, 0S S20 ea06)
0: elsewhere

Letting r 4 A/./EsNo and averaging over the random variable r gives

"CO
fn) = | © exp{—[r 2 — 2r,/ Ft
Es
cos(n — 61) + Es
Ht||dr, Wr 7Reg (4.127)
0, : elsewhere

Since qj is a maximum when the distance |6; — 7| is a minimum, then for 7 in the region
(6; — 7/M, 6; + 1/M), the inequality |0; — n| > |6; — n| is valid for all j #i and hence
gi > q; for all j #1. Based on this result, the probability of correctly detecting the i’” signal
is the probability that 7 is in the region [(2i — 1) /M, (2i + 1) /M}, that is
(2i+1)/M
Ps(C/mi) =|! f(mdn (4.128)
(2i—1)x/M
Letting y 4 n — 2i/M, we see that P;(C/m;) is independent of i; hence, the probability
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 209

20 COS Ot

Figure 4.23 M-PSK receiver requiring two correlators

of correct detection becomes (see Problem 4.22)

2 E, (o-e) m/M Eg
pc) == exp (-=) | rexp (-r?) [ exp 2r,|-cosw )dy |dr (4.129)

Introducing the change of variables u =rcos yw, v=rsinw and noting that v =u tany,
(4.129) reduces to

9) ee) E, A utanz/M 5
p(c)=— | exp4 — (eh No | exp(—v av du (4.130)

The corresponding symbol error probability is derived in Appendix 4A and can be written
as

P.(E) M— | | f Deu Set

‘ = — — — -er i
: M 2 No M
(4.131)
| a/ Es/No sina /M
oe) perf (ycot Da
exp(—y 7) 4 is
fm Jo
For the case of BPSK, M = 2 and the performance given by (4.131) reduces to (4.55). It is
also possible to evaluate the integral in (4.131) in closed form for M = 4, that is, QPSK.
The result is given by

PAE
s(E)|m=4
as Ao =erfCS Es
No amie 4, rfc?
fil Es
2No
gtee. (4.132)
Ske

In the region where E,/No > 1, (4.132) is well approximated by

Es
:
P(E) ~ ere (= (4.133 )

This implies that the symbol error probability performance using four signals is about 3 dB
worse than that of orthogonal signals discussed earlier. For values of P;() less than 10m
210 Coherent Communications with Waveforms Chap. 4

—— mPSK
- - - - MDPSK

Figure 4.24 Symbol error probability for

“5 0 5 10 15 20 25 30 M-PSK signals as a function of bit SNR
EL/No: dB (Reprinted from [12])

the region of great interest in practice, the error probability for M > 2 is well approximated
by

] arate | .

since when an error occurs, one of the two neighbors of the transmitted signal is chosen
with high probability. This approximation becomes extremely tight for fixed M as E;/No
increases and serves as an upper bound on the probability of symbol error for any E;/No
(see Problem 4.23). The performance for various values of M is depicted in Fig. 4.24 for
M =2,4, 8, 16, 32, and 64, as a function of bit SNR.
The computation of the bit error rate for M-PSK signals when M = 2” is considered
for Gray code mapping, as given in Table 4.1. The Gray code has the property that in going
from one symbol to an adjacent symbol, only one bit out of the log, M bits changes. As
a result, an error in an adjacent symbol is accompanied by one and only one bit error. At
large symbol SNRs, we can assume that the only significant symbol errors are those whose
corresponding signal phases are adjacent to that of the transmitted signal. Hence, the bit
error probability can be approximated by

Ps(E)
P,(E) = eee, (4.135)
2
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 211

Table 4.1 Gray Code Bit Mapping and Hamming Weight (Reprinted from [7])

MESSAGE | _ BITS |Win| MESSAGE BITS

=) I oa M i io)fo

OANDOaAARWNH-O

OO
—~NWN—"N

$O
HN
S—~NMNWNHWHWNHH]NHWNH
WA
AA
OC
=
—-$NMNW
WNHHANWNWAWNHH]NMWNH=|

At lower symbol SNRs where the approximation of (4.135) is not valid, it was not until
recently that the bit error rates were able to be evaluated exactly. Only the key results are
presented here and the reader is referred to [7, 8] and its references for additional detail. Let
Rm denote the decision region when message m is transmitted. Let S,, be the probability that
the received signal vector r falls into R,, when message “0,” corresponding to zero phase, is
transmitted (no loss in generality is obtained by considering message “‘0”’). Then, it can be
shown that

ae case {Q [ztan((2m — 1)an)]

= wmalztan(en + 1)an)]}dz Po
m=0,1,:::,7 (4.136)
Io? FR (Ole tan(M— 2m — 1an)]
~Q[ztan((M — 2m + 1)an)]} dz m=M“41,...,4

where d, = ange, An = 10/2" with M = 2” and

212 Coherent Communications with Waveforms Chap. 4

M
Sm = Sm—m for fara ae Net (4.137)

Let W,, be the Hamming weight of the bits assigned to symbol m (see Table 4.1). Since
the probability of a bit error is Wn/logyM when message “0” (all zero bit pattern) is
transmitted and symbol m is decided (that is, re Rm), then
M-1
P(E) = De Wn Sin (4.138)
log, M m=1

For M = 2, 4, it can be shown using integration by parts that

P,(E)= Q(dj sina;) M=2 (4.139)
1 ;
Py(E) = 515i +258)+ 83}= Q(d2sinaz) M=4 (4.140)

Since d) sina, = d2 sin a, the bit error rates for BPSK and QPSK are identical. For M = 8,
we have
1
Pp(E) = ~{S, + 2524+ S3 + 254 + 385 + 2S6 + S7}
3
5 (4.141)

Fis {Q(d3 sin a3) + Q(d3 sin 3a3) [1 — Q(d3 sin a3)]}

For
M = 16,
(2 5
PAE) = 5) Dd)Sm+ D|Sm + $5 +286 + Sy (4.142)
m=1 =o

The exact BERs for M-PSK are shown in Table 4.2 for M = 2,4, 8, 16,32, and 64 and
plotted in Fig. 4.25.
A more recent study [9] presented a technique for evaluating the bit error probability
with any bit mapping for M-PSK signals. A simple technique for evaluating S,, is suggested
which uses closed form expressions for the half-plane and quadrant plane probabilities in
terms of the Q-function. Bit error probabilities for M-PSK with Gray, natural binary, and
folded binary bit mapping are derived.
L-Orthogonal Signals. _L-orthogonal signal sets [10] are generalizations of biorth-
ogonal signal sets and consist of M = NL equal energy, equal time-duration signals. A
given L-orthogonal signal set has the following two properties: (1) the signal set may be
divided into N disjoint subsets, each subset containing L signals; and (2) signals from
different subsets are orthogonal. One instance of an L-orthogonal signal set is realized when
each signal subset is chosen as a set of polyphase signals. A construction of such an L-
orthogonal signal set is given in Reed and Scholtz [10]. The signal correlation coefficients
are given by”

|aA]+[4
cos (22 - j)), a)-[e
vij= 41, i J (4.143)

2. [y] = integer part of y.

Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 213

(E)
P,

-7
10 0 4 8 12 16 20 24 28
E,/N, dB
Figure 4.25 Bit error rate of M-PSK over AWGN channel (Reprinted from [7])

The corresponding correlation matrix 7 —o;th can be partitioned in the following form:

I'7—psk O vee O
O [ne
Pz -orth = nee (4.144)
O BA Bat) sae
where I’; psx is the matrix defined in (4.119) with M replaced by L, and © is the L x L
null matrix whose elements are all zero. Again, the special case L = 2 yields biorthogonal
signal sets. Other special cases appear of little practical importance in the implementation
of phase-coherent systems.
A set of M signals that have the correlation properties of (4.143) can be realized
by combining the orthogonal M-FSK signal discussed in Example 4.2 with the polyphase
signal representation of (4.120). Specifically, the set of M = NL signals

/2P sin|o-t +2 +6)|,, L=1,2

(t) =
SoG ¢ 4.145
2P sin [eet + 24! + 6], 1G =D ( )

is L-orthogonal (see Problem 4.24), where 0; = 2/7 /L andi =nL +1 withn=0,1,---,

N —1 and/=0,1,---,L — 1. Several other examples of L-orthogonal signals sets are
given in Problem 4.25. Note that the minimum frequency separation between the signals
for L > 2 is 1/T as opposed to 1/2T required for L = 1, 2. We will see in a later chapter
that 1/7 is the minimum signal separation required for noncoherent reception to guarantee
orthogonality. When the phase takes on more than two values, it can be argued that it
becomes “unknown” or random and hence, the frequency separation for noncoherent signals
is more appropriate.
Since for the above set of signals, the correlation matrix of (4.144) is singular, the
214 Coherent Communications with Waveforms Chap. 4

Table 4.2 Exact BER Values of M-ary PSK with Gray Code Mapping
(Reprinted from 2

Py (E Ph (E me (E) Pp (E Pp (E
E,/N,.dB} (M= 54) hae M=16) | (M= 32) (M = 6a)
2.132E-01 2.468E-01 2.867E-01 3.185E-01 3.376E-01
1.861E-01 2.217E-01 2.646E-01 3.003E-01 3.242E-01
1.584E-01 1.961E-01 2.420E-01 2.817E-01 3.099E-01
1.306E-01 1.708E-01 2.191E-01 2.629E-01 2.951E-01
1.038E-01 1.461E-01 1.965E-01 2.442E-01 2.800E-01
7.865E-02 1.227E-01 1.745E-01 2.256E-01 2.649E-01
5.628E-02 1.008E-01 1.535E-01 2.073E-01 2.497E-01
3.751E-02 8.061E-02 1.338E-01 1.892E-01 2.345E-01
2.288E-02 6.225E-02 1.155E-01 1.714E-01 2.194E-01
1.250E-02 4.589E-02 9.865E-02 1.538E-01 2.043E-01
5.954E-03 3.186E-02 8.292E-02 1.368E-01 1.895E-01
2.388E-03 2.048E-02 6.816E-02 1.207E-01 1.747E-01
7.727E-04 1.195E-02 5.429E-02 1.055E-01 1.599E-01
1.909E-04 6.181E-03 4.145E-02 9.147E-02 1.452E-01
3.363E-05 2.748E-03 2.998E-02 7.840E-02 1.307E-01
3.872E-06 1.011E-03 2.025E-02 6.614E-02 1.165E-01
2.613E-07 2.937E-04 1.256E-02 5.451E-02 1.030E-01
9.006E-09 6.338E-05 7.010E-03 4.349E-02 9.027E-02
9.417E-06 3.427E-03 3.325E-02 7.848E-02
8.756E-07 1.421E-03 2.406E-02 6.752E-02
4.516E-08 4.789E-04 1.627E-02 5.724E-02
1.246E-04 1.010E-02 4.747E-02
2.342E-05 5.642E-03 3.819E-02
2.925E-06 2.763E-03 2.950E-02
2.187E-07 1.147E-03 2.163E-02
8.573E-09 3.876E-04 1.486E-02
1.011E-04 9.417E-03
1.907E-05 5.394E-03
2.393E-06 2.725E-03
1.799E-07 1.177E-03
7.099E-09 4.176E-04
1.159E-04
2.361E-05
3.264E-06

dimensionality of the integration in (4.77) must be reduced. A procedure for doing so is

described in Reed and Scholtz. The details of the derivation are quite involved, and hence
we merely present the results. For L > 2, the probability of correct detection is given by?

pice ff welt)
(4.146)

| ffPEO Daan
N-1

dx\dx2
A(x1)

where A(x) and B are regions of integration as shown in Fig. 4.26. In particular, B is the
set of points (x1, x2) in the wedge bounded by the straight lines

3. For L = | and L = 2, we have the well-known results of orthogonal and biorthogonal signaling respectively.
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 215

a RADIANS
., REGION
Yt A(x)

5
tH Z Uy)

REGULAR L-SIDED POLYGON

Figure 4.26 Areas of integration (Examples drawn for L = 6)

Xx Gite an :

x X41 an 4.

and A(x;) is the set of points within an L-sided regular polygon circumscribed on a circle
of radius x; + ./E;/No. Substituting (4.147) and (4.148) in (4.146) and making the change
of variables uj = x; + /E;/No gives (for L > 2)

2 (o-e) E; - uy, tan(z/L)

P,(C) = = / exp|—(ui-
JO
/=) ||f
No 0
exp (=x})
dx
(4.149)

x
|A(uj~s/Es/No) a
Since any L-sided regular polygon can be subdivided into L equal area triangles, then from
the symmetry of the double integral, we get
2
2 ee) E uy, tan(a/L)
I(r =f exp |= (.— ‘| | exp (—x3) dx?
mt Jo No 0
N-1
uy y, tan(r/L) exp {— y2 + y2

0 —y, tan(7r/L) wu
(4.150)
Z

= — exp) —
Ta iy
haat U
LT mal hime
Es
No erf(u tan -= )
Now ths
L uy 5 *e N-1
x Fal ee —yj})
exp ( yi)e erf(yi
YI tanSF=)dn| d Uy
216 Coherent Communications with Waveforms Chap. 4

The special case of L = 2 corresponds to the biorthogonal signal set and hence (4.150) re-
duces to (4.102) (see Problem 4.26). When N = 1, we merely have a single set of polyphase
signals, and by inspection (4.150) reduces to (4.130). For other combinations of N and L,
the integrals in (4.150) must be evaluated numerically.
Upper and lower bounds on the probability of symbol error P;(E) = 1 — Ps;(C) may
be readily calculated by circular approximation to the area of integration A(u)). In particu-
lar, an upper bound on P;(£) [lower bound on P,(C)] is obtained by approximating A(u1)
by its inscribed circle of radius u;. Similarly, a lower bound on P;(£) may be calculated
by approximating A(u1) by its circumscribed circle of radius u; sec(2/L). Considering first
the upper bound on P,(£) and letting yj) =r cos w and y2 =r sin y gives
»
2 (oe) E; uy tan(ar/L) >
Pees
5 ( == = | exp Sa
(. 4 | exp (aya
x3) x2

O16 (4.151)
9) ie E; ¢ uy tan(z/L) 5
=— exp | — [ui —,/— / ex (= )dx,
m Jo No 0 ee
N-1
x [1— exp (-«7)] duy

Recall the binomial expansion

v1 N=!
{1—exp [-«7]| =~ y-1Cx(—1) exp (—kut) (4.152)
k=0
where ,,Cx is the binomial coefficient given by

CA n!}
nk Get (4.153)

Substituting (4.152) into (4.151) and making the changes of variables u = uj./1 +k, v=
x2V1+k gives

be Web.) tsi]
PC)
(0) 2 pg CE exp|=
nC l+k (4.154)

te E;/No 2 utanz/L v2
x ~— exp} —lu— - —
Alay hee 1+k | exp a} a]
Recognizing that the k =0 term in (4.154) corresponds to P;(C)|m—psxK as given by
(4.130), we write the upper bound on P,(E) as*

4. P;(E)|1—psx is the symbol error probability of M — PSK signals with M replaced by L, hence the notation
L— PSK.
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 217

Noa
Ct) k kEs/No
5
P(E) (E) = s(E)|~-—psK
PACE) 7 = _ aS pee VtSiG, exp |
— eee (4.155)
:

a} oo yee 2 utanz/L v2
oy = ©,90) || = a du|\d
=i le VeLek | exp | ai 7 ‘i
where P;(E)|z-psk A 1 — Ps(C)|z—psx. Similarly, one can derive a lower bound on P,(E)
with the result

(E)
PEG) >> PCE)\
P(E) |r p= psx _
+D ——
Cb
eee n-1Cx exp {-
k'Es/No
(4.156)
1+k’
2
a co E;/No utanz/L v2
x = [ exp | = |u— / exp }—-—.} dv |du
a Jo 1+k’ 0 LE ke

where k’ 4 k sec? (1/L). The bounds of (4.155) and (4.156) can be simplified by reducing
the double integration involving an infinite limit to a single integration having a finite limit.
The procedure for doing this reduction parallels that given in Appendix 4A. The results are
N-1 Cay k kEs/No
pe)
PCE) == PAE
P(E) YAP.
SP.) hi psk.
ir — Ss ep = Ke cep |
=: cE

6 1 Ee INGps
[45
— + —erf (Ge sno)
0 (ANlioi )

1 / (Es/No)/(k+1) sin@ ( :

+ Jr | exp (—y )erf(ycot Ady

Jo
with 6 4 tan! [see |and

N-1 k }
(1) k'Es/No
P(E) = P; (E) & P;(E)|L—psk = Da (aaa n—-1Ck exp {~<a
k=]

yest | E;/N
Op Natit BasMa
| i Il+k (4.158)
Ge

J(Es/No)/
(+1) sin6!
veal exp (-»’) erf (y cot 6’) dy

with 0 4 tan~! [se |.We note from (4.157) and (4.158) that for N = 1, the two bounds
become equal and give the exact performance; that is P;(E)|, psx. Also for large L (small
N), the bounds are very tight since both the inscribed and circumscribed circles are good
approximations to the L-sided polygon. Figure 4.27 illustrates the upper and lower bounds
of (4.157) and (4.158) respectively as functions of E,/No in dB for M = 64 and various
combinations of N and L.
 218 Coherent Communications with Waveforms Chap. 4

P(E)

\
‘
‘
‘
My
‘

- - LOWER \

10°4+— BOUNDS
‘
iY
‘
\
— UPPER ‘
iY
ir BOUNDS ty
‘
‘ x
|. XxX SUBOPTIMUM x
p RECEIVER . | | | | |
10° Figure 4.27 Symbol error probability of
5 0 Say iG 15 L-orthogonal (M = 64) signals as a function
E,/No.dB of bit SNR (Reprinted form [12])

Optimum detection of the set of signals in (4.145) requires the correlation receiver
of Fig. 4.6 and yields the performance as given by (4.150). When M becomes large, the
number of correlators required to implement the receiver of Fig. 4.6 becomes prohibitive.
Hence, we investigate a suboptimum receiver that requires only 2N correlators and whose
error probability performance is quite close to that of the optimum receiver.
The suboptimum receiver illustrated in Fig. 4.28 uses a noncoherent detector to de-
termine the most probable index n characterizing the received signal [see (4.145)] and si-
multaneously makes a maximum-likelihood decision on the phase 6), or equivalently, the
index /. We note from (4.145) that the index i is correct if and only if n and / are correctly
chosen. The probability that n is correct is the probability of the event {V2 = max, V2} or
equivalently, the probability of choosing the correct frequency from a total of N choices.
This probability P;(C)|now is evaluated in Chapter 5 and can be obtained by taking one
minus the error probability performance as given by (5.77) with M replaced by N. Having
selected the correct frequency, the conditional probability P;(C) of correctly choosing / (or
equivalently, the phase 6; from a total of L choices) is given by (4.130) with M replaced by
L. Thus, the probability of correct reception is

Ps(C)|y signals = Ps(C)|won.n signals Ps(C)|L—psk (4.159)

Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 219

or equivalently, the symbol error probability for the suboptimum receiver of Fig. 4.28 is
given by

P(E) ly signals — P(E) |von. signals + Ps(E)|_—psk (4.160)

— Ps(E) von,n signals ?s(E)t—psk

The error probability performance of the suboptimum receiver as given by (4.160) is su-
perimposed in Fig. 4.27 on the performance bounds for the optimum receiver in Fig. 4.6.
For large Ep/No, the difference in performance between the optimum and suboptimum re-
ceivers is virtually negligible. Finally, note that the receiver implementation of Fig. 4.28
requires N noncoherent frequency detectors and L phase estimators. Since the phase esti-
mate is obtained after the correct frequency has been chosen, an alternate implementation
could employ only a single phase estimator whose input terminals are switched to those of
the noncoherent detector corresponding to the chosen frequency.
M-ary Amplitude Modulation (M-AM). Unlike previous signal sets, M-ary ampli-
tude modulated signals are not equal energy signals. In this case, the messages are transmit-
ted using M distinct amplitudes?, that is

si(t)=JEi@(), OK<t<T, (4.161)

fori =0,1,---, M@— 1 where (tf) is any unit energy function of duration 7; seconds, that
is, es o*(t)dt = 1. Since the energies are unequal, the normalized correlation coefficients
become [see (4.45)]

Ts / ¢. :
pe a (32) (42)dt =1 for alli and j (4.162)
Hence, the signal correlation matrix I" has unity entries and is noninvertible. The optimum
receiver is the minimum distance decision rule and since all the signals are expressed in
terms of one function #(t), the waveform problem can be reduced to an M-ary scalar
equivalent problem. When ./£; =iA, the optimum receiver is given in Example 3.3 and
its performance is evaluated in Example 3.4. Noting that o* = No/2, we obtain

Pry = M = 1 | A?
tx, seen: 4.163
(2) ( M jo 2No

If we further assume that the transmitter is subject to an average power limitation of Eayg/Ts
(Watts), it follows that
M-\1
@ 1 _ Eavg (4.164)

Making use of the inequality

Fe 4 M(M—1)2M -1)
iy — 6 (4.165 )
j=0
5. Quite often, M-AM is defined by a set of amplitudes that are uniformly spaced and symmetrically located
around zero.
 YHAIYYVO
NOLVZINOYHONAS

220
18s. Loa
=()s Q))'s
ONIGN
Ol Od
ONIGNOdS3YHOO
a! %,
» oi}
Yi} Ou {[ou
eat

SNIGNO
Ol d
lane [Pu
y b-Ny
NXOrFrtOrOL

ASWHd YOLVWILSS
ee ee oe ee Roo
AINSI]
P 87 y [eLNdogns
JOALID0I
JO} JUA1AYO9 UoNa!9p
Jo [eUOBOY}S[PUBIS
i0-7
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 221

it is readily seen that the quantity A is constrained to equal

eerie tayy
ee i2 4.166
VG DOM 1 ae)
Consequently, the average probability of error may be written as

P,(E)
aa =2 (=>)
NeM QO Noi —aDQM —1) 4.167
See
When M = 2, (4.167) reduces to the “on-off” keying performance of (4.68) as expected.
Given a symbol-to-bit mapping, the bit error probability can be computed exactly in closed
form using (4.86) and the probabilities of Example 3.4.

4.1.5 Limiting Error Probability for Block Codes as M — oo

We are now in a position to relate the limiting error probability performance of block codes
as the number of codewords M approaches infinity to the notions of channel capacity,
communication efficiency, and the Shannon limit introduced in Chapter 1. We shall deal
with orthogonal signal sets, although the conclusions drawn apply to a more general class
of block codes, including biorthogonal and transorthogonal sets.
Consider the probability of correct symbol detection for orthogonal signals as given
in (4.93), that is

(ee) see)
| jer (-«- i) anleeg (4.168)
el 2
With a simple change of variables, (4.168) can be expressed as
* we: M-1

P,(C) = se) fo(-»- [Perce a) dy 469)

Consider the limiting behavior of P;(C) as M approaches infinity, that is
M-1
© exp (—v7/2) 2E,
i = ]j ——— —v — ,| — log, M dv
Sw Ps(C) Mes 253 Jon 2 No ee
(4.170)
M-1
00 yD 2Ep
= er?) lim Ot =v log, M dv
LAs Jn M->00 No

If we define 6 4 E,/No and consider the limit of the logarithm of the expression in brackets,
we obtain
Mot. In Q (—v — \/2 logy M)
jim,mn[Q(—-»—v26t0e.M)| = Jim“gy ITD
As M > oo, the limit approaches In Q(—0o)/0 = 0/0. Treating M as a continuous variable
222 Coherent Communications with Waveforms Chap. 4

and using I’Hospital’s rule, we have®

gta eB) Ne ae Hf pl oy TR loo 2 \ae WS gcpanay:

B 08> )
x(v + B 082 ) a (v
Fra (ee) GE) |

(4.172)

Defining the constants

bool v?
A,| ee a 4.173

cAv ia (4.174)
In2
and noting that

d 1 aed:
os 28 1025 M |= ——SS 4.175
dM (v+ paces ) Min MV 21n2 oe
In M
exp (—6———In2 ) = m~(6/In2) (4.176)
(4.172) can be expressed as

‘ , 5 1)* exp (-cov In M) 1 4 177)

in (—c)) ———— —YT _- SH :
Moo mit 153) Vin M Q (—v — /2B log, M)
For large M, Q(—v — ,/2B log, M) > 1 and M — 1 ~ M. Hence, (4.177) gives

Rae: cae (—c2 In M)

lim (—c,)M\~ ing? —+—____* 4.178
M-— oo Jin M :

Letting w A J/In M, then M = exp(w?) and (4.178) reduces to

exp (—c2w)
exp ({ 1 — aa w?
lim (—c})
w— oo Ww
(( : ) Ihe <a
Me ©,©)
Bebe i)
B/ In2s<4

The limit is independent of the sign of cz as the exponential in w* always dominates the
exponential in w as w — oo. Since this is the limit of the logarithm in brackets, the limit of
the expression itself is
M-1
: 2Ep 1, Bee >in
lim —v —./— 1 Bead Boxe
ain. ‘ \ No cn) fc E}/No <In2 eee
6. To compute the derivative of Q(x), use Leibnitz’s rule
q
cs
aa Forvayax=ae fgeeage dq - Bayedp
Pp
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 223

so that the limit of (4.168) becomes

1, E,/No>1n2
ee PAC)=| 4 E»/No <\n2 ce
or, in terms of error probability

0, E,/No> In2
jim rey= {9 Ep/No <In2 ae
Allowing M to approach infinity is equivalent to considering the case of unconstrained
bandwidth. Hence, an orthogonal coding scheme produces error-free transmission in the
unconstrained bandwidth, additive white Gaussian noise channel provided that E,/No >
In2. The limiting behavior is particularly significant from an information-theoretic point
of view. Shannon’s formula for the capacity of a channel of bandwidth B perturbed by
additive white Gaussian noise with one-sided PSD No is given by (1.30) and repeated here
for convenience:
P

C= Blog, (:a xs) bits/sec (4.183)

This represents the maximum bit rate for error-free transmission. Letting the bandwidth B
approach infinity and using the expansion In(1 + x) =x — x7/2+x3/3—---, |x| < 1, then
B je B P 12
li = lim
poe gua ere—~In{ n(1+ +a)
— ]=([-—~](——]=
(=) (x) NED 4.184
TMi
Thus, from (4.182) and (4.184), it follows that orthogonal signals achieve error-free trans-
mission in the limit as M — oo, provided that the bit rate 7, is less than

Pee hmG bits/sec (4.185)

O PENG in) bese
4.1.6 Matrix Generation of Binary Signal Sets

In this section, we shall briefly discuss methods of generating orthogonal, biorthogonal, and ° :
transorthogonal signal sets for the binary case, that is
N
Sia) Sippj(t) POE oom AGE E= II (4.186)
j=!
where sj; = +1 for all i and j and {$j (t), j = 1, 2, ---} is any set of orthonormal functions
over the interval (0, 7;). The signals as defined by (4.186) are normalized with energy N.
For any energy E;, the signals should be scaled by /E;/N. The signals can alternatively
be represented by their respective vectors $; = (81, 5i2,+°-, siv)-. Let Hy be an N x N
matrix defined inductively by

FA Soa ik
TEE Wy. alae
Hy (4.187)
;
where

r= ( A) (4.188)
224 Coherent Communications with Waveforms Chap. 4

Such a matrix is called “Sylvester-type Hadamard” matrix and satisfies the property that

Fintan (4.189)
By definition, the construction of such matrices is known when N = 2*. It is also conjec-
tured that when N = 0 mod 4, such matrices exist [11]. As an example, 7112 is given by

1 1 1 1 1 1 | 1 1 1 1 1
1 =| 1M el 1 l 1 =f =1 =1 ||
1 = ee 1 =1 1 1 t=) Si) al 1
1 a | 1 s=! 1 1 |
i ey
1 a 1 -1. -! 1 i 1 1 PRS ee
i =| —I!I ete — 1 ea 1 1 eel
(4.190)
eres |STN Fe Ae oO ny a eee
1 1 =1 =! --! ees Na | geal 1 1
1 1 1 =—1 -1 —] | a oe | 1 =] 1
1 1 1 1 =-1 -l -! ii eal es |
if =i 1 [ =f =b L es 1
1 Po, 1 1 Pe eer (eee | a |

Since Hy HH", = NIn, the columns of these matrices can be used to construct M orthogonal
signals each with N coordinates with M < N. To construct M biorthogonal signals, use any
M/2 columns and augment the set with the negative of the columns to obtain the complete
set of M signals. To obtain M simplex signals, delete the top row (or first column) of Hy
and use the remaining columns as vectors. The number of coordinates will be one less than
the number of constructed vectors as expected.
Example 4.3
Design binary vectors representing eight orthogonal, biorthogonal, and simplex signal sets with
E; = 8. Use Sylvester-type Hadamard matrices to obtain the vectors.
Since M = 8, we need to compute 71g as

[oe pewWitk Eo Up oe
} pee SESS ia est
eis at eed 1 rs eC
Hele H4 ie ee 4 l bk ie 1
Ha —Ha 1 l i 1 ee ee Pe ees
tos aes Drs bel We Re:
| Li sk Sheep er aa 1
1S =besets “lv St Fees
= (So S$} S2 ---S7)

To obtain the orthogonal vectors (E, = 8), use

So S1 S2 -::87

For biorthogonal signals, use

SOUS] (S25S83) 80 ia SSeS

For simplex vectors, delete first row from Hg and use remaining columns to obtain
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 225

1 —1 1 —1
] ] =] —1
1 —1 —] ]
So= [et sp ies = ja) ec ead Rn
1 —1 1 1
1 1 —1 1
1 =| —1 —|
We can check to see that

1 1
ee
1 Sy Fd
s( ee
VE ( 3)
and

4.1.7 Coherent Detection of Differentially Encoded M-PSK Signals

Thus far in our discussion of optimum coherent detection of M-PSK signals, we have
tacitly assumed that the receiver was perfectly synchronized to (had exact knowledge of)
the frequency and phase of the received signal’s carrier. In practice, the receiver derives its
frequency and phase demodulation references from a carrier synchronization loop which,
depending on its type,’ may exhibit an M-fold phase ambiguity. That is, the receiver is
capable of establishing a phase reference only to within an angle @g of the exact phase.
Since the phase ambiguity indeed takes on values equal to those of the M-PSK data phases
themselves, then any value of ¢, other than zero will cause a correctly detected phase to be
erroneously mistaken for one of the other possible transmitted phases, even in the absence
of noise! Clearly, something must be done to resolve this phase ambiguity else the receiver
will suffer a performance degradation. For example, for a phase ambiguity corresponding
to m = M/2, the receiver will be 180° out of phase with the received signal and all data
decisions will be inverted resulting in very poor error probability performance.
One simple solution to the above is to employ differential encoding at the transmit-
ter yet still maintain coherent detection at the receiver. The theory of differential encod-
ing and how it resolves the above-mentioned phase ambiguity is described as follows. Let
5i(t) = /2P cos (wet + 6;) represent® the M-PSK signal to be transmitted during the i‘
transmission interval i7, <t < (i + 1)7; where 6; ranges over the set of allowable values
Bm = 22m/M;m=0,1,...,M — 1. Ordinarily, 6; would be exactly equal to the informa--
tion phase, say A6;, to be communicated in that same transmission interval. At the conclu-
sion of the i‘ transmission interval, that is, at time t = (i + 1)7;, the receiver makes its
M-ary decision 6; on 6; (or equivalently, A@;). Because of the ambiguity associated with the
demodulation reference phase, however, this decision consists of an unambiguous part by.i
plus the ambiguity ¢, itself, that is, 6; = By.i + oq. Suppose now that instead of assigning
0; identically equal to A6;, we instead encode A6; as the difference between two adjacent
7. A complete discussion of the various types of carrier synchronization loops for M-PSK modulation is pre-
sented in Volume II of this book.
8. In this section, it is necessary to introduce a notation to characterize discrete time. For simplicity, we shall
use a subscript, e.g., i, on a variable to denote the value of that variable in the i” transmission interval. To avoid
confusion with the previous notation of 6; to denote the i'” message phase 277i /M (i" hypothesis), we introduce a
new Greek symbol, i.e., B;, to denote the i'" message phase. Whenever it is necessary to represent a variable both
in discrete time and/or by an associated hypothesis, we shall use a parenthetical superscript for the latter.
226 Coherent Communications with Waveforms Chap. 4

transmitted phases, that is, A@; = 6; — 6;—1. If the receiver still performs decisions on the
transmitted phases 6;_; and 6; but arrives at its decision on the information phase based on
the difference decision variable 6; _ 2 = (4. + ba) — (4.11 = ta) = 6,7 —0y.7=15
then clearly the phase ambiguity has been eliminated. Inherent in the above ambiguity res-
olution is the assumption that the phase ambiguity ¢, introduced by the carrier synchro-
nization loop remains constant over the two transmission intervals of interest, that is, for
a duration of 27, seconds. If we continue to encode the information phases in the above
manner and base our decisions on the difference of adjacent decisions made on the actual
transmitted phases, then clearly a sequence of Ns + | transmitted phases is needed to con-
vey N, symbols of information. That is, the first symbol in the transmitted sequence is used
to resolve the phase ambiguity.
As a first example, consider a BPSK modulation (M = 2) and suppose that we desire
to transmit the information sequence A@; = 7, Adz = 1, A63 = 0, A64 = 7, and As = 0.
Assuming 69 = 0 as the reference phase, then the remainder of the transmitted (that is,
the differentially encoded) phases reduced modulo 27 would be 6; = 6) + A@) = 7, 62 =
6; + Adz = 0, 63 = 62 + Ab3 = 0, 04 = 63 + AO4 = 77, 05 = 64 + AOs = 7. In the absence
of noise and assuming no phase ambiguity introduced by the carrier synchronization loop,
the receiver would make perfect decisions on the received sequence, that is, 6=6; i=
0,1,2,...,5 and would then differentially decode the modulo 27 difference of these de-
cisions as 6 — 6) = 1, 6) — 6 = 7, 63 — 65 = 0, 64 —-63 =z, and 6s —64=0 which is
identical to the information sequence. If now the carrier synchronization loop introduced a
phase ambiguity of z radians thus converting the received phase symbols into the sequence
6, = 1, 0; = 0, 05 = 1, 6, = 7, 0, = 0, and 6{ = 0, then in the absence of noise we would
make perfect decisions, that is, 6/ =6', i=0,1,2,...,5 and after differential decoding
(modulo 27) obtain 6; — 6 = 7, 65 — 6; = 7, 65 — 05 = 0, 6, — 6, = 7, 65 — 8, =0, which
is once again identical to the information sequence.
As a second example, consider a QPSK modulation (M = 4) and suppose that
we desire to transmit the information sequence A@; = 21/2, A@2 = 2, AO3 =0, AOy=
32/2, AOs = 3/2. Again assuming 6) = 0 as the reference phase, then the remainder
of the transmitted (that is, the differentially encoded) phases reduced modulo 27 would
be6; = 69 + AQ) = 1/2, 02.= 0; + Ado =31/2,.03=
02 + Ads
= 37/2, 04 = 03 + AO,=
mt, 05 = 64 + A@s= 7/2. In the absence of noise and phase ambiguity, the receiver would
differentially decode (modulo 21) its perfect decisions, that is, 6; =i= mint 6) oa 6 ==
Toe) 6.-6, =n, 6; — 6) = (0), 64 — 63 = 37/2, and 65 — 64 = 3/2 which is identical to
the information sequence. In the presence of a phase ambiguity, say @g = 7/2, the re-
ceived phase symbols would be converted to the sequence 6) = 1/2, 0; = 7, 6, =0, 05 =
0, 0, = 32/2, 05 = 7 and in the absence of noise the receiver would sioke merce decisions
that is, 6/ = of i=0,1,2,...,5 and after differential decoding (modulo 27:) obtain
0, — 6) = 1/2, 0, — 0 = 1, 65 — 65 =0, 0) — 64 = 3/2, 0, — 6 = 3/2 which is once
again identical to the information sequence.
Figures 4.29a and 4.29b are illustrations of a differential encoder and a differential
decoder, respectively, where the inputs and outputs of these devices are the phase symbols
themselves. Alternately, one can represent differential encoding and decoding in terms of
the complex baseband representation of the signals which results in the configurations of
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 227

modulo 2r modulo 2x

Lee

Figure 4.29 Differential encoder and decoder (real form)

asterisk denotes
complex conjugate
multiplication

Figure 4.30 Differential encoder and decoder (complex form)

Figures 4.30a and 4.30b. Which differential encoder/decoder representation is used in a

particular situation is a function of whether the receiver is represented in its real or complex
form. ,
A digital communication system that employs differential encoding of the informa-
tion symbols and coherent detection of successive differentially encoded M-PSK signals is
called a differentially encoded coherent M-PSK system. We now derive the optimum (MAP)
receiver for such a system followed by an evaluation of its error probability performance.
We shall show that the structure of this receiver takes the form of the optimum receiver
for M-PSK signals without differential encoding (see Fig. 4.23) followed by a differential
decoder.
Optimum (MAP) Receiver for Differentially Encoded M-PSK. _ Based on the above
discussion of differential encoding, we observe that the desire to convey the information
phase A6; in the transmission interval i7; < t < (i + 1)7; results in the actual transmission
over the channel of 6;_; in the interval (i — 1)7, <t <iT, and 9; in the interval i7, <
t < (i + 1)T,. Hence, a receiver that makes an optimum coherent decision on A6; should
observe the received signal plus noise, r(t), in the 27;-sec interval (¢ — 1)7; <t < (i+
1)T; and make a joint optimum coherent decision on 6/_, = 6;-| + @a and 6/ = 6; + da,
228 Coherent Communications with Waveforms Chap. 4

namely, (ou; 6r). The ultimate decision on AQ; would then be obtained from A@; =

6! — oe modulo 277.
The design of an optimum MAP coherent receiver for jointly detecting 6/_, and 6;
follows the general likelihood function approach described in Section 4.1.3 and illustrated
in Fig. 4.6 and later applied to conventional (without differential encoding) M-PSK in Fig. _
4.23. Here however, for fixed @,, there are M7 hypotheses (pairs of possible values for
6!_, and 6!) and a joint decision is made based on a maximum correlation over the 27;-sec
interval (i — 1)T, < t < (i + 1)7,. In particular, analogous to (4.38), the receiver computes

(i+1)T;
x) & i) p'(t)s* (tat (4.191)
(i—1)Ts
where the prime on p(t) denotes the fact that the sample function of r(t) includes the phase
ambiguity gg and

sOC Tes € alls atl,

s(t) = 4sq); iT; <t<@+0T, eos:
0; otherwise

with s(t) = /2P cos (wet + By); iT; <t < (i + 1)T,. The decision rule on the transmit-
ted phases 6;-; and 6; in the presence of the phase ambiguity is then made as follows.
Choose Oo, = B; and 6! = f; where x (él) is the maximum value of x“-. Finally, the de-
cision on A@; would be Ad; = 6! _ oS = B; — B;. Since (4.191) can be obviously written
as
iT. +07,
ed p'(t)s(t + T,)dt +f p'(t)s (dt (4.193)
(i-1)Ts iTs
a receiver that implements the above decision rule is illustrated in Fig. 4.31. Furthermore,
because of the white noise model and the assumption of uncoded data [that is, s;_(t) is
independent of s;(t)], then

iTs (i+1)Ts
max x*) — max { p'(t)s(t + roa} + max / p'(t)s(t)dt
( i
a trys Ts (4.194)
A max pat + max mo
k l
and the joint decision on 6:_, and CH can be visualized as successive decisions on 6!_, and
6‘ independently. In view of this, the receiver of Fig. 4.31 can be redrawn in simpler form
(M correlators as opposed to M? correlators) as in Fig. 4.32a. Furthermore, transforming
the problem to polar coordinates as was done in arriving at Fig. 4.23 from Fig. 4.6, we
obtain the familiar form (see [12], Fig. 5-20) illustrated in Fig. 4.32b which requires only
two correlators. We emphasize that the equivalence between Figures 4.31 and 4.32a or 4.32b
would not be valid for coded (correlated) data despite the white noise assumption.
Error Probability Performance. To evaluate the symbol error probability perfor-
mance of the receiver of Fig. 4.32, we begin by considering the possible ways of making
a correct decision on A6j. Clearly, if both 6/_, and @/ correspond to correct decisions,
 Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 229

s(t) |
x! 0,0 )

Delay
T,

Choose

Largest

r'(t)

Delay 6

sO 7)

Figure 4.31 Optimum receiver for differentially encoded M-PSK

230 Coherent Communications with Waveforms Chap. 4

Choose

Largest

Figure 4.32a Optimum receiver for differentially encoded M-PSK equivalent to Figure 4.31

that is, ge, = 6;-1 + bq and 6) = 6; + da, then AO; will be detected correctly. The prob-

ability of this event occurring is PC) ey PAC Mab where P;(C)|y—psx iS given
by (4.130). Furthermore, if both On and 6! are in error but by the same amount, that is,
Oey = 6;-1 + dq + 2k /M and 6! = 6; + da + 2k /M, then A@; will still be detected cor-
rectly. Denoting the probability of this event by POC ), then referring to Fig. 4.32, we
have
k
PL (C) = Pr {(2k — 1)t/M <ni-1 — 6-1 — ba < Qk + I)z/M,
(4.195)
(2k — 1)n/M < ni — 6 — ba < (2k + 1) /M}
 i: —°¢lu

[eyuualassjiq
Japooaq
SOM 4 O[Npow
uz
}SOT[PUIS : —
ul , 9 oh
ainjosqy|
(1)4

A, ane |
9= [u—tgfuia
I)
| Aelaq
l a
ol ee eee
Py

JatIes

“yous
doo]
aaNSIq qzE"p NWINdQ Joatooo1
Joy Ajenuasaryip
Peposua YSq-W JUaTRAINbs
0} SANBIJ BZE'p

231
232 Coherent Communications with Waveforms Chap. 4

where nj—) and 7; are the receiver’s noisy estimates of the phases in the i — 1™ and i®
transmission intervals, respectively. Analogous to (4.127), the pdf’s of nj-1 and n; are

0oo © exp { [r2 — 2r\/Es

x
f (ni-1) = x cos (Ni-1 — 9-1 — ba) +f} ar; O<ni-1 520
0: otherwise
(4.196)

(ees bel eee

f (ni) = x C08 (ni — 61 — ba) + Fe||dr O<nj <2
0; otherwise

Since n;—; and nj; are statistically independent, then

( 6;-\+at (2k+1)2/M 6; +-ba+(2k+1)2/M
|e OF} Finan f f (ni) ani
j 6;-1+¢at(2k—1)2/M 6; +bat(2k—l)2/M
2
es eres
E
/ © bap) |p? on | teas
(2k+1)x/M_ poo E
s i No No
(2k-l)x/M JO 7

a(t") 2

Using the same change of variables as in going from (4.129) to (4.130) and following the
approach taken in Appendix 4A, we arrive at the following result:
2 ;
pes E, u tan[(2k+1)z/M]
p® — ih
exp |—v° }dudu
aur:mone See ae \ene utan{((2k—1)2/M]
No

=a
frei E
tie = sin
aed
$7
1 E
7 =|- ten] in
=
, : (4.198a)

ce a9(-)
Es /No sin{(2k-+1)2/M] 2k+1
ert
|yor SEU
a = |ay

nee, iege Gy ae (-»°)eelasees (2k — 1)x d

2Ja Jo M :
Alternately, using the approach of Section 3.2.8, we obtain the simpler result (see Problem
4.27)
x oOre - 2 (2k—-1
pO wk th i tes _Es (sin
G* de
; 2n Jo No sin?p
(4.198b)
1 m(1—(2k+1)/M) Ey sin2 Git in
eae exp {—-—— | ——_ *—_] }d
2x Jo sin? ?

Since the preceding events are mutually exclusive and exhaustive of the ways to make a
Sec. 4.1 Optimum Waveform Receivers in White Gaussian Noise 233

correct decision, the probability of a correct decision on the i symbol is

M-1
PQ eeP(C) (4.199)
k=0

Furthermore, since (4.199) is independent of i and since we have assumed equiprobable

symbols in arriving at Fig. 4.32, the average symbol error probability is
1 M-1

BE) = lois IS P; }(C)

M-1
s P&C) (4.200)

=2 Ps(E)Im=psk
s(E) |v —psk |1—
5
3 Ps(E)im—psk
1
Ps(E)|m—psk — 25 py
e=|

Ps(E)|m—psk

where P;(E)|y—psk = | — Ps(C)|y—psx is the average symbol error probability for M-

PSK without differential encoding and is given by (4.131).
Some special cases of (4.200) are of interest. For M = 2, we have coherent detection
of differentially encoded BPSK for which (4.200) reduces to

P,(E) =2 P;(E)|y_psx [1 — 2 Ps(E)|v_psx]

( E 1 Ee | (4.201)
= | erfc ,./— 1 — —erfc ,/—
\ No 2 No

For coherent detection of differentially encoded QPSK (M = 4), (4.200) simplifies to

yp Es 3 E, i 4 ES
Es
Pelb
Ce) jee 2 eric’
erfc \ No — 2erfc*
erfc | —— fc?
No + erfc |No
—— — =erf
a Cc INo (4.202 )

Analogous to the formation of a QPSK signal from two independent BPSK signals,
a differentially encoded QPSK signal can also be formed from two differentially encoded
BPSK signals. The optimum receiver would then differentially decode each BPSK signal
independently and then combine these two binary decisions to form the resulting decision on
the QPSK information symbol. The receiver for this system is illustrated in Fig. 4.33 where
the symbol phase decision is made in accordance with the bit to symbol mapping of Fig.
4.12a. To demonstrate that the symbol error probability performance of Fig. 4.33 is identical
to that of Fig. 4.32, we proceed as follows. For Fig. 4.33, the probability of a correct
symbol decision is equal to the probability that both binary decisions are correct. Since each
binary decision is the result of differential decoding a BPSK signal, then the probability of a
correct binary decision is one minus the probability of error given in (4.201) and hence the
probability of a correct QPSK symbol decision is equal to [1 — PE) lees | Finally, the
probability of error for the QPSK symbol is
2
P;(E) |p x. QPSK = a [1 ~ PE)\px. BPSK |
(4.203)
=2 P,(E)\px. ppsk — (Ps(E)Ip.z. BpsK)”
Substituting (4.201) into (4.203) and simplifying gives the identical result to (4.202).
234
dm —1@)so9
(°@

joqurds
aseu

(1)4
UOISIDaq

7A—
(’
—1°@)uts
aaNS1y
¢¢"p
= wind Jantaoo1
10j ATenusayjip
papoouaASdO
 Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 235

4.2 OPTIMUM WAVEFORM RECEIVERS IN COLORED GAUSSIAN NOISE

The problem of detecting signals in colored Gaussian noise can be approached in several
ways, namely, using the sampling approach, the K-L expansion, by performing an intuitive
whitening operation on the received signal or by a direct derivation with a sufficient statistic.
All four methods will be explored in this section to derive the optimum receiver in colored
Gaussian noise. Unlike detection in white Gaussian noise, the detection problem in colored
Gaussian noise and the performance of the optimum receiver depend on the length of the
observation interval. In white noise, the observation of the received signal over the interval
(0, 7), that is, over the signal duration, is sufficient to perform optimum detection. In
colored noise, the observation of the received signal outside the signal duration (and hence,
the observation of the corrupting noise) is of key importance as it provides information on
the noise present inside the signal interval. Limiting the observation interval to the signal
duration will in general result in a receiver performance inferior to that of observing a larger
interval. There are two key issues that will be addressed in this section: (1) the structure
and performance of the optimum receiver assuming a finite observation interval, and (2)
the structure and performance of the optimum receiver assuming an infinite observation
interval. The problem is then to decide between M messages where r(t) = s;(t) + n(t)
under m;, i = 0, 1,---, M — 1, over some observation period (finite or infinite) and n(t)
is Gaussian, zero-mean, nonwhite with autocorrelation function Ry(t1, t2) A E{n(t})n(t2)}.

4.2.1 Time-Sampling Approach

Let’s first start by investigating the time-sampling approach. In this case, the received signal
r(t), observed over the interval [0, 7], is sampled to produce the N-dimensional vector

ry = (r(At), r(2At),---,r(NAD)) (4.204)

consisting of samples taken At seconds apart. This provides a finite dimensional approxima-
tion to the waveform problem and we will then let N become large, improving the accuracy
of the approximation. Define the transmitted signal and noise vectors as
s! = (s;(At), s;(2At),
---, s;(N At) (4.205)
and
n’ = (n(At), n(2At),-+-,n(NAt)) (4.206)
The noise correlation matrix of the noise vector n of dimension JN is thus given by
Re CVst ght) 4*2% UR, (Ar, NAT)
ky = ae (4.207)
R,(NAt, Af) *«- Ra(NAt, NAP)
The density of the received vector ry, assuming message m; is transmitted, is thus Gaussian
with mean s; and correlation matrix 7 as given by (3.61), that is

fey (On/mi) = G (py; Si; Rw) (4.208)

and the unconditional density f-,(o,) therefore becomes
M-1
fry(Py) = >, Pam)G (oy; 8; Ry) (4.209)
i=0
236 Coherent Communications with Waveforms Chap. 4

The MAP receiver decides on that message with maximum a posteriori probability, that is

P (mi) fenOn/mi) (4.210)

ft a more rss I'v
=
P(mi) onan TREE exp {-3(0w —s))’Ry (ey — si)|
z a
Pek P(m)) aan TR exp {-2(0n — $j)"
Ry (On — s;)|
Taking the limit to improve the accuracy of the approximation and simplifying, (4.210)
becomes

Pn/7ToO) =) tt) Sh eee

Pompesp{eQRy'si— 38) Ry'sif (4.211)
N00. SH} Plmy)exp |pyRysi— 387 Ry Si
In arriving at (4.211), we used the fact that PERS: is a scalar and equal to its transpose
s!Ry'p nv as (Rey = Rie due to the symmetry property of the correlation matrix Ry.
Note that the denominator is a constant C(N) independent of 7. If the limit exists and is
finite, then

1
lim
im pURvs
Pn vs = iim
li (4tpyf) cepts
Apne 4,
(4.212)

Define

be Ey
ge Ay en Si (4.213)
or

sj = RygiAt (4.214)

Hence, the j’” element of s; is given by

N
si(jAt) = > Ra At, KAN gi(kAt)At, forall j=1,2,---,N (4.215)
k=)

By continuity of s;(t) and Ry(t, s), we write in the limit as N > oo, s;(j At) > s;(t) and

si(t) = ieRy(t, s)gi(s)ds, Wte[0, T] (4.216)

and g;(t) is determined by solving the integral equation. From (4.213), we find

lim pny Si = slim (At) 08

N

= ene plkAt)gi(kAt)At (4.217)

Tf

= / r(s)gi(s)ds
0
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 237

Similarly

yt = in ade
: Tp-l ‘

T
ii

(4.218)
= / si(s)gi(s)ds
0
Thus, the MAP test reduces to the following: select m(o(t)) = m; iff

P(mj/r(t) = p(t) = is {P(mx/r(t) = p(t))} (4.219)

i ie
= max {o-'Pom) exp {| r(t)gx(t)dt — a / saci} |
k 0 2 Jo
WiherenGa——l My =a) © (LV). Assuming that C exists in the limit, we have derived the opti-
mum receiver. The convergence of C will be treated later when dealing with the K-L expan-
sion. Simplifying (4.219) by taking logarithm, the decision rule becomes: set m(p(t)) = mj
iff
E i
i r(t)gj(t)dt + B; = max {| r(t)eg(t)dt + a,| (4.220)
0 0
where
1 i
B, =1n P(m,) — >| SK(t) gx (t)dt (4.221)
0
and g;(t) is the solution to the integral equation given by (4.216). This test reduces to the
mechanization given in Fig. 4.34. In the special case of white noise, Rp(t, s) = (No/2)d(t —
s) and g;(t) = (2/No)s;(t). The bias B, in the presence of AWGN is related to By of (4.39)
through By = (2/No)By and the receiver reduces to that in Fig. 4.6 as expected.

4.2.2 Karhunen-Loeéve (K-L) Expansion Approach

Let {@;(t), i =1,2,---, co} be a sequence of orthonormal functions defined on the interval
[0, 7]. Assume that the set {s;(t), i =0, 1,---, M — 1} of known signals can be expanded
in terms of the @;(t)’s, that is
[o,@)

HO =)_ siyejG), 1=0,1,-°,M=—1 (4.222)

y=!

where
th
Sij a} Si(t)Pj(t)dt (4.223)
0

Here the definition of orthonormality implies

T aie ‘

i!bi(doj(tdr =| Pee (4.224)

0 , =
238 Coherent Communications with Waveforms Chap. 4

DECISION

AOmMOD>Yrr

(see note)

Note: From here on, “choose largest" implies choose

message corresponding to largest input voltage or value.

Figure 4.34 Waveform receiver in colored Gaussian noise (derived using sampling approach)

The Gaussian noise random process n(t) can also be expanded in terms of the basis func-
tions to yield

n(t) = > njgj(0) (4.225)

j=l

where
th
nj af n(t)pj(t)dt (4.226)

are random variables. For these expansions or sequences to converge, we require that
‘ 2 8,

lim, |si(t) — 2/850 =() (4.227)

for the signal and

N 2

limE } |n() — Ynjgj()| }=0 (4.228)

j=l
for the noise. The latter is referred to as mean-squared convergence. Let r = (r1,72,°-° >
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 239

rn) where

if
rj= | r(t)pj(t)dt (4.229)
0

and

Pp sipnj jf=1,2)---yN (4.230)

when m; is transmitted. The random variables n;, i = 1,2,---, N are Gaussian (since they
are the result of a linear operation on a Gaussian random process) with
if
E{nj}=E {[ no, ar|=) (4.231)
0

since n(t) is zero-mean and

ie
Enna AG = {| / n(e)bi(On(s)0(s)ards}
Tey
ee (4.232)
=/| ifPi(t)Rn(t, s)Pi(s)dtds
0 Jo
The covariances can be thought of as the coefficients in a double orthonormal function
expansion of
CO coo

Rn(t,s) = Y) Y_ di, ji (1)b)(s) (4.233)

i=! j=l

If we set

Aji «> - AIN

Rn = Si ce (4.234)
Ani «+: ANN
then, when m; is transmitted

E{ry/mi} = Sj (4.235)
where s! = (S;], Si2,°**, Sin). Hence, from (4.210), the a posteriori probability P(m;/rjy)
1S

| \Tp-! rar
Pag
b P(mi) exp {3 (oy — Si) Ry (On si)|
cecee
P(m;j/t¥n = py) =
Ne e ON , (4.986)
Ben P(mj) exp {-3(en — sj)
Ry’
(ey — s;)|

To eliminate the problem associated with Big: as N — oo, we use a special case of $j (t)’s,
which render all Ry to be invertible, that is, let A; and ¢;(t) (eigenvalue and eigenfunction)
be solutions to
T
AiPi(t) = / R(t, s)Pils)ds (4.237)
0
240 Coherent Communications with Waveforms — Chap. 4

where A; ; = A;4;;. Thus, the sample function expansion of the noise n(t) = ee njQj(t) is
the Karhunen-Loéve (K-L) expansion. This implies that the n;’s are uncorrelated and
he 708
i,j =Ai9i,j = / i dj (t) Rn(t, S)pj(s)dsdt (4.238)
0 JO
In this case, (4.233) reduces to
CO

Rn(t, 8) =) AiPi(t) ils) (4.239)

i=]
resulting in Mercer’s theorem [13]. Thus, for a fixed NV, we have
Mace TO jie 0 0
Oe ote 0 ee HER Bs
(Pe) (aa : or Ry = ; (4.240)
: bate & ; ee 7
OV Se Oe oN 0 eal (Oras

and therefore (4.236) becomes

P(mj;) exp {ei hee (ris, = 3s2,)|

ens
N->00 hts ee pare
narTEPog)NON PEPSI
exp { MST= 3s2,)}Vee
09212) (rst

The denominator is independent of i and is a constant, say C ~! Hence

Suen (vis - 33)

lim P (mj/ty) = CP (mi) exp mee 13 (4.242)
j=

or
(o,)

In P(m;/r(t))
e =InC +InP(m)
OT + YOa5! |EE
ryjsij eee
ag (4.243)
J=

for alli =0,1,---, M — 1. Therefore, the optimum receiver sets m(r(t)) = m; iff
CO CO

Ba + Bi = max perma + | (4.244)

j=l j=l
where
1 (oe)

By = In P(mi) — 5 aes (4.245)

j=l

represents a bias term. In the presence of AWGN, Aj; = Ay Vj and B/ reduces to B, of

(4.221). This decision rule admits the interpretation depicted in Fig. 4.35 in terms of sig-
nal processing. Implementation of the optimum receiver using the K-L expansion requires
knowledge of the eigenfunctions and eigenvalues, that is, #;(t) and A; for all i. These are
obtained by solving the integral equation given by (4.237). The procedure for doing so in-
volves differentiating the integral equation to transform it into a differential equation which
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 241

SIGNAL
DECISION

AQOMODY>-
MHOOLO

Figure 4.35 Waveform receiver in colored Gaussian noise (derived using K-L expansion)

is easier to solve. The eigenvalues are solved for simultaneously by imposing certain con-
straints on the form of the solution of the differential equation (see Problem 4.31).

Interconnection of K-L and sampling approaches. __Let’s interconnect the K-

L approach to the sampling approach derived in the previous section. Note that
242 Coherent Communications with Waveforms Chap. 4

ed) ee) if
See r(0)bj(Odt
=I j=l (4.246)
T
=i! r(t)gi(t)dt
0

where
oO

Bi(t) A DA; sii) (4.247)

j=l

Multiplying both sides of (4.247) by Ry(t, s) and integrating, we have

T 00 T
/ Rnlt, s)gi(t)dt = DOR a (/ Ra: s)6,(0ar
0) rae
J= 0

= 047!sj(AiGj)) (4.248)
7

Y- sijj(s)
j=l
= si(5)
Recall g;(t) in the sampling approach defined through (4.216). Thus, the K-L and the
sampling approaches are equivalent if

gi (t) = gi(t) (4.249)

Convergence of receiver. Let’s now focus on the convergence of the receiver

and more specifically, on the convergence of the constant C. First note that the convergence
of limp +o yoaus nigj(t) to n(t) is guaranteed by the K-L theorem. The convergence of the
signal expansion limy-+o0 ae ij; (t) to s;(t) is guaranteed whenever the signals are of
finite energy, that is
€
/ s*(t)dt < 00 (4.250)
0
or equivalently

N
Rae
2
Sip < (4.251)

The convergence of the random sums computed in the receiver is handled in the following
way; let Sy denote

N
SrA Aten (4.252)
i=
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 243

This sequence of random variables converges in the mean-square sense as N —> oo when-
ever

lim E{|Sy — Sv+m?| =0 for any

N->0oo
M>0 (4.253)

This is the Cauchy convergence criterion [14] for random sequences. Suppose that m, is
transmitted. Then, the r;’s are independent Gaussian random variables with
T
rj =| r(t)oj(t)dt = sj +n; (4.254)

and

E{r;/m} = sxj and Var{rj} =A; (4.255)

Thus
N+M 2
E {Sv = Sveml? /me| =E Aj rjSij /mk
j=N+1
N+M 2
= 2. my Ae (sey + ny)54 /Mmz (4.256)
j=N+1
2 2

co j=an+1 djJ jJ=N+1

rj4
Suppose we number the eigenfunctions such that A; > A2 > -- -. Noting that
P ie
Average noise power = E tf ncoar} = (| Rn(t, t)dt
0 0
oe)
(4.257)
=p
‘= _

then for finite noise power in the observation interval, we require that

lim’, = 10 (4.258)
Noo

For convergence of original sum in (4.252), we require from (4.256) that

N+M 52
lim Le —/ —0 for any M (4.259)
iliac TS | j

Since the denominator elements are going to zero, we not only require that

lim Sipe 0 (4.260)

j7oo

but additionally, we require

2
Siar.

lim — =0 (4.261)
244 Coherent Communications with Waveforms Chap. 4

If the above is satisfied for all 7, then it is guaranteed that the remaining term in (4.256) also
converges because of Schwartz’s inequality, that is

M NtM 52 1/2 /N+M ,2\ 1/2

Ski Sij <(ys 4
J
(2 1)
Uy
(4.262)
j=N+41 Aj
In summary for the sums to be convergent, we require three conditions, namely

(a) signal waveform energy to be finite, that is

18
i s?(t)dt < 00
0

(b) noise waveform energy to be finite, that is

iD
E {[ roar <0o
0

(c) let Fe/A; be the signal energy-to-noise ratio for signal i in the direction of $;(t), then

=
lim Sj < oo for alli
is oF = 1 Ay

When condition (c) is not satisfied, we arrive at the case of singular detection in which there
exists at least one direction or dimension that is noise free (that is, zero eigenvalue). In this
case, perfect detection is possible provided that the projections of the various transmitted
signals onto that coordinate are distinct, that is, unequal. If the projections are not distinct,
the dimensionality of the problem can then be reduced by the number of zero eigenvalues
and the problem reworked in a less dimensional vector space.

Receiver truncation and theorem of irrelevance. Recall in additive Gauss-

ian noise, the optimum receiver sets m(o(t)) = m; iff

Sey Lisi;
wet + B = max yj LhjSkj
sp: + Byk (4.263)
=

where B;’ is given by (4.245) and the rule results in the functional block discussed pre-
viously in Fig. 4.35. Assume now that the transmitted signals are designed such that they
occupy the first NV dimensions, that is
N
s(t) = x sijj(t) (4.264)
j=l
T ? : :
then r; = fo r(t)b;(t)dt = s;; +-n;, where the last equality assumes that m; is transmitted.
Decompose the vector r into

p= (tigre) (4.265)
where

rf =(r1,72, °°, tN) (4.266)

Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 245

and

Bara TING 28% -, ood (4.267)

where r; =n; for j> N because sij =O for j > N. Thus

r= >, 174;(0)
j=l

N ee)
= DrigiO+ Di 1joj@ Oe
j=l j=N+1
—| ———4”

depends oni does not depend on i

and rp is irrelevant

Since r2 is independent of i and is uncorrelated with rj, it is irrelevant to the detection

process. Mathematically, we have

IAs) R= /)) =
_ P(m) fe(p/mi)
Hasan fr(p) (4.269)
= P(mj) fry,r. (P15 P2/m;)

fr(p)

Since rj), r2 are uncorrelated, then

P(m;/r = p) = (mi) fr; (p,/mi) fr. (P2/mi) (4.270)

fr(p)

But fr,(p2/mi) = fr.(@2) independent of i; hence

P ? h
P(m;/r
= p)= (mj) fr, (@1/Mi) fr. (P2) (4.271)
fr(p)

Therefore, optimum receiver sets m(p(t)) = m; whenever

P(mi)
fr,(p,/mi) = max { P(mk) fr, (91 /mx)} (4.272)

or alternatively iff
N N
| yi eB ee max EG rjSkj + mt (4.273)
j= j=

Thus, the minimum number of correlators required to implement optimum receivers is N as

shown in Fig. 4.36.
In the special case of AWGN with R,(t, s) = (No/2)6(t — s), any orthonormal set
of N waveforms can be used in the receiver resulting in A; = No/2 for all j. Hence, the
optimum receiver sets m(p(t)) = m; iff

2
raat = VrjSij + B;" = max Fe
2~<
5 a B, (4.274)
No = k | No Pa
246 Coherent Communications with Waveforms Chap. 4

SIGNAL
MATRIX

mnooro

DECISION

AOMODDYr

Figure 4.36 Waveform receiver truncation in colored Gaussian noise

In this case
N T
1 Seis te os / r(t)s;(t)dt for alli (4.275)
jel
and ie Side AS Si; = E;. Hence, the receiver reduces to selecting m(p(t)) = m; iff
it T
iEr(t)s;(t)dt + B; = ne {| r(t)sp(t)dt + Br (4.276)
0 0
where By = (No/2) In P(m,) — Ex/2 as given in (4.39).

4.2.3 Whitening Approach

Since the optimum MAP structure of Fig. 4.36 whitens the received signal by projecting it
on the appropriate eigenfunctions, let’s rederive the receiver by imposing a whitening filter
up front.
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 247

DECISION

Figure 4.37 Antipodal receiver in colored Gaussian noise

Infinite observation interval. Let’s first consider the detection of antipodal

binary signals in colored wide sense stationary (WSS) Gaussian noise assuming infinite
(—oo, 00) observation interval. Consider the structure of Fig. 4.37 where the received signal
is first filtered and then sampled at the end of the signal duration. The idea is to whiten the
noise and then perform matched filtering on the new signal. The filter H(f) combines both
the whitening and the matched filtering operations. Recall that for binary signaling in white
Gaussian noise, the bit error probability is given by (4.51) as

P(E)=Q & (4.277)

OAq

where Ago is the mean conditioned on mo and oaq the noise variance at the output of
the matched filter. Since the noise at the output of H(f) is Gaussian in Fig. 4.37, (4.277)
is still applicable with the appropriate mean and variance. Since the output Aq(t) =
Sor@)h(t — t)dt, then

Aqo= iGir / SH (fe af (4.278)

and |

Kg =i! Sn fH (PP f (4.279)

where S,(f) is the PSD of n(t) and S(f) the Fourier transform of the signaling pulse s(t).
Hence
My Jian (pe Tt df (4.280)
AA tae eS
AAMAS
CO AO) patbd)
Defining

Giff) Z A(f)v Sf) (4.281)

Go(f) 4 Bi ere: (4.282)

S.Ct
then
ER GIGsS af ao
se
SIG neat)! : 1/2

Using Schwartz’s inequality, that is

ph ee) (oe)

i Gi(f)Go2(f)df < | Gif) Pdf i IGo(fyPdf (4.284)

248 Coherent Communications with Waveforms Chap. 4

we have
oa a) 1/2

A< (
/ mats ar) (4.285)

with equality if and only if Gj(f) = cG3(f) (c any constant) or

Spier =
H (Av Sh) Sn(f) = JE ~ ((4.286)

or equivalently
K( f\p—J2ufT
Wf) = Aelia an (4.287)
Sn(f)
This is the optimal filter for antipodal signals with colored Gaussian noise with PSD
Sn(f). The filter can be decomposed as a series of two filters, that is

BOY =MQOibg (4.288)

where
1
H
Le Seeaaee
=
Gere
4.289

and
Silent
Hof) = (4.290)
V Sn(f)
The filter Hj(f) whitens the noise? and the filter H2( Ff) is matched to the “whitened”
signals as shown in Fig. 4.38.
Assuming optimal filtering, the bit error probability is given by P(E) = Q(Amax)
where

TSC
Nor a
i Sf) 2 d
Gare
4.29]

Hence, P(E) is now dependent on the choice of S(f), that is, it depends on the form of
the actual signaling pulse s(t), not simply its energy E as in AWGN channels. Then, what

9. When S,,(f) is a rational function, then it can be expressed as

Whee 17)
Sn(f) =c 2
I]jai t = Pj)

where z; and p; are the zeros and the poles, respectively. Since S,(f) is real and nonnegative, then the complex
roots of S,(f) have to occur in complex conjugate pairs and the real roots occur with even order. In this case, the
whitening filter can be constructed using

cl/2 enes —Zj)

A(f)=
ma (f — Pj)

where Im(z;) = 0, 1 <i <n and Im(p;) > 0, 1 < j <m, that is, H\(f) is formed with the right-hand side zeros
and poles of S,(f). In this case, H;(f) is guaranteed to be a causal filter.
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 249

Hy (f) Ho(f)

sre! oT
V Sn (f)

WHITENING MATCHED
FILTER FILTER

Figure 4.38 Decomposition of H(f) into a whitening and a matched filter

would be the optimal signal shape to transmit for a given channel S,(f) or equivalently,
what is the choice of s(t) that maximizes A max? That is

max if ISP af subject to ee (SC rap == (4.292)

S@) JZoo Sn(f) —oo

Consider expanding the received signal r(t) using a K-L expansion as in the previous sec-
tion (remember that only the first N components are sufficient for detection), that is
N
ri(t) =) ri@i(t) and rj = sei +7: (4.293)
i=l
assuming my, is transmitted with

E{njnj}= 1
ce Rr ew,
oe (4.294)

Then, r(t) can be mapped into the vector r; and the whitening operation can be represented
by the equivalent vector transformation!?
S12
di 0) Ses @)

—1/2
2 ot eras (4.295)
nya
0 ip ees
since

hee ak 0 A 0 0
—1/2 S 0 X

; @) \ —1/2 0 Or ty N
ae (4.296)
iS 0 0
=

a e 0
(RS aaa ie ei
10. r},, stands for the whitened received vector and is given by rj» = Sw + Mw.
250 Coherent Communications with Waveforms Chap. 4

and the noise components of rj, are actually white. The signal component Of Lig. Sus
becomes
eg RO Pies? | 0
—1/2
of le ee 5 (4.297)

0 Osan
and
Nae foe) KY iz
Sa =f side = | ST af (4.298)
a Ki —0oo —co Sn(f)

where the first equality is due to the fact that a signal and its expansion have the same
energy and the second to Parseval’s theorem [15]. The optimization problem can therefore
be restated in terms of the vector components as
Ne‘Se N
max
<
) i subject to y s? = iE (4.299)
i=l “" =I)

Let Amin denote the smallest eigenvalue, then

Ns: NTS: E
Dee = (4.300)

with exact equality if we place all the signal energy EF in the dimension with the smallest
noise variance Amin. Therefore, the optimum performance is obtained by transmitting all
the signal power in that coordinate where the noise is weakest. Thus, the optimal antipodal
signal pair in nonwhite Gaussian noise is

+s(t) =+V/Epdmin(t) (4.301)

where ¢min(t) is the eigenfunction of the K-L expansion of the noise having the smallest
eigenvalue. Here again, perfect detection (that is, P(E) =0) can be achieved if Amin = 0
(case of singular detection).

Finite observation interval. When the noise is still WSS but the observation in-
terval is finite, the derivation of the optimum receiver is not as straightforward and requires
solving integral equations. Let (0, 7) denote the observation interval, then the output of the
whitening filter in Fig. 4.39, when message m; is transmitted, becomes (From here on, we
will let H(f) denote the whitening filter. Note that in Figures 4.37 and 4.38, H(f) denoted
the combination of whitening and matched filters)
1
Ha) = i!hG@ — t)r(@)dt=syi@) +N) (4.302)

where
t
Sw i(t) = Si(t) * h(t) =i) si(t)h(t — t)dt (4.303)
0
je
Ny(t) =n(t) *h(t) = / n(t)h(t — t)dt (4.304)
0
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 251

WHITENING
FILTER

Figure 4.39 Whitening filter

Since the noise at the output of the filter, n,,(t), is white, the optimum receiver for equally
likely messages computes
FE Jé
InA(ry(t)/m;) = i!ry (t)Sw,i(t)dt — ;/ s2,,(t)dt (4.305)
0 0

for every transmitted signal and decides on the message with the largest In(r,,(t)/mj;) as
shown in Fig. 4.40(a).!! Substituting for sy j(t) and n»(t), we have
Tig ept i
In A(ry(t)/mi) =i) {| A(t — or(eide | A(t — r)si(e'yar| dt
0 0 0
(4.306)
1 ort et ji
oe > if {| h(t — dsi(eydr [ h(t — r)sirar'} dt
2Jo (Jo 0
Defining!”
7
OG) af h(t —t)h(t — t’)dt (4.307)
0
then (4.306) simplifies to
ToT.
In A(ro(t)/m) = f / r(t) Q(t, t’)sj(t’)dtdt’
ie
1 T af
(4.308)
_ =f | 5i(t) O(c, t')8;(t’)dtdt’
2Jo Jo
and can be implemented as shown in Fig. 4.40(b) with
1 fhe T
Cc; A -; f i 5i(t) Q(t, t’)sj(t)dtdt’ (4.309)
2Jo Jo
Defining
fi
gi(T) af Q(t, t')sj(t')dt (4.310)
0
11. Ey,; denotes the energy of the i‘? whitened signal, that is, Ey; 4 i s2 (t)dt.
12. Note that with a simple change of variables u = t — t and when the observation interval becomes (—00, 00),
then

OG )= i h(u+t' —t)h(u)du A Q(t — 1’)

252 Coherent Communications with Waveforms Chap. 4

(a)

DECISION
WHITENING
FILTER
r(t) ty (t)

MHOOLZO
HOmMODrYrr-

(b)

DECISION

mnoozro
4Homowvrr

Figure 4.40a, b Alternate structures for colored noise receiver (derived using the whitening approach)
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 253

DECISION

Figure 4.40c Alternate structures for

colored noise receiver (derived using the
(c) whitening approach)

then (4.308) becomes

ii
In A(ry(t)/m;) = i r(t)gi(t)dt + B: (4.311)
0

with
1 i
B= -; f si(t)gi(t)dt (4.312)
0

which is identical to (4.221) for equilikely signals. This suggests the alternative implemen-
tation shown in Fig. 4.40(c). The computation of the functions g;(t),i=0,1,---,M—1
require solving (4.310) which depend on Q(t, t’) which is a function of the whitening filter
h(t). The g;(t) functions can also be expressed directly in terms of the transmitted signals
and noise autocorrelation function R,,(t) as follows: since n,,(t) is white, then
ifs Vf
E{ny(t)nw(s)} =| | Rae =F AE = Th HT dt az”
0 0 (4.313)
= d(t —s)

Multiplying the right-hand side by h(t — rT) and integrating over r, then
i a ie
/ b(t —s)h(t —t)dt et / RAG — tt")
0 0 JO
(4.314)
T
x i)A(t — vt)h(t — t)dth(s — t")dt'dt”
0

or
Paint
h(s =) =a) ih Rie’ —t JOG, te)h(s —t")dt'dt” (4.315)
0 0
254 Coherent Communications with Waveforms Chap. 4

The above equation is satisfied only if

i
/ Rie =a O(a dt =s Ge FP) (4.316)
0
This equation relates Q(t, t) directly to the noise autocorrelation function R,(t). Recall
from (4.310) that g;(t) = ee Q(t, t’)s;(t’)dt’. Multiply both sides by R,(t — tT) and inte-
grate over T to get
T Tam
i gi(T)Rn(t — t)dt a / O(t, t’)Ry(t — t)sj(t’)dt'dt (4.317)
0 0 0

Using (4.316), then

T i
i;gi(t)Rna(t — t)dt = ) 8¢ = 7s lar (4.318)
0 0
or

F
si(t) =i) gi(t)Rn(t — t)dt (4.319)
0

This is an integral equation in g;(t) as a function of the observation interval (0, 7), the
signal s;(t), and the noise autocorrelation function R,(t). Note that when the observation
interval is (—oo, oo), then s;(t) = g;(t) * R,(t) and G;(f) = S;(f)/S,(f), which is the
filter obtained earlier!? when discussing antipodal signals.
Example 4.4
Consider the binary communication problem over the colored Gaussian noise channel with
Sn Gi (No/2)(f2 at 1)/(f? + 4). Derive the optimum receiver assuming infinite observation
interval and antipodal signals

s(t) = so(t) = —s1(t) = VER it 172

0, otherwise

From Fig. 4.40(c), we know that the optimum decision rule is

lo) mo
i r(t)g(t) = 0
co my

where

S(f)
G
Sao— =

In this case, S,(f)= meri

Bu and hence

2 3
G(f) = are |S
Since the Fourier transform of e~@!"! is es then

at) =ees
5 [si a ue —2n|t| +(0)|
13. The filter obtained earlier is denoted by H(f) and given by (4.287) as H(f) = S*(f)e/2"/T /S,(f). The
filter h(t) and the pulse g(t) are related through h(t) = g*(T — 1) since h(t) is actually the filter matched to g(r).
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 255

1
t- 3 t t +5 0 nN,

(a) ts-1
2

t eg5 t LEIS, t+ feel

5 ny

1 1
(b) - 3 <t £5

1 1
Ouh= 5 tee 5 r

(c) t oe, Figure 4.41 Computation of f(t) in

2 Example 4.4

or

2 Az!
eO= = [so 16x VES]
where f(t) = fi(t) * fo(t) and

fae rel
Leases
fat) =
Q otherwise

In order to compute f(t), we need to consider three cases as shown in Fig. 4.41, namely
For
t < —1/2
256 Coherent Communications with Waveforms Chap. 4

cil t Figure 4.42 The g(t) function in Example

2 rm 4.4

1/2 :

1-1/2 a

For —1/2
<t < 1/2
0 t+1/2 1 e TE

f(t) =| e™dy +f e**qd) = = = ——cosh(2z7#)

t—1/2 0 RE 4

Fort = 1/2

,o= {Pe ee
fale a

Therefore

a VE sinh (30)?! (23172

g(t)=} q2VElT — 6e-"cosh2nt] —1/2<1<1/2
E sinh (1)e~7" t>1/2
Note that

e” VE ovE
&G) =9(——)
= g( == 126 (ec
er ae No + oon
6—

- VE
0) == ((14
g(0) — 12e7”) — ~ 13.5—
ar) No No

The function g(t) is plotted in Fig. 4.42 as a function of f. Since it rolls off exponentially, the
function can be truncated when it is implemented in a receiver. A “good” engineering judgment
would be to neglect (No/WE)g(t) when it is less than 0.1 (which is less than 1% of its peak
value of 13.5), that is

1 0.1
12 sinh (x) e602" #!.<'0.1 oe = ——1 ~ 1.15
Memes "1 = 5 ln sinh Gr) ee
Therefore, the receiver for all practical purposes observes r(t) over the interval

SSeS

Recall that the optimum receiver requires solving for g;(t) which is given by the
integral equation in (4.319). No attempt is made to derive the solutions to these equations,
but the key results are given and the reader is referred to Van Trees [13] or Helstrom
[16] for additional details. When the observation interval is finite, the noise autocorrelation
Rpt, t2) = E{n(t))n(t2)} can be classified as either rational or separable. In the rational
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 257

case, Ry(ti, 12) = Rn(t) — t2) = Rn(t) and assuming that the noise n(t) can be written as
the sum of two independent noises, that is, n(t) = n,(t) +n¢(t) to separate out the white
from the colored noise contribution, the PSD becomes

No
Sn(f) = o + Sc(f) (4.320)

where S-(f) is the PSD of n,(t) and is given by the ratio of two polynomials

N (f?)
Sf)
4 —— (4.321)
ee)
The numerator is a polynomial of order 2n and the denominator a polynomial of order 2m.
We assume that m — n > | so that n-(t) (colored noise contribution) has finite power. When
No = 0, that is, no white component, (4.319) reduces to
x
5;(t) =i) R(t — t)gj(t)dt (4.322)
0

which is a Fredholm equation of the first kind. The general solution is given by

2n m—n—1 k
dk d
Bi(t) = Boo(t) + 2 aight + aE, eo + G5 - r)| (4.323)
where 2.0(t) is the solution to the integral equation assuming an infinite observation interval
(obtained using Fourier transform techniques), gp.)(t) is the /‘” homogeneous solution to!*
(let p 4 d/dt)

O=N(—p’)gni(t), 0=1,2,---,2n (4.324)

The various constants are solved by substituting g;(t) back into the integral equations and
equating terms.
When the white noise component is nonzero (that is, No # 0), (4.319) reduces to

No uf
si(t) = 5 8) +f Re(t — t)gi(t)dt (4.325)
0

which is a Fredholm equation of the second kind. Its general solution is given by
2n

gilt) = Boolt) + >) agni(t) (4.326)

ill

where the various constants are again obtained by equating terms.

As a final category when the noise is not stationary, assume that the kernel of the
colored noise is separable, that is
K
Re(ti, 2) = > Adi(tigilt), OS, <T (4.327)
i=!

14. The equation is written in operator’s notation.

258 Coherent Communications with Waveforms Chap. 4

Table 4.3 Colored Noise Receiver Solutions

SOLUTION
OBSERVATION WHITE NOISE GIVEN BY
INTERVAL KERNEL COMPONENT TECHNIQUE EQUATION

INFINITE Fi oh FOURIER TRANSFORM (4.287)

FINITE SEPARABLE hee EIGENFUNCTIONS (4.328)

FREDHOLM EQUATION 4.326

RATIONAL PRESENT | OF THE SECOND KIND
FREDHOLM EQUATION
ABSENT OF THE FIRST KIND aces)

where A; and ¢;(t) are the eigenvalues and eigenfunctions of R-(t1, tz), then it can be shown
that
2 K ahi
gi(t) = No [si@) ~ Lint ENo/2% | ms as 1 (4.328)
0 elsewhere

where s; = ie si(t)@;(t)dt. Note that if we had allowed K = oo, all kernels would be
considered separable since we can always write

Re(ti, t2) =) AiPi(ti)di(t2), O<t, 2 <T (4.329)

i=l
All the techniques are summarized in Table 4.3. The reader is referred to references on
integral equations for additional information.
Example 4.5
Consider detecting the presence or absence of a signal s(t), te(0, 7) when the observed wave-
form is

mi >r(t)=s(t)+n(t) te(O, T)

my —>r(t)=n(t) te(O, T)

and n(t) is a zero-mean Gaussian noise with R,,(t1, t2) = exp(—a|t, — t2|) for all t, tre (0, T).
Derive the decision statistic A.
Since R,(t) = exp(—a@|t|), the PSD of the noise becomes

2a
Sn(f) 3 a2 + (nf)?

The function Goo(f) is given by the equation

ee3,(f) -(* + 5AQxf)

§ 2 2
S(f)
Sec. 4.2 Optimum Waveform Receivers in Colored Gaussian Noise 259

or

Boolt)th=—
== [ars(t) ~ -5")]
1 2 y

Since n = 0 and m = 1, (4.323) becomes

8(t) = Zool(t) + bod(t) + cod(t — T)

Substituting back in (4.322), one obtains
wh

= i
0
exp(—a|t — t|)q(t)dt = bye! + coe”)
] t

sara i
0
en [o%s(r) ~ s(t) |dt
] T

? eee
=e" |
as2a [o*s(r) ~ s"(r)| dt
Integrating the terms containing s(t) twice by parts, we get
1
s(t) = bee™ + coe®"—)) — 5, fe“las(0) — s'O)]
a
+ eT [as(T) +.s'(T)] — 2as(t) }
In order for this equation to hold, the coefficients of e®’ and e~® must vanish. Hence, by and cg
must satisfy

bo = [as (0) — s'(0)]/2a

co = las(T) + 8'(T)]/2a
The decision statistic becomes

N= a { [as (0) — s’(0)]r (0) + [as(T) + s'(T)]r(T)

1h

of if[a2s(t) — s(t) |r(t)dt }

0

4.2.4 Sufficient Statistic Approach

We rederive the optimum receiver in the presence of colored Gaussian noise using a fourth
separate approach. !> A received waveform r(t) is observed on a finite interval (0, 7) and
is known to have either the form!® r(t) = n(t) (message mo) or r(t) = s(t) + n(t) (mes-
sage mj) and s(t) is a known real function and n(t) is a zero-mean Gaussian noise with
autocorrelation R,(t, tT). Consider those signals s(t) that can be represented as
ih
s(t) =i] Kyi, teat, O=t=<T (4.330)
0

Consider an arbitrary set of orthonormal functions &)(t), €2(t), --- on the interval (0, T).
Add s(t) to the set. The augmented set {s(t), &1(t), &2(t), ---} is still complete and spans the
same space as the original set. Orthonormalize the augmented set using the Gram-Schmidt

15. This approach is in several references [13, 16, 17]. We follow the paper presented by Brown [18] in 1987.
16. The more general binary detection problem (mo — r(t) = so(t) + n(t) and m; > r(t) =s\(t) +n(t)) can
be handled by defining r’(t) = r(t) — 1 (t) since s\(t) is a known signal and s(t) = so(t) — 51 (t).
260 Coherent Communications with Waveforms Chap. 4

procedure outlined in Appendix 3C to obtain {x;(t) = as(t), x2(t),--:} where xXn(t) is

orthogonal to s(t) for n > 2 and a is a constant to normalize x(t). Next, replace x(t)
by g(t) and call the resulting set {w)(t) = g(t), Wo(t) = x2(t), ---}. Note that y,(f) is still
orthogonal to s(t) for n > 2. The altered sequence is still complete since s(t) and g(t) are
not orthogonal since
T
its(t)g(t)dt =0 » (4.331)
0

would imply that [using (4.330)]

bar
/ / g(t)Rr(t, t)g(t)dtdt =0 (4.332)
0 0

contradicting the positive definite property of R,(t, T) as an autocorrelation function. Pro-

jecting r(t) onto the set {yr (t), w2(t), ---} results in
iA
rj af rOwi@at, T= 192, (4.333)
0

Note that since the set is complete, knowledge of the r;’s allows approximation of r(t) in
the mean-squared error sense with arbitrary accuracy. For j > 2, we have
V4
E{rirj;/mo} = elf neowj(ode f n(a)e(r)dr|
. (4.334)
=f i Ra(t, e(eyj (drat = | s(t)yj()dt =

since yj (t) is orthogonal to s(t) for j> 2. Similarly

T pT
E{rirj/m} =| [ S(t)s(t)Wj(t)g(t)dtdt
ae (4.335)
+/ / Ra(t, Dg(t)pj)dtdt
=0, j>=2
0 JO
Also

E{r;/mo} = elf ntoryj(oatl =0, forall j > 1 (4.336)

and .

E{rj/m} = e{f (s(t) + nconwj(odr} =, [ze (4.337)

so

E{r;/m,}
= E{r;/mo} =0,- j >2 (4.338)
Next consider the cross-covariance terms of the sequence {r2, r3, ---} when either mo orm]
is transmitted:
Tare
Etriry/mo) = [ [ Ralt, tWit)wj(t)dtdt, i,j >2 (4.339)
 Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 261

10. Sh
E(rirj/my}=E |[0 [JO sosmvtovjeacar
[ey th
+ i;[ n(n (oil (e)drdr | (4.340)
0 JO
ih ap
=| / Rit, WpiOwj(t)dtdt, i,j >2
0 0)
Hence

E{rjrj/mo}
= E{rirj;/mi}, i,j >2 (4.341)

Therefore, we have shown that the projections {r2, r3, - - -} are statistically independent of ry
when mo or my is transmitted and that the means and cross-correlations among the terms of
the sequence {r2, r3, -- -} are identical under mo and my.
Forming the likelihood ratio test based on the finite set {7}, 72, ---, 7}, we have

ap Miata
sos NyTA) [Aol rg WE Ort m
(4.342)
f (ri, 172,°-+,1nw/mo)
where the joint densities are Gaussian under either hypothesis. Since rj and {r2, r3,---, rv}
are independent under both mo and m, then

_ f£ri/m1)
f(72, 73, +++, 1N/m1)
N= "FOrJo) fra T3. 4 Fw/m0) one
However, the joint densities of {r2,73,---,rn} are multivariate normal with identical
means and cross-covariances independent of which message was transmitted. Hence,
f(r2,173,-++,7n/mo)
= f(r2,73,-++,7nN/m) and
pecs ada) (4.344)
f(ri/mo)
Since Ay is independent of N, the same test is obtained in the limit as N — oo. Hence,
we conclude that rj = tos r(t)g(t)dt is a sufficient statistic for deciding between mo and
m,. This technique can be expanded to M messages and results in the optimum receiver
depicted in Figures 4.34 and 4.40(c) as expected.

4.3 INPHASE AND QUADRATURE MODULATION AND DEMODULATION

So far, we have considered detection of baseband signals represented by s;(t), i =0,1,---,
M — 1. In real communication channels, the signals are transmitted at radio frequencies
(RF) and the receiver downconverts the received RF signal to baseband, using a coherent
carrier reference, before performing any message detection.

4.3.1 Real and Complex Signal Models

In most communication systems, the transmitted signal can in general be represented by

s(t) =P {I(t) cos wet — Q(t) sin wet} (4.345)

where P represents the transmitted signal power, /(t) and Q(t) the inphase and quadrature
baseband data modulations, and w, the carrier frequency. The cosw-t term is typically
262 Coherent Communications with Waveforms Chap. 4

referred to as the inphase (I) carrier and the sinw,t as the quadrature (Q) carrier. As an
example, antipodal signals are constructed by setting Q(t) = 0 and /(t) = +p7(t) for 0 <
t < T. Alternatively, /(t) can be written as

I(t)= D7) axpi(t — kT) (4.346)

k=—oo

where ay = +1 representing the binary random data and p;(t) is the baseband inphase pulse
of duration T seconds. A specific case of antipodal signals is BPSK where
il Osy ssh
= aes 4.347
Pit) ts otherwise ( )

In this case, we have!’

s(t) =<V2P cos(@-et + &), AT <t#<(K4+1)T (4.348)

where 6; = 0 or z radians, depending on the data polarity. Another example is QPSK where
both J(t) =+p/)(t) and Q(t) =+p/)(t) for kT <t < (k + 1)T, independent of each other.
Alternatively, QPSK can be represented by (4.345) with J (rt) given by (4.346) and

Q(t) = ST bepi(t + kT) (4.349)

k=—oo

where by; = +1 representing an independent binary data stream. For kT <t < (k+1)T,
s(t) can be written as

s(t) = V2P cos(wet + Ox) (4.350)

where 0 = tan—! (by /ax) = 1/4, 37/4, 52/4, 77/4 radians.!® This approach can be easily
generalized to accommodate M-PSK signals with /(t) and Q(t) as given by (4.346) and
(4.349) with ay = cos Oy, by = sin Ox, and 0 = (2k + 1)w7/M,k =0,1,---,M—1. A more
general signal model allows for Q(t) to have a pulse po(t) different from p;(t), as depicted
in Fig. 4.43(a).
Another example is M orthogonal signals, which can be obtained from (4.345) by one
of two methods: (1) set Q(t) = O and let 7 (t) = s;(t),i =0,1,---, M —1for0<t<T;or
(2) divide the signal into two groups and let / (t) be any signal from the first group and Q(t)
any signal from the second group. A specific set of M orthogonal signals is M-FSK where

s(0) = V2P cos(wet + im t

+0), OStST (4.351)
and the frequency separation is fj,; = f; — f; between any two signals is 1/2T (indepen-
dent of i and j), which is the minimum separation required for coherent detection.
At the receiver, the observed signal is given by

r(t) = VP {I(t) cos(wet +0) — Q(t) sin(wet + 4)} + n(t) (4.352)

where @ is a random carrier phase that has to be estimated for coherent reception and n(t) is
a bandpass Gaussian noise process as described in Appendix 4B. Assuming that the receiver

17. Since Q(t) = 0, we replace JP by /2P to maintain a constant signal power of P.

18. The tan~'(x) is a four quadrant inverse tangent.
Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 263

COHERENT
PHASE
REFERENCE

(b) |-Q demodulator/ signal detector

Figure 4.43 Functional diagram of the I-Q modulator/demodulator using real signals:
(a) I-Q modulator/signal generator, (b) I-Q demodulator/signal detector

generates the coherent references (perfect carrier phase estimate)

r(t) = V2cos(wet + 0) (4.353)

and

ro(t) = —V2sin(wet + 6) (4.354)

9
then ignoring the noise contribution, we have!

r(tyri(t)inp = VP/21(t)
(4.355)
r(t)ro(t)|zp = ¥ P/2Q0(t)
which for kT <t < (k + 1)T, reduce to the js signaling baseband pulse s;(¢), assuming that
m; is transmitted. The corresponding I-Q modulator/demodulator is depicted in Fig. 4.43(b)
using real signals.

19. The notation r(t)r;(t)|z p denotes the lowpass component of r(t)r/(t).

264 Coherent Communications with Waveforms Chap. 4

I(t) +) Q()

i(@ct + 8)
e

(a) Complex |-Q modulator

> met) = It) + att)

-|(@gt + @) Figure 4.44 Functional diagram of the I-

V2'e Q modulator/demodulator using complex
signals: (a) Complex I-Q modulator, (b)
(b) Complex |-Q demoduiator Complex I-Q demodulator

Sometimes, it is more convenient to use complex notation to denote the I-Q signal.
Letting m(t) denote the complex baseband modulation, that is, m(t) = /(t) + j Q(t), then
s(t) can be written as

‘Dav Pe {miei| (4.356)

The process of coherent demodulation consists of multiplying the received signal by the
complex conjugate of the signal (and hence, knowledge of the carrier phase is necessary) as
shown in Fig. 4.44. The reader should interchange the real and complex models as they are
equivalent. In the presence of noise, the demodulated I & Q signals become

r(t)r;()|~p =V P/21(t) + nc(t)

(4.357)
r(t)ro(@t)iltp = V P/2Q(t) + ns(t)
where n-(t) and ns(t) are white Gaussian lowpass processes with one-sided bandwidth
B Hz. The optimum receiver structure derived so far assumed white Gaussian noise with
infinite bandwidth. It can be shown that the derived structures are still optimum provided
that B >> 1/T. One approach to verify the latter is to compute the performance of the
integrate-and-dump receiver in the presence of the lowpass noise n-(t) and show that it
reduces to the ideal performance as BT —> oo.
Another more intuitive approach is to consider the detection when implemented using
matched filters. In an AWGN channel, the matched filter has a bandwidth equal to roughly
1/T Hz, which is approximately the bandwidth of the baseband pulse. When the corrupting
noise is white (that is, infinite bandwidth), the output of the matched filter is corrupted by
noise with nonzero power in the filter band, that is, the bandwidth of its corresponding
power spectral density (PSD) is about 1/7 Hz. On the other hand, when the input noise
is bandlimited to B Hz and B > 1/T, the output of the matched filter is still corrupted by
noise with nonzero power in the filter band only, that is, a bandwidth of 1/7 Hz. Hence, in
either the case of infinite or bandlimited noise, the output of the matched filter is corrupted
Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 265

Table 4.4 Coding Rule (N = 4)

Ge ai Say ek a b w
1 1 i 0 1 -3 W- — 3wa
1 -1 1 | -1 Wc — Wd
1 -1 1 2 1 Wc + Wa
1 -1 -1 3 l 3 We + 304
-1 1 1 0 -1 -3 We — 3@¢
-1 1 -1 1 -1 -1 We — Wd
-1 -1 1 2 -] 1 Wc + @d
-1 -1 -1 3 -1 3 Wc + 3w¢

by similar noises and as a result, the receiver provides the same performance in both cases
as long as B >> 1/T.
Example 4.6
One realization of a biorthogonal set of M signals is a combined frequency/phase modula-
tion (FPM) wherein N = M/2 frequencies uniformly separated by 2 fg = k/2T, (k an integer)
are each BPSK modulated, that is, N-FSK/BPSK. Typically, N = 2” (n integer) and the fre-
quencies are symmetrically placed around a carrier, that is, they are located at fe + fu, fet
3 fa,.--, fe + (N — 1) fa where f. = w,/27 is the carrier frequency in Hertz. Such a signal-
ing set is generated as follows. Consider as a data sequence a binary (+1) random sequence
{dj} of rate 1/7) [equivalently a rectangular pulse data stream d(t)]. Blocks of A = log, M =
1 + log, N bits of duration 7; from sequence {d;} are assigned to FPM signals of duration
T; = XT). We specifically use 7; to denote the duration of a symbol so as to be consistent with
what is done later in Chapter 10 where this example is treated as a special case of a more general
form of modulation. The first bit in each of these blocks is assigned to a sequence {a;} that is
used to define the BPSK portion of the biorthogonal modulation. The remaining A — 1 = logy N
bits in each block are assigned to an N-ary [+1, +3,...,+(N — 1)] sequence {b;} which is
used to define the N-FSK portion of the biorthogonal modulation. The coding rule that maps
{d;} into the binary amplitude sequence {a;} and the N-ary frequency sequence {b;}, respec-
tively, is described as follows.
For a given block of A bits of the input sequence, say {do, dj,...,dn}, the first bit (do) is,
assigned directly to the amplitude level a , that is, a = do. The remaining A — | bits in the block
(di, ..., dy) are represented as an N-ary number, say k, in accordance with the natural mapping
rule which is then associated with the frequency number b = 2k — (N — 1) and results in the
actual transmission of the frequency f. + bfa. Note that if a block of 4 bits corresponds to b,
then its complement (negative) also corresponds to b. An illustration of the above coding rule
for N =4 (M = 8) is given in Table 4.4.
Defining the random pulse trains a(t) and b(t) from the sequences {a;} and {b;}, respectively,
foe)

for example, a(t)= >> ajp(t —iT;) with p(t) a unit rectangular pulse, then an FPM signal
i=—0co
is characterized by

So) V2Pa(t) cos [(@e + b(t)@wa) t]

Using the methods presented in Chapter 2, it is straightforward to show that the equivalent
lowpass two-sided power spectral density of FPM with minimum frequency spacing 1/27; is
given by
 266 Coherent Communications with Waveforms Chap. 4

JB
S(/T,,

Figure 4.45a Power spectral density of

(a) N=1, 2, and4 N-FSK/BPSK,N = 1, 2, and4

dB
S(f/T,,

0.5 1.0 1.5 2.0 Paks 3.0

fT,
Figure 4.45b Power spectral density of
(b) N =8, 16, and 32 N-FSK/BPSK, N = 8, 16, and 32

PI Aan i N-1
(N==z X sinc” (1
S(f) = — | fT;tana
+ = — —— )

(1+log,N) PT, , i N-1

= sinc’ | (1 + logy N) fT) + = — ——

which is illustrated in Figures 4.45a,b for various values of N.

Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 267

4.3.2 Effect of Imperfect Carrier Synchronization

In coherent demodulation, the carrier phase should be known precisely. Assuming that the
references r;(t) and rg(t) are not ideal, that is

AG = V2 cos(wet +6)
sf (4.358)
ro(t) = —V2sin(wet + 6)
where @ is the receiver’s estimate of the incoming carrier phase 6, then (4.355) becomes in
the absence of noise

r(t)ri(@)|zp = ¥ P/2[1(t)
cos @— Q(t) sing] 4 I(t)
(4.359)
r(t)re(t)|zp = ¥ P/2[1(¢) sing + Q(t) cosg]4 Qo(t)
where ¢ = 6 — @ denotes the carrier phase error, which is the phase difference between the
received carrier and that generated locally by the carrier synchronization circuitry. Note that
a nonzero carrier phase error results in cross-talk between the I and Q arms, that is, the
observed I channel, /o(t), is a combination of the desired I and Q channels. Similarly for the
observed Q channel, Qo(t). If 6 = 1/2, then

I(t) = —v P/2Q()
(4.360)
Qo(t) = vyP/21(t)
and the I and Q channels are reversed in addition to a polarity inversion in /o(t). If on the
other hand @ = 7, then

I(t) =—JVP/21()
(4.361)
Qo(t) = —-vV P/2Q(t)
and both J and Q channels are inverted, but not reversed. Using complex notation, this
phenomenon can be easily expressed using

Ip(t) + fQott) = VP/2U1(t) + ja(nle!® (4.362)

which clearly indicates an effective rotation by ¢ radians of the desired baseband I-Q modu-
lation. The performance of the detector in the presence of carrier phase error depends on
both the modulation itself and the model associated with the carrier phase error ¢. This
topic will be addressed in great detail in Volume II of this book when considering synchro-
nization loops. Here, we will consider the simplest modulation (antipodal signals) and the
easiest model of ¢, namely, a random variable over the symbol duration. In this case, it is
straightforward to show that the conditional error probability is given by

P(E/¢)=0 ([= cose) (4.363)

The true measure of degradation is the average probability of error P(E) obtained from

P(E) =i) P(E/) f (bdo (4.364)

268 Coherent Communications with Waveforms Chap. 4

10°?

10°3

Figure 4.46 Error probability performance

of antipodal system in the presence of carrier
E/N, dB sync error

where f(@) is the pdf of the phase error that depends on the carrier synchronization (sync)
algorithm employed. Assuming that f(@) is of the form
exp(og00s) 14) <
f@= | 2m 1o(eg) ” (4.365)
otherwise
where pg 4 Wor then using (4.365) in (4.364), P(E) can be evaluated as a function of
E/No for various values of pg. The results are plotted in Fig. 4.46 illustrating the noisy
reference effect on system performance. For pg = 12 dB at P(E) = 10~°, the required
increase in E’//No is about 2 dB.

4.3.3 Effect of Imperfect Bit Synchronization

Assuming that perfect carrier synchronization is achieved with BPSK signaling, the base-
band received signal in the presence of noise is given by (4.357) with P replaced by 2P
[since O(t) = 0]

r(\ri@)\ep =VPI(t) +ne(t)= Y> apt —(k-T - 8) +n-(t) (4.366)

k=—oo

where 5 denotes the bit epoch which needs to be estimated by the bit synchronization
Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 269

circuitry, a, = +1, and p(t) is the baseband pulse (with power P). 20 Consider the situation
where the bit sync is in error by « = 6 — O=AT sec, where 6 denotes the estimate of the
bit epoch. The detector statistic on which a decision is made for the n" transmitted bit
now depends on the received data in the time interval (n — 1)T +6 <i Sh aN Stated
mathematically

nT+6 (ea
an = sgn / r()ri@)\~p pit — (n — 1)T — djdt (4.367)
(n—1)T+6

where G, is the n“ detected bit. Making the change of variables t/ = t — (n — 1)T —6 in

(4.367) and separating the signal and noise correlation components, we get

Gn = sgn O (4.368)

and
If

ope if P(t’) { an-:p(t' + T — AT) +.anp(t' — AT)

: 1h
(4.369)
+ an4ip(t' — T —AT) } at’ +f pit’)ne(t' + (n — 1)T + 8)dt’
0

We note that in general, three input bits are involved in estimating the n" transmitted bit.
Actually, if A is known to be positive (or negative), then only two bits are involved, namely
the n" and one of its adjacent bits. Denoting the signal and noise components of Q by Q,
and Q,, respectively, then

Qs = an-1Rs[—T(1 — A)] + anRs(AT) + angi

Rs [TU + )] (4.370)

where R,(rt) is the signal pulse autocorrelation function given by (4.33) and repeated here
for convenience
ie

R(t) = | p(t’) p(t! — r)dt'

‘ (4.371)
j
bs
mi p(t’+ 7)p(t')dt’
0)

With regard to the input bits affecting the calculation of Q;, one of four situations can occur:
either the n" bit and its adjacent bit are alike (both positive or both negative) or they are
unlike (negative followed by positive, or vice versa). Hence, the overall error probability
conditioned on knowing A is given by
2
P(E/A) = ) [Prob {n'® and adjacent bits = (—1)*}
k=
x Prob {error/n' and adjacent bits = (—1)*} (4.372)
+ Prob {n'® bit = (—1)* and adjacent bit = (—1)**"}
x Prob {error/ n'” bit = (—1)* and adjacent bit = (—1)*+"}]

20. In this section, the /P factor in (4.366) is absorbed in p(t).

270 Coherent Communications with Waveforms Chap. 4

Since the random variable Q is conditionally Gaussian on a,—1, Gn, Gn+1 and fixed A,
with variance

op =E\ G7} =—NoPT (4.373)

and the input bits occur with equal probability and are independent, then
1
P(E/2) = zerfe {VE/Nolrs((AIT) + ro( — ITI]
‘ ; (4.374)
+ erfe |/E/Nolrs(\AIT) — rT — ITI]
where
Rs(T)
rs(t) A ae (4.375)
denotes the normalized correlation function and is pulse shape dependent. For NRZ data,
that is p(t) = JP; O<t<T

=
See
T =
rs(T) 0 int (4.376 )

Substituting (4.376) into (4.374), we get (recall for BPSK, E = E, = E;)

P(E/A) = serte(/E/No) + zerfe[ VE/Notl ~ 2100], IAl<1/2 (4.377)

If the data is Manchester coded, then the conditional error probability is given by (see
Problem 4.38)

P(E/A) = erfe[ VE/No(l — 21a] + zerte|

3 /E/No(l — 41a], IA]
<1/4 (4.378)
When the data format is either RZ or Miller coding, then the expression for P(E/A) as given
by (4.374) is no longer valid. The reasons for this are as follows. From Chapter 1, we recall
that an RZ data sequence contains two basic pulse waveforms, p;(t) and p2(t); thus, the
decision rule of (4.374) must be altered to accommodate this fact (see Problem 4.39). For
Miller coding, there are four elementary pulse waveforms; furthermore, the occurrence of
these waveforms in a data sequence is not independent from bit interval to bit interval (see
Problem 4.40). Despite these differences, one can take an approach similar to that leading
up to (4.374) and produce expressions for the conditional error probability. associated with
RZ and Miller-coded data. For RZ data, we obtain

Pin) jerte[ VEP2No] + zerte|/E/2No(1 — 4a). IAl< 1/4 (4.379)

and for Miller codes, we have
1 3 3
P(E /X) ==se ect (a) + cert (b)

3 [pve erf(x) exp[—(x

awe — /2b)?]dx (4.380)

= “el
Fray es
erf(x) exp[—(x oi/2a)*\dx, |A| < 1/4
Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 271

where

a A VE/2No(1 — 41a)
b A JE/2No (4.381)
c A /E/2No(1 — 2|A))
The average unconditional error probability P(E) is determined from
Amax

P(E) =| fOA)P(E/A)dxr (4.382)

where, as before, f(A) is the probability distribution of the normalized sync error and
Amax is the maximum value for which A is defined; for example, Amax = 1/2 for NRZ and
Amax = 1/4 for Manchester, RZ, and Miller-coded data. The pdf, f(A), depends on the way
in which bit sync is derived at the receiver. For the purpose of comparing the average error
probability performance of the various data formats, we postulate a Tikhonov pdf for f(A)
which is characterized entirely in terms of the variance o of the sync error. Thus, for NRZ
data [let p, 4 1/(270,)7]
exp (~, cos(27 i))
fa) = » [Al<1/2 (4.383)
1o(px)
whereas for baseband signaling that employs transitions in the middle of the symbol inter-
val, f(A) is characterized by

2 exp [(cos 41d) /(420.04)?

fQ)= |A| < 1/4 (4.384)
Ip [C./420)? |
Substituting (4.383) or (4.384) into (4.382) together with the appropriate expression for
P(E/2) and performing numerical integration provides the graphical results of Figures 4.47
through 4.50. In these curves, the average error probability is plotted versus E'/No with the
standard deviation 0, of the normalized sync error 4 as a parameter. The degradation in
system performance relative to a perfectly synchronized system can be expressed in terms
of the amount by which the signal-to-noise ratio must be increased to achieve the same
error probability. As an example, at P(E) = 10~° for an NRZ pulse with o, = 0.06, the
degradation is about 1.4 dB.

4.3.4 Effect of Data Asymmetry

We now discuss yet another system imperfection which causes a degradation of perfor-
mance relative to that of the ideal coherent receiver. This imperfection, namely data asym-
metry, occurs when the rising and falling transitions of the baseband waveform do not occur
at the nominal data transition instants [19, 20]. In Chapter 2, we considered the effects of
such asymmetry on the power spectral density of the data waveform for NRZ and Man-
chester formats. Here we assess the degradation of this type of waveform distortion on the
average error probability performance of the receiver.
NRZ Data. The signal model (referred to as Model 1 in [20]) used to characterize
NRZ data with asymmetry assumes that +1 NRZ bits are elongated by A7’/2 (relative to
their nominal value of T sec) when a negative-going data transition occurs and —1 bits
are shortened by the same amount when a positive-going data transition occurs. When
 272 Coherent Communications with Waveforms Chap. 4

10°2

P(E)

10°3

10°

Figure 4.47 Average probability of error

versus signal-to-noise ratio with standard
deviation of the symbol sync error as a
parameter (NRZ) (Reprinted from [12])

no data transition occurs, the bits maintain their nominal T-sec value. An illustration of
this distortion for a typical sequence is given in Fig. 4.51. Assuming an integrate-and-
dump (matched) filter, then for random NRZ data (50% transitions) with asymmetry, the
integrate-and-dump output depends, in general, on the polarity of the bit over which it
is integrating and that of the preceding and succeeding bits. As such, one must compute
the integrate-and-dump output for each of the eight possible three-bit sequences and their
corresponding conditional (on the particular sequence) error probabilities. Then, averaging
these conditional error probabilities over the eight equally likely three-bit sequences gives
the error probability performance in the presence of asymmetry.
Defining the measure of data asymmetry, 7, by the difference in length between the
shortest and longest pulses in the sequence divided by their sum, that is,

T(i+4)-T(i-4) A
T(+9)+T0-3)
4
2 a4.385
then the average error probability performance described above is given by [20]

5 LE l YS Per |E
Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 273

10°

10" § -|
i
0.10
ie = 0.07 #5

6, = 0.06

a J
0.05 4
10°3

\\ 0, = 0.04 |

10
: “he
1A
| P(E)= 1/2erfc\/E /N,

2 4 6 3 10 12
4

{4
Figure 4.48 Average probability of error
versus signal-to-noise ratio with standard
deviation of the symbol sync error as a
E/N dB parameter (Manchester code) (Reprinted
o’ from [12])

For a given value of n (typically measured in percent), one can evaluate the SNR degrada-
tion by computing the additional E'/ No required (A E’/No) due to asymmetry to produce the
same value of bit error probability in the absence of asymmetry (7 = 0). This degradation
(in dB) is given by

E/N
AE/No = 10 logo (=o (4.387)

where for a given error probability P(£) and data asymmetry n, E/No is computed from
(4.386) and (E/No)¢ is computed from (4.386) with 7 = 0, namely, the value corresponding
to ideal coherent BPSK performance. Table 4.5 gives these degradations for error probabil-
ity values between 10~7 and 10~®°, and asymmetries in the range of 5% to 25%.
Inherent in the above discussion was the assumption of perfect bit synchronization
which in the presence of data asymmetry is tantamount to having a clock with a misalign-
ment equal to half the asymmetry. If in addition to data asymmetry we include the effect
of imperfect bit synchronization (additional clock misalignment) discussed in Section 4.3.3,
then the conditional error probability in the presence of both imperfections is given by [20]
 Coherent Communications with Waveforms Chap. 4
274

10°

10°2

P(E)

10°3

Figure 4.49 Average probability of error

versus signal-to-noise ratio with standard
E/N. dB deviation of the symbol sync error as a
oO parameter (RZ) (Reprinted from [12])

Serte
/E +terfe(/% —n-210)
+ erfc (/za pets 21A1))

P(E =} +rherte
(\/£ (1-20): O<|al<% (4.388)
berfe /£ + ferfe (/£ (1 -n 21)

+herfe (EU
—n +210); P<|al<3}
where as before AT denotes the misalignment of the bit sync clock relative to its nominal
position. Note that for 7 = 0, (4.388) reduces to (4.377) as it should.
The average probability of error, P(E), is obtained by averaging (4.388) over the pdf
of the bit sync misalignment. Postulating, as in Section 4.3.3, a Tikhonov pdf for 4, which
is entirely characterized by the single parameter ao? [see (4.383)], then P(E) is given by
(4.382) where (4.388) is now used for P(E|A). Figure 4.52 illustrates P(E) versus E/No in
dB with oy as a parameter and a fixed value of asymmetry equal to 6%.
Manchester Data. The signal model (Model | in [20]) used to characterize Manch-
ester data with asymmetry assumes that for a +] data bit, the first half is elongated by AT /4
(relative to its nominal value of 7/2 sec) and for a —1 data bit, the first half is shortened
 Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 275

10°

10°2

P(E)

10°3

Figure 4.50 Average probability of error

. . versus signal-to-noise ratio with standard
10-5 5 | | deviation of the symbol sync error as a
parameter (Miller code) (Reprinted from
[12])

by the same amount. When a data transition follows, the second half of the Manchester bit
retains its nominal (7 /2-sec) value. When no data transition occurs, the second half of the
Manchester bit adjusts itself so that the total bit maintains its nominal T-sec value. An illus-
tration of this distortion for a typical sequence is given in Fig. 4.53. Assuming an integrate-
and-dump (matched) filter, then for random Manchester data (50% data transitions) with
asymmetry, the integrate-and-dump output again depends, in general, on the polarity of the
bit over which it is integrating and that of the preceding and succeeding bits. Thus, as for
NRZ data, one must compute the integrate-and-dump output for each of the eight possible
three-bit sequences and their corresponding conditional (on the particular sequence) error
probabilities. Then, averaging these conditional error probabilities over the eight equally
likely three-bit sequences gives the bit error probability performance in the present of asym-
metry.
Defining the measure of data asymmetry, n, relative to the bit time 7 as was done for
NRZ data?! then analogous to (4.385) we have

21. By defining data asymmetry relative to the symbol time for both NRZ and Manchester data, equal percent
asymmetry implies equal amounts of asymmetry (in seconds) as measured by the actual time displacements of the
waveform transitions.
 Coherent Communications with Waveforms Chap. 4
276

(a)

Figure 4.51 (a) Symmetric NRZ data stream; (b) Asymmetric NRZ data stream (Reprinted from [20])

Table 4.5 Degradation Due to Asymmetry for

Uncoded Random NRZ Data

AE / N,, dB

ee
T(i+4)-T(1+4)
ee ee
a 4.389
"“ T(i+4)+7 (1-4) .4 vee
and the average bit error performance becomes [20]

l E 1 sh
BAB) = 17otic (\550-20) + pene([F¢-9) (4.390)

Figure 4.54 illustrates a comparison between the SNR degradation due to data asymmetry
for NRZ and Manchester signals. We see that for analogous definitions of asymmetry as
in (4.385) and (4.389), the Manchester signal always yields a larger SNR degradation for
a given percent asymmetry as one might anticipate since, on the average, the Manchester
waveform has 3/2 as many transitions as the NRZ waveform.
When both data asymmetry and imperfect bit sync timing are present, then the analo-
gous result to (4.388) becomes
 Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 277

10°

1071

10-2

10°3

P(E)
10-4

10°5

10-6

Figure 4.52 Average error probability with

10-7 symbol synchronization error as a parameter
(6% asymmetry), NRZ data (Reprinted from
hl [20))

gerte(\/F (1-2n))
+g erfo(,/# —n—21a0)
+4 erfe(\/£(—n +21A)) O<|al<3
PCE|X)= (4.391)
herte(\/#( a4 AD) +3 erfo(/ #1 =r? ID)

+ erfe(/ +4 —2 iAl)); NS<|Al< -l-—

which reduces to (4.378) when n = 0. Averaging (4.391) over the pdf of the bit sync
alignment”? modeled as a Tikhonov pdf [see (4.384)], P(E) is given by (4.382) where
(4.391) is now used for P(E|A). Figure 4.55 illustrates P(E) versus E/No in dB with
0), as a parameter and a fixed value of asymmetry equal to 6%. Comparing Fig. 4.55 with
Fig. 4.52, we observe that for fixed values of n and A, the error probability performance of
Manchester data is always worse than that for NRZ.

4.3.5 Suboptimum Detection

We conclude this chapter with a discussion of antipodal signaling with rectangular pulses
but using an arbitrary filter H(f) instead of a matched filter (I&D) as in Fig. 4.56. The

22. The nominal bit sync misalignment is now AT/8.

278 Coherent Communications with Waveforms Chap. 4

ae F(t 5) aC 5) if F (1-3)
Dero! yao GY Ear A 2 2

Figure 4.53 (a) Symmetric Manchester data stream; (b) Asymmetric Manchester data stream
(Reprinted from [20])

input modulation takes the form

CO

m(t)= > _ agp(t —kT) (4.392)

k=—oo

Here the a,’s are the independent +1 data bits and p(f) is the rectangular pulse

noe ee
0,
US
otherwise
(4.393)
Zero Signal Distortion. Assume that |H(f)|max = 1 and that H(f) has a single-
sided noise bandwidth By that is sufficiently wide so as to pass m(t) “undistorted.” Typi-
cally, a By of 3 to 4 times 1/T is sufficient. Thus, after passing through the filter H(f), the
signal plus noise is

y(t) =m(t) + n(t) (4.394)

where the “hat” denotes filtering by H(f) and

oA =a, sh Sn( PH (f) Pdf = NoBn (4.395)

The mean-square value of the filter output sample is

E
E* {y(T)} = * (4.396)
 Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 279

E/No
=1.5 dB

MANCHESTER
(n=A/4 =6)
wo

nw

AE/No,
DEGRADATION
SNR
dB

NRZ
(n= A/2 =26)

Figure 4.54 SNR degradation versus

0 5 10 15 20 25 30 35 percent asymmetry for random Manchester
ASYMMETRY (%) and NRZ data (coded) (Reprinted from [20])

Thus, from (4.51), we have

P(E)=Q (= (4.397)

Since for H(f) a matched filter, By = 1/2T and SNRmatched = 2E/No, then the SNR
penalty of using other than a matched filter (assuming undistorted signal) is
SNR
10 logy (Fe) = 10logi)2ByT (dB) (4.398)
Clearly as By get smaller, the distortion of m(t) increases and the mean of y(T)
becomes less than E/T. The value of the mean of y(7T) for any specific sample time
t =nT depends on the past sequence of a,’s, that is, the previous bits interfere with the
mean of the sample corresponding to the present bit. This phenomenon is called intersymbol
interference (ISI). In general, the effect of ISI is difficult to compute since the distortion
depends on the infinite past. Typically, we make some approximation, that is, we assume
 Coherent Communications with Waveforms Chap. 4
280

10°

1071

10-2

10-3

P(E)

10°5

10-6

Figure 4.55 Average error probability

with symbol synchronization error as a
parameter (6% asymmetry), Manchester data
(Reprinted from [20])

that the filter has finite memory, that is, the distortion of the mean of the filter output sample
for a given bit depends on a finite number of previous bits. We now present an example to
illustrate this point.
Worst Case Design—Upper Bound on Performance. Let H(f) be a single pole
Butterworth filter, that is

] 4
HU) = sept Bw = (=) fe (4.399)
20 fe
Assume now that a sequence of all —1 data bits preceding and succeeding the data bit
of interest which is a +1. This input modulation m(t) (assuming rectangular pulses) is
illustrated in Fig. 4.57. Assuming that y(t) has value —1 just prior to the bit interval of
interest, then during this interval we have

E
E{y(t)}= We {-1+2(1 —e~*")} = mets — 2e~@") (4.400)

The value of the mean of y(t) at the end of the pulse interval of interest is
 Sec. 4.3 Inphase and Quadrature Modulation and Demodulation 281

n(t)

DECISION
m(t) J aR

DECISION

t=nT Figure 4.56 Optimum and suboptimum

detection: (a) Optimum detection, (b)
(b) Suboptimum detection Suboptimum detection

mit)

y(t)

Vet-—-p-------=-------

Figure 4.57 |Worst case ISI sequence

E{y(nT)} = fe (1‘ ae) = 2 (1a getant) (4.401)

and the variance is as in (4.395). Thus, the error probability is given by

ey Ge Ae E eis:
eae ( oy )=¢ (,T NoBn (1 He )
(4.402)
a E (1 —2e744n7)
~2\V No VENT
We wish to choose By so as to maximize the argument of the Gaussian tail integral, that
is, produce the best trade-off between ISI and noise power in so far as minimizing error
probability. Differentiating the argument of the Q function in (4.402) with respect to ByT
gives

Bat (Be 2x") - a (1a beget) 0 (4.403)

N
282 Coherent Communications with Waveforms Chap. 4

m(t)

y(t)

Figure 4.58 Random data sequence

which when solved results in

Bul 0162 (4.404)

The corresponding worst case error probability is determined by substituting (4.404) into
(4.402), resulting in

E
P(E) =O (iis] (4.405)
\ No
which when compared with the matched filter result represents a —10 log; 9(1.118/2) = 2.5
dB SNR loss. We now reexamine the design of the same receiver structure in light of a
minimum average error probability criterion.
Average Error Probability Performance Design—Finite Memory of Past Two Sym-
bols. Instead of considering the worst case ISI sequence, we now assume a random data
sequence as shown in Fig. 4.58. We shall make the following approximations to the response
of H(f) to m(t): At any potential data transition point, there is either a transition or no tran-
sition, each occurring with equal probability 1/2.

1. Assume that after two pulse intervals, the response achieves its maximum value. Thus,
if we have ax, = ax—1, then the mean of y(T) equals /E/T.
2. If there is a transition in the waveform, that is, a, # ax—1, then the preceding bit, that
is, ax—2 will either be identical to ay,_, or opposite to ax_}:
(a) ag—2 F ag-1
This case is illustrated in Fig. 4.59(a) and we shall assume that the sequence
has alternated continuously into the infinite past. Then, to determine the mean of
y(T), we have

E
=F lct (= ev) (1esec) = E{y(T)} (4.406)
or
LEU Seed)
E{y(T)} ~( eS (4.407)

(b) ax—2 = ag-1

 Inphase and Quadrature Modulation and Demodulation 283

m(t)

y(t)

y(t)

Figure 4.59 Finite memory of past two

symbols: (a) ax—2 # ax—1, (b) ax—2 = Ag_1

This case is illustrated in Fig. 4.59(b). Here, we assume that the response starts
from —./ E/T. Thus

E{y(T)} = (a eect
yee
E —w-T
) (4.408)

Case | occurs with probability 1/2 (half the time, there is no transition). Cases 2a
and 2b each occur with probability 1/4 (half the time, there is a transition. Half
of that time, the two preceding bits are identical and half of that time, the two
preceding bits are opposite). Thus, the average error probability is

ast |

ahs 30( a ;2( sor (|

(4.409)

# 50 (ah (1a a)

Again, one can choose ByT to minimize P(E) by first rewriting (4.409) as
284 Coherent Communications with Waveforms Chap. 4

Ewin Baa 1 ae {eS

= 22\VNoBut?) * 42 \YNoBut |14 eoeF
(4.410)
+ bay50(
H rah
ee me heat
7]
Note that the optimum ByT depends on E/No. Some typical values for the op-
timum ByT are as follows: E/No =7 dB, (ByT )opt = Ek] Np = Orde
(ByT opt = 256026 /-Neg = 13 06, (ByT opt = .61. Again for sufficiently large
ByT, the above reduces to (4.397) corresponding to the zero ISI result.
 Appendix 4A
Derivation of the Symbol Error Probability for Polyphase Signals
The probability of error for polyphase signals is obtained by taking one minus the
probability of correct detection as given by (4.130), that is

utanm/M
P,(E) =1— = [ exp {-(ua VEIN) || exp (-»”) audu (4A.1)
JO 0

Letting z=u — /E,/No

2 CO

(E)
P,(E) =1-——
I pet (VEN) E;/No, d
a (4A.2)
4A.2

where

(z+x)
tan 1 /M
f (x,z) = exp(—z”) / exp(—v7)du (4A.3)
0
Taking the partial derivative of P;(E) with respect to ./E;/No, we get

OP,(E) --2|{" 2 D c+ $6ER -VETRO) ie

0/Es/No Tal Seeing. 0%, Es) No

ee ie Af VESTN: 2)4,
a | J—/E,jNo 9V/Es/No
From (4A.3)
a 12
Z) _ 2 1A
Ae on
Of (VSEs/No, = exp (-z )tan uM exp |-{(zef Es/No) tan =} (4A.5)

M4 Es 4
= tan = exp|- |2sec? © +22 E, /No tan? WE wn? TI

Completing the square in (4A.5) gives

Of (/Es/No,Z) _ | Es i |
== (EN (2.4 0) 4
0./Es/N M No M
ioe: (4A.6)
2
x exp {-sec” 7 (z+ /Es/No sin® —)

Substituting (4A.6) into (4A.4), we get

— Bio ® 7aS
=-(- )tan(= exp aha
Es/No M
2 (4A.7)
lee) E,; I
x ik expe sec? 24+ |= sin’ — dz
-/E,/No oe)
286 Coherent Communications with Waveforms Chap. 4

Letting t = (z+ /E;/No sin? a /M) sec(z/M)

P(E 2 E oy
ee aes) sin(—) exp |-= sin? a / exp(—t”)dt
0./E;/No a M No M —,/E;/Nocosm/M (4A 8)

= ~
1 E
sin a exp \-= sin? at [P+ erf(./ Es /No cos =)
1

Integrating (4A.8) with respect to ./E;/No between 0 and ./E;/No, we get

: a/ Es/N
P(E) — PE) esy0 =-3er( [sin 4 2 ae) " 4A9)
x exp {= sin? x erf(\/E,/No cos x/M) d./E;/No
0

Letting y= /Es/Nosin(z/M) and noting that Ps(E)| he (M — 1)/M, (4A.9)

simplifies to the desired result
M-1 dA Ort,
P(E) =—— — Serf |VEs/No sin oA
1 pVEs/Nosin/M : :. (4A.10)
peg exp(—y*) erf(y cot wv
 Appendix 4B
Bandpass Random Process Representation

Let x(t) be a real zero-mean bandpass random process given by

x(t) = Re {m(t) exp(j2z fct)} (4B.1)

where m(t) denotes the complex baseband modulation function, consisting of both ampli-
tude a(t) and phase 6(t) modulations, that is

m(t) = a(t) exp(j@(t)) (4B.2)

We can also define x(t) as the complex random process given by
x(t) = m(t) exp(j2z
fet) (4B.3)
such that
x(t) A Re {x(t)} = a(t) cos(2m fct + 6(t)) (4B.4)
Assume that the random process x(t) is wide sense stationary (WSS) with PSD S,(f) as
depicted in Fig. 4B.1 and expressed as

5, (0)

Figure 4B.1 Bandpass and baseband

(b) power spectral densities
288 Coherent Communications with Waveforms Chap. 4

x(t)

Figure 4B.2_ IF to baseband demodulation

SG) =n. = torr S2(—f = fe) (4B.5)

with the condition that

Spf) =O-tor f. <= —fe (4B.6)

Consider the processing depicted in Fig. 4B.2 where the output z(t) of the linear time-
invariant (LTI) filter consists solely of the positive frequency content of the input x(t) by
defining U(f) as

A eel
’ =kena
U(f)S 0, otherwise eae

Since the filter is LTI, the output z(t) is also WSS with PSD

S:(f) == |2U(f)Sx(f/)
|2 a (4B.8)
=4S.(f — fe)
The autocorrelation function of the modulation m(t) is defined as Rym*(t1, t2) = E{m(t))
m*(t2)}. Using m(t) = z(t) exp(—j2z fet), then

Rinm*(th, t2) = E{z(t1)z* (t2)} exp{—j2m felts — t2)} (48.9)

= Rzzx(ty — t2) exp{—j2mfe(ti — t2)}
Therefore, m(t) is also WSS and as a result, we can write Rmm*(t) = R(t) exp(—j27f.T)
or equivalently

Sm(f) = SF + fe) = 481i (Pf) (4B.10)

Expressing m(t) in terms of its real and imaginary parts, we have

m(t)
= mR(t) + jm7(t) (4B.11)

where both mr(t) and m;(t) are real functions. In order to compute second-order statistics
of mr(t) and m;(t), we need an expression for Rz-(t), t2) 4 E{z(t;)z(t2)}, which is dif-
ferent from the autocorrelation function R,,«(t), t2) for complex processes. Consider Fig.
4B.3, which depicts a real WSS process x(t) through a complex LT] filter h(t). The output
z(t) becomes

z(t) =f h(a)x(t — a)da (4B.12)

—o0O
App. 4B Bandpass Random Process Representation 289

COMPLEX
LTI FILTER

x(t) bi => zi)

WSS REAL
PROCESS Figure 4B.3 Complex LTI filter

resulting in

Reha / / h(a)h(B)E
(x(t —«)x(t2 —B)\dodB
See (4B.13)
- / / h(a)h(B)Rex(th— ty — a + B)dadp
Note that R,-(t), t2) is a function of t; — tz only. Therefore

Rze(t) = / ik h(a)h(B) / Se( fyel?7I-°+P)d fdadp

Ez oy Fab (4B.14)
u / Sy(f)H
(fyH(—fyel"I*af
or

SAS ya PILR,()) = S(fA(fPA(—f) (4B.15)

where F.T.{ f(t)} denotes the Fourier transform of f(t). Referring back to Fig. 4B.2,
H(f) =2U(f) and
Sez(f) = S.(f)2U (f)2U (—f) =0 (4B.16)
since U( f)U(—f) is nonzero only at f = 0, but S,(0) = 0 since it is a bandpass function.
Hence, R,,(t) = 0 and

en) Oltor allan (4B.17)

Noting that
Rinm*(t1, 12) = Rig (th, t2) + Ry (th, #2) + JRmimp(t, 12) — JRmem,
(th, t2) (4B.18)
Rim(t1, t2) = Rmg (ti, 12) — Rm, (A, 12) + J Rmpmg (ts 12) + J Rmeym) (th, 22) =90 (4B.19)
we have by adding (4B.18) and (4B.19)
1
Ring(h, 2) = Ret Rmme(t1, t2)} (4B.20)
1
Rinpmg (th, 12) = 5Im{Rmm*(t1, t2)} (4B.21)
where Re(x) and Im(x) denote the real and imaginary parts of x, respectively. Subtracting
both equations, we obtain
1
Rm, (tH, t2) = 5 Ret Rms (ti ,t2)} (4B.22)

1
Rm gm;(thy #2) = — 51m{Rmm*(t1, t2)} (4B.23)
290 Coherent Communications with Waveforms Chap. 4

Hence

Rmg(t) = Rm, (Tt) = 5RelRnm(c)} (4B.24)

and

Rinpm;(T) = —Rinjmg(—t) = —FIm(Rnm*() _ (4B.25)

From (4B.10), Sn(f) =4Sz(f) and since Rinm*(T) = pee Sn(f) exp(j2aft)df, (4B.22)
and (4B.23) become

oe al? fl SL(f) cosQnft)df (4B.26)

and

Rngm0)=—Rnmg(—1)=-2 | Sif sin@xfedf ——_(4B.27)

Using (4B.11) in (4B.1), the bandpass process x(t) can be written as!

x(t) = mp(t) cos(27 f-t) — m;(t) sin(27 ft) (4B.28)

where mp(t) and m;(t) are zero-mean jointly WSS real random processes with second
order statistics given by (4B.26) and (4B.27). Note that if S;(f) is symmetric about f = 0,
then Rmpm,(t) = 0 for all t and mr(t), m;(t) are uncorrelated. Furthermore, if we add the
constraint that

Si(f) =9 for |f1 > Ifel (4B.29)

then m(t) can be recovered by mixing x(t) with the complex exponential exp(— j27 f,t) and
then filtering the product with an ideal lowpass filter of bandwidth f, Hz as depicted in Fig.
4B.4.
An application of great importance is the representation of bandpass random noise
processes. In any receiver, the signal is first amplified and then bandpass filtered (BPF) to
reduce the noise. The noise at the output of the BPF has PSD S,,( f) as shown in Fig. 4B.5(a)
with a two-sided level No/2 Watts/Hz and bandwidth 2B Hz with B < f;. In this case, the
noise can be represented by

n(t) =N¢(t) COS(@ct + YW) — ns(t) sin(@et + Ww) (4B.30)

where n-(t), ns(t) are uncorrelated random processes with PSD as depicted in Fig. 4B.5(b),
that is

Sne(f) = Sn (f) = No |f\|<B (4B.31)

0) otherwise
and w is any arbitrary phase. Since the integral of the spectral density is the power of the
process, we have

Pa. = Pn, = Pn = 2NoB (4B.32)

1. By writing m(t) = mp(t) — jm;(t), x(t) can be expressed as x(t) = m,(t) cos(2z fet) + my(t) sin(27 fet)
with Rmgm,(T) = 2 f°>, SL(f) sin(2xft)df.
App. 4B Bandpass Random Process Representation 291

COMPLEX LTI FILTER

x(t)

-sin2nfot

(b) Real implementation

Figure 4B.4 Baseband recovery: (a) Complex implementation, (b) Real implementation

In some instances, it is more convenient to represent the noise n(t) with an additional = 5)
factor as

n(t) = V2 {n/(t) cos(wet + W) — ni (t) sin(wct + y)} (4B.33)

Since both representations must have equal power, then we need

Nc
Su f) = Sy (f) = {2: fiaeblalge
otherwise
(4B.34)
to produce P, = 2NoB as in (4B.32). It is often convenient to express the noise n(t) in its
complex form as

n(t) =Re bara ey) (48.35)

= a(t) cos(wct + A(t) + W)

where a(t) and @(t) are the envelope and phase noise processes given by
 292 Coherent Communications with Waveforms Chap. 4

5, ()

0
-f, fo f

(a) Bandpass noise PSD

10 = Sn

No

-B 0 B f Figure 4B.5 Receiver noise PSD: (a)

Bandpass noise PSD, (b) Baseband noise
(b) Baseband noise PSD PSD

a(t) = ./n2(t) + n2(t) (4B.36)

ns(t
A(t) =tan-"| s{ 4 (4B.37)
N¢(t)

PROBLEMS
4.1 Consider a binary “on-off” communication system which is used to send one of two equally
probable messages in the face of AWGN (with one-sided PSD No) during the time interval
(0, 7). Assume the “on-signal” has energy E.
(a) Assume N = 2 (dimension of vector space). Construct the vector representation and illus-
trate the optimum vector receiver. Find the probability of error.
(b) What waveform design is compatible with the vector in (a)? Illustrate the optimum wave-
form receiver.
4.2 Show that in an AWGN channel, when the average signal energy is constrained not to exceed
E, that is
T T
Pimo) | so(t)dt + pom) [ s(t)dt <E
0 0
that the optimum signals are given by
Problems 293

Figure P4.3

P(m)
a= S@), OSes
P(mo)

si(t) = — [Feesm, OSn =i

Here s(t) is any signal having energy

T
an s*(t)dt
0

4.3 Consider the following binary communication system. The information is sent using one of
the two signals shown in Fig. P4.3, which are antipodal, nonreturn-to-zero (NRZ) waveforms.
Assuming equilikely signals and an AWGN channel with two-sided PSD No/2, design the
optimum receiver, using the minimum number of correlators, and compute the probability of
error for the case A = 2 volts, T = 1 sec, and No = 2 watts/Hz.
4.4 Consider the pulses depicted in Fig. P4.4.
(a) What is the impulse response for filters matched to pulses (a) and (b)?
{b) What is the output of the matched filters if they are excited by the signals to which they are
matched?
(c) Make comments on the difference in results. Suggest a practical implementation.
4.5 Consider an amplitude-weighted binary FSK system described by the pair of signals

si(t)
= A(t) cos 4 | a + (-1)
i+) 5
A@ Phen
;
Theat Wesyesve

(a) Find the signal cross-correlation coefficient

yofj4 ae So(t)sy(t)dt

Io ss(dt fy spat
vs
(b) If the weighting is exponential, that is, A(t) = Aexp(—Wr), evaluate yo; for the case
where the frequency separation Af = Aw/2z7 is an integer multiple of the data rate 1/7.
(c) Assuming an AWGN channel and an optimum matched filter receiver, what is the error
probability of the system for the amplitude weighting of (b)?
(d) Is the error probability found in (c) better or worse than that for an orthogonal binary FSK
system with A(t) = A?
4.6 For coherent FSK, the transmitted signal in the n" transmission interval nT <t < (n + 1)T is
 Coherent Communications with Waveforms Chap. 4

to 2to ©3tq «4tq §«S5tq 6tq §67to t

(a) Single pulse

(b) Pseudo-noise pulse Figure P4.4

given by

Ao
s(t) = Acos out+ an p(t — nt) |

with a, taking on values +1 with equal probability. Conventionally, p(t) is the unit rectangular
pulse

oe i 0Se= 7
PX’) 0, otherwise
in which case the two possible signals (corresponding to a, = —1 and a, = 1, respectively), are
given by

A@
so(t) = Acos (vx= =) |
2)

Aw
s1(t) = Acos |
(vx
ate =) |

Suppose now that the pulse shape p(t) is the unit square wave

OSRSi2
PM
(t)=a L
ae)Cana fae Pace &

which results in so-called Manchester coded FSK.

(a) Write the equations analogous to the above for so(t) and s(t).
(b) Find the signal cross-correlation coefficient yo; (assume n = 0 for convenience). Does it
differ from that found for conventional coherent FSK (p(t), a unit rectangular pulse)? If
so, how?
Problems 295

Figure P4.11

4.7 The received signal in a binary baseband communication system takes on values +A in each T-
sec interval. The probability of transmitting +A is 1/4 and the probability of transmitting —A
is 3/4. An integrate-and-dump filter is used for detection with the comparison threshold set to
Vo. Assume an AWGN channel with single-sided PSD No.
(a) Derive an expression for the average error probability, P(E), as a function of the detection
threshold Vo.
(b) What is the value of Vo that minimizes P(E)?
(c) What is the corresponding minimum value of P(E)?
4.8 A binary communication system (AWGN with 2-sided PSD No/2) transmits a binary one by
sending either a waveform /)(¢) or fo(t) (energy E, time T), selected by tossing a coin, and
sending nothing for a binary zero. Assuming equally likely bits, sketch the decision mechanism
that minimizes bit error probability.
4.9 (a) Obtain the maximum output signal-to-noise ratio of a matched filter for the following input
signals: A and J, are constants. Assume that the two-sided noise PSD at the input is No/2.
(1) x1(t) = A{u(t) — u(t — 5T¢)}
QO) @) Alu@) — 24 = sic) u(t — 47) — ut — 57,)}
where u(t) denotes the unit step function.
(b) Find the amplitude of an input signal pulse of duration 7, that will give the same maximum
output signal-to-noise ratio as the signals in part (a).
(c) Obtain the autocorrelation functions of the signal in parts (a) and (b). Comment on the
utility of each for resolving delay or relative time differences.
4.10 For the binary coherent communication problem, show that the performance of the optimum
receiver for the unequal energy, unequiprobable signal case is given by (4.64), that is

REEVE uO eee wee AOgg ee

rs V2No dioY 2 P(mo)

+ P(m\)Q (al ae cos vn)

J/2No' digY 2 © P(mo)

4.11 Either of the two signals shown in Fig. P4.11 is transmitted with equal probability in a binary
communication system.
(a) Find the matched filter impulse response h(t) that would be required to implement the
optimum receiver in an AWGN channel.
(b) Calculate the optimum threshold and the error probability performance of the receiver as a
function of fo.
(c) What is the best choice for fo such that the error probability is minimized?
(d) Sketch a correlator receiver for these signals.
 296 Coherent Communications with Waveforms Chap. 4

g(t),
i=0,1
N\
r(t)| WAVEFORM | ™
DECISION

DELAY T/2
Figure P4.12

4.12 Consider the binary communication system shown in Fig. P4.12. Assume the signals are equi-
likely and that n(t) is AWGN with PSD No/2 watts/Hz.
(a) Illustrate the optimum correlator receiver and the optimum matched filter receiver.
(b) If s;(t), i =0, 1 were antipodal signals, evaluate the probability of symbol error.
4.13 A ternary communication system transmits equally likely messages mo, m;, and m2 using three
orthogonal signals with equal energy. The channel adds white Gaussian noise with one-sided
PSD No.
(a) Define vector and waveform signal models.
(b) Illustrate the optimum vector and waveform receivers.
(c) What is the message error probability in terms of E;/No?
(d) What is the minimum number of correlators required to implement the optimum receiver?
(e) Illustrate the optimum waveform receiver in (d) and define all waveforms.
4.14 Consider the following hypotheses testing problem:

Jes AO) ESRO) Fa), OSes iP

fori = 1,---,
8 where
ane A . E
5;(t) =V2E sin (1oS)

This signal set is known as 8-PSK. Assume n(t) is am AWGN with PSD No/2 watts/Hz.
(a) Choose a suitable orthonormal set for the signal space. What is the dimensionality of this
space?
(b) Indicate the optimum decision regions and use the union bound to find an upper bound on
the probability of symbol error.
4.15 A waveform signal set is described by
20
Si(t) = A cos(@ct +1 ae VSS EF

for i = 1,---, 6. Assume all signals are equilikely and that the signals are transmitted over an
AWGN channel.
(a) Find a suitable set of orthonormal basis using minimum number of dimensions. Sketch the
signal vectors and indicate the decision regions for optimum reception.
(b) Use the union bound to find an upper bound for the probability of symbol error.
4.16 Consider the problem of communicating over white Gaussian noise channel (No = 1/2). One
of four possible waveforms is transmitted every second. The source messages are equilikely
and independently chosen. The transmitter has a maximum energy constraint defined by:

|Es| <3
(a) What is the optimum choice for the set of four signals in this situation? Describe the set as
vectors.
(b) Convert the set of vectors in (a) to waveforms.
(c) Draw a block diagram for the optimum waveform receiver. Use the smallest number of
correlators (matched filters).
 Problems 297

90°

1352 45°

AMPLITUDE = yo
180° 0° ABSOLUTE

270° Figure P4.18

(d) Using the result from the union bound, what is an upper bound on symbol error probabil-
ity?
4.17 Consider the 4-ary hypotheses testing problem. Each observation is a 3-dimensional vector

Ld en — Siete ta(al aS

(GC) (4
where

The components of the noise vector are independent identically distributed Gaussian random
variables, with variance No/2.
(a) Choosing an appropriate basis, express the signal vectors as waveforms.
(b) Sketch the optimum receiver assuming equilikely signals using the least number of corre-
lators.
(c) Compute the probability of symbol error.
4.18 Consider the V.29 signal constellation depicted in Fig. P4.18, which transmits 3 bits/symbol.
Assume an AWGN channel and equilikely signals.
(a) Write an I-Q waveform model using rectangular coordinates for square and arbitrary I-Q
pulses.
(b) Write an I-Q waveform model using polar coordinates for square and arbitrary I-Q pulses.
4.19 Consider an M-FSK system that employs the signal set (k an integer)

k
sit) = V2P sin |( + aniaf)), PS Oe laake lt 1

(a) Find the minimum frequency separation spacing that produces orthogonality and the effec-
tive bandwidth occupancy of this set.
(b) Using this orthogonal set, define a biorthogonal signal set and find its effective occupancy.
(c) Repeat for a transorthogonal signal set.
4.20 Consider the V.29 signal constellation depicted in Fig. P4.20, which transmits 4 bits/symbol.
Assume an AWGN channel and equilikely signals.
(a) Write an I-Q waveform model using rectangular coordinates for square and arbitrary I-Q
pulses.
(b) Write an I-Q waveform model using polar coordinates for square and arbitrary I-Q pulses.
298 Coherent Communications with Waveforms Chap. 4

Figure P4.20

4.21 Consider a ternary communication system which transmits one of three equally likely signals

EQ TAC Sie),

Ay: r(t)=Asinw.t
+ n(t)

Ho: r(t)=—Asinwet + n(t)

during 0 < t < T. Assume that n(t) is a white Gaussian noise process with PSD No/2 watts/Hz.
(a) What is the optimum receiver which minimizes the error probability? Sketch this mecha-
nization and show all operations which must be performed in making a decision.
(b) What are the conditional probabilities of correct reception when Ho, H), and HA are true?
(c) Using the results found in (b), write an expression for the average probability of error.
(d) Compare this ternary system with the binary PSK system which uses Asinqw,t and
—A sin wet to transmit data.
4.22 Establish (4.129) by substituting (4.127) into (4.128) and averaging over the equal a priori
probabilities of the transmitted signals. Show that (4.129) reduces to (4.130) and the average
symbol error probability is given by (4.131).
4.23 For large signal-to-noise conditions, Gilbert” has found an approximate formula for the average
symbol error probability of a maximum likelihood receiver in terms of the geometrical config-
uration of the transmitted signals in signal space. Applying his formula to an M-PSK system,
the M signals are represented as uniformly spaced points on the circumference of a circle of
radius R and

rie) =Mete| |()|

N E
0
;

where rmin is the smallest of the set of M(M — 1)/2 Euclidean distances between pairs of signal

2. E. N. Gilbert, “A Comparison of Signaling Alphabets,” Bell System Technical Journal, vol. 31, May 1952, pp.
504-22.
 Problems
299

points, and N, is determined as follows: For each signal s;(t), let Ns, denote the number of
signals that are at the minimum distance rmin. Then

M
Ne asi Ns;
——————
M
(a) Show that rmin = 2R sina /M.
(b) For M > 2, show that N, = 2.
(c) From the results of (a) and (b), verify (4.134).
4.24 A general representation for the L-orthogonal signal set is given by?

si(t) = VP [xn(t)
cos + ya(t) sin@] O<t<T
fof — TN Pt 2 Ltn 1)L + 1 with 6 = 2(/ — 1)z/L and

tf
Ti, Xjedd (t)XK( =a |e (Dat arate
= | oesh
pay:

1 T
a} Xj(t)
yx(t)dt = 0 for all j and k
0

If

xn(t) = V2sin jen lee)

yn(t) = VEe0s|7) fil, cae, IN)

and

Xn(t) = V2 sin peso 2

rae ape
u(j +2n)t
yn(t) = /2cos n=1,2....,.N

Show that the signal set {s;(t)} is L-orthogonal. Assume that / is any integer.
4.25 (a) In terms of the L-orthogonal signal representation of Problem 4.24, let {x,(t)} be a set of
N nonoverlapping sinusoidal pulses defined over the interval 0 < t < 7/2 and {y,,(t)} bea
set of N nonoverlapping sinusoidal pulses defined over the interval 7/2 < t < T, that is

Aa Ms (n — 1)T nT
sn(t) = 2VNsin ik on =! = on

A 2IGHINGE (N+n-—1)T (N +n)T

ya(t) = 2 sin ( iF Poona ye sos ANT

forn = 1, 2,---, N. Show that the resulting set {s;(t)} is L-orthogonal.

(b) If N is a power of 2, then another pair of orthogonal signals for {x,(t)} and {yn(t)} can be

3. See [10], and J. J. Stiffler, A. J. Viterbi, “Performance of N-Orthogonal Codes,” JEEE Transactions on
Information Theory, vol. IT-13, no. 3, July 1967, pp. 521-22.
300 Coherent Communications with Waveforms Chap. 4

modeled after the Hadamard design for an orthogonal code. In particular, we define
2N
; C— yr
xn(t) = ¥ \(-1)" -w (:= ) C= far

2N (i- — 1)T
re
yn(t) Gh)
ce ) aew(ea ne
aN ) taal
SS

for n = 1,2,---,.N where w(t) is an arbitrary waveform of energy T/2N and nonzero
only over the interval 0 < t < T/2N, and bj.» is the i element in the n™ row of a2N x 2N
Hadamard matrix. Show that the resulting signal set {s;(t)} is L-orthogonal.
4.26 Show that for the special case of L = 2 and L = 4, the error probability performance of an L-
orthogonal signal set of M = NL signals defined by (4.145) is equivalent to that of a biorthog-
onal signal set of M signals.
4.27 Using the approach of Section 3.2.5, derive (4.198b) which represents the probability that the
receiver’s noisy estimate of the phase in the i" transmission interval lies in the k" sector of the
unit circle, that is, (6; + @g + (2k — 1)2/M < nj <6; +¢a + (2k + 1)7/M).
4.28 Consider the two sets of orthogonal functions

aie 1 sin(@\t — nz)

514 @\t — ni

@2 Sin(@zt — nv)
6, (t) = — ———_
iE @2t — ni

(a) Show that

‘ ; j — — m2.
iC wy, Sin(@jt — nz) w sin(w2(t — tT) + m7) q@, 1 [ou (n m2) |
MI, = ee NT ee et
6 © @t—nr nN a(t—t)+mn 1 wit —(n-m2) x
@2

Hint: Consider the product of the Fourier transform of ¢,(t) and 6,,(—t).
(b) Suppose that a WSS process x(t) has the correlation function

@2 SIN w2T
R,(t) =
wt

Does this process have a K-L expansion in terms of the orthogonal functions ¢,(t)? If
so, do the coefficients in this expansion represent samples of the random process at the
sampling rate 2 f; = 2w,/27?
4.29 The covariance function of the additive Gaussian noise process present in a communication
channel is characterized by

1
LEC NE miexp{—2|t—s|}, |t—s|<oo

(a) Find the K-L expansion for this process.

(b) Draw the optimum receiver if

N
si(t)= sijPj(t), i=O0,1,---,M-1
y=

and specify the appropriate ¢;(t)’s.

Problems 301

4.30 Let x(t) be a zero-mean random process with correlation function given by

R,(t) = ape

where c > 0 is a real constant.

(a) Write the integral equation that gives the eigenvalues and the eigenfunctions in the K-
L expansion of x(t) over the time interval -T/2 <t < T/2, where T is some positive
number.
(b) The usual approach to solve this integral equation is to convert it into a differential equation
and solve that equation. Show that the related differential equation for this case is:

AP(t) + (1 — c?A) H(t) = 0, for |t| < T/2

where ¢(t) denotes the eigenfunctions, A the eigenvalues, and dots denote differentiation
with respect to f.
Hint: Take derivatives on both sides of your answer in (a) two times. Use the Leib-
niz’s theorem for differentiation of an integral
d £2) bY) T g d d
me f(x, y)dx =) | fee] dx + f (b(y), y)—b(y) — f(aQy), y)——aly)
dy Javy) a(y) LOY dy dy
4.31 Consider the integral equation
it
wate) = f Ri(t, s)oi(s)ds, O<t<T
0
where*
_fl-\t-sl, |t-s|<1
Rts) Ua elsewhere

Find the eigenvalues and eigenfunctions over the interval (0, T) when T < 1.
4.32 Consider the integral equation
T
nat) = | Rn(t, s)bi(s)ds, O<t<T
0

where
6 ; :
2mit 21s
Bad Onda ances (=) cos ( 7 )

Find the eigenvalues and eigenfunctions over the interval (0, 7).
4.33 Consider the binary radar detection problem:
Ay: rH=—s@e)--n), —oo <t <\00

Ho: r@t)=nt), -~o<t<o

where the signal s(t) is nonzero in the interval (0, 7) and the noise n(t) is Gaussian with PSD
given by
4
Saf) = ; Watts/Hz
(2x f)* + 16
4. T. Kailath, “Some Integral Equations with “Nonrational” Kernels,” [EEE Transactions on Information Theory,
vol. IT-12, no. 4, October 1966, pp. 442-47.
 302 Coherent Communications with Waveforms Chap. 4

OPTION 1:

OPTION 2:
Figure P4.34

(a) Determine a suitable realizable whitening filter for the minimum probability of error re-
ceiver.
(b) The optimum receiver processes only that portion of the received signal lying within the
range: (i) [0, 7], (ii) [0, 00), (iii) (—oo, 0], and finally (iv) (—00, 00). Indicate your choice
and explain your answer.
4.34 A noise generator produces a WSS Gaussian random process x(t), whose correlation function
is

Rx(t) = exp(—3|t|)
You must construct a random number generator from x(t), which produces a new independent
random number generator every tenth of a second. You have two options for doing this, as
indicated in Fig. P4.34. Here the samples are instantaneous, ideal, and operate at ten samples
per sec. The variables indexed with the integer n occur at times 0.1n sec.
(a) Compute the correlation coefficient between X,(0) and X;(1).
(b) In option 1, what filter system function G(f) would you use to make w(t) a white Gaussian
random process with PSD S,,(f) = 1?
(c) In option 2, X;(n) is defined on the integers. Is X;(n) a Gaussian random process? Com-
pute E{X,;(n)} and E{X,(n + m)X;5(n)}.
(d) In option 2, what causal filter system function H(f) would you employ to whiten X,(n)
and produce Y,(n)?
(e) Are there any problems (either physical or theoretical) with these designs? Which design
do you prefer?
4.35 Consider the detection of two equilikely signals in additive zero-mean colored Gaussian noise
with autocorrelation R,,(t, s) where

Ae Th) =O) nt), 0S =h

EG TAC) SON sos OSes It

Suppose that the signals are given by

N

sot) = Y> VEndn(t)

n=

si(t) = —s0(t) = — )~ VEngn(t)

n=)

where ¢,(t)’s are eigenfunctions with respective eigenvalues A,, satisfying

i
J Ratt. dbn(s)ds = dnl, O<f= FonSlQecaN
0

(a) Show that the optimum decision rule is

N [E m 0

a ai “yz 0
n=1
An m
1
Problems 303

where ry, = fo r(t)dn(t)dt.

(b) Find an expression for the error probability as a function of Ay, A2,---, Ay and E plop), Goo
En. Simplify your answer for the special case where A, = No/2 for all n and E =
E, + E2+---+ Ey denotes the total signal energy.
4.36 A QPSK signal is characterized by

s(t)=J/P [d) p1(t) cos wet — dgpo(t) sin Wet|

for 0 <t<T where f, is the center frequency of the channel and T represents the symbol
duration.
(a) Determine the PSD if d; and dg are independent, identically, distributed random variables
taking on values +1 and —1 with equal probability.
(b) Repeat (a) if

NE
O=peO=1e
oa AEE
SST
0, elsewhere

(c) Repeat (a) if

1, OSn— 2
pit) = po =|1 LY DISS Oh
0, elsewhere

(d) Repeat (a) if

5 ial SE Dine - aah

Le 0, elsewhere PON) = » ests
0, elsewhere

4.37 An unbalanced QPSK (UQPSK) signal is characterized by

S(t) = J P)d, p;(t).cos

w-t — ./ Podopott) sin Wet

(a) Assume P;/Po = 4 and repeat Problem 4.36.

(b) Specify the transmission bandwidth required and discuss your answer from a system engi-
neering point of view if necessary.
4.38 Consider a data sequence which is Manchester coded.
(a) Show that the correlation function as defined by (4.375) is given by

1-340, |r<t
rs(t) = ie Rae,

(b) By direct substitution of the result of part (a) into (4.374), verify that the conditional error
probability for Manchester coding is given by (4.378).
4.39 Consider a data sequence formed by transmitting with equal probability one of two equal
energy pulse waveforms, that is, pj(t) and p(t) independently in each symbol interval. It is
easily shown that the decision rule [analogous to (4.367)] for this type of sequence is given by

choose p(t) if Q > 0

choose p2(t) if Q <0

where

Ts

oa| rt) {pi[t-—@- DT -3] - pa[t-—(@ — DT -8)} dt

(n—1)T+6
304 Coherent Communications with Waveforms Chap. 4

(a) Show that the conditional error probability for this type of sequence is given by

P(E/A) = serte{v (E/No)' [ru AT) + riui(T — AT) —raQT) — nit > aT)

ae *erfc {v (EY No) ii 1) Fra 2) Sry ke aT))|

[o.2}

an rfc {v (E/No)' [r22(AT) + r22(T — AT) — ry2(AT) — r21(T— aT))|

Co

af gerfe{v (E/ No)’ [r22(AT) + ry2(T aa XT) = rj2(AT) = age = aT))|

for
A>0

and

P(E/4) = gerte|V(E7Noy (ri (IAlT) + riv(Z — JAIL) — ri2(lalT) — rai(P — airy}

a gerfe{v (E/No)' [ru (All) + ria — |AIT) — m2(all) — ra2(P — airy]

38 gente|/(E/Noy! [ro2(lA|T) + rao(T — |AIT) — raQall) — ria — at7)]]

1
~ geite |/(E/Noy’ froatlAll) Aira e= ATE) =a a) — al7))]

forA <0

where

Jo pilt +t) pat

rie)=
PT
BRIN
/, Ves(E/No)
CO i aa
and we have made use of the fact that rj;(t) = rj;(—t) for alli, j = 1, 2.
(b) Show that if p;(t) = —p2(t), the expression for the conditional error probability reduces to
(4.374).
(c) For RZ data, evaluate r;;(t) for i, j = 1, 2, and, by direct substitution of these results into
the expression for P(E /A) given in part (a), verify (4.379).
4.40 For a Miller-coded data format, the appropriate decision rule, in the presence of symbol sync
error, for the n'" symbol is

a& = sgn[|Qil — |Q2l]

where

nT +5
Bij ~r(@)pi[t—(—1)T —8)at
(n—1)T+6

nT +5.
a= | _1(t)p2[t — (a — )T — 8]dt
(n—1)T+6

and
Problems 305

nioa{vF oxrer0 elsewhere

VP O<{t= 1/2
P20)=) —./P 'T/2<1 aT
0) elsewhere
Derive the expression for conditional error probability as given by (4.380).
4.41 Suppose the error probability is a function of random variable x, say P(E/x). Show that the
average P(E) = E{P(E/x)} can be well approximated by
2
P(E) = P(E/%) +“ P"(E/%)
where x is the mean of x.
4.42 Some BPSK modulators exhibit both imperfect phase switching and amplitude imbalance si-
multaneously. In this case, assume that the signal model is given by

2E : A
s(t) =,/ al + ed(t)] sin[w,.t + (1 — a), 4)

where d(t) = +1 and € and @ denote amplitude and phase imbalance, respectively.
(a) Assuming an AWGN channel, equally likely signals and ideal coherent demodulation/de-
tection, find the probability of bit error.
(b) Plot (on one graph) three curves showing the loss in effective E/ No as a function of a for
three values of €, 0.1, 0.2, and 0.3, for a ranging from 0 to 1.
4.43 A QPSK system operates in AWGN with an E,/No = 9.4 GB (resulting in bit error probability
of 107°).
(a) Find the bit error probability if there is a phase mismatch between the transmitted carrier
phase and that stored in the receiver of ¢ degrees. Assume the interval (0, 7] is known at
the receiver.
(b) Assume ¢ is 3°, 6°, and 10° and evaluate the necessary increase in E;,/
No such that the bit
error probability remains at 1075.
4.44 (a) Repeat Problem 4.43 if there is a misalignment of the transmit and receive clocks by € sec.
(b) Assume €/T = 0.05 and 0.1 and evaluate the required increase in E;/ No needed to con-
tinue to operate at a bit rate of 10~>.
4.45 As an approximation to the integrate-and-dump detector of Problem 4.7, consider replacing it
with a lowpass single-pole Butterworth (RC) filter with transfer function

H(f)=
: 2n
aed (3s |

where f;_ gp is the 3 — dB cutoff frequency.

(a) If y(t) is the output of this filter, find the signal-to-noise ratio SNR = [E{y(T)}P/02 ofa
sample of this output taken at time t = 7. Assume that either a single positive or negative
pulse of amplitude A was transmitted and that the filter capacitor is initially discharged.
Also, assume an AWGN channel with single-sided PSD No.
(b) Find the value of the time-bandwidth product f3-¢,7 that maximizes the SNR found in
(a).
4.46 A binary communication system has an input modulation characterized by an antipodal pulse
stream m(t) = pa eek dn p(t — nT) where the a,’s take on the values +A with equal proba-
bility and p(t) = F—'{P(f)} is a known waveform, where F —!{.} denotes the inverse Fourier
transform. The modulation m(t) is transmitted over an AWGN channel with single-sided PSD
306 Coherent Communications with Waveforms Chap. 4

n(t)

m(t) —e(+) an
sample
at t=nT

Figure P4.46

No and the data is to be detected by the receiver configuration shown in Fig. P4.46. Suppose
that P(f) is an ideal rectangular filter with bandwidth equal to half the data rate, that is,

P(f)= VT, \fl<1/@T)

0, otherwise

(a) What is the average power of m(t)?

(b) What are the mean and variance (under each of the two hypotheses) of y, = y(nT)? Do
these quantities depend on n, that is, is there any intersymbol interference (ISI)?
(c) What is the average error probability of the receiver?
4.47 Suppose that in Problem 4.46, the transmitted pulse shape corresponds to a cosine roll-off
characteristic, that is

VT, 0<|fl<
GU -a)
P(A)=4 VT cos{Z [2 fIT -1+e]}, +d-a@)<Ifl<
+ +a)
0, otherwise

Repeat (a), (b), and (c) of Problem 4.46 for this filter.
4.48 A binary communication system, with signals so(t) and s(t) with average power of P watts,
is observed in AWGN (No watts/Hz single-sided). The data rate is 1/T bits/sec and the binary
messages are independent and equally probable.
(a) Draw the optimum coherent waveform receiver.
(b) What is the bit error probability?
(c) Assuming an additive interference characterized by [ (t) = /2A cos(w-t + ) (where ¢ is
a uniformly distributed random variable and A is known), find the optimum receiver for
(k —1)T <t <kT in AWGN.
(d) What is the bit error probability of the receiver used in (a) subjected to AWGN and the
interference characterized in (c) if so(t) = —s,(t)?
(e) What is the bit error probability of the receiver found in (c) subjected to AWGN and the
interference characterized in (c) if so(t) = —s,(t)?
4.49 Consider the bandpass filter illustrated in Fig. P4.49(a), which exhibits an approximate symme-
try about the center frequency f,; the approximation improving as the Q of the resonant circuit
is increased.
(a) Determine the transfer function H(f).
(b) Determine the symmetric part of H(f) about f. = V1 — (1/4Q72)
f,.
(c) Determine the antisymmetric part of H(f) about f..
(d) Given the I-Q modulation-demodulation_ system illustrated in Fig. P4.49(b), develop ex-
pressions for /o(t) and Qo(t) when 6 = 6 and H(f) = 1.
(e) Repeat (d) when H(f) = 1 but 6 #0. Comment on the system engineering meaning of
your result. ‘m
(f) Repeat (d) when 6 = 6 and the bandpass characteristic H (f) is symmetrical about f, (that
is, >> 1):
Problems
307

(a)

MODULATOR DEMODULATOR

I (t)
BANDPASS
CHANNEL

Q (t)

sin (2, +6) sin (o,t+6)

(b)

Figure P4.49

Si2@ %9@ @sg @S.

811@ %10@ ecs7 es

Figure P4.50

4.50 Consider the 16-QAM signal set shown in Fig. P4.50. Assume all signals are equilikely and
that the noise is AWGN with PSD No/2.
(a) Write an I-Q waveform model using rectangular coordinates for square I-Q pulses.
(b) Write an J-Q waveform model using polar coordinates for square I-Q pulses.
(c) Find the probability of error in terms of d and No.
' (d) Find the probability of error in terms of the “peak energy” E, and No.
 308 Coherent Communications with Waveforms Chap. 4

REFERENCES
1. Wozencraft, J. M. and I. M. Jacobs, Principles of Communication Engineering, New York: John
Wiley and Sons, 1965.
2. Papoulis, A., Signal Analysis, New York: McGraw-Hill, 1977.
3. Strang, G., Linear Algebra and Its Applications, 3rd ed., San Diego: Harcourt Brace Jovanovich,
1988.
4. Bracewell, R., The Fourier Transform and Its Applications, 2nd ed., New York: McGraw-Hill,
1978.
5. Gagliardi, R., Introduction to Communications Engineering, New York: John Wiley and Sons,
1978.
6. Weber, C. L., Elements of Detection and Signal Design, New York: Springer-Verlag, 1987.
7. Lee, P. J., “Computation of the Bit Error Rate of Coherent M-ary PSK with Gray Code Bit
Mapping,” JEEE Transactions on Communications, vol. COM-34, no. 5, May 1986, pp. 488-91.
8. Irshid, M. I. and I. S. Salous, “Bit Error Probability for Coherent M-ary PSK Systems,” JEEE
Transactions on Communications, vol. COM-39, no. 3, March 1991, pp. 349-52.
9. Korn, I and L. M. Hii, “Relation Between Bit and Symbol Error Probabilities for DPSK and FSK
with Differential Phase detection,” submitted to EEE Transactions on Communications.
10. Reed, I. S. and R. A. Scholtz, “N-Orthogonal Phase-Modulated Codes,” JEEE Transactions on
Information Theory, vol. IT-12, no. 3, July 1966, pp. 388-95.
11. MacWilliams, F. J. and N. A. Sloane, The Theory of Error-Correcting Codes, New York: North-
Holland, 1977.
12. Lindsey, W. C. and M. K. Simon, Telecommunication Systems Engineering, Englewood Cliffs,
N.J.: Prentice-Hall, 1973. Reprinted by Dover Press, New York, NY, 1991.
13. Van Trees, H. L., Detection, Estimation and Modulation Theory—Part I, New York: John Wiley
and Sons, 1968.
14. Papoulis, A. , Probability, Random Variables and Stochastic Processes, New York: McGraw-Hill,
1965.
15. Cooper, G. R. and C. D. McGillen, Methods of Signals and System Analysis, New York: Holt,
Rinehart and Winston, 1967.
16. Helstrom, C. W., Statistical Theory of Signal Detection, 2nd ed., Oxford: Pergamon Press, 1968.
17. Wainstein, L. A. and V. D. Zubavov, Extraction of Signals from Noise, translated from the Russian
by R. A. Silverman, Englewood Cliffs, N.J.: Prentice-Hall, 1962.
18. Brown, J. L., “Detection of a Known Signal in Colored Gaussian Noise—A New Coordinate
Approach,” JEEE Transactions on Information Theory, vol. IT-33, no. 2, March 1987, pp. 296-
98.
19. Nguyen, T., “The Impact of NRZ Data Asymmetry on the Performance of a Space Telemetry
System,” JEEE Transactions on Electromagnetic Compatibility, vol. 33, no. 4, November 1991,
pp. 343-50.
20. Simon, M. K., K. Tu and B. H. Batson, “Effects of Data Asymmetry on Shuttle Ku-Band
Communications Link Performance,” JEEE Transactions on Communications (Special Issue on
Space Shuttle Communications and Tracking), vol. COM-26, no. 11, November 1978, pp. 1639-
Bl:
 Chapter 5
Noncoherent Communication
with Waveforms

So far, we have concentrated on communication scenarios where a message is transmitted

using a distinct signal that is precisely known to the receiver, except for the corrupting noise.
In Chapter 4, it was shown that coherent demodulation involved mixing the received signal
with inphase (I) and quadrature (Q) sinusoidal references, whose phases were exactly equal
to the incoming received phase. It was shown that lack of knowledge of the incoming phase
can result in cross-talk between the I and Q arms, the amount of which depends on the value
of the carrier phase error ¢, that is, the phase error between the locally generated and the
incoming phase.
In many applications, it is not possible to perform coherent detection due to the in-
ability of the receiver to produce meaningful carrier phase estimates, due, for example, to
significant oscillator phase noise instability. In some scenarios, coherent systems are unde-
sirable due to the additional complexity imposed by the phase estimation circuitry. In this
chapter, we investigate noncoherent waveform communications in AWGN channels. First,
we derive the structure and analyze the performance of optimum and suboptimum receivers
in random phase channels. Both orthogonal and nonorthogonal signals are considered and
the effects of imperfect time and/or frequency synchronization are assessed. We later extend
the concept to both random amplitude and phase channels in which the amplitude of the in-
coming signal is also random due to fading in the link. Communications in both Rayleigh
and Rician channels are examined and their respective performances evaluated.
In noncoherent M-ary communications, the received signal can be expressed by
mo—> Ho: r(t)=so(t, V)+n(t), O<t<T
mA: r(t)=s(t,¥)+tn(t), O<t<T
(xl)

Miyjan
— Hypa Ft) = si 8) +n@), Ox<t<sT
where the vector W denotes the vector of unknown parameters [1] that are not or cannot
be estimated in the receiver. As before, the noise is white Gaussian with one-sided PSD No
Watts/Hz and T denotes the duration of the transmitted message. We assume for now that
perfect symbol synchronization has been achieved. Later, we will consider and assess the
error probability in the presence of symbol sync error. Note that due to the randomness of
the vector W, each message is transmitted using a sample waveform of a random process
rather than a deterministic and known signal. As far as the receiver is concerned, it does not
 310 Noncoherent Communication with Waveforms Chap. 5

distinguish between an unknown random phase introduced in the channel or a random phase
inserted at the transmitter. A message m; can be conveyed using a set or a family of known
waveforms, rather than a fixed and unique waveform. Note that for a fixed vector W, the
received waveforms are known precisely to the receiver. Hypothesis testing as described by
(5.1) is typically referred to as “nonsimple” detection as opposed to the “simple” detection
in coherent communications.

5.1 NONCOHERENT RECEIVERS IN RANDOM PHASE CHANNELS

In random phase channels, the phase @ of the received signal is not known, nor estimated by
the receiver. All the other parameters suchas amplitude and frequency are assumed known
and hence, the vector W of unknown parameters consists solely of the scalar 0, that is,
W = 0. The received signal, assuming m; is transmitted, is thus given by!

r(t) = V2a;(t) cos(wet + @i(t) +O) +n(t), O<t<T (5.2)

where w, is the carrier frequency, a;(t) the amplitude information, ¢;(t) the phase informa-
tion’, @ the unknown phase, and n(t) the bandpass Gaussian noise process given by (4B.33)
as

n(t) = V2 {n¢(t) cos Wet — ns (t) sin wet} (5.3)

The baseband noise components n,(t) and ns(t) are independent Gaussian processes with
PSDs given by

Sa Pick Se (Pe Y otherwise

0
IflsB (5.4)
Now according to the bandpass sampling theorem [2], n(t) is completely characterized by
samples of n-(t) and ns(t) taken every 1/2B sec, that is

n(t) = V2 »e sinc(2Bt — k){ne(ty)

Cos we(t — te) — Ns (th) Sin@e(t — t%)} (5.5)
k=—oo

Moreover, we note that

eS) > i a
/ n(t)dt=— {n2(4) Lsne(tn)}
2
(5.6)

represents the energy in the sample function n(t). The joint pdf of the vector n = (n¢(t}),
Nc(tz), ...,Nc(T), ns (ty), Ns (t2),...,Ns(T))! created from sampling a T-second segment
of n(t) is given by
PNoi 6 VM) 2
=) (Ne (tk) + 5 (tk))
Fne(ty)--sns(T)(M) = C exp | — NB : (5.7)

1. As in Chapter 4, we denote the received random process by r(t) and a sample function of that random process
by p(t). Similarly, the received random vector is denoted by r and a sample vector of that random vector by p.
2. In Chapter 5, ¢;(t) denotes the phase information when message m, is transmitted, but it also denotes the ith
basis function of the K-L expansion (consistent with the notation used in Chapter 4). The reader is therefore alerted
to the dual use of ¢;(t) and special care is taken to redefine ¢;(t) whenever it is used.
Sec. 5.1 Noncoherent Receivers in Random Phase Channels 311

since the 2 x 2BT samples are uncorrelated, Gaussian each with variance NoB. The con-
stant C is the normalization factor such that the multidimentional integral of the density is
unity. Using (5.6) to convert from sums of samples to a time integral, we have
1 T
Friclty),....ns(T)) = C exp {-5 i moar (5.8)
No Jo
which is identical to (4.7). Assuming that message m; is transmitted and conditioning on the
random phase 6, then
Sr(p/m;i, 8) = fn(p — si/mi, 0) = fn(e — Si)
Dy

S(OUS.qae p
fy (0) = V2ai(t) cos(wet + b(t) +)) at] 5.9)
No

and
8

fr(p/mi) = fr(e/mi, 9) fojm;(8/mi)dé (5.10)

Assuming that 6 is random and independent of which message is transmitted, fg/m;(9/mi)
= fg(@) and (5.10) simplifies to
51

fr(p/mi) = fr(p/mi, 9) fo(@)ae (5.11)

Since the receiver does not attempt to estimate the incoming phase 6, it can be assumed
random with a uniform pdf, that is
1
fu(0) = |3 Fh ae (5.12)
0 otherwise
Expanding (5.9), fr(p/mi, @) can be expressed as
ZL E;
fr(e/mi, 0) = C’ exp | a cos@ — Zs; sin } exp (-=) (5.13)
No No
where E; = he a?(t)dt is the energy of a;(t) and the I-Q projections generated by means of
a local oscillator are
ip
oat = p(t)a;(t) V2 cos(wet + oj (t))dt (5.14a)
0
f.
isi = / p(t)ai(t)V2 sin(wet + pi (t))dt (5.14b)
0
As a special case when the transmitted information is contained in the amplitude only,
then r(f) = J/25;(t) cos(w-t +0) +n(t), that is, dj(t) = 0 Vi and s;(t) = a;(t) and (5.14)
can be expressed as the I-Q projections
r
Zci =i! Pc(t)sj(t)dt (5.15a)
0
‘}
Zsi ah Ps(t)sj(t)dt (5.15b)
0
312 Noncoherent Communication with Waveforms Chap. 5

V2 cos
opt

ro

rt)
Figure 5.1 Demodulation with random
V2 sin wt sie

where

pelt) & p(t)V2.cos wet (5.16a)

ps(t) & p(t)V2 sin wet (5.16b)
are the projections [inphase and quadrature (I and Q)] of the received waveform p(t) onto
the orthonormal basis {./2 COS Wet, 2 sin wet}, as depicted in Fig. 5.1.
Transforming z,; and zs; of (5.14) into polar coordinates using the envelope &; and the
phase ¢j as

Zci = § COS Oj (5.17a)

Zsi = &; sin Qi (5.17b)

then
Zci COSA — Zs; sind = (& cos g;) cos 8 — (&; sin g;) sin@ (5.18)
= § cos(0 + gi) |
and (5.13) becomes

2&; E;
fxs 8) ~ exp |= os + 4) |exp a (5.19)
No No

Averaging over the density of 0, we have from (5.11) and (5.13)

| : 2&; E;
i i a Isexp |
Sr(p/mi) mel cos( 0 + 00|
: poy =)
adexp ( sta (5.20)

Such an integral has been encountered in several engineering applications and is referred to
as the “zero-order modified Bessel function of the first kind” [3, 4]. Specifically, it is defined
through

Lutte
f(x) 8 = | e* SsBap (5.21)
=U;

and is plotted in Fig. 5.2 as a function of x. For positive x, it is a monotonically increasing

function and because of the periodicity of the cosine, an arbitrary phase @ can be added to
its argument without altering the value of Jp(x), that is
1 ua

Ip(x) = on /: e* sb+) 78 (5.22)

Sec. 5.1 Noncoherent Receivers in Random Phase Channels 313

Ig

Figure 5.2 Plot of /o(x)

Hence, using (5.22) in (5.20), the optimum decision rule sets m(p(t)) = mx when

ee
0 Noe (-=)
No (23)

is maximum for k =i. Recall from (5.17) that the envelope &; is given by

ag aie: (5.24)
314 Noncoherent Communication with Waveforms Chap. 5

where z,; and Zs; are the projections of p,(t) and ps(t) onto the i signal as given by (5.15a)
and (5.15b), respectively. For equal energy signals, the factor exp(—E; /No) can be ignored
and since J9(2&;/No) is a monotone increasing function, we need only to maximize &; or gE?
where

EP = 22, + 25; (5.25)

The optimum receiver for equal energy signals is depicted in Fig. 5.3 where the
received signal is first mixed with sine and cosine waveforms at the incoming frequency and
the resulting outputs are correlated with each of the M possible transmitted signals.> The
receiver then forms the sum of the squares of the sine and cosine correlations and decides
on the message with maximum resulting envelope. It is clear that the decision process
involves nonlinear operations unlike the coherent receiver discussed in Chapter 4. This is
the penalty paid for not knowing or estimating the incoming carrier phase 9. An alternative
implementation can be obtained with matched filters with impulse responses

hit) =s(F 0); 90)

yea (5.26)

followed by a sampler at multiples of T seconds as shown in Fig. 5.4. When the signals can
be expressed as
N
T(E —)) se), 2 0, Vane (5.27)
j=l

as given by (4.40), then it is possible to correlate r,(t) and r;(t) with the basis functions
gj(t),i =1,2,---,N and apply the appropriate signal weights to form the various Ee.
i=0,1,---,M — 1, as shown in Fig. 5.5.
A significantly different approach for optimum detection of equiprobable equal en-
ergy signals is possible with envelope detectors. Consider Fig. 5.6, which depicts a filter
hpp,x(t) followed by an envelope detector. Since the input to the filter r(t) is centered at fo,
let the filter hp» ,(t) be matched to the signal s,(t), translated to frequency f, (“bp” stands
for bandpass), that is

Nop a(t) & V25,(T — 1) cos@et, k=0,1,-:-,;M—1 (5.28)

The output of the filter becomes

mit) = | r(t) App p(t — t)dt

= (5.29)
= v3| r(t)sx(T —t +7) cos@.(t — t)dt

or

Ug(t) = UcK(t) COS Wet + Usx(t) SIN Wet (5.30)

3. Note that in Fig. 5.3, lowpass filters have been inserted at the output of the mixers. As long as the bandwidth
of these filters is much larger than the data rate, they do not impact the performance of the optimum receiver. It is
however good practice to include them in an analog implementation as the mixers are highly nonlinear devices that
generate more than the desired product at their output.
 Sec. 5.1 Noncoherent Receivers in Random Phase Channels 315

V2 cos w,t

DOAYFDYrVEOO

V2 sin w,t : z (see note)

Lg —sfifoa
foe |fot
M-1

wy,

Sy (t)
Note: From here on, a "comparator" selects the message
corresponding to maximum input value or voltage.

Figure 5.3 M-ary noncoherent receiver for equal energy signals

where we define
lo 2)

Uck(t) a [ r(t)sp(T —t + t)V2coswetdt =| ro(t)sx(T —t+t)dt (5.31a)

—oo Ae, 9)

CO lo @)

Usk (t) 4 | r(t)sy(T —t+ t)V2 sin Wetdt =f rs(t)sx(T —t+1)dt (5.31b)
—o0o —0o

Since r-(t) and r;(t) are low-frequency signals and s,(t) is also a slowly varying signal,
both u,,(t) and us,(t) remain relatively constant over several cycles of cos wet. As a result,
ux(t) is a bandpass signal and can thus be expressed as

up(t) = Re fax(rye /e"} (5.32)

where the complex baseband waveform a,;(t) is given by

a(t) = Ucg(t) + jUuse(t) (5.33)

316 Noncoherent Communication with Waveforms Chap. 5

Cc
O
M
P

A
V2 cosw,t
R
A
7
O
R

V2 sin wt

Figure 5.4 Alternate implementation of M-ary noncoherent receivers for equal energy signals with
matched filters

The output of the envelope detector becomes

ex(t) = lax (t)| (5.34a)
or

ex(t) = \/u2,(t) + u2,(t) (5.34b)

If we sample the output at t = T, we have from (5.34) and (5.31)

ex(T) = ,/u2,(T) + u2,(T)

il 2 T ; (5.35)
= (/ re(t)(e)d) + (/ reou(t dr)
0 0
which is identical to &, as given by (5.24) since u¢x(T) = Zcx and usx(T) = Zsx aS given
by (5.15a) and (5.15b), respectively. Hence, the optimum M-ary noncoherent detector for
equiprobable, equal energy signals can be implemented using matched filters (to the band-
pass rather than the baseband signals) and envelope detectors as shown in Fig. 5.7. Receivers
that examine only the envelope of the matched filter output are termed noncoherent or inco-
herent receivers as they do not exploit the knowledge of the carrier phase; receivers that do
exploit knowledge of the carrier phase are called coherent.

5.1.1 Optimum M-FSK Receivers

A signal set of great interest in noncoherent communications consists of M sinusoids at

distinct frequencies and the set is referred to as Multiple Frequency-Shift-Keying (M-FSK)
signals. In general, the frequency separation can be nonuniform, but equidistant signals are
 Sec. 5.1 Noncoherent Receivers in Random Phase Channels 317

, ()

SIGNAL
; WEIGHTING =
: MATRIX | -
ot M
V2 cos w,t h

Soho} ier [e® sooo uta A

r(t) = T
¢, (t) O

% OX J(at
V2 sin @,t > (t)

) SIGNAL
: WEIGHTING
; MATRIX
(1
.
©

MATCHED
FILTER
Nop x!) t=T
Figure 5.6 Envelope detection

of great practical interest as they occupy the minimum required bandwidth for orthogonal
noncoherent communications. The transmitted signals are characterized by

S(t) =V2P cos(@ct + at +0), OK<t<T (5.36)

where «, denotes the frequency of the k'” tone and @ the unknown random phase. The latter
is independent from message to message because each signal is typically generated by a
unique oscillator whose initial phase is random when the signal is selected. The pairwise
signal correlation coefficient, as defined by (4.45), reduces to
1 Lf 1 Ls ;
= ot si(t)sj(t)dt = al cos(n
fi,jt+0 — 0 )dt 6:37)

where E = PT, fi,; = fi — fj is the frequency separation between the tones, and 6, @’
318 Noncoherent Communication with Waveforms Chap. 5

BANDPASS
MATCHED
FILTERS :
ENVELOPE So
DETECTOR

ENVELOPE
r (t) ‘ DETECTOR

DOHAYFDFYFVEOON
ENVELOPE
DETECTOR i

task
Figure 5.7 Envelope detector noncoherent receiver for equal energy signals

are the random carrier phases when signals s;(t) and s;(¢) are transmitted. Because of
the random phase @ — 6’, the only way to guarantee zero correlation is for the integral to
be carried out over at least a full period of the difference sinusoid. Hence, the minimum
frequency separation to guarantee orthogonality is
1

This separation is equal to twice the minimum separation required for coherent M-FSK
transmission, 1/(27), derived earlier in Chapter 4. Given M tones, the minimum bandwidth
required for orthogonal noncoherent transmission is M/T Hz achieved by uniformly spac-
ing the tones with 1/7 Hz separation. The received signal, assuming that message myx is
transmitted, is given by

r(t) = V2P cos(2zfe tt 2 (=)t+ 6) nd) = OSt=T (5.39)

Typically, the tones are symmetrically spaced around the carrier f-, which can easily be
modeled by (5.39) by considering negative values of k. The optimum receiver is still as
shown in Fig. 5.3 where the various signals are the corresponding sinusoids. The receiver
still computes &2 for each signal using gE?= a + Zz where Z,-; and Zs; are given by (5.14)
with aj(t) = VP Vi and ¢;(t) = a;t Vi, that is
a
en ae | r(t) COS(@e + @;)tdt (5.40)
0
and
ip
Boh = V3 | r(t) sin(@e + @;)tdt (5.41)
0
Hence

il 2 T 2
E? =2P (| r(t) COS(@e + wear] +(/ resin tonya) | (5.42)
0 0
Sec. 5.1 Noncoherent Receivers in Random Phase Channels 319

SPECTRAL
r() CORRELATOR FOURIER OBSERVATIONS
TRANSFORM
i
®, (6+)

D>
>
0
DOA
VUs0

Figure 5.8 Highpass spectrum analyzer receiver

Using the procedure outlined in Appendix 5B, (5.42) is identically equivalent to

ap T-t
é? Sap [ if r(t)r(t + T) cos(@e + w;)tdtdt (5.43)
0
Defining the finite-time autocorrelation function of r(t) as
T—t
oy(t) A / r@)rt+tdt, O<1t<T (5.44)
0

then, (5.43) can be expressed as

T
&?—4P / gr (T) COS(We + @;) TdT (5.45)
0
Following the argument in Appendix 5B, the quantity gE? can be viewed as the Fourier
transform of the even function ¢,.,(t) [constructed from ¢,(t) as in (5B.10)] evaluated at
the discrete frequency f. + fj; thus

} =4P%.,(fet fi) (5.46)

where ®, ,( f) denotes the Fourier transform [5] of ¢.,-(t). This interpretation of the deci-
sion process leads to the the highpass spectrum analyzer receiver in Fig. 5.8.4 In practice,
the received signal is first downconverted to an equivalent lowpass waveform v(t) using
a reference at the carrier frequency. Thus, any practical implementation must be provided
with an estimate of f,. This heterodyning operation reduces the highpass spectrum analyzer
receiver of Fig. 5.8 to its lowpass equivalent illustrated in Fig. 5.9, with the corresponding
observations {®,(i/T), i=0,1,---,M — 1}.

5.1.2 Suboptimum M-FSK Receivers

Exact evaluation of the spectral observations is difficult when one elects to implement
the spectrum analyzer receivers of Figures 5.8 and 5.9. However, it is possible to ap-
proximate the spectral estimate ©,(f) which then results in a receiver performance that
is suboptimum. There are various approaches to approximating the spectral observations
(OAK ak Onl e+, .A4 —1}.
First, one can evaluate the autocorrelation function @,(t) directly and then evaluate
the Fourier transform to obtain ®,(f). This is referred to as the autocorrelation function
4. In Figures 5.8 and 5.9, the Fourier transform is operating on the resulting even function, as discussed in
Appendix 5B.
320 Noncoherent Communication with Waveforms Chap. 5

A SPECTRAL
®v(f) |OBSERVATIONS
CORRELATOR abet ve ;
oy(+)

as)
3
0)
|=
(@)—5
28)
(ONS)
3

Figure 5.9 Lowpass spectrum analyzer receiver

method. A practical realization of this method that is convenient for digital processing is
discussed in [6, 7] and summarized here. Since the Fourier transform of the random process
u(t) is essentially zero outside the frequency interval | f| < M/T, then in accordance with
the sampling theorem we have, to a good approximation
2BT
v(t) ~ vat) Sov (=) sinc [28 («= (5.47)
n \

where B A M/T and vy, = v(n/2B) is a sample of v(t) at t = n/2B. In terms of the above,
the autocorrelation function of vg(t) is given by
Tt 2BT 2BT a
Ont) i Ug(t)ug(t + t)dt = ey = Un UmSiNC E («+ =) (5.48)
tH=lm=l)

which approximates that of ¢,(t). We observe that for r = k/2B, k an integer, the above
expression simplifies to

bulsp) =| Bame! Untntts k=0,41,42,---,4QBT—1)

1 2BT —k
(5 49)
otherwise

Since the spectrum of ¢,,(t) is limited to the frequency range |f| < B, one obtains by
means of the sampling theorem
2BT-1 k k

dy, (T) = a Pv4 (5) sinc ED («= =a) (5.50)

k=—(2BT-1)
Taking Fourier transforms, the spectrum of v(t) is approximated by

1 pera k kf
v(f) vg (Sf) Fp oral Mies 2B Pv, (55) -5( B ). BS fea B (5.51)

These expressions for the correlation coefficients and for the approximate spectral observa-
tions ®, (k/T) form the basis for the computational procedure.
Second, one can calculate the spectrum by the discrete Fourier transform approach.
The latter takes samples v(k/2B), k = {1, 2,---,2BT} and evaluates
2BT E 2BT
X(N=) (55)00s(“L) aw rin= >» (<5) si (=) (5.52)
 Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 321

ENVELOPE
DETECTOR

SAMPLE ONCE EVERY T SEC

ZERO
r(t) THRESHOLD 3>
COMPARE

ENVELOPE
DETECTOR

Figure 5.10 Suboptimum binary FSK receiver

and then computes the spectrum

®5(f) = X7f) +7) (5.53)

which yields the spectral observations ®,(k/T), k = {1, 2,---,2BT}. The discrete Fourier
transform method permits greatly simplified analysis [8].
Third, the receiver can be implemented with a small computer programmed with
the Cooley-Tukey Fast Fourier Transform algorithm [9] and processing the output of an
analog-to-digital converter. The Fast Fourier Transform (FFT) approach produces results at
frequencies that are multiples of 1/7 provided that the number of samples is a power of two.
Although the FFT method does not give better results than the discrete Fourier transform
or the autocorrelation function method, it has the advantage that considerably less time is
required to perform numerical operations.
Suboptimum detection also occurs when there is a bandpass limiter prior to the detec-
tor. In many applications, bandpass limiters are used to provide some form of automatic gain
control (AGC) in the receiver and because of their inherent nonlinearity, they degrade the
performance of the detector. Suboptimum detection of binary FSK signals in the presence of
a bandpass limiter is discussed in [10], where it is shown that a 0.5 dB degradation occurs
ioe Jee ES 10-2. Another form of suboptimum detection occurs when the integrate-and-
dump filters are replaced by bandpass filters as shown in Fig. 5.10 for binary FSK signals,
symmetrically located around the carrier frequency. In this case, the receiver performance is
degraded and is a function of the bandpass filter bandwidth, as discussed later.

5.2 PERFORMANCE OF NONCOHERENT RECEIVERS IN RANDOM PHASE

CHANNELS
It is of great interest to assess the performance of noncoherent systems and compare it to
their coherent counterpart. First, we will derive the performance of orthogonal signals and
determine the resulting penalty paid in not tracking the carrier phase at the receiver. We
will then generalize the results to any arbitrary set of signals and in particular, consider
322 Noncoherent Communication with Waveforms Chap. 5

the effect of frequency and timing synchronization errors. A bound on the performance
of orthogonal M-FSK signals is then presented in the presence of frequency error. Finally,
the performances of two suboptimum M-FSK receivers are evaluated and the effect of time
domain truncation in evaluating the spectral observations is assessed.

5.2.1 Performance of M-ary Orthogonal Signals

The M-ary receiver for equal energy signals is depicted in Fig. 5.3 and the receiver
picks the message mx corresponding to maximum &? for i=0,1,---,M —1. In or-
der to compute the probability of symbol error, we need to characterize the joint pdf
2 2 : 5 A =
Fe2 2,62, (602 815° Sup): Assuming that message m, is transmitted, then r(t) =
J 25,(t) cos(@-t +0) +n(t) with
i if
z= / re(t)s)(t)dt = if(sg (t) cos + n¢(t))s;(t)dt (5.54)
0 0

and
T if
i / rs(t)sj(t)dt = it(sg (t) sinO + ns(t))s;(t)dt (5.55)
0 0

For a fixed 0, the conditional means become

i :
ales ., | £scos@,. tk
E(z<i/8) = 0080| Sk(t)sj(t)dt = te i xk (5.56)

Similarly
T : P
2 Bei). qi
{Zsi/0} = sin@ [ Sk (t) sj (t)dt 8 ay (5.57)
E : = F —_ Ss ’

The conditional variances are computed using

ip 2)
Ni
oF 16 =E |
¢ ne(t(ar |= = Es (5.58a)
0

and
aR 2
N
of 16 =E 10 na(tsi(at = > Es (5.58b)
0
which are also the unconditional variances. Recall that n,(t) and ns(t) are bandlimited
processes; however, assuming that BT >> 1, they can be viewed as nonbandlimited using
the matched filter argument previously outlined in Chapter 4. Since n,(t) and n,(t) are
independent, z-; and zs; are conditionally uncorrelated and hence independent since they
are conditionally Gaussian, that is

iewee : y/@0) ==
1 {-@ — E{zci/6})? + (y — ai"
Rt E;No

| iZ+k (5.59)

I 2?+y?+E2—2E,z cos6—2yE, sin 6 f

TEsNo “XP {eeepc ans |weak
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 323

Consider now the pairwise correlation among the various z,; and zs;, i =0,1,---,
M — 1. Assuming that m, is transmitted, then
ifs T
E{zciZej} = E {|
/ (sg (t) cos @ + nels | (sg(tT) sin 6 + ne(e)sj(e)at]|
0 0
Ae T
E {(/ ne(si(a (| ne(a)sj(e)dr) | (5.60)
0 0

=0, tf ji Fk, jGFk

Moreover
ip IP
E{Zck2cj} = E {(z.cos 8 in| ne(tsu(t )(f ne(n)sj()de) | (5.61)

Fie 4:
Using similar arguments, it can be easily shown that
E{zsiZsj} = ElzsiJE{zsj}=0, iF j
(5.62)
E{zcizZsj} = E{zci}E{zsj}=0, t=j
which holds true even when i = k or j = k but not both. Therefore, the variables z,; and Z;;,
i =0,1,---,M —1 are uncorrelated and hence independent since they are conditionally
Gaussian. Defining 0; = tan! (Zsi/Zci), then (5.59) becomes

I tye
wEsNo ©*P |-54 , ixk
Sccis2si Cap y/0) = 1 z-+y*+E?—2E, [224-2 cos(0—6;) Pee k (5.63)

7EsNo ©XP )— EsNo os gai

Integrating over 9 and using the definition of /(x) as given by (5.22), (5.63) reduces to
1 mem ;
wean? {as | Lat
Feet
s2si2 Y) = | z+y?+E? In {2 z+y? eh: (9.64)
mEsNo =P { E;No 0 No moh

Making the change of variables &| = ety a fee, then

25 ex p |-=, ; ixzk
iZ
ACh
J&j a tte
2& eee)
Bese. I 25
(5.65)
ix=k
EyNo OP E;No O\ No)? ‘7

The Rayleigh and Rician densities of (5.65) can also be obtained using the results of Ap-
pendix 5A, namely, (5A.21) and (5A.27) with 0? = NoE;/2 and sex Ee. Let us proceed
by normalizing é? using
BA
Aj = — (5.66)
E;No

then f,4,(a) is given by

fa;(@) — V EsNoa)
Se (VEsNoa) (5.67)

2S (a/EsNo)
324 Noncoherent Communication with Waveforms Chap. 5

or

faa) =} xP {-(2+ 5)} (4h) u@, i= k (5.68)

Es Es Fe

exp {—a} u(a), LEK

where u(a) is the unit step function. The variables A;, i = 0, 1,---, M — 1 are also inde-
pendent since they are nonlinear functions of independent random variables [11]. Hence

Lt ep Wan LI8 aj,°°° ,amM-—1) = fo (G0) fA; (a1) acre fAy_\(am-1) (5.69)

From this joint pdf, we can now evaluate the probability of symbol error P;(E). Assuming
that m, is transmitted and that the output of the envelope detector matched to s,(t) is Ax,
then the probability of correct symbol detection conditioned on message m, and Ax is given
by
P;(C/mx, Ax) = Prob {Ao < Ax, Ai < Ak, ---, Ag—1 < Ax, Ati < Ak, ++, Am—1 < Ax/my}
M—1 pA,
(5.70)
=[T] | fa@da=— ey"
i=0 —oo
itk
Hence

zs as - (©27))
= exp {=} f e “*Io (sia (1 - ese dag

Using the expansion

M—1
Uae eer e (lence (5.72)
=0

where y—1C was defined in (4.153), (5.71) reduces to

E M-1

Pa(C/m) = exp |=} SY Ota

m=(0
(5.73)

x i} exp {—(m + |)ax} Io (ec) dag

0 No

Using the integral?

je exp {-a?x?} Io(bx)dx = Bi ex

; ; ~ Ig b
dae 5.74
aa
5. This integral is derived from the more general integral [12]

coaT np, dpe TF(4)(5) 2

v

i mn (ate 2peP (vy+1) exp hides

4p 1 Fy Fy (aac
5 +1,v+1; a ap
2
Upon substitution 44 = 2, v = 0 and noting that ; F\(0, 1; x) = 1, the integral of (5.74) is obtained.
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channeis 325

or equivalently (by letting u = x)

(ove) 2
/ exp {-au| Ip (bu) du = 730%?|E ! (5.75)
Fe 4a2
then (5.73) is independent of m, and the probability of correct symbol detection P,(C) is
given by
M-1
PRQ= DD nCn eyexp {- (=) (5.76)

The probability of symbol error, P;(E), is given by 1 — P,(C), or

£ at petal m Ee
aa Seite iGmenp | (=)| (5.77)

The symbol error probability is depicted in Fig. 5.11 versus bit SNR [Ep/No = (Es/No)/n
where n = log, M] for various values of M. Since the signals are orthogonal, the bit error
probability P,(£) is related to P;(E) through (4.96) and is plotted in Fig. 5.12 versus
Ep/ No.
Consider now the case of coded versus uncoded transmission. The symbol error prob-
ability performance of a noncoherent system that transmits an n = log, M bit symbol by
sending one bit at a time (so-called “uncoded transmission’’) is given by

P,(E) = 1— (1 — P,(E)|m=2)" (5.78)

with

P,(E)|
b M=2 pe
>= 5 exp -=)
2No 5.79
(5.79)

obtained from (5.77) by setting M = 2 and E; = Ey. The uncoded symbol error proba-
bility, as given by (5.78), is illustrated in Fig. 5.13 for various values of n as a function
of bit SNR. Coded and uncoded symbol error probabilities are compared in Figures 5.14
and 5.15 for n = 5 (M = 32) and 10 (M = 1024), respectively. Also shown in these two
figures are results given in Chapter 4 for coherent communication employing orthogonal
and biorthogonal signals. For orthogonal codes, we observe that the type of detection per-
formed, whether a coherent or noncoherent detector is mechanized, is not very critical for
error probabilities less than 10~*. As a matter of fact, the use of 5-bit biorthogonal codes
and coherent reception is not appreciably more effective (given a certain energy-to-noise
ratio) than the use of orthogonal codes and noncoherent reception. On the other hand, a
comparison of coded versus uncoded transmission indicates that coding the transmitted sig-
nal into n-bit orthogonal symbols (rather than transmitting an n-bit symbol one bit at a time)
is more effective in noncoherent communications. For example, if five bits of information
are sent with a symbol error probability of 10~? (typical for telemetry systems), the use of
orthogonal codes and coherent detection will reduce the signal-to-noise ratio per bit, that
is, Ep/No, by approximately 2.5 dB under that required for equivalent performance using
bit-by-bit detection. Under the same symbol error probability conditions, the use of orthog-
onal codes and noncoherent detection will reduce the required E,/No by approximately 6
dB under that required for equivalent performance using bit-by-bit detection. For ten bits
 Noncoherent Communication with Waveforms Chap.5
326

15
10
9
8
if
6
5
4
3
2
1

Figure 5.11 Symbol error probability of ideal noncoherent receiver (M = 2”) as a

function of bit SNR (Reprinted from [28])

(M = 1024) and P,(E) = 1073, orthogonal and biorthogonal coding and coherent recep-
tion reduce the required E),/No by approximately 5 dB under that required without coding.
Finally, using noncoherent reception, 10-bit orthogonal codes and P;(E) = 10-3, one real-
izes an improvement of approximately 8.5 dB over the uncoded noncoherent procedure of
bit-by-bit detection.
Finally, by a limiting procedure similar to that employed in Chapter 4, it may be
shown (see Problem 5.3) that the asymptotic (M — oo) error probability performance of the
optimum noncoherent receiver is identical to that for coherent communications as described
by (4.182).
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 327

10°

102!

10m
a a
a

10°?

{o>%

107"
=41(0' 20

Figure 5.12 Bit error probability of ideal noncoherent receiver (M = 2”) as a function
of bit SNR (Reprinted from [28])

5.2.2 Performance of M-ary Nonorthogonal Signals

Nonorthogonal signals can occur in either one of two scenarios; the signals can be chosen
nonorthogonal at the transmitter or the latter sends orthogonal signals but due to imperfec-
tions in the receiver, the signals become correlated and thus nonorthogonal. The second of
the two cases occurs as the result of frequency or timing error in the receiver and will be
treated in more detail in a later section. The key difference between the two cases in terms
of analysis complexity is that when the signals are nonorthogonal at the transmitter, the re-
ceiver correlates the observed waveform with the transmitted signals and the noises at the
output of the integrate-and-dump filters are correlated and dependent. On the other hand,
 Noncoherent Communication with Waveforms Chap. 5
328

—
aAawonarn

(E)
P,

20

Figure 5.13 Uncoded symbol error probability performance of an ideal noncoherent

receiver (Reprinted from [28])

when the signals are nonorthogonal due to frequency or timing errors, the receiver correlates
the observed waveform with the original transmitted orthogonal signals and consequently,
the noises at the output of the integrate-and-dump filters remain uncorrelated and hence in-
dependent in an AWGN channel. Assuming nonorthogonal signals, the received waveform,
when message m; is transmitted, is given by

r(t) = V2a;(t) cos(@ct + (tf) +O) +n(t), O<t<T (5.80)

where aj;(¢) and ¢;(t) are the amplitude and phase information bearing components of signal
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 329

UNCODED

107!
ORTHOGONAL
CODING

BI-ORTHOGONAL
105" CODING

ny
fatase

ORTHOGONAL
CODING
10+.

— — NONCOHERENT
RECEPTION
-4
oe COHERENT
RECEPTION

19%

Figure 5.14 A comparison of coded and uncoded, coherent and noncoherent, symbol
error probabilities (M = 32) (Reprinted from [28])

s;(t) and n(t) is AWGN* with one-sided PSD No Watts/Hz. It is convenient to use complex
notation to denote the transmitted signals. Let

Oya (Hel? (5.81)

be the complex baseband signal associated with signal s;(t). Then

6. It is easier to derive the results when the noise is white. For bandlimited noise, the results are identical, but the
derivation is tedious.
 Noncoherent Communication with Waveforms Chap. 5
330

UNCODED

ORTHOGONAL
CODING

ORTHOGONAL OR
BI-ORTHOGONAL
CODING
w
a”

== —— NONCOHERENT
RECEPTION

COHERENT
RECEPTION

Figure 5.15 A comparison of coded and uncoded, coherent and noncoherent, symbol
error probabilities (M = 1024) (Reprinted from [28])

HOIERE {V25;wet) (5.82)

constitutes the signal component of the received waveform when message m; is transmitted.
The signal s;(t) can be expressed in an alternative form as

oo()= /2 {aj (t) cos $j (t) cos(@ct + 6) — a;(t) sin p(t) sin(@ct + 6)}
(5.83)
= J2 {a¢i(t) cos(wet + 8) — asi(t) sin(wet + 6)}
where
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 331

aci(t)
4 aj(t) cos ;(t)
(5.84)
asi(t) A a;(t) sin g(t)
We will assume that the signals have equal energy E,, that is
r T
E,& / s?(t)dt = / a?(t)dt
18 Ik
(5.85)
:
0 0
Define the complex correlation coefficient /;j [13] as the correlation between any two
complex baseband signals 5;(t) and 5j(t) as (let * denote complex conjugate)

3 Nar Ree
py Zi Si(t)5; (t)dt (5.86)

Substituting (5.81) into (5.86), then

bij = ij + JBij (5.87)

where the real part of the complex correlation is given by
il if

Ga sa aj(t)aj(t) cos(g;(t) — o;(t))dt

- (5.88)
1
= — / (deci(t)aej(t) + asi (t)as;(t)) dt
Es Jo

while the imaginary part is

Vee
Bij & =i aj(t)aj(t) sin(g(t) — pj (0))dt
oo (5.89)
1
= =| (asi (t)acj(t) — aci(t)asj(t)) dt
Es 0

Note that a;; = 1 and 6;; = 0. Furthermore, a;; = aj; and Bj; = —f;;. In the special case
where the signals are orthogonal, then a;; = B;; =0 for i # j. Recall that the optimum
receiver still computes the decision variables &? for k = 0, 1,---, M — 1 where a 2+
ae and
ii
Zeck af r(t)V 2az (t) COS(Wct + Pk (t))dt (5.90a)
0
T
Ze & / r(t)V 2ag(t) sin(wet + Px (t))dt (5.90b)
0
It can be easily demonstrated that E? is independent of the carrier phase angle @ of the in-
coming waveform (see Problem 5.4). Therefore, we will set @ = 0 from here on to simplify
the analysis. Substituting (5.80) into (5.90a) and (5.90b), we have

Zck = Esayi + Nck (5.91a)

Zsk = Es Bei + sk (5.91b)

332 Noncoherent Communication with Waveforms Chap. 5

where
T
Neck & ifn(t)V 2a (t) COS(Wct + by (t))dt (5.92a)
0

i
ys / n(t)V 2ag(t) sin(wet + ox (t))dt (5.92b)
0

It is easily shown that n-, and ns, are zero-mean Gaussian random variables with variance
(No/2) Es and cross-correlations
N
E {ncknej} = = Es Qk; (5.93a)

Ni
E(nsknsj} = > Eset (5.93b)
No
E{ncknsj} = a EsPri (5.93c)

Alternatively, (5.91a) and (5.91b) can be written in normalized form as

OE.
Zck = ,| —— Aki + Xck (5.94)
No

Dade
Zsk = ,| =— Bei + Xsk (5.94b)
No
where x-, and x;, are normalized unit variance random Gaussian variables with

E{xXckXej} = Okj (5.95a)

E{xsxXsj} = Kj (5.95b)

E{xckxXsj} = Bj (5.95c)

Define now the matrices A and B as

Loos ay

A= Ee. bE Ses with aj; = aj; (5.96)

ji l
and
0 Bij
Beet prs (ae with Bj); = —Bji (5.97)
Be

and the vectors X¢, Xs, Ze, and Z, as

Xc0 Xs0 Zc0 Zs0

Xcel Xs Zcl Zs1
X= A ae. ; 2 ange Fes (5.98)

XcM-—1 XsM—-1 ZcM-1 ZsM—1

Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 333

and @; and B; as

QO; Boi
QI Bii
ial and p;=| (5.99)
QM—li Bu -ii
Then. the pdf of the augmented vector (z7 z!), assuming message m; is transmitted, is 2M-
multidimensional Gaussian given by (3.61) as

DER.
fantevimy=0((¥) tei) (5.100)

where the 2M x 2M matrix C is given by

Are
es (31 ) (5.101)
and is the covariance matrix of the augmented vector (z, zs)’. The probability of correct
decision, assuming that message m; is transmitted, becomes

pccimy= ff JJ J] J] i of]

°((5)-Va(3)4) -
dzm—1dym-—1---dzj4+1dyj41dz;-1dyj_-1 ---dzidy,dzodyodzidy;

where R; is the region within the circle of radius Aa + hie centered at the origin. The
unconditional probability of correct symbol detection P;(C) becomes
M-1
PC) =), P.(C)m;) PGni) (5.103)
i=0

where P(m;) is the a priori probability of sending message m;. The 2M-fold integral in
(5.102) cannot be simplified any further in general. In the following sections, we will con-
sider scenarios under which P,(C) can be computed in closed form or in terms of well-
known functions.

5.2.3 Performance of M-ary Equicorrelated Signals

For M-ary signals, Nuttall [14] derived the probability of error for the case when the signals
are equicorrelated and furthermore, the correlation is real and nonnegative, that is

% ee teae oF | 104
Pl; a, ij anda real and nonnegative Chel

Hence, the imaginary component f of the complex correlation is assumed zero. Nuttall
[15] provides a heuristic argument (not a proof) that for a given magnitude of the complex
correlation, the assumption of zero angle (or zero imaginary component of () realizes the
334 Noncoherent Communication with Waveforms Chap. 5

minimum error probability. It is shown [14] that the probability of correct symbol detection
is given by
le,@) [o.@) ]
exp (-3 (x? +»)
P;(C) = (1 — @) exp (-=) | if xy

(5.105)
x o( = - es)Io (Vaxy) [1 — Q (Jay, x)]""' dxay
If a = 0 in (5.105), the integral on y is unity resulting in

lee) 2 aa 2 M-\
pc) =exn (-2) [ reso (-) 0 ([Fe] [1-ex(-5)] dx (5.106)

Using the expansion in (5.72) and the integral of (5.74), (5.106) reduces to (5.76) as ex-
pected. If, on the other hand, a 4 0 but M = 2, use the integral [15]

oo Is > a ac b
[ s exp (-3 (s 4+¢ ))Io(cs) Q(as, b)ds = Q (Ss. =) (5.107)
in (5.105) to obtain

1a | ee
P,(C)
= (1 — @) exp (-=) r exp (-3" (1 —a@)
0

; aba o( ar Fr Je
x r —a - '
"VV No ee wher
(5.108)
af | — a”) exp (-=) [xe (-3* (1-«))

x IIo |x meal
No yy
( — eT (1 — QO (ax,
a?) x)) d xdx

Helstrom [16] integrated that expression to obtain

Ee [Bs
epee BW ieee: i ee
fC) a o(\2No (1 a”), 2No (1+ ; =))
(5.109)

ae (-Be Veale
ane a («m5
which is exactly 1 — P;(E) where P;(E) is the probability of error for two nonorthogonal
signals and will be given by (5.133).

5.2.4 Performance of Binary Nonorthogonal Signals

For the binary case, the probability of error can be simplified much further as follows. The
receiver compares & and & where & = ,/z2,+z?, and z¢i, zsi for i =0, 1 are still given
by (5.94). Mutiplying z.; and zs;, i= 0,1, by /2E;/No, we have (let a = a@19 = ao; and
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 335

B = Bio = —Bo1)

2E; DE
U0 = No Lc) N + UcO

(5.110)
2E,
UVs0 = iNeaee = Us0

and

2E 2E
vei A a lS yao t eel
(5.111)
2E 2E
V5, 4 nar yb + us

assuming mo was transmitted. The Gaussian noise variables, u.0, Uso, Uci, and us},
are zero-mean, unit variance with E{u-ouci} = E{usous1} = (2Es/No)a, E{ucousi} =
—E{u.iuso} = (2E;/No)B. Letting &; A ,/ v2, - v5 we know from Appendix 5A that the
marginal densities of €,9 and €,; are Rician, but due to the nonzero correlation of the Gauss-
ian noises, €)9 and €,; are not independent. The probability of error, assuming message mo
is transmitted, is equal to the probability that €,; exceeds £,9 or

P;(E/mo)
= Prob(&y1 > &v0/mo)= [ i feof(X, y)dydx 6.112)

which requires the computation of the joint pdf f¢,,¢,,(x, y). This approach is outlined by
Helstrom [16] who determined the joint pdf and computed the double integral in (5.112).
Another approach to the computation of the probability of error, taken by Proakis [17],
is based on the observation that a _ ef is a special case of a general quadratic form in
complex-valued random variables, defined by
1b,
A= > (Ailxel? + Arlyel? + Asxeyy + A3xzy«) (5.113)
|g |

where A), A2, A3 are constants and {x,, yx} are pairs of correlated complex-valued Gauss-
ian random variables.
The approach followed in this book (also see Viterbi [18]) is based on whiten-
ing the observations v¢9, Us0, Ucl, Us} SO that the resulting variables Vy.cO, Uw,s0, Uw,cl;
and vy,s1 are uncorrelated. Consequently, the variables &, 9 4 a DE + Cae and &y.; A

[v2 Alaa v2 «| are Rician and independent and the probability of error reduces to the prob-
ability that one Rician variable exceeds another independent Rician variable. Defining the
vector X as

x? =(ve9 Usd Ver si)? (5.114)

then, the mean and covariance of x are given by
336 Noncoherent Communication with Waveforms Chap. 5

(2E;/No) I
= 0 Nae) Ze
(feat 0 se
mx | (2E,/No)a i or a
(2Es/No)B B
and
i 0 a £B
= 2E;s 0 1 —B (64g
= (%) eee 0 (5.116)

B a Oa
Due to the symmetry of the problem, it is easily shown that P;(E/mo) = Ps(E/m1) =
P;(E). Hence, the error probability is given by

P,(E) = Prob(é2, < 2.) = Prob(v2y + v2) — v2, — v2, <0) (5.117)
or

P,(E) = Prob (x" Ox < 0) (5.118)

where
RES Pe 1)
Cintas ol
alle aes eh (5.119)
00), 0,1
In order to whiten the vector x, we need to perform a linear nonsingular transformation J
on x which diagonalizes Kx, that is

Xo x (5.120)

One such transformation is given by

S121)

where

k=,/1— p?-a? 122)

with the result (see Problem 5.5)

Kye TR Tet (5.123)

tf
x! Ox = x" aes su = (5=)k(xx7,OXw) (5.124)
Hence, the probability of error becomes

P;(E) = Prob(x’ Ox < 0) = Prob(x! Ox, < 0)

(5.125)
= Prob; coat Uast Une ate
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 337

where Vy,c0, Vw,s0; Vw.cl, aNd Vy.s; are unit variance uncorrelated Gaussian (and hence
independent) random variables with means

2(Es/No)k(k + 1)
Bieter pixy 2 ols No/Es 0
Be Ba Decpt \i, (hs) Node oa
2(Es/No)kB
The optimum receiver compares ee 9 to ae ,; Where g2 ; 1s the sum of the squares of
Vw,ci and Vy,s;. The pdf of &,; is Rician with parameter s;, that is

ee
Fé, (E /mo) = & exp (-! i 1g (sig) (5.127)

where

E
56 = E*{vw,co} + E7{vw,so} = (=) (+k)
% (5.128)
st SE arc tes (east = (=) (bk)
No
Hence, the probability of error becomes

Ps(E) = Prob(éw,1 = Ew,o/mo)

(5.129)
| fey, dydx
= Fewo(®)

which is similar to (5.112) except that the variables &, 9 and &,,; are now independent. It
can be shown that [19] (see Problem 5.6), given two independent Rician random variables
Ro and Rj, each with parameter s; and variance oe (i = 0, 1), then the probability that Ry
exceeds Ro is given by

P = Prob(Ro < R})

Oo,2 b SISO
=Q (va, vb) _ oe exp fad Io (ab) ( )
a6 +o; 2

where

a=
ae
as eer and b=
eae
5) 5) (5.131)
Og + OF 09 +O;

In our case, a; = Ort = and

(57132)
2No: 1+kh=
E, Es
1 ay 1 — ir)
~
jij rik) = ( |p|?
338 Noncoherent Communication with Waveforms Chap. 5

with p =a + jf. Hence, the probability of error for two nonorthogonal signals is given by

P,(E) = o( + /1- a)
(5.133)
1

which is identical to 1 — P,(C) where P,(C) is given by (5.109). The probability of error
is illustrated in Fig. 5.16 for several values of ||. It is clear that orthogonal signals (that is,
p =0) minimize P,(£). In this case, P;(£) reduces to

jike 1 Ee
| ‘|— 5 exp (-)
P,(E)=@Q (0. (5.134)

But since Q(0, x) = exp(—x*/2), then

= (5.135)
P;(E) zt i ie eee
= = -—exp|{
= exp |-—— } — =exp[— — :
: PONG)! Oo N ONy ONE
which is identical to (5.77), for M = 2, or equivalently, (5.79) with E; = Ep.
Consider the cross-correlation between two FSK signals at frequencies fo and f/f).
Then, the cross-correlation between the corresponding complex baseband signals is

Nps
p=— | V2E,/Tel?"' /2E,/Te lat
2Es 0
- (5.136)
¥ a pidnfo- ft, — IA ASOIT
|jxforT
PI6 m fo,1T
and

5 sin 2 fo,1T
(A= |e 5:137
m fo,1T ( )
The performance for any frequency separation fo,; can then be computed by first evaluating
the corresponding ( and then picking the corresponding curve in Fig. 5.16. In particular, if
fo,1 = 1/T, then |p| = 0 and P,;(E) reduces to (5.135).

5.2.5 Effect of Frequency Error on Orthogonal M-FSK Detection

When the received frequency is not perfectly known, the observed signal, assuming message
m; is transmitted, is given by

r(t) = V2P cos (2n(fe+ fit Af\tt+O)+n(t), O<t<T (5.138)

where Af is the error in carrier frequency. Alternatively, the received signal can be inter-
preted as a carrier f- shifted by the appropriate signal frequency f; + Af, rather than fj.
This scenario is illustrated in Fig. 5.17, where the dashed lines denote start and end of inte-
gration times, with the depicted frequencies. The inphase integrator output z,%, matched to
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 339

10s

110i

40>°

(E)
P,

100

Figure 5.16 Probability of error for noncoherent detection of two nonorthogonal signals (Reprinted from [17])

signal s,(t), becomes’

T
zo f r(t)v2P cos (2n(fce
+ fx)t) dt
0
7 (5.139)
— re) cos (21 (fe + fi + Af)t) cos (2 (fo + fx)t)dt + nex
0

7. We can set 6 to zero without loss of generality.

340 Noncoherent Communication with Waveforms, Chap. 5

(a)
RECEIVED
FREQUENCIES

(b)
LOCAL
FREQUENCIES

Figure 5.17 Effect of frequency error

where n-x is a zero-mean Gaussian random variable with variance o2 = (No/2) Es. Simpli-
fying, (5.139) reduces to

E Ie
Zck sin (27 (fix + AFT) + nex (5.140)
~ On(fix + Af)
where again, fj, denotes the frequency difference between messages i and k, that is

Sika hi— fh (5.141)

Similarly, the quadrature integrator output z;, is given by

Pee ‘| r(t)V2Psin(2( fe + fa)t)dt

0
(5.142)
=5
Thea (o0s (2n(eefin + AFIT)
—1}+i
The envelope statistic & = Rize “2 Ze will then be Rician distributed with parameter se
given by
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 341

P 2

oe aan) {sin? (27 (fin + Af)T) + (cos (27 (fi,x + Af)T) — 1)°|

aa mT) {2 —2cos (27(fix + Af)T)} ‘n e

5.143
: das 2
a

= (seaarap) bs ra
+ann)
a £ : ae

Normalizing z,x and zs5x% by 1/o = ./2/(NoEs), the parameter se is then normalized by
2/(NoEs) and since f; =i/T for orthogonal signals, then

PN)
f= (2) eG =Ee et
aie? CEO)
eye (5.144)
5.144

where ¢ denotes the normalized frequency error [20] given by

pSAFT (5.145)

From (5A.27), the density of & is given by (note that 0? = | after the normalization)

2 QEs/No) fo, &) 2E;

Sa(E) = § exp |-(5ieees) Ip (: iG 0] (5.146)
p 2, No

where we let

1
fpGbHA sin? (1 (i — k + p)) (5.147)
m?(i —k + p)?
First, note that the detector matched to the incoming signal suffers from signal attenuation
equal to

5, i
ue a sin’ (sp) (5.148)

which, as expected, reduces to unity if o = 0. Simultaneously, loss of signal orthogonality

occurs as a result of signal spillover into the remaining M — | detectors; hence, the nonzero
means and the resulting Rician densities. Note that for zero frequency error (that is, o = 0),
then

On ford 54k
fg. = ae sin? (x (i — k)) = ly Morr=k oy)

and signal orthogonality is restored. In this case, the Rician densities reduce to the Rayleigh
densities derived earlier. Despite loss of orthogonality, the variables &, &|,---, €y—1 remain
independent since the Gaussian random variables resulting from the noise integration are
still independent as the local signals remain orthogonal. In this case, the probability of
correct symbol detection, assuming message m; is transmitted, is given by
342 Noncoherent Communication with Waveforms Chap. 5

P,(C/m;) = Prob {iSnares, m

ms Ole Ae |

= 1) Se (Xi) ra: fe (X0) fe, (01) °> + Fe; Oi-1) fei i+) - >
0 0 0
M-1 integrals } (5.150)
x fey _\(Em—1)dxodx1 +++ dxj—dxi41 ++ dxm—14dX;
le.@) M-1 Xj

=} Fe; (Xi) H | Sem (%m)dXm | Xi

m=0
mA

Using the definition of the Marcum Q-function as given by (5A.15), we have

[ MAE ea ae (=0 f3(.m).3;) (5.151)

The probability of symbol error, assuming message mj is transmitted, is then given by

in
P.(E/m;) =1 -| fe, (Xi) I] (:215 (a f3¢.m).x)) dx;
2Es 42. (5.152a)
0 m=0
mAL

and the unconditional probability of symbol error becomes

; Ma!
P(E) == 2 P(E/m;) (5.152)
In order to compute the bit error probability, one can assume that the signals remain
orthogonal in the presence of frequency error, which facilitates the use of the well-known
result, P)(E) = (M/2(M — 1))P;(E), relating bit and symbol error probabilities of orthog-
onal M-FSK. Unfortunately however, this assumption is not valid in the presence of fre-
quency error and hence, the bit error probability results found in [20] are approximate. In
fact, to properly evaluate the bit error probability in the presence of synchronization errors,
one must specify an appropriate mapping, for example a Gray code, of the symbols to bits.
In the perfectly synchronized case, the bit error probability performance is completely inde-
pendent of the symbol-to-bit mapping since all errors are equally likely to occur. To compute
the average bit error probability, one must first compute the probability of a particular sym-
bol error for a given transmitted message. Analogous to (5.150), the probability of choosing
m, when message mj is transmitted is given by

Psx(E/mj) = Prob {& = max ém, m=0,1,---,M— i|

ee) M=1 px (33853)
-[ Se (Xk) H | fen (X%m)dxXm |dx~, k Ai
m=0
mx#k

If (k, i) denote the difference between the codewords (bit mapping) assigned to messages
(symbols) m; and m,, that is, the number of bits in which the two differ, then the average bit
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 343

error probability is as given by (4.86), that is

l M-1M-1
Wo (k, 1) Pse(E/mi) (5.154)
BLO iy arecerray i=—0 k
k rs
It is clear that if a symbol error occurs, it is more likely to occur in an adjacent frequency
than in any other. Thus, a Gray code mapping, in which two codewords from two adjacent
symbols differ by only one bit, is appropriate to this type of modulation. Figure 5.18 depicts
the average bit error probability versus E,/No in dB with p as a parameter for binary, 4-ary
and 8-ary FSK and the conventional Gray code assignment given in Table 4.1.

5.2.6 Effect of Timing Error on Orthogonal M-FSK Detection

When the received frequency is known precisely but the symbol epoch is not, the receiver
integrates using its own estimate of the symbol epoch which is offset from the true epoch
by Ar sec. This phenomenon is depicted in Fig. 5.19 where the integration overlaps two
successive symbol intervals. This results in signal attenuation in the detector matched to
the incoming frequency and moreover, loss of orthogonality due to signal spillover into the
remaining detectors [8]. The received signal can be modeled as

V2P cos(2a(fe+ fi)t +0) +n(t), O<t<T

Tih (S155)
WV2b .cos(2fe + ft +0) +n), 2st s27

where we assumed that message m; is transmitted followed by message m ; and that phase
continuity is maintained. Since the local epoch estimate is not perfect, the receiver integra-
tors operate from At to At + T to obtain at the k"" detector (setting 6 = 0 without loss of
generality)
At+T
Zeck = if r(t)V2P cos(2m(fo + fx)t)dt
At
A
=2P {| cos(2 (fe + fi)t) cos(2m (fe + fx)t)dt (5.156)
At
At+T
+ / cos(2m(foe+ fy)t) cos(2a(fo + fx)t)dt —+ nck
T
and
At+T
Zsk =f r(t)V2P sin(2a(fo + fx)t)dt
At
;
=2P {| cos(27 (fe + fit) sin(21 (fe + fx)t)dt (5.157)
At
At+T
+ / cos(2m(fe+ fy)t) sin(2m (fe + fonat| + nsx
de

Normalizing z-x and zs, by 1/o, the pdf of &; can be expressed as

E> fx(i,i.)
(2Es/No) im Ame
 ALA MU
T
MTabANE I
TT
344 Noncoherent Communication with Waveforms Chap. 5

i
LE a SO HA
1072 EN

ET
PEAS
tT
10

10

10

10
10 11 12 13

Figure 5.18a Bit error probability of M-FSK with frequency error, M 2

Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 345

! N
-o

awe

105

107

107
10 11 12 13 a 15 16

Figure 5.18b Bit error probability of M-FSK with frequency error, M =4

346 Noncoherent Communication with Waveforms Chap. 5

o ! N
—
°
bl
oO

100°

10

1077

Figure 5.18¢ Bit error probability of M-FSK with frequency error, M = 8

Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 347

where

LC
aay es 1
rears ‘
sin? [7(i 1
— (1 — A) + Gab? sin” [1(j — KA]
< 1 ol ee 4 ae ;
POG Sh {x [@ — fj) -— (Gi — kal}
, (5.159a)
rer h)sin” {7 [@ —k)A— (j —k)]}
1
+ G2 HG Tm SY LE -D- GO, ixk, j#k
i ® ey 1 oy A
OS =e m(i —k) ;

5.159
elec
TOG OK

eG =) Ae sin Ao
‘ (i al : Ces) (5.159c)
+a — jf) —A)snQzrG-j)A+1))} ifs

ARG.
Die
B= pH
ae
[sir(i
Sao) b
— HL —)
5.159d
sgt Ac ak)
— mA(i —k)sin(2wrA(Gi —k))}, ifXk

Gi, . Vi (5.159e)
and i denotes the normalized timing error given by

At
Pa
Ships (5.160)
As acheck, we have

1
heats ened {2+4 2) 1 — 2)?
K Oe eee : 2 — 2cos(2mxa
CALS asp woa) ater ioe (5.16 1a)
4 Anx(1 — A) sin(2rx(A + 1)) — 4ax(1 —A) sin(2xx)|
which is 0/0 ! Using L’ Hopital’s rule once, we have
I
lim f2G, j,i) = lim —— {[8x?x(1 — 2)? + 2(2mA)sin(20xA)
imj x30 872x

+ 4n(1 — A) (sin(2ax(a + 1)) + x20 (A + 1) cos(2rx(a + 1)))O-161b)

— 4r(1 — A)(sin(27x) — x(277) cos(27x))

which is still 0/0 ! Applying L’Hopital’s rule one more time, we obtain
348 Noncoherent Communication with Waveforms Chap. 5

1
lim f2@, j,i) = lim — {8771 —Aa)*?+4nr(1-—A) +4204 1)
i>j x0 87
(5.161c)
— 4n(1 —2)4m + 8272"
=i
as expected. Using a similar procedure, it can be shown that

lim FG, kk) = 1 (5.162)

i>

From (5.152), the probability of symbol error, assuming message m; is sent followed by
message m j;, is then given by
ee) 2 2E, N, 2 Pr Oe

(2h ew
«to (s efttasd) (5.163)

ia 2B ya
m=0 No
mA

The unconditional probability of symbol error, P;(£), is then

(ee!
P,(E) = = 2 Ps(E/mi)
(5.164)

As was the case for frequency error, the presence of timing error produces a lack of
orthogonality which results in the symbols not being equally likely. Hence to compute the
average bit error probability, one must once again compute the probability of a particular
symbol error for a given transmitted message. Analogous to (5.153), the probability of
choosing m, when message m; was sent followed by message mj, is given by

Pal (mi)
sk /m;,mj; = [exp
= 4 Xj ©Xp |— (2+= Ba. j.) |to (mn = 120.
5) Noes ) 0 | Xi No es )

oh e (5.165)
x I] 1i 0 (FAG. jem.) a
m=0
m#k

with f,(i, j,m) as in (5.159). Finally, the average bit error probability is analogous to
(5.154)
M-1M-1M-1
P(E) = W108, M ; (k, i) Psk(E/mj, m;) (5.166)
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 349

(a)
RECEIVED
FREQUENCY
AND TIME

At At -— At

(b)
LOCAL
FREQUENCY
AND TIME

Figure 5.19 Effect of timing error

Here again, the evaluation of (5.166) will depend on the mapping of the symbols to bits.
For the conventional Gray code mapping of Table 4.1, Fig. 5.20 plots the average bit error
probability versus E,/No in dB for binary, 4-ary and 8-ary FSK with 4 as a parameter.
Digital computer simulations were used to confirm some of the results.

5.2.7 Effect of Timing and Frequency Errors on Orthogonal M-FSK Detection

When both the incoming frequency and the symbol epoch are not known [21], the received
signal is still given by (5.155) with f. replaced by f. + Af, as depicted in Fig. 5.21. The
inphase and quadrature integrator outputs become
T

cen=2P || cos(an(fe+ Af + fi eosn(fe-+ fia

ss
At+T
(5.167)
+ i) cos(2a(fc + Af + fj)t) cos(2m(fo + fx)t)dt }+ nek
7
and
Ih
cn =2P |[ cost2a(fe + Af + fi)SinnCfe + far
os
At+T
(5.168)
ota / cos(2a(fe + Af + fj)t) sin(2m (fe + fit + Ask
T
 AIL IE
350 Noncoherent Communication with Waveforms Chap. 5

ee
in a

é:

I
ULINN
l

(oS EN
Gane ee
ns
BS
ae ae
e

a Ee SAN ae eB
r=0 XN A= So

5
10
aan
2 a el ST

SSS

10

=
10
-7 ieSA NN
10 11 12 13 14 15 16

Figure 5.20a Bit error probability of M-FSK with timing error, M= 2

 FAN
GE

C
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 351

EM ALLL
10-1

SENT
4=0.3

100-2
2

j
10%

eI IIE
ZiT r=0 2

rail
ae 10c

105"
OIE =0.1

pee
10-8

10-
10
iN
Wee
zit
(AO
NY, 11 a 13
i SIMULATION POINTS
14 15 16

Figure 5.20b Bit error probability of M-FSK with timing error, M = 4

 : HiL
352 Noncoherent Communication with Waveforms Chap. 5

07

UMC PTL
aa

iiaHE

Olme

07°

10-
11 12 13

Figure 5.20c Bit error probability of M-FSK with timing error, M = 8

Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 353

fj + Af fj + Af

(a)
RECEIVED
FREQUENCY
AND TIME

FREQUENCY
AND TIME

Figure 5.21 Effect of both frequency and timing errors

Normalizing as before and following a similar procedure, we have

£2 (2E;/No) f7,@, 3,4) Py Wb ae

fig(E) = & exp |-(5+ — lo (6 foae 1% ) 6.169)
where
i eee te nD ae in
forG, j,k) = aS
7 aay sin’ [z(t —k + p)(1 —A)]

e
Se in fe (jf — Kk + 2)A]
i Gee pp eo i
1 (5.170)
{cos(2z(i Fu phyB)
+ 292 —k
+p)G —k +p)
— cos(2m(i — j)) — cos(2a((i — j)A —j —k — p))

+ cosQar((i —k + p)k—U —k+p)))}, p#0,A 40

If o = 0, then Eat J, k)\p=o0 reduces to iG, j,k) as expected. Similarly, if A = 0, then
fe aG, J, k)la=o reduces to iG, k). The probability of symbol error is still given by (5.163)
and (5.164) with f?(i, j,k) replaced by f>,(i, j,k). The probability of bit error is still
given by (5.165) and (5.166) with FG J, k) replaced by eas j,k). For the conventional
Gray code mapping, Fig. 5.22 depicts the average bit error probability versus E,/No in
354 Noncoherent Communication with Waveforms Chap. 5

dB for binary, 4-ary and 8-ary FSK with both p and A as parameters. Digital computer
simulations were used to confirm some of the results. Note that when timing and frequency
errors occur simultaneously, the losses are not additive.

5.2.8 A Bound on the Performance of Orthogonal M-FSK Detection

in the Presence of Frequency Error
A number of years back, Jim K. Omura developed a Chernoff-type bound on the error prob-
ability performance of certain types of M-ary communication systems. This unpublished
result has particular application in noncoherent M-FSK communications.
In this section, we apply a slightly generalized version of the bound to predicting the
error probability performance of noncoherent M-FSK with frequency error. Upper union
bounds for this performance have been previously obtained in [22] in terms of the exact re-
sult for the performance of binary (M = 2) FSK with frequency error (see Problem 5.7). The
latter is expressed in terms of the Marcum’s Q function, which in general is cumbersome to
compute. Here, we shall derive a simple-to-compute bound on this performance that will en-
able system comparisons to be readily made. The result is obtained in a form that is similar
to the exact error probability performance of noncoherent M-FSK with no frequency error.
Since, as mentioned above, Omura’s bound was never published but rather privately
communicated to the authors, Appendix 5C presents the derivation of the bound in its
generalized form. Assuming that message m, is transmitted, then the detector matched to
Jn produces &, with pdf as given by (5.146), that is

£2 (2E;/No) f2(n, n) aoe

2. ©) =e exp |-(5a cele Io (: = a; ») 63191)

while the remaining M — | detectors produce independent &;’s with

£2 (2Es/No) f2(n, i) DE.

fie) = Fe |- (5+ See} Io (: No 3d) (5.172)

As required by the results in Appendix 5C, we need to evaluate the characteristic functions
of the pdfs in (5.171) and (5.172). In particular, letting 8; = ae then

ees i ef fe,(E)dE
@ Gees, (5.173)
= exp ———
1 — 2x No 1—2x

and

w(x gad) af—0o eferas

UEC ioe: (5.174)
ele P No 1—2x

The bound on P;(E/my,) is given by (5C.8) and repeated here for convenience
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 355

10°-

105°

P(E)5
i} >

as
PN
rie ss
Zz
as
10
7a ae eee

Sa
Saar
Be
aes
law wee
EeRar a
(ae Pee

10
Hie ae
ae
ee

PS
[ea
Dee
a
a fr ey
eee
ae
eee

ae
eee
l
ETT
MOAI
UTI
TT
VTA
ii se Se

noth) ae
Ep
Ns , dB

Figure 5.22a Bit error probability of M-FSK with both frequency and timing errors, M = 2 and p = 0.1
356 Noncoherent Communication with Waveforms Chap. 5

UII ti
TL VAT PME
IN
= oO

IAAT
°
a
'o
I

Lan N.
- Oo
wo

— Oo
©

- oOI!
~

10 12 13 14 15 16 17

Figure 5.22b Bit error probability of M-FSK with both frequency and timing errors, M = 2 and p == ()22
Sec. 5.2

at
Performance of Noncoherent Receivers in Random

A Phase Channels 357

5
Mal
ry
Ik
Biiiee TL
A
euil

a-
=a
oo

107 —————
aaa
Rae
Seay
er

i
107
if

—=4
Hi a

Aas!
iT
nou

Hi SIMULATION POINTS
4
[our

Son
“og is
oa
Fa

10%
10 14 15 16

Figure 5.22¢ Bit error probability of M-FSK with both frequency and timin g errors, M 4andp=0.1
358 Noncoherent Communication with Waveforms Chap. 5

Mi SIMULATION POINTS

Figure 5.22d Bit error probability of M-FSK with both frequency and timing errors, M 4 and p = (2
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 359

107!

V,

'
Pe
A foo
(E) 3
P,

Figure 5.22e Bit error probability of M-FSK with both frequency and timing errors, M = 8 and p = 0.1
 LA
aaML
IA

MN al iM Ul
i
360 Noncoherent Communication with Waveforms Chap. 5

i
Om

:
107

INT HOTS HLL

i
ITT TTT TIE
105
AML
10°
ALTE

Om

(Om

Om
i

Figure 5.22f Bit error probability of M-FSK with both frequency and timing errors, M = 8 and p —=i 02
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 361

M-1

Bee
P,
mat ok)
k+1

J] son: gin) (9.175)

cea gis
i] .1Q,4ipAn and
1] <i <-<ig

Letting x9 = 2x and using (5.173) and (5.174) in (5.175) gives after some manipulations

P,(E/mp) < mi Sane

Ps a 4eee \, (ame
fre) = minsDk (1— xo ) ae!
Xx exp
Ex (n,n)
kxo fj

SE Beale) any
awe 1+kxo
(5.176)

i1—=0 i =0

i <ig<---<i, and
iy tgs tpn

Finally, the desired upper bound on P;(£) is given by

1
M-1
P(E) S = eB P,(E/mn) (5.177)

Figure 5.23 illustrates the upper bound on P,(E) as given by (5.177) versus bit
energy-to-noise ratio E,/No for M = 4 and various values of p (the normalized frequency
error), assuming minimum spacing for orthogonality. It is to be emphasized that the results
in Fig. 5.23 are upper bounds and thus should not be used to predict the true error prob-
ability performance. Rather, their value is for making system comparisons and trade-offs
since the relative tightness of the bound to the exact result should be about the same in all
cases considered. As is true for most Chernoff-type bounds, they are asymptotically loose
by about | to 1.5 dB. For other bounds, see [29].

5.2.9 Performance of Two Suboptimum Receivers

The spectrum analyzer receiver with autocorrelation function trunca-

tion. Here, we consider the error probability performance of the spectrum analyzer re-
ceiver when spectral estimates are evaluated by means of time domain truncation of the
autocorrelation function. This approach reduces the amount of computation time required
but increases the error probability.
The simplest form of time domain truncation of the autocorrelation function is to
apply a rectangular window to the function; that is, compute only ,,(j/2B) for j =
0,1,2,---, J where J < 2BT — 1. When the correlation function is truncated in this man-
ner, the computation of the approximate spectral observation ®,,(k/T7’) is based on the ex-
pression
dbv,(t; J) A by, (t) W(t) (5.178)
 WC
inveaica
beth

HINE
362 Noncoherent Communication with Waveforms Chap. 5

EXACT
===: BOUND

107 *
‘IN

107 ec
ail
eas ha
if
|
eC
107
a
iiihags! He

wo \
= a So-
q
S

af
}
nee

|
o
|
-

|
DS
~-

oO
oO
o 10 11 iE 15) 16

Figure 5.23. A bound on the symbol error probability of 4-FSK in the presence of frequency error

where W(t) is the truncating window defined by

(S179)

Hence, from (5.51)

Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 363

Dif) =~, AOvq (0 4 -

a (5) cos (77). fi Ba Gus0)

The quantity 1 — [(J + 1)/2BT] is a measure of the fractional truncation. With truncation,
the frequency separation fj, ; (between tones i and j) must be adjusted so that they remain
envelope orthogonal at the receiver. Thus

aa ORT ain
PASSTON eI
yields orthogonal signals when truncation is employed. Hence, the minimum total transmis-
sion bandwidth B,, required for M tones must now satisfy the inequality

MCE) OM ere
aes se aaa 6 Picea
Using a result from the theory of sequences of j-dependent random variables [23], Springett
and Charles [7] are able to approximate the pdf of the spectral estimates ©, (fi; J),
1 =0,1,---,M —1 and hence the error probability performance under the assumption
P/NoB «<1 and J + 1<2BT. Under these conditions and assuming that the signal cor-
responding to f, was in fact transmitted, the pdf of ®, (f; J) at f = fi is Gaussian with
mean
{6
oy A (Ai J)} = =a
2B IN0B + Rs, (9)]

J FOBT j : (5.183)
+ oe ae Rs,(j) cos (27 (5)

where

J
Ria (jf).2P cos Baz (5) (5.184)

The variance of ®,,(

fj; J), that is

PA E{®? SpOia E*{@y, (fis J)} (5.185)

is quite tedious to evaluate and hence we outline only the major steps. The first term in
(5.185) can be written as a double sum, namely

sete} «Sgn HS

cgi
al! Z

(5.186)
; J ait
x E {bv (A) brv.(h)} cos 27 fi (5) cos 27 f] (=)

where €9 = 1 and €, = 2 (n £0). The elements E{dy,(j)¢,,(A)} are now considered for the
following four cases: (1) j =h =0; (2) j=0, h £0; (3) j =h £0; and (4) j Fh £0,
364 Noncoherent Communication with Waveforms Chap. 5

J <h. The results are:

Eldry,D) Pr (h)}
(35) (NoB)? [2a (1+
+ 55 i 2(1 +288)|, bag (ere
2BT—H (NoB) |
[287 (1+ #0 )olRy (h), , Re Cae
» ae (NoB)? (! [oar - hn)(ay + 4h |oa (5.187)

~ (4 fn (Sa 0) j=na0
ae ee) {ear — i)) (8) + 25¢5 (2+ 8) |ea
(2-+fa)Fat?)
— (fe) tbAU Te
where

5A = (5.188)
The variance of the spectral estimates can now be obtained by weighing each j, h element of
the covariance matrix by €j€n(1/2B)?[cos 2 fi(j/2B) cos 27 fi(h/2B)], summing the re-
sulting (J + 1)? elements and subtracting the square of the mean. Although closed form
expressions can be obtained for both (5.183) and (5.185), it is simpler to evaluate them
numerically on a digital computer. Assuming that the M Gaussian distributed spectral es-
timates ®,, (fo; J), By, (fis J), >>, Pv, (fm-—1; J) are uncorrelated, an expression for the
probability of correct detection is given by

a ia 2\ {1 Of be
— w]e
PG) ce ei (= ){Serte| a aoa || dx (5.189)

where the jzz’s and o,’s are defined in (5.183) and (5.185), respectively. Also in (5.189), any
value of / #k is suitable since jz; and o; are independent of / for 1 4k (see Problem 5.8).
Numerical integration of P,;(E) = 1 — P;(C) was carried out in [7] for various values of
the normalized window width (J + 1)/2BT in percent. The results are illustrated in Fig-
ures 5.24 and 5.25 where P;(E) is plotted versus E,/No, for M = 2”, n= 1, and 5 bits.
Although the results were plotted for small percentage truncations (including no truncation),
one should be suspect of applying these results to a practical system, since the assumptions
regarding the independence and Gaussian nature of the spectral estimates break down when
J — 2BT. Comparing the no-truncation curve of Figs. 5.24 and 5.25 with the ideal nonco-
herent results of Fig. 5.11, one observes that the approximation breaks down in such a way
as to give a pessimistic estimate of the correct answer.

The bandpass filter receiver. Let’s consider the performance of the subopti-
mum receiver (binary case) when the integrate-and-dump filters are replaced by bandpass
filters as illustrated in Fig. 5.10. The two possible transmitted signals are given by

s\(t) = V2P cos fio.= yi - |- so(t) = V2P cos io.“a yt un | (5.190)

Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 365

102

10 | 5 = 1/256

(E)
P;

NO TRUNCATION

4075
6.7 8.7 10.7 12.7 14.7 16.7 18.7 20.7

Figure 5.24 Probability of error performance for the spectrum analyzer receiver with
autocorrelation function truncation; n = | (Reprinted from [7])

The bandpass filters in Fig. 5.10 each have a noise bandwidth B « f, and are centered,
respectively, at the frequencies corresponding to the two possible transmitted signals. As-
suming further that Aw is large enough (B is small enough) that the BPFs don’t spectrally
overlap, then the noises at the filter outputs are uncorrelated (see Problem 5.9). Finally,
assume that B is also wide enough to assume negligible intersymbol interference (ISI).
We express the noises at the filter outputs as narrowband expansions around the ap-
propriate bandpass frequency. Thus

r(t)=sj(t)+ni@), i=0,1 (5.191)

 Noncoherent Communication with Waveforms Chap. 5
366

8 = 1/256

P(E)

Figure 5.25 Probability of error performance for the spectrum analyzer receiver with
autocorrelation function truncation; n = 5 (Reprinted from [7])

where

Aw : Aw
n(t) =Nne\(t) cos |
(x~ >) t+ | —ns\(t) sin IG ~ =) t+ | (5.192a)
Z
A
no(t) =Nn-o(t) cos |
(vx
+. >) t+ | — nso(t) sin |
(0+ >) (Pa | (5.192b)
2
Rewriting (5.191) in polar form gives
Sec. 5.2 Performance of Noncoherent Receivers in Random Phase Channels 367

A@
r(t) = &(t) cos (2.ot >) t+. + wo e Od (5.193)

where

g(t) = iG 2P me (£))? + (nsi(t))? (5.194)

(0) = tan a > Nsj(t)

(5.195)
V2P + n¢i(t)
and the “—” sign in the “+t” notation is associated with the subscript i = 1 and the “+” sign
with the subscript i = 0. The process &;(t) is the envelope of r(t) (assuming message m; is
transmitted) and any sample of it has a Rician pdf, namely

fe (€) = |52 &XD ea ln(2F), §>0 (5.196)

0, otherwise

where o? = NoB is the noise power at the output of either BPF. Since the noises nj(t)
and no(t) are uncorrelated (and thus independent since they are Gaussian), €; and & are
independent random variables, that is, at any sample time, the envelopes of the processes
r,(t) and ro(t) are independent.
To compute the probability of error, we first compute the conditional probability of
error assuming s;(t) was transmitted. Under this assumption, BPFo has no signal in it. Thus,
f(x) should be computed with P = 0, yielding the Rayleigh pdf

x2

feg(x) = exp (—32), 720 (5.197)

0 ’ otherwise

The probability of choosing so(t) when indeed s;(t) was transmitted is the probability that
&, — & is less than zero (the threshold) at the sample time, that is

P;(E/mj) = Prob{é, — & < 0} = Prob{&o > &1}

0° oo 5.198
\ == [ fa) | feg(x)|p=o0dxdy
y

Similarly if so(t) were transmitted, then f¢,(y) should be computed with P = 0 and the
conditional error probability would be identical to (5.198) with the subscripts “1” and “0”
reversed. Thus, because of the symmetry of the problem, the two conditional probabilities
are equal and the average error probability P;(£) is equal to either one of them. Substituting
(5.196) and (5.197) into (5.198) gives
 368 Noncoherent Communication with Waveforms Chap. 5

lee)
y y~2 +2P V2P y /(ee) x ( 52 )
PUR)
3° (E)= [ peesa exp (ey 452 )o( 2 hae — expPile |——
seo ] dx dy
———
ae
exp(—y2/207)

(C2) Poe eie )s P\O

FT
i

(1/2) exp(P/207)
J2P
ae
i ie 1 ]
= 5 exp (-3) a 5 exp (-35NRer)

where SNRzp is the ratio of the signal power P to the bandpass (input) noise power NoB.
Comparing (5.199) with (5.79), we observe that the suboptimum implementation gives
“essentially” the same performance if the bandpass bandwidth B of the filters in Fig. 5.10 is
chosen equal to 1/7, that is, the bandpass noise bandwidth of an integrate-and-dump filter.
By “essentially,” we imply that we are ignoring the portion of the data spectrum that falls
outside the bandpass bandwidth B.

5.3 NONCOHERENT RECEIVERS IN RANDOM AMPLITUDE AND PHASE CHANNELS

In some communication channels, the amplitude of the received signal is random and un-
known to the receiver. In this case, the optimum decision rule maximizes the likelihood of
the signal over all possible amplitudes and phases. Sometimes, the amplitude and phase are
independent variables. However, in other cases, they are correlated. We will provide a model
and a physical intuition of the statistics of the amplitude and phase for both Rayleigh and Ri-
cian channels. But first, let’s consider the following problem, which will be related to both
channels later. Assume, when message mz, is transmitted, that the received signal is given by
df

r(t)=) >gisia(t)+n@), O<t<T (5.200)

t=1

where the g;’s are independent Gaussian random variables each with mean jz; and variance
of and n(t) is AWGN with one-sided PSD No Watts/Hz. We further assume that the var-
ious signals {8}, j(t), 52,;(¢), +++, 8s,;(t)} are orthogonal with equal energy Ej, which is a
function of the message m ;, that is
1h
/ Si, j(t) 8), jdt = E 6.1 (5.201)

where 6; is the Kronecker delta function. The MAP decision rule selects message mj,
which maximizes the general likelihood function A(r(t)), given by
ie.) foe) 9) ap J
ar) = | of Foe festenexe |ef r(t) >> gisi, (at
=—6o) —0o 0 0 i=l

J) Sie (5.202)
J

iosD> gigisi, jt) S81,joa]

i=l 7=1
Sec. 5.3 Noncoherent Receivers in Random Amplitude and Phase Channels 369

The exponential term is the likelihood function assuming a perfectly known signal in
AWGN. The general likelihood function is its average over the unknown parameters, the
Gaussian amplitudes in this specific instance. Defining
if
uae LOse@ar,t = 1,2,..7, Jandy =0,1,2.7,M@—1 (5.203)
0
then

J ee)
ps 8;2
A(r(!) = I]ih
ae fc;(gi)exp
p }—giWi,;
No ND — E;ae } dg;
§ (5.204)

Substituting for the Gaussian density, (5.204) can be written as

sow=[]fxHG; JiL
—
M8 dg, (5.205)

where h(g;) is the argument of the exponential function and is given by

2 a; 2
eit
2 gers eal ati ae Caerele a4 Wa (5.206)
Tai Ne No ” No 20?
Using the integral [12]

(5.207)

then (5.205) becomes

2

Nepean | (= Nisbiae (5.208)

t=] Ue)

where

(5.209)

l o?
7S — (2u,+4—-wW;,; (5.210)
at ne Deda init NG wt
Hence, the MAP decision rule selects message m;, which maximizes

(5.211)
ileJe (2)
i=1

Assuming that the signals have equal energy, then E; = E and asa result, p;,; = pi. In this
case, we need to only maximize

(5.222)
370 Noncoherent Communication with Waveforms Chap. 5

Taking the natural logarithm, we have

Z a2 4
SeMs ge (5.213)

J
op
(#5‘ Mis) (5.214)

As a special case, let the means be zero (that is, 4; = 0 for all i.) Then the MAP rule reduces
to maximizing
df 2
ue tJ (x73)
No
(5.215)

2.
On the other hand, when the means are nonzero but the variances are equal (that is, 07 =
o), then we need to maximize

Sy (4 + aM) (5.216)
J Li 1 2

Both rules given by (5.215) and (5.216) will be very useful in deriving the optimum detec-
tion structures in Rayleigh and Rician channels, respectively.

5.3.1 Optimum Receivers in Rayleigh Channels

In Rayleigh channels, the received signal, when message m, is transmitted, is given by

r(t) =rv2ay(t) cos(wet + P(t) +O) +n(t), O<t<T G2)

where r and @ are independent random variables, r Rayleigh distributed [that is, fr(r) =
(r/o7) exp(—r?/207), r > 0], and @ uniformly distributed over (0, 27). In real physical
channels, the fade r(t) is a random process and its statistics vary with time. In slowly fading
channels, the fade is assumed constant over a symbol duration and can thus be denoted by
r. At this point, it would be very helpful to examine the scenarios or circumstances under
which the signal model given by (5.217) reflects the actual output of a physical channel.
Consider the case where a “large” number of scatterers exist in the propagation path and are
furthermore randomly located, as depicted in Fig. 5.26. Let the received signal from the I"
scatterer be (no noise case)

ri(t) = cw 2ag(t — 1%) cos (we(t — t1) + x(t — T))) (5.218)

with attenuation c; and delay t;. The composite received signal is the sum of the signals
from all scatterers, that is

=a) (5.219)
all /
Sec. 5.3 Noncoherent Receivers in Random Amplitude and Phase Channels 371

a TROPOSPHERE
b C, ey
cS G e a
a MI ~ ie ee

RECEIVER —
TRANSMITTER

Figure 5.26 Scattering of the transmitted signal by the troposphere

Assuming that the delays {t;} are all small relative to the time constant (which is approxi-
mately the inverse of the bandwidth) of the complex signals {a,(t) exp(j@x(t))}, then r(t)
can be approximated by

r(t) ~ V2ax(t) }>c1 cos (c(t — 11) + x(t)) (5.220)

all /
As an example, consider the case of a baseband BPSK signal. The bandwidth of the signal is
roughly 1/T sec where T is the symbol duration. Hence, its time constant is approximately
the symbol period, which is intuitive since the signal is constant during a symbol period. If
the delays from the various scatterers are fractions of the symbol period T (that is, t) < T),
then for all practical purposes, the delays can be ignored without any significant distortion
of the signal. The received signal of (5.220) can be alternatively written as

r(t) ~ V2ax(t) ¥ cy cos (wet + bx(t) — $1) (5.221)

all/
where ¢; 4 w-t;. If the delays are comparable to 277/w,, the phases can not be ignored
and are essentially independent (each is a function of its respective path) and uniformly
distributed over (0, 277). Expanding the cosine with trigonometric identities, r(t) can be
written as

aU) row 2ag(t) COS(Wct + Px(t)) + 7s V2a,(t) sin(Wet + Px(t)) (5:222)

where

i we CL COS dj (5.223a)
all /

rs =) cr sin gy (5.223b)
all /
Using the independence of the {7}, it is easily shown that

E{r:} = E{rs} =0 (5.224)

372 Noncoherent Communication with Waveforms Chap. 5

1 (5.225)
On = 9%, = 5 > Elci} = o*
all /

and
E{rers} =0 (5.226)
Assuming that the number of scatterers is “large,” it can be argued via the central limit
theorem that r, and r, are Gaussian, independent with zero-mean and equal variance o*.
Rewriting r(t) as

r(t) = rv
2ag(t) cos(Wct + x(t) + 9) (5.227)
where

f= /ferre and @ Se (5.228)

lc

then from Appendix 5A, r is Rayleigh distributed and independent of 0, which is uniformly
distributed over (0, 277). Thus, the Rayleigh model is the result of a channel that has many
independently located scatterers in the propagation path.
In order to derive the optimum decision rule in a Rayleigh channel, consider the signal
model of (5.222) in the presence of noise, namely

r(t) = rev 2az(t) COS(@ct + by (t)) + rsV2ag(t) sin(Wet + de (t)) + n(t) (5.229)
where n(t) is AWGN with one-sided PSD No Watts/Hz. Comparing (5.229) with (5.200),
the optimum decision rule for equal energy signals can be written by inspection of (5.216),
with xj =0, J =2 and s1;(t) = V2a;(t) cos(wet + o;(t)), 82,;(t) = V2a;(t) sin(wet +
gi (t)); namely, the receiver sets m(r(t)) = mx if and only if
2 2 2
& = Wr, + Wo;
if 2
= (/ r(t)v 2a; (t) COS(@ct + ; (dr)
0 (5.230)
T 2
=f (/ r(t)V2a;(t) sin(wet + dr)
0
is maximum for i =k. Note that W,,; = ze; and W2,; = zs; where Z¢i, Zsi are defined by
(5.14). The decision rule of (5.230) is identical to the optimum detector in noncoherent
channels with a fixed amplitude. The optimum receiver can thus be alternatively derived by
noting that the noncoherent detector with a known amplitude is the optimum noncoherent
detector for random amplitude channels, if we condition on a known amplitude. Once we
condition on the amplitude, it can be ignored as it does not contribute to the maximization
process and as a result, averaging over its distribution does not alter the structure of the
optimum detector. Note however that even though the structure of the detector is not altered,
its performance will be a function of the amplitude pdf, as will be shown shortly.

5.3.2 Optimum Receivers in Rician Channels

In numerous fading channels, there is a direct line of sight between the transmitter and the
receiver. As a result, the observed signal is composed of a fixed (or specular) component in
Sec. 5.3 Noncoherent Receivers in Random Amplitude and Phase Channels 373

SATELLITE

Figure 5.27 Example of a Rician channel

addition to the Rayleigh component, as depicted in Fig. 5.27. Such channels are referred to
as Rician and the received signal, assuming message m, is transmitted, is given by

r(t) = av2ax(t) cos(wct + dx (t) +4)

6.231)
+ rvV2az(t) cos(@ct + dx (t) +0) tn(t), OK<t<T
where @ and 6 are the amplitude and phase of the specular signal. At this point, we dif-
ferentiate between two distinct receiver structures, namely, the coherent and noncoherent
receivers. In both scenarios, we assume that @ is known and fixed. In the coherent case, the
phase of the fixed component 6 is estimated perfectly while in the noncoherent structure, 6
is not tracked. In either case, the phase of the Rayleigh component @ is unknown and as-
sumed uniformly distributed. We will thus derive both coherent and noncoherent receiver
structures and compare their performances.

Coherent detection. When the phase 6 is known (that is, coherent detector), the
received signal can be expressed as

r(t) =reV2ax(t) cos(wet + e(t)) + rsV2ag(t) sin(wet + de (t)) tnt) (5.232)

where r, and r; are independent Gaussian random variables with nonzero means given by
E{r-}=acosd and E{r;}=asinéd (02233)

and
Var(re} = Var{rs} = 07 (5.234)
Comparing (5.232) with (5.200), the optimum receiver can be written down immediately
using (5.216) with J = 2, 4; =acosd, 42 = a@ sin 4, that is, set m(r(t)) = mg if and only if

acosé
——
|
+ —W,;]
: a sind \*
( ro2 46 No is) +(—W2;+
( Det ) (3.239 )

is maximum for k=i, where Wj = ie r(t)/2aj(t) cos(wet + oj(t))dt and W2,; = ie r(t)
2a; (t) sin(@cet + ;(t))dt . Equivalently, the decision rule can be modified to maximize

ONO f(a) v2ai(¢) coset + gilt) — 8)dt + W?; + W3,

he

(5.236)
 Noncoherent Communication with Waveforms Chap. 5
374

DOHPYPDPVEO

V2a_ (t) sin(at+oy_,(t) _asind

M-1 262

Figure 5.28 Optimum coherent receiver for Rician channels

which clearly shows a coherent and a noncoherent contribution. One implementation of

coherent receivers in Rician channels is depicted in Fig. 5.28 and requires knowledge of
o”/cNo. Keep in mind that the term “coherent” refers to the phase of the fixed component
only. As expected when a = 0 (that is, no specular component), the receiver reduces to the
classical noncoherent receiver of Fig. 5.3.

Noncoherent detection. In the noncoherent receiver, the phase 5 is unknown

and thus can be assumed uniformly distributed. In this case, the optimum decision rule is
derived by expressing the received signal in the form of (5.200), by conditioning on 5. Then,
the likelihood function conditioned on 6 and message m ; (is transmitted) is given by (5.211)
with J = 2, that is

a qt
A(r(t)/5, mj) =] | (—-) exp (#4) (5.237)
eh a 4pi,j
Averaging over a uniform 4, (5.237) reduces to
P4 2
A(r(t)/m;) = Koki ¥2s) 7 (Ka,/W?, BSWw?) (5.238)
 Sec. 5.4 Performance of Noncoherent Receivers 375

where K,, K2, and K3 are positive constants and are independent of j for the equal energy,
equilikely case. Since the likelihood function is a monotonic increasing function of the en-
velope (or equivalently, envelope squared), the optimum receiver makes its decision on the
basis of samples of the envelope alone, that is, Ww? ae Ww? j (see Problem 5.10). The receiver
need not know o7/aNo, which is a great practical advantage over a coherent receiver.

5.4 PERFORMANCE OF NONCOHERENT RECEIVERS IN RANDOM AMPLITUDE

AND PHASE CHANNELS .
We shall now evaluate the performance, in terms of probability of error, of the optimum
receivers in Rayleigh and Rician channels. Note that since the received signal energy is
random itself, we will evaluate the performance as a function of the “average” received
energy Egy. Moreover, results are presented for orthogonal and, when possible, correlated
signals.

5.4.1 Performance in Rayleigh Channels

M-ary orthogonal signals. Since the optimum receiver in Rayleigh channels is

identical to its noncoherent counterpart in AWGN channels (that is, both seek to maximize
2 = me + Zz), then the performance of the receiver in Rayleigh channels, conditioned on a
fixed fade r, is identical to that of noncoherent signaling in AWGN as given by (5.77) with
Ey replaced by r7E,, that is

PAE se
5 ( eG” y+ M-1 3cm(——5)e7| awe
EET (e=)}
r No (5.239)

Averaging over the Rayleigh fade r, we obtain

P(E)y=
3 ( 4)[|
P,(E/r)—
5 ( jn) Z exo(-7)a
Pons
5.240
M-\ pet ( )
= _ |Cy
dm m+
1+ m(Eav/ No)
where Egy/No & 207(E;/No). Note that P,(£) in Rayleigh channels is no longer an ex-
ponentially decreasing function of Eg,/No as in AWGN channels, but is a function of the
inverse of Eg,/No. This means that for a fading channel, the transmitter must provide a large
amount of power in order to provide a reasonably low probability of error at the receiver. In
many applications, the required transmitted power is not permissible for various reasons.
An alternative solution is obtained by means of diversity techniques [24, 25], which insert
redundancy in the channel. Depending on the nature of the fade, time, frequency, spatial
diversity, or combinations thereof may be employed (see Problems 5.11 to 5.17).
For the binary case (that is, M = 2), P;(E) as given by (5.240) reduces to

PEs M =2, Rayleigh Channel (5.241)

2+ (Eav/No)

It would be interesting to assess the effect of signal fading, in the presence of perfect carrier
coherency. As a result, assume that the receiver performs an exact estimate of the carrier
376 Noncoherent Communication with Waveforms Chap. 5

phase and uses that information in the detection process. Then, the conditional probability
of error, assuming orthogonal signals, is given by (4.59), that is

piein=o(r E
| (5.242)
No,
and
ee if r2 jks 1 My 210)
PCE = — exp |-—; —_—e™’ '*dydr (5.243)
ae i o? o( 20°] SrENV20
Performing the integration, one obtains

JOA) = U 1- EEO , M=2, Random Amplitude Channel (5.244)

2 Nei Eav/No

and for comparison, we list the other two previously derived results for orthogonal signals,
namely

Es
JE3)) =O) ae M =2, Coherent Channel (5.245)
0

and
1
P,(E) = = exp'| — *), M=2, Noncoherent Channel (5.246)
2 2No

The performance of these four channels is depicted in Fig. 5.29. Note that at high E,/No,
coherent detection offers about 3 dB advantage over noncoherent detection. Simultaneously,
detection in random amplitude channels (Rayleigh channel with perfect phase measure-
ment) offers about a 3 dB gain over detection in random amplitude and phase channel.

Binary correlated signals. Other results have been derived in the literature for
correlated signals in Rayleigh channels. Turin [26] has derived the performance of corre-
lated signals for M = 2 and is given by

te
P(E) = 5 (a oe) (5.247)
“Ve pe
with = pio denoting the correlation between the complex baseband signals as given by
(5.86) and
A (Eav/No)
Se 5.248
(Eav/No) See ( )

5.4.2 Performance in Rician Channels

In Rician channels, two different receivers can be employed depending on whether the phase
of the specular component is known or not.

Noncoherent detection. _Let’s consider the noncoherent receiver that selects the
signal corresponding to maximum &? = ja + roe
Sec. 5.4 Performance of Noncoherent Receivers 377

1.0

107 RAYLEIGH CHANNEL

RANDOM AMPLITUDE CHANNEL

(E)
P, NONCOHERENT
CHANNEL

COHERENT
CHANNEL

Figure 5.29 Probability of error for the coherent, noncoherent, random amplitude and Rayleigh channels

M-ary Orthogonal Signals. Here again, if we condition on the fade r, then the
performance is identical to (5.77). Averaging over the Rician variable r, one obtains

M-1 1 CO >

P(E) = 0 (-1)"4
1m-1Cm m+ 1 0 o2
m=1 (5.249)

eae 1 2 2
= m
5)No we 0(5)
2 je
ey pela
ar

fi
x exp | pa +a) exp] ar

Using the integral in (5.74), then P;(E) becomes

378 Noncoherent Communication with Waveforms Chap. 5

M-\
P(E) =) (-1)" |
thy-1Cm sien xe | 5.250
m=1
| Hrd Ste aby |e
where Bf is the SNR in the scatterlike signal, that is
E, '
BSA 207No (5.251)

and y? is twice the ratio of the power in the specular component to the power in the Rayleigh
component, that is

yas (5.252)
oO

As acheck, if y = 0, the Rician channel reduces to a Rayleigh channel and (5.250) reduces
to (5.240), as expected. Note that the average symbol SNR in a Rician channel is given by

for =(a?+20)Ft=(1+2)e
Z
Eav Es We
(5.253)
which for y = 0 reduces to f, the average symbol SNR in a Rayleigh channel. In the binary
case, (5.250) reduces to

IRUa v= ——.ex
Casde
Depo. {-acne
2(B +2) M =2 Rician Channel (Noncoherent) (5.254)

Here again, when y = 0, (5.254) reduces to (5.241) as expected. The performance in Rician
channels for M = 2 is depicted in Fig. 5.30 as a function of 8 for various values of y. For
y =0, the performance reduces to that in a Rayleigh channel. On the other hand, when
y = 00, it reduces to the performance of binary orthogonal FSK in an AWGN channel.
Binary Correlated Signals. The performance of binary noncoherent reception in
the presence of correlated signals has also been derived by Turin [26] and the probability
of error is given by [with 6 defined by (5.86)]

Aares CLS TNE

1), +
wvt—Pe
ape
— Kp
es ae (5.255)
(u- + w*)v 5
x exp Bicis Ses
5 Sati |
byfo
(wan?)

where

Te
— V0 = lad = wa?) 5.256
1 —plp\? eee

nd ie IA) — w2 1A)
w= 5.257
tial ( )

by? 1— plo? aren,

2 1-wpP
 Sec. 5.4 Performance of Noncoherent Receivers 379

0.5

107"

102,

a
a”

107°

10-2

107” \

0 10 20 30 40 50
B, dB
Figure 5.30 Probability of error for binary orthogonal signals, Rician channels (Reprinted from [27])

and uw A B/(B +2). Note that when |f|=0, u=0, w = V2 and v=y/p/2. Since
COO X= exp(—x?/2), (5.255) reduces to

2 2
yu I yu
P(E) =exp | = 501+ wexp |-7 |
, (3.259)
Se hate {-;255 |
pd 2(B + 2)

which is identical to (5.254). On the other hand, when y = 0, v = 0 and (5.255) reduces to
 Noncoherent Communication with Waveforms Chap. 5
380

(E)
P,

2
(a +20 )E, dB
No
(a) (b)
Figure 5.31 Probability of error, noncoherent receiver: (a) y = 0 and y = 00, (b) y =2
(Reprinted from [26] © IEEE)

1 rf L = bole
P(E) = 0(0,0) ~ 5 fry
(5.260)
=53(1- H er |
2 v1 — p?|p/?
which is identical to the performance of the Rayleigh channel as given by (5.247). The
performance for noncoherent detection is depicted in Fig. 5.31 for y =0, oo and y = 2 for
various values of |p|, as a function of Egy/ No. Note that orthogonal signals (that is, || = 0)
provide the optimum performance for all y and Egy/No.

Coherent detection. In the coherent Rician receiver, the decision rule is modi-
fied by the addition of the appropriate biases as given by (5.235). Since the phase 6 is known
perfectly, it can be set to zero without any loss in generality.
Binary Orthogonal Signals. _ For the binary case, the rule reduces to

a pO)
Wye 2 en 28
Woo\? 2, 0 (a Wis 2
Was 2
etree le | aed 5.261
(5 72) (=22) a Ea No ~ No ( )
Sec. 5.4 Performance of Noncoherent Receivers 381

where

Wen =tol
ee i,” r(t)V2a0(t) (cos
ee )(Wet + do(t))dt (5.262)

and

Win. eer: cos

ae = ifr(tyv2ai(t) ey (wet + p(t)dt (5.263)

Assuming that mo is transmitted, then from (5.232), r(t) is given by

r(t) =rev2a0(t) cos(@et + o(t)) + rs 2ag(t) sin(wet + go(t)) +n(t) (5.264)

and E{r.} =a, E{rs} =0 with Var{r.} = Var{rs} = 0. Assuming that the signals are or-
thogonal, then

Wio=rcEs +m (5.265a)

W20=PrsEs + n2 (5.265b)

Wii =73 (5.265c)

W211 =14 (5.265d)

where n;, i = 1, 2,3, 4 are zero-mean Gaussian independent random variables with vari-
ances NoE,/2. From Appendix 5A, we know that R; is Rician where

(5.266)

with parameters

(5.267a)

> a \2
(5.267b)
nf bla a

and
| D0) NoE Ss Esk
2
= — | E’o* + = 1+ (5.268a)
°0 No ( : D) ) IN| P)

(5.268b)

The conditional probability of error, P;(E/mo), is then the probability that R; > Ro, which
is given by (5.130) for two general Rician random variables. In our case, the parameters a
and b from (5.131) become
Tee ane
(5.269)
“opto? BIB +2)
rad ss yp +1)2 (5.270)
~optor B(B+2)
382 Noncoherent Communication with Waveforms Chap. 5

and

a ee (S271)
og +o; +2

Substituting all these results in (5.130) and noting that P;(E) is identical to P;(E/mo)
because of symmetry, then
P(E) = (ats
y 4
y(B +1)
: JB(B +2) J/B(B +2)
(5.272)
(S) y? (6? +26 +2) ; (v B+ )
- (55) ex ~9BUS
ED) tle aoa)
which was also derived in [27].
Binary Correlated Signals. The performance of binary coherent reception in the
presence of correlated signals has been derived by Turin [26] and is given by (5.255) with
[p defined by (5.86)]

pV 1 — pl?
u= | —- ———_—_ (5.273)
v1— w*lal?
Peed pea/ 1 — |p|?
pk eo (5.274)
v¥ l= p7lpr
and

» —_ |BY2 2 [8G =
SW
1A?) + 20 az.—Re())]
i
eo
467(1 — |p|?)
Here again when p = 0, thenu = 1 — w=2/(8 +2), w=1+ m= (28 + 2)/(B6 + 2), and
v = (y/2)./(B
+ 2)/B and P;(E) reduces to (5.272) since uv = fa, wv = Vb, and so on.
The performance for coherent detection is depicted in Fig. 5.32 for y = 0, co and y =2.
For comparison purposes, various curves from Fig. 5.31 and Fig. 5.32 are replotted in Fig.
5.33. The results indicate that when fading does exist, there is vary little gain obtained by
the use of a coherent receiver. Of course, for a specular signal alone (that is, y = 00), the
use of a coherent receiver can result in up to 4 dB improvement over a noncoherent receiver.
But the performance of the coherent receiver for y = oo is already excellent so that the extra
advantage obtained from a coherent detection might not be needed.
In conclusion, a system that is to operate through fixed and fading paths will be for
all practical purposes optimal if it is designed to use uncorrelated signals and noncoherent
reception. The latter has the practical advantages that it does not require estimates of 5 or
o*/aNo at the receiver.
Sec. 5.4 Performance of Noncoherent Receivers 383

Re@)= +0.95
+0.9
+0.8
+0.6
+0.4

Re(p) >0
Re@)= 0.95
0.9

50

Figure 5.32 Probability of error, coherent receiver: (a) y = 0 and y = 00, (b) y =2
(Reprinted from [26] © IEEE)
384 Noncoherent Communication with Waveforms Chap. 5

0.0005

0.0002

1. ¥=0, Re(p) =0, C,NC

10 2.y=2, Re(p) =9 NC
3.¥=2, Re(f) =9, ¢

5. Y = 0, Re(f) =0, Cc

0.0002 6. Y=oo ,Re(P) =-1,C

C = COHERENT
ae NC = NONCOHERENT
10
a
a”

0.0005

0.0002

10°

0.0005

0.0002

Figure 5.33 Comparison of probability of error curves for several important cases (Reprinted from [26] © TEEE)
 Appendix 5A
Useful Probability Density Functions
In various optimum receivers, we encounter nonlinear operations on Gaussian random vari-
ables. Specifically, in noncoherent and later on in Chapter 6 partially coherent communi-
cations, the optimum receiver requires squaring operations on Gaussian or conditionally
Gaussian random variables. In this appendix, we summarize key results in this area and de-
fine useful densities, namely, the Chi-square, the noncentral Chi-square, the Rayleigh and
finally, the Rician density. A more detailed treatment of this topic can be found in general
probability theory books [11]. Here, we merely present the results as taken from Proakis
[17].

5A.1 CENTRAL CHI-SQUARE DENSITY

Define the random variable Y as
n

Vie Xe (5A.1)

where the X;, i = 1, 2,---,n are statistically independent and identically distributed (1.1.d.)
Gaussian random variables with zero-mean and variance o°. Then, the characteristic func-
tion of Y is given by
1
Wy(jo) 2 E {el = $e ee RS (SA.2)
(1 = j2wo?)"”

which is the consequence of the statistical independence of the X;’s. The pdf of Y is ob-
tained by the use of the inverse Fourier transform [5], that is

i n/2—1,—y/20
fyry)= onan (nJ2)> ae gy) (SA.3)

where I"(q) is the gamma function, defined as

[o@)

T'(q) =i! fi ee dt, q>0

0

T(q)=(q—1)!, q anintegerand gq >0 (5A.4)

1 3 aft
iGan r="
This pdf is called the Gamma or Chi-square pdf with n degrees of freedom and is illustrated
in Fig. 5A.1 for several values of n. The first and second moments of Y can be computed to
give

E{Y} =no*
E{Y?} =2no4 +n?o* (5A.5)
pan= 2no
oy
 Noncoherent Communication with Waveforms Chap. 5
386

ty
(y)

Figure 5A.1 Pdf of a chi-square-distributed random variable for several degrees of freedom

The cumulative distribution function (cdf) of Y is obtained by integrating (5A.3), that is

MM 1 a2) a >
Fy(y) 4 Prob{Y <= Se te" au (5A.6)
For the special case when n is even (let k=n/2), (SA.6) can be computed in closed form to
give

pa ee
Fy(y) =1—e 2/9 2 i (Gen) (5A.7)

5A.2 NONCENTRAL CHI-SQUARE DENSITY

When the Gaussian random variables of (5A.1) have nonzero means m;, i = 1, 2,---,n but
still identical variances o7, the characteristic function of Y is no longer given by (5A.2) but
becomes

n 2
Wy (jw) = Mae
Di | (5A.8)
(1 — j2woa2)n/2 o| 1 — j2wo?

As expected, (5A.8) reduces to (SA.2) when all the means are zero. The inverse Fourier
transform of this characteristic function yields

1 py \(-2/4 st+y s
fr) = = (5) exp |- as |5-1 (v=). y>0 (A9)
 Sec. 5A.3 Rayleigh Density 387

where s* is the sum of means squared, that is

By ny (54.10)

and I,(x) is the k'*-order modified Bessel function of the first kind [3] which can be ex-
pressed by the infinite series [4]

eA =) Fire
K(x) dLene
ee ac+740’ x>0 (5A.11)

The pdf, as given by (5A.9), is typically referred to as the noncentral Chi-square density
with n degrees of freedom. The parameter s? is called the noncentrality parameter of the
pdf. The cdf is obtained by integration as

Y 1 suy(n-2/4 stu 5
Fro)= f pee) exp |- =p |1-1 (va) du (5A.12)

Here again, when n is even (k = n/2), the cdf can be expressed in terms of the generalized
Q-function Q;(x, y) as

Fy(y) =1— Qx (<.2) (5A.13)

where

ou(a,b)& f° x(=) ‘exp {—*

2 2
Iy_\(ax)dx

(5A.14)
2 b2 gall b J

= 01(a,b) +exp |* 5 1° (2) I;(ab)

j=
The function Q (a, b) is typically denoted by Q(a, b) [not to be confused with the Gaussian
tail integral Q(x) which is a function of one variable] and is referred to as the Marcum Q-
function and is given by

O(a, b) a [sexe {-7$| Io(ax)dx (5A.15)

The first and second moments of Y can be computed in closed form to give

Ey, = no> + s*

E{Y?} = 2no* + 4025? + (no? vi2)" (5A.16)

oy = 2no* + 4075”

which reduce to (5A.5) when the means are zero (that is, s = 0).

5A.3 RAYLEIGH DENSITY

Suppose we define a new random variable R which is
 388 Noncoherent Communication with Waveforms Chap. 5

(5A.17)

where the X;, i = 1, 2,---,n are as defined earlier and the means are zero. Then, the pdf of
R can be easily obtained from the Chi-square pdf with a simple transformation to give

pal r2 }
fa) = saan {ae} r>0 (5A.18)

Here again, when n is even (let k = n/2), the cdf of R can be obtained in closed form as

pa ee ee
Fair) =1~ 00 |-3-3 07 (53) ts!) (SA.19)

and the m'" moment of R is given by

m/2T ((n+m)/2)
E(R™) = (207 ,
(R"} =(207) P(n/2) annO 5A.20
ene
for any integer m. As a special case of (5A.18) when n = 2, we obtain the familiar Rayleigh
pdf

r i
RON SE Vere (Re ee (5A.21)

with corresponding cdf

ee
Fe(r) = 1-exp |}. Te) (5A.22)
20

5A.4 RICIAN DENSITY

When R is still defined by (5A.17) and the X;’s are nonzero means (each X; has mean m;)
each with variance o7, then the pdf of R becomes

rn/2 r2 + 52 rs
Fal) = yaaa | = ce (=), r>0 (5A.23)
with corresponding cdf for n even (k = n/2)
Ser
Fr(r) =1— Ox (=,
0 0
=) (5A.24)
where s2 is defined in (5A.10). The m= moment of R is

m/2 s? ) TP ((n+m)/2) n+mn_ s?

E{R™}= (20? ex {-,| “Gee Fi eee
(207) exp 2a2| 7salGein wk Oar eeenan ean
where | F(a, B; x) is the confluent hypergeometric function [3] given by

Fila6 x Sy oem (5.26)

Teretpy PT
 Sec. 5A.5 On the Joint Density of Amplitude and Phase 339

When n = 2, (SA.23) reduces to the familiar Rician pdf

pert
Simei arya
e: rs :
PRO =
70x |-ee lo (=) x fee Oe)

where s = m+ + ms and corresponding cdf

Som:
Far) =1-0(=,-)
ST 1G
(5A.28)
When s = 0, the Rician pdf and cdf of (5A.27) and (5A.28) reduce to the Rayleigh pdf and
cdf of (5A.21) and (5A.22), respectively.

5A.5 ON THE JOINT DENSITY OF AMPLITUDE AND PHASE

So’far, we have been concentrating on the marginal densities of Y or R, basically ignoring
any phase information. Let’s define

R=,/xX?+X2 and tan@=

x (5A.29)
1

where X, and X> are i.i.d. Gaussian random variables with variance o2 and m; =m but
mz = 0. We already know that R is Rician distributed if m 4 0 but Rayleigh distributed if
m = 0. It would be interesting to derive the joint pdf fr@(r, 8) for both cases. Since R and
© are two functions of two random variables, it is easily shown [11] that

tr OU, 0) =ix (cose,rsn?), r=0,<2 0-7

] “
(5A.30)
= exp 552 ((rcos6 — m)- + (r sin a°)|

or equivalently
1
Jr.00,9)= a exp
&x |-a2 (r?— 2mr cos @ + m*I. r>0,-—7 <6 <7(5A.31)
210

To obtain the marginal densities, we integrate the joint density

1X r2>+m 2mr rade
fr) = fret, 0)dé = a exp 1-5 | [iw >|| dé
—1 —1

if r2 +m? I mr St)
rhc 2 se ola). Views

which is the Rician pdf as expected with s = |m|. On the other hand, the marginal pdf of ©
becomes

fo(@) = i! fro(r, 9dr

(5A.33)
1 m2 IP r2 — 2mr cos@ q
= epee re —_——___.—-_} dr
oe ci ae 0 *P 202
390 Noncoherent Communication with Waveforms Chap. 5

Completing the square in the integral and then substituting x = r — mcos 6, (5A.33) reduces
to

1 m? sin? 0 oe) 2
fo@) = agi OP | |f (x +meos@)exp |- 25
2
|ds (5A.34)
Miss —m cos6 oO

or
m2

m cos @ m? sin?@ mcos 6 exP {-25]

= Si es NN S| ee 5A.35

Note that when m = 0, the joint pdf in (5A.31) becomes

2
TRO, 0) = Seon 55) = fr(r) fo) (5A.36)
where fr(r) = (r/o) exp(—r?/207),r > 0 and fe(6) = 1/27, 1 <0 <2. Hence, R
and © are independent random variables with R Rayleigh distributed and © uniformly
distributed. On the other hand, when m ¥ 0, R and © are dependent with joint pdf as given
by (5A.31). The marginal pdf of R is Rician and f@(9) is given by (5A.35) and is plotted in
Fig. 5A.2 as a function of |m|/o. As one would expect, the pdf of 8 becomes quite peaked
as |m|/o increases.
Sec. 5A.5 On the Joint Density of Amplitude and Phase 391

Watt

(9)

0 0.17 0.2n 0.31 0.41 0.51 0.6n O0.7n 0.81 0.97 1.07
A)

Figure 5A.2 Pdf for phase angle, Rician channel

 Appendix 5B
Derivation of Equation (5.43)

From (5.42) (let w; 4 w. + @;)

2 Tit
a = / / r(t)r(s) {cos w;t cos w}s + sin w;t sin ws} dtds
0 0
rater (5B.1)
=|! ifr(t)r(s)
cos w;(s — t)dtds
0 JO

Let

f(t, s) Ar(t)r(s) cos w;(s — t) (SB.2)

then (5B.1) can be written as

e- In yoih
ap =| [ f(t, s)dtds (5B.3)

Divide the integration on'‘T? square region in the , s plane into two integrations on triangu-
lar regions as shown in Fig. 5B.1. As such, (5B.3) can be written as the sum of two integrals,
namely

&? lst T ps
c= | / fit.sydsdt + | / f(t, s)dtds
DP o Jo 0 JO
(5B.4)
Tieet Pom
=i) i S(t, s)dsdt +f ifJ (s, t)dsdt
0 0 0 0

where the second integral is obtained by switching ¢ and s. But from (5B.2), f(t,s) =
f(s, t), thus

Figure 5B.1 Integration on T? region

App. 5B Derivation of Equation (5.43) 393

dt’

Figure 5B.2 Integration on triangular

dt v i t region

E? et
52 f | f.srasat
eye OBS)
= 2 | / r(t)r(s) cos w,(s — t)dsdt
0 0

Letting s =f + T, then
Es T 0
AD a | / r(t)r(t + tT) cosw,tdtdt
Bago (5B.6)
T ft
= 2f / r(t)r(t — t') cos w;t’dt'dt
0 JO

This is an integration in the triangular region of the r, t’ plane using a vertical strip of width
dt extending from t = 0 to t =T. If, instead, we choose for our integration in the same
triangular region a horizontal strip of height dt extending from t = t’ to t = T as shown in
Fig. 5B.2, then we get
£2 She Hl
or = / ifr(t)r(t — t') cosw,t'dtdt’
0 v!
z (5B.7)
ii
= 2 | COs ojr'at’ | r(t)r(t — t’)dt
0 ue

Let v =t — 7’ in the inner integral, then

£2 if T—t’
a 2 f cos ojr'ae’ | r(vu+t)r(v)dv (5B.8)
Qe 0 0

which is identical to (5.43).

Consider the finite-time autocorrelation function ¢,(t) defined by (5.44) as
T-t

gy (T) =i rOre+ di, O<r<T (5B.9)

0
394 Noncoherent Communication with Waveforms Chap. 5

Note that ¢,(t) is not an even function due to the finite duration. Define the modified
function @, ,(T) as

Or(t), teO (5B.10)

be,r(T) = fas a
Then, ¢¢,,(t) is an even function and its Fourier transform is given by

®. (f) = | bes(acosanfedt =i) or (t) cos 22 f tdt (5B.11)

0 0

Hence, the Fourier transform of ¢.,,(t) (which is identical to its cosine transform) is iden-
tical to the cosine transform of ¢,(t). This technique provides an efficient computational
method for evaluating the cosine transform of ¢,(t) using fast Fourier transform (FFT) rou-
tines on @¢ -(T).
 Appendix 5C
Generalized M-ary Symbol Error Probability Bound
Consider an M-ary communication system whose decisions are made based on a relation
among M outputs xo, x1, ---, xw—1. Let these outputs be represented by independent
random variables with pdfs as follows:

Rice hi Xan)
: (5C.1)
he ar Cin) for B=; l,---,n—I1,n+1,---,;M—1

That is to say, for some particular n, the random variable x; has a fixed pdf whereas the
remaining M — | r.v.’s x;, i #n, all have the identical form pdf (perhaps different than that
for x,)! which, however, depend on a parameter ¢;, that varies with both i and n. Assuming
that message m,, is transmitted, then a correct decision is made at the receiver when x, > x;
for all i #£n. Then, the conditional probability of a correct decision is

P;(C/m,) = Prob{Correct decision/m,}

= Prob{x,,> x; -for.all’t 4 /m,}
CO

/ Prob {x, > x; for alli 4 n/my,Xn =a} fi(a)da

—oo

i Prob {x; < @ for alli #n/m,} fi(a)da

(SE22)
coo M-1
=p | | Problxi < w/mn} fi(ada
CO i=0
isn

coo M-1
jeu all[1 — Prob{x; > o/mn}] fi(oda
ae

If Prob{x; > a/my} is hard to evaluate, then use the Chernoff bound
| A(xj—a@)
Prob{x; = a@/mn} < E {e /m, | (G3)
=e *E {ei /m,} for any 4 > 0

Define

gi(a)
SE {e*"/m,} = iL e*? f\(B)dB
(5C.4)

g0(A; Lin)A E {e’"/mn} = ie e*? fo(B; in)dB

1. If the two pdfs have identical form, then we shall ignore the “0” and “1” subscripts on them and simply
write f (x) or f(x; ¢), as appropriate. An example of where the two pdfs are, in principle, different in form would
correspond to the case of ideal (zero frequency error) noncoherent detection of M-FSK, in which case f)(x) would
be Rician and fo(x) would be Rayleigh. Also in this ideal situation, fo(x) would not be dependent on a parameter
¢ which varies with the random variable being characterized.
396 Noncoherent Communication with Waveforms Chap. 5

Then

Prob {x; > a@/mn} < e ** g0(A; Sin) (5C.5)

and
oo M-1 '
P,(C/mn) > i) T] [1 —e*s0(a: tind] fede (5C.6)
oe tee
Finally
P(E /mn) = 1— Ps(C/mn)

(oe) M-\

ef no, 2)
ft [] [esos tw] ffil@de

= f: [+ Septem (SC)
oe) M-\

rz k=1

k
¥ 7 lil20(A; 7 fi(a)da
ij=0 i2=0 =O

ip .igsipAn and
ip <ig<ir<iz

or using (5C.4), simplifying and minimizing over the Chernoff parameter, we get
M-\
Ps(E/mn) < min oS(181
gi(—Ank)no

M—\M-\1 ell is
(5C.8)
es ‘
i;=0 i2=0 ix=0 j
a
80(An; Sijn)

play ipaAn and

ip <ig< <i

Assuming equiprobable signals, then the average probability of symbol error is given by
1 M-1
P(E) = = DT Ps(E/mn) (5C.9)
n=0

Note that if the parameter ¢;, is independent of i, that is, all x;, i #4, have identical pdfs
fo(x), then
—1M-1 M-\1 ek

ou ODER cea Ge ') boo (5C.10)

i;=0 i2=0 n=O =)

fy tasetgAn and
i] <ig<-+<ig

where

oe | oP fy(B)aB (5C.11)
 Problems 397

Thus, (5C.9) together with (5C.8) simplify to

M—\ M—1
P(E) <min )|(—1)*! ( i )ex—aesa (SC. 12)
k=1

PROBLEMS
5.1 Let z be a complex Gaussian random vector of dimension N, with

DIX,

where x and y are the real and imaginary parts of z. Let m, = E{z} denote the mean and
Kyz* = E{zz*"} denote the covariance matrix. Assume further that E {zz" }= [0] (the all-zero
matrix). Show that the probability density function of the random vectors x and y is given by

exp {—(w - m,)*? Kh (w — m,)}

Fx,y( Wi,1 W2) 2) = aN det(Kea*)

where w = wi + jW2.
5.2 Consider the following binary detection problem:

Ay: r=sin

Jaye jeer

where r is the observed vector, n is a zero-mean, complex Gaussian noise vector with

E {nn*”} =2NoL

E{nn"} = [0], (all-zeros matrix)

All vectors are N-dimensional and s is composed of independent, complex random variables
m,n =1,---, N of the form
eg J2Eanel*

where a, is a Rayleigh random variable with unit second moment, that is

2
20682 a0
(a) = :
Say 0, otherwise
and ¢, is an independent phase, uniformly distributed over 277, that is
1
s-, O<@<2n
= 27 — ~3
fon) |0, otherwise
(a) Show that r, is a complex Gaussian random variable when H is true. Compute its mean
and covariance matrix.
(b) Compute the likelihood ratio
s fr(r/ HM)
A(r)
fr(t/ Ho)
and simplify the likelihood test to derive the optimum detector.
(plket a= es Ir, |*.What is the pdf of z when Hp is true? Repeat when H, is true.
(d) Derive an expression for the probability of detection.
(e) Derive an expression for the false alarm probability.
398 Noncoherent Communication with Waveforms Chap. 5

5.3 Show that the asymptotic (M —> 00) error probability performance of the noncoherent receiver
as given by (5.77) is identical to that for coherent communications as described by (4.182).
5.4 With r(t) given by (5.80), verify that the decision variables &, k =0,1---, M — 1 are inde-
pendent of the incoming carrier phase 0, as claimed below (5.90b).
5.5 Perform the matrix multiplication TK,x7! and verify that the whitening transformation T as
given by (5.121) does indeed produce Kx, = Z.
5.6 Consider two independent Rician random variables Ro and Rj, each with parameters s; and
variance Ge (i = 0, 1), that is

2 2
ij Sea stele Tj{Sj ;
intr) = Fes] - eos |(22), r, >0
1

We wish to evaluate

P = Prob {R7 > Rj} = Prob(Ri > Ro}

(a) Show that

Prob{R; > Ro/Ro} = Q (2. *)

where O(a, 6) is the Marcum Q-function. Therefore

Oo) Ss ro
P =| Q (2. *) fRo(ro)dro
x 0 oO) OF}

(b) Show that the Laplace transform (L.T.) of /o(2./Bt) is given by (1/s) exp(B/s), that is
[o@)
1
i In(2./ Bt)e ‘dt = —e8/5, Re(s) > 0
0 AY

or, equivalently

1 ot+joo b/s
ts ©
Io(2/Bt) = al Gis GPS)
2} Ja—joo Ss
(c) By inserting the expression for /o in part (b) into the definition of the Q-function, show that

o+joo past&
Q (v2a, v2B) = etary | () ds, o>-l
2nj -—jo §— 1

(d) By substituting the above expression in (c) for the Q-function into the expression for P in
(a), interchanging the order of integration and simplifying, show that

(e) Define
2 2
aA a] A SQ
== ee 2 2 2}
09 + OF 09 +9;

2
ey0 Oo i
D
A A
o= 2 5) cc) = 5) 5)
Problems 399

Use the partial-fraction expansion

( 1 )( s 1 co
c, |—— = =
OS Al S—Cco S— Ll sto

to separate the expression for P in (d) into the sum of two integrals. Make the change of
variable eG — 1)=(p — 1) in both integrals. Using the results of (c), simplify the first
integral to show that it reduces to Q(./a, Vb). Using results from (d), show that the second
integral is given by

Therefore

P=10 (Va, vb) — ¢1 exp (-“=*) Io (vab)

=H7/ Consider the noncoherent reception of an M-FSK signal with a common frequency error Af in
all of the possible input tones. It is desirable to compute an upper union bound on the average
symbol error probability in terms of the exact bit error probability of noncoherent binary FSK in
the presence of the same frequency error. Let P,(E; 0, m) denote the exact bit error probability
of noncoherent binary FSK when the tones are spaced by m/T (recall that 1/T is the minimum
spacing required for orthogonality) where p = Af T is the normalized frequency error. Assume
that the tones in the M-FSK signaling set are spaced by 1/T.
(a) For M = 4, compute a union bound on the average symbol error probability P,(£).
(b) Repeat (a) for M = 8.
(c) By induction, generalize the results of (a) and (b) to the case of arbitrary M.
5.8 Using the trigonometric identities

3 cos? po n ie = sinn@ an + 1)6]

am; sin 6

pad nsin{{(2n — 1)/2]6} 1 —cosné

ye kcosk@é = : a5)
rex 2sin6/2 4 sin“ 0/2

show that the mean of the spectral estimate ®,, (fi; J) from (5.183) is given by

Mi 2 EAPy,
(ft; DY= ) yor
ME(t+ abe [1- ee), n=}
P_ J+ pee
oo Listen [air — 1 as
where k/T is the signaling frequency. A similar independence on / for /#4 k can be shown for
the variance of ®, (fi; /), but the mathematics required is cumbersome.
5.9 Consider a WSS random process x(t) which is input to two different linear filters G(f) and
H(f) resulting, respectively, in outputs y(t) and z(t). If G(f) and H(f) do not overlap in
frequency, that is, G(f)H(f) = 0, show that y(t) and z(t) are uncorrelated.
5.10 Derive the optimum noncoherent receiver for the Rician channel. Follow the approach as out-
lined by (5.237) and (5.238). Compute the constants K}, K2, and K3 for the unequal energy
case. Simplify your results for the equal energy case.
5.11 One form of combating fading is to employ diversity. In this case, multiple channels are used to
transmit the signal using different frequencies (frequency diversity), multiple antennas (spatial
diversity), different time slots (time diversity), or different polarizations (polarization diversity).
400 Noncoherent Communication with Waveforms Chap. 5

ny (t)

Rayleigh channel 1
(rq, 84)

Rayleigh channel 2
(rp, 82)
DECISION
i COHERENT
HO), Uneees WS RECEIVER

Rayleigh channel L
(TL; 81)

Figure P5.11

The signal propagates through L channels for L'>-order diversity to produce

ri(t) = riW2ag(t) cos(wet + belt) +61) + nit), O<t<T

for i =1,2---,L, when message mx is transmitted (out of M messages). The parameters
(rj, 6;) are Rayleigh (with parameter 0”) and uniform distributed and are independent from
each other and from the parameters of other channels, as shown in Fig. P5.11.
(a) Assuming that the channel attenuations (7;, i = 1, 2,---, L) and phase shifts are perfectly
known at the receiver, derive the optimum coherent receiver for L = 2 when n;(t) is white
Gaussian noise with one-sided PSD No (independent of 7) and the noises are independent.
(b) Generalize your results for an arbitrary L and draw the optimum coherent receiver.
5.12 Using the optimum receiver derived in Problem 5.11
(a) Show that the probability of error for BPSK is given by

cee Gaal 2 Caen)

ia ee de ee ee

where

i (Eav/ No)

1 + (Egy/ No)
and (Eqy/No) is the average SNR per channel and is given by

Bo E
No No
(b) When (Eqy/No) > 1, use the approximations (1 + w)/2 ~ 1, (1 — )/2 ~ 1/4(Eqy/ No)
and
Ll
> eee
= k L

to show that

8 1 Eo
mi2= (ea) L )
Problems 401

Rayleigh channel 1
(ry, 84)

Rayleigh channel 2
(Tp, 89)
i=0,.1,..., M-1 NONCOHERENT DECISION
RECEIVER

Rayleigh channel L
(rL, 8)

Figure P5.15

5.13 Repeat Problem 5.12 using binary, orthogonal FSK and show that the same results hold, with
(Eav/No) replaced by (Eav/No)/2.
5.14 The transmitted electric field of a communication signal is circular polarized and used to pro-
— dual diversity in a Rayleigh fading channel, using binary orthogonal FSK signal. Assume
Ora 2:
(a) Assuming a coherent receiver, find the error probability if E,/No = 20 dB.
(b) Compare the result in (a) to the case where no diversity is used.
(c) Repeat (a) if the channel adds only white noise, that is, does not fade.
5.15 Repeat Problem 5.11, assuming that the phase information about the various channels is not
available at the receiver. In this case, the receiver combines the signals noncoherently as shown
in Fig. P5.15. Draw the optimum noncoherent receiver.
5.16 Show that the performance of the noncoherent receiver (derived in Problem 5.15) for binary,
orthogonal FSK is given by

1 \" A (L+k-1\ (14+ BY)

rie= (55) >,( eras)
where B = Egy/No.
Hint: See [25].
5.17 A binary noncoherent receiver is required to operate in a Rayleigh fading channel with a bit
error probability of 10°, using binary orthogonal FSK signal.
(a) What is the required E,,/No assuming 0? = 1/2?
(b) Design a diversity system such that the 10-5 bit error rate can be met if an E,/No < 15 dB
is available.
(c) What is the required E;,/ No in your design and what is the power savings over that required
in (a)?
5.18 Consider the M-ary system in Problem 5.11. All the assumptions remain the same except now
we assume that the channels are independent Rician [with parameters (r;, 9;; @;, 6;)] instead of
Rayleigh, as depicted in Fig. P5.18. Assume that the amplitude a; and phase 4; of the specular
component of channel i is known fori = 1, 2,---, L.
(a) Derive the optimum coherent receiver for L = 2.
(b) Generalize the result to an arbitrary L and draw the optimum coherent receiver.
 Noncoherent Communication with Waveforms Chap. 5
402

n4(t)

Rician channel 1
(r4, 04; O44, 54)

Rician channel 2
(ro, 89; a9, 52)
cd COHERENT DECISION
ne RECEIVER

Rician channel L
(r, OL; OL, 5)

Figure P5.18

Rician channel 1
(r4, 64; Oy, 34)

Rician channel 2
(rp, 89; a, 59)
NONCOHERENT DECISION
i=0, 1,. ., M-1 RECEIVER
se ign

Rician channel L
(r_, 8; &, d,)

Figure P5.19

5.19 Repeat Problem 5.18 assuming that the phase of the specular component in channel i is not
known for all i and that the 4;’s are independent from channel-to-channel. Draw the optimum
noncoherent receiver as shown in Fig. P5.19.
5.20 In noncoherent combining,” one decision rule in M-ary communications is to decide on mes-
sage m; (j =0, 1,---,M — 1) if and only if
yi {iG 1h 2
>a:Sp = > |
(| rg(t)V2a;(t) Ccos(wet + bi(opat)
0
T 2

+ (/ ryp(t)V 2a; (t) sin(@et + oi(0)at) |

0

2. W. C. Lindsey, “Error Probabilities for Rician Multichannel Reception of Binary and N-ary Signal,’ JEEE
Transactions on Information Theory, October 1964, pp. 339-50.
Problems 403

is maximum for i = j, where r;(t) is the received signal in channel k. The combining rule is
optimum for Rayleigh channels but not for Rician channels. Assuming that the channels are
Rician with respective parameters (rx, 0,3 ax, 6x) fork = 1,2,---,L
(a) Show that the probability of error for binary (M = 2) orthogonal FSK is given by

I Ss =
BO- Gea) ?\-s5]2g\ +) k=0

—_—— F, | —k, L; -————_———

eS se en,
where B = 207(E;/No), nk = a?(E;/No), r= Seri nz and ; F\(a, c; z) is the confluent
hypergeometric function defined by (5A.26).
(b) Let a, =0 for k=1,2---,L and show that the result in (a) reduces to that in Prob-
lem 5.16, that is

1 ib,dbp ee 1+ k
P(EY = —— eats
oe (eA) Dal k Nees
where 6 = Ey/No as all the signal power is in the Rayleigh component [see (5.253)].
Hint: see [25].
BPA Given the M-FSK signal set

s(t) = Eee OSE

otherwise

with $;(t) = /2/T cos[(w, + w;)t + 4] fori =0, 1,---, M — 1 and 6 a uniformly distributed
random variable. Find the minimum frequency spacing required to obtain orthogonal signals if
(a) Coherent detection is used,
(b) Noncoherent detection is used,
(c) When T = 10~’ sec and M = 1024, what is the required minimum bandwidth for coherent
detection?
(d) Repeat (c) if noncoherent detection is used.
5.22 A BFSK transmitter sends one of two orthogonal signals

Sj(f)' =../ 2B / T cos(a;t + @)

for 0 <t <T. Here, @ is a uniformly distributed random variable. The receiver knows @;, E,
T, and No exactly.
(a) Sketch the optimum waveform receiver.
(b) What is the message error probability and the energy per bit required if E,/No = 2 dB and
T = 290° Kelvin?
5.23 For the binary signals defined in Problem 5.22, determine and sketch the optimum receiver if
(a) £@) = [8@ + %) + 8@ — §)].
(b) Repeat Problem 5.22(b).
5.24 Assume the same signal model as in Problem 5.22. The channel model is as shown in Fig.
P5.24. Assuming the multiplier “r” is a Rayleigh distributed random variable,
(a) Determine the optimum decision rule for noncoherent detection.
(b) Sketch the optimum waveform receiver.
(c) Derive an expression for the average bit error probability.
(d) Sketch the average bit error probability if o = 1 (o is the parameter which characterizes
the Rayleigh random variable).
 404 Noncoherent Communication with Waveforms Chap. 5.

5,(t) r(t)

r n(t) Figure P5.24

REFERENCES
1. Helstrom, C. W., Statistical Theory of Signal Detection, 2nd ed., Oxford: Pergamon Press, 1968.
2. Papoulis, A., Signal Analysis, New York: McGraw-Hill, 1977.
3. Abramowitz, M. and I. A. Stegun, Handbook of Mathematical Functions, New York: Dover
Publications, 1970.
4. Watson, G., Theory of Bessel Functions, London and New York: Cambridge University Press,
1944.
5. Bracewell, R., The Fourier Transform and Its Applications, New York: McGraw-Hill, 1978.
6. Charles, F. and N. Shein, “A Preliminary Study of the Application of Noncoherent Techniques
to Low Power Telemetry,” Pasadena, CA: Jet Propulsion Laboratory, Technical Memorandum
3341-65-14 (Reorder No. 65-815), November 1965.
7. Charles, F., N. Shein and J. Springett, “The Statistical Properties of the Spectral Estimates Used
in the Decision Process by a System Analyzer Receiver,’ Proceedings of the National Telemeter-
ing Conference, San Francisco, CA, May 1967.
8. Chadwick, H., “Time Synchronization in an MFSK Receiver,’ Pasadena, CA: Jet Propulsion
Laboratory, SPS 37-48, vol. III, 1967, pp. 252-64. Also presented at the Canadian Symposium
on Communications, Montreal, Quebec, November 1968.
9. Cooley, J. and J. Tukey, “An Algorithm for the Machine Calculation of Complex Fourier Series,”
Math. of Computation, vol. 19, pp. 297-301, 1965.
10. Simon, M. K. and J. C. Springett, “The Performance of a Noncoherent F SK Receiver Preceded
by a Bandpass Limiter,’ JEEE Transactions on Communications, vol. COM-20, no. 6, December
1972, pp. 1128-36.
11. Papoulis, A., Probability, Random Variables and Stochastic Processes, New York: McGraw-Hill,
1965.
12. Gradshteyn, I. S. and I. M. Ryshik, Table of Integrals, Series and Products, corrected and en-
larged edition, Florida: Academic Press, 1980.
13. Weber, C. L., Elements of Detection and Signal Design, New York: Springer-Verlag, 1987.
14. Nuttall, A. H., “Error Probabilities for Equicorrelated M-ary Signals under Phase-Coherent and
Phase-Incoherent Reception,” JRE Transactions on Information Theory, July 1962, pp. 305-14.
15. Nuttall, A. N., “Error Probabilities for Nonorthogonal M-ary Signals under Phase-Coherent and
Phase-Incoherent Reception,” Waltham, MA: Litton Systems Inc., Tech Report TR-61-1-BF, June
1961.
16. Helstrom, C. W., “The Resolution of Signals in White Gaussian Noise,” JRE Proceedings, vol.
43, September 1955, pp. 1111-18.
17. Proakis, J. G., Digital Communications, New York: McGraw-Hill, 1983.
18. Viterbi, A. J., Principles of Coherent Communications, New York: McGraw-Hill, 1966.
19. Stein, S., “Unified Analysis of Certain Coherent and Noncoherent Binary Communication Sys-
tems,” [EEE Transactions on Information Theory, vol. IT-10, January 1964, pp. 43-51.
References 405

20. Nakamoto, F., R. Middlestead and C. Wolfson, “Impact of Time and Frequency Errors on Pro-
cessing Satellites with MFSK Modulations,” JCC 8] Conference Record, Denver, CO, June 1981,
pp. 37.3.1-37.3.5.
Die Chadwick, H., “The Error Probability of a Spectrum Analysis Receiver with Incorrect Time and
Frequency Synchronization,” International Symposium of Information Theory, Ellenville, NY,
January 1969.
22 Wittke, P. and P. McLane, “Study of the Reception of Frequency Dehopped M-ary FSK,” Re-
search Report 83-1, Queen’s University, Kingston, Ontario, March 1983.
23% Frasen, D., Nonparametric Methods in Statistics, New York: John Wiley and Sons, 1957.
24. Lindsey, W. C., “Error Probabilities for Rician Fading Multichannel Reception of Binary and N-
ary Signals,” IEEE Transactions on Information Theory, vol. IT-10, October 1964, pp. 339-50.
2d: Lindsey, W. C., “Error Probabilities for Incoherent Diversity Reception,’ JEEE Transactions on
Information Theory, vol. IT-11, October 1965, pp. 491-99.
26. Turin, G. L., “Error Probabilities for Binary Symmetric Ideal Reception through Nonselective
Slow Fading and Noise,” Proceedings of the IRE, vol. 46, September 1958, pp. 1603-19.
Dike Van Trees, H., Detection, Estimation and Modulation—Part I: Detection, Estimation and Linear
Modulation Theory, New York: John Wiley and Sons, 1968.
28. Lindsey, W. C. and M. K. Simon, Telecommunication Systems Engineering, Englewood Cliffs,
N.J.: Prentice Hall, 1973. Reprinted by Dover Press, New York, NY, 1991.
Hug). Hinedi, S., M. K. Simon and D. Raphaeli, “The Performance of Noncoherent Orthogonal M-
FSK in the Presence of Timing and Frequency Errors,” to be published in JEEE Transactions on
Communications.
 Chapter 6
Partially Coherent
Communication
with Waveforms

When the incoming carrier phase is estimated in the receiver, the latter can employ coherent
detection schemes as discussed in Chapter 4, provided the estimation is performed perfectly.
In many practical communication systems, the estimation results in a residual carrier phase
error, the amount of which depends on various parameters such as signal-to-noise ratio, fre-
quency stability, the estimation algorithm itself, and so on. Given the statistical description
of the phase error, the communication engineer is faced with two choices: design the de-
tector assuming a perfect carrier reference (that is, coherent detection) and then account for
any degradation due to the nonideal nature of the estimate, or as a second alternative, imple-
ment the optimal detector which exploits the available statistical information on the phase
error to obtain an improved detector. Such a scheme, which utilizes the carrier phase error
information, is referred to as partially coherent detection as it is neither coherent nor non-
coherent in the absolute sense. If the estimation is improved to the point where it is perfect,
then we expect the partially coherent receiver to approach the coherent receiver of Chap-
ter 4. On the other hand, if the estimation is degraded to the extent of providing no credible
information about the phase error, then we expect the partially coherent receiver to reduce
to the noncoherent detector, as discussed in Chapter 5.
In partially coherent communications, the structure of the optimal detector depends
heavily on the statistics of the phase error. This is to be expected since the optimal detector,
by definition, fully exploits the available information. The phase error itself is typically
a random process that we assume is slowly varying over a symbol (or multiple symbols)
and can thus be modeled as a random variable. Assuming that the communication system
transmits an unmodulated carrier for the sole purpose of aiding synchronization in the
receiver, a phase-locked loop (PLL) can then be used to track the incoming carrier phase
0. For a first-order PLL, the stationary pdf of the carrier phase error ¢, defined as 9 — A)
where @ is the locally generated carrier phase estimate, is given by the Tikhonov density [1 )

2], depicted in Fig. 6.1 and can be expressed as

ea
AO) ’ |p| < os
fo) = Io (pg)
27
; (6.1)
0F otherwise
where pg denotes the carrier loop signal-to-noise ratio (SNR) and is given by
 Partially Coherent Communication with Waveforms 407

4.0

3.5

3.0

{,) 2.0

1.0

0.5

(6.2)

In (6.2), P- denotes the power of the unmodulated carrier (which is unrelated to the power
of the information bearing signal), No is the one-sided PSD of the white Gaussian noise,
and B, is the one-sided bandwidth (in Hz) of the PLL. The synchronization aspect of the
receiver will be the topic of Volume II of this book and it suffices to state here the tracking
performance as characterized by the pdf of (6.1). Note that (6.1) is the exact pdf for a first-
order analog PLL but can also serve as an approximation for a second-order PLL, provided
that py is sufficiently large. Moreover, since the phase reference is arbitrary, we may set
@ =0 without loss in generality so that the received phase 6 is identical to @ (that is, 6 = @),
as depicted in Fig. 6.2. Partially coherent reception is a more general mode of operation than
previously considered since
 408 Partially Coherent Communication with Waveforms Chap. 6

V2 s(t) cos (ag t + 8)

V2 cos(w,t + 6)

(a) Practical Implementation

V2 s(t) cos (ae t+) s(t) cos

ey

V2 cos wet

(b) Equivalent Mathmatical Model Figure 6.2 Phase error generation

exp(pgcosd) _ 1
i = — (6.3)
pe>0 27 I10(pg) 20

corresponds to noncoherent reception, while

pg>
Sane DeOe 5
=27 Io(pg)
(6.4)
corresponds to coherent communication as ¢ = 0 with probability one. Since, in principle,
p@ can take on any value between 0 and oo, all intermediate cases are possible.

6.1 OPTIMUM PARTIALLY COHERENT RECEIVERS BASED ON ONE SYMBOL

OBSERVATION
In partially coherent channels, the received signal is still given by (5.2) with the understand-
ing that @ is no longer uniformly distributed, but rather distributed according to (6.1) since
0 = @. In this case, the conditional pdf of r is still given by (5.19), namely

fr/o(e/mi, 9) ~ exp a cos(@ + 6| exp (-=) (6.5)

No No
Assuming equal energy signals and averaging over the pdf of 6, (6.5) becomes

fr(e/mj) ~ feet : exp ea cos(@ + 6| exp(pg cos 6)dé (6.6)

2m Io(P¢) J—n No
Using trigonometric identities, (6.6) reduces to
1 If

fr(p/m;) ~ Dine) iSexp {x;cos(@ + vi)| dé (6.7)

where

26; Su Qe 4
ie A (= cos @; + vs) + (= sin 6) (6.8)
 Sec. 6.2 Performance of Partially Coherent Receivers/One Symbol 409

and

(2&;/No) sin ¢;
tan yj; & (6.9)
(2&;/No) cos ; + pg
Using the definition of Jo(x) as given by (5.22), (6.7) can then be expressed as

1o(x;)
Sr(p/mi) ~ To(og) (6.10)

Recalling once again that Jo(x) is a monotonically increasing function of its argument, the
MAP receiver sets m(p) = m; if and only if x Ais maximum when k = i. Using z,; and Zs;
as given by (5.17), the rule reduces to maximizing

oD) 2 2 2
Xx; = No ci + Po No si .

For the case where ¢;(t) = 0 Wi and a;(t) = s;(t), then the I-Q projections z,; and Zs;
are given by (5.15) and are repeated here for convenience:
T
la = / Pe(t)si(t)dt (6.12a)
0
me
Zsi = / Ps(t)si(t)dt (6.12b)
0

Alternatively, the rule can be simplified to maximizing

2 2
Zci Zsi Zci
= — — 6.1
Ga) ae + b6 (#) Ora
which clearly indicates that the optimum partially coherent detector is simply a linear com-
bination of the optimum coherent and optimum noncoherent detector with the weighing fac-
tors pg and 1, respectively.! Two alternate implementations of the partially coherent detector
are shown in Fig. 6.3. In both cases, knowledge of No as well as og are required for optimal
detection. Clearly, as pg — 0, the partially coherent receiver reduces to the familiar non-
coherent receiver and as py — oo, the noncoherent receiver contribution may be ignored,
resulting in the coherent receiver.

6.2 PERFORMANCE OF PARTIALLY COHERENT RECEIVERS BASED ON ONE

SYMBOL OBSERVATION
It would interesting to compute the performance of partially coherent receivers for various
signals and determine their advantages over noncoherent reception. In this section, we will
consider binary antipodal, binary orthogonal, binary nonorthogonal, and finally, M-ary
orthogonal signals.

1. Note [3] that if we define X; = (2/No)zci and Y; = (2/No)zs;, then the rule of (6.11) becomes (X; + po)” +
Yy or after expansion, (Xx? + Vf} + 2p4X;i, which indicates a weighting of 2g and 1 for the coherent and
noncoherent contributions.
 Partially Coherent Communication with Waveforms Chap. 6
410

r%
(ing)80"

C
fe)
M
=)
A
R
A
T
r fe)
V2 cos(w.t + 6) R

r.(t)
S?)
>
LPF :

r (t) ®

KA Pr

V2 sin(wet + 6)

Figure 6.3a Two alternate implementations of the optimal partially coherent receiver

6.2.1 Performance of Binary Antipodal Signals

When the signals are antipodal, so(t) = —s,(t) and as a result, the noncoherent contribution
(that is, i + Za) to the partially coherent detector is independent of which signal was
transmitted since the squaring operation destroys the sign of the signal. Thus, the optimal
receiver for antipodal signals is identical to the coherent receiver whose performance in the
presence of carrier phase jitter was computed in Chapter 4 and given by (4.363) and (4.364),
Sec. 6.2 Performance of Partially Coherent Receivers/One Symbol 411

DOHYFVDYVEOO

V2 cos(w,t + 8)

(b)
Figure 6.3b Two alternate implementations of the optimal partially coherent receiver
412 Partially Coherent Communication with Waveforms Chap. 6

that is

ote IT Sa r “Ecos dé (6.14)

O\Wd

The performance is actually plotted in Fig. 4.46 for various values of loop SNR, pg. When
the latter is less than 14 dB, significant degradation starts to occur and gets worse with
decreasing py. For the limiting cases of noncoherent (pg = 0) and coherent (pg = oo) de-
tection, (6.14) reduces to
1
lim P,(E) = = (6.15a)
po 2

IE
lim P,(E)=@Q ( ) (6.15b)
po oo No

the latter being identical to (4.55).

6.2.2 Performance of Binary Orthogonal Signals

For binary orthogonal signals, the receiver compares x, and x, and selects the message
corresponding to the largest value. Conditioning on the phase error @¢, the variables Xo and
x, become Rician distributed (see Appendix 5A) with parameters 5A and sa which can be
computed using (5.56) and (5.57) to yield (assuming message my, is transmitted)

2 2Es cos @ + ie Zh, ¢ ji

Sy 1 = No Po ——
No sin
(6.16)
= aE
No ie:Po244 (=
No
=-_ Po cos o
and

59 = p3 (6.17)
Since both z,; and zs; have identical variance given by (5.58), then oO; and oe [equivalent to
o” in (5A.27)] are equal and given by

2, 2 2 y NoE;s
Ofs E Ss
0; = =9% =i (=)
— = 2s (=)
aatek (6.18)

Because of the symmetry of the problem, P;(E/m}) is identical to P,(E/mo). The condi-
tional probability of error, P;(E/@), is then the probability that one Rician variable exceeds
another independent Rician variable, which is given by (5.130). Hence

P(E /) = Prob {xy > x,/p, mi}

(6.19)
=@Q (Va, vb) a exp (-S *)Io (ab)

where
2 2
P04 waa
ih Daas ea (32) oe
Sec. 6.2 Performance of Partially Coherent Receivers/One Symbol 413

and
2
GE)
2Es
a1; P5 +4 (#)
s
po cos
i

= SPC gar Seay Ee ea (6.21)

The probability of error is then obtained by averaging over the pdf of ¢, that is

P(E)= |
base
PO uo | _ (PsE/G) dip,
27 Io(p¢)
62D)
If we take the limit of (6.22) as pg approaches zero, we obtain the result for noncoherent
reception of binary orthogonal signals, namely

lim P,(E) = —: exp |———

= } = limlim P,(E/) (6.23)
‘
pp0 2 P 2No ee

which is identical to (5.79). To obtain the limit as pg — oo, it is more convenient to return
to the optimal detection metric as given by (6.13) and note that we are only interested in zo
and z,1. In this case, it can be shown that (see Problem 6.9)

line. (E) =O (=| (6.24)

pg>oo No

which is identical to (4.59). Note that for large E;/No, the limiting cases are not very
far apart, since asymptotically Q(./E;/No) approaches exp(—Es;/2No)//27 Es/No [see
Appendix 3B, (3B.12)], so that for transmission with orthogonal signals, the amount of
coherence is not a crucial matter for performance, as it is with antipodal signals.

6.2.3 Performance of Binary Nonorthogonal Signals

To compute the performance for arbitrary binary signals, we need to evaluate Z¢0, Zs0, Zc1,
and z,1 in the presence of carrier phase error ¢. Assuming that mo is transmitted, then from
(5.86) through (5.90), we have (let 6B4 Boi, a A a1)

Zc0 = Es cosh + neo (6.25a)

250 = Es sing + nso (6.25b)
Ze] = E;acos¢ — E;B sing + ne} (6.25c)
Zs1 = E;Bcoso + Esa sing + ns (6.25d)

where E{n-one1} = E{nsonsi} = (EsNo/2)a and E{neinso} = —E{nsinco} = (EsNo/2)B.

We will follow the same procedure as in Chapter 5, where the observations are first whitened
and then the probability of error is computed using (5.130). Rewriting x; in (6.11) as

XL =Zit+Y; (6.26)
where
é)
Z A —Zck + bp (6.27a)
No
5
ae Now (6.27b)
414 Partially Coherent Communication with Waveforms Chap. 6

then we define the vector X as

X? =(Z Yo Zi Yi)" (6.28)

with mean
po + (2Es/No) cosd
= (2Es/No) sing : (6.29)
mx | py + (2Es/No)(a cos$— Bing)
(2E;/No)(B cos¢ + a sind)
and covariance matrix
140 7Saal
2Es OL Bre)
Kx = (No ) Vis esha eK (6.30)
6.30
Bie eA Oa
As in Chapter 5, we whiten the vector X using the transformation
Wye 0 —xr —B
_ VNo/Es QO I+k B —k (6.31)
2kV1+k\ —A Peta 0
—B —ir ONAC
to get the vector

Zw,0

KPa 7Tx Ayeso

Zw,l
(6.32)
Yo,

with (see Problem 6.10)

Kx, =TKxT' = (6.33)

and
2E
X7 OX =k (=) es OX.) (6.34)
where k is given by (5.122) and Q by (5.119). Due to the symmetry of the problem, the
conditional probability of error is given by

P,(E/) = Prob(X!OX < 0/mo) = Prob(X! OX, < 0/mo) (6.35)

The components of X,, are conditionally uncorrelated and independent with unit variance
and means
a po(1 +k — a) + 2(E;/No)k(k+ 1) cos¢
_ VNo/Es PoB + (2Es/No)kU +k) sing
Se
E{X,}
= 7
FEMEET |p61 +k — at) +2(E,/Noyk(a cos — psing) |©2
E{X} = ———

—pgB + (2Es/No)k(a sind + Bcos d)

Letting or A Le oe Vee then x ,i, i = 0, 1 are Rician variables with parameters

8) = (E{Zw0})” + (E{%w,0})” (6.37a)

st = (E{Zw,1})? + (E{Yw,1)) (6.37b)
Sec. 6.2 Performance of Partially Coherent Receivers/One Symbol 415

and since the variables have been normalized, n=ao Cn= 1. The conditional probability of
error is then given by (5.130), that is

P(E/#) = Q (Va, Jai) ~ 51 exp (-* 7 Tee (6.38)

where

a;Ree
= 5 {a? + 6?Ben + 2a orfi cos(o — ¢)}, oul
we (6.39)
and

jp
og =a,
= /_—_8 (6.40a)
kV (2Es/No)
Bo = (1 + k)(Es/No) (6.40b)
By = —V/(1 —k)(Es/No) (6.40c)
= B
=t —— 6.4
SS at G4) (One)
-| B
¢) = tan (4) (6.40e)

The unconditional probability of error is then given by

Pe(E)= [Py(E/6)
My
nx
POPP
exp(0g COS f)
27 Ip(p¢)
ag 6.41)
which is evaluated numerically in Figures 6.4 and 6.5 for 8 =0 and @ between 0 and
—1. The error probability is plotted versus pg where E;/No is selected so that the error
probabilities for coherent reception of antipodal signals are 10-7 and 10~°, respectively.
These correspond to E;/No = 6.8 dB and 9.58 dB, obtained from Fig. 4.46 with pg = oo
The curve a = —1 corresponds to antipodal signals while the case w = 0 corresponds to
orthogonal signals. The effect of nonzero imaginary signal correlation (that is, nonzero £) is
illustrated in Fig. 6.6 for E;/No = 9.58 dB. It is clear that the best performance is obtained
when B = 0.

6.2.4 Performance of M-ary Orthogonal Signals

With M-ary orthogonal signals, the receiver selects the message corresponding to the maxi-
mum x,. Assuming that message m, is sent, then x,, 1 = 0,1,..., M — 1 are all condition-
ally Rician distributed with parameters as given by (6.16) and (6.17), that is

2Es 2 1 yee
9? = (78)+ 6 + ACR) PG COS, ae (6.42)
Po l - k

Furthermore, z,; and zs; have identical variances for i = 0, 1,..., M — 1, given by o2=
2E;/No. The conditional probability of correct detection is then given by

Ps(C/mg, @) = Prob{x, = max x,,i=0,1,...,M—1J (6.43)

U
416 Partially Coherent Communication with Waveforms Chap. 6

ul

=oooooo
bh
©
ODN
ReRRRRRR
0
-0.
-0.
owou -0. wr
nou
RRee

"%
Figure 6.4 Performance of the partially coherent receiver for M = 2 and real nonpositive
a (B =0, E;/No = 6.8 dB) (Reprinted from [3] © IEEE)

Since the signals are all orthogonal, the variables x, are all independent. Hence
CO

P;s(C/mx, &) = / Fx OK)

0
Xk Xk
x fof fcrOd Fy Ou fe,ORHD fe,, COND
M-1 integrals
dxodx +--+ dxp—jdxp41 +++ dxy—idx (6.44)

The variables x,, i =0,1,...,M —1 and i #& are all identically distributed. Therefore,
(6.44) becomes

fone) Xk M-\
P;s(C/mx, &) = [ fx,(XK) | f,,(s04x0| dx, (6.45)
0
Using the definition of Q(a, b) as given by (5A.15), the inner integral reduces to

SQ Xk
i fy, (0) dxpx0 =1—- Q (= a )
—,— (6.46)
Substituting (6.46) into (6.45) and simplifying, we have
Sec. 6.2 Performance of Partially Coherent Receivers/One Symbol 417

Figure 6.5 Performance of the partially coherent receiver for M = 2 and real nonpositive
a (B =0, E;/No = 9.58 dB) (Reprinted from [3] © IEEE)

Ps(C/mx,
&) = Ps(C/)
00 pie 6.47
=i erp (-£ = )vent Ove, 14a oy
0

where
2)

EAI, vee 2p4 cos (6.48)

i 2(Es/No)
p Zz
ee (6.49)
~ 2(Es/No)
 Partially Coherent Communication with Waveforms Chap. 6
418

%
Figure 6.6 Effect of nonzero B (M =2, E;/No = 9.58 dB) (Reprinted from [3] © IEEE)

The unconditional probability of symbol error becomes

o4 (o°e) 24 2
P;(E)= 1 - | [ y exp & = )wen
(6.50)
_1 €XP(o¢g Cos P)
x [1 — Q(c2, y)|” dydo
27 I9(P¢)
The probability of symbol and bit error [obtained using (5.154)] can then be computed
numerically for various values of M and loop SNR pg.
 Sec. 6.3 Optimum Partially Coherent Receivers/Multiple Symbol 419

6.3 OPTIMUM PARTIALLY COHERENT RECEIVERS BASED ON MULTIPLE SYMBOL

OBSERVATION
In this section we consider the structure of the optimum receiver for the partially coherent
channel defined by the phase error statistics of (6.1) when the observation interval is not
restricted to a single symbol interval [4]. Since the presence of a constant phase error
between the incoming carrier phase and the receiver’s estimate of it introduces memory
into the observable (assumed to be greater than T sec in duration), one should exploit this
property in designing an optimum decision algorithm. In particular, one should resort to
a form of sequence estimation wherein a joint decision is made on a group, say Ns, of
information symbols, the length of the sequence being proportional to the duration over
which the phase error can be assumed constant.
As in Section 6.1, we take a maximum-likelihood (ML) approach to this problem. All
our attention will be focused on the case of M-PSK modulation. As such, the analogous
results to be presented in Chapter 7 for optimum differential detection of M-PSK can be
viewed as a special case of the results derived here corresponding to pg = 0. We shall say
more about this later on.

6.3.1 Maximum-Likelihood Block-by-Block Detection of M-PSK Signals

Since M-PSK is a two-dimensional modulation, it shall be convenient to deal with the

complex representation of signals. Thus, the baseband representation of the transmitted
signal in the kth transmission interval kT, < t < (k + 1)T; has the complex form?

5p = J2E;/ Tet (6.51)

where 6; denotes the transmitted phase in this interval which for M-PSK takes on values
from the set Bm = 27m/M; m=0,1,2,...,M— 1. Assume that the phase @ introduced
by the channel can be assumed constant (independent of time) over a duration of N; data
symbols but is otherwise unknown. The bandpass representation of the received signal in
this same time interval is, in accordance with (4B.1), given by

r(t) = Re {r(t)} = Re {5, exp (j@ct + 9) + n(t)} = sx (t; 0) +n) (6.52)

where 7i(t) is a complex additive white Gaussian process with single-sided PSD 2No. The
receiver demodulates 7(t) with the complex reference signal exp { 4] (cet = 6) to pro-

duce the complex baseband signal S,e/? + n(t)e Je +4) which is then passed through an
I&D, resulting in the complex detected sample?

2. As in Section 4.1.6, it is necessary to introduce a notation for characterizing discrete time when dealing with
observations that span more than one symbol interval. As such, we shall once again use a subscript, for example,
k, on a variable to denote the value of that variable in the k"* transmission interval. To avoid confusion with the
previous notation of 6, to denote the k' message phase 27k /M (k'® hypothesis), we introduce a new Greek
symbol, that is, Bx, to denote the k'h message phase. Whenever it is necessary to represent a variable both in discrete
time and/or by an associated hypothesis, we shall use a parenthetical superscript for the latter.
3. It is convenient in what follows to introduce a normalization of 1/7, into the definition of the complex I&D
output sample. This normalization is denoted by the use of a script typeface.
420 Partially Coherent Communication with Waveforms Chap. 6

f . 1 P&tDTs | ‘
Vy, = Spel? + 7 / n(tye /°'dt |e-/
s k JkT, (6.53)

= spel? + N.e/?

where NM, is a zero-mean complex Gaussian random variable with variance 2No/T;, or
equivalently, a variance a, = No/T; for the real and imaginary components. For an obser-
vation of the received signal over the interval (k — N; + 1) T; < t < (kK + 1)T;, the sequence
v= (Viewer Ve-N, +21 tes View Vx)conditioned on ¢ is a sufficient statistic for making
an ML decision on the transmitted sequence § = (SKN, +15 SEs Niland Pika 5k).
For the assumed AWGN model, the a posteriori pdf of v given S and ¢ is
Fe 22

fy gVis.
9)= (2702) exp |— |b — 5e’9|
Vie Do

1 1 > [fe 2 at ate (6.54)

Sea TD. Voaaes \-i| + |S—i|
(2702) {a2 | :
Ne-1~
—2 Ve-i5Z_; |COS(@ — a|
ji==10)

where
Nl
Iim-fu oe Ve-iSy_;
a A tan aes (6.55)
Re|ss nat
i=0

Averaging (6.54) over the pdf in (6.1) gives, upon simplification

Fy P= | Fy 08.0 foods
~ uw ~

1 l a Raigad
= an Seer i Y =H 5 =i] :
Io(Po) (27202)™ of 20; 2, | ith che Sa ] (6.56)

a it iste Nox! :
Xda b=
o2
yy Vai},
TI kj
|cosa+ P¢ pgos
n | + Vi_.5*
k-iSp_; :
sing

Sau i=0
Since
N-1 N-1
ye Ve-i5f_;| COS = Re 3 Pesa fy
i= i=0
N-1 ees (6.57)
» Ve-iS¢_;| Sina = Im se
= i=0
Sec. 6.3 Optimum Partially Coherent Receivers/Multiple Symbol 421

Equation (6.56) further simplifies to

Ce
Wh l Se
l if +vei?
ty eee(eS |pee (6.58)
\ 1o(P$) (2702) P| 22 a+ |Sk—i| ki]

1 Ned : heels :
x Ip — Re ey Vea |
+ wet) ae (is ate mst.)
; eae i=0

Note that for M-PSK, Kak is constant for all possible transmitted phases B,.. Thus, maxi-
mizing Sys) over S$is equivalent to finding

Ne=1_ 2 N,—I 4
max (Re|1 ni ae ooo) + (is & ist) (6.59a)
1=0

or equivalently

Ne=1 2
max Se Ve-iS¢_, + 060°
: i=0
(6.59b)
Ne—1_ 2 Ne
= max ys Vets, Ae 20907 Re |SD aid
e i=0 i=0

Dividing (6.59a) by of and noting that

ease, 1 p&-itDTs
Saas = — F(t)Sf_; exp (—j@ect)dt (6.60)
on No J(k=i)T,
where we have arbitrarily set 6 =0 and thus 6 = , we see immediately that this met-
ric is the generalization of (6.11) for the case of multiple symbol observation (N, > 1).
As for the single symbol observation case, the first term in (6.59a) represents the compo-
nent of the decision metric associated with noncoherent (differential) detection, whereas the
second term represents the component of the decision metric for ideal coherent detection.
Thus, the partially coherent decision metric represents a hybrid of the two extremes with
a weighting of the two terms proportional to py. Note that for any nonzero value of pg, a
decision rule based on this metric is unique. For pg = 0 (differentially coherent detection),
there is a phase ambiguity associated with it since the addition of an arbitrary fixed phase,
say 0q, to all Ns estimated phases Oy N41, aie 6r_1, 6, results in the same decision for
Ok—N,+15 +++» Ok—1, %. In Chapter 7 we shall see that by letting 6, = 6,—n,+1 and differen-
tially encoding the input phases at the transmitter, that is,

Ox, = On—1 + AO (6.61)

where now A; denotes the input data phase corresponding to the k'® transmission interval
and 6; the differentially encoded version of it, then we can turn the decision rule into one in
which the phase ambiguity is resolved. For the moment, we assume pg # 0 and thus there is
no formal requirement for differentially encoding the data phase symbols.
 422 Partially Coherent Communication with Waveforms Chap. 6

Figure 6.7 is an illustration in complex form of a receiver implemented on the basis of

(6.59b). Note that this receiver requires knowledge of the loop SNR pg (as well as the sig-
nal power P = E,/T; and the noise variance a.) The accuracy of this knowledge, which
must be obtained by measurement, will have an impact on the ultimate performance of this
receiver. Later we shall investigate the sensitivity of the receiver to a mismatch between the
true loop SNR and the value supplied to the receiver implementation in Figure 6.7. For the
moment, we assume that the receiver has perfect knowledge of pg and thus should outper-
form a conventional bit-by-bit correlation receiver that does not make use of this knowledge.
The next matter under investigation is to determine just how much the optimum partially co-
herent sequence receiver outperforms the conventional bit-by-bit correlation receiver, which
was shown in Section 6.2.1 (at least for binary signals) to be the same as the optimum par-
tially coherent receiver based on a single symbol observation interval.

6.4 PERFORMANCE OF PARTIALLY COHERENT RECEIVERS BASED ON MULTIPLE

SYMBOL OBSERVATION
To obtain a simple upper bound on the average bit error probability, P,(E), of the pro-
posed block-by-block detection scheme, we use a union bound analogous to that used
for upper bounding the performance of error correction coded systems (see Chapters 12,
13). In particular, the upper bound on P,(£E) is the sum of the pairwise error proba-
bilities associated with each N,-symbol error sequence. Each pairwise error probability
is then either evaluated directly or itself upper bounded. Mathematically speaking, let
6 = (O—Nn, +1, k—N, +2, +--+» O—1, O%) denote the sequence of N, data phases and 6=
(6,_ Nee ts By dep eC O—1, 6x) be the corresponding sequence of detected phases. Let
u be the sequence of b = N; log, M information bits that produces @ at the transmitter and
ii the sequence of b bits that result from the detection of @. Then, since M-PSK is a sym-
metric signaling set, that is, it satisfies a uniform error probability (UEP) criterion,’ we get
the upper bound on conditional (on #) probability of error?

I a
WO esl rey Dd) w(u, ti) Pr {i}F > n16, 9} (6.62)

where w(u, U) denotes the Hamming distance between u and G, @ is any input sequence,
(for example, the null sequence (0, 0,...,0) = 0), and Pr {7 > ld, >} denotes the condi-
tional pairwise probability that 6 is incorrectly chosen when indeed @ was sent.
Finally, the decision statistic 7 is defined from (6.51) and (6.59b) by
Ny] 2 |?
= y ¥
Y-je
—jORj Poon
JPk-i + —— .
0 V2Es/Ts (662)

and the corresponding error statistic 7 is identical to (6.63) with each 6, replaced by By.

4. By UEP we mean that the probability of error conditioned on having sent a particular phase is independent of
which phase is indeed sent.
5. The result in (6.62) is a special case of the more general result in (4.88) wherein M is replaced by MNs and
because of the UEP condition, the double sum collapses to M times a single sum. Furthermore, a dependence on ¢
is now included in the notation for the bit and pairwise error probabilities.
 r
poyoieyy % Aryaq my, Aejaq Avlaq ttn, ABloq
| 'n-4
aN L L = J t

SSouppY
= af 6 hee “yWN

t9/-? aS
ZS gS See oO
e e e

AMvy-orenbs
adoyaaua 10}99)}9p

¥490TO oredwoa
pur
9S 199]24}
1SOBIE]
“WN sajaAo
dod "LN
gqeuq ae399 aun
[ea
: : E ;
Aejap

) 0 Og)
*N-¥61+ N-F,
TH 062 60 A 61-8Vp
aanS1q
"9 OONG UONRIUOW[dWI
JOB JOATOo0I
Jo} a[dry[nus joquiés uonoajep
yo Aqened ls19Yo9 MASd-W

423
424 Partially Coherent Communication with Waveforms Chap. 6

6.4.1 Evaluation of the Pairwise Error Probability

To compute Pr {7 > ni|6, >}, we use the approach taken in Section 6.2.2 for evaluating the
probability of one Rician random variable exceeding another independent Rician random
variable. In particular, letting 7 = |z) ? and 7 = \z2|? [see (6.63) and the statement below it
for the definitions of z; and zz], then®

Pr {4 > nid, >} = Pr{{z2| > |z11 16,d}

1 (6.64)
=5[1- a(vb, va) + Ova, Vy] 8 fla.b)
where the arguments of the Marcum-Q function are given by

{7} — 1 Sj1++ SpS2—2|p2| VS1 es ©, I -—0 te

—2|p2| V S18. a eee
ieee (6.65)
aj ' 2N; 1 — |pz 1 — |pz|
with

8s
bol a=
| es
=
E
are) |N4 ie
Gore je] (nee
Es
Ty \' MEL[Ngee ot
Po Po
EOS
?

S 2 |e
|
A ={z\*=
E
— 16 a+—*—e e J?|
pee aes =
5 [a2 =|

ee TN oree a No
N-
A =|z.
2 - HP = =l22
D - Zl? = Ns —; (6.66)

eC
SON oe
NEO v faeLarg p,
ee= args

6) 2 areaoa =arg |
Neel?+ Po
PI A arg =z2 =ar
@, {se!#+ Po
; 2Es;/No its S 2E;/No

and
Nsw) :
5 A ae of ($x-i- 4-1) (6.67)

Substituting (6.66) into (6.65) results, after considerable simplification, in

b Es I Po
= Ns|1+— ;
te 2No |‘| TN tana po
| p W) d
- (N; — |6| cos v)
2Ns (N2 — |5|*) (aaa) é (6.68)

Es
+ ON V Ne lol
2
ae
I
(22Po
) iwsc050 — bl eos + vp
yf NSO loa
We now consider some special cases of practical interest.

6. The equivalence between (6.19) and (6.64) is shown in [5].

Sec. 6.4 Performance of Partially Coherent Receivers/Multiple Symbol 425

6.4.2 Special Cases

Case 1. BPSK with Two-Symbol Observation and Detection (M = 2, N, = 2)

In this case, E;/No = E,/No. There are M* — 1 = 3 possible error sequences each of
length 2. The pertinent results related to the evaluation of (6.67) and (6.68) are given below:

be — k br—1 — Pk-1 )
0 TU 0
4 0 0
a4 4 —2

For the first two error sequences, (6.68) evaluates to

E i
b= — 442 Bt eos + 7 Bo
2No E,/No 4 \ Ey/No
(6.69)
a LY
poe
2No | 4 \ Ep/No

where E,/No = (Es/No) / logM is, as always, the bit energy-to-noise spectral density
ratio. For the third error sequence, both a and b approach infinity (the ratio a/b, however, °
approaches unity) as 6 approaches —2. Thus, we must evaluate (6.64) separately for this
case. It is straightforward to show that

ae J(G5 0) = 5erfe(3— ‘A (6.70)

a/b—>1

Furthermore, in the general case where 5> —WNs, (6.70) evaluates to

I | Oe
, lim, f(a, b) = 5 erfe (Ne cos ‘) (6.71)

which for NV, = 2 and M = 2 becomes

/2E
glim, f(a, b) = 5 erfe (ae cos ) (6.72)

Finally, noting that the Hamming distance w(u, @) is equal to | for the first two error
sequences and is equal to 2 for the third sequence, then substituting (6.72) and (6.64)
combined with (6.68) and (6.69) into (6.62) gives

Eel (: Pos \" | Eb ( po ) 1 ( hg )

eg aa 5 (ab) |No as E»/No ae as E»/No
11 [2Eo
2Ep
f ; erfc ( No cos ) 6.73
(6.73)

Finally, the upper bound on average bit error probability P, (E) is obtained by averaging the
upper bound in (6.73) over the pdf in (6.1). Figures 6.8 and 6.9 are plots of this upper bound
426 Partially Coherent Communication with Waveforms Chap. 6

Coherent

Probability
Error
Bit
Average
Bound
Upper
on

Figure 6.8 Upper bound on average bit

error probability versus E,/No for optimum
block-by-block detection receiver with N; as
a parameter; M = 2, pg = 7 dB

on average probability versus E»/No in dB for values of pg = 7 dB and 10 dB. For purposes
of comparison, we also show the exact results for the conventional ideal coherent metric (or
optimum partially coherent metric with NV; = 1) operating in a noisy carrier synchronization
environment as given by (6.14). We see that even with only one additional observation
bit interval, considerable savings in E,/No can be achieved at a fixed error probability,
particularly in the region of the knee of the curve where the system begins approaching its
irreducible error probability’ asymptote.

Case 2. BPSK with Ns-Symbol Observation and Detection (M = 2, Ns arbitrary)

For Ns arbitrary, 6 takes on values —(N, — 2i);i =0,1,2,..., Ns — 1. The number
of error sequences corresponding to each of these values of 6 is binomially distributed,

that is, there are >) sequences that yield a value 5 = —(N, — 27). Furthermore, the

Hamming weight associated with each of the (ae) sequences that yield a value 6 =

7. We have seen in Chapter 4 (for example, Fig. 4.46) and also earlier in this chapter that conventional BPSK
systems exhibit an irreducible error probability (i.e., a finite error probability in the limit as E,/No approaches
infinity) when a noisy carrier synchronization reference with fixed power is used as a demodulation signal. This is
observed by examining a curve of P,(E) vs. E,/No with loop SNR, pg, held fixed. The value of this irreducible
error probability is given by P,(E)|;,, = Ser fo(@)d@. Unfortunately, increasing the observation interval and
using sequence estimation does not reduce this error. Thus, any gain due to performing block detection must take
place prior to approaching this asymptotic limit.
Sec. 6.4 Performance of Partially Coherent Receivers/Multiple Symbol 427

10°.
\
\

Coherent

\ Ng

Bound
Upper
Probability
Error
Bit
Average
on

10° Figure 6.9 Upper bound on average bit

5 6 7 8 9 10 11 12 13 14 15 — error probability versus E,/No for optimum
E./N.. dB block-by-block detection receiver with N, as
b 0 a parameter; M = 2, py = 10 dB

—(N, — 2i) is w(u, @) = N, —i. Finally then, using the above in (6.68) and (6.71) and
substituting the results in (6.62), the conditional bit error probability is upper bounded by

PE
» (Cay
eyeSete
on PINee Leaale gh s —t)
fee f (a, b Bi)
ny (6.74)
= 1see tee ( 1 )Fab
S'(N,-
witha) 4 aE

where

Pre. Ee po ih( Po if
ln|ao ik: Gand 08 + ai \2B /No (6.75)
: 5 Ns —1 Po
se E i(Ns SO aves, (-2%,) |

The upper bound on unconditional average bit error probability is now obtained by averag-
ing (6.74) over the pdf in (6.1). The numerical results are illustrated in Figures 6.8 and 6.9
for values of VN, = 4, 6, and 8. We see that as N, gets large, the curves appear to approach
an asymptote. This asymptotic behavior is analytically evaluated as follows.
For N, large, the first term in (6.74) when integrated over the pdf in (6.1) approaches
428 Partially Coherent Communication with Waveforms Chap. 6

the irreducible error probability Py (£)|;,-- = Sere fo(~)d¢. Also, the dominant term in the
summation term of (6.74) corresponds to i = N, — 1, that is, 6 = Ns — 2. Thus, for large Ns,
the second term of (6.74) approaches f (ay—1, by—1) where

paaee
an,-1
ee
2No
[w, #2V(,—D} (6.76)
Since, from (6.76), ,/by,—1 >> //bn,—1 — /@n,—1, then using the asymptotic form of
(6.64) for a and b large (see [5], Appendix A), namely

1 =
f(a, b) = ~ erfe vb- Ja (6.77)
D 2

we obtain for large N,

1 E
f (an,-1, On,-1) = 5 erfc || (6.78)

independent of @. Finally then, for large N,, the asymptotic behavior of the average bit error
probability is approximately upper bounded by

Py tia /: fo(o)dd (6.79)

2 No m/2

namely, the sum of the performance for ideal coherent detection and the error floor.
Equation (6.79) is in very close agreement with the curves for N, = 8 in Figures 6.8
and 6.9.

Case 3. QPSK with Two-Symbol Observation and Detection (M = 4, N,; = 2)

In this case, E,/No = 2Ep/No. There are now a total of M2 — 1 = 15 possible error
sequences, each of length 2. Of these, only 8 produce distinct combinations of |5| and v.
These are tabulated below:

Err. Seq. Case be — $k bk—1 — Pe-1 3} v

5 Ps SoH 1 7, O, 30/2, eye ONEs 230072 0 0
5,6 Ds 0, 2/2 r/2,0 V2. a/4
728 3 0, 3/2 32/2, 0 V2 ana
9, 10 4 n/2, 1 m, 1/2 J/2 3n/4
[haz 5 m, 3/2 32/2, 1 J/2 —3n/4
13 6 m/2 m/2 2 m/2
14 a 37/2 32/2 2 =7/2
15 8 4 T 2 T

The corresponding values of a and b for each of the first five cases are given as follows:
Sec. 6.4 Performance of Partially Coherent Receivers/Multiple Symbol 429

A ee: Pg )co po_\”

te| No | ‘ (saa Pia "AGrarTy
= — {1/2

Po
+[2+ (sarin, )e# ||
b\ _£» Peo 1s tek
E |~ No 2 . Gane eer Gad) |
+|v4 —( ef -)cos + sind) ||

(6.80)

; [v2+F Cam) iia sna]

b Ep Po 3 Po
: No | ‘4lane! ae atts
= — oD) —

ae v2 i “ (=) (3cos@ + sin 6)|

{2} = 2 [a+(spite) cose? (ae)

+|va+ vaJ ea/No e ge - sina]
For cases 6 and 7, we have analogous to (6.71)

1 (INsEp
lim f(a, bp) =—Pine (cos@ + sind) (6.81)
b> +) Ns 2No

which for NV; = 2 and M = 4 becomes

1 2Ep .
ae 5 J4 b)=-=Ree (cm (cos @ + sin ») (6.82)
\ 0

Finally, for case 8, we use (6.71) to obtain

1 4Ep
slim ie b= one ( : oO (6.83)
\ No

Evaluating the Hamming distances for the 15 error sequences and substituting the above
results into (6.62) gives
 Partially Coherent Communication with Waveforms Chap. 6
430

1071

' ine)
=<oO

Ideal Coherent

a =, w Block-by-Block
Detection

Probability
Error
Bit
Average
Bound
Upper
on

Figure 6.10 Upper bound on average bit

10-4 error probability versus E,/No for optimum
5 6 7 8 9 10 11 12 13 14 15 — block-by-block detection receiver and
E/N. dB comparison with exact performance of ideal
b 0 coherent metric; M = 4, N, = 2, pg = 13 dB

|
PL(E |?) < y {6f (a1, b1) + 2f (az, br) + 2f (a3,b3) + 4f (a4, b4) + 4f (as, bs)}

1 of 2Ep , f 2Ep ;
+7 yertc No (cos @+ sing) } + erfc Ne (COS SSID) (684)

_( /4Ep
rete([oo

Figures 6.10 and 6.11 are the comparable plots to Figures 6.8 and 6.9 for the M = 4
(QPSK) case. The analytical exact result corresponding to the ideal coherent metric operat-
ing in a noisy carrier synchronization environment is now

| LE
PRES / aoe (ie(cos @ + sin ®)fo(g)dg

(6.85)

+fee jen
erfc eb (cos@ — sing) } fo(d)dd
No

6.4.3 Performance Sensitivity to Mismatch

Here we investigate the sensitivity of the average bit error probability (in terms of its upper
bound) of Fig. 6.7 to a mismatch between the true loop SNR, pg, and the estimate of it,
 Problems 431

107!

‘S ro

—sS; x)

Ideal Coherent

axae aN

Block-by-Block
Detection
' oO
Bound
Upper
Probability
Error
Bit
Average
on oO
—_

Figure 6.11 Upper bound on average bit

error probability versus E,/No for optimum
5 6 7 8 9 10 11 12 13 14 15 — block-by-block detection receiver and
E/N. dB comparison with exact performance of ideal
by 20 coherent metric; M = 4, N, = 2, pg = 16 dB

/¢, supplied to this implementation. In particular, we shall evaluate for the special cases of
Section 6.4.2 the upper bound
8

P(E) < ‘| Pou (Ed, bo) fo(p)do (6.86)

aes

where P,, (E |, Ay) is given by the upper bound in (6.73) or (6.74) with pg replaced by
be = Po [1+ (bo — p9)/P¢| & eg [1 + €9]. Figures 6.12 and 6.13 are illustrations of (6.86)
for M = 2, pg = 10 dB, and N; = 2 and 8, respectively, with fractional mismatch eg as a
parameter. We observe that even with mismatches as much as 50% (€g = £0.5), there is
negligible effect on the error probability performance. Thus, we conclude that the block-by-
block detection receiver is quite insensitive to mismatch in the parameter /g.

PROBLEMS
6.1 The random phase inserted into the transmitted signal s;(t) = aj(t) cos(wet + 6j(t) + @) is
governed by the probability density function

]
f@)= 4 |sin gl, lp| <7

for all i=0,1...,M — 1. Determine the optimum receiver when white Gaussian noise is
present.
432 Partially Coherent Communication with Waveforms Chap. 6

By = 4 0B, €g = -0.75

Ay = 10 dB, eg = 0

Coherent

By = 16 OB, eg = 0.75

Probability
Error
Bit
Average
Bound
Upper
on

Figure 6.12 Upper bound on average

bit error probability versus E,/No for
optimum block-by-block detection receiver
in the presence of loop SNR mismatch;
M—2e Ne 2 p¢ — 10GB

6.2 Consider the following binary hypothesis testing problem:

lee reyen@) Osrsi

Ai: r(t)=v2Ea(t)cos(@-t + 6(t) +6) +n(t) O<t<T

where @¢ is a random variable with the Tikhonov density given by (6.1) and n(t) is AWGN with
PSD No/2.
(a) Derive the decision rule that minimizes the probability of error.
(b) Express the probability of false alarm (choosing H; when Ap is true) in terms of the
Marcum Q-function.
(c) Express the probability of detection as an integral of the Marcum Q-function.
6.3 Consider the M-PSK communication system and assume that

Ay: r(t) = V2P cos (wet + i= +) +n(t), OSti

fori =0,1,..., M — 1 where @ is a random variable with the Tikhonov density given by (6.1)
and n(t) is AWGN with PSD No/2.
(a) Show that the optimum receiver is independent of pg.
(b) What is the minimum number of correlators required in implementing the optimum re-
ceiver?
(c) Express the probability of error as a multidimensional integral of the multidimensional
Gaussian density.
6.4 (a) Derive the optimum receiver in an AWGN channel (No one-sided PSD) for M-DPSK
Problems 433

Ay = 10 dB, eg =0

By = 16 cB, Ey = 0.75

By = 4 dB, €g = -0.75

Coherent

Bound
Upper
Average
Probability
Error
Bit
on

Figure 6.13 Upper bound on average

bit error probability versus E,/No for
6 7 8 9 10 141 #12 #143 14 15 — optimum block-by-block detection receiver
E/N. dB in the presence of loop SNR mismatch;
a0 M=2,N;
='8; pg = 10 dB

signaling if the phase uncertainty is given by the Tikhonov density. Assume signal models
of the form

sj-1(t) = V2P sinwet +09" +6), O<t<T

sj(t)=V2P
sin(wet +0! +9), T<t<2T
where 0“) — @9-)) = 277i /M fori =0,1,...,M—1.
(b) If pg — 0, what is the optimum receiver?
(c) If pg — 00, what is the optimum receiver?
6.5 (a) Derive an expression for the bit error probability when M = 2 for the receiver in Prob-
lem 6.4(a).
(b) What does the expression in (a) reduce to when pg = 0? Discuss your answer.
(c) What does the expression in (a) reduce to when pg = 00? Discuss your answer.
6.6 A binary source produces equilikely symbols mo and mj. The transmitter emits so(t) or 5) (t)
defined by

SH=H=VZEcosaoct. OUSts Tf

s(t) =—vV2Pcosw,t, O<t<T

and the channel randomly phase shifts so(t) or s;(t) through @ radians with probability density
function

rd) = 5[5(6- 5)+8(6+5)]

434 Partially Coherent Communication with Waveforms Chap. 6

and then adds white Gaussian noise (No one-sided PSD) to the resulting signal to produce the
received signal

r(t)=s(t,@)+n(t), O<t<T

(a) Derive the optimum decision rule.

(b) Illustrate the optimum receiver.
(c) Derive the error probability performance.
6.7 Repeat Problem 6.6 if FSK signals are used

so(t) = V2P cos wot

s(t) =V2P cosa jt

where |fo — fi] =1/T Hz.

6.8 Consider the Rician diversity channel described in Problem 5.19. Assume that the phase of the
specular component is not known exactly but is distributed according to

fa(8) = “PCSES”
7 19(p5)
alc
and that the 6;’s are independent from each other and all other random variables in the model.
(a) Derive the optimum receiver when L = 2.
(b) Generalize your result to an arbitrary L.
(c) Simplify the optimum receiver if ps = 0.
(d) Simplify the optimum receiver if ps = oo.
6.9 Verify for binary orthogonal signals that the limit of (6.22) as pg — oo is indeed given by
(6.24).
6.10 Verify that the whitening transformation as given by (6.31) produces Kx, = TZ as claimed in
(6.33).
6.11 Consider the received signal of (6.52) corresponding to the transmission of M-PSK over an
AWGN channel with unknown phase @. Assuming binary modulation (M = 2), then passing
r(t) through an I&D filter produces at time t = (k + 1)T), the received sample

ee
pe 2h e
sine A
Th
where 6, denotes the transmitted information phase which takes on values of 0 or z and nx is
a sample of a zero-mean complex Gaussian noise process with variance 0,2 = No/ Tp. Consider
a received sequence r = (Feehy 7x-2,-.-,F%k—n,) Of length N, and assume that the unknown
channel phase @ is constant (in time) over the length of this sequence.
(a) Write an expression, analogous to (6.54), for the conditional pdf of F given the information
phase sequence @ = (0x—1, 9-2, ..., ®%—N,) and channel phase @. Show that this pdf can
be expressed in the form
Ny
fro © 10,0) = Fexp fRe lieuer ee)
i=]

where F is a constant independent of the data and

ee VJ 2E,/Tp 2) V2E,Tp

Oo; N0

(b) Show that the average of the conditional pdf feo F|8,0) over the information phase
Problems 435

sequence is given by

Ns
Fro © 19) = Eg {fro F 10, Oy) =F | [cosh (a Re {7-1e 7°})
t=!

For a given observation (received sequence), we would like to determine the maximum-
likelihood estimate, Our. of the channel phase 6 and then use this as a demodulation
reference phase resulting in a pseudocoherent receiver (a form of partial coherence).
(c) Show that 6u L defined by

Ou = es ~!fag (E10) = max ~' In fro (F 10)

is the solution to
Ns
)- tanh (a Re {7-1e7/°} Re {jr ie /?}) =0
i=1

(d) For small SNR, we can approximate tanh x by x. Making this approximation in the result
of part (c), show that the maximum-likelihood phase estimate is given by
1/2

(e) Sketch in block diagram form the complex representation of a pseudo-coherent receiver
that implements the above phase estimation and performs detection of 6; using Out (actu-
ally e/ but) as a demodulation reference.
6.12 Redo parts (b)-(e) of Problem 6.11 for the case of arbitrary M = 2”. In particular:
(a) Show that the average of the conditional pdf ffi9 (f|@,0) over the information phase
sequence is now given by

N, 3-1
fr 10) = Eg {fe @18,6)} = FT | Y>cosh (aRe fixie 104271} )
i=! /=0
where F’ is a constant independent of the data and a is as previously defined.
(b) Show that bur is the solution to

Me
M

O+2"/M) |
_je
Ds sinh (a Re {7_je-/0+27"/)}) Re {j7,J
Ns

4
DY cosh (a Re {7—j;e7/@+27!/M)})
i=1 2

=O

(c) For small SNR, we can make a first-order approximation of sinh x by x and cosh x by 1.
Making these approximations in the result of part (b), show that it reduces to

IS ! Ny i
% Re tfB pede Re {ifBs et (Oui+2n1/M) =
1=0 i=1

which does not yield a unique solution for 6u L since the left-hand side of this equation is
identically equal to zero for any @y, and M > 2 (show this).
436 Partially Coherent Communication with Waveforms Chap. 6

(d) To get an approximate solution for 6u_, we must take the next higher order terms in
the Taylor series expansions of sinh x and cosh x, namely sinh x = x + x>/6 and cosh
x = 1+x?/2. Using these approximations in the result of part (b) and recognizing that
(1 + x2/2)~! = (1 — x?/2) for small x, show that, for QPSK (M = 4), the maximum-
likelihood phase estimate is given by
1/4
Ny /
a mm Vi
clout — i=l
Ns
<4
Thi
ji=1

which for the case of arbitrary M = 2” generalizes to

1/M
Neat /
. ye,Peei
elOmL — i=l
N65,
om Tey
i=]

6.13 Reconsider Problem 6.11 where now the unknown channel phase 9(t) varies with time in
accordance with a frequency offset, Af, due to Doppler. In particular, 6(t) at time t = kT),
is related to 6(t) at time t = (k —i)Tp by 0(kT,) = 6 ((k —i)T,) + 207i Af Tp. As before, it is
desired to find an ML estimate of 0(t) now, however, in the presence of frequency offset.
(a) Show that the conditional pdf of the received vector r = (Fe-1, tee 7k—N,) given
the information phase sequence 6 = (6-1, Cpa e wr Glee N,)s channel phase 6(f) at time
t = kT), and frequency offset Af can be expressed in the form

Ns
fioar (£10, 6 (kT) ,Af) = F exp feRe ogame mle
=

where F is a constant independent of the data.

(b) Show that the average of the conditional pdf fyj9 af (F 0,0 (kT,), Af ) over the informa-
tion phase sequence is given by

fioaf (FIO KTs), AF) = Eg {feoar (F 10, 0 (KT),

Af )}
Ns

| [oosh(aRe|
=F | | cosh(aRe |Ry_je
1° 4™
!)
where

Reyeipge
(c) Assume that the frequency offset is completely unknown and as such has a uniform pdf
between — fax and fmax Where fmax corresponds to the maximum expected Doppler fre-
quency shift. Compute the conditional pdf fi9 (if|6 (k7;)) by averaging the result of part
(b) over this uniform distribution.
(d) Show that the ML estimate, 6uL, of the channel phase 6(t) at time t = kT} as defined by

9A (kTs)a1
Seneaat
= max -le
~'feo (F(z 19 (KT)) == max -1 "In fay (F10
=
(KTS)
 References 437

is the solution to
a ; io ‘
Smax Ns

/ (aRe ey
B sinh ( TAO (Re Wee ol)
Simax i=

x ilcosh (wRe {Ri me ape 0

sy

(e) For small SNR, we can approximate sinh x by x and cosh x by unity. Making these approx-
imations in the result of part (d), show that the ML phase estimate is given by
1/2
, Yi
ie_;, Sine (4 fmaxi Tp)
el? &To) nx — | t=!

Fp_; sinc (4 fmaxi Tb)

REFERENCES
1. Lindsey, W. C., Synchronization Systems in Communication and Control, Englewood Cliffs, N.J.:
Prentice-Hall, 1972.
2. Lindsey, W. C. and M. K. Simon, Telecommunication Systems Engineering, Englewood Cliffs,
N.J.: Prentice-Hall, 1973. Reprinted by Dover Press, New York, NY, 1991.
3. Viterbi, A., “Optimum Detection and Signal Selection for Partially Coherent Binary Communi-
cation,’ JEEE Transaction on Information Theory, vol. IT-11, pp. 239-46, April 1965.
4. Simon, M. K. and D. Divsalar, “Multiple Symbol Partially Coherent Detection of MPSK,” sub-
mitted to the JEEE Transactions on Communications.
5. Schwartz, M., W. Bennett, and S. Stein, Communication Systems and Techniques, New York:
McGraw-Hill, 1966.
 Chapter 7
Differentially Coherent
Communication
with Waveforms

In Chapter 4, we discussed coherent communication systems wherein the receiver provides

a means (e.g., a carrier synchronization loop) for establishing a carrier demodulation ref-
erence signal that is synchronized in phase and frequency to the received signal. The im-
plementation and operation of a coherent receiver is, in general, a complex process involv-
ing many ancilliary functions associated with the carrier synchronization loop, for exam-
ple, acquisition, tracking, lock detection, false lock prevention, etc. In many applications,
simplicity and robustness of implementation take precedence over achieving the best possi-
ble system performance. In these cases, differential detection, which obviates the need for
implementing the carrier synchronization function, is an attractive alternative to coherent
detection. Aside from implementation considerations, it is also possible that the transmis-
sion environment may be sufficiently degraded, for example, a multipath fading channel,
that acquiring and tracking a coherent demodulation reference signal are difficult if not im-
possible. Here again, differential detection is a possible solution that still provides better
performance than noncoherent detection. While, in principle, differential detection can be
used with many different modulation schemes, the most widespread applications pertain to
constant envelope modulations. We will focus in this chapter on the most common of these,
namely, multiple phase-shift-keying (M-PSK). In Chapter 10, we shall discuss the appli-
cation of differential detection to a nonconstant envelope modulation, namely, quadrature
amplitude modulation (QAM).
Over the years, differential detection of M-PSK traditionally has been accomplished
by comparing the received phase in a given symbol interval (of duration T; seconds) with
that in the previous symbol interval and making a multilevel decision on the difference be-
tween these two phases. Implicit in this process is the assumption that the phase introduced
by the channel is constant over these same two symbol intervals and thus cancels when the
above difference is taken. Since, in the absence of noise, the decision is equivalently being
made on the difference between two adjacent transmitted phases, then in addition a suitable
coding must be applied at the transmitter to allow this phase difference to represent the input
information phase to be communicated. The encoding of the input information phase into
the difference between two adjacent transmitted phases was previously referred to as differ-
ential encoding (see Section 4.1.6 of Chapter 4). There, differential encoding was included
 Sec. 7.1 Classical M-DPSK 439

in the system solely as a means of resolving the phase ambiguity associated with the demod-
ulation reference signal supplied by the carrier synchronization loop. Stated another way,
differential encoding is not an essential element of the coherent detection process provided
that another means for resolving the phase ambiguity of the carrier demodulation reference
signal is employed. On the other hand, for differential detection, the inclusion of differential
encoding at the transmitter is absolutely essential! This will become clear shortly.
Perhaps the earliest known (to the authors) citation of the above ideas, herein re-
ferred to as classical multiple differential phase-shift-keying (M-DPSK), is due to Fleck and
Trabka and appears in a 1961 document edited by Lawton [1]. The receiver structures pre-
sented in [1], which were based on observation of the received signal plus noise over an
interval of two-symbol duration, were obtained without regard to their sense of optimality.
We begin our discussion by showing that classical M-DPSK is indeed optimum (based on
maximum-likelihood (ML) considerations) for a two-symbol observation interval. Later in
the chapter, we shall show that if the channel phase characteristic can be assumed constant
over N,7T; seconds, N, > 2, then by extending the observation interval to N,7; seconds and
applying an appropriate maximum-likelihood detection rule, one can further improve the
performance of the system. In fact, by making N, larger and larger, one can approach the
performance of a coherent M-PSK system with differential encoding (see Section 4.1.6).
More about this later on.

7.1 CLASSICAL M-DPSK

Consider the desire to communicate in the i" transmission interval iT, < t < (i + 1)T, an
information phase, A6;, which ranges over the set of allowable values 8, = 27m/M;m =
0,1,..., M — 1. In accordance with the above discussion, the first step is to differentially
encode A6; (see Figure 4.29a) into the phase 6; according to the relation 6; = A6; + 6;-1.
The phase 6; is then modulated onto the carrier and transmitted over the channel in the
form of an M-PSK signal, namely, s;(t) = J/2P cos (w-t + 9;). In addition to additive white
Gaussian noise, n(t), of single-sided power spectral density No Watts/Hz, the channel in-
troduces an unknown phase @ that is assumed to be constant for a duration of at least 27;
seconds. Thus, the received signal in the i" transmission interval has the form

r(t)=s;(t; 0) +n(t) =V2P cos (@ct + 6; + 8) + n(t) (C/alt)

For classical symbol-by-symbol differential detection of M-PSK, the optimum re-

ceiver takes the form of Fig. 7.1. The derivation of the optimality of this receiver follows
along the lines of that presented in Chapters 4, 5, and 6 for coherent, noncoherent, and par-
tially coherent detectors. The details of the derivation for the differentially coherent receiver
are presented in Appendix 7A. At first glance, Fig. 7.1 resembles Fig. 4.32b. However, there
are some important differences. In both figures, an estimate, n;, is made of the phase! asso-
ciated with the i transmission interval. In Figure 4.32b, a hard decision 6; is made based
on this phase estimate and then compared (in the differential decoder) with the hard deci-
sion 6;_1 based on the previous phase estimate, ;_1, to arrive at the final decision on the
information phase A6;. In Fig. 7.1, the difference of the two phase estimates 7; and n;—1 1s
first taken and then a hard decision is made based on this difference (modulo 277) to arrive

1. It is understood that the tan! operation used to define the phase nj; in Fig. 7.1 is meant in the 4-quadrant sense
and not in the principal value sense.
440
O[Npour
uz esooyD
ISoTTeUIS
ur
(4)4 ainjosqy
oneA
[apoio

[eo
IO}eT[DSE

eansiy
TL woaundg J9AIVO0I
JO} [edtsse[d ASdd-W
Sec. 7.1 Classical M-DPSK 441

at the final decision on the information phase A6;. As such, the differential detector intrin-
sically includes the differential decoding operation and thus no separate differential encoder
is required.
The receiver of Fig. 7.1 can also be redrawn in a different form as in Fig. 7.2a, which
takes on the simpler form of Fig. 7.25 when M = 2. This receiver was suggested in [4] as
well as several other places and its equivalence with Fig. 7.1 was demonstrated in [5] (see
Problem 7.1).

7.1.1 Symbol Error Probability Performance

The decision algorithm used in the receiver in Fig. 7.1 is quite analogous to that in Fig. 4.32b
in that the decision regions are still defined as wedges uniformly spaced around a circle (see
Fig. 4.22). Here, however, the phase associated with the signal that defines each of these
regions is the value of the phase difference 6; + 6 — (6;-; + 9) = 0; — 6;-1 = AQ; and the
decision variable that falls into these regions is the phase difference yy; = nj — ni_1 modulo
27. In particular, if 0; — 0;_; = AO; = Bx 1s the value of the phase communicated in the ith
transmission interval, then a correct decision is made when y; falls in the Ee wedge defined
by the angular interval (8, — 27/M, By +72/M). In mathematical terms, the conditional
probability of a correct symbol decision given that 6; — 6; = A0; = Bx is

ca
TACO)Wea 0 io0 eal a) Cea 2) |e
Cai Bee, M
Bx G2)

apr(ii—alsi}=] fxooax
a/M

where x = Wj — Bx. Thus, analytical evaluation of symbol error probability reduces to find-
ing the pdf of X.
In [1], the pdf fx (x) was determined by first expressing it in the form of a convolu-
tion integral based on the pdf’s of n; and nj;—) (since n; and n;— are statistically indepen-
dent) and then evaluating this integral using the method of characteristic functions with the
result

5 or? (sin a) [!+ eC + cosx sin a)|

fx) = x exp {—FE(1 — cos x sina)}dar xi<a (7.3)

; otherwise

Since this pdf is an even function of x and independent of the particular phase 6, transmit-
ted in the i" interval, then P,(C|k) is independent of k and hence the average symbol error
probability is given by
442
- YLT
UIS W I-W""“TT0=7'

{74}

YLT
eye) W I-W“ZTT0=7'

[eo]
JO}e]]DSO

aaNS1yBZ" Uy aJeusoye
WO}JO emBL{
I'L
 Teo]
10}eT[DSO

BANBIYqZ*Z,
= UY JUgTeAINDaJoATa001
0} OINB1JZL IO} W = Z

443
444 Differentially Coherent Communication with Waveforms Chap. 7

1 M ma/M ;

(E) —>) P,( (E|k)

PE)= Elk) =1- [a v
ut m/M
fu (Ww) dy — i fu (w)dw (7.4)
=

Safi fu (Ww) dy
m/M

where we have arbitrarily assumed 6, = 0 and hence x = —. Substituting (7.3) into (7.4),
we observe that P,(E£) is in the form of a double integral, which must be evaluated numeri-
cally.
More recently, the evaluation of (7.2) has been carried out as a special case of the
more generic problem of finding the probability distribution of the phase angle between two
vectors, each perturbed by additive Gaussian noise. The solution to this problem is discussed
in great detail in [4] and summarized here in Appendix 7B. For the M-DPSK receiver of Fig.
Hel @asenl applies” with yy; = 2/M and w2 = z. Thus, from (7B.3) and (7B.4)

if fu ()) dw = F(x) — F(t/M)

ma/M

sin
:
Z 7/2 exp |- 7g
45[1 — cos 7,
It
cost]
=0= dt} (7.5)
4x J_n/2 1 — cos 4, cost

2
sin = (eo exp -4 [1 = COS 7,cos ]|
ce on a eR a ee ca
4x J_x/2 I — cos ¥, cost

Since A*/2c7 represents the symbol SNR, then

Laem cE my 2 ae
202 No 62 No fEo)

Substituting (7.6) into (7.5), the symbol error probability of (7.4) becomes

Es
sin Z #/2 eXp {—# [1 —cos %I cos]
PCE) = Sit f i es eee es (oT!)
1 —cos 47 cost
Uv
21 —7 {2

which is in the form of a single integral of simple functions.

Many approximate expressions have been found for the symbol error probability per-
formance of M-DPSK. One form due to Fleck and Trabka that originally appeared in [1] is
accurate for large E; /No and is given by

A exp (— A?)
PAB) = ee ee (7.8)
ae lee a1 EH
2. Since, as mentioned above, P; (C|k) is independent of k, then for convenience we again take k=0 (Bk =
Bo = 0), which implies that the transmitted signal is the same in both the (i — 1)*' and i" intervals.
Sec. 7.1 Classical M-DPSK 445

Table 7.1 Ratio of Symbol Error Probabilities Corresponding to Eq. (7.13)

(Reprinted from [4])

M Q
3 0.9186
4 0.9768
8 0.9987
16 0.9999

where

A == J2E,/N
5 0 sin
1 (——)
=e
2M (7.9)
é

Another approximation due to Arthurs and Dym [5] is

P,(E) & erfc (j= sin Fa) (7.10)

This expression is simpler to evaluate but less accurate than (7.8). Bussgang and Leiter [6]
derived an upper bound on P;(E£) which for large EF /No is itself a good approximation. The
result is

P,(E)
& erfe — sin a) (7.11)
ell JI+2E;/No M
When both E,;/No and M are large, (7.11) resembles (7.10) (since for small x, sinx = x);
however, for smaller values of E/No, (7.11) is the more accurate of the two.
Although all three of the above approximations are good for large E/No, none are
asymptotic in the sense that the error in using them becomes vanishingly small as E;/No
becomes infinitely large. Pawula et al. [4] obtained a true asymptotic (for large E;/No)
formula for symbol error probability. Their result is

M E
P;(E) Ile 1+ cos7/M erfc {| _/— [1— cos =| wo M>3 (7.12)
2cosm/M No M
The Fleck and Trabka approximation, that is, (7.8), is closest in form to (7.12). To compare
the two, we form the ratio Q of (7.8) to (7.12) by first neglecting the 1/8 factor in the
denominator of (7.8) and then employing the asymptotic expansion of the erfc function in
both (7.8) and (7.12). The result is

loos & 1 gin?

7 P;(E)\|(7.8) a cos # [1+ 3 sin a | (7.13)
~ P(EYI7.12) COS
347
which is independent of E;/No. Table 7.1 tabulates the ratio Q for values of M = 3,4, 8,
and 16. Thus, although Fleck and Trabka’s approximation is not, in the strict sense, an
asymptotic formula, it nevertheless differs in its leading term from the asymptotic formula
by less than 9% in the worst case when M = 3 and by less than 3% in the more typical
situation when M is a power of 2 (M > 2).
446 Differentially Coherent Communication with Waveforms Chap. 7

The special case of binary (M = 2) DPSK, often simply called DPSK, is a case where
the error probability performance can be found in closed form (as opposed to an integration).
In particular, letting M = 2 in (7.7) gives the simple result

PC — EbBC) ;exp (-3) (7.14)

This result was independently obtained by Lawton [7] and Cahn [8]. Figure 7.3 is an il-
lustration of the exact expression for symbol error probability of M-DPSK as obtained
from (7.7) or (7.4) versus E,/No in dB for M = 2, 4, 8, 16, 32, and 64. Superimposed on
these curves are the analogous results for the symbol error probability performance of co-
herent M-PSK as obtained from (4.131).? Asymptotically for large M and large E,/No,
we see that differential detection of M-PSK requires 3 dB more E,/No than coherent de-
tection of M-PSK. This result can be obtained quite easily by comparing the effective
SNR’s for these two detection schemes as determined from the arguments of the erfc func-
tions in (4.139) and (7.13). In particular, for coherent detection the effective symbol SNR
is given by (Es/No)er¢ = (Es/No) (sin? a /M) = (Es;/No) (1 /M)* whereas for differen-
tial detection (Es/No)ege = (Es/No) (1 — cos /M) & (Es/No) [1 — (1 — (t/M)? /2)] =
(Es/2No) («/M)°.
7.1.2 Bit Error Probability Performance

Assuming a Gray code bit to symbol mapping as was done in Chapter 4, the bit error prob-
ability performance of M-DPSK can be related to its symbol error probability performance
evaluated in the previous section by the simple approximate relation given in (4.140). It is
also possible to obtain an exact expression for the bit error probability performance using
the method of Lee [11] combined with the results of Appendix 7B. For example, for 8-
DPSK, Lee obtained the result

2 137 TU
Po(E) = 5 lF(=) bs 3 (=)| (7.15)

where F(y) is given by (7B.4) with U = E,;/No. This result was later generalized by Korn
[12] to other values of M. In particular

P,(E) =F (=) Ope (=) - M=4 (7.16a)

Pes |ad(<2)
r(e)=5[F el Vly F(=)
peas Vek F(3)-F()]:
of eV ie M=16 (7.16b)

3. Using the results of Appendix 7B in the degenerate case where one of the two vectors is noise free, Pawula
et al. [4] have been able to arrive at a simpler expression for the symbol error probability performance of coherent
M-PSK, namely

1 m/2—n/M E 1
P(E) = — | ex |— sin? &sec? ao
: X J_x/2 : No M

which is essentially the same result obtained by Weinstein [9].

Sec. 7.1 Classical M-DPSK 447

1071

10-2
me
WW
qa»

10°3

10-4
—— COHERENT
DETECTION
-—- M-DPSK
10-5 Figure 7.3 Symbol error probability
performance of M-PSK and M-DPSK
b’ 0’ (Reprinted from [10])

2 297 231 19x 171

moey=
5 (ar)
+? Gr)" Gr)" (Gr)
# B A (7.16c)
1 be TU 4
F | —)-F(|—)-F({—]-F(=)); =
re (=) ($) ($) (5): yorre
Note that the last term in each expression in (7.15) and (7.16) is
2 TU P,(E)
P,(E)|; =—- F — i a
wel log, M lea) log, M ( )

which represents the Gray code approximation of (4.135). When compared with coherent
detection, some typical values of E,/No degradation at P, = 10-5 are 0.75 dB for M =2,
2.2 dB for M = 4, and 2.5 dB for M = 8.

7.1.3 Effect of Noise Correlation and Power Imbalance

Suppose now that because of channel bandlimitation, the amplitude of the received sinu-
soid is time-varying and the power spectral density of the received noise is not flat for all
frequencies. In particular, the received signal of (7.1) now takes the form

r(t) = A(t) cos (wet + 6; + 8) +n'(t) (7.18)

where n’(t) is additive Gaussian noise with power spectral density [S,(f — fo)+
Sa(f + fo)]/2 and S,(f) = F {Rn(t)} is a symmetric function. After demodulation with
inphase and quadrature reference signals as in Fig. 7.1 and passing through the integrate-
and-dump (I&D) circuits, we get the lowpass signals
448 Differentially Coherent Communication with Waveforms Chap. 7

Ve = Ai cos; +0) +N. ; Vs, = Ai sin@; + 8) + Ns; (7.19)

where

] (i+1)Ts
Aj = — A(t)dt

(7.20)
(407, G+T,
Ni j= / n'(t) (v2cos wt) dt; Nij= / n'(t) (v2sin wt) dt
: ie F Hie

If, as in (7B.1), we relate the complex form of these I&D outputs to the geometry in
Fig. 7B.1, then the association is now

A, =Aj-1; A2=Aji; $1=6)-1+9; G2=6 +80

m=Nji-3 WN W=N, BY y2= Ny;

1
maf f- Rn (u — v)dudv = 07;
¢ GE) TesGaN) E CT24 5
(i+1)T; (i+1)Ts
of = =e ie R,(u — v)dudv =o"

iT i+1)Ts
G_DT. ho
iT;
Ts R(u — v)dudv
; =0
o2 (
For binary DPSK, Pawula and Roberts [13] have been able to evaluate the bit error
probability performance in closed form (in terms of tabulated functions). For the binary
case, the error probability can be written in the form

—n/2
P, (BE) = Pri Ag=0} ([ fu wav)
32/2

m/2
+ Pr{Ad=n) ( fu Wav) (7.22)
—n/2

—n/2 m/2
a (/—3n /2 e wav) eR (=a) fe way)
where Po and Pa are the a priori probabilitiest of the two possible information phases trans-
mitted in thei" interval. Evaluation of the integrals in (7.22) follows the results correspond-
ing to Case 2 in Appendix 7B. For the case where the limits of (7B.3) are separated by z
radians, as is true in (7.22), Pawula and Roberts [13] have been able to evaluate the integrals
in closed form. In particular

4. In their analysis, Pawula and Roberts [13] allow different a priori probabilities for the information phases 0
and zr.
Secaval Classical M-DPSK 449

Wo
Bl) ee wl =i} fu (Wd
Wo-m

5 1red le Ble (E.4)|

ane sin yo e “Io(B) .
2/1=r2 cos?Yo’ (7.23)
AD—r<W<AP

rsin Yo e€“Io(B) .
+ ’
24/1—r2 cos? Wo
elsewhere
where Je(k, x) is the Rice Je-function [14] defined by

Te(k, x) af exp (—t)Io(kt)dt (7.24)

0

and tabulated in [14, 15], and

pias U —rWcos wWocos (A® — Wo)

ee 1 — r2cos?Wo
025)
> __ W?sin* (A — yo)
Sa.
1 — r2cos? Wo
Since in our application A® = A@6j, and noting that for AD = 0, Wo = —2/2 while for
A® =7, Wo = 1/2, then combining (7.22) and (7.23) gives the desired result

] Ww V r _u
PoE) = 5 {1
Sate (Gu) ]+5 0% - Ave Io(V) (7.26)

where U, V, and W are defined in (7B.2) using (7.21). From (7.26), we observe that the error
probability does not depend on the noise correlation r so long as the a priori probabilities
Po and P, are equal, whether or not there is a power imbalance. When there is no power
imbalance, V = 0, and (7.26) reduces to

|
P(E) = ; (1+ (Pr — Po) rjerZ (7.27)

which was previously obtained by Lee and Miller [16]. When in addition Po and P, are
equal, then (7.27) reduces to (7.13) as it should.

7.1.4 Effect of Imperfect Timing

The effect of bit timing error on the bit error probability of an otherwise ideal DPSK receiver
was first considered in [17] and later in [18]. We present only the results from [18] since
they are based on the approach taken in [4] and are thus in simpler form. In particular, if we
denote the bit sync error, as in Section 4.3.3, by ¢ = AT» where A represents the normalized
(with respect to the bit time) error, then the block diagram of Fig. 7.1 applies if we replace
450 Differentially Coherent Communication with Waveforms Chap. 7

the lower and upper limits of the I&D’s by (i — 4) Tp and (i + 1 — A)T), respectively. Having
done this, the outputs of these circuits become

Vei = PT;
V [(1 — A) cos (6; + 8) + Acos (6-1 + 8)] + Ne,i
(7.28)
V,,i = WPT, [(1 — A) sin (6; + 8) + Asin 6;-1 + )] + Ns,i

where

(i+1-A)Tp
Noi & } n(t) [v2cos wet dt,
(i—A)Tp
(7.29)
(i+1-A)Tp
Nowe / n(t) |-v2 sin wet dt
(i—A)Tp

are independent zero-mean Gaussian random variables with variance o*= NoTp/2. The
equivalent phase nj; for the i" transmission interval is found from the arctangent of the ratio
of V, ;to V..; and using (7.28) can be expressed in the form

nj ctan~! |YPET s sinGe +8) +Noi on

J/PT,T; cos (§& +0) + Neji :

where

ie (1 — A) sin; + Asin 6;_|

ey (1 — A) cos 6; + A cos 6-1
(7.32)
lr; = Ja — A)? + 22 + 2A(1 — A) cos (6; — 6:1)

= V(1 — A)? + A2 + 2A(1 — A) cos AG;

A similar expression can be written for the equivalent phase, n;_1, in the (i — 1)‘ transmis-
sion interval. Once again the geometry of Fig. 7B.1 applies, where now

At=VPTli-1; A.=V PRE

gi=§&-1+0; ow =§& +06

X1=Nei-1; yi =Nsi-1; | X2= Nei; yo=Ns,i (7.32)

DAW. to)

r=p=0

Also, from (7B.2)

Sec. 7.1 Classical M-DPSK 451

v=3(H +H) -B[ ever)

2 \ 202 Oy od pa No 2

9! ?
ee aisAil) Fh|se)202 No 2
(7:33)
2\202

A® = (¢2 — $1) modulo 27 = (&; — &|_}) modulo 27

Substituting (7.31) into (7.33) and simplifying gives

E [a — 4)? +42 +.2(1 — A) (cos
y= < AG; + cos A6i-1)|
0

V= No [AC — A) (cos A6; — cos AGj-1)] (7.34)

aoe (1 — A)* sin AG; + A? sin AG;_1 + AC — A) sin (AG; + AG;-1)

(1 — A)? cos AG; + A? cos AGj_-1 + A(1 — A) [1 + cos (AG; + AGj_-1)]
modulo 27

For DPSK, the information phase, A6;, to be communicated in the ith transmission
interval takes on a value of 0 or z. For each of these values we must compute the conditional
bit error probability when A6;_; takes on these very same values (with equal probability).
To do this, we apply the results of Case 2 in Appendix 7B, in particular, the special case
of these results given in (7.23)—(7.25). To begin, we observe from (7.34) that for A6;_;
equal to either 0 or 7, AD = A@;. Furthermore, for A® = 0, we have wo = —2/2 while
for A® = 2, we have wy = 7/2. Thus, the above conditional bit error probabilities are
evaluated from (7.23) as

P, (E|0,4)=Pr{-—32 <y<-~}=-|1-
2
1
1 ‘
Bre
ao
(2.
ao
ao)
(7.35)
Pp (E| 2,/Ay=Pr{->
A) r a<
1 1
<tl=5|i=
eae
I |, Be, ol(
pee)
Pite
(Bx 2)
ee

where from (7.25)

wie A6;-1 =0
Ng lb 2a Aye, Abeay ser
i Ep 0; AG;-7 = 0
Pig Nea Sin A Gi) arr
(7.36)
Ep. {V5 20G= 2); -AG-1=0
EBaciingg: os [iden) AGj-1=7
En [200 = a)} AG) =0
Pe = 0; AGj;-1 =
The average error probability conditioned on a given timing offset A is then obtained by
452 Differentially Coherent Communication with Waveforms Chap. 7

averaging (7.35) over equiprobable events A; = 0 and A@;_; = 7, namely

l 1
Pp (El A) = i Py (E|9, A) a6,_;=0+ m1 Pp (E|0, A)|a6;_:=2
1 1
Be riPy (E|2, 2)|a6,_,=0 + riPy (E| 1, A) lao;_,=2

=;[1- (0 x) 737)

+5 ce ee)

+5 i fy - POO te ( 1-200 -2 F)

a ; Ae (0.(ha |

Using the identity [13, Eq. (A3-5)]

baled 1 ~Fre(2.0) = Oe?

|18 exp etlees—_ BeOpON H Taig,
2 a2 \a on 0 a — Bcosé

we can rewrite (7.37) as

1 Ee
=|- —— || D;
uk Boal el
1 1 pt exp {Fe[4a(1 — A) cos? Al
D=-—)1+(—2)) i!
4 TW JO 1 — 40(1 — A) cos? §
(7.39)
1p exp {Fe[4a(1 — a) sin? al
+= 2 Tw fJO Fay1 Ey
— 44(1 eae
— A) sin® 5

Ep
+ exp |
x [4A(1 — ai}

which clearly places in evidence the degradation in performance due to bit timing error
relative to the ideal (perfectly timed) performance of (7.14). Note that when A = 0, D
reduces to unity as expected. Assuming a Tikhonov pdf for the normalized timing error as
was done in Chapter 4 [see Eq. (4.324)], the average bit error probability performance of
DPSK in the presence of timing error is obtained by averaging (7.39) over this pdf and is
illustrated in Fig. 7.4 for various values of rms normalized timing error o}.

7.1.5 Double Error Rate Performance

In designing receivers that employ error detecting and error correction codes, the occur-
rence of multiple consecutive errors, that is, burst errors, is an important consideration. For
Sec. 7.1 Classical M-DPSK 453

yamaihigt ET Tela Gl tiara) 14 igure 7 4 Error probability performance

of DPSK in the presence of bit timing error
&/No, dB (Reprinted from [18]) © IEEE

uncoded coherent receivers that make bit by bit decisions, such as those discussed in Chap-
ter 4, the probability of a burst error is simply the product of the bit error probabilities over
the length of the burst, or equivalently, the average bit error probability raised to the power
of the burst length in bits. The reason for this is that in a coherent system, the bit decisions
are independent of one another. For uncoded DPSK, since in effect the phase of one signal
acts as the phase reference for demodulation of the next signal, the decision statistics are not
independent from bit to bit and hence the probability of two or more consecutive errors is
not the product of the individual bit error probabilities over the length of the error sequence.
The analysis of double error rates in DPSK systems was first considered by Salz
and Saltzberg [19]. Although their results were derived for a suboptimum receiver (to be’
discussed later in this chapter), the application of their approach to the optimum DPSK
receiver of Fig. 7.2b follows directly. In particular, let

Zi A Vej—1 Veg EVs 321V 5,5

denote the decision variable in the receiver of Fig. 7.2b where from (7.28) and (7.29) with
0

Veg = PT,
V 0086; + Nei, Vo,i = PTp
V sin; + Ny.i
Here again, N..; and N, ;are independent zero-mean Gaussian random variables with vari-
ance 0” = NoTp/2. The decision on the information phase, A6;, communicated in the me
transmission interval (or, equivalently, the decision on the binary bit a; = cos A@;) is then

a; = cos Ad; = sgn Z; (7.40)

The approach taken in [19] is to first determine the joint pdf of Z;-; and Z; condi-
tioned on V,,;—; and V, ;-; being fixed. This conditional joint pdf is Gaussian and, since for
454 Differentially Coherent Communication with Waveforms Chap. 7

binary modulation E {V,;-1} = E {Vs,i} =0, it is given by

1
f (Zi24: Zi\ Ve.i—1, V5) =
2x0?| (Ve,i-1) + (Vsi-1) |
2 y (741)
1 (Zi-1 = VP Tp 008 6;Ve,i-1) + (Zi — /PTp 0s6;-2Ve,i-1)
x exp 4 ———s>
20? (Vei-1)° fe (Vea)?

Averaging (7.41) over the joint pdf of V.,;-; and V;;-; which is also Gaussian,
namely

2
TAY iain) ee exp |x (V1 -— VPT»c0s6}-1) + (Ves) }(7.42)

we obtain the unconditional joint pdf f (Z;~;, Z;). This pdf is a function of the three
transmitted phases 6 = (6;—2, 6;—-1, 9;) which for binary modulation takes on eight possi-
ble values. Associated with each of these transmitted phase sequences is a double error
probability, Pp, (Fi-1, Fi); k= 1,2,...,8. For example, if @ = (0, 0, 0), or equivalently,
A6;-; =0, AG; = 0, then a double error (two consecutive bit errors) occurs when Z;_; < 0
and Z; < 0. This double error probability is given by

0 0
Pp (Ei-1, Ei)lo=c,0,0) & Poi (Ei-1, Ei) = f (Zi-1, Zi) dZj-1dZ; (7.43)
—oo J—00

which is an integral over the lower left quadrant of the Z;_,, Z; plane. Salz and Saltzberg
[19] show that, with appropriate changes of variables, (7.43) together with the other seven
integrals of this type can all be expressed in the form of two integrals each of which in-
tegrates over the upper right quadrant of the Z;_;, Z; plane. Since each phase sequence
6 occurs with probability one-eighth, then averaging P,, (E;-1, E;); k=1,2,...,8 over
this equiprobable distribution gives the desired result for average double error probability in
the form of a quadruple integral, namely

Pt Ee >» Pil
oyEe Bi) =e Fo/Ma(& aide: fae
, (7.44)
Ey.) Cia): a Ga) 2, .2
x exp |_5 [fc aa + x4 + yp + 2x1 | } dxidx2dz\dz2

Finally, it is also shown [19; Appendix I] that this quadruple integral can be simplified to a
single integral, which yields the desired result
Sec. 7.1 Classical M-DPSK 455

Pp EEE =e
RNG | 1—erf-|
Ep
./—cosé
;
4 0 No
2
+ |am cos getEe/Nayeos?6 11 _ ore (|Be cos @ (7.45)
No No

E
x +a (soo) |
\ No

Oberst and Schilling [20] have also considered the double error probability perfor-
mance of DPSK and arrived at (7.45) in a somewhat simpler manner. Although they too
considered the same suboptimum configuration as Salz and Saltzberg [19], their approach
and mathematics is nevertheless applicable to the optimum DPSK receiver of Fig. 7.2b.
Oberst and Schilling [20] base their approach on the observation that DPSK can be viewed
as the signal in the (i — 1)*' transmission interval (i — 1)T, < t < iT; acting as a noisy de-
modulation reference for the signal in the ith transmission interval iT, <t < (i + 1)Th. They
referred to these two signals, respectively, as the “reference” and the “data.” The clever trick
employed in [20] is to switch the names of the “data” and “reference” signals for the in-
tervals (i — 2)T, <t < @@ — 1)7T; and (i — 1)T, <t <i7}; that is, one should think of the
signal in (i — 2)T, < t < (i — 1)Tp as the “data” and the signal in (i — 1)T, < t <iTy) as the
“reference.” This allows one to view the two demodulations ordinarily occurring at t = iT)
and t = (i + 1)7) as equivalently having been performed independently (with regard to the
additive noise) at t = (i — 1)7, and t = (i + 1)T) by acommon reference signal. In mathe-
matical terms, the probability of error for the i" bit conditioned on the phase estimate nj;_1
associated with the previous bit is given by

1 E
Pp (Ej \ni-1)= ~yyerfc (Ji
No
cos n-1] (7.46)

where

es
i Mirae eet?

I Ep nj-1e —(E,/No)
sin? PV ni—1 hl ert Ep COS nj-]
a —,_2 |xNo
—— cos Ni-1 —No Ni ; (7.47)

= Snare
Based on the above discussion, the double error probability conditioned on 7;— is then

1 E =
Pp (Ei-1, Ei lni-1) = ;
2;
erfc (2
No
cos n-t)] (7.48)

Averaging (7.48) over the pdf in (7.47) produces, after some manipulation, the identical
result to (7.45).
456 Differentially Coherent Communication with Waveforms Chap. 7

FEE;

THEORETICAL
O SIMULATION

ste 6 8 1 2 3.4 5 Figure7.5 The ratio of double fo single

error probability for DPSK (Reprinted from
Fi [20] © IEEE)

The conditional error probability Py, (FE; |Ei—1) = Pp (Ei-1, E;) /Pp (E;—1) repre-
sents the ratio of double to single error probabilities, which, in view of (7.14), is given
by
Pp (Ej-1, Ei) _
P, (E; |Ei-1) = 2e%/Nop, (E;_, Ei) (7.49)
Py (Ei-1)
Figure 7.5 illustrates this ratio as a function of E,/No obtained theoretically from (7.49)
together with (7.45) as well as by simulation results [20]. The monotonically decreasing na-
ture of P, (E; |E;—1) with Ep/No indicates that the ratio of double errors to single errors
does not increase as SNR increases. As a final note, we also point out that using Oberst and
Schilling’s viewpoint, it is straightforward to show that nonadjacent decisions are mutually
exclusive and thus the conditional probability P, (£; |E;—,%) , k > 1 is equal to the uncondi-
tional bit error probability, namely, (7.14).

7.1.6 Error Probability Performance Comparison of DPSK and Noncoherent

FSK

In Chapter 5, we examined the bit error probability performance of noncoherent FSK

(NCFSK), in particular, that of the optimum receiver;Itwas found that whereas optimum
coherént detection of FSK (see Chapter 4) resulted in an error probability that varied as
a complementary error function of E,/2No, the comparable behavior for NCFSK was an
exponential function of E,/2No. If we make a similar comparison between optimum co-
herent PSK (see Chapter 4) and DPSK, we also note that the error probability performance
of the former varies as a complementary error function of E,/No, whereas that of the lat-
Sec. 7.1 Classical M-DPSK t 457

ter varies as an exponential function of E,/No. Aside from the 3 dB differences in SNR,
the performance similarity between optimum coherent detection of PSK and FSK and the
performance similarity between optimum differential detection of PSK and noncoherent de-
tection of FSK is more than accidental.
We saw in Chapter 4 that both optimum coherent detection of PSK and optimum co-
herent detection of FSK resulted in a correlation receiver with the same form of decision
variable, the difference between the two being the correlation coefficient, y, of the signals
[see (4.52)] which affects the mean of this variable. In the case of FSK, y = 0, whereas for
PSK, y = —1. Since the effective SNR in both cases is proportional to (1 — y), then the
above difference in the value of y accounts for the 3 dB difference in E,/No performance.
The explanation of the similarity in performance between optimum differential detection
of PSK and noncoherent detection of FSK is perhaps not as obvious since indeed the two
receiver structures appear to be quite different. In what follows, we shall expose the under-
lying thread that ties these two different modulation and detection forms together. The basis
of this thread is a special case of a unified analysis of certain coherent and noncoherent bi-
nary communication systems performed many years ago by Stein [21], which still stands as
a classic contribution in the field.
Consider the optimum DPSK receiver of Fig. 7.2b and the optimum NCFSK receiver
discussed in Chapter 5, redrawn here as Fig. 7.6 for the purpose of ease of reference and
similarity of notation. We are interested in comparing the error probability performances of
these receivers for both messages mg and m,, which are defined below for the two different
modulation/detection schemes.”

|6; — 6;-; = AG; =0 DPSK

message mo:
fi= f© NCFSK
(7.50)
| |.0; — 67-1 = AG; =a DPSK
message m : ae NCFSK

The rules for deciding between these two messages (see Figs. 7.2b and 7.6) are

mo

Z; = VeiVei-1 + VeiVet1 0 DPSK (7.51a)

m|\

mo
2 2 a\
Z} = [(Wer)? + (Veni) |= [(Veou)” + (Yori) ]= 0 = NCFSK (7.516)
mM

where

5. We use a parenthetical superscript notation here to characterize the particular message being referred to so as
to avoid confusion with the subscript notation used for denoting discrete time.
 458 Differentially Coherent Communication with Waveforms Chap. 7

v2 cos 27f ue

—¥2 sin2afr

fo <
r(t) Ja Sti f
<
x

Figure 7.6 Optimum noncoherent FSK receiver

Sec. 7.1 Classical M-DPSK 459

Voi = VPTp cos (6; + 0) + Nei, V;,i = VPTp sin (6; + 8) + Nj

2n (f — fi)

+ Nei; k=0,1 (7.52)

an (fF — fi)
Vises & = 0,1

For DPSK, V.,; and V;; are independent Gaussian random variables with variance o*=
NoTp/2, while for NCFSK, Ve1,i, Vc2,i, Vsi,i, and Vs2,; are independent Gaussian random
variables also with variance o? = NoT,/2. The means of these random variables when
messages mo and m, are transmitted are as follows:

V PT, cos @; k=
Eat 2 J mo orm
ci] eae Chae

V PT» sin@; C—O

JO Veeee pe 2 : mo orm
si] —JPT,sin@; 6=7 :
C53)
th a Wales mo
susp aaa ’

25,10; mo heel’:
E Waal = {yen con my” E (Vad = [Yen sine i
It is at this point that we introduce a clever algebraic trick suggested by Stein [21] that
allows (7.51a) and (7.51b) to be more readily identified as being equivalent. In particular,
we rewrite (7.51a) as

doVei-1 aPmame
Voi
oe ae a
Vs i-1 Ar Vs.i oi

at, Vo,i-1 — Voi : a Vs,i-1 ay Vs.i =

2 2 (7.54)

The random variables V.4,;, Vs+,i; Ve—,i, Vs—,i are all independent Gaussian with variance
a? /2 = NoT)/4 and means as follows:
 460 Differentially Coherent Communication with Waveforms Chap. 7

J PT; cos 0; 6; =0
E{Ve+,i} = |—J/PT,cos0; 6; =n J"
0; mM},

J PT, sin 6; 6; =0
ee —VPT, sind;
= | 6 =
0; mM),
(7.535)
0; mo
ELVem) =| VPnycos 6; =0 pe
—/PT;, cos 0; GP 6 Ee

0;
B{V,2c} =| VPipsine G=0|
—/PTy sin 0; Ch .

Comparing (7.53) and (7.55), we observe the similarity between the squares of V.+,; and
Vetis Vs+,i and Vs1,;, Ve—; and Vo2i, Vs—; and V;2,;. Finally then comparing Z; of (7.54)
with Z/ of (7.51b), it is clear that they have the same probability statistics except that No
should be replaced by No/2 in the former since, as shown above, each Gaussian component
in Z; has half the variance. Equivalently, DPSK and NCFSK have the same error probability
performance except for a factor of two in SNR.

7.2 OPTIMUM WAVEFORM RECEIVERS BASED ON A LARGER THAN TWO SYMBOL

OBSERVATION INTERVAL
Thus far, we have discussed symbol-by-symbol differential detection of M-DPSK based on
an observation of the received signal plus noise over two symbol intervals. We have ob-
served that although differential detection eliminates the need for carrier acquisition and
tracking in the receiver, it suffers from a performance penalty (additional required SNR at
a given bit error rate) when compared with ideal (perfect carrier phase reference) coherent
detection. The amount of this performance penalty increases with the number of phases, M,
and is significant for M > 4. For example, at a bit error probability P,(E) = 10~>, DPSK re-
quires about 0.75 dB more E»/No than coherently detected BPSK (see Fig. 7.3). For QPSK
(M = 4), the difference in E;,/No between differential detection and ideal coherent detec-
tion at P,(E) = 10~> is about 2.2 dB. Finally for 8PSK, the corresponding difference in
E»/ No performance between the two is greater than 2.5 dB. The latter two E,/No penalties
may be obtained from Fig. 7.3 and the approximate result of Eq. (4.134).
Thus, it is natural to ask: Is there a way of enhancing classical (two symbol observa-
tion) M-DPSK so as to recover a portion of the performance lost relative to that of coherent
detection and still maintain the advantages of a simple and robust implementation? Further-
more, if this is possible, what is the trade-off between the amount of performance recovered
and the additional complexity added to the classical M-DPSK implementation? The answers
to these questions stem from the idea of allowing the obervation interval to be longer than
two symbol intervals while at the same time making a joint decision on several symbols si-
multaneously, that is, block-by-block detection, as opposed to symbol-by-symbol detection.
Sec. 7.2 Optimum Waveform Receivers 461

In order for this approach to have meaning, one must extend the previous assumption on the
duration of time over which the carrier phase is constant to be commensurate with the ex-
tended observation interval. For observations on the order of three or four symbol intervals,
this is still a reasonable assumption in many applications.
The theoretical framework in which we shall develop this so-called multiple-symbol
differential detection technique [22, 23] is once again the ML approach to statistical detec-
tion and is analogous to that used in Chapter 6 for deriving the optimum receiver of partially
coherent M-PSK based on an observation of N, symbols. As such, the resulting receiver
configurations derived here will be the optimum block-by-block detection receivers for dif-
ferentially coherent M-PSK. In the next section, we derive the appropriate ML algorithm.
Clearly, classical M-DPSK must result as a special case of this more general model when
Ne.

7.2.1 Maximum-Likelihood Differential Detection of M-PSK—Block-by-Block

Detection

Consider as before the transmission of M-PSK signals over the AWGN channel. The chan-
nel introduces the unknown phase, 6, which in the absence of any information is assumed to
be uniformly distributed in the interval (—z, 7). The received signal in the k"” transmission
interval then takes the form of (7.1) (with i replaced by k) which can be expressed in terms
of the complex baseband signal s, = V2 Pel % by [see (6.52)]

r(t) = Re {Sj exp (jact + 0) + A(t)} = Re {r(t)} (7.56)

where v(t) is a complex additive white Gaussian process. The receiver demodulates 7r(t)
with the complex reference signal exp(—jw@-t) to produce the complex baseband signal
5peJ? + n(t)e~/%", which is then passed through an I&D, resulting in the complex detected
sample 6

ee eo At Se
Vp = Spel” + T n(b)ew dt = spe 4 Ne @eai)
8 JKT;

Consider now an observation of the received signal over an interval of length N;7; (Ns is
herein referred to as the block length) and assume that the unknown channel phase, 6, is con-
stant over this interval. In particular, for an observation of the received signal over the inter-
val (k — N, +1) T; <t < (kK + 1)T,, the sequence v = (Viena Ve—Ny+2s Arar Ves Vi)
conditioned on @ is a sufficient statistic for making a ML decision on the transmitted se-
quence $s= (Se—N,+1) Sk=N, 42) We yee sk =

For the assumed AWGN model, the a posteriori probability of V given s and @ is

6. We apply the same normalization of 1/7, to the complex I&D output sample as was done in Section 6.3.
Again, we denote this normalization by a script typeface.
462 Differentially Coherent Communication with Waveforms Chap. 7

Pere~ s,9) span = ————— elie

(2707) a
exp dd {-————>_—
hace
2g
_
Ae

“pear EeaL) Di i + iP]

> [fa ae 7.58
(7.58)
(270? n | i=0
Ng=1 |
—2 »e Vu-iSp_;| CoS(O — a)
i=0
where
Nels
Im we Ve=iSf_;
i=0
I (7.59)
Nee
Re; >” Vu iS;
i=0

Averaging (7.58) over the uniform pdf of @ and simplifying gives the conditional probability

fy V8) = if_ ApeVIB. 9) f(a (7.60)

1 Ns~1 Ns—1

=(pean Qo? D [eii) ++ [Skil ‘lI.(|ye Ve- si)

exp {-5

noting as before that for M-PSK, 5,|? is constant for all transmitted phases. Then, since
eae
La is independent of the transmitted signal and since /o(x) is a monotonic function
2
Ns-1 |
of x, maximizing AVS) over S is equivalent to finding max | )> Vy~iSz_;| » which
S a
is identical to (6.59b) with p set equal to zero. Defining the transmitted phase sequence
6= (Ox—N,+15 Ou Nek One so, Ok Ts Ox), we obtain the decision rule:

A Ns=l . >
choose @ corresponding to max Se Ve_je
1 =!
i=0

where 9 = (net GeON EDs curbs 6,-1, 4x)is the decision sequence. Note that the actual

complex I&D output in the (k — i)" interval, Vg, is related to Ve-i by (see Fig. 7.2)

r, (k-i+1)T;
Ve—i = Vek—-i + I Vs,k-i= = Ves Va
ae awe r(t)exp{—ja@,t}dt (7.61)

and thus the above decision rule can equivalently be rewritten as

Niel 2
choose @ corresponding to max Sy Vj, je 1 %-
70
Sec. 7:2 Optimum Waveform Receivers 463

Note that, as for the two symbol observation case in Appendix 7A, the differentially
coherent decision rule for an N,-symbol observation also has a phase ambiguity associ-
ated with it since the addition of an arbitrary fixed phase, say @,, to all estimated phases
Ox N.+1> Oy NALD gas 6, ileby results in the identical decision rule, that is,

N,-1 2
choose @ corresponding to Ws se Vy exp {—j (@-; + 9g)}
t=)

Thus, as for classical M-DPSK, we must employ differential encoding at the transmit-
ter to resolve this phase ambiguity. Letting 6, = —0,—N,+1, then the above decision rule
can be rewritten in terms of the input information phases, AO,_; = O,—; — O—j-1; i=
Dal sa Ng = 2.88:

choose the information sequence A@ = (AOx—n,+2, AOk—N, +3, +--+» AOx—1, AO)

that maximizes the statistic

INe—1 N,=i=2 2

ae sexe i> si-in)t

(7.62)
: Ns=2 _ N;—i—2 2

= Vent + Da Vii Exp -;(= 101m)

1=0 m=0

A receiver that implements the above decision rule is illustrated in Fig. 7.7. Note that
an observation over N, symbol intervals is used to make a sequence (block) decision on
N, — 1| input information phases. As such, the Ns-symbol observation blocks overlap by
one symbol since the last symbol of a given block serves as the phase reference for the next
block. Thus, for this type of block-by-block detection, the rate at which decisions are made ~
is identical to that of the symbol-by-symbol detection scheme of classical M-DPSK.
Some special cases of (7.62) are of interest. For Ns = 1, that is, an observation of the .
received signal over one symbol interval, (7.62) simplifies to
$5 eA
A= i (7.63)
which is completely independent of the input data phases and thus cannot be used for
making decisions on differentially encoded M-PSK modulation. In fact, the statistic of
(7.63) corresponds to the classical case of noncoherent detection (see Chapter 5) which is
not applicable to phase modulation.
Next, let NV, = 2, in which case (7.62) becomes
i 2s es 2 San is ;
alent Vie I A% 2 =|%-1 +|%| +2Re {HiVg oe/4% (7.64)
This results in the decision rule

choose A6; corresponding to max Re |YeVi_,e 2%

k

which has a receiver implementation (see Fig. 7.8) that is the complex form of Fig. 7.2b, as
one would expect based on the results in Appendix 7A.
 Aeyod 1-3 | Ny Kejoq
| inn
L L Aejaq a kejoq

464
cS
=? < eV
1 w—1-¥
o=u

odojaaua
10}9919p

ae oreduioa
pue
199195
ay?
IsoZIe'T]

W |» sapoAo
tad ‘L(1—“N)
a]qeuo i 2,949 oun
Avyap
S.9V
UO1s

DY: SEN EHNA

oc co 19v''9
Jo

jo

§=L*L

wo
uonoa

sans1
jooiq

©[Zz
ajdyj
paiuti

jenuas

joquiA
ySdq-

(4a
uonej
 0/9

YSII9}se
Sa}OUaPp

xatduroo
a}e8n{uoo

uoneoydynur
u

asooy.)
x

*L(I+3)
Ip(e)

.
=
J='ev

{“Stoy
ST SOBIE]

aang
gt, UNUINdQ JoAIg091 xa[dwOd)
(WiI0J
10j [edIssejo y¥Sdd-W payurday)
wos [ZZ]© (JAI

465
466 Differentially Coherent Communication with Waveforms Chap. 7

Next, we consider (7.62) for N; = 3. Here we have

F 2
pre Vx-2 fe Vije (0 a6 Ve F(A%-1+ 4%)

2 3 2 wf aes
B
oF i +2Re {VaVite /A%} (7.65)
= \%-2| “4 x1

+ 2Re {Vi- Vee Re 7 Ut ge J(a%-14Am)|

Thus, the decision rule becomes

choose Aby and A6z-1 corresponding to

max Re {V1Vine Ta + Vy-1 Viner Pet “Ve Vie Ine)

AO
1, AOE

Note that the first and second terms of the metric used in the decision rule of (7.65) are
identical to those used to make successive and independent decisions on AO and A®&_},
respectively, in classical M-DPSK. The third term in the optimum metric is a combination
of the first two and is required to make an optimum joint decision on AO and A@_1.
Clearly, a receiver implemented on the basis of (7.65) will outperform a classical M-
DPSK receiver. Before demonstrating the amount of this performance improvement as a
function of the number of phases, M, we first present the implementation of the optimum
Ns = 3 receiver in accordance with the above form of the decision rule. Figure 7.9 is a
parallel implementation and Fig. 7.10 is a series implementation that although simpler in
appearance requires envelope normalization and additional delay elements. It should also
be noted that in the parallel implementation of Fig. 7.9, the apparent M? phasors needed to
perform the phase rotations of the output V;V,"_, can be accomplished with only M phasors
since the sum angle A@ + A6,—; when evaluated modulo 27 ranges over the same set of
M values as does AO; and AO;_1.

7.2.2 Bit Error Probability Performance

To obtain a simple upper bound on the average bit error probability, P,(£), of the opti-
mum block-by-block M-DPSK scheme, we use a union bound analogous to that used in
Chapter 6 to derive the performance of the optimum partially coherent block detection
receiver. In particular, the upper bound on P;(£) is the sum of the pairwise error prob-
abilities associated with each (N, — 1)-bit error sequence. An exact evaluation of these
pairwise error probabilities is carried out in the next section. Mathematically speaking,
if A@ = (AOx—N, +25 AGkoN.
23; «05, AOpen AO) is the sequence of N, — 1 information
phases and A6 = (Adi—n42, AOk_N, 435 ae A6z-1, Ab) is the corresponding sequence
of detected phases, then if u denotes the sequence of b = (Ns — 1)log)M information bits
that produces A@ at the transmitter and a the sequence of b bits that result from the detec-
tion of A@, then the bit error probability has the upper bound [see (4.88) with M replaced
by M™s~1}
irae
PAE) SF aamct 2, 2 wee, ) Pr {fi > 0140} (7.66)
A0# A6
 “ev
S‘g
gsoouD

"eV
= “d
aN

BEN |
SOSIE] ST

467
aINBI
GL [a[feseg ueyeyuaWe[dun
Jo afdynw foquiAs JeNUsIaysIp‘1oONIEp
SV = ¢ pajutiday)
woy [ZZ]© ‘(aaal
468
Ogs—?

st 1SoBIR]

MeT-arenbs
adojaauq
10}93}2q

aansiy
OTL SIUIS uoneyuswajd
Jo ajdninw
ut [oquiAs [enuasaygip‘10j0019p
“yy = ¢€ pautiday)
wos [ZZ]© (AqagI
Sec. 7.2 Optimum Waveform Receivers 469

where w(u, ) denotes the Hamming distance between u and &i and Pr {4 > n|A@} denotes
the pairwise probability that A6 is incorrectly chosen when indeed A@ was sent. The
decision statistic 1 is defined in (7.62) and the corresponding error statistic 7 is identical
to (7.62) with each AO replaced by Aby. For symmetric signaling sets (such as M-PSK),
(7.66) satisfies a uniform error probability (UEP) criterion, that is, the probability of error
is independent of which input phase sequence A@ is chosen as the correct sequence. Under
these conditions, (7.66) simplifies to [see (6.62) with ¢ = 0]
1
P(E) < W,Sane Sy w(u, G) Pr {7 > n|A@} (7.67)
= D
A64A6
where A@ is any input sequence (e.g., the null sequence (0,0, ..., 0) = 0).

Evaluation of the pairwise error probability. To compute Pr {7 > n|A6},

we take the approach of Section 6.4.1 which, for the particular case under consideration,
was evaluated in detail in [22, Appendix AJ. In particular, letting n = lz1/? and 7 = \zo|",
then analogous to (6.64) we have
1
Pr {> n|A@} = Pr{\z2l > |z1||A0} = 5 [1— Q(vb, Ja) + Q(Va, vb)| (7.68)
where Q(q, f) is as before Marcum’s Q-function, whose arguments are given by (6.65) and
evaluate to (see Problem 7.8)

{?|i aRea [net

b Es 9)
— yy |8|*
2
(7.69)
7.6 |
with
N,-1 N anf) Ss

5A ae exp f > (S05 = sih-in)|

i=0 m=0
(7.70)
Nix Nei? |
\|> y— exp J a 50K —i-m
1=( m=0

In (7.70), it is understood that the summation in the exponent evaluates to zero if the upper
index is negative.
Note that for any given N,, M, and data sequence A@, 5 can be evaluated for each
error sequence A@. We now demonstrate the evaluation of (7.67) and (7.69) for some
special cases of interest.
Special cases.
Case 1. Classical DPSK (N; = 2, M = 2)
From (7.70), we immediately get 5 = 0 and thus from (7.69)
b 2Es
Paes CHAD

Substituting (7.71) into (7.68) gives

2 I 2E; 2E,
mi niao}=3|1-0( a0) +0(0 =)| CAP)
470 Differentially Coherent Communication with Waveforms Chap. 7

From the definition of the Q-function

2
O(@;0)=1; “OO; B) =exp (-5) (73)

Since for the binary case the pairwise error probability is indeed equal to the bit error
probability, then we have from (7.72) and (7.73) that

P,(E) = = exp (-2) (7.74)

which is identical with (7.14).

Case 2. DPSK with 3-Symbol Observation (NV; = 3, M = 2)

Here, there are three possible error sequences of length two. The pertinent results
related to the evaluation of (7.70) and (7.71) are given below:

AQ — AOy AOy 1 — A6x-1 r) fs

(ae || 33

x 0 +1 & [343]
x 1 +1 &[34v2]
Since the Hamming distance w(u, W) is equal to unity for the first two error sequences
and is equal to two for the third sequence, then using the above results in (7.69) and (7.70),
the upper bound on bit error probability as given by (7.66) is evaluated as

niey<1-0( 5 (5+¥2) #(3-~2)]

-o((8E-9 EE)
To see how much performance is gained by extending the observation interval for
differential detection from N; = 2 (classical DPSK) to Ns = 3, we must compare (7.75) to
(7.74). Due to the complex form of (7.75), this comparison is not readily obvious without
resorting to numerical evaluation. On the other hand, by examining the asymptotic (large
Es/No) behavior of (7.75), we can get an immediate handle on this gain.
When both arguments of the Q-function are large, the following asymptotic approxi-
mations are valid [24]:

Re 2
O(@,6) =1- fF 500 |-ol,
e a>Bp>|
(7.76)
ee Wea et (B —a)?)-
O10 ON ae F ol#2 B>a>l
Sec. 7.2 Optimum Waveform Receivers 471

Using these approximations, (7.68) becomes

2
z 1 |
Pr {if}> |A0} = 5 eae eeb i; a ao
lb = (1.77)

or, from (7.69)

Pr {7 > n|A@} =
— y NP 18)?
2/2 Fn, — 181) \ in,+ \/N2 — 18?
1/4
Ns Sey +4/Neal}
N2 — |6|? MEA hee (7.78)
+
SSS= KP leaan Dia plon
Ns — ,/N2 — |6|? Y

Nsrt | ex {=sein - 1b}

mae 51(N;— 18D} | 2No-
For N; = 3 and M = 2, |5| = 1. Thus, (7.75) becomes the approximate upper bound
2/2 [1 E
P(E) S ve Eexp \-= i (7.79)
5 ||
V7 No
Comparing (7.79) and (7.74), we observe that the factor in front of the term in brackets in
(7.79) represents a bound on the improvement in performance obtained by increasing the
memory of the decision by one symbol interval from N, = 2 to N,; = 3.

General asymptotic results. In the general case for arbitrary N,, the dominant
terms in the bit error probability occur for the sequences that result in the minimum value
of N, — |6|. One can easily show that this minimum value will certainly occur for the error
sequence Aé having N; — 1 elements equal to the correct sequence A@ and one element
with the smallest error. Thus

min (N, — |6|) = 5 [Ns 1 gh A%-Ab min

Ad, A6
(7.80)
+ (Ns — 1)(2—2.)+1
= N, — /(Ns — 1)?

= Ns — |6|max

where

2 = ae A 6k) min ee ok
dain Ne Waa AY = 4sin 7; (7.81)

and (A& — A6x) min denotes the smallest possible value between the iit symbols of the data
and error phase sequences, that is, 27 /M. Also note that for |5| = |6|max, (7-70) reduces to

12} Es [Ns + 2V/N,= Isin =| (7.82)

ak
472 Differentially Coherent Communication with Waveforms Chap. 7

Thus, the average bit error probability is approximately upper bounded by

1 A
PRE is= aT
a 1ylogy M Se w(u, &)
Ad4AO (7.83)

1 Ns+ |6|max E;
is Ie
as ee ———
(Ng — |8lmax)
2 ime5 ( 15 |max (Ns ae lorax) 2No ‘eae

where w(u, ti) corresponds only to those error sequences that result in |6|max.
For the binary case (M = 2), we have from (7.81) that Bae = 4 and hence N; —
\5|max = 2. Similarly, it is straightforward to show (see [22, Appendix B] and Problem 7.10)
that the sum of Hamming distances required in (7.83) is given by

> w(u, a)
i
= ZN
he G1):
eae
Ns = 2
(7.84)
Ad4AO
Thus, (7.83) simplifies to

2 N, — 1 1 Es
Pi (BE) = ied ( me ;)E exp =] (7.85)
No

which is the generalization of (7.79) for arbitrary N,.!

Equation (7.85) has an interesting interpretation as N, gets large. Taking the limit of
(7.85) as Ns — 00, we get

1
P)(E) exp {= (7.86)
[e.
7 No
which when expressed in terms of the first term of the asymptotic expansion of the comple-
mentary error function, namely
]
erfe x & exp (—x?) (7.87)
JX
becomes

|E
P, Serfc a (7.88)

From the results in Chapter 4, the exact bit error probability performance for coherent
detection of BPSK with differential encoding and decoding [see Eq. (4.206)] has an asymp-
totic upper bound identical to (7.88). Thus, as one might expect, the performance of multiple
symbol differentially detected BPSK approaches that of ideal coherent detection BPSK with
differential encoding in the limit as the observation interval (decision memory) approaches
infinity.

7. Note that (7.85) is not valid for N, = 2 since in that case |5|max = 0 [see (7.80) and (7.81)] and thus the
inequalities of (7.76) are not satisfied.
Sec. 7.2 Optimum Waveform Receivers 473

A similar limiting behavior as the above may be observed for other values of M. In
particular, it can be shown (see [22, Appendix B] and Problem 7.9) that for M > 2 anda
Gray code mapping of bits to symbols, the sum of Hamming distances corresponding to
I5|max iS given by

Ds naa) =r)
A De
Res
Ns = 2
(7.89)
A64A6
Using (7.89) in (7.83), we get (for N; > 2)

mn (log M) H Fe ( 1Ol max

Hs)
(Ns ox (Ola 7 90)

Es
x exp aN NS —="1S5 bax)

where, from (7.80) and (7.81)

[5|max = (Ns — 1)? + 2(N, — 1)(1 — 2sin?2/M) +1 (7.91)

For N,; = 2, the asymptotic upper bound on bit error probability becomes

1 cos 4,
Og Se
S ‘ IU ate
(log, M) 21 x, SiN 347 ,/COS 44 (7.92)

x exp 4——— sin? zee |

P No 2M

As Ns gets large, |6|max > Ns — 2 sin? z1/M and (7.90) reduces to

P,(E)
b < ———— e expp j - —No
Es sin? —M=
(logy M),/2%¢sin? %
(7.93)
1 Big)
= erfc — sin —
log» M No M

which is identical to the asymptotic bit error probability for coherent detection of M-PSK
with differential encoding and decoding [see Eq. (4.134), (4.135) together with the asymp-
totic form of (4.200)].°
Figures 7.11, 7.12, and 7.13 are illustrations of the upper bounds of (7.90) and (7.92)
for M = 2, 4, and 8, respectively. In each figure, the length (in M-PSK symbols) of the
observation interval is a parameter varying from Ns = 2 (classical M-DPSK) to N; = oo
8. It should be noted that the result in (7.93) can be obtained by observing that, for large Ns, (7.82) satisfies
Jb > Vb — Ja > 0. In this case, (7.69) can be approximated as in (6.67) (also see [25], Appendix A) by

Pr {7 > n|Ae} = 5 erfe (

I ao)
J2
Using this relation in (7.67) gives the asymptotic bit error probability in (7.93).
474 Differentially Coherent Communication with Waveforms Chap. 7

SIMULATION POINTS
BN 2
O Ns =3
A N,=5
Ng = 2

Figure 7.11 Bit error probability versus

E»/No for multiple symbol differential
6 7 8 9 fOnb esi detection of M-PSK; M = 2 (Reprinted
E,/No, dB from [22] © IEEE)

(ideal coherent detection). Also indicated on the figures are computer simulation results
corresponding to the exact performance. We observe that, for example, for DPSK, extending
the observation interval from N,; = 2 to N; = 3 recovers more than half of the E,/No loss
of differential detection versus coherent detection with differential encoding. For M = 4, the
improvement in E,/No performance of N; = 3 relative to Ns; = 2 is more than 1 dB, which
is slightly less than half of the total difference between differential detection and coherent
detection with differential encoding.

7.2.3 Implementation Algorithms

In the previous section, we showed that for sufficiently large block size, Ns, the bit error
probability performance approaches that of coherent detection of differentially encoded M-
PSK. Unfortunately, the complexity of a receiver implementation of the type illustrated
in Fig. 7.9 for Ns =3 grows exponentially with N;. To see this, we first observe that,
for general N,, the decision metric [such as that described by (7.65) for N; = 3] must be
evaluated for Ms~! hypotheses (the number of possible A@ vectors of length N, — 1).
For each of these hypotheses there are ee terms in the metric [see, for example, the

decision rule below (7.65) which contains S = 3 terms within the braces]. Thus, we

N;—1
must calculate ( 5 )phase angle sums of the form AO_, + A@—g41 +... + AG—p

Ns F Ns Di.
for 0 < p<q< WN; —2, perform 7 phase rotations and DON fos 1 real additions.

Ns \ .
Since (Ye)is on the order of N?, the total number of operations required to implement

the decision rule is on the order of 3N2 x MNs-1,

Sec. 7.2 Optimum Waveform Receivers 475

SIMULATION POINTS

Figure 7.12 Bit error probability versus

6 + 8 9 10 1 12 13 E,/ No for multiple symbol differential
detection of M-PSK; M = 4 (Reprinted
E,/No, dB from [22] © IEEE)

SIMULATION POINTS
Gl Ne. =\2
O Ns, =3
A Ns, =5

N, = 2

Figure 7.13 Bit error probability versus

E,,/No for multiple symbol differential
Oe Onme tsne e igus alo Aime to geil et -7
detection of M-PSK; M = 8 (Reprinted
E,,/No, dB
from [22] © IEEE)

Recently, Mackenthun [25] suggested an algorithm for implementing multiple sym-

bol differential detection of M-PSK, which reduces the number of operations for each N,-
symbol block from order 3N2 x M%s~! to order N; logy Ns. The impact of this fast com-
putational algorithm is that it allows for a scheme that is a practical and viable alternative
to coherent detection of differentially encoded QPSK without the need for carrier recovery
476 Differentially Coherent Communication with Waveforms Chap. 7

at the receiver. We present here without proof (see [25] for the details) a description of this
algorithm.
Consider the vector of complex I&D samples” V= (v1,V2, RSE Vw.) corresponding
to a particular observation (block) of Ns; symbols. Let vj = tan! (Vs riVeni) denote the
argument of each component of this vector [see (7.61)]. For each component in the vector
V, define the complex quantity

Zz, = Ve =|, Y= vi eli (7.94)

where for each value of i= 1, 2,..., Ns one finds the phase angle 6; € {0,27/M,4n/M,
...,2(M — 1)2/M} such that ¢ A arg {Zi} = v; — 6; € [0, 27 /M). Next, list the values of
¢; in order of largest to smallest. Define the function k(i) as giving the subscript k of ¢ for
the i" list position, that is,

O < SkN,)) SK(Ns—1)9 + ++» $k(2)> Sk) < = (7.95)

Let

8 = Zu) (7.96)
The complex quantities {g;;i = 1,2,..., Ns} are referred to by Mackenthun as the remod-
ulations of the received complex samples {V;;i = 1, 2,..., Ns} in [0, 277/M) ordered by
the magnitude of their phase angles. Augment the set of values of g; by including a 217/M
phase shift of each as follows:

Bi =e JIM
Gg, ys i=Nst+1,Ny+2,...,2Ns (7.97)
The extended set of 2N; values of g; represent all possible remodulations of {V;; i = 1, 2,
.., Ns} in (—22/M, 21/M) indexed by their phase angle, going clockwise around the
circle starting at e/?”/”. Next form the set of N, sums

q+Ns-1 2
figA eg, | ge eee (7.98)
l=q

Let q’ be the value of q corresponding to the largest value of Ng. 1.e.,

q =max ng (7.99)

and define the phases

Axi) =O; g' SEX Ng

2 . on (7.100)
. /
Oki —N;) = OG —N,) + Vi Ns <i<q+WN;-1

Arranging the phases 6x(i); i= 1, 2,...Ns in order of subscript value k(i) produces the joint
decision @ = (41,Coins 6.) on the sequence of transmitted phases. Finally, the decision

9. For simplicity of notation, we shall choose the discrete time k equal to N, — 1.

 Sec. 7.3 DPSK 477

A6 = (Ad, NOs ct Ady, )on the information data phase vector has components A6; =
6; — 6;-1; t= 42h Ny where 6;: i=1,2,...Ns are the components of6 found in
(7.100).
As an example, suppose that M = 4, N, = 6 and the ¢;’S are ordered such that 0 <
62 < 66 < 05 <b1 <3 < f4 < 2/2. Then
81 = Z4, 82 = Z3, 83 = Z1, 84 = Zs,Bs= Zo, 86 = Zr

BF Dae a= Zoer go = Ze)? gig Zee?) I (AON)

B= Zeer gi Zoe 1%?
Suppose now for the values of g; in (7.98) we find that g’ = 4. Then 6= (6 + Oe 6, 63 +
5,04 + 5,05, Os) and A@ = (24, 5,0, 3%, 0).

7.3 DPSK—NONCOHERENT DETECTION OF OVERLAPPED BINARY ORTHOGONAL

MODULATION
Earlier in the chapter, we demonstrated that the decision variables of DPSK and noncoherent
FSK are statistically equivalent except for a 3 dB difference in effective E/No. Here we
show an interpretation of DPSK as noncoherent detection of overlapped binary orthogonal
signals from which we can immediately glean the 3 dB difference in E/No performance
between the two techniques. This interpretation of DPSK also has the advantage that it can
be generalized to M-ary (M > 2) modulation.
Consider a pair of complex signals so(t) and s;(t) defined over the interval 0 <t <
2Ty as follows:

5e(t) = V2P px(tyel?"; k=0,1 (7.102)

where
(pia NS 5S ie Sp
PO)= 0; elsewhere
(7.103)
(= Ile OS,
PAM) = —1; T,<t<2T,
Clearly, so(t) and s(t) are orthogonal. Suppose now that we were to transmit a DPSK signal
corresponding to the binary information sequence {a;} = 1, —1, 1,1, —1,1,—1, —1, —-1, 1
which occurs at a data rate 1/7}. The actual transmitted baseband waveform (normalized to
unit power)
9
S(t) =) bip GIT»), bi = aibi-1 (7.104)
i=0

appears as in Fig. 7.14 (assuming a +1 as a reference bit, i.e, b-1 = 1) where p(t) is a
unit rectangular pulse of 7;,-sec duration. We observe that S(t) of (7.104) can be viewed
as a sequence of overlapped 27)-sec pulses of the form given in (7.103). Thus, if s(t) =
J/2P S(t)e/%! denotes the complex transmitted DPSK signal, then from Fig. 7.14 we see
that S(t) can be represented as the overlapped sequence So(t), 51(t), —So(t), —So(t), —81(0),
478 Differentially Coherent Communication with Waveforms Chap. 7

50(t), 51(t), —51(t), 51(t), —So(t) where the interval of overlap has duration 7p sec. In short,
the DPSK signal has been represented as a sequence of 100% overlapped binary orthogonal
signals.
Note that except for the polarity of the p,(t)’s in the overlapped pulse sequence
of Fig. 7.14, there is a one-to-one correspondence between the input data bit a; in the
interval i7, < t < (i + 1)T> and the transmitted pulse p,(t) in the interval (( —1)7T, <t <
(i + 1)Tp. In particular, a; = +1 corresponds to transmitting po(t) (or —po(t)) anda; = —1
corresponds to transmitting p(t) (or —pi(t)). As such, an optimum decision rule for the
information data sequence {a;} can be obtained by noncoherently detecting the sequence of
overlapped signals. In particular, based on an observation of the received signal, r(t), over
the 27)-sec interval (i — 1)7, < t < (i + 1)Tp, the £1 decision on a; is made according to
the rule
Pre 235 2
(i+1)T) (i+1)T,
i; r(t)s{(t)dt Hl r(t)S5g(t)dt (7.105)
(i-1)T, a (i—1)T,
The decision rule in (7.105) represents an optimum noncoherent decision on a pair of binary
orthogonal signals [So(t) and s;(t)] whose duration is 27) sec and is illustrated in Fig. 7.15.
Thus, the error probability associated with this decision is identical to (5.79) with Ey = PT,
replaced by 2E, = P (2T;), namely

Braye Lee (-3) (7.106)

However, (7.106) is identical to the bit error probability performance of DPSK as given by
(7.14). Thus, we have demonstrated the desired equivalence between DPSK and noncoher-
ent detection of overlapped binary orthogonal modulation. Stated another way, Fig. 7.15 is
an alternate receiver implementation to Fig. 7.2b.

7.3.1 Generalization to Overlapped M-ary Orthogonal Modulation

The notion of signaling with overlapped sequences of orthogonal signals can be generalized
to M-ary modulation. !° Here, however, the overlap is less than 100%. In particular, consider
a set of M complex signals {5,(t); k=0,1,..., M — 1} defined as in (7.102) in terms
of a set of M pulses {px(t); k =0,1,...,.M — 1} which in addition to being orthogonal
over the interval 0 <t < Pots (T; = Tp logy M) have the following properties. The
px(t)’s are all identical in the interval 0 < t < mals and differ only in their polarity in
the interval T; <t < Wot. A set of functions that has these properties can be generated
from a Hadamard matrix [1] by using the rows of this matrix as generating sequences for
rectangular-shaped waveforms. An example of such a set of pulse waveforms for M = 4 is
illustrated in Fig. 7.16 corresponding to the 4 x 4 Hadamard matrix
Geawiones
(7.107)

10. This generalization was suggested by Dan Raphaeli while a Ph.D. student at Caltech and appears in an
unpublished document.
 479
[euosoy}I0
sas[nd
Aseuiqg paddeypsoao Jo Suto} Ul ULIOJOARM YSqq puegaseg e jo uoNRlussaiday py sansiy
——

(1)'d
tea (yd
——e}—— ('d-
—e}=— ('c-
—>}=—_ er
—e}— (y- —e}=— u- —>}=— a —> |
480
A-
7rous
21+)
of 1p(0)

(J)4

sploeq
'p = °z}u3s
— {'z
11+)
we),

dp
P'Sso9
aansiyqsz uonrjusu
JO sjdwy
YSdq JOAINA1
se juardYOouoU
uoNoa}9p
Jo Areulg Jeuosoyrio
sjeusis
Sec. 7.3 DPSK 481

Po(t)
q

p(t)

Pp, (t)

p(t)

-1

Figure 7.16 Pulse shapes for overlapped binary orthogonal modulation; M = 4

Suppose now that we were to transmit the same information data sequence as be-
fore (see Fig. 7.14). Consider grouping these bits into symbols consisting of n = log, M
bits apiece corresponding to (qj, dj+1,...,4@j+n—1) in the interval i7, <t < (i + n)T) and
assign px(t) (or —p,(t)) as appropriate to the interval (i— 1) Tp <t < i +n)T} ac-
cording to the following rule. Let (k, k2,..., k,) be the binary (0, 1) representation of the
integer kwhere | < k < M. Then, p,x(t) (or — px(t)) is assigned to the binary (+1) sequence
aj = 1 — 2k, aj4, = 1 —2kp, ..., Gj4n—1 = 1 — 2k,. For the n = 4 example, we would have

a; = 1, 4341 =1> polt)

ai = 1,4a;41 =—-1=> pit)

(7.108)
Q; = —1, aj4) =1=> ptt)

ay = —1, a3. = 1 => p3(t)

482 Differentially Coherent Communication with Waveforms Chap. 7

Using the above assignment, the overlapped pulse sequence for the data sequence of
Fig. 7.14 is illustrated in Fig. 7.17. Note that the overlap interval here is T;/3 = 27/3
Sec:
An optimum decision rule for choosing groups of n information bits (i.e., data sym-
bols) is again obtained by noncoherently detecting the sequence of overlapped signals. In
particular, based on an observation of the received signal, r(t), over the n (#4) Tp-sec

interval (i_ wa) Ty <t < (i +n)Tp, the joint +1 decisions on a; and a;+) are made ac-
cording to the rule. Decide a; and a;+) corresponding to p;(t) as determined by (7.108)
where
2
(i+n)Tp
aan r(t)sy(t)dt (7.109)
ab 5| )T

The above decision rule represents an optimum noncoherent decision on a set of n orthog-
onal signals, each of duration n (wir)(i (Cai)T; sec. Thus, the symbol error prob-

ability associated with this decision is identical to (5.77) with E; replaced by (4) 1B
namely

PAE
3(E) = ba(= UN
es jMol C vex | aLad
- (; mana(=)}
\ (33 Es (7.110)

The bit error probability is related to the symbol error probability of (7.110) through (4.96).
In view of the above, we see that overlapped noncoherent M-ary orthogonal modulation
can achieve an improvement in E/ No by a factor of (4). Clearly, as M increases, the
improvement factor diminishes and is maximum for M = 2 where it corresponds to 3 dB as
previously shown.
Another way of viewing the Pray improvement factor in E/No for overlapped non-
coherent orthogonal signaling is to consider the equivalent detection problem for nonover-
lapped signals. In particular, suppose that we ignore the overlap region, that is, the region
where all transmitted signals are defined to be identical, and just define the complex signals
5,(t) as in (7.102) but only in the interval (Gat:)ote (#45)
Ye T;. Then a noncoherent
decision on this new signal set {si (t)} defined over a 7;-sec time interval should produce
the same error probability as the overlapped orthogonal set {5,(t)} defined on a (4) Ts-
sec interval. The difference is that the members of the nonoverlapped set {5,(t) } are not
orthogonal. To see what their correlation properties are we proceed as follows. We can write
5,(¢) in terms of 5, (t) as
l
Wow) 2s
H(t) = ( nate (7.111)

Using (7.111), it is straightforward to show (see Problem 7.13) that the 5,(t)’s are transorth-
ogonal (see Chapter 4), that is, their off-diagonal normalized correlation is equal to -Wh.
As such, {5,(t)} represents a simplex set that produces an wo improvement factor in E /No
performance relative to an orthogonal signaling set.
 ae
a
Oe
fo—
a
i.)
=
aoe

«LIL,
ejep
Jo aouenbas
poddejisao
yeuosoyyio
asnd
Areulg
ANSIYy
Ase-p
uONRIUasaIday
aouanbes
&se

(1)'d

483
 484 Differentially Coherent Communication with Waveforms Chap. 7

r(t) |Wideband| 7 : a,=cosA@,

Bandpass
Filter

Figure 7.18a A suboptimum DPSK receiver—wideband IF, 1&D detector

sample once
per bit
Lowpass
Bandpass
Filter Filter

Figure 7.18b A simpler suboptimum DPSK receiver—narrowband IF, sampling detector

7.4 SUBOPTIMUM WAVEFORM RECEIVERS

To many people, the acronym “DPSK” implies the optimum receiver of Fig. 7.2a while to
many others it implies a simpler, but suboptimum, receiver. One such suboptimum receiver
is illustrated in Fig. 7.18a. The primary difference between Fig. 7.2a and Fig. 7.18a is the or-
der in which the differential detection and matched filtering (I&D) operations are performed
since ordinarily the bandpass filter in Fig. 7.18a is assumed to be sufficiently wideband so
that signal distortion and intersymbol interference (ISI) are not an issue. Quite often, how-
ever, to simplify the implementation still further, the I&D is replaced by a lowpass filter (to
remove second harmonics of the carrier) and sampler (one sample per bit typically taken at
the center of the bit). Furthermore, the input bandpass filter is now made narrowband with a
bandwidth optimized to trade signal distortion and ISI against noise power reduction. This
configuration, which is illustrated in Fig. 7.18b, has been suggested in many places in the
literature [26, 27, 28].
The suboptimum receivers in Fig. 7.18 behave similarly in that they both demodulate
the received (and filtered) bandpass IF signal in the current transmission interval with a ref-
erence corresponding to the same signal in the previous transmission interval. The fact that
this demodulation reference is perturbed by additive noise is what distinguishes this form of
differential detection from true coherent detection wherein the demodulation reference is as-
sumed to be noise-free. Furthermore, since this noisy demodulation operation is in principle
a correlation of the received (and filtered) bandpass signal with itself T,-sec ago, these im-
plementations are often referred to as autocorrelation demodulator (ACD)-type differential
detectors.
In what follows, we present approximate analyses of the performances of these
two suboptimum receivers and compare them with that of the optimum configuration in
Sec. 7.4 Suboptimum Waveform Receivers 485

Fig. 7.2b. The details of these analyses follow along the lines given in Park [2]. A more
exact evaluation of the performance of ACD-type differential detectors is given in Chapter
8 as a prelude to the discussion of double differential detection.

7.4.1 Wideband IF Filter, Integrate-and-Dump Detector

As stated above, we assume that the IF bandwidth, W (B = W/2 is the equivalent lowpass
bandwidth) of the input bandpass filter is sufficiently wide that the desired signal will pass
through this filter undistorted and with negligible ISI, that is, 1/T, << B <,/2z. Thus,
from Fig. 7.18a, and (7.1) _
r(t)=~V2P cos (@,t + 6; + 0) +:n(t)
(7.112)
= J/2P cos (wet +6; +0) + V2[n¢(t) cos (wet + 6) —ng(t) sin (wet + 0)]
where the hat denotes filtering by the input bandpass filter. For ease in analysis, we shall
assume that the product of the radian carrier frequency and the bit time, a7), is an integer
multiple of 27 and B is an integer multiple of 1/27;,. Under these conditions, the output, Vj,
of the I&D is (ignoring second harmonic terms)
(i+1)F
V; = / r(t)r(t — Tp)dt = PT, cos AG;
iTp

(i+1)T,
4 vP | [n-(t)
cos6-1 + ne(t — Tp) cos 6;] dt
ee (7.113)
(i+1)Tp
~ vP | [ns (t) sin 6-1 + ns(t — Tp) sin 6;| dt
iTp

(i+1)T,
ee ‘| [nc(t)nc(t — Th) + ne(t)ne(t — Tp)] dt
iTp

Since for DPSK the transmitted phase only takes on values of 0 and zr, then (7.113) simpli- 7
fies to
(i+1)Tp
V; =, r(t)r(t — Tp)dt = PTp cos AG;
iTp

(i+1)T,
3 vP | [tne(t) tne(t — T,)]dt (7.114)
1Tp

(i+1)Tp
+/ [ne (t)nc(t — Tp) + ns(t)ns(t — Tp)| dt
iT)

where the plus sign is associated with a transmitted phase equal to 0 and the minus sign with
a transmitted phase equal to 2. Assuming an ideal (unit height rectangular frequency charac-
teristic) filter for the input bandpass filter, then n,(¢) and n;(t) are statistically independent
with autocorrelation function
sin 27 Bt
Ry(t) = E {ne(t)nc(t + t)} = E {ns (t)ns(t + T)} = NoB ( aT BT ) (7s 103)
486 Differentially Coherent Communication with Waveforms Chap. 7

In view of the assumption that B is an integer multiple of 1 /2Ty, then R, (Tp) = 0 and hence
the first two moments of V; are given by

ee Ad; =0
Ue = sli eNO} 70

/sin 27 B(t — t’)‘)

[ sin sin 27 B(t — t' — Tp) ;
of =2PBNo |G+DT /TepO+DI 2a Bit
On B(t ——t’)
ind
id fps a1
2x B(t — t' — Tp)
2)
(7.116)
G+DT p+)T | (sin 22 B(t — t’)
+2(BNo)* ile Gace
27 Bit — t')
sin 27 B(t — t' + Tp) sin 27 B(t — t’ — Ty)
aL dtdt’
2a B(t — t’+ Tp) On Bit —t'—Tr)

Since we have assumed that B >> 1/7», then the above integrals can be approximately
evaluated, resulting in

oy, = PT»No + BT,Ng (7.117)

Both noise terms in (7.114) are zero mean and for sufficiently high SNR, the first of the
two, which is Gaussian, dominates. Thus, the bit error probability is evaluated in the same
manner as for an antipodal signal set on an AWGN channel, namely

(7.118)

Comparing (7.118) with (7.14) we can readily identify the degradation in performance of
the suboptimum configuration with respect to that of the optimum one.
We remind the reader that care must be exercised when using (7.118) that all of the
above assumptions are met. In particular, if B is made too large, then V; becomes decid-
edly non-Gaussian. Similarly, if B is made too small, then the assumption of zero signal
distortion becomes inappropriate. In addition to these conditions, it is also important not
to let Ep/No become too large relative to BT), else the Gaussian assumption again be-
comes incorrect. In particular, if for BT), fixed (and large) we were to let E,/No — oo, then
from (7.118) we reach the erroneous conclusion that the suboptimum differential detector of
Fig. 7.18a asymptotically performs 3 dB worse in SNR than the optimum detector of coher-
ent PSK, the latter being asymptotically equivalent in performance to the optimum detector
of DPSK (see Fig. 7.2b). The range of BT, and E,/No values over which the Gaussian
assumption provides approximately valid results for error probabilities of practical interest
will be illustrated in Chapter 8, where the performance of this configuration is revisited in
the context of a discussion on the behavior of differential detectors in the presence of fre-
quency error.
Sec. 7.4 Suboptimum Waveform Receivers 487

7.4.2 Narrowband IF Filter, Sampling Detector

Since in this configuration (see Figure 7.18b) there is no I&D to narrowband filter the noise,
the input bandpass filter must be used for that purpose. Once again, we will assume an ideal
(unit height rectangular frequency characteristic) filter with bandpass bandwidth W = 2B,
which will now be chosen on the order of half the bit rate. More about that in a moment. It
is convenient here to represent the received DPSK signal in the form of a binary rectangular
pulse train amplitude modulated on the carrier. Thus, in the i transmission interval, (GAD)
can be rewritten in the form

r(t) = V2Pbjp(t — iTy) cos (ct

+ 8) + n(t) (7.119)

where b; = cos 6; is the binary (+1) amplitude equivalent of the transmitted phase modu-
lation and p(t) is a unit rectangular pulse in the interval (0, 7). After passing through the
bandpass filter, we get

F(t) =V2P ) > dip (t —iT,) cos (wet +0) + A(t)

i=—co
(7.120)
(oe)

= /2 vp \- bip(t — iT») + ntoCos (Wet +6) — V2ng(t) sin (wet + 0)

i=—0oOo

where

A sind. Br’
b(t) =2B thesesan BY
p(t) LEY, f (7A )

Note that because p(t) extends before and beyond the 7;-sec duration of the current symbol,
ISI exists and we must sum over all past and future data bits. To obtain a detection sample
for making a decision on the information bit a; = bj-1b; = cos A@;, we multiply 7(t) by
7(t — Ty), filter out the second harmonic of the carrier, and sample the remaining lowpass
component at t = (i + 0.5) Tp, giving

Ve vp \ bep (G —& + 0.5) Ty) + ne (i +0,5) n|

k=—oo

(7122)
x lveKB (G — k — 0.5) Ty) + ne (i = 0.5) Th)
k=—0o

+ ns (i + 0.5) Tp) ns (i — 0.5) Tp)

Computing the bit error probability based on comparing (7.122) (which accounts for
the total ISI) with a zero threshold is in general a difficult if not impossible analytical
task. Thus, we must make some simplifying assumptions from which we can obtain an
approximate result. One such simplification, suggested in [28], assumes that B = 1/27} and
that the bandpass filter is cleared or reset at the end of each bit time to eliminate ISI. Under
this assumption, (7.122) simplifies to
488 Differentially Coherent Communication with Waveforms Chap. 7

bib(15/2) + nc
Vj = [VP + 0.5) T»)]
(7.123)
x [VPbi-1P (Th /2) + ne (i — 0.5) Ty)

+ ny ((i + 0.5) Ty) ns (G — 0.5) Th)

where from (7.121)
Tr/2 gin 27 Bt’ 2 2
P(Tp/2)
p(1,/2)=4B [ a ay 2= Sia
i( BI) b) = _— ob (1/2)
eye (7.124)

with S(x) the sine integral function defined by [30]

* sin
Si(x) af ee y (7.125)
On ae
Recalling from (7.115) that for B = 1/27), the filtered noise samples n, ((i + 0.5) Tp),
Ne (i — 0.5) Ty), ns (Ci + 0.5) Ty), and ns ((i— 0.5) T,) are statistically independent zero-
mean Gaussian random variables each with variance NoB = No/2T», then the decision
variable of (7.123) has the same form as that for the optimum DPSK receiver. Hence, we
can immediately write the bit error probability in the form of (7.14) where the energy per
bit is now degraded by (p(T;/2))’, that is,
P(E) -exp |-= E Ye cn/2)| = ;exp {0.765 (7.126)

Again comparing (7.126) with (7.14), it is a simple matter to see that, under the above
assumptions, the degradation in the performance of the suboptimum receiver relative to that
of the optimum receiver is a reduction in SNR by the factor 10 logy9 (4/27) Si? (1/2) =
—1.2 dB.
A second approach to approximately computing the bit error probability performance
of Fig 7.18b is to still make the assumption of neglecting ISI to simplify (7.122) but us-
ing a bandwidth B = 0.57/7,, which in [29] is shown to minimize ISI. In this case, the
noise variables n- ((i + 0.5) 7,) and n,. ((i — 0.5) T;) are still zero-mean Gaussian, each
with variance NoB =0.57No/Tp, but they are no longer independent and likewise for
ns (i + 0.5) Tp), and ns ((i — 0.5) T,). However, the noise variables n; ((i + 0.5) T,) +
Nc ((i — 0.5) Ty) and n- ((i + 0.5) Ty) — ne ((i — 0.5) Tp) are indeed independent and like-
wise for ns ((i + 0.5) Tp) + ns (i — 0.5) Tp) and ns ((i + 0.5) T,) — ns (Ci — 0.5) T). Thus,
remembering the analogy between the noise terms in DPSK and those in noncoherent FSK
as discussed in Section 7.1.6, and noting from (7.124) that p(7,/2) = (2/7) Si (0.577), we
can again immediately write down the bit error probability as
|
,(E) ==|
P(E) exp |="=Ep | 2
———_
se: ,
$2
i2( 1
(ara || se et Ey | (7.127)
en| 0.8052
which represents a degradation of 10 logy (2/0.5727) Si* (0.57) = —1.0 dB in SNR rel-
ative to the optimum DPSK detector.

7.4.3 Multiple Decision Feedback Symbol-by-Symbol Detector

A simplification of the optimum multiple symbol block-by-block detector discussed in Sec-
tion 7.2.1 is suggested in [31, 32] wherein the previous information phases, A@;,_;; i=
Sec. 7.4 Suboptimum Waveform Receivers 489

1,2,..., Ns — 2 in the metric of (7.62), are replaced by past decisions on these quantities,
that is,
N,;—2
7 = Ve_—-Nn,+1 + Ve exp |-;(asaE ye sin)
m=1
} (7.128)
N,—2 & Ns—i-—2
=- SS Ve; exp |-;(DS sii
i=] m=0

Since the only unknown in (7.128) is now the current symbol A@,, the detector is once again
of the symbol-by-symbol decision type. While suboptimum, this multiple decision feedback
type of detector should still improve bit error probability performance relative to that of
classical M-DPSK.
As a simple example of the above, consider a three-symbol observation interval (NV, =
3). The optimum decision rule for this case is given below (7.65), which for the suboptimum
detector now becomes
choose AG corresponding to

max Re 4V,V;*_,e J4% + Y_1V*_,e J4% 4 UVF. od (44-1+44)

Ab; el k-2 k-2

where Ab _1 represents the previous decision that is fed back. Since the middle term in the
braces doesn’t depend on Ag, it can be eliminated, resulting in the yet simpler decision rule

choose A@& corresponding to

ae NS — j (Ab +AO
max Re |Va see 8 Vi Vee j (Ab 1+ |
AG

The bit error probability performance of the multiple decision feedback receiver was
analytically computed in [32] for DPSK and 4-DPSK under the optimistic assumption of no
errors in previous decisions, that is, perfect knowledge of the previous transmitted symbols.
The results are given by an expression analogous to (7.68), namely

Pye) = 5[1- 0 (vb, Va) + 0 (va, vb)] (7.129)

where for DPSK

{?| RD [Ns +2VN, —1] (7.130)

roid (Oe 2No

and for 4-DPSK

b Ep
(21= [Ne 2(N, — | (7.131)

In the limit as Ns — oo both performances approach that of coherent M-PSK with differ-
ential encoding/decoding. This is the same conclusion as was reached for the optimum ML
block-by-block detector.
Figures 7.19 and 7.20 are illustrations of the above theoretical results for N, = 2, 3,
and 4. Also shown are two sets of simulation results; one that assumes perfect decision
 490 Differentially Coherent Communication with Waveforms Chap. 7

10-15
--=- THEORY
40g SIMULATION POINTS

10-2

ae
ne 10°F EQS. (7.111)
AND (7.112)

PSK

10-4

; Figure 7.19 Bit error probability

10° performance of the multiple decision
es Faa a LOSI estan Aye feedback symbol-by-symbol detector;
bias0! M =2 (Reprinted from [32] © IEEE)

feedback as does the analysis, and one that takes into account the effects of errors in the past
decisions, that is, the true situation. For the case of perfect feedback, the theoretical and sim-
ulation results agree exactly as one would hope. For the case of errors in the decisions being
fed back, the simulation results, of course, yield a poorer performance than the theoretical
results which assume perfect decision feedback. However, over the range of system param-
eters illustrated in Figs. 7.19 and 7.20, no error propagation was observed in the simulation.
Furthermore, the results are quite comparable to those of the optimum ML block-by-block
detector. For example, for Ns; = 3 and P,(E) = 10-5, the difference between the two is only
about 0.3 dB.

7.5 SYMMETRIC DPSK

A variation of conventional DPSK that is particularly useful in communication systems
possessing a nonlinear channel is symmetric DPSK (SDPSK). This modulation scheme,
which is described by Winters [33] and Chethik [34], produces a radian phase shift in
the transmitted carrier of +2 /2 or —z/2 (in constrast to 2 or 0 of conventional DPSK)
depending, respectively, on whether a logical 1 or 0 is to be sent. Mathematically speaking,
the transmitted phase 6; in (7.1) is still the differentially encoded version of the information
phase, that is, 6; = A@; + 6;-1; however, A@; now ranges over the set (—2/2, 1/2) rather
than (0, 7). As a result, the optimum receiver for SDPSK is similar to that for DPSK except
that the complex decision variable is found from a cross product rather than a dot product
of the complex I&D output and a T)-sec delayed version of itself; that is, the noise-free
decision variable for making a +1 data decision is proportional to sin AO; rather than cos
A@;. Analogous to Figure 7.2b, an I-Q receiver for demodulating SDPSK is illustrated in
Fig. 7.21.
Sec. 7.5 Symmetric DPSK 491

10-1
=-—- THEORY
, 400 SIMULATION POINTS

10-2

DETECTED
SYMBOLS
= FED BACK
Ww 49-3 FED BACK \
a2 EQS. (7.111) \ ea 4-DPSK
AND (7.113)

10-4

108 Figure 7.20 Bit error probability

01234 567 89 101112131415 performance of the multiple decision
E/N. dB feedback symbol-by-symbol detector;
a0 M =4 (Reprinted from [32] © IEEE)

By choosing an initial phase 6) equal to 7/4, the set of allowable values for 0; in
SDPSK is restricted to (1/4, 37/4, 52/4, 77/4) in every symbol interval and since the
magnitude of the maximum phase change is limited to 2/2, we see that either the I or
Q components of the transmitted signal, but not both, must change polarity from symbol -
to symbol. Thus, the envelope of the transmitted signal never undergoes a complete phase
reversal (change of z radians) as is typical of conventional DPSK. The upshot of this mod-
ification is that SDPSK suffers from negligible spectrum “regrowth” after being filtered
and passed through a nonlinear device. Specifically, laboratory tests [33] show that when
SDPSK and DPSK are bandpass filtered and then passed through a hard limiter, spectrum
regrowth for DPSK is nearly 100%, whereas for SDPSK it is significantly lower. On a lin-
ear channel, unfiltered DPSK and SDPSK have identical bit error probability performance
[see (7.14)] since in both cases the two possible symbols are separated by z radians. Fur-
thermore, the two modulations have identical power spectral densities. Finally, Winters [33]
shows that SDPSK produces a lower average bit error probability than DPSK in the presence
of timing error and frequency uncertainty.
Although the desirable attributes of SDPSK with regard to minimum spectral re-
growth after bandpass filtering and hard limiting were demonstrated almost a decade ago
[33], it has only recently been suggested by Alexovich, Kossin, and Schonhoff [35] to in-
clude the filtering and hard limiting as an inherent part of the modulation scheme itself.
Such a bandpass limited symmetric DPSK (BPL-SDPSK) produces a bandwidth efficient,
constant envelope modulation. By choosing the bandwidth-time product B7) equal to unity,
it was found by simulation that the power spectral density resembled that of minimum-shift-
keying (MSK) (see Chapter 2). Furthermore, the bit error probability performance of BPL-
SDPSK with B7; = 1 was shown to be comparable to and in some cases better than that of
various forms of differentially detected MSK [36], [37], [38], [39]. A detailed discussion of
MSK performance will be given in Chapter 10.
492
P@soo Sn

@vuts
(3)4

Voutsz7p—

[R00]
10}e][1IISO

ens1Z7°L UW C-] IATV90I

Jo} onawiwAs
MSdq
 Appendix 7A
Derivation of the Optimum Receiver for Classical (2-Symbol Observation)
M-DPSK
Consider the a priori pdf of the received signal given the transmitted message as given by
(4.8). This pdf is the basis for deriving all of the optimum waveform receiver structures
in Chapters 4, 5, and 6 and will be once again used here as the starting point for deriving
the optimum receiver for differential detection of M-PSK (M-DPSK) using an observation
interval of two symbols. For the purpose of the derivation of the optimum receiver, the
observation interval shall be taken as 0 < t < 27;,.
For M-DPSK, the transmitted signal waveform in the interval 0 < t < 27; correspond-
ing to message mj; can be expressed, analogous to (7.1), as

Sj(t; 0) =V2P cos (wet + 96;(t) +0); O<t <2T, (7A.1)

where @ is an unknown random phase uniformly distributed in (—z, 7) and

Gig; O<t<T;
an =|
i. Tere in Ose
IND

is the transmitted data phase waveform with 6;,, 0;, independently taking on values from the
set ofM possible phases B,, = 27/M; m=0,1,...,M —1. Since the signal now depends
on an additional unknown random variable, that is, 9, the conditional pdf of (4.8) is written
as

GG @ ORG) dt
f-(o1m,0) =Cexo|
No

2 aa) Dy Parte Pe
= Comp) Bans
ORT loisee
ne (7A.3)
No
a
en |
2 Jo ; p(t)
si
i; (t; 8)dt
(t; 6)

No

Since fo p (t) dt is independent of i and ie s; (t; 0) dt =2PT;, we can rewrite (7A.3)

in the equivalent form

oT,
Dia pt) sp 50)
at
fr (p ino) =Ciexp| POE (7A.4)
0

where C’ is a new normalization constant that is independent of i. Using (7A.1) and (7A.2),
the numerator of the exponent in (7A.4) can be expressed in the following convenient
form:
 Differentially Coherent Communication with Waveforms Chap. 7
494

OT; ge
2exp {j (wet + 6;(t) + 8)} ar|
/ p(t)s;(t; 0)dt = VP Re {| p(t)
0 0

Ts
= /PRe [ oervdexp {i loet + 6)} at
$$

Uy (7A.5)

OT,
+| p(t)V2exp {j(@et + 6;,)} dt |exp (j9)

U;

= VP |p + 01) cos (6 +a)

where

(7A.6)

Substituting (7A.5) into (7A.4) gives

DU) oie S
f-foim,0)=C'sxp| 20 lip + Osos +o (74.7)
Averaging (7A.7) over the uniformly distributed phase, 6, gives

f(oim)
=| fr (p mi.)
epi

fo(9)d0
alee ORR Te :
= 015[e227 ldo + ifm 0 +09}a (7A.8)

ON Rae D,
-'tn( |e + a)
No

where as before /o(x) is the zero order modified Bessel function of the first kind with
argumentx.
For equiprobable messages, the optimum (maximum-likelihood) receiver chooses that
message as having been transmitted corresponding to the maximum of f; (p |m; ) over 1.
In particular, letting m (r) = m; denote the decision made by the receiver as a result of
observing r, then since Jo(x) is a monotonically increasing function of its argument, we
have
App. 7A The Optimum Receiver for Classical M-DPSK 495

*s
i=max |— f,(p|m;)
= max atsdo + H1|
Sale
l Ll

Ts
= max! [ p(t) V2exp {7 (Wet + Gig)}dt (7A.9)
10,1]

8 2
+f p(t)/2 exp {i (@ct + 6;,)} dt
Ts

where max! g(i) denotes the value of i that maximizes the function g(i). We note that
Ll

adding an arbitrary phase, say 6,, to the decision statistic in (7A.9), that is,

Ts
i = max! / p(t)/2 exp {ji(Wet + Big + Oa)} dt
io,t} 0)

ote 2
+/ p(t)/2 exp {j (wet + 6), + Oa)} dt

Ts
= max ~' exp (ja) if
10,1] 0
p(t)V2exp {j (wet + Gig)}at
(7A.10)
hip
+f pia) VZexp{J (vet +6,)} dt
Ts
Ts
= max = if p(t)/2 exp {j (wet + Gi) }dt
10,1) 0

ye 2

+| p(t) 2exp {j (wet

+ 6;,)} dt
Ts

doesn’t change the decisions made on 6;, and 6;,. Thus, the decision rule has a phase
ambiguity in that the receiver is as likely to choose the pair (6j) + 0a, 61, + 9a) as it is to
choose the pair (Gi9, Oras To resolve this ambiguity, we employ differential encoding in
which the information (data) phase is the difference of two adjacent transmitted phases.
Letting Ad; A 6;, — 9;, and also setting the arbitrary phase 6, = —6io, (7A.10) becomes

Ts 2Ts 2

i = max! i p(t) V2exp joxt) dt + f p (t) V2 exp {j (wet

+ A6j)} dt
i 0 Ts
2

Ts 2Ts
= max dt +exp ae)
~' i! p (t) V2exp (jact) | p(t) V2exp (jact)dt|74-11)
i 0 Ts
Vo Vi

ye ait
= max "|| +h + 2Re[oVi'exp (7060 |
Ll
496 Differentially Coherent Communication with Waveforms Chap. 7

ee po Ac
Since Vo and v1 are independent of i, we get the equivalent decision rule

i =max—!Re {VV" exp (— 6)

l
(7.12)
Finally, letting Vo 4 Vo e/"0 and V; 4 i |
e-/™, we have

i =max~'Re {|Vo]|| exp in — no - 44D]

1

= max ~! Vo] LA GosGhee ng NO} (7A.13)

Ll

=min~" |n1 — no — AG;

Figure 7.1 is an implementation of the decision rule in (7A.13) corresponding to the 27,-
observation interval ((i — 1) 7, <t < (+ 1)T7,).
 Appendix 7B
Probability Distribution of the Phase Angle
between Two Vectors Perturbed by Additive Gaussian Noise
Consider the geometry illustrated in Fig. 7B.1. Here A,, Az and 1, ¢2 represent the am-
plitudes and angles, respectively, of two known signal vectors and x), y; and x2, y2 are the
inphase and quadrature components of the complex Gaussian noises n; , nz added to each
of these vectors. The noises are each zero mean and, in general, have different variances,
that is, var(x;) = var(y)) = oP and var(x2) = var(y2) = Ge! The resulting signal plus noise
vectors have amplitudes R; , R2 and phases 1, 62. The specific application of this geometry
to the problem at hand corresponds to the case where the signal plus noise vectors represent
the complex form of the integrate-and-dump outputs for the (i — 1)‘‘ and i" transmission
intervals, namely

Vit SVeu—vt J Vedat = Vii git Riele)

(7B.1)
Vi Voi + IVs = vi gua Roel?
We are interested in the pdf (actually the probability distribution function) of the modulo 27
difference wy = (B2 — B,) modulo 27, —7 < Bi, Bo <7.
Defining the parameters!
A® = (¢2 — ¢}) modulo 27 (7B.2a)

ued Ai, Aj nee Az. tA? Gas

ee Noe oe
1. The modulo 277 is to be interpreted so that the resulting A® lies within the 27 interval of interest. Also, A®
ordinarily includes the carrier phase shift w,7. However, this term is frequently eliminated by receiver processing
and thus we do not include it in Fig. 7B.1.

Figure 7B.1 Geometry for angle between

vectors perturbed by Gaussian noise
(Reprinted from [4] © IEEE)
 498 Differentially Coherent Communication with Waveforms Chap. 7

W=VJU2- V2 (7B.2c)

_ Etim) _& (yiy2) sis E {xiy2} _ E tyix2} (78.24)

0102 0102 0102 0102

then it is shown in [4] that

v2
Priwi<¥<
wy} = fu(w)dy
"1 (7B.3)
Ey =F Oh 1 veAD
~ | Fhe)
— FQ); A®
> 2 or AD
< Wy
where F (yy) is an auxiliary function that is periodic with period 27 and discontinuous at the
point y = A® with a jump discontinuity F (A&*) — F (A®~) = —1. Specific evaluation
of F(y) depends on the assumptions made regarding the relative amplitudes and phases
of the signal vectors and the correlation between the two noise processes as well as their
relative variance. Three different cases are considered in [4]. The results are summarized as
follows.

Case 1. Equal Signals (A® = 0, V = 0) and Uncorrelated Noises (r = p = 0)

siny [7/2 exp{—U[1 — cos ycost]}

WG a) We dt (7B.4)
4x J_n/2 1 —cos
y cost

Case 2. Unequal Signals and Uncorrelated Noises

W sin(A® — wv) [7/? exp{—[U — V sint — Wcos (A® — ) cost]}

F = dt(7B.5
) 4n —n/2 U —Vsint — Wcos(A®
— w) cost ( )

Case 3. Unequal Signals and Correlated Noises

FW) = ke as W
sin (A® — p)
ue _n/2, 4% LU —Vsint — Wcos (A® — wv) cost
(7B.6)
rsinw — pcosw
1—(rcosyw+psiny) cost

where
_ U-—Vsint
— Wcos(A®— wv) cost
E= Mee SW SOA eS (7B.7)
1 — (rcos yw + psin yw) cost
Note that for all three cases, the probability distribution function is expressable as a single
integral of simple (exponential and trigonometric) functions. Further simplications of the
integrands of F(y) are considered in [4] at the expense of complicating the integration
limits.

PROBLEMS
7.1 The optimum receiver for M-DPSK is illustrated in Fig. 7.1, which has the complex form of
Fig. 7.8. With reference to Fig. 7.8, the decision rule can be written as
Problems 499

Ad; = Bn

Hoc ~~ iB:
where n = min : {argMetre
sede ~ ~ 2m1
= min {arg1% — arg \%-1| = fi|; B= a
(! I i

An equivalent receiver to the above that makes direct use of the inphase and quadrature outputs
Vox and V;x is illustrated in Fig. 7.2a. The decision rule for this receiver is given by

A6; = Bn

where n = max
I
~! {(Vs.Vine + Vik Ver-1) cos Bj + (Wx Ven—-1 — Vek Vs4-1) sin pi|:

201
ie
(a) Show that the above two decision rules are completely equivalent. Hint: Note that the value
of i that results in ae {©;} is equal to the value of i that results in max {cos @;}.
(b) Show that the Peceiier of Fig. 7.2a reduces to that of Fig. 7.2b for ke DPSK (M= 2),
that is, they implement the same decision rule.
7.2 (Wideband IF Filter, Sampling Detector) A differential detector for binary DPSK is illustrated
in Fig. 7P.2. Assume that the IF bandwidth of the input bandpass filter is sufficiently wide that
the desired signal will pass through this filter undistorted and with negligible ISI. As such, the
output of the bandpass filter can be represented as

r(t) = V2P cos (@ct + O(t) + 8) +n (t) cos (Wet + 6) — n2(t) sin (wet + 8)

where 6@ is the unknown phase introduced by the channel, 0(t) = 6; is the transmitted phase in |
the i” bit interval i7, <t < (i+ 1)Tp corresponding to the information phase A@; = 6; — 6;—1, ‘
and n\(t), n2(t) are uncorrelated zero-mean Gaussian noise processes each with variance
o* = NoB. Here B is the equivalent lowpass bandwidth of the input bandpass filter. Also
assume that the transfer function of this filter is symmetric around its center frequency, w-,
and that the lowpass equivalent of this transfer function is denoted by H(w).
(a) Assuming that the lowpass filter merely eliminates second harmonic terms of the carrier, :
show that a sample of the output of this filter taken in the ith bit interval can be written in’
the form

z= >Re {ziz2}
l *

where

r(t) |Wideband
Lowpass
Bandpass
Filter
Filter

Figure P7.2
500 Differentially Coherent Communication with Waveforms Chap. 7

<1 = (V2P cos6; +m) + j (V2P sin6;+ n2)

2= (v2P cos 6;-) + n3) + j (v2P sin 6; 1 +n4)

with 13(t) 4 ni(t — Tp), n4(t) A no(t — Tp), and nj; 2, n3, n4 respectively denoting sam-
ples of the noise processes nj(t), 2(t), n3(t), n4(t). Also, we have further assumed that
Onlp== 2.
(b) Show that the correlation matrix of the noise samples 11, n2, n3, 4 is given by

PO yo @
— Ta Oe ln 0
[ninj] mee Oren te ea)
OR Onn Oana

where o7p = Su Io |H(@)? el@% da.

(c) Compute the mean, variance, and cross-correlation of the complex variables z; and z2
defined by

AVertuise (6 Dig eae eS 2

1 — = :
pz0\07 & 5 1 —Z)* (22 — 22); Pz = Pez + JPsz

7.3 For threshold decisions made on random variables, z, of the form given in part (a) of Prob-
lem 7.1, Stein [21] has shown that the average bit error probability P,(£) is given by

PE) = 5[1- 0 (vb,va) + @ (Va, vb)|- 5exp(—“F*) 1 (vad)

where Q (a, 6) is the Marcum Q-function, A = pcz/./1 — p2,, and

{?| _ 1 J Si + So + (8) — Sg) cos v + 24/8; 82 sin (O; — @2) sinv

ee ie Ni+ No+ (Mi — No)? +4NiNop2,

a S| + S2 — (S; — Sp) cos v — 24/S] Sp sin (O; — @2) sinv

Ni + M2 — JM — Np)? + 4NiN2p2,

2/8 S2 cos (©; — ©)

N\N2 (1 es p32)

with

Se
nei?
5 lex > NeSog;
2 af
Ox Sarg {z}=tan~*
Shek
——-; k=1,2
Mek

v A arg (p,}
=tan7! 2
Pz

Recognizing that P,(E) = 5 Pr {Re {z}z2} <0|A0; =0} + 4 Pr {Re {ztzo} > 0|AG; =z},
evaluate a, b, and A, and show that P,(E) simplifies to

B=eg @ih) ( Ng )f (1 — p)
Ravrcest2 Ey 1
Ey
‘ a Soi: 2 exp ( No am)
Problems 501

sample at t= iT, + T,

r(t) |Wideband | 7 (2)

Bandpass Lowpass Threshold
Filter Filter Decision

Figure P7.4

7.4 In addition to coherent and noncoherent detection, binary FSK can also be differentially de-
tected. The receiver structures are quite similar to those for DPSK, the primary differences
being: (1) the value of delay, Tz, used to produce the delayed demodulation reference should
be chosen relative to the frequency modulation index, h; and (2) a 90° phase shift should be
applied to the delayed demodulation reference. As such, consider the suboptimum (wideband
IF, sampling) differential detector for binary FSK illustrated in Fig. 7P.4.
Assume that the IF bandwidth of the input bandpass filter is sufficiently wide that the
desired signal will pass through this filter undistorted and with negligible ISI. As such, the
output of the bandpass filter can be represented as

F(t) = V2P sin (@-t + O(t) + 8) +n, (t) cos (wet + 0) + n2(t) sin (wet + 8)

where 9 is the unknown phase introduced by the channel, 0(t) = meat is the effective phase
modulation transmitted in the i" bit interval i7, <t < (i + 1)T) with a; denoting the +1 i®
data bit, and n;(t), n2(t) are uncorrelated zero-mean Gaussian noise processes, each with
variance o* = NoB. Here B is the equivalent lowpass bandwidth of the input bandpass filter.
Also assume that the transfer function of this filter is symmetric around its center frequency, @¢,
and that the lowpass equivalent of this transfer function is denoted by H(w). Finally, we point
out that for FSK, differential encoding is not employed at the transmitter; hence, the transmitted
sequence {a;} is indeed the information data sequence.
Assume that the lowpass filter merely eliminates second harmonic terms of the carrier
and furthermore that w,.Ty = 2kz (k integer).
(a) Show that, in the absence of noise, the sample of the output lowpass filter taken at t =
iT, + Tg is given by

whTy
eta Sin
b

(b) Although the noise at the input to the threshold decision device will not be Gaussian,
one would still expect to get the lowest average bit error probability when the separation
between the signal components of z corresponding to a; = 1 and a; = —1 is a maximum.
Using the result of part (a), show that the value of delay, Tz, for which this condition occurs
is given by

(2n + 1)T,
LE caer yore iW? Oa Pc

(c) Since the delay must be restricted to lie in the interval 0 < Tz < Tp, show that for h < 0.5,
there is no suitable value of delay such that the signal components of z are maximally
separated and thus the performance of differentially detected FSK will be inferior to that of
DPSK.
7.5 Consider the receiver of Problem 7.4.
502 Differentially Coherent Communication with Waveforms Chap. 7

(a) Show that, in the presence of noise, a sample of the output of the lowpass filter taken at
t =iT» + Ty can be written in the form
1
z= 5 Re {ziza}

where

ne (v2P sin6;+ ni)=, (v2P cos 6; +n)

mwha;T,
2= (V2P cos (6; — &) +n4) saul (V2P sin (6; — &;) +n) ; § 4 Se

with n3(t) 4 ny(t — Ta), na(t) 4 no(t — Ta), and nj, n2,n3, n4, 6; respectively denoting
samples of the noise processes nj(t), n2(t), n3(t), n4(t) and the phase process @(t) taken at
P=SvIES SS 16%,
(b) Show that the correlation matrix of the noise samples 11, n2, 13, 14 1s given by

0 0
1 p
[ninj]=o 0 0
SS)
SiS
= vo) Si 1
ies)

where o*p = ye ined |H(w)|* e/°"dw.

(c) Compute the mean, variance, and cross-correlation of the complex variables z; and z2
defined by

— ; ; 2
te A Wer siee: io

1 et =
p,0\07 & 5 — 21)" (22 — 22); Pz= Pez + JPsz

7.6 Using the specific results of Problem 7.5 in the general bit error probability results given in the
description of Problem 7.3:
(a) Show that for differential detection of FSK, the average bit error probability is given by

Ps(E)= 5[1- 0 (vb. va) + 0 (va, ve)

where

Gerilscr |
a 1 — p2

(b) Show that for p = 0, the average bit error probability reduces to

ace P
= = exp {-——
a 2 *P\” Now
which agrees with the comparable result for DPSK given in Problem 7.3.
Teil A special case of Problem 7.6 corresponding to h = 0.5 and a phase modulation that is contin-
uous in time from bit interval to interval is referred to as minimum-shift-keying (MSK). (This
modulation is discussediiymore detail in Chapter 10). In particular, the elicit phase modu-
lation transmitted in the i"bit interval iTy <t < (i+ I)T is now O(t) =a; oi + x; where x;
is a phase constant whose value is determined by the phase continuity requirement at the bit
transition instant t = i7). Consider a modulation a(t) starting at t = 0 corresponding to an in-
put data sequence ay = —1, a; = 1, a2 = 1, a3 = —1, ag = —1,as= —1,a6= 1, a7 = —1, that
Problems 503

8
is, a(t) = }° ajp (t —iT,) where p(t) denotes a unit rectangular pulse of width T,. Assum-
i=0
ing that noise is absent, zero initial phase, that is, 9(0) = 0, and a delay Ty = Tp, sketch the
following waveforms over the time interval 0 < t < 87:
(a) a(t)
(b) 6(¢)
(c) A(t — Tp)
(d) A@(t) =O) — Ot — Th)
(e) sin A@(t)
Compare sin A@(t) with a(t). How do they differ?
7.8 For maximum-likelihood block-by-block differential detection of M-PSK, the pairwise error
probability is given by (7.68) where from (7.62) we have

Aes |-;(ss 10m}

Ns=l Newi2

i=0 m=0

7 Nyx? N,—-i-2
= Ve_-n,+1 + SS Vai Exp |-;(ss so-im)t
i=0 m=0

N,—1 cs Ns—t—2 rd
ce Si Ves; exp |-;(ye sin] |
i=0 m=0

i N,—2 i N;—i-2 ne
= Veen + > Vy_—i exp |;(De sion)
i=0 m=0

(a) Compute the mean, variance, and cross-correlation of the complex variables z; and 22
defined by

( =o
Brae tien nore 1 a lee
1 — — ,
p,0102 & 5 —Z1)* (22 — 22); Pz = Pez + JPsz

Wherever possible, express your answers in terms of the parameter 6 defined in (7.70).
(b) The arguments of the Marcum Q-function in (7.68) are obtained from [see (6.65)]

a 1 |S a et SS
a DN, 1 — |p, vt
= (psi
where

1 ea peu
SA =ll2;
D
OAarg(z)=tan!
="; 4 =1,2
Mek

v 4 arg {p;} = tan 1 Psz

CZ

N, Sof =03

Using the results found in part (a), show that a and b are given by (7.69).
7.9 In evaluating the upper bound on the bit error probability performance of maximum-likelihood
block-by-block differential detection of M-PSK, the sum of Hamming distances required in
504 Differentially Coherent Communication with Waveforms Chap. 7

(7.83) must be evaluated. To accomplish this, we start by observing that 5 of (7.70) can be
written in the form
N,-1 i
S=1t Yel; oj A>) 56-N,41-1
i=l n=1

We are interested in determining the various possible solutions for the 66’s such that the maxi-
mum value of the magnitude of 4, namely, |5|max, is achieved. There are four cases to consider.
Case 1: All N, — 1 vectors e/%; i =1,2,..., Ns; — 1 must be collinear and equal to efen/M
Case 2: All Ny — 1 vectors e/@; i =1,2,..., Ns — 1 must be collinear and equal to e~/27/™.
Case 3: Any N, — 2 vectors e/“' must be collinear and equal to 1 and the remaining vector
must be equal to e/?7/™.
Case 4: Any N, — 2 vectors e/“' must be collinear and equal to 1 and the remaining vector
must be equal to e~/27/™,
(a) Show that the solutions for the 50’s to Cases 1-4 are as follows:

Case1: b0;2Niyo =") 0Oponpip =O; T= 3,4).Ne

Case 2: 60-42 =—5F; SOx-n,4i = 0; 1 =3,4,...,. Ne
Case 3:
Days Qn
dOk—N, +141 = ve 50x—N, 4142 = are 50—n,+i= 0;

baal oO) Peraeye Lav tee3 Liste aor estes eee IN tees
dl yen

20 (
sealSeales = OP b= 1 2y es Ne
Open

Case 4:
20 oA
JbOk—N, 4141 = Pg? SOk—N, +142 = i’ 504—N, +i =0;

eae
et On a CS aoe Ns ee eee

20 :
Ses ay a OO k= et eh == eee ee

(b — For the solutions of part (a) show that for a Gray code bit to symbol assignment, the
accumulated Hamming distances are as follows:
Case 1: w (u, a) = 1
Case 2: w (u, a) = 1
Case 3: w (u, a) = 2(N, — 2) +1
Case 4: w (u, @) = 2(N, — 2) +1
(c) From the results of part (b), obtain the result in (7.89) corresponding to M > 2.
7.10 For the special case of M = 2, show that Cases 1 and 2 of Problem 7.9 are one and the same
and likewise for Cases 3 and 4, that is, only half the number of solutions exist. As such, show
that (7.84) is now the appropriate solution for the accumulated Hamming distance.
7.11 For maximum-likelihood block-by-block differential detection of M-PSK, the conditional like-
lihood ratio is given by (7.58). In (7.60), the average-likelihood ratio was found by averaging
(7.58) over the pdf of the unknown phase, 6. The decision statistic was then found by max-
imizing this average-likelihood ratio over the signal sequence. Suppose now that we instead
 References 505

first find the maximum-likelihood estimate, bu L, Of @, then substitute this value into (7.58) to
produce a maximum-likelihood ratio, and finally maximize this ratio over the signal sequence.
(a) Find the maximum-likelihood estimate of 6 defined by

Our. A See aiy (v Is, 6)

Note that 6,7, will be a function of the signal sequence.

(b) Show that the decision rule obtained by maximizing fy . (v s,6u L) over the signal se-

quence § is identical to that obtained by maximizing the average likelihood ratio fy (v is)
of (7.60) over this same sequence.
7.12 Consider a DPSK system where the signal has a residual carrier component. As such, it can be
written in the form

s(t) = <V2P sin (»+ 6m ye by p(t —nT) + ”

n=—oo

where 6, < 90° is the modulation index, 6 is the unknown phase introduced by the channel,
p(t) is a unit rectangular pulse, and {b,} is a differentially encoded data sequence, that is,
bn = anbn—| with {a,} the true input data sequence. The signal is sent over an AWGN channel
and the receiver is as illustrated in Fig. 7.2b.
(a) Find the mean of the sample Z; under the hypothesis a; = 1 and a; = —1, that is, how does
the presence of the residual carrier affect these results relative to the case where it is absent
(Om 902)7
(b) Is it reasonable to expect that the receiver will work even in the presence of the residual
carrier? If so, what might you expect to choose for the decision threshold?
(c) How might you expect the value of 6,, to affect the system performance, that is, what is the
best choice of 6,,?
7.13 For the equivalent nonoverlapped signal set defined in (7.103), show that the signals have the
normalized correlation matrix given by (4.116). Hint: Write 5;(t) as 5,(t) = 5, (t) — 5/(t) in the
interval 0 <t < (#4) T; where

todff
K(t); O<t< (7) ii
5, (tf) =
06 (gh)nsts(Wh)t
and make use of the orthogonality properties of the 5;(t)’s over this same time interval.

REFERENCES
1. Fleck, J. T. and E. A. Trabka, “Error Probabilities of Multiple-State Differentially Coherent
Phase-Shift-Keyed Systems in the Presence of White, Gaussian Noise,” in Investigation of Digital
Data Communication Systems, J. G. Lawton, ed., Buffalo, NY: Cornell Aeronautical Laboratory,
Inc., January 1961. Available as ASTIA Document No. AD 256 584.
2. Park, J. H., Jr., “On Binary DPSK Detection,’ JEEE Transactions on Communications, vol.
COM-26, no. 4, April 1978, pp. 484-86.
3. Simon, M. K., “Comments on ‘On Binary DPSK Detection’,” JEEE Transactions on Communi-
cations, vol. COM-26, no. 10, October 1978, pp. 1477-78.
4. Pawula, R. F,, S. O. Rice and J. H. Roberts, “Distribution of the Phase Angle between Two
Vectors Perturbed by Random Noise,” /EEE Transactions on Communications, vol. COM-30,
no. 8, August 1982, pp. 1828-41.
506 Differentially Coherent Communication with Waveforms Chap. 7

2 Arthurs, E. and H. Dym, “On the Optimum Detection of Digital Signals in the Presence of White
Gaussian Noise—A Geometric Interpretation and a Study of Three Basic Data Transmission
Systems,” IRE Transactions on Communication Systems, vol. CS-10, no. 4, December 1962, pp.
336-72. Also TM 3250, Bedford, MA: Mitre Corporation.
. Bussgang, J. J. and M. Leiter, “Error Rate Approximations for Differential Phase-Shift-Keying,”
IEEE Transactions on Communication Systems, vol. CS-12, no. 1, March 1964, pp. 18-27. Also
SR-103, Bedford, MA: Mitre Corporation.
. Lawton, J. G., “Comparison of Binary Data Transmission Systems,” Proceedings of the National
Convention on Military Electronics, 1958, pp. 54-61.
. Cahn, C. R., “Performance of Digital Phase-Modulation Communication Systems,’ IRE Trans-
actions on Communications Systems, vol. CS-7, no. 1, May 1959, pp. 3-6.
. Weinstein, F. S., “A Table of the Cumulative Probability Distribution of the Phase of a Sine
Wave in Narrow-Band Normal Noise,” IEEE Transactions on Information Theory, vol. IT-23,
September 1977, pp. 640-43.
10. Lindsey, W. C. and M. K. Simon, Telecommunication Systems Engineering, Englewood Cliffs,
NJ: Prentice-Hall, 1973. Reprinted by Dover Press, New York, NY, 1991.
ike Lee, P. J., “Computation of the Bit Error Rate of Coherent M-ary PSK with Gray Code Bit
Mapping,’ JEEE Transactions on Communications, vol. COM-34, no. 5, May 1986, pp. 488-91.
. Korn, I. and L. M. Hii, “Relation between Bit and Symbol Error Probabilities for DPSK and FSK
with Differential Phase Detection,” to appear in the JEEE Transactions on Communications.
. Pawula, R. F. and J. H. Roberts, “The Effects of Noise Correlation and Power Imbalance on
Terrestrial and Satellite DPSK Channels,’ JEEE Transactions on Communications, vol. COM-31,
no. 6, June 1983, pp. 750-55.
. Rice, S. O., “Statistical Properties of a Sine Wave Plus Random Noise,” Bell System Technical
Journal, vol. 27, January 1948, pp. 109-57.
. Roberts, J. H., Angle Modulation, England: Peregrinus, 1977.
. Lee, J. S. and L. E. Miller, “On the Binary Communication Systems in Correlated Gaussian
Noise,” [EEE Transactions on Communications, vol. COM-23, no. 2, February 1975, pp. 255—
59.
If. Huff, R. J., “The Imperfectly-Timed Demodulation of Differential Phase Shift Keyed Signals,”
Report 2738-1, ElectroScience Laboratory, Department of Electrical Engineering, Ohio State
University, June 1969.
. Simon, M. K., “A Simple Evaluation of DPSK Error Probability Performance in the Presence of
Bit Timing Error,’ to appear in the February 1994 issue of the JEEE Transactions on Communi-
cations.
. Salz, J. and B. R. Saltzberg, “Double Error Rates in Differentially Coherent Phase Systems,”
IEEE Transactions on Communication Systems, vol. COM-12, June 1964, pp. 202-05.
. Oberst, J. F. and D. L. Schilling, “Double Error Probability in Differential PSK,” Proceedings of
the IEEE, vol. 56, June 1968, pp. 1099-1100.
. Stein, S., “Unified Analysis of Certain Coherent and Noncoherent Binary Communications Sys-
tems,” IEEE Transactions on Information Theory, vol. IT-10, no. 1, January 1964, pp. 43-51.
. Divsalar, D. and M. K. Simon, “Multiple-Symbol Differential Detection of MPSK,” JEEE Trans-
actions on Communications, vol. COM-38, no. 3, March 1990, pp. 300-08.
. Wilson, S. G., J. Freebersyser and C. Marshall, “Multi-Symbol Detection of M-DPSK,’” GLOBE-
COM °89 Conference Record, Dallas, TX, November 27-30, 1989, pp. 47.3.1-47.3.6.
- Helstrom, C. W., Statistical Theory of Signal Detection, England: Pergamon Press, 1960, p. 164.
References 507

25: Mackenthun, K., “A Fast Algorithm for Multiple-Symbol Differential Detection of MPSK,” sub-
mitted for publication in the IEEE Transactions on Communications.
26. Schwartz, M., W. R. Bennett and S. Stein, Communication Systems and Techniques, New York:
McGraw-Hill, 1966.
Dale Carlson, A. B., Communications Systems, 2nd ed., New York: McGraw-Hill, 1975, pp. 396-404.
28. Downing, J. J., Modulation, Systems and Noise, Englewood Cliffs, NJ: Prentice-Hall, 1964, pp.
184-87.
Pas): Park, J. H., Jr., “Effects of Band Limiting on the Detection of Binary Signals,” IEEE Transactions
on Aerospace and Electronic Systems, vol. AES-5, September 1969, pp. 868-70.
30. Abramowitz, M. and I. A. Stegun, eds., Handbook of Mathematical Functions with Formulas,
Graphs, and Mathematical Tables, New York: Dover Press, 1965 (originally published by the
National Bureau of Standards in 1964).
Sile Leib, H. and S. Pasupathy, “The Phase of a Vector Perturbed by Gaussian Noise and Differentially
Coherent Receivers,’ JEEE Transactions on Information Theory, vol. IT-34, no. 6, November
1988, pp. 1491-1501.
a2r Edbauer, F., “Bit Error Rate of Binary and Quaternary DPSK Signals with Multiple Decision
Feedback Detection,’ JEEE Transactions on Communications, vol. COM-40, no. 3, March 1992,
pp. 457-60.
335 Winters, J. H., “Differential Detection with Intersymbol Interference and Frequency Uncertainty,”
IEEE Transactions on Communications, vol. COM-32, No. 1, January 1984, pp. 25-33.
34, Chethik, F., “Analysis and Tests of Quadrature Advance/Retard Keying Modulation and Detec-
tion,’ Aeronutronic Ford Corp. (internal technical memorandum), August 1974.
35. Alexovich, J., P. Kossin and T. Schonhoff, “Bandpass-Limited Symmetric DPSK: A Bandwidth-
Efficient Modulation for Nonlinear Channels,’ MJLCOM’92 Conference Record, San Diego, CA,
October 11-14, 1992, pp. 35.4.1-35.4.5.
36. Davarian, F., M. K. Simon and J. Sumida, “DMSK—a Practical 2400 bps Receiver for the Mobile .
Satellite Service,’ Pasadena, CA: Jet Propulsion Laboratory, JPL Publication 85-51, June 1985. ~
Sie Simon, M. K. and C. C. Wang, “Two-Bit Differential Detection of MSK,’ GLOBECOM’S4
Conference Record, December 1984, Atlanta, GA, pp. 22.3.1—22.3.6.
38. Kaleh, G. K., “Differentially Coherent Detection of MSK,” JCC’88 Conference Record, Philadel-
phia, PA, June 12-15, 1988, pp. 23.7.1-23.7.5.
Shep Kaleh, G.K., “A Differentially Coherent Receiver for Minimum Shift Keying Signals,” JEEE
Journal on Selected Areas in Communications, vol. SAC-7, no. 1, January 1989, pp. 99-106.
 Chapter 8
Double Differentially Coherent
Communication with
Waveforms

In our discussion of differentially coherent communication in Chapter 7, we were quick

to point out that although this form of detection is quite robust in the presence of slowly
varying channel phase errors, it degrades quite rapidly when in addition slowly varying
channel frequency errors (offsets) are present. A simple example of such a channel is one
characterized by communication with a moving vehicle, in which case the frequency offset
is that associated with the well-known Doppler effect. As a reminder to the reader, the
problem with using classical! differential detection in an environment subject to frequency
offset is that the phase of the detector output complex signal contains a bias proportional
to the product of the frequency offset Af and the detector delay T. This bias produces an
asymmetry (the direction of which depends on the polarity of Af) in the decision between
the two equally likely messages (antipodal signals) which ultimately degrades the average
bit error probability performance of the receiver. Since ordinarily the detector delay T is
chosen equal to the reciprocal of the data rate, then equivalently we are saying that classical
differential detection as discussed in Chapter 7 will yield poor performance if the frequency
offset introduced by the channel is an appreciable fraction of the data rate.
One solution to the problem is to include in the receiver a means for estimating,
either continuously with a frequency tracking loop or at discrete time intervals with an open
loop technique, the frequency offset and using this estimate to cancel the phase bias in the
differential detector output. Clearly, the success of this technique depends heavily on the
accuracy of the frequency offset estimate, which is ultimately limited by how narrow one
can make the tracking loop bandwidth or how long one can smooth the open loop estimate.
These system parameters must in turn be commensurate with the rate of variation of the
frequency offset itself. An example of a mobile communication receiver that employs such
open loop Doppler estimation and phase correction is discussed in [1].
The focus of this chapter is on techniques that do not attempt to estimate the frequency
offset but instead modify the classical differential detection scheme in such a way as to make

1. Herein, we shall refer to this form of differential detection as first-order differential detection (recall that the
input information phases are encoded as a first-order difference of two successive transmitted phases) to distinguish
it from second-order and higher-order differential detection schemes, whose definitions will be introduced shortly.
 Sec. 8.1 Definition of Higher Order Phase Difference Modulation (PDM) 509

it insensitive to this offset, thereby maintaining the previously discussed advantage of imple-
mentation simplicity. The means by which this is achieved lies in encoding the information
data phases as a second-order difference phase process” and using a two-stage differential
detection process. The class of schemes that accomplish this goal are herein referred to as
second-order differential detection or more simply double differential detection [2,3]. When
specifically applied to phase-shift-keying, we shall refer to this class of schemes as double
differential phase-shift-keying (DDPSK or D*PSK).
From a historical perspective, the Russian literature [4-12] beginning around 1972
was first to introduce the notion of using higher order phase differences for modulation and
detection when the channel introduced higher order phase variations. Later in the 1970s,
Pent [13] contributed much additional insight to these ideas and indeed produced a result
that allowed for analysis of several extensions of the theory given in [2] and a comparison
with the work of [3]. Since much of the work in [4-13] is not commonly known in this
country, in fact the book (in Russian) by Okunev [2] is, to the authors’ knowledge, the
only textbook that even mentions higher-order differential detection, we shall make every
effort in this chapter to expose the reader to this body of literature, giving due credit where
appropriate.

8.1 DEFINITION OF HIGHER ORDER PHASE DIFFERENCE MODULATION (PDM)

Consider an M-ary communication system in which the information in the i" transmission
interval iT; < t < (i + 1)T, is conveyed by one of M phase angles. For example, for M-PSK,
these phase angles would take on values from the set {By =27m/M; m=0,1,2,...,
M — 1}. Let 6; denote the actual signal phase transmitted over the channel in the i trans-
mission interval. Then a first-order phase difference modulation (PDM-1) is one that en-
codes the input information phase, denoted by Aj, in accordance with the operation

AG; 4 6; — O;-1 (8.1)

namely, the phase difference’ between the current and previous transmitted signal phases.
Turning this relation around, we express the signal phase transmitted in the 7th transmission
interval as

6; = Ad; + 6-1 (8.2)

The operation in (8.1) may be viewed as the finite difference representation of the first
derivative of @(t) at t= (i + 1)7, and we saw in Chapter 7 that it is accomplished by
including a differential encoder between the information source and the phase modulator.
Suppose now that we encode the information phases with two back-to-back differen-
tial encoders. Since the ouput of the second differential encoder is to be as before, namely,
6;, then its input (also the output of the first differential encoder) is A6;, and thus the input

2. If A26; denotes the data information phase to be communicated in the i transmission interval defined by
iT; <t < (i+ 1)T,, then the actual phase 6; transmitted over the channel in this same interval is related to A?6; by
the second-order difference A26; = (6; — 0;-1) — (0:1 — 6;-2) = 9; — 20;—-1 + 6;—2. This will be expanded upon
later in the chapter. For now, it is sufficient to understand this simple definition of the term second-order difference
modulation.
3. Unless otherwise noted, it is assumed that the result of first and higher order difference and addition operations
performed on phase angles are evaluated modulo 277.
510 Double Differentially Coherent Communication with Waveforms Chap. 8

exp{ jA76;} exp{ jA6;} exp{ j0;}

Differential Encoder Differential Encoder

exp{ jA6,_)} exp{ j4;_1}

AG; = A@;_| +A76; => Aé; = AG,_; = A’6,

finite difference approx. to

first derivative of AQ; (t)

6; = 6,_; + A6; — 6; = 63-4 = A6;

0-201 +02 =A°6,

finite difference approx. to
second derivative of @(t)

Figure 8.1a Transmitter configuration for second-order phase difference modulation

exp{ jA°6;} exp{ j,_1}

exp{ j6;_2}
Figure 8.1b Equivalent transmitter to
Figure 8.1la for binary PSK

of the first differential encoder would be

A*6; A AG; — A6;—-1 = (6; — 6-1) — (-1 — 93-2)

(8.3)
= 6; — 26;-1 + 6-2
This defines a second-order phase difference modulation (PDM-2) and may be viewed as
a finite difference representation of the second derivative of @(t) at t = (i + 1)T,. The
operation in (8.3) is illustrated in complex form in Figure 8.la. Note that for binary PSK,
wherein the allowable information phases are 0 and zr, the term 26;— in (8.3) can only take
on values 0 or 27. Thus, for this case, (8.3) simplifies to

A*6; = 6; + 6-2 = 6; — 6;-2 (8.4)

where the latter equality in (8.3) again holds because the allowable values for 6; are only
0 and z. Figure 8.1b is an illustration in complex form of (8.4).
 Sec. 8.2 Structures Motivated by the Maximum-Likelihood (ML) Approach 511

exp{ j[O0+ 27Aft;]}} n;

Figure 8.2 Classical differential encoding/detection system based on two-symbol observation

Clearly the above ideas can be extended to higher (than two) order phase differences.
In particular, define the n“" order phase difference modulation (PDM-n ) by
n

Ales Soi)" (Ese = n

(8.5)
k=0

where, as in earlier chapters

ah ps a aos 8.6
@E- ess Se
is the binomial coefficient. The modulation of (8.5) is implemented by including n back-
to-back differential encoders between the information source and the phase modulator and
would be of interest in situations where one is attempting to mitigate the effects of an
(n — 1)-order phase derivative introduced by the channel. When applied to binary PSK, the
acronym D” PSK is used.
What is important in the above is that PDM-n is a modulation with memory in that
the current transmitted phase depends on the immediate n previous transmitted phases. This
property will be exploited later on in the chapter when consideration is given to optimum
detection schemes for such modulations.

8.2 STRUCTURES MOTIVATED BY THE MAXIMUM-LIKELIHOOD (ML) APPROACH

To understand in what sense Fig. 8.1b is an appropriate encoding for channels that introduce
frequency offset, consider first a classical differential detection scheme (see Fig. 8.2) that,
for a two symbol observation, was shown in Chapter 7 to be optimum (in the ML sense) for
a channel with constant (time-invariant) but unknown phase 6. We shall now demonstrate
how to augment this scheme to account for the presence of a channel frequency offset Af.
As before, the complex baseband signal transmitted over the channel in the i” trans-
mission interval is given by (assuming unit amplitude for simplicity)

Spel! (8.7)
and the corresponding received complex signal is
Fj es Ge O27 ATK) + ii; A piel Cine ren Af ti+ni) (8.8)

where nj; is a sample of complex Gaussian noise. From Fig. 8.2, we have that the output of
the differential detector in the same transmission interval is given by
512 Double Differentially Coherent Communication with Waveforms Chap. 8

Figure 8.3. Large SNR equivalent channel

for system of Figure 8.2

Rep = FP = (SieOHA) 4 tap)(Gf ye JOHAR) ty) A ele Ny ES)

5; a e/ Gi—G-1) = el AG. O'=27 Af (#— G21) =2r Aft;

(8.10)
Nee nyst_e JC +27 4fa-1) as wee ee + fiji,

For large SNR, we can neglect the term njn;_, and thus Ni can be approximately modeled
as a Gaussian random variable. Assuming that the frequency offset is constant but unknown,
then the phase © = 2x Af T, can be modeled as being uniformly distributed in the interval
(—z, 2). Hence, we can look at the combination of differential encoder, channel, and dif-
ferential detector in Fig. 8.2 as an “ equivalent” channel with additive white Gaussian noise
N, and random (but time-invariant) phase ©. This is depicted by the illustration of Fig. 8.3.
The problem of finding the optimum receiver for this equivalent channel with input e/4%
is now the same as that previously treated in Chapter 7, namely, apply a differential encod-
ing/detection scheme identical to that used for an AWGN channel with zero frequency offset
and random (uniformly distributed) time-invariant phase. For an N;-symbol observation,
the solution is again the maximum-likelihood sequence receiver derived in Section 7.2.1.
For N; = 2, we have the classical differential detector based on symbol-by-symbol deci-
sions which requires first-order differential encoding at the transmitter. This simplest case
is illustrated in Fig. 8.4. Finally, putting Figures 8.2—8.4 together, we see that for large
SNR, the optimum transmitter/receiver configuration equates to back-to-back differential
encoders at the transmitter (see Fig. 8.1a) and back-to-back differential detectors at the re-
ceiver (Fig. 8.5). The output of the receiver in Fig. 8.5 is from (8.8) and (8.9) given by

Rike, = (F774) (7i-17}_»)

os (oipi-1 ane eA ( a Gv anns Gal Ea)

Pi-1Pi-2 (8.11)

= pip? ire! \a+4°n)

which is clearly independent of the frequency offset Af as desired. Note that whereas con-
ventional DPSK schemes require that the channel phase be constant over two transmission
intervals, D?PSK schemes such as Fig. 8.5 require that the channel phase and frequency be
constant over three intervals.
If one reviews the literature, double differential detection schemes fall into two cat-
egories depending on whether their two stages are implemented serially or with a parallel
architecture. Furthermore, they can be classified according to whether they have an auto-
correlation demodulator (ACD) or inphase-quadrature (I-Q) type of first stage. Finally,
Sec. 8.2 Structures Motivated by the Maximum-Likelihood (ML) Approach 513

(asterisk on multiplier denotes complex conjugate multiplication)

Figure 8.4 Optimum transmitter/receiver for equivalent channel of Figure 8.3

piexpt6; +O+27A ft; + ;}

[ |
PiPi-1 EXP) jl 9; — O-1 + 22ST, +7; — Ni |
A@ i An;

Differential Differential
Detector Detector

[ |
PiP;-1Pi-2 ia} AG, - A0,-, +An; An;
A* 6; An; |

Figure 8.5 Optimum (large SNR) receiver for second-order phase difference modulation

depending on the specific implementation, the final detector output will, in the absence of
noise, either be entirely independent of the input frequency (the detection scheme is abso-
lutely frequency-invariant) or vary as a function of frequency within some specified small
frequency range (the detection scheme is relatively frequency-invariant). It is important to
emphasize that even though a detector is absolutely frequency-invariant according to the
above classification, its error probability performance will in most cases depend on the fre-
quency offset. Typically (as we shall see shortly), this comes about because of the depen-
dence of the signal x noise terms in the detector ouput on the frequency offset.
 514 Double Differentially Coherent Communication with Waveforms Chap. 8

n(t)

jae (at
iT,

Figure 8.6 Autocorrelation demodulator of DPSK (PDM-1) signals

We begin with an examination of well-known PDM-1 receivers (at least one of which
was presented in Chapter 7) with emphasis on their behavior and performance in the pres-
ence of carrier frequency offset. The primary purpose of this discourse is to expose the
configurations that will later represent the first stage of serial-type receivers for PDM-2
modulation.

8.3 AUTOCORRELATION DEMODULATION (ACD) OF BINARY PDM-1 SIGNALS

Consider the autocorrelation demodulator for binary DPSK illustrated in Fig. 8.6. The input
signal s;(t) in the i transmission interval iT}, <t < (i + 1)Tp is of the form

si(t) = Asin (@ct +0 + 6;) (8.12)

where w, is the radian carrier frequency, 6; is the differentially encoded information phase
in that interval taking on values 0, 2, and @ is the unknown phase introduced by the
channel (assumed to be constant over an interval of duration 27), sec). The sum of s(t)
and white Gaussian noise n(t) (with single-sided power spectral density No) are passed
through a bandpass filter (BPF) of single-sided bandwidth W Hz (B = W/2 is the equiv-
alent lowpass bandwidth) producing the signal x(t). Assuming, for convenience, that the
center radian frequency, w-o, of the bandpass filter is an integer multiple of the data rate,
that is, wo = MQ; 2 = 27 / Tp, m integer, then the actual carrier radian frequency in (8.12)
can be expressed as mw, = m&2 + Aw where Aw = 27 Af represents the radian frequency
uncertainty about the nominal carrier frequency. Note that the bandwidth of the BPF is
chosen commensurate with the maximum value of Aw. Assuming that the signal s(t) re-
mains undistorted after passing through the BPF (this is equivalent to assuming a wide
enough bandwidth W relative to the data rate 1/7) that intersymbol interference (ISI) can
be ignored), then expressing the filtered noise in a truncated Fourier series* in the interval
iT, <t < Gi +1)Tp, we have

4. This form of the noise expansion was used in [10] and in the context of this problem was first introduced
in [5].
Sec. 8.3 Autocorrelation Demodulation (ACD) of Binary PDM-1 Signals 515

F(t) = si(t) +24 (t) = Asin (mQt + Aat +6 +6;)

ky . (8.13)
+ So (ns1,¢ sin KQt + Awt +) + n¢1,4C08 (Kt + Aot + 6))
k=k,

where the number of harmonics in the sum is given by kz — kj + 1 = 2BTy, and the carrier
frequency integer, m, lies in the range kj < m < kp. Thus, for the results that follow, 2B 7;
is assumed to be an integer.
Similarly, in the interval iT, < t < (i + 1)Tp, the BPF output corresponding to the
previous interval is given by

F(t — Th) = sj_-1(t — Ty) + N2(t) = Asin (mMQt + Aw (t — Ty) + 6 + 6-1)

ka (8.14)
+ S> (12,4 sin (kQt + Act + A) + 1¢2,¢ 008 (KQt + Act + 6))
k=k

The quantities 51,4, M52,k, Nc1,k, %c2,k are independent Gaussian random variables with zero
mean and variance No/ Tp.
The ACD algorithm for the DPSK signal is to make the decision A6; on the dif-
ferentially encoded data phase AO; = 6; — 6;_; transmitted in the ith transmission interval
according to

ns (i+1)T,
d; = sgn / r(t)r(t — T,)dat (8.15)
iT,

where, as in previous chapters, sgn {-} denotes the signum (sign) function of its argument.
Using (8.13) and (8.14), the integral in (8.15) can be written as

(i+1)Tp A2Tp AT;

/ r(t)r(t — Tp)dt = 5 cos Ad; cos AwT7p, + M82, cos 6;
Tp

AT,
ar Msn cos 6;_; cos AwTp ;
8.16
—_ a Nelm cos 6;-; sin AwT,

k2
Tp
ae > a (ns1,KMs2,k + Nel ,kNc2,k)

Substituting 51,4 = Vsi,k + Vs2,k, Ms2,k = Vstk — Vs2,k> Nelk = Velk + Ve2,k, and Ne2,k =
Vel.k — Ve2,k into (8.16) allows it to be expressed as the sum of the squares of Gaussian-
distributed independent random variables (the v,’s and the v,’s). Arbitrarily assuming that
the data phase A6; = 0, then an error occurs when the integral in (8.16) is negative and the
error probability is given as
516 Double Differentially Coherent Communication with Waveforms Chap. 8

AoT,
Py CB) Pr {(m + Acos6;— 1 cos? aah

kz
Asin cl 2 2
+
Vel,m — ee Sue ys (214 vet)
k=k
k#m

<i (vam + Acos 6;—1 COS

;a (8.17)
2

kz
Asin Sot 2
“hE
Ve2,m + see) 48 SS (v4 + v2.4)

=Pr{aj <5}
where A; and Az have noncentral x2 distributions with 4B7, degrees of freedom. The
details of the evaluation of (8.18) are carried out in [12] with the following results. Letting
a = AwTh, that is, the ratio of radian frequency offset to data rate, then

1 E
PEC) = FBT, exp |- a (2sin? S + 00s? S)|

ie
x, (2fsin?g0
n! Me ro
1, (Roots) 0
!
m

(
kL
Bd
k—m )
(8.18)

where Ep = A*T;/2 is the energy in the signal and : is the combinatorial coefficient
defined in (8.6). When the frequency error is absent, that is, a = 0, then (8.18) simplifies to
a result previously obtained in [4, 11], namely
i
2BT,—1 (£2) 257-1
1 E, b (#) ay lk PE ORT Ea
PE) = saa ex? |-5"| > aries Ds a’ soar ) (8.19)

For 2BT» = 1, (8.19) becomes

P(E) = = exp (-3) (8.20)

which is equal? to the well-known result for optimum detection of DPSK [see (7.14)]. For
2BT, =2, 3, 4 we have

| E LE
P(E) |\2a7%,=2 = Ds
= exp (-<)
No
1usZ| (8.21a)
5. Strictly speaking, for 2B7;, = 1, the assumption that the signal s(t) remains undistorted after passing through
the BPF is not valid. Thus, (8.20) is optimistic for the configuration of Fig. 8.6.
Sec. 8.3 Autocorrelation Demodulation (ACD) of Binary PDM-1 Signals 517

ekp(E) |237)=3 Te P,(E) {

l2B7,=2 + exp ( =)
s Mo)
_B\|(8) 74(8)(7) 9125
(8.21b)
1 =) 3 (Ep Ih flap 2
=-exp| —-— }/1+-{— — |—
2 P( No + 3(2)+5(2)
P(E) |2e7,=4 = Pp(E) |287,=3

ey +30 (#)
+ exp (-3) (f) ve
3127 (8.2 1c)

1 ( =) 29) Ej 1 @ eng) YARNS

ete — he ee) 8 ee ee
2 No 64 \ No 16 \ No 384 \ No

which shows the degradation of performance caused by widening the input BPF bandwidth.
Again, these results are somewhat optimistic because of the zero ISI assumption, but im-
prove in their accuracy as 2B7), increases.
Figure 8.7 illustrates Pp(E) as computed from (8.19) versus E,/No in dB for values
of 2BT) equal to 1, 2, 4, 8, 16, and 32. Also illustrated are the approximate values obtained
from (7.118) of Chapter 7 corresponding to the analysis of this detector assuming Gaussian:
statistics at the output of the [&D. We observe that for small values of 2B7),, the Gaussian
approximation analysis yields very poor results relative to (8.19) over the range of P,(F)
between 107! and 10~°. As 2B7; increases, the Gaussian approximation analysis becomes
better and the range of values of E,/No values over which (7.118) tends to agree with the
exact (assuming zero ISI) result in (8.19) is increased. However, for any fixed and finite
value of 2BTp, it is clear from Fig. 8.7 that for sufficiently large E,/No, the results in
(7.118) and (8.19) deviate significantly from each other and thus it is inappropriate to use
(7.118) to predict asymptotic (E,/No — oo) performance.
An approximate account of the effect of frequency error on the noise immunity of the
ACD is suggested in [4] and is based upon replacing E,/No by Ep/No cos* a in (8.19) (the
result for zero frequency error), which yields

b-! | 5° cos* a 2BT 1 (4p 2 i 2BT—1

gg On ae ra
1 Ep 2 0 (ee ' (8.22)
P(E) = S357, exp cos a} » ; eae OD ie
= Vet

The advantage of this approximate expression is that, for computational purposes, one
does not have to compute an infinite number of terms. The degree of accuracy to which
(8.22) approximates (8.18) was investigated in [4] with the following results. For 2B7p <
10, Epy/No = 4 (6 dB), and a = AwT) < 45°, there is less than | dB inaccuracy in the
assessment of the E,/No requirement for fixed P,(£) using the approximate result in (8.22)
relative to that in (8.18).
 518 Double Differentially Coherent Communication with Waveforms Chap. 8

10°

10°

ioe
5

Figure 8.7 Bit error probability performance of ACD of PDM-1! signals based on exact
(zero ISI) and Gaussian approximation analyses

8.4 |-Q DEMODULATION OF BINARY PDM-1 SIGNALS

The optimum demodulator for binary DPSK with known carrier frequency is illustrated
in Fig. 7.2b and is repeated here in Fig. 8.8 for convenience of reference. As was shown
in Chapter 7, the error probability performance of this receiver is given by (8.20). It is
tempting to use this configuration in the presence of frequency offset despite the fact that the
integrate-and-dump (I&D) circuits are no longer matched filters under these circumstances.
We now examine the impact of this mismatch on error probability performance.
Let the signal plus noise in the i" transmission interval be expressed by

r(t)=s(t) +n(t) = Asin (mQt + Awt + 6 + 4;) + n(t) (8.23)

The inphase and quadrature demodulation reference signals are given by

rs(t)-=2snmQy rety=2cosmQr (8.24)

Hence, the I&D outputs (ignoring double harmonic terms) become

Sec. 8.4 |-Q Demodulation of Binary PDM-1 Signals 519

J, Ode
(i+1)7,

Hy

Figure 8.8 [-Q demodulator of DPSK (PDM-1) signals

(i+1)7, (i+1)T,
vie rind =A | cos (Awt +04 6;)dt+ N;
iTh iTp
(8:25a)*
sin Aol 1

and

(i+1)Tp (i+1)T)
Voi ih r(t)re(t)dt = A | sin (Awt + 6 + 6;) dt + N3
iTp iTp
(8.25b)
sin AwTp 1

= AT, ee EA ( [(2i — 1) AwT), + 20; + 201)a 7N3

AwTp
y)

Here N and N2 are independent zero-mean Gaussian random variables with variance No7p.
Analogous to (8.15), the decision algorithm here is
520 Double Differentially Coherent Communication with Waveforms Chap. 8

d; = sgn {Vc,i-1Ve,i + Vs,i-1Vs,i}

! 2
sin
= sgn AZT; (Aa |cos (6; — 6-1 ipA@Tp) + Nett (8.26)

Din HORN 7
2 sin 5
= sen {A 7, = cosa cos AG; + Neg¢
Z
where again a = Aw7) and we have furthermore made the observation that, for binary
phase modulation, sin Ad; = 0. Also, Neg is the effective noise consisting of signal x noise
and noise x noise components as follows:

sin 5 \ | saa
Neff = ATy a me COS > {(2i = l)a ia 26; =F 20)
2

1
+ N; cos (5
2 [(2i — 3) + 26,1 + 20]); (8.27)
1 1
+N4 sin (5[(2i — l)a + 26; + 201)+ N3 sin ( [(2i — 3)a + 20;-; + 201)

+ Ni N2 + N3N4

Here again; N2, N4, which represent the noises in V, ;-; and V.;—1, respectively, are inde-
pendent zero-mean Gaussian random variables with variance No7) and are independent of
N, N3. The effective noise in (8.27) can be written in a more compact form, viz.

sin 5
Nett = ATp | —~ }Nsxn + Nuxn (8.28)
D

where Nsxy iS a zero-mean Gaussian random variable with variance 2No7, and Nyx =
N,N2 + N3Nz4 is a non-Gaussian random variable that is uncorrelated with Nsyy.
Following the approach taken by Stein [14], the error probability of the decision
algorithm given in (8.26) can be shown to be
|
P(E) = 5 [1- 0(vb, va) +0 (va, vb) (8.29)
where

2E! 2E!
a= a sinta: b= on cos? a (8.30)

with

sin % \?
B, 8 Bs ( =) (8.31)
)

and Q(a, b) is as before the Marcum Q-function. This result was originally obtained by
Henry [15].
Sec. 8.4 |-Q Demodulation of Binary PDM-1 Signals 521

Note that for zero frequency error (a = 0), we have from (8.30) and (8.31) that a = 0
and b= 2E>»/No. Then, using the special cases of the Marcum Q-function as in (7.76),
it is easy to see that (8.29) reduces to (8.20) as it should. On the other hand, there are
many values of a, the smallest of these being a = 2/2 (Af = 1/47,), for which the error
probability goes to one-half. Even at Af = 1/67>, there is already a loss of 3 dB in SNR
relative to the case of no frequency error. This rather rapid deterioration of performance
with @ is a direct consequence of using I&D filters that are not matched to the inphase and
quadrature output signals when frequency error is present.
The error probability performance of the I-Q demodulator in Fig. 8.8 can also be
obtained in several other equivalent forms. Using a geometric approach [3] to be discussed
later in this chapter in connection with the demodulation of higher order PDM, the authors
obtained the following integral form which is identical to (8.29)-(8.31), namely

(8.32)
ih e Ey ( Eos a
x F € No
ANS n Or COs n ex exp SSS
No n} an

Also, replacing the error function and exponential functions by their expansions in terms of
modified Bessel functions of the first kind, namely

ape) [o-e) 2
erf(zcos¢) = a EY ye (—1)‘en ly (=)
4 k=0 Z
Ka (8.33a)
cos [(2k + 1)¢] _ COS [(2k — 1)¢]
ip eorex 2k —1
and

CO

exp (zcos¢) = ) >Ik(z) cos kp (8.33b)

k=0

the following equivalent series form is obtained [3]:

PAE | E,, =)
p(E) = > 2No exp No
> (8.34)
2 E’ E; cos [(2k + l)a]
—1)* pawl: |fal PE |e(iseaeSeas
. 2 | ” E (52) Se (=*)| 2k+1

Miller [16] obtained a similar and equivalent form to (8.32), which is given by
 522 Double Differentially Coherent Communication with Waveforms Chap. 8

perpee E, cos” a
PAB) Se Beato) |elk Mie alae )|67)
nie =) exp | St (oe

1 m/2 E!
—.,/E, erf > cos ncosa
ON No (8.35)

2E; Bam
— erf —® cos (n+ a) cos n exp (-= sin? ")dn
No No

It is interesting to study the limit of (8.29) in the limit of large E»/No. In particu-
lar, using the asymptotic forms of the Marcum-Q function given in (7.76), we obtain for
cosa > 0 (e.g., 0 < |a| < 2/2)

P,(E) a =< [ole1 + |sina|

(onan sinc
signel exp
:
ieee .
(—E;,(1—|sina|)) 8.36
— (8.36a)
2,/2m (5)
No

and for cosa < 0 (e.g., 7/2 < |a| < 7)

= me else |e E,
ie )
ee (jsina)) (1 — [sine exp(
~ mei. 36b

which for E,/No — oo becomes

0O- cosa = 0
Ep/No>co le Ccosa ra) (8.37)

8.5 COMPARISON OF ACD AND I-O DEMODULATION OF PDM-1 SIGNALS

Another way of visualizing the receivers of Figures 8.6 and 8.8 is to observe that the
ACD performs differential detection prior to “matched filtering” where the I-Q demodula-
tor performs “matched filtering” prior to differential detection. The quotations on the words
“matched filtering” refer to the fact that for nonzero frequency error, the [&D’s are not truly
matched filters, as mentioned above.
In the case of zero frequency error, the I-Q demodulator in Fig. 8.8 is the optimum
scheme and thus should give the best performance. Nevertheless, we have seen that, for
small values of 2B7), the behavior of the ACD of Fig. 8.6 closely approximates the opti-
mum performance under the assumption that ISI is ignored. The assumption of zero ISI is
a direct consequence of assuming rectangular pulses for the information modulation (i.e.,
constant phase over the symbol interval) at the output of the BPF in Fig. 8.6, and as such the
error probability performance results as given in (8.21) are only approximately valid with
the degree of approximation improving as 2 BT) increases.
If, instead of rectangular pulses for the phase modulation, one were to use a Nyquist
pulse shaping, that is, one that achieves zero ISI (see Chapter 9), then it might be anticipated
Sec. 8.5 Comparison of ACD and I-Q Demodulation of PDM-1 Signals 523

that the results of (8.21) would exactly apply. In [1], this was verified for the case 2BT; =
2 where a 100% excess bandwidth raised cosine transfer function was employed for the
overall (transmitter plus receiver) Nyquist filter response. Since, for an AWGN channel with
no ISI, the average error probability is minimized (with respect to the noise) by splitting the
total filtering (transmit filter, receive filter, and matched filter) equally between transmitter
and receiver (see Chapter 9), then for the above this can be accomplished as follows.
The transmit filter (see Fig. 8.9a) is a root raised cosine filter with transfer function

Ty. 2
Pay = V2T, cos“; lol < 7 (8.38)
0; otherwise

The input receive filter is a brick wall filter whose equivalent lowpass version has a transfer
function magnitude Pr(w) given by

Pr(w) = |
AE lol < F (8.39)
0; otherwise

Note that the single-sided bandwidth of this lowpass filter is B = 1/Tp.

The combination of Pr(w) and Pr(w) presents a root raised cosine characteristic to
the matched filter input, namely

ols. 2x
P(o) = Pr(w) Paw) = {0;Trees a otherwise
(OS 7, (8.40)
The impulse response of such an overall characteristic has the property that two of its
samples taken at intervals of 7/2 sec are equal and nonzero with the remaining samples
all having zero value, that is,

BG) Pla)
Th Tp ; (8.41)

T;
p (+ok +1) 2) =0; k=1,2,...
The matched filter (assuming zero frequency offset) for the pulse shape of (8.41) has
a transfer function whose magnitude is given by (see Fig. 8.9)

(62)Tp

Thus, we see that the cascade of (8.40) and (8.42) achieves the Nyquist pulse shaping
corresponding to a 100% excess bandwidth raised cosine transfer function and also the equal
split (except for a scale factor) of the total filtering between transmitter and receiver.
Since zero ISI is achieved by this implementation, the optimum detector (for zero
frequency error) would be to perform the matched filtering prior to the differential detection
524
1e ojdureg

=] ae
Ao
cle

—-1)9"¢
(yu
osu
(1)u

0}
[enusraszI

D}
10}09}0p

uy.
4 (1)= (4yu—1)!d"0"{
coo—=u

ye o[dures
€
=] Loe
be)

aniBG" YOO[g WeISeIP

JO oY} Wa\sks
 Sec. 8.6 Serial (Absolutely Frequency Invariant) Structures 525

(see Fig. 8.9b). When this is done, the error probability performance is indeed given by
(8.20). If on the other hand, differential detection is performed prior to matched filtering
(see Fig. 8.9c), then it can be shown [17] that the error probability is exactly given by
P,(E) loBT%=2 of (8.21) which corresponds to the case B = 1/Tp.

8.6 SERIAL (ABSOLUTELY FREQUENCY INVARIANT) STRUCTURES

8.6.1 Autocorrelation Demodulation (ACD) of PDM-2 Signals

When carrier frequency error is present, it is possible to design a differential detection-type

receiver whose performance is absolutely invariant to this error in the sense that the sig-
nal component of the demodulator output is independent of the carrier frequency. There
are several possible schemes for accomplishing this, all of which depend on encoding
the input information phases as a second-order phase difference modulation (PDM-2) as
in (8.3).
An ACD for binary PDM-2 signals in the presence of frequency error is illustrated in
Fig. 8.10 (originally shown as Fig. 5 in [8]). This configuration is a hybrid of Figures 8.6
and 8.8 in the following sense. The first stage accomplishes inphase and quadrature ACD of
the type illustrated in Fig. 8.6 to convert the frequency error Af into an equivalent phase
Af Ty, appearing as cos Af T, and sin AfTp degradations at the I&D outputs. The second
stage then is identical to Fig. 8.8 which, as before, is the optimum processing for signals
with unknown but constant phase.
This scheme and others of its type are difficult to afialyze because of the successive
multiplications of signal plus noise at the first and second stage I and Q multipliers. An
approximate analysis of the ACD in Fig. 8.10 is given in [10] by modeling the I&D outputs
as Gaussian random variables similar to what was done in Section 7.4.1. In particular, if the
decision algorithm is given by

: iTy (i+1)T,
d; = sgn i r(‘)r(t — Tat [ r(t)r(t — Tp)dt
(i~1)Tp iT

iT (i+1)Ty A ate (8.43)

+/ F(t)Foo(t — tat [ r(t)roo(t — Tp)dt
(i—1)Tp iTp

= sgn {Vc,i-1Ve,i + Vs,i-1Vs,i}

where f(t), 7(t — T,) are modeled as in (8.13) and (8.14) and roo9(t — Tp) is the Hilbert
transform of F(t — Ty), then the I&D outputs V..;, Ve,i-1, Vs,i, and Vs ;-1 are given by

V.,; = E,cos AG; cosa + Ni; Ve,i-1 = Epcos AGj-1 cosa + N2;
(8.44)
V; ;= Epcos Aé; sina + N3; Vs,i-1 = E,cos AG;-1sina + N4

where from (8.16) and analogous relations

526
(*1u-1)9"¢
ou
(1)u sea ahahaAt 1e g[dures
a| a =3 tee
(@)D FV OP;
[eHuUsrTass
eyed 10109}9p
201nos

tu (1)
=
cou
: —1)4d""¢
(4qu

aaNnSIy
qg6"g jUSTeAIND| [eoNeUIYJeUI
Jopow
JO OY) Wd}sAs
Ul QINSIY26°8
 ojdures
Je

JO poyney

|
jenuaiayjiq
1019919

|
-1)4d*v"{
(*yu
OM

4 (1)
=

9p
+
JOI[Y
ANSI
I¢6"g
payoieul
10}99]
JOLId
YM
[BNUSIAJJIP
YOO[G
WLISLIP
JO
Oy]
0}
W}sks
Cru—1)9"¢

527
528
SIL] I8¥JS | puodasa3e}S

f m),

UC1+1)
aans1yq
OT'g
= uonejauosoiny
Jow[npowlsp
Jo ASdzC (7-Wwdd)speusis payuriday)
wor[Z] © (AAAI
Sec. 8.6 Serial (Absolutely Frequency Invariant) Structures 529

AT, b AT)
i a 752.m cos6; + a Mtstam cos 6;_; cosa

AT; Th
oy ee Bel,m<08 6;—1 sina + > Sy (1s1,4752,k + Me1,kNc2,k)
k=ky
ATp ATp
No = a Ts3im cos 6-1 + a [ts2,m cos 6-2 COs a

ATp The
= a tem cos 6; 2 sina + > de (ns2,4Ns3,k si Nc2,kNc3,k)
k=k,
(8.45)
AT) AT» 4
N3 = ek cos 6; + a Msi.m cos 6;_1 sina

ATp lip ka
+ —-Nei,mcos6j—1 cosa + >2 }7 (nsi,krter,k — Nel, bs2,k)
k=k,

AT» AT), ’
N4= a Me3.m cos 6;-; + a s2m cos 6;_2 sina

ATs Tiek
Bite a Tem Cos 6;-2 cos a + oy ye (152,403,k oe Nc2,kN53,k)
k=ky
The noise random variables N;; i = 1, 2, 3, 4 are all zero mean with variance [10]

On OE
= ING |it oe (8.46)
é
M et E»/No
Furthermore, N; and N3 are uncorrelated and likewise for Nz and N4. However, the other
pairs are correlated with normalized correlation coefficients®

NiNz _ N3Nq4 __ cosa cos (A76;)

P12 os 2 - 5) . ’

on on 2 [1ca ft]
(8.47)
, NiNa — N2N3 _ sina cos (A76;)
Oe Gre me ay em aoe PRAY Fe
on oN 2 [i¥ rote |

For 2BT, > 1 and E;/No not too large relative to 2BTp,, Ni, N2, N3, and Nq can be approx-
imated by Gaussian random variables. Substituting (8.44) into (8.43) gives

d; = sgn{ E; cos AG; cos A6;-; + Ni Ep, cos AO;—1 cosa + N2E, cos AG; cosa

+ N3Ep cos AG sina + N4Ep cos AG; sina + NiN2 + N3N4} (8.48)

= sen{E; cos A*6; + ExNsxn + Nx}

6. This result was first pointed out by Shouxing Qu while at Technical University of Nova Scotia in a private
communication to the authors.
530 Double Differentially Coherent Communication with Waveforms Chap. 8

where Nsyy is a zero mean Gaussian random variable with variance 2az and Nvnxn =
N\N> + N3N4. The error probability may be evaluated using Stein’s approach [14]. In
particular, it is straightforward to show that

P(E) = ;[1- o(vb, va) + 0 (Vb, va)] - 5exp (-“=*) Ip (ab) (8.49)
where

1 BT aye
F (i 5 sin «(1 ar a)

For zero frequency offset (a = 0), (8.49) simplifies to

1 1 Bisa, E, Bin v=
Sova ee ®iy Bore I 8.51
foes |! ncaa |e |a ne) ren
The result in (8.51) differs from that given in [8], namely

=
nie) = 5000 2No
se (14 E»/No
ele ) (8.52)
since in (8.52) the correlations of (8.47) were not taken into account. It is felt that this
difference between (8.51) and (8.52) will be small over the region of parameter values where
the Gaussian assumption is valid.
It is difficult if not impossible to obtain an exact (even assuming zero ISI) expression
for P,(E), analogous to (8.18) and (8.19), for the ACD scheme of Fig. 8.10. Thus, we
cannot check the absolute validity of the Gaussian approximation results obtained in (8.49)—
(8.51). However, based on the results for PDM-1 in Fig. 8.7, we intuitively expect that for
sufficiently large BT,, over a range of P,(E) values of practical interest, the above results
can be used with a fair degree of accuracy. As a measure of some confidence, the authors in
[10] do compare the Gaussian analysis results (not accounting for the correlations of (8.47),
however) to those obtained from a computer simulation and show reasonable agreement (see
Table 8.1).
The PDM-2 demodulator of Fig. 8.10 can be generalized to apply to M-ary (M > 2)
D*PSK (M-D?PSK) with information phase angles f,, as defined in Section 8.1. The result-
ing configuration is illustrated in Fig. 8.11. Letting Vi = Vc; + jVs,i, then the appropriate
decision rule is:

Choose A760; = B; where k= max many (8.53a)

Sec. 8.6 Serial (Absolutely Frequency Invariant) Structures 531

Table 8.1 A Comparison of Simulation and Theoretical Results for

Error Probability Performance of ACD for Binary PDM-2 Signals

P,(E)
FT, E,/ No (dB) Simulation Results Eq. (8.49)

ys 8 4.03 x 1072 3.28 x 1072

2 10 7.67 x 1073 5.31 x 1073
2 12 6.37 x 10-4 2.90 x 1074
4 8 5.61 x 1072 4.56 x 1072
4 10 1.10 x 1072 TIS <AOe?
4 12 9.08 x 1074 4.39 x 1074
8 8 8.26 x 1072 725 < 105
8 10 1.85 x 1072 1.41 x 1072
8 12 1.64 x 1073 8.93 x 1074
16 8 1.37 x 107! 1.24 x 107!
16 10 3.80 x 1072 3.11 x 107-2
16 12 4.02 x 1073 2.58 x 1073

with

Ve = Re {&%} = Re {ViVive |
(8.53b)
= (VeiVei—1 + VsiiVs,i-1) COS Be + Wei Vs,i-1 — Vs,iWe,i—1) Sin By

Since the exact symbol error probability performance of differentially detected PDM-1
modulation with M signals and no frequency offset is given by (7.7), then by analogy with
the relation of (8.53) to the exact bit error probability when M = 2, that is, (8.20), the
symbol error probability performance of Fig. 8.11 in the presence of frequency offset can
be approximated by

-1
Es
(1re te)
BTs
[1 — cos 4,cos p|
ree Bees ties gis
sin 7/2 st {-2
P,(E) = —* Be (8.54)
Es od 1 — cos 47 cos

In (8.54), E; denotes symbol energy and 7; denotes symbol time. Furthermore, assuming a
Gray code bit-to-symbol mapping, the bit error probability, P,(E), would approximately be
given in terms of the symbol error probability by (4.135).
Finally, in Chapter 7 we also presented an easy-to-evaluate asymptotic (large E/No)
approximation to the symbol error probability performance of M-DPSK, namely, (7.12). For
the PDM-2 demodulator of Fig. 8.11, this would become

btcey 1) 1 Es BT, i. 1
eae ME ect i (1 = cos =
Be eaeoe N3 a E,/No a) - M>3(8.55
ee
532
AN AU SaJAN
I-ro,19is I-1's

ef *IK1+1) POL

.
wm),

“11+1)
| tad a7 at
te Sg

d Ror Bs I-W
/ Pees W gf_ bry Coeebarby

ainsi
["g
= Joye[np
uonrjauosoiny Jo owep WASdzd-W (7-Wdd)sjeusis payuuida
wouy
y)[Z]
© (AaaI
Sec. 8.6 Serial (Absolutely Frequency Invariant) Structures 533

8.6.2 |-Q Demodulation of PDM-2 Signals

An I-Q demodulator for PDM-2 signals is illustrated in Fig. 8.12. An equivalent’? demod-
ulator [3] which performs the double differencing operation directly on the detected phase
angle rather than working with the sine and cosine projections of it is illustrated in Fig. 8.13.
The exact error probability performance of this implementation® specialized to the case of
binary D?PSK was first obtained in [8] and [13] under the assumption of zero frequency
error with the result?

Pye ee (2? epee

1 HE Ep aie yy oy

aga €a ep ( 4a
[o-e) »
Gp? Ep
Teese Ve Ep
meet 8.56
ca {oe malo Cay

aa 2m+1/2 EP
2No ae 2m+1+4+1/2 Eb
2No

It is also shown in [8] that (8.56) represents a good lower bound on the performance of the
ACD of Fig. 8.11 in the presence of frequency offset when 2 BT; is small.
For M-DPSK (M > 2), an exact expression cannot be found. However, a union bound
on symbol error probability P;(£) in the form of (8.56) can be obtained with the result [13]

ney zt (2) o(-322) 0

3/2

y oe ied
K i {im( Te)+ Int (=)| (8.57 )
m=0

lie Es xn f{M—2
x |lom+1/2 2No + Iym4+141/2 2No cos ik a Pama

The bit error probability upper bound can then be obtained from (4.140) and (8.57).
The exact error probability performance of Fig. 8.13 (specialized to the binary D?PSK
case) in the presence of frequency offset was first carried out in [3] using a geometric
approach analogous to that taken by Arthurs and Dym [20] in evaluating the performance
of DPSK. The result obtained in [3] is in the form of a double integral and is given by

7. Figures 8.12 and 8.13 are equivalent (produce the same decision algorithm) in the same sense that their
counterparts for PDM-1 modulation, namely, Figs. 7.1 and 7.2a, are equivalent [19].
8. The actual configuration analyzed in [8] and [13] replaced the I&D’s with distortion-free (zero ISI) lowpass
filters whose outputs were sampled once per bit. However, due to the zero ISI assumption, the error probability
performance results obtained there as a function of the lowpass filter output SNR, 7 = A?/o? apply directly to
Fig. 8.13 if 7 is replaced by E,/No.
9. There is a discrepancy between the error probability results obtained in [8] and [13]. Based on a derivation of
(8.56) supplied to the authors by Mario Pent, we have concluded that the result in [13] is indeed the correct one
and this is the one presented here in (8.56).
534
n="A
a" nat eeACA AAS
Tees: [I-12 2,12 I's,

i
"i141
140)

ele
oi
g
Co

5W
i,

bys

=u
l=W

Solo
cebay

ainsi
ZI{"g
AANA

ee

O-] Joyw[npowsp
Jo WSdzd-W (7-Wwdd)speusis pojuriday)
wor[Z] © (AAAI
Bai
 535
poywtidoy)
(ASI
wosy
[¢]
©
sjeusis (7-Wd) ASdzC-W Jo Jowjnpowep O-] Jo uonRjuswedul aeusay ET" g ainsi]
-Wo TTS
bry Gy 6
tomes
‘ W
I
536 Double Differentially Coherent Communication with Waveforms Chap. 8

ae 3
5 ONT
E! E!

xf [ect [2 ¢08 2m — erf Vn Os” (8.58)

xX COS 7] COS N2 Exp (-£; sin? ni + sin? m}) dnidnz

with E defined as before in (8.31). Upon careful examination of the mathematical model
used in arriving at (8.56) in the absence of frequency offset and the characterization of the
I&D outputs used to obtain (8.58) in the presence of frequency offset, it becomes readily
apparent that the two results can be made equivalent if in (8.56) E, is replaced by E;,, that
is,

2
3 E;, (—1)™ E\, (E\,
-=— In |—— I —— (8.59)
aad ae Ie m=0
seh Ng oN

x 9 Il2m+1/2 Ey
oar + I12m+1+1/2 ae
2No

Indeed, the equivalence between (8.58) and (8.59) can be formally derived. Thus, we con-
clude that (8.59) is also an exact result for the error probability performance of Fig. 8.12 or
Fig. 8.13 specialized to the binary (M = 2) case.
Figure 8.14 illustrates the error probability as given by either (8.56) or (8.58) for the
special case of no frequency error (E ae Ep). Also plotted in this figure for the purpose
of comparison are the corresponding results for coherent BPSK and DPSK as given by
(4.55) and (8.20) respectively. We observe that using D*PSK instead of DPSK requires
approximately 4 dB more E,/No and thus a sizable penalty is payed for using a double
differential scheme when indeed no frequency offset is present. Figure 8.15 illustrates (8.58)
or (8.59) for various normalized frequency offsets A f7). Also shown for comparison are the
corresponding results for DPSK with frequency offset as given by any of the forms (8.29)—
(8.31), (8.32), (8.34), or (8.35). Here we see that the D?PSK scheme is quite insensitive to
frequency offsets over a large range of values, whereas the DPSK performance degrades
rapidly. Thus, it is natural to ask: At what value of Af 7), is D°PSK preferable to DPSK,
that is, where do the performance curves of the two intersect? The answer to this question
was obtained in [3] and is illustrated in Fig. 8.16 where the values of Af 7) at which
P,(E)|ppsk = Pb(E)|p2psx are plotted versus E,/No in dB. The region above the curve
corresponds to the situation where D?PSK would be preferable over DPSK and vice versa
for the region below the curve. Thus, for a given E;/No, Fig. 8.16 offers one an idea of
how large Af must be relative to the data rate 1/7, before having to resort to a double
differentially coherent detection scheme such as Fig. 8.12 or 8.13.
Sec. 8.6 Serial (Absolutely Frequency Invariant) Structures 537

-4

109.49,
(E)
-5

-6

0 Pras 6 8 10 12 1416
E,/No,dB

Figure 8.14 A comparison of the error probability performances of BPSK, DPSK, and
D?PSK for the case of zero frequency offset (Reprinted from [3] © IEEE)

Before moving on to double differentially coherent detection schemes that make

use of longer (than three symbols) observation intervals, we wish to point out that all of
the above techniques can readily be extended to n" order differential detection of PDM-
n modulation. For example, the scheme of Fig. 8.13 would employ n (rather than two
as shown) differential detectors with a corresponding error probability performance given
by [3]

|Oe iid uf | E; 2n+ 1

PoE) == f of erfc Ng 8 (1 = aM tn mat 5) «)
—1 —1
(8.60)

x |] f (i 10) ani
(el

where A”nn+1 is the n™ order differential operator applied to n, +1 as defined in (8.5) and
538 Double Differentially Coherent Communication with Waveforms Chap. 8

0.0

-0.5

: AER
\,

-1.0

(E)
10949,

-2.0

25 \
2 0 2 4 6 8 10 12
Ey/No, dB

Figure 8.15 A comparison of the error probability performances of DPSK and D°PSK
in the presence of frequency offset (Reprinted from [3] © IEEE)

10 a
1 ENS VE rae
ete, = Bes ;- —_——
F (10) 5 exp ( ae = cos (1 2 «)
ober 1-1
x exp Sate acres (8.61)

< et
|E Di)
e rer i) wi
= c0s(n
re 5 «)

is the pdf of n; given that the message selected for transmission corresponds to zero phase.
This result can be evaluated numerically for any n and reduces to (8.32) for n = 1 and (8.57)
join = 72.

8.6.3 Multiple Symbol Differential Detection of PDM-2 Signals

In Chapter 7, we introduced the notion of multiple symbol differential detection [21] which
used a multiple (more than two) symbol observation interval for improving the performance
of PDM-1 signals relative to that obtained with classical differential detection. Specifically,
a metric was derived that made a maximum-likelihood sequence estimate of N,; — 1 M-PSK
symbols based on an observation over N, symbol intervals. For N; = 2, it was shown that
Sec. 8.6 Serial (Absolutely Frequency Invariant) Structures 539

0.17
2
DPSK HALF-PLANE

0.16

AfTy
0.15

0.14

0.13

DPSK HALF-PLANE

0.11 _
sf Onde 2804 56 671 38 NON AODTIHID
19! 1401546
E,/No, dB
Figure 8.16 Frequency offsets that yield P,(E)|ppsk = Pp(E)|p2psx (Reprinted from [3] © IEEE)

this metric reduced to the classical differential detection algorithm whose performance is
given by (7.14) for M = 2 or (7.7) for M > 2.
It is of interest here to extend these notions to PDM-2 signals. Since the first stage
of a PDM-2 demodulator (Figures 8.10—8.13) in effect converts the frequency error Af to
an unknown phase shift 27 Af 7, for processing by the second stage, then, in principle, one
should be able to improve performance by replacing the second stage of the detector with
a multiple symbol detection algorithm analogous to that considered for PDM-1 signals in
Chapter 7. It is to emphasize that the resulting total configuration is not truly maximum-
likelihood; however, conditioned on the first stage being implemented in the classical
way for PDM-1 signals, the remaining portion of the receiver is maximum-likelihood as
before.
Mathematically speaking, applying multiple symbol differential detection to the ACD
in Fig. 8.11, the decision rule of (8.53) generalizes as follows. We assume that the in-
put phase A76;_ 1,41 now acts as a phase reference for the sequence A?@ = (A76;_y,+2,
A76;_ jets Wesco A76;_1, A’6;). Then, second order differentially encoding these phases for
transmission over the channel, the appropriate sequence decision rule is:

Choose A726 = A76 (8.62a)

540 Double Differentially Coherent Communication with Waveforms Chap. 8

where A76 is the particular vector, with components ranging over the allowable set {Bm =
27m/M;m=0,1,2,..., M — 1}, that maximizes
No k— 2 2
Ns~2 -j DD A°G-e—m
n=|Vien+i+ >) Vine (8.62b)
k=0

For example, for N; = 3 (one additional symbol observation), (8.62) becomes

Choose (A76;_1, A76;) = (A76;-1, A76;) that maximizes

F A a A235 2
n Ss V5 Vie F(A +4761) Ae Vai a

ss (8.63a)
= |Vi-2l? + 1Vi-1? + Vi? + 2Re {ViVjt ge H+ "4-0
[ViiVjige 1AM +2Re {Viviaje fa)
+2Re

or, equivalently

Choose A76; and A*6;_1 if

ise mh ee (8.63b)
Re {Vivit ye4 ht Vi=1 Vice 189-14 ViVi
oe 1 a a is maximum
An implementation of (8.63b) in complex form is illustrated in Fig. 8.17. It is to be em-
phasized that the enhancement of Fig. 8.11 with the multiple symbol detection algorithm
(as in Fig. 8.17) does not change its status as a demodulator that is absolutely invariant to
frequency error.
To determine the performance of multiple symbol detection of PDM-2 signals in the
presence of frequency error, we proceed along the lines of the approximate (Gaussian)
analysis leading to (8.52) mixed with the approach taken in [21]. In particular, noting that
(8.52) is identical to (8.20) if E,/No is replaced by the “effective E,/No” defined by

(=) -($) (14 AL fe (8.64)

No) ete \2No Ep/No
then for large 2B7); and binary PDM-2 signals, we have from (7.85) that

@), V sf:
PE ae eee ( v3] E exp {~(<2) i (8.65)

No eff

where again Ns — | denotes the number of simultaneous decisions being made. As N, >
00, (8.65) approaches

| E

(Oe one
Pp(E) S eee eS {-(=) (8.66)
No }ett

which, using the asymptotic expansion of the complementary error function [see (7.87)],
 541
(¢ = SN) UoNsa\ap JoquiAs o[dnynuw
yum speusis (7-Wdd) ASdzA-W JO Joye[npowlap uoHe[aUO90Ne JO WO} xgfduioy Lg ans
soe]
{“‘S}oy
St

asooyD

(i)
542 Double Differentially Coherent Communication with Waveforms Chap. 8

Differential Encoder Differential Encoder

Figure 8.18 Baseband implementation of

modified transmitter in complex form

becomes

E
P,(E) < erfe (=) (8.67)
No eff

Thus, in the limit as the observation interval (decision memory) approaches infinity, the
large 2BT}, large SNR behavior of the performance of an ACD demodulator employing
multiple symbol differential detection of PDM-2 signals in the presence of frequency offset
approaches that of ideal coherent detection of BPSK (with differential encoding) with an
effective E,/No given by (8.64).
The above can be extended to M-ary PSK, again using the results of Chapter 7 to-
gether with (8.64). Analogous to the relation between (8.65) and (4.206) as N; + o9, it
is straightforward to show that in the limit as the observation interval (decision memory)
approaches infinity, the large 2BT), large SNR performance of an ACD demodulator em-
ploying multiple symbol differential detection of M-ary D’PSK signals in the presence of
frequency offset approaches that of ideal coherent detection of M-PSK (with differential
encoding) with an effective E,/No given in terms of E;/No by a relation identical to (8.64).

8.6.4 Further Enhancements

The conceptual double differential detection scheme illustrated in Figures 8.la and 8.5 can
be enhanced (i.e., its performance improved) by choosing the delay in the second stage of
the transmitter and correspondingly the delay in the first stage of the receiver to be 27, rather
than 7; (see Figures 8.18 and 8.19). The reason for doing this is to eliminate the high noise
correlation in the receiver output caused by the sharing of a common noise sample between
the two stages of the receiver as in Fig. 8.5. This can be noted in Fig. 8.19 where the receiver
output signal consists of four different noise samples each spaced T;-sec apart. Despite the
fact that the transmitter no longer implements a true second-order difference operation, the
receiver output is still independent of the frequency offset present in the received input
and as such Fig. 8.19 generically represents an implementation of an absolutely frequency
invariant demodulator.
To see the amount of improvement in performance gained by the above modification,
consider Fig. 8.12 or 8.13 (assuming binary modulation, i.e., M = 2, T; = Ty) where the
Tp-sec delays in the first stage are now 27)-sec in duration. The exact performance for this
modified implementation of Fig. 8.12 can be obtained by the method employed in [13] with
the result
 Sec. 8.6 Serial (Absolutely Frequency Invariant) Structures 543

Differential Differential
Detector Detector

(asterisk on multiplier denotes conjugate multiplication)

Figure 8.19 Baseband implementation of
At output of second differential detector, the Doppler has been removed modified receiver in complex form

Pik)
Ve LE" See
2E'
a 2.8 Ge) exp ( aa
(8.68)
lo) / ! 4
(-—1)” E, E,
if — I,
ésm=0
D3 2m + 1 , (2 et (+)

Alternately, using the geometric approach taken in [3], the bit error probability can be
expressed in integral form as

No cos cos (n3 + 2 — m1) |COS N1 COs 72 COS 3

— i
: te
2, era paar
No Ge lamse e
(8.69)

x exp — Ey (2 ni + sin“
<2(sin <2) n2 + sin
Pay) ns) 1 + erf E,
ene dnidn2dn3
0 0

Figure 8.20 plots P,(E) of (8.59) and (8.68) versus E;,/No in dB. The large improvement
of the configuration using a combination of 7,-sec and 27)-sec delays over that using T;-
sec delays in both stages is clearly evident. The same kind of improvement in performance
can be obtained with an ACD, for example, modifying the autocorrelator (first stage) of
Fig. 8.11 to have a 27)-sec delay. Figure 8.21 illustrates simulation results for the error prob-
ability performance of Fig. 8.11 and its modification as per the above. Over a wide range of
values of 2BT), the advantage of a 7,, 27) delay combination is strongly demonstrated.
Finally, for M > 2, a union bound on the symbol error probability of the modified
version of Fig. 8.12 described above can be obtained in a form analogous to (8.57) with the
result
n (E‘\* PEON 2 aye E' ee 2he
Ds ASE (Balt pa a ASE ig a Og Ey s
bee 4 (5) exp ( No )2s 2m + | |‘i(A paint (ae)
Way (8.70)
, nz M-—-2
x 008 {2m 4 3 (=F )

The structures of Figures 8.18 and 8.19 are actually special cases of a form of D?PSK
544 Double Differentially Coherent Communication with Waveforms Chap. 8

Ty» 27, DELAY

COMBINATION - EQ. (8.68)

Te Tips DELAY,
COMBINATION -
EQ. (8.59)

P(E)

6 7 8 9 10 11 42 13 Figure 8.20 Evaluation of Equations

(8.59) and (8.68) as a function of E,/No in
E,/N,, dB dB (Reprinted from [2] © IEEE)

referred to as nonadjacent D?PSK (NA-D2PSK) [22]. To see this, we recognize that, anal-
ogous to (8.3), the input, A26;, of the first differential encoder in Fig. 8.18 can be written
in terms of the transmitted phase, 6;, as AiO; = (0; — 0;-2) — (0;-1 — 9-3) = (6; — 8-1) —
(6;-2 — 6;-3) = AO; — AO;—2, which is recognized as the difference of two first-order phase
differences that are separated from each other by a single symbol interval. In the more gen-
eral case, if one replaces the 27; delays in Figures 8.18 and 8.19 by LT, delays (L > 2),
then the input, Me of the first differential encoder and the transmitted phase, 6;, are
related by AGO: = (6; — 0;-1) — (6;-L — 9;-L-1) = AO; — AG;_,. Here the two first or-
der phase differences are separated by (L — 1) symbol intervals. With reference to Fig. 8.13
ig oeriately modified by replacing the first 7,-sec delay by an LT;-sec delay), a deci-
sion on Bee ,9; would be made by comparing the phase difference AD ini= (Ni — Ni-1) —
(ni-L — Ni-L—1) = Ani — Ani—z with all possible transmitted phases around the circle and
choosing the one that is closest. Since for any L > 2, the random variables An; and Anj—
are statistically dependent, then from the standpoint of error probability performance on the
AWGN, there is nothing to be gained!? by choosing a value of L larger than 2; specifi-
cally, the bit error probability performance of NA-D?PSK is given by (8.68) or (8.69) for
all L > 2. Moreover, Qu [22] shows that asymptotically as E,/No becomes large, the dif-
ference in error probability performance between D*PSK (L = 1) and NA-D2PSK (L > 2)

10. For fast fading channels, it is shown in [22] that the correlation of the fading process can be utilized to
improve error probability performance by choosing L larger than 2. In fact, depending on the fading channel model,
e.g., satellite-aircraft, land-mobile, etc., an optimum (in the sense of minimum error probability) value of L can be
determined.
Sec. 8.6 Serial (Absolutely Frequency Invariant) Structures 545

—— T,,27,, DELAY COMBINATION

; i ease I,» T,, DELAY COMBINATION

“."s, 2B= IF BANDWIDTH OF BPF

ns “ie 2BT, =8

RE)

Figure 8.21 Computer simulation results

for the error probability performance
of Figure 8.10 and its modified version
(Reprinted from [2] © IEEE)

is attributed entirely to the removal by the latter of the correlation between the two first-: .
order phase differences making up the decision variable. Specifically, the asymptotic E,/No
degradation of D?PSK relative to NA-DPSK is given by

(1 am DAR IMAC DORSE

A Ep/No (dB) = 101og19 (8.71)
( * Pan) lpopsK

where

py E {Ani Ani-v}
ae (8.72)
[An

is the normalized correlation between the two first order phase differences making up the
decision variable. Since for NA-D?PSK (L > 2), PAn = 0, whereas for D2PSK (L = 1)

2
E{(ni — ni-1) (ni-1 — mi-2)}_ —E{-njy} 1
eis 2a? mB gI in ae ot
then from (8.71) we find that AE,/No (dB) = 10 log; (1.5) = 1.77 dB, which is in full
agreement with the numerical results of Fig. 8.20. The asymptotic E,/No degradation of
D?PSK relative to NA-D?PSK as given by (8.71) is also valid for M > 2 [22].
 546 Double Differentially Coherent Communication with Waveforms Chap. 8

8.7 PARALLEL (RELATIVELY AOI SIEN INVARIANT) STRUCTURES—

MULTIPHASE ACD
In an effort to improve the noise immunity of autocorrelation demodulators of PDM-2 sig-
nals with frequency error, while maintaining to some degree their invariance with respect
to the actual signal frequency itself, attention is now directed toward a class of relatively
invariant ACD’s. These ACD’s process the received PDM-2 signal plus noise in a parallel
bank of L ACD’s similar to Fig. 8.6, each of which now introduces an appropriate phase
shift 6) = (1 — 1)/L; 1=1,2,..., L between the direct and delayed signals before mul-
tiplication (hence the name multiphase ACD). The purpose of this phase shift is to try to
match the ACD to the equivalent phase shift a = AwT>) introduced by the frequency error
in the input.
One configuration of this type is illustrated in Fig. 8.22 and was first suggested in [4,
7]. The outputs Vis 1=1,2,..., L from the I&D’s corresponding to the ith transmission
interval are squared!! (to remove the modulation) and compared to Sige pees the maximum.
Letting i denote the value of / corresponding to the maximum (Ag) , then only the pa

switch is closed and as such the /'" ACD output, namely, a” (which represents a hard
decision on the first-order phase difference A6; = 6; — 0;—1) is passed on to the differential
decoder which performs the second stage of differential detection. Note that the range of
values for the L phase shifters in Fig. 8.22 only covers the interval (0, zr) [rather than (0, 277)
as one might first expect]. The reason for this is that a frequency error-induced phase shift
of @ in the region (zr, 277) will result in a 180° phase shift at the I&D output of the ACD
“matched” to the phase shift a — 2. However, since the “matched” ACD is determined
by the square of the I&D output, the ACD corresponding to the phase shift a — 7 will
ideally be selected. Furthermore, the 180° phase shift at the I&D output of this “matched”
ACD, which results in an inversion of the hard decisions at the “sgn” device output, will be
resolved by the differential decoder.
Intuitively, one would anticipate that the above scheme would yield an improvement
in noise immunity over the absolutely invariant type of ACD since: (1) first stage demodula-
tion is achieved by only a single ACD that is “matched” (within the resolution of the number
of ACD’s used in the bank) to the input frequency error; and (2) the second phase difference
is calculated from the adjacent hard decision outputs of the first stage as is done in coherent
detection with differentially encoded input symbols. As such, we would expect that as the
number, L, of ACD’s in the bank increases and the probability of correctly selecting the one
that matches the input approaches unity, the noise immunity of Fig. 8.22 should approach
(within approximately a factor of two due to the additional differential encoding operation)
that of an ACD of PDM-1 signals with no frequency error (Fig. 8.6). Practically speaking, it
has been shown [7] that a bank of L = 6 ACD’s gives performance quite close to the above
limit. This and other conclusions will be demonstrated by the results of the analysis that
follows.
Assume that the signal of (8.12) is received (in noise) where 9; is now the double
differentially encoded information phase in the i transmission interval. The output of the
multipler in the /th ACD is (in the absence of noise)

e)(t) = A* cos (AwT) + 6; — 6;—1 — 61) = A’ cos (A6;) cos (aw — Bi) (8.74)
11. Equally, one could replace the square-law devices with absolute-value devices (full-wave rectifiers). The
choice between the two would depend on the SNR as is typical in situations of this type.
 547
s[eusis (Z-Wdd) 3Sdzd-W $0 Jowjnpowsep uonejasosone oseydnjnu y 77"g aaNSLy
,S09
‘@V
1
548 Double Differentially Coherent Communication with Waveforms Chap. 8

where we have again set « = Aw7}, and recognized the fact that for binary modulation,
Ad; = 6; — 9; only takes on values 0 and 7. After passing through the I&D, we get the
(noise-free) decision variable

V.° = AT; cos (AG;) cos (a — By) = Epcos (A6;) cos 1 Gre)
where ©; = a — f; denotes the phase error associated with the mismatch of the i“ ACD
to the input signal. Assuming that based on a comparison of the squared values of Vo the
correct ACD (the jth corresponding to the one whose value of ; is minimum) is always
chosen,!* then the hard decision outputs from the correct ACD are described by

d\? = sgn |A°7,cos (AG;) cos (w — Bj) + N

: ”}
(8.76)
= sgn 2 cos ®; cos (A6;) + ny?

where Ne? is a noise variable that accounts for S x N and N x N degradations. More
important, however, is the fact that the correct choice among L ACD’s reduces the range
of the phase error, ®;, to the interval (—2/2L,7/2L). As such, the 1&D output signal for
the correctly selected ACD will dip by no more than a factor of cos 1/2L relative to that of
the same ACD of PDM-1 signals in the absence of frequency error. For example, for L = 6,
this dip is less than 4%.
The error probability associated with the hard decisions a is obtained by noting that
(8.76) represents the decisions in Fig. 8.6 if the energy E, is replaced by Ep cos ®;. As such,
this error probability, now denoted by P,)(E |®;) to indicate its dependence on ®;, is given
by (8.19) with E,/No replaced by (E,/No) cos? ®;. For example, for the special case of
2 BT, = 2, we would have

1 Ja es 1Ep 4
Poi(E |®;) = 5 XP (-2 Cos ®i) af aN cos” ®; (8.77)

The relation between the error probability, P,;(E |®; ), associated with the hard de-
cisions d; on the first-order phase difference A@; and the error probability, P,(E |®;),
associated with the ultimate hard decisions on the second-order phase difference A206; =
AG;
— AG;—1 18

Pp(E |) = 2Ppi(E
|®;) (1— Poi
(E|®;)) (8.78)
Finally, the average probability of error of the receiver of Fig. 8.22 is given by
n/2L
P,(E) = / —7/2L
Pr(@;) f(@))d®; (8.79)
where f(®;) is the pdf of De.
Assuming that the frequency error is completely unknown and as such the equivalent
phase a = AwT) is uniformly distributed in (—z, 7), then the pdf of ®; is given by

IE 4 4
@:) led et Opa
f(%)) {a otherwise Sey
12. Practical conditions under which this assumption is valid are discussed in [4, 7].
 Sec. 8.7 Parallel Structures—Multiphase ACD 549

RE)

E,/N,, dB

Figure 8.23 Bit error probability performance of the multiphase ACD in Figure 8.22
with the number of ACD’s in the bank as a parameter (Reprinted from [2] © IEEE)

Thus, for example, the average error probability performance of the receiver of Fig. 8.22 for
2BT} = 2 would be found from!

A ee Ep ts a et
P,(E) = — exp cee ®; MN ees ®;
HM J—n/2L
(8.81)

x Wes
— —exp |——Sree cos* co 9; lie
——Eeeeeaae
cos* el
®; aw ‘
2 P No l 4 No l l

Evaluation of (8.81) in closed form is difficult if not impossible. Nevertheless, we observe

that as L gets large, we can approximate the integral by the value of the integrand at ®; = 0
multiplied by the interval of integration 7/L. Thus, for large L, we get

P,(E)~& exp |-—Eb Nilah-—Bibel dy — 4d ~exp (|——WO ps—— (8.82)

4 aA Ny 4 No OPS Ne) eed Ni
which is approximately twice the first result in (8.21). Figure 8.23 is a plot of P,(E) versus
E,,/No (in dB) as determined from (8.81). Also shown is the limiting result of (8.82) which
corresponds to L approaching infinity. We observe that a value of L = 6 gives performance
within a few tenths of a dB of the limiting result of (8.82).
In addition to the above analytical results, a computer simulation has been written
to assess the error probability performance of Fig. 8.22. Numerical results obtained from

13. References [4, 7] incorrectly evaluate this error probability since they first average (8.77) over the pdf of ©;
and then use a relation analogous to (8.78) to convert this result to the final error probability. Also, Figure 1 in [7]
is incorrectly drawn. Figure 4.14 in [4] is the correct figure.
 550 Double Differentially Coherent Communication with Waveforms Chap. 8

this simulation for the case 2B7), = 2 and L = 6 lie within 0.1 dB of the corresponding
theoretical curve in Fig. 8.23.
Finally, we point out that the configuration of Fig. 8.22 can be clearly generalized to
apply to M-DPSK (PDM-2) signals with M > 2 in the same manner that Fig. 8.6 generalizes
to the first stage of Fig. 8.11. The details are left to the reader.

PROBLEMS
8.1 A binary phase information source whose phases take on values (0, 77) is to be communicated
using a second-order phase difference modulation (PDM-2).
(a) If the input information sequence {A76;} is 2,7, 7,0, 0, 7,0, 7, find the sequence of
complex baseband signals {e/%} transmitted in the intervals (0, 7;), (Ts, 2Ts), -.., (9Ts,
107,). Assume that the first two transmitted phases are given by 6) = 6; = 7 and since the
source is binary, 7; = Tp.
(b) If now the information source is 4-ary with phases that take on values (0, 1/2, 7, 32/2),
repeat part (a) for the information sequence 1/2, 7, 2, 0, 32/2, 1/2, 0, 37/2.
8.2 A binary phase information source whose phases take on values (0, 77) is to be communicated
using an n'*-order phase difference modulation (PDM-n). Show that the appropriate encoding
scheme relating the information phases {A”6;} to the transmitted phases {6;} is given by
APO iO) = On
8.3 A ternary phase information source whose phases take on values (0, 27/3, 47/3) is to be
communicated using a third-order phase difference modulation (PDM-3).
(a) Show that the appropriate encoding scheme relating the information phases {A”6;} to the
transmitted phases {6;} is given by

A36; = 6; — 6-3
(b) For the ternary source, does the relation found in Problem 8.2 also hold, in general, for
values of n greater than 3? Explain.
8.4 A binary amplitude (+1) information source is to be communicated using a second-order
phase difference modulation (PDM-2). Show that the two encoders illustrated in Fig. P8.4
are equivalent, that is, show that the relation between the information sequence {a;} and the
transmitted sequence {c;} is the same for both.
8.5 Consider the communication system of Problem 8.4 wherein the data source {a;} is char-
acterized as an independent, identically distributed (i.i.d.) sequence with Pr {a; = —1} =

Figure P8.4
Problems 551

Figure P8.6

p, Pr{a;=1}=1-— p. Assuming that the first two bits in the output sequence are co =
1, cCjh= ie

(a) Evaluate the statistical mean of the output sequence {c;} for all i.
(b) Evaluate the autocorrelation function of the output sequence R,(m) A E {cjcj+m}. Is the
sequence te;} in general, wide sense stationary (WSS)? Explain.
(c) If p = 1/2, that is, {a;} is a purely random data sequence, is the sequence {c;} WSS? Is it
an uncorrelated sequence, that is, R. {m} = 0 form #40?
8.6 A binary amplitude (+1) information source is to be communicated using a nonadjacent
second-order phase difference modulation (NA-PDM-2). The encoder for such a system is
illustrated in Fig. 8.18 and is repeated in Fig. P8.6 for convenience. The data source {a;} is
characterized as an independent, identically distributed (i.i.d.) sequence with Pr {a; = —1} =
p, Pr{a; =1}=1-— p. Assuming that the first three bits in the output sequence are co =
jig (5) AN O=l,
(a) Evaluate the statistical mean of the output sequence {c;} for all i.
(b) Evaluate the autocorrelation function of the output sequence R,(m) A E {cjcj+m}. Is the
sequence {c;} in general, wide sense stationary (WSS)? Explain.
(c) If p = 1/2, that is, {a;} is a purely random data sequence, is the sequence {c;} WSS? Is it
an uncorrelated sequence, that is, R, {m} = 0 form 4 0?
8.7 A binary phase information source whose phases take on values (0, 7) is to be communicated
using a second-order phase difference modulation (PDM-2). Based on Equations (8.3) and
(8.4), an equivalent transmitter implementation is obtained by replacing the pair of back-to-
back differential encoders in Fig. 8.l1a each having delay 7, by a single differential encoder
having delay 27, (also see Problem 8.4). Is it possible to make the same type of replacement
in the receiver of Fig. 8.5, that is, can the two differential detectors each having delay 7, be
replaced by a single differential detector having delay 27,;? Answer the question separately for
the case of Af = 0 and Af #0.
8.8 Consider a differential PSK system with I-Q detection that transmits an all-zero phase angle
sequence of length five, that is, (9 = 0; = 62 = 03 = 64 = 0. Assume that there is a channel fre-
quency offset of Af 7, = 1/87), and a channel phase offset 6 = 7/5. Also assume for simplicity
that there is no additive noise.
(a) Find the sequence of measured angles {n;};i =0,1,..., 4 seen by the phase measurement
portion of the I-Q detector as shown in Figures 7.1 and 8.13.
(b) If the system under consideration is binary DPSK (Fig. 7.1), find the receiver’s estimate of
the sequence of first-order differences {w; = An;} fori = 1,..., 4. Also find the receiver’s
estimate of the sequence of data information angles aéi| LOT ers 4. (Assume that
4 = 0 is known.) How large can 6 and Af be before the first-order differential decoding
results in an error in A6;?
(c — If the system under consideration is binary D*PSK (Fig. 8.13), find the receiver’s estimate
of the sequence of second-order differences {A?n;} for i = 2, 3, 4. Also find the receiver's
estimate of the sequence of data information angles {476} for i = 2, 3, 4. (Assume that
> = 6; = 0 is known.) How large can 6 and Af be before the second-order differential
decoding results in an error in A6;?
8.9 Suppose that signals s2(t), s3(t), and s4(t) as given by Eq. (8.12) are transmitted and received
eee ee eee rr eee Meee re hl ete ee ee IDO “at waalacae tlihactentad in
552 Double Differentially Coherent Communication with Waveforms Chap. 8

Fig. 8.10. Trace the signals through the receiver block diagram to find the summer output (input
to the signum operator) V...4V¢.3 + Vs.4Vs.3. Assuming that the transmitted phase angles for all
three signals have been second-order differentially encoded, explain how the signum operator
can be used to make the decision for the binary signal alphabet (0, 77).
8.10 For a D*PSK system, suppose that, in addition to the frequency offset Af and phase offset
6, there is a higher-order offset between the received signal carrier and the receiver refer-
ence signals given by at?, that is, the complex channel rotation shown in Fig. 8.2 becomes
exp{j[0+22Aft; + at?]}.
(a) In the absence of noise, find the output of the summer (input to the signum operator)
Voi Ve.i-1 + VsiVs,;-1 in the ACD of Fig. 8.10. Also find A?n; in the I-Q demodulator
of Fig. 8.13.
(b) Design a third-order differential encoder and a corresponding demodulator with third-order
differential decoding that would, in the absence of noise, eliminate the effect of this higher-
order offset.
8.11 It is desired to evaluate the channel capacity (see Chapter 1) of a second-order phase difference
modulation (PDM-2) system using a transmitter as in Fig. 8.la, that is, a double differential
encoder, and an I-Q receiver as in Fig. 8.13. Such a system can be modeled as an M-ary
symmetric channel (see Chapter 3) for which the channel capacity (in bps) is given by
M-1 , M-1
os rr D> Pen (E |m) logs Pen (E |m) + logyM
m=0 n=0

where P,,, (E |m) is the conditional probability of choosing the n'" symbol (phase) when the
m" symbol (phase) was transmitted, that is
i Qn
Pn (E \m) & Pr {A76; = By esis Bn = =
Note that Ps, (E |m) is actually the conditional probability of a correct decision given that the
mh symbol was transmitted. Since the £;’s are equiprobable, P;,, (E |m) is also the average
probability of correct symbol detection.
With reference to Figures 8.la and 8.13, the pdf of the second-order phase difference
Ax; & A?n; — A?6; (see Problem 8.12) is denoted by Fax; (Axi), which is independent of i
and the transmitted information phase A’6;, and is distributed in the interval —7 5S ANG Sok
(a) Evaluate the channel capacity of the PDM-2 system. Express your answer in terms of the
set of probabilities
(21I-M+1)2/M
ee) idx CAN Ne ene ee M-1
(QI—M—1)x/M
—(M—1)x/M +
Po =i)
=
fax, (Axi) dAXi +f fax; (Axi) dAxXi
(M—1)x/M
(b) Suppose now that we have a PDM-1 system using a single differential encoder as a trans-
mitter and a receiver as in Fig. 8.8 (or Fig. 7.1). Here, the conditional pdf of x; 4 An; —
A6;, which is denoted by fx, (x;), is independent of i and the transmitted information phase
A6;, and is distributed in the interval —z < x; < 2. Also, denote the probabilities corre-
sponding to the P;’s of part (a) by
(2l-M+1)x/M
a= | PaCOehden US MNees se. 54 ll
Ql-M-1)x/M
ag
OV= fy. (x) dx; + [ fy. (y;) dy;
 References 553

In principle, how would your answer found in part (a) change, if at all; that is, express the
channel capacity in terms of the Q)’s.
8.12 Consider a second-order phase difference modulation (PDM-2) system composed of the trans-
mitter of Fig. 8.la and the I-Q receiver of Fig. 8.13. The pdf of the second-order phase
difference Ax; 4 A?n; — A6; is denoted by fax, (Axi) which is independent of i and
the transmitted information phase A*6; , and is distributed in the interval —7 SING GS He,
In terms of the equivalent first-order phase differences x; 4 An; — AO;, we have Ax; 4
An; — Anj—; — 46; — AO;-; = xi — xi-1. The pdf of Ax; can, in principle, be obtained if
———P3 OO”
A?ni A6;
the joint pdf of x,;-; and x;, namely fx,_,.x;(Xi-1, Xi), is available. In particular
8

Fax(Axi) = i at at AX ail
Say

Although the joint pdf fx,_,,x;(Xi-1, Xi) can be found and then f,x,(Ax;) evaluated as above,
the resulting expression becomes quite complicated and is in the form of multiple integrals.
In the high SNR case, the matter can be simplified by approximating fx,_,x,(Xi-1, Xi) by a
two-dimensional Gaussian form. In particular

: ieetten 1 fe ee DO Gas & Cee eee

Xj, Xi-1, Xi ” 2ro2/1 — p2 p 2(1— p2)o? ’ SXi-1: IES ’

where p = —1/2 is the correlation between the first-order phase differences in the two adjacent
symbol intervals and o* = No/ST; = R~' denotes the inverse of the symbol SNR.
(a) Show that the second-order phase difference approximate (high SNR) pdf is given by

Ffax(Axi) = Sa
ZONE
|
Ot a
= (Ax; + 2am)?
20
| UE SING GSE

Hint: In the case where x;_; and x; are not restricted to values in the interval (—z, z)
as in the approximate pdf above, one can first compute the pdf of Ax;g 4 x; — x;-1, the
“actual” phase difference without the modulo 27 condition by applying the above integral
with limits —oo and oo instead of —z and z. Then, the pdf of the modulo 27 reduced Ay;
is given by

AROS De FAXig
(AXi + 20m)
m>=—CO

where fax,, (Axia) is the pdf of Axia.

(b) Using the result found in part (a), evaluate the asymptotic (high SNR) average symbol error
probability, P;(E), of the I-Q D?PSK receiver of Fig. 8.13. Express your answer in terms
of “erfc’”’ functions with appropriate arguments.

REFERENCES
i Divsalar, D. and M. K. Simon, “Doppler-Corrected Differential Detection of MPSK,” JEEE
Transactions on Communications, vol. 37, no, 2, February 1989, pp. 99-109.
. Divsalar, D. and M. K. Simon, “On the Implementation and Performance of Single and Double
Differential Detection,’ IEEE Transactions on Communications, vol. 40, no. 2, February 1992,
pp. 278-91.
. VanAlphen, D. K. and W. C. Lindsey, “Higher-Order Differential Phase Shift Keyed Modulation,”
to appear in the JEEE Transactions on Communications.
. Okuney, Yu. B., Theory of Phase-Difference Modulation, Moscow: Svyaz Press, 1979.
. Fink, L. M. , Theory of Digital Data Transmission, Moscow: Sovetskoye Radio Press, 1970.
554 Double Differentially Coherent Communication with Waveforms Chap. 8

6. Korzhik, V. I., L. M. Fink and K. N. Shchelkunov, Calculation of the Noise Immunity of Digital
Data Transmission Systems (A Handbook), Moscow: Radio i Svyaz Press, 1981.
. Okunev, Yu. B., V. A. Pisarev and V. K. Reshemkin, “The Design and Noise-Immunity of Mul-
tiphase Autocorrelation Demodulators of Second-Order DPSK Signals,” Radiotekhnika, vol. 34,
no. 6, 1979 [Telecomm. Radio Engng., Part 2, vol. 34, no. 6, 1979, pp. 60-63].
. Okunev, Yu. B. and L. M. Fink, “Noise Immunity of Various Receiving Methods. for Binary
Systems with Second-Order Phase-Difference Modulation,” Radiotekhnika, vol. 39, no. 8, 1984
[Telecomm. Radio Engng., vol. 39, no. 8, 1984, pp. 51-56].
. Okunevy, Yu. B., N. M. Sidorov and L. M. Fink, “Noise Immunity of Incoherent Reception with
Single Phase-Difference Modulation,” Radiotekhnika, no. 11, 1985 [Telecomm. Radio Engng.,
no. 11, 1985, pp. 103-106.
10. Okunev, Yu. B. and N. M. Sidorov, “Noise Immunity of a Carrier-Frequency-Invariant Demodu-
lator of DPSK-2 Signals,” Radiotekhnika, no. 6, 1986 [Telecomm. Radio Engng., no. 6, 1986, pp.
81-83].
iit, Gut, R. E., “Noise Immunity of Autocorrelation Reception of Single PSK Signals,” Radiotekhni-
ka, no. 10, 1972 [Telecomm. Radio Engng., no. 10, 1972, pp. 120-122].
. Gut, R. E., Yu. B. Okuney and N. M. Sidorov, “Effect of Carrier Detuning on the Noise Immunity
of Autocorrelation Reception of DPSK Signals,’ Radiotekhnika, no. 7, 1986 [Telecomm. Radio
Engng., no. 7, 1986, pp. 96-98].
13. Pent, M., “Double Differential PSK Scheme in the Presence of Doppler Shift,’ Digital Commu-
nications in Avionics, AGARD Proceedings no. 239, 1978, pp. 43-1-43-11.
14. Stein, S., “Unified Analysis of Certain Coherent and Noncoherent Binary Communication Sys-
tems,” [EEE Transactions on Information Theory, January 1964, pp. 43-51.
. Henry, J. C., Ill, “DPSK versus FSK with Frequency Uncertainty,’ JEEE Transactions on Com-
munications Technology, vol. COM-18, December 1970, pp. 814-817.
. Miller, T. W., “Imperfect Differential Detection of a Biphase Modulated Signal—An Experi-
mental and Analytical Study,’ Ohio State University, Technical Report No. RADC-TR-72-16,
February 1972.
. Divsalar, D. and M. Shahshahani, “The Bit Error Rate (BER) of DPSK with Differential Detector
(DD) Followed by Matched Filter (MF),” private correspondence.
. Pawula, R. F., S. O. Rice and J. H. Roberts, “Distribution of the Phase Angle Between the Vectors
Perturbed by Gaussian Noise,” JEEE Transactions on Communications, vol. COM-30, no. 8, pp.
1828-1841.
. Simon, M. K., “Comments on ‘On Binary DPSK Detection’,” JEEE Transactions on Communi-
cations, vol. COM-26, no. 10, October 1978, pp. 1477-1478.
20. Arthurs, E. and H. Dym, “On the Optimum Detection of Digital Signals in the Presence of White
Gaussian Noise—A Geometric Interpretation and a Study of Three Basic Transmission Systems,”
IRE Transactions on Communications Systems, December 1962, pp. 336-372.
Ne Divsalar, D. and M. K. Simon, “ Multiple Symbol Differential Detection of MPSK,” JEEE Trans-
actions on Communications, vol. 38, no. 3, March 1990, pp. 300-308. Also see “Multiple Symbol
Differential Detection of Uncoded and Trellis Coded MPSK,” Pasadena, CA: Jet Propulsion Lab-
oratory, JPL Publication 89-38, November 15, 1989.
yp), Qu, S., “New Modulation Schemes: Principles and Performance Analyses,” Ph.D. Dissertation,
Technical University of Nova Scotia, 1992. Also see S. Qu and S. Fleisher, “Double Differential
MPSK on the Fast Rician Fading Channel,” submitted for publication in the IEEE Transactions
on Communications.
 Chapter 9
Communication
over Bandlimited Channels

Except in a few instances, all of the communication scenarios we have discussed thus far
have corresponded to an AWGN channel with unconstrained bandwidth. As such, the only
effect the channel was assumed to have on the transmitted signal (aside from the addition
of Gaussian noise) was attenuation of its amplitude and the addition of a phase shift to the
carrier, the latter being assumed constant in time over the duration of the data symbol or
longer. In reality, the communication channel is never infinite in bandwidth; however, in
certain applications, it is considerably wideband when compared to the bandwidth of the
modulation being communicated over it. Thus, for all practical purposes, one can in these
situations still model the channel’s effect on the transmitted waveform as multiplication by
a fixed gain (actually an attenuation) and the addition of a constant carrier phase shift. Al-
ternately, the equivalent baseband signal when passed through the channel is multiplied by
a fixed gain and delayed by an amount that is constant for all frequencies in the waveform.
For all practical purposes, this is the case that we have treated thus far in the book.
In this chapter, we discuss the design of modulation/demodulation schemes for send-
ing digital information over a bandlimited channel whose bandwidth is on the same order as
that of the modulation. Our treatment will be brief by comparison with other fine textbooks
devoted to this subject [1, 2, 3], and will only focus on the key issues. We shall consider
only linear channels, that is, those modeled as a linear filter with bandpass transfer function
F.( f), and linear modulations of the type corresponding to a baseband pulse stream

m(t) =A) Gp —iTs) (9.1)

i=—0Oo

modulated onto a carrier where A is the signal amplitude and c; = a; + jb; is, in general, a
complex discrete random variable representing the data symbol to be communicated in the
i transmission interval. The key issue here is the design of the pulse shape p(t) to allow
communication at the highest possible data rate while achieving the best trade-off between
signal power and noise power. As we shall see, this pulse shape design is dictated by the
characteristics of F.(f) over its bandwidth occupancy. In particular, if F.(f) is an ideal
(rectangular transfer characteristic) filter of IF bandwidth W Hz (lowpass bandwidth B =
W/2), then, at least theoretically, p(t) can be designed to allow distortion-free transmission
at a rate equal to W bps with an SNR (E»/No) equivalent to that of the unconstrained
bandwidth channel. If the channel is not ideal over its bandwidth occupancy, then signaling
556 Communication over Bandlimited Channels Chap. 9

at a rate W bps or higher results in a form of distortion known as intersymbol interference

(ISI) wherein a number of adjacent received pulses interfere with each other in so far as
the detection process is concerned, thus producing a degradation in SNR. These last two
statements are the basis of the well-known Nyquist criterion [4], which will be reviewed
early in this chapter. Even when the channel is not ideal over its passband, it may still satisfy
the Nyquist criterion in the sense that it allows distortion (ISI)-free transmission but at a
communication rate less than W bps. The pulse shapes that satisfy the Nyquist criterion are
referred to as Nyquist pulses, and systems that employ such pulse shapes achieve the best
level of performance.
Assuming first that the overall pulse response is designed according to the Nyquist
criterion, the next issue is to determine how to best apportion it between the transmit and
receive filters so as to maximize the detection SNR. We shall see that, under the constraint of
a given amount of power available to the channel, the solution to this problem is to equally
split the total filtering between the two locations. Once again, this comes about because of
the optimization in the trade-off between signal plus distortion power and noise power at the
receiver, which is consistent with the notion of matched filtering introduced in Chapter 4.
Quite often, it is not possible or practical to design a Nyquist pulse and thus the system
will be corrupted by ISI. When the ISI is a controlled amount, it is possible to exploit this
occurrence in the design of the transmitter and receiver. This leads to a class of systems
known as partial response (PR) systems [5], which have the advantage that they are simple
to implement yet produce an efficient trade-off between signal plus distortion and noise
powers. The simplest case of a PR system is one where the ISI results only from a single
adjacent symbol and is referred to as duobinary [6]. We shall discuss this case in detail.
In other instances, one might resort to the more complicated case of designing a re-
ceiver that is optimum in the presence of both ISI and additive Gaussian noise. The solution
to this problem takes several different forms depending on the optimization criterion spec-
ified and the nature of the detection and decision processes. Of particular interest are the
structures resulting from application of the maximum-likelihood principle. These structures
are based on the notion of maximum-likelihood sequence estimation (MLSE) wherein a joint
estimate is made on a sequence of transmitted information symbols with the goal of min-
imizing the sequence error probability.. The reason for estimating a sequence rather than a
single symbol at a time is that ISI has the effect of introducing memory into the modulation
in the sense that the detection of a given symbol depends on the knowledge of past (and
possibly future) data symbols. This idea of making a joint decision based on an observation
of the received signal over more than one symbol interval is not unlike that discussed in
several of the previous chapters in connection with optimum block-by-block detectors. We
shall discuss an efficient algorithm for performing MLSE, namely, the Viterbi algorithm [7],
which was originally proposed in the context of decoding convolutional codes.
The ability to design a receiver as in the above to mitigate the deleterious effects of ISI
relies on having complete knowledge of the channel transfer characteristic F.(f). In certain
practical channels, for example, the telephone channel, this characteristic is time-varying,
and thus the design of the demodulator must be made adaptive so as to remain optimum as
a function of time. The general process by which the receiver attempts to cancel the ISI
produced by the channel to restore a Nyquist pulse shape is referred to as equalization.
In the case of a time-varying channel, the receiver achieves the equalization process by
continuously adjusting a number of its parameters based on measurements performed on the
 Sec. 9.1 Definition and Characterization of a Bandlimited Channel 557

channel characteristics. As such, the process is called adaptive equalization. A large body
of literature has been developed that is devoted to the design and convergence of algorithms
for adjusting the equalizer’s parameters. Since this is not the major focus of this textbook,
we shall be relatively brief in the coverage of this subject.
We begin the chapter with a brief characterization of what is meant by a bandlimited
channel and a practical application where it is encountered.

9.1 DEFINITION AND CHARACTERIZATION OF A BANDLIMITED CHANNEL

By “bandlimited channel,” we infer a strictly bandlimited channel corresponding to one
whose frequency response is nonzero only in a finite portion of the spectrum. If the channel
is bandpass with frequency response F,(f), then we require that F.(f) = 0 for |f — f.| >
W/2(f > 0) and for |f + f-| > W/2(f <0). If the channel impulse response f(t) =
F-'{F.(f)} is written in the form

felt) = Re {gc(ryel"*| (9.2)

where g,(t) is the equivalent lowpass impulse response, then in terms of the lowpass equiva-
lent channel frequency response G.(f) = F {gc(t)} =2Fo (f + fc), we have that G,(f) =
0 for | f| > W/2 = B. The notion of baseband equivalence of an IF function used here is
analogous to that discussed for random processes in Appendix 4B. For a baseband chan-
nel, the bandlimiting requirement is identical to that imposed on G,(/f). In what follows,
we shall find it convenient to work with G,(f) rather than F.(f) with the understanding
that if we are dealing with a baseband channel they are one and the same. Furthermore, for
simplicity, we shall assume that F.(f) is symmetric around its center frequency, f., so that
g-(t) is a real function of t.
In view of the above, a bandlimited AWGN channel is one that, in addition to band-
limiting the signal, also adds white Gaussian noise. Thus, if a bandpass signal s(t) =
Re {m(t)e/ aH Sct } is transmitted over this channel, the received signal is given by r(t) =
Re {m’ (t)e/ an fot } where m’(t) is obtained from the convolution of the lowpass modulation
m/(t) with the lowpass channel impulse response g,(t). For the assumed random pulse train
form of modulation, m’(t) becomes
o@)

mit=A >. &p'(t —iT;);

i=—0Oo
(9.3)

= T)dt =) P(f)Ge(fye!*f'df
p(t) =! P(T)gclt
65 5x

with P(f) the Fourier transform of p(t).

It is important to understand that both the envelope |G,-(f)| and phase @.(f) =
arg {G_(f)} characteristics of the channel lowpass frequency response are influential in
determining the distortion introduced into the transmitted pulse shape. Assuming that P(f)
is bandlimited to (—B, B) (since no energy outside that interval will get through the channel
anyway), then if |G.(f)| is constant, that is, Go9 = G-(0), and @.(f) is a linear function of
frequency, that is, —Kg f over this same interval, then the channel introduces no distortion
other than a pure time delay Ty = Kg/27 and from (9.3) the received pulse shape becomes
 558 Communication over Bandlimited Channels Chap. 9

p(t) = Gop(t — Ta). Note that since P(f) is bandlimited, p(t) cannot be timelimited (the
uncertainty principle [8] prevents a signal from being limited in both time and frequency)
and likewise for p’(t). Thus, the pulses in m’/(t) must, by necessity, overlap. This does not,
however, mean that ISI is present at the receiver since by proper design of p(t), it may still
be possible to separate out the various pulses in the data stream. What is important here is
that if p(t) is designed to produce zero ISI at the receiver, then the distortion-free channel
above does not change that condition, that is, p’(t) also yields zero ISI.
In the more general case where |G-(f)| is not constant and 6,(f) is not a linear func-
tion of frequency over (—B, B), the channel will introduce amplitude and phase distortion
into the transmitted pulse shape. It is common to define the group delay, Ty(f), by the neg-
ative of a scaled derivative of the channel phase characteristic, 6.(f), that is,

Lf) =
1 d6.(f) (9.4)
DT ats
This definition is consistent with the above discussion for the distortion-free channel, where-
upon the group delay becomes a constant (fixed) delay, Ty, independent of frequency. For
the channel that introduces phase distortion, Tg(f) will be a function of frequency. The ef-
fect of amplitude and phase distortion on a transmitted pulse stream modulation, m(f), is to
smear the pulses in the received modulation, m’(t), so that they cannot be separated by the
receiver. It is indeed this additional smearing of the pulses that accounts for what is referred
to as intersymbol interference.
Perhaps the most common bandlimited channel that, in addition to many other impair-
ments (e.g., impulse noise, phase jitter, frequency offset, etc.), exhibits both amplitude and
phase distortion over its frequency range of definition is the telephone channel, whose us-
able band is typically in the range of 300 Hz to 3 KHz. A detailed discussion and illustration
of the telephone channel characteristics are given in [2]. As an example, an average medium
range telephone channel of the switched telecommunications network [9] has an impulse re-
sponse of about 10 ms; thus, for communication over this channel at a rate on the order of
2.5 Ksps (a symbol duration of 0.4 ms), the ISI would extend over 25 symbols.

9.2 OPTIMUM PULSE SHAPE DESIGN FOR DIGITAL SIGNALING THROUGH

BANDLIMITED AWGN CHANNELS
Consider the block diagram of a baseband bandlimited communication system illustrated in
Fig. 9.1. This system is also the baseband equivalent of a bandpass bandlimited communi-
cation system in the context of our discussion above. The source generates a pulse stream
modulation m(t) of the type described by (9.1). This modulation is passed through the trans-
mit filter G;(f) and the channel G,(f). The filtered modulation is perturbed by additive
complex Gaussian noise n(t) with power spectral density (PSD) S,(f) and the sum of the
two is passed through the receive filter G,(f).! Thus, the detected pulse is shaped by the
combination of G;(f), Gc(f), and G,(f) whereas the additive noise spectral density is
shaped only by G,(f). Samples of the output of G,(f) are taken once per symbol interval,
T;, and processed to make decisions on the information sequence {¢,}. The problem at hand

1. Since the noise process 7(t) is, in general, complex, its PSD is defined as Sjax(f) 4 FE {a(t)a*(t + T)}}.
However, to keep the notation simple, we shall define it with only a single subscript, namely, S;(f), as is done for
the PSD’s of real random processes.
 =()u
v (L1-)d'2"0
oo-=t

(1)u

=
dSd
(f)"s
1+ "74 =11e gJdues

be JeuueyD EDS EYNCR IsDeq

30m (f)'D C9 (f)‘9

aang
[6 WY YOOIG WeIBeIP
JO& PoyUUI[pueQ UONROTUNWIWIOS
W9}SAS

559
560 Communication over Bandlimited Channels Chap. 9

is to design G;(f) and G,(f) [for a given G,(f)] to minimize the error probability of the
receiver.”
The output y(t) of the receive filter can be expressed as

IO=A YO Fig —iTs) HO (9.5)

i=—0O

where g(t) = F-!{G(f)} = F-!{Gi(f)Gce(f)G,(f)} is the overall impulse response,

that is, the detected pulse shape, and /i(t) is the receive-filtered complex noise process
with PSD Sz(f) = Sa(f) IGSCPI*. Suppose that j(t) is now uniformly sampled at times
t =kT, + to (k integer) to produce the complex sample values

H(KTs +0) =A D> Gig((k — i) Ts + t0) + W(KTs + 10) (9.6)

i=—0COo

where fo is selected to match the delay of the channel (more about this in a moment). In
shorthand notation, (9.6) can be rewritten as
[o,@) CO

y= A ~ Ci8k—i + He = AChgo + A 2a Ci8k—i + Mk (OF)

i=—0Oo ==——(0.6)
ith
where, in general, vz, denotes the sample of the waveform v(t) at time kT; + fo. The first
term in (9.7) represents the desired signal sample corresponding to the k® transmission
interval and thus tg should be selected to maximize go, that is, we should choose our
sampling time epoch so that the desired sample corresponds to the peak of the overall pulse
shape g(t). The second term represents ISI and ideally should be made equal to zero. The
third term is the sample of the filtered complex Gaussian noise process.
Assuming for the moment that we can indeed achieve zero ISI, then the optimum
decision rule is identical to that considered in Chapter 4, namely, choose the data symbol
CE that is closest in distance to yg. For example, for binary modulation with ¢c; real (c; = a;)
and taking on equiprobable values +1, the decision variable is real, that is, yy = yz and
the decision rule is implemented as a threshold comparison, that is, a, = sgn {yx}. The
corresponding bit error probability is given by

a! E (cleat » a Ago
P,(E) = 5)erfc [ital = 3 erfc |
ae (9.8)

which is the minimum achievable without further specification on G;(f) or G,(f). As

noted above, the variance of the filtered noise sample depends only on G,(f) and thus

oy = ( Sn(f) IGP) Pdf (9.9)

2. Although, in a real communication system, the modulation pulse shaping and transmit filtering are separate
entities, from a mathematical standpoint, their cascaded effect on the transmitted waveform is represented by the
product of their respective transfer functions P(f) and G;(f). Thus, for analysis purposes, we shall denote the
product P(f)G;,(f) by simply G;(f) with the understanding that in the physical system G;(f) must be partitioned
between the modulation pulse s..ape and the transmit filter.
Sec. 9.2 Optimum Pulse Shape Design for Digital Signaling 561

Also, as noted above, the desired signal sample go depends on G;(f), G-(f), and G;(f),
ViZ.

«= | Gi(f)Ge(f)G-( fel"! df (9.10)

We will return to a consideration of the selection of G;(f) and G,(f) to minimize P(E) of
(9.8), that is, maximize go/oy, after first deriving the criteria necessary to achieve zero ISI,
that is, the Nyquist criterion. For simplicity, we shall let go = 1 since an arbitrary gain can
be included by scaling G(f) later on. Also, we shall continue to consider the case of real
(but not necessarily binary) data, that is, ¢; = a;, until further notice.

9.2.1 Nyquist Criterion for Zero ISI

We start out by recalling the well-known sampling theorem, namely, if X (f) = F {x (t)} is
bandlimited to (—B, B), then x(t) is uniquely determined by samples taken at 1/2B, viz.

= n \ Sin 27B (t — 35)

x(t) = oe ee (9.11)
pe (35) 27 B (1 a opi)

and taking the Fourier transform of x(t)

ACf) = te 2 x(Py)exp(—J2ef sy): Ifl<B

2B page 2B 2B = (9.12)

0; otherwise

The lowpass bandwidth B is referred to as the Nyquist bandwidth and is equal to one-half
the sampling rate.
From (9.7), for zero ISI, we desire a pulse waveform g(t) such that its samples taken
at a rate 1/7; (the data symbol rate) satisfy?

j2n fnT. 1; n=0

on / G(f)e! ‘df = bn = 0: on Z 0 (9.13)
[o,@) 2

We can rewrite the above integral in terms of a partition of adjacent Nyquist bandwidths of
width 2B = 1/T,, viz.

oo (2k+1)/2T; es
&n= i G( fel" df =, (9.14)
k=—o0o (2k—1)/2Ts

Using the change of variables v = f — k/Ts, (9.14) becomes

3. For the purpose of the discussion here, we shall assume without any loss in generality that the sampling epoch
to equals zero. Whatever conditions we determine on the overall pulse transfer function G(f) can later be modified
by replacing G(f) with G(f)e/?7/".
562 Communication over Bandlimited Channels Chap. 9

W275
2n= = /; G (vay =)el 2A Utk/Ts)nTs qyy

a er 3
ine) 1/2T; k ; x
=e i G €+ =)eltmunts dy (9.15)
pene ek 8
1/21; [ & Peer
=i) ) G (vSe =) eltmunts dy
ml TN pee Ts
Define the equivalent Nyquist channel characteristic

Geq(Sf) cal \ar

eS G(f+#)s isd Le — 275 (9.16)
0; otherwise

that is, the sum of all of the translates of G(f) folded into the interval (—1/2T7;, 1/2T;).
Substituting (9.16) into (9.15) gives
1/2T,
8n = i Galpe ay (9.17)
—1/2T;
But the inverse Fourier transform of Ge,(f) is by definition

00 1/2T, 4
8eq(t) = / Gel felt dfi= Geg(fyel*"f'df (9.18)
=e0 =W2re
with samples
1/2T;
8eqn = if GH (pe ar Pap (9.19)
S/T:
Thus, from (9.18) and (9.19), we see that the Nyquist rate samples of g(t) are also the
Nyquist rate samples of geq(t), that is, g¢g, = &n-
Since, by the definition of (9.16), Geg(f) is a strictly bandlimited function on the
interval (—1/2T7;, 1/2T7;), it can be expressed in the form of (9.12), namely

Tic) cents ne aaae 2Ts1

ee) .

Geq(f) = nace (9.20)

OR otherwise

However, for zero ISI we require that g, satisfy (9.13). Thus, using (9.13) in (9.20) gives

G a Ts; \flson27; 9.21

qf) |0; otherwise ( )
that is, an ideal brick wall (rectangular) filter. Finally, combining (9.16) and (9.21), we see
that the combined (transmitter, channel, and receiver) overall system frequency response
G(f) must satisfy

= k 1
oy o(s+5) =n: lar (9.22)
k=—0o
Sec. 9.2 Optimum Pulse Shape Design for Digital Signaling 563

that is, the superposition of all of the translates of G(f) in the interval (—1/2T,, 1/2T;)
must yield a flat response.
It can also be shown (Problem 9.1) that the superposition of all of the translates of
G(f) in the interval ((2n — 1) /27; < f < (2n + 1) /2T;) must also yield a flat response
for any integer n. Thus, the zero ISI condition on G(f) as specified by (9.22) generalizes to

ye G = =) =T, forall f (9.23)

k=—co

Note that the criterion for zero ISI does not uniquely specify the pulse spectrum G(f)
unless its bandwidth is limited to (—1/2T7;, 1/27;), in which case it must be flat since the
sum in (9.23) reduces to one term (k = 0). In this special case, G(f) itself has a rectangular
transfer function in the interval (—1/27;, 1/2T;), or equivalently

sin wt /T;
(9.24)
g(t) sca aiidbk

The interpretation of this result is that for a digital modulation of the form in (9.1), the
minimum one-sided lowpass bandwidth needed to transmit R; = 1/7; symbols per second
(sps) without ISI, is R;/2 = 1/27; Hz. When the data symbols {a,} are independent and the
noise samples (spaced 7;-sec apart) are uncorrelated, then each symbol can be recovered
without resorting to past history of the waveform, that is, a zero memory system.
Since in the above system R; sps are transmitted without ISI over a bandwidth R,/2
Hz, then the throughput (efficiency of transmission) is Rs sps/(R;/2) Hz = 2 sps/Hz. To
achieve this efficiency, one must generate the sin x/x pulse shape in (9.24). Unfortunately,
this pulse shape is impractical because its very slowly decreasing tail will cause excessive
ISI if any perturbations from the ideal sampling instants occur.
To reduce this sensitivity to timing (sampling instant) offset and at the same time
ease the filter design problem in the transmitter and receiver, we use more practical shapes
for g(t) whose Fourier transform G(f) has smoother transition at the band edges yet still
satisfies the zero ISI criterion of (9.23). One such class of channels that has achieved a
good deal of practical application is the set of raised-cosine channels defined by the overall
transmission characteristic

Gf) =) Toco {Z(21flts-1t+e]}; Fe sills 7 (9.25)

0; otherwise

where 0 < a@ <1 is defined as the excess bandwidth factor. The special case of a = 0
corresponds to the minimum Nyquist bandwidth brick wall (rectangular) filter as discussed
above. Values of a > 0 trade increased bandwidth (and hence reduced throughput for a fixed
data rate) for a reduction in the rate of roll-off of the frequency response. In fact, for the class
of systems defined by (9.25), the throughput is reduced to Rs/[(Rs/2)(1 + @)] =2/(. + @)
sps/Hz. Figure 9.2a is a plot of G(f) for several values of a. The corresponding overall
system impulse response is
iG (Se) ( cosamt/Ts ) (9.26)
Ae L=—\(2at/ T.)7
———— SS

go(t) 8a(t)
564 Communication over Bandlimited Channels Chap. 9

1/2T, VTs

(b)
1.0 sin (xV/T,) cos (zat/T;)
(nT) 1 —(2a/T,)?

Figure 9.2 The raised cosine pulse: (a) frequency function, (b) time function

which is illustrated in Fig. 9.2b for the same values of a as in Fig. 9.2a. For a = 0, we get
the time response for the minimum bandwidth brick wall filter as given in (9.24). For values
of a > 0, the tails of the response decay more rapidly than for the sin x /x response of (9.24).
Note that (9.26) is written in the form of a product of go(t) (the overall system impulse
response corresponding to the minimum bandwidth brick wall Nyquist filter) and another
response, ga(t). In fact, (9.26) can be generalized to include any function gg(t) that is
bounded at the zero crossings of go(t) since under this condition, the zero crossings of g(t)
will be the combined zero crossings of go(t) and gq(t) and thus the zero ISI condition is
maintained. Stated in the frequency domain, the convolution of the minimum bandwidth
brick wall filter Go( f) = F {go(t)} and an arbitrary filter Gg(f) whose inverse Fourier
transform gq(t) is bounded at the zero crossings of go(t) is a Nyquist filter.

9.2.2 Apportionment of Filtering between Transmitter and Receiver

Having just seen how to design the overall system frequency response G(f) to achieve
zero ISI, the next issue concerns how to optimally apportion this response between the
transmitter and receiver. By optimum, we mean that the apportionment should result in
Sec. 9.2 Optimum Pulse Shape Design for Digital Signaling 565

minimum error probability or equivalently maximum detected sample SNR. From (9.8), we
see that this is equivalent to maximizing the ratio*

ge (SR GX NGAAG- (Pela fy” SO27)

ea OO I 7) eC
oy fsa PIG) "af
which is identical to the matched filtering problem considered in Chapter 4. Thus, we
conclude that the transmit and receive filters are related by (see Problem 9.2)

GUy=K GG peri (9.28)

Sn(f)
where K is a real constant, or, in terms of the overall frequency response, we should choose

CONE oe
ys |ee
ace ( Sa(f) Lane
G ee ent ee Soe
eee a
Pie ( KIGAA)F
Since S,(f) is real, then we must have arg {G(f)} = e~/7ft in order for (9.29) to have
any meaning. Moreover, since go = 1, then from (9.27)-(9.29), the constant is evaluated as
a
K=(f°2, IGP PIGCAPMSaAaf) .
If we now assume that n(t) corresponds to white Gaussian noise with PSD S,(f) =
No/2, then we get
V2
G,(f)|
IG-(f)| = (2K ee — ieee!
| eG ryiGu Ald
IG(f)|
No
(9.30)
1/2 Pe 1/2
IGe(AIIGAA)| = (pre i (in | GNP GAP af)
which implies that, except for a scaling constant, the overall pulse response should be
split equally between the combined transmit/channel and receive filters. Furthermore, the
maximum SNR achieved by this apportionment is from (9.27) given by

> 82 A? 2A? [IGA IGA) af

aialciiigs
ES BPG.
ak a char”
No (9.31)
DT Ay.pan,ge : (fl 5 df )ONE
a (4 a [ie
een al Yaar
4 ——

where P, is the signal power at the input of the receiver. For binary amplitude modulation,

4. Note that for multilevel (M-ary) amplitude modulation, the symbol error probability is given by P,(E) =
aot erfe V Wa
M=1 “|
V0, (see Chapter 4) and thus the optimum apportionment is still obtained by minimizing
(9.27).
 566 Communication over Bandlimited Channels Chap. 9

Figure 9.3 Throughput performance of M-

OANWHLADNDWO P, ,dB ary amplitude modulation on a raised cosine
0 55 105 15 205 225) 430) 355,740
NoB Nyquist channel

the minimum error probability is obtained by using (9.31) in (9.8), resulting in

1 heed i
P»(E)|min = =5 erfe ,/ No
— 9.32
(9.32)

where we have also substituted 7, for T,. Comparing (9.32) with the comparable result for
the unrestricted bandwidth case (that is, rectangular pulses transmitted over an infinite band-
width channel) as discussed in Chapter 4, we conclude that the two are completely equiv-
alent if we associate P,7; with the energy per bit Ey. For M-level amplitude modulation
with a Gray code bit-to-symbol mapping, the result would be approximately given by (9.32)
with a factor of ,/(M2 — 1)/3 included in the argument of the erfc function. Similarly, the
throughput for M-level raised cosine channels would now be 2 log, M/(1 + @) bps/Hz.
It is of interest now to compare the throughput of the various Nyquist systems as
a function of the SNR in the Nyquist bandwidth, P,/NoB required to achieve, say, a
fixed symbol error rate equal to 10~>. For M-ary amplitude modulation, we have approx-
imately (ignoring the factor of (M — 1)/M in front of the erfc function) that [3/(M@ Pi
1)] P,T;/No = 9.12 (9.6 dB). Thus
P, Pets R; Pare Rp 1
NoB =( No )e 3 (No ee log,M
(9:33)
mmc (Mirek) (M*=1) 18.24
3 (lta) 3 (l+qa)
Figure 9.3 is a plot of throughput R,/B = 2log, M/(1 + a) bps/Hz versus SNR in the
Nyquist bandwidth, P,/NoB for Ps(E) = 10-> with a and M as parameters. Also indicated
on this plot is the throughput corresponding to channel capacity [see Chapter 1, Eq. 1.30)],
which is repeated here (letting C = Rp) as

Rp | PS
p 7 082 (:te i) (9.34)

As an example, suppose we wanted to achieve a throughput R,/B = 2 bps/Hz. If we were

to use a binary (M = 2) system, we would require an ideal brick wall Nyquist filter (a = 0)
with P,/NoB = 18.24 (12.6 dB). If instead we were to use a 4-level system (M = 4), we
could relax the filtering requirement by using a 100% excess bandwidth (a = 1) Nyquist
Sec. 9.2 Optimum Pulse Shape Design for Digital Signaling 567

BEST SAMPLING TIME

A SSAA

AN)
EETSAAAATD gO en ARTGSASEEEFTTT
DISTORTION OF
SENSITIVITY TO ZERO CROSSINGS
TIMING ERROR NOISE MARGIN

DISTORTION AT
SAMPLING TIME

Figure 9.4 The eye pattern and its characteristics

raised cosine filter, but the required P,/NoB would now be 18.24(5/2) = 45.6 (16.6 dB).
Thus, the penalty in P,/NoB for using higher level modulation with practical filtering
is 4 dB.

9.2.3 The Eye Pattern of a Bandlimited Communication System

A useful tool for qualitatively assessing the severity of ISI degradation is the so-called eye
pattern. To see how this comes about, we proceed as follows. Consider the bandlimited
communication system of Fig. 9.1 wherein the transmitted modulation is of the form in
(9.1) and a received signal is of the form in (9.3). For a specific data symbol sequence
{ay}, m(t) takes on a particular waveform. Suppose we consider the superposition of all
m(t) waveforms corresponding to all possible sequences {a,}. The resulting pattern when
viewed on an oscilloscope is called an eye pattern, so-called because of its resemblance
with the shape and characteristics of the human eye (see Fig. 9.4 for a qualitative example
corresponding to binary amplitude modulation). Since the above superposition considers
all possible sequences (each theoretically of infinite length), the resulting waveform is the
same in every 7;-sec interval, that is, the eye pattern is periodic with period T,. Hence, it is
sufficient to examine the eye pattern and its properties only in a 7;-sec interval.
The two most important properties of an eye pattern are its maximum opening (the
largest unobstructed interior vertical distance at any sample time instant) and the broadness
(horizontally) of the opening in the neighborhood of its maximum. The significance of these
two eye pattern measures is as follows.
The maximum eye opening is a measure of the maximum degradation of the signal
component of the detected sample due to ISI. Equivalently, it is an indication of the ability
of the system to protect itself against bit errors when noise is present since, indeed, half the
maximum eye opening is the size of the smallest output noise sample that will cause an error
(assuming that sampling is performed at the point of maximum eye opening). Clearly, the
smaller the eye opening, the more ISI is present and the smaller the noise sample need be to
cause a bit error.
The broadness of the eye in the neighborhood of the maximum eye opening is an in-
dication of how timing jitter tolerant the system is. Clearly, the steeper the eye pattern is
568 Communication over Bandlimited Channels Chap. 9

Figure 9.5a An eye pattern for binary

amplitude modulation (Reprinted from [3])

re
Was; os Cs
8 INTO,
Ny, LBEOSSS
C4 RSX
ESEZESN\
= ZZ, S—S>
<> . —_FA
Se ———, - SS
Sioa F Figure 9.5b An eye P pattern for multilevel
s amplitude modulation (Reprinted from [3])

in this vicinity, the quicker the eye closes with any perturbation around the desired sample
point. Thus, a broad eye is desirable so that an offset in the sampling instant will produce
minimum eye closure (relative to its maximum value) and minimum error probability degra-
dation. Based on this consideration, it is now possible to clarify a statement made earlier
with regard to the reduction in sensitivity to timing offset of the raised cosine channel pulse
response relative to the sinx/x response of the brick wall channel. For the latter, the eye
will be much narrower in the vicinity of the peak because of the slowly decreasing tail.
A more realistic eye pattern diagram that results from a digital stream with pulses
having a shape that extends over many 7)-sec intervals is illustrated in Fig. 9.5a. We observe
that at the point of maximum opening, the eye is “well-focused” with little degradation
due to ISI. Figure 9.5b illustrates an eye pattern for multilevel (MM = 4) pulse amplitude
modulation. Here the maximum opening of any of the M — | = 3 eyes is a measure of the
resistance against making an error (due to noise) by choosing a symbol adjacent to the one
transmitted.

9.2.4 Systems with Controlled ISI—Partial Response

In many transmission channels, bandwidth is at a premium and thus the question naturally
arises: how can we increase the throughput of the system, for example, for a fixed band-
width, how can we increase the bit rate? If we restrict ourselves to zero ISI systems, that
is, channels that satisfy the Nyquist criterion of Section 9.2.1, then the only solution is to
increase the number of modulation levels by going to a larger signaling alphabet. While this
Sec. 9.2 Optimum Pulse Shape Design for Digital Signaling 569

does indeed produce a higher bit rate for a given symbol rate (or equivalently, a given band-
width), the penalty paid is an increase in the required transmitter power. On the other hand,
if we remove the zero ISI constraint and allow a finite and controlled amount of ISI, then
it is possible to achieve a more efficient trade-off between bandwidth and power than that
achievable by the above solution.
Lender [6] was the first to show that using binary modulation but introducing depen-
dencies or correlations between the amplitudes (that is, nonzero ISI) and suitably changing
the detection procedure, one could still achieve the maximum efficiency of 2 bps/Hz with,
however, realizable and perturbation-tolerant filters. The method was called duobinary en-
coded modulation, the prefix duo referring to the fact that the overall system time response
g(t) has two equal nonzero sample values. As such there exists 100% ISI which stems,
however, from only one adjacent symbol. Duobinary is the simplest form of controlled ISI
signaling; hence, we start our discussion with a description of this correlative encoding
technique.

The duobinary encoding technique. Consider the block diagram of Fig. 9.1
where the block labeled “modulation source” that generates (9.1) is modeled as a binary
data source represented by a uniform train of impulse functions at the data rate 1/7, fed to
a pulse shaping filter P(f). Consider now encoding the binary data source prior to pulse
shaping with a simple digital filter as shown in Fig. 9.6. The digital filter delays the input
and then adds it to itself. Thus, the n'" impulse into the pulse shaping filter has a strength
proportional to the algebraic sum of the n“ (current) input bit and the (n — 1)** (preceding)
input bit. As such, the effective encoded bit stream {a/,} = {ay + ay,_1} at the channel input
has dependencies among its members. Note that because of the algebraic addition, the
symbol sequence {a/,} is now ternary (rather than binary) in accordance with Table 9.1.
Note that when a, = —1, there are two possible levels for a}, namely, 0, —2, depend-
ing on the value of a,—1. Similarly, when a, = 1, there are again two possible levels for a,,,
namely, 0, 2, also depending on the value of a,_). We can think of this dependence on the
previous symbol as introducing a controlled amount of ISI. Thus, whereas Nyquist (zero
memory) signaling is based on eliminating ISI, signaling with duobinary or correlative en-
coding schemes is based on allowing a controlled amount (finite memory) of ISI, that is, a
decision on ap based on a threshold test performed on the sum of aj, plus noise depends on
Gn-A-
We can write the signal component of the receive filter output y(t) as

ysth=A )> alg(t—iTs)=A )_ (ai +ai-1)8¢

—iT)
t=—00 t=—0O

[e,¢)

=A )~ [aig(t —iTs) + aig(t — G+ 1)T))] (9.35)

1=—09

(oe) (oe)

=A a aj (g(t —iTp) + g(t —(@i + 1)T))J=A ye ajh(t

— iTp)
i=—0o i=—oo

Thus, as far as the modulation is concerned, the combined transmitter/receiver of Fig. 9.6
can be represented as in Fig. 9.7 where
570
(1)u
(Lt-ad'v’¢v Sd= (f)'s
Ke (LH y odures
1 %7y 14+

ons
ed Dg. *11—
< Ng’ cy tga x yrursuely, 5
pee are Jay UOIST
(f)'9
wyo3Ty

(L0+2)-Dgo<
co-=!

aansIq
9°6 YOo[g WeIZeIP
JO payIWI[PULG UONRITUNUILOD
W}sksYM AreuIgONppepoousaoInos
Sec. 9.2 Optimum Pulse Shape Design for Digital Signaling 571

Table 9.1 Data Encoding

ane ‘ape a

| =
zSoLS

A ¥ a,(t—i7,)

sample at kT, + %

Decision
Algorithm
Figure 9.7 Equivalent block diagram to
Figure 9.6 for modulation

H(f)=G(f) [1at esters _ G(f)e—i7F% [ears a gaint

= 2G(f)cosaf
Ty exp (— jf Ty) (9.36)

|H(f)| =2|G(f)| |cos

x fTp|
Suppose that G(/) is an ideal brickwall (minimum bandwidth Nyquist) filter, that is,
. plies
0; otherwise
Then,

IAP) = |poimretpig
:
:
oie mie
otherwise
(9.38)
We can implement H(f) directly and thus no separate digital filter is required in front of
P(f) as in Fig. 9.6. Rather, as we shall see shortly, we merely split H(f) (equally for
additive white noise) into H,;(f) and H,(f). Note that since no excess (over the mini-
mum Nyquist) bandwidth is used, the efficiency of transmission is 2 bps/Hz. Thus, we are
able to achieve the maximum efficiency for binary transmission using, however, practical
(perturbation-tolerant) filters.
The impulse response corresponding to (9.38) is

h(t) = (9.39)

and is illustrated in Fig. 9.8. Note that h(t) has two equal nonzero samples, that is, one at
t = 0 and one at t = Tp. Thus, as previously stated, the detected samples will have 100% ISI
due to one adjacent pulse.
572 Communication over Bandlimited Channels Chap. 9

h(t)

Figure 9.8 Duobinary pulse shape

Receive
Filter
G,(f)

Figure 9.9 Duobinary receiver with perfect

A decision feedback

If we could always make perfect decisions, we could, in principle, subtract out the
ISI and use a binary decision scheme with a single threshold. The appropriate receiver for
doing this would be as shown in Fig. 9.9. If no error is made on the decision for a,_, that
is, Gk—1 = ax—1, then z_ = yy — AGg—| = Aag + fix and the probability of error for ax would
be 4 erfc (4/V20,) which is the identical result as that for binary Nyquist signaling. On
the other hand, if an error was made on the decision for a,—1, that is, @,_] = —ax_}, then
Zk = Vk — AGg—1 = A(ax + 2ax—1) + nz. Thus, depending on the relative polarities of a,x
and az—1, the double ISI can help or hurt. The problem with this detection scheme then is
that once an error occurs, it tends to propagate further errors, that is, if a,—, is in error, then
its effect is incorrectly compensated for when deciding ax; consequently, a, is likely to be
in error also.
To get around the problem of error propagation, Lender [6] suggested a means of
avoiding the necessity for ISI compensation (subtracting out the single ISI) at the receiver.
His scheme suggested that we precode the data {a;} at the transmitter into the sequence {d;}
as follows. If we think of the binary data as 0’s and 1’s, then the precoder is described by
dn = an ® dyn— 1where the symbol @ denotes modulo-2 addition. Note that a, now indicates
the change in the output symbol d,. If a, = 0, then there is no change in d,. If a, = 1, then
there is a change in d,. This simple encoding scheme is identical to differential encoding
described in Chapter 7 in connection with differential detection of BPSK modulation. The
precoder can also be reinterpreted in terms of “1”’s and “—1’’s for the input bits {ay}.
In particular, the modulo-2 summer is replaced by a multiplier resulting in the relation
dn = an X dy—). Note that d, is still a binary random sequence with equiprobable 1’s and
—1’s.
Using the +1 representation for the a,’s, we obtain the implementation for the duobi-
nary system illustrated in Fig. 9.10. The mean of yz, the sample of y(t) at t = kT, + to,
is now equal to A[dy + dy_]. Table 9.2 gives the values for this mean in relation to the
possible values of the input (information) binary symbol ax.
Thus, we see that when the information bit is a “—1”, then the detected signal sample
is “O” (independent of ay_;). When the information bit is a “+1”, then the detected signal
Sec. 9.2 Optimum Pulse Shape Design for Digital Signaling 573

n(t) sample at kT, +t

5;d,5(t —iT,)
i=.

Figure 9.10 Duobinary system implementation

Table 9.2 Possible Values for the Mean of the Detected Sample

ak dx-| dy A[dy + dy—1] = Y;

1 1 1 2A
1 —] —| —2A
=i) 1 —1 0
Mery 1 0

sample is +2A with the polarity depending on az_. If we examine the magnitude of the
detected signal sample, then, independent of ax_1, the detected output sample will be 2A
when the information bit is a “+1”. Thus, in the absence of noise, we can decode perfectly
according to the following rule:

Decide ax = 1 if |yx| = 2A Decide

a, = —1 if |yx| =0

By choosing a threshold? halfway between “0” and “2A,” then in the presence of noise
we have the decision rule a, = sgn {|yg| — A}. This decision rule involves rectification of
yz followed by a single threshold binary decision. Since no knowledge of a sample other
than yy is involved (that is, the decision rule requires no knowledge of previous symbol
decisions), error propagation cannot occur. The threshold comparison and decision logic
described above can be implemented as in Fig. 9.11.
What remains is to compute the probability of error for the duobinary scheme with
precoding. For a flat AWGN channel, we saw in Section 9.2.2 that the total system response
[in this case H(f)] should be split equally between the transmitter and receiver. Thus, we
choose the two filters H;(f) and H,(f) to be identical and equal to the square root of H(f).
Under these circumstances, the input power to the receiver is

1 (2T))~
5 iP df=a? if =
27, cosnf Thdf
ae [o,2)

P, = A*d? i =(2T, —1 Tp
—0o0
pau! (9.40)
-i(f
Me Tp

The variance of the detected sample, yx, is

5. This choice of threshold can be shown (see Prob. 9.5) to minimize the average bit error probability in the limit
of large SNR.
574 Communication over Bandlimited Channels Chap. 9

sample at kT, +t

Figure 9.11 Threshold comparison and decision logic

ayaaNo flf% APas => (=) ae

No (4
ane
The receiver levels are 0, 2A, and —2A. The level “0” occurs with probability 1/2 and
the levels “2A” and “—2A” each occur with probability 1/4. Thus, the probability of error
is

1
Pp(E) = 7 Pr {—A < ye < Alay = 15k = 2A)]
]
Hey Pr{—A < yx < Alag = 1(ye = —2A)}
|
iy EEE = A Or ye < —Alde
=— Wye OY)

=-—
| Ip 1
exp
(y — 2A)? o
eles /2n0? 203
(9.42)
piper ty y42A)*
oF if exp Ot 5 ) d
Todd Phiters 20;

PY ee as 2
=f / exp
2A lon ae 2a5

In terms of P, of (9.40), (9.42) becomes

Sec. 9.2 Optimum Pulse Shape Design for Digital Signaling 575

3 2 P.T; 2
Pi Eyeene) | eat SEED
4 16 No 4 16 No
(9.43)
3 jihad aed
= 4 erfc ae i for large SNR

Thus, when compared with binary amplitude modulation using ideal (brick wall) fil-
ters to achieve 2 bps/Hz, the SNR penalty (ignoring the factor of 3/4 versus 1/2) is
—10log,y(77/16) = 2.1 dB.
Generalization of duobinary to raised cosine Nyquist channels. Suppose
now that instead of G(f) being a flat channel as in (9.37), G(f) is the raised cosine channel
characterized by (9.25). Then, the magnitude response of the equivalent channel H(/) is
given by the product of (9.25) and 2 cos 2 f Ty. Since g(t) has a single nonzero sample when
sampled at 1/7), then h(t) will once again have two nonzero samples when sampled at the
same rate. Hence, the implementation of Fig. 9.10 and the detection scheme of Fig. 9.11 still
apply here. In fact, an evaluation of the input power, P,, and variance of the detected sample,
eS, produce results identical to (9.40) and (9.41) respectively (see Problem 9.6). Thus, the
probability of error for the duobinary raised cosine channel is also given by (9.42).
What is different then about the raised cosine channel from the flat channel is the
trade-off between throughput and signal-to-noise ratio. Ignoring the factor of 3/4 in (9.42),
then for P,(E) = 10~>, we require (77/16) P,Tp/No = 9.12 (9.6 dB). The throughput is,
as before, R,/B = 2/(1 + a). Thus
Fy. P,Ty / Rp 16 2 16 —)
= pe Nae (9:12 — ne 9.44
NoB No ($) \ (3)(<5) m2 \1l+a )
Once again, we see the trade-off between throughput and SNR as a function of the excess
fractional bandwidth qa. In fact, a plot of R»/B versus P,/NoB with a as a parameter
would be identical to that determined from (9.33) evaluated for M = 2, except it would be
displaced horizontally (along the P,/NoB axis) by 10 log;9(16/27) aed Bille

Other partial response techniques. Generalizations of the duobinary encod-

ing scheme to multiple levels is possible. For M-level duobinary, that is, ay = +1, +3,...-£
(M — 1), one would replace the binary differential encoder with an M-ary differential en-
coder (the modulo-2 summer becomes a modulo-M summer). Asymptotically for large
SNR, the resulting symbol error probability performance of this scheme becomes

| deal BopFe 3
oy ( ras \ 16 (No Nie ) eee
Thus, ignoring the (1 — 1/M7) factor, the SNR degradation for any M relative to M-ary
amplitude modulation with Nyquist pulses is again 10 log (16/77) See) OS:
Aside from extension of duobinary to multiple levels, other generalizations involve
correlative encoding techniques that make use of overall pulse responses with more than
two nonzero samples. These techniques fall into what is commonly referred to as partial
response signaling [5, 10, 11]. A generalized partial response digital filter, T(f), has the
576 ~ Communication over Bandlimited Channels Chap. 9

form of a tapped delay line, that is,

Nz N-1
T(Dy= > 7D T)=T ON py its = nee OO
n=0 =0
(see Fig. 9.12) which, when combined with a Nyquist filter, G(f) produces the overall
response characteristic H(f). By choosing various combinations of integer values for the
coefficients {jy} in T(f), one can obtain a variety of different useful overall responses,
H(f), which have a controlled amount of ISI over a span of N bits. Each of these generates
a different partial response scheme. As in duobinary, the separation of H(f) into T(f) and
G(f) is artificial in that the actual implementation may consist of simply a single transmit
filter and a single receive filter corresponding to an equal partition of H(f). The selection
of the coefficients {7,} is governed by the following considerations.
If H(f) and its first K — 1 derivatives are continuous and the K th derivative is dis-
continuous, then h(¢) decays asymptotically as 1/|r|**!, that is, the larger K is, the less
energy there exists in the tails of h(t) and thus the smaller the ISI is [12]. Clearly, for min-
imum bandwidth systems wherein G(f) is the brick wall Nyquist filter, 7(f) must have a
zero (spectral null) at f =1/2T7) or equivalently, J (D) must have a zero at D = —1 (where
G(f) has a discontinuity) in order for H(f) to be a continuous function. Stated another
way, for H(f) to be a continuous function, 7 (D) must contain the factor (1 + D). In the
more general case, the conditions on J (D) to assure the continuity of H(f) and its deriva-
tives are described by the following proposition. The first K — 1 derivatives of a minimum
bandwidth system response H(f) are continuous if and only if T(D) has (1 + D)¥ asa
factor. Note that as the multiplicity of the (1 + D) factor increases, the number of output lev-
els in the detected signal sample increases and thus error probability performance degrades.
Thus as K increases, a trade-off exists between reduced sharpness of spectral roll-off (easing
of practical filter design) and increase in error probability.
In addition to the K™ order spectral null at f = 1/2T7;, it is often desirable to have
a spectral null at the origin (f = 0), for example, for the possibility of inserting a carrier
pilot tone. To achieve a spectral null at the origin, J (D) must contain the factor (1 — D).
Higher order (1 — D) factors, such as (1 — D)”, cause a more gradual roll-off of frequency
components just above dc, which may also be desirable from the standpoint of allowing
easier extraction of the pilot tone at f = 0 at the receiver. Combining the two spectral null
requirements, we see that J(D) should contain the factors (1 + D) and (1 — D). Most
common partial response systems have a 7 (D) of the form

T(D) = (1+ D)* (1— D)4 (9.47)

Some common examples are as follows:
Duobinary

R= 1. £30), FD) =i1--D

Dicode

KaseOpe Ei SCD

Modified Duobinary (see Prob. 9.7)

K=1, L=1, T(D=Q=D)G+D)=

i027
 577
waysks asuodsai [ensed pezije1ouay 716 ant]
UN
wyyLOSTy
uoIsPDeq
+4 47 ye a[dures
oo—=t
(L1-2)9'0"< v
 578 Communication over Bandlimited Channels Chap. 9

Table 9.3. Characteristics of Minimum Bandwidth Partial Response Systems

Number 0}
Output Leve
T(D) H(f) for |f| < 1/2Ts h(t) (M-ary Inpt
1+D 2T, cos mf Ts Aer a 2M —1
iD j2T, sin xf T; alts eS 2M —1
f= Dp? j2T, sin 2x f T, 2 a 2M —1

(1+ Dy? 4T, cos* rf Ts ibe ea AM = 3

(1+ D)2(1—D) 47, sin2x fT, cos fT; Aa: Gon pS) 4M —3

(1+ D)(1—D)? —4T, sin 2x fT; sina fT; a ee

pt apy 4m 3
(1+ D)?(1— D)? —4T, sin? 2x fT; str SS 4M —3
De). T.(1-+ cos 2a fT; + j3sin2efT) — zp Sonat 4M —3
2 DD T.(—1 + cosdnfT, + j3sin4nfT) xq Canale 4M —3

Table 9.3 gives expressions for the minimum bandwidth H(f) and A(t) corresponding to
the above examples as well as some others due to Kretzmer [5].°

9.3 OPTIMUM DEMODULATION OF DIGITAL SIGNALS IN THE PRESENCE OF ISI

AND AWGN
As we have seen in Section 9.2.4, the presence of ISI in a digital pulse stream introduces
memory into the decision process in that the detected sample y,; for estimating the data
symbol ax depends on knowledge of the previous data symbols az—1, ax—2, .... If the ISI
extends only over a finite number of symbols, then the memory is finite. In the case of
small amounts of controlled (intentionally introduced) ISI, we saw that by proper pulse
shape design and encoding of the data information symbols, we could still obtain a rea-
sonable error probability performance using partial response techniques. In many instances,
the ISI introduced is unintentional and extends over a large number of past symbols, thus
potentially causing severe performance degradation. In these cases, it becomes necessary
to design an optimum receiver with memory which takes into account the presence of both
ISI and the additive noise. The approach to designing such optimum receivers is based on
the maximum-likelihood principle and is referred to as maximum-likelihood sequence esti-
mation (MLSE) [13]. As the acronym suggests, optimum decisions are made on sequences
of information symbols rather than symbol-by-symbol with the goal of minimizing the se-
quence error probability.
Consider the block diagram illustrated in Fig. 9.1 which sends a complex digital data
sequence {c,} over a bandlimited AWGN channel. As before, the transmitter filtering is
accomplished by G;(f) [impulse response g;(t)] and the channel G,(f) [impulse response

6. In this table, for mathematical convenience, we have shifted the time origin to the center of the nonzero
samples, that is, t = (N — 1)T;/2 is the new time origin. Thus, for example, the duobinary frequency response
H(f) is given by (9.36) multiplied by e~/*/7 and h(t) corresponds to (9.39) with t replaced by t + T,/2.
Sec. 9.3 Optimum Demodulation of Digital Signals 579

8c(t)] being bandlimited distorts the pulse shape g;(t) and introduces ISI.’ The received
pulse shape is given by h(t) = g;(t) * g(t). Thus, the received signal plus noise is expressed
as
[e.@)

F)= D> eh nT.) +A) (9.48)

n=—-OW

Suppose now we expand 7(t) in a Karhunen-Loéve series, namely

K
ho= im, 2 Feoe(0 (9.49)
where {¢;(t)} is any set of orthonormal functions (since the noise is white) and
(oe) oOo & ie“)
rk = / 7(t)og(t)dt = Cnh(t —nTs)bg(t)dt + ik A(t)
py (t)dt
—oo —CO n=—00 —CO

(9.50)
lo @)

Gln ie
n=—CO

with
le 2) CO

hkn =4/ h(t —nT;)@j(t)dt; rig =)) A(t) (t)dt (9.51)

06 —oo

The n,;’s are in general independent complex zero-mean Gaussian variables with equal
variance (because of the white noise assumption all eigenvalues are equal)

E {lx | = 26? (9.52)

that is, o* = No/2A? represents the variance of either the real or imaginary part of /i.
Consider the K-dimensional observable rx = (71, 72,...,7K) which depends on the
data sequence ¢ = (..., C2, C-1, Co, C1, C2, . .) through (9.48) and (9.49). Since from the
above the 7;’s are independent complex Gaussian random variables conditioned on ¢, then
the joint pdf of the components of rx is given by
2
]
Mie) 2 2
(9.53)

= ————;
1 exp
Kes ae ae
~ 353 2 Viel

The optimum (in the sense of maximizing the probability of correct detection) decision rule
is obtained from maximizing the a posteriori probability of the transmitted sequence given
the observable, namely, f (¢|rx). Using Bayes rule (see Chapter 3), this probability can be
expressed in terms of the probability f (rx |c) of (9.53). For independent data, maximizing

7. As before, we again assume that G,(f) includes the modulation pulse shaping P(f), and thus g;(t) is, in
reality, the convolution of g;(t) with p(t).
8. Herein it shall be convenient to renormalize the problem such that the signal amplitude A = 1. Thus, the PSD
of n(t) now becomes No/A?.
580 Communication over Bandlimited Channels Chap. 9

f(€lFx) is equivalent to maximizing f(Fx|é). Alternately, since the optimum (maximum

a posteriori) decision rule is: choose the sequence ¢ that maximizes f (¢|rx) for the given
observable rx, then the equivalent optimum (maximum-likelihood) decision rule is:

Choose the sequence é that maximizes f (Fx|€) or equivalently, In f (rx |e).

Taking the natural logarithm of (9.53) gives

2
K 1X
In f(x |@) =—In (270) - 35 aig sani CnNkn (9.54)
=| n=—CO

Since the first term of (9.54) is independent of ¢, then equivalently we can maximize the
function
K
Fx@ = -» rk — 3 eel =-) axl?
ee (9.55)
k= n=—OCoO k=1
In the limit as K — ov, we get

ib iePdr=f limDae oa
Aj oe(t)dt=limDe (9.56)

Thus, taking the limit of (9.56) as K — oo gives

00 PD

F@ = jim Fe@=- f fice? de = — f F(t)— Yo Eq h(t —nTs)

=| rears 2Re|NY =) r(t)h(t - nT (9.57)

a x s Cnc Aly h(t —nT,)h(t

— mT;)dt
n=—C M=—CO

The maximum-likelihood estimate of the data sequence ¢ is the particular sequence that
maximizes F(¢) of (9.57). We can simplify this result still further by noting that the first
term in (9.57) is independent of ¢; hence, it can be discarded with regard to the maximiza-
tion operation. For the second term in (9.57), we define

Yn & W(nTs) = / r(t)h(t —nT,)dt (9.58)

where

y(t) i r(t)h(t —t)dt (9.59)

The interpretation of (9.58) is that it represents uniform samples at t = nT, of the output,
y(t), of a matched filter with impulse response g,(t) = h(—t) excited by r(t).
In the third term of (9.57), define

Ginn) i he — nT ale = mTat (9.60)

Sec. 9.3 Optimum Demodulation of Digital Signals 581

where

g(t) =i! h(t)h(t + t)dt = i h(t’ — t)h(t')dt’

CO
(9.61)
= if h(t’ —nT, — t)h(t' — nT,)dt’
10,9)

The interpretation of (9.60) is that it represents uniform samples at t = (m —n)T, of the

output, g(t), of a matched filter with impulse response g,(t) = h(—t) excited by A(t),
or equivalently, uniform samples at t = nT, of this same output when the excitation is
h(t — mT;). Thus, the set {gm-—n} represents the ISI in the output of the receive filter G,( 9)
in Fig. 9.1 at the sample time t = n7;. ° Note from (9.61) that g(t) is an even function of f
and is in fact the time autocorrelation of h(t). Thus, the Fourier transform of g(t) is given
by G(f) =|A(f)|? and is a real and even function of frequency. Finally, substituting (9.58)
and (9.60) in (9.57) gives the decision rule:

Choose the sequence € that maximizes

ie) e.2) ©O (9.62)

Fo®)=2Re| 9 cn]A Cas
n=—CO n=—-W M=—-CO

We point out that we are not particularly concerned with the noncausality of h(t) since, in
practice, we can introduce a sufficiently large delay to make it causal.
Note that for g(t) corresponding to a Nyquist channel (that is, g(n7,) = 0 for n £0),
(9.62) simplifies to .

Fx@)=2Re |>I cn]— >> lénl?go (9.63)

n=—CO n=—Co

Hence, the decision rule becomes:

> oS ” Fae
Choose the sequence ¢ that maximizes yy (Re{er yn} he leal? so)
n=—OCO

Furthermore, we can write yy = >> Cm&8m—k + x where

Se8 ~ ~

m=—oo

lo.@) (oe)

[Lk =a n(t)g, (kT; — t) dt = n(t)h (t — kT;)dt (9.64)

—0O =X)

is the detected Gaussian noise sample at time k7; with complex correlation function

Ra @—k) = E {ieii'} = 207 g1- (9.65)

For the Nyquist channel, the noise samples are uncorrelated (also independent because they
are Gaussian) and hence the probability of a correct decision for the sequence ¢ is equal
to the product of the correct decision probabilities for each symbol detected independently.
Thus, the above decision rule results in the same average error probability as symbol-by-
symbol detection, that is, in the absence of ISI on an AWGN channel optimum sequence
detection offers no improvement in performance over optimum symbol-by-symbol detection.
9. We assume here that the arbitrary sampling epoch fo = 0.
582 Communication over Bandlimited Channels Chap. 9

9.3.1 Evaluation of the MLSE—The Viterbi Algorithm

The brute force method for determining the optimum sequence ¢ is to exhaustively try
all possible sequences!® in (9.62) and then select the one that yields the largest value of
Fo(@). A more efficient method for processing the sampled complex matched filter outputs
{Sn} is the so-called Viterbi algorithm (VA) [7] which was originally proposed for optimum
decoding of convolutional codes. This algorithm performs maximum-likelihood estimation
of a digital sequence and thus is often called a maximum-likelihood decoding algorithm
[14]. Note again that the optimum decision rule in the presence of IS] and AWGN makes a
decision on a sequence € rather than symbol-by-symbol detection, which is a suboptimum
decision rule. The MLSE technique is capable of producing performance approaching that
of an optimum symbol-by-symbol demodulator operating in the presence of AWGN alone.
The VA may be viewed as a solution to the problem of maximum a posteriori (MAP)
estimation of the state sequence of a finite-state discrete-time Markov process observed in
memoryless noise. Before starting a formal explanation of the VA, we must first cast the
MLSE in the presence of ISI and AWGN in this form. The formulation for accomplishing
this is due to Forney [13, 15]. For the purpose of our discussion, we shall assume that the ISI
exists over a finite time interval corresponding to v discrete times on either side of go, that
is, g; = 0 for |/| > v. Although g(t) is bandlimited and thus cannot also be timelimited, the
assumption of a finite length is a reasonable approximation in most practical applications.
We begin by first writing the detected sample yz of (9.58) as

Je = D> Cx-1g + ite (9.66)

l=-v

The discrete-time process {yx} is called a shift register process since it can be modeled by a
shift register of length 2v whose input is the sequence ¢ (see Fig. 9.13a). The problem with
the representation of Fig. 9.13a is the correlation of the noise samples [see (9.65)] which
makes the performance analysis of the MLSE difficult. To circumvent this problem, we shall
introduce a noise whitening discrete filter (see Chapter 4) after the sampling operation in
Fig. 9.1 so that the detected samples will take the form of (9.66) with, however, uncorrelated
Gaussian noise samples.'! The noise whitening filter is determined as follows.
Since from (9.60) and (9.61), 2m—n = 8n—m, then gx can be expressed as the discrete
convolution

v—k

ge) Whee (9.67)

1=0

where f; = 0 fori <0 andi > v. Note that 9, = 0 for |k| > v as required above. In terms of
the z-transform of these sequences, namely

10. This, of course, only has practical meaning if the sequence length and thus the number of possible sequences
is finite. We shall be more specific about this shortly.
11. The introduction of a noise whitening filter is made merely for convenience but is not necessarily required
to implement the Viterbi algorithm. A description of the operation of the Viterbi algorithm for combined ISI and
AWGN channels without using a noise whitening filter is given in [17].
Sec. 9.3 Optimum Demodulation of Digital Signals 583

Figure 9.13a Shift register representation of bandlimited system in the presence of ISI
and AWGN

OSE TS NCO ie (9.68)

k=-v k=0
we have G(z) = F(z)F(z~!). Since the sequence {f;} is causal, F(z) is a physically realiz-
able digital filter but F(z~') is not. Despite the fact that F(z~!) is not physically realizable,
we can still choose it to be minimum phase. By doing so, its inverse 1/F(z7') is a physi-
cally realizable digital filter. By selecting 1/F(z~') as the noise whitening digital filter, then
passing ), through this filter produces the sample ,, which can be expressed in the form
v

He = > eeufi+ Hy A tye + ity (9.69)

H=40)

where ji, is an element of a zero mean white Gaussian noise sequence with variance
E ||; ‘|= 20°. In terms of the sequence {5}, the decision rule of (9.62) can be rewritten
as
Choose the sequence ¢ that maximizes

(o @) (o@)

Fo(€) =2Re wry} — \Wnl?

ost ak d i (9.70a)

9)
v

=aRe (S22) ate 9

[e.6) Vv [o,@)

De oie
ni=—CcoO i=0 n=—OO i=0

or equivalently minimizes
CO [o.2) v
2

Fes). |i a,| = >-

~/ ~ ~
Yn 9 y° Cn-i Si (9.70b)
n=—CO n=—OOoO i=0

The j,’s form a set of sufficient statistics for estimation of the input data sequence ¢ [15].
584 Communication over Bandlimited Channels Chap. 9

Figure 9.13b Shift register representation of bandlimited system in the presence of ISI
and AWGN (noise whitening filter included)

Figure 9.13b is a model of the shift register process corresponding to (9.69). !2 Note that the
length of the register is now equal to v in accordance with the duration of the f; sequence.
Also, {5x} is now a v" order M-ary Markov process.
The state, x, of the process is defined by

Kee (Chetek 2s Chow)

arid the transition (change of states), &&, at discrete time k is defined as the pair of states
(Xk+1, Xk), Or ifi terms of the inputs!3

Ep = (Cp, Chea Ch 2) en Cep)

Thus, the maximum number of states is M” and the maximum number of transitions is
M°’*!. Note that there is a one-to-one correspondence between sequences of states and
sequences of transitions. Also if the input sequence “starts” at time zero (that is, Co is the first
transmitted symbol) and “ends” at time K — v — | (that is, Cx—,— is the last transmitted
symbol), then the contents of the shift register in Fig. 9.13 is zero prior to the start, that
is, the state xp = (C_1, C=, C—3, +.., Coy) = (0,0, 0, ... ..,0);-ahd is again zero’ Ki symbols
later; that is, the state Xn" = (Cx). Ceo; CK 3, os CR) nO OO NO).
Finally, in the context of the above, we can restate the problem to which the VA
is a solution. Given a sequence y of observations of a discrete-time finite-state Markov
process in memoryless noise, find the state sequence x for which the a posteriori probability
f (&|¥) is maximum. Alternately, since x and the transition sequence € have a one-to-one
correspondence, we can restate the problem as: find the sequence transition € for which
f (é\y) is a maximum. For the shift register model, this is the same as finding the most
probable input sequence ¢€ since x and ¢ also have a one-to-one correspondence. This MAP

12. Since herein we shall be dealing with the respresentation of (9.69) rather than (9.66), for simplicity of
notation, we shall drop the prime on jy, and ji; .
13. Despite the fact that the state and transition are defined by sequences, it is customary not to denote them with
a boldface typeface.
 Sec. 9.3 Optimum Demodulation of Digital Signals 585

Aes Figure 9.14a State diagram of a four-state

shift register process

10

11

Figure 9.14b Trellis diagram of a four-state shift register process

decision rule, which minimizes the error probability in detecting the whole sequence, is thus
optimum in this sense.
Associated with a discrete-time finite-state Markov process is a state diagram in
which nodes represent states and branches represent transitions. As discrete time progresses,
the process traces a path from state to state in this diagram. An example of such a state dia-
gram is illustrated in Fig. 9.14a for a binary!* system with ISI memory v = 2. Since M = 2,
there are M” = 4 states and a total of M’*+! = 8 transitions among them. Note that from
any given state, there are only two allowable transitions to the next state. For example, if the
current (at time k) state is “Ol” corresponding to a,_; = 0 and ay_—2 = 1, then the next state
will be “OO” if a, = 0 or “10” if ax = 1. Note further that the structure of the state diagram
depends only on the input sequence a and the memory of the ISI, not the values of the ISI
(shift register tap weights) themselves.
An alternate and more redundant description of this same process is a trellis diagram
[14], so-called because of its resemblance to a garden trellis. Figure 9.14b is the trellis
diagram corresponding to the state diagram of Fig. 9.14a. The trellis diagram illustrates
the progression of the process with time, a property that is lacking in the state diagram
representation. Each node in the trellis diagram corresponds to a possible state of the state
diagram at that particular instant of discrete time. Thus, the maximum number of states at

14. For this example, C, is real, that is, ¢, = a, and takes on values 0 or 1 with equal probability.
586 Communication over Bandlimited Channels Chap. 9

any particular time in the trellis is equal to the number of states in the state diagram. In fact,
the only time the trellis will have fewer than this maximum number of states is for the first
few time instants at the beginning and at the end of the input sequence. Each branch in the
trellis diagram represents a transition from a particular state at a given time to a new state
at the next instant of time. The inital state (I) of the trellis corresponds to Xo and the final
state (F) corresponds to Xx, both of which will be “00” if the initial contents of the shift
register was empty. Associated with each sequence of states, X, is a unique path through the
trellis and vice versa. Since the most probable sequence is the one for which f (x|y) [or
equivalently f (x, y) = f (xly) f (¥) is maximum, if we assign to each path x the quantity
—In f (X, y), then since the negative logarithm is a monotonically decreasing function of its
argument, the most probable sequence is also the one for which — In f (x, y) is minimum.
We shall see in a moment that we can attach to the quantity — In f (x, y) the significance of
path “length.”
Since the process y is observed in memoryless noise, then
K-1
f 91%) = F (HE) = [TF (Sul&e) = TT] f Celie, 9.71)
k=0 k=0

Furthermore, since the states obey a first-order Markov behavior, that is,

F Xxss[h0, x1, -.- 5%) = Sf Oni |e) (9.72)

then since the initial state occurs with probability one, that is, f (xo) = 1, the probability of
the state sequence becomes
Kea
eoaa NAc eT (9.73)
k=0

and thus the joint probability f (x, y) is given by

K-1 Kel
fRN=f Mf G® = [] FfGerled) [|] FfGelFer1,%) (9.74)
k=0 k=0

Taking the natural logarithm of (9.74) converts the products into sums; thus
K-1 K=1
=In fH) =— Yn f Geile) +10 f Guléev, 40) = 2 (E) 0.75)
k=0 k=0

where the weight A (&) is assigned to each branch (transition) and represents its length.
Hence, the sum in (9.75) represents the length of the entire path.
Finding the shortest length path through the trellis [path with minimum — In f (X, ¥)]
in an efficient way is what the VA accomplishes. Consider a segment of the state sequence
X = (x0, %1,..., XK) Consisting of the states from discrete time zero to discrete time k, that
iS, xf = (Xo, X1,..., Xx). In the trellis diagram, xh corresponds to a path from state x9 to
state x,. For any of the M” (or fewer) nodes at discrete time k, there will be in general
several such path segments entering that node, each with a length

2 (%§) = (&) (9.76)

Sec. 9.3 Optimum Demodulation of Digital Signals 587

that depends on the particular transitions €;; i=0,1,...,k—1 along that path segment.
One of these path segments, x (X;), entering the node will have the shortest length T (x,) A
min A (x6) = (x(%)). This particular shortest length path segment is referred to as the
survivor corresponding to that particular node. The reason for this terminology stems from
the following observation:

The shortest state sequence x (the one we are looking for) must at any discrete time k along its
path correspond to a survivor path segment.

If this were not true, then at any node along X we could replace the path segment from
state xg to that node by the survivor and thus achieve a smaller length which leads to a
contradiction that x is the shortest length path. The result of all this is that at discrete time
0 <k < K, it is only necessary to save (keep track of) the survivor paths and their lengths,
one each for the M” (or fewer) nodes at that time instant. In proceeding to discrete time k +
1, one merely appends the survivors with the branches leading to the next set of states and
increments the survivor lengths with values related to the corresponding transitions. At time
k + 1, one again chooses the M” survivors and the process continues recursively. Finally,
at time K, the trellis converges to a single node with one survivor that must correspond to
the shortest state sequence x being sought after. The length, I" (x) ule (x(x x), of this
shortest path must from (9.76) be equal to the minimum value of (9.75) as desired.
Figure 9.15 illustrates the above recursive evaluation of the shortest path for the trellis
diagram of Fig. 9.14b and a value of K = 5. The values assigned to the trellis branches are
not intended to be typical of any particular application but are merely selected for simplicity
of the distance evaluations. Since M = 2 and v = 2, there are 4 (or fewer) nodes and 8 (or
fewer) transitions at each discrete time. At each of these nodes, the solid lines entering
the node indicate the survivor path while the dashed lines indicate the nonsurvivor paths.
Finally, the shortest path through the trellis has length equal to 3.
Although the VA as described above will theoretically find the shortest length path
through the trellis, the presence of very long or infinite state sequences requires a buffer
memory beyond practical limits. Hence, in practice, one must invoke a modification of the
algorithm wherein survivors are prematurely truncated to a manageable length, say 5. That
is, at time k the algorithm makes a final decision on the first k — 5 branches of the shortest
path (that is, nodes up to time k — 4) and retains only the length and path information
on the most recent 6 branches (the truncation depth) of the survivors. The ability to make
such a premature decision relies heavily on the extent to which the time-& survivors share a
common path for the first k — 5 nodes. For example, in the example of Fig. 9.15 we observe
that at k = 4, the two surviving paths shared a common path up to time k = 2. Since one
of these two survivors will ultimately be selected as the shortest path, nothing is lost by
making a final decision at time k = 4 on the first two branches (5 = 2) and truncating the
survivors to two branches. Clearly, for sufficiently large 5, there is a high probability that all
survivors at time k will pass through the same set of nodes for the first k — 6 time instants.
In the rare cases where survivors disagree in their first k — 6 branches, one must adopt a
time-(k — 5) decision strategy that has negligible effect on the ultimate error probability
performance. One such possible strategy is to choose the time-(k — 6) node corresponding
to the shortest time-k survivor hoping, of course, that this survivor will eventually wind up
588 Communication over Bandlimited Channels Chap. 9

I(x, =1) =2

T(x, =2) =3

I(x, =3)=2

P(x =1)=2 T(x, =4)=1

I(x, =2)=2

I(x, =3) =3

P(% =4)=2

I'(x,=1)=2

I(x, =2) =3

FG; =)=I)=35

Figure 9.15 Recursive determination of shortest length path for four-state trellis diagram with K = 5
Sec. 9.3 Optimum Demodulation of Digital Signals 589

to be the shortest length path. Another possibility is to choose the time-(k — 5) node by a
majority vote, that is, the one through which most of the time-k survivors pass.
From a computational point of view, as time k gets large, the values of survivor path
length, T (x,), for each of the M” states also become large and thus from time to time it is
necessary to subtract a constant from all of them.
A few words about complexity are in order. Since the algorithm requires M” storage
locations, one for each state (after the initial transient) capable of storing a survivor path
and its length, then from a memory standpoint, the complexity grows exponentially with the
length v of the shift register. The complexity associated with the computation of the incre-
mental lengths, (&), is assessed as follows. Since from (9.74), one component of A (E.)
is equal to — In f (X,+1|x;), then since for a shift register process f (%4+1|X,x) is either equal
to 1/M or 0 depending upon whether state x; +) is allowable after the occurrence of state x;,
all allowable transitions have the same value of —In f (x,+1|X,). Thus, this component of
Xr (&) may be ignored. The computation of the remaining component of A (&), namely,

In f (Silke) depends on the data sequence. However, from Fig. 9.13b and (9.69), we ob-
serve that, in general, many different values of & produce the same value of shift register
output wz. Since yy, = we + fx, then for a given yx, all of these values of & produce the
same value of Gaussian probability f (5xl&) and thus — In f (nlx) need be computed
only once. Because of the above, the complexity associated with the computation of the
A (&)’s tends to be insignificant relative to that associated with the memory requirement.
Example 9.1
To illustrate the application of the Viterbi algorithm specifically to the ISI problem, consider
a binary PAM (cx = a, = +1) system with memory v = 2. The appropriate representation of
the system is given in Fig. 9.13b where the shift register now has only two stages. The state,
the transition, and the shift register output are real and are given by x, = (ag—1,.Ax—2), &k =
(Ap, Ak—1, Ak—2), and wy = foag + fiax—1 + foax—2, respectively. The state transition diagram
for this shift register process is identical in form to Fig. 9.14a. We redraw it here in Fig. 9.16a
where we have now labeled the branches with the input (a,)/output (w;) corresponding to each
transition. The conditional pdf f (yz |&%) is Gaussian and given by

l (ye — we)?
f (ye 1E&&) == F OK we) = a exp p 4————_——_
No
u (9.77)
ete ale [yx — (foak + fiar—-1 + frax—2)]
% /m No P No

Taking the natural logarithm of (9.77), we get

4
1 — (foax+ firag-1+ frax—2)|
In f Olé) = Ino + ue i0 wake) (9.78)
Since the first term in (9.78) is independent of the data, then dropping the 1/ No scaling constant
in the second term, the normalized incremental length (weight) to be assigned to transitions
between the k'" and k + 1* state in the trellis diagram is [yx — (foax + fiax—1 + fra»),
that is, the squared distance between the output signal plus noise sample and the shift register
output. Figure 9.16b is an illustration of the appropriate trellis diagram (identical in form to
Fig. 9.14b) where the branches have now been labeled as per the above.
590 Communication over Bandlimited Channels Chap. 9

OUT 6

Figure 9.16a State diagram for binary

Wfythth transmission with ISI memory v = 2

ly.- fl’ Yi :
State , k=2 vans se AER So ha oI k=3
0,0 eocee

0,1 eoce

1,0 eoce

iLJt eoce

ae Di- Got ft Ay
D2-Goth }

Da-G+h) %,-(¢ +4
Figure 9.16b Trellis diagram corresponding to state diagram of Figure 9.16a
Sec. 9.3 Optimum Demodulation of Digital Signals 591

9.3.2 Error Probability Performance

To evaluate the error probability performance of the VA, we introduce the notion of an
error event at time k, which is defined based on the following observation. Consider that the
correct path through the trellis (that is, the path corresponding to the transmitted sequence
¢) is defined by the state sequence x. Suppose, however, that the shortest distance path
computed by the VA corresponds to a different state sequence, say X. For sufficiently large
sequence lengths, the paths x and x will agree over much of their length with an occasional
discrepancy wherein X deviates (diverges) from x for a number of branches and then returns
to (remerges with) x. These deviant paths are referred to as error events. Along each of
these deviant paths (unmerged segments), the accumulated weight (metric) is larger than
that along the correct path. When the time instant at which a deviant path begins corresponds
to time k, we refer to these paths as error events at time k. Assuming that the error event
terminates at time k + L, then for such an error event, we have x; = aie See thy L, and
te Xi; k+1<i<k-+L-—1. The length of the error event is defined as being equal to
L. Clearly L > v + 1 but has no upper bound, although it can be shown [13] that L is finite
with probability 1. Letting e = x — X denote the total error event path between x and x, then
é is composed of a number of error events, each of which starts at a different time instant
along the path and, in general, have different lengths and number of branches with respect
to one another and different probabilities of occurrence. Note that the weight assigned to an
error event branch is the difference of the weights assigned to the corresponding branches
of X and x.
For a given time instant k, there are many possible paths that diverge from the correct
path at the node x, and remerge for the first time with the correct path L branches later at
time k + L. When one of these paths corresponds to an error event, the probability of its
occurrence relative to the correct path is computed as a pairwise error probability, that is,
the probability that the segment of x between time k and time k + L is more probable than
the corresponding segment along x between the same two time instants. For a given input
sequence, this pairwise error probability can be evaluated as a closed form solution and is
a decreasing function of L. Evaluation of the pairwise error event probability is essential to
computation of error event probability at time k and also average symbol error probability
since, as we shall soon see, upper union bounds on these quantities can be obtained as
appropriately weighted sums of all pairwise error probabilities. Thus, we now evaluate
the pairwise error probability for the more general case of the correct sequence and any
incorrect sequence (not necessarily corresponding to an error event).

Evaluation of the pairwise error probability. Consider two paths x and x

which are identical up to time k and from time k + L on but are not constrained in any other
way between times k and k + L. We have seen that the metric used for choosing one path in
the trellis over another is the sum of the squared Euclidean distances between the detected
sample j, and the shift register output w, along all of its branches. Let mx = |yx — wel?
Ae? ’ ; >
and 7, = |
Ve — ie denote the above squared Euclidean distances at time k where wx and
Wk respectively denote the shift register outputs corresponding to the correct and incorrect
paths. Then, given that ¢ is the transmitted sequence, the probability of choosing x rather
than x, denoted by P(x — x|C), is given by
592 Communication over Bandlimited Channels Chap. 9

P(X > X|® =Pr

|
keL- | [re(5 (y= 4)"}
:Gs = |p
ie | (9.79)

|
Note the similarity between (9.79) and the metric used in the decision rule of (9.70).
To evaluate (9.79), let

x= )* Re {5 (i, stiin) } (9.80)

The random variable X is Gaussian with mean

Keel
eae Re {iin(in — tin)} (9.81)
n=k

and variance

k+L-1 7 aS k+L—1 5
ot =E (d> Re {iin(ton — tn) ) =o? > |tin—tn| (9.82)
n=k n=k

Hence, the pairwise error probability is evaluated as

a mee . i =
P(K—> 1® =Pr{X>y|é}=~erfe| ” (9.83)
2 202
VCE
where
y kt! >
+ ~ 12
ya-2 ({i, — |Wn| ) (9.84)
n=k

But, from (9.81) and (9.84) we have that

nlp Rlaaa D0 hg
yi Dy X=Ces E.5 2 pee |)
Ve Se
AG? (9.85)
Sec. 9.3 Optimum Demodulation of Digital Signals 593

Thus, (9.83) becomes

a 1

(9.86)
kK+L—-1| v - Z 2
Sa ph eF ees = cri)
= ae n=k_ l|i=0

D, 802

Since the correct and incorrect paths share the same set of states for discrete times less
than or equal to k and greater than or equal to k + L, that is, x; = ti BSS kao
then from the definition of the system state in terms of the input data sequence, we have that
C; —c;=0; i<k—1, i>k+L -—v. Then, in terms of the definition of the actual ISI
coefficients given in (9.67), we can rewrite (9.86) as

P(K > X|®) = (9.87a)

RAE=1 Aa
v , ‘i
5erfe aa ae c Cn — Cn * +2Re poe (er ~ in) (c= ~txi)'}
n=k i

which in view of the above properties of c; — Cj can be rewritten as

P(K > &|®) = (9.87b)

1 1 k+L—v-1 in 2 v k+L—v-1 de &: A - *

= erfc ame 2 80 |Cn — Cn| +2Re i 8 ya (é, a én) Ge ae ii)

2 80 n=k jab nek
Despite the fact that (9.86) is in closed form, it is convenient from the standpoint of
average symbol error probability evaluation to find an upper bound on P(X — X|€) in the
form of a product of terms. To achieve this desired form, we overbound the complementary
function by erfc y < exp {- y*} which when used on the right-hand side of (9.86) gives
A 7
y k+L—1 Wn — Wp

n=k

Aid} ie 2 (9.88)
ae _ Yatfi (Ga
(Cnt = n Cri )

2 80?

This result (except for the factor of 1/2) can also be obtained by applying a Chernoff bound
directly to (9.79).
594 Communication over Bandlimited Channels Chap. 9

Evaluation of an upper bound on the average symbol error probability

for PAM. We consider here the case of PAM wherein c; and ¢; are real, that is, a; and dj,
respectively, and take on values from the set +1, +3,... + (M — 1) where M is an even
integer. Also, the real sequences X and x are now such that they correspond to a single
error event in the time interval k to k + L. We define the input error sequence &, associated
with the error event by & = (Ec,k—vs Ec,k—-vtly +++ Ecsks Echt1 «++» Ec,k+L—1) Where &¢,; A
(4; — a;) /2and e.; =0; i<k—1,i>k+L—v. 6 has dimension!> L + v and has the
property that it contains no sequence of v consecutive zeros (else we would have x; = x;
for some intermediate value of i between k and k + L resulting in two distinct error events).
Also, the elements of ¢, take on values from the set 0, +1, +2,..., +(M — 1). Associated
with the input error sequence &, is a signal error sequence &,, defined by the corresponding
sequences of shift register outputs, that is, ¢y = (wk Ob RAs wes Cu Rees myers)
where éy ;& (0; — wi) /2. €w has dimension L. The squared Euclidean weight of an error
event, d2 (E), is defined as the squared norm of the associated signal error sequence, that is,

PO= Dv ban D(A) =D (Liens

k+L-1 FELD pay soy: \oe hea (a g

n=k n=k n=k i=0

(9.89)
ee Gd | v
= 2 2 : ;
= UNS SiEcnEc,n—-i
n=k t=1

The number of errors in &,, that is, the number of nonzero elements in &,, is defined as the
Hamming weight wy (€) of the error event.
Following Forney [13], we consider a particular error event € as being composed of
three subevents, €,, €2, and &3, all of which must occur in order for € to occur. Subevent €
corresponds to the requirement that both the incorrect and correct path must have the same
state at time k, that is, x, = Xz. Subevent 2 requires that for 0 <i < L — v — 1, the input
sequence must be such that, for the particular error sequence assumed, ¢; = 2€,,; + cj is an
allowable error sequence, that is, it takes on values from the set +1, +3,...+(M — 1). Fi-
nally, subevent €3 corresponds to having noise terms nj; k <i <k +L —1 such that the
likelihood of the segment of X over this interval is greater than for the corresponding seg-
ment of x. The probability of €3 occurring, Pr {3}, is precisely the pairwise error probability
discussed in the previous section applied now to the specific case where the incorrect path
corresponds to an error event.
Based on the above partitioning of € into €), €2, and €3, we can write the probability
of the particular event € occurring as

Pr{ép= Pr{&) | 3} Pr (Es) Prfeoicy.

ea} (9.90a)
However, since €2 is independent of €|, and €3 being only dependent on the message
ensemble, then (9.90a) can be simplified to

Pr {€} = Pr{ €]| 3} Pr {&3} Pe {&>} (9.90b)

As mentioned above, the subevent probability Pr {€3} may be computed from the
exact result in (9.86) or the upper bound in (9.88). In particular, using (9.89) in these

15. Actually, €- has a true dimension equal to L — v in terms of the number of elements that may or may not be
zero. However, it is convenient to define the extended sequence &, as we have done.
Sec. 9.3 Optimum Demodulation of Digital Signals 595

equations gives

Bs ] é
Pr {€3} = P(x > x|e,) = — erfc au)
2 202

and
v Ds
er 2 elo ( fitcn-i)
Pr {3} = P(x > Sie) < 5 Ul exp
1 |4 -»- 1 i=0

262 a ae 202

v
1 ate 80ES n +2 d 8iEc.nEcn—i
i=

52 Jide a (9.92)
n=k 20

where we have now changed the conditioning event to the error event sequence. It is con-
venient to normalize (9.91) and (9.92) in terms of the received signal-to-noise ratio. From
(9.60), we have

w= | me —ntar= | n(pras = f IG(A)P IGA)? af (9.93)

Analogous to (9.31), the average received power for M-ary PAM is
M2 pat Az (ove) M2 will
P, = —AN /ae IGA)? IGA)
Pdf=
AZ
s0 .94)
0

3 3 T.

Since o* = No/2A?, then solving for A? from (9.94) and substituting into (9.91) and (9.92)
gives

M2—1) No
and
: k+L—1 3 Per 5 2

ee ey NPN ge) ox (>fiseri

(9.96)
1 ktLe! 3 pT, ;
I] ee M2=] No eon anid Me 8iEc nEcn—i
i 9
n= =I

where, as earlier in the chapter, we have assumed go = | without loss in generality.

To compute Pr {€2}, we observe that the subevent £2 is independent of €) and &3.
Specifically, for any particular input error sequence of length L — v, in each of the L —
v — wy (E) positions where &,,; = 0, there are M possible values for the correct sequence
element and likewise for the incorrect sequence element. In each of the wy (€) positions
where &.,; #0, there are M — lec,i| possible combinations for the values of the correct and
incorrect sequence elements. Thus, if /,,,, denotes the set of values of i where €,,; # 0, then
596 Communication over Bandlimited Channels Chap. 9

the number of input data sequences that produce a particular error sequence with Hamming
weight wy (E) is M'~’-“#©) TT (M — |ec,;|). Since the total number of possible input
i€lwy
data sequences of length L — v is M/~”, then the probability of subevent €2 occurring is

Pr (Ea)-= MT" [|] (M— |ec.i|)

1Elw py

(9.97)

TT (SE) (Ge
k+L—v-1 M — |ec.i| Reo M — |e,i|

i=k r=K

Finally, computing the conditional probability Pr {€)| €3} is difficult because of the
possible dependence of €; on the noise terms involved in €3. However, the unconditional
probability that €; does not occur, namely, 1 — Pr {€)} is on the order of the symbol error
probability. Thus, in the normal operating region, the conditional probability Pr {€;| €3}
is closely approximated as well as overbounded by 1. Substituting this upper bound on
Pr {€)| €3} together with (9.97) and (9.95) [or (9.96)] into (9.90b) gives the desired upper
bound on the probability of the particular event €, namely

1 3 P,T RL! (iM —|ec,i| Ct

Pr {€} < Bes
5 erfe (eS No d ©) I] ("of
ee Ke +S 2

1 HLT IM — ee.n| 3 pT, (e 2

woelt (“fe exp Ge :) i de fiken- (9.98)
n=k

_
1 k+L—1 M — lec 3 BTA
I ( M oa es (szcP :) No Eon + 2 ~- 8iEcnEc,n—i
F 2

For binary PAM where e,.; can only take on values 0, +1, (9.97) simplifies to
1 k+L—1 1

Qwu(E) ae I] 2H (Ec,n) (9.99)

Pr {&2} =
hye

where wy(Ec,n) is the Hamming weight of e€.,,, that is, wH(€on) = 1 if en 40 and
WH(Ec.n) = 0 if &¢., = 0. Thus, (9.98) becomes

Pe(e} < sete ( No d ©)sane

l > P,Tp 2 l

j tee 1 1 2,

(9.100)
Ves is =
= I] SACHS ass ene (35fron
r=

pRticl 4 PT» | 5 r
cs) I] wu (Ec,n) Dn No Een +2 y scat
n= i=1
Since the probability of a union of events is less than or equal to the sum of their
individual probabilities, if E denotes the set of all possible error events € starting at time k,
Sec. 9.3 Optimum Demodulation of Digital Signals 597

then the probability of any error event starting at time k is union bounded by

Al k+L—1 Abt f
P(E) < ye 5 erfic(Gea)7
PTs ©] leh (“5
EEE 1=k

1 k+L—1 Me een 3 2
= 2.5 I] (“f= exp Ga) NoE (55fee )

SIGS ~ leen|
= 5 nt (“ie (9.101)

x exp |-(@-)
Ne 9oeeal es
me 2,+
2
2 2 Sitenton-)

Similarly, the average probability of a symbol error is obtained by weighting each error
event probability by the number of decision errors wy (E) resulting in the upper bound

3 Puli iL ed) Mle

P(E) i< So wi jee C= -)7aa))
= I] (“S24
EEE t=K

y KtLel M om ae

K De WH ©)5 fel (“SF)

EEE n=k

2
3 Pe tuhes
x exp = (Ga—3) No (>:
fist (9.102)

k+L-1 tee
= wn); I] (* a
ECE n=k

s eles -
x exp | (Gz a7 :) oa le,2 Ptieatent|
hell

Evaluation of (9.101) or (9.102) is facilitated by the use of a so-called error-state

diagram. The error-state diagram is similar to the conventional state diagram, the ma-
jor difference being the labeling of the transitions. In particular, we define an error state,
&;, at time i as the difference between the incorrect and correct states X; and x; at the
same time instant, namely, ¢; 4 X; — x; = (Eo j=1; Bate Mea &c,t—): Since ¢,; takes on
values 0,+1,2,...+(M — 1), then the maximum number of possible error states is
(2(M — 1) +1)” = (2M — 1)”. The initial and final states are the all zeros state since by
definition the correct path and incorrect path start and end at identical states. There are
2(M — 1) branches leaving the initial all zeros state (at time k). To each of these branches,
we assign the weight ao,;/ (M — |€c,x|) /M where €¢.4 = j; j =+1,+2,...,(M — I) is
the input error that causes the transition to the next state, / is an indicator variable whose
presence indicates that the next state is not the all zeros state, and ao,; is the exponen-
tial term in the product of (9.102) corresponding to e, = (Se.221, So RED ehion Sky) =
598 Communication over Bandlimited Channels Chap. 9

(ORO O) thats
Bree
ay,j= exp {-(@—) ele
7a 2, | (9.103)
There are 2(M — 1) branches entering the final all zeros state (at time k + L). To each of
these branches, we assign the weight | since the input error €¢,,4z—1 that causes the tran-
sition to this final state is zero (thus (M — |éc,x4z—1|) /M = 1), J is absent (since the next
state is the all zeros state), and the exponential term in the product of (9.102) correspond-
ing to €¢.44,—-1 = 0 is unity. To the remaining branches between error states, we assign the
weight an—«,j1 (M — |écn|) /M where &¢,n = j is the input error that causes the transition
and aj, is again the exponential term in the product of (9.102) corresponding to &¢,n = j
and the current state ¢, = (GRSEi, yee pea ay that is,

3 P,T,
Tso: -
An—k, j= €XP |-(@—) No le,+2 De tren (9.104)
i=

A pictorial representation of the error-state diagram rapidly becomes complicated as

v grows. Even for M = 4 and v = 1, there are already 7 states in the diagram. Thus, for the
purpose of example, we shall restrict ourselves to the binary case with v = 1, in which case
the number of states is 3. In terms of the above, there are two branches leaving the initial all
zeros state each of which is assigned the weight ag/ /2 where now

P,Th
ay = exp {- (9.105)
No
There are two branches entering the all zeros final state each of which is assigned the weight
1. The branches between other pairs of states are assigned weights a_//2 where

as = exp |
Tpie
- hi (1+ 2¢| (9.106)
0
where the plus sign corresponds to a transition from a given state to the same state, that is,
1 to 1 or —1 to —1, and the minus sign corresponds to a transition from a given state to the
opposite state, that is, | to —1 or —1 to 1. The complete error-state diagram is illustrated in
Fig. 9.17.
To evaluate the error event error probability in (9.101) or symbol error probability in
(9.102), we must compute the generating function of the error-state diagram, which can also
be regarded as the transfer function of a signal flow graph with unity input. This function can
be most directly computed by simultaneous solution of the state equations obtained from
the error-state diagram itself. In particular, from Fig. 9.17 letting &,, &, and &4 represent
dummy variables for the partial paths to the intermediate nodes, and T (ao, a_, a4, 1) = €g
the desired generating function, then with €, = 1 (since the graph has unity input) the set of
state equations is given by

T (ao, ax, ay; 1) =& + &,

I I I
Sb = a0; gee nh tte ee (9.107)
I I I
ge a0 + 4458 oh 4-5 5b
Sec. 9.3 Optimum Demodulation of Digital Signals 599

Figure 9.17 Error-state diagram of a

ad /2 binary system with memory one

The solution of (9.107) is

agl/2
Paneer Vee ayer
aa (9.108) )
IEGita Gh) =
|Beee |(2.07Se
toe)Oy4
Finally, the desired upper bounds on P.(E) and P;(E) are obtained from T (ao, a_, a4, 1)
as [16]
ag/2
P.(E) < T (ao, a_-, a4, I|j-1 = 1G
= Gi aan
Pas

too A
b—exp {57 cosh (24z21)

ag/2 (9.109)
7
P,(E)< 1 Al)
b as =
of i=1 [1—-(@,+a_) /2}°

5 exp {—|

[0A] oan (0)

An upper bound on the exact bit error probability performance of binary PAM with no IST,
2

namely, P;,(E) = (1/2) erfe (/P,T,/No), is given by the numerator of the upper bound on
P,(E) in (9.109). Thus, we see that as long as the “eye” is open, that is, gj < g9 = 1, the
upper bound on the performance of the Viterbi algorithm for binary PAM with ISI of memory
] asymptotically approaches the no ISI bit-by-bit detection result as P,7),/No gets large.
Clearly, the larger the value of g;, the slower will be the approach to the asymptotic result.
Figure 9.18 is an illustration of this behavior where the upper bound on P,(£) as given in
(9.109) is plotted versus P,7)/No in dB with g as a parameter.
 600 Communication over Bandlimited Channels Chap. 9

107k

10-2

10°3

(E)
F
ON
BOUND
UPPER

104

Figure 9.18 Upper bound on the bit error

10-5 probability performance of the Viterbi
algorithm for binary PAM with ISI of
memory one

9.4 EQUALIZATION TECHNIQUES

In the broadest sense, the term “equalization” has been used to describe any technique em-
ployed by the receiver to deal with ISI. As such, one would include the MLSE scheme de-
scribed in the previous section in the class of equalization techniques. More often than not,
however, equalization refers to the introduction of a linear digital filter between the sam-
pler and the decision algorithm of Fig. 9.1 that compensates for the non-Nyquist behavior
of the overall system response. Such a linear equalizer tries to undo the ISI distortion intro-
duced by the channel. It is in this narrower context that we shall now discuss the process of
equalization.
In the simplest of terms, a linear equalizer tries to introduce gain at those frequencies
where the overall system frequency response has a loss so that the resulting cascade of the
two is either identically or approximately constant. By introducing gain at the receiver,
the total noise at the equalizer output is greater than if the equalizer were not present,
resulting in what is commonly called noise enhancement. Clearly then, the challenge is to
design the equalizer so as to produce as flat a total frequency response as possible for the
overall system while at the same time keeping the noise enhancement to a minimum. The
manner in which this is accomplished, that is, the criterion used for optimization, defines
the type of equalizer. Needless to say, the most desired criterion would be to minimize
 Sec. 9.4 Equalization Techniques 601

D = unit delay

eee eee

+ D 1B) ~ Dd ] 'D| - D /D. i

ece eee

(X) - eS és = (x) : eS -

Figure 9.19 Tapped delay line representation of zero-forcing linear equalizer

the probability of error. Since however, the ISI is not Gaussian distributed, designing an
equalizer to satisfy this criterion is a difficult task and is typically not attempted. Instead, one
considers optimization criteria that lead to analytically tractable yet practically meaningful
results. In discussing the various equalizer configurations, we shall not place any constraint
on their implementation complexity.
The simplest (but not necessarily the best in performance) class of equalizers are the
so-called linear equalizers, which have a z-transform of the form
K

O@S Seige (9.110)

k=—K

which is implemented as either an infinite impulse response (IIR) filter (K — oo) or a

finite impulse response (FIR) filter (K finite). In the latter case, the filter is referred to
as a transversal filter or tapped delay line'® and the coefficients {gn} are called the tap
weights (see Fig. 9.19). Determination of the coefficients (tap weights) is made based on the
optimization criterion selected. The input to the filter is the sequence of detected samples
{yx} in (9.66) and the output of the filter is the sequence {vx} representing estimates of
the complex input data symbols. Figure 9.20 is a discrete-time representation of the entire
system including the equalizer.
The two most common types are the zero-forcing linear equalizer (LE-ZF) and the
mean-squared error linear equalizer (LE-MSE). The former completely removes ISI by
identically creating a flat overall system frequency response without regard to the resulting
amount of noise enhancement produced. The latter attempts to mitigate the noise enhance-
ment by allowing a residual amount of ISI. We begin with a presentation of the LE-ZF and
its performance.

16. We note that, from the standpoint of a mathematical characterization, the tapped delay line representation of
a digital filter as in Fig. 9.19 is no different than the shift register representation such as that in Fig. 9.13.
602 Communication over Bandlimited Channels Chap. 9

Decision
Algorithm

Figure 9.20 A discrete-time block diagram of a linearly equalized communication system

9.4.1 Zero-Forcing Linear Equalizer

As mentioned above, the optimization criterion used by a zero-forcing equalizer is to com-
pletely remove the ISI without regard to the noise enhancement that results. Thus, if G(z) of
(9.68) represents the overall system response,!” then to obtain an ISI-free signal sequence
Q(z) must be chosen to produce a flat channel response for all frequencies, that is,

1
Q(z) = ——
G@ (pny
fa )

The output of such an equalizer then has the form

De = Ce + Vy (9.112)
where {5x is a zero-mean complex Gaussian noise sequence with discrete power spectral
density
oo 2
S3(2)= > R5(m)z-" = EES PN ERIHK ahs {5e5f.m} O.113)
m=—00 G()
Note that although we have completely removed the ISI, the noise sequence is not white.
Thus, symbol-by-symbol detection is not the optimum decision algorithm; nevertheless, it
is still the most common algorithm used in receivers employing an LE-ZF because of its
simplicity.
The mean-squared error at the equalizer output is simply the noise power in 3, or
equivalently the contour integral of (9.113) around the unit circle defined by |z| = 1. If G(z)
is Zero Over some interval, then the equalizer Q(z) becomes infinite over that same interval
and hence the mean-squared error at the equalizer output also becomes infinite. Thus, a
condition on the existence of the LE-ZF is that 1 /G(z) be integrable. From (9.110), (9.111),
and (9.113), the above mean-squared error is given by
eee Po!
ei ncar LD {ax — | = E |
9.114
= Ry tO) ogee
q
v
—
90
— Co —_—
In n/t,
———ea
G (ef°%) @ fe

where qo is the zero" equalizer tap.

17. We shall assume here the most general case where the memory v can be infinite.
Sec. 9.4 Equalization Techniques 603

Example 9.2
By way of example, we shall consider a received pulse shape, A(t), that is causal and of the
following form:

h(t) = V2ae sin Bt u(t) (9.115)

From (9.60) the ISI coefficients can be obtained as

ee “os Ik| BT; + Boos |k|BTs] ine.

oo) at 2 e (9.116)

Letting BT, = 27, then (9.116) simplifies to

2
ok = we Ae ut balbie: (9.117)
2 (a? + B?)
with transfer function

2 ee
G(z)= oan lente 5 naehs (9.118)
(ee) ili ye) 42)
The corresponding equalizer is described by

2 (a? + B’) (1— nz si )a — z= Pn)

Q(z) = rea)
2 (a? + 6?) (-nz eT +1l+n°-
2 nz) (9.119)

that is, a finite impulse response (FIR) filter. The corresponding mean-squared error is from
(9.114) given by

2
SLE-ZF -~ 4a? (a? + B) (1 + e787)
B2 (1 — e287) (9.120)

which approaches infinity in the limit as w approaches zero.

9.4.2 Mean-Squared Error Linear Equalizer

The mean-squared error (MSE) linear equalizer allows for a residual amount of ISI at
its output as a trade against minimizing the mean-squared error between the equalizer
output and the transmitted symbol. Thus, Q(z) of (9.110) is chosen so as to minimize
ae lo.@)

E {lx - el? het Cm = 5: ¢z* denote the z-transform of the transmitted data sym-
k=—0o
5 (ee)
bol sequence and M(z)= >> jinz—* the z-transform of the noise sequence at the equal-
k=—0o
izer input. Since the data symbols are independent of one another, that is, the sequence {Cx}
is a zero mean “white” sequence with variance E {{cxl| A a then since

Da ea,
Ren) = E {eect = {@-' iC25 (9.121)

the power spectral density of {c,} is flat and is given by

CO

eS Rie "= 0; (9,122)

m=—OOoO
604 Communication over Bandlimited Channels Chap. 9

Furthermore, from the correlation properties of the noise sequence {/1,} given in (9.65), the
power spectral density of this sequence becomes
CO [e.@) 2 ;

Sill = So Riel = 207 Yo gael = ois ana

m=—CoO m=—CoO

lo.@)

Finally, writing the z-transform of the equalizer output sequence V(z)= > vUmz”™
m>=—OoO

as V(z) = C(z)G(z)Q(z), then the power spectral density of the error sequence {&,} =
{Vz — Cx} is given by
2 ; Te
85 (e/0™) = 5; (eT) |G(el™ Q(el#™) — 1] + 55 (e/°™) |Oe™)
f (9.124)
; 2
= 02 |Gel™) (ei) — 1] + 207G (e/™) |O(e!*™)
and hence the mean-squared error is

es Tk m/Ts , 2,
LE MSE 9NU /J_—zx/T, jo?G(el#™) Q(el*) — 1]
(9.125)
2
+20°G (ef®) |Q(eJ@1) |e

Completing the square in (9.125) with respect to G(e/®?s) and, for simplicity of notation,
omitting the argument e//* in the various transfer functions, we get

iy
éfe_wse = 5rn] (2ean IG? nam
+2076)
2
o-G

wt —1/Ts ' ab o2|G|? + 202G

(9.126)
—_——.———
(o2G)"
+0* | daw
2
o2|G|* + 202G 4]

Since G(z) is real and even (since 9, = g_,), then G(e/®@") = G*(e/@"s), that is, IG/? = G?
and (9.126) simplifies to

sa
eas ae 5AY Nae tp
BadLE-MSE
2
bl ici= eae
57 (| G(o2G+2
2
(0: a”) ) |2 wih bese
2 c
02G +202
(9.127)
207o?2 r
o2G + 20? ri
Since both terms in the integrand of (9.127) are positive, the mean-squared error is mini-
mized by choosing Q so as to force the first term equal to zero at every value of w. This is
accomplished by selecting

af 1
(9.128)
2@)= o2G(z) + 202 <a G(z) + 202/02
Sec. 9.4 Equalization Techniques 605

which results in the minimum mean-squared error

(ase ) Ts ceTs 2676"

rs
LE—MSE aan ——on :
iE. o2G (e/27s) ae Io2
a)

(9.129)
2 Ts Bits 1
= 20° — : dw
2n J_x/1, G (e/27s) ae 202/02

Comparing (9.129) with (9.114), we clearly see that

(ste-mse),
2
,,<te—ze (9.130)
as would be expected. At high signal-to-noise ratio as measured by a2 /o”, the LE-ZF and
the LE-MSE give approximately the same performance. Also, for the LE-MSE the noise
enhancement is finite, and hence this equalizer always exists.
For Example 9.2, the LE-MSE would be

(a* + B*) (nz7! +14? + nz)

(92131)
ee FO ea ay 2 (a2 102) (a2 wR ee)
9.4.3 Decision-Feedback Equalization

The performance of the linear equalizers considered in the previous two sections can be im-
proved upon by introducing nonlinearity into the design. One such structure is the decision-
feedback equalizer (DFE) [17, 18] which uses decisions on the previous data symbols in an
attempt to cancel out a portion of the ISI affecting the decision on the current symbol. The
primary advantage of the DFE is that it achieves its purpose with substantially less noise
enhancement than its linear counterpart. The main disadvantage of the DFE is that it suffers
from error propagation caused by feeding back incorrect decisions. Fortunately, however,
the error propagation effect is usually far less significant than the benefit obtained from the
reduction in noise enchancement.
The structure of a DFE is composed of two parts: a forward filter, Q(z), which
typically includes one of the two linear equalizers previously discussed, and a feedback
filter, P(z), which attempts to further reduce the ISI (see Fig. 9.21). Each of these filters
has a distinct purpose in the way it handles the ISI. The forward filter acts on the precursor
ISI (the samples of the overall system response occurring prior to the present time) in a
manner similar to that of the linear equalizer discussed previously. It makes no attempt to
equalize the postcursor ISI (the samples of the overall system response occurring after the
present time). Since the feedback filter operates on past decisions, it must be causal. Thus,
it has a discrete-time representation with only positive delay terms
K
P=) pez* (9.132)
k=1

and as such aids in cancelling the postcursor ISI. Since the forward filter only attempts to
equalize the precursor ISI, it can accomplish this task with less noise enhancement that if it
tried to equalize the entire system response. On the other hand, assuming that the decision
algorithm produces virtually error-free decisions, then the feedback filter produces no noise
at its output and thus the overall effect is a reduction in total noise enhancement.
 wia}sks UOTBOTUNUWIWIOS
pazienbs yorqpady uoIsIoap SuIdoJ-019z & JO WIRIZIP YOo[G SWIN-a1019SIP YW 176 ANSI

eyeq g0In0S
Joqutés
wyzOZ[y
uoIsbaq

606
Sec. 9.4 Equalization Techniques 607

DFE’s are classified in much the same way as LE’s, that is, in accordance with the
optimization criterion applied to the forward filter. In this regard, we shall discuss the
so-called zero-forcing decision feedback equalizer (DFE-ZF) and the mean-squared error
decision feedback equalizer (DFE-MSE).

Zero-forcing decision feedback equalizer (DFE-ZF). The idea behind the

DFE-ZF is for the forward filter to accomplish what the LE-ZF did but only for the precursor
ISI. Conceptually, the way it does this is to equalize the entire overall response (precursor
and postcursor ISI) and then reinsert the postcursor ISI. Thus, the forward filter may be
viewed as a cascade of two filters, that is, Q(z) = Q1(z)Q2(z), the first being identical to
the LE-ZF, namely, Q;(z) = 1/G(z) [see (9.111)] and the second defined in terms of the
feedback filter by Q2(z) = 1 + P(z). The idea here is that at the output of Q)(z) [concep-
tually just an intermediate point within Q(z)] there is no ISI (the signal component of these
samples is precisely cx) while at the output of Q(z), there is once again ISI but just due to
the postcursor. If now the decision algorithm were to produce totally noise-free decisions,
that is, C, = cx, then the output of the feedback filter P(z) would exactly cancel the postcur-
sor ISI introduced by Q2(z) leaving once again a totally ISI-free set of samples on which to
make a decision.
To see how the reduction in noise enhancement comes about, we can write the input
to the decision algorithm in Fig. 9.21 as

H+ (9.133)
where CE has the z-transform

C(z) =E@GWO®) —C@P@ =C@ A+ P@)—C@P@ 9.134)

with C (z) denoting the z-transform of the symbol decision sequence {cx}. If we assume

error-free decisions, then (9.134) simplifies to C(z) = C(z) and likewise (9.133) becomes

Dp = Ce + & (9.135)
that is, the input to the decision algorithm is once again ISI-free as predicted above. The
noise sequence a rhowever, now has the discrete power spectral density [see (9.123) and
Fig. 9.21]

Eyer
Az)
° G(z)
ieet z ) (1+ PE) (9.136)
:

where we recall that G(z) = G(z7!) since gx is real and even. Comparing (9.135) and
(9.136) with (9.112) and (9.113), respectively, we see that the DFE-ZF has an identical sig-
nal component to the LE-ZF, that is, it is independent of the feedback filter P(z); however,
the noise component of the DFE-ZF is clearly affected by this filter. As such, we should be
able to choose the causal filter P(z) to reduce the noise at the input to the decision algorithm
relative to that of the LE-ZF wherein P(z) = 0.
The mean-squared error of the DFE-ZF at the decision algorithm input, namely,
2
edenop= E {li 2 el], is from (9.135) simply given by erp7p= E {| |To de-

termine the optimum P(z), that is, the one that minimizes this mean-squared error, we
608 Communication over Bandlimited Channels Chap. 9

first reinterpret (9.136) as the power spectrum of the output t of a filter Q2(z) =1+ P(2)
whose input is 3, with power spectrum as in (9.113). Based on the orthogonality principle
[19, Chapter 13], it can be shown that the particular filter P(z) that minimizes fore has
the property that the output of Q(z) is uncorrelated with the past of its input, that is,

EGO Oa taalie (9.137)

Furthermore, since from the above interpretation

K
Sian Oc a a ey (9.138)
a=

then using (9.137), we see that

Re(m) SE {Eben} = 0; Papaay eee (9.139)

and since R; (m) = Rz(—m), it follows that (9.139) is true for all m 4 0, that is,

AP
Rz(m) SE |50m (9.140)

where again 5(m) is the Dirac delta function. Thus, we conclude that the equalizer output
roa,
noise sequence that minimizes Dae. =E a is a “white” sequence, that is, its
00 te
discrete power spectrum is a constant given by Sz Ges R; (m)z "SE |
Ee |
m>=—CO
In our discussion of noise whitening filters in Section 9.3.1, we considered partition-
ing the overall system response G(z) into a minimum phase, physically realizable part F(z)
and a nonminimum phase, nonphysically realizable part F(z~'), the former having all its
zeros and poles inside the unit circle and the latter having all its zeros and poles outside the
unit circle. To find the optimum forward filter, we again consider this same partitioning of
G(z) (actually a normalized version of it, as we shall describe shortly) and then associate
the filter (1 + P(z)) with the normalized minimum phase, physically realizable part and the
filter (1 + P(z*)) with the normalized nonminimum phase, nonphysically realizable part.
In particular, let

T] dG = Az74) [[ GQ — Az)
Z l
GQ =FOFR 2K SS (9.141)
[Jd — Xmz7") he — XmZ)
m m

where the A;’s and the Xm’s respectively represent the zeros and poles of G(z). Since from
the definition of P(z) in (9.132), (1 + P(z))|,~o9 = 1, then the correct association referred
to above would be

T]d= Aiz7!)
1
(9.142)
Mieka iho es
Sec. 9.4 Equalization Techniques 609

and thus from (9.136), the sequence {x has the power spectrum

2 2
Sp(2) = = (9.143)
which is white in accordance with the above discussion. Finally then, the minimum mean-
squared error is

fern JM 202 i
DFE-ZF) i, OK (9.144)

Note that (9.142) also determines the feedback filter by simply subtracting 1 from the right-
hand side of this equation.
Example 9.3
Considering Example 9.2, we immediately see from (9.118) that

K= B* (1 -n’)
7a? 3 Iba (gh ee) per ; ai= entat (9.145)
145

corresponding to the case of no zeros and one pole. Thus, the forward and feedback filters of
the DFE-ZF are given by

2 (a? + p?) nz}

Q(z) = —.——— (1 - ; P(z)=——— i

and the minimum mean-squared error is

2D of Hah ea)CE Des (9.147)

(bre-ze) Ag B2 (1 — ea)

When compared with (9.120) for the LE-ZF we see that we have achieved a reduction in mean-
squared
q error by the factor 1 + e Ts The fact that (e205
DFE-ZF) <?,,_>,
LE-ZF can be shown [3]
to be true in a more general context than just this example.

Mean-squared error decision feedback equalizer (DFE-MSE). The dis-

crete-time block diagram of a communication system employing a mean-squared error
decision-feedback equalizer (DFE-MSE) [20] can be represented as in Fig. 9.21 except that
now Qj(z) is not constrained to be equal to 1/G(z) as in the DFE-ZF case but rather is
to be determined along with the feedback filter P(z) to minimize the mean-squared error
at the decision algorithm input, namely, E {ax ~ I. By inspection of Fig. 9.21, the z-
transform of the signal component of vz , namely, cx, can be written as

Cz) = CGO 1+ P@) —C@P@ (9.148)

which, assuming again that there are no errors in the decision process, that is, cx = Cx,
becomes
EG,
C@O= {ew (ere ~ —______—_
ae |}1+P
+ P(z)) (9.149 )

The z-transform of the signal component, &s, A cx — Cx, of the total error signal &, is then

E,(2) = C(z) — C@ =C@[E@O1(@ — 10. + PQ) (9.150)

610 Communication over Bandlimited Channels Chap. 9

Finally, since the signal and noise components of &% are uncorrelated, then analogous to
(9.124), the power spectral density of the total error signal is
; 2 , ae Pu ie
ce {22G(elM)Q,(e/°%) — 1] + 207G(e!°™) |O1(e/2%)
| (9.151)
ENG
x 1ae (el )

and hence the mean-squared error to be minimized is

ee Tf m/Ts ihe ¥ a)

eDFE—MSE = = | Qi(e/@") = i
{02G(el@*)
nes (9.152)
i ji+ P (ei) dw
é ; 2)
+207G (el?) | 1 (e/2%)
Because of the way we have partitioned the integral in (9.152) into a part (between
braces) dependent only on Q, (e/®7) and a part dependent only on P (e/®"'), we can
perform the minimization of ene a in two steps, that is, we first choose Q; (e/ots) to
minimize at each frequency the power spectrum between braces and then choose P (e/oTs)
to minimize the total power spectrum corresponding to the error signal itself. Comparing
the power spectrum between braces in (9.152) with the integrand in (9.125), we see that
we are dealing with identical minimization problems. Thus, without further ado, we can
immediately write the desired solution for the first section of the forward filter as [see
(9.128)]
|
Q\ (z) = (9.153)
G(z) + 202/ a2

which results in the term between braces in (9.152) being equal to the integrand in (9.129),
or equivalently
oS
2 xete iei‘ 7 20 2 L 1p (it) 2
DFE—MSE ™ 97 ever. G(eleTs) aE 207/02 dw (9.154)

The second step involves a spectral factorization similar to that performed for the DFE-ZF,
namely, we factor G’(z) 4 G(z) + 257 Ja? into

[1G -Ajz7!) []dG—ajz)

Oa FOF =) kh 9.155
Td -xi275 [TG = xiz
m m

and then make the association

TT (i —Aj271)
l
1+ P(z)=————— 9.156
Td —x,2-5
m
Ce
from which we can determine the feedback filter P(z). Finally, the minimum mean-squared
error for the DFE-MSE is given as

(cire—wse),.,, = gr (9.157)
2 202
 Sec. 9.5 Further Discussion 611

Example 9.4
Considering again Example 9.2, we have that

eae) gee
2 (a? + 62) (1—nz)(1—nz) o2
(9.158)
ei) (uz)
(1 — z-!) (1 = 072)
where

aE
6? (1=n?

1+ p2
UtA)_ KAM
a?

1+ pe?
o

_ P=) (a2) 1 (Ler?

O 8n(a2+ B) Lo?) 721 eee
2
_ em (2) +3 ie) eee
87 (a? + B?) \o? Dy n

It can be shown [3] that in general Saal PR (eer) _ (that is, K’> K) and
in min

furthermore (e
DFE-MSE} )i, <— (€?
\LE-MSE) )Jin ‘

9.5 FURTHER DISCUSSION

In our discussion of equalization, we have made several idealistic assumptions that, al-
though theoretically satisfying, may not be met in practice. First, we implicitly assumed that
the channel impulse response is completely known to the receiver and furthermore is fixed,
that is, it does not vary with time. This is tantamount to assuming that the ISI is completely
known and time-invariant and as such the filters that made up the various equalizer config-
urations have fixed coefficients determined in accordance with the particular optimization
criterion that defined the equalizer type. Second, we placed no constraint on the complexity
(number of coefficients) in the equalizer; that is, we assumed, in general, an implementation
in the form of an IIR rather than an FIR (transversal filter).
The first of the above two idealizations is handled by making the equalizer dynamic.
In particular, the channel impulse response is continuously (in discrete time increments)
estimated and the equalizer coefficients are then adjusted to minimize ISI and noise in
accordance with these estimates. This combined process is referred to as adaptive equal-
ization and was first introduced by Lucky [21] in connection with the telephone channel
application. A simple block diagram of an adaptive linear equalized receiver is illustrated
in Fig. 9.22. The error signal used to drive the filter coefficient adjustment algorithm is
obtained from the difference of the data symbol decisions and the equalizer output. For ex-
ample, in the absence of noise and ISI, the two would be identical and the error signal would
be zero, resulting in no change in the equalizer coefficients. When there is noise but no ISI,
then the error signal is zero whenever a correct decision is made and nonzero whenever an
error (due to noise) occurs. However, since the equivalent additive noise at the decision al-
gorithm input is zero mean, the effect of decision errors on the adaptation of the equalizer
612 Communication over Bandlimited Channels Chap. 9

sample at t = kT, + t,

received Transversal
signal | Receive Filter Decision
ne Equalizer Algorithm
G,(f) (2)

Coefficient
Adjustment
Algorithm

Training
Sequence

Figure 9.22 A simple block diagram of an adaptive linearly equalized receiver

will be averaged (smoothed) out by the large time constants typically built into the adjust-
ment algorithm resulting in no net change in the equalizer coefficients. Finally, when both
noise and ISI are present, the adaptive equalizer will function on the average to bring the
input and output of the symbol decision algorithm into agreement.
The error signal produced from the difference of the data symbol decisions and the
equalizer output will be successful in adapting the equalizer provided that the data symbol
estimates are reasonably accurate, which in turn requires that the ISI not be excessively
large. While this condition is typical of the steady-state (tracking) behavior of an adaptive
equalizer, it is not typical during initial acquisition when the equalizer is learning the state of
the channel. Thus, at the outset of transmission (and perhaps periodically for reacquisition
purposes), it is common to send a training data sequence (assumed to be known at the
receiver) over the channel during which time the adaptation error signal is formed as the
difference of the known training symbol sequence and the equalizer output (see Fig. 9.22).
Upon conclusion of the training interval, actual data transmission begins, whereupon the
training sequence is replaced by the decision algorithm output for the formation of the error
signal.
Since the complexity of the equalizer and thus the speed with which it adapts is
directly related to the number of coefficients in its z-transform representation [for example,
the parameter K in (9.110)], it is only meaningful to talk about adaptive equalizers in
the context of an FIR. Hence, the second idealization mentioned above, namely, the lack
of a constraint on complexity, is automatically handled as a consequence of the need for
adaptation. However, one must re-solve the problem of finding the filter coefficients under
the various optimization criteria discussed in Section 9.4 subject now to the constraint
 Problems 613

of a finite number of degrees of freedom. The solutions to these constrained complexity

equalizer problems require methods quite similar to those used in designing other adaptive
systems such as adaptive antenna arrays [22]. For example, designing an LE subject to an
MSE criterion requires a so-called mean-squared error gradient (MSEG) algorithm [3].
This and other methods of its type lead directly to methods for designing the adaptation
algorithm. For example, by substituting a time average for the statistical (ensemble) average
in the MSEG leads to an adaptation algorithm called the stochastic gradient (SG) algorithm.
Since, as mentioned in the introduction of this chapter, these subjects are not the main
focus of this textbook, we shall not continue their discussion beyond this point. A detailed
presentation of the subject of constrained-complexity equalizers and their corresponding
adaptation algorithms may be found in such excellent references as [2] , [3], and [23].

PROBLEMS
9.1 Following an approach analogous to (9.14)-(9.21), show that the Nyquist condition on G(f)
can alternately be expressed as

— Eee Cn 4) (2n +1)

Dita) Set srs
for any integer n [including n = 0 which corresponds to (9.22)].
9.2 For a Nyquist (zero ISI) overall system frequency response, show that the choice of transmit
and receive filters as per (9.28) maximizes the detected sample SNR of (9.27), or equivalently
minimizes the average bit error probability of (9.8).
9.3 Let G(f) characterize the overall channel response of a bandlimited system sending data at a
rate 1/7,. Then, if g(t) = F—! {G (f)} is to have nonzero sample values only at t = 0, T,, and
both of these values are required to be equal to unity, show that G(f) must satisfy

x (+z)
k=—oo
== AR soosm fT: IflS srl
tik ges

9.4 Consider a frequency response G(f) characterized by the convolution of an arbitrary (not
necessarily bandlimited but bounded impulse response) filter Gg(f) and a strictly bandlimited
filter
Gr f) = |2T costs
Tee Ih; IFSa7,
0; otherwise

that is, G(f) = Ga(f) * Go(f). Show that using G(f) for the overall channel response of a
communication system sending data at a rate 1/7; results in zero ISI from all but adjacent
symbols.
9.5 For the duobinary system illustrated in Fig. 9.10, show that choosing a detection threshold
halfway between “0” and “2A” minimizes the average bit error probability in the limit of large
SNR.
9.6 Show that, for the raised cosine class of Nyquist channels as described by the overall system
frequency response of (9.25), the received power P, and amplitude A are related by (9.40), and
the variance of the detected sample, yx, is given by (9.41).
9.7 A modified duobinary system is illustrated in Fig. P9.7a. As in the conventional duobinary case,
this system has an equivalent, as illustrated in Fig. P9.7b.
(a) Letting P(w) = P\(w) P2(w), and H(w) = H)(w) H2(o), find H(@) in terms of P(w) such
that the above equivalence is established, that is, in the absence of noise, y(t) is the same
in both.
614 Communication over Bandlimited Channels Chap. 9

AY, 6 kT) n(t)

k=-00

y(t)
Precoder
to threshold
comparison
and: decision
logic

Figure P9.7a

AY) 5(t- kT) n(t)

k=-0o

to threshold
comparison
and decision
logic

Figure P9.7b

(b) If P(@) is the ideal brick wall Nyquist filter, that is, P(@) =7T; |w| <2/T and P(w) =0;
otherwise, find the overall channel response h(t) = F-'{H()}. (Hint: make use of the
fact that p(t) = F-!{P(o)} is a sin x/x function and then use the relation between H (w)
and P(w) found in part (a)]. Sketch H(w) and h(t).
(c) What are the values of h(t) at the sampling points t = nT, that is, how many have nonzero
value and what are these values?
(d) Based on the result of part (c), in the absence of precoding, which data bit(s) contribute ISI
to the detection of ag?
(e) As in duobinary, error propagation will also occur for modified duobinary unless precoding
is used. Find the appropriate precoder and sketch its configuration.
(f) For the precoder of part (e), what is the appropriate decision rule for determining {a,,} from
the samples of y(t)?
(g) Assuming H(q) is split equally between Hj(w) and H2(w), find the error probability for
the modified duobinary system using the results found for parts (b)-(f). Assume n(t) is
white Gaussian noise with single-sided spectral density Ng Watts/Hertz.
9.8 A system is designed for a quadrature partial response (QPR) modulation, in particular,
quadrature modified duobinary, that is, modified duobinary (Problem 9.7) on quadrature car-
riers transmitted over a common channel. The transmitter for such a modulation scheme is
illustrated in Fig. P9.8a. Binary (+1) data at a rate 1/7) is split into its even and odd compo-
nents, a, and b,,, respectively, each occurring at a rate 1/T = 1/27). These two data bit streams
are precoded and formed into baseband streams with pulse shaping equal to that produced by a
modified duobinary transmit filter, that is

-( oT ua
Hiloy= JOT sate Ne): lo| <4iP
otherwise

The two baseband data bit streams are amplitude modulated onto quadrature carriers and the
sum of the two sent over the channel as the transmitted signal
Problems 615

AY 5(t-nT)
n=—0o

Binary
Source A Ss5(t—nT)
Rate 1/T, n=-0o

Figure P9.8a

sample at t= nT

Figure P9.8b

CO [o,@)

S(t)=A > Cnhi(t — | COS Wet — A Da d,h\(t — | SiN Wet

n=—o0o n=—0o

Assume an additive white Gaussian noise channel with single-sided power spectral density No
Watts/Hz. The received signal plus noise is first demodulated with quadrature carriers

re(t)=2cos@ct; rs(t) = —2 sin wet

and then passed through the modified duobinary receive filters H2(w) = H\(@). The outputs
of these filters are sampled at times tf = nT (n integer) and passed through full-wave recti-
fied threshold comparison devices typical of modified duobinary baseband modulation. The
detected bits are denoted by a, and Dns respectively, as shown in Fig. P9.8b.
(a) Assuming w,T >> 1 (that is, ignore all second harmonic carrier frequency terms), find the
bit error probabilities for the in-phase and quadrature data sequences, that is, P(E)|q =
616 Communication over Bandlimited Channels Chap. 9

n(t) sample at

¥,4,8(0-n2)
Figure P9.9a

Figure P9.9b

Pr {Gn # an} and P(E)|, = Pr {brSOx |.Express your answer in terms of the bit energy-
to-noise ratio E,/No of the system where Ep, = P;,,7; is the bit energy with P;, the input
power to the channel, that is, the average power of s(t).
(b) If a pair of bits a, and b, is thought of as having sent the complex symbol x, =a, +
jbn, evaluate the symbol error probability of the system, that is, P(E)|x = Pr {X, # xn}.
Express your answer in terms of the bit energy-to-noise ratio E,/ No of the system.
9.9 A pulse g;(t) = ie {G,(f)} is designed for transmission over a given channel G,(f) per-
turbed by additive white Gaussian noise so that the overall system response go(t) = F—!
{Go(f)} = F7! {Gi(f)Ge(f)G,(f)} has no ISI. In particular, let Go(f) satisfy the Nyquist
condition of Eq. (9.22). Furthermore, assume that G,(@) = (G;(w)G_-(@))* so that the er-
ror probability is minimized. Suppose now that the channel has a somewhat different transfer
function, say Gf ), than that expected in the design such that the overall system response
g(t) =F Gif} = FUNG CAG AGG} now contains ISI at the sample time instants.
It is desired to pass y(t) through a zero-forcing linear equalizer (LE-ZF) with transfer function
Q(f) to restore the zero ISI condition at a finite number of sample times in the equalizer filter’s
output pulse response h(t) = F “1G f)Q(f)}. The LE-ZF has the form of a transversal filter
with three taps as illustrated in Fig. P9.9b. Samples of the pulse shapes are illustrated in Fig.
P9.9c.
(a) Find a matrix relation that when solved yields the tap coefficients g_}, go, gq; in terms of
the samples ho = h(O), hj = h(T;), hz = h(2T,) and an appropriate set of samples of g(t)
taken at integer multiples of T;, namely, g; = g(i7;). How many and which samples of g(t)
and h(t) are needed?
(b) If the equalized pulse is to have the same nominal delay as the equalizer, that is, h(t) =
go(t — T;) at the three sample points t = 0, 7;,27;, solve the matrix relation found in
Problems 617

8 (t)

Figure P9.9c

part (a) for g_1,

go, qi when g-; £0, go=1, gi; AOandg;=0; i~-1,0,1. Express
your answers entirely in terms of g_; and g}.
(c) Numerically evaluate the tap coefficients found in part (b) when g_; = .2, g9=1, g1=
—.1.
9.10 The design of an LE-MSE equalizer can be carried out in the discrete time domain as opposed
to the frequency domain (z-transform) approach described in Section 9.4.2. In particular, con-
sider a linear equalizer described by (9.110) whose input is the sequence {y,} and output is the
K
sequence {v,} where vk = > Gnyk—n. For simplicity of notation, we shall assume that all se-
n=—K

quences are real. Define the MSE for the k" equalizer output sample by € = E {(vx — ax)*}
where a, is the desired k'" data symbol. Note that if a, is binary (+1), then oa? = |. Since € is
a concave function of the tap weights {q,}, a set of sufficient conditions for minimizing € is

aE OR je) Sell Peis ht a mil

Om

(a) Show that the above set of conditions leads to the following set of relations:

Ryy(m) = Rya(m), és = 10) el REZ Re syed

618 Communication over Bandlimited Channels Chap. 9

where Ryy(m) A E {yx—mvx} is the cross-correlation of the input and output equalizer se-
quences and Ryq(m) A E {yx—mag} is the cross-correlation of the equalizer input sequence
with the data symbol sequence.
(b) Using the convolution relation between the input and output equalizer sequences in terms
of the equalizer tap weights, show that the set of conditions of part (a) can be expressed in
the following matrix form:

Ryq= Ryu

where

q-K Rya(—K)
q-K+1 Rya(—K + 1)
q= : » Ry= : ,
qK-\ Rya(K = 11)
qK Rya(K)
Ry(0) Ry() vee Ry(2K)
R,(-1) Ry (0) Ry(2K — 1)
R, = : : ; :

Ry(=-2K)) RyG2k + 1) Ry (0)

with Ry(m) & E {yx—myx} equal to the autocorrelation function of the equalizer input se-
quence.
(c) Assume now that the equalizer input sequence is as specified in (9.66), namely y, =
v

Y> giax—1 + Ux, where 2v is the number of ISI components), {a,x} is an uncorrelated data
l=—v
sequence with variance GF and {jx} is a zero-mean Gaussian correlated noise sequence
with correlation function R,,(m) = E {ux—mULk} = O'S n2 Also, assume that go = 1. Show
that the elements of the matrix Ry and the column vector Ry, are given by

Rya(m) = ae

Ry(m) = 07 D> 818i4m +078m

l=-v

from which the tap weight vector q can be uniquely found using the result in part (b),
that is

q— Rak.

9.11 Consider the design of an LE-MSE in accordance with the development of Problem 9.10.
Letting Ko=p'= 1, o = 1, o? =.2 and an ISI channel described bype= 220) ee
—.1, find the optimum tap weights q_1, go, 41-
9.12 Consider a zero-forcing linear equalizer (LE-ZF) for use in a system with an overall system
response of the form

G(z)=g-12 '+g04+ g12

that is, ISI is contributed only by the pulses adjacent to the one of interest. Letting g_; = g1,
we define

p |>
81
80
as the normalized ISI contributed by each of the two adjacent pulses.
 References 619

(a) Show that the necessary Fourier transform condition G (e/®") > 0 for all w implies that
lo| < 1/2.
(b) Show that G(z) can be factored in the form
¥ mare
A eS ae a)
a)

where p A — (a + aa)?
(c) Assuming without loss in generality that |a| < 1, find a partial fraction expansion for the
equalizer filter Q(z).
(d) Expanding each of the partial fraction terms in the answer of part (c) in a power series in z
or z_' l as appropriate, show that the zero" equalizer tap go can be expressed as
Je a
pgo (1 — a?)
(e) Evaluate the mean-squared error of the LE-ZF using your answer found in part (d). Express
your answer only in terms of go, a7, and p.
(f) What is the value of the mean-squared error as p > +1/2 or equivalently a > +1? What
happens to the gain of Q(z) at these limiting values and what does this result in in terms of
noise enhancement?
9.13 Consider now a mean-squared error linear equalizer (LE-MSE) for use in a system with the
same overall system response as in Problem 9.12 and the same ISI condition, namely g_; = g}.
(a) Show that the equalizer Q(z) is now given by the reciprocal of the equivalent overall
system response

G'(z) = pgo
(c-a’)(2-a'-!)
where p’ A p (1+ 22") A (a’ ally .
(b) Expressing Q(z) asa partial fraction expansion and expanding each of the partial fraction
terms in power series in z or z~! as appropriate, show that the zero" equalizer tap go can
be expressed as

qi = — SO
a /
pFO
/

© pgo(1— a) — pgo/1 — 4p?

(c) Evaluate the mean-squared error of the LE-MSE using your answer found in part (b).
Express your answer only in terms of go, oe ae, and p’.
(d) What is the value of the mean-squared error now as p > +1/2? Is it bounded? Explain.

REFERENCES
1. Lucky, R. W., J. Salz, and E. J. Weldon, Jr., Principles of Data Communication, New York:
McGraw-Hill, 1968.
2. Proakis, J., Digital Communications, 2nd ed., New York: McGraw-Hill, 1989.
3. Messerschmitt, D. G., and E. A. Lee, Digital Communication, \st ed., Boston: Kluwer Academic
Publishers, 1988,
4, Nyquist, H., “Certain Topics in Telegraph Transmission Theory,” A/EE Transactions, vol. 47,
1928, pp. 617-44.
5. Kretzmer, E. R., “Generalization of a Technique for Binary Data Communication,” JEEE Trans-
actions on Communication Technology, vol. COM-14, no. 1, February 1966, pp. 67-68.
620 Communication over Bandlimited Channels Chap. 9

6. Lender, A., “The Duobinary Technique for High-Speed Data Transmission,” JEEE Transactions
on Communication Electronics, vol. 82, May 1963, pp. 214-18.
. Viterbi, A. J., “Error Bounds for Convolutional Codes and an Asymptotically Optimum Decoding
Algorithm,” JEEE Transactions on Information Theory, vol. IT-13, April 1967, pp. 260-69.
. Papoulis, A., The Fourier Integral and its Applications, New York: McGraw-Hill, 1962, Chap-
ter 4.
. Duffy, F. P. and T. W. Tratcher, “Analog Transmission Performance on the Switch Telephone
Communications Network,’ Bell System Technical Journal, vol. 50, April 1971, pp. 1311-47.
. Pasupathy, S., “Correlative Encoding: A Bandwidth Efficient Signaling Scheme,’ JEEE Commu-
nications Magazine, July 1976, pp. 4-11.
. Kabal, P. and S. Pasupathy, “Partial-Response Signaling,’ JEEE Transactions on Communica-
tions, vol. COM-23, September 1975, pp. 921-34.
. Bennett, W. R., Introduction to Signal Transmission, New York: McGraw-Hill, 1970, p. 16.
. Forney, G. D., Jr., “Maximum-Likelihood Sequence Estimation of Digital Sequences in the Pres-
ence of Intersymbol Interference,’ IEEE Transactions on Information Theory, vol. IT-18, May
1972, pp. 363-78.
. Forney, G. D., Jr., “Coding System Design for Advanced Solar Missions,” Final Report on Con-
tract NAS 2-3637 for NASA Ames Research Center, Watertown, MA: Codex Corp., December
1967.
. Forney, G. D., Jr., “The Viterbi Algorithm,” Proceedings of the IEEE, vol. 61, no. 3, March 1973,
pp. 268-78.
. Viterbi, A. J., and J. K. Omura, Principles of Digital Communication and Coding, New York:
McGraw-Hill, 1979.
. Austin, M. E., Decision Feedback Equalization for Digital Communication over Dispersive
Channels, Lexington, MA: MIT Lincoln Laboratory, August 1967.
. Belfiore, C. A. and J. H. Park, “Decision Feedback Equalization,’ Proceedings of the IEEE, vol.
67, no. 8, August 1979, pp. 1143-56.
. Papoulis, A:., Probability, Random Variables, and Stochastic Processes, 2nd ed., New York:
McGraw-Hill, 1984.
. Salz, J., “Optimum Mean-Square Decision Feedback Equalization,” Bell System Technical Jour-
nal, vol. 52, October 1973, pp. 1341-73.
. Lucky, R. W., “Automatic Equalization for Digital Communications,” Bell System Technical
Journal, no. 44, April 1965, pp. 547-88.
. Widrow, B. and S. D. Stearns, Adaptive Signal Processing, Englewood Cliffs, NJ: Prentice-Hall,
1985.
- Qureshi, S. U. H., “Adaptive Equalization,” Proceedings of the IEEE, vol. 73, no. 9, September
1985, pp. 1349-87.
 Chapter 10
Demodulation
and Detection of Other
Digital Modulations

Thus far in the book, we have focused our attention on the characterization of traditional
digital modulation schemes and their performance over the AWGN channel. By traditional
modulations we mean those that are treated in most textbooks on the subject of digital com-
munications, although not perhaps in the depth to which they are covered here. While in
many of today’s applications these schemes still provide adequate performance, there ex-
ist other applications for which modulations having greater flexibility and sophistication are
absolutely essential. An example of an application requiring increased flexibility might be a
satellite channel over which it is desired to simultaneously send two independent and gener-
ically different types of information. In this instance, we would need a modulation scheme
capable of handling two asynchronous data streams with, in general, different rates, dif-
ferent powers, and different formats. Both the Tracking and Data Relay Satellite System
(TDRSS) and Space Shuttle make widespread use of such modulation schemes for com-
municating various combinations of operational, scientific, and telemetry data. An example
of an application requiring increased sophistication of design is today’s commercial mo-
dem market where baud rates as high as 19.2 Kbps are being transmitted over a nominal 5
KHz bandwidth telephone channel. Here the focus is on modulation schemes that achieve
a high degree of bandwidth conservation with as little expenditure of additional power as
possible.
For the most part, all other digital modulations that meet the above requirements fall
into two primary categories: inphase-quadrature (I-Q) modulations and continuous phase
modulations (CPM), the latter most often being represented by a constant envelope signal
with digital frequency modulation. In certain special cases, CPM also has an I-Q represen-
tation and thus the two categories are not always divorced from one another. Nevertheless,
it is convenient to separate them this way since the generic implementation for conven-
tional CPM is in the form of a single carrier which is frequency modulated by a waveform
that maintains phase continuity from symbol interval to symbol interval. More about this
later on.
 622 Demodulation and Detection of Other Digital Modulations Chap. 10

pn
ig std
v2 cosa,t

VP, (1) < Figure 10.la An unbalanced QPSK

modulator

10.1 |-Q MODULATIONS

10.1.1 Unbalanced QPSK (UOQPSK)

Following the form of the generic I-Q modulation introduced in Chapter 4, an unbalanced
QPSK modulation (see Fig. 10.1a) can be written as

Ope D [VPidi(e) cos wet — Pdp(t)

J sin wt (10.1)
where w, is the radian carrier frequency, P|, P2 are, respectively, the powers of the inphase
and quadrature components of s(t), and d)(t), d2(t) are unit power inphase and quadrature
binary (+1) modulations. In general, d,(t) and d(t) are allowed to have different data rates,
that is, Ry; = 1/7; and R2 = 1/7», different data formats (e.g., NRZ, Manchester), and
different data types (e.g., coded or uncoded). Furthermore, the two data streams may be
asynchronous or synchronous the latter requiring that their data rates be integer related. In
the special case where P; = P2, T; = 72, the data streams are synchronous, and both are
uncoded NRZ waveforms, (10.1) characterizes balanced QPSK modulation as discussed in
Chapter 4.
The channel introduces the bandpass Gaussian noise process

n(t) = V2 [ne(t) COS Wet — ns (t) Sin wet} (10.2)

where n,(t) and ns(t) are lowpass bandlimited “white” Gaussian noise processes with flat
spectral density S,(f) = No/2 in the frequency interval ~B < f < B where BX fo =
w-/2m. Also, assume that the channel introduces a constant (time independent) but un-
known phase shift 6 into the transmitted signal resulting in

sh Oya [VPiai) Cos (wet + 6) — /Podp(t) sin (wet + 6) | (10.3)

Assuming initially the case of ideal coherent detection, then at the receiver, the sum of (10.2)
and (10.3), namely, r(t) = s(t,@) + n(t), is first demodulated! by the quadrature pair of
unmodulated carriers

re(t) = V2cos (wet +0); rs(t) = —V2 sin (wet + 0) (10.4)

1. We assume as before that the inphase and quadrature demodulators (multipliers) include lowpass filters that
eliminate second harmonics of the carrier.
Sec. 10.1 |-Q Modulations 623

sample att = (i+ 1)7,

Matched
Filter a;
(P, (t))

v2 cos @,t + 8)

r(t)

sample at t= (i+ 1)T,

—y¥ 2 sin(@,t+ 0)
Matched A
Filter b;
(Pp, (t))

Figure 10.1b An optimum coherent receiver for unbalanced QPSK

to produce the I and Q lowpass signals

€\(t) = V Pid\(t) + n-(t) cos@ + ns(t) sin@

(10.5)
€2(t) = Podo(t)
V/ + n¢(t) cos @— ns(t) sin é
Representing d)(t) and d2(t) as binary pulse streams, that is,
co CO

a(t)= )) aipit—-iN—vi); bi) = DY? bipt—it—v) (10.6)

1=—0O 1=—0O

where p;(t) and p2(t) are unit power pulses of duration 7| and 7, respectively, {a;} and {b;}
are +1 independent sequences at rates 1/7 and 1/7, respectively, and v1,v2 are random
independent timing epochs to allow for asynchronization, then after passing through the
I and Q matched filters and assuming 1/7), 1/72 « B, we obtain the output samples (at
times t = (i + 1)7; and t = (i + 1)7)

Vii =v Piai + Ne; Voi =v P2bi + Ns (10.7)

where N,. and N, are independent zero-mean Gaussian noise samples with variances No/27;
and No/27>, respectively. Suppose now that the sequences {a;} and {b;} correspond to
uncoded random data. Then, making independent hard decisions on the I and Q channels
(see Fig. 10.1b), results in the bit error probabilities

Pyi(E)
pI — feDace
oewoi eeL
No en
=a 2,) (E) =— 5| orferic |E22
No (10.8)
.

where Ep; = Pi 7; and Ey2 = P2T> are the respective energies per bit on the I and Q chan-
nels. Thus, in the ideal phase coherent case, the bit error probability performance per chan-
nel for UQPSK is identical in form to that of balanced QPSK.
We now consider the case where a carrier phase error exists between the received
signal and the I and Q demodulation reference signals. In particular, assume now that the
624 Demodulation and Detection of Other Digital Modulations Chap. 10

quadrature pair of demodulation reference signals of (10.4) are described by

7) = 4/2008 (wet 4 6): 7, (t) = —V2sin (cetan.6) (10.9)

where @ is an estimate of the received signal phase 9 such as that produced by a carrier syn-
chronization subsystem. Demodulating r(t) with the I and Q references of (10.9) produces
€1(t) = J Pid\(t) cos d _ J Prdx(t) sind
+: Nc(t) cos 0 + ns(t) sin 6 (10 10)

€9(t) = J Prdx(t) cos@ + J Pid (t) sing +ns(t) cos6 — n-(t) sin 0

where ¢ 46 — 6 is the carrier phase error assumed to be constant” over the observation
interval pertinent to making decisions on d\(t) and d2(t).
Consider representing d)(t) and d(t) as in (10.6). After passing through the normal-
ized I and Q matched filters, we obtain the output samples [at time ¢ = (i + 1)T]

Vii = Vv Pia; cos @ — ¥ P2d2,; sing + Ne,1 cos6 + Ns\1 sin 0

= J Pia; cos@ — ¥ Pod2,; sing + Ny

(10.11)
V2,i = J P2bj coso + V Pidi,2 sing + Ns,2cos@ — Ne,2 sin@

= / Pobj cos + V Pidj,.2 sing + N2

where
1 (i+1)7T) +, ;
Hi=— dy(t)pi(t — iT, — v)dt
Ti Jit, +)
1 (i+1)T2+v2
i= s- d\(t) p2(t — iT2 — v2)dt
P T2 JiT+v2
| (10.12)
1 (i+1)T+v~
Nek = => Ne(t) pelt — iT; — vy)dt; k=1,2
Tk UTK+V

1 (+1) Ty +v%
Ns k => ns(t)px(t
—iT; —vyjdt; k=1,2
T Th +V%

Here N, and N>2 are zero-mean independent noise samples with variances Oe = No/2T; and
os = No/27?, respectively. We note from (10.11) that, as for balanced QPSK, the presence
of carrier phase error introduces crosstalk into the detection process, that is, the data in one
channel affects the detection of the data in the other channel. The nature of the crosstalk,
however, now depends on both the powers in the I and Q channels as well as the I and Q
data formats through the definitions of d),2 and d2,; in (10.12). To assess the effect of this
crosstalk on uncoded bit error probability performance, we proceed as follows.

2. In Volume II of the book, we shall introduce other phase error variation models wherein the phase error can
vary in time over the duration of a data decision. For now, we consider only the so-called slowly varying phase
error case.
Sec. 10.1 |-Q Modulations 625

The data estimates a; and b; are found from

a; =sgn{V,;}; b; = sgn {V>,;} (10.13)

Hence, the I and Q bit error probabilities conditioned on a fixed phase error @ are given by

2 1 Pris PoT,
oe a eet Fi sno
\ No \ No
(10.14a)
i _ [Pe
ae qerfe
ote| pastue
ese— |
nd
anne]

and

: #bi|o} = je
Po( E1) = Pr {6; 1 PoT cos
[7 Pi
+ Feaiasn|
(10.14b)
1 | P2T> | P;T> :
+ —erf
Te | a No cos @ — .|—d
No 1,.2sing

where the overbar denotes averaging over the data sequences and uniformly distributed
random epochs v, and v2 in d; 2 and d2 1.
In principle, (10.14a) and (10.14b) can be evaluated exactly for given pulse shapes
pi(t) and p2(t), that is, data formats on the I and Q channels. This, however becomes a
tedious calculation when, for example, the ratio of data rates R2/7R, is large? since the
statistical average over the data sequence required in Pp( E| d) of (10.14a) will then require
evaluation of dz; for each possible sequence d2(t) in the interval nT) + vj <t <(n+
1)T; + v;. Furthermore, in order to ultimately calculate the I-channel average probability
of error [by averaging (10.14a) over the probability density function f(@) of @], these
statistical averages over d2(t) must be computed for each value of @ in the domain of f(@).
To get around this computational bottleneck, a scheme was proposed in [1] that ex-
pands Pp;(E| ); i = 1, 2 into a power series in @ whereupon maintaining only the first few
terms, the average error probability P,; for the I and Q channels is easily evaluated as a func-
tion of E4;/No, E2/No, the mean-squared crosstalks dS and rae and the mean-squared
phase error, O%. For small values of OG, as is typically the case of interest, this scheme gives
credible results. We now proceed to describe this simplified evaluation technique.
From (10.14a) and (10.14b), we note that P,;(E| @); i = 1, 2 is of the form

———— en
PACE oC) = gore [Aj cosf + &; sing] + gone [Aj cos@ — & sing] (10.15)

where A; is a constant and &; is a random variable in the sense that it is both data sequence

3. With no loss in generality, we shall assume that the data rate 722 = 1/7} is arbitrarily chosen to be greater
than or equal to the data rate R, = 1/7}.
626 Demodulation and Detection of Other Digital Modulations Chap. 10

and random epoch dependent. Differentiating P,;( |) twice with respect to @ and evalu-
ating the results at ¢ = O gives
APHCEIO) | we Ae aye
oe a gion (-A7) =°
; (10.16)
a>2 Pyi(El~)|
Py; | _ Aj Aiexp(~Aj)
— A; [1+287]
as
de> ey 35 Ja
Furthermore, from (10.15)
1
Pri( E| ’)|g=0 = moe Aj ONT)

which corresponds to the bit error probability performance in the absence of carrier phase
error. Thus, expanding (10.15) in a Taylor series about ¢ and keeping only the first three
terms gives, upon substitution of (10.16) and (10.17)

d et——)
Pyi(E o a
51 d* Py; E Pe
= Pr
E| @)SS
Pyi( b)|g-0 +——
Eli(
es (10.18)
=a _, Aiexp(—A 4)
rie AGS 25a
Comparing (10.15) with (10.14) and identifying A; and &; in terms of the receiver parame-
ters, (10.18) becomes

Ep, E oe E a=
Pri (E| 9) = erfe |= + sy nen )
Bt) f+ 2S yeh |
: (10.19)

PioCEh li) c= ence) —

I Eng Aa En exo} Ey2 |
= ——
Ey 431 | 15
b2( EO) = 5 \ No 2V xNo No Me aera
where
Tht) ARGS}
Y=Aes
tae
eS] (10.20)

denotes the ratio of the I and Q data rates.

Finally, averaging (10.19) over the phase error probability density function f(@) gives
the desired result for I and Q channel average bit error probability performance, namely

Eo, Epi Fu) 1 pad

En =2 2
‘ Wenge Vane? ir et
a (B) = 5erfe

(10.21)
Pia(E) % 5erfe Epba E
Fb2 ie2) rey Ene
eee o3
Noa No No yr

Note that (10.21) consists of the sum of a term representing perfect carrier synchronization
performance [see (10.8)] and a term proportional to the mean-squared crosstalk and mean-
squared phase error. What remains is to evaluate the mean-squared crosstalks ds , and a. 2
Sec. 10.1 |-Q Modulations 627

Table 10.1 Evaluation of Mean-Squared Crosstalk for NRZ and Manchester

Data Formats
i a a ee ee SO es
d2(t)
Manchester NRZ

Manchester : EE = ae+ 12—- 6yr| ; 1 aad = ihe ae)

Ri < R2 < 2R, Reha =2Ry
d\(t) mri R2=2Ry Uae eR
NRZ 1
bn 1
1— ao

for various combinations of data formats in the I and Q channels. From the definitions of
dz,; and d;,2 in (10.12) and the approach taken in Chapter 2, Section 2.1, it is straightfor-
ward to show that

Ba= af Smilf)Snat Pods

Bacnn 1 one)

zz (10.22)

Bi= > iE SCM

San Hah =ody 2
Ss 1 CO

where Sini(f) and Sm2(f) are the power spectral densities of the modulations d)(t) and
dz(t), respectively. In [1], closed form expressions for a > are evaluated for all combina-
tions of Manchester and NRZ coding on the I and Q channels. The results are summarized
in Table 10.1.

10.1.2 Quadrature Amplitude Modulation (QAM)

An I-Q modulation formed by quadrature carrier multiplexing two M-AM modulations is

called quadrature amplitude modulation (QAM).‘ The signal constellation corresponding
to such a signal takes the form of a rectangular array (grid) of points in the form of a
square lattice (see Fig. 10.2 for M = 16). If each M-AM modulation corresponds to a signal
set of size 2” (m integer), then the QAM signal set contains M = (2”)* = 4” signals. In
accordance with the above description, a QAM signal can be mathematically represented by
(see Fig. 10.3a)

s(t) = V2A [d,(t) cos wet — ds(t) sin wet] (10.23)

where A is the signal amplitude (which will shortly be related to the average signal power),
and d,(t), d;(t) are, respectively, the inphase and quadrature data pulse streams, that is,
lo.@) oe)

dty= )° apt iT); de(t)= \ > bip(t —iT) (10.24)

i=—0O i=—0O

In (10.24), p(t) is a unit power pulse of duration 7;, 7; is the symbol time which is related
to the bit time by 7; = 2m7), and the data amplitudes {a;} and {b;} range over the set of
values +1, +3,...,+/M — 1. Note that for M = 4, QAM is identical to QPSK.

4. In the special case where the I and Q amplitude modulations correspond to amplitude-shift-keying (ASK), the
quadrature multiplexed form is often referred to as quadrature amplitude-shift-keying (QASK).
628 Demodulation and Detection of Other Digital Modulations Chap. 10

Figure 10.2 The signal constellation for

| QAM with M = 16

The average power P of s(t) is easily computed as

Pie ae 1 | FMD 40-1 )

P= A? (a? +?) =A? (+1) |=> (M —1) 4? (10.25)
|gael eco hind fone ad ee rece

Assume now that (10.26) is transmitted over an AWGN as characterized by (10.2) and
in addition the channel introduces a constant phase into s(t) producing
s(t, 0) = A [d,(t) cos (Wet + 8) — ds(t) sin (wet + 8)] (10.26)

Thus, the received signal is the sum of (10.2) and (10.26), namely, r(t) = s(t, 0) + n(t).

Coherent detection. In the case of ideal coherent detection, the received signal
is first demodulated (see Fig. 10.3) by the pair of quadrature carriers in (10.4) producing the
lowpass signals
E-(t) = Ad-(t) + n-(t) cos @ + ns(t) sind
(10.27)
&s5(t) = Ad, (t) +ns(t) cos @ — n-(t) sin@

After passing through the I and Q matched filters and assuming 1/7; < B, we obtain the
output samples [at time t = (i + 1)T;]

Vo,i = Aaj + Ne,i

(10.28)
Vs.i = Abi + Ns,i
where N,,; and N;j are independent zero-mean Gaussian noise samples with variance
No/2T;. A multilevel threshold decision (see Fig. 10.3b) is made on V,.; and V; ; to produce
the I and Q data estimates a; and b;. A correct QAM symbol decision then corresponds to
Sec. 10.1 l-Q Modulations 629

Ad, (t)

2 COS @,t
S(t)

Ad,(t)
Figure 10.3a A QAM modulator

sample at t = (i+ 1)T,

Matched
Filter &

(p(t)

v2 cox @,t + A)

r(t)

sample att = (i+ LT,

—y¥2sin(@,t+ 6)
Matched
Filter
(p(t))

Figure 10.3b An optimum coherent receiver for QAM

aj = a; and b; = b;. The average probability of a correct symbol decision is computed as

follows.
For a QAM signal constellation with M points, there are 4 corner points, 4 (vM— 2)
2
exterior points, and (VM — 2) interior points. Because of the symmetry of the constella-
tion, it is sufficient to examine only the points in the first quadrant. For each of the interior
points, the probability of a correct decision is

Pya(C) = Pr {Aq — 1) < Voi < AC + 1); AG — 1) < Ves < AG + D)}
=Pr{-A< Noi <A;—-A< Ng < A}
2
1 vf AZT; 1 f AZT;
= =5 eric a No —=—s—5 CLL No (10.29)

2
A2T.
Lo lithcveete =} | =('—2P;(B))*
630 Demodulation and Detection of Other Digital Modulations Chap. 10

For each of the exterior points, the probability of a correct decision is

Pyg(C) = Pr {A(ai — 1) < Voi < 00; Ai — 1) < Vs,5 < Ai + ‘DY
=Pr{—A< Noi <00; -A<Nyi < A}
AT. 1 AT, 1 55 Wer .(10.30a)
=| —5 erfc eee
No =5 erfc
erfi No 5 e c | ,/——
No

= (1 — Ps (E)) i — 2Ps1(E))
or

P,p(C) = Pr {A(ai — 1) < Voi < A(ai + 1); Ai — 1) S Vs,i < 00}
— Pr {—A SING SAP SANG oo}

1 A?T, \— 1 A2T, I ACT, Vie)

= JeClic erfc = —o
2, No 2 No 2 No

= (1 — 2Ps5\(E)) (1 — Psi (E))

Finally, for the single corner point, the probability of a correct decision is

P3c(C) = Pr {A@; — 1) 3 Ve; = 00; Ahi — 1) 5 Vsi = 00}

=Pr{—-A= Nj = 00; -A = Ng; = 00}
5 (10.31)
AT,
5
= | —erfc |— No =.(1— Psi (E))”

The average probability of symbol error is then

P,(E) =1 — P,(C)
=1—— | {4 — P(E)? 2 +4 (VM 2)d= P(E) = 2P (BD)

+ (vm -2) (1 -2P,(E)?

Ail (Pa)|=) Pe “e)|Mes (:re=a)

1
ees hyPate]
2 10.3
(10.32)

1 3 E,
( ra) (cara) #)
=) >) ——— F =
Se ee

< [1-3(1- Fe te(( 3 a)

2 M 2(M —1)/ No
Sec. 10.1. ~- |-Q Modulations 631

where we have substituted for A? in terms of P from (10.25) and as in previous chapters
E;/No & PT;/No denotes the average symbol energy-to-noise ratio.
It is interesting to observe that (10.32) can also be expressed in terms of the average
error probability performance of M-AM , that is, the individual modulations on the I and
Q channels. Since a correct QAM decision is made only when a correct decision is made on
both the I and Q channels, then

Ps(C)lm-cam = [Ps(C)| yazan | (10.33)

2

or, equivalently
2
P;(E)|m—qam = 1 — [1= FAC ae
(10.34)
1
=2P;(E)| az_am (1~ 5 PAE apna)
which, using the result for the symbol error probability of M-AM [see (4. 167)}° and replac-
ing M by VM and E, by E,/2, demonstrates the equivalence.
To obtain the average bit error probability of QAM from the average symbol error
probability of (10.32), we observe that QAM can be assigned a perfect 2-dimensional Gray
code [2], that is, for each symbol error in the set, all adjacent symbol errors cause only a
single bit error. Thus, since the number of bits/symbol is log, M, then for large SNR we
have the simple approximate result that

P,(E) = 5
Go) US) (10.35)
og,M sm

where in addition the symbol energy-to-noise ratio E;/No is equal to 2m times the bit
energy-to-noise ratio E,/No. The values of E»/No corresponding to P, = 10-> are tabu-
lated in Table 10.2.
An exact evaluation of bit error probability can be obtained by considering the number
of bit errors that are associated with the QAM symbols having each of the symbol error
probability types, P;4(E), Psg(E), and Psc(E) and then averaging over all possible cases.
For arbitrary M, it is tedious to express the result in closed form. However, for any particular
value of M, an answer can be obtained. For example, for M = 16 and M = 64 and a Gray
code mapping (see Chapter 4), it can be shown that

3 PU|e atte lia ieee

anny i 2 Es | (10086
fe ateet ( eae,
ait tye ( 3 t= jt erie) ( Sums
Hoty Beet
5. The reader is cautioned that (4.167) applies to asymmetric M-AM where all of the signal points lie on the
positive real line. For symmetric M-AM (the signal points are symmetrically located about the origin) as referred
to here in drawing our analogy with QAM, the equivalent symbol error probability expression would be

P(E) Mam = (=F *)ere ((a ‘|

where we have furthermore replaced Ey, by Es for consistency with the usage in this chapter.
632 Demodulation and Detection of Other Digital Modulations Chap. 10

Table 10.2 Bandwidth and Power Efficiencies of Various Modulation Schemes (P;, = 10-5,
Values of E,/No are in dB)
I Ne Re a GS SS iad =A a Sl naeLt Bat 2
Constant Envelope
QFPM CPFPM N-FSK/
QFPM (N-FSK/QPSK) CPQFPM (MFMSK) BPSK QAM _ M-PSK

Se Aa Ce MR aie E E; E
ee E
LIEy

41.0 9.60 1.0 9.60 0.67 9.60 0.67 9.60 0.76 9.60 1.0 9.6 1.0 9.6
Gree, ae ee kG 8.31 2. "0.60. 83 0830831. = salem
16 1.33 9.60 0.8 7.36 0.80 9.60 0.44 7.36 0.71 7.36 2.0 13.4 2.0 17.4
BOG ae yt RKONSSE 67! = (Sv"0.39" 6.700.5206.7 1 =o = es
6401222831 035. rele 0.67 8.31 0.18 6.16 0.34 6.16 3.0 17.8 3.0 27.5
256 0.89 7.36 0.12 5.36 047936 ee ee
1024 0.59 6.71 0.04 4.80 0:30. 6.71 = See 5 RTS es Onion
4096 0.36 6.16 0.01 4.40 O18 6.16) =) —) 22260. 57.61a0NGes

and

Ey 1 Ep
P,(E) = — erfc — }+ —erfc {|3,/ —
lB) = 74 (i, 4 ( i,
(10.36b)

_ bah (sey = Date ( | _ ere (13it]

24 7No 24 7No 24 7No

respectively, both of which simplify to (10.35) for large SNR.

An alternate and simpler procedure is to observe® that because the I and Q channels
operate independently, the average bit error probability for M-ary QAM is equal to the
average bit error probability for /M-AM. Thus, substituting /M for M and E,/2 for Es
(the same substitutions as were used to demonstrate the relation between the symbol error
probabilities of QAM and /M-AM) in the expression for average bit error probability of
M-AM, we again obtain (10.36). Figure 10.4 is an illustration of (10.36) together with the
results for M = 4, that is, QPSK. The type of relation in (10.34) and (10.36) between the
average error probability performance of an I-Q modulation and that of its I and Q channels
individually is typical of quadrature carrier multiplexed modulations.
We now consider the case where a carrier phase error exists between the received
signal and the I and Q demodulation reference signals. Demodulating r(t) with the I and
Q reference signals of (10.9) and passing the results through the I and Q matched filters, we
obtain the output samples [at time t = (i + 1)T;]

Voi = Ala; cos¢ — bj sing] + Noi

(10.37)
Vs,i = A[b; cos@+ a; sing] + Ni;

where Noi and Nii are independent zero-mean Gaussian noise samples with variances
No/2Ts.

6. This observation was made in [5] for the special case M = 16.
Sec. 10.1 |-Q Modulations 633

To compute the conditional probability of symbol error we again consider the proba-
bility of a correct decision (conditioned on @) for each of the three different types of points
in the first quadrant of the constellation. Defining

P.(C| j,k) Apr| 4,= ab = bila: 0; af, (10.38)

then for each of the interior points, the conditional probability of a correct symbol decision
is

P,(C|j,k, @) =Pr {AG —1)< Voi < AG H+ 1); AK—-1) < Vi < AK +)
=Pr{A[—1+ j(1 —cos$) +ksing < N,; <1+ j(1—cos¢) +ksing];

A[—1+k(1 —cos¢) — jsing < Ni; <1+k(1 —cos¢) — jsing]}

1 | A2T, A‘ ;
= 5 erfc No 1411-6) +4sne

1 AZT; :
— 5 erf No (1+ jC —cos¢)
+k sing) (10.39)

1 AZT;
(—1+k(1 —cos¢) — j sing)
» No

1 A2T;
(1+k(1 —cos@¢) — j sing)
2 No

jrkdy 3, es (Vm -3)

For each of the exterior points, the conditional probability of a correct decision is

Py(Clivks @) = Pr [A(/M ~ 2) < Voi $00; Ak 1) < Voi SAK + D}

b xe . an (-1 + (VM ~ 1)(1 ~ cos) + k sing)

x xe ‘ oa (—1 + (1 — cos) — (VM —1) sing) (10.40a)

—5 erfe [AB ¢ +k(1—cos¢) — (vm - 1)sin)

k=1,3,...,(VM -3)
 634 Demodulation and Detection of Other Digital Modulations Chap. 10

1071

10-2

10-3

10-5

0 2 4 6 See t0M 124 Sal eS ae20 Figure 10.4 Exact bit error rate of 4, 16,
and 64 QAM

or

P,(Clj,k, 6) =Pr {AG 1) < Voi SAG +); AWM - 2) < Vs, < 00]
2 ferfe JAE (-1+ id ~ cos) + (VM ~ 1)sind)

—5erfe [AE (14 10 cos) + (VM ~ 1)sing) (10.40b)

1 AZT;
x | —erfc (—1+ (VM — 1)(1 ~ cos) — j sing) ;
2) No

j=1,3;...,(¥M-3)

Finally, for the single corner point, the conditional probability of a correct decision is
Sec. 10.1 |-OQ Modulations 635

P,(C| j,k.) =Pr {ACM - 2) < Voi <00; A(WM — 2) <V,; < 00]

= 5erfe saat(—1+ (VM - 1)(1 ~cos¢ + sing)) at

3 1crf AX (-14 (V# 1) d ~cos9

~ sing)

From the symmetry of the signal set and the decision regions, the following properties
hold:

Ps (C| j,k,
b) = Ps (C| — j, —k,)
Ps (C| j,k, 6) = Ps (C| j, —k, —¢) (10.42)
P;(C| j,k, 6) = Ps (Clk, j, —?)

The average probability of correct detection (conditioned on @) is given by

P.(C|) = PE SEECh ko) (10.43)

Bs, VM-\1/M-1

us j=1,3,..k=1,3....
isk

Gh,
10.45
Deter a= (O ( )

Substituting (10.40a), (10.41), and (10.42) into (10.44) and combining like terms gives the
final form of the desired result, namely [3]
636 Demodulation and Detection of Other Digital Modulations Chap. 10

; +(/M-1) +(VM— 2) 4
P.(E|d) = 1 — P(C1$) = = ps oe
j=

x ate Gren. e+ oe

oo — 1 (10.46)
M hale A lOc Dyers 2

a 5 a
«ete (<a)
la
Ne aa
No (k+(U—k)cos¢+
’
(1 vt ind)

1 3 :
< fete Gaa)e d+ —1)cos¢ — sing)

The conditional error probability of (10.46) is plotted in Fig. 10.5 for M = 16 with A=
- wo i a as a parameter. The limiting cases of (10.46), namely, A — 0 and A — ov, are
also illustrated in this figure and are mathematically described by
1
Ae P,(E|¢) =1— v (10.47)

+(VM-1) +(/M-2)
lim pws) ee Sy [5 588m + (100s ~ jsind)|
Aes ee Mee iat ea eee
+(VM-2) +(/M-2)
4 1 1
— — Ly [p- pen +0 Heose-+d— Dsing)|
M 0, 50,44,...1=0,4044,.

x E — 5sen +d —1)cos¢ — (k — Ising)

Note that the limiting curve for A — oo exhibits discontinuities at values of phase error
at which the four signal points in the first quadrant move across a decision boundary (see
Fig. 10.6), namely, @; = 16°53’, 2 = 18°26’, 3 = 20°48’, and ¢5 = 45°. The absence of a
discontinuity at 4 = 32°20’ is explained as follows. For a phase error of value 3, the signal
point at (3A, A) moves across a decision boundary and into the decision region of the signal
at (3A, 3A), thus resulting in an error in its detection (see Fig. 10.6). When the phase error
increases to #4, the signal point at (3A, A) moves across another decision boundary and into
the decision region of the signal at (A, 3A). However, this still results in an error insofar as
detecting the transmission of the signal point at (3A, A); thus, there is no discontinuity in
the conditional error probability at dg.
In order to calculate the average error probability performance of QAM in the pres-
ence of carrier phase error, we must characterize the pdf f(@) of this phase error. Since this
characterization in turn depends on the particular carrier synchronization loop used, we shall
delay our discussion of average error probability performance until Volume II of this book,
 Sec. 10.1 |-Q Modulations 637

Se Soe ee

(E\9)
Pg

0 0.2 0.4 0.6 0.8 1.0 Figure 10.5 Conditional error probability
4 d performance of 16-ary QAM (Reprinted
Tt from [3] © IEEE)

Figure 10.6 Values of phase error at

which quadrant signal points cross decision
boundaries (Reprinted from [3] © IEEE)

where we shall present carrier synchronization loops for QAM and their associated f()’s.
At that time we shall be in a position to average (10.44) over these pdf’s.

Differential detection. Ordinarily, one thinks of differential detection only in

the context of constant amplitude modulations such as M-PSK. However, by applying
maximum-likelihood decision theory in the presence of unknown channel phase to com-
bined amplitude-phase modulations, it is possible to develop a differential detection deci-
sion metric for the special case of QAM. As was done for M-PSK in Chapter 7, the results
will be obtained for the general case of observation over N;(Ns > 2) symbol intervals. This
leads to an MLSE algorithm that performs multiple symbol differential detection of QAM,
herein referred to as DOAM.
638 Demodulation and Detection of Other Digital Modulations Chap. 10

Without reference to maximum-likelihood theory, one might think that, at least in

principle, differential detection of combined amplitude-phase modulations could be accom-
plished by treating the phase and amplitude of each signal point as separate entities and as
such applying differential detection to their phase and noncoherent (envelope) detection to
their amplitudes. Such a scheme was proposed for DQAM in [4] in the context of its appli-
cation in frequency-hopped spread spectrum systems. Unfortunately, because this scheme
is quite suboptimum from a maximum-likelihood standpoint, its performance is very poor.
Without further ado, we now present the maximum-likelihood approach to DQAM, fol-
lowed by an evaluation of its performance over the AWGN.
If we assume that the receiver has no knowledge of the channel phase, 0, then without
loss in generality we can assume that the quadrature demodulation references have zero
phase. This is equivalent to setting 6 =0 in the reference signals of (10.9). Then, using
complex notation, we obtain analogous to (10.37)

Vi A Vek + UVs.k = A (ae + fbx) cosO + A (—b_ + jax) sind + Neg a ING,
(10.48)
A Acpe*el® + N;

where cye/% A a, + jb, is the amplitude-phase complex representation of the transmitted

QAM signal in the k" transmission interval and N,2N/,+ jN,, is a sample of a zero-
mean complex Gaussian noise process with variance ae = No/T;. Consider now a sequence
of V;’s of length N;. This sequence, represented as a row vector, V, is expressed as

V =se/® +N’ (10.49)

where Vz_;, Sx—; = Acz_je/**-i, and Ni; are, respectively, the i elements’ i = 0, 1, 2,
.., Ns — 1 of the Ns-length row vectors V and N’. For the assumed constant (but unknown)
phase channel, the conditional probability of V given s and @ is

1 1 =19
fy\s,o V\|s,(Vs, 6)
@) = ———-
Q DN, exp ——|\VEV
5 7 | | (10.50)

Using the definition of V, the argument of the exponential in (10.50) can be expressed as

lV —¥|° =IVI? + Isl? — 2 |vs*? |cos(@ — @) (10.51)

where

Im {Vs*"}
of Atay ee 10.52
Re {Vs*"} ( )
Substituting (10.51) into (10.50) and averaging over the uniform pdf of 6, we get
4

fyis(V|s) = fvjs,a(V\s, 8) fo(0)de

1

(10.53)
= a
Qo)"
x {515
20;
(IMP? (05-2 vs"? )
+I?)|10aes
7. For notational convenience here, we number the elements of an N,-length sequence vector by 0,1,...,N;—1
rather than by 1, 2,..., Ns. A similar notation applies in two dimensions to matrices.
Sec. 10.1 |-Q Modulations 639

where as before /o(x) is the zero order modified Bessel function of the first kind. Since the
natural logarithm is a monotonically increasing function of its argument, then maximizing
fvjs(V|s) over s is equivalent to maximizing In fyjs(V|s) over s. Hence, eliminating all
irrelevant terms (that is, those that are independent of the transmitted data), the following
decision rule results: Choose the sequence s that maximizes the metric

n= saa?
e +In Io (o,,*)vstT

1 N;-1 Ns—1
(10.54)
Fane? pure, 4 int (00 DOV pcicni
On i=0 i=0
Unfortunately, as was true for the analogous result for M-PSK derived in Chapter 7, the met-
ric of (10.54) is ambiguous in phase in the sense that the argument of the Bessel function
is unchanged by a fixed phase rotation. Thus, even in the absence of noise, it is possible to
make an erroneous sequence decision. The simplest case of this to envision occurs where the
phase rotation is a multiple, say J, of 90° in which case a particular sequence of QAM sym-
bols is mistaken for another sequence whose symbols correspond to a rotation of the true
symbols by J quadrants. To resolve this ambiguity, we again propose the use of differential
encoding at the transmitter. Thus, if AQ, now denotes the phase of the k" transmitted com-
plex QAM symbol, that is, s, = cye/4%, then the actual phase transmitted over the channel
in this same interval is given by the usual differential encoding relationship

Ox = Ox_1 + AOz (10.55)

Note that unlike M-PSK, the phases {6,} do not necessarily take on values corresponding
to QAM symbols. This, however, is of no consequence since indeed (10.54) holds for
any AM-PM constellation. Finally, then, substituting (10.55) into (10.54) gives the desired
unambiguous metric
Nea
Ns =i) 3 AO—i—m
(10.56)
1 iba

(Es eR Dec lola, oy) Vecwaie

r=()
on =

Using the approximation (valid for large arguments, that is, high SNR) In Jo(x) = x, then
after normalization by 0,7, (10.56) becomes
Ns—i-2
| Ns—1 j >y A®k—i-m
n=—5 Dy cgay eS Vi-ice-ie —"™=9 (10.57)
ji i=0

Analogous to Fig. 7.7, a direct implementation of a receiver that employs a decision rule
based on the metric in (10.57) is illustrated in Fig. 10.7. Note that the receiver uses one
extra complex amplitude, namely, cx—n, + 1, to make its decision on cy—j, A@-i} 1 =
Ge SINae:
Analysis of the average bit error probability performance of (10.57) is difficult. Thus,
we present in Fig. 10.8 computer simulation results typical of this performance for the
case of 16-ary DQAM. Also illustrated for comparison is the performance of ideal coherent
detection of QAM as given by (10.36). We see that despite the use of a maximum-likelihood
 a[dure
13s= “11
peyoie,y 7) Aeylaq "yy Aejaq
49A!4 : l Aepoq myPr keoq Ny
Zi

640
SSoI
= ppYCal
‘ boned “IN

Y w—1-¥
Ip= x4 9
;

WOU |
———

edojaaua
10}99}9p

oredwo5
pue
199T9S
91

WN sapako D,
JadN ‘L aqeue kyafoA9 out)
Aeyap

SOV. a VV
7

v A
v
v
vv

PInoy_L“OT WY AIP UONRJUS

JoB WIE[dUTJoAro9eI
T paseg
UO ay) omeu!
Jo Uonenbg(L¢‘01)
 Sec. 10.1 l-Q Modulations 641

10-1

10-2},

10-4

Figure 10.8 Average bit error probability

8 OR Ore ecm ass © 115) S168 a7” 18
performance of multiple symbol detection of
16-ary DQAM

receiver, the cost (in E,/No) of using DQAM relative to coherent QAM is quite significant,
although one can achieve a good reduction in this cost by increasing the observation interval
from two to three symbols times. Further increase in the observation interval at the expense
of increased complexity of the receiver processing brings one closer and closer to the ideal
coherent performance.

10.1.3 Quadrature Biorthogonal Modulation

In Chapter 4, we defined a biorthogonal modulation as one that generates an N-dimensional

signal set of size M = 2N consisting of M/2 = N pairs of antipodal signals, each pair being
orthogonal to all‘other pairs. We also discussed a particular realization of such a scheme as
a combined frequency/phase modulation (FPM) wherein N frequencies uniformly separated
by 2 fa = k/2T, (k an integer, 7; the signaling interval) are each BPSK modulated, that is,
N-FSK/BPSK. Typically, N = 2” (n integer) and the frequencies are symmetrically placed
around a carrier f-(= w,/2m), that is, fo+ fa, fe +3fa,---, fe + (N — 1) fa. Also, the
minimum frequency spacing required to guarantee orthogonality, namely, 2 fg = 1/27;(k =
1), corresponds to a modulation index h = 2 faT; = 0.5.
Consider now a modulation scheme that results from quadrature carrier multiplexing
two N-dimensional biorthogonal signal sets as characterized above. This scheme is referred
to as quadrature biorthogonal modulation (QBOM) [7] and yields a 2N-dimensional sig-
nal set of size M = (2N)*. Analogous to FPM, a particular embodiment of QBOM results
642 Demodulation and Detection of Other Digital Modulations Chap. 10

from modulating the orthogonal carriers cos w-t and sin @-t with N-FSK/BPSK modula-
tions. Here the minimum frequency spacing required for orthogonality corresponds to se-
lecting h = 1(faT; = 0.5) for each of the two N-FSK/BPSK modulations and the resulting
modulation is referred to as quadrature frequency/phase modulation (QF PM) [7]. A math-
ematical representation of a QFPM signal and its bit error probability performance on an
AWGN are presented in the following section.

Quadrature frequency/phase modulation (QFPM). Consider as a data

source a binary random sequence {d;} of rate 1/7» [equivalently, a rectangular pulse data
stream d(t)]. Next subdivide this sequence into two subsequences {d,,;} and {d;;} corre-
sponding, respectively, to the even-numbered and odd-numbered bits in {d;} and generating
rectangular pulse data streams d,(t) and d;(t). Since for each QFPM symbol of dura-
tion 7; there are log, M = log,(2N y=2(1+ log, N) bits of duration 7;, then blocks of
4 = 1+ log, N bits of {d-;} and {d; ;} are respectively assigned to the inphase and quadra-
ture biorthogonal components of this symbol. The first bits in these blocks are assigned to
binary (+1) sequences {a¢,;} and {as ;}, respectively, which are used to define the BPSK por-
tions of the biorthogonal modulations. The remaining A — 1 = log, N bits in each block are
assigned to N-ary [+1, +3,..., +(N — 1)] sequences {b,,;} and {bs;}, respectively, which
are used to define the N-FSK portions of the biorthogonal modulations.® Defining random
rectangular pulse trains a-(t), as(t), b-(t), and bs(t) from the sequences aej, asi, bc,i, and
(oe)

bs.i, for example, ac¢(t)= >> dei p(t —iTs), then a QFPM signal is characterized by
1=—0O

s(t) = VP {ac(t) cos [(we + be(t)wa) t] + as(t) sin [(e + bs (t)@a) t]}
(10.58)
= (Soh) Sst)

where P is the average signal power and wg = 27 fq is the radian frequency deviation about
the carrier radian frequency w,. Note that (10.58) represents a constant energy (E; = PT;)
but nonconstant envelope signal. To see this, we rewrite (10.58) as

s(t)
s(t) = Pa¢(t)
V {cos(en se baea) te cos [we Bey Aya aI
ac(t)

= 2/ Pa,(t) cos [oO at + *| (10.59)

be(t) + bs (t) T
X COS OEE TTS ens

where the upper signs correspond to a;(t)/a¢(t) = 1 and the lower signs to as(t)/a¢(t) =
—1. The cosine factor in (10.59) having a radian frequency [(b¢(t) — bs(t)) /2] wa rep-
resents the envelope of s(t), whereas the second cosine factor of radian frequency w, +
[(be(t) + bs(t)) /2] wa represents the carrier. Clearly, the envelope is constant only when
b-(t) = bs(t), that is, the inphase and quadrature component frequencies in (10.58) are iden-
tical. In this special case, the (2N)*-ary QFPM signal reduces to a 4N-ary N-FSK/QPSK
signal.
8. The coding rule used to map the data sequence {d,,;} into the binary amplitude and N-ary frequency sequences
{ac,i} and {b-,;}, and likewise the data sequence {d,;} into the binary amplitude and N-ary frequency sequences
{as,;} and {bs,;}, is identical to that given in Chapter 4 for realizing a biorthogonal signal set with an N-FSK/BPSK
modulation (see Example 4.6).
Sec. 10.1 |-O Modulations 643

a, (t)

eM a
4 FM f=120.0N ©)
ile Encoder Modulator
J

Demultiplexer
sin @,t

0, =@,-[N-(2n-l]@,; n=iorj

Figure 10.9a A block diagram of a modulator for (2N)?-ary QFPM

Block diagrams of modulators for (2N)*-ary QFPM and 4N-ary N-FSK/QPSK are
illustrated in Figures 10.9a and 10.9b, respectively. Since for the former, the quadrature
components of s(t), namely, s-(t) and s;(t), are orthogonal over the signaling interval T,,
they can be demodulated and detected independently. As such, an optimum coherent re-
ceiver for a QFPM signal (see Fig. 10.10) is a parallel combination of two optimum coherent
receivers for biorthogonal signals of the type illustrated in Fig. 4.17. Exact bit error proba-
bility results for VN-FSK/BPSK biorthogonal modulation are given by Eq. (4.107) together
with (4.102), (4.105), and (4.106). Since the two constituent N-FSK/BPSK signals in the
M-ary QFPM signal of (10.58) are demodulated and detected independently, then the bit
error probability performance of the latter is also given by Eq. (4.107), keeping in mind that
now N = /M/2 and thus E, = (1 + log, N) Ep = (3logy M) Ep. This equivalence be-
tween the average bit error probability of the resulting modulation and that of its individual
inphase and quadrature components is typical of quadrature multiplexed modulations, for
example, QPSK and QAM.
An upper bound on the bit error probability performance of QFPM that is simple to
evaluate can be obtained from the same upper bound on the performance of N-FSK/BPSK
and the equivalences noted above. In particular, since the probability of error of the first
kind, P;;(E), of (4.105) can be upper bounded by the error probability for antipodal signal-
ing, that is, 4 erfc ./E;/No, then from (4.107) together with (4.103) and (4.106) we have

1 |
P(E) = 5 Psi(E) te 5 Ps(B)

Es
te
ae |No
— er "|] [(— 9)) arte
erfc
Es
Dignity
Eig
fe.( |=4 (10.60)

ea)
]
ree ) ctleerfc it
E;
ane
1
oe Ccert |No Ee

where M = 2N. For (2N )?-ary QFPM with the same number of frequencies as the N-
FSK/BPSK scheme, the bit error probability has the same upper bound as in (10.60) which,
644 Demodulation and Detection of Other Digital Modulations Chap. 10

a, (t)

tH
5)
oO
=—
a.
om
—
:O
a)

@,= @,-(N-(2j-Nlo,
Figure 10.9b A block diagram of a modulator for 4N-ary N-FSK/QPSK

COs @,t
b=12 5-0. ,N

Optimum
Biorthogonal Decoder
Detector
Output
Received Data
Signal

Convert

Optimum
Biorthogonal Decoder
Detector

j =1,2,...,N

Figure 10.10 Block diagram of an optimum receiver for (2N)?-ary QFPM

when written explicitly in terms of NV, becomes

E
p(E) <<
Pp(E) erfc ,/(1+ log, N) se
No +<erfe
2 /(1+log)N)—No (10.61)
For the constant envelope 4N-ary N-FSK/QPSK modulation, the receiver in Fig.
10.10 is not optimum. The reason for this is that the two biorthogonal components s,(t)
and ss(t) are no longer independent of each other (they share the same frequency) and
thus should not be detected as if they were. More specifically, in Fig. 10.10, the common
frequency of s-(t) and s;(t) is redundantly detected twice (once in each biorthogonal detec-
tor) each time using only half the total power available. Rather, an optimum receiver (see
Fig. 10.11) for this special case of QFPM should coherently combine both quadrature sig-
nal components prior to detection of the common frequency thus utilizing the entire signal
 Sec. 10.1 |-Q Modulations
645

92 sin @,t

y(t)

d, (t)

2 cos @yt

Figure 10.11 Block diagram of a receiver for 4N-ary N-FSK/QPSK

power for this purpose. The idea behind such a configuration is generically the same as that
used in constructing the optimum biorthogonal (V-FSK/BPSK) receiver, namely, remove
the phase modulation for detecting the frequency, and then use the frequency decision to
determine which of the N parallel channels the phase modulation should be extracted from.
Removal of the QPSK modulation in Fig. 10.11 is accomplished at a given frequency by co-
herently adding and subtracting the I&D output samples (projections of the signal onto the
quadrature coordinate axes) and choosing the larger in absolute value of these two quanti-
ties. The maximum of the N largest absolute values so produced then provides the desired
frequency decision, and a binary hard decision on the amplitude in each of the quadrature
components corresponding to this frequency channel provides the QPSK phase decision.
To compute an upper bound on the bit error probability performance of Fig. 10.11,
646 Demodulation and Detection of Other Digital Modulations Chap. 10

we observe that for the N-FSK/QPSK signal of (10.58) with b(t) = b-(t) = b(t), ihe.sum
and difference I&D outputs of the i" frequency channel (at frequency @;), namely, V,; and
V._.,
n,i?
corresponding to an input signal at the same frequency, that is, wp + b(t)wa = @j, are
given by

+ EC ~ -
a 5 (45,0 + den) + (Men + 75.0)
(10.62)
Pe
=! iP oe ay.
Vie = 9 (den wt as.n) * (Ee 5 ne)

where for the noise representation of (10.2)

(n+1)T
nes = Vi| n-(t) sin [(@j — @e)t +2 /4] dt;
nT

(n+1)T
oe = v3| ns(t) sin [(@j — We) t + 72/4] dt
i (n+1)T
(10.63)
Non = v3| Nc(t) sin [(@j — we) t— 1/4] dt;
nT

(n+1)T
Non= vi| Ns (t) sin [(@j — We) t— 1/4] dt
Tiel

In (10.62), the noise components n* and n~ are zero mean and uncorrelated (also indepen-
dent since they are Gaussian) and have equal variance o,2 = No/T. The sum and difference
I&D outputs, V,"; and V,,;, of the j™ frequency channel (j 47), have zero signal com-
ponents (due to the orthogonality of the signaling frequencies over the symbol interval) and
noise components analogous to those in (10.62). Note that depending on whether as, = dan
OF ds. n FAc.n, the signal component of Vee CU ae)will have a magnitude equal to /2P.
Thus, the effective SNR of the signal used for frequency detection is 2PT,/No, which is a
3 dB improvement over the equivalent signal in Fig. 10.10. Once again, this improvement
comes about because of the coherent combining of the I and Q signals prior to frequency
detection which is possible when the two quadrature components of s(t) share a common
frequency.
An upper union bound on the symbol error probability performance of the receiver
in Fig. 10.11 can be obtained by observing that a symbol error results when any of the
following occur: a wrong frequency is detected, the sign of a¢(t) is erroneously detected,
the sign of a(t) is erroneously detected. As such

ie
|
Wo4)erfc i./(2+log,
5
B N ) Fes —ete
Ep 1
erf (2+
2+ log,oe N)
) ——
ao
P(E) Ep
< = (M —

(10.64)
1 Ep Ep
+ = éric _/ (2+log, N) == ON— I
2 i 62 ) ONG ( ye erkoy 2+ ine?
log, N) ——
) ON
Sec. 10.1 l-O Modulations 647

The first term of the first line in (10.64) corresponds to the detection of an incorrect fre-
quency, while the second and third terms correspond to erroneous detection of the signs of
as(t) and a,(t).
An upper bound on the bit error probability performance is obtained by employing
reasoning analogous to that used in relating bit error probability to symbol error probabil-
ity for biorthogonal (in particular, N-FSK/BPSK) signals. In particular, with reference to
Table 4.4, if a frequency error occurs, then on the average, half of the bits are in error. If an
error occurs in detecting the sign of as(t), then all but one [that corresponding to a¢(t)] of
the total number of bits are in error. Finally, if an error occurs in detecting the sign of a,(t),
then only one bit is in error. Since the total number of bits per symbol is 2 + log, N, then
from (10.64) and the above arguments, the average bit error probability has the upper bound

PoE)
ane < =)AS(=) =(M= (Mt —4)erfc
—4)erfe, |(2 + logy
fe) N) 22
52°") 3No
1+1 Iss
abses (2 + logy N
2+ log, N oe
(10.65)
] eh
2+ log, N
+ (Scher )ee= B2 ) aN

=5ON SDNate I(
(2 + logy N

which is simply one-half of the symbol error probability upper bound in (10.64). Finally,
we shall see later in this chapter that the structure of the receiver in Fig. 10.11 forms the
basis of the optimum receiver for continuous phase FPM, which is constant envelope by
construction.
Using the methods presented in Chapter 2, it is straightforward to show that the
equivalent lowpass two-sided power spectral density (PSD) of QFPM is given by [7]

N-—1
FS sine? (sr +i BS ~~)
Ss(f) = —_
He)
(10.66)
2 (1 +log, N) PT, X Ne
= * De sitte? (21 + log, N) fTp+i-~—)
i=0

which is illustrated in Fig. 10.12 for various values of N. Note that since QFPM corresponds
to a modulation index h = 1, its PSD is not? simply a frequency scaled (by a factor of two)
version of the PSD for N-FSK/BPSK (h = 0.5). This is readily seen by comparing (10.66)
with the comparable result in Example 4.6 and Fig. 10.12 with the comparable figure in that
example, that is, Fig. 4.45.
The null-to-null bandwidth, Wyn, (Ban = Wnn/2 is the single-sided lowpass null
bandwidth) of QFPM is easily obtained from (10.66) by setting the argument of the sinc

9. The exception to this statement occurs for the trivial case of N = | whereupon QFPM becomes QPSK and
N-FSK/BPSK becomes simply BPSK.
 648 Demodulation and Detection of Other Digital Modulations Chap. 10

dB
S,(/PTp,

-40 1.00 1.25 1.50 1.75 2.00

fb

dB
S,(f)/PTp,

Figure 10.12 Power spectral density of

40 000 O:252 10!50)0 0.7551. 00 falcon OOmmall 70m es00 QFPM with WN as a parameter (Reprinted
lp from [7] © IEEE)

function equal to unity when i = 0 and letting f = W,,,/2 with the result that
1 New 1
Weg Ne) Se (10.67)
Ts; 2(1+ log, N) T
Table 10.2 illustrates the bandwidth efficiency 5 = 1/WnnTp = 2(1 + logy N)/(N + 1) of
QFPM for various values of M = (2N)?. Also tabulated are the exact values of Ep/No
(in dB) for a bit error probability P, = 10~-° as determined from the exact N-FSK/BPSK
results of Chapter 4, that is, Eq. (4.107). We observe that in comparison with QAM and
MPSK, the bandwidth efficiency of QFPM is significantly lower for M > 16 as one might
expect; however, the power efficiency is significantly better. In particular, QFPM yields a
power savings of 3.8 dB for M = 16, 9.5 dB for M = 64, and 15.1 dB for M = 256 when
compared with QAM. When compared with M-PSK, the power savings are even larger;
however, the reader is reminded that M-PSK is a constant envelope modulation whereas,
in general, QFPM is not. For the special case of constant envelope QFPM (N-FSK/QPSK)
as discussed above, Table 10.2 also gives the exact values of E,/No (in dB) required for a
bit error probability P, = 107°.
If instead of using null bandwidth, we were to use a fractional out-of-band power defi-
nition of bandwidth (see Chapter 1) for computing bandwidth efficiency, then QFPM shapes
 Sec. 10.1 |-Q Modulations 649

Hon
i
z22e2
onnM—
ie NEG
dB
n,

- Figure 10.13 Fractional out-of-band

power versus normalized bandwidth W,T;
for QFPM with WN as a parameter (Reprinted
from [7] © IEEE)

Table 10.3. A Comparison of the Bandwidth Efficiencies of Various Modulation Schemes

Using a 90% Power Containment Bandwidth Definition (5 = 1/ Woo7,)

CPFPM QAM CPFSK —-N-FSK/2PSK

M QFPM CPQFPM (MFMSK) M-PSK (h=0.5) (h =0.5)
2 as “ hy 0.58 1.25 0.58
4 1.16 1.25 1.25 1.16 1.18 1.00
8 as hi 0.93 1.74 a ia
16 1.74 1.24 0.59 2.31 0.56 0.95
32 = = 0.36 2.90 = 0.65
64 1.52 0.89 0.21 3.47 0.21 0.41
128 = we 0.12 4.07 =
256 1.07 0.58 = 4.62 =
1024 0.68 0.35 we 5.78 fe
4096 0.42 0.21 Ee 6.93 =

up considerably better when compared with QAM and M-PSK. Figure 10.13 illustrates the
fractional out-of-band power [see (2.125)] 7 = (1 — x/100) in dB of QFPM versus normal-
ized bandwidth W,7; where x is the percentage of the total power that lies in the band-
width W,. Tables 10.3 and 10.4 [7] present the comparable data to the column labeled 6 in
Table 10.2 for 90% and 99% power containments, that is, 6 = 1/Wo97} and 6 = 1/ Woo97)
where

wwe S(faf ox (10.68)

Je Sclpidpe) 10D
In Section 10.2 we will discuss other forms of QBOM that exploit the desirable
aspects of carrier phase continuity.

Quadrature-quadrature phase-shift-keying (Q*PSK). Ina 1987 patent [8]

and later in the open literature [9], Saha described an I-Q modulation scheme that, using a
pair of quadrature (orthogonal) carriers and a pair of quadrature (orthogonal) pulse shapes,
aimed at forming a signal set that would make more efficient spectral utilization of the
 650 Demodulation and Detection of Other Digital Modulations Chap. 10

Table 10.4 A Comparison of the Bandwidth Efficiencies of Various Modulation Schemes

Using a 99% Power Containment Bandwidth Definition (5 = 1/ Wo97;)

CPFPM QAM CPFSK N-FSK/2PSK

M QFPM CPQFPM (MFMSK) M-PSK = (h=0.05) (h = 0.5)
2 = = * 0.05 0.83 0.05
0.1 0.83 0.83 0.1 0.79 0.1
8 as =. 0.73 = = 0.15
16 0.2 0.97 0.51 0.2 0.47 0.20
32 = ae 0.32 = as 0.23
644 03 0.77 0.19 0.3 aa 0.25
128 a4 ae 0.11 aa Ss.
256 9) 1037 0.51 2 0.4 =
1024 0.43 0.32 a 0.5 =
4096 ~—0.35 0.19 i 0.6 =m

available signal dimensions than conventional two-dimensional signal sets such as QPSK.!°
The modulation scheme dubbed quadrature-quadrature phase-shift-keying (QPSK) places
two QPSK signals in quadrature to form a four-dimensional signal set that, in a given
bandwidth, ideally allows for an increase of the bit rate by a factor of two (i.e., twice the
throughput) relative to QPSK. Mathematically speaking, a Q?PSK signal takes the form
CO

sip =Re|VPY> pit — ise! + S° pat — aS (10.69)

i=—oo i=—oo

where T; = 47; is the signaling interval, p;(t) and p2(t) are unit power real pulse shapes
of duration 7; and orthogonal to one another over the interval 0 <t < 7;, and 6, ;;n =
1,2 denotes the n' data phase (constant over the n" signaling interval nT, <t < (n+
1)T;) from each of the two QPSK signals with values ranging over the set 6», = (2m +
— 1)m/4; m=0, 1, 2,3. A particular case of interest is where p)(t) = J/2cos (zt/T;) and
Prt) = J2 sin (rt /T;), whereupon (10.69) becomes (in the zero” signaling interval)

wt t
s(t) =V2P co (F) Cos (wet + 61,0) + sin (+) cos (wet + 620)|

ut t
=/P jarcos (+) COS Wet — by.9 cos (=) SiN Wet (10.70)

See Wy Cine
+az2,9 sin Tr COS Wt — b2,9 sin TT SIN @ct

with dno, bn,o; n = 1,2 representing the equivalent +1 inphase and quadrature pairs of
data bits corresponding to the data phases 6,9; n = 1, 2, that is, an,9 = J/2cos On.0. Ono =
J/2 sin 6n,0; N= 1,2. The structure of (10.70) clearly identifies this particular form of
Q’PSK as the sum of two sinusoidally pulse-shaped QPSK signals that are in quadrature
(with respect to their pulse shaping) to each other.

10. By “ QPSK” here and in the discussion that follows, we mean the more generic case that allows for arbitrary
(rather than rectangular) pulse shaping.
 Sec. 10.2 Continuous Phase Modulation (CPM) 651

It is interesting to note that Q?PSK described by (10.70) is a special case of QFPM

discussed in the previous section corresponding to N = 2. To see this, we note that for NV=
2, bc(t) and b(t) of (10.58) are binary (+1) waveforms. Then, for the minimum frequency
to guarantee orthogonality, namely, wg = 2/T;, we have

(4) — JP factocos (+) COS Wet — ac (t)b-(t) sin (=) sin vt |

(10.71)
+ /P Eo cos (+) Sin Wet + as(t)bs(t) sin (=) cos vt |

Since a-(t), ac(t)bc(t), as(t), and as(t)bs(t) are all binary waveforms, that is, rectan-
gular pulse data streams of rate 1/7; = 1/47), then letting a,(t) = a(t), ac(t)b-(t) =
b2(t), as(t) = —bi(t), and as(t)bs(t) = az(t), we obtain the desired equivalence with
(10.70). Thus, Q?PSK of (10.70) can be generated, demodulated, and detected as a QFPM
signal with N = 2 in accordance with Figures 10.9 and 10.10.

10.2 CONTINUOUS PHASE MODULATION (CPM)

Continuous phase modulation (CPM) refers to a generic class of modulations wherein the
carrier phase is maintained continuous from symbol interval to symbol interval. Tradition-
ally, this acronym has been assigned to schemes that are constant envelope and convey
their information in the form of a digital frequency modulation represented as a doubly
infinite random pulse train. Recently, new nonconstant envelope CPM schemes have been
developed (see Section 10.2.3) which convey their information in the form of combined fre-
quency and phase modulation and have certain advantages over their predecessors. Thus, to
allow for this distinction, in this book we shall assign the acronym continuous phase fre-
quency modulation (CPFM) to what has in the past been referred to as simply CPM.
By allowing the phase to remain continuous from symbol interval to symbol interval,
one introduces memory into the transmitted signal which can be exploited at the receiver by
making decisions on an observation of the signal longer than one symbol interval. Perhaps
the first recognition of this idea was in the 1971 conference paper by Pelchat, Davis, and
Luntz [10] which considered coherent detection of continuous phase binary FSK (CPFSK)
signals and showed that the high SNR error probability performance improved with increas-
ing observation interval. Since that time, there has been a myriad of publications on the
subject of coherent and noncoherent detection of CPM with consideration of both spectral
and power efficiencies. An excellent documentation of much of this work appears in a text-
book by Anderson et al. [11] that is entirely dedicated to this subject. Since the focus of our
book is on a considerably broader coverage of modulation schemes, the treatment of CPM
given here shall by comparison be quite brief. For a more extensive expose on the subject,
the reader is referred to [11] and the host of references cited therein.

10.2.1 Continuous Phase Frequency Modulation (CPFM)

In the most general case, a CPFM signal is described by

s(t) =V2P cos e +27 » ajhiq(t — | inl, sts FT 0:72)

1==00
652 Demodulation and Detection of Other Digital Modulations Chap. 10

where a, denotes the n'® transmitted data symbol which takes on values from the M-ary
alphabet {A, =-M+(2k+1); k=0,1,2,...,M—1}, hn is the modulation index,
which may vary from symbol interval to symbol interval, and q(t) is the phase pulse, which
has length LT, (L integer) and is expressed as the integral of the normalized frequency pulse
g(t) by
t LTs 1 woes
q(t) =i) g(t)dt; q(LTs) =} g(dt= 5 (10.73)
0 0
CPEM schemes characterized by (10.72) fall into several categories depending on the vari-
ation of h and the shape and duration of g(t). Typical shapes for g(t) are rectangular and
raised cosine which lead, respectively, to the modulation types LREC and LRC. Multi-h
modulations refer to the class of CPFM’s where h,, varies from interval to interval (typically
the variation is cyclic through a set of indices). Full response CPFM is one where the fre-
quency pulse duration is one symbol, that is, L = 1. Otherwise the modulation is referred to
as partial response CPFM. Continuous phase FSK (CPFSK) is the acronym associated with
a full response CPM having a rectangular frequency pulse, that is, 1REC. For our purposes
in this chapter, we shall consider only the case where /; is independent of i, that is, the mod-
ulation index is constant over the duration of the transmitted signal, and also L = 1. For this
case, (10.72) reduces to
n—1
S(t) =v 2P.cos ee+ 2mhang(t —nTs;) +h a a
1=—CO (10.74)

=vV2Pcos[wt+O6(t,@)]; nI;<t<(+1)T;

The first term in 6(t, @) represents the contribution to the total phase in the n™ signaling
interval due solely to the n'” data symbol a,. The second term represents the contribution
due to all previous data symbols a,j; i = 1, 2,..., 00 and as such is responsible for the
memory associated with this form of modulation.
Analogous to (10.74), the excess phase in the (n — 1)*' transmission interval (n —
LQTS 1 AS CIVEDY

M2
A(t, @) = 2thonig(t
—(n-1)T.) +h Day; (10.75)
i=—oo

Comparing (10.75) with (10.74), we observe that @(t, a) is continuous at the symbol transi-
tion instant t = nT; [since q(0) = 0] which is required for CPFM.
Transmitting s(t) over a constant phase AWGN channel results in the received signal
r(t) =s(t, 0) +n(t) = Acos (wet + O(t, @) + 8) +n(t) where as in previous sections n(t)
denotes the additive white Gaussian noise process.
Since, as previously mentioned, the continuity of phase from symbol interval to sym-
bol interval introduces memory into the modulation, the optimum receiver (coherent or non-
coherent) should base its decisions on an observation interval, say N,7;, commensurate with
the memory duration. In this regard, two types of optimum receivers have been considered
in the literature. One type calculates an estimate of the data sequence in the observation in-
terval and performs block-by-block (Ns-symbol) detection. The second type decides on only
one symbol (e.g., the first one) per N,-symbol observation by averaging over the remaining
Sec. 10.2 Continuous Phase Modulation (CPM) 653

Ns — 1 symbols in the sequence, thus performing symbol-by-symbol detection. In what fol-

lows, we shall present the structure and performance of both receiver types for the special
case of CPFSK modulation.

10.2.2 Continuous Phase FSK (CPFSK)

As previously mentioned, continuous phase FSK!! (CPFSK) or, as it is sometimes, called

IREC CPM, is a full response continuous phase modulation with a rectangular frequency
pulse defined by

1
ee eyae,
9 OST,
=f 10.76)
k
Bt) {2 otherwise
Thus from (10.73), the phase response pulse, q(t), is a linear function of t over its T,-sec
duration and likewise the excess phase in the n" transmission interval nTy <t<(n+1)Ty
is also linear, that is,

ha,(t — nT,
iia) = wee +h) a; (10.77)

where we now assume that the observation interval starts at t = 0.

A plot of 6(t,q@) versus t resembles a trellis diagram (see Fig. 10.14) where the
number of possible states [values of 6(t, @)] at any integer multiple of the symbol time 7,
depends on the value of h. For example, if h = 0.5 [later we shall refer to this special case as
minimum-shift-keying (MSK)], then there are four possible states (reduced modulo 277) for
O(nTp, @), namely, 0, +7/2, and z. In general, for arbitrary h, there are an infinite number
of states.

Optimum bit-by-bit detection. An important follow-on to the work of Pelchat

et al. [10] appeared in the early paper by Luntz and Osborne [12], who derived structures
for optimum bit-by-bit coherent and noncoherent detection of CPFSK. Furthermore, the
performance of these structures was determined (in the form of upper and lower bounds)
for low SNR as well as high SNR. What follows here is a brief presentation of the results
in [12].
Coherent Detection. For ideal coherent detection, the channel phase 6 is assumed
to be absolutely known and thus the received signal becomes simply r(t) = s(t) + n(t).
In what follows, it will be important to distinguish between signals s(t) observed in the
interval 0 < t < N,T7T, whose first bit (ao) is a +1 and those whose first bit is a —1. To
this end, we replace s(t) with the notation s (t, a, a’) where a’ denotes the data sequence
(a1, @2,...,@N,—1), that is, the sequence @ excluding the first data bit. Thus, a signal trans-
mitted in the interval 0 < t < N,7) corresponding to a data sequence whose first bit is a +1
would be denoted by s(t, 1, w’) and likewise a signal transmitted in the same interval corre-
sponding to a data sequence whose first bit is a —1 would be denoted by s(t, —1, a’). Repre-
senting r(t) in its Karhunen-Loéve expansion with coefficient vector r, then in terms of the

11. Unless otherwise noted, we shall consider only the case of binary FSK and thus the terminology “symbol”
is replaced by “bit.” Also, the generalization of the results in this section to continuous phase M-FSK is given by
Schonhoff in [13].
654 Demodulation and Detection of Other Digital Modulations Chap. 10

Figure 10.14 Excess phase trellis diagram for CPFSK

above, the conditional probability of r given Ch ao, a’), denoted simply by ce |, a’ ) is

|
is (r |, a ) = Tee exp {7 [
Lope
[r(t) — s(t, a0, @ )} at

(10.78)
9) NsTp
= C exp (= | r(t)s(t, ao, @’) ar)

where C is a constant that is independent of s(t, QQ, a’). Averaging (10.78) over the pdf of
a’, namely

5c No=1
fla) = I] f(@)= I] Faces b+ 58a +0]
1 I
(10.79)
i=1
Sec. 10.2 Continuous Phase Modulation (CPM) 655

gives the conditional probability

A) NsTp
f(r lao) =| [- i!C exp Ga r(t)s(t, ao, a!) ar)f(a’) da’
@ a2 =any-y
(10.80)
1 ie 2 NsTo
eee C exp (=f r(t)s(t, (do,ai)ar)
INs—1 = No 0

where a’; i =1,2,...,2'~! represents the set of possible sequences for @’. A decision
between a = | and ag = —1 is then made by comparing the likelihood ratio
QNs-l ; A QNs-l
ae > exp (# Jo r()s(t, 1, a) de) my, X1,i

ae ee Se Ot
oy 1 2Nse! aNss!
ES ae X exp (i Ost, —1,a') dt) Dees
i= i=l
with unity, that is, if A > 1 then decide ap = 1, otherwise decide ag = —1. Figure 10.15
is an illustration of a receiver that implements this decision rule. Unfortunately, the exact
error probability performance of this receiver cannot be evaluated analytically. However, it
is possible to approximate and bound this performance at low and high SNR.
At low SNR, we may use the approximation e* = | + x, in which case the decision
rule dictated by (10.81) and Fig. 10.15 becomes
QNs-!
2 NsTp
ao = sgn ss (:+ a, r(t)s(t, I.a{)dr)

2 NsTp
-)> (:+o f rids(t.~1a{) dt)
No Jo

ae ( r(t)s(t, —1, @') ar) (10.82)

NsTp
r(t)s(t, 1, @;) dt — i
NsTb
/
0 0

NsTp :
= sgn / r(t) 5 s(t, 1,a@;) — »s, s(t, —1,a;) |dt
0

NsTp
a sen|| r(t) [s(t, b 5-1
0
Here, s(t, 1) and S(t, —1) denote the average transmitted waveform given that the first bit
is a “1” or a “—1”, respectively. A block diagram implementation of (10.82) is illustrated
in Fig. 10.16, which takes the form of a conventional matched filter receiver for binary
communication over the AWGN. This implementation is referred to as an average matched
filter (AMF) receiver in [11]. Other variations of the AMF are also discussed in Chapter 8
of [11]. Using the results of Chapter 4 and recognizing the symmetry of the present system,
656 Demodulation and Detection of Other Digital Modulations Chap. 10

D
— s(t,1,@,)
No

Figure 10.15 Optimum receiver for symbol-by-symbol coherent detection of CPFSK

we immediately have that the average bit error probability performance for low SNR is
QNs-!

]
P(E) =aa 2 Poi(E) (10.83)
i}

where
NgT,
Ppi(E) = serte| r ua a, LejEG Ese ila:
J20 0
(10.84)
20No
Oo Rr
pt =
[s(t, 1) —
=
s(t,
2
—1)] dt
0
Sec. 10.2 Continuous Phase Modulation (CPM) 657

S(t,1) —S(t,-1)

r(t)

Figure 10.16 An equivalent receiver to Figure 10.15 for low SNR

Using (10.74) together with (10.77), it can be shown [13] that

2 = Note
NoTi
5
; 1 — cos 2h)
[— sinc 2h + SES
(1 + sinc 2h) (cos?%s~2
TOUUE Se Coen
rh — 1
ter) (10.85a)
2 (cos?
xh — 1)

and

Soares
; ale 1 — sinc 2h + sinzth s (cost * rh)
k=2
: (10.85b)
=
x sinOx, + te (cos Ox — cos (27hai,x + e))|}; Q,A S 5whoi,j
j=l

where aj4; k =2,3,..., N, is the k™ component of the data vector Qe.

At high SNR, we can upper union bound the error probability performance by first
finding the pairwise error probability between each output of the upper group and each
output of the lower group of x;,;’s (see Fig. 10.15) and then averaging over all of them.
Again recognizing the symmetry of the problem, we can write
QNs—1 2Ns-1

P,(E) < ~ p> 2 By (10.86)

where

P;,;= Pr {x11 < x216;|8 (t, 1, a’) }

s) NsTp
= Pr {exp(=f r(t)s(t, Ia) dr)
0/0

Z NsTp /
< exp (=f r(t)s(t, -1,a))dr)|s(t, 1, o)/ ae,

NsTp NsTp
=Pr{ f r(t)s(t, 1, a5) dt < | r(t)s(t,
—1, o;) dt |s(¢, Ia)
0 0

1 N; Ep
= 5 erfc 7; (1 — pi,;)
 658 Demodulation and Detection of Other Digital Modulations Chap. 10

with E,/No = PT)/No and

aon / /
pi, & cE
I
f s(t, —1, a) s(t, 1, a) dt

1 : Ns h (aii, = j,k)
= N, [sscos7h + ms (Aas (10.88)

mh (ai,~ — j,k) kl
x COS [em + ma: (ai. — @;,1)

A lower bound on average bit error probability for all SNR can be obtained by suppos-
ing that for each transmitted sequence the receiver needs only decide between that sequence
and its nearest neighbor, that is, the sequence most correlated with it. Then
QNs—l
1
P(E) > ot ye cue (10.89)
r=)

where

N; Ep
Pi,jmax = = erfc s (1 aia max); Pi, jmax = Max Pj,j (10.90)
Z No j

The lower bound of (10.89) and the upper bound of (10.86) asymptotically approach
each other at high SNR and thus represent the true bit error probability performance in that
region. Thus, recalling that (10.83)-(10.85) represents an approximation of the true per-
formance at low SNR, then a composite curve formed from (10.83)—(10.85) and (10.86)—
(10.88) can be looked upon as an approximation of the performance for all SNR. Fig-
ure 10.17 is an illustration of the behavior of these various bit error probability bounds and
the composite approximation for an observation interval of five bits (Ns = 5) and a modula-
tion index h = 0.715. This value of modulation index was chosen since in [10] it was shown
that it achieves the maximum value of the minimum distance between ali pairs of signals
s(t, 1,q@‘) and s(t, —1, a’), i, j =1,2,..., 16 which dominates performance at high SNR.
By distance between signals s(t, 1,q@) and s(t, —1, a’), we mean
57h le
dij = / [ste 1,ai) —s(t, -1,¢")| dt (10.91)
0

It can also be shown [12] that little or no further improvment in performance can be achieved
by extending the observation interval beyond five symbols. This behavior appears to be
typical of CPFSK systems independent of modulation index.
Noncoherent Detection. _Here we assume that the channel phase 6 is unknown and
uniformly distributed in the interval (—z, 2). Also, in contrast with the coherent detec-
tion case, we shall be making decisions here on the middle bit of the observed sequence
as explained in [12]. Thus, to simplify notation, the length of the observed sequence N,
shall be equated to 2n, + 1 and the decision will be made on the bit Qn,. If a’ now de-
. notes the transmitted sequence excluding the bit a,,, and s(t) is notationally replaced by
 Sec. 10.2 Continuous Phase Modulation (CPM) 659

eee UPPER BOUND USING

UNION BOUND
10-1R
APPROXIMATION USING
AVERAGE MATCHED
FILTER

10-2 COMPOSITE
APPROXIMATION

P(E) LOWER BOUND

ON PERFORMANCE
10-3
COHERENT PSK

10-4
OBSERVATION
INTERVAL (N,) =5 BITS

MOD INDEX (h) = 0.715

Figure 10.17 Approximations to and
10-5
Beret Oar) 82). 34) 6 67: B9" 10 bounds on the average bit error probability
E,/No, dB performance of CPFSK (Reprinted from [12]
© IEEE)

S(t, &p,,@’) and s(t, 0) by s(t, &p,, @’, 0), then analogous to (10.81), the likelihood ratio is
now

Q2ns

Xo Jone xP (aeJoe ro)s(t, 1, a, 8) dt)do

pyc
ae

aD te I exp (ae get r@)s(t, —1, a, @) dr)do

(10.92)
Q2ns

> 10 (21,1)
_ isl
a, Q2ns

Bla (z=1,1)
1=1

where
 660 Demodulation and Detection of Other Digital Modulations Chap. 10

2
, 2 (2ns+1)Tp F
LE ha af r(t)s(t, 1, @;,0) dt

2
2) (2ns+1)Tp
ae =| r(t)s(t, 1,a;, 2/2) dt
No Jo
(10.93)
2 p2nst1)Tr 2
)
21,1 = =f s(a,
r(t)—1, 0)dt
t,

9 pQns+l)Ts 2
a val r(t)s(t,
—1, o;, 2/2) dt
0 Jo
A block diagram of a receiver that implements the decision on ap, based on (10.92) is
illustrated in Fig. 10.18. As was true for the coherent receiver of Fig. 10.15, the exact error
probability performance of Fig. 10.18 cannot be evaluated analytically. Thus, once again we
resort to suitable approximations and bounds for high and low SNR.
At low SNR, the Bessel functions in (10.92) can be approximated by Jp(x) = 1+
x?/4. Making this approximation in (10.92) results in the simplified decision rule

22ns

On, = sgn ss (zi. ~ z24,) (10.94)

P=

If we now switch to a complex signal representation, that is,

SiG An,» a’, 0) af js(t, Qn,» a’, 1/2) BS S (é Qn,» a’) = A exp G [wet ot A(t, a)]) (10.95)

then (10.94) may be rewritten as

Q2n 5

On, = Sgn Ss (ane = rave :

t=1
(10.96)
2 pQnst)Tp
2e,,,1 = =f r(t)S(t, on,, a’) dt
0
In order to calculate the error probability performance of a receiver that implements
the decision rule of (10.96), it is convenient to rewrite it in a different form. Suppose
that, analogous to the coherent detection case, we define the average complex transmitted
waveform

92n s

S(t, On,) = ae Se Sana) (10.97)

i=l

_ Then
Sec. 10.2 Continuous Phase Modulation (CPM) 661

Bei
No Lan0
(t,1, 1»Y)

2
Feta
Oe ,0)

Figure 10.18 Optimum receiver for symbol-by-symbol noncoherent detection of CPFSK

662 Demodulation and Detection of Other Digital Modulations Chap. 10

2
yp het) ae 2° -\ hp One ge
=| r(t)S(t, 1) dt} — x | r(t)S(t, 1) dt
0 No Jo
2
(2ns+1)Tp
= | r(t)S(t, 1, wt) dt
No Jo

2
5) (2ns+1)Tp
=| r(t)S(t, —1,
a) dt
0 JO

22ns Q2ns
2 enstlTp ;
r(t)S(t, 1, @;) dt (10.98)
as 2uk xh
I= J=
if]

(2 ostna)a)|
(2ns+1)T,

Q2ns 72ns 2 (2ns+1)Tp

af Soon — | r(t)S(t,
—1, @;) dt
i=l j=l 0/0
iF]
y) (2ns+1)Tp : ;
x al TENS) (1,-l,a')) at

For any given sequence a; in the double summations of (10.98), the complement (negative)
of that sequence, say a; = —a@j;, also exists. Hence, it can be shown that these double
summation terms cancel each other and what is left (that is, the single summation) is the
argument of the sgn function in (10.97). Finally, we conclude that the decision rule of
(10.97) can alternately be written as (eliminating the 2/ No scaling)

2
OnvttT, (mani 2
= |= sgn
gAn, il r(t)S(t, 1) dt} — it r(t) S(t, —1) dt
0 0
(10.99)

I|> sen {|Z2I? — |Zi1°|

which can be implemented by the simple binary noncoherent AMF receiver of Fig. 10.19.
To obtain the average bit error probability performance of Fig. 10.19, we observe
‘that (10.99) is in a form that allows application of the results in [14] for the solution of
‘the general binary noncoherent problem. In particular, the bit error probability given the
transmitted sequence a; is
Sec. 10.2 Continuous Phase Modulation (CPM) 663

S$(t,1)

S(t,-1)

Figure 10.19 An equivalent receiver to Figure 10.18 for low SNR

Poi(E) = Pr {|Zil? > [Zal? fern, = 1, at}

(10.100)
1

where Q(x, y) is again the Marcum Q-function and

ij 1 S; + So —2 |p| / 5) S2 cos (O2 — ©; + ®) S;

— So
f= a eat eS Pe (10.101)
bj 20 1 — |p| V1l—lpl

with

| 2)
Se = 5 ELAN: Kaa, 2

o? Dy =atteSE
1
(iZ1-at E (Zi)?og |ae)=58
1
{lZ2—~ EZ)?2 |
1 (10.102)
p= x58 {(Z1 ~ E (Zi) (Za ~ E (Za))"}

In (10.102), the values of the various parameters are computed conditioned on the trans-
mitted middle bit, a»,, being a +1 and the remaining bits belonging to the vector'@;. The
evaluation of these parameters is carried out in [12] under the assumption that all signals
have zero phase at the beginning of the middle bit interval. The results are as follows:
664 Demodulation and Detection of Other Digital Modulations Chap. 10

NoAZT) 1 — cos?" th
o? =~?
4
41.4 (1 + sinc 2h) ||
1 —cos* zh

oe NoA2T;, 1 : 1 —cos”" rh FZ ,
{5 + sine 2m) (1+ -7**) | }+ 7 sinc |
eed , 1 — cos? 2h
(10.103)
AGT:ie
ATT |
Si ao ye ith inch + e7 jah a, At cee 2 |Ay +14 Aol?

©; =arg im+e J" sinch + eh Aa]> @2=arg{A; +1+ A}

where

= ;So (-
jah ye fines)(cos!
= th) (1+ e IThAins—l sinc h)
f=!
a (10.104)
1 Ns ia 5
=~ ; Son —jmh re
ELS nat) (cos’! rh) )((1 + € eI 7 hist! sinch)

Finally, the average error probability is found from (10.83) with N, = 2n; + 1.
As a simple example of the above, consider the case of N; = 1(n; = 0). For this case

A‘TZ A‘T2
AlE—vA |) te Z sinc’ h: Y= Us
8 8
NoA2T, NoA2Th _; 10.105
ae po’ = ae sinc h ( )

O,=—th: OQo=0: O=—7h

Substituting (10.105) into (10.101) gives

ae al
A7T; y V1 sine?) = E
5%[1 yi=sine?| (10.106)
0
0

for all sequences a@;. Combining (10.106) with (10.100) gives the well-known result [14, p.
207] for the error probability performance of noncoherent FSK with modulation index h. If
in addition the signals are orthogonal, that is, h = 1, then this result simplifies to

P,(E) = = exp (-=¢) (10.107)

which 1s identical to (5.79).

As a second example, let VN; = 3 (ns = 1) andh = 1. Then

1 Acre Atre
At=Ao=-;
I Sool Sj=
1 7 Le S} ee ean
5

2 NoA*T, 2 NoA*Ty (10.108)

2S nn aesCas ar
Sec. 10.2 Continuous Phase Modulation (CPM) 665

and

ale Eb J3
i = No [+ + (10.109)

for all sequences @;. Combining (10.109) with (10.100) results in an improved error proba-
bility performance over the case N; = 1.
For high SNR, we may use the approximation
Q2n s

D5 Loi) & Lo(Xmax) (10.110)

i=1
where Xmax is the largest of the set {x;}. With this approximation, the decision rule corre-
sponding to (10.92) becomes

Gn, = sgn {Io (Z1,max) — Jo(Z—1,max)} (10.111)

or, since Jg(x) is a monotonic function of x

Gn, = sgn {Z1,max — 2—-1,max} (10.112)

As in the coherent case, we resort to a union bound for characterizing the error proba-
bility performance of the receiver. In particular, (10.86) applies with N, = 2ns + 1 and P;, ;
now defined by

|
Pein aca
(2ns+1)Ty
2.
/ r(t)S(t, 1, @;)dt
No 0 (10.113)

|
9) (2ns+1)Tp
a, —
/ rns(t, =1,¢”;) dt
No 0
The pairwise error probability of (10.113) can be expressed in the form of (10.100)
where
. IE /
Di,j 2No

with
2ns+1 k—]
1
Pl,m ne pe exp (isLd,(ev o-n))
(10.115)
h h
x exp GF (aK.,1 _ 7) sinc (5(ax, _ aun)

where a, is the v'h bit in the sequence a; whose middle bit is a —1 and a@),m is the v" bit
in the sequence @,, whose middle bit is a +1. Note that for Ns = 1, (10.1 14) together with
(10.115) reduces to (10.106) and hence the high SNR upper bound of (10.86) agrees with
the low SNR approximation, both of which yield the exact performance.
Figure 10.20 is an illustration of the behavior of the composite performance approx-
imation made up of a combination of the low SNR performance of (10.83) together with
 666 Demodulation and Detection of Other Digital Modulations Chap. 10

10°

BINARY
ORTHOGONAL

1071

3 BIT OBSERVATION
10-2
COHERENT PSK ————>

(E)
P
10°3

10-4
5 BIT OBSERVATION

Figure 10.20 Error probability

-4 -2 0 2 4 6 8
performance of noncoherent CPFSK receiver
E,/No, dB (Reprinted from [12] © IEEE)

(10.100) and the high SNR approximate bound of (10.86) together with (10.113). The nu-
merical results are plotted for 3-bit (ns; = 1) and 5-bit (ns; = 2) observations and h = .715.
Also, plotted for the purpose of comparison is the performance of noncoherent orthogo-
nal (h = 1) FSK [Eq. (10.107)] and coherent PSK. We observe that for E,/No greater than
about 7.3 dB, noncoherent CPFSK with 5-bit observation outperforms coherent PSK. For
bit error probabilities below 10-3, the performance of noncoherent CPFSK with 3-bit ob-
servation is within | dB of that corresponding to coherent PSK.

Optimum block-by-block detection. CPFSK receivers that perform optimum

block-by-block detection derive their structure based on a maximum-likelihood estimate of
the data sequence in the observation interval. In principle, for an N;-bit observation this
entails an implementation that calculates the correlation of the received signal with each
of the possible 2s transmitted sequences and selects the largest among these correlations.
For the case of coherent detection, one further requires knowledge of the signal phase at
the beginning of each block. Approximate knowledge of this can be obtained from the
data decisions on the previous blocks; however, the resulting configuration would then be
suboptimum. In the noncoherent case, knowledge of the signal phase at the beginning of
each block is not required and thus the resulting configuration is indeed optimum from
Sec. 10.2 Continuous Phase Modulation (CPM) 667

a sequence detection standpoint. Hence, we shall consider only the noncoherent case and
furthermore show that considerable simplification of the generic implementation with 2s
correlators can be achieved. The results to be presented here for CPFSK are taken from [15]
where the more general case of full response CPM is considered.

Noncoherent Detection. Since we are now interested in detecting a sequence rather

than just one bit in the sequence, we start by rewriting the conditional probability of (10.78)
as
1 NsTp
f(rle, 6) =exp |= | ir) —s(t,a, 09P ar}
0 JO
(10.116)
2 N;Tp
= (2 Exp Gal r(t)s (46) )

where as before we have incorporated into the notation the arbitrary (uniformly distributed)
phase @ introduced by the channel. For the CPFSK signal of (10.74) together with (10.77),
the correlation in the argument of the exponential in (10.116) can be expressed as

NsTp Ns Tp
/ r(t)s (t,@,0)dt= (| r(t)s (40) cos 0
0 0
(10.117)
NsTp
_ (| r(t)s (.a,1/2)a) sin@
0

Substituting (10.74) into (10.117), we get

NsTp NsTp
if TA) SEN CORO aL — E / r(t) COS (Wet + A(t, a)) ar|cos
0 0

NsTh ne LOSELS)
= i r(t) sin (@ct + O(t,
@)) dt |sind
0

= A |B| cos (argB — @)

where the complex baseband correlation f is defined by

NsTp
seal r(t)exp[—j(@ct + O(t, a))] dt (10.119)
0

Evaluating (10.119) separately in each bit interval we get

Tb r(the
b * sf ;
Jet e—Imhoot/ Todt 2. (ends) /2Tp (tye Jeet eI hat —1)/ Th dt

0 Tb
(10.120)
Ns—2 NsTb F
4. and + il e I mhax / nitye Ite ithon,-10-Ws-D)15)/ Thge

rae, (Ns—1)Tp
Substituting (10.119) into (10.116), averaging over 6, and taking the natural logarithm gives
668 Demodulation and Detection of Other Digital Modulations Chap. 10

In f(r la) =In E iE C exp (= |B| cos (arg B -0))ao]

Paige |foxes No
(10.121)
= InCn +In n I0 ee 1)
No
—

Note that any signal phase (due to prior data bits) that is present at the start of the obser-
vation interval can be included in (added to) 6 and because of the averaging performed in
(10.121) will have no bearing on the decision rule. This affirms a previous statement that for
noncoherent detection, the structure of the optimum receiver does not require knowledge of
the signal phase at the beginning of each block.
Since In Jo(x) is a monotonic function of x, then maximizing In f(r |@) leads to the
appropriate decision rule as follows. For each finite sequence (block) @; = (@;,0, @i,1,.--,
Qj,N,—1), the optimum receiver computes
Ns-1
Bo Poa ceri nent Cie (10.122)
k=0

where
(Ns—m)Tp ‘
Tr} n= i; r(te Jet e—Jtht—mT)/ To gy

5 (N,—1—m)Tp
(10.123)
(Ns—m)T), ;
Css = if r(the Jeet eI ™ht—mT5)/To dy. m=0,1,..., Nei
(Ns—1—m)T,
and the complex constants C;,, are defined recursively by

Cio=1, Cin = Ci pre J; =k = 1,2,...N,—-1 (10.124)

A decision as to which sequence was indeed transmitted is then made in accordance with
max |8;|. An implemention (in complex form) of this decision rule is illustrated in Fig.
l

10.21. We note that only two correlators are required, each of which is followed by a tapped
delay line and suitable processing to make the sequence decision.
To evaluate the performance of the receiver in Fig. 10.21, we make use of the tech-
nique used in the discussion of multiple symbol detection of MPSK in Chapters 6 and 7
to obtain an upper bound on the bit error probability, P,(£). In particular, we use a union
bound analogous to that used for upper bounding the performance of error correction coded
systems. This bound is expressed as the sum of the pairwise error probabilities associated
with each N;-bit error sequence. Exact evaluation of the pairwise error probabilities is again
made possible by the results in [14] analogous to (10.100)-(10.102).
Mathematically speaking, let @ = (ao, a1, ..., @y,—1) denote the sequence of N, in-
ot bits and @ = (a, @, ..., &y,—1) be the corresponding sequence of detected bits.
Then

P,(E) < _ >>) we@, @) Pr {181 > '6i]«| (10.125)

ata

12. The result in (10.125) is a special case of (4.88) corresponding to M = 24s,

 ayndurop
T=2N:
of iS);
V0 51-4-'N'
41

4)
yl
a
(1)4 aouenbas yore I0j
"ny= 0't)
6 2699
e+ (TO'N'gg

*/(41-"N)
3urpuodsai09
0}
12s00yD
fo|xeu

Pol- "114 I—*N)9 yal

2 a

ens [ZO UY UONeJUSWIE]dUI

Jo& pooyTey!]-wnuIxew
sdUoNbes JOJLUIS9
JOJ JUdIOYOOUOU
uoNoa}ap
JO YS4d\paseg
UO Ue joquiAs-SAy uoneasasgo payutiday)
wos O[S]1] (Adal

669
670 Demodulation and Detection of Other Digital Modulations Chap. 10

where w(@,@) denotes the Hamming distance between @ and @ and Pr ||B| > || a|
denotes the pairwise probability that & is incorrectly chosen when indeed @ was sent.
The decision statistic || is defined in (10.120) and the corresponding error statistic |B|
is identical to (10.120) with each a, replaced by a. Unlike M-PSK, CPFSK is not a
symmetric signaling set and thus it does satisfy a uniform error probability (UEP) criterion
where the probability of error is independent of which input phase sequence @ is chosen as
the correct sequence. Hence, without any additional characterization of Pr ||B| > |Blal,

further simplification of (10.125) is not possible. Later we shall show that Pr |B| > |B |a|
depends only on the error sequence ¢ = @ — @ and thus further simplification is possible.
To compute Pr ||BI > || al,we use the approach taken in [14] for evaluating the
probability that one Rician random variable exceeds another. In particular, analogous to
(10.100)
1
Pr{ IAI > IBl|@} == [1- ovo, va) + Ova, Vb)| (10.126)
where

b| _= Fo |v.+ niane
2} | (10.127)
with

5 =c(@,@) + js(@, @) (10.128)

In (10.128), c(@, @) and s(@, &) are defined analogous to [13] by

NsTp
c(a,@) = x | cos [A(t,@) — A(t,
&)|dt,
vse
1 NsTp
(10.129)
s(@,a@) = al sin [A(t, a) — O(t, @)| dt
Ty Jo
Reinterpreting c(@, @) and s(@, @) as sums of N; integrals each corresponding to a single
data bit interval 7, then using (10.77) in (10.129) and substituting the result into (10.128),
we obtain after some simplification
n—l
N;-1 Th jah aj;—Q;
LEO, (=f ReuCR OE 2, (ara)
sai
n=O Tp 0
(10.130)
n—1
Nel jah Ss (aj—a;)
Ith A
— we el 5 (Onan) sinc (5(ap 23 an) )e i=0

n=0 2

Before proceeding to the evaluation of (10.126)-(10.130) for specific cases, we first present
an expression for the pairwise error probability under the conditions a, b >> 1 which from
(10.127) corresponds to the case of large E,/No. Using the appropriate asymptotic expres-
sions for the Marcum Q-function and the complementary error function, (10.126) can be
Sec. 10.2 Continuous Phase Modulation (CPM) 671

written as

P No lope |
P{iAl> laa] = ( a )se |set} (10.131)

Casel. Ns; =2, h arbitrary

Since in evaluating (10.127) or (10.131) we are interested in the magnitude of 6 (rather
than 4 itself), there are only four cases of interest. A tabulation of these evaluations is given
below:

‘ m {°|
a) — a ao
— aQ
a

(1+/1- l+e/7
ansinch

'

0 2 Ep (ae 1 ei
2
}

2 2 fe (14 v1- lcoszth sinc h|**)

2 =2 Ep
(1 pe jsinc h|?)
The first two cases in the above tabulation correspond to error sequences having a Hamming
distance w (a, &) = 1 while the third and fourth cases have w (a, &) = 2. We note further
that the values of a and b for cases 1 and 2 are identical. Finally, computing the number
of times that each entry in the above table occurs corresponding to the 4 x 3 = 12 possible
cases for which a # @, then from (10.125)-(10.127) and the above tabulation we arrive at
the desired upper bound on bit error probability given by
1
PoC) = Pi Ais (Parte) (10;732)

where

P= 5[1- obi. Vai)+ O(a, Jo]: i=1,3,4 (10.133)

by _ Fe rat |1 + e/™ sinch |°

ay No
(10.134)

|b3 |= Ep 2
(ityi- |cos 7th sinc) oa (1 ae
l sine h?)
a3 No
Figure 10.22 is a plot of the upper bound on P;(E) [as given by (10.132)] versus E,/No in
dB for various values of h between 0 and 1. Figure 10.23 is the associated plot which shows
the optimization of P,(E) as a function of h with E,/No in dB now as the parameter. We
note that the optimum value of h [corresponding to minimum P)(£)] is quite insensitive to
Ep/ No over the large range of values considered.
 672 Demodulation and Detection of Other Digital Modulations Chap. 10

10-2

10°3

10-4

PF
ON
(E)
BOUND
UPPER

10°5

106

Figure 10.22 Upper bound on bit error

7 probability versus E,/No with modulation
10 9 10 11 12 index as a parameter; binary modulation with
E/N. dB rectangular frequency pulses; two-symbol
jay Aa observation (Reprinted from [15] © IEEE)

Case 2. WN, =3,h arbitrary

Again recognizing that we are only interested in the magnitude of 5, we can identify
certain symmetries and arrive at a table analogous to that for the case N; = 2 which will
contain only eight distinct entries. Finally, computing the number of times that each of the
eight entries in this table occurs corresponding to the 8 x 7 = 56 possible cases for which
a #a@, then from (10.125)-(10.127) we arrive, after much simplification, at the desired
upper bound on bit error probability given by
| 1
Py ES Pit 3 (2P2 + P3 +2P4+ Ps) + ri(Po + 2P7+ Pg) (10.135)

where P;;i = 1, 2,..., 8 is given by (10.133) with

bj |_ 3Eb
fe ate (10.136)

and
 Sec. 10.2 Continuous Phase Modulation (CPM) 673

10°

E,/Ny = 6 dB

E,/Ny = 8 dB
10-2

E,/No = 10 dB
10°3

10 -4 rau

10°5

MINIMUM
OF
BOUND
UPPER
ON
(E)
F,
10°6

E,,/Ng = 14 dB ———>
Figure 10.23 Minimum of upper bound
on bit error probability versus modulation
index with E),/Npo as a parameter; binary
10°8 modulation with rectangular frequency
0.1 02 03 04 05 06 07 0.8 0.9 1.0 pulses; two-symbol observation (Reprinted
h from [15] © IEEE)

i Id

1 {1+ 2e/"" sinc h|

2 [1+ (e/7" + e/97") sinc h|
3 | + (e/7" + e- J") sinc h|
4 |2+e/"" sinch|
PS) [1 + 2/27" +e" sinc hl
6 |(1+e/27"
+ e/47*) sinc h|
7 |(1 + 2e/2"*)
sinc h|
8 |3 sinc h|

Figure 10.24 illustrates the analogous results to Fig. 10.22 where the upper bound on Pp(E)
is now computed from (10.135) together with (10.133), (10.136), and the above table.
Comparing these two figures, one can assess the improvement in performance in going from
a two-symbol to a three-symbol observation interval.
 674 Demodulation and Detection of Other Digital Modulations Chap. 10

100

1071

10-

10°3

10-4

(E)
P
ON
BOUND
UPPER

10°5

10°6
Figure 10.24 Upper bound on bit error
probability versus E,/No with modulation
5 index as a parameter; binary modulation
10° ith rectangular frequency
S10. Ae a2 46 ae tig 8 i il pulses; we three-
: : y E ny dB symbol observation (Reprinted from [15] ©
fay (0) IEEE)

To extend these results to values of N; > 3, we make use of the fact that from
(10.130), |5| is only a function of the error sequence ¢ = @ — @ rather than @ and @ indi-
vidually. Thus, from (10.126) and (10.127), Pr |\B| Ss i6||a| is likewise dependent only on
the error sequence ¢ = a — @. Hence, analogous to [16, Eq. (4.9.21)], we can write (10.125)
in the simpler form

PoE) < aD peterPr {Il > 16 (10.137)

where we have now indicated the dependencies of the Hamming weights and conditional
pairwise error probabilities on the error sequence. Since the number of terms in the double
sum of(10.125) is 2%» x (2s — 1) and the number of possible error sequences in the single
sum of (10.137) is 34° — 1, we see that we have indeed achieved a simplification. Also, since
from (10.130), 6(—e) = 6*(é), then in (10.137) we need only sum over half the number of
possible error sequences and double the result.

Case 3. N; arbitrary, h = | (orthogonal signals)

In the general case of arbitrary N;, the upper bound on bit error probability is domi-
nated by the terms that correspond to the sequences resulting in the maximum value of |6|
Sec. 10.2 Continuous Phase Modulation (CPM) 675

(minimum value of equivalent squared distance N, — |5|) provided that the other values of
|5| generated by the remaining sequences are sufficiently smaller than |5|max. For arbitrary
h, one cannot show that this always occurs. In fact, for certain values of h, the values of
|5| may indeed be bunched around |5|max. However, for h = 1, the values of 65 are all real
and integer and there is sufficient separation between them that one can evaluate the upper
bound on error probability based on the above approximation. Thus, we proceed with an
analysis of the behavior for this special case.
For h = 1 (orthogonal CPFSK), it is easy to show that the maximum value of |4| oc-
curs for each error sequence & having Ns — 1 elements equal to the correct sequence @ and
one element in error. From (10.130), |6|max is easily shown to have value N, — 1. Since the
element in error can occur in any one of the N; positions of the sequence, then the total
number of pairs of sequences yielding |5|max is Ns x 2s. For each of these sequence pairs,
the Hamming distance is w (a, @) = 1. Thus, from (10.125)-(10.127), an approximate up-
per bound on the performance of orthogonal CPFSK with N,-symbol observation (N, > 2)
is given by

Po(E) <Pr |1B| > (Bllar}= [1— O6/bmaxs Vaimax) + Q(/aimax> VPmax)| (10.138)
where

b max =eEp E +,/N2— bax =oEy [s+ V2N5—1] (10.139)

Omax

For large SNR, we can use (10.131) instead of (10.126), which gives the asymptotic upper
bound

IN; +|dlmax \ 1 Ep
Pp( bE)
) (—————
Dine ] -—erf
5° c ——(N,
No: Ss — ||6 lmax)

(10.140)
2N;-—1)\ 1 Ep
= ; — erf —}; Ns >2
(4) Sie (a :
As Ns — oo, (10.140) reduces to

1
PRE) 5 erfc si (10.141)
2No

which is identical to the performance of orthogonal FSK with coherent detection (see Chap-
ter 4).

10.2.3 Minimum-Shift-Keying (MSK)

The special case of CPFSK corresponding to a modulation index h = .5, or equivalently a

frequency spacing Af = 1/27), is referred to as minimum-shift-keying (MSK). The reason
for this acronym stems from the fact that this value of Af is the minimum frequency spacing
that allows for coherent detection of orthogonal signals. As mentioned in the introduction
to this chapter, certain forms of CPM also have an I-Q representation. MSK is one such case
that is partially responsible for the popularity of this form of modulation. To demonstrate
this alternate representation, we proceed as follows.
676 Demodulation and Detection of Other Digital Modulations Chap. 10

Substituting = .5 in (10.77), the excess phase for MSK in the n™ transmission

interval nT, < t < (n+ 1)T, becomes

t—nT
a(t, a) = SR th eetSa =a (=)
mt
+ Xp (10.142)
2Tp

where x, is a phase constant required to keep the phase continuous at the transition points
t =nT» andt = (n + 1)T7p. In the previous interval, (n — 1)Tp, < t < nT), the excess phase is
given by

aa
Tanae
(t — (nHamma
— 1)Tp)
Bega
— mt
als ais +Xn—-1 (10.143)

For phase continuity at t = nT, we require that

1 1
Qn (= (nTp) + Xn = Qn-1 (=) (nT) + Xn-1 (10.144)

or
wn
Xn =Xn-1 + in — An) (10.145)

Equation (10.145) is a recursive relation for x, in terms of the data in the n'® and the n — 1°
intervals.
We observe that (@,—1; — a@,)/2 takes on values 0, +1. Thus, for a, = a,_1, we have
Xn = Xn—1 while for a, = —dy,_1, we have x, = x,—-1 + nz. If we arbitrarily set x9 equal to
zero, then we see that x, takes on values of 0 or z (when reduced modulo 277). Note from
(10.142) that x, represents the y-intercept (reduced modulo 277) of the line representing
O(t,@) in the interval nT, <t < (n + 1)Tp. This is easily verified by letting h = .5 in the
trellis diagram of Fig. 10.14.
Let us now apply trigonometric identities to (10.74) with @(t,@) defined as in
(10.142). In particular, we rewrite (10.74) as

S(t) = V2P [cos wet cos O(t, @) — sin wet sin O(t, a)] (10.146)

Now
wt se cule
cos O(t, a) = cos oT, COS Xn — Gy Sin oT, SiN Xp (10.147)

Recalling that x, only takes on values 0 or 2(modulo 271), then sinx, = 0 and (10.147)
simplifies to
t
cos O(t, @) = dy COS ae =O} = Gell (10.148)
Qi,

Similarly

; : wt RELA A TUUAS .
sin 6(t, @) = sin (ane ae ‘|= My, Sin oT, COS Xn + COS aT SIN Xp

(10.149)
A URS
= Pa sin 57 by = @y COS An = Ondn = =k
Sec. 10.2 Continuous Phase Modulation (CPM) 677

Finally then

S(t) = V2P [anC(t) cos wet — by S(t) sin@et]; nTy <t <(n+1)T,

An =COSXn; by =SiINX, =Andn => effective I&Q binary data sequences (10.150)

wt 5 CULE
C(t) =cos ——; S(t)=sin— = effective 1&Q waveshapes
Dilip dip

To relate this representation back to the notion of a frequency modulation, we observe

that

C(t) cos wet

eS ae oS SF OT= t wee
Fee se ze
OT
= (0) — — — —

(10.151)
S(t) sin wet =--— cos aF a || Se ze
oc ae) OTR Bee BN BITS he
QO, —— _ —- —

ons (olla)
](85%)[le-2)
Thus

panes (an = 1) > 5(2) = Aay cos |(0.+ 5) (10.152)

Me lip

1
ay = OaOn = —1) => S(t) = Aa, cos |
(o- =| |

Note that in (10.150), C(t) is S(t) shifted by 7, seconds. Thus, it appears that s(t)
might be in the form of offset QPSK with half sinusoidal pulse shaping. To justify that
this is indeed the case, we must examine more carefully the effective data sequences {ay}
and {b,}. Since the actual data sequence a, can change every 7) seconds, it might appear
that a, and b, can also change every 7, seconds. To the contrary, it can be shown that as
a result of the phase continuity constraint [Eq. (10.145)], an = cos x, can change only at
the zero crossings of C(t) and by = a, cos x, can change only at the zero crossings of S(t).
Since the zero crossings of C(t) and S(t) are each spaced T; = 27) apart, then a, and by are
constant over 27p-sec intervals (see Fig. 10.25 for an illustrative example). Noting that C(t)
and S(t) alternate in sign every 27) sec, we can strengthen the analogy with sinusoidally
pulse-shaped offset QPSK by defining the basic pulse shape

j&)
singe:
2Tp’
OSSey f= 27)
PE 10.153
pw) |0; : otherwise )

and then rewriting s(t) of (10.150) in the form

s(t) = V2P [d-(t) cos wet — ds(t) sin at] (10.154)

where

de(t) =) \cnp(t — (2n—1)Ts), ds(t)= )dnp(t— 2nTp) (10.155)

with
678 Demodulation and Detection of Other Digital Modulations Chap. 10

x, (mod 27) a time interval

0 O0<t<T,
1 T, Sts 21,
1 | — PYhe SEES BIW.
0 3T, <t $47,
— 0 4T, st SST,
0 =
ee
el
oe
oe
el
Ln 5T, <t $67,
0 -1 67, <t< 77,
1 -1 -1 TBIPSES ESOL:
FWY
UN
nA
OS
NY
oN 4 -1 1 87, <¢ <9T7,

Figure 10.25 An example of the equivalent I&Q data sequences represented as rectangular pulse streams

Cn = (—1)" Grete Ons (—1)"boy (10.156)

To complete the analogy, we point out that the sequences {a2,—} and {b2,} are the odd/even
split of a sequence {v,} which is the differentially encoded version of {a}. This is easily
verified (see Fig. 10.26 for an illustrative example) by recalling that for differential encod-
INZ, Vp = An Vp_— and then noting from (10.150) that b, = a@,a,. The alternate representation
of MSK as in (10.152) can also be expressed in terms of the differentially encoded bits, vy.
In particular
Sec. 10.2 Continuous Phase Modulation (CPM) 679

Figure 10.26 Equivalence between differentially encoded input bits and effective 1&Q bits

Forn odd

s(t) =V2P (=) COS (xay a |

+(25)
5 c ( - oT,
os || We = )]| 10.157a)
(10.157a

For n even
*

s(52)=[-8)])
Un—1 + Un ry,

mom
‘an VP ( 2 Joo (o. a ie,(
Un—1 — Un IU
680 Demodulation and Detection of Other Digital Modulations Chap. 10

MSK
Modulator

Yc, At— (2n-1)T,)

1)" cos@,t

7) a-1
Pulse
Shaping
(p(2)
Data
Splitter
Pulse
Shaping
(p(t)

(=1)" sin@,t

> 4,A(t— 2nT,)

Figure 10.27 1-Q implementation of MSK

Combining these two results we get

Un + Un=1 IU
S(t) =a LID) (=) COs |
(x=F =) |

(10.158)
tee Wa (252) COs |
(vx— ) |}SWAT SSE IMG
b
Finally, the I-Q implementation of MSK is as shown in Fig. 10.27. We observe that this
figure resembles a transmitter for offset QPSK except that here the pulse shaping is half-
sinusoidal [see (10.153)] rather than rectangular and in addition a differential encoder is
applied to the input data sequence prior to pulse shaping and modulation onto the quadrature
carriers.

Coherent detection. Based on the above I-Q representation of MSK, an opti-

mum receiver would consist of an I-Q coherent receiver for offset QPSK with I&D filters
replaced by matched (to the half-sinusoidal pulse shape) filters and followed by a differen-
tial decoder (see Fig. 10.28). Because of the differential encoding/decoding operation, the
bit error probability of the combination of Figures 10.27 and 10.28 has a penalty of approx-
imately a factor of two [see (4.202) for the exact result] relative to the same performance of
 Sec. 10.2 Continuous Phase Modulation (CPM) 681

ee
(2n-1)T, (e)dt

Carrier
Recovery Data
r(t) Diff.
Circuit Combiner Decoder

Figure 10.28 An I-Q receiver implementation of MSK

Unity Transmission

S(t) . MSK
Modulato

Figure 10.29 Two equivalent MSK transmitters

offset QPSK (or QPSK or BPSK). Thus, the often quoted statement that “optimal coherent
detection of MSK and optimal coherent detection of BPSK yield the same bit error probabil-
ity performance” is, strictly speaking, incorrect. Again we remind the reader that the reason
behind this inaccuracy is that MSK is defined as a frequency modulation as per (10.74) to-
gether with (10.142) and its equivalent I-Q representation requires a differential encoder at
the transmitter and hence a differential decoder at the receiver.
It is possible, however, to modify MSK so that the above statement becomes valid.
Since a differential encoder followed by a differential decoder is equivalent to a unity trans-
mission device, that is, input = output, then preceding the MSK modulator with this combi-
nation (see Fig. 10.29) still produces s(t) as an output. Comparing Figures 10.27 and 10.29,
we then have the equivalence illustrated in Fig. 10.30, namely, that preceding an MSK mod-
ulator with a differential decoder has an I-Q representation identically equivalent to that of
offset QPSK with, of course, half-sinusoidal pulse shaping. The modulation implemented in
Fig. 10.30 is referred to as precoded MSK and as such its receiver requires no differential
decoder following the data detection. /t is this modulation that has the exact same bit error
probability performance as BPSK, QPSK, or offset QPSK. While all of this may seem like a
moot point, it is nonetheless important to point out the discrepancy. In this text, we shall be
careful to make the distinction between MSK, and precoded MSK both of which have the
identical power spectral density given in Chapter 2.
682 Demodulation and Detection of Other Digital Modulations Chap. 10

Differential MSK
Decoder Modulator

Yc, At-(2n-1T)
(-1)" COs @.t

2) n-1
Pulse
Shaping
(p(t))
Data
Splitter
Pulse
Shaping
(p(t))

(-])" sin @,t

» 4,At- 2n7,)

Figure 10.30 I-Q implementation of precoded MSK

Noncoherent detection. Since MSK is nothing more than a special case of

CPFSK corresponding to h = .5, then the noncoherent techniques discussed in Section
10.2.2 also apply here with, however, some simplification in implementation. For exam-
ple, consider the simplification of the optimum block-by-block noncoherent receiver of
Fig. 10.21 when the transmitted signal is MSK and for example, VN; = 2. Under these condi-
tions, the complex weights [see (10.124)] needed to compute the decision statistic from the
matched filter (and delayed matched filter) outputs are:

Gio mle wn! aah = la (10.159)

which leads to the simplified complex receiver of Fig. 10.31.

A real implementation of Fig. 10.31 in the form of an I-Q structure can be obtained
as follows. Letting w», = 2 /2T» denote the radian frequency of the equivalent I and Q half-
sinusoidal baseband pulse shaping, then the four values of 6; can be expressed as:
 ul
a
ss00yD
1

(4s)
(1)4
3utpuodsai09
0}
[xews
Id

rot-? -? uf 172(%-1) fe

eansig[¢E-O, uy uonejuswedui
Jok Yoo]q 10j99}ap
JO} Wasayoouou
YS paseg
UO& [oquiks-om}
uonearasqo paquday)
wos [$1]© aI (Aa

683
684 Demodulation and Detection of Other Digital Modulations Chap. 10

‘.r(tye Je! elm! dt +if ae r(the J” Wm (t—Th)dt

,| Pm
ict es
By =i}
0 Tb
Th ; ; 2Tp ; ‘wm(t—Ts)
Bo = r(tye
108 osm! dt + if r(tyer tee 1OON a
: : (10.160)
Tp d ; 2Tp . ‘ tT)
B3 =) r(the Jere Jom dt — afl r(the
12 elom@—Th) Gy
0 Tp

Tp 27, gon
Ba = (| r(tye
[ete Jom dr — if rite
eg a dt
0 Ty

Taking the squared!? magnitudes of the four complex quantities in (10.152) gives terms of
the form:

|Bil? = (arRs+ aaIe +.a3Rer + aalsr)” + (a5Re + 461s +a7Rst t+agler)? (10.161)

where
2T, 2T)
Rs =I) P(t) COS @ct SIN@m (t — Th) dt, Re = / r(£) COS wel COS Wm (t — Ty) dt
Tp Th

2T, 2Tp
Noe / r(t) sin @ct sin@p, (t — Tp) dt, I-= / r(t) SiN Wet COS @m (t — Tp) at
3 ci (10.162)
Tp Th
Rsr =|! r(t) COS Wet SiIN@mtdt, Rer =! r(t) COS @ct COS Wmt dt
0 0
Tp Th
[37 a r(t)Sin@ct sin@,t dt, Icr =i] r(t) Sin @ct COS Wmt dt
0 0
and the a;’s take on values of +1 depending on the value of i. For example

[Bil? = (Rer + Isr — Ry + Ic)? + (Rev — Ter + Re + Is) (10.163)

Figure 10.32 illustrates a portion of the I-Q receiver implementation, namely that which
generates |, |*. The remainder of the receiver that generates the other three |; |2’s makes
use of the same front end structure (the part enclosed in a box in Fig. 10.32) which is, in
fact, identical to that found at the front end of a coherent receiver of MSK.
The bit error probability performance of the receiver in Fig. 10.32 can be obtained
from (10.132) simply by substituting the value A = .5 in (10.130). In particular, we get

di are ie 3 —4/x? : bs _ ro 2|.

a| No 4 d a3 No 0}?
(10.164)
ba _==Fs (13 Fae
{ot 1 Ds\ s
an)

13. Note that a decision based on the maximum |8;| is equivalent to one based on the maximum |Bi\*.
 0}
UOISIDap
s01Aap

ensiq EOL ay Q-] uoNRUUIa|

Joev duIT
Yo[q 10}99}9p
Joy IWessayoou
ASWou paseq UO
joquiks-omi uoneasssqo pajutday)
wos [$1]© (AAAI

685
686 Demodulation and Detection of Other Digital Modulations Chap. 10

Since
p*

Q(a,0)=1; Q(0, B) =exp (-5) (10.165)

we get the simplified result

1
PoE) <5 [1= O(/b1, fai) + OL/ai: vb]
E; (10.166 )
+5[1- O(vbs, Jaa) + (Vas, Vou)|+ 7 exp (-=)
1

where a), a4 and bj, bq are as defined in (10.164).

If the large SNR approximation of (10.131) is applied to the first two terms of
(10.166), then we get

2+ /71+4/n?1 =. (
Ppl) Soh terpenes
3 2/1+4/n2 2 (2No
(10.167)

=F
[2+4/x
8/7
1 | E, L, Gee,
jet si0~4/x)) + exo =)

What we see from (10.167) is that the bit error probability performance (at least in terms of
its asymptotic upper bound) of noncoherent MSK with two symbol observation and block-
by-block detection is a scaled mixture of the performance of DPSK and coherent PSK’s
with effective E/No’s determined from the arguments of the erfc functions. Figure 10.33
illustrates the upper bound of (10.167) as well as the accompanying results for N, > 2
as determined from (10.137). From this figure, one can see the dramatic improvement in
performance obtainable with moderate values of Ns.

10.2.4 Sinusoidal Frequency-Shift-Keying (SFSK)

As mentioned at the beginning of Section 10.2.1, another common pulse shape for full
response CPM is the raised cosine (RC) pulse, which is a special case of LRC corresponding
to L = 1. In particular, the frequency pulse g(t) has the form

| a Qnt\.
et) =4 EF (1 COS Sr i Ostet (10.168)
0; otherwise

For h = .5, this modulation scheme has independently been referred to as sinusoidal
frequency-shift-keying by Amoroso [17]. The primary advantage of this form of CPM over,
say, MSK is its increased rate of spectral roll-off as discussed in Chapter 2. It also possesses
an I-Q representation in the form of offset QPSK with I and Q pulse shapes as in (2.122).
As such, coherent and noncoherent detection techniques are quite similar to those for MSK
and will not be discussed here in detail. As an example, for noncoherent block-by-block de-
tection, the upper bounds on bit error probability given by (10.132) [together with (10.133)]
 Sec. 10.2 Continuous Phase Modulation (CPM) 687

10-1f

10-2

10°3

10-4

RP
UPPER
ON
BOUND
(E)
10°5

10°6

Figure 10.33 Upper bound on bit error

probability versus E;,/No with the number
7 of observation symbols, N,, as a parameter;
10 h =0.5, binary modulation with rectangular
frequency pulses (Reprinted from [15]
© IEEE)

for Ns = 2, (10.135) [together with (10.136)] for N, = 3, and (10.137) for arbitrary N, still
apply with 6 now given by [15]
n—\

Ns~1 i% Dd (i-4)
Ri Sh E,e i=0 (10.169)
n=0

where
Ie h , h i h me
&j = 5 Deedat (5(a; = x) c (5(a; = a) + 2)=H A (5(a; fae ai) — 2)

oe< x (h (a - 4i)) (10.170)

|ex cs
688 Demodulation and Detection of Other Digital Modulations Chap. 10

and

A(x) = e/™* sincx (10.171)

Also in (10.170), ex is again the Neumann factor, that is, e, = 1 ifk =O and e, = 2 ink 505
and Jx(x) is again the k"* order Bessel function of the first kind with argument x. As such
the values of a and 3, that is, the arguments of the Marcum-Q function, given previously for
CPFSK still apply to SFSK if sinc h is replaced by sinch x f(h) where

= h so h(2k — 1)
fies Dderau(h) lean |—2 2 Jnx—1(h) eed (10.172)

and then evaluated ath = .5.

10.2.5 Continuous Phase FPM and QFPM

Continuous phase QF PM (CPQFPM) [7] is a modification of the QFPM scheme discussed

in Section 10.1.3 that allows for sinusoidal pulse shaping and an offset of 7; /2 in the relative
alignment of the baseband modulating waveforms for the two quadrature components. Anal-
ogous to the I-Q representation of MSK as an offset QPSK modulation with sinusoidal pulse
shaping, the above modification of QFPM produces a continuous phase modulation. Contin-
uous phase FPM (CPFPM) or alternately Multifrequency Minimum-Shift-Keying (MFMSK)
[18] is a special case of CPQFPM wherein the frequencies of the two quadrature compo-
nents are constrained to be identical, thereby producing a constant envelope signal set. In
both schemes, the minimum frequency spacing required for orthogonality corresponds to
selecting f= 2 (f/g1, = 1).
A (2N)?-ary CPQFPM signal is represented as

s(t) = V2P {actocos (= cos [(@e + be (t)@g) t]

s

€ mt
+a;(t) sin (=)sin [(@e + bs(t)wa) n} (10.173)
AY

= Sc(t) + Ss(t)

where the even- and odd-numbered data waveforms d,(t) and d,(t) that generate the
amplitude-frequency pairs a;(t), b-(t) and as(t), bs(t), respectively, are offset by 7;/2 in
their relative alignments.'* This staggered alignment between a;,(t) and a;(t), and simi-
larly between b,(t) and b;(t) together with the sinusoidal pulse shaping by cos (xt/T;)
and sin (7t/T;) results in the phase of s(t) at data transition instants t = (2i — 1)T, /2 and
t = iT; being continuous. To see this, we rewrite (10.173) as follows:

14. The data streams a,(t) and b,(t) have data transition instants at t = (2i — 1)T, /2 and the data streams a, (t)
and b;(t) have data transition instants at t = iT,.
Sec. 10.2 Continuous Phase Modulation (CPM) 689

IP
s(t) = [eat {cos|
(«.+ be(t)@g + =)|+ cos |
(x+ b-(t)@g — x)|

aie as(t) cos |

(«+ b;(thag — x)|
ac(t) 5

as(t) IU
Brass |
(o.+ bs (thog + =)|
(10.174)
= V2Pa¢(t) {cos(OO out)

bs(t) + be(t
xX COS |
(2Ar OO 4 a =)|

: bs a be < S Cc
+ sin eo
5 Put sin |
(e+ bs(t) :belt) @g + x)|}

where the upper signs correspond to a,(t)/ac¢(t) = 1 and the lower signs correspond to
as(t)/ac(t) = —1. Evaluating (10.174) at t= (2i — 1)T,/2, we get

s G = Dz) = +(-1)'"!/2Pa, G = ey)

x sin {[o.+ bs (a - ey) va (ca- v)|

(10.175)
= (-1)'"!/2Pa, (ci- D2)

x sin {[o.+ bs (2 _ 02) va (2 = 03)

Since a,;(t) and b,(t) cannot make a transition at t = (2i — 1)7;/2, then s(t) will be contin-
uous at these time instants. Similarly, at t = i7;, (10.174) evaluates to

s Ts) = (—1)'V2Pae (iTs) cos {le + be (iTs) Wal iTs} (10.176)

Since a,(t) and b,(t) cannot make a transition at t = i7;, then s(t) will be continuous at
these time instants.
The special case of (10.173) corresponding to N = 1, that is, b.(t) = bs(t) = 0, yields

SW) = J2P factcos (+) COS Wet + as(t) sin (=) sin vat| (10.177)
AY AY

which is in the form of an offset QPSK signal with sinusoidal pulse shaping, or equivalently,
precoded MSK.
A block diagram of a modulator for (2N )?-ary CPQFPM is obtained from the block
diagram of a (2N )?-ary QFPM modulator (Fig. 10.9a) by introducing a 7; /2 offset between
the alignments of d,(t) and d,(t) and adding sinusoidal pulse shaping to the s,(t) and s,(t)
signals prior to their summation. The result is illustrated in Fig. 10.34a. Similarly, the block
690 Demodulation and Detection of Other Digital Modulations Chap. 10

Demultiplexer

ane!
@, =@,-[N-(2n-])]@,; n=iorj sin

Figure 10.34a A block diagram of a modulator for (2N )?-ary CPQFPM

(|
Bile FM

Demultiplexer
sin @,t

@,= @, —[N-(2j—)]@,

Figure 10.34b A block diagram of a modulator for 4N-ary CPFPM

diagram of a modulator for 4N-ary CPFPM is the same modification of the modulator for
4N-ary N-FSK/QPSK (Fig. 10.9b). The result is illustrated in Fig. 10.34b. The optimum
receiver for CPQFPM is a modification of the optimum receiver for QFPM (Fig. 10.10)
wherein the I and Q integrate-and-dump filters of the optimum biorthogonal detector blocks
are replaced by matched filters, that is, amplitude weighting by cos (zt/T;) and sin (zt/T;)
followed by I&D’s with integration limits as well as the sampling instants staggered by T, /2
with respect to one another (see Fig. 10.35). It should be clear from Fig.10.35 that the ideal
coherent error probability performance of CPQFPM on the AWGN is identical to that of
QFPM (see Table 10.2).
Using the methods presented in Chapter 2, it is straightforward to show that the
equivalent lowpass two-sided PSD of CPQFPM is given by [7]
 Sec. 10.2 Continuous Phase Modulation (CPM) 691

oni tcoess
i eenT:
i=1,2,...,N

Optimum
Biorthogonal
Detector
Received
Signal

Optimum
Biorthogonal
Detector

sinossin=

J=1,2,...,N

Figure 10.35 Block diagram of an optimum receiver for (2)?-ary CPQFPM

Ss(f) = 5
T. alae
Sf i oe 3\72

1+ lo 1
Aesoes (1 + logy N) fT+2i-—-N+
8)PTp D[sine(2( 5) (10.178)

3\72
+ sinc (2(1+ log, N) fT) +2i-N +3)

which is illustrated in Fig. 10.36 for various values of N. The curve labeled N = 1 is
identical to the PSD of MSK. Comparing Fig. 10.36 with Fig. 10.12 we observe that,
analogous to the PSD’s of QPSK and MSK, the main lobe of the CPQFPM spectrum is
wider than that of QFPM; however, the PSD of CPQFPM falls off at a faster rate than that of
QFPM. This can also be seen by comparing the fractional out-of-band behavior of CPQFPM
(Fig. 10.37) with that of QFPM (Fig. 10.13). Finally, the power and bandwidth efficiencies
of CPQFPM are included in Tables 10.2—10.4 along with the comparable data for the other
I-Q modulations previously discussed.
As previously mentioned, CPFPM is a special case of CPQFPM wherein the I and Q
signal component frequencies are identical. With reference to (10.173), it is mathematically
modeled as a 4N-ary constant envelope signal of the form!>

15. We remind the reader that, as for N-FSK/QPSK, the number of bits per symbol, i.e., 7;/ 7p, is now logy 4N =
2 + log, N since two bits are required for the pair of binary amplitudes a,(t) and a;(t) and log, N bits to determine
the frequency common to both channels.
692 Demodulation and Detection of Other Digital Modulations Chap. 10

Figure 10.36 Power spectral density of

2.00 2.50 CPQFPM with N as a parameter (Reprinted
from [18] © IEEE)

S\(f)) = J2P {actcos (+) COS [(@e + b(t)wg) t]

Ss

+ b(t)wa) t] (10.179)
+ a;(t) sin (+) sin [(@e
AY

= Sc(t) + S5(t)

Here a,(t) and a;(t) are offset in their relative alignment by 7,/2, that is, a;(t) has transi-
tions at t = i7; and a,(t) has transitions at t = (2i — 1)T;/2. However, b(t) being common
to both I and Q signal components now has the same transitions as b,(t) of CPQFPM, that
is, att =iT;. Also, s-(t) and s;(t) are orthogonal over the interval (i7;/2, (i + 1)T,/2) of
duration T;/2 sec, which allows for demodulating and detecting them separately.
An equivalent representation to (10.179) is
 Sec. 10.2 Continuous Phase Modulation (CPM) 693

dB
n,

Figure 10.37 Fractional out-of-band power

versus normalized bandwidth W,7), for
CPQFPM with WN as a parameter (Reprinted
from [18] © IEEE)

s(t) = V2Pa,(t) cos (os+ b(t)wa)t — celal) |

(10.180)
= /2Pac(t) cos ou + (2b(t) — ac(t)as(t)) =|

which clearly elucidates the constant envelope nature of the CPFPM signal. Moreover, the
phase continuity of s(t) is easily identified since a transition in b(t) and/or a;(t) at t = iT, or
a transition in a,(t) at t = (2i — 1)T;/2 produces a phase change equal to 277 (modulo 277).
An optimum receiver for coherently demodulating and detecting CPFPM is given in
[18] and is illustrated here in Fig. 10.38. Its structure resembles the optimum receiver for
constant envelope QFPM (see Fig. 10.11) . Because of the coincident alignment of b(t) and
a;(t), that is, both transition at integer multiples of 7;, the bit stream, d;(t), corresponding
to the signal frequency is estimated only in the receiver block, Ds, associated with s,(t).
This estimation, however, is aided by the receiver block, D,, associated with s,(t) so as to
improve performance. This will be explained shortly. Thus, at the output of block Ds, as(t)
and d,(t) [or equivalently, b(t)] are recovered.
In the block D,, only the binary amplitude data stream a,(t) (or a delayed ver-
sion of it) needs to be recovered since as mentioned above no frequency decisions are
made in this block. The particular correlator in D,, that is, the ith, from which a,(t)
is estimated should be chosen in accordance with the frequency w; decided on in the
D; block. However, since the time at which this frequency estimate is made, namely,
t =(n+1)T;, occurs in the middle of the pulse a,(t)cosmt/T; associated with s,(t),
then it is possible that a frequency transition may occur at this very same time instant.
Stated another way, the first half of the pulse a,(t)cosat/T; may correspond to one
frequency while the second half of this pulse may correspond to another frequency. Be-
cause of this, the integration interval of the I&D’s associated with the correlators in block
D, must be limited to a duration of 7;/2 sec so that the appropriate correlator for the
 VisinE sin@,t

OL
(at)T,

(a+),
i, Coat

Pets Ler
x2 sin—sinayt
ty

y(t)

sample at
t=(n+ DT,

v2 cose cos @,t (+) (+)

Ci

Delay
D, T,/2

a,(t—T, /2)
a2 cos COS Wyt
’

Figure 10.38 Block diagram of the optimum receiver for 4N-ary CPFPM (Reprinted from [18]
© IEEE)

694
 Sec. 10.3 Further Discussion 695

first and second halves of the symbol a,(t) can be separately chosen in accordance with
the frequency decisions at t = (n + 1)T, and t = (n+ 2)T,, respectively. For example,
in order to detect a,(t) [the sign of s,(t)] in the interval T, /2 <t <3T,/2, the particular
correlator corresponding to I&D integration in the interval T,/2 < t < T, should be selected
by the frequency decision at t = T, whereas the particular correlator corresponding to I&D
integration in the interval T; < t < 37,/2 should be selected by the frequency decision at
t = 2T;. To effectively utilitize both of these components in making the decision on a,(t),
the first-half symbol correlator output should be delayed by 7; sec and then added to the
second-half symbol correlator output (see the lower portion of block D, in Figure 10.38).
By taking the sign of this sum, we obtain an estimate of a,(t — T,/2).
As mentioned above, an improvement in detection of the signal frequency can be
obtained by utilizing the correlator outputs corresponding to the quadrature signal compo-
nent s-(t). Since midway in the detection interval for s,(t) a transition in a,(t) may occur,
we must separately add and subtract the first- and second-half symbol correlator outputs
corresponding to s,(t) (this is accomplished in the D, block of Fig. 10.38 by the blocks
Ci;i = 1,2,..., N) to obtain pairs of signals that can be used to aid in the formation of the
block D, output magnitudes, w;;; i=1,2,...,N, j=1,2,3,4. As in Fig. 10.11, the
pair of signals from each of the blocks C;;i = 1,2,..., N is coherently combined (added
and subtracted) with the corresponding I&D output of the D, block to produce a 3 dB im-
provement in SNR relative to the case where the frequency estimate is obtained without
aiding from the D, block.
In determining upper bounds on the symbol and bit error probability performances of
the optimum 4N-ary CPFPM receiver of Fig. 10.38, we follow the same arguments as used
to arrive at upper bounds on the performance of the optimum M-ary receiver for 4N-ary N-
FSK/QPSK (see Fig. 10.11). In particular, it is straightforward to show that the symbol and
bit error probabilities of Fig. 10.38 are upper bounded by (10.64) and (10.65), respectively.
Figure 10.39 is an illustration of the upper bound of (10.65) for values of N =
1,2,4,8, and 16. We observe from these numerical results that the power efficiency of
CPFPM is a monotonically increasing function of N and that relative to MSK (N = 1) sig-
nificant power savings (on the order of 1.5 dB at Pp(E) = 10->) can be achieved with only
N =2. A more detailed numerical comparison is made in Table 10.2.

10.3 FURTHER DISCUSSION

In this chapter we have attempted to cover the demodulation and detection of a variety of
digital modulations that have achieved a certain degree of popularity by finding their way
into modern communication system design. Needless to say, there exist a host of other
digital modulations that for one reason or another have received attention in the literature.
Our purpose here is to mention some of these and briefly discuss their merits.
In our discussion of I-Q modulations we have focused entirely on full response
schemes, that is, those whose pulse duration is equal to a single symbol interval. Analo-
gous to partial response CPFM schemes, such as, LREC and LRC, it is possible to achieve
additional bandwidth compression by considering I-Q modulations whose pulse duration
extends over more than one symbol interval. An example of this is so-called Quadrature
Overlapped Raised Cosine (QORC) modulation [19] wherein the I and Q pulse shape is a
raised cosine pulse that extends over two symbol (four bit) intervals. Mathematically speak-
ing, QORC is described by
696 Demodulation and Detection of Other Digital Modulations Chap. 10

P(E)
ON
BOUND
UPPER

E,,/No, dB

Figure 10.39 Upper bound on the bit error probability performance of CPQFPM with
N as a parameter (Reprinted from [18] © IEEE)

s(th=A ( So ap 201) COS Wet — (bape on) sno (10.181)

n=—0Oo n=—oOo

where

1 Tote Ven
p(t) = (1 — cos Hi,5 Ostet (10.182)
: otherwise

Note that for QORC, the amplitude A is related to the average power of the modulation P
by A = /4P73, whereas for QPSK the relationship is A = /P and for precoded MSK,
A = V2P. While the power performance of QORC is quite comparable to that of QPSK
and precoded MSK, the real advantage of QORC lies in its spectral occupancy. Indeed it
can be shown [19] that, except for a normalization constant, the equivalent lowpass PSD
of QORC is the product of the equivalent lowpass PSD for QPSK and MSK (or precoded
MSK), that is,

sinin 27
2xffTy||” Fe DfT,
nf 4 (10.183)
Ss(f) = ZPD
2nfTp 1— (4fT,)°
Thus, asymptotically the PSD of QORC rolls off at a rate f~° as compared with that of
MSK, which rolls off as f~*, and that of QPSK, which rolls off as f~*. Offset (staggered)
versions of QORC have also been considered for application on nonlinear amplification
channels [20].
Another I-Q technique that has merit on nonlinear amplification channels is a variation
of conventional QPSK referred to as 7/4-QPSK and was initially described by Baker [21].
Sec. 10.3 Further Discussion 697

The idea here is to reduce the maximum 100% envelope fluctuations due to possible 180°
phase reversals in QPSK without resorting to an offset form of modulation. While indeed
offset QPSK reduces the envelope fluctuation by limiting carrier phase changes to 90°, its
performance suffers relative to conventional QPSK when differential detection is employed.
By limiting the maximum phase change to 135°, 2 /4-QPSK achieves a compromise be-
tween QPSK and offset QPSK in the spectral spread caused by nonlinear amplification. The
following brief description of 7/4-QPSK explains how this reduction in maximum phase
change is accomplished without offsetting the I and Q channels.
The basic difference between 2 /4-QPSK and conventional QPSK is the set of values
assigned to the phases representing the four possible two-bit information sequences. In
conventional QPSK, the information phases used are 0, +7/2, 2 whereas in 1/4-QPSK
we use +77/4, +37/4. Since with differential encoding (which is required for differential
detection) the information phases represent the phase changes between adjacent transmitted
symbols, then we see that whereas with QPSK a maximum phase change of z radians is
possible, with 7/4-QPSK the magnitude of the maximum phase change is limited to 37/4
radians. In terms of the actual phases transmitted over the channel, with QPSK the allowable
set remains the same in every transmission interval, namely, 0, +2 /2, 2 whereas with 7 /4-
QPSK it alternates between 0, +7/2, a and +7/4, +37/4.
Finally, we bring the reader’s attention to multi-amplitude minimum-shift-keying
(MAMSK), a generalization of MSK introduced in the late 1970’s [22, 23]. As the name
of the modulation indicates, MAMSK is a multilevel version of MSK in that the effec-
tive inphase and quadrature data sequences are multilevel rather than binary. Specifically,
MAMSK assumes the form of (10.154) where the data sequences c, and d, take on values
B=) esi See +./M — 1. An alternate representation of MAMSK is in the form of the sum
of n = log, (log, M )MSK signals whose powers are distributed such that the i" component
of the sum, i = 1, 2,..., has an amplitude that is 2‘! times that of the first component
[23]. For example, for M = 16, the signal

wt wt
s(t) = Acos (ou+ Ay (=) te *»)+2Acos (ou+ Bn (=) oF mn):
2T, 2Tp (10.184)
nlp <t <(n+1)Tp

is identically equivalent to

mt
S(t) = ACOS wet Cos a7, [cos Xn + 2 COS yn]
b
, eee
— Asin wet sin aT, [Qn COS Xn + 2B, COS Yn] (10.185)
b

wt . a EES
= ACy COS Wet COS rT, — Ad, Sin Wet sin oT,

where the effective inphase and quadrature data sequences cy and d, take on values'E1, +3.
The probability of error performance of precoded MAMSK is identical to that of
QASK with the same value of M in the same sense that the probability of error performance
of MSK is identical to that of PSK. The advantage of the former in both cases is, of course,
the faster spectral roll-off achieved by using I and Q half-sinusoidal pulse shaping.
698 Demodulation and Detection of Other Digital Modulations Chap. 10

Table 10.5 A Table of Various Modulation Types and Their Properties

Description and Relation to Continuous or Constant

Modulation Type Other Modulation Types Discontinuous Phase —_Envelope

Unbalanced Quadra- QPSK with, in general, different N/A Yes

ture Phase-Shift-Keying power, data rate, and data format
(UQPSK) on I&Q channels
Quadrature Amplitude Independent M-AM modulations N/A No
Modulation (QAM) of the same power, data rate, and
data format on I&Q channels
Quadrature Amplitude- Independent M-ASK modula- N/A No
Shift-Keying (QASK) tions (rectangular pulse shape
implied) of the same power, data
rate, and data format on I&Q
channels
Quadrature Biorthogonal Independent N-dimensional Discontinuous No
Modulation (QBOM) biorthogonal modulations of the
same power, data rate, and data
format on I and Q channels
Quadrature Frequency/ Particular embodiment of QBOM _ Discontinuous No
Phase Modulation using, for the biorthogonal mod-
(QFPM) ulations, N-FSK/BPSK with
minimum frequency spacing re-
quired for orthogonality
Quadrature-Quadrature Independent QPSK modulations Discontinuous No
Phase-Shift-Keying on I&Q channels (special case
(Q?PSK) of QPFM when N = 2, pulse
shaping is sinusoidal, and I&Q
channels are offset by half a
symbol)
Continuous Phase Mod- Digital frequency modulation Continuous Yes
ulation (CPM) or Con- with continuous phase from sym-
tinuous Phase Frequency bol interval to symbol interval
Modulation (CPFM)
Full Response CPFM CPFM wherein frequency pulse Continuous Yes
is one symbol in duration
Partial Response CPFM CPFM wherein frequency pulse Continuous Nes
is greater than one symbol in
duration
Continuous Phase Fre- Special case of full response Continuous Yes
quency-Shift-Keying CPFM using a rectangular fre-
(CPFSK) quency pulse
Multi-h CPFM CPFM wherein the modulation Continuous Yes
index varies from symbol interval
to symbol interval
Minimum-Shift-Keying Special case of CPFSK wherein Continuous Yes
(MSK) the modulation index equals 0.5.
Also equivalent to precoded
offset QPSK with sinusoidal
pulse shaping
$$$ Se a ee a ee eee
 Probiems 699

Table 10.5 A Table of Various Modulation Types and Their Properties, continued
ee
EE aE a ee a ee ae Se

Description and Relation to Continuous or Constant

Modulation Type Other Modulation Types Discontinuous Phase —_Envelope
i eT See i SM ee Hen ea 3
Continuous Phase QFPM Modification of QFPM wherein Continuous No
(CPQFPM) pulse shaping is sinusoidal and
I&Q channels are offset by half a
symbol
Continuous Phase FPM Special case of CPFPM wherein Continuous Yes
(CPFPM) or Multifre- frequencies of the two quadrature
quency Minimum-Shift- components are constrained to be
Keying (MFMSK) identical
Quadrature Overlapped QPSK with raised cosine pulse N/A No
Raised Cosine (QORC) shapes of two-bit duration
mz /4-Quadrature Phase- QPSK where the signal phases N/A Yes
Shift-Keying (77/4-QPSK) take on values +7/4, +37/4
(rather than 0, +77/2, 77)
Multi-amplitude Minimum- Multi-amplitude version of I-Q Continuous No
Shift-Keying (MAMSK) form of MSK

We conclude this chapter by presenting, for quick reference purposes, a table listing
the various modulations considered in this chapter together with their acronyms, relation to
other modulation types, and key properties such as discontinuous versus continuous phase
and constant versus nonconstant envelope. These results are presented in Table 10.5.

PROBLEMS
10.1 Consider a digital communication system transmitting an offset QPSK modulation. As such,
the received signal is of the form

r(t) = VP [de(t) cos Wet — ds(t) sin wet] + n(t)

where P is the received power, w, is the carrier radian frequency, n(t) is an additive white
Gaussian noise source with single-sided power spectral density, No W/Hz, and d,(t),
d,(t) are, respectively, the in-phase (I) and quadrature (Q) data streams defined by

de(t)=.)) deip(t—iTs)
i=—0o0

— 1
d;(t) — ye dsi p (:
7 (:
at ;)r.)
i=—0o

with p(t) a unit amplitude rectangular pulse of duration 7; sec. The two data streams {d.;} and
{d,;} are independent, identically distributed (iid) +1 sequences.
The carrier reference signals generated at the receiver (see Fig. P10.1) are not ideal in
that their phase differs from that of the received signal by ¢ radians. That is

r-(t) =2cos (wt —), rs(t) = —2 sin (wct — d)

(a) Find the conditional error probability, P;(E£; 6), associated with detecting the sequence
{d.i} in the presence of the phase error $. For simplicity, consider only the detection of the
data bit do and ignore any second harmonics of the carrier frequency.
700 Demodulation and Detection of Other Digital Modulations Chap. 10

Carrier Parallel
r(t) Recovery to Serial
Circuit Convert

(n+1/2)T,

Figure P10.1

(b) If the I and Q data sequences {d,;} and {ds;} were originally obtained as the odd and even
bits of an iid +1 data sequence {d;} occurring at a bit rate of 1/7, = 1/27;, what is the con-
ditional bit error probability, P,(E; ¢) |ogpsk ; of the sequence {d;}? Express your answer
in terms of the conditional bit error probability of QPSK (non-offset), P,(E; ¢) |Qpsk »
and the conditional bit error probability of PSK, P,(E; ¢) |psx, in the presence of the
same phase error, ¢. That is, write P,(E; ¢) |ogpsk as a function of P,(E; ) |apsk and
P,(E; &) Ipsk-
(c) For any ¢ £0, how does P,(E; ¢) |ogpsk compare with P,(E; ¢) |psx? Is it better or
worse? How does P,(E; ¢) |ogpsk compare with P,(E; ¢) lapsk ? Is it better or worse?
How do these conditional bit error probability performances compare at ¢ = 0?
10.2 Consider an unbalanced QPSK receiver with imperfect carrier phase synchronization. Because
of the presence of a carrier phase error, @, between the received modulation s(t, 9) of (10.3) and
the I and Q demodulation reference signals, r.(t) and r;(t) of (10.9), the I and Q integrate-and-
dump outputs will each contain a term due to crosstalk [see (10.11)]. Normalized versions of
these crosstalk terms denoted as d,2 and d2,; are defined in (10.12). Using the approach taken
in Chapter 2, Section 2.1 to derive the PSD of a random pulse train, show that the mean-squared
values of d;,2 and dz; can be expressed in terms of the PSDs of the I and Q data streams as in
(10.22).
10.3 Using (10.22), derive the mean-squared crosstalk de, for the case where d(t) is an NRZ
waveform of rate 1/7; and d2(t) is an NRZ waveform of rate 1/7.
10.4 Using (10.22), derive the mean-squared crosstalk a for the case where dj (t) is a Manchester
waveform of rate 1/R, and d2(t) is an NRZ waveform of rate 1/R2.
10.5 (Gray Code for QAM) In Chapter 4 we defined the notion of a Gray code as applied to M-
PSK modulation. In particular, a bit assignment was made to each symbol (signal point on
the circular constellation) in such a way that adjacent symbol errors resulted in only a single
bit error. Since each point on the circle has only two adjacent neighbors, when viewed in
polar (amplitude-phase) coordinates, the Gray code for M-PSK is one-dimensional, that is,
it relates only to the phase coordinate. For QAM modulation, a two-dimensional Gray code
can be assigned in such a way that for each of the two rectangular coordinates, an adjacent
symbol error results in only a single bit error. Construct such a 2-dimensional Gray code for
Problems 701

Acos(@,t + 0)

d(t)= Sd,p(e - kr)

k=-<0

PSK
Modulator

Figure P10.6

16-QAM, that is, show the appropriate bit to symbol assignments for each of the 16 points in
the constellation.
10.6 (Serial MSK) Consider the transmitter implementation illustrated in Fig. P10.6. Binary (+1)
data {dx} is PSK modulated onto a carrier Acos (wt + 6) to produce the signal x(t) =
co

A >> dkp(t —kT) cos (wt + 6) where p(t) is a unit rectangular pulse shape and @ is an
k=—oo
arbitrary phase. The PSK output is then passed through a linear filter H7(@) with finite impulse
response

Zsin@t; O<t<T
mys yi ; aie
re) {3 otherwise

Suppose that we choose the carrier frequency w2 = w, + Aw/2 and the filter impulse response
frequency w; = w- — Aw/2 where AwT = 72.
(a) Show that the output signal s(t) in the k transmission interval kT <t <(k+1)T depends
only on the input bits d, and d,_;, that is, write s(t) as a convolution of x(t) with h7(t)
and express the result as two integrals with appropriate limits involving only dy and dy_.
Ignore double harmonic terms.
(b) Evaluate your result of part (a) and show that it can be written in the form of an MSK-type
signal, that is

s(t) = A {ug
cos wzt + (—1)* vx cos wt} het Ska T

where {u,} and {vz} are (—1, 0, +1)-valued sequences that depend on dx and dy_; such
that uz, = 0 when v; = +1 and vice versa. Indicate your choice of 0.
10.7 In addition to coherent detection, MSK can be differentially detected as illustrated in Fig. P10.7.
Assume that n(t) is white Gaussian noise with PSD No/2, and s(t) is the usual MSK input, that
is
t
s(t) = Acos(wet + 6(t)); @(t) = deo xy KT SPER EDT
The MSK signal plus noise is multiplied by itself delayed one bit and phase shifted 90°.
The resulting product is passed through a lowpass filter H(f) which, for simplicity, we shall
assume does nothing more than remove second harmonics of the carrier frequency terms, that
is, assume that H(f) is wide enough not to introduce any distortion into the signal s(t). Also
assume that the carrier frequency and data rate are related such that w-T = 2kz (k an integer).
(a) Write expressions for A@(t) = 0(t) — 0(t — T) and the statistical mean of y(t) in terms of
the appropriate data bits {d,} that are valid over the interval kT <t < (k + 1)T.
(b) Show that a single sample of y(t) taken at t = (k + 1)T yields a suitable decision variable
702 Demodulation and Detection of Other Digital Modulations Chap. 10

n(t)

sample att = (k+ IT

Threshold
Vi Compare

y(k+1)T)

Figure P10.7

for detecting d;, that is, express the statistical mean of y((k + 1)T) in terms of d; and give
its values for d, = —1 and dj = 1.
(c) Consider an input data sequence dp = —1, d) = 1, dz = 1,43 = —1,d4 = —1,d5 = —1,
dg = 1, and d7 = —1. Letting p(t) denote a unit rectangular pulse of width 7, and assuming
that noise is absent, sketch, over the interval 0 < t < 87, the following waveforms:
[o,2)

1. d(t)= 2 dgp(t —kT) (the input pulse stream)

k=—0O
2. O(t)
3. O(t —T)
4. AO(t) =O(t) — A(t — T)
5. sin A@(t)
CO

6. d(t —T)= D> y(kK+1)T) p(t — (k + 1)T) [the sample and held values of y(t)].
k=—00
You may assume A = | for the purpose of illustration. Assume an initial phase value
xo = 0. Compare d(t) and de) and justify your previous results.
10.8 Consider an offset QPSK signal with half sinusoidal pulse shaping, that is, precoded MSK. The
conventional receiver for this modulation (see Fig. 10.28 without the differential encoder) uses
I and Q matched filters that are implemented as follows. The I and Q demodulated signals are
multiplied by the half sinusoidal pulse shape p(t) = sin 7t/27) and then passed through I&D
filters spanning the appropriate 7;-sec time intervals. An approximation to the above matched
filter design that avoids the multiplication by p(t) is a windowed I&D filter, namely, one whose
integration time is shortened by an amount ¢ at the beginning and at the end of the integration
interval. In particular, with reference to Fig. 10.28, the multiplications by p(t) in the I and
Q channels are removed and the remaining I&D filters have integration intervals respectively
given by ((2n — 1)7T, + €, (2n + 1)T) — €) and (2nT, + «€, (2n + 2)T; — €).
(a) Evaluate the error probability of the resulting (suboptimum) receiver in terms of the rela-
tive gate parameter &/ Tp.
(b) Find the value of ¢/ 7), that minimizes the error probability of part (a).
10.9 Redo part (a) of Problem 10.8 for the case where the I and Q carrier reference signals are not
ideal and are given as in Problem 10.1. Express your answer in terms of the carrier phase error,
od, and €/ Tp.
10.10 An MFMSK signal can be viewed as an MSK signal with carrier frequency which may vary
from symbol to symbol and can take on one of N equiprobable values. Consequently, the
PSD of the MFMSK signal can be derived as the average of the PSD’s of N MSK signals
with carrier frequencies @; = w, + [2i-—(N +1)]a/T;, i=1,2,...,N. Starting with the
equivalent lowpass PSD for an MSK signal as given by (2.148), derive the equivalent lowpass
PSD for an MFMSK signal in a form analogous to (10.178), namely
&
Problems 703

Figure P10.12

catpyey N N\72

i=0
=

(2+log,N) AIT. Coy

= Tey ae My, sinc(c+ log, N) fT, +i- >)
i=0

N Dp

+ sinc (2(1+ log, N) fT] +i+1- >)

10.11 For MAMSK modulation with M = 16, show the equivalence between the I-Q representation
of (10.185) and the multiple MSK representation of (10.184).
10.12 Consider a 16-signal amplitude-phase-shift-keyed (APSK) modulation whose signal point con-
stellation is composed of two circles (one having radius R; and one having radius Ry) each
containing eight uniformly spaced points (see Fig. P10.12). The phase angles associated with
the points on each circle are identical, that is, for each of the eight possible phase angles,
B; = (i — 1)7/4, 1 =1,2,..., 8, the signal can be at one of two amplitude levels. As such,
it is possible, in principle, to independently detect the amplitude and phase of the modulation.
It is desired to detect this modulation differentially both in amplitude and phase, hence
the acronym 16DAPSK (sometimes called 16 star QAM). Assume an ISI-free environment, for
example, Nyquist filtering equally shared between transmitter and receiver. Then the complex
samples of the receive filter output (taken once every 7; sec) are given by rz = 5ye/° + A, where
5. = V2Paze!™ A /2PS; is the transmitted amplitude-phase signal in the k'" signaling inter-
val, 9 is the unknown channel phase, and nz is the AWGN sample with variance a. Writing sx
in differential form, namely 5, = Sx—1 (+) exp {j (0% — O—1)} = 8-1 Aag exp (j Ady), then
the differential phase Ad, ranges over the set of values {6; = (i — 1)w/4, i1=1,2,...,8}
and the differential amplitude, Aax, ranges over the set of values {y~', 1, y} where y =
Ry/Rz > | denotes the ratio of the circle radii.
704 Demodulation and Detection of Other Digital Modulations Chap. 10

(a) Show that if P = S Y} \5,|* denotes the average power of the signal set, then the radii

of the circle must satisfy the relations

2 2y2
Ri ey ee
‘ 1+y? é 1+y?

(b) Show that the conditional pdf of 7, given Aa,, Ady, Sp_1, 6 is expressed by
F S . 2
lik— Acne (V2PS.-12!*)|
vy (7 |Aax, Agx, 5,-1.6) ae
2no2 exp y— 20;2

(c) It would be desirable to detect the information pair Aaz, Ad, using maximum-likelihood
detection, namely, select that particular pair of values, Ad, Ady, that maximize the pdf
in part (a). However, S,_; and @ are both unknown. Thus, to obtain an implementable
but suboptimum solution, the term /2P S;_,e/® in the argument of the exponential is first
replaced by an estimate, cps (see Problem 10.13 for the characterization of the estimate)
and then the maximization of the pdf is performed. With this replacement, show that the
joint decision can be decomposed as follows:

Ag, = max~'Re [rere 88)

Ad

. es TERRA yaa
Aa,
= min ~' |Aa, — Re {———~e-/ 4%
Aa, s 2
rk-1

Note that the decision rule for Adx follows the form of that corresponding to pure M-
DPSK, whereas the decision rule for Ady depends on the decision Adx.
10.13 In Problem 10.12 the unknown received signal in the k — 1® interval, 5,e/9 = /2P S;e/®, is
replaced by an estimate, r~—1, before maximum-likelihood detection of Aa, and Ady is per-
formed. One possibility for rh | is the maximum-likelihood estimate of 5,e/° using the set of L
past samples r,_1, 7x2, ..-, Tk Wherein the past decisions Adx-1, SGee alee ere
arrived at by the decision rules in Problem 10.12 are substituted for the true past values
IN Spy INGE), TES NE wk 1G
(a) Show that S,_) and S,—1 are related by
I-1
Spor = Se ht Aa, e-i4e-
f=l1

(b) Using the relation in part (a), and defining the past received complex sample vector r 4
Geetha rk—L), Show that the joint conditional pdf of F given Aa, Ad¢,, Seay 6 is
expressed by
1
f (F|Aa, Ad,, Sk 1,6) =
(2102)”

t=

(c) Substituting Ady, AG), |= 1,2,...,

Lfor Adx—1, Ags, 1=1,2,...,
Linthe pdf
Problems 705

of part (b), show that the maximum-likelihood estimate of 5,e/9 = /2P S,e/®, that is, the
value that maximizes the joint pdf is given by

r,—1 = ML estimate of 5;_,e/° =

Upon substitution of this result in the decision rules of part (c) of Problem 10.12, one
obtains a subeptimum detection scheme for 16DAPSK based on decision feedback.
10.14 An approximation (asymptotically tight for high SNR) to the average symbol error probability
performance of an M-ary two-dimensional (amplitude-phase) modulation scheme transmitted
over an AWGN channel is given as
1 M-1

P(E) So) Nina Pim

i=0

where P7;,,,, denotes the maximum of the pairwise error probabilities between signals s;(t) and
sj(t), j =0,1,2,...,M—1, i j which is expressed in terms of the phasor representation
of these signals by

fh | eee
E, is the average symbol energy, and N7,,,, is the number of signals that result in P;;,,,,Tmax for each
value of i, that is, the number of signals at the minimum distance from signal point 7.
Evaluate the approximate average symbol error probability for the 8-ary signal constel-
lations illustrated in Fig. P10.14.
10.15 A quadrature overlapped raised cosine (QORC) signaling scheme is illustrated and defined as
shown in Fig. P10.15 where p; (t) is a unit rectangular pulse of duration 7; and p(t) is a unit
rectangular pulse of duration 27;. The output can be expressed as:

S(t) =A (a anp(t— 2m) COS Wet — (pe byp (t — ar) sin oa]
n=—oo n=—OoO

with p (t) a raised cosine pulse, that is

nim |3u (1 em) Be

=i 2) wp;
eis O<t<4T,
0; otherwise

Note that the transmission rate on the in-phase and quadrature channels is 1/27) but the pulse
width is 47}, that is, the adjacent pulses overlap in time; hence the name of the modulation
QORC.
(a) If we consider the pulse shape p (t) as being generated by passing a rectangular unit pulse

pu (ty Ostsct,
P2\') =) 0; otherwise

through a filter H(f), find the impulse response h(t) = F{H (f)} of the filter.
(b) Show that the average power P of s(t) is related to the amplitude A by P = 3A/4.
706 Demodulation and Detection of Other Digital Modulations Chap. 10

Figure P10.14

(c) From your result of part (a), which choice, when filled in the blank space, correctly com-
pletes the following sentence: A raised cosine pulse of duration 47) is obtained by con-
volving a square pulse of duration 27; with the pulse shape of modulation.
(a) PSK (b) QPSK (c) MSK (d) OQPSK (e) SFSK
(d) Using the results of parts (a) and (c), derive the two-sided equivalent low-pass PSD S;,(f)
of s(t) as given by Eq. (10.183).
(e) If the modulation is now changed to an offset QORC scheme, which is now characterized
by

s(t)=A (Ms an p(t — anny)COS Wet — (» by p(t — 2(n+ »n))sin oe

n=—O0O n=—OO

how does the result of part (d) change if at all?

10.16 A generalized modulation scheme that includes MSK and offset QPSK as special cases, and
has the practical advantage of not requiring any RF filtering, is described as follows. Consider a
modulator in the form of a sequence transducer followed by an RF carrier selector as illustrated
in Fig. P10.16a. The sequence transducer (see Fig. P10.16b) converts the binary (+1) random
data sequence {d;} into a pair of ternary (0, +1) sequences {a,;} and {agi} in accordance with
the relations

l
ais (dj+di-1), agi= an)
Laie (d; — dj_1)
Problems 707

> a, p(t-2n7,) AcOs@,t

n=—0o

¥ 4,7,0-207)
A>-0o

d(t)= )'d,p,(t—nT,)
n=—2

¥
b,p, (t-2nT,)
a>

¥
bp
n=-co
2nf,) Asin @,t

Figure P10.15

{a;;}

{d;} Sequence RF Selector

Transducer Switch SU

—C(t) C0) -Co(t) Colt)

Figure P10.16a

{d;}

(-1)"' (alternating sequence)

{ao; }

Figure P10.16b

(a) Show that for each i, either a7; = 0 or ag; = 0, but not both; that is, in each bit interval,
one of the two transducer outputs is zero while the other takes on value +1 or —1.
(b) Show that the transducer illustrated in Fig. P10.16c, whose outputs satisfy the relations

1 l / / / I
an = 5 (di +di-1), agi = 5 (di +i), di = (-1)'*"d;
 708 Demodulation and Detection of Other Digital Modulations Chap. 10

{d;}

{29; }

Figure P10.16c

and thus have identical statistics, is equivalent to the transducer shown in Fig. 10.16b.
The inputs C;(t) and Cg(t) to the RF selector are carrier waveforms that are as-
signed to the modulator output s(t) in accordance with the following rule:
CMutivan— a
—C,(t) if aj; =—1
s(t) = » ifnst<@+)T
Co(t) if agi = ]
= Cot) if agi = —]

If the carrier waveforms are selected as

Ci(t) = V2P sin IG + =) t +0 , Co(t)=V2P sin |

(v~ =) t +6]
Die 20p

(c ~ Show that the phase of s(t) is continuous at the potential carrier switching transitions, that
is, at times t = iT, for all i and for all choices of the data sequence {d;}. Consider even and
odd values of i separately.
(d) With the choice of carrier waveforms as in part (c), show that the modulator realizes
differentially encoded MSK, that is, show that the output signal, s(t), is equivalent to what
one would get from an MSK modulator preceded by a differential encoder.
10.17 For the modulator of Problem 10.16, assume that the carrier waveforms are selected as

Ci (t) =V2P cos(w@et-+ 8), Colt) =V2P sin\(@ct + A)

(a) Show that the phase of s(t) can never have a transition of z radians. Equivalently, show
that s(t) can never make a transition from C;(t) to —C;(t), or Cg(t) to —Cg(t).
(b) Show that the modulator realizes differentially encoded offset QPSK, that is, show that the
output signal, s(t), is equivalent to what one would get from an offset QPSK modulator
preceded by a differential encoder.

REFERENCES
1. Simon, M. K., “Error Probability Performance of Unbalanced QPSK Receivers,” JEEE Transac-
tions on Communications, vol. COM-26, no. 9, September 1978, pp. 1390-97.
2. Weber, W. J., Ill, “Differential Encoding for Multiple Amplitude and Phase Shift Keying Sys-
tems,” [EEE Transactions on Communications, vol. COM-26, no. 3, March 1978, pp. 385-91.
References 709

. Simon, M. K. and J. G. Smith, “Carrier Synchronization and Detection of QASK Signal Sets,”
IEEE Transactions on Communications, vol. COM-22, no. 2, February 1974, pp. 98-106.
. Simon, M. K., G. K. Huth and A. Polydoros, “Differentially Coherent Detection of QASK for
Frequency-Hopping Systems—Part I: Performance in the Presence of a Gaussian Noise En-
vironment,” JEEE Transactions on Communications, vol. COM-30, no. 1, January 1982, pp.
158-164.
. Simon, M. K. and A. Polydoros, “Coherent Detection of Frequency-Hopped Quadrature Modu-
lations in the Presence of Jamming—Part 1: QPSK and QASK Modulations,” JEEE Transactions
on Communications, vol. COM-29. no. 11, November 1981, pp. 1644-60.
. Divsalar, D. and M. K. Simon, “Maximum-Likelihood Differential Detection of Uncoded and
Trellis-Coded Amplitude-Phase Modulation over AWGN and Fading Channels—Metrics and
Performance,’ JEEE Transactions on Communications, vol. 42, no. 1, January 1994, pp. 76-89.
. Fleisher, S. and S. Qu, “Quadrature Frequency/Phase Modulation,” submitted for publication
in the JEEE Transactions on Communications. Also see S. Fleisher, S. Qu and S. Periyalwar,
“Quadrature Biorthogonal Modulation,’ Canadian Conference on Electrical and Computer Engi-
neering, September 25—27, 1992, Quebec, Canada, pp. 61.2.1-61.2.4.
. Saha, D., “Quadrature-Quadrature Phase Shift Keying,” U.S. Patent 4,680,777, July 1987.
. Saha, D. and T. G. Birdsall, “Quadrature-Quadrature Phase Shift Keying,” EEE Transactions on
Communications, vol. 37, no. 5, May 1989, pp. 437-48.
. Pelchat, M. G., R. C. Davis and M. B. Luntz, “Coherent Demodulation of Continuous Phase -
Binary FSK Signals,” Proceedings of the International Telemetry Conference, Washington, D.C.,
OTe
. Anderson, J. B., T. Aulin and C.-E. Sundberg, Digital Phase Modulation, New York and London:
Plenum Press, 1986.
. Osborne, W. P. and M. B. Luntz, “Coherent and Noncoherent Detection of CPFSK,” JEEE Trans-
actions on Communications, vol. COM-22, no. 8, August 1974, pp. 1023-36.
. Schonhoff, T. A, “Symbol Error Probabilities for M-ary CPFSK: Coherent and Noncoher-
ent Detection,’ IEEE Transactions on Communications, vol. COM-24, no. 6, June 1976, pp.
64452.
. Proakis, J., Digital Communications, New York: McGraw-Hill, 1983.
. Simon, M. K. and D. Divsalar, “Maximum-Likelihood Sequence Detection of Noncoherent Con-
tinuous Phase Modulation,” /EEE Transactions on Communications, vol. 41, no. 1, January 1993,
pp. 90-98.
16. Viterbi, A. J. and J. K. Omura, Principles of Digital Communications and Coding, New York:
McGraw-Hill, 1979.
We Amoroso, F., “Pulse and Spectrum Manipulation in the Minimum (Frequency) Shift Keying
(MSK) Format,” JEEE Transactions on Communications, vol. COM-24, no. 3, March 1976, pp.
381-84.
. Fleisher, S. and S. Qu, “Multifrequency Minimum Shift Keying,’ JEEE Journal on Selected Ar-
eas in Communications, vol. SAC-10, no. 8, October 1992, pp. 243-53. Also see S. Qu and S.
Fleisher, “Continuous Phase Frequency/Phase Modulation,” AIAA 14th International Communi-
cations Satellite Systems Conference, Washington, D.C., March 22-24, 1992, pp. 500-06.
. Austin, M. C. and M. U. Chang, “Quadrature Overlapped Raised-Cosine Modulation,” JCC’80
Conference Record, Seattle, WA, June 1980, pp. 26.7.1—26.7.5.
20. Simon, M. K., J. K. Omura and D. Divsalar, “Performance of Staggered Quadrature Modulations
over Nonlinear Satellite Channels with Uplink Noise and Intersymbol Interference,’ GLOBE-
COM’82 Conference Record, Miami, FL, December 1982, pp. A7.1.1—A7.1.8.
710 Demodulation and Detection of Other Digital Modulations Chap. 10

21. Baker, P. A., “Phase-Modulation Data Sets for Serial Transmission at 2000 and 2400 bits per
Second, Part I,” AEE Transactions on Communication and Electronics, no. 61, July 1962, pp.
166-71.
22. Weber, W. J., II, P. H. Stanton and J. Sumida, “A Bandwidth Compressive Modulation System
Using Multi-Amplitude Minimum Shift Keying (MAMSK),” NTC’77 Conference Record, Los
Angeles, CA, December 1977, pp. 05:5-1—05:5-7.
23. Simon, M. K., “An MSK Approach to Offset QASK,” JEEE Transactions on Communications,
vol. COM-24, no. 8, August 1976, pp. 921-23.
 Chapter 11
Coded Digital Communications

In previous chapters, we studied the reliability and communications efficiency of various

M-atry modulation-demodulation and signal detection techniques. For the most part, the
techniques that have been expounded upon are rather exhaustive of those that are of practical
interest. In particular, in Chapters 3-8, we studied communications reliability and efficiency
without a channel bandwidth constraint. On the other hand, Chapters 9 and 10 consider
communication via channels where constraints are placed on bandwidth occupancy and
information throughput is maximized. In other words, spectrum utilization efficiency was
considered to be a design constraint. The question naturally arises as to what else can
be done to improve communications reliability and efficiency. What is the ultimate limit;
the so-called “thermodynamic limit?” This question can be answered by introducing the
information coding applica at the transmitter and decoding of information applica at the
receiver. As we shall see, a coded! digital communication system is a good example of a
complex system.
This chapter is organized so as to introduce the systems engineer to the notion of
coded communications, top-down system engineering, and design. We begin by introduc-
ing channel coding and decoding using the level 2 and 3 system architectures (see Figs. 1.3,
1.6, and 1.7). This has the advantage that all signal processing functions in the signal path
are combined (levels 4 through 8) into the baseband channel function (see Figs. 1.3, 1.6,
and 1.7). A clear picture of the coded information flow can be presented to the reader from
this system perspective. Next, we introduce the need for a communications carrier which
requires that the modulation and demodulation functions be introduced into the coded sys-
tem architecture; hence, the level 5 system architecture, seen in Figs. 1.6 and 1.7, must be
considered. From an analysis perspective, this function differs from that in an uncoded sys-
tem in that it is mathematically convenient to combine the signal generation-modulation
functions and the signal detection-demodulation functions into the coded-modulation and
coded-demodulation function. At level 5, the IF channel model includes channel features
realized in hardware as well as the propagation media. This IF channel model is time and

1. Did you know? The first code to be used in digital communications was the Morse Code. The Morse Code
System lasted as the primary communication code until the end of the nineteenth century. The insightful Samuel
B. Morse used what is called a “variable-length code”; however, the “variable-length bit pattern” of the Morse
Code was a problem to interpret so a new code was developed. This was done by the brilliant Frenchman Jean
Maurice Baudot. Baudot’s code was based on five bits of information with each alphabet character represented by
combinations of binary digits “1” and “0”; there are 32 such combinations (codewords). It is from Baudot’s name
that the speed (information rate) of information transmission called the baud rate evolved. It is frequently used in
computer communications. Donald A. Huffman was apparently the first to qualitatively investigate variable-length
codes (see Proc. IRE, 40, 1952).
 712 Coded Digital Communications Chap. 11

amplitude continuous, but the demodulator output is continuous in amplitude at discrete

points in time. Since decoders are realized in digital hardware, some form of quantization
of the demodulator output samples is required. This process is characterized in terms of the
quantization levels and a quantization interval for use in the chapters that follow. The quan-
tization converts the continuous IF channel model into an equivalent discrete IF channel
model.
Next, the question of channel memory is addressed. To mitigate the deleterious ef-
fects that channel memory has on communications efficiency, the concept of an interleaver-
deinterleaver system is introduced; hence, the notion of interleaved coding for burst noise
channels is introduced. From this, we are then able to understand the notions of a continu-
ous memoryless and a discrete memoryless channel (CMC and DMC). Using the approach
of top-down engineering, the systems engineer thereby gains the insight needed to design
an error control system for use in a coded digital communication system.
We now proceed by evaluating the reliability and efficiency of transmitting informa-
tion in packets of K, bits; bandwidth is not constrained. This allows us to introduce the
notion of sequential message transmission (bit-by-bit) with the passage of time; hence,
the requirement for a stable clock to mark the boundaries of the information bits at both
the transmitting and receiving end.

11.1 INFORMATION TRANSFER USING BIT-BY-BIT BINARY SIGNALING

Previous chapters have dealt with the problem of transmitting a single message from the
M--ary alphabet” (mo, m1, +-+,my,—1). This has been dubbed the “one-shot” approach to
information transmission in communication theory. In practice, information is generally
transmitted serially in time so systems that communicate a sequence of messages—one right
after the other—are of practical interest.
Consider now the problem of transmitting a message from the M,-ary alphabet by
transmitting bit-by-bit sequentially in time. We need K, = log, M- one-bit transmissions
for each message. Let Tp denote the transmission time per information bit. Then, the trans-
mission time per message is T = K,.7), seconds. The information data rate can be defined
as
DG
Re ies oa bits/sec Chia)

Now the K,-bit binary sequence which represents message m,, can be represented by the
“carrier waveform”
K.-1
sn@)= Yo pibokhh)y ma 0aMe (11.2)
k=0
where 7 is a random variable taking on values “0” and “1” with equal probability and
pi(t), i =0, 1, are the corresponding pulse waveforms. When viewed as vectors in a space
of K, dimensions, the M- points represent the vertices of a hypercube. Assuming the use of
the most efficient modulation-demodulation and detection system (BPSK) operating in the

2. For coded digital communications, we use M, as the number of messages to be coded and reserve M for the
number of distinct waveforms the modulator generates. This is consistent with the notation used in the first ten
chapters.
 Sec. 11.2 Coded Digital Communication System Architectures 713

face of no bandwidth constraint, we have then that po(t) = — Pit) = /2E,/T; sin wet for
0 <t < 7; and zero elsewhere. With equal energy signals given by
Tp Tp
E, = PT, = if s*(t)dt = / p;(t)dt joules/bit (11.3)
0 0

we know from Chapter 4 that the bit error probability is given by

2E
niey=0( 7") =0( am) a (11.4)
where we have assumed the AWGN channel and P is the average received power. Since the
errors in the K,-bit message occur independently, then the probability of correct symbol or
message detection is

Ps(C)
=(1— pp)*¢
=(1— pp)? (11.5)
Therefore, the message or symbol error probability (MEP) is

P,(E)
=1 —(1 — pp)* (11.6)
Since pp < 1/2 for all values of E,/No, then from (11.4) and (11.6), we have

oe Py = Zim. P,(E) > 1 aL

which says that the MEP tends to one as the amount of information transmitted (K, bits)
becomes large. The reason for this behavior is that the Euclidean distance ./2E;, between
nearest neighbors remains fixed as K, increases, whereas the number of nearest neighbors
and the number of dimensions occupied by the signal set increase linearly with K,. The
probability that at least one of the K, relevant noise components will carry the received
vector closer to a neighbor than the transmitted signal vector becomes large as K; increases,
that is, there are K, chances for this to happen. On the other hand for a fixed T and No, the
MEP can be made small by either decreasing the data rate R, or by increasing the received
power of P Watts. This fate of an uncoded communication system performance can be
alleviated by employing encoding and decoding techniques at the transmitter and receiver
respectively; we now investigate methodologies of practical interest.

11.2 CODED DIGITAL COMMUNICATION SYSTEM ARCHITECTURES

Our journey into the application of error-correcting coding and decoding of information as
a technique to further improve communications reliability, P,(£), and efficiency, E»/No,
begins by establishing the coding-decoding viewpoint from the system engineering perspec-
tive. In 1948, the brilliant and insightful Claude E. Shannon demonstrated, via his celebrated
work Mathematical Theory of Communication [1], that by proper encoding and decoding of
information, errors induced by a noisy channel could be reduced to any desired level, subject
to a code rate constraint. Since the appearance of that work, a great deal of effort has been
expended on the problem of designing encoding and decoding techniques for error control
in a communications channel. Documentation of these efforts has led to the development of
the subjects of Information and Coding Theory.
Our treatment of this subject begins by introducing coded communications from the
systems perspective; this serves to take advantage of the system architecture presented in
714 Coded Digital Communications Chap. 11

Chapter 1. In particular, level 3 and level 5 coded-system architectures will now be used to
further our investigation from a systems point of view (see Fig. 1.6).

11.2.1 Coding-Decoding System Architecture (Level 3)

Figure 11.1 illustrates the level 3 system architecture discussed in Chapter 1; this architec-
ture will lead us to our mathematical model of coded information flow between the informa-
tion source and information sink (user). As discussed earlier, we will not treat the problem
of source encoding and decoding? from an information redundancy perspective, that is, the
techniques associated with data compression should not be confused with channel coding
which is used for error control.
For most practical purposes, the channel encoder (see Figs. 11.1a and 11.1b) is char-
acterized by the three parameters M,, N-, and r-. Here, M, serves to characterize the number
of codewords in the codeword dictionary. These M, words are frequently called the code.
The parameter N; serves to characterize the dimensionality (codeword length) of the code
while r¢ specifies the code rate defined as r. = log M-/N¢ bits/dimension. Assuming that
the source encoder output is characterized by a Q-ary alphabet and that the messages that
are generated from this alphabet contain K, letters, then we require that M, = QO. This as-
sures that the source encoder and channel encoder* are matched with respect to the informa-
tion to be transmitted. Clearly, each message contains log, M, = K; log, Q bits. In the bi-
nary case, a case of great practical interest, Q = 2 and M, = 2*< so there are K, = logy M-
bits per message. In general, we will consider the class of (M., N-) Q-ary codes and, in par-
ticular, the class of (N., K-) binary codes.> In the binary case, the code rate is re = K-/Ne
bits/dimension. As we shall later see, r,' serves as the parameter which characterizes the
bandwidth expansion due to coding; both r, and r- | play key roles in what follows.
In the case where we are working with Q-ary codes, we will find it convenient to
refer to the codeword elements as channel or coded symbols. When we talk about binary
codes, we will refer to the NV. elements which make up the codewords as coded message bits,
message bits, or codeword symbols (there seems to be no standard in the literature). Thus,
the log, M, information bits, each of duration Tp sec, inputted into the encoder are converted
into N, coded message bits, message bits, or codeword symbols. The exact prescription for
this mapping defines the encoding algorithm (the type of code) and constitutes the subject
of Coding Theory. Later, in Section 11.6, the classification for the error correction and
detection coding techniques of greatest practical interest is presented.
When time is introduced, we require that each codeword duration T sec satisfy the
relationship T = T, log, M. = T,K- logy Q sec. For a binary code, this means that T =
K-Tp = N-T,- where T, is the duration of the coded message bit or code symbol. From this
relationship, we see that the message bit duration is related to the bit duration via the code
rate, viz., Tc = KeTp/Nc =1rcTp» sec. Thus, the channel symbol rate R = 1/T: = Rp/Tre is
r,' times the data rate Ry. This is why the inverse code rate rz! is referred to as the
bandwidth expansion. For example, a binary block encoder accepts data in blocks of K;

3. The term codec used in Fig. 11.1 was defined in Chapter 1 and refers to the channel coder-decoder pair.
Similarly, the term modem refers to the channel modulator-demodulator pair.
4. In general, N. > K. and the encoder adds redundancy in the data stream for error correction purposes. When
Nc < Ke, the data is compressed to remove redundancy in the information and thus, the term data compression.
For N. = Ke, the encoder acts like a scrambler for secure communications.
5. We use two different notations for binary and Q-ary codes to emphasize the different alphabets.
 Sec. 11.2 Coded Digital Communication System Architectures 715

Channel M, Codewords (M, = 2c)

Encoder N, Dimensions
K, Information Bits
logo M
Code Rate r, = “2 bits/dimension
c

(a) Binary Code (N,, K,)

K, Symbols Channel M, Codewords (M, = Qc)

Encoder N, Dimensions

| M
Code Rate r, = dae = a bits/dimension
Nc c

(b) Q-ary Code (M,, N,)

AAA eneanneneeennesensenannassaseeesuseeseesseeneseneaeeuegenanessusneuansnannsasansunsusasesnacsseaacsnsansenensesenpeasnsenseneeesusnesescessanssseesenensseestensnsansacensegensensesensnsusinansensnteaensenentstensensenensenerenes “

Information Source Codec : Channel Codec Channel

Information Flow Path

Code Rate: l, = 10a MIJN, = K, logs Q/N, bits/dimension Baseband

Channel
Bandwidth Expansion: dite

(c) System Functional Architecture (Level 3)

Figure 11.1 Simplified coded-digital communication system functional model: (a)

Binary code (N;, K,), (b) Q-ary code (M,, N-), (c) System functional architecture
(level 3)

bits and outputs coded message bits in blocks of N, coded bits (see Chapter 12). Since
redundancy r = N, — K- bits is added, N, will be larger than K,. On the other hand, we
shall see in Chapter 13 that binary convolutional coders can also be described as providing
N, output coded message bits for each group of K, input bits. The major difference is that
any channel symbol output consisting of N; coded message bits depends, not only on the
last set of K, input bits, but also on several preceding sets of input bits. Furthermore, with
binary convolutional encoders, K, and N, are usually much smaller than their block code
counterparts (see Chapters 12 and 13). In either case, we will see that low-rate codes have
716 Coded Digital Communications Chap. 11

a greater error-detection-and-correction potential than high-rate codes; as a consequence,

they also require a greater bandwidth expansion.
Refer now to the coded baseband system illustrated in Fig. 11.1. We assume that the
information source is discrete and outputs a message symbol m from the M--ary alphabet
m=0,1,--:,M-— 1. The source encoder transforms the information source output into
the random sequence u = (uj, U2,--+,WxK,); in particular, if message symbol ‘m is trans-
mitted then u = Uy, = (Umi, Um2,°**,UmK,). When the source output symbol alphabet is
composed of the binary digits u;=“0” or “1”, the symbols are generally called information
bits; this is the case of greatest interest. We will think of these binary digits as corresponding
to “+1” volt or “—1” volt. y
The channel encoder, according to known rules, transforms the source encoder out-
put sequence u into N. channel input symbols (coded message bits) denoted by the vec-
tor x = (x1, X2,-++,xXN,); X is called a codeword for u. This transformation, u — x, is
called channel encoding.° In particular, if message m is transmitted then m > Up, > Xm =
(Xm1,Xm2,°**,Xmwn,). When the elements of X, SAY Xmn, are drawn from a Q-ary alphabet,
then the number of possible codewords in the vector space is Q%* words; only M, = O* of
these are needed and are included in the codeword dictionary. We also note that each channel
input symbol x, € x can depend on the source encoder output symbols u; € u in any desired
way. The channel symbols (coded message bits) are denoted by x» during the known n’ :
symbol time (n — 1)T- <t <n7-; m=0,-:-, Me —1;n =1,2,--*, Ne. We assume that
the source and channel clocks at the transmitter and receiver are synchronized.
Example 11.1
Consider a 4-ary message source (M, = 4) whose output is to be encoded using a binary (Q =
2) rate 2/5 code (K; = logy Me = 2, Nc = 5). The four possible messages mo, m , m2, m3 are
respectively represented by the four possible binary sequences 00, 01, 10, 11 which serve as
the input to the channel encoder in each T-sec interval. Of the 2%« = 2° = 32 possible binary
sequences of length 5, we choose the following four sequences to act as channel encoder
outputs corresponding to the four possible input sequences.

Output Sequence
Message Input Sequence (Coded Bits)

mo 00 00000
my 01 01010
m2 10 10100
m3 11 11111

This one-to-one mapping between input and output channel encoder sequences defines the
code. Note that the above mapping produces a nonlinear code, that is, the modulo-2 sum of
two output sequences (codewords) does not necessarily generate another possible codeword.
Example 11.2
Consider a 16-ary message source (M, = 16) whose output is to be encoded using a 4-ary (OQ =
4) rate 2/3 code (K, = logg M. = 2, N- = 3). The 16 possible messages mo, m1, m2, ..., 15
are respectively represented by the 16 possible 4-ary sequences 00, 01, 02, 03, 10, 11, 12,
13, 20, 21, 22, 23, 30, 31, 32, 33 which serve as the input to the channel encoder in each
T-sec interval. Of the Q%: = 43 = 64 possible 4-ary sequences of length 3, we choose the

6. The codes used in the examples in this chapter are for illustration purposes only and are not necessarily “good”
codes.
Sec. 11.2 Coded Digital Communication System Architectures 717

following 16 sequences to act as channel encoder outputs corresponding to the 16 possible input
sequences.

Output Sequence
Message Input Sequence (Coded Bits)

mo 00 000
my 01 010
m2 02 202
m3 03 030
m4 10 102
ms 11 020
m6 12 302
m7 13 110
mg 20 002
mo 21 300
m0 22 133
m\\ 23 210

m2 30 321
M13 31 111
mi4 32 322
m15 33 213

Once again the above mapping represents a nonlinear code since the modulo-4 sum of any two
output sequences (codewords) does not necessarily result in another codeword.
Based on the coded system architecture discussed thus far, we observe that, as N,
gets large, the decision regarding the message is delayed for T = N.T, = K-Tp log, Q sec.
This delay is called coding delay and represents a constraint which must be considered by
the system engineer. Moreover, since M, = Q*« grows exponentially with K, we would
expect the complexity of the encoder and decoder to grow rapidly; this represents another
issue to be faced by the system engineer. In practice, there are other issues which affect the
system engineer’s choice of a particular coding-decoding technique, for example, it is driven
by implementation complexity, performance requirements, and cost. We will not be able to
address these issues from a theoretical point of view.

Baseband channel characterization. The physics of the baseband channel is

that, for analysis and synthesis purposes, it can be modeled as a CMC. For decoding pur-
poses (using digital hardware), this channel can be converted into a DMC. The baseband
channel maps the input vector x into the output vector (observables) y = (y1, y2,°*+, YN.) €
RNe where R< is the N,-dimensional Euclidean space. In practice, the vector components
Yn are continuous random variables.
In general, the continuous baseband channel is characterized by the known transition
probability density function f(y|x); the continuous memoryless baseband channel is char-
acterized by
Ne
;

fax) =| | fOnlzn). (11.8)

n=1
718 Coded Digital Communications Chap. 11

Furthermore for the AWGN channel, if codeword x, is transmitted, then x, = Xmn and

1] (Yn — sony]
Xmn) = ——=—= exp |-—————_ (11.9)

for alln =1, 2, .-+, Noand m=0, 1,+--, M.—1.

Now, the channel output sequence y = (y1, y2,---, yN,) Serves as the input to the
channel decoder whose purpose is to produce the output sequence v = (vj, v2,--+-,U Ke)
such that the effects of the channel disturbances are minimized, that is, the number of errors
between the output sequence v and input sequence u is minimized. The transform y > v
is called channel decoding. From a system performance perspective, we are interested in
block errors, that is, errors that occur at the receiver in the transmitted block (x1, ---, Xy,)-
Thus, one can conveniently drop the sequences u from consideration and regard the channel
encoding simply as a mapping from the m integers, m = 0, ---, M, — 1, onto the codewords
Xo, °::, Xm,—1. If message m enters the source encoder, X is transmitted, and on the basis
of the channel output sequence y, the channel decoder outputs v = (v),---, vx.) and the
source decoder produces m. An error occurs if and only if m #~m. We will demonstrate
the conversion of the CMC into the DMC after we introduce the reader to the Level 5
coded system architecture (see Fig. 1.6) which incorporates a communications carrier. This
level of modeling requires careful consideration of the interfaces between the coder and the
modulator and between the demodulator and the decoder.

11.2.2 Coded-Modulation and Demodulation-Decoding System Architecture

(Level 5)

Level 5 of our information transportation system architecture (see Fig. 1.6), serves as the
coded system model when a communications carrier of intermediate frequency (IF) f,
Hz is introduced to transport the information (see Fig. 11.2). Introducing the carrier into
the system picture requires the use of a Modulator and a Demodulator and this requires
separate attention. Unlike the uncoded system, for analysis and synthesis purposes only, the
modulator function in a coded system can be viewed as incorporating the two functions of
signal generation and modulation. Similarly the Demodulator in a coded system can also be
viewed as incorporating the two functions of demodulation and detection (see Fig. 11.3). We
adopt this point of view here.’ In practice, the demodulator projects the received IF signal
to baseband; this projection is accomplished by a carrier synchronization circuit.
In one simple form, the modulator can transform, via a known technique, the code-
word x from the channel encoder output into the transmitted signal vector S,, [waveform
Sm(t)] for presentation to the IF channel. In yet another simple form, the modulator can
transform each coded bit or symbol from the codeword into the transmitted waveform. In
between, the modulator can transform a group of coded bits or symbols into the transmitted
waveform. In general, this process is tedious and is best accomplished in two steps using a
signal mapper which we now discuss.

7. Demodulation involves four major tasks: (1) synchronization of the local communication carrier with the
received carrier component used at the modulator, (2) use of this reference to convert the received modulation
to baseband, (3) extraction of channel symbol timing so that the “matched filtered” version of this baseband signal
can be sampled and the recovered clock synchronization maintained with the transmit clock, and (4) detection of
the demodulated channel symbols and decoding them in accordance with the appropriate decoding algorithm.
 Sec. 11.2 Coded Digital Communication System Architectures 719

Prttcsceeneeenessenscenncnececnsacccncenerscececsncccccscccccensencccccacececs.

Information Source Codec Channel Codec Modem : Channel

Information|
Source He
i™1 Source
Encoder
Channel
Encoder
:Xm| Modulator Salt
(M,, No, T,)

farhdZ

Source Channel Local IF

Clock : Clock : Oscillator Channel

Frequency and] Timing S) stem

Information ee Source 7 Channel a Demodulator

Sink : Decoder Decoder

Figure 11.2 Coded-digital communication system functional architecture (level 5)

Modulator characterization. For our purposes, a digital modulator is charac- -

terized by the three parameters M, N, and 7;. Here M represents the number of distinct
waveforms in the set {s,_(t); m =0, 1,---, M — 1} which the modulator can generate and
T; represents the time duration of each waveform (signal). We emphasize that the set
{Sm(t); m =0, 1,---,M — 1} of modulator waveforms is not in general equal to the set
{Sm(t); m =0,1,---, M- — 1} of coded modulation waveforms. As demonstrated in Chap-
ter 4, these waveforms can be decomposed into N pieces using the orthonormal expansion
N
aD=
> onan 0 Sts Te=NTy (11.10)
f=

and zero elsewhere; we will refer to s,,(t) as the transmitted signal when codeword Xj,
is to be transmitted if and only if N = N, and M = M,. When this is not the case, we
will consider mapping methods to provide codeword transmission. From the above repre-
sentation, it is clear that the waveform set {s,,(t),m =0,---,M — 1} can be represented
by the N-dimensional vector 8 of its coefficients, that is, Sm = (Sm1, Sm2,°°°,SmN),m =
0, ---, M — 1. In other words, there exists a one-to-one correspondence between the wave-
form s(t) and the vectors s,,. AS we shall see, s(t) = Sm(t) if and only if M = M, and
N= Ne.
It is convenient to identify Ty 4 7;/N as the time used per dimension by the mod-
ulator and D = 1/Ty as the number of dimensions used per second by the modulator. In
Chapter 12, we will need to introduce the notion of bandwidth as it relates to signal dimen-
sionality. The interconnect will be made (Chapter 12) via the Shannon-Nyquist sampling
rate defined as B = 1/(2Ty) Hz; there, we identify dimensionality D = W = 2B Hz as the
required channel bandwidth.
There are two modulator rate parameters of practical interest. They are defined using
720 Coded Digital Communications Chap. 11

eeneceencnsessacenoscscncecesenss
eeeceveeseeesvesesescenecccssccscccsceaccscncscsecccscecesececsscsssesascsacassecssescseseassvaconsssevoesenenscusscsscasssocsnooaussssnoasooonavasseassesgseeeseeeey

Encoder ( Sm(t) §
(Mg No,T)

T = Tp loggM, = Ng Ty = kTgi Ty = Te OG2M = NTY

logoQ
k = logo>M,log2Q/r, logoM = N./RyN = N, logoQ/logaM.

(a) Modulator Model for Coded System Design

To
Decoder

(ob) Demodulator Model for Coded System Design

Figure 11.3 Combining transmitter and receiver functions in a coded-digital

communication system: (a) Modulator model for coded systen design, (b) Demodulator
model for coded system design

the parameters which characterize a digital modulator. These are the channel rates defined
by R A log, M/T; bits/sec or by Ry A logy M/N bits/dimension. Thus, the modulator
output rate is R = Ry/Ty = DRy bits/sec. These two rates will play a key role in what
follows. We will be interested in coded digital systems which are implemented using the
(M., Nc, Tc) code and the (M, N, T,) modulator. Thus, the modulation/demodulation tech-
niques studied in the earlier chapters are applicable here. In the coded digital communica-
tion system architecture of Fig. 1.6, the signal generator and signal detector serve as bridges
which respectively connect the channel encoder to the modulator and the demodulator to
the decoder. We will describe these bridges next and present combined coding-modulation
waveform designs which can support the transmission of binary and multilevel codes using
the modulation-demodulation techniques (BPSK, QPSK, M-PSK, M-FSK, QASK, QAM,
and so on) presented in Chapters 3 through 10. In this chapter, we deal with rectangular
baseband pulses; generalization to arbitrary pulse shapes is straightforward (see Chapter 9).
Finally, the energy in the m'” waveform output of the (M, N, T;) modulator is

Ts
E,, = / s*(t)dt (11.11)
0

for m=0,---, M — 1. For equal energy signals, E;,,= Es and the energy generated per
dimension is Ey = E;/N joules. Since the time per dimension is Ty = T,/N, then E, =
PT, = PNTy = NEv joules. If the (M, N, 7;) modulator is used k times per codeword
transmission and we denote this composite waveform by S,,(t), then the energy required per
Sec. 11.2 Coded Digital Communication System Architectures 721

codeword is E = PT = kE, =kPT, = NcEn, = (Ke logy Q) PT, = (Ke logy Q)Ep Joules.
These various energy relationships will be useful in what follows.
Example 11.3
Consider mapping the channel encoder output into a ternary (M = 3) signal set of dimension-
ality N = 2. Such a signal set can be generated by an M-PSK modulation with normalized
signal points located on the perimeter of a unit radius circle at 9 = 0, 27/3, and 47/3 radians.
The orthonormal basis functions in (11.10) are defined by o1(t) = /2/T,cosw.t, d2(t) =
V/2/T; sin wet and the coefficients, 5.,, for each signal are tabulated below:

Signal 7] Coefficients

So(t) 0 Soi=1, so =0
51 (t) 21/3 Stl /2s 512 = V3/2

S2(t) 4x /3 $23 = —1/2, 527 = —V3/2

Signal generator mapper characterization. Clearly, the output of a (M;., Ne,

T.) channel encoder cannot directly serve as the input to a (M, N, T;) modulator. The prob-
lem is that, in general, the codewords in the code dictionary cannot be mapped one-to-one
with the modulator output waveforms s,,(t) because M, 4 M; a matching problem occurs
if M. > M. In this sense, we will use the signal generator in a coded digital communica- -
tion system as a mapper, viz., one that maps codewords (vectors) onto the signal vectors
or modulator output waveforms; coding theorists frequently call this a mapper. For exam-
ple, the mapper could be a demultiplexer which partitions the coded message bits into two —
coded message bit streams. One stream could be used to modulate the I-channel and the
other to modulate the Q-channel of the modulator.
We begin our discussion regarding this mapping process by considering the class of
(M_, Nc, Tc) Q-ary codes. In this class, log, M; = K. log, Q information bits are mapped
into one of the M,. codewords containing N, coded symbols (or coded bits). Since Tp is
the time per information bit, then K, log, Q bits require T = (K, logy Q)T> sec. These bits
must be mapped into the N,. coded symbols in the same amount of time, that is, T = N-T;
sec. At this time, we define also the coded bit duration JT.» = T,/ log Q. Note that TJ,» is not
necessarily a physical parameter and differs from the information bit 7). For purposes of dis-
cussion, let us assume that the modulator is capable of producing M points (vectors) in N-
dimensional Euclidean space; each point contains log, M bits of information. If M > Mc,
the modulator produces enough waveforms to map one-to-one with the codewords. The
problem then is to select the best subset of waveforms. If M < M-, a problem exists and
something must be done to match the modulator output to the channel encoder output. In
particular, let us assume that we group the N, log, Q coded message bits into k groups of
log, M coded bits such that N, log, Q = k log, M. In the special case of binary codes, we
group N, coded bits into k groups of logy M coded bits such that N. = k logy M. Physically
speaking, k=N- log, Q/ log, M = (logy M,) (logy Q)/r- logs M = (N-/N) (log, Q/Rn)
characterizes the mapper. The parameter k is the ratio of the number of coded message bits
per codeword to the number of bits per message that the modulator can represent in T; sec.
hens = Nela= kT. (logy M/ log, Q) which means that, in T sec, we must use the mod-
ulator k times. This says that within T sec, the modulator can produce one of Mk > M-
waveforms (vectors); we only need M,. which the encoder output selects. Since we re-
quired that T =kT, = kT, log, M/ logy Q, then T; = T; logy M/ logy Q. Thus, the channel
rate R = (log, M*)/T = log, M/T; = log, O/T. = log, Q(R»/r-) bits/sec is matched to
 722 Coded Digital Communications Chap. 11

F Um1 af Um2 a x | UmKe |

u:

(a) Encoder Input

Kr x m2 Xm3 < Fr
XmNo =

0 tH 20s 3T, t—> T=NoTo=KeTp

(b) Encoder Output

Cn) nC) {S(t

c: (1 of M) i (1 of M) t cc t (1 of M) :
0 T, 21, t—> T=kT,logsM = kT, = KNTy
(c) Modulator Output

{Sy(t)}
(1 of Mo)

- T
(d) Coded Waveform

Figure 11.4 Encoder-Modulator mapping and timing diagram when Q = 2

the information rate as required. In this sense, the (M;., N., Ts) code is mapped onto the
(M, N, T) modulator.
More generally, Table 11.1 defines all time intervals, rates, energies, bandwidths,
and so on. Table 11.2 summarizes the coder-modulator time-interval and rate relation-
ships, while Fig. 11.4 demonstrates the time scale mapping and summarizes the encoder-
modulator mapping for some examples which follow. Some examples will clarify the role
of this mapper (see Table 11.2)
Coded BPSK: Consider first the (N., K;) binary code and the (2, 1) BPSK modulator. Here
M =2 and N =1 s0 thatk = N., with T = N.Ty and Ty = T,. In principle the n!” coded
message bit is used to modulate the basis function g(t) = /2/Ty cos wet, (n — 1)Ty <
t <nTy, which is used serially in time for all n = 1, 2,---, N-. In other words, the set of
Sec. 11.2 Coded Digital Communication System Architectures 723

Table 11.1 Fundamental Parameters of a Coded Digital Communication System

T-—Time used per codeword dimension, sec

T;—Modulator waveform duration; T, = NTwy sec
Time intervals T,—Time per message bit, sec
T—Codeword duration; T = (log, M-)Tp = kT, =I Peel iol paSeC
Ty—Time used per modulator dimension by modulator, sec
A
k—Number of times modulator is used per codeword transmission
M, = Q*’«—Number of codewords in code dictionary
K,-—Number of information symbols per codeword (message
Codec/modem length)
N-—Number of coded message symbols per codeword
(Codeword dimensionality = block length)
M—Number of waveforms modulator generates per duration 7, sec
N—Dimensionality of modulator; N = WT, = WT /k

Rp = (logy M.)/T =re/T: = 1/Th, bits/sec

R = log, M/T; = Rn /Tw; modulator output rate, bits/sec
re = logy M./N-¢; code rate, bits/dimension
Ry = (logy M)/N; modulator output rate, bits/dimension
Ro—Channel cutoff rate, bits/dimension
Ro = RoW—Channel cutoff rate, bits/sec

E = PT; received energy per codeword

E, = PT;; modulator waveform received energy per use of modulator
Ey, = PT,; received energy per codeword dimension
Energy (joules) E, = PTp; received energy per information bit
En = PTy; received energy per modulator dimension used
No = kT°; noise energy or single-sided PSD
P = E/T; received power, Watts

W =N,/T = 2B; two-sided IF Nyquist bandwidth

Bandwidths B =1/2T-,; single-sided Nyquist bandwidth
(Hz) D =1/Ty = N/T; number of dimensions used per sec by modulator
D = W = 2B for binary coded BPSK and binary coded QPSK
r~! and Ry'—Coder and modulator bandwidth expansion, respectively
: R =Blog, Ma oss Men) URN 4: Peis i «
Normalized WwW
“Wi = NWTy = Wy Dits/sec/Hz = bits/dimension
throughputs Re = 1c, bits/sec/Hz = bits/dimension

Frequency fc - IF channel center frequency, Hz

basis functions are time-orthogonal within the transmission interval if @, is chosen to be

some multiple of 2/Ty. If x» is to be transmitted, then the BPSK coded waveform model
becomes, form =0,1,---,M.—1

j
Sm(t)
= /2/Tn > Smncoswct; O<t<T=NTy (11.12)
nae
724 Coded Digital Communications Chap. 11

Table 11.2 The (M-, Nc, T-) Code Mapped into the (M, N, T,) Modulator at a
Bit Rate of R, = 1/7), bits/sec

Rate (bits/dim) e= log, M./Ne Ry = log, M/N

= K. logy Q/Ne
when M, = Q*«
Time intervals T= (log, M.)T» T; = NTy = (log, M)T./ log, Q
= (K; log, Q)T, T. = NTy log, Q/ log, M
= Ved = Ty log, O/Rn
SOM == ell, so Ty = rcRnT)/ log, Q

Rates (bit/sec) Code rate: Channel rate:

R = log, M/T;
= log, M/NTy = log, O/T
or R = Ry
/Tn logy O/T:
= (Rp/re) log» Q

Coder—modulator | T = kT, = kT.(logy M/ log, Q) =NcTe = (Ke log Q)Th so

time interval 1h 1hel| °82 M/ logy Q =rcT, log, M/ log, Q. Therefore
matching £ logy Mc = Se Q (N.\ _. Nelog, Q

and Sinn = /EnXmn for (n — 1)Ty < t <nTw and zero elsewhere. The parameter Ey rep-
resents the energy per dimension. This coded waveform is generated using the (2, 1) BPSK
modulator k= N- times. If one now identifies the coded message bits x.» with the voltage
levels +1 or —1, we have a waveform carrying N,. coded message bits (see Table 11.2 and
Fig. 11.4). Figure 11.5(a) demonstrates the coded waveform mapping. Needless to say, bi-
nary phase-shift-keying combined with coherent detection can always be used to transport
coded information. As we shall see shortly, QPSK can also be used.
Coded QPSK: An alternate approach to transmitting the (N., K;) code is to use the (4, 2)
QPSK modulator. Here M = 4, N = 2 and one uses the modulator k= N-/ log, M = N,/2
times to generate the coded waveform S,,(t) corresponding to one codeword. In this case we
introduce an I-channel and a Q-channel which supports the transmission of the coded mes-
sage bits. Thus, T = kT, ss 2kT = N.T, and T; = 2Ty = 2T,. In order to map the code-
words onto the (4, 2) QPSK modulator, we demultiplex the coded message bits out of the
encoder for codeword x,, into the | and Q channel components Xmz = (Xm2, Xm4, +++, XmN,)
and Xm@Q = (Xm1, Xm3,***,XmN.—1). The orthonormal basis functions are $7 (t) = d2n(t) =
J/2/Ty cos wet and Pg(t) = d2n-1(t) = /2/Ty sin wet with w, a multiple of 27/Ty and
(n—))ivest snine n= 1,206, k. The resulting QPSK coded waveform model be-
comes (for N, even)

A
k k
Sin(t)= /2/Tw pe, Smn
COS Wet +) SmnSinwct | O<t<T (11.13)
n odd

1 channel Q channel

where Sn A / ENXmn for QPSK and xm» takes on value +1 or —1 volts. Figure 11.5 illus-
 Sec. 11.2 Coded Digital Communication System Architectures 725

l- or Q-Channel py (t) = V2/Ty COS wet; (Nn -1) Ty StS nTy and zero elsewhere
Sint :
J Yi “
Yy A yj 4,

(NZ
if SLE,

T=

(a) For Binary Coded BPSK, k = N,

I-Channel Pon-4(t) = V2/Ty COS wt; (N -1) Ty <t<nTy and zero elsewhere

i. «
Sint

0F Og
Y Z
a Lia LLL

Ts =2Ty 21. fe tk Ay. T =kT, = 2kTy

Q-Channel sin wet; (n -1) Ty < t <nTy and zero elsewhere

on(t) = V2/Ty

a(t) ‘7

7 Ot he

T; T = kT, = 2kTy
(b) For Binary Coded QPSK k = N,/2, 8-PSK k = N,/3, and M-PSK k = N,/ogoM
Figure 11.5 Codeword transmission using a set of time-orthogonal functions

trates the use of time-orthonormal and time-orthonormal quadrature functions to transport

coded information.
Coded 8-ary PSK: Finally, consider the transmission of the (N;, K;) code using the (8, 2)
8-ary PSK modulator. Then log, M = 3, k = Nc/3 and T =kT, = 2kTy = 3kTe = K¢Th.
The coded message bits are sequentially mapped three at a time onto the eight phase states
of the modulator. Table 11.3 summarizes various coded modulation mappers for the class
of binary and multilevel codes. Table 11.4 summarizes the corresponding coded modulation
waveforms S,,(t) corresponding to the transmission of codeword x»,m = 0, 1,---, Me —
1. The coded modulation waveforms needed for transmission of nonbinary codes will be
given in the next chapter, that is, aside from the use of time-orthonormal functions one
726 Coded Digital Communications Chap. 11

can use frequency-orthonormal functions. Table 11.5 summarizes numerous communication

carriers of great practical interest; these carriers represent basic building blocks needed in
the generation of {S,,(t)}.
Example 11.4
Consider a 4-ary message source (M, = 4) whose output is to be encoded using a binary
(Q =2) rate 2/3 code (K- = logy M, = 2, N. = 3) and transmitted over the channel using
an M-PSK modulation. Define the mapping of the four possible messages mo, m1, m2, m3,
which are respectively represented by the four possible binary sequences 00, 01, 10, 11, into
codewords as below:
Message Input Sequence Output Sequence

mo 00 000
mM, 01 001
m2 10 010
m3 11 111

(a) Coded BPSK Modulation (M = 2, N = 1)

Here the modulator is used k = N./ log, M = 3 times to generate the coded waveform Sj»(t)
corresponding to each codeword. That is, each of the N. = 3 coded bits per codeword is as-
signed to one BPSK waveform. These waveforms are transmitted sequentially over the channel
every T; = Ty = T/3 sec. Also note that 7, = 3Ty/2 = 37, /4. Letting so(t) = /2/Tp cos wet
be assigned to a “0” coded bit and s;(t) = —./2/T7;, cos wet be assigned to a “1” code bit, then
the mapping from a typical message sequence to the transmitted BPSK modulation would ap-
pear as below:

Message Sequence: mi,Mo,M3.m,,mM2, M0, ..-

Coded Bit Sequence: 001000111001010000...
Modulator Output: so(t), So(t), s1(t), so(t), so(t), so(t), si(t), s1(t), 51 (0),
Coded Modulation Waveform S\(t) So(t) S$3(0)
So(t), So(t), s1(t), So(t), 81(t), So(t), so(t), So(t), so(t), ...
es

Si (0) S2(1) Sot)

(b) Coded QPSK Modulation (M = 4, N = 2)

Here the modulator is used k = N-/ logy M = 3/2 = 1.5 times to generate the coded wave-
form S,,(t) corresponding to each codeword. That is, two of the N. = 3 coded bits per code-
word are assigned to one QPSK waveform. These waveforms are transmitted sequentially
over the channel every 7; = 2Ty = 27/3 sec. Also note that 7, = 3Ty/2 = 37,/4. Letting
so(t) = /2/T,; cos wet be assigned to the coded bit sequence “00,” 5, (tf) = —/2/Ty sin wet be
assigned to the coded bit sequence “01,” s2(t) = —./2/T, cos wet be assigned to the coded bit
sequence “10,” and s3(t) = ./2/T, sin ct be assigned to the coded bit sequence “11,” then the
mapping from a typical message sequence to the transmitted QPSK modulation would appear
as below:

Message Sequence: m,™Mo, m3.m,,m2,mMo,...

Coded Bit Sequence: 00 10 00 11 10 01 01 00 00...
Modulator Output:
a)
So(t),82(t),
—e—_ en
So(t), OO
s3(t),$2ee
(t), s1(t),
eee
s1(t),So(t),
eee SS
so(t), ...
Coded Modulation Waveform Si(0) So(t) S$3(t) S(t) S(t) So(t)

(c) Coded 8-PSK Modulation (M = 8, N = 2)

Sec. 11.2 Coded Digital Communication System Architectures 727

Table 11.3 Coded Modulation Waveform Mappers for Binary and Multilevel Codes

Codec-Modem Coded Message Bit

Modulator Parameters Communication Carrier to Modulator

On(t)= /2/Ty cos@ct | Smn = /EnXmn With

= y(t) for all Xmn = +1 or —1 for all m
leo peer and ANOh = Ip Was cay he
(n —1)Ty St <nTy;
@c is a multiple of 7/Ty

I-channel: @2,_1(t) Sip A) EIN Xp wien

= 4/2)Tn COS @,t Xmn = +1 or —1 for all m
Q-channel: ¢2,(t) AOD = Np Bo sso ols
= /2/Ty sin wet,
Te BY seen
TNT ee ks
and @, is a multiple of
2n/Tyn

I-channel: Same as I-channel: (n odd)

QPSK Smn = VEn cos(m/4)
Q-channel: Same as Q-channel: (n even)
QPSK Sinn = JEn sin(rm/4)
and @, is a multiple for all m and |
of 27 /Ty le oooh

Same as QPSK I-channel: (n odd)

Sinn = VEn cos(1m/M)
Q-channel: (n even)
Smn = VEn sin(m/M)
for all m and |
Ti Ne envhG

n(t) = /2/TN cos wt, Smn = JV En, Xmn with

(n—1)Ty St <nTy, Xmn = +1 or —1 for all m
and w, is a multiple of elie Oe oe oa hs
n/Ty

Same as QPSK I-channel: (n odd)

Sn = V En,Xmn
Q-channel: (n even)
Sian = V En, Xmn
Xmn = +1 or —1 for
all m
ance Dee

Quadrature Same as QASK Same as QASK Same as QASK except

Amplitude nonrectangular pulse
Modulation shapes are used in IJ- and
(QAM) Q-channels (see Chapter
10)
728 Coded Digital Communications Chap. 11

Table 11.4 Coded Modulation Waveforms for Transmission of Binary and Multilevel Codes

Coded Modulation Waveform Communication Carrier

Modulator 0<t<T=hkT, Coded Message Bit Map

(2, 1) BPSK) %p< Smif) = op) Sn (Gn) for See Table 11.3,
or(M,
1) M-| m=0,1,...,Me—1;
k= Ne columns
2 and 3_
ASK
k k
(4, 2) QPSK, Xb) — Sy Sin P2n—1() + De SmnP2n(t) See Table 11.3,
(M, 2) n odd n even columns 2 and 3
ASK, M-
fee M-FSK I-channel _ Q-channel
Or =O, Nn bsog ie lle Kei

The expansions assume unit energy rectangular pulse shapes; if other baseband pulse shapes
are used these models can be readily modified (see Chapters 9 and 10) to accommodate
arbitrary I and Q pulse shapes

Table 11.5 Communication Carriers of Greatest Practical Interest

Communication Carrier Channel Waveform

Time-orthogonal (w, a Gn(t) = V2/Ty sin@ct, (n—1)Ty St <nTy

i 0, else
multiple of 2/ Ty)

Time-orthogonal (I-Q Q2n(t) = . 2/Ty sin wet, . —1)Ty <t <nTy

modulation) (@, a multiple ’ else
of 2z/Ty) donai(t) =} V2/TN coset, (n—1)Ty <t <nTy
’

Frequency-orthogonal (w. | Y(t) = x 2/T sin[(we +2n/T)t], O<t<T

a multiple of z/T) :

Frequency-orthogonal Gon (t) = /2/T sin [(@e + 2mn/T)t),

(I-Q modulation) (@;, a Pon+i(t) = /2/T cos [(we + 20n/T)t],
multiple of 27/T)

Here the modulator is used k = N-/ log. M = 3/3 = | times to generate the coded waveform
Sm(t) corresponding to each codeword. That is, all three of the N. = 3 coded bits per code-
word are assigned to one 8-PSK waveform. These waveforms are transmitted sequentially
over the channel every T; = 2Ty = T sec. Also note that 7, = Ty = T, /2. Letting So(t) =
/2/T; cos wet be assigned to the coded bit sequence “000,” S)(t) = /2/T, cos (wet + 1/4)
= J/1/T, coswt — /1/T; sinw-t be assigned to the coded bit sequence “001,” S(t) =
J2/T; cos (wet + 1/2) = —./2/T,
sinwet be assigned to the coded bit sequence “010,” etc.,
then the mapping from a typical message sequence to the transmitted 8-PSK modulation would
appear as below:

Message Sequence: m1, Mo, M3.mM,,m7, Mo, ...

Coded Bit Sequence: 001 000 111 001 010 000...
Modulator Output: SEE), Soe), S3(0), ssn(sob).
so )s cee
— we eee Se
Coded Modulation Waveform S(t) So(t) SOY SiG). BAG So(t)
Sec. 11.2 Coded Digital Communication System Architectures 729

(d) Coded 16-PSK Modulation (M = 16, N = 2)

Here the modulator is used k = N-/ log, M = 3/4 = .75 times to generate the coded wave-
form S,,(t) corresponding to each codeword. That is, all three of the coded bits of one code-
word and one bit of the next codeword (or two coded bits of one codeword and two coded
bits of the next codeword) are assigned to one 16-PSK waveform. These waveforms are
transmitted sequentially over the channel every T; = 2Ty = 47/3 sec. Also note that T, =
3Ty /4 = 3T,/8. Letting so(t) = /2/T, cos wet be assigned to the coded bit sequence “0000,”
s(t) = /2/Ts cos (wet + 1/8) = (./2/T; cos 1/8) cos wet — (./2/T; sin 7/8) sin wet be as-
signed to the coded bit sequence “0001,” s2(t) = /2/T, cos (wt + 1/4) = /1/T, cos wet —
/1/T, sin wet be assigned to the coded bit sequence “0010,” etc., then the mapping from a
typical message sequence to the transmitted 16-PSK modulation would appear as below:

Message Sequence: m,,Mo,m3.m,,m2, M0, ...

Coded Bit Sequence: 0010 0011 1001 0100 00...
Modulator Output — : SiG Gal(@))aae, (2) 5 SHEN cae
——— —_—_— Seen Sen eer
Coded Modulation Waveform Si(t) Sot) S30) S(t) S2(1)
Example 11.5
Consider a 2*-ary message source (M, = 2") whose output is to be encoded using a binary (Q =
2) rate k/k + 1 code (K. = logy Me =k, Ne =k + 1) and transmitted over the channel using
an M-PSK modulation where M = 2‘+!, This is a generalization of case (c) in Example 11.4.
Here the modulator is used k= N-/ log, M = (k + 1) /(k + 1) = 1 times to generate the coded
waveform S,,(t) corresponding to each codeword. That is, all kK + 1 of the N. =k + 1 coded
bits per codeword are assigned to one M-PSK waveform. These waveforms are transmitted
sequentially over the channel every 7; = T sec.
Example 11.6
Consider a 16-ary message source (M, = 16) whose output is to be encoded using a quaternary
(Q =4) rate 2/3 code (K- = logg Mc = 2, N. = 3) and transmitted over the channel using an
M-PSK modulation. Define the mapping of the 16 possible messages mo, mj, m2, ...,™15,
which are respectively represented by the 16 possible quaternary sequences 00, 01, 02, 03, 10,
11, .. . , 33, into codewords as in Example 11.2.

(a) Coded BPSK Modulation (M = 2, N = 1)

Here the modulator is used k = N, log, Q/ log, M = (3)(2)/1 = 6 times to generate the coded
waveform S,,(t) corresponding to each codeword. That is, each of the N, = 3 coded 4-ary
symbols per codeword is converted to two coded bits [the mapping from coded symbols to
coded bits is assumed here to be binary coded decimal (BCD)] each of which is assigned to
one BPSK waveform. These waveforms are transmitted sequentially over the channel every
T, = T/6 sec. Also note that 7, = 3Ty/2 = 3T;/2. Letting so(t) = /2/T;, cos wet be assigned
to a “0” coded bit and s\(t) = —./2/T) cos wet be assigned to a “1” code bit, then the mapping
from a typical message sequence to the transmitted BPSK modulation would appear as below:

Message Sequence: m1,Mo, M3, M6, ..-

Coded Symbol Sequence: 010 000 030 302...
Coded Bit Sequence: 000100000000001 100110010
Modulator Output: so(t), So(t), so(t), s1(t), So(t), so(t),
Coded Modulation Waveform S,(t)
so(t),
a
So(t), So(t),
,
so(t), So(t),
’
So(t),
So(t)
730 Coded Digital Communications Chap. 11

so(t), So(t), si(t), si(t), So(t), So(t),

NY

S3(t)
si (t), 81(t), so(t), So(t),
SSeS ee
s1(t), So(t), ---
So(t)

(b) Coded QPSK Modulation (M = 4, N = 2)

Here the modulator is used k = N- log, Q/ log, M = (3)(2)/2 = 3 times to generate the coded
waveform S,,(t) corresponding to each codeword. That is, each of the N- = 3 coded 4-ary
symbols per codeword is directly assigned to one QPSK waveform without the need for first
converting it to coded bits. These waveforms are transmitted sequentially over the channel
every T; = T/3 sec. Also note that 7, = 3Ty/2 = 3T,/4. Letting So(t) = /2/T; cos wet be
assigned to the coded symbol “0,” 5\(t) = —./2/T; sin wet be assigned to the coded symbol
“1” S(t) = —./2/T, cos wet be assigned to the coded symbol “2,” and $3(t) = ./2/T; sin wct
be assigned to the coded symbol “3,” then the mapping from a typical message sequence to the
transmitted QPSK modulation would appear as below:

Message Sequence: m1, Mo, M3.M6, ...

Coded Symbol Sequence: 010 000 030 302...
Modulator Output: so(t), s1(t), so(t), so(t), so(t), so(t),
Coded Modulation Waveform Si (t) So(t)
So(t), 53(t), So(t), 53(t),
So(t), S2(t), ...
—_—_—_——Ja eo”
S3(t) S6(t)

(c) Coded 8-PSK Modulation (M = 8, N = 2)

Here the modulator is used k = N, log, Q/ logy M = (3)(2)/3 =2 times to generate the coded
waveform S,,(t) corresponding to each codeword. That is, each of the N. = 3 coded 4-ary
symbols per codeword is converted to two coded bits [the mapping from coded symbols to
coded bits is assumed here to be binary coded decimal (BCD)] and each three coded bits
is assigned to one 8-PSK waveform. Thus two 8-PSK waveforms are generated from each
codeword. These waveforms are transmitted sequentially over the channel every T; = T/2
sec. Also note that 7, = Ty = T;/2. Using the same coded bit to waveform mapping as in
Example 11.4, part (c), then the mapping from a typical message sequence to the transmitted
8-PSK modulation would appear as below:

Message Sequence: m1, M0, M3.mM6, ...

Coded Symbol Sequence: 010 000 030 302...
Coded Bit Sequence: 000 100 000 000 001 100 110 010...
Modulator Output: So(t), S4(t), So(t), so(t), 51 (t), s(t), S6(t), S2(t),...
—_—_——_—— ——_-_—_—_—"r— eeeeee 2 eee een SS
Coded Modulation Waveform S(t) So(t) S3(t) S6(t)

(d) Coded 16-PSK Modulation (M = 16, N = 2)

Here the modulator is used k = N- logy Q/ log,M = (3)(2)/4 = 1.5 times to generate the
coded waveform S,,(t) corresponding to each codeword. That is, each of the N- = 3 coded 4-
ary symbols per codeword is converted to two coded bits [the mapping from coded symbols to
coded bits is assumed here to be binary coded decimal (BCD)] and each four coded bits is as-
signed to one 16-PSK waveform. Thus three 16-PSK waveforms are generated from each pair
of codewords. These waveforms are transmitted sequentially over the channel every T, = 27 /3
sec. Also note that 7, = 37/4 = 3T;/8. Using the same coded bit to waveform mapping as in
Example 11.4, part (d), then the mapping from a typical message sequence to the transmitted
16-PSK modulation would appear as below:
Sec. 11.2 Coded Digital Communication System Architectures 731

Message Sequence: mM1,Mo0,M3.mM6,...

Coded Symbol Sequence: 010 000 030 302...
Coded Bit Sequence: 0001 0000 0000 0011 0011 0010...
Modulator Output: 51(t), So (t), Sot), s3(t), 53 (t), 52(t),...
——— ————— Se eee SS
Coded Modulation Waveform Si(t) So(t) S$3(t) S(t)

(e) Coded 16-QASK Modulation (M = 16, N =2)

Here the modulator is used k = N- log, Q/ logy M = (3)(2)/4 = 1.5 times to generate the
coded waveform S,,(t) corresponding to each codeword. That is, each of the N, = 3 coded 4-
ary symbols per codeword is converted to two coded bits [the mapping from coded symbols
to coded bits is assumed here to be binary coded decimal (BCD)] and each four coded bits is
assigned to one QASK waveform. Thus three QASK waveforms are generated from each pair
of codewords. These waveforms are transmitted sequentially over the channel every T, = 2T /3
sec. Also note that 7, = 3Ty/4 = 3T,/8. Using the +1, +3 coordinates of a 16-point QASK
grid in two-dimensional signal space, we let so(t) = ./2/T, cos wet — /2/T, sinw,t be as-
signed to the coded bit sequence “0000,” 51 (t) = /2/T; cos wet — 3./2/T; sin wet be assigned
to the coded bit sequence “0001,” s2(t) = 3,/2/T7; cos wet — ./2/T; sin wet be assigned to the
coded bit sequence “0010,” 53(t) = 3./2/T; cos wet — 3./2/T; sin wt be assigned to the coded
bit sequence “0011,” etc. Then, the mapping from a typical message sequence to the transmitted
16-PSK modulation would appear as below:

Message Sequence: m|,™M0,M3.mM6, ...

Coded Symbol Sequence: 010 000 030 302...
Coded Bit Sequence: 0001 0000 0000 0011 0011 0010...
Modulator Output : Sih) S0.(E)) So(b)s) 53 (bss (bE)ssp (h)arae
—_———_, —_—__—_—_—_— See

CodedModulation Waveform Si(t) So(t) S$3(t) So(t)

This mapping is identical to that in part (d) of this example. The reason for this is that the
code to waveform mapping is only dependent on the values of M and N associated with the
modulation and not on the actual form of the signal constellation, that is, the values of s,,,, in its
orthonormal representation.
Example 11.7
In this example, we consider the use of a ternary M-PSK (M = 3, N = 2) modulator in combi-
nation with binary (Q = 2) and ternary (Q = 3) rate 2/3 codes.
(a) Binary Code (Q = 2)
Since each symbol of the ternary PSK modulator would have to be assigned to logy M = 1.585
coded bits, the use of a binary code does not lend itself naturally to this sort of mapping since
the conversion of coded symbols to coded bits (in accordance with BCD) has an integer rela-
tionship.
(b) Ternary Code (Q = 3)
Here it is not necessary to convert the coded symbols to coded bits since the ternary code al-
phabet represents a natural match to the ternary modulator alphabet. In particular, each coded
symbol of the ternary codeword, regardless of the value of N,, would be assigned to one of the
three signals in the modulator of Example 11.3.

IF channel characterization. This channel transforms the coded waveform

Sm(t) into the received noise-corrupted waveform r(¢) = Sm(t) +n(t); O<t<T = Kes
As we have seen in the previous section, S(t) can be rewritten as a linear combination of
a set of N, orthonormal functions; furthermore, S,,(t) is representable by the vector S,, =
732 Coded Digital Communications Chap. 11

(Sm1+ Sm2s°**+ Smn,) Of its coefficients in this expansion for 0 < ¢ < T. The demodulation-
detection process in Fig. 11.3 reduces r(t) to its projection r = (71,72, °°, ry,); this process
is covered in Chapter 3 and this chapter. The vector r characterizes the output of the demod-
ulator (demodulation processor followed by matched filter detection). The channel decoder
then transforms r into the codeword output v = (v1, v2, *--, UK.) from the codeword dic-
tionary. The source decoder, based on the rules of source encoding, transforms v into an
estimate m of the information source output. Source decoding (data compression) tech-
niques are not covered herein.
The IF channel model consists of modeling numerous subsystems; the subsystems
which make up the model are demonstrated in levels 6 through 8 of the system architecture
shown in Fig. 1.6 of Chapter 1. For the continuous IF CMC it is characterized by the known
channel transition probability density f(r|S) and, if the channel is memoryless, then
Ne
fOISm) =| |£ralsmn), (11.14)
n=!

when message m is transmitted, that is, message m — X, — Sm. For the AWGN channel

] Ca aa
Sa
(Tn\Smn) = Hor Xp P|
f rn\S No (11S)

for all m and n. Frequently, sn» = NE Near for all n and m so that the IF channel input-
output is directly related to the baseband channel input-output via the energy Ey. used per
dimension. So much for the generic baseband and IF channel models of a coded digital
communication system. They are both referred to as CMCs.

11.2.3 Quantizing the Demodulator/Detector Outputs

Practical demodulator outputs are defined on the continuum, that is, re R%¢ where Re de-
notes the N,-dimensional Euclidean space. However, before the demodulator output can be
processed by a digital signal processor (decoder), some form of amplitude quantization on
the observable r must be performed. The quantization process is accomplished in practice
using some form of vector quantization circuit (see Fig. 11.6). In essence, the quantizing
circuit converts the CMC into the DMC which is then an approximation to the CMC. We
will discuss the details of this probability transformation shortly. For the moment however,
we note that this circuit transforms each observable r, € r from a random variable (r.v.) of
the continuous type into a discrete r.v., say R,. Strictly speaking, we will need N, quantiz-
ing circuits; however, this can be avoided by using BPSK modulation, requiring one quan-
tizer, and using the modulator and demodulator k = Nz times to create r (see Fig..11-7);
In this case, the components of r are output serially in time every Ty sec. If M-PSK or
QAM modulation is used, we need two quantizing circuits; one for the I-channel and one
for the Q-channel. In addition, one will need to compute the tan! (ro, /r2n—1) sequentially
in time for M-PSK, M > 4. The QASK or QPSK modulator and demodulator will only
need to be used k = N, /2 times to transmit the coded message bits serially in time (see
Fig. 11.7). The quantizer converts the continuous random variable r, into the discrete ran-
dom variable R,. In particular, a J-point scalar quantizer Q) is a mapping Q| : rp, > Rp
where ry is the real tine in the n™ direction of an N--dimensional vector space and R, =
Qi(rn) € {Rno, Rni, +++, Rn,s—1} in the quantizer output set of size J forn = 1,2,---, Ne.
The output values are referred to as output levels, Associated with each J-level quantizer is
a partition of the real line r, into J cells of width A; for j=0,1,---,J—1.
 Sec. 11.2 Coded Digital Communication System Architectures 733

Ue P[R, = Rael Stn

f(rAlSon)

f(rAlS4,)

eee ete oe Rno4},000

patie Soft-Decision Symbols
Confidence Symbols
Hard-Decision Symbols

Figure 11.6 Quantization of the observable r, using a uniform eight-level (3-bit)

quantizing circuit;n=1,2,..., Ne

The vector quantizer of dimension N, and size J illustrated in Fig. 11.3 is a mapping
from a vector r = (r},72,---,7n,) in N--dimensional Euclidean space Rc into a finite
set R = (Rj, R2,---, Rn.) where each component takes on one of J output levels; R, €
{Rno, Rni,**-, Rn,s—1) for all n = 1,2, -++, Nc. Thus, R = Qy,(r): RNe > R.
The design, that is, selection of the number of quantization levels, of the quantizing
circuit is an important part of the demodulator design. Assume that R, takes on one of J
values (a J = 27 level quantizer), say (Rno, Rni,---, Rny—1). For example, an 8-level (¢ =
3 bit) quantizing circuit is illustrated in Fig. 11.6; this figure demonstrates the quantization
regions, and the input-output transformation law.
In order to discuss the process further, two more concepts involved in error control
system design must be presented, viz., the concept of hard decisions and soft decisions.
From the point of view of codec design, creation of the DMC from the CMC outputs is
most important. When the demodulator outputs r, and this value is quantized into two levels
(J = 2), for all n, the demodulator is said to make hard decisions and the decoding process
is termed hard-decision decoding. When the demodulation outputs are quantized into more
than two levels (J > 2), the demodulator is said to make soft decisions and the decoding
process is called soft-decision decoding. A soft-decision demodulator is more complex to
implement when compared with a hard-decision demodulator. Soft-decision demodulators
require automatic gain control in the signal path and qg = log, J bits have to be manipulated
for every channel symbol interval of duration Ty sec. Infinite bit quantization is frequently
referred to by system engineers as no quantization, that is, perfect representation of the am-
plitude voltages associated with r is retained, no quantizer is used, and no information is
lost at this point in the decoding process. We defer until later the presentation of the details
regarding codec performance as a function of the quantizer design parameter J = 27 and the
 734 Coded Digital Communications Chap. 11

Quantizer

H Pp (t) =N2/T n COS @,t

'
(n=1)Ty stent; n =1, 2, ..., k and zero elsewhere, Ty = T,
aos

Quantizer

Quantizer

Qon-4(t) = V 2/T yy sin Wet

n-1)Ty N Stsnty; N n =1, 2, ..., k and zero elsewhere, Ty,N = T c

(b) |-Q Demodulator for QPSK, M-PSK, QASK, or QAM

Figure 11.7 Optimum demodulation of the coded message bits: (a) BPSK and ASK
demodulator, (b) I-Q demodulator for QPSK, M-PSK, QASK, or QAM

chosen coding-decoding technique. Suffice it to say, we shall show in the next chapter that 3-
bit soft-decision demodulators improve the communication efficiency by approximately 2.2
dB for coherent detection and 4.7 dB for noncoherent detection when compared with hard-
decision demodulators. On the other hand, 3-bit quantizers loose little in efficiency when
compared with the unquantized demodulator output (see Fig. 11.6). Figure 11.8 demon-
strates the DMC model for the quantizer shown in Fig. 11.6. The confidence thresholds
illustrated in Fig. 11.6 will be useful in creating erasures in Reed-Solomon decoding (see
Chapter 12).
As we shall see later, there is only a small difference in the communication perform-
ance efficiency achievable between eight-level quantization and the unquantized case. The
eight-level quantized outputs involve one decision threshold and three pairs of confidence
Sec. 11.2 Coded Digital Communication System Architectures 735

Decoder Input
7 Ta

P 6 110
My, =0-® So, = VEN ¢ Cee 5 pomdonce
101 | Thresholds
4 100
Hard-Decision
3 011 Threshold
Via
mM, = 1->5,, = -VEy LN 010 Confidence
c BN 2
Thresholds
1 001
Channel Transition Probabilities:
LaF SR a ee ee oO 0 000
ee Od OS a Vay, Confidence Bits
qe OOn ls Nat, 2,033; KeeiN. Hard-Decision Bit

Figure 11.8 DMC for BPSK demodulation using 8-ary (3-bit) soft-decision
demodulation (for QPSK one has an I-channel and a Q-channel; however,
Pale) er IN-/D)

thresholds for the (M = 2, N = 1)‘BPSK Modulator-Demodulator of a (N,, K-) code; this

3-bit quantization process is illustrated in Fig. 11.8.

QAM or QASK. To accommodate multilevel codes, that is, the case where the coded mes-
sage bits x, can take on more than two values, quadrature amplitude modulation (QAM)
can be used and the coefficients s,,, can be adjusted to give the appropriate mapping. The
telephone channel is a good example of this. Here the data rate requirement is 9,600 bits/sec
and the modulator can support the transmission of 2,400 symbols/sec (bauds). Without
channel coding, a 16-point QAM modulator (Chapter 10) can be used to accommodate
the data rate. With coding and a code of rate r,=4/5, a 32-point signal structure will sup-
port the 9,600 baud data rate. In practice, the CCITT V.32 convolutional encoder is used in
combination with a QAM modulator. We consider the transmission of multilevel codes in
Chapter 12.

11.2.4 Creating the DMC for Channel Decoding

Now the channel transition probabilities {P|Rn = Rnj\m] for 7 =0,2,---,J —1;m=0
or 1; n=1,2,---, N-} which define the IF DMC are found from the IF CMC transition
probability densities f (rn|Smn) defined in (11.15). In particular, using (11.15), we have

P([Rn = Rnj|Smn) = f (Tn|Smn)arn A P(j|m) (11.16)

meQ;

for = 0) 1,5)-5 J — 13m = 0 orl; n=T1,2,---,-Ne. Here Q; is the j'" quantization

interval of width A (see Fig. 11.4). Selection of the design values for A and J will be
considered in the next chapter.
The simplest DMC which can be created using the demodulator output is the bi-
nary symmetric channel (BSC) for which the coded message bits are transmitted using
BPSK, with J =2 (hard-decision demodulation). We only need to consider quantizing
the observable r, 4 r into R, 4 R for all n where R =0 or 1. For the BSC with inputs
 736 Coded Digital Communications Chap. 11

m =0;S = VEn R=0

m=1;S,=-VEy R=1

(a) BSC Model

P[0|0] R=0
m = 0; S9 = VE

P(0|1] <> R=1

m=1;8,;=-VEy_ a

a4 Figure 11.9 BSC and DMC model for hard

and J-ary soft-decision demodulation of
BPSK coded message bits: (a) BSC model,
(b) DMC Model (b) DMC model

mo =0—> so=./En, or mj = 1 > 5; = —/Ey,, for all coded message bits, the channel
transition probabilities are given respectively by
0
Pir <0\sol=f_f(rlsopdr = p
—0Oo
(11.17)
CO

Pir > ols f f(r|sydr = p (11.18)

0
where p = Q (/ 2En,/No). Figure 11.9 demonstrates the BSC channel model and the J-
ary DMC for M = 2; this channel is used k = N, times per message transmission. These
two quantized channel models of the IF channel of Fig. 11.2 are of greatest interest in what
follows.
When the physical channel is not memoryless, it can usually be made so by adding an
interleaver-deinterleaver system into the information flow path. Interleaving-deinterleaving
techniques are discussed next.

11.3 INTERLEAVING-DEINTERLEAVING TO MITIGATE CHANNEL MEMORY

We have completed our discussion regarding the problem of quantizing the demodulator
output so as to create the DMC needed to accomplish the channel decoding function using
digital hardware. In doing so, we tacitly assumed that the channel is memoryless, that is,
the received channel symbols are independently contaminated by noise. In practice, many
channels exist such that this is not the case; noise bursts are common. Examples include
Sec. 11.3 Interleaving-Deinterleaving to Mitigate Channel Memory 737

multipath, interference, jammed, and fading channels. Telephone channels typically have
impulses on the line which last for approximately 10 ms.
We now address techniques whereby channels with memory can be engineered such
that they appear to the channel decoder as memoryless. Interleaving or scrambling is the
common technique used to make the channel appear to the decoder as memoryless when
burst noise is present. In practice, the system designer begins by specifying some maximum
noise burst length which the system must tolerate—the so-called channel memory time.
Once this is determined, the design of the interleaving system can proceed.
To mitigate channel memory, thereby taking full advantage of coding, an interleaver-
deinterleaver (I-D) system is inserted into the signal path. We will discuss the placement or
logical position of the I-D system in the functional system architecture in the next section.
For the moment, we present various interleaver types and discuss the trade-offs between
their performance measures of interleaving depth (span), interleaving (channel) delay, and
interleaving memory size. The I-D system is critical to being able to realize the performance
gain provided by coding in channels with memory. The function of the I-D system is to ef-
fectively reduce the degradation in bit error probability by randomizing the distribution of
the burst errors (or correlated bits) caused by channels with memory. The interleaving proc-
ess serves to scramble the order of the code symbols x,,, in time while the deinterleaving
process unscrambles the recovered symbol stream into the original sequence for application
to the decoder. Error bursts which would overwhelm the decoder are spread out by the dein-
terleaver over a number of channel symbols longer then the duration of the error burst. Thus
for purposes of error control, the burst or memory characteristics of the channel must be
specified.
There are at least five I-D system techniques of practical interest. They include: (1)
the periodic block I-D system, (2) the helical I-D system, (3) the periodic convolutional I-D
system, (4) the pseudorandom I-D system, and (5) the hybrid I-D system (see [2], [4], [6],
[7], and [8]).

11.3.1 Periodic Block Interleavers-Deinterleavers

Figure 11.10 depicts a block of length L = 4 symbols which is block interleaved to depth
D = 3. In essence, an (L, D) periodic block interleaver is characterized by the block inter-
leaver length L and interleaver depth D. Symbols moving across the channel are obtained
by scanning rows from left to right; symbols leaving the channel encoder, or entering the
decoder, are obtained by scanning succesive columns from top to bottom. The codewords
occupy the positions labeled ABCD, EFGH, IJKL, ... . The transmitted code symbols are
AEI, BFJ, CGK, DHL, and so on. Suppose a burst of noise “wipes out” the transmitted code
symbols AEI, and converts these into XXX (see Fig. 11.10) . The decoder input consists of
(XBCD, XFGH, XJKL). A single error correcting code which corrects one error will allevi-
ate this error. Also a single error detecting code will detect but not correct this error. Thus,
an interleaving depth of D units means that a noise burst must have length (D + 1)7y sec
in order to strike any codeword more than once.
Figure 11.11 demonstrates a block of length L symbols which is block-interleaved
to a depth of D units of time. We follow the convention that the block symbols x;; to
be transmitted over the channel are read into the interleaver columns and then transmitted
over the channel row by row from left to right. On the receiver side, the action is reversed.
The deinterleaver inputs are the rows of the interleaver output. The rows are read into the
738 Coded Digital Communications Chap. 11

ie vas
TIME Noise Burst

| = | b ae?) See.4

DEFAY]\— Sian eened BE rere

L=4
| Crain Gig nats Coe. Gaak

De oahloene | Diorieeee le

(a) Interleaver Action (6) Channel Action

TIME
eld eee
| A. BaG.D xX BC D
DELAY
pes Eye Geek Xe a Gar

| lena Kane be Xo> JiswKeaak

Burst Noise Present
(c) Deinterleaver Action (d) Deinterleaver Action in Burst Noise

Figure 11.10 Block I-D system of length L = 4 and depth D = 3: (a) Interleaver action,
(b) Channel action, (c) Deinterleaver action, (d) Deinterleaver action in burst noise

columns of the deinterleaver and are read out of the deinterleaver (decoder inputs) by rows
as shown in Fig. 11.10(b). Thus, the D x L deinterleaving matrix is the transpose of the
L x D interleaving matrix.
In practice, the encoder output symbols are written into L vertically contiguous mem-
ory cells in a column and the outgoing symbols are read from D horizontally contiguous
memory cells in a row. The interleaver span is a measure of the time across which pre-
viously contiguous data is spread. It serves to characterize the ability of the interleaver to
randomize burst errors or spread out adjacent symbols over time intervals greater than the
noise burst durations. The span of the block interleaver is given by

CD NSEC (11.19)

where Ty is the time duration of each channel code symbol. The interleaving (channel)
delay introduced into the recovered code symbols is

Tq = ZL DTN sec (11.20)

whereas the required memory at the transmit and receive end is M, = 2LD symbols for
hard-decision demodulator/decoders. ;

11.3.2 The Helical Interleaver

Figure 11.12 demonstrates a helical interleaver for a block of length L = 4 with a depth
D = 3. The codewords are seen from this figure to be ABfa, CDbc, and EFde. As a function
of time the output of the encoder is ABfaCDbcEFdeABfa, the output of the interleaver
(channel input) is abFAcdBCefDEabFA, and the input to the decoder (deinterleaver output)
Sec. 11.3 Interleaving-Deinterleaving to Mitigate Channel Memory 739

Pee D
Ke Aeine PeX.1\4|
ieee D
Delay, L Mo eAL ees Xp {ta

1 2 D
DE OL, (ees) Ah

(a) Interleaver Delay vs. Time

Time
=|

X; 1 Xo 1 eee XL 1

2 2 2
Atte Sbs2 * 2° Xr
Delay, D
e e

X(p-1)L +1
D e e e
Xo
D

Figure 11.11 Periodic block interleaver-

d y deinterleaver: (a) Interleaver delay versus
(b) Deinterleaver Delay vs. Time time, (b) Deinterleaver delay versus time

is the same as the encoder output; no errors are assumed. The total interleaving delay
is seen to be 4 x 3 = 12 characters. If a burst of noise in the channel were to convert
the transmitted codeword abFA into the word XXXX, then the deinterleaver action would
scramble the X symbols such that the input to the decoder would be the codeword sequence
ABfXCDXcEXdeABfxX. Again a single error correcting code would be able to correct these
errors or a single error detecting code would be able to detect the errors.
Figure 11.13 depicts a code of length 8 which is helical-interleaved to a depth of 7.
This means that a noise burst must have a length of at least 8 in order to hit any code block
more than once. Each codeword lies in a column and the symbols are fed into the interleaver
by columns and then read out by rows. A block interleaver requires the receiver to know the
synchronization modulo LD; however, the helical interleaver requires the receiver to know
synchronization modulo L. In general, a code of length L can be helical-interleaved to a
depth of L — 1. The helical-interleaver requires a memory size about one-half the size of
the corresponding block interleaver of the same depth.

11.3.3 The Periodic Convolutional Interleaver-Deinterleaver

Variable delay is the essence of interleaving. Hence, it is natural to view an interleaver as a

collection of delay lines of various lengths with multiplexers shuffling data through suc-
cessively different routes. An (L, D) periodic convolutional I-D system which uses this
implementation concept is shown in Fig. 11.14 for the case L = 30, D = 116. When the
740 Coded Digital Communications Chap. 11

—Channel Time—>

re
A Cc d
faa orl
B Cc e
wae
f D E

la b F
Sars
Interleaver A Cc d
Time oar
B Cc e
ie eet
f D Ee

la b F
ayal
A Cc d
Sanam
B Cc e Figure 11.12 A (4,3) helical interleaver

delay lines are tens of thousands of symbols (bits) long, then they are most effectively imple-
mented with random access memory (RAM). We note from Fig. 11.14 that the convolutional
interleaver has a staircase geometry, as opposed to the rectangular geometry of the block
type. This staircase geometry may be thought of as a series of shift registers of graduated
lengths, as indicated in Fig. 11.14. Actual implementations are more likely to use a random
access memory (RAM) and stepped address pointers as opposed to shift registers.
The shift register model facilitates discussion of the operation. In contrast to the block
case where the data was stationary and a two-dimensional commutator was used, a one-
dimensional commutator is used in the convolutional interleaver and the data in each row of
the interleaver and deinterleaver is periodically shifted to the right. The commutator at the
interleaver input is stepped row by row from top to bottom inserting one bit in each row.
After the commutators reach the bottom position, each data bit in every row is shifted one
bit toward the output, and the commutators are reset to the lowest row. The output or last
bit in each row is overwritten after it is read. In contrast with the block interleaver, there is
no conflict between input and output, so long as both input and output precede shifting on a
consistent basis for all rows.
The convolutional interleaver has several advantages over the block interleaver. The
memory and delay required are lower and it lends itself to rate buffering of the data. When
inserted in the encoded symbol sequence, the system guarantees separation (of any two
symbols within LTy sec of each other in the interleaver symbol sequence) to be at least
[LD/(D — 1)]Tw sec between each other in the deinterleaved sequence. Thus, the inter-
leaver span is given by

t; =[LD/(D — 1)]Ty sec (11.21)

and the interleaving delay is tq = 2t; sec. The memory required for the convolutional
interleaver is the sum of the row lengths. For a typical interleaver, it becomes
 Sec. 11.4 Synchronization of I-D Systems 741

Channel Time ———>

Pomme CeaaOar: iat WAZ

et e f g h Xx >

Boe eek eto

Cre i Oe ee One D

oe M| q r Ss

eat Pies Q | u Vv

Wel On RS, | U4) y

Interleaver 15 SO Hl
ee eaten? iGo g Tr WZ
rch si ooleng
: as‘2 ee
axe
EOE a ee Ores

eS as “ erty

kat Q | u,v

Vener mMoovA: a U|

z 6 ® P S V_ YI Figure 11.13 A(8,7) helical interleaver

Mi DAE ne 1) (ie22)
biel
characters or symbols.

11.3.4 Other Interleavers-Deinterleavers

McEliece [2] discusses pseudo-random interleavers; it is also possible to consider concate-

nated I-D systems. Berlekamp [7] discusses concatenating block and helical interleavers to
obtain deep-staggered and shallow-staggered |-D systems; the description and operation
of such systems are outside the scope of our presentation. The hybrid interleaver [4] was
perceived for use in coded systems employing information packets.

11.4 SYNCHRONIZATION OF I-D SYSTEMS

Coded digital communication systems that employ interleaving require a level of synchro-
nization to operate the I-D commutators. In particular, decoders for block codes need to
know when codewords begin and end; convolutional decoders need to know when the nodes
in the code tree begin and end. It is also necessary to ensure that the receiver’s deinterleaver
is synchronized with the transmit side interleaver. For the most part, the methodologies of
 742 Coded Digital Communications Chap. 11

Endoder Symbols or a ae : To Channel

ee (4)
Coded Message Bits

Cover . P
Sequence

oe ee

Note: D = Delay Element

o—fienX{}20
o—frexgC +0
108D 2 @
Interleaved O
Encoder To Channel
—
Symbols © 104D 30 Decoder
from E | 1040 _| ;

Channel

" f Cover
O 286 Sequence

29

Note: D = Delay Element; The cover sequence can be used to provide data privacy.

Figure 11.14 Periodic convolutional interleaver-deinterleaver

synchronization and resynchronization of I-D systems which accept hard or soft decisions
from the demodulator largely remain the trade secrets of organizations who design and build
error control systems.

11.5 CODED SYSTEM ARCHITECTURE FOR CHANNELS WITH MEMORY

As we have just discussed, the common strategy for eliminating the deleterious effects of
noise bursts on coded system performance is to introduce into the information flow path an
interleaver-deinterleaver system. Figure 11.15 shows the logical position of interleavers in
 Sec. 11.5 Coded System Architecture for Channels with Memory 743

Information Source Channel

Source Encoder Codet Interleaver Modulator

: Transmit :
Local
)
Clock
Distribution Oscillator

Frequency and Timing System IF

: Channel
Receive
Clock Local
Distribution Oscillator

oleh Seacuns oe es Deinterleaver K@-| Demodulator

ecoder

Figure 11.15 Coded-digital communication system architecture for channels with memory

the system architecture. The interleaver lies between the channel encoder and the modulator;
the deinterleaver is inserted between the demodulator output and the channel decoder. Both
units frequently require substantial memory.
Suffice it to say, block decoders tend to be less complex when driven by hard deci-
sions from the demodulator; convolutional decoders can perform using either hard- or soft-
decision demodulator outputs. In order to gain the advantages of both error control meth-
ods in the design of an overall error control system, a convolutional decoder is driven by
the soft decisions from quantized demodulator outputs. This decoder then outputs hard de-
cisions to a block decoder for further error reduction. Combining the attributes of block and
convolutional codes into cascaded codes is known as concatenated coding.
Concatenated coding, introduced by Forney [5], represents an alternate technique to
mitigate noise bursts when channel bandwidth is not constrained. Figure 11.16 demonstrates
the architecture of a system that employs an inner code, an outer code, and interleavers and
deinterleavers designed to randomize the location of the burst errors in the symbol stream
seen by the channel decoders.
Typically, the inner code, that is, the code that interfaces with the channel, is designed
to correct most of the channel errors while the higher rate outer code reduces the error rate to
the desired level. The purpose of concatenated coding is to obtain a small error probability
with an end-to-end coded system implementation complexity which is less than that which
would be required by a single codec. Such coding techniques could also be applied to
improve the error probability performance of an existing coding system. One of the most
successful concatenated coded systems is one that employs a convolutional inner code and
a Reed-Solomon outer code with interleaving-deinterleaving [8]. The inner codec can take
maximum advantage of soft decisions from the demodulator output while the outer Reed-
Solomon codec can easily deal with the hard-decision output from the inner convolutional
codec.
 744 Coded Digital Communications Chap. 11

In many cases, the analog signals being processed in the front end of the receiver or
the confidence symbols shown in Fig. 11.6 can be used to generate a “fade detector” or
“erasure indicator” signal (shown dotted in Fig. 11.16). When an error-correcting code is
used as the outer code, this erasure signal may enable the decoder to correct more severe
noise bursts than it could if there were no such control signal. We will discuss the error
and erasure correction capacity for RS codes in the next chapter. This will be followed by
presenting the performance of certain convolutional and concatenated coding systems.

11.5.1 Coding and Modulation Matching for Concatenated Coding Systems

Consider the binary concatenated coding-modulation system depicted in Fig. 11.17 which
cascades the binary codes (Nj, Kei) of code rates rej with M;j = QXci for all i = 1, 2,---,
L. At each encoder output, we must have ( let T;1 = Tp)

T= Kila = Kolo =-c2= Karlin see (11.23)

where 7,; are the duration of the coded bits at the output of each coder. Now the signal
generator must map N-, coded bits of the L'” coder output to one of the M available
modulation waveforms (signals). Then

T = (logy M)Te (11.24)

and since the channel rate 7? = (log, M)/T;, we can use (11.23) and (11.24) to write
L
T = (logy M)Tp || rei (11.25)
t=]

so that the channel rate becomes R = Ry/re bits/sec where

L
le I> ee (11.26)
i=1

is the equivalent concatenated code rate. The bandwidth expansion of the coded system over
that of an uncoded system is r;!, that is
Ren al
—=r;! (Lean
Rp Ve
A BPSK modulator can be used to transmit the codewords one coded bit at a time, while
demultiplexing the coded message bits; a QPSK modulator can transmit them two at a
time, one in the I-channel and one in the Q-channel (see Fig. 11.5). In the next chapter,
we investigate block-coded digital communication techniques; this is followed by looking
at convolutional coded digital communication techniques.

11.6 CLASSIFICATION OF ERROR CORRECTION AND DETECTION CODES

For our purposes, the class of error correction/detection codes to be considered in the next
two chapters can be broken into two types—tree codes and block codes. Both code types
can be further classified as linear or nonlinear depending on the hardware mechanization
required to generate them. Herein we are concerned with the class of linear codes. Fig-
ure 11.18 depicts the Venn diagram for error correction/detection codes of interest. The
major characteristic of linear codes is that they have a block or a tree architecture with an
optional systematic versus nonsystematic structure. As seen from Fig. 11.18, cyclic codes
 UONEWIOJU| 301nN0S 121NO Jauu| JauU|
S sojino poou3
J9poou ‘Sa
J9po : yaqul JOACS]IJ9]U poo4J9po lic oe
JBAC9|J9}U

| yiWSuUeI
je007
490/D,
J0}BII9SO
uoNNqisiqg

i Aouanbe4
pue Bulwiy Waj}skS I
jouueyD
: BAISOOY
49019 [2007
uoRNgISsIG JOJeIIIOSO

eee
SP aeaneeseeeconeecesseceucnoeoeseecescssecenasonsesenseeey

eeecesoeeees soneccononse
errrry Coe

UOHEWO}U| 80iNOS ed 431NO JOUU] Jeu} powg0re

HUIS Japo08q Japooeq JOACBLBJUIAG Japooeq JaAeBLB]UIBG

YOS sudIsiog9q

et ferree
Secuscereceecceveeeuconensency
pocuececoncvncceoGe

aans1q
TT] gy
= Popoo-payeusjwou0
[eIISIP D UONROIUNUTWIOD
WII\sKs aM{oaqTYOIe
JO} sauULYO
YIM AJOUIOUL

745
746
18p0D
Z 48p09
7 Joyeijnpo;
9 — wg uoneing
: (yy) Pry) (Py ‘W) (S1'N
ae eeeeeeeeeasevece.

aanB1gLTT] payeuawwouog
SuIpoos
pue UoNe[Npow
waisXs Jopow
 Sec. 11.6 Classification of Error Correction and Detection Codes 747

TREE ERROR CORRECTION CODES

CONVOLUTIONAL CODES

BLOCK ERROR CORRECTION CODES

LINEAR BLOCK CODES Orthogonal, Transorthogonal, Biorthogonal, L-orthogonal
CYCLIC CODES
PRIMITIVE, NON-PRIMITIVE, Golay
q-ARY BCH CODES Codes q-ARY BCH CODES Code

NONBINARY BCH CODES NONBINARY BCH CODES

REED-SOLOMON CODES REED-SOLOMON CODES

REED-SOLOMON CODES REED-SOLOMON CODES

OVER GF(2m) OVER GF(2™)

Figure 11.18 Venn diagram for error-correction codes

are a subset of linear block codes [9, 10]. The text book by MacWilliams and Sloane [10]
probably contains the most comprehensive discussion of block codes to date while Michel-
son and Levesque [11] gives a clear discussion of the details of codec design using block or
convolutional codes.
Bose, Chaudhuri, and Hocquenghem (BCH) codes are a subset of cyclic codes. BCH
codes can be classified as P-ary or binary (P = 2). P-ary BCH codes have codeword
symbols and codeword generator polynomial coefficients which come from Galois field
(GF) algebra, viz., GF(P) where P is prime. BCH codeword generator roots are from
GF(q)=GF(P”) where gq is the order of the field and m is an integer greater than one. Binary
BCH codes are most often used in practice; they are from GF(2). The codeword generators
of binary BCH codes are nonbinary; they come from GF(g) = GF(P™) = GF(2”). These
codes have inherent error detection capability without sacrificing any of the error-correction
capability. Berlekamp [12], Lin [13], Lin and Costello [14], Peterson [15], and Peterson
and Weldon [16] treat the Galois field algebra of linear block codes and present numerous
decoding algorithms. This important material lies outside the scope of this book.
Reed-Solomon (R-S) codes are special cases of q-ary BCH codes (see Fig. 11.18).
The code symbols of a binary R-S code are also nonbinary; they are generated from GF(q)
= GF(P”). R-S codes are very powerful burst and random error correcting codes. An ex-
cellent tutorial on R-S error correcting coding can be found in Geisel [17]. Discussions of
current applications of codes and codec design can be found in Odenwalder [18], Berlekamp
[19], and Berlekamp, Peile, and Pope [20]. Odenwalder [18] presents a clean presentation
from a system engineering and design point of view. Finally, Chapter 12 covers block coded
communication system design and performance, while Chapter 13 presents the use of con-
volutional codes and their performance.
 748 Coded Digital Communications Chap. 11

PROBLEMS
11.1 (a) For the one-bit quantizer characteristic illustrated in Fig. P11.1, characterize the channel
transition probabilities when the input channel symbols are given by so = +/En, and
5; = —,/Ey, with equal probability. Assume additive white Gaussian noise and BPSK
signaling.
(b) Sketch the DMC (binary input—binary output) model and label all input to output connec-
tions.
11.2 Repeat Problem 11.1 for the quantizer characteristic illustrated in Fig. P11.2. The DMC model
for this case represents the binary erasure channel (BEC), discussed in Chapter 3. Here, M = 2
and = 3:
11.3 We are given the demodulator output random variable r to quantize. The pdf for r is given by

Apap a4
f(p)= ,Atp —-l<p<0
0 otherwise
Assuming a 2-level (1-bit) quantizer for this input signal with output levels R; = 1/2 and
Ry =—1/2
(a) Find the constant A.

Figure P11.1 Quantizer characteristic

required to create the BSC

Figure P11.2 Quantizer characteristic

required to create the BEC
Problems 749

(b) Find the pdf f.(o) of the quantizer error e(r) = Q(r) —r= sgn(r), where sgn(r) = 1 if
r >Oand sgn(r) = —1 ifr <0.
(c) Find the distortion D = E{e?(r)}.
(d) Now redesign the quantizer so that R; = 1/3 and Ro = —1/3. Find the new distortion
D = E{e*(r)}. Compare this to the result obtained in (c).
11.4 The span of an interleaver is a measure of the time duration over which previously continuous
bits are spread, and is a metric of the ability of the interleaver to randomize burst errors. For the
(10, 20) block interleaver, determine the span for symbol rates of 2.4, 4.8, and 9.6 kb/sec.
11.5 (a) For the block interleaver, show that the memory required at the transmitter or receiver is
given by

M, =2LDAq
where Ag = 1 for block interleavers and Ag = 1 for deinterleavers with Demodulators
using hard decisions and Ag = 3 for deinterleavers with Demodulators using 3-bit soft
decisions.
(b) For the (10, 20) block interleaver, evaluate the required transmitter and receiver memory
for both hard- and soft-decision Demodulation.
(c) Repeat (b) for the (39, 116) interleaver.
11.6 (a) Given a coded symbol rate of R = 1/Ty = 19.2 kb/sec, find the span for a block inter-
leaver of length L = 8 and depth D = 8.
(b) Determine the channel delay ty.
(c) Determine the required memory size M, for hard-decision demodulators.
(d) Using Fig. 11.11 and the numbers 1, 2, 3, ... to fill the interleaver input rows, demonstrate
the interleaver output matrix and the deinterleaver output matrix.
11.7 Repeat Problem 11.6 assuming a convolutional interleaver.
11.8 Show that the memory M, required for the convolutional interleaver is given by

Kab + De
a Be -1)

where Aq is defined in Problem 11.5.

11.9 Repeat Problem 11.5 (b) and (c) for a convolutional interleaver-deinterleaver system. Compare
the results found in Problem 11.5 with those obtained for the convolutional I-D system.
11.10 For the (4, 3) helical interleaver illustrated in Fig. 11.12
(a) Specify the symbol sequence versus time across the channel.
(b) Specify the symbol sequence versus time out of the decoder assuming no channel errors.
(c) Repeat (a) and (b) for the (8, 7) helical interleaver specified in Fig. 11.13.
11.11 (a) Assuming coded BPSK modulation, develop the channel transition probability mappings
for the hard- and soft-decision demodulator thresholds illustrated in Fig. P11.11.
(b) For each of the scalar quantization methods, illustrate the equivalent DMC model for use
in the decoding operation.
11.12 (a) Given a noise burst of 20 msec duration, design a block I-D system to accommodate a
bit rate of 9, 600 bits/sec such that all errors into the decoder due to this burst noise are
independent.
(b) What is the interleaver span, delay, and memory required?
(c) Repeat (a) and (b) for a convolutional interleaver.
 750 Coded Digital Communications Chap. 11

no
fr |
So) f(r \s,) quantization

5 ee ie

oe | ae Ey
matched filter output y
een) See hard
quantization
<q——_—_——- 0) 1 2 ———— 2-bit
quantization
1

<—__—_ 0 1°22, [23% A|)5 8 |05 ee eae

quantization

Figure P11.11 Soft decision demodulator thresholds

REFERENCES
1. Shannon, C. E. and W. Weaver, The Mathematical Theory of Communication, Urbana, IL: Uni-
versity of Illinois Press, 1949.
2. McEliece, R. J., “Communication in the Presence of Jamming—An Information-Theoretic Ap-
proach,” Chapter 3 in Secure Digital Communications, ed. G. Longo, Wien, NY: Springer-Verlag,
1983.
3. Berlekamp, E. R., Algebraic Coding Theory, New York: McGraw-Hill, 1968.
4. Olson, D., “Hybrid Interleaver,’ International Télemetering Conference, vol. 19, San Diego, CA,
pp. 703-04, 1983.
5. Forney, G. D., “Burst-Correcting Codes for the Classic Bursty Channel,’ JEEE Transactions on
Communication Technology, vol. 19, pp. 772-81, 1971.
6. Ramsey, J. L., “Realization of Optimum Interleavers,” JEEE Transactions on Information Theory,
vol. IT-16, pp. 338-45, 1970.
7. Berlekamp, E. R. et. al., Interleaved Coding for Bursty Channels, Cyclotomics Technical Report,
April 12, 1983.
8. Berlekamp, E. R. and J. L. Ramsey, “Readable Erasures Improve the Performance of Reed-
Solomon Codes,” JEEE Transactions on Information Theory, vol. IT-24, pp. 632-33, 1978.
9. Clark, Jr., G. C. and J. B. Cain, Error-Correction Coding for Digital Communications, New York,
NY: Plenum Press, 1981.
10, MacWilliams, F. J. and N. J. A. Sloane, The Theory of Error-Correcting Codes, Amsterdam:
North Holland, 1977.
11. Michelson, A. M. and A. H. Levesque, Error-Control Techniques For Digital Communication,
New York: John Wiley and Sons, 1985.
12. Berlekamp, E. R., Algebraic Coding Theory, New York: McGraw-Hill, 1968.
References 751

13: Lin, S., An Introduction to Error-Correcting Codes, Englewood Cliffs, N.J.: Prentice Hall, 1970.
14. Lin, S. and D. J. Costello, Error Control Coding: Fundamentals and Applications, Englewood
Cliffs, N.J.: Prentice Hall, 1983.
15. Peterson, W. W., Error-Correcting Codes, New York: MIT Press, 1961.
16. Peterson, W. W. and E. J. Weldon, Error-Correcting Codes, New York: MIT Press, 1972.
iN/E Geisel, W. A., “NASA Technical Memorandum 102162,’ Houston, TX: Lyndon B. Johnson
Space Center, June 1990.
18. Bartee, T. C., Data Communications, Networks, and Systems, Chapter 10 by J. P. Odenwalder,
IN: Howard W. Sams & Co., 1985.
19) Berlekamp, E. R., “The Technology of Error-Correcting Codes,” Proceedings of the IEEE, vol.
68, no. 5, pp. 564-93, May 1980.
20. Berlekamp, E. R., R. E. Peile and S. P. Pope, “The Application of Error Control to Communica-
tions,’ IEEE Communications Magazine, vol. 25, no. 4, pp. 44-57, April 1987.
 Chapter 12
Block-Coded Digital
Communications

In previous chapters, we have optimized the design of a digital communication system from
the viewpoint of the receiver; that is, we have determined optimum receiver structures and
evaluated their performance for an arbitrary set of modulation waveforms which represent
the messages in the channel. The following two chapters are primarily concerned with the
problem of optimizing the transmitter design, that is, determining the set of coded modula-
tion waveforms that minimizes the bit error probability. Optimizing the transmitter design
for a given channel characterization represents the signal design or coding and modulation
selection problem. We shall equivalently pose this problem as one of finding the best codec
and modem design which minimizes the bit error probability when the receiver used is opti-
mum. This leads us to presenting the Shannon capacity theory.
In Chapter 11, the functional architecture of a coded digital communication system
was presented. In this chapter, we shall exercise this model by developing the performance
of block-coded digital communication systems. The “Ro-criterion” for modem and codec
design will be introduced as the methodology for separating the design of the codec from
that of the modem; hence, the material presented in Chapters 3 through 10 can be used in
the modulation-demodulation design process for uncoded systems, while the theory given
in Chapters 11 through 13 can be used in the design of the coder-decoder (codec) and
Modulator-Demodulator (modem).
The analysis revolves around the fundamental parameter called the cutoff rate, Ro (“R
zero”). As we shall see, Ro (in bits/dimension) is a Shannon capacity-like quantity defined
for any CMC or DMC, whose value is less than the Shannon capacity, Cy., by approxi-
mately 3 dB in the energy-efficient region where coding is of interest. From a system engi-
neer’s perspective, Ro sets the practical limit on the rate at which information can be reliably
transmitted through the channel. Apparently, Wozencraft and Kennedy [1] were the first to
propose the use of the Ro criterion for modulation-demodulation system design; however, as
pointed out later by Massey [2], their proposal fell on deaf ears. They observed that a mo-
dem for use in a coded system should be designed to achieve the highest possible value of
Ro rather than the lowest value of post-demodulation bit error probability for the compara-
ble uncoded system. In this context, Ro is optimized not only with respect to the probability
distribution of data symbols entering the channel and the modulation-demodulation choice,
but also with respect to the quantization boundaries which define the decision regions at the
quantized demodulator output; that is, for a given number, J, of quantized demodulator out-
 Sec. 12.1 Block-Coded Digital Communication System Architecture 753
Saas MIN eerie mice caSraco anos no cnucb seus soz aM Nosy aves zcaiohicoasansonXenasusueninsssnyshnyrohaqnscnscohreansatnasonsSuesneseascencosstassovsc«qbad
snaseaecouassdustessvbivexsokesbtassogncoseoterivonevastertses cavenessusvedseestoveg

Encoder
VEC IN 17)

Saks nusnuosbsauadpaneabuahsaabeaseuussaasdear? Suupaunsisasnuseseseobsenssdussvessueesusbeas poarch nsecuasnnissnseachands>suatersoveus seuasnuuasusesandsaaeaod

ioysergovecseresvosesieevaseaserasv
sneaveveeedorevnnsensesensesssasaneesseesnenseunnesssssssaseneesseneeeesensenesseeseHPaHOsOSEOOSOSSSGRN SESSLER OE SESESEOOSEESSAEONPSSSSAOALSSSSDNORO
eoueestouoesoeessesaveseusscuscseeu SSSUAOROOESESSENSSOESESeReSeeSDESOS
ssarebergoseszecossacceescasoesavoa csessectE

T =T,log,M, =N.T.=kT,; T, =T.log, M/log,Q= NTy

de Vlog. log, O/ N. N, log, Q
= =—|—— Jlog, Q=——— |S |= —
Tt 3 t0g5,M Ry XN log, M

Figure 12.la Block-coded digital communication system model: Coded modulator

model for coded system design

puts (quantization intervals—see Chapter 11), the decision boundaries will be chosen so as
to maximize the value of Ro subject to a hardware complexity constraint. This criterion is
also applicable when convolutional and/or concatenated codes are used for error control. In
fact, in the development and analysis of powerful decoding algorithms for long constraint-
length convolutional codes, Ro has been given a more descriptive name—the computational
cutoff rate, sometimes denoted Rcomp. We proceed by developing an analytic understand-
ing of this important parameter with respect to the coding and modulation process.

12.1 BLOCK-CODED DIGITAL COMMUNICATION SYSTEM ARCHITECTURE

The functional architecture of a block-coded communication system is illustrated in Fig.
12.1. System engineers accept the perspective that coding and modulation are both compo-
nents of the signal design problem and that demodulation and decoding are likewise com-
ponents of the signal detection problem. The (M., Nc, T-) encoder, mapper, and (M, N, T;)
modulator of Fig. 12.1a work with a binary alphabet as follows: every T sec logy M. mes-
sage bits are encoded into N, coded message bits, and every T; = T/ k sec, the mapper maps
logy M of these N, coded message bits onto the M-ary modulator waveform s,,(t); thus,
bey JM — kT, sec, the coded modulation waveform S(t),m=0,...,M-—-—1,0<t<T,
contains log, M, bits of information. This process is functionally illustrated in Fig. 12.1a
and described in detail in Chapter 11. Numerous coded modulation examples using BPSK,
QPSK, QASK, QAM, and so on, are given in Chapter 11. See Tables 11.2, 11.3, 11.4, and
11.5 for the details of S,,(t).
Block-coded communications proceed by taking blocks Um» = (Um1, --., UmK,) Of Ke
source output symbols and presenting these to the channel encoder as one of M¢- possi-
ble messages. Here M. = Q*« and Q represents the alphabet size of the encoder input;
that is, wu; € u can take on one of Q values while u can take on one of QO¥*e values. The
baseband channel codeword vectors {Xm = (Xm1, Xm2,---,XmN,),m=0,..., Me — 1} are
then mapped, one-to-one, by the modulator onto the IF coded modulation vectors {S, =
(Smi, Sm2) +++» SmN,)} OF, equivalently, the IF channel input waveform S,,(t). For purposes
754
Jepoouz Jo}e|NPOW ge Joyejnpoweg Jepooeq
|eUBIS) /uolyesOUSH
: jauueug
> /uoie;npowsq)
:
(uole;npoyy : jeuBis (uoloa}eq

aan3rq qI'ZT papoo-yoo[g

[eIISIP UONeoTUNUIUIOD
wW9}sXs :Japow wayshg [opoul
40}N = YT= °N
Sec. 12.1 Block-Coded Digital Communication System Architecture 755

of being precise, in this section we assume a Modulator which produces the coded modula-
tion waveform
Ne
Sm(t)
= D7 SmnPn(t), m=0,1,...,Me—1, O<t<T (12.1)
n=1

so that k = N- and T, = T;. As discussed in Chapters 3, 4, and 11, the building blocks
{@n(t)} form a known orthonormal set over the time interval T sec. Each of these coded
signals has energy
Ie
Em = PmT =i) S? (t)dt =| Sm |? (12.2)
0

for all m=0,...,M_ — 1. Notice that smn =./En.Xmn where Ey, is the peak energy
available per codeword dimension. For constant envelope signals (for example, M-PSK),
Ey, is also the average energy used per codeword dimension.
A particular mechanization of the encoder-mapper-modulator is shown in Fig. 12.1b
for the special case k = N., N = 1. Two cases of particular interest are binary coded BPSK
and Q-ary coded Q-ary ASK. We note that while the actual coded message bits assumed
in coding theory for BPSK are xm, =0, 1, we often write x, = +1, —1 corresponding
to voltage ievels realized in hardware. Thus, for binary coded BPSK, we have xm, = +1,
while for Q-ary coded Q-ary ASK we have Xmy = qj = (—1) + (2i — 2)/(Q — 1) with
probability p;, P= 1,...:,:Q;/ where Q is: even, and: Sap = ENOnn: For both of these
coded modulation techniques,

10
Wee | ees De eel
0, else

for all n.

12.1.1 Coding from the Baseband and IF Channel Model Perspectives

When the IF channel model is assumed to add white Gaussian noise (No watts/Hertz
single-sided) to the transmitted signal, it was shown in Chapters 3 and 4 that the optimum
transmitter-receiver configuration for equally probable transmitted messages reduces to that
illustrated in Fig. 12.1. Notice that the signal generator and modulator functions and the
demodulation and signal detection functions in the uncoded system have been combined
into the Modulator and Demodulator functions of the coded system; see Chapter 11. We
use bold, capital letters in “Modulator” when referring to the combined mapper/modulator,
and in “Demodulator” when referring to the combined demodulator/detector. In addition,
the interleaver and deinterleaver functions (if needed) in a coded system are also combined
into the channel encoder and decoder functions respectively.
Referring now to Fig. 12.1, we note that if hypothesis H,, is true, that is, if message
Mm is to be transmitted, then the code vector x,, implies coded modulation vector S,,, which
implies the IF channel waveform S,,(t). We then observe r(t) = Sm(t) + n(t), from which
the received signal plus noise projection r = (r},...,7N,) is generated by the demodula-
tor; Pn = Smn + Mn for all n. The decoded vector v = (v1, ..., UK,) 18 generated from r and
transformed into the recovered message m corresponding to hypothesis Hj». Table 12.1 sum-
marizes several sets of time and frequency orthonormal functions (communication carriers)
of great practical interest. The I-Q time-orthonormal functions are of greatest interest in the
756 Block-Coded Digital Communications Chap. 12

Table 12.1 Communication Carriers of Greatest Practical Interest

Communication
Basis Functions (n = 1, 2,..., k)

oy = xe COS Wet, (n—1)Ty <t <nTy

BPSK or Time orthonormal
’ else
ASK (@e a multiple of
m/TN)
1-Q modulation: Q2n(t) = |Kcr Ty Sin Wel, We 1)Tn Si nTn

time orthonormal y
(n—1)Tw St <nTw
(we a multiple of | gy_\(t) = {V2Tw cos wt, else
2n/Tn) 0,

Coherent M- Frequency or-

ee ee ae +n/T)t], 0S
ae
FSK thonormal (w, a
multiple of 7/7)

Noncoherent I-Q modulation: Gon(f)—~/ 2/8 sn i@e-+ 2an/ Tt), Osta

M-FSK frequency orthonor- @mn—1(t) = V/2/T cos[(@- + 2an/T)t],0<t<T
mal (w, a multiple
of 27/T)

implementation of coherent, partially coherent and differentially coherent systems, while

the I-Q frequency-orthonormal functions are used as building blocks in noncoherent com-
munication systems.

IF channel perspective. When the IF channel is considered, we wish to select

the coded modulation vector set {S,,,m =0,...M- — 1} so as to minimize the probabil-
ity of error. This points immediately to the problem of optimally mapping the vector space
generated by the encoder onto the IF waveform space (generated at the modulator output)
such that the distance properties of the code are not degraded. In what follows, we shall
address this coded modulation problem from a system engineer’s perspective. We shall con-
sider the coded communication design problem assuming that the mapping of the encoder
output onto the modulator output is isometric when properly normalized. A metric space
(X, d) is isometric [19] to a metric space (Y, e) if there exists a one-to-one, onto function
f : X — Y which preserves distances; that is, for all p,q €¢ X, d(p,q) =e(f(p), f(q)).
Here d is a “distance function” defined on the Cartesian Product X x X, so that d(p, q) is
the distance (as defined by d) between two points p, q*e X. Similarly, e(f(p), f(q)) is the
distance (as defined by e) between two points f(p), f(q) € Y.

Baseband channel perspective. For the baseband channel model (see Chap-
ter 11), we can equivalently pose the problem of communicating with block codes from a
geometric point of view. A block code consists of a subset of M. vectors (points or code-
words) chosen from an N,-dimensional vector space. It may be thought of as a way of
transmitting an integer m = 0 to M, — | to the receiver by sending the corresponding code-
word X. A decoding system for such a code is a partitioning of the N,-dimensional ob-
servation space into M, subsets or disjoint decision regions Aj». If the received signal plus
noise projection r belongs to A, the transmitted message is taken to be message m. See
Chapters 3 and 4 for all the details. An optimal decoding system for a code is one which
minimizes the probability of error for the code. Since the Gaussian density is monotone
 Sec. 12.2 Performance of Block-Coded Communication Systems 757

decreasing about its mean, an optimal decoding system for a given code is one which de-
codes any received signal as the integer m corresponding to the geometrically nearest code-
word, assuming equally probable, equal energy signals. If there are several codewords at
the same minimum distance, any of these may be used without affecting the probability of
error. As shown in Chapter 3, a decoding system of this sort is called minimum distance
decoding or maximum-likelihood decoding; it amounts to partitioning the N--dimensional
space into N,-dimensional polyhedra or decision regions “centered” around the different
signal points. Each polyhedron is bounded by a finite number (not more than M, — 1) of
(N_ — 1)-dimensional hyperplanes. In this context, there are M. = Q*< different codewords
to be assigned to points in our N.-dimensional signal space, in which there are Q< points
available; it follows that there are QNe“ different ways (codes) to assign the M, codeword
vectors. This is the usual approach to studying coding system design; that is, the baseband
channel model is assumed, and the modem design problem does not explicitly enter into the
system model. The question of isometric mapping of the encoder output onto the modulator
output does not arise in the baseband model; the modem remains hidden in the baseband
channel and is assumed to be ideal. In this chapter, we work with the IF CMC and DMC
models which take the modem outside the channel.

12.1.2 Fundamental Parameters of a Block-Coded Communication System

Before proceeding, it seems appropriate to summarize the communication parameters and

fundamental relationships of our coded system model.! It is highly recommended that the
reader gain a good physical understanding of each; they not only play a key role in the
theory to be presented, but they also enter into the specification of the subsystems illustrated
in the functional models shown in Figures 11.2 and 12.1. Understanding of these models
and their associated parameters will be required if one wishes to comprehend the system
engineering implications associated with Shannon’s random coding bounds, his channel
capacity theorem, and the physical meaning of efficient and reliable communications.
Tables 12.2 and 12.3 provide a summary of the basic parameters. Study the contents
of these two tables; they are of great importance to system engineers. For convenience,
the parameters are divided into several categories; some are physical, yet others take on
numerical values. The first category in Table 12.2 consists of time intervals. These intervals
are generated in practice by means of frequency generators called clocks when dealing with
baseband waveforms and called local oscillators when generating the IF and RF waveforms
at their respective nominal center frequencies. Without loss of generality, we shall assume
the existence of a single clock which can be used to generate the various time intervals
defined in Table 12.2. We introduce bandwidth via the sampling theorem (or the so-called
Nyquist interval of 7, sec), that is, B A 1/27, Hertz—the single-sided Nyquist bandwidth.

12.2 PERFORMANCE OF BLOCK-CODED COMMUNICATION SYSTEMS

The exact performance expression of an uncoded M-ary, coherent communication system
with equal energy signals was given by (4.80). Here we wish to establish the interplay
between the channel, the codec, and the modem illustrated in Figures 11.2 and 12.1. This
is best accomplished using the union bound result developed in Section 3.2.7, which allows

1. A coded digital communication system is complex and consists of many parameters involving different time
intervals, rates, energy, energy-to-noise ratios, and bandwidth.
758 Block-Coded Digital Communications Chap. 12

Table 12:2 Fundamental Parameters of a Coded Digital Communication System

T-—Time used per codeword dimension, sec

T;—Modulator waveform duration; 7; = NTy sec
Time intervals T,— Time per message bit, sec
T—Codeword duration; T = (log) Mc)T, = kT; = NcTe, Te = reTb, sec
Ty—Time used per modulator dimension by modulator, sec

k—Number of times modulator is used per codeword transmission

M.. = Q*:—Number of codewords in code dictionary
K,.—Number of information symbols per codeword (message length)
Codec/modem N-—Number of coded message symbols per codeword
(Codeword dimensionality = block length)
M—Nunmber of waveforms modulator generates per duration 7; sec
N—Dimensionality of modulator; N = WT, = WT /k

Rp = (logy M-)/T =re/Te = 1/Tp, bits/sec

R = log, M/T; = Rn/ Tw; modulator output rate, bits/sec
re = log, M./ Nc; code rate, bits/dimension
Ry = (log, M)/N; modulator output rate, bits/dimension
Ro—Channel cutoff rate, bits/dimension
Ro = RoW—Channel cutoff rate, bits/sec

E = PT; received energy per codeword

E, = PT;; modulator waveform received energy per use of modulator
Energy (joules) Ey, = PT; received energy per codeword dimension
Ey = PTp; received energy per information bit
En = PTy; received energy per modulator dimension used
No =kT°; noise energy or single-sided PSD
P = E/T; received power, Watts

W = N./T = 2B; two-sided IF Nyquist bandwidth

Bandwidths B = 1/2T,; single-sided Nyquist bandwidth
(Hz) D=1/Ty = N/T; number of dimensions used per sec by modulator
D = W = 2B for binary coded BPSK and binary coded QPSK
r~! and Ry'—Coder and modulator bandwidth expansion, respectively

Normalized “Tr = NTs = Wi bits/sec/Hz = bits/dimension

throughputs b — r., bits/sec/Hz = bits/dimension
fc—IF channel center frequency, Hz

us to introduce the reader to the important concept called the channel cutoff rate, Ro. As we
shall see, this rate is a simple function of the modem and the communication parameter
En,/No. In Section 3.2.7 and in (4.81), we showed that the conditional message error
probability performance of an M,-ary communication system with equally likely messages
over an AWGN channel is union bounded from above by a sum of the pairwise error
probabilities (PEP’s) discussed in Appendix 3D; that is,
Sec. 12.2 Performance of Block-Coded Communication Systems 759

Table 12.3. Processing Gains and Energy-to Noise Ratios

PGy = WTy
= 2BTy
Processing PG,=WT,=2BT,
Gains PG=WT =2BT=N-

Energy-to-noise ratios xe—Received energy per bit-to-noise density ratio

E J ; 2 ; : 5
Wy Received energy per modulator dimension-to-noise density ratio
EE
No —Received energy per message-to-noise density ratio
E ; : : : ;
i —Received energy per coded message bit-to-noise density ratio

Received signal power-

to-noise density ratio
for constant envelope
modulation, Hertz

M.-1
c di;

P(E |8;) < 12.3

he | Cree eee
(PEP)
where the Euclidean distance between the coded modulation waveforms S;(t) and S;(t) is
given by

dij = |S; — Sj|= [ ts@ - sora

0 (12.4a)

= /EF;+ E; —2/E;,;E; cos; = Ej; + Ej —2/

Ei Ejyij

From (12.3) we see that good coded modulation techniques employ coded modulation vec-
tors such that the Euclidean distance between any pair is large enough to ensure that the
channel noise will not cause the message that was sent to be confused with another message.
The Hamming distance D(x;, Xj) 4 Dj; between the two binary codewords x; and x;
is the number of positions in which the two codewords differ. For example, the Hamming
distance between (0, 0, 1, 1, 1) and (1, 0, 1, 0, 1) is 2. Considerable insight into the coded
communications problem is obtained by writing (12.3) in terms of the Hamming distance
Dj; when the coded modulation waveforms have equal energy E. In Section 12.8.1, we will
show that for binary coded BPSK modulation, the squared Euclidean distance is related to
the Hamming distance through

dj, = 2E(1 — yij) =4En,Dij = 4E Dij/Ne (12.4b)

Apparently, for most other coded modulation techniques, the relationship between the Ham-
ming distance between codewords and the corresponding Euclidean distance between coded
modulation vectors cannot be readily identified.
Using (12.4b), (12.3) can be rewritten as

ely 2E ( Dij |
BES) o( 72(34) (12.5)
J=014)
760 Block-Coded Digital Communications Chap. 12

If we denote the minimum Hamming distance by Din and the minimum Euclidean distance
by dmin, then (12.5) can be used to upper bound the message error probability for a block-
coded system with equally probable messages which uses binary coded BPSK modulation.
The resulting bound is

denn 2E ( Dmin
P(E) < (Mc ~ 1) (2%.) = (Me - no( =z(N ))
(12.6)

=(M,-1
=(M,.-1)Q sine”
re ma

where we have used the facts that E = PT = P(N,T,-) = PN¢-(rcTp) = Ep(Nerc) and the
definitions

dnin & min dj; and Dyin A min Dij Cl27)

isi isl
The result in (12.6) will be very useful in what follows because the minimum Hamming
distance Dmin plays a key role in error-correction code selection. From (12.6), we also see
that the use of binary coding has effectively enhanced the value of E,/No by the factor
rc Din Over that of an uncoded BPSK system. ;

12.2.1 Transmission of Orthogonal Modulation Vectors

Since (4.80) cannot in general be simplified, we first illustrate key features of (12.3) by
assuming transmission of N.-dimensional, equal energy, orthogonal coded modulation vec-
tors, namely

Si = VEQ; (12.8)
with @; - ?;= 6;; for all i, j =0,1,...,M_- — 1 (these codes will be considered further
in Section 12.3.1). In this case, dj; = /2E. Thus (12.3), when averaged over equal prior
probabilities, reduces to

P(E) < (Mc — 1)0 (/E/No) (12.9)

Using the fact that Q(x) < exp (—x*/2); then

P(E) < M.exp (—E/2No) (12.10)

If we assume a binary code with Q = 2, then M, = 2*-. Since the bit rate Ry = log, M-/T
= K,/T or K- = RpT (see Table 12.2), we can write (12.10) as

P(E) < 2*¢ exp (—E/2No) = 2! exp (—PT/2No) (12.11)

or

PEE) << Exp |-7(= - Ryin2) (12.12)

2No
SO

P(E) < exp[—K,-(Ep/2No

— 1n2)] (12-13)
 Sec. 12.3 Special Types of Binary Block Codes 761

From (12.12) we see that the decoded message error probability can be made to approach
zero exponentially with increasing T as long as the bit rate Ry satisfies the bound

R. <a aad 4 t e

Alternatively, from (12.13) we see that P(E) approaches zero exponentially as K, increases,
as long as the communication efficiency parameter

Ep
— >2In2~1.4dB (12215)
No
Since No = kT°, it follows that the minimum received energy per bit per degree Kelvin is

Ep/°K = 2kin2~ 1.91 x 107” joules/°K (12.16)

where k = 1.38 x 107? joules/°K is Boltzmann’s constant. At room temperature (T° =
290° K), this requires an energy per bit of Ey > 5.55 x 10-2! joules! The results established
in (12.12) and (12.13) differ from the bit-by-bit signaling results given in (11.7) for large K,
when E,/No > 21n2. We next consider the reason for such contrasting behavior.
In the bit-by-bit signaling (signal points on the vertices of a hypercube) of K; bits
using uncoded BPSK, increasing T (or K,) forced P(E) toward unity regardless of how
large the value for E,/No. In the case of signaling with orthogonal codes, by increasing
K. we can force P(E) to approach zero (perfectly reliable communication) provided that
E,/No exceeds 1.4 dB! The reason for this interesting asymptotic behavior is that the
minimum Euclidian distance between the signals (vectors) used to accomplish bit-by-bit
signaling is d = 2,/E,, no matter how large we make K, or T. In the case of block-
orthogonal coded modulation, d = ./2K,Ep, so that as K, increases, so does the distance to
the nearest coded modulation vector. Finally, we note that we have exploited two different
signal designs; however, in both cases, we use the optimum matched filter receiver for
signal detection (demodulation and decoding). The important role which signal design has
on communication reliability and efficiency is clearly manifested by comparing these two
signaling techniques. In the next section, we consider the generation of three types of binary
codeword dictionaries.

12.3 SPECIAL TYPES OF BINARY BLOCK CODES

In this section, we describe methods of generating binary orthogonal, biorthogonal, and
transorthogonal (simplex) codes. The codewords are assumed to be represented by se-
quences of “Os” and “1s”, the case of greatest interest. Block codes have the interesting
feature that the minimum Hamming distance between any two codewords is related to the
error-correcting capability, say t (not to be confused with time) of the code. In general, it is
known [3] that a code with minimum Hamming distance

min & min Dj; (12:17)

iZj
has the t error-correcting capability

ACD 2 bog hte ES)

762 Block-Coded Digital Communications Chap. 12

where the brackets denote the largest integer no greater than (Dmin — 1)/2. It is also true (see
Section 12.8.1) that the cross-correlation coefficient between any two binary codewords, say
x; and xj, of length N¢, is given by

siete oa fe geet (12.19)

Ne Ajj + Dij

where Aj; represents the number of term-by-term agreements and Dj; represents the number
of term-by-term disagreements (or the Hamming distance) between the two codewords. We
shall use these facts in what follows by considering the transmission of specific block codes
using BPSK. For this case, we note that Ry = 1 and that the coding to modulation mapping
is isometric.

12.3.1 Orthogonal Binary Codes

Let Hx, be a 2*« x 2%¢ matrix defined inductively by

C ».¢
Hp) Hee Ss
HySSt)|) Mee aay ee ae (12.20)
Hx-1 : Hkx,-1 XN,-1
where we have assumed N, = M, and

Hie i AE (12.21)

Here the overbar denotes the complement of the “H” matrix, wherein the ones are replaced
by zeros and vice versa. Note that the rows of Hx, represent the M, orthogonal code-
words {Xm = (Xm1,Xm2,---,Xmn,)} (see Figures 11.3 and 12.1). Such a matrix is called
a Hadamard matrix [4] and the codeword dictionary generated by (12.20) is known as a
first-order (24°, K-) Reed-Muller code. For an orthogonal code, Aj; = Dj; so that y;j =
0, for alli # j. Thus, from (12.4b), yj; =0 = 1 — 2D;;/N¢ since Ajj = N- — Djj, so

Dij = Dain = Ne/2=2"°", i Aj (12.22)

and

= i Cia » 1)/2] (12.23)

errors are correctable.
Now that the codeword dictionary has been formed, all that remains to completely
specify the code is to define the mapping between encoder input vectors uy, = (Um, ...,
UmK,) and codeword vectors Xm. Define a matrix Gx, to be a 2Ke x K, matrix whose
m'" row is the binary representation of the integer (m — 1). The columns of Gx, that is,
81, 82,---, 8K,, represent the code generators of the codeword dictionary Hx,. Also define
the 2Ke x K, data symbol matrix Ix, whose it? row ib= 1,2, ...; N. = M¢ is the m" data
vector Um. Since the elements of u,, correspond to the binary representation of the integer
(m — 1), we notice that Ix, is identical to Gx. In terms of the above matrix definitions, the
codeword dictionary is easily generated from

Hx, = Gx Ty. (12.24)

Sec. 12.3 Special Types of Binary Block Codes 763

where the superscript “7” denotes a matrix transpose. It is understood that the multiplication
and addition operations involved in the matrix multiplication are reduced modulo-2.
Example 12.1
(8,3) Reed-Muller code

o7"000
O00
Oo 1 0
artodl arom On kl tomy
G= pees
t
O0L 1 Shi0080 es | be t
me
Dea Om iO 1 Qual one : a
pe beat)
i tee
so that

ORO" 400 On OLaROME O

Oe Ve hOment CGY WOE va
Test ads Nl | Oeeae at
eae a ine EN Oh lo tat, 0
H3 = —

Va OesOn WOU si)- le tt ot x7

(I Oo ee OL een
(Sula ena te ee paLs afl be Opi
Ge Ne MLO Tie eho tecoy ay
The message and bit error probabilities have been derived and are given by (4.93)
and (4.96), respectively. We note that for coded communications, there exists a one-to-
one correspondence between messages and codewords, so that the phrases “message error
probability (MEP)” and “codeword error probability” are synonymous.

12.3.2 Biorthogonal Binary Codes

The codeword dictionary of a (2*°—!, K,) biorthogonal code is easily defined in terms of
the dictionary of an orthogonal code through the partitioned matrix

Hx.-1
Bx.=|= *
ii ae
The message and bit error probabilities have been derived and are given by (4.102) and
(4.107), respectively. Here M, = 2N- and for any particular codeword x;, Djj = N-/2 for
M,. — 2 of the codewords x; (i # j), and Dj; = N- for the complementary codeword x; = Xj.

12.3.3 Transorthogonal Binary Codes

The codeword dictionary of a (2Ke — 1, K-) transorthogonal code can be generated from an
orthogonal codeword dictionary by deleting the first column in Hx,. In practice, this can
be accomplished by deleting the first digit of each codeword. The message and bit error
probabilities are given in (4.116) and (4.96) respectively. The minimum Hamming distance
764 Block-Coded Digital Communications Chap. 12

Table 12.4 Comparison of the MEP for Various Binary Coding Techniques

Uncoded BPSK
(signal points on
hypercube)

is Dmin = M-/2 with N. = M, — 1. Also note from (12.19) that y,; = —1/(M_- — 1) since
Ajj = Djj — 1 and Ne = Me — 1 = Ajj + Dij-
A comparison of the message error probability bounds, as found from the union bound
result given in (12.5), is given in Table 12.4 for various binary coding techniques. The case
of bit-by-bit signaling (uncoded BPSK) is also given for comparison purposes. For this case,
we note that r- = 1 and Dmin = | since the signal coordinates disagree in at least one place.

12.3.4 The Bandwidth Issue for Binary Coded BPSK Modulation

In Chapters | and 2, definitions and techniques for assessing the bandwidth requirements of
a communication system were presented. Although no constraint was placed on the band-
width of the binary coded waveforms discussed in the previous section, it is necessary in
practice to be able to allocate and control bandwidth. Assuming that the transmitter uses the
(2, 1, T;) binary coded BPSK modulator, then D A N/T; = 1/Ty = 1/T; = 1/T; dimen-
sions per sec; however, we know from the sampling theorem that a signal of bandwidth B
Hertz can be uniquely represented by NV, = 2BT samples taken at the Nyquist sampling rate
of 2B = 1/T, = 1/Ty = D samples per second [15]. Thus

BR=l2T.= DOM: (12.25)

serves as the relationship which we will adopt between the parameters D, T,, and B. We
will later see that (12.25) also holds for Q-ary coded Q-ary ASK.
Since N, = 2*¢ for orthogonal binary codes, we observe that B = 2Kc—!/T Hertz,
which demonstrates that B expands exponentially with K, = R,T! This is undesirable be-
cause the bandwidth required also expands exponentially with T. However, the orthogonal
binary codes constitute a small subset of the class of codes, binary or otherwise. This leads
one to ask if there exist codes whose bandwidth requirements do not expand exponentially
with T and yet have the property that the message error probability (MEP) asymptotically
goes to zero as T gets large. This question is answered in the next few sections.

12.3.5 Channel Throughput and Time-Bandwidth Requirements

In selecting modulation and coding techniques, one important consideration is channel

throughput. This is usually defined as the ratio of the bit rate 7, to the two-sided Nyquist
bandwidth W = 2B Hz. From Tables 11.1 and 12.2, we have the modulator throughput
R _ log.M _ logyM _ Rn
bits/sec/Hz (12.26)
W WT, WTyN WTy
 Sec. 12.4 The Ensemble of Block-Coded Modulation Systems 765

Table 12.5 Channel Throughput Comparison for Binary Coded

Modulation Techniques

2a a
wesconr=1] €[—[—_
Trans-_ codes BPSK orQpsk [1

and for coded systems we have the normalized channel throughput

R
- = log, Mc/WT = logy Mc/Ne = re bits/sec/Hz (12.27)
Equations (12.26) and (12.27) demonstrate that the units “bits/sec/Hz” and “bits/dimension”
are equivalent since WTy is unitless. Also note that for binary coded BPSK, T. = Ty
and WTy = 1. For the case of an uncoded system, we recall that Rp = log, M/T, so
Rp/W =log, M/WT = (log, M)/N = Ry bits/sec/Hz.
Table 12.5 provides a comparison of channel throughputs achievable for various un-
coded and binary coded modulation techniques. Notice that as K, (the number of bits per
message) increases, the normalized channel throughput R»,/W increases linearly with K,
for M-PSK without coding, while for the three binary coded modulation techniques, the
normalized throughput approaches zero exponentially as K; increases. This effect is due to
the fact that the bandwidth requirement W = N,/T grows exponentially with K, (since N.
is proportional to 2*¢). This severely limits the applicability of block-orthogonal codes in
practice.

12.4 THE ENSEMBLE OF BLOCK-CODED MODULATION SYSTEMS

As we observed in the previous section and in Chapter 3, it is difficult to evaluate the de-
coded MEP for an arbitrary set of code vectors, binary or otherwise. Yet a more difficult
signal design (coding) problem is to find the matrix of correlation coefficients T = [y;;]
which minimizes the MEP. Even if this matrix were known, it would still be difficult to
specify the design rules for the channel encoder and decoder. Furthermore, recall from Sec-
tion 12.3.4 that, while binary orthogonal codes have MEP that can be made to approach
zero, the bandwidth W = 2*«/T grows exponentially with message length K,. Such a band-
width requirement quickly exceeds any bandwidth allocated by the Federal Communication
Commission (FCC) for any physical channel. Perhaps Shannon [5] recognized that nature
has hidden good codes (channel codec algorithms) with MEP as demonstrated in (12.10) to
(12.16), and devised an astute argument to show that, as T becomes large, many codes exist
such that bandwidth does not increase exponentially with T even though the MEP decreases
exponentially with 7. Shannon’s derivation of the exponential error bound was based on an
analysis called the random coding argument. The bound is obtained by averaging the mes-
sage error probability over an ensemble of all possible codes, as we shall subsequently do.
766 Block-Coded Digital Communications Chap. 12

We begin by considering the ensemble of block coded systems illustrated in Fig.

12.2. As we shall presently see, the Q-ary block code ensemble consists of the QNeMc
distinct block codes x A ({x;}1, {xi}2,..., {Xi} QNcMc) which we now define. Each code in
the ensemble consists of M, codewords, not necessarily distinct, such that each message
m(m=0,..., M~— 1) can be mapped to a codeword. The Q™ Mc coded systems in the en-
semble operate in parallel as follows. If message m is sent, then u = Um = (Umi, .--, UmK,)
for all m=0,..., M,. — 1. We further assume a Q-ary code with umx € {g1,---,qQ} for
allm =0,...,M@e—1 and k=1, 2,..., Ke. Thus, there exist M; = Oke possible source
encoder outputs for M, distinct input messages.
Now the ensemble of block encoders maps u,, into {x}¢ = {xo, X1,.-., Xw,—1}¢ con-
sisting.of codewords x,,, 7m =0,...,M,—1for@=1,2,..., QNeMe. These codes are now
described along with the operation of the system ensemble in Fig. 12.2. If message m is sent,
then u=u,, for all @ and {x}; = {(Xm1, ---, Xmn,)}1. Thus, there are QNe different code
vectors available and M, = Q*« = Q®#T = Q’Ne messages to be assigned to code vectors;
it follows that there are (QN°)@ce = QNeMe distinct ways (codes) to assign the messages to
codewords; thus, the communication system consists of @N““< block-coded systems oper-
ating in parallel; see Figures 12.2 and 12.3.
Assume that the ensemble of modulators isometrically maps the ensemble of block
codes x into the ensemble of coded modulation vectors S A ({Sm}1, {Sm}2,---, {Sm} gnemc)
or waveforms S(t) (see Figures 12.2 and 12.3). The ensemble of channels maps the trans-
mitted coded modulation ensemble S(t) into the received waveform ensemble r(t). The
ensemble of demodulators maps the received waveform ensemble into the demodulator out-
put vector ensemble of projections r = ({r}1,..., {}gncm-). The continuous random vector
ensemble r emerging from the IF CMC ensemble can be mapped into the quantized ran-
dom vector ensemble R = ({R}1,..., {R} gemc) using a J-level vector quantizer (see Fig.
11.3). Here the ensemble of IF CMC’s is characterized by the ensemble of conditional chan-
nel transition probability density functions {f(r|S»)} A {fr(e1, ---, Pn.|Sm)} and the quan-
tized IF CMC (that is, the IF DMC) is characterized by the ensemble of transition probabili-
ties {P(Rn = Raj |Smn = Gi) = pij & PUM} for alli =1,...,Q, forall 7=0, ..., J —1,
for alln =1,..., Ne, and for all m =0,..., M. — 1 (see Section 3.1.1). Later in this sec-
tion we will show how the IF CMC is reduced to the IF DMC. Finally, the ensemble of
block decoders maps ensemble r (or R) into the ensemble vector U, the estimate of u,,; and
the source decoder maps 8 into the recovered message m. Thus, our ensemble of block-
coded systems can be conceptually and functionally represented as in Fig. 12.3. Notice, the
ensemble of AWGN channels in Fig. 12.2 has been replaced by the ensemble of IF chan-
nels. In this chapter, we will only consider “matched” coding and modulation techniques.
By this we mean that the Hamming distance between codewords is preserved in the Eu-
clidean sense at the modulator output, or that the mapping from codewords to waveforms is
isometric. Binary coded BPSK and QPSK are examples of isometric modulation techniques.

12.4.1 Performance of the Ensemble of Binary Coded Systems Using BPSK

Modulation

Consider now the binary coding case (Q = 2) and assume that m = m,,. This implies that
the code vector is x,,, and the coded modulation vector is Sin = (Sm1,..-, Sm n,)- With BPSK
modulation we assume that Sm, takes on +,/Ey, or —/En, with equal probability for all
m=0,...,M-—1,andn=1,..., Ne. For these equal energy signals, the received signal
energy is
 Japoouy Joye|NpoW |(;)1, 40VeINpOW
iy] 40/4
SQ
is} Jepooeq

‘{?x}
(a)'u .

Joye|npoW

yoo|g
SS

'{!s}
yoo|q

'f!x}
Jepoouz
Jepo0eq

'u (2) ;

((7)“c)
wn Oy apn (J) ‘ Oy 2wn

my)
*w'nO(
y90|q
Japoouy J0}e|NPOW Joyeinpowag 490/q
{ s} Oh, u ?WN Jepo0eq
*w?nO {1} >wnOf!

Oy wn (2)

eins
"ZI sjquissuq
Jo pepoo-yoo]q
[eIISIP UOTBITUNWIWOD
SUIDISAS

767
768 Block-Coded Digital Communications Chap. 12

Source
Encoder
“a, fer hf
Ensemble Modulators

Ensemble
of IF
Channels

Source
oder |
Decoder Ensemble Demodulators ll

Figure 12.3. Ensemble of block-coded modulation systems

Ne
E=PT=) $s? =N-Ey, joules (12.28)
n=1

and since T = N-T;, we have that Ey, = PT, joules. Since the calculation of the decoded
message error probability (MEP) is not feasible for an arbitrary block code, we consider
instead the problem of upper bounding the ensemble average MEP for the class of bi-
nary block codes. Examples of binary codes in this class include the Hamming and Golay
codes, while examples of Q-ary codes are the Reed-Solomon and BCH codes. Applica-
tions of these coding techniques to the error-control problem will be given later on in this
chapter.
Now the conditional MEP, P(E|{S,,}¢), for the ¢'” system in the ensemble can be
expressed using (4.80) as a function of E/No and the é'" code cross-correlation matrix
I'e = [yjx]e; however, as we have seen, it is difficult to find the global minimum of P(£)
as a function of Iz. In principle, the ensemble average MEP

P(E)= E[P(E|{Sm}2)] (12.29)

can be upper bounded using the union bound result (4.81) for the conditional MEP (assum-
ing message m is sent and all messages are equally likely) given by

M.-1
:
P(E\{Sm}e,Sm)< D> ({dink
ae}e (12.30)
kém,k=0 2No

where {dx}¢ denotes the Euclidean distance between the coded modulation vectors Sj, and
S, in the £’" codeword dictionary, £=1,...,2%ec. In the context of random coding, we
note that dmx is a random variable since the components S,,,, and sxn,n = 1,2,..., Nc, of
the vectors S,, and S, are binary random variables taking on values ,/Ew, or —\/E Nn. With
equal probability over the ensemble. To further characterize the random variable d,,x, let
Dmx be a random variable denoting the Hamming distance between S,, and S,;, taking on
value h. If S,, and S; differ from each other in h coordinates, then
Sec. 12.4 The Ensemble of Block-Coded Modulation Systems 769

ane, ={IS»— Sil?) = ite Stak (2/En.) Seve elon

n=

for all £=1,2,...,2N«Mc

whereh =0, 1,2,...,.N. Ck
In general, we point out that the Euclidean distance d,,x is the variable of interest
in the design of coded modulation systems, while the Hamming distance has been used
as a measure of the goodness of a code in coding theory. For binary coded BPSK, the
Hamming distance is related to the Euclidean distance via as = 4En.Dmx from (12.4b).
Now substitution of this relationship into (12.30) leads to

P(E\{Sm}esSm) < =)~ o( Pails:

{Dink}eEn,
(12.32)
k=0,kém No
Since Dmx counts the number of places in which S,, and S, differ, it takes on value h with
the same probability as that of getting heads in N, tosses of a fair coin; that is

1\%
PCDane= b= (4) (5) (12.33)

forh=0,1,...,N- and £=1,2,..., 2NeMe, Using (12.33) to perform the ensemble aver-
age over the Hamming distances in order to eliminate the conditioning on {S,,}¢ in (12.32)
produces

ref Y o(FE)()O
M.-! N
Nee Ma 2hE eas

=0 k=0,k xm

Ome ae)
(12.34)
Ne M--1 Ne hEy,

2 k=0,k4m h=0 ;
where the last line follows by changing the order of summation and upper bounding Q(x) by
e~*’/2. Note that the inner sum is independent of k and reduces to (1 + e~2Ne/ Noy ¢, Upper
bounding M, — 1 by M;, and summing over h and k yields the closed form expression

P(E|Sm) < Mc{1 + exp (—Evn,/No)\%°/2™ (12.35)

Then, since the right-hand side is independent of S,,, we have the exponential MEP bound
for the ensemble,

P(E) < Mc2~Ne®o = M277 (12.36)

where we have introduced the binary coded BPSK channel cutoff rate
RoA log,{2/[1 + exp (—Ey,/No)]} bits/dimension (1237)

and Ro 4 Ro/T- = W Ro, the normalized channel cutoff rate in bits/sec. Since M;== 2Ke=
QRoT— 2"Nc, (12.36) can be rewritten as
P(E) <2 NelRo~re] — 9—KelRo/re—1] — 9-T(Ro-Ro) (12.38)

For the system engineer this is a very striking result: it says that as long as the code rate
re < Ro, the ensemble average error probability can be made arbitrarily small by taking
the block length N- (and hence message length K,) sufficiently large. Equivalently, as long
770 Block-Coded Digital Communications Chap. 12

as Ry < Ro, the MEP can be made arbitrarily small by taking the codeword duration T
sufficiently large. Since the ensemble of binary codes satisfies (12.38), it follows that at least
one code in the ensemble must have an error probability P(E) no greater than the ensemble
average, P(E). Notice from (12.37) that Ro approaches zero for small Ey,/No and unity
for large Ey./ No.
Now let us determine the form of Ro in the energy-limited region, Ey./No < 1,
where, with unlimited bandwidth, coding can be used to reduce the required signal energy
to the lowest possible level. To do this, we differentiate (12.37) with respect to x = Ey,/No
and take the limit as x approaches 0 to obtain
lim{d Ro/dx] = 1/21n2
>

Thus, Ro ~ (En,/No)/21n2 = (reEp/No)/21n2. When the coded system operates at a

code rate re = Ro and Ey,./No « 1, then Ro © Ro(Ep/No)/21n2. Again, this says that the
minimum value of E;/No to achieve the cutoff rate is E,/No = 21n2 = 1.4 dB.

12.4.2 MEP Performance Comparisen of Orthogonal Codes

with the Ensemble of Binary Codes Using BPSK
For binary orthogonal codes, (12.13) can be rewritten as

P(E) = or ae oe (12.39)
by noting that e* = 2*/!"?. We observe that the bound on the MEP tends to zero with
increasing message length K; if E,/No > 21n2 = 1.4 dB.
Since T, = rceT) (see Table 11.2), then for binary coded BPSK

—Ro =_ Th Ey Ro
— Ry = —— , _= ———_
Eb/No Ro (12.40)
rc T, En, En./No

and we can use this in (12.38) to write

P(E) < 2 Kel(Eb/No)/B-1] (12.41)

where
En./N
garni (12.42)
Ro
Thus, the bound on P(E) goes to zero with increasing K- if Ep/No > B for the
ensemble of binary coded BPSK versus E»/No > 21n2 for orthogonal modulation vectors
(signals). Notice that for small Ew,/No, where coding is most useful, we have shown that
Ro = (En,/No)/2 1n 2, so that B in (12.42) approaches 21n 2. This says that for very noisy
channels, binary coded BPSK is essentially equivalent to binary coded orthogonal signals,
which (from Chapter 4) are equivalent to binary coded transorthogonal signals for large M,
which are known to be optimum for the AWGN channel.
The binary coded BPSK result can also be obtained another way. Since Ey,/No =
rc(Ep/ No), then we can use this in (12.37) to write

Ro = Ro/ W = 1 — logy[1 + exp (—r-Ep/No)] bits/sec/Hz (12.43)

This says that operating the system at r, = Ro requires the communications efficiency of
Ep
wari Inf) a aire (12.44)
Sec. 12.4 The Ensemble of Block-Coded Modulation Systems 771

Using the restriction r- < Ro corresponds to requiring that

Ep
ates =
Qrre— Dre (12.45)
which says that, as the code rate re approaches zero (N, approaches infinity or the band-
width expansion 1/r, approaches infinity), that

Ep
— >2ln2=1.4dB (12.46)
No
for re < Ro. Can one do better? If yes, then the question is how much better? We proceed to
answer this question.

12.4.3 Performance of the Ensemble of Q-ary Codes Using Q-ary ASK

and OASK

In Section 12.4.1, we used the (2, 1, T;) BPSK modulator to generate the binary coded
BPSK waveform
Ne
Sm(t) =) 7)Smnn(t), O<t<T (12.47)
n=1

in which Sn is a binary-valued random variable taking on values +,/Ey, or —\/En, with

equal probability. We then demonstrated that over the AWGN channel the ensemble error
probability delivered by an optimum receiver approaches zero exponentially with block
length N, for rates Ro greater than r,. There are two issues of concern: (1) the bandwidth
W = N./T required to achieve this rate grows with the block length VN. = WT, and (2) the
cutoff rate Ro saturates at one bit/sec/Hz for large Ey./No due to the two-level nature of
the BPSK coded waveform S,,(t). One would intuitively expect that large values of Ey./ No
would permit reliable communication at rates above one bit per dimension so that binary
codes (binary-waveform sequences) will not be exponentially optimum when Ey,/No is
large. As we shall see in this section, both of these concerns can be significantly alleviated
using Q-ary Coded Q-ary Amplitude-Shift-Keying (Q-ary Coded ASK).
Consider now the (M, 1, 7;) Q-ary coded ASK modulator used k= Ne times per
codeword (that is, 7. = T,;) to transmit the class of Q-ary codes where M = Q. The coded
message symbols x,» are determined from the Q-ary alphabet Q = {q1, q2, ..-,qQ}; that
is, Xmn is a random variable taking on value g; from Q with probability P(g;) A pi. We
further assume q? < 1 for all i. The modulator waveform coefficients sj, in (12.47) are then
given by Smn = \/En.Xmn- Figure 12.4a demonstrates the possible coded message symbols
for the case Q = M =8, N = 1, and T, = T;, assuming the points are equally spaced on
(—1, 1).
The first concern mentioned above is addressed by comparing the channel through-
puts R,/W for Q-ary and binary coded (M, 1, 7;) modulators. For Q-ary coded ASK,
the throughput is given by Rg/ Wo = log, M-/WoT = K- logy Q/Ncg; for binary coded
BPSK, the throughput is given by ?2/W2 = K-/Nc2 bits/sec/Hz. Thus, the throughput ra-
tio is given by (Rg/Woa)/(R2/W2) = Nc2 log, Q/Nco. For a fixed data rate requirement,
Ro = R2, and equal block lengths, N-2 = Neg, the throughput ratio reduces to bandwidth
ratio W2/ Wg = logy Q or W2 = Wg log, Q. On the other hand, for a fixed bandwidth re-
quirement, W2 = Wg, and equal block lengths, the throughput ratio reduces to the data
772 Block-Coded Digital Communications Chap. 12

2/(Q-1) = 2/7 Smn = /EN, %mn

Figure 12.4 Q-ary ASK and Q?-ary QASK signal vector models: (a) Possible set of
values permitted the sy,; M = Q =8, N = 1, (b) Possible set of values permitted the
(Sinn (2); Smn(Q)); M= Q? =O Nat

rate ratio Ra /R2 = log, Q or Rg = R2 log, Q. These relationships define the throughput-
bandwidth trade-off of interest to the systems engineer.
The second concern mentioned above is addressed by obtaining a bound on the en-
semble MEP and examining the resulting Ro. Our approach is identical with that used for
the case of binary codes; that is, we determine the Ro (as a function of Q and Ey,/No)
which exponentially bounds the ensemble of codes average error probability P(E). There
are Q*< signal points available (potential codewords), and M, messages to be assigned to
the available points. Thus, the total number of ways to make the assignment, and hence the
total number of possible codeword dictionaries in the ensemble, is @%«“<. This says that the
ensemble of communication systems illustrated in Fig. 12.3 is now QNeMc systems; each
system uses different coded modulation vectors {S}¢,@=1,..., ONcMe: together with the
optimum receiver.
Notice from Fig. 12.4a that the end values gg = | and gq; = —1 have neighbors only
on one side. In the presence of noise these values have a higher noise immunity (or are
Sec. 12.4 The Ensemble of Block-Coded Modulation Systems 773

more distinguishable) and it is intuitively clear that the error probability will be smaller for
systems which use the coded modulation vectors {S,,}¢ in which the values g; near the ends
are used more frequently than the interior values.
If we let N; count the number of times that g; is used in the entire ensemble of
encoder-modulator pairs and if we assume that the message symbols occur independently,
then to the coded modulation vectors {S,,}¢ we assign the a priori probability
N
Sep, Dy ee lnc Oe (12.48)
where pj + p2+...+ po =1. Since each code yields M, coded modulation vectors S,n,
made up of N. symbols S,, then Nj + No+...+ Nog = N-M.. In order to bound the en-
semble MEP we use the union bound as before. Using the fact that O(x) < ent /2 we note
that the pairwise error probabilities in (12.3), conditioned on using the £th code, become

{dnk}e )
{P2(Sm,Si)}e = Q (
J/2No
(12.49)
-¢o (Ss - sh) E ples [-«s aa )2/4No]
J2No ahs mn n

Since the coefficients s,,, of S,, are random variables, we will need to average over all pos-
sible values in order to get the ensemble error probability, P(E). If Sinn = qi JEN, and sxp, =
qjVEn, then we have that the coordinate distances are |Smn — Skn| =|gi — aj|VEn, &
gij/En.- Moreover, the squared coordinate distance (Smn — Skn)* = ij Ey, occurs with
probability p;pj; since the levels q; and qj; are chosen independently and at random from
the Q-ary alphabet. Thus
Q
E {exp |— (mn z sin)”/4No| =) pipjexp(—)jEw./4No) (12.50)
i=) =i)

and the conditional PEP in (12.49) can be averaged over all codes in the ensemble to give
Ne
OF :
Px(Sm, Sx) < bsY= pip exp (-ahen./4m) (12.51)
i=1 j=1
Averaging the union bound over the ensemble of codes, we have
Tiss BUEN ey heat Sd ror =
P(E|Sm)< >. P2Sm, Sk) = (Mc — 1) P2(Sm, Sk) (12.52)
k=0,k#m
where the last equality holds since the right-hand-side of (12.52) is independent of k. Since
it is also independent of m, the MEP for the ensemble of codes is P(E) = P(E\S»m).
Bounding M, — 1 by M. = Q*c, and assuming equally likely messages, we obtain
RUE) (12.53)
where
Q Q 5
Ro A — logy Pi pj exp (—q;; En,/4No) (12.54)
I Let
774 Block-Coded Digital Communications Chap. 12

and where r, = K- logy Q/ Ne. Since Ro = T-Ro = Ro/ W and re = Rp/ W, we can rewrite
(12.53) as
-
P(E) < 27 Nel(Ro—Rs)/W] _ 9-T(RoRe)
Equation (12.53) demonstrates that the error probability can be forced to zero by using long
block lengths as long as Ro > re. Furthermore, there exists at least one code in the ensemble
of codes with P(E) < P(E).
If we further assume that the levels gj and qj occur with equal probability, then
pi = 1/Q@ foralli=1,..., Q, so the channel cutoff rate in (12.54) reduces to

1 Q oO ’
Ro = Ro/
W = —log, FEye) xn -ah n/N bits/sec/Hz (12.55)
i=l j=l

Lor. O=.2, ij = 4 fori # j and qj; = 0; substitution of these conditions in (12.55) leads to
(12.37) for coded BPSK as it should. Upper and lower bounds for the cutoff rate Ro can be
derived from (12.55) by noting that 4Ey./(Q — Ly qi EN < 4Ey, for all i A j. Using
these conditions, we find from (12.55) that

—Evn,/No
og: |+ (Q — l)exp || < (Ro
— log, Q)

< — log, [1 + (Q — 1) exp (—Ey,/No)]

with equality holding for Q = 2. This says that for large Ey./No, we have Ro < log, Q
bits/sec/Hz.
Given any choice of the amplitude levels {g;}, (12.55) can be evaluated numerically.
For equally probable levels and equally spaced amplitudes, as shown in Fig. 12.4a, the
curves represented by (12.55) for Ro are illustrated in Fig. 12.5 as a function of En,/No for
various values of Q [7]. Notice that as Ey./No approaches infinity, Ro approaches log, Q,
which represents the source entropy for equally probable message symbols; equivalently,
Ro approaches W log, Q bits/sec. This alleviates the concern that existed for binary codes,
where Ro saturated at | bit/sec/Hz. Furthermore, as Ey./No approaches infinity, there is
no need for adding redundancy in the codewords to achieve error control; hence, M;, ap-
proaches Q™ and the code rate r- = R»/ W also approaches the source entropy, log, Q.
At lower values of E'y,./ No (the region where coding is of greatest practical interest),
we see the effects that the interior signal points (q2, ..., q7 in Fig. 12.4a) have on Ro when
Q > 2. The curves for Q > 2 bend over later in the range of values for Ey./No and also
saturate at larger values of Ey,/No. Note that Ro increases as the number Q of amplitude
levels increases and that the noise immunity of the Q-ary code vectors that lie to the interior
of the outermost points is reduced. As Q increases, so does the number of interior points.
The Ro curves represented by (12.54) for equally spaced amplitude levels using an optimum
probability assignment are shown in Fig. 12.6 for various values of Q. We note that the
improvement in Ro due to the optimality of the probability assignment is negligible.
In Section 12.7.6, we will employ (12.54) to demonstrate how the Ro criterion can
be used to simultaneously optimize the design choice for the coding and modulation. For
low Ey./No, where coding is most useful, we will show that the optimum code is the
(M, = 2*, N. = 2*¢ — 1, T,) binary transorthogonal code, and the optimum choice for the
modem is compatible with the (M = Q =2, 1, T;) coded ASK or BPSK modulator.
Sec. 12.4 The Ensemble of Block-Coded Modulation Systems 775

10.0
8.0

6.0

4.0

2.0
N
=1)
v
AD

g
S +0
=So 08
fe
0.6
pe
So

0.4

0.2

0.1
-10 0 10 20 30 40 50 dB
Energy per Dimension to Noise Ratio, 10 log;y Ex, /No

Figure 12.5 Ro versus Ey,./No for equispaced coded Q-ary ASK; p; = 1/Q (Reprinted from [7])

To complete the discussion of Q-ary coded ASK, we revisit the bandwidth considera-
tions in Section 12.3.4 pertaining to use of the (2, 1, 7;) binary coded BPSK modulator. We
note that the argument that D = 1/Ty = 1/7; = 1/T; is true as long as k = Ne and N = 1.
This condition holds for the (M, 1, 7;) Q-ary coded ASK modulator of this section. Thus
(12.25) holds for Q-ary coded ASK.

Q?-ary coded QASK. The cutoff rate for Q?-ary coded QASK can be derived
easily starting with the I-Q generalization of (12.49). Figure 12.4b demonstrates the ex-
tension of the Q-ary coded ASK signal design to the Q?-ary coded QASK signal design
using the (16, 2, 7;) modulator. For the Q?-ary coded (M = Q?, 2, T;) QASK modulator,
the modulation vectors S,, are broken into in-phase, s,,(/), and quadrature, sm,(Q), com-
ponents with Sn(1) = /En,/2XmnU) and smn(Q) = J En,/2%mn(Q), where Xnn(1) and
Xmn(Q) take on values g; = (—1) + (2i — 2)/(Q — 1),i=1,..., Q for Q even, indepen-
dently and at random with probability p;. In other words, we transmit one of M = Q?
776 Block-Coded Digital Communications Chap. 12

10.0
8.0 a
64
6.0 |-
SZ

4.0 |- 16
* 8
Ro

32.0 4
ea@ 3
“we
2
=
= 10- Q=2 4
ce 0.8 1- =
M
oc 0.61 a
#o
Re ——— Upper envelope of Ry
0.4 |- curves for {p;)
optimized

02) a

0.1
-10 0 10 20 30 40 50 dB
Energy per Dimension to Noise Ratio, 10 log; Ex /No

Figure 12.6 Ro versus Ey,/No for equispaced coded Q-ary ASK; {p;} optimized (Reprinted from [7])

QASK signal vectors (Smn(Z), Smn(Q)) for each of the k= N,. codeword coordinates x,
with energy constraint |(Smn(1), Smn(Q))|* < En,
Thus, the I-Q generalization of (12.49) becomes

Ne
(P2Sm. So} < ] [exp {(smn) ~ sin VP + [5nn(Q) ~ Sin Q)P)/(—4No)|
n=1

Furthermore, since the I and Q channel random variables are independent,

E (exp {(smn(1)— sen (VP + [5nn(Q) — sin(Q)1?)/(—4No)|)

Q o : :
= sy Pip &xp (—4)) En,/4No)
i=l j= !
Sec. 12.4 The Ensemble of Block-Coded Modulation Systems Vii

where the expected value is computed over the ensemble. We use the above expression to
find the ensemble average PEP, P2(S,,, S,), which when substituted into the union bound
yields
2N,
Q Q
P(E|Sm) < (Mc — 1) b Pip; exp (—4;, EWN
t= 1

Bounding (M, — 1) by M. = Q?%« and assuming equally likely messages yields

P(E) Zz 9—Ne(Ro-re)

Q @Q
where Ro = Ro/W = —2 log, b»Y- pip exp (—47, En, ano bits/sec/Hz denotes the
i=1 j=l
cutoff rate for Q?-ary coded QASK and where r, = 2K, logy Q/Nc. We note that both r,
and Ro are twice the values obtained for Q-ary coded ASK, as we would expect.

12.4.4 Comparison of the Ensemble Performance of Q-ary Codes with Energy

Constrained Codes
Using a derivation we shall not repeat here, Shannon [6,7] considers the AWGN channel
and the N,-dimensional coded modulation vectors {S;} constrained only in energy, that is,
\S;|? < N-Ey, = E joules. He shows that
P(E) BZ Nel Ro re) ae 27 Ke log) O(R5/re—1) (12.56)

where the channel cutoff rate is defined by

1 En
Rji455
& 1 + No <—,/1+ i(En,/No)*
(En,/ 0? |
(12.57)
| 1
=f ; log, E (1+,/1+ (En./N0?) bits/dimension

Here Rj denotes the channel cutoff rate for a continuum of possible transmitted signal
points on or within the hypersphere of radius VE =./NcE N,- This derivation assumes
that the vector space generated by the modulator is isometric to the vector space generated
by the encoder. For small values of Ey,/No, where coding is most useful, we note that
R5 ~ (Ey,/No)/21n2 = (reEp/No)/21n2, which is the same approximation obtained for
the channel cutoff rate, Ro, for binary coded BPSK in Section 12.4.1. Operation of the
system at code rate r, = Ro in this energy-efficient region requires E,/No > 21n2.
Figure 12.6 plots Rj versus Ey,/No together with the Ro of (12.54) for the ensemble
of Q-ary codes with equally spaced amplitude levels and an optimum probability assign-
ment. We note that the richness of the continuum of possible signal points from the trans-
mitter results in a larger channel cutoff rate; that is, Ro > Ro for all O > 2. This in turn
leads to a tighter bound on the ensemble error probability, as seen by comparing (12.56) and
(12.53).
Some insight into the physics of the parameter Ey./No will be useful before we
further discuss Fig. 12.6. This can be obtained by using the bandwidth defined via the
Nyquist sampling rate of 2B = 1/T.. Thus, 2Ey,/No = P/NoB & CNR, which can be
interpreted as the receiver input signal-to-noise ratio in a single-sided Nyquist bandwidth of
778 Block-Coded Digital Communications Chap. 12

B Hz. For low input CNR’s, we see from Fig. 12.6 that Ro nearly coincides with Ro for all
Q, but for CNR’s greater than zero dB the binary codes are less desirable. In fact for CNR
> 10 dB binary codes are exceedingly undesirable; however, for high input CNR’s, coding
is generally not required to achieve the desired performance. At low input CNR’s (power
limited channels) where coding is usually applied, binary codes are essentially optimum.
The increase in achievable Ro as Q increases for large CNR’s is due to the fact that the
signal vectors (codewords) that lie inside the signal hypersphere of radius ,/N-En, can
be detected without error, while at low input CNR’s only the signal vectors (codewords)
that lie on or near the surface of this sphere can be detected without error. This behavior
tends to justify the fact that binary codes are optimum for use in channels that are power
limited, that is, those with low values of CNR. For large values of CNR, it should be clear
why Ro ~ log, Q bits/dimension can be made to approach Ro for large Q. In this region
of operation, those codewords in the Q-ary codes which lie inside the hypersphere have
enough signal-to-noise ratio to be detected with essentially the same reliability as the set
of codewords constrained only in energy. In fact, as Q approaches infinity, the discrete
constrained-energy signal points approach a continuum, and Ro approaches Rj. For large
CNR’s, the channel throughput is usually limited by bandwidth so that the application of
coding (which tends to expand bandwidth) is of lesser interest.

12.4.5 Q-ary Coded M-PSK

As discussed in Chapter 4, M-PSK modulation is frequently used to reduce the bandwidth

requirement. Consider now the use of the (M, 2, 7;) modulator k= N- times with T, = T, =
2Ty and M > 3. Assume further that the coded message symbols x, are M-ary; that is,
Q = M. Then the modulator maps each codeword x,, onto the coded M-PSK waveform
Ne
Sm(t) = > Smalt — (2 — 1) TOs a7
n=1

where

Perea |V2Es/Ts cos (wot + 20i/M),

mn .
(n—1)T, <t <nT,
; else
if the random variable x, = i(i =0,...,M — 1), and where E, = E,. For example, if
M = Q=4, Smn(t) models one of the four QPSK signals for each one of the coded message
symbols xXmn € {0, ..., 3} in the codeword (Xm1, ..-,XmN,)-
Following the approach used for Q-ary coded ASK, we again have M, = Q*« mes-
sages to be assigned to Q¢ available signal points, for a total of QNMe possible codes in
the ensemble. Assuming equally probable codewords and transmission of the coded M-PSK
waveform over the AWGN channel, it is possible to show that the ensemble message error
probability is bounded by

PLE] < 2~Ne(Ro-re) — 9—Nel(Ro—Re)/W] _ 2-T(Ro—Ro) (12.58)

l hoe E k
Ro =Ro/W A ao log, FeSy exp(- sin? 4] bits/sec/Hz (12.59)
Sec. 12.4 The Ensemble of Block-Coded Modulation Systems 779

for M > 3. For small values of Ep/No, it is easy to show from (12.59) that

Ro& 4 =
. sm (7!) 2In2 (=) Oe
If we insist on operating the system at Ro > rc, then this requires a communications effi-
ciency of
E
ere in?) (12.61)
No
equivalent to that given for block orthogonal codes in (12.15).
It was shown in Chapter 4 that the M-PSK symbol error probability is well approxi-
mated by

2 wen ot
P(E) 220 (\vA sin ) (12.62)

for E;/No > 1. For large M and E;/No, we have, using the small-angle formula and the
fact that E,; = Ey, for this model, that

P,(E) ~ 20 (ey (12.63)

so that the coded symbol energy-to-noise ratio required to maintain a given error probability
must increase as the square of the number of modulation states, M.

12.4.6 Further Discussion Regarding Ro and Rj

Except for the special cases of M-PSK, orthogonal, biorthogonal, and transorthogonal
coded modulations, we have seen that the problem of expressing the message error prob-
ability performance of a coherent communication system in closed form and evaluating the
expression is a formidable problem. The problem of optimum code selection centers around
being able to determine that signal cross-correlation matrix, namely I’ of equation (4.78),
which minimizes P(E). Shannon no doubt recognized this and in order to gain insight into
the coding problem proposed the astute method of bounding the MEP by using the random
coding techniques presented in the previous sections.
This method uses the pairwise error probability and the union bound to evaluate the
MEP of an ensemble of communication systems which use amplitude modulation and the
class of all possible Q-ary codes to communicate over the AWGN channel. The significant
result, an existence proof, is that this ensemble of communication systems has the ensemble
message error probability union bounded for isometric coded modulation mapping by
P(E) < q—Ne(Ro—re) — 9—Nel(Ro—Ro)/W] _. g—~PG(Ro—Ro/W) _ 9-T(Ro—Ro) (12.64)

where we have used the facts that PG = WT = N, (see Table 12.3). This says that the
ensemble MEP approaches zero as N. approaches infinity provided that Ro = Ro/W >
re = Ry/ W or as T approaches infinity provided that Ro > Ry. From this, we can conclude
that there exists at least one codeword dictionary among the ensemble of Q-ary codes with
MEP less than or equal to the average value. Indeed, such code(s) are of great practical
interest; however, it appears that nature has hidden them and the search for such codes
780 Block-Coded Digital Communications Chap. 12

has led to the development of Coding Theory. Later in this chapter, we investigate specific
block-coding techniques for error control from a system design perspective. In the next
chapter, we will investigate the use of convolutional codes as a coding alternative.

12.4.7 The Bhattacharyya Bound and Distance for Isometric Coded

Modulation

When message m is sent over an AWGN channel, the pairwise error probability has been
shown (in Appendix 3D) to be given by
dm
PGiSi=o (=) (12.65)
for all m 4k =0,1,...,M- — 1. More generally, it can be shown [8, 20] that, for the
arbitrary IF CMC

Ps(Sm, Sx) < / a / JF elSm) FalSpar (12.66)

ee!
N.—fold
when the coder output is isometrically mapped onto the modulator output. For the arbitrary
IF DMC which outputs one of J symbols, Rnj(j =0,..., J — 1), we have
R\(y-1) Rne(J—-1)
Po(Sm,Sk)< D>... D> VPRISn) PRIS) (12.67)
R}=Rio0 Rn-=Rn-0

under the same assumptions, where a vector quantizer produces the quantized random vec-
tor R = (R,..., Rv.) whose coordinates are the quantized random variables R,, taking on
values Rj. Here P(R|S,,) and f(r|S;) denote, respectively, IF channel transition probabil-
ities and probability densities. The bounds in (12.66) and (12.67) are called Bhattacharyya
bounds and the negative logarithms are the Bhattacharyya distances, say By. For the BSC
with crossover probability p, we have from (12.67)

iin (V4pc a ?)) (12.68)

while for the AWGN channel we have, from (12.66)
1 Ne/2 poo fore) ir—S 2 + |r — S|?

aot Palen ee f_e( =2No ;

SN < ara cae aoe fe ee eee d
P S

oe (12.69)

Thus, application of the total error probability formula, the union bound, and the PEP result
from the AWGN channel yields
M-—1 Mc-1
P(E) <(1/Mc) }) }> exp(—dpy/4No) (12.70)
k=0 m =0
m#k

for equally likely messages.

 Sec. 12.5 Shannon's Channel Coding Theorem 781

Now let dmin denote the minimum Euclidean distance as defined in (12.7). Using
this in (12.70) clearly demonstrates the important role dmin plays in the design of error-
correcting codes, that is, P(E) < M, exp (—d?., /4No). Recall further from (12.4b) that the
Hamming distance is proportional to the square of the Euclidean distance for binary coded
BPSK.

12.5 SHANNON’S CHANNEL CODING THEOREM

We now present and discuss the most basic theorem of information theory, the achievability
of the ultimate in channel capacity. The basic argument was first stated by Shannon in
his original 1948 paper [5]. The result is very striking as well as counter-intuitive, for if
the channel introduces errors randomly, how does one correct all of them? Shannon used
a number of key ideas to prove that information can be sent reliably (error-free) over a
channel at all rates up to what he called the channel capacity. These ideas include: (1)
allowing for an arbitrarily small, say €, nonzero MEP; (2) using the channel many times in
succession; (3) evaluating the average MEP over a set of all possible codeword dictionaries
(the ensemble); and (4) showing the existence of at least one good code.

Shannon’s Channel Coding Theorem. The rate r, of an (M;, Nc, T-) code is
log, M.
ro= a bits/dimension (12.71)
Cc

and all code rates below the channel capacity Cy, are achievable. Specifically, for Gaussian
channels and M, sufficiently large, an optimum receiver, and code rate
1 1 ; : :
re< Cy, = 5 log» ( + )= 5 logs[1 + 2r-(Ep/No)] bits/dimension (12.72)
No
there exists an (M., N-, T-) code with MEP approaching zero. The restriction re < Cy,
implies that Ep,/No > (22re — 1)/2r. for all r. < Cy,. If re > Cy, and M,¢ is sufficiently
large then P(E) approaches one for every possible (M., Nc, T-) code. It can be shown [7]
also that 5C Ne < Ro < Cn,. The code rate inverse rks 4 1/r, is frequently identified as the
bandwidth expansion factor by system engineers.
The proof of this famous theorem is given in numerous books on information theory
[7, 8, 9] and in Shannon’s original paper [5]; however, it is not our purpose to present this
important proof here since it is too lengthy to reproduce. Shannon’s approach to proving
this famous theorem included the modulator and demodulator as a part of the channel; that
is, he used the baseband channel model to prove the theorem. This is equivalent to assuming
an isometric mapping from coder output to modulator output when the IF channel model is
used. In this sense a codec and modem exist such that the theorem is true.

12.5.1 Shannon Limits and Information Transfer Physics

In Shannon’s famous paper [5], it was shown that all information sources, for example,
produced by people speaking, television cameras, telegraphy key, and so on, have a rate
(r-) associated with them which can be measured in bits/sec. Communication channels have
a capacity (Cy,) that can be measured in the same units. Shannon’s theorem states that a
codec-modem exists such that information can be transmitted essentially error-free over the
channel (baseband or IF) if and only if the rate does not exceed the channel capacity. To
782 Block-Coded Digital Communications Chap. 12

gain further insight, let us investigate the communication efficiency requirement implied by
(12.72) for the AWGN channel. Since Ey,./No = rcEp/No, we can write

1
re=< Cy, = 5 logy(1 + 2r-Ep/No) (12:73)

Thus

Eb _ (22m — 1) /2r, forre < Cy, (12.74)

No
Operation at r- = Cy, implies that
E
ORs 1) Ge. (12.75)
No
and in the limit as the code rate r, approaches zero (N; approaches infinity), we have
E
—? 5in2=-1.6dB forr,
<Cy, (12.76)
No
This compares with
E>
iis >2In2= 1.4 dB for re < Ro C277)
0
as given in Section 12.4.2 for binary orthogonal codes; note the 3 dB difference. From a
system perspective, this says that the best codec-modem technique will require an efficiency
of E,/No > —1.6 dB, a large coding delay, and complex coder-decoder (codec) imple-
mentation; but for small error rates of practical interest and a reasonable coding delay and
codec complexity, E,/No = 1.4 dB is required. In conclusion, since No = kT°, we see from
(12.76) that

E,/°K>kin2 _ forre < Cy, (12.78)

and, we see from (12.77) that

Ep/°K > 2k\n2 =1.91 x 10773 joules/°K for re < Ro (12.79)

so that a 3 dB difference exists between Ro for binary orthogonal codes and Cy, for all
codes in the region of small code rates.
We can gain further insight into information transport by transforming the units of
Cy, from bits/dimension to the more useful physical units of bits/sec. Since
En.= E/Nce= PT, = PT/Nz joules (12.80)

then, with N. = 2BT per our bandwidth discussion of Section 12.3.4, we write

2E 2PT P
NoNN SAEED
No(2BT) NoB~A CNR (12.81)
Thus, we have a physical interpretation of the parameter 2E y,./No as merely the input signal
power-to-noise power ratio! Substituting the above relationship into (12.72), we see that the
temporal channel capacity is given by
P
C ACy,/Te = B log, | 1 + —— } bits/sec (12.82)
NoB
 Sec. 12.5 Shannon's Channel Coding Theorem 783

P
Sareeae
NoIn2 bit
(bits/sec)

(bits/sec)
C

P/N,
B (Hz)

Figure 12.7 AWGN channel capacity versus bandwidth B = 1/27;

Equivalently, Cy. = C/ W bits/dimension = bits/sec/Hz. A plot of the capacity C versus B

is illustrated in Fig. 12.7.
It is now interesting to see what happens as the single-sided bandwidth B = 1/27,
Hertz approaches infinity, that is, 7. — 0. Now

Coe lm C= hin 10 ee
Sey ees B00 82 No B
(12.83)

= log, eP/No — Pere bits/sec

No In2

The parameter Cy is the so-called infinite bandwidth Shannon capacity limit for the AWGN
channel. Furthermore, the maximum amount of information that can be transported error-
free in T sec is
|RR aie
Le Col = Noin2 bits (12.84)

The parameter /,o is the so-called infinite bandwidth Shannon information transport limit!
In practice, most radio communication systems are designed to operate with 10 dB-Hz
< P/No < 120 dB-Hz. Thus, with P/No = 120 dB-Hz, Io = 10!2/ In 2 bits of information
can be delivered error-free in one second; this is more information than stored in the Library
of Congress.
One question that remains as a consequence of the existence proof is to specify the
ultimate codec, modem and their implementation. Unfortunately, nature seems to have hid-
den the particular coded modulations; they remain to be the slippery eel which motivates the
further development of coding theory from the perspective of coded modulation.
784 Block-Coded Digital Communications Chap. 12

12.5.2 The Channel Capacity, Matched Filter, and Spread Spectrum

Communications Interconnect
For the bandlimited AWGN IF channel, there exist a codec and a modem which allow one to
communicate error-free provided that the rate r- = logy M-/Nce = log, Mc/WT = Rp/ W is
less than the capacity
1 2En
Cy. =——-—]
5 082 (1+ No : ) (12.85 )

in bits/dimension where the available energy-per-dimension to noise density ratio for each
coded message bit is
P 2P P
SENS ee yw Sets ae ee es ONY (12.86)
No No NoW NoB
and W is the double-sided Nyquist bandwidth. Thus, (12.85) can also be written as
1
Oe 5 log,(1 + CNR) (12.87)

As before, converting Cy, to bits/sec requires that one normalize Cy, by T,; that is, define
C2 Cy.) dz,,so that
Cy.
CA= oe = B log,(1 + CNR) bits/sec (12.88)

SINCE. 2 Beale
Let us now relate the capacity theorem to the matched filter theory developed in
Chapter 4. We note that the input signal power to noise power ratio can be written as
22 fogs Tt |
CNR Litesbe Pa eee (12.89)
NoW No WT), No WT,
Now, with the 7p-sec processing gain defined by PG, A WT), and noting that the output
signal power-to-noise power ratio of a matched filter is pp,; = 2E,/No, then
2Ep
= (PGp) (CNR) = (PGp) pin (12.90)
0
where ~jn = CNR is the input signal-power-to-noise power ratio to the matched filter. Thus

C—o5 logs(1 ae Pout} P Gp) (12.91)

serves to relate the “matched-filter-theory” to the channel capacity formula. Moreover

Pe
lim CA¢ja= 5 bits/sec (12.92)
B>oo og in

so that in T sec, the maximum amount of information that can be transported error-free is
E
Noin2 bits!
its (12.93)

In summary, there exists a codec-modem design such that error-free communications

can take place (even in the presence of noise) provided that the channel throughput

(12.94)
Sec. 12.5 Shannon’s Channel Coding Theorem 785

This requires that the communication efficiency E,/No satisfy

Ey _(2Re/®
No >
RaW
en
— 1) (12.95)

while operation at R, = C requires that

Ey (22e/W ase)
AAC =
4 (12.96)
No 2C/W
Thus, three basic parameters which define the ability to communicate error-free over the
bandlimited AWGN channel are E,/No, R,/ W, and C/W. The region of C/B < 1 is called
the power-limited region of operation and the region C/B > 1 is called the bandwidth-
limited region. The boundary between the two regions occurs at CNR = 1. The greatest
communication efficiency (minimum E;/No) is achieved when the bandwidth can be made
large in comparison with the information rate R». In the limiting case of very large band-
width, and operating at C = Rp = 1/Tp, then from (12.94) and (12.96), C/B approaches
zero and E,/No approaches the Shannon limit, In2 = —1.6 dB. Thus, by invoking the cod-
ing theorem, one can say that for E,/No > —1.6 dB, the probability of message error in
information delivered to the user can theoretically be made arbitrarily close to zero by use
of a suitably chosen error-control code.
The fact that pou; = PG», - CNR = PGppin in (12.90) is very interesting. For exam-
ple, if the channel input signal power to noise power ratio is CNR = —20 dB (signal power
one hundred times smaller than the noise power) with PG, = 1000, which is easily attain-
able, then 2E,/No = 10 dB! Using this together with (12.88), we can relate the notion of
information transfer efficiency, C/B, to the communication efficiency, E,/No. In fact, for
small CNR, we see that C/B ~ CNR/In2 = (P/NoB)/\n2 = 2(Ep/No)/( PG» In 2).
Shannon’s formula for the capacity of the bandlimited AWGN channel (baseband or
IF) was one of the earliest results found in information theory [5]. This remarkable the-
orem demonstrated to system engineers that enormous improvements in the performance
of digital communications systems could be achieved. The initial era of excitement (circa
1950) was followed by an era of discouragement since practical schemes which realized the
improvement were difficult to find. This feeling was summarized in an often quoted folk the-
orem [16] which has been paraphrased as: “All codes are good, except those that we know
of.” For example, voice grade telephone channels can be modeled as bandlimited AWGN
channels. The capacity of typical channels exceeds 20,000 bits/sec and the commercial in-
centive for achieving the highest possible data rates has a high dollar value. Yet 20 years
after Shannon, the upper practical limit was about 2,400 bits/sec. Thirty years after Shan-
non, the upper practical limit was about 9,600 bits/sec and this advancement had little to do
with Shannon. Ironically enough, this improvement over telephone channels was primarly
due to the implementation of digital adaptive equalizers.

12.5.3 Spread Spectrum Communications

Spread spectrum communications is motivated by the need for reliable communications in
the heat of battle. The basic concept can be summarized as follows: Suppose we transmit
an N--dimensional signal in the presence of a jammer with finite power J Watts. If this
signal is disguised as an N.-dimensional signal, N./ > Nc, then the jammer must distribute
his power over N. dimensions and his power will be effectively reduced by the factor
N-:/Nc. If the jammer’s power is J Watts, the average single-sided jamming noise power
 786 Block-Coded Digital Communications Chap. 12

density is Ny = J/B Watts/Hz where W = 2B Hz is the spread spectrum bandwidth. The

jamming power is effectively added to the noise No from the channel. Now since in the
jamming channel, Ny >> No, we usually approximate the spread spectrum communication
efficiency Ep/(No + Ny) by Ep/Ny. Since Ey = PTp = P/Rp, we have the relationship
J/P = (B/Rp)/(Ep/Ny) = PGp/(2Ep/Nz7). The quantity J/P represents the received
jamming power to signal power ratio. In antijam systems, it is desirable to have J/P as large
as possible so as to minimize the vulnerability to jamming. From the relationship for J/P
we see that this can be accomplished in two ways: one can increase the processing gain,
PG) = WT), which in essence increases the dimensionality of the transmitted signal, or
decrease the required E,/N, for a particular bit error probability. In practice, the jamming
margin can be 30 to 50 dB, or more.
We now return to the concept of channel capacity and relate it to spread spectrum
communications in the presence of jamming. From (12.82) we have

€ = tog, (1+ =)
Cc P
bits (12.97)
which says that C/B ~ (P/J)/1n2 if the jamming margin is large (J/P > 1), a case of
great practical interest.

12.5.4 The Union Bound for Binary Orthogonal Codes Written in Terms
of the Shannon Capacity Limit

Using the fact that 2* = e*!"*, (12.12) can be rewritten in terms of the Shannon capacity
limit, namely

<2 Ry)
P(E) <2 (F (12.98)
which says that, as long as 0 < Rp < Coo/2, then P(E) can be made to approach zero by
making T large. This remarkably simple expression serves to upperbound (4.80); unfortu-
nately this bound is not very tight.

12.6 GALLAGER’S UPPER BOUND ON THE ENSEMBLE ERROR PROBABILITY

One major purpose of information theory has been that of understanding Shannon’s coding
theorem by developing upper and lower bounds on the decoded error probability. Tremen-
dous insight into the problem of transmitter design (coding trade-off) has been obtained
using various error bounding methodologies on the ensemble of block codes defined in Sec-
tion 12.4. It is not our purpose to develop this detailed theory here; rather, we will present
Gallager’s bound [8] and discuss its meaning from a system perspective. The reader who is
interested in the details associated with the development of this bound should consult Gal-
lager [8]. A special case of his bound is the Bhattacharyya bound.
We wish to gain insight into the minimum probability of decoding error as a func-
tion of the code rate r,, the block length N,, and the channel, by analyzing the performance
achievable with the ensemble of block codes discussed in Section 12.4. This peculiar ap-
proach is dictated by the fact that there appears to be no systematic way to find block codes
that minimize the probability of decoding error, P(E), perhaps with the exception of those
treated earlier in the chapter.
Sec. 12.6 Gallager’s Upper Bound on the Ensemble Error Probability 787

Gallager has developed a tight bound on the ensemble error probability P(E) for
block codes. In particular, consider the case of the Q-ary coded (M, 1, T;) modulator used
k = Nz times per codeword to transmit the class of Q-ary codes where M = Q, as discussed
in Section 12.4.3. For this case, he has shown that

P(E) < 27 NelEr(re)] (12.99)

where E,(r-) is the random coding exponent defined via

E,(rc) & max, EO Q) — pr-¢] bits/dimension (12.100)

<p<
and where Q denotes an arbitrary probability assignment for the coordinates of the coded
modulation vector, which are chosen independently with probability P[sjn = q,,/E woke
px. If OQ = M =2 (binary coded BPSK), for example, we have

= JEN.) = po, k=0

(smn RY
Ok) = IeP (Sa EN.) =pi, k=1,

where the probability is computed over the ensemble.

For the IF CMC model defined in Section 11.2.2
I+p
co f 2 ane
Eo(p, Q) = — log, / 2° (rtae =a Ex)" | dr, |} (12.101)
So

for all observables r,,n = 1,..., Nc, and all messages m =0,..., M, — 1. Here f (rn|smn
= qk JEN.) denotes the known channel transition probability density function (which char-
acterizes the channel) for the n‘ h received observable, given that message m is sent.
For the IF DMC (quantized IF CMC) model defined in Section 3.1.1, Section 11.2,
and used in Section 11.2.4
sea ge a l+p
Eo(p, Q) = — log, | 3? [A= Rate En.|' “ (12.102)
+

j=0 Lk=1

for all quantized observables Rp, n=1,..., Nc, and all messages m=0,...,M. — 1.
Here P(Ry = Rnj|Smn = gkx/En,) = prjA& P(j|k) denotes the known channel Panction
probability for the n’ h received quantized observable (that is, the quantized version of ry),
given that message m is sent.
For modem and codec design purposes, it is of interest to connect E;(r-) with the
channel cutoff rate Ro for the IF CMC model and with the quantized Ro for the IF DMC
model. This is accomplished using Gallager’s function Eo(p, Q) and setting o = 1, that
is Ro = Eo(1, Q). In regions of low Ey,/No, then E;(rc) is maximized at p = 1 [20], so
E;(re) X (Ro — re) for equally likely, independent message symbols. For the IF CMC we
have, from (12.101)

Ro = Ro/W = Eo(1, Q)
2
oo f @
= — log, if bsPka| f (Fabs ~ alex) dry | bits/sec/Hz
—0oO k=1
788 Block-Coded Digital Communications Chap. 12

while for the IF DMC model, we have, from (12.102)

Ro = Ro/W = Eo(1, Q)
2
J=1 (29
=— log, a bsDEAL. (Rr = Rnj|Smn = 4k Fx) bits/sec/Hz
j=0 Lk=1

where the channel transition probabilities for the quantized observables Ry are related to the
channel transition probability density functions f (7n|Smn = Gk./ En.) through

1 (fn — Qkx/En,)?
j
for Q-ary coded ASK and A; is the observation region which gets mapped into the observ-
able Rj for all j =0,1,..., J — 1 (see Fig. 11.6). This defines the connection between the
IF CMC and the quantized IF CMC (IF DMC).
Properties of the random coding exponent E,(r-) have been studied with much suc-
cess; from our system perspective, one important feature is that E,(r-) > 0 for allO <r, <
Cy,. Thus, (12.99) says that for appropriately chosen codes and modems, the probability
of decoding error can be made to approach zero with increasing block length N, for any
rate less than capacity. In expression (12.100), E,(r-) is a positive function of the channel
transition probability via Eo(o, Q). This function yields a fundamental characterization of
the channel for coding purposes. It brings the relationship between decoding error proba-
bility, data rate, block length, and channel conditions into focus. In fact, as we shall later
see, the question of what modulation-demodulation techniques one should use when coding
is applied to system design can be addressed on the basis of E;(r-). Moreover, since the
ensemble of block codes satisfies (12.99), it follows that at least one block code in the en-
semble must have P(E) no greater than the ensemble average P(E). From a system design
perspective, this bound is not useful in accessing design point requirements with respect
to communications efficiency (E,/No), supportable data rate (72,), or power requirements
which achieve a specified error probability. The insight provided is philosophical in the
sense that a code(s) exist(s) such that if r. < Cy, then the error probability can be made
arbitrarily small by employing large block lengths N,. Unfortunately, the complexity of the
encoder and decoder which accommodate long (or short) block lengths is not addressed by
this interesting bound. In fact, the complexity issue remains to be the important unsolved
problem in information theory. The random coding exponent versus rate curve is of inter-
est to the system engineer, especially for comparing the Ro values corresponding to specific
codec-modem pairs. We will consider an example of how such a comparison can be useful in
the design of the demodulator in Section 12.7.5, for binary coded DPSK. The general char-
acteristics of Ro are discussed extensively in Gallager. However, for our purposes, we are
particularly interested in its behavior for very noisy channels, the case of greatest practical
interest.

12.6.1 Block versus Convolutional Coding for the Very Noisy Channel

Without elaborating the details, let us consider the class of very noisy channels (low values
of E»/No, where coding is of interest) and present the random coding exponent versus rate
curve E,(r-) which results and discuss the implications from a coding viewpoint. For block
 Sec. 12.7 Demodulator Design Using the Ro Criterion 789

E,(r,)

Ne Convolutional Codes

Block Codes

CN,

Figure 12.8 Random coding exponent versus code rate for the very noisy channel

codes, it can be shown that [8]

2 Cy,
ast (/Cn, — V7 ; os Tig.EGON:
E-(rc) = ol ; c) aa ; a Cn, (12.103)
2 (on) aes CE. 4

Figure 12.8 demonstrates plots of the random coding exponent given in (12.103) versus
code rate. For comparison purposes we also demonstrate the random coding exponent of
convolutional codes (to be discussed in Chapter 13) found by Viterbi [10]. From a systems
perspective, we conclude from Fig. 12.8 that for a given code rate, there exist convolu-
tional codes which yield a lower error probability than that of block codes. (This statement
assumes a bandwidth unconstrained, memoryless channel.) Notice too that the maximum
value of E,(r-) predicted by this bound is Cy,,/2. The result in (12.103) was derived using
the baseband system architecture which absorbs the modulator and demodulator as parts of
the channel model.

12.7 DEMODULATOR DESIGN USING THE Rp CRITERION

We now make use of Gallager’s bound (12.99) and develop a philosophy for the design of
the modulation-demodulation system in a coded BPSK system with equally likely messages.
Practical demodulator outputs, r, are continuous random vectors, that is, the projections of
r(t) onto the signal set basis functions by the combined demodulator-detector. Moreover,
digital signal processing algorithms require that the observable r be vector-quantized into
say J = 2" levels.
790 Block-Coded Digital Communications Chap. 12

To develop a quantizer design philosophy for processing the demodulator output, we

again consider the special case of Eo(o, Q) when p = 1. Then, with Ro = Eo(1, Q), the
ensemble error probability bound in (12.99) reduces to
P(E) < 27Ne(Ro-re) — 9-T (RoR) (12.104)
when we use (12.100). The number Xo not only establishes a limit on the rate at which one
can communicate with arbitrarily small error probability, but also gives an exponent of error
probability. No other communication parameter gives such insight into the communication
process—not even the channel capacity, Cyv,, which only establishes a limit on the rates at
which reliable communication is possible. The “unquantized Ro” given by (12.101) with
p = 1 is also the computational cutoff rate of sequential decoding (see [7], Chapter 6), that
is, the rate above which the average number of decoding steps per decoded symbol becomes
infinity.
The idea that the quantizer design used in the demodulation system can be based
on Rg for a coded digital communication system was first suggested by Wozencraft and
Kennedy [1]. A modem design based on the fact that r is the input to the decoder is im-
practical because operation of digital hardware requires that r be quantized. Therefore,
the unquantized Ro = Eo(1, Q) found from (12.101) yields an upper bound for the cut-
off rate that is unachievable by any practical demodulator. On the other hand, the quan-
tized Ro = Eo(1, Q) found from (12.102) can yield an upper bound for the case of a finite
demodulator output alphabet which is almost as tight. In other words, the choice of the
modulation-demodulation quantizer design to be used in a coded system is not based on
the error probability (symbol, bit or message) derived in Chapters 3 through 10; this er-
ror probability serves only as input to the computation of Ro. The choice of the isometric
modulation-demodulation technique and quantizer design should proceed on the basis of
maximizing Ro [as seen from (12.104)], as opposed to minimizing the M-ary symbol er-
ror probability of an uncoded system. In the next section we will demonstrate this system
engineering design philosophy for coded BPSK for hard-decision and soft-decision demod-
ulators assuming that the IF CMC has been converted to the IF DMC using the methodology
presented in Sections 11.2.3, 11.2.4, and 12.6.
Finally, from a system engineer’s perspective, the Ro parameter can be considered to
be characteristic of both the codec (since it is a function of r.) and the modem. For hard
decisions, we will see from (12.107) that Ro is a function of the coded message bit error
probability, p, given by Q (,/2En,/ No) for binary coded BPSK. In this sense, the bit error
probabilities derived in Chapters 3 through 10 for uncoded systems serve as a tool in the
design of coded systems.

12.7.1 Design of Hard-Decision Binary Coded BPSK Demodulators

Consider first the IF DMC with binary coded BPSK and with equally probable messages
mo = 0 and m, = 1. The rate Ro, found from (12.102) with p = 1 for soft-decision demod-
ulation, reduces to
l
Ro = — log, {5 (12.105)

qe]
x 1+ >> P (Rn = Rujlsnn = VEN.) P (Rn = Rnjlsnn = —VEn.)
j=0
Sec. 12.7 Demodulator Design Using the Ro Criterion 791

for J = 2" output quantization levels or L bits of quantization. This compares with the
capacity of the same IF DMC [8]
2e-1
Cy. = Ss [P (Rn = Rajlsmn = —VEn.) (12.106)
j=0

x log 2P(Rn = Rnj|Smn AY En,)

Dy —_—_—— <<< <<_<—_- ern ooor———————_—_——_—_—______—

P(Rn = Rnj|Smn =—,/En,) + P(Rn= Rnj|Smn = /En,.)

For hard-decision demodulators (L = 1), (12.105) reduces to

1
Ry= log) E (1Ee Apr »)| (12.107)
for the BSC with crossover probability p of the uncoded message bits. The quantized Ro
given in (12.107) for hard-decision decoding compares with the channel capacity of [8]

Cy, = 1+ plogy p+ (1 — p) log,(1 — p) (12.108)

for the same BSC. Figure 12.9 demonstrates the channel symbol error probability p required
to operate at rates of r, = Ro and r, = Cy, using (12.107) and (12.108) versus code band-
width expansion r- ' — 1/r,. For example, from Fig. 12.9, we observe that for hard-decision
demodulators (one-bit quantizers) and a code rate r¢, it appears that coding is only useful
when the channel symbol error probability p is less than 0.11.
For BPSK modems operating in the AWGN channel,

p = O(/2En,/No) = QV 2rcEp/ No).

Notice here that p not only depends on the modulation but also on the code via the code
rate. For this case, the limiting value (small Ey./No) of r- can be increased from that of
(12.107) to the limiting value of the unquantized error exponent given by Ro in (12.37),
with a 10 log; (7/2) ~ 2 dB improvement in the Ey./No efficiency of energy utilization.
This is the 2 dB loss due to the two-level quantization process required for hard-decision
decoding.

12.7.2 Design of Soft-Decision Coded BPSK Demodulators

As we have mentioned, the Ro criterion is natural for use in the design of the modem as
it can be optimized as a function of the number J of quantization levels, allowing the
system engineer to trade off system performance with the implementation complexity of
the decoder. For example, in making this trade, the engineer would examine the quantized
Ro attainable with J = 2, 4, 8, 16, and so on, until he has obtained an Ro sufficiently close
to the unquantized (J = 00)Ro which justifies the decoder cost and complexity versus the
improved bit error probability performance. As we have already pointed out, the absolute
upper limit on the code rate r, is the channel capacity Cy,, and the upper limit for all
practical purposes (considering cost and implementation complexity) is the computational
cutoff rate Ro; the unquantized Ro appears to be approximately 3 dB smaller than Cy, for
the cases of greatest interest.
The procedure used in the early design of modems for selecting the quantization cell-
size parameter A demonstrated in Fig. 11.6 (see Sections 11.2.3, 11.2.4, and 12.6) was to
choose it to minimize the E,/No required to operate at a code rate of Ro bits/dimension.
792 Block-Coded Digital Communications Chap. 12

System operation
atr,=R ‘

System operation
atr, = Cy
€

Figure 12.9 Coding limits for the BSC

1 2 3 4 5 6 7 8 9109 _ Using hard-decision demodulation of
binary coded BPSK or QPSK modulation
c (Reprinted from [17])

This philosophy is intuitively justified since one is attempting to maximize the coding gain
(the reduction in E,/No required when coding is used to that required in an uncoded BPSK
system for the same BEP) when operating at rate r, near Ro. Usually computer simulations
of the coding system are required to make this trade, and the results have shown good
agreement with those determined using the theoretical considerations due to Massey [2]
which we now describe.
Consider the case of binary codes with equally likely message symbols using BPSK
modulators, that is, Sinn =./En, or —,/En,, and J output symbols; see Fig. 11.6 for
the case L = 3 and J = 8. Recall that each received observable r,, must be assigned one
of J quantization levels Rnj € {Rno, .--, Rn(y—1)}. To do this, the quantizer needs (J —
‘ 1) thresholds 7;, 7 =0,..., J — 2. Here T; denotes the threshold between two quantizer
Sec. 12.7 Demodulator Design Using the Ro Criterion 793

T) T, T,
ee bere os peed

Figure 12.10 Quantization regions and levels for the case J = 4

output levels Rn; and Ry(j+1). The sketch in Fig. 12.10 illustrates the decision regions or
quantization levels for the case J = 4.
Our goal is to select these threshold levels in an optimum way; specifically, we seek
a selection that will maximize the quantized cutoff rate Ro, given by (12.105) and repeated
below for convenience:
1
Ro = — log, {> (12.109)

aE
x 1+) > P (Ru = RnjlSnn = VEN.) P (Rn = RnjlSnn = —VEn.)
j=0
Hence the demodulator design problem is to choose each threshold 7; such that
dRo Nes
0 (12.110)
ony
Applying the above equation, Massey found that the optimum threshold levels were given
by
T; = ,/ A(Rnj)A(Rn,j+1) (12.111)
where the likelihood ratio is given by

P (Rn = Rnj\Smn = V En.)

LAA BS goo pert ener ceca ne (12.112)
les(Rn = Rnj|Smn = —V En.)

for a given output quantization level Ryj, j =0,..., J — 1. This says that each optimum
threshold is the geometric mean of the likelihood ratios for the quantizer outputs whose
decision regions the threshold value separates.
Odenwalder [17] considers the use of equally spaced thresholds at 0,+7,...,
+(24—! — 1)T, with the spacing A; =T chosen to maximize Ro. For L = 3 (3-bit quanti-
zation) and Ey,/No = 1.5 dB, we obtain threshold spacing T of roughly .5 or .6 times the ©
standard deviation of the unquantized demodulator outputs [17]. Decreasing (or increasing)
En,/ No results in only a minor decrease (or increase) in 7. Thus, a fixed value for T can
be used over the entire dynamic range of Ey./No, and an automatic gain control (AGC)
circuit can be used to estimate the noise variance No/2. System performance is not sensitive
to small variations in the AGC output, and the value of T is usually selected for the worst
case value for Ey./No or Ep/No.
 794 Block-Coded Digital Communications Chap. 12

1-bit quantization

2-bit quantization, T = 1.0

3-bit quantization, T = 0.5

no quantization

E./No
dB
,

1 2 3 4 5 6 7 8 910
Bandwidth Expansion = 1/r,

Figure 12.11 E),/No required to operate at re = Cy,, for coded BPSK modulation (Reprinted from [17])

12.7.3 Choice of the Number of Quantization Levels

The results presented so far in this section are now combined to demonstrate the effects
of the number of quantization levels on demodulator performance. We assume that coher-
ent demodulation of coded BPSK or QPSK is used in the AWGN channel. First, we con-
sider the problem of demodulator design when we wish to operate at code rates re = Cy..
Equation (12.72) for infinite quantization and (12.108) for 1-bit quantization establish this
performance by relating the bandwidth expansion r- ' — 1/r, and the required communi-
cation efficiency to operate at r- = Cy. Plots of this relationship are demonstrated in Fig.
12.11. In addition, comparable results found from (12.106) for soft-decision demodulators
are demonstrated for 2- and 3-bit quantizers with equally-spaced threshold levels.
Figure 12.12 gives the corresponding curves of required E,/No versus 1/r, for system
operation at r, = Ro using (12.37), (12.105), (12.107) and equally spaced threshold levels.
From Figures 12.11 and 12.12 we see that 3-bit soft quantization is almost as good as
- infinite quantization (L = oo), and that 1-bit quantization (hard-decision) requires about
2 dB more E»/No than infinite quantization for comparable performance. It can also be
concluded by comparison of Figures 12.11 and 12.12 that for small values of required
E»/No corresponding to large bandwidth expansions, approximately 3 dB more Ep,/No is
required to operate the system at r- = Ro than is required to operate at re = Cy,.
 Sec. 12.7 Demodulator Design Using the Ro Criterion 795

1-bit quantization

2-bit quantization, T = 1.0

3-bit quantization, T = 0.5

no quantization

E,,/No
aB,

1 Z 3 4 5 6 7 BO) 1
Bandwidth Expansion = 1/r,

Figure 12.12 £;/No required to operate at r, = Ro for coded BPSK modulation (Reprinted from [17])

12.7.4 Coded DPSK and DDPSK Modems with Hard Decisions

For binary coded DPSK modulation, p = (1/2) exp (—r-Ep/No) is the coded message bit
error probability as given by (7.14). For binary coded double differential phase shift keying
(DDPSK), the expression for p is given by (8.58). In both cases, we have assumed optimum
demodulators and no frequency offsets. Furthermore, the Ro for both coded DPSK and
coded DDPSK with hard decisions is given by (12.107). Figure 12.13 illustrates the E,/No
required to operate at code rate r, = Ro versus bandwidth expansion 1/r, for coded BPSK,
DPSK, and DDPSK. These curves reveal the minimum value of E,/No for which coding is
useful for each of the coded modulation techniques. From the DDPSK curve, for example,
we see that if the E,/No is less than 10.45 dB, no practical code exists at any rate that will
allow us to achieve arbitrarily small bit error probabilities. Figure 12.13 also illustrates the
E,,/ No needed to achieve a bit error probability of 10~> without coding for each of the three
modulation schemes; these curves enable us to compute achievable coding gains. From the
DDPSK curve, for example, we see that if r- = 1/2, then the needed E,/No is 10.45 dB.
Thus, the maximum coding gain achievable (for BER = 10~> and hard-decision decoding)
is 3.55 dB. To achieve a greater gain we must increase the complexity of the receiver to
allow for soft decisions.
We further note from Fig. 12.13 that for both coded DPSK and DDPSK it is not the
 796 Block-Coded Digital Communications Chap. 12

10°5 DDBPSK, no coding

10-5 DBPSK, no coding

dB
Ey/Np,

Bandwidth Expansion = I/r,

Figure 12.13 E,/No required to operate at re = Ro for hard decision decoding of

various coded modulations (Courtesy of Debbie van Alphen)

case that decreasing the code rate (say by increasing the redundancy of the code) automat-
ically increases the achievable coding gain, as was the case for BPSK. This is because the
energy in each coded message bit is given by Ey. = r-Ep; thus, decreasing the code rate r,
decreases the channel symbol energy, effectively making the demodulation phase reference
noisier. This effect is not evident in the case of BPSK because a perfect phase reference is
assumed [17]. Both DPSK and DDPSK appear to achieve maximum coding gain with code
rates near r, = 1/2.

12.7.5 Binary Coded DPSK with Infinite Quantization

We now consider the case of binary coded DPSK with infinite quantization (L = oo); that
is, we assume that the demodulator output random variable is handed over to the decoder
unquantized. For this case, Odenwalder [17] has shown that

1 om ;
Ro = — log, E (:+/ Vy f(vlA! = 0) Fwle =mav)| (12.113)

‘where f (y|0! = 60) denotes the conditional probability density function for the demodula-
tor output random variable y, given that the information phase angle 6! = A"6; is equal to
69. (See Chapters 7 and 8.) Clearly, the form of the conditional probability density function
:. depends on the detector implementation, via y.
In order to illustrate our thesis that the Ro criterion should be used to design the
Sec. 12.7 Demodulator Design Using the Ro Criterion 797

modem in a coded system versus using the post-demodulation bit error probability sug-
gested for the comparable uncoded system, we shall consider two demodulation techniques,
namely the optimum I-Q demodulator and the suboptimum autocorrelation detector (ACD)
demodulator (delay received signal one coded message bit and multiply; see Fig. 7.18). In
the case of the autocorrelation detector, the demodulator output random variable w is equal
to the sampled output of the lowpass filter, and the resulting Ro is given in [17]. For the
optimum I-Q detector the random variable w; is equal to the difference of the two adjacent
transmitted phase angle measurements, nj — ni—1. The conditional probability density func-
tion f (|! = 0) is given by (7.3), while f(W|6! = 7) is simply a shift of the same density
function by z(mod 277).
In Fig. 12.14, we plot the E,/No necessary to operate at rate r, = Ro versus band-
width expansion for binary coded DPSK with infinite quantization using both the ACD and
I-Q demodulators. For comparison, we also show the curve for binary coded DPSK with
hard decisions using the I-Q demodulator. It is very intriguing to note that, while in the
absence of coding the I-Q demodulator is known to be optimum, with coding the autocor-
relation detector (ACD) actually gives better Ro results. Thus, for binary coded DPSK, it is
preferable to use an autocorrelation detector rather than an I-Q detector. This confirms the
thesis that, from a communication efficiency perspective, the Ro criterion is better to use
in the design of a modem for coded systems; the post-demodulation bit error probability is
better to use in the design of an uncoded system.

12.7.6 Optimization of Codes and Modem Design Using the Ro Criterion

In (12.105) through (12.108), we have presented the cutoff rate Ro and capacity Cy, for soft-
and hard-decision binary coded BPSK demodulators. We know from Shannon’s theorem
that channel coding will be of no value unless r, < Cy,; we also assert that, for most
practical purposes, operation at r- < Ro will be required in order to realize the decoder
mechanization with acceptable hardware complexity.
As in Section 12.4.3, we now consider the use of the (M, 1, 7;) modulator with
k= N, to transmit M = Q-ary (M-, N-, T-) codes. This implies that the Q-ary coded ASK
modulator is used k = N, times for the transmission of a codeword. If we use (12.54) and
optimize over the a priori probabilities of the coded message symbols, we have

Q
Ro < - min log bsye PjPk exp (-aivens400) (12.114)
j=l k=l

where the minimization is taken over all possible probability distributions P = {pj}, j =
1,..., Q, on the coded message symbols. (Notice we have assumed an isometric mapping
Beaeen the coder and modulator.) Since the minimization occurs when p; = 1/Q for all /,
then one can show [2] that (12.114) reduces to

+(0 Do (3-2)
ore mae
(12.115)
fy sop O— Wor, |1

where the equality holds only for equally likely Q-ary transorthogonal (simplex) signals and
where Ey, denotes the average energy per codeword dimension:
798 Block-Coded Digital Communications Chap. 12

4B
E,/No,

1 2 3 4 5 6 7 8 9 10
Bandwidth Expansion = 1/r,

Figure 12.14 £,/No required to operate coded DPSK at r, = Ro (Courtesy of Debbie van Alphen)

hae
En.
= 5 D4
Oa
This inequality can be used to gain considerable insight into the design of coded modulation
systems for low and high values of En,/No.
Now for transorthogonal (simplex) signals, the equality in (12.115) holds and we
differentiate with respect to x = Ey,./ No to obtain
=—- =
dRo _ Q(En./No) a
Te ae) sse( x0 —1) +@Q J (12.116)

which decreases monotonically with increasing x. Thus, for small values of x, where coding
is most useful, (12.116) yields

dR at
ee ~ 1/21In2or Ro ~ (Ew, /No)/21n2 (12117)
which is independent of Q! This demonstrates that one should use binary coded BPSK
(Q = M = 2) in the design of coded modulation systems since this technique is easiest
 Sec. 12.8 Linear Block Codes 799

to implement. In this sense, the optimum value for Q is Qopt= 2. From a systems engi-
neering perspective, this value for Q simultaneously implies two things: (1) it says that
M =2= Qop which implies that the (2, 1, T;) binary coded ASK or BPSK modulator is
optimum, and (2) that the binary transorthogonal (simplex) code (M, = 2K¢, N. = 2Ke — 1)
is simultaneously optimum! Further, with Q = 2 in (12.115), we observe for transorthogo-
nal codes and the (2, 1, T;) coded BPSK modulator that

Ro = logy[2/(1 + exp (—Ewn,/No)]

which agrees with (12.37), as it should.
Finally, from (12.117) and (12.83) we see that as T, approaches zero (B approaches
infinity)

ue Ro A ee = = Co/2 bits/sec

This says that 70 is approximately 3 dB less than the infinite bandwidth channel capacity,
Coo. This 3 dB backoff from channel capacity is frequently taken to be the achievable region
of operation by practicing system engineers.
For high values of Ey./No, then we readily observe from (12.115) that Ro < log, Q
and dRo/dx = 0. This serves to demonstrate why one might choose a large Q in system
designs which require a large value of Ey,./No, where (12.115) holds with near equality.
In conclusion, in this section we have demonstrated how the Ro-criterion can be
used to accomplish signal design, that is, the simultaneous choice of both the coding and
modulation techniques.

12.8 LINEAR BLOCK CODES

In the past several sections of this chapter, we have demonstrated the implications of a very
striking theorem regarding the transmission of information over a noisy channel, namely,
Shannon’s coding theorem. This theorem says that every communication channel has a def-
inite capacity Cy, (or C) and that for any rate re < Cy, (or Rp < C), there exist codes (and
modems) of rate r- which, with maximum likelihood decoding, render an arbitrarily small
probability of codeword decoding error P(E). Shannon’s theorem only demonstrates the ex-
istence of such codes; unfortunately, it does not indicate how these codes can be constructed.
The system engineer who wishes to use error control coding is faced with three problems:
(1) finding good codes, (2) finding practical encoding methods (encoders), and (3) find-
ing practical decoding methods (decoders). By practical we mean small coding delay and
minimum hardware complexity. Transmission errors created in the channel are frequently
referred to as random errors. Codes devised by coding theorists to combat independent er-
rors are called random error-correcting codes. Sometimes channel errors cluster together
and occur in bursts. Codes devised to correct burst errors are called burst._error-correcting
codes. In this section, we will discuss and present, from a systems perspective, coding tech-
niques for error correction—in particular, the Hamming, Golay, Reed-Solomon, and BCH
techniques. See the Venn diagram of Fig. 11.18. We will not present the algebraic theory
leading to the structure of such codes; this can be found in eEOOKS of algebric coding
theory [3, 11, 13] and information theory [8, 9].
800 Block-Coded Digital Communications Chap. 12

K,—Bit Message Message Parity-Check or

Redundant Digits

fo Ke —+|«—__—_—-N,- K,-—+|

ne
Figure 12.15 The codeword for a systematic block code

We will primarily be concerned with the class of (N-, Ke) codes.” An (Nz, Ke) block
code is a code in which each codeword contains N, symbols; K- of these are information
symbols and r = N. — K¢ are redundant parity check symbols. In this context, the dimen-
sionality of the codeword space refers to the number of encoder output symbols for each
block of K,. encoder input symbols. It is possible to encode the K,-symbol message into
a codeword such that the first K. symbols of the codeword are exactly the same as the
message block and the last N; — Ke symbols are redundant parity check symbols which
are functions of the information symbols (see Fig. 12.15). A code of this form is called a
systematic code. The redundant symbols are used to combat (control) errors which occur
during transmission. Systematic codes are of greatest practical interest and we shall con-
centrate on the use of such codes for error control in what follows. Linear block codes can
be described by a K, x N- generator matrix G. If the K, information symbols into the en-
coder are represented by u = (uj, u2,..., UK), the encoder output word is represented by
the N,-dimensional row vector x. The encoder input-output relationship is characterized by
+= uG (12.118)
so that the N,--symbol encoder output blocks (that is, the codewords) are linear combinations
of the rows of the generator matrix. For the binary case, the output blocks are bit-by-
bit modulo-2 sums of the appropriate rows of G. Usually block codes are decoded using
algebraic decoding techniques; the important and elegant details can be found by the system
engineer in many coding and information theory books [3, 11, 12, 13]. The Venn diagram
for error-correcting codes is illustrated in Fig. 11.18. In what follows we shall assume that
the mapping of code vectors x to coded modulation vectors S is isometric.

12.8.1 The Union Bound Expressed in Terms of the Hamming Distance

for Binary Codes

From the coding perspective it is very instructive to relate the Euclidean distance, dj;, to the
Hamming distance, D;;, and use this result in the union bound result given in (12.3) to gain
further insight into block coding. To do so we must make sure that the modulation technique
chosen to transmit the coded message bits does not destroy the Hamming distance properties
of the code; otherwise, the effectiveness of the coding scheme to minimize the required
Ep»/No is lost. As was demonstrated in Chapter 11, coded BPSK and QPSK modulators
can be used since they map the coded message bits into the modulator output waveform
space without modifying the effective distance properties of the code. In this section we
will quantify this fact for the class of binary codes.
2. In the coding literature, it has been fashionable to refer to such codes as (n, k) codes. We prefer to use (Ne, Ke)
here so as to minimize notational confusion; that is, k is Boltzmann’s constant, and n and k are summing indices.
Sec. 12.8 Linear Block Codes 801

Recall that the Hamming distance D(x;, x pi SD; j between the two binary codewords
x; and x; is the number of positions in which the two codewords differ. On the other
hand, the Hamming weight of a codeword (N-,-tuple) x is defined as the number of nonzero
components in x. Using coded BPSK, the Euclidean distance d;; between two binary coded
signal vectors (S;) and (S;) can be written in terms of the Hamming distance D; j- Now

d;, = |S; — Sj|? = E; + Ej —2/EiEjyij (12.119)

where
Sp:Sj
COs 6; (12.120)

measures the normalized cross-correlation between the two coded modulation vectors. For
binary codewords

(12.12%)

where A;;(Dj;) represents the number of term-by-term agreements (disagreements) be-

tween the two binary vectors.
For binary coded BPSK we have from (12.119), (12.120), and (12.121)

dj, = 2E(1 — yj)

=2NcEn,(1 — vis) (2122)
and using the facts that Ne = Ajj + Djj and Aj; = Ne — Dj; in (12.121) we find that 1 —
Vij = 2Dj;;/Nc. Substitution of this into (12.122) yields

d?,BS =2N-Ey,(1 — yj) = 4En,Dij (12.123)

Equation (12.123) implies that Dj; = N-/2 for binary orthogonal codes since yj; = 0. This
result does not readily generalize to other coded modulation techniques.
Using (12.123) in the union bound result given in (4.81), we have
M.—1
Oe D
P(EIS;))< >> O(,/—== (12.124)
i#j,i=0 No
This demonstrates the key role the Hamming distance plays in specifying good error control
codes. With
Dmin = min{
Dj;} 25125)
i)
we can bound the decoded error probability for block length N, by

P(E) <M.Q (etchk Dai)=M.Q (a5 Dai) (12.126)

No No

since Ey,/No =rcEp/No. This result clearly demonstrates the importance of selecting
codes with large values for the minimum Hamming distance. As we shall subsequently see,
the minimum Hamming distance Dyin plays a key role in defining the error-correcting and
detecting capability of a good block code. Notice from (12.126) that the minimum Ham-
ming distance parameter serves as a multiplier on the energy-to-noise ratio, thereby increas-
ing the effective communication efficiency E,/No. From this discussion, we see that one
 802 Block-Coded Digital Communications Chap. 12

must be careful in selecting the modulation-demodulation technique when coding is used to

improve system performance. The main reason for this is that the modem can modify the
distance properties of the code in such a way that the coding gain is destroyed. Fortunately,
coded BPSK and QPSK modems do not change the distance properties of the code. This
combined problem of designing the coding and modulation so as to maintain the codec dis-
tance properties and minimize the bandwidth expansion due to coding falls under the topic
of Coded Modulation.
Finally, since M, = 2”Ne and Q(x) < exp (—x*/2) then (12.126) can be written as
P(E) Be 9—NelRo—re] = 9—-WT[Ro—Ro/ wi (12.127)

where
[Din 1G Dea JEip
Re Secs é})= aes 12.128
NcRo S In2 (5) in Ne ( )
and, as N. approaches infinity, P(E) approaches zero if Ro > re. From (12.128) we can
write the ratio
Pas| i
Ro = = 12.129
CC Gea NaN Cee
which says that the product of the bandwidth expansion factor and the cutoff rate is propor-
tional to both the relative minimum Hamming distance, Dmin/Nc, for any particular binary
code and the communication efficiency, E,/No. Thus, the communication efficiency must
be
E
a > In2/(Dinin/
No) (12.130)
in order to achieve arbitrarily small P(E) with large N.. For orthogonal codes, Dmin =
N,/2, so that (12.130) simplifies to (12.15) as it should.

12.8.2 The Plotkin Bound for Binary Codes

Now Plotkin has shown [8] that the minimum Hamming distance is bounded by
1 NM.
D (42.131)
SMT)
for binary codes with M, codewords and block length N,. This is called the Plotkin bound.
Thus
Dyin Z M.
(12.132)
Ne ~ 2(M. — 1)
Direct substitution of (12.132) into (12.130) demonstrates the requirement that E,/No >
(21n2)(M, — 1)/M- in order for P(E) to approach zero for large N,. Moreover, for large
M_-, we see from (12.132) that Dmin < N¢/2 with the equality being achieved by orthogonal
codes.

- 12.8.3 The Role of the Minimum Hamming Distance in Specifying Good Error
Control Codes
Much insight into the problem of selecting a good error control code can be obtained by
_ the systems engineer by noting that Din is directly related to the error correcting and error
Sec. 12.8 Linear Block Codes 803

detecting capability of any block code. In particular, it can be shown [3] that a block code
with minimum Hamming distance Dyin has t error-correcting capability

t= [(Doin — 1/2] (122933)

This means that any occurrence of t or fewer symbol errors per codeword will be corrected.
On the other hand, it can be shown [3] that if a block code with minimum Hamming distance
Dyin is used for error detection only, the decoder can detect all binary error patterns of
(Dmin — 1) or fewer errors per codeword.
As previously discussed, codes can be characterized as systematic or nonsystematic.
For a systematic code, the encoding process takes the K,-tuple of bits and maps them
into the N.-tuple by appending the additional redundant bits for error correction and/or
detection; see Fig. 12.15. If the error-correcting capability of the code is not exceeded,
then the decoder is guaranteed to correctly decode the N,-tuple into the transmitted K,-
tuple. If the error-correction capability is exceeded, then the decoder will usually do one
of two things. It will either decode the received N,-tuple into incorrect information bits or
it will detect that the error-correction capability was exceeded. If the decoder error cannot
be detected then it is an undetectable decoder error. If the number of errors detected is in
excess of the codes error-correction capability, then when a feedback channel is available,
the decoder might be designed to send an automatic repeat request (ARQ) [12]. Otherwise,
the decoder will output the errors to the user or source decoder. With an ARQ approach, the
decoder detects errors and makes use of a feedback channel to the transmitter to have the
incorrectly received codewords (frame) repeated. Presumably these requests can be repeated
until the codewords are received correctly. This procedure has the effect of reducing the
information rate. In an ARQ system which employs error-correcting codes, a bit error occurs
when the decoder incorrectly accepts a frame of data with errors. Lin and Costello [12] give
a good summary of numerous approaches to error control design using ARQ approaches.
We now investigate the error-correcting and detecting performance characteristics of other
block codes of great practical interest.

12.8.4 Error Probability Performance for Error-Correcting Block Codes

As we have observed in Section 12.8.1 for the AWGN channel, the performance of block
codes can be estimated in terms of the minimum Hamming distance if x is mapped onto
S isometrically. Furthermore, for a fixed code rate r, and block length N., the goal is to
choose a code with a large minimum Hamming distance. Usually, block codes are decoded
using algebraic decoding techniques which require a demodulator-detector to make a hard
decision on each received channel symbol, that is, “1” or “0.” Let p denote the probability of
a channel symbol (coded message bit) error on the resulting binary symmetric channel. The
performance of block codes using hard decisions (one-bit quantization of the demodulator
outputs occurring every Ty = T; sec) on the observables rj, r2,..., Nn, can be characterized
in terms of their ¢ error-correcting ability. In particular, suppose that the decoder can correct
any combination of t or fewer channel symbol errors, and no combination of more than
t errors. Such a decoder is called a hard-decision bounded distance decoder. Thus, the
codeword error probability is

P(E) =Problé >t +1] (12.134)

 804 Block-Coded Digital Communications Chap. 12

where € is a random variable representing the number of coded message bit errors in the
codeword. Since there are ts different ways of having k errors in N, detected coded
message bits, the hard-decision codeword error probability is

Ne
{eos (Xi) ofa agi (12.135)
k=t+1

On the other hand, when hard-decision decoding techniques are not used (that is, with
finer quantization of the observables r; see Fig. 11.6), the decoder implementation com-
plexity is usually much greater than that of hard-decision (one-bit) quantization. Forney
[13] has investigated such block decoder quantization techniques; they are sometimes called
erasure-and-error decoders. For infinite quantization and binary coded BPSK, (12.126) up-
per bounds the codeword error probability for the AWGN channel in terms of Dmin and
Ey/No.
The bit error probability for systematic block codes can be found from (12.135) to be
[17,18]

P pall Do
= ee(Nev ifae
° |pl —p) nee (12.136)
vEV==-
© k=t+1

where e; is the average number of channel symbol errors that remain in the corrected N--
tuple when the channel caused k symbol errors. In practice, it has been fashionable to use
ex =k for k > t, which appears to be a good approximation for the majority of codes of
greatest interest, thereby making (12.136) easy to evaluate numerically.
If the block code is designed for the sole purpose of error detection, the decoder fails
to detect an error in the observed block if and only if the channel transforms the encoder
output into another valid codeword. The undetected error probability P(E) is given by

Ne
PED= >° wp'a—py (12.137)
k=t+1

where w is the number of codewords of Hamming weight k.

12.8.5 Hamming Codes and Their Error Probability Performance

Hamming codes are perhaps the simplest class of block codes. The (N¢, Ke) Hamming
codes are characterized by the parameters defined in Table 12.6. Using this table, (12.135),
(12.136), and (12.137), one can numerically evaluate the codeword error probability, the
bit error probability, and the undetected error probability. Figures 12.16, 12.17, and 12.18
demonstrate these probabilities versus the hard-decision coded message bit error probability,
p, for m = 3,4, and 5 respectively. Figure 12.19 demonstrates the bit error probability ver-
sus E;,/No for various Hamming codes on an AWGN channel, assuming a BPSK modulator
_ so that p = Q(./2r-E,/No). This figure also demonstrates the coding gain as a function of
_ parameter m. Coding gain, G, is the difference between the E,/ No (in decibels) required for
uncoded communication and the E, /No (in decibels) required for coded communication to
achieve the same bit error probability; that is, G = (Ep/No)uncoded — (Ep/ No) coded:
Sec. 12.8 Linear Block Codes 805

Table 12.6 Hamming Code Parameters

Minimum distance

12.8.6 Golay Codes and Their Error Probability Performance

Golay codes are more powerful than the Hamming codes discussed in the previous section;
however, the decoder complexity and the bandwidth expansion factor are greater. The ubiq-
uitous Golay (23,12) code [3] is the only known multiple-error correcting binary perfect
code that can correct any combination of three or fewer errors per 23-bit codeword. From a
system perspective, the extended Golay code is more useful. This code is formed by adding
one parity bit to the (23,12) Golay code to form the (24,12) extended Golay code. This par-
ity bit increases the minimum Hamming distance of the code from seven to eight and yields
a code rate 1/2 which is easier to decode in hardware than the (23,12) Golay code. The code-
word, bit, and undetected error probabilities for hard-decision decoding can be determined
from (12.135), (12.136), and (12.137) to be [17,18]

P(E) =)(24)
(A )p¢ (t— p) 24-k (12.138)

1SDat (
PE) = 35 24| )p
\ ok (1 — p) 24—k (12.139)
and

= (1— p)*4 c (2) zu1

P(E,) (12.140)
where function A(z) is defined in detail in [17]. The codeword, bit, and undetected error
probabilities versus p are illustrated in Fig. 12.20 for the extended (24,12) Golay code.

12.8.7 Cyclic Codes

Cyclic codes are a subset of linear block codes (see the Venn diagram in Fig. 11.18)
with the additional property that every cyclic shift of a codeword is also a codeword. If
X = (x},..., Xy,) is a codeword, then the shifted version x = (Xy,, X1,---,XN.-1) is also
a codeword. Cyclic codes are usually easier to encode and decode than all other linear block
codes. The encoding operation is easily performed using shift registers; furthermore, imple-
mentation of the decoder becomes more practical due to their simpler algebraic structures
[12]. Cyclic block codes require that synchronization words be attached to the. beginning of
the transmitted codewords. a
 806 Block-Coded Digital Communications Chap. 12

107 oe,

10

10

P(E)

10

10

107 107 10° 10

a a -4

Channel Bit Error Probability = p

Figure 12.16 Codeword error probability versus channel bit error probability for block
length N. = 2” — 1 Hamming codes with m = 3, 4, and 5 (Reprinted from [17])

12.8.8 BCH Codes and Their Error Probability Performance

Bose-Chaudhuri-Hocquenghem (BCH) codes are a special case of powerful cyclic codes

which have well-defined decoding algorithms [3, 12, 18]. In this section, we will consider
P-ary (P prime), binary, and Q-ary BCH codes. For each type, we will identify character-
istics and present typical achievable performance results.

P-ary BCH codes. Here the codeword symbols belong to the Galois Field
G F(P). Every codeword is associated with a code polynomial with coefficients in GF (P).
The roots of the generator polynomial lie in an extension field GF (P”) of GF (P) for some
_ integer m > 1. BCH codes can either be primitive or nonprimitive; see the Venn diagram in
~ 3. An introduction to the algebra of Galois Fields (also known as finite fields) may be found in Lin and Costello
{12].
 Sec. 12.8 Linear Block Codes
807

107 F

P,()

10 : : :
107 107 107 10°
Channel Bit Error Probability = p

Figure 12.17 Bit error probability versus channel bit error probability for block length
N,. = 2" — | Hamming codes with m = 3, 4, and 5 (Reprinted from [17])

Fig. 11.18. Primitive BCH codes are defined as codes whose block length is N. = P™ — 1.
Given any error-correction capability rt, there exist choices of the message length (K, sym-
bols) and block length (NV, symbols) for which a t-error correcting P-ary (N., K-) code
exists [8, 12, 18]; see Table 12.7 with Q = P.

Binary BCH codes. Binary BCH codes are a special case of P-ary-BCH codes
with P = 2. The codeword symbols (coded message bits) of binary BCH codes are from the
Galois Field GF (2). The roots of the generator polynomial of the code lie in the extension
field GF (2”) for some integer m > 1. In particular, a t-error correcting (t < 2”~') prim-
itive, binary BCH code having generator polynomial with roots in GF (2”) can be found
which achieves the parameters listed in Table 12.8 for any m > 3. There is no simple for-
mula for the system engineer to use for calculating VN; — K-; however, N. — K- is less than
or equal to mt. The parameters for all binary BCH codes of length NV. = 2" —-1 with m < 10
808 Block-Coded Digital Communications Chap. 12

10

10

10

10

Undetected
of
Error
Probability
an

10

10
Channel Bit Error Probability = p

Figure 12.18 Probability of an undetected error versus channel bit error probability for
block length N, = 2” — 1 Hamming codes with m = 3, 4, and.'5 (Reprinted from [17])

are summarized in Table 6.1 of [3] and Table 4.6 in [18]. The codeword error probability
performance versus channel symbol (coded message: bit) error probability, p, is illustrated
in Figures 12.21, 12.22 and 12.23 for several binary BCH codes. Figure 12.24 provides a bit
error probability comparison versus E»/No for various BCH codes.
These codes often have some inherent error detection capability beyond their error-
correction capability. If some error-correcting capability is sacrificed to gain some addi-
tional error detection capability, say tg, then the resultant ¢ error correcting, tg additional
error detecting (tg is an even number), primitive, binary BCH code is characterized by the
parameters summarized in Table 12.9.

Q-ary BCH codes. Q-ary BCH codes are a generalization of the P-ary codes.
The codeword symbols of a Q-ary BCH code come from GF(Q) = GF(P‘), for some
integer c > 1, while the roots of the generator polynomial lie again in an extension field
 Sec. 12.8 Linear Block Codes
809

€)
P,

4 5 6 7, 8 9 10 11
E,/Noy dB

Figure 12.19 Bit error probability versus E,/No for block length N. = 2” — |
Hamming codes with m = 3, 4, and 5 on an AWGN channel (Reprinted from [17])

GF(Q”) =GF(P”™), for some m > 1. If c= 1, Q-ary codes become P-ary. There exist Q-
ary BCH codes having the code parameters listed in Table 12.7. A discussion of the theory
of finite fields underlying Q-ary BCH codes can be found in [3, 8, 11, and 18].

12.8.9 Reed-Solomon Codes and Their Error Probability Performance

Reed-Solomon (RS) codes have found widespread applications beginning with the 1977
Voyager’s Deep Space Communications System [21]. At the time of Voyager’s launch,
efficient encoders existed, but accurate decoding methods were not available! Jet Propulsion
Laboratory scientists and engineers gambled that by the time Voyager II would reach Uranus
in 1986, decoding algorithms and equipment would both be available for decoding of the
810 Block-Coded Digital Communications = Chap. 12

Codeword
error
probability

Bit error
probability to

Probability
Error

Probability
of an
undetected
error

1 107 107 10
4 <4 3

Channel Bit Error Probability = p

Figure 12.20 Codeword, bit, and undetected error probabilities versus channel symbol
error probability with extended Golay coding (Reprinted from [17])

data transmitted back to Earth. Fortunately, they were correct! Voyager’s communications
system was able to transmit at a data rate of 21,600 bits per sec from 2 billion miles away
and successfully extract the data from a signal whose received power was 100 billion times
weaker than that generated by an ordinary wrist watch battery!
Other areas of application of RS codes include: (1) communications over telephone
systems with such applications as advanced facsimile machines or faxes (which send and
receive imaging data) and high-speed modems (which usually send and receive computer
data); (2) satellite communications with such applications as the Hubble Space Telescope,
Mobile Satellite Terminals, and the 300 megabits per sec (Mbps) return link of the Space
Station Freedom via the Tracking Data Relay Satellite System (SF/TDRSS); (3) Deep Space
Communications with such applications as Voyager, the Galileo Orbiter, the Mars Observer,
Sec. 12.8 Linear Block Codes 811

Tabie 12.7. Parameters for Primitive Q-ary BCH Codes

Block length N,- = Q” — 1 code symbols

Number of parity checks N. ioe— K- code symbols

Minimum distance Dyin = 2t + 1 code symbols

Table 12.8 Parameters for Binary BCH Error-Correcting Codes

Block length N. = 2" — 1 code symbols

Number of parity checks N- — Ke < mt code symbols

Minimum distance Dymin = 2t + 1 code symbols

Table 12.9 Parameters for Binary BCH Error Detecting Codes

Block length Ne = 2” — 1 code symbols

Number of parity checks Ne — Ke < m[t + (ta/2)] code symbols

Minimum distance Dymin = 2t + tg + 1 code symbols

and the Cassini Titan Orbiter/Saturn Probe; (4) Space Shuttle Communications; (5) enter-
tainment, with applications such as compact disks (CDs) and CD-ROM, etc.
In general, Reed-Solomon (RS) codes may be regarded as a special subclass of Q-ary
BCH codes that results when parameter m = 1. Thus, the coefficients of the code polyno-
mials and the roots of the generator polynomial all lie in GF (P‘). A primitive RS code has
block length N, = P* — 1. For any RS code, the error-correction capability t and the mes-
sage length K, are simply related by K, = N; — 2t. For example, if Q = P© = 2° = 23 = 8,
then this can either be a (7,5) or a (7,3) or a (7,1) RS code depending on whether it is de-
sired to correct t = 1, 2, or 3 symbols respectively. Table 12.10 summarizes the achievable
paramaters for a ¢ error-correcting, primitive RS code.
Binary RS codes are most frequently used in practice. A “binary RS code” is simply
an RS code where P = 2. In the nonbinary case, P is an odd prime. Note that the code
symbols of a binary RS code are not in fact binary, but belong to the field GF (2°). The
parameters for a binary, ¢ error-correcting, primitive RS code are summarized in Table 12.10
with: P= 2.
RS codes meet the Singleton bound (Dmin < Nc — Ke + 1) because for an RS code,
Dmin = 2t + 1 = N- — Ke + 1. Therefore, RS codes are optimal with respect to this bound,
and the term maximum distance separable (MDS) is used to describe optimality in this

Table 12.10 Q-ary RS Code Parameters

Block length N. = Q — 1= P* — 1 code symbols

Number of parity checks N. — Ke = 2t code symbols

Minimum distance Din = 2t + 1 code symbols

 812 Block-Coded Digital Communications Chap. 12

(127, 120)
t=1

JEXGE)

(127, 36)
t= 15
(127, 22)
t = 23

Figure 12.21 Codeword error probability

1071 4072 10°3 49-4 versus bit error probability for some binary
BCH codes of block length 127 (Reprinted
P from [18])

sense. It is also worth pointing out that the block length N, of an RS code over GF(Q) =
G F(P‘) can be extended to either Q or Q + | while still maintaining the MDS condition. It
can also be shown [17, 18] that the block (codeword) and 2”-ary symbol error probabilities
for RS codes are given respectively by
Qgm_} Qm fe 1 f

PEyS ye ( ; )pha = py oot (12.141)

k=1+t

and

P(E) = Lol Sec

aye (Daa
: )p (1 — p) 2m 1k (12.142)
where we have used the approximation eg, ~ k for k > t, and where p denotes the channel
symbol error probability. The bit error probability can sometimes be upper bounded by
P;(E). For certain cases, namely for orthogonal codes, the relationship is
 Sec. 12.8 Linear Block Codes 813

P(E)

10°

10° Figure 12.22 Codeword error probability

1072 - : versus bit error probability for some binary
BCH codes of block length 255 (Reprinted
P from [18])
K.-1

where K; is the number of bits per codeword; see Chapter 4. Figures 12.25, 12.26, 12.27,
and 12.28 demonstrate the codeword error probability (MEP) versus p profiles for various
RS codes of practical interest.
Interleaving of the channel encoder output is required when the channel errors occur
in bursts; see Section 11.3 in Chapter 11. Inserting an interleaving-deinterleaving system
into the signal path introduces another new system error control system parameter, namely
the interleaving length or depth. The value of this parameter is specified in terms of the
prespecified channel memory length (sec), say T-. Physical considerations which lead to
specifying t- are the expected burst-error duration and the multipath duration which are
frequently characterized in terms of the temporal correlation time of the channel.
RS codes also have applications in error control system designs as outer codes in
concatenated coding systems (see Section 11.5 and Chapter 13). In such designs, an inner
code is also used to provide a portion of the error control, thus the name concatenated code
 iN
814 Block-Coded Digital Communications — Chap. 12

107 ‘i Fy

10°

WA,
10°3

P(E)

104
(511, 157)
t= $1
(511, 448)
(511, 211) t=7
t= 41
(511, 259)
5 {a=230
10°
(511, 340) (511, 466)
te =<20

(511, 376)
p t=ot5)

_—- (511, 421)

t = 10

10°° Figure 12.23 Codeword error probability

1071 10°72 10°3 versus bit error probability for some binary
BCH codes of block length 511 (Reprinted
p from [18])

[13]. The inner code is usually a binary code which maps each symbol in the RS alphabet to
a unique binary codeword. Decoding of a concatenated code typically proceeds as follows:
First, the demodulator-detector outputs are soft quantized (soft-decision) and presented as
inputs to the inner channel decoder. The output of this decoder are hard decisions which
serve to drive the outer RS decoder. The inner code protects the RS code against random
channel errors, leaving the correction of burst errors to the outer RS code.
The performance of concatenated coding systems which use an RS outer code are
given in the next chapter. Concatenated codes were first introduced by Forney [13], and their
performance for special coding techniques will be presented in Chapter 13.
Sec. 12.8 Linear Block Codes 815

(31, 15) t = 3 Soft Decoding

gPp
a
on>)
¥
7
N

PI
gurpoo2d

E p/Nov dB

Figure 12.24 Bit error probability versus E,/No for coherently demodulated BPSK
over a Gaussian channel using BCH codes (Reprinted from [22])
 816 Block-Coded Digital Communications Chap. 12

(31, 13)
t=9
(31,9)
C= sii

P(E)

Figure 12.25 Codeword error probability

10°3 versus channel symbol error probability
for various RS codes defined on a 32-ary
p (32-ary) alphabet (Reprinted from [18])
 Sec. 12.8 Linear Block Codes 817

1071

10°2

10°3'

P(E)

104

10°5

Figure 12.26 Codeword error probability

versus channel symbol error probability
for various RS codes defined on a 64-ary
p (64-ary) alphabet (Reprinted from [18])
 Block-Coded Digital Communications Chap. 12

P(E)

(127, 97)

Figure 12.27 Codeword error probability

1071 107 10°3 versus channel symbol error probability
for various RS codes defined on a 128-ary
p (128-ary) alphabet (Reprinted from [18])
 Sec. 12.8 Linear Block Codes 819

1071

a Nes
\\\
Seay

ee
1072

10°3

P(E)

104

10°5

Figure 12.28 Codeword error probability

versus channel symbol error probability
for various RS codes defined on a 256-ary
p (256-ary) alphabet (Reprinted from [18])
 820 Block-Coded Digital Communications Chap. 12

PROBLEMS
12.1 Consider the discrete memoryless channel (DMC) modeled by the diagram in Fig. P12.1. Write
an expression for the cutoff rate Ro and the channel capacity Cy, assuming equally probable
input messages. :
12.2 For the discrete memoryless channels illustrated in Fig. P12.2, write expressions for the channel
cutoff rate Ro and capacity Cy. Assume equally probable input symbols.
12.3 (a) Develop an explicit expression for the achievable Ro performance for coded QPSK, M =
adh,
(b) ie the result in (a) with that for coded BPSK and comment on your results.
12.4 Consider the M-ary signal set
N
si(t) =) sindn(t)
n=

fonUkpe—eandi | Ow lereer. M — 1. Assume an AWGN channel with spectral density No/2.

Suppose we randomly select codewords by choosing each s;, independently from the ensemble
of random variables. Using the union bound, one can show that the ensemble message error

My

Ms Figure P12.1 4-ary input with four output

levels

l-p 1
Ry My My Ro
P
Pp
P
R, a R,
l-p l-p

(a) BSC (b) BEC (c) Z Channel

Figure P12.2 DMC models

Problems
821

Figure P12.5

probability is bounded by

Pe(M) < M2-NRo

(a) Determine Ro when s;,, is a zero-mean Gaussian random variable with variance E.
(b) Repeat (a) if sj, is a binary random variable taking on values VE or —VE with equal
probability.
12.5 (a) Calculate the quantized cutoff rate Ro (in bits/per channel use or per clock cycle) for the
DMC model shown in Fig. P12.5.
(b) Derive a way to use this DMC at rate r, = 1 with P(E) = 0.
12.6 For block orthogonal signals, the union bound states that

Ep
P(E) < exp [-K.(= —In 2)

Calculate how large the number of messages M, must be to guarantee that P(E) < 10~° when
E,/No = 0, 3, and 6 dB.
12.7 (a) For M equilikely transorthogonal signals disturbed by additive white Gaussian noise, de-
rive a bound similar to (12.12) and (12.13) which is valid for orthogonal signals.
(b) Repeat (a) for M equiprobable biorthogonal signals.
(c) Using the bounds, discuss the relative advantages of the three signaling systems from a
systems engineering viewpoint.
Hint: See Chapter 3 for the union bound result and Table 12.4.
12.8 Consider communicating over the AWGN channel using binary codes and antipodal signals.
(a) Using the fact that 2B7,. = 1, write an expression for Ro as a function of P/NoB.
(b) Show that as B increases, Ro in bits/dimension decreases for a fixed P/No.
(c) Now consider the channel throughput Ro = Ro/ T; = 2B Ro in bits/sec and show that Ro
increases monotonically with increasing bandwidth B.
(d) Explain these results.
12.9 Show that the channel capacity C in bits/sec increases monotonically toward its maximum
value Ca. as B increases. Show that the capacity Cy. in bits/dimension decreases monotoni-
cally as B increases. Explain this phenomenon.
822 Block-Coded Digital Communications Chap. 12

12.10 Show for block-orthogonal codes that (12.12) can be written as

P(E) <exp |-7In2 (S - Rs)|

so that as T approaches infinity, P(E) approaches zero if Rp < Co./2. This bound is not tight
but we know that orthogonal signals actually achieve Co. in the limit (see Chapter 4).
12.11 Derive an expression for the channel cutoff rate for the ensemble of binary codes when orthog-
onal, equal energy, equiprobable signal waveforms (BFSK) are used to transport information.
(a) Derive the channel cutoff rate for the ensemble of binary codes when equal energy,
equiprobable waveforms having cross-correlation coefficient y are used.
(b) Compare and discuss the results found in (a) and (b) with the channel cutoff rate in (12.37)
for antipodal signals (BPSK).
12.12 Consider a voice-grade telephone circuit with a bandwidth of B = 3 kHz. Assume that the
circuit can be modeled as an AWGN channel.
(a) What is the capacity of such a circuit if the CN R is 30 dB?
(b) What is the minimum CAN R required for a data rate of 4,800 bits/sec on such a voice-grade
circuit?
(c) Repeat part (b) for a data rate of 19,200 bits/sec.
12.13 Consider that a 100-kbits/sec data stream is to be transmitted on a voice-grade telephone circuit
(having a bandwidth of 3 kHz). Is it possible to achieve error-free transmission with a CNR
of 10 dB? Justify your answer. If it is not possible, suggest system modifications that might be
made.
12.14 Telephone modems operating at 19.2 kbits/sec are available in the market place.
(a) Assuming AWGN and an available E,/No = 10 dB, calculate the theoretically available
capacity in the 2,400-Hz single-sided bandwidth.
(b) What is the required E,/No that will enable a 2,400-Hz bandwidth to have a capacity of
19.2 kbits/sec?
12.15 Consider the set of J orthonormal waveforms {@,(t)}, k = 1, 2,..., J over the interval [0, T,].
These waveforms are used to construct the signal set {s;(t)} of the form [7]
N

s(t) = En, )>b(t — iTe)

i=]

in which the {k;} are integers between | and J. Thus, each signal in the set {s;(t)} is specified
by a vector of the form (kj, k2,..., kz).
(a) Assuming N = 5, J = 4, and

P(t) = ¥2/Te sin (2akt/T-), O<t<T

and zero elsewhere, sketch the signal specified by the vector (2, 1, 4, 2, 3).
(b) How many waveforms are there in the set {s;(t)} for arbitrary N and J? Are all these
waveforms orthogonal?
(c) Using the union bound, show that the ensemble MEP is bounded by
es Epa di
P(E) < exp ;—K, |——- 1
bles P| elas V+ (P= Dew in2]}
where x = PT,/2No.
12.16 An upper bound for the performance of linear block codes is given by (12.127). Notice from
(12.127) that the performance depends on both code parameters rz and Din aS well as the
energy per bit-to-noise ratio. Assume that the bit error probability and message error probability
are approximately equal.
Problems 823

(a) With M, = 2%, show that the coding gain G, defined in Section 12.8.5, is given by

G 4 1010849 [(Eb/No)uncodea/(Eb/
No)coded]
== 10 log jo {re-Dmin — K,/[In 2(E,/No)coded]}

if one uses uncoded BPSK (antipodal) transmission. Use the asymptotic expansion for
Q(x), (4.55), and (12.126).
(b) What is the limiting value for G as E;,/No approached infinity?
(c) Assuming the use of a Hamming code defined in Table 12.6, write an expression for the
code rate r, and plot the asymptotic gain achieved with m = 2, 3,4, 5,..., 10.
12.17 Using the coding gain formula provided in Problem 12.16
(a) Develop an explicit expression for the coding gain achievable for the class of binary BCH
codes defined in Table 12.8.
(b) From Problem 12.16, we see that the asymptotic coding gain for large E;/No is given
by G = 10logjg7¢Dmin. Repeat (a) and write an expression for G in terms of the error-
correcting capability t, N., and Ke.
12.18 Repeat Problem 12.17 for the binary Reed-Solomon codes defined in Table 12.10.
12.19 Consider the coding gain formula given in Problem 12.16.
(a) For the binary primitive (15,9) RS code, write an expression for coding gain.
(b) Plot this gain as a function of E,/No.
(c) What is the maximum achievable gain using this code?
(d) Repeat (a), (b), and (c) for the (154,130) RS code.
12.20 (a) Repeat Problem 12.19 (b) and (c) for the (255,235) RS code with t = 10.
(b) Repeat Problem 12.19 (b) and (c) for the binary (255,179) BCH code with t = 10.
12.21 Figure 12.5 shows a family of channel cutoff rate (Ro) curves for the set of equispaced, equally
probable Q-ary ASK signal points as shown in Fig. 12.4a.
(a) Find the expression for the distance g;; in terms of Q,i, j, and VEN. é
(b) Use the expression for the distance q;; found in (a) to find Ro for the special case of Q = 4.
(c) Graph your result as a function of Ey,./No.
(d) Verify your answer by checking for agreements with Figure 12.5.
12.22 Consider a set of QPSK signal points assigned to the coded message symbols {xm} = {00, 01,
10, 11}.
(a) Find an isometric mapping from coded message symbols to coded modulation vectors
{Sm}. For each pair of points in your signal set, compare the Hamming distance to the
Euclidean distance.
(b) Find a nonisometric mapping from coded message symbols to coded modulation vectors
{Sm}. For each pair of points in your signal set, compare the Hamming distance to the
Euclidean distance.
12.23 Heuristically, justify use of the signal space dimensionality to bandwidth relationship N. =
WT =2BT where T is the signal duration and B (in Hz) is the single-sided bandwidth mea-
sure.
Hint: See Chapter 5 of [7], Nuttall and Amoroso, /T, July 1965, and Landau and Pollak,
Bell System Tech. Journal, 41, July 1962, pp. 1295-1336.
12.24 A customer wants to send information through a channel with single-sided bandwidth 10 kHz.
Channel noise is white and Gaussian with single-sided power spectral density No = 107?”
Watts/Hz. He/she can achieve a received signal power of | pW.
(a) What is the maximum theoretical rate at which error-free transmission is possible?
(b) If the customer decides to send at rate 100 kbps using uncoded quaternary PSK, what will
the actual symbol probability be?
 824 Block-Coded Digital Communications Chap. 12

12.25 The received signal power in a coherent AWGN channel is —102 dBm and the power spectral
density of the noise is —228 dBm.
(a) What is the maximum rate (bits/sec) at which one can send information error-free?
(b) What is the corresponding value of E,/No?
(c) What is the maximum amount of error-free information that can be sent in two seconds?
12.26 For QPSK signaling and M = 4
(a) Evaluate Ro using the appropriate Q-ary coded M-PSK model from Section 12.4.5.
(b) Evaluate Ro using the appropriate QASK model from Section 12.4.3.
(c) Compare your answers from parts (a) and (b). If they agree, explain why; if they disagree,
explain why.

REFERENCES
1. Wozencraft, J. M. and R. S. Kennedy, “Modulation and Demodulation for Probabilistic Coding,”
IEEE Trans. Information Theory, IT-12, July 1966, pp. 291-97.
2. Massey, J. L., “Coding and Modulation in Digital Communications,” Zurich Seminar, vol. E2,
1974, pp. 1-4.
3. Lin, S., An Introduction to Error-Correcting Codes, Englewood Cliffs, NJ: Prentice Hall, 1970.
4. Lindsey, W. C. and M. K. Simon, Telecomunication Systems Engineering, Englewood Cliffs, NJ:
Prentice Hall, 1973. Republished by Dover Publications, 0-486-66838-x, 1991.
5. Shannon, C. E., “A Mathematical Theory of Communication,” Bell System Tech. J., 27, 1948,
pp. 379-423 (Part I), pp. 623-56 (Part II). Reprinted in book form with postscript by W. Weaver,
Urbana, IN: Univ. of Illinois Press, 1949; also see Claude E. Shannon Collected Papers, edited
by N. A. Sloane and A. D. Wyner, IEEE Press, 1992.
6. Shannon, C. E., “Probability of Error for Optimal Codes in a Gaussian Channel,” Bell System
Tech. Journal, vol. 38, May 1959, pp. 611-56.
7. Wozencraft, J. M. and I. M. Jacobs, Principles of Communication Engineering, New York: John
Wiley and Sons, 1965.
8. Gallager, R. G., Information Theory and Reliable Communication, New York: John Wiley and
Sons, 1968.
9. Cover, T. and J. A. Thomas, Elements of Information Theory, New York: John Wiley and Sons,
1991.
10. Viterbi, A. J., “Error Bounds for Convolutional Codes and an Asymptotically Optimum Decoding
Algorithm,” JEEE Trans. Information Theory, vol. IT-13, April 1967, pp. 260-69.
11. Berlekamp, E. R., Algebraic Coding Theory, New York: McGraw-Hill, 1968.
12. Lin, S. and D. J. Costello Jr., Error Control Coding: Fundamentals and Applications, Englewood
Cliffs, NJ: Prentice Hall, 1983.
13. Forney, G. D. Jr., Concatenated Codes, Cambridge, MA: MIT Press, 1966.
14. “Troposcatter Interleaver Study Report,’ CNR, Inc., Report Number RADC-TR-75-19 for Rome
Air Dev. Center, Air Force Systems Command, Griffiss AFB, NY, February 1975.
15. Nyquist, H., “Certain Topics on Telegraph Transmission Theory,” Trans. AIEEE, vol. 47, April
1928, pp. 617-44.
16. Wozencraft, J. M. and B. Reiffen, Sequential Decoding, Cambridge, MA: MIT Press, 1961.
17. Odenwalder, Joseph P., Error Control Coding Handbook, Final Report under contract No.
F44620-76-C-0056, Linkabit Corporation, July 1976. Also see Chapter 10 on Error Control in
Data Communications, Netwoks and Systems by Thomas C. Bartee, ed., Indianapolis: Howard
W. Sams, 1985.
References 825

18. Michelson, A. M. and A. H. Levesque, Error-Correcting Techniques for Digital Communications,

New York: John Wiley and Sons, 1985.
19. Lipschutz, Seymour, Schaum’s Outline of Theory and Problems of General Topology, New York:
McGraw-Hill, 1965.
20. Viterbi, Andrew J., and Omura, Jim K., eS of Digital Communication and Coding, New
York: McGraw-Hill, 1979.
21. Wicker, Stephen and Vijay, Bhargava, editors, Reed-Solomon Codes and Their Applications, New
York, NY: IEEE Press, 1993.
22. Weng, L. J., “Soft and Hard Decoding Performance Comparisons for BCH Codes,’ [CC’79
Conference Record, Boston, MA, pp. 25.5.1—25.5.5, 1979.
 Chapter 13
Convolutional-Coded Digital
Communications

In the previous chapter, we saw that a linear (N;, K-) block code can be used for error
control purposes by partitioning the data stream into blocks containing K, bits each. Then,
using an encoding rule to associate the r = N, — K¢ parity check bits with each K; data bits,
an N.-symbol codeword was formed for transmission. However, with convolutional codes
(originally called recurrent codes), the encoded data do not have a simple block structure,
and, as we shall see, a convolutional encoder operates on the input bit stream such that each
information bit can affect a finite number of consecutive symbols (coded message bits) in
the encoder output.
Apparently convolutional codes were first introduced by Elias in 1955 [1] as an alter-
native to block codes. In this chapter, we shall introduce convolutional codes, methods for
representing such codes, and present the maximum-likelihood (ML) decoding algorithm due
to Viterbi [2]. Detailed performance results achievable with maximum-likelihood decoding
of convolutional codes are given for a set of particular convolutional codes and key design
issues are treated. The chapter concludes by presenting the applications and performance
of concatenated codes introduced by Forney [3] for error control. Punctured convolutional
codes are also discussed.
Other decoding techniques such as threshold decoding, sequential decoding, etc., lie
outside the scope of this book; however, the details of these techniques can be found in
Gallager [4], Michelson and Levesque [5], Berlekamp [6], Anderson and Mohan [7], and
Lin and Costello [8].

13.1 CONVOLUTIONAL ENCODERS

As we observed in Chapter 12, a linear block code is characterized by two integers N, and
K, and a generator polynomial. The integer K, is the number of data bits that form an input
to a block encoder while the integer N, represents the block length of the encoder output.
The codeword N,-tuple is uniquely determined by the input message K,-tuple and the ratio
re = K-/N-¢ is called the code rate which measures the amount of added redundancy; re !
also serves to measure the bandwidth expansion due to coding.
A convolutional code is characterized by three integers N., K-, and K where the ratio
rc = K-/N¢ has the same physical significance that it has for block codes; however, N, does
not define the block length as it does in block codes. Moreover, the parameter K is used
to denote the constraint length of the convolutional code. The parameter K, plays the same
 Sec. 13.2 Representation of Convolutional Codes 827

role for block or convolutional codes. An important difference between binary block codes
and binary convolutional codes is that the encoder has memory. As we shall see, the Ne-
tuple output of a convolutional coder is not only a function of the input K,-tuple but is also
a function of the previous K, — | input K,-tuples. Usually K, and N, assume small integer
values and K is varied to control redundancy in the transmitted data stream.
The functional architecture of a (K-, Ne, K) convolutional encoder is mechanized
with a kK-stage linear shift register and N, modulo-2 adders (see Fig. 13.1). From this
figure, we observe that a convolutional code is defined by the number, kK, of stages in
the shift register, the number of modulo-2 adders (number of outputs), and the connections
between the shift register and the modulo-2 adders. The encoder state is defined to be the
contents of the rightmost kK — | stages in Fig. 13.1. Knowledge of the next K, input
bits is therefore necessary and sufficient to determine the next output sequence. In other
words, an input sequence of k data bits (“ones” and “zeros”) of time duration T = K-Tp
sec are shifted into the first k stages of the kK register and all bits in the register are
shifted k stages to the right, thereby establishing a new state of the kK-stage register.
The outputs of the N. modulo-2 adders are sequentially sampled to yield the N, output
channel code symbols (coded message bits) which are identified as the codeword branch
vectors X; = (Xj1, Xj2,.-.,XjN,) Of N- channel symbols. In what follows, we shall consider
the most commonly used code for which k = 1. Generalization to higher-order alphabets
can be found in [8]; however, treatment of such alphabets here lies outside the scope of
this book. The N, code symbols occurring at the encoder output during the time interval
tj+1 — tj = N-T- comprise the ih codeword branch xj; = (Xj1,...,Xjn,). Here xji(i =
1,..., Nc) is the i binary channel symbol in the j" codeword branch and is assumed
to take on values “O” or “1.” When K, = 1, the convolutional coder of rate re = 1/N¢ is
implemented with a (K — 1)-stage register and the constraint length K can be referred to in
units of bits.

13.2 REPRESENTATION OF CONVOLUTIONAL CODES

Several methods can be used for representing convolutional codes. These include: (1) the
connection diagram, (2) generator polynomials, (3) the tree diagram, and (4) the trellis
diagram. In what follows, these techniques will be described using a rate r- = 1/2 convo-
lutional code with K. = 1, Nc = 2, and with constraint length K = 3. Fortunately, convolu-
tional codes are most easily described by example and any generalizations should become
intuitively obvious. Figure 13.2 shows a shift-register circuit that generates this code. Input
bits are clocked into the circuit from left to right at a rate of Rp = 1/7, Hz. After each input
is received, the coder output is generated by sampling and multiplexing outputs of the two
mod-2 adders. Thus, for a rate 1/2 code, we observe that 7, = 27; sec.

13.2.1 Polynomial Representation

The encoder of Fig. 13.2 performs two polynomial multiplications. The output orthe up-
per mod-2 adder is the product of the oe -2 aie Bi Na eee eer x? and the
input sequence u(x) = ug + ux + nx? +n. + uj ix/ +..., where uj; € {0, 1} represents
the input information bit in the j™ bit time iieral Gi- DT. < 5< jT;. The output of
the lower modulo-2 adder is the product of u(x) and g2(x) = 1+ x*. The output sequence
is represented by the two (in this case) semi-infinite polynomials X (CY)= X10, lees
 (1-4)
- eBeysylus Jejsibes

I 4

828
€ L-y

‘ln=n
‘en ieahe In Sa

yndu|
jiq eouenbes
peyiys
ul °y syiq
Auane
°y 4 ‘spuoses

jy—
2 c-O|Npow
sioppe

promepon eouanba
x = s‘tyx) x («lx
OJ8UM
ly= ‘tX) aly Cae NN
(

a P1oOMAapod
youesg

& aes Aveuigepoo joqwAs

Jo
e

= I
youelg

plomapod

aanBrgq
"ET [eUONNJoAUOD
Jopooud
Jo jUTeNsUOD
YISug]
yy pur ayes°4 = °N/?Y
 x= ysul4 apog indjno
joquiAS
jo apoyo ebessay
sig
yy! youesg
=X X)
‘ Sy Goes (Caay

jig
Beg

yndu;
=a cIneacta)
2

iG
eouenbas
|

2x
Jo
apog

puosas

= yy!
joquiAg

youelg

ens7Z'ET aey Z/{ jeuonnfoauos

Jopooua
WMY = €

829
830 Convolutional-Coded Digital Communications Chap. 13

xox +...+ x4 jx) + ...= g1(x)u(x) and X2(x) = x20 + x21% + x22Xx eo) haga oe
.. = g(x) w(x) which are interleaved to produce the encoder output sequence. The con-
volutional code is completely defined by the set of generator polynomials g(x) and g2(x).
These polynomials define the length of the shift register, the connections to the modulo-2
adders, and the number of modulo-2 adders. For example, the encoder of Fig. 13.2 maps
each information bit into N, = 2 coded message bits using a rule which depends on the pre-
vious two information bits. Specifically, an input 1 produces the outputs 11, 01, 00, or 10
when the previous inputs are 00, 10, 01, or 11 respectively. The state of the encoder in Fig.
13.2 corresponds to the contents in the shift register, viz., the previous two input bits. The
constraint length K of the code is one plus the number of shift registers required to im-
plement the encoder. The number of stages in the shift register required to implement the
encoder is K — | and the number of encoder states is 2X~!. The encoder of Fig. 13.2 has
constraint length K = 3 since the current output pair is a function of the current input plus
the two previous inputs. Note that different definitions of constraint length can be found in
the literature on convolutional coding. In all cases, however, constraint length is a measure
of the memory required to implement the encoder.
In general, convolutional codes produce N, output aneerel symbols for each K, input
symbols and have rate re = K-/N-. Although this would appear to make convolutional
codes identical to block codes, they are significantly different. For instance, the number of
past inputs which affect the mapping of the current K, inputs into N; outputs is an important
parameter of convolutional codes. This parameter affects convolutional code performance
and complexity much as the block length affects block code performance and complexity.
As already pointed out, consideration of the more complex (K;., Nc, K) codes lies outside
the scope of this book and the reader is referred to Lin and Costello [8] for a through
treatment. With this limitation, however, convolutional codes can be defined by a set of N,
generator polynomials {g1(x), g2(x),..., 9N.()}, where there is one polynomial for each
of the N, output channel symbols. The constraint length K of the code is equal to one plus
the degree of the highest-degree generator polynomial. The length of the shift register for
the encoder is equal to the degree of the highest-degree generator polynomial. Table 13.1
summarizes the code generator vectors gj,i = 1,..., Nc, for a set of optimum rate 1/2
and 1/3 codes. The free distance, dy, indicated in this table will be further discussed in
Section 13.4. For example, if the code generator vector g; = 111 then gj(v) =1+x+ x"
and if gy = 101 then go(x) = 1 + x? as required in the implementation of the encoder
illustrated in Fig. 13.2.
Finally, we point out that convolutional codes can be characterized as systematic
or nonsystematic depending on whether or not the information sequence appears directly
within the code sequence. For a convolutional code to be systematic, at least one of the
code generator polynomials must equal | (for example, g;(x) = 1) so that one of the N;
outputs is simply the current information bit. Unlike block codes, the performance of the
best systematic convolutional codes is not equivalent to the performance of the best nonsys-
tematic convolutional codes. For the same code performance, a systematic code will have a
longer constraint length and be more complex to generate than the performance-equivalent
nonsystematic code. For this reason, most of the convolutional codes used in practice are
nonsystematic.
Sec. 13.2 Representation of Convolutional Codes 831

Table 13.1 Optimum Short Constraint Length

Convolutional Codes (Rate 1/2 and Rate 1/3).
Source: J. P. Odenwalder, Error Control Coding
Handbook, San Diego, Calif.: Linkabit Corp.,
July 15, 1976

101111
110101

1 1001111
1101101

1 10011111
11100101

110101111
100011101

101111
110101
111001

1001111
1010111
1101101

11101111
10011011
10101001
832 Convolutional-Coded Digital Communications Chap. 13

13.2.2 The State Diagram

In order to introduce the reader to the state diagram, it is necessary to understand the concept
of state. The state of a rate 1/N, convolutional encoder is defined as the contents of the
shift register (in this case two). Since 2* —! — 4 for the encoder of Fig. 13.2, there are four
encoder states A = 00, B = 10, C = 01, and D = 11. One way to represent simple encoders
is with a state diagram. Such a representation for the encoder in Fig. 13.2 is shown in Fig.
13.3. The states, shown in the circles of the diagram, represent the possible contents of the
shift register, and the paths between the states represent the output branch words (codeword
branches) resulting from such state transitions. The diagram shown in Fig. 13.3 illustrates
all the state transitions that are possible for the encoder in Fig. 13.2. There are only two
transitions emanating from each state, corresponding to the two possible input bits. Next to
each path between states is written the coder output-branch codeword associated with the
state transition. In drawing the path, we adopt the convention that a solid line denotes a path
associated with an input bit, zero, and a dashed line denotes a path associated with an input
bit, one. Notice that it is not possible in a single transition to move from a given state to any
arbitrary state. As a consequence of shifting-in one bit at a time, there are only two possible
state transitions that the register can make at each input bit time. For example, if the present
encoder state is 00, the only possibilities for the state at the next shift are 00 or 10, whereas
if the state is 11 then the next state can either be 11 or 01. Thus, the state diagram depicts the
output branch codeword as a function of the encoder input data bit stream and demonstrates
that convolutional coders can be thought of as finite state machines that change states as a
function of the input bit sequence.

13.2.3 The Tree Diagram

The disadvantage of the state diagram as a method for representing a convolutional encoder
is that it suppresses the time scale. The tree diagram, although a little more complex, adds
the time dimension to the state diagram. Figure 13.4 illustrates the tree diagram for the
encoder shown in Fig. 13.2. At each successive input bit time, the encoding procedure can
be described by traversing the diagram from left to right. Here, in Fig. 13.2, each tree branch
depicts the encoder output codeword branch x;. The approach used to generate the tree
diagram is described as follows: If the input bit is a zero, its associated codeword branch
output is found by moving to the next rightmost branch in the upward direction. If the
input bit is a one, its codeword branch output is found by moving to the next rightmost
branch in the downward direction. Assuming that the initial contents of the encoder is all
zeros, the diagram shows that if the first input bit is a zero, the output branch word is 00
and, if the first input bit is a one, the output branch word is 11. Similarly, if the first input
bit is a one and the second input bit is a zero, the second output branch word is 10. Or,
if the first input bit is a one and the second input bit is a one, the second output branch
word is a 01. Following this procedure, it is easy to see that the input data bit sequence
u = 11011 traces the heavy line illustrated in Fig. 13.4. This path corresponds to the output
codeword sequence x = (Xj, X2, X3, X4, Xs) where xj = 11, x7 = 01, x3 = 01, x4 = 00, and
x5 = 01. From this tree diagram, we notice that the tree repeats itself (period K in general)
after three input bits which corresponds to the constraint length; however, for large K one
readily sees the limitation of using the tree diagram to represent the encoder output. Noting
this periodic feature of the tree diagram suggests an alternate representation, viz., the use of
a trellis diagram first suggested by Forney [10].
 Sec. 13.2 Representation of Convolutional Codes
833

00

Output Branch Word

alak

-or- weer rere-en wee.

-oo* "sa -

01

Legend

GaobKs j2plio (9)

Sees - - - Input
Bit 1

10

Figure 13.3 State transition diagram for rate 1/2, K = 3 encoder

13.2.4 The Trellis Diagram

To produce the trellis diagram for the encoder shown in Fig. 13.2, we take advantage of
the fact that the tree structure repeats itself after K branchings, that is, it is periodic with
period K. We therefore begin by labeling the tree nodes in Fig. 13.4 to correspond to the
four possible code states A, B, C, and D defined by the state diagram illustrated in Fig. 13.3.
To incorporate time in the diagram, let t; denote the time epochs of the codeword branches.
In this case t;+; =1t; + 27; sec, for all j = 1. We arbitrarily set t; = 0. More generally, fora
code rate re = 1/N-, then tj+; =t; + NcTc sec. We now develop the trellis diagram shown
in Fig. 13.5 for the encoder shown in Fig. 13.2.
Assume that the encoding operation begins in state A = 00 shown on the far left of
the trellis at ¢ = t; = 0. If the first information bit is a zero, the encoder moves along the
solid line out of state A = 00 arriving again at state A = 00 at t = fy. The encoder output
is the symbol pair 00 which is labeled on the trellis branch between the two states. On the
other hand, if at t = ft; the first encoder input bit were a one, the encoder would move along
the dashed branch out of state A = 00 arriving at state B = 10 at time t = f2 = 27,. In this
case, the encoder output is 11 which is indicated on the trellis branch connecting states A
and B. The second possible branching takes place at t = f2 where the encoder is in either
834 Convolutional-Coded Digital Communications Chap. 13

Codeword
branch ‘

oo A

Figure 13.4 Tree diagram representation

ty to tg tq ts for encoder shown in Figure 13.2

state A or B. If state A exists and a zero is inputted to the encoder, then it moves along the
solid branch outputting 00 and arriving in state A at t = fp; if a one is input, the encoder
moves to state B along the dashed line outputting 11 at t = &. Similar comments apply if
at t = f2 the encoder is in state B. This process of moving from left to right every bit time
through the trellis and outputting the trellis branch labels continues as long as the encoder
input is clocked with data bits. Using our representation, input information bits which equal
zero cause the encoder to move along the solid trellis branches to the next state while input
bits which equal one cause the encoder to move along the dashed branches. The codeword
branch output sequence x = (x), X2,..., Xj, ...) consists of the sequence of branch labels
Xj = (xj1,Xj2) encountered as the encoder follows the path from left to right. A typical
input output sequence and associated modulator input waveform for this convolutional code
is shown in Fig. 13.6. For L input bits into the encoder, there are 2%</ possible output
sequences that can be generated.
Since convolutional encoding is best understood using illustrative examples, we fur-
ther demonstrate the concepts of the state and trellis diagrams using a rate r, = 1/3 convolu-
 Sec. 13.2 Representation of Convolutional Codes
835

State A=00

B=10

C=01

D=11

Legend
Input bit O
— Input bit 1

Figure 13.5 Trellis diagram for the rate 1/2, K = 3 convolutional encoder of Figure 13.2

tional of constraint length K = 4. Figure 13.7 illustrates the encoder mechanization for the
code generator polynomials defined by

gi(x)=lt+xtx7?
$x
g(x) =14+x4+x3 (13.1)
g3(0) z= 1 ex
For this case, three channel output symbols defined by the output branch word x; =
(xj1,%Xj2, xj3), are produced at multiples of 7, = 37, sec for each input bit. The output
from the 3 mod-2 adders in Fig. 13.7 are sampled to produce the components of x;. The
specific registers summed to produce an output are defined by the generater polynomials in
(13.1). None of the generators equal | so that the code is nonsystematic. Since all powers
of x between 0 and 3 occur in the generator g;(x), the first output symbol is the modulo-2
sum of the input bit and all three previous inputs stored in the shift register. The number of
stages of the shift register is equal to 3, which is the degree of the largest-degree generator
polynomial (the degrees of all generators are the same). The constraint length of the code
is K =4 which is 1 plus the degree of any of the generator polynomials. Thus, an input
bit influences the output due to the next three input bits. The number of states in the trellis
representation of the code is 2* ~! — 8. The state transition diagram for this convolutional
encoder is illustrated in Fig. 13.8 while the corresponding trellis diagram is shown in Fig.
13.9. The trellis is found using the procedure discussed earlier for the rate re = 1/2 code
 836 Convolutional-Coded Digital Communications Chap. 13

Information
Sequence, m | m,=0 Mo= 1 M3= m,4=0 m,=1 mg=0 m7=0

Codeword 00 i eal Ol 01 00 uf19) aes

aes Ai) gad [St _ 52 GT. 62h 7h oe

Xs Ko Xo [xr Xe (kT XT [Xo Xo [xt Xt 1X0 Xo [Xo Xo

sora SendMps H O n
+1
Waveform

t2
-1

t,=T, =2T,

Figure 13.6 Encoder input and output sequences and waveform for L = 7

where the trellis state is the contents of the shift-register in Fig. 13.7. The three components
of a trellis branch label represent the mod-2 sums defined by the generator polynomials or
by the operations indicated in Fig. 13.7. Since x° = 1 is a component of all three generator
polynomials, the labels on the two branches leaving the trellis state are complements of each
other. The careful reader should take time to verify the state and trellis diagrams shown in
Figs. 13.8 and 13.9.

13.3 DECODING CONVOLUTIONAL CODES

Consider now the convolutional coded digital communication system model illustrated in
Fig. 13.10. This model is a special case of the coded system model discussed in Chapter 11,
Fig. 11.2. In order to discuss convolutional decoding techniques, we must introduce the
reader to the notation required to keep track of the codeword branches, codeword (channel)
symbols, and codeword paths that are generated by a random (equal probable ones and ze-
ros) binary data source. In addition, we must be specific about the output of the demodulator
which drives the convolutional decoder. We tacitly assume that the decoder and demodulator
have acquired the various levels of synchronization needed in accomplishing the decoding
task, viz., the demodulator is coherent with respect to the carrier modulation process and has
achieved channel symbol, bit, and codeword branch (node) synchronization. This knowl-
edge, together with that of the convolutional encoder choice, is equivalent to providing the
decoder with the knowledge contained in the trellis diagram. The transition from Fig. 11.2
 yIlUuS Jaysibay

Be
fy= Hy Ay (ffx
jndjno sjoquiAs
pepod)
aBessoyy
(sig

ai
ely

eansiq
LE] AeY ‘¢/[y p= [euolnfoAuos
Japoous

837
838 Convolutional-Coded Digital Communications Chap. 13

Figure 13.8 State diagram for r, =

011 1/3, K =4 convolutional encoder

to 13.10 is accomplished by adding the sub or superscript “j” to binary source bits, channel
oo 299

codeword branches, modulator and demodulator outputs to denote that the j" branch in the
trellis diagram is being processed in the decoder. This branch occurs during the j th time in-
terval tj41 —t; =NcTc, j = 1, where knowledge of t; corresponds to codeword branch or
node synchronization we assume ft; = 0. Assuming further that the IF channel in Fig. 13.10
is memoryless, that is, a CMC, then the demodulator functions to create the output vec-
tor rj; = (rj1,.-.,7jN,) whose components are the projections of received waveform r ;(t)
onto the basis functions of the coded modulator output waveform Si.(t); see Chapters 3 and
4 and Fig. 13.10. Furthermore, by quantizing the demodulator outputs the modulator, IF
channel, and demodulator can be easily converted into the DMC discussed in Section 11.2
of Chapter 11. This model will be useful when we wish to operate the decoder using hard-
or soft-decision outputs from the demodulator. This important design issue will be discussed
further when we present the so-called Viterbi algorithm.

13.3.1 Maximum-Likelihood Decoding of Coded BPSK Modulation

For purposes of being precise, consider now the class of binary convolutional codes of rate
rc = 1/Nc; specifically consider encoding a binary message sequence of L message bits. In
 Sec. 13.3 Decoding Convolutional Codes 839

State

A=000 5 ee ee

B=001

C=010

D=011

E=100

F=101

G=110

H=111

Figure 13.9 Trellis diagram for r. = 1/3, K = 4 convolutional encoder

this context, one of M, = 2! messages sequences exist and can be represented by the binary
L-tuple m = (mj, ..., my). Here m A m; is a binary random variable taking on values one
and zero. Let ui, (with m = 0 or 1) characterize the source encoder output of Fig. 13.10 dur-
ing the What known data bit interval (j — 1)T) <t< < JT =v= Ned conjucrile Otle eaeeretn OUIe re
words, if m = 0 then We= 0 and if m = 1 then ul 11, The Cee Ae be in Fig.
13.10 maps the ee bit ul, into the j'"codeword branch a — (xj; ais Xn) of the trel-
lis diagram where xj; is a binary random variable taking on values ay = 0or ee = | during
the i*° channel symbol interval of the j"” bit interval. To connect the codeword branches to
the physics of the modulator, then as veces x), are mapped one-to-one onto the coded
BPSK modulation vectors $j,= (5) Pa se “) which are equivalently represented by the
coded modulation waveform

Sn eRe pj(t —iTo) (13.2)

i—

for (j — 1)T, < t < jTp sec. Here the i coded message bit (channel symbol) coefficient sj,
takes on values a = Ses if m= 1 during the j'" codeword time interval tia. —t7 =Nele
and sd’ = —,/Ev, if m = 0 during the same time interval; Ey, represents the “energy” of
the BPSK channel symbols and Ty = 7;. Furthermore, the pulse shape pj(t — i7;) of unit
energy serves to characterize the modulator output waveform during the known i” channel
symbol time of the ies codeword branch and is zero elsewhere. Thus, the IF channel coded
modulation waveform for a sequence of L input bits (the message m) can be represented as

Uist=NebTy=LT, (13.3)
L .

2) aSe
er)
j=1
 840 Convolutional-Coded Digital Communications Chap. 13

eet Pe Camce a nnceeren ls Sante See

Codec Codec :

Information Source
Source Encoder

Channel IF
Clock Channel

Information
Sink Decoder Decoder modulator

Figure 13.10 Convolutional coded-digital communication system model

where for each j, m=O or 1 with equal probability and m is one of Pligbinary message
vectors. Now Sm(t) encodes L bits into N,L binary channel symbols si; or one of 2Ne4
ae ee viz., the L trellis codeword branches containing N, channel symbols
x),= (x; x)JNey Substitution of (13.2) into (13.3) leads to the channel waveform

Soepie pj(t iT¢) (13.4)

alot

j** codeword branch

which is observed to be the sum of L waveforms which represent the codeword branches in
the trellis diagram. The vector equivalent to the coded modulation waveform S(t) is de-
fined by the coded modulation vectors Sm = (Soe S27; age Ss‘) where Sj, = (si) SHAKY.JNey
If r;(t) denotes the noise corrupted received waveform occurring during the aie inter-
val tj41 —tj =NcTc, j =1,...,L, then the demodulator creates the output projection
rj =(rji,.--,17jn.) of this waveform using the equivalent basis functions for the modulator
output waveform sj, (t).
We are now in a position to establish the maximum-likelihood decision function for
the j" trellis codeword branch. Assuming infinite “leepear eaeifeh m “Na 0 or 1) is
to be transmitted, then the codeword branch xj, implies Si, A (si! a JNey so that the
Sec. 13.3 Decoding Convolutional Codes 841

decision function can be written as!

Na

p(rjlSh)= [[p (rion) (13.5)

t=

j® trellis branch metric

where we have assumed the noise samples in r;j are independent and p(r; iS, yA Sr; (0; SZ, )
characterizes the conditional joint probability density function of the demodulator output,
or equivalently, the IF channel transition probabilities. If we assume the L bits transmitted
consist of the message m = (m , m2, ..., m,), then the corresponding maximum-likelihood
decision function (conditional upon m being transmitted) becomes

atm =[] »(r)'8h) = 1(|TT>(cts) (13.6)

j trellis branch metric

Since the logarithm is a monotone increasing function of an increasing argument, the de-
coder could equivalently use the decision metric In A(m). In this case, the decoder metric
(decision function) becomes

L Ne ae
In A(m) = 5 In inp Go) (13.7)
j=! j=l

and the maximum-likelihood decoder outputs the vector m = m which maximizes In A(m).
When the IF CMC is modeled as the AWGN channel, then

ij 1 (ri fal 7)
any
p (riilsm) = a exp inane (13.8)

where sii =,/Ey,

= if m= 1 and si! = Ey, if m=0. From (13.8), it is clear that the
probability of making an error in the channel BPSK symbol si is

2En. 2roEE
ee | a 13.9

for all i, 7, and m. Substitution of (13.8) into (13.7) leads to a path (decision) metric Uae
tion) which can be written equivalently in terms of the squared Euclidean distances dtjm=
rj — S/,|2, that is

fom a Gere > ir,= S| aC (13.10)

L .

j=]

1. Note again that m= 0 or 1 for each j = 1,2,..., L; hence, no reason to subscript m in the probability
densities.
842 Convolutional-Coded Digital Communications Chap. 13

This says that one selects the message sequence m which minimizes the sum of the squared
Euclidean distances, that is, set i = m if

min{dz (m)] = d,(m) (1311)

where i is now one of the 2” binary vectors representing the binary message sequence.
This says that the maximum likelihood decoder evaluates the Euclidean distance between
the received vector r = (r},¥2,...,1%z) and all possible coded modulation vectors Sm 4
(Ss! S2 LoS) and selects that binary vector mm = m = (m1, m2,..., mz) which min-
pnived. d(m). We are using the fact that maximum likelihood decoding of convolutional
codes is equivalent to minimum distance decoding; see Chapter 3. Equivalently, for the
AWGN channel the minimum distance decoding can be shown to reduce to that of maximiz-
ing a set of 2'«“ cross-correlation operations; see Chapters 3 and 4. Further simplification of
(13.10) leads to the equivalent decision rule of selecting mm = m if and only if m maximizes
the cross-correlations
16
ngs | Dor +c (13.12)
ele

where for each value of i and j, m =0 or | and C’ is a constant bias. This says that to
perform maximum-likelihood decoding we see that the test in (13.7) requires that one eval-
uate the 2%" cross-correlations defined by (13.12). For large L this algorithmic approach to
decoding becomes too complex to implement, that is, exponential growth in decoder com-
plexity occurs. To avoid this exponential buildup in decoder complexity and still achieve
the performance offered by maximum-likelihood decoding, Viterbi [2] recognized two key
facts: first, that the tree diagram representation of a convolutional code was periodic with pe-
riod K; and second, that paths in the tree diagram actually merge and hence this fact could
be used to simplify the decoding process. Forney [10] clarified this point by introducing
the trellis diagram from which it is clear that certain codeword paths remerge as a function
of the state of the encoder. In what follows, the Viterbi algorithm for use in maximum-
likelihood decoding of convolutional codes is presented.

13.3.2 Maximum-Likelihood Decoding Using the Viterbi Algorithm

As the L-bit message grows in length we observed that the numbers of correlations (decoder
complexity) required for a rate r, = 1/N,- code grows exponentially. The Viterbi algorithm
(VA) is an elegant and efficient method for performing ML decoding of convolutional codes.
The algorithm was first described mathematically by Viterbi [2] in 1967. Since that time
the algorithm has been described many times by many authors; most notably Forney [10]
provides a highly readable and insightful description of the algorithm and its performance.
Recall from the previous section that the function of a ML decoder is to find the
codeword vector sequence S,,= (Se 5 An So) which was most likely to have been
transmitted given the received channel output vector sequence r = (rj,..., rz) (see Fig.
11.2). Corresponding to each Sy there is a binary message vector m = (m ,..., mz), which
corresponds to a path through the trellis that gives the largest value for the log-likelihood
function In A (m) established by (13.7). Maximizing this is equivalent to selecting the binary
vector m, say m, such that the Euclidean distance
Sec. 13.3 Decoding Convolutional Codes 843

dy (m) = |r — S| = (13.13)

is minimum at m=m. Here d;jm 2 Ej — si, | is the distance between the demodulator
output r; during the j" bit interval and the hypothesized coded modulation vector S/, =
(S745 ees Sin.) transmitted during the j" bit interval. In the context that follows, the variable
L is used to truncate the “length” of the cross-correlation operations which in practice will
be a few constraint lengths.

13.3.3 Hard-Decision and Soft-Decision Viterbi Decoding

In order to illustrate this decoding procedure, it is simplest to consider the case of hard-
decision decoding, that is, the case where R j =(Rj1,..., Rjn,) is one of QNec binary vectors
with components taking on values Rj; = 1 or 0. The vector R; is generated from rj by
making hard decisions on each component of rj, that is, Rj; = sgn(rji); see Figs. 11.3,
11.6, and 11.7. As opposed to S$”, these hard-decision vectors are to be compared with all
possible encoder output codeword branch vectors x), = Oi, wees a iNey, generated in the
trellis. From (12.122) we know that the Euclidean distance squared is proportional to the
Hamming distance; this important fact will now be used to rewrite (13.13) and illustrate the
Viterbi algorithm. Thus, for hard-decision decoding, the ML decoding algorithm in (13.13)
uses the decoding components qd‘, =|rj- si,|? == 4Ey,Dj(m) where Dj(m)4 |Rj — x),|
for all j = 1,2,..., L with m =0 or 1. Here the distance measurement is now made in the
Hamming sense, that is, Dj(m) 4 |R; — x),| is the number of places in which the binary
vector R; differs from x},. In other words, for hard-decision decoding of coded BPSK,
the Euclidean distance measure is “equivalent” to the Hainnnng distance. For example, if
the hard- rc BPSK demodulator output during the j codeword branch is R;j=ill
and the j'4 codeword branch is x), == 00 then the distance D;(m) = |Rj — Si,f= 25 Inkthis
context, the ML decision rule for hard-decision decoding is to select that message vector
=(mj,...,mz), say m, selected such that the accumulated Hamming distance

Ib L
Dim) = )/ Dj Gn) = > IR; — Xml (13.14)

is minimum. The Hamming distances Dj(m) = |Rj — x},| are sometimes called branch
metrics and D,,(m) is called the accumulated path metric (consisting of the sum of the
branch metrics) due to transmission of data vector m. Specifically, each codeword branch
vector x), in the trellis diagram is at a Hamming distance

Dj(m) = [Rj — x},| (13.15)

away from the demodulated codeword vector and as such serves as the branch metric for m.
The sum of the branch metrics create the so-called path metric value Dj(m) at time t = tj+1
for j > 1,2,..., L and for each path in the trellis diagram. Using the sum of the branch
metric values to create the path metric value as a function of the state of the encoder and
the demodulator outputs Rj, then ML decoding can be greatly simplified using the Viterbi
844 Convolutional-Coded Digital Communications Chap. 13

algorithm. We now demonstrate the Viterbi algorithm using the encoder of Fig. 13.2 and its
corresponding trellis diagram shown in Fig. 13.5.

Hard-decision Viterbi decoding. Since hard-decision decoding requires that

the IF CMC be converted into the BSC (binary symmetric channel) with transition prob-
ability given by (13.9), it is convenient to convert the encoder trellis in Fig. 13.5 into the
decoder trellis shown in Fig. 13.11. Here we have assumed thatm=110 1 1. The cor-
responding values for the transmitted codeword branches x), and assumed hard-decision
demodulator output vectors R; are illustrated in Fig. 13.11; here the hard-decision demod-
ulator output vectors are called the “received sequence.” As we shall see, the VA essentially
performs maximum-likelihood decoding by reducing the cross-correlation computational
load by taking advantage of the special structure in the code trellis. The advantage of Viterbi
decoding is that the complexity of a Viterbi decoder is not a function of the number of
symbols in the codeword sequence. The algorithm involves calculating the distance (branch
metric) between the demodulator output time ¢; and all the trellis paths entering each state
of the encoder at time t;. The Viterbi algorithm removes from consideration (eliminates
cross-correlation paths) those trellis paths entering each state that cannot possibly be can-
didates for the maximum-likelihood choice. This is accomplished as follows: Consider the
two paths which enter the same state. The one having the largest path metric is chosen; this
path is called the surviving path. This selection of surviving paths is performed for all the
states for each t; = tj-1 + NcT¢. The decoder continues in this way to advance deeper into
the trellis, making decisions by eliminating the least likely paths. The early rejection of
the unlikely paths reduces the decoding complexity, that is, the number of required cross-
correlations which have to be maintained every bit time.
The decoding procedure discussed above can be best understood by simultaneously
considering the encoder trellis diagram in Fig. 13.5 and the decoder trellis diagram illus-
trated in Fig. 13.11. Along with the pypeticsad input message sequence m, the encoder
codeword output sequence x = (x), = Ge bl?) j=1,2,...,5} and the hard-decision
outputs R = (Rj, Ro, ..., Rs) of the demodulator; see Figs. 11.3, 11.6, and 11.7 for a sys-
tem perspective. The encoder codeword branch outputs for each state (see Fig. 13.5) are
known a priori by the encoder and decoder. The labels on the decoder trellis of Fig. 13.11 at
time tj41 =tj; + 27, are the accumulated Hamming distances D ;(m) between the transmit-
ted sequence and the received sequence, viz. with L = j in (13.14) then

Dj(m) = (13.16)

At t =t; we assume that the encoder is initialized to the state A = 00. In fact, (13.16)
serves to define the cumulative Hamming distance (path metric) as the sum of the branch
Hamming distances relative to the demodulator outputs up to time t = t;. For example, at
t = t7 the demodulator output is R2 = 11 and the Hamming distance between R> and the
encoder output of XG = 00(m = 0) is 2, while for x} = 11(m = 1) and the distance is 0.
These accumulated distances are labeled on the decoder branches at t = fy in Fig. 13.11,
that is, going from encoder state “A” at t = t) to state “A” at tf = tf we label the decoder
trellis branch with 2, while going from state “A” at t; to state “B” at t2 we label the decoder
trellis with an accumulated distance of 0. One continues in this manner by labeling the
accumulated distances on the decoder trellis branches as the demodulator outputs occur
 Sec. 13.3 Decoding Convolutional Codes 845

Input data sequence m 1 1 0) 1 1

Transmitted codeword x 11 01 01 00 01

Received sequence R 11 01 01 10 01

ty to
State A=00

Branch
metric
B=10 = >

ay

C=01 =

D=11 rT]

Figure 13.11 Trellis diagram and channel input-output sequences

through t = f. The VA uses these accumulated Hamming distances to find the accumulated
path metric D;(m) which gives rise to the minimum distance, that is, the shortest path
through the trellis relative to the demodulator outputs R = (Rj, ..., Rs).
Viterbi made the following creative and innovative observation which in essence elim-
inates the need to create all cross-correlations required in (13.12). If any two paths in the
trellis merge to a single state at any time f = t;, one of them can always be eliminated in
the search for the shortest (minimum distance) path. For example, in Fig. 13.12 two paths
merge at t = f5. The upper path has the accumulated distance (path metric) Ds = 4 while
the lower accumulated path metric Ds = 1. Thus, the upper path need not be given further
consideration with respect to finding the shortest path through the trellis because the lower
path, which enters the same decoder state, has a lower path metric. This lower path is called
the survivor path for this state. Needless to say, this was Viterbi’s astute observation and is
primarily responsible for eliminating the need to compute, store, and accumulate all 2Nel.
correlations required in (13.12). Technically speaking, this is justified on the basis of the
Markov nature of the encoder state, that is, the present encoder state summarizes the en-
coder history in the sense that past encoder states do not affect future trellis branch outputs.
More generally, we can state that a (N., K-, K) convolutional encoder has 2*-! states in
the trellis at time t =¢;4; =1t; + NcTc, j = 1. Each of these states can be entered by one
of two paths. The VA consists of computing the accumulated branch metrics for the two
paths entering each state and eliminating one of them. This computation is done for each of
846 Convolutional-Coded Digital Communications Chap. 13

ty to tg tg ts
Path metric = 4
1 ia
State A=00

Figure 13.12 Accumulated Hamming distances at t = ts for two merging paths

the 2“—! states at time ¢; and the survivors are retained; then the decoder moves to time tj+1
and repeats the process of updating the accumulated path metrics and creating the survivors.
To further demonstrate the process, we consider decoding the message for the demod-
ulator outputs and trellis diagram given in Fig. 13.11. Following Sklar [27], we assume that
the decoder is synchronized with the encoder and knows that the initial state at t = f), is
A = 00. In practice, knowledge of the initial state is not necessary; however, for purposes of
discussion there is no loss in generality. From state A = 00 the only possible transitions are
to state A = 00 or B = 10; see Fig. 13.13(a). At time ¢ = fy state “A to A” has branch metric
Aa = D2(A) = 2 while state “A to B” has branch metric 4p = D2(B) = O. At t = fo there
are two possible branches leaving states A = 00 and B = 10; see Fig. 13.13(b); the cumu-
lative path metrics of the codeword branches are labeled A4 = D3(A), AR = D3(B), Ac =
D3(C), Ap = D3(D) terminating at t = £3. At time t = & in Fig. 13.13(b) there will be two
possible trellis branches diverging from each state. As a result there are two paths entering
each state at t = f4; see Fig. 13.13(c). For each state, the path giving rise to the largest cu-
mulative can be eliminated with regard to further computation and the resulting decoding
trellis is illustrated in Fig. 13.13(d). Here we observe that 44 = D4(A) = D4(B) = Ag (ac-
cumulated path metrics) equal at t = t4. When this occurs in the decoding process, one path
can be chosen at random for elimination without affecting the average error probability. The
surviving path for each state is illustrated in Fig. 13.13(d).
Thus far in the decoding process, the cumulated branch distance metrics have been
the major concern and no mention has been made regarding the decoded bit stream. With
respect to this we note from Fig. 13.13(d) that there is only a single surviving path at time fy
and this path (shown dashed) has been produced by the input bit m = 1; the decoder outputs
a | as its first decoded bit. Here we can see how the decoding of the surviving branch is
facilitated by having drawn the lattice branches with solid lines for input zeros and dashed
lines for input ones. Note that the first bit was not decoded until the path metric computation
Sec. 13.3 Decoding Convolutional Codes 847

had proceeded three bit times in practice into the trellis. This so-called coding delay can be
as much as five times the constraint lenght in bits, that is, 5K Ty sec.
Now that we have demonstrated the decoder bit decision process and the concept of
coding delay, let us continue with the decoding process. At each succeeding step (t > t4)
in the decoding process there will always be two possible paths entering each state of
the encoder and decoder; one of the two will be eliminated by comparing cumulated path
metrics leaving the survivor paths. At t = t4, Fig. 13.13(e) demonstrates the two departing
paths which arrive at the four states at time fs. At t = fs there are two paths entering each
state and one of each pair can be eliminated; Fig. 13.13(f) shows the survivors. At this point
we note that we cannot make a decision on the second message bit because there still are two
paths leaving the state 10 node at time fy. At time f@ in Fig. 13.13(g) we see the pattern of
remerging paths, and in Fig. 13.13(h) we see the survivors at time fg. Also, in Fig. 13.13(h)
the decoder outputs one as the second decoded bit, corresponding to the single surviving
path between f2 and f3. The decoder continues in this way to advance deeper into the trellis
and to make decisions on the message bits by eliminating all paths but one.

Soft-decision Viterbi decoding. For soft-decision decoding the VA described

here, using the Hamming distance decoding metric, works in exactly the same manner using
soft-decision decoding metrics [18]. A minor difference is that for soft-decision decoding
the path with the largest log-likelihood metric, In A(m) given in (13.12), is being found
rather than the path with the smallest Hamming distance. Thus, discarded path segments are
the segments with the smaller (rather than the larger) of the two metrics. In both cases, the
path most likely to have been followed by the encoder is found.

13.3.4 Path Memory and Synchronization Requirements

The storage requirements of the Viterbi decoder grow exponentially with constraint length
K. For a code with rate 1/N,, the decoder retains a set of 2X—! paths after each decoding
step. With high probability, these paths will not be mutually disjoint very far back from the
present decoding depth. All of the 2“ ~! paths tend to have a common stem which eventually
branches to the various states. Thus, if the decoder stores enough of the history of the 2 ~!
paths, the oldest bits on all paths will be the same. A simple decoder implementation, then,
contains a fixed amount of path history and outputs the oldest bit on an arbitrary path each
time it steps one level deeper into the trellis. The amount of path storage required, S, is [13]
5 =h2*—! where h is the length of the information bit path history per state. A refinement,
which minimizes the value of h, uses the oldest bit on the most likely path as the decoder
output, instead of the oldest bit on an arbitrary path. It has been demonstrated, [13], that a
value of h of 4 or 5 times the code constraint length is sufficient for near-optimum decoder
performance. The storage requirement, S, is the basic limitation on the implementation of
Viterbi decoders. The current state of the art in microelectronics technology limits decoders
to a constraint length of about K = 15. For K > 10 such decoders have been called big
Viterbi decoders.
Codeword branch synchronization is the process of determining the beginning of a
branch word in the received sequence. Such synchronization can take place without new
information being added to the transmitted symbol stream because the received data appear
to have an excessive error rate when the branch synchronization is incorrect. Therefore, a
simple way of accomplishing synchronization is to monitor some concomitant indication of
this large error rate, that is, the rate at which the path metrics are increasing or the rate at
 848 Convolutional-Coded Digital Communications Chap. 13

ty to Path Metric
2
A=00 ———— Ma =2
\0

B=10 ‘ N. i
d,,=0

(a)

Path Metric
ty to tg
2 1 2
Be eee A, =3
XO N1
B=10 - A373
2

C01 \
rare Nese
“
:
D=11_ . : \: do - 0

A=00

B=10

C=01

D=11

Surviving Path Metric

dX, =3
A
i =3
B
Xr. =0
< Figure 13.13 Selection of survivor paths:
(a) Survivors at t = f, (b) Survivors at
d=? t = 13, (c) Metric comparisons at t4, (d)
Survivors at t = t4 (Reprinted from [27])
Sec. 13.3 Decoding Convolutional Codes 849

ty to
A=00 \ 0 .

; “N
=10 - ~*~

C=01

D=11 :
(e)

Surviving Path Metric

A=00 da =1

de =i
B=10

Ne =3
C=01

D=11 x D7

A=00 Nn

B=10

C=01

D=11

Surviving Path Metric

4 to

~0
B=10 - aN d,=2

C=01 he a

D=11 x 5 =i
(h)

Figure 13.13 continued Selection of survivor paths: (e) Metric comparisons at fs, (f)
Survivors at t = ts, (g) Metric comparisons at f, (h) Survivors at t = t¢ (Reprinted from
[27])
 850 Convolutional-Coded Digital Communications Chap. 13

which the surviving paths in the trellis merge. The monitored parameters are compared to a
threshold, and synchronization is acquired accordingly. The important details of this process
lie outside the scope of this book and are usually best understood in practice where decoders
are built to work in practical digital communication systems. :

13.4 PROBABILITY OF DECODING ERROR FOR CONVOLUTIONAL CODES

Analogous with block codes, the probability of error in decoding an L-branch Q-ary con-
volutional tree code transmitted over the equivalent IF DMC (see Chapter 12) is upper
bounded by [2]

L(Q—1) exp [—N.-E(re)] (13.17)

where
Ro; 0<re=
Ro —€ < Ro
E(rc) = [email protected]).. Ro eak = =v.0)* < Cy, (13.18)

for 0 < op < 1 and Ro = Eo(1, Q) is the quantized channel cutoff rate defined in (12.102).
Analogous with block codes, the probability of error in decoding convolutional codes
can be made arbitrarily small by allowing N, to approach infinity in the region where
E(r-) > 0.
Figure 12.8 illustrates the random coding exponent E(r,) achievable in very noisy
channel for convolutional and block codes. The relative improvement increases with code
rate; in particular for r. = Ro = Cy,./2, the exponent for convolutional codes is almost six
times that for block codes. Since exact expressions for probability of error are difficult to
come by, we consider alternate approaches in what follows.

13.4.1 Distance Properties of Convolutional Codes

Like block codes, the distance properties of convolutional codes play a key role in under-
standing the error-correcting capability of the code. This point can best be made using the
encoder shown in Fig. 13.2 to demonstrate the distance between all possible pairs of code-
word sequences that can be generated by a random bit stream. In block codes, it was the
minimum distance Dmin that determined the f error-correcting capability of the code. For
convolutional codes, Dmin is replaced by the so-called minimum free distance D f and re-
lated to t via
Des
t= 5 (13.19)

where [x] denotes the largest integer which does not exceed the bracketed ratio.
To demonstrate the meaning of Ds, we make use of the fact that a convolutional
code is a linear code; hence, there is no loss in generality in simply finding the minimum
distance between each codeword sequence (a cascade of code branch word) and the all-zeros
sequence. First, we redraw the trellis diagram in Fig. 13.5 and relabel the branch outputs
by the Hamming distance from the all-zeros codeword. The resulting trellis is shown in
Fig. 13.14. Assuming that the all-zeros input sequence was transmitted, the trellis paths
of interest are those that start and end in state A = 00 anywhere in between. According
to the VA discussed in the previous section, at least one bit error will occur whenever the
 Sec. 13.4 Probability of Decoding Error for Convolutional Codes
851

State A=00

B=10

C=01

D=11

Legend
Input bitO
ee Input bit1

Figure 13.14 Trellis diagram labeled with distances from the all-zeros path

distance of any other path that merges with state A = 00 at time ¢ = f; is less than that
of the all-zeros path up to time f;, thereby causing the all-zeros path to be eliminated in
the decoding process. The minimum distance for making such an error can be found by
exhaustively evaluating the distance between all paths from the 00 state to the 00 state in
trellis relative to the all-zeros path. For those paths that remerge to the 00 state that are
shorter in time length than the longest path, the sequence will be made of equal length by
appending the necessary number of zeros so as to make all paths have equal length. Using
Fig. 13.14, consider then all paths that diverge from the all-zeros path and then merge for
the first time at some arbitrary trellis node ending at t,. There is one path that departs at
= ft, and remerges to the all-zeros path at t = f4; this path has distance 5 from the all-zero
path and corresponds to the input data sequence of 100. Similarly, there are two paths at
distance 6 from the all-zeros path; one departs at f; and merges at time ts, while the other
departs at t; and merges again at t = f¢, etc. The input message bits for these two paths are,
respectively, 110 and 10100. Each of the message vectors differs in two bits from the all-
zeros path. Thus, the minimum distance is seen to be 5 since the distances for all arbitrarily
long paths in greater than or equal 5. Setting Dr = 5 in (13.19) verifies that Viterbi decoder
corresponding to the encoder in Fig. 13.2 can correct any two channel symbol errors.
For an arbitrary convolutional coder of rather short constraint length, the distance
properties can be examined qualitatively using the so-called code generating function
or transfer function; however, for long constraint lengths the complexity of the transfer
852 Convolutional-Coded Digital Communications Chap. 13

00
iy
: ale

Figure 13.15 State diagram labeled with distances from the all-zeros path

function increases exponentially and other methods must be explored [8], [12]. Table 13.1
summarizes the free distance for several optimum short constraint length rate 1/2 and 1/3
convolutional codes. The details associated with using the transfer function to study the dis-
tance properties of convolutional codes lie outside the scope of this chapter; however, in the
next section we will present one example to demonstrate the approach.

13.4.2 The Transfer Function Approach

Usually the best way to find Dy for larger constraint lengths is by computer simulation
methodology. The analytical tool is to use the transfer function which can in principle be
found from the state diagram. In addition, a rich by-product of establishing the transfer
function is that it can also be used as a tool to calculate the bit error probability performance
for convolutional codes. First, we illustrate the methodology for finding the transfer function
for the encoder shown in Fig. 13.2 and its corresponding state transition diagram shown in
Fig. 13.3. This state transition diagram is redrawn in Fig. 13.15 so as to clearly demonstrate
the all-zeros state into two states which correspond to the state from which the trellis path
of interest leaves and the state to which the path of interest returns. In this figure, the states
A, B, C, and D are identified with A = E. Additionally, the loop from state A = 00 to itself
has been discarded since paths that follow this loop fall outside the span of interest.
Now redraw the trellis diagram of Fig. 13.11 by labeling each codeword branch with
its Hamming distance from the all-zero codeword sequence; see Fig. 13.14. By redrawing
the trellis diagram in this way, we can clearly depict and determine the Hamming distance
from the all-zeros path and the number of information one bits associated with all trellis
paths through the modified bit equivalent state diagram. It is recognized that there are an
infinite number of paths through this state diagram; however, we will enumerate the most
important paths for application to the bit error probability analysis.
To accomplish this, we label the modified state diagram in Fig. 13.16 with the letters
D,L, and N so as to account for additional information contained in the trellis. The letter
D raised to power n is used to characterize the number of ones in the codeword symbols
associated with the particular trellis branch. In particular, the power of D represents the
number of ones. The letter L is associated with every branch and is used to characterize the
Sec. 13.4 Probability of Decoding Error for Convolutional Codes 853

00
/ ee LN Ss ™» ‘

ce D°LN
a eee
Ue 10
DL
seis 11
D*L
ate
401 10
DLN \_ yy DL
aan de

10. i
/ DLN

Figure 13.16 State diagram labeled according to distance, length, and number of input ones

number of branches traversed by a path. Letter N characterizes the number of ones in a path.
The branch label includes N only if the information bit associated with that path is a one. ~
To demonstrate how this notation accounts for a codeword sequence ones, information bit
ones, and path lengths consider the path passing through the state sequence 00, 10, 01, 10,
11, 01, 00. The product of the labels on this path is D’L°N* = D?LN-DL-LN-DLN-
DL - D*L and the codeword sequence for this path is 11, 10, 00, 01, 11 which contains»
seven ones corresponding to the power of D in the overall path product. The corresponding ~
transmitted bit sequence for this path is 1 0 1 1 0 which contains three ones corresponding
to the power of N in the path product. On the other hand, the total number of branches
traversed is seen in the power of L and the path product is six. Consequently, we conclude
that the path label products for the parameters D, L, and N for all paths through the state
diagram characterize the desired features from which we will now take advantage.
Let S;(00), S(10), S(01), S(11), and S.(00) denote a sum of the path label products
for all paths starting with 00 and proceding with 10, 01, 11 and ending with 00. Then the
generating function, which also characterizes the transfer function of a signal flow graph,
can be found by simultaneous solution of the state equations written from Fig. 13.16, viz.
S,(00) = 1
S(10) = S,(00) - D?LN + S(01)LN
S(11) = S(10) - DLN + S(11)DLN (13.20)
S(01) = S(11)DL + S(10)DL

$-(00) = S(01) - D?L

so that the transfer function becomes

TiD/N. Ly Se(00) _ D°L7N Po

eS (O0) ok = DNL(1-L)
13.21
Spin DL + L)N? + D15(1 + L)2N? +... ( )
4+ pit5 p31 Lyne Ss 4
 854 Convolutional-Coded Digital Communications Chap. 13

Each term in the above sum corresponds to the branch label sequence for a path which di-
verges and then merges with the all-zero path. The first term corresponds to the shortest
path which goes through the state sequence 00, 10, 01, 00. This path has length three cor-
responding to L* with Hamming weight five corresponding to D° and a message sequence
Hamming weight of one corresponding to N. The second and third terms in (13.21) enumer-
ate the two paths with Hamming weight of six and a message Hamming weight of two. This
analysis methodology which has just been described by example is applicable in principal
to all convolutional codes; however, for large K it is difficult to write down.

13.5 ERROR PROBABILITY PERFORMANCE OF CONVOLUTIONAL CODED BPSK

MODULATION
The decoded error probability and bit error probability associated with the Viterbi algorithm
does not lend itself to an exact analysis; usually for codes of practical interest the error prob-
ability performance must be established using the method of computer simulation. However,
like maximum-likelihood decoding of block codes it is possible to establish bounds on error
probability performance. In what follows, the decoded error probability and the bit error
probability are given in terms of the transfer function T(D, L, N). Further details of the
methodology used to establish error probability bounds can be found in [2, 12, 18].

13.5.1 Hard- and Soft-Decision Decoded Error Probability

For convolutional codes with Viterbi decoding, the appropriate performance measure is the
probability that the correct path through the trellis is discarded at depth j (time ft = fj).
That is, the performance measure is the probability of making an error at each opportunity
for decoding error. An opportunity to discard the correct path exists each time the decoder
moves one step deeper into the trellis. This measure is comparable to the measure used for
block codes where an opportunity for error exists each time a new block is decoded and
the appropriate performance measure is the block decoding error probability. Note that the
probability of selecting an incorrect path over the full extent of the trellis is an inappropriate
performance measure. That is, it would not be correct to evaluate the probability that the
total semi-infinite decoder output sequence is not continuously merged with the correct
sequence. Over all time, the probability that the decoded path diverges from and remerges
with the correct path at least once approaches unity. Since convolutional codes are linear,
overall average decoding error probability is then equal to the decoding error probability
calculated assuming that the all-zero sequence is transmitted. For the remainder of this
discussion assume, therefore, that the correct path through the trellis is the all-zero path.
The decoding error probability is then the probability that a nonzero path entering state
00 at depth j of the trellis is selected as the survivor. In order to evaluate decoding error
probability, all potential decoding error events at depth j are enumerated, and their error
probabilities are calculated and summed. The summation is the result of a union bounding
argument similar to that used for the error probability bound for block codes. The reader is
referred to [2, 12, 18] for details of the error proabability derivation. There it is shown that
the decoding error probability P(E) is overbound by
[e,2)

P(E) =") ae (13.22)

k=Df
Sec. 13.5 Convolutional Coded BPSK Modulation 855

where a, is the number of nonzero paths with Hamming weight k that merge with the all-
zero path at depth j and diverge once from the all-zero path at some previous depth, and
P, is the probability of incorrectly selecting a path with Hamming weight k. The values a,
can be found by differentiation from the generating function T(D, L, N) for the code with
L=1 and N = 1. Specifically

T(D, 1,1) = Ss a, D* (13.23)

k=Drf
where

ka TOs 1)
ak = ki BRR yk ae, |D=0

The probability of incorrectly selecting a path which is Hamming weight k from the all-
zero path is a function of the channel and the received signal-to-noise ratio. For the binary
symmetric channel, P; is equal to the probability that symbol errors in the k positions where
the correct and incorrect paths differ cause the received sequence to be closer in Hamming
distance to the incorrect path than to the correct path. In particular, it can be shown that [12]

: 1
P= DO (5) reap) toe Ga pe py] (13.24)
e=(k/2)+1
for k even and
k
m= DO (E) pa = pk (13.25)
e=(k+1)/2

for k odd. Furthermore, for infinite quantization and convolutional encoded BPSK modula-
tion it can be shown that [2, 12]

2kreE
n=0/ ees) (13.26)
No
For soft-decision decoding, the calculation of Py, is much more complex and the reader is
referred to [2, 12] for the details.

13.5.2 Bit Error Probability

Bit error probability is calculated using essentially the same techniques as used for the
decoding error probability. To calculate bit error probability, each decoding error event is
weighted by the number of bit errors associated with that event. Since the all-zero path is
the correct path, the number of bit errors associated with a nonzero path is equal to the
number of information ones associated with that path. The generating function is used to
enumerate error events along with the number of attendant bit errors. Recall that the number
of information ones on a nonzero path through the trellis is equal to the power of N in the
generating function term corresponding to that path. In order to change the exponent of N
into a multiplier to weight the path by the number of bit errors, the generating function is
differentiated with respect to N and N is then set to unity.
Consider the generating function given in (13.21). This function enumerates all
nonzero paths through the trellis of Fig. 13.5. Differentiating 7(D, L, N) with respect to
 856 Convolutional-Coded Digital Communications Chap. 13

N and settingN = 1 and

L = 1 results in
dT Dia ee
OS = D+ 2-2 DI 4+B-4- DY +... + G4) DIE...
00 (13.27)
= Ges
k=Df

since the path length does not affect the result. This summation accounts for all possible
nonzero paths through the trellis which merge with the all-zero path at time t;. The first term
indicates that there is a single bit error associated with all nonzero paths which are Hamming
distance five from the all-zero path. The second term indicates that there are four bit errors
associated with all nonzero paths which are Hamming distance six from the all-zero path.
These four bit errors come from two different paths, each of which is associated with two
bit errors. In general, there are a total of cg bit errors associated with all paths which are
Hamming distance k from the all-zero path. Using the generating function to weight error
events by the number of bit errors, an upper bound on the bit error probability is
[o.@)

P,(E) < > cy Py (13.28)

k=Dyf

where Py is calculated using (13.24) for the hard-decision channel and (13.26) for the con-
tinuous output additive white Gaussian noise channel. The coefficients cz, can be found from
(13.27) by differentiation. Unfortunately, the performance analysis of a decoder which uses
the VA does not lend itself to closed form analysis and computer simulation techniques have
to be used. We shall present several such results in the next section for a set of convolutional
codes of greatest practical interest to date.

13.6 ERROR PROBABILITY PERFORMANCE OF CONVOLUTIONAL CODES USING

THE VITERBI ALGORITHM
Using the methodology discussed in the previous section, Ziemer and Peterson [18] evaluate
the bounds on bit error probability performance. Figures 13.17 and 13.18 demonstrate the
hard- and soft-decision bit error probability performance versus E}/No for various rate 1/2
convolutional codes. Figures 13.19 and 13.20 demonstrate the bit error probability perfor-
mance for rate 1/3 convolutional codes. While both upper bounds in (13.22) and (13.28) for
error probability are reasonably good for large signal-to-noise ratios, computer simulation
methods must be used to get results over the signal-to-noise ratio region of interest. Indeed,
extensive computer simulation results have been published for use in practice and can be
found in the frequently used work of Heller and Jacobs [13] and Odenwalder [14]. Typical
results are demonstrated in Figs. 13.21 and 13.22 for rate 1/2 codes and soft-decision de-
coding. Both upper bounds (solid curve) and simulation results (dotted lines) are compared.

13.7 CONCATENATED REED-SOLOMON AND CONVOLUTIONAL CODES

Powerful error-correction capability can be obtained by concatenating a Viterbi decoded
convolutional code with a Reed-Solomon block code. Figures 11.16 illustrates a block dia-
gram of a coded communications system which uses both a convolutional and a block code.
 Sec. 13.7 Concatenated Reed-Solomon and Convolutional Codes
857

Constraint
length 3

ae
SNASGAR EEE
Log
Error
[Bit
Probability]
9;

3 4 5 6 i 8 10 11 12
E,/No, dB
Figure 13.17 Bit error probability performance for various rate 1/2 convolutional codes
using hard-decision decoding (Reprinted from [18])

Information bits from the data source are grouped into blocks of K, to create the 2Ke_ary
symbol alphabet used by the Reed-Solomon code. The coding system which appears first in
the chain, in this case the RS code, is called the outer code of a concatenated code system.
The RS output symbols are converted to their K,.-bit binary representations and input to a
symbol interleaver. The purpose of this interleaver is to enable the spreading of bursts of
channel symbol errors over RS codewords as uniformly as possible. As discussed in Chap-
ter 11, the interleaver changes the order in which the RS symbols are transmitted over the
channel. The symbols are put back in their original order in the deinterleaving buffer in the
receiver. This process of changing the transmission order, transmitting, and reestablishing
the original order causes contiguous bursts of channel errors to be spread uniformly over a
number of RS codewords rather than occurring all in a single codeword. RS codes, as well
as all of the other codes which have been studied, are designed to perform best when chan-
nel errors occur independently from one symbol to the next and do not occur in bursts. For
our purposes, we assume that the interleaver/deinterleaver combination may be considered
processors which take bursts of channel errors and spread these errors over some number of
codewords. The outer interleaver output of Fig. 11.16 serves as the input to a convolutional
coder whose output is sent over the channel. The convolutional code and Viterbi decoder op-
erate exactly as discussed previously; the convolutional code is called the inner code of the
system. The output of the Viterbi decoder is an estimate of the binary sequence which was
858 Convolutional-Coded Digital Communications Chap. 13

aulRi
se
Ae
NN
NWN
WO
lon

NAN rd

9
,
Log
Probability]
Error
[Bit

ee
es
ee
aA7
e177
ies
Pe
oe St
se
see
ENN
COL
A NS
ASAE
HHH Ae
MABHRONS: 6
[\q

[TTTTTT iiratrersa'ae
MRS
4/Ny, 4B
Figure 13.18 Bit error probability performance for various rate 1/2 convolutional codes
using soft-decision decoding (Reprinted from [18])

input to the encoder. The Viterbi decoder output is input to the outer deinterleaver. Detailed
measurements and simulations of Viterbi decoder performance have shown that output er-
rors tend to occur in bursts. It is for this reason that the interleaving scheme is required.
Without the interleaver, the channel would not appear to be memoryless. The outer deinter-
leaver output binary symbols are grouped into blocks of K, to construct the 2*¢-ary symbols
for input to the RS decoder. The RS decoder performs the processing required to estimate
the data source sequence. The RS decoder is able to correct some of the decoding errors of
the Viterbi inner decoder, thus improving communication reliability. Figure 13.23 illustrates
the bit error probability performance of the concatenated coding system of Fig. 11.16. The
convolutional code used for Fig. 13.28 has re = 1/2 and constraint length K = 7. Results
are shown for a number of different RS codes with K = 6,7, 8, and 9. Observe the cod-
ing gain at the bit error probability of P,(E) = 10~> of aproximately 7.0 dB relative to an
uncoded BPSK system.
Concatenated codes consisting of an inner convolutional code and an outer Reed-
Solomon (RS) code are used in several current and planned deep-space missions. The Voy-
ager spacecraft, for example, employs a concatenated coding system based on a (7, 1/2)
convolutional code as the inner code and an 8-bit (255, 223) RS code as the outer code. The
Galileo mission will use an experimental (15, 1/4) convolutional code concatenated with the
same 8-bit RS code. Future missions may use the recently discovered (15, 1/6) convolutional
code together with a 10-bit (1023, 959) RS code [19].
 Sec. 13.7 Concatenated Reed-Solomon and Convolutional Codes
859

~ SCCEET EEE
ee Ed Uncoded He

Ov
Constraint
length 3

Log
Error
[Bit
Probability]
,

FEET SAS
CCCP
3 8
ANE
is/No, dB
Figure 13.19 Bit error probability performance for various rate 1/3 convolutional codes
using hard-decision decoding (Reprinted from [18])

Figures 13.24 to 13.32 show the performance of concatenated codes with ideal and
nonideal interleaver, and point out the difference in performance between 8-bit and 10-bit
RS codes [19]. The E,/No (bit SNR) shown in these figures is the bit signal-to-noise ratio
of the concatenated system, which includes a penalty of 0.58 dB due to the rate of the 8-bit
RS code, or 0.28 dB for the 10-bit RS code.
Figures 13.24 and 13.25 show the bit error rate of the (7, 1/2) code concatenated with
the 8-bit and 10-bit RS codes, respectively. Interleaving depths 7 = 2 and J = 4 are shown
along with ideal interleaving. Higher values of 7 had nonmeasurable performance degrada-
tion relative to ideal interleaving. Larger constraint length codes suffer more from shallow
interleaving since the average length of bursts grows with the constraint length. It must be
observed that results for nonideal interleaving need very large amounts of data (decoded
bits) and are not very accurate even with runs of 10 million or more bits. Figure 13.26 shows
a comparison of the bit error rates of the 8-bit and 10-bit codes concatenated with the (7, 1/2)
code and ideal interleaving. The advantage of the 10-bit RS code becomes apparent only at
very low bit error rates. A different 10-bit RS code specifically optimized for concatena-
tion with the (7, 1/2) code could offer a larger improvement over the 8-bit RS code than the
(1023, 959) RS code, which was optimized for the (15, 1/6) code.
Figures 13.27 and 13.28 show the bit error rate of the (15, 1/4) code concatenated
with the 8-bit and 10-bit RS codes, respectively. Figure 13.29 shows a comparison of the bit
error rates of the 8-bit and 10-bit codes concatenated with the (15, 1/4) code and ideal inter-
 860 Convolutional-Coded Digital Communications Chap. 13

2B
8AS o
nif
EPacbe=
oI
6 N
a
5 aa SAE
VCE
eS
a
& V
ea)
a

S
=
a I
X=
Bs
a
EIFale
Slee
Pehl
SA FatDaw \

Figure 13.20 Bit error probability performance for various rate 1/2 convolutional codes
using soft-decision decoding (Reprinted from [18])

leaving. Now the advantage of the 10-bit RS code over the 8-bit RS code is approximately
0.2 dB at a bit error probability of 10~°. This advantage grows to approximately 0.3 dB at
ieee
Finally, Figs. 13.30 and 13.31 show the bit error rate of the (15, 1/6) code concate-
nated with the 8-bit and 10-bit RS codes, respectively. Figure 13.32 shows a comparison of
the bit error probability of the 8-bit and 10-bit codes concatenated with the (15, 1/6) code
and ideal interleaving. The advantage of the 10-bit code is slightly less than 0.2 dB at a bit
error probability of 10~°, and approximately 0.3 dB at a bit error probability of 107!2.

13.8 OTHER CONVOLUTIONAL DECODING TECHNIQUES

The scope of this book places a constraint on the ability to cover the details of all decoding
algorithms that have been proposed and reported upon in the literature. In particular, prior
to the 1971 paper by Viterbi convolutional codes there were two other decoding algorithms
that were of greatest interest and used in decoding convolutional codes, viz., the sequential
decoding algorithm suggested by Wozencraft [15] and subsequently modified by Fano [16],
and the threshold decoding algorithm suggested by Massey [17]. The two algorithms are
discussed briefly in what follows; however, the reader who is interested in the details should
see [5], [12], [15], and [16].
 Sec. 13.8 Other Convolutional Decoding Techniques 861

10°
=O—oO- Simulation
aa De?

10%
=
a
Ss
te
e
=ro 10°
°
S
s5)
=
Q

10°

Figure 13.21 Bit error probability versus

E»/No for rate 1/2 convolutional codes—
1 0” simulation with 8-level quantization and
32-bit path memory (solid lines); upper
bounds for unquantized AWGN channel
E,/No, 4B (dashed lines) (Reprinted from [13] © IEEE)

13.8.1 Sequential Decoding

A sequential decoder works [15], [16] by generating hypotheses about the transmitted code-
word sequence; it computes a metric between these hypotheses and the received signal. It
goes forward as long as the metric indicates that its choices are likely; otherwise, it goes
beckward, changing hypotheses until, through a systematic trial-and-error search, it finds a
likely hypothesis. Sequential decoders can be implemented to work with hard or soft deci-
sions, but soft decisions are usually avoided because they greatly increase the amount of the
required storage and the complexity of the computations.
While the error probability delivered by the VA decreases exponentially with K, the
number of states of the decoder complexity grows exponentially with constraint length.
Yet the computational complexity of the VA is independent of characteristics in the sense
that, compared to hard-decision decoding, soft-decision decoding requires a small increase
in the number of computations. The error probability associated with sequential decoding
algorithm is asymptotically equivalent to that of maximum-likelihood decoding and does
not require searching all possible codeword paths; in fact, the number of paths searched
is essentially independent of the constraint length, thereby making it possible to use large
constraint lengths (K = 4). This important factor provides for the ability to operate at very
small values of error probability. The major disadvantage of sequential decoding is that the
number of state metrics is a random variable, while the average number of poor hypotheses
and backward searches is a function of channel symbol energy-to-noise ratio, Ey./No =
862 Convolutional-Coded Digital Communications Chap. 13

=-O—O- Simulation

Probability
Error
Bit

Figure 13.22 Bit error probability versus

E,,/No for rate 1/2 convolutional codes—
simulation with 8-level quantization and
32-bit path memory (solid lines); upper
bounds for unquantized AWGN channel
(dashed lines) (Reprinted from [13] © IEEE)

rcEp/No. When Ey./No is small, more hypotheses must be tried than with high values
of Ew./No. Because of this variability in computational load, buffers must be provided to
store the arriving sequences. For small Ey./No the received sequences must be buffered
while the decoder is laboring to find a likely hypothesis. If the average symbol arrival rate
exceeds the average symbol decode rate, the buffer will overflow, no matter how large it
is, causing a loss of data. The sequential decoder typically puts out error-free data until the
buffer overflows, at which time the decoder has to go through a recovery procedure. The
buffer overflow threshold is a very sensitive function of E,/No. Therefore, an important
part of a sequential decoder specification is the probability of buffer overflow, see [5], [15],
and [16].
Figure 13.33 compares the bit error probability achievable versus E;,/No when Viterbi
and sequential decoding are used to decode convolutional codes. Their comparative perfor-
mance is made with BPSK operating in the AWGN channel. The curves compare Viterbi
decoding (rates 5 and 4 soft decision, K = 41). One can see from Fig. 13.33 that coding
gains of approximately 8 dB at Pg = 10~° can be achieved with sequential decoders. Rel-
ative to the approximate 11 dB of coding gain in E,/No promised by the “Shannon limit”
relative to uncoded BPSK, one might conclude that the major portion of that has already
been achieved. Sequential decoding has been used in various early (circa 1970) important
deep space missions.
 Sec. 13.8 Other Convolutional Decoding Techniques 863

1073

t) = Number of symbol
errors that the R-S
| decoder is capable
1044 \ of correcting.

10-5

Probability
Error
Bit
10-6

Figure 13.23 Summary of concatenated coding bit error probability performance with
a K =7,r = 1/2 convolutional inner code and various RS outer codes (Reprinted from
[18])

13.8.2 Threshold Decoding

Certain convolutional codes can be decoded using circuitry similar in complexity to that
used to decode Hamming codes. In order to be decoded so simply, the codes must have
additional structure as defined in detail by Massey [17] and by Lin and Costello [8]. The
simplicity of threshold decoders makes their implementation at very high speeds possible.
Although the performance of these special convolutional codes is not outstanding for the
AWGN channel, the structure is particularly convenient for applications where interleaving
is required; see Chapter 11. For channels where the interference is of the burst type, large
values of coding gain can be achieved.
 864 Convolutional-Coded Digital Communications Chap. 13

INTERLEAVING DEPTH, |
Oo 1=2
a1=4
@ |=o

Probability
Error
Bit

1.9 2.0 2a 2.2 23 2.4 25 2.6

E,/No, dB
Figure 13.24 Concatenated code performance (7, 1/2) convolutional code and 8-bit
(255, 223) RS code (Reprinted from [19])

13.9 PUNCTURED CONVOLUTIONAL CODES

The class of high-rate, r, © 1, punctured convolutional codes [20] are of interest in bandlim-
ited channels, for example, transport of high definition television. These special high-rate
codes are more easily decoded using either Viterbi decoding, [21], [22], [23], or sequential
decoding, [24], [25], because they have the underlying structure of low-rate codes. They also
lend themselves to an easy implementation of variabie-rate coding-decoding, [22], [25], as
well as rate-compatible coding-decoding [26]. Begin and Haccoun [24] present some prop-
erties of punctured codes in addition to providing an upper bound on the free distance which
indicates that high-rate punctured codes are good codes. The reader is referred to the refer-
ence material since the details fall outside the scope of this book.
Sec. 13.9 Punctured Convolutional Codes 865

INTERLEAVING DEPTH, |
Oni
4 1=4
@ l=o0

Probability
Error
Bit

1.7 1.8 1.9 2.0 al 2?) 2.3

E,/No, dB
Figure 13.25 Concatenated code performance (7, 1/2) convolutional code and 10-bit
(1023, 959) RS code (Reprinted from [19])
866 Convolutional-Coded Digital Communications Chap. 13

4 8-BIT RS CODE
B 10-BIT RS CODE

Probability
Error
Bit

ad) 1.9 2.1 2.3 2.5 2 /

E,/N, 6B
Figure 13.26 Concatenated code performance comparison: 8-bit (255, 223) and 10-bit
(1023, 959) RS code concatenated with (7, 1/2) convolutional code (Reprinted from [19])
Sec. 13.9 Punctured Convolutional Codes 867

INTERLEAVING DEPTH, |
Ol=2
a4l1=4
4 1=6
o!1=8
@ |=0

Bit 10-4
Error
Probability

0.5 0.7 0.9 et Ties

E,/No, dB
Figure 13.27 Concatenated code performance (15, 1/4) convolutional code and 8-bit
(255, 223) RS code (Reprinted from [19])
 Convolutional-Coded Digital Communications Chap. 13

INTERLEAVING DEPTH, |
o1=2
1=4
1=6
1=8
| = 00

Probability
Error
Bit

0.2 0.4 0.6 0.8 1.0

E,/No, dB
Figure 13.28 Concatenated code performance (15, 1/4) convolutional code and 10-bit
(1023, 959) RS code (Reprinted from [19])
Sec. 13.9 Punctured Convolutional Codes 869

1072 O 8-BIT RSCODE

4 10-BIT RS CODE

Probability
Error
Bit

10-15

10-16
0.2 0.4 0.6 0.8 1.0 epg lid 1.4

E,/No, dB

Figure 13.29 Concatenated code performance comparison: 8-bit (255, 223) and 10-bit
(1023, 959) RS codes concatenated with (15, 1/4) convolutional code (Reprinted from
[19])
870 Convolutional-Coded Digital Communications Chap. 13

INTERLEAVING DEPTH,|
1=2
1=4
1=6
{=8
| ="e0

Error
Bit
Probability

0.4 06 08 i025) 12
E,/No, dB
Figure 13.30 Concatenated code performance (15, 1/6) convolutional code and 8-bit
(255, 223) RS code (Reprinted from [19])
Sec. 13.9 Punctured Convolutional Codes 871

10°!
1072 INTERLEAVING DEPTH, |
1073 Oo 1=2

1=4
10-4 1=6

10-5 Pak
| = 00
10-6
1077
10-8
10-9
Probability
Error
Bit 10°19
10711
10712

10-13

10°14

OME
0.1 0.3 0.5 0.7 0.9

Figure 13.31 Concatenated code performance (15, 1/6) convolutional code and 10-bit
(1023, 959) RS code (Reprinted from [19])
 872 Convolutional-Coded Digital Communications Chap. 13

O 8-BIT RS CODE
@ 10-BIT RS CODE

Fd Probability
Error
Bit

0.2 0.4 0.6 0.8 1.0 1.2

E,/No, dB
Figure 13.32 Concatenated code performance comparison: 8-bit (255, 223) and 10-bit
(1023, 959) RS codes concatenated with (15, 1/6) convolutional code (Reprinted from
[19])
Sec. 13.9 Punctured Convolutional Codes 873

10°!
RATE = 1/3,K=41,
SEQUENTIAL
RATE =1/2,K=41,
SEQUENTIAL
10°2 RATE 1/2, K=7,
VITERBI, SOFT DECISION

RATE 1/3,K=7,
VITERBI SOFT DECISION
107 RAT ESE 25 Ki= 7
VITERBI
HARD DECISION
RATE 1/3,K=7,
VITERBI,
HARD DECISION
Error
Bit 10°4
Probability
UNCODED BPSK

10°?

=6
1giG 2 4 6 8 10 12 14
E,/No. dB
Figure 13.33 Bit error probability performance for various Viterbi and sequential
decoding schemes using coherent BPSK over an AWGN channel (Reprinted from [28]
© IEEE)
 874 Convolutional-Coded Digital Communications Chap. 13

PROBLEMS
13.1 Assuming one input bit per clock cycle:
(a) What is the constraint length K and code rate for the convolutional encoder illustrated in
Fig. P13.1?
(b) Draw the tree, trellis, and state diagram generated by this encoder.
13.2 Assuming one input bit per clock cycle:
(a) What is the constraint length K and code rate for the convolutional encoder illustrated in
Fig. P13.2?
(b) Draw the tree, trellis, and state diagram generated by this encoder.
13.3. Assuming one input bit per clock cycle:
(a) What is the constraint length K and code rate for the convolutional encoder illustrated in
ign Piss,
(b) Draw the state diagram generated by this encoder.
13.4 Given the encoder illustrated in Fig. P13.2 and suppose the decoder employs hard decisions
to create the binary symmetric channel. Suppose that the received sequence for the first eight
branches is [12]

00 01 10 00 00 00 10 10

Trace the decisions on a trellis diagram labeling the surviving path’s Hamming distance metric
at each node. When a tie occurs in the decision metric, always choose the upper path.
13.5 Given the K =3,r = 1/2 code in Fig. P13.2, suppose the code is used on a BSC and the
transmitted sequence for the first eight branches is [12]

Bits
Coded Message
Bits

Figure P13.1

Bits Coded Message

Bits

Figure P13.2
Problems 875

11 01 01 00 10 11 00

Suppose this sequence is received error free, but the first bit is lost and the received sequence is
incorrectly synchronized giving the assumed received sequence

10 10 10 01 01 10

This is the same sequence except the first bit is missing and there is now incorrect bit syn-
chronization at the decoder. Trace the decisions on a trellis diagram, labeling the survivor’s
Hamming distance metric at each node.
Hint: Note that when there is incorrect synchronization, path metrics tend to remain’, = -<~
relatively close together. This fact is used to detect incorrect bit synchronization.
13.6 Draw the state diagram, tree diagram, and trellis diagram for the K = 3, rate 1/3 convolutional
code generated by

gi(x) =x+ x?

82(x) =1+x
g3(x) =1 te

13.7 Suppose that you were trying to find the quickest way to get from London to Vienna by boat
or train. The diagram in Fig. P13.7 was constructed from various schedules. The labels on each
path are travel times. Using the Viterbi algorithm, find the fastest route from London to Vienna.
In a general sense, explain how the algorithm works, what calculations must be made, and what
information must be retained in the memory used by the algorithm.

, . : Coded Message
Bits Bits

Figure P13.3

London 10 Amsterdam Munich 8 Vienna

Figure P13.7
 876 Convolutional-Coded Digital Communications Chap. 13

13.8 Assuming the coded system operates at capacity Cy,, use Fig. 12.11 and determine the max-
imum coding gain (relative to uncoded BPSK operated at a bit error rate of 10-5) achievable
for a rate 1/2 code. Obtain this gain for hard-decision decoding (J = 2), 3-bit quantization
(J = 4), and infinite quantization. Recall that the communication efficiency required for BPSK
is E,/No = 9.6 dB.
13.9 Assuming that the coded system operates at Ro, repeat Problem 13.8 using Fig. 12.12. Compare
your results with those found in Problem 13.8.
13.10 What is the free distance, D+, achievable with the rate 1/3 convolutional coder illustrated in
Fig. 13.7? The free distance Df is the minimum Hamming distance between two distinct trellis
paths through the code trellis.
13.11 Assuming rate 1/2 convolutional coded BPSK and operation at a bit error probability of 1072;
what is the coding gain relative to uncoded BPSK if:
(a) arate 1/2, K =7 convolutional code and hard-decision decoding is used.
(b) arate 1/2, K =7 convolutional code and 3-bit soft-decision decoding is used.
(c) Repeat (a) and (b) assuming a rate 1/3, K = 7 convolutional code is used.
(d) Compare the coding gains achievable for the two code rates and comment on why the
results are different.
13.12 A concatenated Reed-Solomon/convolutional code is used to improve system performance in
the AWGN channel at a bit error probability of 10~>.
(a) Assuming a K =7, rate 1/2 convolutional code and coded BPSK, use Fig. 13.23 and find
the coding gain achievable (relative to uncoded BPSK) when the 2”-ary RS symbols are
constructed from n-bit blocks of input data bits, n = m = 6, 8.
(b) Compare your results with those found in Problem 13.11.

REFERENCES
. Elias, P., “Error-Free Coding,” JRE Trans. Information Theory, IT-4, 1954, pp. 29-37.
. Viterbi, A. J., “Error Bounds for Convolutional Codes and an Asymptotically Optimum Decoding
Algorithm,” JEEE Trans. Information Theory, vol. IT-13, April 1967, pp. 260-69.
. Forney, G. D., Jr., Concatenated Codes, Cambridge, MA: MIT Press, 1966.
. Gallager, R. G., “Information Theory and Reliable Communication,” New York: John Wiley and
Sons, 1968.
. Michelson, A. M. and A. H. Levesque, Error-Control Techniques for Digital Communications,
New York: John-Wiley and Sons, 1985.
. Berlekamp, E. R., Algebriac Coding Theory, New York: McGraw-Hill, 1968.
.J. Anderson, J: B. and S. Mohan, Source and Channel Coding, Boston, MA: Klumer Academic
Publishers, 1991.
. Lin, S. and D. J. Costello, Jr., Error Control Coding: Fundamentals and Applications, Englewood
Cliffs, NJ: Prentice Hall, 1983.
. Forney, G. D., Jr., “Convolutional Codes I: Algebraic Structure,’ IEEE Trans. Information The-
ory, vol. IT-16, 1970, pp. 720-38.
. Forney, G. D., Jr., “The Viterbi Algorithm,” Proceedings of the IEEE, vol. 61, March 1973, pp.
268-78.
. Forney, G. D., Jr., “Convolutional Codes, II: Maximum Likelihood Decoding,” Information and
Control, vol. 25, July 1974, pp. 222-26.
. Viterbi, A. J. and J. K. Omura, Principles of Digital Communication and Coding, New York:
McGraw-Hill Book Company, 1979.
References 877

13: Heller, J. A. and I. M. Jacobs, “Viterbi Decoding for Satellite and Space Communication,” IEEE
Trans. Communication Technology, vol. COM-19, no. 5, pp. 835-48, October 1971.
14. Odenwalder, J. P., “Error Control” in Data Communications, Networks, and Systems, Thomas
Bartee, ed., Indianapolis, IN: Howard W. Sams, 1985. Also see Error Control Coding Handbook,
San Diego, CA: Linkabit Corp., July 1976.
LS: Wozencraft, J. M. “Sequential Decoding for Reliable Communication Record,” vol. 5, part 2,
1957, pp. 11-15.
. Fano, R. M., Transmission of Information, New York: MIT Press and Wiley, 1961.
. Massey, J. L., Threshold Decoding, Cambridge, MA: MIT Press, 1963.
. Ziemer, R. E. and R. L. Peterson, Jntroduction to Digital Communication, New York: Macmillan
Publishing Company, 1992.
. Pollara, F. and K. M. Cheung, “Performance of Concatenated Codes Using 8-Bit and 10-Bit
Reed-Solomon Codes,” Telecommunications and Data Acquisition Progress Report 42- ce Jan.—
Mar. 1989, pp. 194-201, Pasadena, CA: Jet Propulsion Laboratory.
20. Begin, G. and D. Haccoun, “High-Rate Punctured Convolutional Codes: Structure Properties arid
Construction Technique,” IEEE Trans. on Communications, vol. 37, no. 12, December 1989, pp: ~:
1381-85.
Dc. Cain, J. B., G. C. Clark and J. Geist, ““Punctured Convolutional Codes of Rate (n — 1)/n and
Simplified Maximum Likelihood Decoding,’ JEEE Trans. Inform. Theory, vol. 1T-25, January
1979, pp. 97-100.
fuphe Yasuda, Y., Y. Hirata, K. Nakamura and O. Otani, “Development of a Variable-Rate Viterbi
Decoder and Its Performance Characteristics,” 6” Int. Conf. Digital Satellite Commun., Phoenix, 2
AZ, September 1983.
23. Yasuda, Y., K. Kashiki and Y. Hirata, “High-Rate Punctured Convolutional Codes for Soft Deci-
sion Viterbi Decoding,” IEEE Trans. Commun., vol. COM-32, March 1984, pp. 315-19.
24. Begin, G. and D. Haccoun, “Decoding of Punctured Convolutional Codes by the Stack Algo-
rithm,” Abstracts of Papers, 1986 IEEE Int. Symp. Inform. Theory, Ann Arbor, MI, October
1986, p. 159.
De Haccoun, D. and G. Begin, “High-Rate Punctured Convolutional Codes for Viterbi and Sequen-
tial Decoding,” IEEE Trans. Commun., vol. COM-37, November 1989, pp. 1113-25.
26. Hagenover, J., “Rate Compatible Punctured Convolutional Codes and Their Applications,” JEEE. |
Trans. Commun., vol. COM-36, April 1988, pp. 389-400.
Pay Sklar, B., Digital Communications, Englewood Cliffs, NJ: Prentice Hall, 1988.
28. Omura, J. K. and B. K. Levitt, “Coded Error Probability Evaluation for Antijam Communication
Systems,” JEEE Trans. Commun., vol. COM-30, no. 5, May 1982, pp. 896-903.
: ii Hae 7 ie
&
—

Index
———<§$§| ———— eh Ss

A maximum-likelihood sequence estimation

(MLSE):
Absolutely frequency-invariant structures, 525-45 defined, 578
defined, 513 error probability performance, 591-600
PDM-2 signals: evaluation of, 582-90
ACD of, 525-33 optimum demodulation in presence of ISVAWGN,*
I-Q demodulation of, 533-38 578-600
multiple symbol differential, detection of, zero ISI, Nyquist criterion for, 561-64
538-42 Bandpass limited symmetric DPSK (SDPSK), 491
Adaptive equalization, 557, 611-12 Bandpass random process representation, 287—92
Additive white Gaussian noise (AWGN), 43, 117-19 Bandwidth expansion, 714, 744, 781, 826
channel, 35, 100 Baseband channel, 35
binary vector receivers in, 117-18 function, 711
M-ary vector receivers in, 118-19 Baudot’s code, 711n
optimum design for digital signaling through, Bayes criterion, 141-43
558-78 BCH codes (Bose, Chauhuri, and Hocquenghem),
American Mobile Satellite Corporation (AMSC) 747, 768, 799
systems, 18 binary, 807-8
American Standard Code for Information Interchange error probability performance, 806-9
(ASCI), 13 primitive vs. nonprimitive, 806-7
Amplitude quantization, 732 Q-ary, 808-9
Antenna gain, 35-38 Bhattacharyya bounds/distances, 780, 786
Antipodal receiver, in colored Gaussian noise, 247 Big Viterbi decoders, 847
Antipodal signals, 188-89 Binary BCH codes, 807-8, 811
A posteriori probability, 102-4 Binary block codes:
Application layer, 13 special types of, 761-65
Archimedes, 19, 24-27, 30 biorthogonal binary codes, 763
Autocorrelation demodulator (ACD), 797 orthogonal binary codes, 762-63
of PDM-2 signals, 525-33 transorthogonal binary codes, 763-64
-type differential detectors, 484-85 Binary coded DPSK with infinite quantization,
Automatic gain control (AGC), 2 796-97
Binary cross-correlation receivers, 180-81
Binary phase shift keying (BPSK), 2, 727
coded, 790-93, 838-42, 854-56
Bandlimited channels, 555—620 Binary scalar receivers, 1O8—Th= sr ge
AWGN, optimum design for digital signaling implementation of, 110 ; “een
through, 558-78 performance of, 113-14
controlled ISI-partial response systems, 568-78 Binary symmetric channel (BSC), 735-36
definition/characterization of, 557-58 Binary vector receivers, in the AWGN channel,
duobinary encoding technique, 569-75 117-18
equalization techniques, 600-611 Biorthogonal binary codes, 763
decision-feedback equalization, 605-11 Biorthogonal signals, 198-203
mean-squared error linear equalizer (LE-MSE), Biphase format, 46-47 :
603-5 Biphase (Manchester) baseband Biealine? 66
zero-forcing linear equalizer (LE-ZF), 601, Bit-by-bit binary signaling, 712-13
602-3 Bit error probability, 7, 854
eye pattern, 567-68 block-by-block M-PSK, 466-74
filtering between transmitter/receiver, apportion- classical M-DPSK, 446-47
ment of, 564-67 convolutional codes, 855-56
 880 Index

Block-by-block M-PSK: Channel codec, 7

bit error probability performance, 466-74 Channel cutoff rate, 758, 769
implementation algorithms, 474-77 Channel decoder, 7, 718
maximum-likelihood differential detection, Channel decoding, 718
461-66 Channel encoding, 716
Block-coded digital communications, 753-825 Channel hierarchy, 3
baseband channel perspective, coding from, Channel memory, 712
756-57 Channel memory time, 737
binary block codes, special types of, 761-65 Channel rates, 720-21
binary coded BPSK modulation, bandwidth issue Channel symbol error, 803
for, 764 Channel symbols, 714
block-coded modulation systems, 765-81 Channel throughput, 764
channel throughput/time-bandwidth requirements, Chi-square pdf, 385
764-65 Circular polarization, 38
demodulator design using Ro criterion, 789-99 Classical M-DPSK, 439-60
binary coded DPSK with infinite quantization, bit error probability performance, 446-47
796-97 ; double error rate performance, 452-56
coded DPSK/DDPSK modems with hard error probability performance comparison, 456-60
decisions, 795—96 and imperfect timing, 449-52
hard-decision binary coded BPSK demodula- and noise correlation and power imbalance,
tors, 790-91 447-49
optimization of codes/modem design, 797—99 symbol error probability performance, 441-46
quantization levels, 794 Classical multiple differential phase-shift-keying
soft-decision binary coded BPSK demodulators, (M-DPSK), 439-60
791-93 optimum receiver for, 493-96
fundamental system parameters, 757 Coded digital communications, 711-51
Gallager’s upper bound on ensemble error bit-by-bit binary signaling, information transfer
probability, 786-89 using, 712-13
Golay codes and their error probability Coded digital communication system, 7
performance, 805 architectures, 7-11, 713-36
IF channel perspective, coding from, 756 for channels with memory, 742-44
linear block codes, 799-819 coded-modulation and demodulation-decoding
BCH codes and their error probability system architecture, 718-32
performance, 806-9 coding-decoding system architecture, 714-18
cyclic codes, 805 quantizing the demodulator/detector outputs,
error probability performance for error- 732-35
correcting block codes, 803-4 channel decoding, creating DMC for, 735-36
good error control codes, 802-3 error correction/detection codes, classification of,
Hamming codes and their error probability 744-47
performance, 804—5 interleaving-deinterleaving, 736-41
Plotkin bound for binary codes, 802 synchronization of I-D systems, 741-42
union bound for binary codes, 800-802 Coded DPSK/DDPSK modems with hard decisions,
orthogonal modulation vectors, transmission of, 795-96
760-61 Coded 8-ary PSK, 725-26
performance of, 757-61 Coded message bits, 714
Reed-Solomon (RS) codes and their error Coded-modulation and demodulation-decoding
probability performance, 809-14 system architecture, 718—32
-™.
Shannon’s channel coding theorem, 781—86 modulator characterization, 719-21
‘gh * system architecture, 753-57 signal generator mapper characterization, 721-31
Block codes, 744 IF channel characterization, 731—32
Block errors, 718 Coded symbols, 714
Burst error-correcting codes, 799 Code generating function, 851-52
Code rate, 714, 826
Cc Codeword branch synchronization, 847—50
Codeword symbols, 714
‘Carrier phase error, and crosstalk, 624-27 Coding-decoding system architecture, 714-18
Carrier synchronization circuit, 718 Coding delay, 717, 847
Central Chi-square density, 385-86 Coding gain, 804
Channel architecture, communication system Coding theory, 714
architecture vs., 6 Coherent communications with waveforms, 178-308
Channel bandwidth, 719 binary cross-correlation receivers, 180-81
Channel capacity, 43 performance, 186-91
Index 881

differentially coherent communication, 438-507 continuous phase frequency modulation (CPFM),

differentially encoded M-PSK signals, coherent 651-53
detection of, 225-34 full response CPFM, 652
error probability performance, 186-221 partial response CPFM, 652
inphase/quadrature modulation and demodulation, continuous phase FSK (CPFSK), 653-75
261-84 power spectral density of, 88-91
M-ary waveform receivers, 184-86 continuous phase frequency-shift-keying
performance, 191-201 (CPFSK), 89-91
matched filter receivers, 181-84 full response CPM, 89
partially coherent communication, 406-37 partial response CPM, 89
waveform receivers: sinusoidal frequency-shift-keying (SFSK), 686-88
in colored Gaussian noise channel, 235-61 Continuous phase QFPM (CPQFPM), 688-95, 699%. .
in white Gaussian noise, 179-234 Controlled ISI-partial response systems, 568-78
Coherent receivers, defined, 316 Convolutional codes, 827-77
Colored Gaussian noise channel: bit error probability, 855-56
antipodal receiver in, 247 concatenated Reed-Solomon (R-S) and, 860
finite observation interval, 250-59 decoding, 836-50
infinite observational interval, 247—50 hard-decision/soft-decision Viterbi decoding,
M-ary vector receivers in, 119-24 843-47 ;
optimum waveform receivers in, 235-61 maximum likelihood of coded BPSK -.
convergence of receiver, 242-44 modulation, 838-42
Karhunen-Loeve (K-L) expansion approach, using Viterbi algorithm, 842-43
237-46 path memory/synchronization requirements, oe

receiver truncation/theorem of irrevelance, 847-50

244-46 decoding error, probability of, 850-54
sufficient statistic approach, 259-61 distance properties of, 850-52
time-sampling approach, 235-37 error probability performance: 3
whitening approach, 246-59 of convolutional coded BPSK modulation,
waveform receiver truncation in, 246 854-56
Communication equation, 34 of convolutional codes using Viterbi algorithm, ~ ~ bat
Communication receiver, 3 856
Communications carrier, 711 hard-decision/soft-decision decoded error
Communications channel, 3—7 probability, 854-55
Communication transmitter, 2—3 punctured, 864
Computational cutoff rate, 790 representation, 827-36
Concatenated coding systems, 743, 813-14, 826 polynomial representation, 827-31
coding/modulation matching for, 744 state diagram, 832
Concatenated interleaver-deinterleaver systems, tree diagram, 832
741 trellis diagram, 833-36
Concatenated Reed-Solomon (R-S), and convolu- sequential decoding, 860, 861-62
tional codes, 860 threshold decoding, 860, 863
Confidence thresholds, 734—-35 transfer function approach, 852-54
Constraint length, 826 Convolutional inner code, 743
Continuous memoryless channel (CMC), 99, 107, Convolutionally coded modulations:
az power spectral density of, 72-83
baseband characterization, 717-18 convolution encoder model, 73-75
scalar communications over, 107-8 encoder output spectrum for first-order Markov
vector communications over, 115-40 input, 80-83 2 ea
Continuous phase FPM (CPFPM), 688-95, 699 encoder output spectrum for indeperidehibinary ;
Continuous phase frequency-shift-keying (CPFSK), symbol input, 75-80
89-91 Cooley-Tukey Fast Fourier Transform algorithm,
Continuous phase FSK (CPFSK), 652, 653-75, 698 B21
minimum-shift-keying (MSK), 675-86 Correlated convolutional codes, 77
coherent detection, 680-81 Cyclic codes, 744-47, 805
noncoherent detection, 682-86 Cyclostationary sequence, 59-63, 72, 74-75
optimum bit-by-bit detection, 653-66
coherent detection, 653-58 D
noncoherent detection, 658-66
optimum block-by-block detection, 666-75 Data asymmetry, 271-77
Continuous phase modulation (CPM), 621, 651-95, Manchester data, 274-77
698 NRZ data, 271-74
continuous phase FPM and QFPM, 688-95 Data-link layer, 13
 882 Index

Data pulse streams with data asymmetry: double, 509-54

power spectral density for, 68-72 DPSK, 477-84
Manchester data, 69-72 overlapped M-ary orthogonal modulation
NRZ data, 68-69 generalization, 478-84
Data rate, 712-13 symmetric, 490-92
Decision-feedback equalizer (DFE), 605-11 multiple symbol differential detection technique,
mean-squared error decision feedback equalizer 460-77
(DFE-MSE), 609-11 suboptimum. waveform receivers, 484-90
parts of, 605 multiple decision feedback symbol-by-symbol
zero-forcing decision feedback equalizer (DFE- detector, 488-90
ZF), 607-9 narrowband IF filter, sampling detector, 487-88
_ Decision rules, 98, 126-28 wideband IF filter, integrate-and-dump detector,
Decision threshold, 734-35 485-86
Decoded error probability, 854 Differentially encoded M-PSK:
Decoding: coherent detection of signals, 225-34
convolutional codes, 836—S0 optimum (MAP) receiver for, 227-28
hard-decision/soft-decision Viterbi decoding, error probability performance, 228-34
5 843-47 Digital communication system:
maximum likelihood of coded BPSK data formatting, 43-47
modulation, 838-42 defined, 2
using Viterbi algorithm, 842-43 functional architecture, 2-13
path memory/synchronization requirements, channel architecture vs., 6
847-50 coded, 7-11
Decoding error, convolutional codes, probability of, communications channel, 3—7
850-54 uncoded, 7
Deep Space Network (DSN), 14 ISO-OSI network architecture, 11-13
Deep-staggered interleaver-deinterleaver systems, link budgets, 32-42
741 performance criterion, 27-42
Deinterleaver, 7 power and bandwidth, relationship between, 27-32
Deinterleaving, 736-41 Shannon’s capacity theorem, 42-43
Delay modulation or Miller coding, 66-68 subcarriers, communication with, 43-47
Demodulation, tasks involved in, 718” Discrete memoryless channel (DMC), 99, 103, 712
Design specificatons, 3 examples of, 104
Detectability, 99 scalar communications over, 102—7
Detection criteria: Diversity communications (vector receivers), 124—26
Bayes criterion, 141-43 Diversity reception, 124-26
min-max criterion, 143-45 Double differentially coherent communication with
Neyman-Pearson criterion, 145-46 waveforms, 509-54
Detection system, elements of, 100 binary PDM-1 signals, 514-25
Differential detection, 438 autocorrelation demodulation (ACD) of, 514-18
Differential detection of QAM (DQAM), 637-41 I-Q demodulation of, 518-22
Differential encoded M-PSK, optimum (MAP) first-order phase difference modulation (PDM-1),
receiver for, 227—28, 230-31, 234 509-10
Differential encoding, 225, 438 maximum-likelihood (ML) approach, structures
Differentially coherent communication with motivated by, 511-14
waveforms, 438-507 nth order phase difference modulation (PDM-n), 511
autocorrelation demodulator (ACD)-type second-order phase difference modulation
A .. differential detectors, 484-85 (PDM-1), 510
% ae ae
ae block-by-block M-PSK: Double differential phase-shift-keying (DDPSK),
bit error probability performance, 466-74 509, 795-96
implementation algorithms, 474-77 Double error rate performance, classical M-DPSK,
maximum-likelihood differential detection, 452-56
461-66 Duobinary, defined, 556, 569
classical M-DPSK, 439-60 Duobinary encoding technique, 569-75
bit error probability performance, 446-47 Duobinary raised cosine channel, 575
double error rate performance, 452-56
error probability performance comparison, E
456-60
and imperfect timing, 449-52 Effective Isotropic Radiated Power (EIRP), 35
and noise correlation and power imbalance, 8-ary PSK, 725-26, 727
447-49 Encoder output spectrum:
symbol error probability performance, 441-46 for first-order Markov input, 80-83
Index 883

for independent binary symbol input, 75-80 Generalized M-ary symbol error probability bound,
Encoder state, 827 395-97
Equalization, 556 Global Positioning System (GPS), 15-16
techniques, 600-611 GlobalStar, 18, 24, 29
decision-feedback equalization, 605-11 Golay codes, 768, 799
mean-squared error linear equalizer (LE-MSE), error probability performance, 805
603-5 extended, 805
zero-forcing linear equalizer (LE-ZF), 601, Good error control code, and minimum Hamming
602-3 distance, 802-3
Equal a priori probabilities, binary communication Gram-Schmidt orthogonalization procedure, 154-56
with, 126 Gray code mapping, 210, 211, 214, 349
Equivalent noise bandwidth, 29-32 Group delay, 558
Equivalent Nyquist channel characteristic, 562
Erasure-and-error decoders, 804 H
Error correction/detection codes:
classification of, 744—47 Hamming codes, error probability performance,
Venn diagram for, 747 804-5
Error probability performance: Hamming distance, 759-60, 763-64, 769, 799,
coherent receivers, 186—221 800-802, 843, 847
convolutional codes, 854-56 and good error control code, 802-3
maximum-likelihood sequence estimation Hard-decision bounded distance decoder, 803
(MLSE), 591-600 Hard-decision demodulators, 733, 838
scalar communications, 113-15 binary coded BPSK, design of, 790-91
vector communications, 128-31 Hard-decision/soft-decision Viterbi decoding,
Excess bandwidth factor, 563 843-47
Exponential error bound, 765 Helical interleaver, 738-39, 741
Extended Binary Coded Decimal Interchange Code Hybrid interleaver, 741
(EBCDIC), 13 Hypothesis testing, 98, 100-102
Extended Golay code, 805
Eye pattern, 567-68

FE
Imperfect bit synchronization, 268-71
Imperfect carrier synchronizaiton, 267-68
False alarm, 145 Infinite bandwidth Shannon capacity limit, 783
Fast Fourier Transform (FFT), 321 Infinite bandwidth Shannon information transport,
Feedback filter, 605 783
Finite memory, 280 Information bits, 716
First-order phase difference modulation (PDM-1), Information symbols, 800
509-10 Information transportation system:
Forward filter, 605 basic building blocks of, 2-3
Fredholm equation of the first kind, 257 defined, 2
Fredholm equation of the second kind, 257 Information Transportation System (ITS) Pyramid, 7,
Frequency bands, electromagnetic spectrum, 33 9-11
Full response CPFM, 652, 698 Inmarsat, 18, 19-20, 24, 25
Full response CPM, 89 Inner code, 743
concatenated code system, 857-58
G Inphase-quadrature (I-Q) modulation and
#
demodulation, 261-84, 621° ++: nw. ae
Gabor bandwidth, 29 data asymmetry, effect of, 271-77 "pes
Gallager’s upper bound on ensemble error imperfect bit synchronization, effect of, 268-71 a
probability, 786-89 imperfect carrier synchronization, effect of,
very noisy channels, block vs. convolutional 267-68
coding for, 788-89 real and complex signal models, 261-66
Gamma pdf, 385 suboptimum detection, 277-84
Gaussian integral Q(x), 147-53 See also I-Q demodulation; I-Q modulations
series approximation for, 147-53 Integrated Services Digital Network (ISDN), 16
Generalized M-ary Markov sources: Interface control documents (ICDs), 3
power spectral density of, 63-68 Interleaver, 7
biphase (Manchester) baseband signaling, 66 Interleaver span, 738
delay modulation or Miller coding, 66-68 Interleaving-deinterleaving, 736-41
NRZ baseband signaling, 65 concatenated interleaver-deinterleaver systems,
RZ baseband signaling, 65 741
 884 Index

Interleaving-deinterleaving (cont.) Linear vector space, connecting to physical waveform

deep-staggered interleaver-deinterleaver systems, space, 157-59
741 Link budgets, 32-42
helical interleaver, 738-39, 741 L-orthogonal signals, 212-19
hybrid interleaver, 741 suboptimal receiver for coherent detection of, 220
periodic block interleavers-deinterleavers, 737-38 symbol error probability of, 218
periodic convolutional interleavers-deinterleavers, Low noise amplifier (LNA), 35
739-40, 742
pseudo-random interleavers, 741
shallow-staggered interleaver-deinterleaver
systems, 741
synchronization of I-D systems, 741-42 Mapper, 721
Interleaving delay, 738, 740 M-ary amplitude modulation (M-AM), 219-21
Intersymbol interference (ISI), 279, 556, 558 M-ary scalar receivers, 111-13
I-Q demodulation: implementation of, 112
of binary PDM-1 signals, 518—22 performance, 114-15
of PDM-2 signals, 533-38 M-ary symbol error probability bound, generalized,
I-Q modulations, 622-51 395-97
quadrature amplitude modulation (QAM), 627-41 M-ary symmetric channel, 102
quadrature biorthogonal modulation, 641-51 M-ary vector receivers:
unbalanced QPSK (UQPSK), 622-27, 698, 727, in the AWGN channel, 118-19
728 in the colored Gaussian noise channel, 119-24
Iridium, 18, 20-22, 26, 27 multiple (diversity), 124-26
Isometric coded modulation, Bhattacharyya bound M-ary waveform receivers, 184-86
and distance for, 780-81 implementation of, using matched filters, 187
ISO-OSI network architecture, 11-13 structure of, 185
model, 12 Matched filter, 7187
defined, 183
J implementation of M-ary waveform receiver using,
187
Jamming margin, 786 Matched filter receivers, 181-84
Joint density of amplitude and phase, 389-90 Mathematical Theory of Communication (Shannon),
713
Matrix generation of binary signal sets, 223-25
K Maximum distance separable (MDS), 811-12
Maximum-likelihood (ML) decision rule, 104
Karhunen-Loeve (K-L) expansion approach, 178,
Maximum-likelihood (ML) decoding, 757
237-46, 653-54
of coded BPSK modulation, 838—42
interconnection of time sampling approach and,
using Viterbi algorithm, 842-43
241-42
Maximum-likelihood sequence estimation (MLSE),
556, 578
L defined, 578
error probability performance, 591-600
Likelihood ratio, 109 evaluation of, 582-90
Limiting error probability for block codes as M Maximum-likelihood vector receiver, 116
approaches infinity, 221-23 Maximum a posteriori (MAP) probability decision
Linear block codes, 799-819, 826 rule, 104~7, 128
«> cyclié codes, 805 Maximum a posteriori (MAP) probability vector
© error probability performance: receiver, 116
BCH codes, 806-9 Mean-squared error decision feedback equalizer
for error-correcting block codes, 803-4 (DFE-MSE), 609-11
Hamming codes, 804-5 Mean-squared error linear equalizer (LE-MSE),
good error control codes, 802-3 603-5
Plotkin bound for binary codes, 802 Message bits, 714
union bound for binary codes, 800-802 Message error probability (MEP), 713, 764, 781
Linear codes, 744 Miller coding, 66-68, 270-71
Linear equalizers: Miller format, 46-47
defined, 600 MILSTAR, 16-17, 20
mean-squared error linear equalizer (LE-MSE), Minimum distance decoding, 112-13, 118, 126, 180,
603-5 842
zero-forcing linear equalizer (LE-ZF), 601, 602-3 Minimum-shift-keying (MSK), 85, 675, 698
Linear polarization, 38 I-Q implementation of, 680
Index 885

multi-amplitude minimum-shift-keying bound on performance of, 354-61

(MAMSK), 697, 699 and frequency error, 338-41, 349-54
precoded MSK, 681 and timing error, 343-54
Min-max criterion, 143-45 performance:
Mobile and personal communications, 17-27 binary nonorthogonal signals, 334-38
Archimedes, 19, 24—27, 30 M-ary equicorrelated signals, 333-34
GlobalStar, 18, 24, 29 M-ary nonorthogonal signals, 327-33
Inmarsat, 18, 19-20, 24, 25 M-ary orthogonal signals, 322-27
Iridium, 18, 20-22, 26, 27 in random amplitude and phase channels, 368-75
mobile satellite communication systems (MSCS), performance of, 375-84
17-18 in random phase channels, 316-21
Odyssey, 18, 22-24, 28 performance of, 321-68
personal communications services (PCS), 18 suboptimum M-FSK receivers, 319-21" ©? = aw
Mobile satellite communication systems (MSCS), performance of, 361-68
17-18 See also Rayleigh channels; Rician channels
architecture, 21 Nonorthogonal signals:
objective, 22 binary, 334-38
Mobil Radio Services (MRS), 24 M-ary, 327-33
Modulation: Nonprimitive BCH codes, 806-7 ;
continuous phase modulation (CPM), 621, 651-95 Non-Return to Zero-Level (NRZ-L) data format, 46 *
I-Q modulations, 622-51 NRZ baseband signaling, 65
Modulation loss, 44 Nth order phase difference modulation (PDM-n), 511
Morse Code, 711n Nyquist.bandwidth, 561, 757, 784
Multi-amplitude minimum-shift-keying (MAMSK), Nyquist criterion, 561
697, 699 Nyquist pulses, 556
Multifrequency minimum-shift-keying (MFMSK),
See Continuous phase FPM (CPFPM) O
Multilevel codes, 735
Multipath, 33, 34 Odyssey, 18, 22-24, 28
Multiphase ACD, See Relatively frequency-invariant One symbol observation:
structures partially coherent receivers based on, 408-9
Multiple amplitude modulated signals (M-AM), 191 performance of, 409-18
Multiple frequency-shift-keying (M-FSK) receivers, Optimum M-FSK receivers, 316-19
316-19 Orthogonal M-FSK detection:
Multiple phase-shift-keyed (M-PSK) signals, 191, bound on performance of, 354-61
206-12, 438, 727 and frequency error, 338-41, 349-54
Multiple symbol differential detection, of PDM-2 and timing error, 343-54
signals, 538-42 Orthogonal binary codes, 762-63
Multiple symbol observation: Orthogonal signals, 196-98
partially coherent receivers based on, 419-22 Outer code, 743
pairwise error probability, 424 concatenated code system, 857
performance of, 422-31
performance sensitivity to mismatch, 430-31 P
special cases, 425-30
p% fractional bandwidth, 29
Pairwise probability of error (PPE), 160-62, 194, *
591-93 J «
NAVSTAR, Global Positioning System (GPS), 15-16 defined, 160 te
Negative equally probable (NEP) signals, 64 Partially coherent receivers, 406-37 ‘ ae
Network layer, 13 based on multiple symbol observation, 419-22. *-™
Neyman-Pearson criterion, 145-46 pairwise error probability, 424 xyh +
Noise bandwidth, 29-32 performance of, 422-31
Noise bursts, 736-37 performance sensitivity to mismatch, 430-31
Noise enhancement, 600 special cases, 425-30
Noiseless channel, 102 based on one symbol observation, 408-9
Noncentral Chi-square density, 386-87 performance of, 409-18
Noncoherent communications with waveforms, M-PSK signals, maximum likelihood block-by-
309-405 block detection of, 419-22
Noncoherent receivers: performance: ws.
defined, 316 of binary antipodal signals, 410-12 “ee
optimum M-FSK receivers, 316-19 of binary nonorthogonal signals, 413-15 ‘
orthogonal M-FSK detection: of binary orthogonal signals, 412-13
 886 Index

Partially coherent receivers (cont.) central chi-square density, 385-86

performance (cont.) Gaussian density, 109
of M-ary orthogonal signals, 415-18 joint density of amplitude and phase, 389-90
Partial response CPFM, 652, 698 multivariate Gaussian density, 119
Partial response CPM, 89 noncentral Chi-square density, 386-87
Partial response (PR) systems, 556, 575-78 normal density, See Gaussian density
PDM-2? signals: Rayleigh density, 387-88
ACD of, 525-33 Rician density, 388-89
I-Q demodulation of, 533-38 Probability of Message Error, 99
multiple symbol differential detection of, 538-42 Pseudo-random interleavers, 741
Performance criterion, 27-42 Public Switched Telephone Network (PSTN), 17--18,
link budgets, 32-39 24
relation between power and bandwidth, 27-32 Punctured convolutional codes, 864
signal-to-noise ratio (SNR) as, 40-42
Periodic block interleavers-deinterleavers, 737-38 Q
Periodic convolutional interleaver-deinterleaver,
739-40, 742 Q-ary BCH codes, 808-9
Physical channel, as propagation media, 34-35 primitive, parameters for, 811
Physical layer, 13 Q-ary coded M-PSK, 778-79
Physical tilted phase, 90 Q-ary coded ASK, 755, 771, 797
Plotkin bound for binary codes, 802 Q-ary RS code parameters, 811
Polarization, defined, 38 Quadrature amplitude modulation (QAM), 2, 178,
Polynomial representation, of convolutional codes, 438, 627-41, 698, 727, 735
827-31 coherent detection, 628-37
Polyphase signals, symbol error probability for, differential detection, 637-41
285-86 and multilevel codes, 735
Postcursor ISI, 605 Quadrature amplitude-shift-keying (QASK), 627n,
Power spectral density, 27 698, 727
of continuous phase modulation (CPM), 88-91 Quadrature biorthogonal modulation (QBOM),
continuous phase frequency-shift-keying 641-51, 698
(CPFSK), 89-91 quadrature frequency/phase modulation (QFPM),
full response CPM, 89 642-49, 698
partial response CPM, 89 quadrature-quadrature phase-shift-keying
of convolutionally coded modulations, 72—83 (Q*PSK), 649-51, 698
convolution encoder model, 73-75 Quadrature overlapped raised cosine (QORC)
encoder output spectrum for first-order Markov modulation, 695—97, 699
input, 80-83 Quadrature phase-shift-keying (QPSK), 84-85,
encoder output spectrum for independent binary 696-99
symbol input, 75-80 Quality of service (QOS), 13
for data pulse streams with data asymmetry, 68-72 Quantization levels, 733
Manchester data, 69-72 Quantizing circuit, 732
NRZ data, 68-69
of generalized M-ary Markov sources, 63-68 R
biphase (Manchester) baseband signaling, 66
delay modulation or Miller coding, 66-68
Radio Determination Satellite Services (RDSS), 18
NRZ baseband signaling, 65 Raised-cosine channels, 563
RZ baseband signaling, 65
Random amplitude and phase channels, noncoherent
spectral occupancy, I-Q modulation techniques receivers in, 368-75
"baer compared to, 83-88 Random coding argument, 765
of synchronous data pulse streams, 56—63 Random coding exponent, rate curve vs., 788
generated by binary, nonzero mean,
Random coding techniques, 779
cyclostationary sequence, 60-63 Random errors, 799
generated by binary, zero mean, cyclostationary
Random phase channels, noncoherent receivers in,
sequence, 59-60
316-21
generated by binary, zero mean, WSS sequence, Rayleigh channels:
57-59 optimum receivers in, 370--72
Precoded MSK, 681
performance in, 375-76
Precursor ISI, 605
binary correlated signals, 376
Presentation layer, 13
M-ary orthogonal signals, 375-76
Primitive BCH codes, 806-7
Rayleigh density, 387-88
Probability of buffer overflow, 862
Receiver input signal-to-noise ratio, 777-78
Probability density functions:
Recurrent codes, 826
Index 887

Redundant parity check symbols, 800 vector communications, 124

Reed-Solomon (R-S) codes, 734, 743, 747, 768, 799 waveform communications, 244
error probability performance, 809-14 Sinusoidal frequency-shift-keying (SFSK), 86,
Relative efficiency, communication scheme, 43 686-88
Relatively frequency-invariant structures, 546-50 Slowly varying phase error, 624n
defined, 513 Soft-decision binary coded BPSK demodulators,
Residual carrier, 45 791-93
Rician channels: Soft-decision decoding, 733
optimum receivers in, 372-75 Viterbi, 847
coherent detection, 373-74 Source codec, 7
noncoherent detection, 374-75 Source encoder, 7, 716
performance in, 376-84 Space Station Freedom via Tracking and Data Relay
coherent detection, 380-82 Satellite System (SF/TDRSS), 810
noncoherent detection, 376-80 Spectral occupancy, I-Q modulation techniques
Rician density, 388-89 compared to, 83-88
RZ baseband signaling, 65 Spectrum utilization efficiency, 711 —
RZ coding, 270 Spread spectrum communications, 785-86
State diagram, 585, 832
S Statistical Decision Theory, 99
Suboptimum detection, 277-84
Sampling theorem, 757 average error probability performance design,
Scalar communications, 98-115 282-84
binary scalar receivers, 108-11 worst case design, 280-82
performance, 113-14 zero signal distortion, 278-80
error probability performance, 113-15 Suboptimum M-FSK receivers, 319-21
M-ary scalar receivers, 111-13 performance of, 361-68
performance, 114-15 bandpass filter receiver, 364-68
optimum scalar receivers, 108-13 spectrum analyzer receiver with autocorrelation
over the memoryless channel, 100-115 function truncation, 361-64
continuous, 107-8 Suppressed carrier system, 45
discrete, 102-7 Surviving path, 844
Scrambling, See Interleaving Survivor paths, 845, 847
Second-order differential detection, 509 Symbol error probability, 713
Second-order phase difference modulation (PDM-2), for polyphase signals, 285-86
Symmetric DPSK (SDPSK), 490-92
510
Sequential decoding, 860, 861-62 bandpass limited, 491
Session layer, 13
Synchronization function, 2
Shallow-staggered interleaver-deinterleaver systems, Synchronous data pulse streams:
power spectral density of, 56-63
741
generated by binary, nonzero mean,
Shannon capacity, 752, 781, 785
cyclostationary sequence, 60—63
Shannon limit, 43, 781-85, 862
generated by binary, zero mean, cyclostationary
Shannon-Nyquist sampling rate, 719
sequence, 59-60
Shannon’s capacity theorem, 42-43
Shannon’s channel coding theorem, 781—86 generated by binary, zero mean, WSS sequence,
channel capacity/matched filter/spread spectrum 57-59
communications interconnect, 784-85
System requirements, 3
Systems Engineering and Integration (SE& I), 3
Shannon limits and information transfer physics,
781-83
spread spectrum communications, 785—86 T tr teems.
union bound for binary orthogonal codes, 786
Shift register process, 582
Signal detection, 99 Tapped delay line, 601
Signal generator mapper characterization, 721—31 Telecommunication networks/services, 2, 14-27
coded BPSK, 722-24 Deep Space Network (DSN), 14
coded 8-ary PSK, 725-26, 727 Global Positioning System (GPS), 15-16
coded QPSK, 724-25 Integrated Services Digital Network (ISDN), 16
Signal power-to-noise power ratio, 782 MILSTAR, 16-17, 20
Signal-to-noise ratio (SNR), 40-42 mobile and personal communications, 17—27
threshold, 41 Archimedes, 19, 24-27, 30
Simplex signal set, 204-6 GlobalStar, 18, 24, 29
Union bound for, 205-6 Inmarsat, 18, 19-20, 24, 25
Singular detection, 124 Iridium, 18, 20-22, 26, 27
888 Index

Telecommunication networks/services (cont.) V

mobile and personal communications (cont.)
mobile satellite communication systems Vector communications, 115—40
(MSCS), 17-18 binary vector receivers:
Odyssey, 18, 22-24, 28 in the AWGN channel, 117-18
personal communications services (PCS), 18 performance of, 128-29
Tracking and Data Relay Satellite System decision rules, 126-28
(TDRSS), 14-15 error probability performance, 128-31
Theory of Signal Decisions, 98 of two-dimensional M-ary vector receivers,
Theory of Signal Detection, 98 137-40
3-dB bandwidth, 28-29 M-ary vector receivers:
Threshold decoding, 860, 863 in the AWGN channel, 118-19
Time-sampling approach, 235-37 in the colored Gaussian noise channel, 119-24
interconnection of K-L approach and, 241-42 multiple (diversity), 124-26
Top-down system engineering, 711 performance of, 129-31
Tracking and Data Relay Satellite System (TDRSS), union bound on performance, 135-37
14-15, 621, 810 message error probability, and translation/rotation
Transducer, 2—3 of signal vectors, 131-33
Transfer function, 851, 852—54 over the CMC, 115—40
Transorthogonal binary codes, 763-64 a posteriori probability computer, 116-17
Transorthogonal signals, 203-6 signal energy, translation to minimize, 133-34
Transparent code, 82 Virtual channel, 3
Transport layer, 13 Viterbi algorithm, 582, 589, 838, 842
Transversal filter, 601 Viterbi decoders, 847, 851, 857
Traveling Wave Tube (TWT), 35, 100
Tree codes, 744 WwW
Tree diagram, 832
Trellis diagram, 585, 833-36 White Gaussian noise:
waveform receivers in, 179-234
U binary cross-correlation receivers, 180-81
M-ary waveform receivers, 184-86
Unbalanced QPSK (UQPSK), 622-27, 698, 727, 728 matched filter receivers, 181-84
Uncoded digital communication system, 7 Whitening operations, 120-21, 178, 246-59
architecture, 7 Wideband IF filter, integrate-and-dump detector,
Uncoded transmission, 325 485-86
Unconstrained bandwidth, 43
Uncorrelated convolutional codes, 76-77 Z
Unequal a priori probabilities, binary communication
with, 126-27 Zero-forcing decision feedback equalizer (DFE-ZF),
Union bound, 193 607-9
for binary codes, 800-802 Zero-forcing linear equalizer (LE-ZF), 601, 602-3
defined, 136 Zero ISI, Nyquist criterion for, 561-64
on M-ary vector receiver performance, 135—37 Zero-order modified Bessel function of the first kind,
for simplex signal set, 205—6 312
Useless channel, 102
 ed
4

“at

¢
~*~

2.*
<"
os

io ’ .
 +

Information Information
Channel
Source Sink

Information Source Source /

Source Encoder Decoder

Information Source Channel Baseband Channel

Source Encoder Encoder Channel Decoder

Information Source Channel Signal Baseband Signal

Source Encoder Encoder Generator Channel

Information Source Channel Signal \F

Source Encoder Encoder Generator Modulator Demodulato
Channel

Information Source Channel Signal \F RF IF

Encoder Generator Modulator
Source Encoder Transmitter Channel Receiver

Information Source Channel Signal \F RF RE RF

Source Encoder Encoder Generator Modulator Transmitter Transmitter fated) Receiver

Information Source Channel Signal \F RF Transmit Physical Receive”

Encoder Generator Modulator
Source Encoder Transmitter Transmit Antenna Channel Antenna

Systems Engineer Pers:

Figure 1.6 Information Transportation System (ITS) pyramid; coded digital communication system functional architecture
 Information
Sink

Source Information
Decoder Sink

Channel Source Information

Decoder Decoder Sink

Channel Source Information

Decoder Decoder Sink

Source Information s xe
Signal. Channel
Sink

ee Rog
Decoder Decoder
Detector

he ae “>
0
yy Rey
Source Informatio n ae? ee
Demodulator Signal Channel
Decoder Decoder Sink
i
Receiver Detector
2)

oe

——-
 DIGIT
COMMUNICATION
TECHNIQUES
Signal Design and Detection
Marvin K. Simon »Sami M. Hinedi = William C. Lindsey
The global transition from analog to digital communication infrastructures makes it necessary
for practicing engineers, professors, and students to understand the theory associated with
Digital Communication Techniques. This timely volume, written by professionals with over 60
years of combined teaching, research, and product development experience, provides an
in-depth state-of-the-art treatment of the important concepts of signal design and detection,
modulation and demodulation, coding and decoding, and synchronization written from the
systems architecting, engineering, analysis, and simulation perspectives.

Digital Communication Techniques presents the methods of coherent, noncoherent, partially coherent,
and differentially coherent detection for various binary and M-ary coded and uncoded modulation
techniques (BPSK, QPSK, MSK, MFSK, QASK, QAM, etc.) that can be applied to the design of
uncoded systems while the Ro-criterion (channel throughput in bits/sec/Hz) and bit error
probability are used as the methodology for designing efficient coded digital communications.

Key concepts include scalar, vector, and waveform communications over the additive white
and nonwhite (colored) Gaussian noise channels. In addition, communications via Rayleigh
and Rician fading (due to multipath) channels are also covered.

The authors present:

* Optimum scalar, vector and waveform receiver principles
* MAP, ML, Bayes and Neyman-Pearson signal detection methods
* Likelihood functions, decision functions, decision boundaries, decision regions
* Cross-correlation and matched filter detection
* The ultimate communications reliability and efficiency achievable by way of Shannon’s
channel capacity or coding theorem
* The union bond and its applications to coded system performance evaluation and the
development of random coding arguments to maximize channel throughput
* The communications efficiency of various practical uncoded and coded modulation,
demodulation, and decoding techniques
* The subjectof block (Reed-Solomon, etc.) and convolutional coding and decoding, and
interleaving-deinterleaving techniques from a systems perspective

From the systems perspective, Digital Communication Techniques offers a complete,

comprehensive, and modern treatment of the various disciplines and algorithms that form
the foundations for implementing advanced digital communication systems.

ISBN O-13-200610-3

* PRENTICE HALL | | | (|
9'780132/006101 ||
- Englewood Cliffs, NJ 07632