Codes and Turbo Codes

Codes and Turbo Codes
Springer
Paris
Berlin
Heidelberg
New York
Hong Kong
Londres
Milan
Tokyo
Claude Berrou (Ed.)

Claude Berrou
Télécom Bretagne
CS 83818
29238 Brest Cedex 3
France
ISBN : 978-2-8178-0038-7 Springer Paris Berlin Heidelberg New York
© Springer-Verlag France, Paris, 2010

Printed in France
Springer-Verlag France is a member of the group Springer Science + Business Media
First edition in French © Springer-Verlag France 2007

ISBN: 978-2-287-32739-1
This work is subject to copyright. All rights are reserved, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and
storage in data banks. Duplication of this publication or parts thereof is permitted only
under the provisions of the German Copyright Law of September 9, 1965, in its current
version, and permissions for use must always be obtained from Springer. Violations are
liable for prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
Product liability: The publishers cannot guarantee the accuracy of any information about
dosage and application contained in this book. In every individual case the user must
check such information by consulting the relevant literature.
Cover design: Jean-François MONTMARCHÉ

Cover illustration: Jean-Noël JAFFRY
under the direction of Claude Berrou (Télécom Bretagne)
The following have contributed to this work:

- Karine Amis,
- Matthieu Arzel,
- Catherine Douillard,
- Alain Glavieux †,
- Alexandre Graell i Amat,
- Frédéric Guilloud,
- Michel Jézéquel,
- Sylvie Kerouédan,
- Charlotte Langlais,
- Christophe Laot,
- Raphaël Le Bidan,
- Émeric Maury,
- Youssouf Ould-Cheikh-Mouhamedou,
- Samir Saoudi,
- Yannick Saouter,
all at Télécom Bretagne,
- Gérard Battail,
at Télécom ParisTech,
- Emmanuel Boutillon,
at the Université de Bretagne Sud,
with the invaluable assistance of Josette Jouas, Mohamed Koubàa and Nicolas
Puech.
Translation: Janet Ormrod (Télécom Bretagne).
Cover illustration: Jean-Noël Jaffry (Télécom Bretagne).
Any comments on the contents of this book can be sent to this e-mail address:
turbocode@mlistes.telecom-bretagne.eu
"The oldest, shortest words — yes and no —
are those which require the most thought"
Pythagoras, fifth century BC
To our late lamented colleagues and friends,

Alain Glavieux and Gérard Graton.
Foreword
What is commonly called the information age began with a double big bang.
It was 1948 and the United States of America was continuing to invest heavily
in high-tech research, the first advantages of which had been reaped during the
Second World War. In the Bell Telephone Laboratories, set up in New Jersey, to
the south of New York, several teams were set up around brilliant researchers,
many of whom had been trained at MIT (Massachusetts Institute of Technology).
That year two exceptional discoveries were made, one technological and the other
theoretical, which were to mark the 20th century. For, a few months apart, and
in the same institution John Bardeen, Walter Brattain and William Shockley
invented the transistor while Claude Elwood Shannon established information
and digital communications theory. This phenomenal coincidence saw the birth
of near-twins: the semi-conductor component which, according to its conduction
state (on or off), is able to materially represent binary information ("0" or "1")
and the Shannon or bit (short for binary unit ), a unit that measures information
capacity.
Today we can recognize the full importance of these two inventions that en-
abled the tremendous expansion of computing and telecommunications, to name
but these two. Since 1948, the meteoric progress of electronics, then of micro-
electronics, has provided engineers and researchers in the world of telecommuni-
cations with a support for their innovations, in order to continually increase the
performance of their systems. Who could have imagined, only a short while ago,
that a television programme could be transmitted via a pair of telephone wires?
In short, Shockley and his colleagues, following Gordon Moore’s law (which
states that the number of transistors on a silicon chip doubles every 18 months),
gradually provided the means to solve the challenge issued by Shannon, thanks
to algorithms that could only be more and more complex. A typical example
of this is the somewhat late invention of turbo codes and iterative processing
in receivers, which could only be imagined because the dozens or hundreds of
thousands of transistors required were available.
Experts in micro-electronics foresee the ultimate limits of CMOS technology
at around 10 billion transistors per square centimetre, in around 2015. This
is about the same as the number of neurons in the human brain (which will,
x Codes and Turbo Codes
however, remain incomparably more powerful, due to its extraordinary network

of connections - several thousand synapses per neuron). Billions of transistors on
the same chip means that there will be easily enough room for algorithms that
require the greatest calculating resources, at least among those algorithms that
are known today. To repeat the slogan of one integrated circuit manufacturer,
"the limit lies not in the silicon but in your imagination". Even so, and to be
honest, let us point out that designing and testing these complex functions will
not be easy.
However, we are already a long way from the era when Andrew Viterbi,
concluding the presentation of his famous algorithm in 1967, showed scepticism
that matched his modesty: "Although this algorithm is rendered impractical
by the excessive storage requirements, it contributes to a general understanding
of convolutional codes and sequential decoding through its simplicity of mecha-
nization and analysis" [1]. Today, a Viterbi decoder takes up a tenth of a square
millimetre in a cellphone.
Among the results presented by Shannon in his founding paper [2], the fol-
lowing is particularly astonishing: in a digital transmission in the presence of
perturbation, if the average level of the latter does not exceed a certain power
threshold, by using appropriate coding, the receiver can identify the original mes-
sage without any errors. By coding, here and throughout this book, we mean
error-correcting coding, that is, the redundant writing of binary information.
Source coding (digital compression), cryptographic coding, and any other mean-
ing that the term coding might have, are not treated in Codes and Turbo codes.
For thousands of researchers and engineers, the theoretical result established
by Shannon represented a major scientific challenge since the economic stakes are
considerable. Improving the error correction capability of a code means, for the
same quality of received information (for example, no more than one erroneous
bit out of 10,000 received in digital telephony), enabling the transmission system
to operate in more severe conditions. It is then possible to reduce the size of
antennas or of solar panels and the weight of power batteries. In space systems
(satellites, probes, etc.), the savings can be measured in hundreds of thousands
of dollars since the weight of the equipment and the power of the launcher are
thus notably reduced. In mobile telephone (cellphone) systems, improving the
code also enables operators to increase the potential number of users in each
cell. Today, rare are those telecommunications systems that do not integrate an
error-correcting code in their specifications.
Another field of application for error-correcting codes is that of mass mem-
ories: computer hard drives, CD-ROMs, DVDs and so on. The progress made
in the last few years in miniaturizing the elementary magnetic or optical mem-
orization patterns has been accompanied by the normal degradation of energy
available when the data is being read and therefore a greater vulnerability to
perturbations. Added to this are the increased effects of interference between
neighbours. Today, it is essential to use tried and tested techniques in telecom-
Foreword xi
munications systems, especially coding and equalization, in order to counter the

effects induced by the miniaturization of these storage devices. Although Codes
and Turbo codes does not explicitly tackle these applications, the concepts de-
veloped and the algorithms presented herein are also a topical issue for mass
memory providers.
This book therefore deals mainly with error-correction coding, also called
channel coding, and with its applications to digital communications, in associa-
tion with modulation. The general principles of writing redundant information
and most of the techniques imagined up until 1990 to protect digital trans-
missions, are presented in the first half of the book (chapters 1 to 6). In this
first part, one chapter is also dedicated to the different modulation techniques
without which the coded signals could not be transported in real transmission
environments. The second part (chapters 7 to 11) deals with turbo codes, in-
vented more recently (1990-93), whose correction capability, neighbouring on
the theoretical limits predicted by Shannon, have made them a coding standard
in more and more applications. Different versions of turbo codes, as well as
the important family of LDPC codes, are presented. Finally, certain techniques
using the principles of turbo-decoding, like turbo-equalization and multi-user
turbo-detection, are introduced at the end of the book.
A particular characteristic of this book, in comparison with the way in which
the problem of coding may be tackled elsewhere, is its concern with applications.
Mathematical aspects are dealt with only for the sake of necessity, and certain
results, which depend on complex developments, will have to be taken as given.
On the other hand, practical considerations, particularly concerning the pro-
cessing algorithms and circuits, are fully detailed and commented upon. Many
examples of performance are given, for different coding and coded modulation
schemes.
The book’s authors are lecturers and researchers well-known for their exper-
tise in the domain of algorithms and the associated circuits for communications.
They are, in particular, the inventors of turbo codes and responsible for general-
izing the "turbo principle" to different functions of data processing in receivers.
Special care has been taken in writing this collective work vis-à-vis the unity
of point of view and the coherence of notations. Certain identical or similar
concepts may, however, be introduced several times and in different ways, which
– we hope – does not detract from the pedagogy of the work, for pedagogy is
the art of repetition. The aim of Codes and turbo codes is for it to be a book
not only for learning about error-correction coding and decoding, a precious
source of information about the many techniques imagined since the middle of
the twentieth century, but also for addressing problems that have not yet been
completely resolved.
[1] A. J. Viterbi, "Error Bounds for Convolutional Codes and an Asymptoti-

cally Optimum Decoding algorithm", IEEE Trans. Inform. Theory, vol. IT-13,
pp. 260-269, Apr. 1967.
[2] C. E. Shannon, "A Mathematical Theory of Communication", Bell System
Technical Journal, Vol. 27, July and October 1948.
Contents
Contributors v
Foreword ix
1 Introduction 1
1.1 Digital messages . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 A first code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Hard input decoding and soft input decoding . . . . . . . . . . . 7
1.4 Hard output decoding and soft output decoding . . . . . . . . . 11
1.5 The performance measure . . . . . . . . . . . . . . . . . . . . . . 11
1.6 What is a good code? . . . . . . . . . . . . . . . . . . . . . . . . 15
1.7 Families of codes . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Digital communications 19
2.1 Digital Modulations . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.2 Linear Memoryless Modulations . . . . . . . . . . . . . . 22
2.1.3 Memoryless modulation with M states (M-FSK) . . . . . 29
2.1.4 Modulations with memory by continuous phase frequency
shift keying (CPFSK) . . . . . . . . . . . . . . . . . . . . 31
2.2 Structure and performance of the optimal receiver on a Gaussian
channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2.1 Structure of the coherent receiver . . . . . . . . . . . . . . 37
2.2.2 Performance of the coherent receiver . . . . . . . . . . . . 42
2.3 Transmission on a band-limited channel . . . . . . . . . . . . . . 59
2.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.3.2 Intersymbol interference . . . . . . . . . . . . . . . . . . . 60
2.3.3 Condition of absence of ISI: Nyquist criterion . . . . . . . 63
2.3.4 Expression of the error probability in presence of Nyquist
filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.4 Transmission on fading channels . . . . . . . . . . . . . . . . . . 69
2.4.1 Characterization of a fading channel . . . . . . . . . . . . 69
2.4.2 Transmission on non-frequency-selective slow-fading chan-
nels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
xiv Codes and Turbo Codes
3 Theoretical limits 83
3.1 Information theory . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.1.1 Transmission channel . . . . . . . . . . . . . . . . . . . . 83
3.1.2 An example: the binary symmetric channel . . . . . . . . 84
3.1.3 Overview of the fundamental coding theorem . . . . . . . 86
3.1.4 Geometrical interpretation . . . . . . . . . . . . . . . . . . 87
3.1.5 Random coding . . . . . . . . . . . . . . . . . . . . . . . . 88
3.2 Theoretical limits to performance . . . . . . . . . . . . . . . . . 91
3.2.1 Binary input and real output channel . . . . . . . . . . . 91
3.2.2 Capacity of a transmission channel . . . . . . . . . . . . . 92
3.3 Practical limits to performance . . . . . . . . . . . . . . . . . . . 96
3.3.1 Gaussian binary input channel . . . . . . . . . . . . . . . 96
3.3.2 Gaussian continuous input channel . . . . . . . . . . . . . 97
3.3.3 Some examples of limits . . . . . . . . . . . . . . . . . . . 99
3.4 Minimum distances required . . . . . . . . . . . . . . . . . . . . 100
3.4.1 MHD required with 4-PSK modulation . . . . . . . . . . 100
3.4.2 MHD required with 8-PSK modulation . . . . . . . . . . 102
3.4.3 MHD required with 16-QAM modulation . . . . . . . . . 104
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4 Block codes 109

4.1 Block codes with binary symbols . . . . . . . . . . . . . . . . . . 110
4.1.1 Generator matrix of a binary block code . . . . . . . . . . 110
4.1.2 Dual code and parity check matrix . . . . . . . . . . . . . 112
4.1.3 Minimum distance . . . . . . . . . . . . . . . . . . . . . . 113
4.1.4 Extended codes and shortened codes . . . . . . . . . . . . 114
4.1.5 Product codes . . . . . . . . . . . . . . . . . . . . . . . . 115
4.1.6 Examples of binary block codes . . . . . . . . . . . . . . . 115
4.1.7 Cyclic codes . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.2 Block codes with non-binary symbols . . . . . . . . . . . . . . . 130
4.2.1 Reed-Solomon codes . . . . . . . . . . . . . . . . . . . . . 130
4.2.2 Implementing the encoder . . . . . . . . . . . . . . . . . . 132
4.3 Decoding and performance of codes with binary symbols . . . . 132
4.3.1 Error detection . . . . . . . . . . . . . . . . . . . . . . . . 132
4.3.2 Error correction . . . . . . . . . . . . . . . . . . . . . . . 134
4.4 Decoding and performance of codes with non-binary symbols . . 143
4.4.1 Hard input decoding of Reed-Solomon codes . . . . . . . 143
4.4.2 Peterson’s direct method . . . . . . . . . . . . . . . . . . 144
4.4.3 Iterative method . . . . . . . . . . . . . . . . . . . . . . . 151
4.4.4 Hard input decoding performance of Reed-Solomon codes 159
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Appendix: Notions about Galois fields and minimal polynomials . . . 161
Contents xv
5 Convolutional codes and their decoding 167

5.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
5.2 Representations of convolutional codes . . . . . . . . . . . . . . 169
5.2.1 Generic representation of a convolutional encoder . . . . 169
5.2.2 Polynomial representation . . . . . . . . . . . . . . . . . . 172
5.2.3 Tree of a code . . . . . . . . . . . . . . . . . . . . . . . . . 173
5.2.4 Trellis of a code . . . . . . . . . . . . . . . . . . . . . . . . 173
5.2.5 State machine of a code . . . . . . . . . . . . . . . . . . . 176
5.3 Code distances and performance . . . . . . . . . . . . . . . . . . 178
5.3.1 Choosing a good code . . . . . . . . . . . . . . . . . . . . 178
5.3.2 RTZ sequences . . . . . . . . . . . . . . . . . . . . . . . . 178
5.3.3 Transfer function and distance spectrum . . . . . . . . . . 180
5.3.4 Performance . . . . . . . . . . . . . . . . . . . . . . . . . 183
5.4 Decoding convolutional codes . . . . . . . . . . . . . . . . . . . . 186
5.4.1 Model of the transmission chain and notations . . . . . . 187
5.4.2 The Viterbi algorithm . . . . . . . . . . . . . . . . . . . . 187
5.4.3 The Maximum A Posteriori algorithm or MAP algorithm 192
5.5 Convolutional block codes . . . . . . . . . . . . . . . . . . . . . . 192
5.5.1 Trellis termination . . . . . . . . . . . . . . . . . . . . . . 193
5.5.2 Puncturing . . . . . . . . . . . . . . . . . . . . . . . . . . 196
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
6 Concatenated codes 201

6.1 Parallel concatenation and serial concatenation . . . . . . . . . . 203
6.2 Parallel concatenation and LDPC codes . . . . . . . . . . . . . . 206
6.3 Permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
6.4 Turbo crossword . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
7 Convolutional turbo codes 213

7.1 The history of turbo codes . . . . . . . . . . . . . . . . . . . . . 213
7.2 Multiple concatenation of RSC codes . . . . . . . . . . . . . . . 215
7.3 Turbo codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
7.3.1 Termination of constituent codes . . . . . . . . . . . . . . 221
7.3.2 The permutation function . . . . . . . . . . . . . . . . . . 222
7.4 Decoding turbo codes . . . . . . . . . . . . . . . . . . . . . . . . 235
7.4.1 Turbo decoding . . . . . . . . . . . . . . . . . . . . . . . . 235
7.4.2 SISO decoding and extrinsic information . . . . . . . . . . 238
7.4.3 Practical considerations . . . . . . . . . . . . . . . . . . . 245
7.5 m-binary turbo codes . . . . . . . . . . . . . . . . . . . . . . . . 249
7.5.1 m−binary RSC encoders . . . . . . . . . . . . . . . . . . 249
7.5.2 m-binary turbo codes . . . . . . . . . . . . . . . . . . . . 251
7.6 Analysis tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
7.6.1 Theoretical performance . . . . . . . . . . . . . . . . . . . 256
xvi Codes and Turbo Codes
7.6.2 Asymptotic behaviour . . . . . . . . . . . . . . . . . . . . 256

7.6.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . 259
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
8 Turbo product codes 271

8.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
8.2 Product codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
8.3 Hard input decoding of product codes . . . . . . . . . . . . . . . 273
8.3.1 Row-column decoding . . . . . . . . . . . . . . . . . . . . 273
8.3.2 The Reddy-Robinson algorithm . . . . . . . . . . . . . . . 274
8.4 Soft input decoding of product codes . . . . . . . . . . . . . . . 277
8.4.1 The Chase algorithm with weighted input . . . . . . . . . 277
8.4.2 Performance of the Chase-Pyndiah algorithm . . . . . . . 280
8.4.3 The Fang-Battail algorithm . . . . . . . . . . . . . . . . . 280
8.4.4 The Hartmann-Nazarov algorithm . . . . . . . . . . . . . 285
8.4.5 Other soft input decoding algorithms . . . . . . . . . . . . 289
8.5 Implantation of the Chase-Pyndiah algorithm . . . . . . . . . . 291
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
9 LDPC codes 297

9.1 Principle of LDPC codes . . . . . . . . . . . . . . . . . . . . . . 297
9.1.1 Parity check code . . . . . . . . . . . . . . . . . . . . . . . 298
9.1.2 Definition of an LDPC code . . . . . . . . . . . . . . . . . 301
9.1.3 Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
9.1.4 Decoding LDPC codes . . . . . . . . . . . . . . . . . . . . 308
9.1.5 Random construction of LDPC codes . . . . . . . . . . . 312
9.1.6 Some geometrical constructions of LDPC codes . . . . . . 315
9.2 Architecture for decoding LDPC codes for the Gaussian channel 318
9.2.1 Analysis of the complexity . . . . . . . . . . . . . . . . . . 318
9.2.2 Architecture of a generic node processor (GNP) . . . . . . 319
9.2.3 Generic architecture for message propagation . . . . . . . 322
9.2.4 Combining parameters of the architecture . . . . . . . . . 325
9.2.5 Example of synthesis of an LDPC decoder architecture . . 328
9.2.6 Sub-optimal decoding algorithm . . . . . . . . . . . . . . 330
9.2.7 Influence of quantization . . . . . . . . . . . . . . . . . . . 333
9.2.8 State of the art of published LDPC decoder architectures 335
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
10 Turbo codes and large spectral efficiency transmissions 343

10.1 Turbo trellis coded modulation (TTCM) . . . . . . . . . . . . . 343
10.2 Pragmatic turbo coded modulation . . . . . . . . . . . . . . . . 347
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
Contents xvii
11 The turbo principle applied to equalization and detection 359

11.1 Turbo equalization . . . . . . . . . . . . . . . . . . . . . . . . . . 360
11.1.1 Multipath channels and intersymbol interference . . . . . 360
11.1.2 The equalization function . . . . . . . . . . . . . . . . . . 362
11.1.3 Combining equalization and decoding . . . . . . . . . . . 366
11.1.4 Principle of turbo equalization . . . . . . . . . . . . . . . 369
11.1.5 MAP turbo equalization . . . . . . . . . . . . . . . . . . . 372
11.1.6 MMSE turbo equalization . . . . . . . . . . . . . . . . . . 381
11.2 Multi-user turbo detection and its application to CDMA systems 396
11.2.1 Introduction and some notations . . . . . . . . . . . . . . 396
11.2.2 Multi-user detection . . . . . . . . . . . . . . . . . . . . . 397
11.2.3 Turbo CDMA . . . . . . . . . . . . . . . . . . . . . . . . . 403
11.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Index 413
Chapter 1
Introduction
Redundancy, diversity and parsimony are the keywords of error correction cod-
ing. To these, on the decoding side, can be added efficiency, that is, making the
most of all the information available. To illustrate these concepts, consider a
simple situation in everyday life.
Two people are talking near a road where there is quite heavy traffic. The
noise of the engines more or less disrupts their conversation, with peaks of
perturbation noise corresponding to the vehicles going past. First assume that
one of the people regularly transmits one letter chosen randomly: "a", "b" ... or
any of the 26 letters of the alphabet, with the same probability (that is, 1/26).
The message does not contain any other information and there is no connection
between the sounds transmitted. If the listener doesn’t read the speaker’s lips,
he will certainly often be at a loss to recognize certain letters. So there will be
transmission errors.
Now, in another scenario, one of the two people speaks in full sentences, on a
very particular topic, for example, the weather. In spite of the noise, the listener
understands what the speaker says better that when he says individual letters,
because the message includes redundancy. The words are not independent and
the syllables themselves are not concatenated randomly. For example, we know
that after a subject we generally have a verb, and we guess that after "clou",
there will be "dy" even if we cannot hear properly, etc. This redundancy in the
construction of the message enables the listener to understand it better in spite
of the difficult transmission conditions.
Suppose that we want to improve the quality of the transmission even further,
in this conversation that is about to take an unexpected turn. To be sure of
being understood, the speaker repeats some of the words, for example "dark
dark". However, after the double transmission, the receiver has understood
"dark lark". There is obviously a mistake somewhere, but is it "dark" or "lark"
that the receiver is supposed to understand? No error correction is possible
using this repetition technique, except maybe to transmit the word more than
twice. "dark lark dark" can, without any great risk of error, be translated as
"dark".
2 Codes and Turbo Codes
More elaborate coding involves transmitting the original message no longer

accompanied by the same word but by a synonym or an equivalent: "dark
dusk", for example. If we receive "dark dust" or "park dusk", correction is
possible by referring to a dictionary of equivalences. The decoding rule would
then be as follows: in the case of an error (if the two words received are not
directly equivalent) and if we can find two equivalent words by changing at
most one letter in the received message, then the correction is adopted, ("dust"
becomes "dusk") and the first of the two words ("dark") is accepted as the
original message. The same would be true if "park dusk" was received, where
the first word would now be corrected. Of course, if a large number of errors
alter the transmission and we receive "park dust", we will probably no longer
understand anything at all. So there is a limit to this error correction capability.
This is the famous Shannon limit, that in theory no correcting code can exceed.
Compared to simple repetition, coding by equivalence, which is more effi-
cient, uses the diversity effect. In this analogy with conversation, this diver-
sity is expressed by a lexicographical property: two distinct words with a close
spelling (dark and lark) are unlikely to have two equivalents ("dusk" and "bird"
for example) that also have a close spelling. Diversity, as presented above, thus
involves constructing a redundant message in a way that minimizes any ambigu-
ity on reception. This is also called temporal diversity as equivalent words in the
message are transmitted at different instants and undergo perturbations of un-
equal intensities. For example, "dark" and "dusk" could be transmitted when a
motorbike or a bicycle, respectively, passed by. In telecommunications systems,
we can search for complementary diversity effects. Frequential diversity involves
cutting up and sending the message in frequency bands that are not perturbed at
the same instant in the same way. As for using several emission and/or reception
antennas, this offers spatial diversity as the paths between antennas do not have
the same behaviour. Jointly exploiting these three types of diversity: temporal,
frequential and spatial, leads to highly efficient communications systems.
Finally, the desire for parsimony, or economy, is imposed by the limitation
of resources, either temporal or frequential, of the transmission. The choice of
"dark dusk" is thus certainly more judicious, from the concision point of view,
than "dark night-fall". However, we sense that the latter message might be
more resistant to multiple errors because it is more redundant (the reception
and resolution of "dirk might-fall" is not problematic if we use the decoding law
mentioned above and extend the correction capability to two errors). Searching
for performance via the redundancy rate, and the parsimony constraint are
therefore in total opposition.
Redundant coding is generally simple to implement and the corresponding
software or hardware has low complexity. Decoding, however, requires compu-
tation techniques that can be costly, even if, in fact, the number of instructions
in the program (typically several hundred, in high level computing language) or
the silicon surface occupied (typically several square millimetres) remains low.
1. Introduction 3
1.1 Digital messages

A digital message is a sequence of characters or symbols taken from an alphabet
of finite size. Genetic information (DNA), to take a natural example, uses an
alphabet of four characters, denoted A, T, G and C, that stand for the initials
of their nitrogen bases (adenine, thymine, guanine, and cytosine). The first
digital transmission technique was Morse code (1832), with its two-character
audio alphabet: TIT or dot, a short tone lasting four hundredths of a second
and TAT or dash, a long tone lasting twelve hundredths of a second. Samuel
F. B. Morse could well have called these characters 0 and 1, which today are
the universal names used in any alphabet with two elements or binary alphabet.
The binary elements, 0 and 1, were first called bits by J. W. Tukey (1943), as a
contraction of binary digit, after rejecting bigit and binit. Shannon borrowed the
term when he wished to introduce the concept of unit of information. Today, it
is preferable to refer to this unit of information as the Shannon, to distinguish
it from the bit, which has acquired a more electronic meaning.
An alphabet having M symbols is called an M -ary alphabet. It can be
transcribed into a binary alphabet by representing each of the M symbols by a
word of m bits, with:

m = log2 (M ) + 1 if M is not a power of 2
or: (1.1)
m = log2 (M ) if M is a power of 2
where x denotes the whole part of x. Multimedia messages (voice, music,
fixed and moving images, text etc.) transiting through communication systems
or stocked in mass memories, are exclusively binary. However, in this book
we shall sometimes have to consider alphabets with more than two elements.
This will be the case in Chapter 4, to introduce certain algebraic codes. In
Chapters 2 and 10, which deal with modulations, the alphabets, which we then
more concretely call constellations, contain a number of symbols that are a power
of 2, that is, we have precisely: m = log2 (M ).
Correction coding techniques are only implemented on digital messages.
However, there is nothing against constructing a redundant analogue message.
For example, an analogue signal, in its temporal dimension, accompanied or fol-
lowed by its frequential representation obtained thanks to the Fourier transform,
performs a judiciously redundant coding. However, this technique is not very
simple and the decoder remains to be invented.
Furthermore, the digital messages that we shall be considering in what fol-
lows, before the coding operation has been performed, will be assumed to be
made up of binary elements that are mutually independent and taking the values
0 and 1 with the same probability, that is, 1/2. The signals that are produced
by a sensor like a microphone or a camera, and then digitized to become bi-
nary sequences, do not generally satisfy these properties of independence and
equiprobability. It is the same with text (for example, the recurrence of "e" in
an English text is on average 5 times higher than that of "f"). The effects of
dependency or disparity in the original message, whether they be of physical,
orthographical or semantic origin or whatever, cannot be exploited by the digital
communication system, which transmits 0s and 1s independently of their con-
text. To transform the original message into a message fulfilling the conditions of
independence and equiprobability, an operation called source coding, or digital
compression, can be performed. Today, compression norms like JPEG, MPEG,
ZIP, MUSICAM, etc. are well integrated into the world of telecommunications,
the Internet in particular. At the output of the source encoder, the statistical
properties of independence and equiprobability are generally respected and the
compressed message can then be processed by the channel encoder, which will
add redundancy mathematically exploitable by the receiver.
1.2 A first code

Figure 1.1 shows an electronic circuit that performs a very simple, easily decod-
able correction encoding. The code implemented is an extended Hamming code,
which was used in teletext (one of the first digital communication systems),
which will also be presented in Chapter 4.
Figure 1.1 – Extended Hamming encoder: a simple, easily decodable code.
The encoder contains four identical operators performing the exclusive-or

function. The exclusive-or (XOR) of k binary values b0 , b1 , ..., bk−1 is calculated
as:

k−1
XOR(b0 , b1 , ..., bk−1 ) = b0 ⊕ b1 ⊕ ... ⊕ bk−1 = bp modulo 2 (1.2)
p=0
It is therefore quite simply 0 if the number of logical 1s appearing in the input

sequence is even and 1 in the opposite case. In the sequel, when modulo 2
1. Introduction 5
additions are performed and if there is no ambiguity in the notation, the term
modulo 2 may be omitted.
The encoder transforms the message containing four data bits: d =(d0 , d1 ,
d2 , d3 ) into a word of eight bits: c = (d0 , d1 , d2 , d3 , r0 , r1 , r2 , r3 ), called code-
word. The codeword is therefore separable into one part that is the information
coming from the source, called the systematic 1 part and a part added by the
encoder, called the redundant part. Any code producing codewords of this form
is called a systematic code. Most codes, in practice, are systematic but there is
one important exception in the family of convolutional codes (Chapter 5).
The law for the construction of the redundant part by the particular encoder
of Figure 1.1 can be simply written as:

3
rj = dj + dp (j = 0, ..., 3) (1.3)
p=0
Table 1.1 shows the sixteen possible values of c, that is, the set {c} of codewords.
d0 d1 d2 d3 r0 r1 r2 r3 d0 d1 d2 d3 r0 r1 r2 r3
0 0 0 0 0 0 0 0 1 0 0 0 0 1 1 1
0 0 0 1 1 1 1 0 1 0 0 1 1 0 0 1
0 0 1 0 1 1 0 1 1 0 1 0 1 0 1 0
0 0 1 1 0 0 1 1 1 0 1 1 0 1 0 0
0 1 0 0 1 0 1 1 1 1 0 0 1 1 0 0
0 1 0 1 0 1 0 1 1 1 0 1 0 0 1 0
0 1 1 0 0 1 1 0 1 1 1 0 0 0 0 1
0 1 1 1 1 0 0 0 1 1 1 1 1 1 1 1
Table 1.1 – The sixteen codewords that the encoder in Figure 1.1 can produce.
We can first note that the coding law is linear : the sum of two codewords is
also a codeword. It is the linearity of relation (1.3) that guarantees the linearity
of the encoding. All the codes that we shall consider in what follows are linear as
they are all based on two linear operations: addition and permutation (including
shifting). Since the code is linear and the transmission of a codeword might be
affected by a process that is also linear (the addition of a perturbation: noise,
interference, etc.), the choice of a codeword, to explain or justify the properties
of the code, is completely indifferent. It is the "all zero" codeword that will
play this "representative" or reference role for all the codewords vis-à-vis the
general properties of the encoder/decoder pair. At reception, the presence of 1
will therefore be representative of transmission errors.
1 We can also say information part because it is made up of bits of information coming from
the source.
The number of 1s contained in a codeword that is not "all zero" is called the
Hamming weight and is denoted wH . We can distinguish the weight relating
to the systematic part (wH,s ) and the weight relating to the redundant part
(wH,r ). We note, in Table 1.1, that wH is at least equal to 4. Because of
linearity, this also means that the number of bits that differ in two codewords
is also at least equal to 4. The number of bits that are different, when we
compare two binary words, is called the Hamming distance. The smallest of all
the distances between all the codewords, considered two by two, is called the
minimum Hamming distance (MHD) and denoted dmin . Linearity means that
it is also the smallest of the Hamming weights in the list of codewords excluding
the "all zero". dmin is an essential parameter for characterizing a particular
code, since the correction capability of the corresponding decoder is directly
linked to it.
We now write a decoding law for the code of Figure 1.1:
"After receiving the word c’ = (d0 ’, d1 ’, d2 ’, d3 ’, r0 ’, r1 ’, r2 ’, r3 ’) transmitted
by the encoder and possibly altered during transmission, the decoder chooses
the closest codeword ĉ in the sense of the Hamming distance".
The job of the decoder is therefore to run through Table 1.1 and, for each of the
sixteen possible codewords, to count the number of bits that differ from c’. The
ĉ codeword selected is the one that differs least from c’. When several solutions
are possible, the decoder selects one at random. Mathematically, this is written:

3

3

ĉ = c ∈ {c} such that dj ⊕ dj + rj ⊕ rj is minimum (1.4)
j=0 j=0
A decoder capable of implementing such a process is called a maximum likelihood

(ML) decoder as all the cases, here the sixteen possible codewords, are run
through in the search for a solution and no other more efficient decoding process
exists. With codes other than the very simple Hamming code, the message to
be encoded contains far more than four bits (in practice, it goes from a few tens
to a few tens of thousands of bits) and ML decoding is impossible to execute as
the number of codewords is too high.
Assume that the Hamming encoder transmits the "all zero" word that we
have chosen as the reference and that some of the 0s have been inverted be-
fore reception, on the transmission channel. How many errors can the decoder
correct, based on law (1.4)? If c’ contains a single 1 (a single error), the "all
zero" word is closer to c’ than any other codeword that possesses at least four
1s. Correction is thus possible. If c’ contains two 1s, for example in the first
two places (d0 = d1 = 1), the decoder is faced with a dilemma: four codewords
are at a distance of 2 from the received word. It must therefore choose ĉ at
random, from among the four possible solutions, with three risks out of four
of being erroneous. In this situation, we can also ensure that the decoder does
not give a solution but merely indicates its dilemma: it then plays the role of
1. Introduction 7
non-correctable error detector. Finally, if c’ contains three 1s, the decoder will
find a single codeword at a distance of 1 and it will propose this codeword as
the most probable solution, but it will be erroneous.
The error correction capability of the extended Hamming code is therefore
t = 1 errors. More generally, the error correction capability of a code with a
minimum distance dmin is:
dmin − 1
t= (1.5)
2
Note that the correction capability of the code given as the example in this
introduction is not decreased if we remove one (any) of the redundancy symbols.
The MHD passes from 4 to 3 but the correction capability is still of one error.
This shortened version is in fact the original Hamming code, the first correcting
code in the history of information theory (1948).
In a given family of codes, we describe a particular version of it by the
shortcut (n, k, dmin ) where n and k are the lengths of the codewords and
of the source messages, respectively. Up to now, we have thus just defined
two Hamming codes denoted (8, 4, 4) and (7, 4, 3). The second seems more
interesting as it offers the same error correction capability (t = 1) with a τ =
n−k
k
redundancy rate of 0.75, instead of 1 for the first one. However, the code
(7, 4, 3) cannot play the role of an error detector: if the received word contains
two errors, the decoder will decide in favour of the single, erroneous codeword
that is to be found at a Hamming distance of 1.
Rather than redundancy rate, we usually prefer to use the notion of coding
rate, denoted R, and defined by:
k 1
R= = (1.6)
n 1+τ
The product Rdmin will appear in the sequel as an essential figure of merit
vis-à-vis a perturbation caused by additive noise with a Gaussian distribution.
1.3 Hard input decoding and soft input decoding

We continue this presentation of the basic principles of coding and decoding by
using the example of the extended Hamming code.
The decoding principle defined by (1.4) assumes that the received codeword
c’ is composed of binary values, that is, the transmission of data is carried out
according to the law of perturbation given by the diagram of Figure 1.2. A
0 becomes a 1 and vice-versa, with a probability p and the binary values are
correctly received with the complementary probability 1 - p. Such a transmission
channel is said to be a binary symmetric channel and the decoder performs
what we call hard input decoding. In certain communications systems (optical
fibres, switched networks, etc.) and in most storage hardware, the decoders can
effectively exploit only binary information.
Figure 1.2 – Binary symmetric channel with error probability p.
When the signal received by the decoder comes from a device capable of
producing estimations of an analogue nature on the binary data transmitted,
the error correction capability of the decoder can be greatly improved. To show
this using the example of the extended Hamming code, we must first change
alphabet and adopt an antipodal (or symmetric) binary alphabet. We will make
the transmitted values x = -1 and x = +1 correspond to the systematic binary
data d = 0 and d = 1, respectively. Similarly, we will make the transmitted
values y = -1 and y = +1 correspond to the redundant binary data r = 0 and
r = 1, respectively. We then have:
x = 2d − 1 = −(−1)d
(1.7)
y = 2r − 1 = −(−1)r
Figure 1.3 gives an example of a transmission during which the transmitted

values -1 and +1 are altered by additive noise of an analogue type. The values
at the output of the transmission channel are then real variables, which must in
practice be clipped then quantified when the decoder is a digital processor. The
number of quantization bits, denoted Nq , does not need to be high: 3, 4 or 5
bits are sufficient to closely represent analogue samples. One time out of two on
average, the noise is favourable as it has the same sign as the transmitted value.
In the other case, the amplitude of the signal is attenuated and, when this
unfavourable noise is large, the sign can be inverted. An immediate decision
taken per threshold (that is, is it larger or smaller than the analogue zero?)
would then lead to an erroneous binary value being given.
Figure 1.3 – Transmission channel with additive noise of an analogue type.
Since the decoder has information about the degree of reliability of the values
received, called soft or weighted values in what follows, the decoding of the
1. Introduction 9
extended Hamming code according to law (1.4) is no longer optimal. The law of
maximum likelihood decoding to implement in order to exploit these weighted
values depends on the type of noise. An important case in practice is additive
white Gaussian noise (AWGN).
u is a random Gaussian variable with mean μ and variance σ 2 when its
probability density p(u) can be expressed in the form:
1 (u − μ)2
p(u) = √ exp(− ) (1.8)
σ 2π 2σ 2
The AWGN is a perturbation which, after adapted filtering and periodic sam-
pling (see Chapter 2), produces independent samples whose amplitude follows
probability density law (1.8), with zero mean and variance:
N0
σ2 = (1.9)
2
where N0 is the noise power spectral density.
A transmission channel on which the only alteration of the signal comes from
an AWGN is called a Gaussian channel. At the output of such a channel, the
ML decoding is based on the exhaustive search for the codeword that is at the
smallest Euclidean distance from the received word. Denoting X and Y the
received values corresponding to the transmitted symbols x and y respectively,
the soft input decoder of the extended Hamming code therefore chooses:

3
2

3
2
ĉ = c ∈ {c} such that (xj − Xj ) + (yj − Yj ) is minimum (1.10)
j=0 j=0
Since the values transmitted are all such that x2j = 1 or yj2 = 1 and all the
Euclidean distances contain Xj2 and Yj2 , the previous law can be simplified as:

3
3
ĉ = c ∈ {c} such that − 2xj Xj − 2yj Yj is minimum
j=0 j=0
or as:

3
3
ĉ = c ∈ {c} such that xj Xj + yj Yj is maximum (1.11)
j=0 j=0
Minimizing the Euclidean distance between two codewords c and c’ therefore

3
3
means maximizing the scalar product x, X + y, Y = xj X j + yj Yj
j=0 j=0
where x, X, y and Y represent the transmitted and received sequences of the
systematic and redundant parts.
In a Digital Signal Processor (DSP) or in an Application Specific Integrated

Circuit (ASIC), it may be useful to have only to deal with positive numbers.
Law (1.11) can then be implemented as:

3
ĉ = c ∈ {c} such that (Vmax +xj Xj ) + (Vmax +yj Yj ) is maximum (1.12)
j=0
where [−Vmax , Vmax ] is the interval of the values that the input samples Xj and
Yj of the decoder can take after the clipping operation.
In Figure 1.4, the "all zero" codeword has been transmitted and received
with three alterations in the first three positions. These three alterations have
inverted the signs of the symbols but their amplitudes are at a fairly low level:
0.2, 0.4 and 0.1. Hard input decoding produces an erroneous result as the closest
codeword in terms of the Hamming distance is (1, 1, 1, 0, 0, 0, 0, 1). However,
soft input decoding according to (1.11) does produce the "all zero" word, whose
maximum scalar product is:
(−1)(0.2) + (−1)(0.4) + (−1)(0.1) + (−1)(−1)

+ (−1)(−1) + (−1)(−1) + (−1)(−1) + (−1)(−1) = 4.3
in comparison with:
(+1)(0.2) + (+1)(0.4) + (+1)(0.1) + (−1)(−1)

+ (−1)(−1) + (−1)(−1) + (−1)(−1) + (+1)(−1) = 3.7
for the competitor word (1, 1, 1, 0, 0, 0, 0, 1).

This simple example shows the interest of not excluding the reliability infor-
mation in decision taking, whenever possible.
Figure 1.4 – The "all zero" word (emission of symbols with the value -1) has been
altered during the transmission on the first three positions. The hard input decoding
is erroneous, but not the soft input decoding.
The ML decoding rules that we have just used in a specific example are easily
generalizable. However, we do realize that beyond a certain length of message,
such a decoding principle is unrealistic. Applying ML decoding to codewords
1. Introduction 11
containing 240 bits in the systematic part, for example, would mean considering
as many codewords as atoms in the visible universe (1080 ). In spite of this, for
most of the codes known, non-exhaustive decoding methods have been imagined,
enabling us to get very close to the optimal result of the ML method.
1.4 Hard output decoding and soft output decod-

ing
When the output of the decoder is not directly transmitted to a recipient but
must be used by another processor whose performance is improved thanks
to weighted inputs, this upstream decoder can be required to elaborate such
weighted values. We thus distinguish the hard output when the decoder provides
logical 0s and 1s from the soft output. In the latter case, the decoder accompa-
nies its binary decisions with reliability measures or weights. The output scale
of weighted values is generally the same as the input scale [−Vmax , Vmax ].
For the extended Hamming code decoder, it is relatively easy to build
weighted decisions. When the decoder has calculated the sixteen scalar prod-
ucts, it lists them in decreasing order. In the first position, we find the scalar
product of the most likely codeword, that is, the one that decides the signs of
the weighted decisions at the output. Then, for each of the four bits of the sys-
tematic part, the decoder looks for the highest scalar product that corresponds
to a competitor codeword in which the information bit in question is opposite
that of the binary decision. The weight associated with this binary decision is
then the difference between the maximum scalar product and the scalar product
corresponding to the competitor word. A supplementary division by 2 puts the
weighted output on the input scale. This process is optimal for an AWGN per-
turbation. Taking the example of Figure 1.4 (which is not typical of an AWGN)
again, the weight associated with the decisions on the first three bits would be
identical and equal to (4.3 – 3.7)/2 = 0.3.
From a historical point of view, the first decoding methods were of the hard
input and output type. It was the Viterbi algorithm, detailed in Chapter 5,
that popularized the idea of soft input decoding. Then turbo codes, which are
decoded by repeated processing and require weighted values at every level of
this processing, made soft input and output decoders popular. The generic
abbreviation used to qualify these decoders is SISO for Soft-Input/Soft-Output.
1.5 The performance measure

The performance of an encoder/decoder pair is first judged in terms of residual
errors at the output of the decoder, when we have fixed a specific evaluation
framework: type of perturbation, length of message, rate of redundancy or cod-
ing rate, etc. Other aspects, like the complexity of the decoding, the latencies
introduced by the encoder and the decoder, the degree of flexibility of the code
(in particular its ability to conform to different lengths of message and/or to
different coding rates) are also to be considered more or less closely, depending
on the specific constraints of the communication system.
The residual errors that the decoder has not managed to correct are measured
by using two parameters. The binary error rate (BER) is the ratio between the
number of residual binary errors and the total number of bits of information
transmitted. The word, block or packet error rate (PER) is the number of
codewords badly decoded (at least one of the bits of information is wrong) out
of the total number of codewords transmitted. The ratio between BER and PER
is the average density of errors δe in the systematic part of a badly decoded word:
w̄ BER
δe = = (1.13)
k PER
where w̄ = kδe is the average number of erroneous information bits in the
systematic part of a badly decoded block.
Figure 1.5 gives a typical example of the graphic representation of the per-
formance of error correction coding and decoding. The ordinate gives the BER
Eb
on a logarithmic scale and the abscissa carries the signal to noise ratio N0
, ex-
pressed in decibels (dB). N0 is defined by (1.9) and Eb is the energy received
per bit of information. If Es is the energy received for each of the symbols of
the codeword, Es and Eb are linked by:
Es = REb (1.14)
The comparison of different coding and decoding processes or the variation in

performance of a particular process with the coding rate are always defined
with the same global reception energy. When there is no coding, the energy
per received codeword is kEb . With coding, which increases the number of
bits transmitted, the energy kEb is to be distributed between the n bits of
the codeword, which justifies relation (1.14). The reference of energy to be
considered, independent of the code and of the rate, is therefore Eb .
In Figure 1.5 are plotted the curves for the error correction of the (8, 4, 4) and
(7, 4, 3) Hamming codes that have been dealt with throughout this introduction
on a Gaussian channel. Hard-input decoding according to (1.4) and soft-input
decoding according to (1.11) are considered. Also shown in the diagram is the
curve for the binary error probability Pe that is obtained on this channel without
using coding2 . This curve is linked to the complementary error function erfc(x)
1 Eb
given by the relation (2.74) of Chapter 2: Pe = 2 erfc N0 . With a low error
2 The distinction between Pe and BER is only traditional: the value of Pe is given by an
equation whereas the BER is obtained by measuring or simulation.
1. Introduction 13
Figure 1.5 – Error correction capability of (8, 4, 4) and (7, 4, 3) Hamming codes on a
Gaussian channel, with hard input and soft input decoding.
rate, the asymptotic performance of Pe is approximated by:

1 exp − Eb
N0
Pe ≈ (1.15)
2 π Eb
N0
To evaluate the probability Pe,word that the soft-input decoder of a code with
rate R with minimum distance dmin produces an erroneous codeword, in the
Eb Eb
previous equation we replace N 0
by Rdmin N 0
and we introduce a multiplicative
coefficient denoted N (dmin ):

exp −Rd Eb
1 Eb 1 min N0
Pe,word = N (dmin )erfc Rdmin ≈ N (dmin ) (1.16)
2 N0 2 πRd Eb
min N0
The replacement of Eb by REb comes from (1.14) for the energy received by
symbol is Es . The multiplication by dmin is explained by the ML decoding
rule (relation (1.11)), through which the decoder can discriminate the correct
codeword and its closest competitor codewords thanks to dmin distinct values.
Finally, the coefficient N (dmin ), called multiplicity, takes into account the num-
ber of competitor codewords that are the minimum distance away. For example,
in the case of the extended Hamming code, we have N (dmin = 4) = 14 (see Ta-
ble 1.1).
To obtain the error binary probability Pe,bit , it suffices to multiply Pe,word

by the mean error density δe defined by (1.13):

1 exp −Rdmin N Eb
0
Pe,bit ≈ δe N (dmin ) (1.17)
2 πRd Eb
min N0
Reading Table 1.1, we note that average number of errors in the 14 competitor
words of weight 4, at the minimum distance from the "all zero" word, is 2.
Equation (1.19) applied to the extended Hamming code is therefore:

1 2 exp − 1
2 × 4 × Eb
N0 exp −2 Eb
N0
Pe,bit ≈ × × 14 = 3.5
2 4 π ×4×
1 Eb
2π Eb
2 N0 N0
This expression gives Pe,bit = 2.8 10−5, 1.8 10−6 et 6.2 10−8 for N Eb
0
= 7, 8
and 9 dB respectively, which corresponds to the results of the simulation of
Figure 1.5. Such agreement between equations and experimentation cannot be
found so clearly for more complex codes. In particular, finding the competitor
codewords at distance dmin may not be sufficient and we then have to consider
words at distance dmin + 1, dmin +2 etc.
For a same value of Pe and Pe,bit provided
by the relations
(1.15) and (1.17)
Eb Eb
respectively, the signal to noise ratios N0 and N0 without coding (NC)
NC C
and with coding (C) are such that:
⎛ ⎞
Eb

Eb Eb ⎜ N0 N C ⎟
Rdmin − = log ⎝δe N (dmin ) ⎠ (1.18)
N0 C N0 N C Eb
Rd
min N0
C
If δe N (dmin ) is not too far from unity, this relation can be simplified as:

Eb Eb
Rdmin − ≈0
N0 C N0 N C

Eb
The asymptotic gain, expressed in dB, provides the gap between N and
0
NC
Eb
N0 :
C ⎛ ⎞
Eb
N0
Ga = 10 log ⎝ N C ⎠ ≈ 10 log (Rdmin ) (1.19)
Eb
N0
C
As mentioned above, Rdmin appears as a figure of merit which, in a link budget
with a low error rate, fixes the gain that a coding process can provide on a
1. Introduction 15
Gaussian channel when the decoder is soft input. This is a major parameter
for communication system designers. For types of channel other than Gaussian
channels (Rayleigh, Rice, etc.), the asymptotic gain is always higher than what
is approximately given by (1.19).
In Figure 1.5, the soft input decoding of the (8, 4, 4) Hamming code gives
the best result, with an observed asymptotic gain of the order of 2.4 dB, in
accordance with relation (1.18) that is more precise than (1.19). The (7, 4, 3)
code is slightly less efficient since the product Rdmin is 12/7 instead of 2 for
the (8, 4, 4) code. On the other hand, hard input decoding is unfavourable to
the extended code as it does not offer greater correction capability in spite of
a higher redundancy rate. This example is atypical: in the very large majority
of practical cases, the hierarchy of codes that can be established based on their
performance on a Gaussian channel, with soft input decoding, is respected for
other types of channels.
1.6 What is a good code?

Figure 1.6 represents three possible behaviours for an error correcting code and
its associated decoder, on a Gaussian channel. To be concrete, the information
block is assumed to be length k = 1504 bits (188 bytes, a typical length for
MPEG l compression) and the coding rate 1/2.
Curve 1 corresponds to the ideal system. There are in fact limits to the cor-
rection capacity of any code. These limits, whose first values were established
by Shannon (1947-48) and which have been refined since then for practical situ-
ations, are said to be impassable. They depend on the type of noise, on the size
of codeword and on the rate. Their main values are given in Chapter 3.
Curve 2 describes a behaviour having what we call good convergence but with
mediocre MHD. Good convergence means that the error rate greatly decreases
close to the theoretical limit (this region of great decrease is called the waterfall )
but the MHD is not sufficient to maintain a steep slope down to very low error
rates. The asymptotic gain, approximated by (1.19), is reached at a binary error
rate of the order of 10−5 in this example. Beyond that,the curve remains parallel
Eb
to the curve of error rates without coding: Pe = 12 erfc N 0
. The asymptotic gain
is here of the order of 7.5 dB. This kind of behaviour, which was not encountered
before the 1990s, is typical of coding systems implementing an iterative decoding
technique (turbo codes, LDPC, etc.), when the MHD is not very high.
Curve 3 shows performance with mediocre convergence and high asymptotic
gain. A typical example of this is the concatenation of a Reed-Solomon code and
of a convolutional code. Whereas the MHD can be very large (around 100, for
example), the decoder can benefit only relatively far from the theoretical limit.
It is therefore not the quality of the code that is in question but the decoder,
which cannot exploit all the information available at reception.
Figure 1.6 – Possible behaviours for a coding/decoding scheme on a Gaussian channel

(k = 1504 bits, R = 1/2).
The search for the ideal encoder/decoder pair, since Shannon’s work, has
always had to face this dilemma: good convergence versus high MHD. Excellent
algebraic codes like BCH or Reed-Solomon codes were fairly rapidly elaborated
in the history of the correction coding (see Chapter 4). MHDs are high (and even
sometimes optimal) but it is not always easy to implement soft input decoding.
In addition, algebraic codes are generally "sized" for a specific length of codeword
and coding rate, which limits their fields of application. In spite of this, algebraic
codes are of great use in applications that require very low error rates, especially
mass memories and/or when soft information is not available.
It is only recently, with the introduction of iterative probabilistic decoding
(turbo decoding), that we have been able to obtain efficient error correction
close to the theoretical limit. And it is even more recently that we have been
able to obtain sufficient MHDs to avoid a change in slope that is penalizing for
the performance curve.
It is not easy to find a simple answer to the question posed at the beginning
of this section. Performance is, of course, the main criterion: for a given error
rate, counted either in BER or in PER and for a fixed coding rate, a good code is
first the one whose decoder offers a good error correction capability close to the
corresponding theoretical limit. One preliminary condition of this is obviously
the existence of a decoding algorithm (random codes do not have a decoder, for
1. Introduction 17
example) and that the software and/or hardware of this algorithm should not be
too complex. Furthermore, using soft inputs could be an imperative that might
not be simple to satisfy.
Other criteria like decoding speed (that fixes the throughput of the infor-
mation decoded), latency (the delay introduced by the decoding process) or
flexibility (the ability of the code to be defined for various word lengths and
coding rates) are also to be taken into account in the context of the application
targeted.
Finally, non-technical factors may also be very important. Technological
maturity (do applications and standards already exist?), the cost of components,
possible intellectual property rights, strategic preferences or force of habit are
elements that carry weight when choosing a coding solution.
1.7 Families of codes

Up until the last few years, codes were traditionally classified in two families that
were considered to be quite distinct due to their principles and their applications:
algebraic codes (also called block codes) and convolutional codes. This distinction
was mainly based on three observations:
• algebraic codes are appropriate for protecting blocks of data independent

one from the other whereas convolutional codes are suitable for protecting
continuous flows of data,
• the coding rates of algebraic codes are rather close to unity, whilst convo-
lutional codes have lower rates,
• block code decoding is rather of the hard input decoding type, and that
of convolutional codes is almost always soft input decoding.
Today, these distinctions are tending to blur. Convolutional codes can easily
be adapted to encode blocks and most decoders of algebraic codes accept soft
inputs. Via concatenation (Chapter 6), algebraic code rates can be lowered to
values comparable with those of convolutional codes. One difference remains,
however, between the two sub-families: the number of possible logical states of
algebraic encoders is generally very high, which prevents decoding by exhaustive
state methods. Decoding algorithms are based on techniques specific to each
code. Convolutional encoders have a limited number of states, 2 to 256 in
practice, and their decoding uses a complete representation of states, called
a trellis (Chapter 5). It is for this reason that the book is structured in a
traditional manner, which, for the time being, makes the distinction between
algebraic codes and convolutional codes.
Modern coding requires concatenated or composite structures, which use sev-
eral elementary encoders and whose decoding is performed by repeated passages
in the associated decoders and by an exchange of information of a probabilistic

type. Turbo codes (1993) opened the way to this type of iterative processing
and, since then, many composite structures based on iterative decoding have
been imagined (or rediscovered). Three of these are detailed in this book: turbo
codes, both their original version and their recent evolution, iteratively decoded
product codes and LDPC (Low Density Parity Check ) codes. All these codes
have been adopted in international norms, and understanding their coding pro-
cesses and their decoding algorithms is a basis wide enough to tackle any other
principle of distributed coding and of associated iterative decoding.
Chapter 2
Digital communications
2.1 Digital Modulations

2.1.1 Introduction
The function of modulation has its origin in radio-electric communications. An
emitter can radiate an electromagnetic wave only in a limited, and generally
narrow, portion of the spectrum, which can roughly be described as a frequency
"window" with Δf centred on a frequency f0 , with Δf << f0 . The messages to
be transmitted, that can be either analogue (for example speech) or digital (for
example Morse code), are represented by signals that occupy only the bottom of
the frequency spectrum. The spectrum of the signal, coming from a microphone
in the case of speech, does not extend beyond a few kilohertz. The same thing
applies for a signal that represents the two short (Tit) or long (Tat) elements of
Morse "code", since the speed of handling several dozen signs per second is very
small compared to the frequency f0 that is measured in hundreds of kilohertz
or in megahertz. Another use for modulation is frequency multiplexing which
enables several simultaneous communications on the same wideband (cable or
optical fibre) support, that are easily separated due to the fact that they each
occupy a specific bandwidth, not connected to that of any other.
A sinusoidal wave f0 can be represented by the function
s(t) = a cos(2πf0 t + ϕ) (2.1)
where t denotes the time and f0 is constant. Modulation involves making one or
other of parameters a the amplitude, and ϕ the phase, depend on the signal to
be transmitted. The modulated signal s(t) then has a narrow spectrum centred
on f0 , which is what we want.
The signal to be transmitted will in the sequel be called the modulating
signal. Modulation makes one of the parameters a and ϕ vary as a function of
the modulating signal if the latter is analogue. In the case of a digital signal, the
modulating signal is a series of elements of a finite set, or symbols, applied to the
modulator at discrete instants that are called significant instants. This series
is called the digital message and we assume that the symbols are binary data
applied periodically at the input of the modulator, every Tb seconds, therefore
with a binary rate D = 1/Tb bits per second. The binary data of this series are
assumed to be independent and identically distributed (iid). A given digital, for
example binary, message can be replaced by its "mth extension" obtained by
grouping the initial symbols into packets of m. Then the symbols are numbers
with m binary digits (or m-tuples), the total number of which is M = 2m ,
applied to the modulator at significant instants with period mTb . If the mth
extension is completely equivalent to the message (of which it is only a different
description), then the signal modulated by the original message and the signal
modulated by its mth extension do not have the same properties, in particular
concerning their bandwidth, since the larger m is, the narrower the bandwidth.
The choice of integer m thus allows the characteristics of the modulated message
to be varied.
Consider the complex signal
σ(t) = a exp[j(2πf0 t + ϕ)] (2.2)
whose s(t) is the real part, where j is the solution to the equation x2 + 1 = 0.
We can represent σ(t) as the product
σ(t) = α exp(2πjf0 t), (2.3)
where only the first factor

α = a exp(jϕ) (2.4)
depends on the parameters a and ϕ that represent the data to be transmitted.
The values taken by this first factor for all possible values of the parameters
can be represented by points in the complex plane. The set of these points is
then called a constellation and the complex plane is called a Fresnel plane. The
modulated signal (2.1) is the real part of the complex signal σ(t) defined by
(2.3).
If the correspondence established between the modulating signal and the
variable parameters is instantaneous, the modulation is said to be memoryless.
It can be useful for this correspondence to be established between the variable
parameters and a function of the values taken later by the modulating signal. For
example, if the latter is analogue, a conventional modulation process (called fre-
quency) involves making ϕ vary proportionally to the integral of the modulating
signal in relation to time. In the same way, in the case of a digital modulation,
the constellation point can be chosen to represent the symbol present at the in-
stant considered, and the modulation is then said to be memoryless, or a symbol
obtained by combining it with other later symbols. A modulation can therefore
2. Digital communications 21
be either analogue or digital, with memory or memoryless. In all that follows

below, we restrict ourselves to digital modulations that differ from each other
according to their form or number of points in their constellation, and perhaps
by a memory effect. When the latter is obtained by combining the symbol ap-
plied with later symbols, this can be interpreted as a preliminary transformation
of the digital message. Moreover, it is often necessary to ensure the continuity
of the phase to improve the shape of the spectrum of the modulated signals,
which implies a memory effect.
The choice of a modulation system depends on many factors. Modulated
signals are emitted on an imperfect channel, perturbed by the addition of par-
asitic signals collectively called noise and often, in radio-electricity, affected by
variations in the amplitude of the received signal, due for example to a rapid
change in the propagation conditions, a phenomenon called fading. In spite of
these channel faults, we want to receive modulating messages with a small er-
ror probability, which implies that the signals associated with them should be
as different as possible. On the other hand, whether it be radio-electricity or
multiplexing, the radio-electric spectrum is common to several users, each of
which perturbs the others. Therefore we wish to concentrate the power emitted
in a frequency interval that is limited as far as possible. This implies choosing
the modulation parameters in order to give the spectrum the most appropriate
shape. The signal spectrum of the general (2.3) shape is made up of a main
lobe centred on f0 that concentrates most of the power emitted, and tails or
sidelobes where the spectral density decreases more or less rapidly in relation
to the central frequency f0 . Whatever the modulation system, the width of the
main lobe is proportional to the symbol rate R = D/m = 1/mTb , expressed in
symb/s. The decrease in the spectral density far enough away from the cen-
tral frequency depends only on the discontinuities of the modulating signal and
its derivations. It varies in 1/(f − f0 )2(d+1) where d is the smallest order of a
derivation of the discontinuous signal (d = 0 if the modulated signal itself is
discontinuous). We note that it is only possible to increase the value of d by
introducing an increasing delay between the instant when a modulating symbol
is applied to the modulator and the characteristic instant corresponding to it.
The main parameters associated with modulation are therefore the size and
the shape of the constellation used (on which the error probability depends, on
a given channel), the width of the main lobe of the spectrum of the modulated
spectrum and the decrease in its spectral density away from the central fre-
quency. They are largely in conflict: for example, we can only reduce the width
of the central lobe by increasing the size of the constellation, to the detriment
of the error probability for the same power. The choice of a modulation system
can therefore only result from a compromise adapted to a particular application.
Apart from the parameters indicated, the complexity of implementation must
be taken into account. For example, shaping that improves the decrease in the
secondary lobes of the spectrum by the increase of order d in the first discontin-
uous derivation, or increasing the size of the constellation to lower the central
lobe of the spectrum, mean an increase in the complexity of the modulator.
2.1.2 Linear Memoryless Modulations

Amplitude-shift keying with M states: M-ASK
Let us first note a particular case of amplitude modulation: On Off Keying
(OOK), for which a in the expression (2.1) takes the value 0 or A. One of the
binary states therefore corresponds to an extinction of the carrier. The presence
or absence of the carrier can be recognized independently of the knowledge of the
phase ϕ, by measuring the energy of the received signal during a short interval
of time on the scale of period 1/f0 of the carrier wave, which is called incoherent
detection. Figure 2.1 shows the time interval of a signal modulated by On Off
Keying.
Figure 2.1 – On Off Keying (OOK).
In the general case of amplitude-shift keying (ASK) with M states, the am-
plitude of the carrier is the modulated value aj = Aj h(t) for j = 1, 2, ..., M
where Aj takes a value among M = 2m values according to the group of data
presented at the input of the modulator and h(t) is a rectangular pulse with
unit amplitude and width T . The modulator thus provides signals of the form:
sj (t) = Aj h(t) cos(2πf0 t + ϕ0 ) (2.5)
with Aj = (2j − 1 − M )A j = 1, 2, · · · , M where A is constant (2.6)
and h(t) = 1 for t ∈ [0, T [

(2.7)
= 0 elsewhere
Amplitude Aj of the modulated signal is constant for a width T then changes
value; the modulated signal thus transmits log2 (M ) binary data every T seconds.
We can note that half the nominal amplitudes Aj are negative. The identity
a cos ϕ = −a cos(ϕ + π) thus implies coherent demodulation where the phase
ϕ is known. The phase often only being known to a multiple of π, a binary
message where the symbols are ±1 will only be demodulated to one sign.
For an M-ASK signal, the different states of the modulated signal are situ-
ated on a straight line and its constellation is therefore one-dimensional in the
Fresnel plane. There are many ways to make the association between the value of
the amplitude of the modulated signal and the particular realization of a group
of data of m = log2 (M ) data. In general, we associate with two adjacent values
taken by the amplitude, two groups of data that differ by only one binary value.
This particular association is called Gray coding. It enables the errors made by
the receiver to be minimized. Indeed, when the receiver selects an amplitude
adjacent to the emitted amplitude because of noise, which corresponds to the
most frequent situation, we make only one error for m = log2 (M ) data trans-
mitted. We show in Figure 2.2 two examples of signal constellations modulated
in amplitude by Gray coding.
Figure 2.2 – Example of 4-ASK and 8-ASK signal constellations with Gray coding
The mean energy Es used to transmit an M -ary symbol is equal to:

T

Es = E A2j cos2 (2πf0 t + ϕ0 )dt
0

where E A2j , the expectation of A2j has the expression A2 (M 2 − 1)/3.
Making the hypothesis that f0 >> T1 , the previous relation gives the mean
energy Es :
A2 T (M 2 − 1)
Es = (2.8)
2 3
The mean energy Eb used to transmit a bit is:
Es
Eb = (2.9)
log2 (M )
For a transmission with a continuous data stream, the amplitude modulated
signal can be written in the form:

S(t) = A ai h(t − iT ) cos(2πf0 t + ϕ0 ) (2.10)
i
where the {ai } are a sequence of M -ary symbols, called modulation symbols,
which take the values (2j − 1 − M ), j = 1, 2, · · · , M . In the expression of the
modulated signal, i is the time index.
The signal S(t) can again be written in the form:
S(t) =
e {se (t) exp (j(2πf0 t + ϕ0 ))} (2.11)
where se (t) is the complex envelope of the signal S(t) with:

se (t) = A ai h(t − iT ) (2.12)
i
Taking into account the fact that the data di provided by the source of infor-
mation are iid, the modulation symbols ai are independent, with zero mean and
variance equal to (M 2 − 1)/3.
It can be shown that the power spectral density (psd) of the signal S(t) is
equal to:
1 1
γS (f ) = γse (f − f0 ) + γse (f + f0 ) (2.13)
4 4
with: 2
M2 − 1 2 sin πf T
γse (f ) = A T (2.14)
3 πf T
The psd of se (t) expressed in dB is shown in Figure 2.3 as a function of the
normalized frequency f T , for M = 4 and A2 T = 1.
Figure 2.3 – Power spectral density (psd) of the complex envelope of a signal ASK-4,
with A2 T = 1.
The psd of S(t) is centred on the frequency carrier f0 and its envelope
decreases in f 2 . It is made up of a main lobe of width 2/T and of sidelobes that
periodically have zero crossing at f0 ± k/T .
Note
The bandwidth is, strictly speaking, infinite. In practice, we can decide only
to transmit a percentage of the power of the signal S(t) and in this case, the
bandwidth is finite. If, for example, we decide to transmit 99% of the power of
the modulated signal, which results in only a low distortion of the signal S(t),
then the bandwidth is about 8/T where 1/T is the symbol rate. We shall see in
Section 2.3 that it is possible to greatly reduce this band without degrading the
performance of the modulation. This remark is valid for all linear modulations.
Phase Shift Keying with M states (M-PSK)

For this modulation, also called Phase Shift Keying (PSK), it is the phase of
the carrier that is the modulated value. The modulator provides signals of the
form:
sj (t) = Ah(t) cos(2πf0 t + ϕ0 + φj ) (2.15)
where f0 is the carrier frequency, ϕ0 its phase and h(t) a rectangular pulse of
unit amplitude and width T . The modulated phase φj takes a value among
M = 2m with:
π
φj = (2j + 1) + θ0 0 ≤ j ≤ (M − 1) (2.16)
M
The different states of the phase are equidistributed on a circle of radius A.
The phase θ0 is fixed at −π/2 for a 2-PSK modulation and at 0 for an M-PSK
modulation with M > 2 states.
The modulated signal can again be written in the form:
sj (t) = Ah(t) [cos φj cos(2πf0 t + ϕ0 ) − sin φj sin(2πf0 t + ϕ0 )] (2.17)
In this form, the M-PSK signal can be expressed as the sum of two quadrature
carriers, cos(2πf0 t + ϕ0 ) and − sin(2πf0 t + ϕ0 ), the amplitude modulated by
cos φj and sin φj with cos2 φj + sin2 φj = 1. We can check that when M is
a multiple of 4, the possible values of the amplitude of the two carriers are
identical.
In Figure 2.4 we show two constellations of a phase modulated signal with
Gray coding. The constellations have two dimensions and the different states of
the modulated signal are on a circle of radius A. We say that the constellation
is circular.
Figure 2.4 – Examples of constellations of a phase modulated signal with Gray coding.
The energy Es for transmitting a phase state, that is, a group of log2 (M ) binary
data, is equal to:
T
A2 T
Es = A2 cos2 (2πf0 t + ϕ0 + φj )dt = if f0 >> 1/T (2.18)
2
0
Energy Es is always the same whatever the phase state transmitted. The energy
used to transmit a bit is Eb = Es / log2 (M ).
For the transmission of a continuous data stream, the modulated signal can
be written in the form:

S(t) = A ai h(t − iT ) cos(2πf0 t + ϕ0 )
i (2.19)

− bi h(t − iT ) sin(2πf0 t + ϕ0 )
i
where the modulation symbols ai and bi take their values in the following sets:

ai ∈ cos (2j + 1) M
π
+ θ0 0 ≤ j ≤ (M − 1)
(2.20)
bi ∈ sin (2j + 1) M π
+ θ0 0 ≤ j ≤ (M − 1)
The signal S(t) can again be written in the form given by (2.11) with:

se (t) = A ci h(t − iT ), ci = ai + jbi (2.21)
i
Taking into account the fact that the data di provided by the source of infor-
mation are iid, the modulation symbols ci are independent, with zero mean and
unit variance.
The psd of the signal S(t) is again equal to:
1 1
γS (f ) = γs (f − f0 ) + γse (f + f0 )
4 e 4
with this time: 2
2 sin πf T
γse (f ) = A T (2.22)
πf T
the psd looking like that of Figure 2.3.
Quadrature Amplitude Modulation using two quadrature carriers (M-

QAM)
For this modulation, also called Quadrature Amplitude Modulation (M-QAM),
it is two quadrature carriers cos(2πf0 t + ϕ0 ) and − sin(2πf0 t + ϕ0 ) that are
amplitude modulated. The modulator provides signals of the form:
sj (t) = Acj h(t) cos(2πf0 t + ϕ0 ) − Asj h(t) sin(2πf0 t + ϕ0 ) (2.23)

where f0 is the frequency of the carrier, ϕ0 its phase and h(t) a rectangular
pulse of unit amplitude and width T .
Two situations can arise depending on whether the length m of the groups of
data at the input of the modulator is even or not. If m is even, then M = 2m is
a perfect square (4, 16, 64, 256, . . .); in the opposite case, M is simply a power
of two (8, 32, 128, . . .).
When m is even, the group of data can be separated into two sub-groups of
length m/2, each being associated respectively with amplitudes Acj and Asj that
√ √
take their values in the set (2j − 1 − M )A, j = 1, 2, · · · , M . In Figure 2.5
are represented the constellations of the 16-QAM and 64-QAM modulations.
These constellations are said to be square.
Figure 2.5 – Constellations of two QAM-type modulations.
When m is odd, the M-QAM signal can no longer be obtained as a combina-

tion of two quadrature amplitude-modulated carriers. However, we can build the
M-QAM signal from an N-QAM signal modulated classically on two quadrature
carriers, where N is a square immediately higher than M by preventing (N −M )
states. For example, 32-QAM modulation can be obtained from 36-QAM modu-
lation where Acj and Asj take the values (±A, ±3A, ±5A) by preventing the four
states of amplitude (±5A, ±5A) for the pairs (Acj and Asj ). The constellation of
the 32-QAM modulation is shown in Figure 2.6.
The M-QAM signal can again be written in the form:
sj (t) = Vj h(t) cos(2πf0 t + ϕ0 + φj ) (2.24)
with: Asj
Vj = (Acj )2 + (Asj )2 φj = tan−1
Acj
In this form, the M-QAM modulation can be considered as a modulation
combining phase and amplitude. Assuming that the phase takes M1 = 2m1
states and the amplitude M2 = 2m2 states, the modulated signal transmits
log2 (M1 M2 ) = m1 +m2 data every T seconds. Figure 2.7 shows the constellation
Figure 2.6 – Constellation of the 32-QAM modulation.
of a modulation combining phase and amplitude for M = 16 (M1 = 4, M2 = 4).
Figure 2.7 – Constellation of a modulation combining phase and amplitude for M = 16.
The average energy Es to transmit the pair (Acj , Asj ), that is, a group of
log2 (M ) binary data, is equal to:
T

Es = E Vj2 cos2 (2πf0 t + ϕ0 + φj )dt (2.25)
0

For a group of data of even length m, E Vj2 = 2A2 (M − 1)/3 and thus, for
f0 >> 1/T , the average energy Es is equal to:
M −1
Es = A2 T (2.26)
3
The average energy used to transmit a bit is Eb = Es / log2 (M ).
For a continuous data stream, the signal can be written in the form:

S(t) = A ai h(t−iT ) cos(2πf0 t+ϕ0 )
i (2.27)

− bi h(t−iT ) sin(2πf0 t+ϕ0 )
i
√
where the modulation
√ symbols ai and bi take the values (2j − 1 − M ), for
j = 1, 2, · · · , M and for M = 2m with even m. The signal S(t) can be
expressed by the relations (2.11) and (2.21):
S(t) =
e {se (t) exp j(2πf0 t + ϕ0 )}

with se (t) = A ci h(t − iT ), ci = ai + jbi
i
The binary data di provided by the information source being iid, the modulation
symbols ci are independent, with zero mean and variance equal to 2(M − 1)/3.
The psd of the signal S(t) is again given by (2.13) with:
2
2(M − 1) 2 sin πf T
γse (f ) = A T (2.28)
3 πf T
The spectral width of a modulated M-QAM signal is therefore, to within an
amplitude, the same as that of M-ASK and M-PSK signals.
2.1.3 Memoryless modulation with M states (M-FSK)

For this modulation, also called Frequency Shift Keying (M-FSK), it is the fre-
quency that is the modulated value. The modulator generates signals of the
form:
sj (t) = Ah(t) cos(2π(f0 + fj )t + ϕj ) (2.29)
where fj = jΔf , j = 1, 2, · · · , M and h(t) is a rectangular pulse with unit
amplitude and width T . The ϕj are random independent phases with constant
realization on the interval [0, T [. The sj (t) signals can therefore be generated
by independent oscillators since there is no relation between the phases ϕj .
Let us compute the correlation coefficient between two modulated signals taking
different frequency states.
T
ρj,n = A2 cos(2π(f0 + jΔf )t + ϕj ) cos(2π(f0 + nΔf )t + ϕn )dt (2.30)
0
After integration, and assuming f0 >> 1/T , we obtain:

A2 T sin(2π(j − n)Δf T + ϕj − ϕn ) sin(ϕj − ϕn )
ρj,n = − (2.31)
2 2π(j − n)Δf T 2π(j − n)Δf T
Choosing Δf = 1/T , ρj,n = 0 ∀j = n and the modulated M signals are orthogo-

nal. Orthogonal signals are generally chosen since signals of different frequencies
can easily be separated at reception. At instants iT where the M-FSK signal
changes frequency, the modulated signal presents a discontinuity since the phases
ϕj are independent. We then speak of discontinuous-phase frequency modula-
tion. Figure 2.8 gives an example of a 2-FSK signal.
Figure 2.8 – Frequency modulated signal with discontinuous phase.
The energy Es used to transmit a group of data is equal to:

T
j A2 T
Es = A2 cos2 (2π(f0 + )t + ϕj )dt = if f0 >> 1/T (2.32)
T 2
0
and the energy Eb used to transmit a bit is equal to Es / log2 (M ).

For a continuous flow transmission of data, the modulated signal can be
written in the form:
ai
S(t) = A h(t − iT ) cos(2π(f0 + )t + ϕi ) (2.33)
T
i
where the modulation symbol ai is equal to 1, 2, · · · , M, M = 2m . Every T

seconds, the modulated signal S(t) transmits a group of log2 (M ) binary data.
For 2-FSK modulation, the power spectral density of the signal S(f ) is equal
to:
1
γS (f ) = (γ(f − f1 ) + γ(f + f1 ) + γ(f − f2 ) + γ(f + f2 )) (2.34)
4
where f1 = f0 + 1/T and f2 = f0 + 2/T and:
2
A2 T sin πf T A2
γ(f ) = + δ(f ) (2.35)
4 πf T 4
where δ(f ) is the Dirac distribution. The psd of a 2-FSK signal has a continuous
part and a discrete part. Limiting ourselves to the two main lobes of this power
spectral density, the band of frequencies occupied by a 2-FSK signal is 3/T , that
is, three times the symbol rate. Let us recall that at a same symbol rate, an
M-PSK or an M-QAM signal occupies a bandwidth of only 2/T . The discrete
part corresponds to two spectral lines situated in f1 and f2 .
2.1.4 Modulations with memory by continuous phase fre-

quency shift keying (CPFSK)
For Continuous Phase Frequency Shift Keying (CPFSK), the modulated signal
does not show any discontinuities at the instants of frequency changes. Figure 2.9
shows a time interval of a CPFSK signal for M = 2. The CPFSK signal has the
expression:
S(t) = A cos(2πf0 t + φ0 + φ(t)) (2.36)
where f0 is the frequency of the carrier and φ0 its phase.
Figure 2.9 – Time interval of a 2-CPFSK signal.
The instantaneous frequency deviation is:

1 dφ
f (t) = =h ai g(t − iT ) (2.37)
2π dt
i
where h is called the modulation index and the M -ary symbols ai take their
values in the alphabet {±1, ±3, · · · , ±(2p + 1), · · · , ± ( M − 1) }; M = 2n .
The function g(t) is causal and has a finite pulse widtht.

g(t) = 0 t ∈ [0, LT [ L integer
(2.38)
= 0 elsewhere
Putting:
t
q(t) = g(τ )dτ
0
and for considerations of normalization, by imposing that:

1
q(t) = if t ≥ LT
2
the phase φ(t) on the interval [iT, (i + 1)T [ is equal to:

i
i−L
φ(t) = 2πh an q(t − nT ) + πh an (2.39)
n=i−L+1 n=−∞
When L = 1, the continuous phase frequency modulations are said to be full

response whereas for L > 1, they are said to be partial response.
To illustrate continuous phase-frequency-shift keying, we are going to con-
sider three examples, Minimum Shift Keying (MSK) modulation, L-ary Raised
Cosine (LRC) modulation and Gaussian Minimum Shift Keying (GMSK) mod-
ulation.
Continuous phase-frequency-shift keying with modulation

index h = 1/2: Minimum Shift Keying (MSK)
For this modulation, the index h is equal to 1/2 and the symbols ai are binary
(±1). The function g(t) is a rectangular pulse of amplitude 1/2T and width T .
Thus the function q(t) is equal to:
q(t) = 0 t≤0
q(t) = t
2T 0≤t≤T (2.40)
q(t) = 1
2 t≥T
MSK modulation is full response continuous phase frequency shift keying mod-
ulation (L = 1).
On the interval [iT, (i + 1)T [, the phase φ(t) of the MSK signal has the
expression:
π (t − iT ) π
i−1
φ(t) = ai + an (2.41)
2 T 2 n=−∞
The evolution of the phase φ(t) as a function of time is shown in Figure 2.10.
We can note that the phase φ(t) varies linearly over a time interval T and that
there is no discontinuity at instants iT .
0.5
1
Phase (rad)
1.5
2
2.5
3
3.5
0 T 2T 3T 4T 5T 6T 7T 8T 9T 10T
Figure 2.10 – Evolution of the phase of an MSK signal as a function of time.
Using the expressions (2.36) and (2.41), the MSK signal can be written in the
form:
ai π
S(t) = A cos 2π(f0 + )t − i ai + θi + ϕ0 iT ≤ t < (i + 1)T (2.42)
4T 2
with:
π
i−1
θi = ai (2.43)
2 n=−∞
The MSK signal uses two frequencies to transmit the binary symbols ai = ±1.
1
f1 = f0 + 4T
if ai = +1
(2.44)
f2 = f0 − 1
4T
if ai = −1
We can verify that the two frequency signals f1 and f2 are orthogonal and
that they present a frequency deviation Δf = f1 − f2 = 1/2T minimal. This
minimum deviation is at the origin of the name of MSK modulation.
The modulated MSK signal can also be written in the form:

S(t) = A c2i−1 h(t − 2iT ) cos 2T
πt
cos(2πf0 t + ϕ0 )
i (2.45)

− c2i h(t − (2i + 1)T ) sin 2T sin(2πf0 t + ϕ0 )
πt
i
where the symbols ci are deduced from the symbols ai by transition coding.
c2i = a2i c2i−1 and c2i−1 = a2i−1 c2i−2 (2.46)

and h(t) is a rectangular pulse of unit amplitude, but of width 2T .
h(t) = 1 if t ∈ [−T, T [
(2.47)
= 0 elsewhere
MSK modulation can be seen as an amplitude modulation of the terms
πt πt
cos cos(2πf0 t + ϕ0 ) and − sin sin(2πf0 t + ϕ0 )
2T 2T

by two bit streams uc (t) = c2i−1 h(t − 2iT ) and us (t) = c2i h(t − (2t + 1)T )
i i
whose transitions are shifted by T . Each bit stream enables a bit to be trans-
mitted every 2T seconds and thus, the binary rate of an MSK modulation is
D = 1/Tb with T = Tb .
MSK modulation is a particular case of CP M-FSK modulation since it is
linear. The psd is given by:
1 1
γS (f ) = γ(f − f0 ) + γ(−f − f0 ) (2.48)
4 4
with: 2
16A2 T cos 2πf T
γ(f ) = (2.49)
π2 1 − 16f 2 T 2
Figure 2.11 shows the psd of the complex envelope of an MSK signal expressed
in dB as a function of the normalized frequency f Tb . We have also plotted the
psd of the complex envelope of a 4-PSK signal. In order for the comparison of
these two power spectral densities to make sense, we have assumed that the rate
transmitted was identical for these two modulations (that is, T = 2Tb for the
4-PSK modulation).
The width of the main lobe of the power spectral density of an MSK mod-
ulation is 3/2Tb whereas it is only 1/Tb for 4-PSK modulation. Thus, for a
same rate transmitted, the main lobe of MSK modulation occupies 50% more
bandwidth than that of 4-PSK modulation. However, the envelope of the psd
of an MSK signal decreases in f −4 whereas it only decreases in f −2 for 4-PSK
modulation. One of the consequences of this is that the bandwidth B that con-
tains 99% of the power of the modulated signal for MSK is 1.2/Tb whereas it is
around 8/Tb for 4-PSK.
L-ary Raised Cosine modulation (L-RC)

For this modulation, the function g(t) has the expression:
g(t) = 1
2LT
(1 − cos 2πt
LT
) for 0 ≤ t ≤ LT
(2.50)
= 0 elsewhere
The larger parameter L is, the faster the power spectral density of this modu-
lation decreases. For example, the psd is -40 dB for a 2RC (h = 1/2, ai = ±1)
Figure 2.11 – Power spectral density of the complex envelope of MSK and 4-PSK
signals.
modulation for f T = 1.2 whereas for a 4RC (h = 1/2, ai = ±1) modulation,

this level is reached for f T = 0.7.
The function q(t) is equal to:
q(t) = t
2LT
− 1
4π
sin 2πt
LT
0 ≤ t ≤ LT
(2.51)
1
= 2
t > LT
Gaussian minimum shift keying modulation (GMSK)

The index h of this modulation is equal to 1/2 and the symbols ai are binary
(±1). The function g(t) is defined as follows:
g(t) = h(t) ∗ χ(t) (2.52)
where h(t) is a rectangular pulse of amplitude 1/T (T = Tb ) on the interval
[−T /2, T /2] and χ(t) is the impulse response of a Gaussian filter with passband
at -3 dB equal to Bg :

2π
χ(t) = Bg exp(−2π 2 Bg2 t2 )/ ln 2 (2.53)
ln 2
After resolution of the convolution product, the function g(t) can be written in
the form:
! " # " #$
1 2 T 2 T
g(t) = erf πBg (t + ) − erf πBg (t − ) (2.54)
2T ln 2 2 ln 2 2
where erf(x) represents the error function defined by:

x
2
erf(x) = √ exp(−u2 )du
π
0
In Figure 2.12, we have plotted g(t) as a function of the normalized variable

t/T , for different values of the normalized passband BN = Bg T . We note that
the graph of the function g(t) has been shifted by 2T for BN = 0.2 and by 1.5T
for BN = 0.3.
Figure 2.12 – Variation of the function g(t) for two values of BN .
The term BN allows the time spreading of function g(t) to be fixed. Thus for
BN = 0.2, this function is approximately of width 4T whereas its width is only
3T for BN = 0.3. When BN tends towards infinity, it becomes a recatngu-
lar pulse with width T (the case of MSK modulation). GMSK modulation is
therefore a partial response continuous phase modulation (L > 1).
On the interval [iT, (i + 1)T [, the phase φ(t) of the GMSK signal is equal to:

i
π
i−L
φ(t) = π an q(t − nT )+ an (2.55)
2 n=−∞
n=i−L+1
where L = 3 if BN = 0.3 and L = 2 if BN = 0.2.

Thus on an interval [iT, (i + 1)T [, the phase φ(t) of the GMSK signal depends on
symbol ai but also on the symbols prior to symbol ai (ai−1 , ai−2 , · · · , ai−L+1 ).
This non-linear modulation presents an memory effect that gives it good spectral
properties. GMSK modulation, with a normalized passband BN = 0.3 was
chosen for the GSM (Groupe Spécial Mobile and later Global System for Mobile
communications) system. We note that there is no simple expression of the
power spectral density of a GMSK signal. For values of the normalized passband
of 0.3 or of 0.2, the power spectral density of the GMSK signal does not show
sidelobes and its decrease as a function of frequency is very rapid. Thus at -10
dB the band occupied by the GMSK signal is approximately 200 kHz, and at
-40 dB 400 kHz for a rate D = 271 kbit/s.
2.2 Structure and performance of the optimal re-

ceiver on a Gaussian channel
The object of this chapter is to determine the structure and the performance
of the optimal receiver for memory and memoryless modulations on an additive
white Gaussian noise (AWGN) channel. The type of receiver considered is the
coherent receiver where the frequency and the phase of the signals transmitted
by the modulator are assumed to be known by the receiver. Indeed, a coherent
receiver is capable of locally generating signals having the same frequency and
the same phase as those provided by the modulator, unlike the so-called non-
coherent or differential receiver.
Generally, the receiver is made up of first a demodulator, the aim of which is
to convert the modulated signal into a baseband signal, and second, a decision
circuit in charge of estimating the blocks of data transmitted. The receiver
is optimal in the sense that it guarantees a minimal error probability on the
estimated blocks of data.
2.2.1 Structure of the coherent receiver

Let sj (t), j = 1, 2, · · · , M be the signals transmitted on the transmission chan-
nel perturbed by a AWGN b(t), with zero mean and of power spectral density
equal to N0 /2. On the time interval [0, T [, the signal received by the receiver is
equal to:
r(t) = sj (t) + b(t)
The M sj (t) signals define a space of dimension N ≤ M , and can be represented
in the form of a series of normed and orthogonal weighted functions νp (t).

N
sj (t) = sjp νp (t)
p=1
where sjp is a scalar equal to the projection of the signal sj (t) on the function
νp (t).
T
sjp = sj (t)νp (t)dt
0
The noise can also be represented in the form of a series of normed and orthog-
onal functions but of infinite length (Karhunen Loeve expansion). When the
noise is white, we show that the normed and orthogonal functions can be chosen
arbitrarily. We are therefore going to take the same orthonormed functions as
those used to represent the sj (t) signals, but after extension to infinity of this
base of functions:
∞

N
b(t) = bp νp (t) = bp νp (t) + b (t)
p=1 p=1
where bp is a scalar equal to the projection of b(t) on the function νp (t).
T
bp = b(t)νp (t)dt
0
The quantities bp are random non-correlated Gaussian variables, with zero mean
and variance σ2 = N0 /2.
TT
E {bp bn } = E {b(t)b(t )}νp (t)νn (t )dtdt
00
The noise being white, E {b(t)b(t )} = N0

2
δ(t − t ) and thus:
T
N0 N0
E {bp bn } = νp (t)νn (t)dt = δn,p (2.56)
2 2
0
where δn,p is the Kronecker symbol, equal to 1 if n = p and to 0 if n = p.
Using the representations of the sj (t) signals and of the b(t) noise by their
respective series, we can write:

N ∞

N
r(t) = (sjp + bp )νp (t) + bp νp (t) = rp νp (t)+b (t)
p=1 p=N +1 p=1
Conditionally to the emission of the signal sj (t), the quantities rp are random
Gaussian variables, with mean and variance N0 /2. They are non-correlated to
the noise b (t). Indeed, we have:
% ∞
&

E {rp b (t)} = E (sjp + bp ) bn νn (t) ∀ p = 1, 2, · · · , N
n=N +1
Taking into account the fact that the variables bn , whatever n is, are zero mean
and non-correlated, we obtain:
∞

E {rp b (t)} = E {bp bn } νn (t) = 0 ∀ p = 1, 2, · · · , N (2.57)
n=N +1
The quantities rp and the noise b (t) are therefore independent since Gaussian.
In conclusion, the optimal receiver can base its decision only on the quantities
rp , p = 1, 2, · · · , N with:
T
rp = r(t)νp (t)dt (2.58)
0
Passing the signal r(t) provided by the transmission channel to the N quantities
rp is called demodulation.
Example
Let us consider an M-PSK modulation for which the sj (t) signals are of the
form:
sj (t) = Ah(t) cos(2πf0 t + ϕ0 + φj )
The signals sj (t) define a space with N = 2 dimensions if M > 2. The normed
and orthogonal functions νp (t), p = 1, 2 can be expressed respectively as:

2
ν1 (t) = cos(2πf0 t + ϕ0 )
T
2
ν2 (t) = T
sin(2πf0 t + ϕ0 )
and the signals sj (t) can be written:

T T
sj (t) = A cos φj h(t)ν1 (t) − A sin φj h(t)ν2 (t)
2 2
After demodulation, the observation R = (r1 , r2 ) is equal to:

T T
r1 = A cos φj + b1 r2 = A sin φj + b2
2 2
The observation R = (r1 , r2 ) depends only on the states of phase φj and on the
noise. We say that observation R = (r1 , r2 ) is in baseband since independent of
the carrier frequency f0 .
The demodulation operation requires knowledge of the frequency f0 and
the phase ϕ0 of the carrier, the signals νp (t) having to be synchronous with
the carrier generated by the modulator. That is the reason why we speak of
synchronous demodulation or coherent demodulation.
The N integrators of the demodulator can be replaced by N h(T − t) impulse

response filters, each followed by a sampler at time t = T .

+∞
sj (t)νj (t) ∗ h(T − t) = sj (τ )νj (τ )h(T − t + τ )dτ

−∞
where ∗ represents the convolution product.

Sampling at t = T , we obtain:
T
sj (t)νj (t) ∗ h(T − t) |t=T = sj (τ )νj (τ )dτ
0
which is equal to the output of the integrator.

The filter h(T − t) is called the filter matched to waveform h(t) of width T . We
can show that this filter maximizes the signal to noise ratio at its output at time
t = T.
For a continuous data stream, the integration is performed on each interval
[iT, (i + 1)T [ i = 1, 2, · · · and, if we use matched filters, the sampling is realized
at time (i + 1)T .
After demodulation, the receiver must take a decision about the group of
data transmitted on each time interval [iT, (i + 1)T [. To do this, it searches
for the most probable signal sj (t) by using the maximum a posteriori (MAP)
probability criterion:
ŝj (t) if Pr {sj (t) | R} > Pr {sp (t) | R} ∀ p = j p = 1, 2, · · · , M
where ŝj (t) is the signal that was transmitted and R = (r1 · · · rp · · · rN ) the
output of the demodulator. To simplify the notations, the time reference has
been omitted for the components of observation R. Pr {sj (t)/R} denotes the
probability of sj (t) conditionally to the knowledge of observation R.
Using Bayes’ rule, the MAP criterion can again be written:
ŝj (t) if πj p(R|sj (t)) > πp p(R|sp (t)) ∀ p = j p = 1, 2, · · · , M
where πj = Pr {sj (t)} represents the a priori probability of transmitting the sig-
nal sj (t) and p(R|sj (t)) is the probability density of observation R conditionally
to the emission of the signal sj (t) by the modulator.
Taking into account the fact that the components rp = sjp +bp of observation
R conditionally to the emission of the signal sj (t) are non-correlated Gaussian,
with mean sjp and variance N0 /2, we can write:
'
N '
N
ŝj (t) if πj p(rp |sj (t)) > πn p(rp |sn (t))
p=1 p=1
∀n = j p = 1, 2, · · · , M
Replacing the probability densities by their respective expression we obtain:

" #

N
N
ŝj (t) if πj √ 1
πN0
exp − N01
(rp − sjp )2
p=1
" #

N
N
> πn √πN 1
exp − N0 1
(rp − snp )2
0
p=1
After simplification:

N
N
ŝj (t) if rp sjp + Cj > rp snp + Cn ∀ n = j n = 1, 2, · · · , M (2.59)
p=1 p=1
Ej
N
where Cj = N0
2
ln(πj ) − 2
with Ej = (sjp )2 .
p=1
Noting that:
T T
N
N
r(t)sj (t)dt = rp νp (t) sjm νm (t)dt
0 0 p=1 m=1
and recalling that the functions νp (t) are normed and orthogonal, we obtain:
T
N
r(t)sj (t)dt = rp sjp
0 p=1
In the same way:
T T
N
N
s2j (t)dt = sjp νp (t) sjm νm (t)dt
0 0 p=1 m=1
and finally:
T
N
s2j (t)dt = s2jp
0 p=1
Taking into account the above, the MAP criterion can again be written in the
form:
(T (T
ŝj (t) if r(t)sj (t)dt + Cj > r(t)sn (t)dt + Cn
0 0
(2.60)
∀ n = j n = 1, 2, · · · , M
Ej (T
where Cj = N0
2 ln(πj ) − 2 with Ej = s2j (t)dt.
0
If all the sj (t) signals are transmitted with the same probability (πj = 1/M ),
the term Cj reduces to −Ej /2. In addition, if all the sj (t) signals have in
addition the same energy Ej = E (the case of phase or frequency shift keying
modulation), then the MAP criterion is simplified and becomes:
T T
ŝj (t) if r(t)sj (t)dt > r(t)sn (t)dt ∀ n = j n = 1, 2, · · · , M (2.61)
0 0
2.2.2 Performance of the coherent receiver

Amplitude shift keying with M states
For an M-ASK modulation, the sj (t) signals are of the form:
sj (t) = Aj h(t) cos(2πf0 t + ϕ0 )

with:
Aj = (2j − 1 − M )A j = 1, 2, · · · , M
They define a space of dimension N = 1 and thus observation R at the output
of the demodulator reduces to a component r.
T
r= r(t)ν(t)dt
0

with ν(t) = T2 cos(2πf0 t + ϕ0 ).
For a continuous data stream, the estimation of the symbols ai is done by
integrating the product r(t)ν(t) on each time interval [iT, (i + 1)T [. If a matched
filter is used rather than an integrator, the sampling at the output of the filter
is realized at time (i + 1)T .
On the interval [0, T [, assuming iid information data di , all the amplitude
states have the same probability and decision rule (2.59) leads to:
1 1
Âj if rsj − s2j > rsn − s2n ∀ n = j (2.62)
2 2
where Âj represents the amplitude estimated with:
T
T 1
sj = Aj cos(2πf0 t + ϕ0 )ν(t)dt = Aj if f0 >> (2.63)
2 T
0
The coherent receiver, shown in Figure 2.13, takes its decision by comparing
observation r to a set of (M − 1) thresholds of the form:

−(M − 2)A T2 , · · · , −2pA T2 , · · · , −2A T2 , 0,
(2.64)
2A T2 , · · · , 2pA T2 , · · · , (M − 2)A T2
Figure 2.13 – Coherent receiver for M-ASK modulation.
Example
Consider a 4-ASK modulation, the three thresholds being −2A T2 , 0, 2A T2 .
The decisions are the following:

Âj = −3A if r < −2A T2

Âj = −A if −2A T2 < r < 0

Âj = A if 0 < r < 2A T2

Âj = 3A if r > 2A T2
The emission of a state of amplitude Aj corresponds to the transmission of a

group of log2 (M ) binary data di . The error probability on a group of data can be
calculated as the average value of the conditional error probabilities P e2j−1−M
given by:
) *
P e2j−1−M = Pr Âj = (2j − 1 − M )A | Aj = (2j − 1 − M )A
The mean error probability on the symbols, denoted P es, is equal to:
1
M
P es = P e2j−1−M
M j=1
The conditional error probabilities can be classified into two types. The first
type corresponds to the probabilities that the observation is higher or is lower
than a certain threshold and the second type, to the probabilities that the
observation does not fall between two thresholds.
TYPE 1: Probabilities that the observation is higher or is lower than a threshold

) *
P e(M−1) = Pr Âj = (M − 1)A | Aj = (M − 1)A
) *
= Pr r < (M − 2)A T2 | Aj = (M − 1)A
) *
P e−(M−1) = Pr Âj = −(M − 1)A | Aj = −(M − 1)A
) *
= Pr r > −(M − 2)A T2 | Aj = −(M − 1)A
TYPE 2: Probabilities that the observation does not fall between two thresholds
) *
P e2j−1−M = Pr Âj = (2j − 1 − M )A | Aj = (2j − 1 − M )A
)
P e2j−1−M = 1− Pr (2j − 2 − M )A T2 < r < (2j − M )A T2 |
Aj = (2j − 1 − M )A}
Observation r is+Gaussian conditionally to a realization of the amplitude Aj ,
with mean ±Aj T /2 and variance N0 /2. The conditional probabilities have
the expressions: ,
1 A2 T
P e(M−1) = P e−(M−1) = erfc
2 2N0
,
A2 T
P e(2j−1−M) = erfc
2N0
where the complementary error function is always defined by:

+∞
2
erfc(x) = 1 − erf(x) = √ exp(−u2 )du
π
x
To calculate the mean error probability on the groups of data, we have two
conditional probabilities of type 1, and (M − 2) conditional probabilities of type
2. ,
M −1 A2 T
P es = erfc
M 2N0
2 (M 2 −1)
Introducing the average energy Es = A2T 3
received by group of data, the
mean error probability is again equal to:

M −1 3 Es
P es = erfc
M M 2 − 1 N0
The symbol error probability P es can also be expressed as a function of the

average energy Eb = Es / log2 (M ) received by transmitted bit di .
,
M −1 3 log2 (M ) Eb
P es = erfc
M M 2 − 1 N0
or again as a function of the received mean power P and of the transmitted bit
rate D = 1/Tb : ,
M −1 3 log2 (M ) P
P es = erfc (2.65)
M M 2 − 1 N0 D
Figure 2.14 provides the mean error probability P es as a function of the signal
to noise ratio Eb /N0 for different values of the parameter M .
Figure 2.14 – Mean error probability P es as a function of the signal to noise ratio
Eb /N0 for different values of parameter M of an M-ASK modulation.
The bit error probability P eb can be deduced from the mean error proba-
bility P es in the case where Gray coding is used and under the hypothesis of
a sufficiently high signal to noise ratio. Indeed, in this case we generally have
an erroneous bit among the log2 (M ) data transmitted. (We assume that the
amplitude of the received symbol has a value immediately lower or higher than
the value of the transmitted amplitude).
P es Eb
P eb ∼
= if >> 1 (2.66)
log2 (M ) N0
Phase shift keying with M states

For M-PSK modulation , the sj (t) signals are of the form:
sj (t) = A cos(2πf0 t + ϕ0 + φj ) (2.67)
with:
π
φj = (2j + 1) + θ0 j = 0, 1, · · · , (M − 1)
M
The sj (t) signals, for M > 2, define a two-dimensional space. Observation R at
the output of the demodulator is therefore made up of two components (r1 , r2 )
with:
T T
r1 = r(t)ν1 (t)dt r2 = r(t)ν2 (t)dt
0 0

where ν1 (t) = T2 cos(2πf0 t + ϕ0 ) et ν2 (t) = − T2 sin(2πf0 t + ϕ0 ).
Using decision rule (2.50) and assuming the information data iid, all the states
of phase have the same probability and the decision is the following:

2
2
φ̂j if rp sjp > rp snp ∀n = j (2.68)
p=1 p=1
with:

T T 1
sj1 = A cos φj and sj2 = A sin φj if f0 >> (2.69)
2 2 T
Taking into account the expressions of sj1 and of sj2 , the decision rule can again
be written:
φ̂j if r1 cos φj + r2 sin φj > r1 cos φn + r2 sin φn ∀n = j (2.70)
The coherent receiver for an M-PSK modulation is represented in Figure 2.15.

It is made up of two components
+ called the phase component (projection of the
received signal on ν1 (t) = 2/T cos(2πf0 t+ϕ +0 )) and the quadrature component
(projection of the received signal on ν2 (t) = 2/T sin(2πf0 t+ϕ0 )) and a decision
circuit.
The emission by the modulator of a phase state corresponds to the transmis-
sion of a group of log2 (M ) information bit. The error probability on a group of
Figure 2.15 – Coherent receiver for M-PSK modulation.
binary data, whatever the value of M , does not have an analytical expression.
However, at high signal to noise ratios, this probability is well approximated by
the following expression:
! $
∼ Eb π Eb
P es = erfc log2 (M ) sin if >> 1 (2.71)
N0 M N0
Noting that Eb = P Tb and D = 1/Tb , the relation Eb /N0 is again equal to

P/N0 D where P is the received power of the modulated signal.
For Gray coding, the bit error probability with a high signal to noise ratio
is equal to:
P es Eb
P eb ∼
= if >> 1 (2.72)
log2 (M ) N0
Case of 2-PSK modulation
For this modulation, phase φj takes the values 0 or π. Each phase state is
therefore associated with a bit. Adopting the following coding:
φj = 0 → di = 1 φj = π → di = 0
the decision rule for 2-PSK modulation is simple:
dî = 1 if r1 > 0 dî = 0 if r1 < 0 (2.73)
Observation r2 is not used for decoding the data di since the space defined by
the signals modulated with two phase states has dimension N = 1.
For 2-PSK modulation, there is an exact expression of the bit error proba-
bility P eb. Assuming the binary data iid, this error probability is equal to:
1 1
P eb = Pr {r1 > 0|φj = π} + Pr {r1 < 0|φj = 0}
2 2
Output r1 of the demodulator is:
+
r1 = ± Eb + b
where Eb = A2 T /2 is the energy received per information bit transmitted and b

is an AWGN, with zero mean and variance equal to N0 /2.
1√1
(∞ √
P eb = 2 πN0
exp(− N10 (r1 + Eb )2 )dr1
0
(0 √
+ 21 √πN
1
exp − N10 (r1 − Eb )2 dr1
0
−∞
Introducing the complementary error function, the bit error probability P eb is

equal to:
1 Eb
P eb = erfc (2.74)
2 N0
Case of 4-PSK modulation
For this modulation, phase φj takes four values π/4, 3π/4, 5π/4, 7π/4.
With each state of the phase are associated two binary data. For equiprob-
able phase states, the MAP criterion leads to the following decision rules:
π
φ̂j = 4 if r1 > 0; r2 > 0
3π
φ̂j = 4
if r1 < 0; r2 > 0
5π
φ̂j = 4
if r1 < 0; r2 < 0
7π
φ̂j = 4 if r1 > 0; r2 < 0
Considering the following Gray coding:

π 3π 5π 7π
→ 11 → 01 → 00 → 10
4 4 4 4
The estimation of the binary data can be performed by separately comparing
outputs r1 and r2 of the demodulator to a threshold fixed to zero. The coherent
receiver for a 4-PSK modulation is represented in Figure 2.16.
Figure 2.16 – Coherent receiver for 4-PSK modulation.
An exact expression of P eb can however be given by observing simply that

the coherent receiver for a 4-PSK modulation is made up of two components
identical to that of a 2-PSK receiver. The bit error probability P eb for a 4-PSK
modulation is thus the same as for a modulation 2-PSK, that is:

1 Eb
P eb = erfc (2.75)
2 N0
The symbol error probability P es is equal to:
P es = 1 − (1 − P eb)2
For high signal to noise ratios, the error probability P eb is much lower than
unity and thus, for 4-PSK modulation:
Eb
P es = 2P eb if >> 1 (2.76)
N0
Figure 2.17 shows the error probability P es as a function of the relation Eb /N0
for different values of the parameter M 1 .
Amplitude modulation on two quadrature carriers – M-QAM

For M-QAM modulation, the sj (t) signals provided by the modulator are of the
form:
sj (t) = Ajc h(t) cos(2πf0 t + ϕ0 ) − Ajs h(t) sin(2πf0 t + ϕ0 ) (2.77)
Two situations can arise depending whether the length m of the groups of data
at the input of the modulator is even or not. When M = 2m with even m,
1 For M = 2, it is the exact relation 2.74 that is used since P es = P eb, for M > 2, P es is
provided by Equation (2.71).

Figure 2.17 – Error probability P es as a function of the relation Eb /N0 for different
values of M of an M-PSK modulation.
the group of data can be separated into two sub-groups of length m/2, each
sub-group being associated respectively with amplitudes Ajc and Ajs , with:
√ √
Ajc = (2j − 1 − M )A j = 1, 2, · · · , M
√ √ (2.78)
Ajs = (2j − 1 − M )A j = 1, 2, · · · , M
√
In this case, the M-QAM modulation is equivalent to two M -ASK modulations
having quadrature carriers. The coherent receiver for an M-QAM modulation is
made up of two components called phase
√ and quadrature, and each component,
similar to a receiver for modulation M -ASK, performs the estimation of a
group of m/2 binary data. The receiver is shown in Figure 2.18.
The error probability√on a group of m/2 binary data is equal to the error
probability of an ASK− M modulation, that is:
√ , √
M −1 3 log2 ( M ) Eb
P em/2 = √ erfc (2.79)
M M − 1 N0
Figure 2.18 – Coherent receiver for M-QAM modulation with M = 2m and even m.
The symbol error probability P es on the group of m binary data is therefore

equal to:
P es = 1 − (1 − P em/2 )2
When m is odd, the probability of error P es on a group of m binary data can
be upper bound by the following expression:
, √
3 log2 ( M ) Eb
P es ≤ 2erfc (2.80)
M − 1 N0
The bit error probability P eb can be deduced from P es if Gray coding is used
and for a sufficiently high signal to noise ratio:
P es
P eb = (2.81)
log2 (M )
For the values of M ≥ 8, the performance of M-PSK and M-QAM modulations

can easily be compared. Indeed, by making the following approximation:
π
π
sin ∼
= if M ≥ 8
M M
the bit error probability for an M-PSK modulation can be written:
,
1 π 2 Eb Eb
P eb = erfc log2 (M ) 2 if >> 1 (2.82)
log2 (M ) M N0 N0
Neglecting the coefficients that weight the complementary error functions, M-

PSK modulation requires the relation Eb /N0 to be increased by:

3M 2
10 log dB
2(M − 1)π 2
to obtain the same error probability as M-QAM modulation. If, for example, we
compare the performance of 16-PSK modulation with that of 16-QAM modula-
tion, we note that the former requires about 4 dB more for the relation Eb /N0
to obtain the same error probabilities.
Frequency shift keying – M-FSK

For an M-FSK modulation, the modulator provides signals of the form:
sj (t) = Ah(t) cos(2πfj t + ϕj ) (2.83)
where the frequencies fj are chosen in such a way that the M sj (t) signals are
orthogonal. The space defined by these signals therefore has dimension N = M
and the vectors νj (t) are of the form:

2
νj (t) = cos(2πfj t + ϕj ) j = 1, 2, · · · M (2.84)
T
Assuming the information data di iid, the sj (t) signals have the same probability.
In addition, they have the same energy E and thus, the MAP criterion leads to
the following decision rule:

M
M
ŝj (t) if rp sjp > rp snp ∀n = j (2.85)
p=1 p=1
where sjp is equal to:
T
T
sjp = sj (t)νp (t)dt = A δjp (2.86)
2
0
Taking into account the expression of sjp , decision rule (2.85) is simplified and
becomes:
ŝj (t) if rj > rn ∀n = j (2.87)
The optimal coherent receiver for an M-FSK modulation is shown in Figure 2.19.
Conditionally to the emission of the signal sj (t), the M outputs of the de-
modulator are of the form:
+
rj = Es + bj rp = bp ∀p = j
Figure 2.19 – Coherent receiver for M-FSK modulation.
where bj and bp are AWGN, with zero mean and variance equal to N0 /2.
The probability of a correct decision on a group of binary data, conditionally to
the emission of the signal sj (t) is equal to:

+∞
P cj = Pr {b1 < rj , · · · , bp < rj , · · · , bM < rj } p(rj )drj

−∞
The noises being non-correlated and therefore independent, since they are Gaus-
sian, we have:
⎛ rj ⎞M−1
2

1 b
Pr {b1 < rj , · · · , bp < rj , · · · , bM < rj } = ⎝ √ exp − db⎠
πN0 N0
−∞
and thus the probability of a correct decision is equal to:

⎛ ⎞M−1

+∞ rj
1 1 1 1 +
P cj = ⎝ √ exp − b db √2 ⎠ exp − (rj − Es ) drj
2
πN0 N0 πN0 N0
−∞ −∞
After changing variables, the probability of a correct decision can be expressed

as a function of the relation Es /N0 .
⎛ ⎞M−1 " #

+∞ y 2
1 ⎝ 1 x 1 E
dx⎠ exp − (y −
s
P cj = √ √ exp − 2
) dy (2.88)
2π 2π 2 2 N0
−∞ −∞
The probability of a correct decision is the same whatever the transmitted signal.
The signals sj (t) being equiprobable, the mean probability of a correct decision
P c is therefore equal to the conditional probability P cj . The symbol error
probability is then equal to:
P es = 1 − P c
The error probability can also be expressed as a function of the relation Eb /N0
where Eb is the energy used to transmit a bit with Eb = Es / log2 (M ).
We can also try to determine the bit error probability P eb. All the M − 1
groups of erroneous data appear with the same probability:
P es
(2.89)
M −1
In a group oferroneous
data, we can have k erroneous data among m and this
m
can occur in possible ways. Thus, the average number of erroneous data
k
in a group is:
m m
P es 2m−1
k =m m P es
k M −1 2 −1
k=1
and finally the bit error probability is equal to:
2m−1
P eb = P es (2.90)
2m − 1
where m = log2 (M ).
The error probability for an M-FSK modulation does not have a simple
expression and we have to resort to digital computation to determine this prob-
ability as a function of the relation Eb /N0 . We show that for a given error
probability P eb, the relation Eb /N0 necessary decreases when M increases. We
also show that probability P es tends towards a value arbitrarily small when M
tends towards infinity, and for Eb /N0 = 4 ln 2 dB, that is, -1.6 dB.
For a binary transmission (M = 2), there is an expression of the error probability
P eb.
Let us assume that the signal transmitted is s1 (t), we then have:
+
r1 = Eb + b1 r2 = b2
The decision can be taken by comparing z = r1 − r2 to a threshold fixed to zero.

The error probability P eb1 conditionally to the emission of s1 (t), is equal to:
P eb1 = Pr {z < 0 | s1 (t)}
Assuming the two signals identically distributed, error probability P eb has the
expression:
1
P eb = (P eb1 + P eb2 )
2
The noises b1 and b2 are non-correlated Gaussian, with zero mean and same
variance equal to N0 /2. The variable z, conditionally to the emission of the
√
signal s1 (t), is therefore Gaussian, with mean Eb and variance N0 . Thus
probability P eb1 is equal to:
0 √
1 (z − Eb )2
P eb1 = √ exp − dz
2πN0 2N0
−∞
Introducing the complementary error function, P eb1 is written:

1 Eb
P eb1 = erfc
2 2N0
It is easy to check that the error probability conditionally to the emission of
signal s2 (t) is identical to the error probability conditionally to the emission of
signal s1 (t) and thus, we obtain:

1 Eb
P eb = erfc (2.91)
2 2N0
If we compare the performance of 2-FSK modulation to that of 2-PSK we note
that the former requires 3 dB more signal to noise ratio to obtain the same
performance as the latter.
Continuous phase frequency shift keying – CPFSK

For continuous phase frequency shift keying, the modulated signal has the ex-
pression:
S(t) = A cos(2πf0 t + φ(t)) (2.92)
where the phase φ(t), on the interval [iT, (i + 1)T [, is equal to :

i
φ(t) = 2πh an q(t − nT ) + θi−L (2.93)
n=i−L+1
with:

i−L
θi−L = πh an
n=−∞
h is the modulation index and the symbols ai are M -ary in the general case.
They take their values in the alphabet { ±1, ±3, · · · , ±(2p + 1), · · · , ± ( M
−1 ) }; M = 2m .
If the modulation index h = m/p where m and p are relatively prime integers,
phase θi−L takes its values in the following sets:
) *
(p−1)πm
θi−L ∈ 0, πm p , p , ··· ,
2πm
p if m even
) * (2.94)
(2p−1)πm
θi−L ∈ 0, πm 2πm
p , p , ··· , p if m odd
The evolution of phase φ(t) can be represented by a trellis whose states are
defined by (ai−L+1 , ai−L+2 , · · · , ai−1 ; θi−L ), that is:
(M L−1 p) states if m even

(2.95)
(M L−1 2p) states if m odd
Note that the complexity of the trellis increases very rapidly with the parameters
M and L. For example, for a modulation with quaternary symbols (M = 4)
with a partial response with modulation index h = 1/2 and with parameter
L = 4, the trellis has 256 states. For MSK and GMSK, the symbols ai are
Figure 2.20 – Trellis associated with phase φ(t) for MSK modulation.
binary (M = 2) and the modulation index h is 1/2, that is, m = 1 and p = 2.

Phase θi−L therefore takes its values in the set {0, π/2, π, 3π/2} and the trellis
associated with phase φ(t) has 2L−1 × 4 states. Figure 2.20 shows the trellis
associated with phase φ(t) for MSK modulation.
To decode symbols ai we use the Viterbi algorithm whose principle is recalled
below. For each time interval [iT, (i + 1)T [, proceed in the following way:
• for each branch l leaving a state of the trellis at instant iT calculate metric
zil as defined later, that is, for MSK and GMSK, 2L × 4 metrics have to
be calculated;
• for each path converging to instant (i + 1)T towards a state of the trel-
lis, calculate the cumulated metric, then select the path with the largest
cumulated metric, called the survivor path;
• among the survivor paths, trace back along s branches of the path having
the largest cumulated metric and decode symbol ai−s ;
• continue the algorithm on the following time interval.
Branch metric zil has the expression:

(i+1)T
zil = r(t) cos(2πf0 t + φli (t) + ϕ0 )dt

iT
where r(t) = s(t) + b(t) is the signal received by the receiver and b(t) is a
AWGN, with zero mean and power spectral density equal to N0 /2. Quantity
φli (t) represents a realization of phase φ(t) associated with branch l of the trellis
on time interval [iT, (i + 1)T [.
Taking into account the fact that the noise can be put in the form b(t) =
bc (t) cos(2πf0 t + ϕ0 ) − bs (t) sin(2πf0 t + ϕ0 ) and that f0 >> 1/T , the branch
metric can again be written:

(i+1)T
(i+1)T
zil = rc (t) cos φli (t)dt + rs (t) sin φli (t)dt (2.96)
1T iT
where the signals rc (t) and rs (t) are obtained after transposition into baseband
of r(t)(multiplying r(t) by cos(2πf0 t + ϕ0 ) and − sin(2πf0 t + ϕ0 ) respectively,
then lowpass filtering).

i
l
cos φi (t) = cos 2πh an q(t − nT ) + θi−L
l l
n=i−L+1
(2.97)

i
sin φli (t) = sin 2πh aln q(t − nT ) + l
θi−L
n=i−L+1
Putting:

i
ψil (t) = 2πh aln q(t − nT )
n=i−L+1
branch metric zil can again be written in the form:

zil = cos θi−L
l l
Al + sin θi−L Bl (2.98)
with:

(i+1)T
Ali = (rc (t) cos ψil (t) + rs (t) sin ψil (t))dt
iT

(i+1)T
Bil = (rs (t) cos ψil (t) − rc (t) sin ψil (t))dt
iT
For MSK modulation, it is possible to decode symbols ai by using a receiver
similar to that of 4-PSK modulation. Indeed, the MSK signal can be written in
the following form:

S(t) = A c2i−1 h(t − 2iT ) cos 2T
πt
cos(2πf0 t + ϕ0 )
i (2.99)

− c2i h(t − (2i + 1)T ) sin 2T
πt
sin(2πf0 t + ϕ0 )
i
where symbols ci are deduced from symbols ai by a coding by transition.

c2i = a2i c2i−1 and c2i−1 = a2i−1 c2i−2 (2.100)
and h(t) is a unit amplitude rectangular pulse shape of width 2T :
h(t) = 1 if t ∈ [−T, T [
= 0 elsewhere
The coherent receiver for MSK is shown in Figure 2.21. It comprises two matched
filters at h(t) with waveform h(2T − t). Symbols c2i−1 and c2i are decoded by
comparing samples taken at the output of the matched filters, at time 2iT and
(2i + 1)T respectively.
Figure 2.21 – Coherent receiver for MSK modulation.
It is easy to show that the error probabilities on binary symbols c2i−1 and
c2i are identical and equal to:

1 Eb
P eci = erfc (2.101)
2 N0
where Eb is the energy used to transmit a binary symbol ci .
To obtain binary data ai from the symbols ci , at the output of the coherent
receiver we have to use a differential decoder given by the following equations:
a2i = c2i c2i−1 and a2i−1 = c2i−1 c2I−2
The bit error probability P eb on ai is:
P eb = 1 − (1 − P eci )2
thus for P eci << 1, a good approximation of the bit error probability P eb is:
P eb ≈ 2P eci (2.102)
As a first approximation, the performance of the MSK modulation is identical
to that of the 4-PSK modulation.
2.3 Transmission on a band-limited channel

2.3.1 Introduction
In this chapter, so far we have assumed that the bandwidth allocated to the
transmission is infinite. We will now envisage a more realistic situation where
a bandwidth W is available to transmit the modulated signal. In this band W
the channel is assumed to be frequency non-selective. We shall restrict ourselves
to the case of linear modulations of the M-ASK, M-PSK and M-QAM types
that have a power spectral density made up of a main lobe of width 2/T , where
1/T is the modulation speed, and sidelobes with zero crossing at n/T . The
bandwidth of a linearly modulated signal is therefore, strictly speaking, infinite.
The modulated signal must consequently be filtered by an emission filter before
being placed at the input of the transmission channel. We are now going to
determine what the minimum band W is necessary to transmit the modulated
signal without degradation of the performance compared to a transmission with
an infinite bandwidth. The frequency response of the emission and reception
filters will also be established.
In what follows, we will consider the complex envelope of the modulated
signal and the equivalent baseband response of the emission filter. Without
compromising the generality of our remarks, this avoids introducing the carrier
frequency, which complicates the notations.
The complex envelope of an M-ASK, M-PSK and M-QAM signal has the
expression:
se (t) = A ci h(t − iT ) (2.103)
i
where h(t) is a unit amplitude rectangular pulse shape of width T , and ci =

ai + jbi is a modulation symbol with:
M-ASK ci = ai bi = 0
M-PSK ai = cos φi bi = sin φi

√
M-QAM ai and bi symbols M − ary
Let g(t) be the impulse response of the emission passband filter centred on
the carrier frequency. This waveform can be written:
g(t) = gc (t) cos(2πf0 t + θ0 ) − gs (t) sin(2πf0 t + θ0 ) (2.104)
or equivalently:
g(t) =
e {ge (t) exp [j(2πf0 t + θ0 )]} (2.105)
where ge (t) = gc (t) + jgs (t) is the baseband-equivalent waveform of the emission
filter. The output e(t) of the emission filter is equal to:

e(t) = A ci z(t − iT ) (2.106)
i
where z(t) = h(t) ⊗ ge (t) is, in the general case, a complex waveform while h(t)
is real.
2.3.2 Intersymbol interference

After passing through the emission filter, the modulated signal has a bandwidth
W and, thus, the signal at the output of the transmission channel has the
expression:
r(t) = e(t) + b(t) (2.107)
where b(t) is a complex AWGN, with zero mean and power spectral density
equal to 2N0 .
The coherent receiver uses a reception filter followed by a sampler at time
t0 + nT , where t0 can be chosen arbitrarily. The output of the reception filter
with impulse response gr (t) has the expression:

y(t) = A ci x(t − iT ) + b (t) (2.108)
i
where:
x(t) = z(t) ⊗ gr (t)
b (t) = b(t) ⊗ gr (t)
Sampling signal y(t) at time t0 + nT , we obtain:

y(t0 + nT ) = A ci x(t0 + (n − i)T ) + b (t0 + nT ) (2.109)
i
Considering that in the general case x(t) = p(t) + jq(t) is a complex waveform,
the sample y(t0 + nT ) can again be written in the form:

y(t0 + nT ) = Acn p(t0 ) + A cn−i p(t0 + iT )
i=0
(2.110)
+jA cn−i q(t0 + iT ) + b (t0 + nT )
i
The first term Acn p(t0 ) represents the desired information for the decoding of
the symbol cn , the following two terms being Intersymbol Interference (ISI)
terms. Let us examine the outputs of the two components of the receiver, called
the in-phase and quadrature components, corresponding to the real part and
to the imaginary part of y(t0 + nT ) respectively. We can notice that the in-
phase component (respectively the quadrature component) depends on symbols
ai (respectively on symbols bi ) but also on symbols bi (respectively on symbols

ai ). We sometimes say that there is crosstalk between the two components of the
receiver. Of course, ISI is a phenomenon that can only degrade the quality of the
transmission. This is the reason why it is important to define the condition to
be satisfied in order to cancel the ISI. But before that, we are going to indicate
a simple way to characterize the ISI by tracing the eye diagram, thus called by
analogy with the shape of the human eye, at the output of the reception filter
of the receiver’s in-phase and quadrature components.
The eye diagram is the figure obtained by superposing all the plots or realizations
of the non-noised signal yc (t) where yc (t) is the real part of y(t). We also obtain
the eye diagram from non-noised ys (t) where ys (t) is the imaginary part of y(t).

yc (t) = A ai p(t − iT ) − A bi q(t − iT )
i i
(2.111)
ys (t) = A bi p(t − iT ) + A ai q(t − iT )
i i
Let us analyse, for example, output yc (t) of the reception filter of the in-phase
component on time interval [t1 + nT, t1 + (n + 1)T [ where t1 represents an ar-
bitrary time. Replacing t by t + t1 + nT , signal yc (t) can be written:

yc (t + t1 + nT ) = A an−i p(t + t1 + iT )
i (2.112)
−A bn−i q(t + t1 + iT ) for 0 ≤ t ≤ T
i
Assuming x(t+t1 +iT ) to be negligible outside the interval [t1 − L1 T, t1 + L2 T ],

each sum of the previous expression includes (L1 + L2 + 1) terms, that is,
4(L1 +L2 +1) possible plots for the eye diagram. If we shift the observation in-
terval by a quantity T , the number of plots that make up the eye diagram is
always 4(L1 +L2 +1) , the eye diagram is therefore a figure that is repeated every
T seconds. Its analysis can, thus, be limited to an interval of time T .
The eye diagram can be visualized on an oscilloscope. Indeed, the different
plots of p(t) and q(t) can remain on the screen if the scanning speed of the
oscilloscope is fast enough compared to the remanence time of the cathode ray
tube or, better, if it is a deep memory oscilloscope. We show the eye diagram of
an ISI and ISI-free 2-PSK modulation in Figure 2.22. For this modulation, the
symbols ci are real and the receiver has a single component.

yc (t + t1 + nT ) = A an−i p(t + t1 + iT ) for 0 ≤ t ≤ T (2.113)
i
In the absence of ISI, at the sampling time, all the plots of p(t) pass through a
single point. The more open the eye diagram at the sampling time, the greater
the immunity of the transmission to noise. In the same way, the greater the
horizontal aperture of the eye diagram, the less sensitive the transmission is to
Figure 2.22 – Eye diagram of an (a) ISI and (b) ISI-free 2-PSK modulation.
positioning errors of the sampling time. In the presence of ISI, the different
plots of p(t) no longer pass through a single point at the sampling time and the
ISI contributes to closing the eye diagram.
The output of the reception filter at time t0 + nT of the in-phase component
of the receiver is equal to:

yc (t0 + nT ) = Aan p(t0 ) + A an−i p(t0 + iT )
i=0 (2.114)
−A bn−i q(t0 + iT )
i
√
For a M-QAM
√ modulation, the useful signal is A(2j − 1 − M )p(t0 ), for j =
1, · · · , M and the decision is taken by comparing signal yc (t0 + nT ) to a set
of thresholds separated by 2p(t0 ). There will be errors in the absence of noise if
the ISI, for certain configurations of symbols ci , is such that the received signal
is situated outside the correct decision zone. This occurs if the ISI is higher in
absolute value to p(t0 ). This situation is translated by the following condition:

Maxan−i ,bn−i an−i p(t0 + iT ) − bn−i q(t0 + iT ) > |p(t0 )|
i=0 i
Taking
√ into account the fact that the largest value taken by symbols ai and bi
is M − 1, the previous condition becomes:
" #

|p(t0 + iT )| + |q(t0 + iT )|
√ i=0 i
Dmax = ( M − 1) ≥1 (2.115)
|p(t0 )|
Quantity Dmax is called the maximum distortion. When the maximum distor-
tion is greater than unity, the eye diagram is closed at the sampling time and
errors are possible, even in the absence of noise.
2.3.3 Condition of absence of ISI: Nyquist criterion

The absence of ISI is translated by the following conditions:
p(t0 + iT ) = 0 ∀ i = 0 (2.116)
q(t0 + iT ) = 0 ∀i (2.117)
which can again be written by using the complex signal x(t) = p(t) + jq(t)
x(t0 + iT ) = p(t0 )δ0,i ∀i (2.118)
where δ0,i is the Kronecker symbol.

Let us introduce the sampled signal xE (t), defined by:

xE (t) = x(t) δ(t − t0 − iT ) (2.119)
i

We can notice that the impulse train u(t) = δ(t − t0 − iT ) is periodic, of
i
period T . It can therefore be decomposed into the Fourier series:

1 i i
u(t) = exp −j2π t0 exp −j2π t (2.120)
T i T T
Since we are seeking to determine the minimal bandwidth W necessary to trans-

mit the ISI-free modulated signal, it is wise to work in the frequency domain.
Taking the Fourier transform denoted XE (f ) of relation (2.119) and taking into
account the previous expression of u(t), we obtain:

1 i i
XE (f ) = exp −j2π t0 X(f − ) (2.121)
T i T T
The sampled signal, according to relation (2.119), can also be written:

xE (t) = x(t0 + iT )δ(t − t0 − iT ) (2.122)
i
which, after Fourier transform and taking into account the condition of absence
of ISI, becomes:
XE (f ) = p(t0 ) exp(−j2πf t0 ) (2.123)
The equality of the relations (2.121) and (2.123) gives:

i i
exp j2π(f − )t0 X(f − ) = T p(t0 )
i
T T
Putting:
X(f )
X t0 (f ) = exp(j2πf t0 )
p(t0 )
the condition for the absence of ISI can be expressed from X t0 (f ) by the follow-
ing relation:

+∞
i
X t0 (f − ) = T (2.124)
i=−∞
T
This condition of absence of ISI is called the Nyquist criterion.

Let us recall that the transmission channel with spectrum C(f ) has a pass-
band W .
C(f ) = 0 for |f | > W (2.125)
Let us consider relation (2.124) for three situations.
1. X t0 (f ) has a bandwidth W < 1/2T . Relation (2.124) being a sum of

functions shifted by 1/T , there are no functions X t0 (f ) that enable the
Nyquist criterion to be satisfied. The bandwidth W necessary for ISI-free
transmission is therefore higher than or equal to 1/2T .
2. X t0 (f ) has a bandwidth W = 1/2T . There is a single solution that satisfies
the Nyquist criterion:
X t0 (f ) = T |f | ≤ W
(2.126)
= 0 elsewhere
or again:
X(f ) = T p(t0 ) exp(−j2πt0 ) |f | ≤ W
(2.127)
= 0 elsewhere
which, in the time domain, gives:
sin [π(t − t0 )/T ]

x(t) = p(t0 ) (2.128)
π(t − t0 )/T
This solution corresponds to a strictly speaking non-causal waveform x(t).

However, since the function sin y/y decreases fairly rapidly as a function
of its argument y, it is possible, by choosing t0 large enough, to consider
x(t) as practically causal. With this solution, the eye diagram has a hor-
izontal aperture that tends towards zero and thus, any imprecision about
the sampling time can lead to errors even in the absence of noise. In con-
clusion, this solution is purely theoretical and therefore has no practical
applications.
3. X t0 (f ) has a bandwidth W > 1/2T . In this case, there are many solutions
that enable the Nyquist criterion to be satisfied. Among these solutions,
the most commonly used is the raised-cosine function defined by:
X t0 (f ) = T if 0 ≤ |f | ≤ 1−α
2T
- πT .
T
2
1 + sin α
1
( 2T − |f |) if 1−α
2T
≤ |f | ≤ 1+α
2T
(2.129)
0 if |f | > 1+α
2T
or again:
X(f ) = p(t0 )X t0 (f ) exp(−j2πf t0 ) (2.130)
whose waveform is:
sin π(t−t
T
0)
cos πα(t−t
T
0)
x(t) = p(t0 ) π(t−t0 ) 2 (2.131)

0)
T
1 − 4α2 (t−t
T2
Figure 2.23 – Frequency and time domain characteristics of a raised-cosine function

for different values of the roll-off factor α.
Figure 2.23 shows the frequency domain X t0 (f ) and time domain x(t) char-
acteristics of a raised-cosine function for different values of α, called the roll-off
factor.
The bandwidth of the raised-cosine function is W = (1 + α)/2T ; 0 ≤ α ≤ 1.
Function x(t) is again non-causal in the strict sense but the more the roll-off
factor increases, the greater this function decreases. Thus, by choosing t0 large
enough, implementing a raised cosine becomes possible. Figure 2.24 plots the
eye diagrams obtained with raised-cosine functions for different values of roll-off
factor.
All the plots of x(t) pass through a single point at the sampling time t0 + nT ,
whatever the value of the roll-off factor. Note that the larger the roll-off factor,
the greater the horizontal aperture of the eye diagram. For α = 1, the aperture
of the eye is maximum and equal to T ; the sensitivity to any imprecision about
the sampling time is thus minimum.
Figure 2.24 – Eye diagrams for modulations with (2-PSK or 4-PSK) binary symbols
for different values of roll-off factor α (0.2, 0.5, 0.8).
Having determined the global spectrum X(f ) that the transmission chain
must satisfy in order to guarantee the absence of ISI, we will now establish the
frequency characteristics of the emission and reception filters.
Optimal distribution of filtering between transmission and reception

We have seen that the reception filter must be matched to the waveform placed
at its input, that is, in our case:
gr (t) = z(t0 − t) (2.132)
where z(t) results from filtering h(t) by the transmission filter:
z(t) = h(t) ⊗ ge (t)
the frequency characteristic Gr (f ) of the reception filter being equal to:
Gr (f ) = Z ∗ (f ) exp(−j2πf t0 ) (2.133)
where ∗ represents the conjugate operator.
Of course, the global characteristic of the transmission chain must satisfy
the Nyquist criterion, which is translated by:
Z(f )Gr (f ) = p(t0 )CSα (f ) exp(−j2πf t0 ) (2.134)
where CSα (f ) = X t0 (f ) is the raised-cosine spectrum of roll-off factor α. In the
previous relation we considered that the channel transmits the signal placed at
its input in its entirety.
Expressing function Z ∗ (f ) from the previous relation,
p(t0 )CSα (f )
Z ∗ (f ) = exp(j2πf t0 ) (2.135)
G∗r (f )
then replacing Z ∗ (f ) by its expression in relation (2.133), we obtain the magni-
tude spectrum of the reception filter:
+ +
|Gr (f )| = p(t0 ) CSα (f ) (2.136)
The magnitude spectrum of the emission filter is obtained without difficulty

from:
|Z(f )|
|Ge (f )| = (2.137)
|H(f )|
Using relation (2.135), we can determine the module of |Z(f )|.
+ +
|Z(f )| = p(t0 ) CSα (f ) (2.138)
Replacing |Z(f )| by its value in relation (2.137), we obtain the magnitude spec-
trum of the reception filter:
+ +
p(t0 ) CSα (f )
|Ge (f )| = (2.139)
|H(f )|
We have obtained the magnitude spectrum of the emission and reception filters.
These filters are therefore defined to within one arbitrary phase. Distributing
time t0 between the emission and reception filters, we obtain:
+ +
Gr (f ) = p(t0 ) CSα (f ) exp(−j2πf t1 )
√ √ (2.140)
p(t0 ) CSα (f )
Ge (f ) = H(f ) exp(−j2πf (t0 − t1 ))
where t1 is a time lower than t0 that can be chosen arbitrarily.

The characteristics of filters Ge (f ) and Gr (f ) show that Nyquist filtering
must be equidistributed between emission and reception. The term H(f ) that
appears at the denominator of the emission filter means that the output of this
filter no longer depends on the waveform used by the modulator. Spectrum
H(f ), a Fourier transform of a rectangular pulse shape of width T , has spectral
zeros crossing at frequencies n/T which, strictly speaking, makes it impossible
to implement filter Ge (f ). However, filter Ge (f ) must be implemented only in
the [−(1 + α)/2T ; (1 + α)/2T ] band where H(f ) has no spectral zero-crossing.
To finish, let us determine the PSD of the modulated signal at the output of the
emission filter and check that it has bandwidth W = (1 + α)/2T .
The PSD of the baseband modulated signal at the input of the emission filter
is equal to:
2
(M − 1) 2 sin πf T
γse (f ) = 2 A T (2.141)
3 πf T
The PSD of the baseband modulated signal at the output of the emission filter
has the expression:
2
γe (f ) = γse (f ) |Ge (f )|
Replacing |Ge (f )| by its expression, we obtain:
(M − 1) A2
γe (f ) = 2 p(t0 )CSα (f ) (2.142)
3 T
Considering the modulated signal on a frequency carrier, its PSD is given by

expression (2.13):
1 1
γ(f ) = γe (f − f0 ) + γe (f + f0 )
4 4
In conclusion, the modulated signal on a frequency carrier uses the f0 −(1+α)/2T
to f0 +(1+α)/2T band, that is, a bandwidth 2W = (1+α)/T or again (1+α)Rm
where Rm is the symbol rate. The spectral efficiency of the M-QAM modulation
expressed in bit/s/Hz is then equal to:
D Rm log2 (M ) log2 (M )
η= = = (2.143)
2W (1 + α)Rm (1 + α)
where D is the bit rate of the transmission expressed in bit/s.
The spectral efficiency increases as a function of the number of states M
of the modulation but the performance of the modulation, in terms of error
probability, decreases as a function of this parameter M .
2.3.4 Expression of the error probability in presence of

Nyquist filtering
Let us determine the bit error probability provided by the source by considering,
for example, a 4-PSK modulation. In this case, each symbol ai (respectively bi )
transmits a bit di every T seconds. The error probability on the data di is
therefore identical to the error probability on the symbols ai or bi . Let us
calculate, for example, the error probability on symbol ai .
The output of the reception filter of the in-phase component at sampling
time t0 + nT is equal to:
yc (t0 + nT ) = Aan p(t0 ) + bc (t0 + nT ) (2.144)
where bc (t0 + nT ) is the real part of the noise b (t0 + nT ).
Assuming the data dn iid, the error probability on the symbols an has the
expression:
P ean = 1
2
Pr {yc (t0 + nT ) > 0|an = −1}
(2.145)
1
2
Pr {yc (t0 + nT ) < 0|an = +1}
Since yc (t0 + nT ) is Gaussian, with mean Aan p(t0 ) and variance σb2c , the error
probability P ean is equal to:
1 Ap(t0 )
P ean = erfc √ (2.146)
2 σbc 2
The variance σb2c of the noise bc (t0 + nT ) is equal to:

+∞
+∞
2
σb2c = N0 |Gr (f )| df = N0 p(t0 )CSα (f )df = N0 p(t0 ) (2.147)
−∞ −∞
Let us introduce the power transmitted at the output of the emission filter:

+∞
+∞
+∞
1 1 1
P = γe (f − f0 )df + γe (f + f0 )df = γe (f )df (2.148)
4 4 2
−∞ −∞ −∞
Replacing γe (f ) by its expression, we obtain:

+∞
A2 A2
P = p(t0 )CSα (f )df = p(t0 ) (2.149)
T T
−∞
Using expressions (2.147) and (2.149), the error probability is equal to:

1 PT
P ean = erfc (2.150)
2 2N0
The energy Eb used to transmit an information bit dn is:
Eb = P Tb (2.151)
where Tb is the inverse of the bit rate of the transmission.

For a 4-PSK modulation, T = 2Tb and the bit error probability dn is finally:

1 Eb
P edn = erfc (2.152)
2 N0
The error probability in the presence of Nyquist filtering for a 4-PSK modulation
is identical to that obtained for a transmission on an infinite-bandwidth channel.
This result is also true for the other linear modulations of the M-ASK, M-PSK
and M-QAM type.
To conclude this section, we can say that filtering according to the Nyquist
criterion of a linear modulation makes it possible to reduce the bandwidth nec-
essary for its transmission to (1 + α)Rm , where Rm is the symbol rate. This
filtering does not degrade the performance of the modulation, that is, it leads to
the same bit error probability as that of a transmission on an infinite bandwidth
channel.
2.4 Transmission on fading channels

2.4.1 Characterization of a fading channel
Let us consider a transmission over a multipath channel where the transmitter,
which is mobile compared to the receiver, provides a non-modulated signal s(t) =
A exp(j2πf t) with amplitude A and frequency f . Signal s(t) propagates by being
reflected on different obstacles and, thus, the receiver receives M copies of the
signal transmitted, each copy being affected by an attenuation ρn (t), with delay
τn (t) and a Doppler frequency shift fnd (t). The attenuations, delays and Doppler
frequencies are functions of time in order to take into account the time-varying
channel. To simplify the notations, in the following we will omit variable t for
the attenuations, the delays and the Doppler frequencies.
Let r(t) be the response of the transmission channel to the signal s(t), which
is generally written in the form:

M

r(t) = ρn exp j2π(fnd + f )(t − τn ) (2.153)
n=1
Making s(t) appear, the received signal can again be written:

M

r(t) = ρn exp j2π(fnd t − (fnd + f )τn )) s(t) (2.154)
n=1
and thus the frequency response of the transmission channel is defined by:

M

c(f, t) = ρn exp −j2π(f τn − fnd t + fnd τn ) (2.155)
n=1
The multipath channel is generally frequency selective, that is, it does not trans-
mit all the frequency components of the signal placed at its input in the same
way, certain components being more attenuated than others. The channel will
therefore create distortions of the transmitted signal. In addition, their evolution
over time can be more or less rapid.
To illustrate the frequency selectivity of a multipath channel, we have plotted
in Figure 2.25 the power spectrum of the frequency response of this channel for
M = 2, in the absence of a Doppler frequency shift (fnd = 0) and fixing τ1 to
zero.
2
|c(f )| = ρ21 (1 + α cos 2πf τ2 )2 + α2 sin2 2πf τ2 (2.156)
with α = ρ2 /ρ1 .
Two parameters are now introduced: coherence bandwidth Bc and coherence
time tc that allow the transmission channel to be characterized in relation to
the frequency selectivity and its evolution speed.
Coherence bandwidth
There are several definitions of the coherence bandwidth but the most common
definition is:
1
Bc ≈ (2.157)
Tm
Figure 2.25 – Frequency response of a multipath channel.
where Tm is the multipath spread of the channel as a function of the delays τn

of the different paths.
Two modulated signals whose carrier frequencies are separated by a quantity
higher than the coherence band of the channel are attenuated by the latter in
a non-correlated way. Thus, at a given instant, if one of the two signals is
strongly attenuated by the channel, it is highly probable that the other will be
little affected by the channel.
When band B occupied by a modulated signal is lower than the coherence band,
the channel is frequency non-selective. It shows a flat frequency response and a
linear phase on the bandwidth of the modulated signal. The modulated signal
just undergoes attenuation and dephasing on passing through the channel. We
will illustrate this point by considering the transmission of an M-QAM signal
by a frequency non-selective channel.
Let s(t) be an M-QAM signal provided at the input of the frequency non-
selective multipath channel.
! $

s(t) = A ai h(t − iT ) cos 2πf0 t − bi h(t − iT ) sin 2πf0 t
i i
and let r(t) be the response of the channel at s(t):

M - .
r(t) = A ρn ai h(t − τn − T ) cos 2π(f0 + fnd )(t − τn )
n=1 i
- .
− bi h(t − τn − T ) sin 2π(f0 + fnd )(t − τn )
i
For a frequency non-selective channel, we have B < Bc . Noting that band B

is proportional to 1/T , this leads to T > Tm , or again T >> τn ∀ n. In the
expression of r(t), we can therefore neglect τn in front of T which gives:

M
r(t) = A ρn ai h(t − T ) cos(2πf0 t + ϕn (t))
n=1 i

− bi h(t − T ) sin(2πf0 t + ϕn (t))
i
with:
ϕn (t) = fnd t − (f0 + fnd )τn
Putting:

M
M
ac (t) = ρn cos ϕn (t) and as (t) = ρn sin ϕn (t)
n=1 n=1
and:
ac (t) as (t)
cos φ(t) = + and sin φ(t) = +
a2c (t) + a2s (t) a2c (t) + a2s (t)
signal r(t) can again be written:

r(t) = Aα(t) ai h(t − iT ) cos(2πf0 t + φ(t))
i (2.158)

− bi h(t − iT ) sin(2πf0 t + φ(t))
i
+
with α(t) = a2c (t) + a2s (t).
For a frequency non-selective multipath channel, the modulated M-QAM
signal only undergoes an attenuation α(t) and dephasing φ(t).
Modelling the attenuations ρn , the delays τn , and the Doppler frequencies fnd
by mutually independent random variables then, for large enough M and for a
given t, ac (t) and as (t) tend towards non-correlated random Gaussian variables
(central limit theorem). The attenuation α(t), for a given t, follows a Rayleigh
law and the phase φ(t) is equidistributed on [0, 2π[.

2α α2
p(α) = 2 exp − 2 α≥0 (2.159)
σα σα

with σα2 = E α2 .
The attenuation α(t) can take values much lower than unity and, in this
case, the information signal received by the receiver is very attenuated. Its level
is then comparable to, if not lower than, that of the noise. We say that the
transmission channel shows deep Rayleigh fading.
If band B occupied by the modulated signal is higher than the coherence
band, the channel is frequency selective. Its frequency response, on band B, is
no longer flat and some spectral components of the modulated signal can be very
attenuated. The channel introduces a distortion of the modulated signal which
results in the phenomenon of Intersymbol Interference (ISI). In the presence of
ISI, the signal at a sampling time is a function of the symbol of the modulated
signal at this time but also of the symbols prior to and after this time. ISI
appears as noise that is added to the additive white Gaussian noise and, of
course, degrades the performance of the transmission.
Coherence time
The coherence time tc of a fading channel is defined by:
1
tc ≈ (2.160)
Bd
d
where Bd is the Doppler band of the channel that is well approximated by fmax
with:
d
fmax = Maxfnd (2.161)
The coherence time of the channel is a measure of its evolution speed over time.
If tc is much higher than the symbol period T of the modulated signal, the
channel is said to be slow-fading. For a frequency non-selective slow-fading
channel, attenuation α(t) and phase φ(t) are practically constant over one or
more symbol periods T .
A channel is frequency non-selective slow-fading if it satisfies the following
condition:
T M Bd < 1 (2.162)
2.4.2 Transmission on non-frequency-selective slow-fading

channels
Performance on a Rayleigh channel
For this channel, the modulated signal undergoes an attenuation α(t) and a
random dephasing φ(t) of constant realizations over a duration higher than or
equal to T . Considering a coherent receiver, the error probability per binary
data, conditionally to a realization α of the attenuation α(t), is equal to:
,
1 α2 Eb
2-PSK or 4-PSK modulation P eb(α) = erfc (2.163)
2 N0
,
1 α2 Eb
2-FSK modulation P eb(α) = erfc (2.164)
2 2N0
We obtain the error probability P eb by averaging P eb(α) over the different

realizations of α(t).
∞
P eb = P eb(α)p(α)dα (2.165)
0
where p(α) is the probability density of α.

Taking into account the fact that, for a given t, α(t) is a Rayleigh random
variable with probability density

α α2
p(α) = 2 exp − 2 α≥0
σα 2σα
the probabilities P eb have the expressions:

" , #
1 Ēb /N0
2-PSK or 4-PSK modulation P eb = 1− (2.166)
2 1 + Ēb /N0
" , #
1 Ēb /N0
2-FSK modulation P eb = 1− (2.167)
2 2 + Ēb /N0
where Ēb is the average energy per transmitted bit :

/ 2
0
2 A Tb
Ēb = E α = A2 Tb σα2 (2.168)
2
For high Ēb /N0 , the error probabilities can be approximated by:
1
2-PSK or 4-PSK P eb ≈ (2.169)
4Ē/N0
1
2-FSK P eb ≈ (2.170)
2Ē/N0
On a Rayleigh fading channel, the performance of the different receivers is
severely degraded compared to those obtained on a Gaussian channel (with
identical Eb /N 0 at the input). Indeed, on a Gaussian channel, the error prob-
abilities P eb decrease exponentially as a function of the signal to noise ratio
Eb /N0 whereas on a Rayleigh fading channel, the decrease in the probability
P eb is proportional to the inverse of the average signal to noise ratio, Ēb /N0 .
To improve the performance on a Rayleigh fading channel, we use two tech-
niques, which we can combine, diversity and, of course, channel coding (which
is, in fact, diversity of information).
Performance on a Rayleigh channel with diversity

Diversity involves repeating the same message (or copies coming from channel
coding) several times by using the carrier frequencies separated by a quantity
higher than or equal to the coherence band Bc of the channel. In this case,
we speak of frequency diversity. An alternative to this approach involves trans-
mitting a same message several times on a same carrier but on time intervals
separated by a quantity that is higher than or equal to the coherence time tc
of the channel. This is time diversity. Finally, we can transmit a message a
single time and use several sensors spaced typically by a few wavelengths from
the carrier of the modulated signal. In this case, we have space diversity.
Let us assume that we use a 2-PSK modulation to transmit the information
message and a diversity of order L. On the time interval [iT, (i + 1)T [ and
considering a coherent reception, after demodulation we have L observations of
the form: +
rin = αni Eb cos ϕi + bni n = 1, 2, · · · , L (2.171)
where αni is a Rayleigh attenuation, ϕi the phase (0 or π) carrying the informa-
tion to transmit and bni a white Gaussian noise, with zero mean and variance
equal to N0 /2. The L attenuations αni are mutually independent as well as the
L terms of noise bni . These L attenuations can be seen as L independent sub-
channels, also called diversity branches. Eb is thus the energy used to transmit
one bit per diversity branch.
To take a decision in the presence of diversity, we construct the variable Zi
in the following way:
L
Zi = rin · αni
n=1
The bit error probability P eb in presence of diversity is then equal to:

1 1
P eb = Pr (Zi > 0|ϕi = π) + Pr (Zi < 0|ϕi = 0) (2.172)
2 2
Conditionally to one realization of the L attenuations αni , the decision variable
Zi is Gaussian with mean:
√
L
E {Zi } = Eb (αni )2 if ϕi = 0
n=1
(2.173)
√
L
E {Zi } = − Eb (αni )2 if ϕi = π
n=1
and variance:
N0 n 2
L
2
σZ = (α ) (2.174)
2 n=1 i
Under this hypothesis, the error probability P eb(ρ) is equal to :
1 √
P eb(ρ) = erfc ρ (2.175)
2
with:
Eb n 2
L
ρ= (α )
N0 n=1 i
By averaging the probability P eb(ρ) on the different realizations of the random

variable ρ, we obtain the bit error probability in presence of diversity of order
L:
∞
P eb = P eb(ρ)p(ρ)dρ
0
The random variable ρ follows a χ2 law with probability density:

1 ρ
p(ρ) = ρL−1 exp − (2.176)
(L − 1)!m̄ L m̄
where m̄ is equal to:

Eb n 2
m̄ = E (αj )
N0
After integration, the bit error probability is equal to:

1−η
L L−1
L−1+n

1+η
n
P eb = (2.177)
2 n 2
n=0
with: ,
Ēb /LN0 L−1+n (L − 1 + n)!
η= and =
1 + Ēb /LN0 n n!(L − 1)!
where Ēb is the average total energy used to transmit an information bit (Ēb =
LEb ).
For a high signal to noise ratio, an approximation of the bit error probability
P eb is given by:
L
2L − 1 1 Ēb
P eb ≈ pour >> 1 (2.178)
L 4Ēb /LN0 LN0
In the presence of diversity, the bit error probability P eb decreases following the
inverse of the signal to noise ratio to the power of L.
Figure 2.26 – Performance of a) 2-PSK and b) 2-FSK modulations in the presence of

diversity.
For an 2-FSK modulation, calculating the bit error probability in the pres-
ence of coherent reception is similar to that of 2-PSK modulation. We obtain
the following result:

1−η
L L−1
L−1+n

1+η
n
P eb = (2.179)
2 n 2
n=0
with this time: ,

Ēb /LN0
η=
2 + Ēb /LN0
With a high signal to noise ratio, a good approximation of the error probability
P eb is given by:
L
2L − 1 1 Ēb
P eb ≈ for >> 1 (2.180)
L 2Ēb /LN0 LN0
Note that diversity like the type presented here is a form of coding that uses
a repetition code and weighted decoding at reception. Figure 2.26 shows the
performance of 2-PSK and 2-FSK modulations in the presence of diversity.
Transmission on a slow-fading frequency-selective channel

Different transmission strategies are possible. We can use a waveform at emission
that is only slightly, or not at all, sensitive to the selectivity of the channel, or
we can correct the effects of the selectivity of the channel at reception.
Multicarrier transmission
Multicarrier transmission uses a multiplex of orthogonal carriers that are dig-
itally phase modulated (M-PSK), frequency modulated (M-FSK) or amplitude
and phase modulated (M-QAM). This waveform, called a "parallel" waveform
and known as Orthogonal Frequency Division Multiplexing (OFDM), enables us
to avoid the frequency selectivity of transmission channels.
We have seen that a channel is frequency selective for a modulated signal
with bandwidth B if its coherence bandwidth Bc ≈ 1/TM is lower than B. Let
us recall, on the other hand, that the bandwidth B of a digitally modulated
signal is proportional to its symbol rate 1/T . To build a type of waveform that
is not sensitive to the frequency selectivity of the channel, we can proceed in the
following way.
Let us divide the bit rate D = 1/Tb to be transmitted into N sub-rates
D = 1/N Tb , each elementary rate feeding a modulator with M states, with
carrier frequency fi . The symbol rate of the modulated carriers is then equal
to R = 1/T with T = N Tb log2 (M ). Choosing N large enough, the symbol rate
of the modulated carriers can become as low as required, and the bandwidth B
of a modulated carrier then becomes far lower than the coherence bandwidth
Bc of the channel. Proceeding thus, the channel is frequency non-selective for
the modulated carriers of the multiplex. For a multipath channel each mod-
ulated carrier is weighted by an attenuation that follows a Rayleigh law and
with a dephasing equidistributed on [0, 2π[. Of course, not all the modulated
carriers are affected in the same way at the same time by the channel; some
are strongly attenuated whereas others are less so. The performance in terms
of error probability per modulated carrier is that of a non-frequency-selective
slow-fading Rayleigh channel.
To avoid large packets of errors at reception, we can make sure that informa-
tion data that follow each other are transmitted by carriers affected differently
by the channel, that is, spaced by a quantity at least equal to the coherence band
Bc of the channel. We can also transmit these data by using the same carrier but
at a time separated by a quantity at least equal to the coherence time tc of the
channel. This way of proceeding amounts to performing frequency interleaving
combined with time interleaving.
Implementing multicarrier transmissions
Considering an M-QAM modulation on each carrier, the OFDM signal has

the expression:

+∞ N −1
s(t) = A
{cn,i h(t − iT ) exp(j2πfn t)} (2.181)
i=−∞ n=0
where cn,i = an,i + jbn,i is a complex modulation symbol, h(t) a unit ampli-
tude rectangular pulse shape of width T , and N is the number of carriers with
frequency fn .
Considering time interval [iT, (i + 1)T [, the signal s(t) is equal to:

N −1
s(t) = A
{cn,i exp(j2πfn t)} ∀ t ∈ [iT, (i + 1)T [ (2.182)
n=0
The implementation of an OFDM signal requires N M-QAM modulators with

carrier frequency fn to be realized. We can show that these N modulators can
be realized from an inverse discrete Fourier transform, which allows a reasonable
complexity of implementation.
Considering N orthogonal carriers, the frequencies fn must be separated by
at least 1/T .
n
fn = n = 0, 1, · · · , (N − 1) (2.183)
T
The power spectral density γOFDM (f ) of an OFDM signal in baseband is pro-
portional to:
N−1 2
sin π(f − n/T )T
γOFDM (f ) ∝ (2.184)
n=0
π(f − n/T )T
which gives a flat spectrum in the frequency band B = (N − 1)/T .
The signal s(t) can be sampled at frequency fe on condition that fe satisfies
the Nyquist criterion, that is:
2(N − 1)
fe ≥ (2.185)
T
We can choose fe = 2N/T , and thus signal s(t) sampled at time lTe with
Te = 1/fe is equal to:

N −1 / 0
nl
sl = s(lTe ) = A
cn,i exp j2π l = 0, 1, · · · , (2N − 1) (2.186)
n=0
2N
In this form, sl is not an inverse discrete Fourier transform. Introducing virtual

symbols in the following way:
c2N −n, i = c∗n, i n = 1, · · · , (N − 1) cN,i = {c0, i } (2.187)
and replacing c0,i by

{c0,i }, we can show that the sampled signal sl can effec-
tively be written in the form of an inverse discrete Fourier transform:

2N −1
nl
sl = A cn, i exp j2π (2.188)
n=0
2N
On each interval of duration T we perform an inverse discrete Fourier transform

on 2N points that enable the N M-QAM baseband signals to be obtained.
A digital-analogue converter followed by a frequency transposition enables the
modulated signals to be obtained on a carrier frequency.
At reception, after amplification and transposition of the OFDM signal into
baseband, the decoding of the modulation symbols cn,i = an,i + jbn,i can also
be performed by a discrete Fourier transform.
We have seen that the duration T of the modulation symbols increases with
N and, for large values of N , can become comparable to the coherence time tc of
the transmission channel. In this case, the hypothesis of a slow-fading channel is
no longer satisfied. There are therefore limits to the choice of parameter N . For
a multipath channel, if the choice of N does not enable B < Bc to be obtained,
the channel remains frequency-selective in relation to the modulated carriers,
and intersymbol interference appears.
To avoid residual intersymbol interference without increasing N , we can use
the guard interval principle. For a multipath channel, the propagation paths
are received at the receiver with delays τn . Calling τMax the largest of these
delays, guard interval Δ must satisfy the following inequality:
Δ ≥ τMax (2.189)
We put:
T = ts + Δ
In the presence of a guard interval, the modulation symbols are still of duration T
but the discrete Fourier transform at reception is realized on the time intervals
[iT + Δ, (i + 1)T [. Proceeding thus, we can check that on this time interval
only the modulation symbol transmitted between iT and (i + 1)T is taken into
account for the decoding: there is therefore no intersymbol interference.
The introduction of a guard interval has two consequences. The first is that
only a part of the energy transmitted on emission is exploited on reception.
Indeed, we transmit each modulation symbol over a duration T and we recover
this same symbol from an observation of duration ts = T −Δ. The loss, expressed
in dB, is therefore equal to:

1
10 log (2.190)
1 + Δ/ts
The second consequence is that the orthogonality of the carriers must be ensured
so as to be able to separate these carriers at reception, that is, separating them
by a quantity 1/ts . The band of frequencies occupied by the signal OFDM with
a guard interval is therefore:
N −1
B= (2.191)
ts
that is, an expansion of the bandwidth compared to a system without a guard
interval of 1 + Δ/ts .
In the presence of a guard interval, we should therefore choose Δ so as
to minimize the degradations of the signal to noise ratio and of the spectral
efficiency, that is, choose the smallest possible Δ compared to duration ts .
Transmission with equalization at reception

In the presence of a frequency selective channel, we can use a linear single-
carrier (M-PSK, M-QAM) modulation at emission and correct the ISI created
by the channel by an equalizer. Some equalizer architectures are presented in
Chapter 11.
If we compare the OFDM approach and single-carrier transmission with
equalization, in terms of complexity of implementation, the difficulty of real-
izing an OFDM transmission lies in the modulator whereas with equalization,
it lies in the receiver.
Chapter 3
Theoretical limits
The recent invention of turbo codes and the rediscovery of LDPC codes have
brought back into favour the theoretical limits of transmission which were re-
puted to be inaccessible until now. This chapter provides the conceptual bases
necessary to understand and compute these limits, in particular those that cor-
respond to real transmission situations with messages of finite length and binary
modulations.
3.1 Information theory

3.1.1 Transmission channel
A channel is any environment where symbols can be propagated (telecommuni-
cations) or recorded (mass memories). For example, the symbols 0 and 1 of the
binary alphabet can be represented by the polarity of a voltage applied to one
end of a pair of conducting wires, stipulating for example that +V corresponds
to 1 and −V to 0. Then, the polarity measure at the other end will show which
binary symbol was emitted. At the emitter side, the polarity is changed at reg-
ularly spaced intervals to represent the bits of a message and will enable this
message to be reconstituted at the receiver side. This scheme is far too simple
to illustrate modern telecommunications systems but, generally, it is the sign
of a real physical value that represents a binary symbol at the output of the
channel. Usually, a binary symbol is represented by a certain waveform and the
operation that associates a sequence of waveforms with the sequence of symbols
of the message is the modulation. Modulation was the subject of the previous
chapter.
We consider a situation where the channel is not very reliable, that is, where
the observation at the receiving end does not enable the bit really emitted to
be identified with certitude because an interference value, noise, independent
of the emitted message and random, is added to the useful value (spurious
effects of attenuation can also be added, like on the Rayleigh channel). Thermal
noise is well represented by a Gaussian random process. When demodulation
is performed in an optimal way, it results in a random Gaussian variable whose
sign represents the best hypothesis concerning the binary symbol emitted. The
channel is then characterized by its signal to noise ratio, defined as the ratio
of the power of the useful signal to that of the perturbing noise. For a given
signal to noise ratio, the decisions taken on the binary symbols emitted are
assigned a constant error probability, which leads to the simple model of the
binary symmetric channel.
3.1.2 An example: the binary symmetric channel

This is the simplest channel model, and it has already been mentioned in Sec-
tion 1.3. A channel can generally be described by the probabilities of the tran-
sition of the symbols that are input, towards the symbols that are output. A
binary symmetric channel is thus represented in Figure 3.1. This channel is
memoryless, in the sense that it operates separately on the successive input
bits. Its input symbol X and its output symbol Y are both binary. If X = 0
(respectively X = 1), there exists a probability p that Y = 1 (resp. Y = 0). p
is called the error probability of the channel.
Figure 3.1 – Binary symmetric channel with error probability p. The transition prob-
abilities of an input symbol towards an output symbol are equal two by two.
Another description of the same channel can be given in the following way:
let E be a binary random variable taking value 1 with a probability p < 1/2 and
value 0 with the probability 1 − p. The hypothesis that p < 1/2 does not restrict
the generality of the model because changing the arbitrary signs 0 and 1 leads
to replacing an initial error probability p > 1/2 by 1 − p < 1/2. The behaviour
of the channel can be described by the algebraic expression Y = X ⊕ E, where
X and Y are the binary variables at the input and at the output of the channel
respectively, E a binary error variable, and ⊕ represents the modulo 2 addition.
Configurations of errors on the binary symmetric channel

Let us now suppose that we no longer consider a particular single symbol,
but a set of n symbols (consecutive or not) making up a word, denoted
3. Theoretical limits 85
x = (x1 x2 . . . xn ). The operation of the channel is described by the vector

addition modulo 2 of x and of an error vector e = (e1 e2 . . . en ):
y=x⊕e (3.1)
with y = (y1 y2 . . . yn ), the notation ⊕ now designating the modulo 2 addition of

two words, symbol to symbol. The hypothesis that the binary symmetric chan-
nel is memoryless means that the random variables ei , i = 1...n, are mutually
independent. The number of configurations of possible errors is 2n, and their
probability, for an error probability p of the given channel, depends only on the
weight w(e) of the configuration of errors e realized, defined as the number of 1
symbols that it contains. Thus, the probability of the appearance of a particular
configuration of errors of weight w(e) affecting a word of length n equals:
Pe = pw(e) (1 − p)n−w(e) (3.2)
As p was assumed to be lower than 1/2, probability Pe is a decreasing function

of the weight w(e), whatever n.
The probability of the appearance of any configuration of errors of weight w
equals:
n
Pw = pw (1 − p)n−w (3.3)
w
The weight of the error configurations thus follows a Bernoulli distribution whose
mathematical expectation (or mean) is np and the variance (the expectation of
the square of the difference between its effective value and its mean) is np(1 − p).
Mutual information and capacity of the binary symmetric channel

To characterize a channel, we first have to measure the quantity of information
that a symbol Y leaving a channel provides, on average, about the corresponding
symbol that enters, X. This value called mutual information and whose unit is
the Shannon (Sh), is defined for a discrete input and output channel by:
Pr(X |Y ) Pr(X, Y )
I(X; Y ) = Pr(X, Y ) log2 = Pr(X, Y ) log2
Pr(X) Pr(X) Pr(Y )
X Y X Y
(3.4)
In this expression, the sums are extended to all the discrete values that X and
Y can take in a given alphabet. Pr(X, Y ) denote the joint probability of X
and Y , Pr(X|Y ) the probability of X conditionally to Y (that is, when Y is
given), Pr(X) and Pr(Y ) are the marginal probabilities of X and Y (that is, of
of the variables X and Ywhatever the value taken by the other: Pr(X) =
each
Y Pr(X, Y ) and Pr(Y ) = X Pr(X, Y )). These different probabilities are
linked according to Bayes’ law:
Pr(X, Y ) = Pr(X |Y ) Pr(Y ) = Pr(Y |X ) Pr(X) (3.5)

The first equality in (3.4) defined I(X; Y ) as the logarithmic increase of the
probability of X that results on average from the data Y , that is, the average
quantity of information that the knowledge of Y provides about that of X. The
second equality in (3.4), deduced from the first using (3.5), shows that this value
is symmetric in X and in Y . The quantity of information that Y provides about
X is therefore equal to what X provides about Y , which justifies the name of
mutual information.
Mutual information is not sufficient to characterize the channel because the
former also depends on the entropy of the source, that is, the quantity of in-
formation that it produces on average per emitted symbol. Entropy, that is,
in practice the average number of bits necessary to represent each symbol, is
defined by:
H(X) = Pr (X) log2 (Pr (X))
X
The capacity of a channel is defined as the maximum of the mutual informa-

tion of its input and output random variables with respect to all the possible
probability distributions of the input variables, and it could be demonstrated
that this maximum is reached for a symmetric memoryless channel when the
input variable of the channel, X, has equiprobable values (which also causes the
entropy of the source to be maximum). For example, for the binary symmetric
channel, the capacity is given by:
C = 1 + p log2 (p) + (1 − p) log2 (1 − p) (Sh) (3.6)
This capacity is maximum for p = 0 (then it equals 1 Sh, like the entropy of the
source: the channel is then "transparent") and null for p = 1/2, which is what
we could expect since then there is total incertitude.
3.1.3 Overview of the fundamental coding theorem

The simplest code that we can imagine is the repetition code that involves
emitting information bits in the form of several identical symbols. Hard decoding
is performed according to the principle of a majority vote, and soft decoding by
the simple addition of the samples received. If the channel is Gaussian, repetition
coding provides no gain in the case of soft decoding. For example, transmitting
the same symbol twice, allocating each of them half of the available energy and
then reconstituting the emitted symbol by addition does not give a better result
than transmitting a single symbol with all the energy available. As for the
majority vote, it can only be envisaged from a triple emission and in all cases,
on a Gaussian channel this procedure degrades the budget link in relation to
the non-coded solution. It should however be noted that repeating messages is
a widespread technique, not as a procedure for error correction coding, but as a
technique for recovering packets of erroneous messages or messages lost during
transmission. This technique called ARQ (Automatic Repeat reQuest) cannot
be implemented in all systems, in particular in point to multipoint links (e.g.

television broadcasting).
The codes are ideally efficient only if their codewords are long, in the sense
that the error probability can be made arbitrarily small only if the length of
these codewords tends towards infinity. In addition, a good code must keep an
emission or coding rate R = k/n non-null when the number k of information bits
tends towards infinity. That an error-free communication is effectively possible
asymptotically for a non-null emission rate is a major result of information the-
ory, called the fundamental theorem of channel coding, which preceded attempts
to construct practical codes, thus of finite length. This theorem was a powerful
incentive in the search for ever more efficient new codes. Moreover, it presented
engineers with a challenge, insofar as the proof of the theorem was based on
random coding, whose decoding is far too complex to be envisaged in practice.
Although the mathematical proof of the fundamental theorem in its most
general form contains fairly difficult mathematics, we believe that it can be easily
understood with the help of the law of large numbers. This law simply says that
experimental realizations have frequencies, defined as the ratio of the number
of occurrences noted to the total number of attempts, which tend towards the
probabilities of the corresponding events when the number of attempts tend
towards infinity. Let us consider, for example, the game of heads and tails.
With an "honest" coin, after 10000 throws we could theoretically arrive at the
sequence consisting exclusively of all heads (or all tails), but with a probability
that is only 2−10000 ≈ 10−3010 (in comparison, one second represents about 10−18
of the time that has elapsed since the creation of the universe). In stark contrast,
the probability that the frequency of the heads (or tails) is close to the mean
1/2 (belonging for example to the interval 0,47-0,53) is in the neighbourhood
of 1. In a similar way, an error configuration with a weight close to np when
n symbols are emitted on a binary symmetric channel of error probability p is
very likely, on condition that the message sent is sufficiently long.
3.1.4 Geometrical interpretation

Consider the finite space Sn of the codewords of n bits having the minimum
Hamming distance d. It contains 2n elements that are said to be its points.
In geometrical terms, saying that the number of errors is close to np with high
probability means that the received word is represented by a point that, with
high propability it is very close to the surface of a hypersphere with n dimensions
in Sn , centred on the emitted word and whose radius is equal to the expected
mean number of errors np. If the minimum distance d of the code is higher than
twice this number, the point on the surface of this hypersphere is closer to the
word effectively emitted than to any other codeword and therefore identifies it
without ambiguity. The optimal decoding rule, which was presented in Chapter
1, can therefore be stated thus:
"Choose the codeword closest to the received word"
The larger n is, the smaller the probability that this rule has an erroneous
result is, and this probability tends towards 0 (assuming that p is kept constant)
when n tends towards infinity, provided that d > 2np. So d has also to tend
towards infinity.
Still in geometrical terms, the construction of the best possible code can
therefore be interpreted as involving choosing M < 2n points belonging to Sn
in such a way that they are as far away as possible from each other (note that
the inequality M < 2n implies that the code is necessarily redundant). For a
given value of the error probability p of the channel (still assumed to be binary
symmetric) it is clear that there is a limit to the number M of points that can
be placed in Sn while maintaining the distance between these points higher than
2np. Let Mmax be this number. The value
log2 (Mmax )
C = lim
n→∞ n
measures in shannons the greatest quantity of information per symbol that can
be communicated without any errors through the channel, and it happens to
coincide with the capacity of the channel defined in Section 3.1. No explicit
procedure making it possible to determine Mmax points in Sn while maintaining
the distance between these points higher than 2np is generally known, except in
a few simple, not very useful, cases.
3.1.5 Random coding

Random coding, that is, the construction of a code by randomly choosing its
elements, is a way of choosing M scattered points in the space Sn . This method is
optimal for the distribution of distances, when n tends towards infinity. Random
coding enables the points to be, on average, equally distributed in all the n
dimensions of Sn and it reaches a mean emission rate equal to the capacity of
the channel. For a code containing M codewords of length n, it means randomly
drawing each bit of a codeword independently of the others with the probability
1/2 that it is 0 or 1, the M codewords that make up the code being drawn in
the same way independently from each other. The probability of a particular
codeword c is Pc = 2−n . We thus obtain codewords whose weights follow a
Bernoulli distribution and the probability of obtaining any codeword of weight
w is given by (3.3) for p = 1/2, that is:

n
Pw = 2−n (3.7)
w
The mathematical expectation, or mean, of this weight is n/2 and its variance
equals n/4. For very large n, a good approximation of the weight distribution
of the codewords obtained by random coding is a Gaussian distribution. If
we replace w/n by the continuous random variable X, the probability that

X ∈ (x, x + dx) is pX (x)dx, where:

2n
pX (x) = exp 2n(x − 1/2)2 (3.8)
π
This function has a maximum at x = 1/2, therefore for w = n/2, and takes
symmetric decreasing values when x diverges from 1/2. It is centred around its
maximum x = 1/2 and √ the width of the region where it takes non-negligible
values decreases as 1/ n, and therefore tends towards 0 when n tends towards
infinity.
Unfortunately, decoding a code obtained by random coding is impossible in
practice since decoding a single received word would imply comparing it to all
the codewords. Since long words are necessary for good performance, the number
of codewords (2Rn ), and therefore the number of necessary combinations, is
considerable if Rn is large, which is the case in practice. This is why research
on error correcting codes has been directed towards non-random coding rules
offering the path to decoding with reasonable complexity.
No general way is known for constructing a code having Mmax codewords, for
an arbitrary value of n and a given error probability p. We know with certitude,
or we conjecture, that a small number of schemes are optimal for given values of
M and n, for a few simple channels. In the absence of a general rule for building
optimal codes, research has focused on codes satisfying a simpler criterion: that
of minimum distance, that is, the greater a code’s minimum distance, the better
it is. The pertinence of this criterion was not questioned until the end of the
1980’s. This criterion does not take into account the number of codewords at
the minimum distance from a given word (or multiplicity), whereas a large value
for this number leads to a degradation in performance. Turbo codes, which will
be examined in the following chapters, were not initially built to satisfy this
criterion. Their minimum distance can be small (at least if we compare it to the
known bounds on the largest minimum distance possible and in particular the
Gilbert-Varshamov bound which we shall see later) but their multiplicity is also
very small. These properties mean that there can be an error floor, that is, a far
less rapid decrease in the error probability of their decoding as a function of the
signal to noise ratio when the latter is large, than when it is small. This error
floor phenomenon can also be visible with LDPC codes, although the latter can
be designed on the criterion of minimum distance. Be that as it may, in the
case of turbo codes like in that of LDPC, since the finality of correction coding
is to improve communications when the channel is bad, we could say that these
codes are good when they are useful and mediocre when they are less useful.
Codes imitating random coding

A simple idea is to try to build codes "imitating" random coding, in a certain
sense. Since the performance of a code depends essentially on the distribution
of its distances, and that of a linear code on the distribution of its weights,
we can undertake to build a linear code having a weight distribution close to
that of random coding. This idea has not been much exploited directly, but we
can interpret turbo codes as being a first implementation. Before returning to
the design of coding procedures, we will make an interesting remark concerning
codes that imitate random coding.
The probability of obtaining a codeword of length n and weight w by ran-
domly drawing the bits 0 and 1 each with a probability of 1/2, independently
of each other, is given by (3.7). Drawing a codeword 2k times, we obtain an
average number of words of weight w equal to:

n
Nw,k = 2−(n−k)
w

n
Assuming that n, k and w are large, we can express approximately,
w
using the Stirling formula:

n 1 nn+1/2
≈√
w 2π ww+1/2 (n − w)n−w+1/2
The minimal weight obtained on average, that is wmin , is the largest number
such that Nwmin ,k has value 1 for the best integer approximation. The number
Nwmin ,k is therefore small. It will be sufficient for us to put it equal to a constant
λ close to 1, which it will not be necessary to detail further because it will be
eliminated from the calculation. We must therefore have:
1 nn+1/2
2−(n−k) √ wmin +1/2 n−wmin +1/2
=λ
2π wmin (n − wmin )
Taking the base 2 logarithms and ignoring the constant in relation to n, k and
wmin that tend towards infinity, we obtain:
k
1− ≈ H2 (wmin /n)
n
where H2 (·) is the binary entropy function:
H2 (x) = −x log2 (x) − (1 − x) log2 (1 − x) for 0 < x < 1

= 0 for x = 0 or x = 1
The weight wmin is the average minimal weight of a code obtained by drawing
at random. Among the set of all the linear codes thus obtained (weights and
distances therefore being merged), there is at least one whose minimum distance
d is higher than or equal to the average weight wmin , so that we have:
k
1− ≤ H2 (d/n) (3.9)
n
This is the asymptotic form of the Gilbert-Varshamov bound that links the
minimum distance d of the code having the greatest minimum distance possible,
given the parameters k and n. It is a lower bound but, in its asymptotic form,
it is very close to equality. A code whose minimum distance verifies this bound
with equality is considered to be good for the minimum distance criterion. This
shows that a code built with a weight distribution close to that of random coding
is also good for this criterion.
3.2 Theoretical limits to performance

3.2.1 Binary input and real output channel
Only the case of the binary symmetric channel, with constant error probability p,
has been considered so far. Instead of admitting a constant error probability, we
can consider that the error probability in fact varies from one symbol to another
because the noise sample that affects the received value varies randomly. Thus,
in the presence of Gaussian noise, the value leaving the optimal demodulator
is a Gaussian random variable whose sign represents the optimal decision. We
will consider the channel that has this real random variable as its output value,
that we denote a. It can be shown that this value is linked to the optimal
decision x̂, that is, to the best hypothesis concerning the emitted bit x, and to
the "instantaneous" error probability pa , according to the relation:

x̂ 1 − pa
a = − (−1) ln (3.10)
pa
which means, assuming pa lower than 1/2:

1 1
pa = = (3.11)
exp (−(−1)x̂ a) + 1 exp (|a|) + 1
We mean by instantaneous error probability the error probability pa that affects
the received symbol when the real value measured

at the output of the channel is
1−pa
a. The inequality pa < 1/2 makes ln pa positive and then the best decision
is x̂ = 1 when a is positive
and

x̂ = 0 when a is negative. In addition, the
absolute value |a| = ln pa 1−pa
is a decreasing function of the error probability
of the decision, and it therefore measures its reliability. It is null for pa = 1/2 and
tends towards infinity when error probability pa tends towards 0 (the decision
then becomes absolutely reliable). The real quantity that (3.10) defines is called
relative value or more often log likelihood ratio (LLR) of the corresponding
binary symbol.
The capacity of the channel thus defined can be calculated as the maximum
with respect to X of the mutual information I(X; Y ), defined by generalizing
(3.4) to real Y = a. This generalization is possible but the expression of the
capacity thus obtained will not be given here. We merely note that this capacity
is higher than that of the binary symmetric channel that is deduced from it by
taking a hard decision, that is, restricted to the binary symbol Y = x̂, by a
factor that increases when the signal to noise ratio of the channel decreases. It
reaches π/2 when we make this ratio tend towards 0, if the noise is Gaussian.
For a given signal to noise ratio, the binary input continuous output channel is
therefore better than the binary symmetric channel that can be deduced from
it by taking hard decisions. This channel is also simpler than the hard decision
channel, since it does not have any means to take a binary decision according
to the received real value. Taking a hard decision means losing the information
carried by the individual variations of this value, which explains that the capacity
of the soft output channel is higher.
3.2.2 Capacity of a transmission channel

Here we will consider the most general case where the input and the output of
the channel are no longer only scalar values but can be vectors whose dimension
N is a function of the modulation system. For example, we will have N = 1
for an amplitude modulation and N = 2 for a phase modulation with a 4-point
constellation. X and Y are therefore replaced by X and Y.
Capacity was introduced in Section 3.1 for a discrete input and output chan-
nel, and is defined as the maximum of the mutual information of its input and
output variables, with respect to all the possible probability distributions of the
variables. For any dimension of the signal space, the law remains:
C = max I (X; Y) (3.12)

p(X)
where I(X; Y) is the mutual information between X and Y. When the input
and the output of the channel are real values, and no longer discrete values, the
probabilities are replaced by probability densities and the sums in relation (3.4)
become integrals. For realizations x and y of the random variables X and Y,
we can write the mutual information as a function of the probabilities of x and
y:
+∞
+∞
p (y |x )
I (X; Y) = ··· p (x) p (y |x ) log2 dxdy (3.13)
p (y)
−∞ −∞
1 23 4
2N times
To determine C, we therefore have to maximize (3.13) which is valid for all types
of inputs (continuous, discrete) of any dimension N . In addition, the maximum
is reached for equiprobable inputs (see Section 3.1), for which we have:
1
M
p (y) = p (y |xi )
M i=1
where M is the number of symbols or modulation order. (3.13) can then be

written in the form:
M +∞
+∞ " M #
1 j=1 p (y |xj )
C = log2 (M ) − ··· p (y |xi ) log2 dy (3.14)
M i=1 p (y |xi )
−∞ −∞
1 23 4
N times
According to the additional information available about the transmission, such

as the type of noise on the channel, possible fading, the type of input and output
(continuous, discrete) and the modulation used, (3.14) can be particularized.
Shannon limit of a band-limited continuous input and output Gaus-

sian channel
Consider the case of a Gaussian channel, with continuous input and output.
The Shannon bound [3.3] giving the maximum capacity C of such a channel is
reached taking at its input a white Gaussian noise of null mean and variance σ2 ,
described by independent probabilities on each dimension, that is, such that:
5
N
p(x) = p(xn )
n=1
where x = [x1 x2 . . . xN ] is the input vector and p(xn ) = N (0, σ2 ). The mutual
information is reached for equiprobable inputs, and denoting N0 /2 the variance
of the noise, (3.14) after development gives:

N 2σ 2
C= log2 1 + .
2 N0
This relation is modified to make the mean energy Eb of each of the bits and
Eb
consequently the signal to noise ratio N 0
. For N=2, we have:

Eb
Cb = log2 1 + R (3.15)
N0
the capacity being expressed in bit per second per Hertz and per couple di-
mension. Taking R = 1, this leads to the ratio Eb /N0 being limited by the
normalized Shannon limit, as shown in Figure 3.2.
Capacity of a discrete input Gaussian channel

The discrete input, denoted x = xi , i = 1, · · · , M , is typically the result of a
modulation performed before transmission. The inputs xi belong to a set of M
Figure 3.2 – Normalized Shannon limit.
discrete values, M being the modulation order, and have dimension N , that is
xi = [xi1 , xi2 , · · · , xiN ]. The transition probability of the Gaussian channel is:
" #
5N 5N
1 − (yn − xin )
2
p (y |xi ) = p (yn |xin ) = √ exp
n=1 n=1
πN0 N0
and we assume inputs taking

√ the M different possible values equiprobably. De-
noting dij = (xi − xj )/ N0 the vector of dimension N relative to the distance
between the symbols xi and xj , and t an integration vector of dimension N , we
obtain a simplified expression of (3.14), representing the capacity of a discrete
input Gaussian channel, for any type of modulation:
C = log2 (M )

+∞ ! $
+∞

( π)−N
√ M M
2 2
− M · · · exp − |t| log2 exp −2tdij − |dij | dt (3.16)
i=1 j=1
−∞ −∞
1 23 4
N times
C being expressed in bit/symbol. We note that dij increases when the signal
to noise ratio increases (N0 decreases) and the capacity tends towards log2 (M ).
The different possible modulations only appear in the expression of dij . The
discrete sums from 1 to M represent the possible discrete inputs. For the final
calculation, we express dij as a function of Es /N0 according to the modulation,
Es being the energy per symbol, and the capacity of the channel can be deter-
mined using a computer. Figure 3.3 gives the result of the calculation for some
PSK and QAM modulations.
Figure 3.3 – Capacity of some modulations.
Capacity of the Rayleigh channel

Let there be a Rayleigh channel whose attenuation is denoted α. For discrete
equiprobable inputs (a case similar to the Gaussian channel treated above),
(3.14) is always applicable. There are two cases, conditioned by the knowledge
of the attenuation α of the channel, or not.
In the case where α is not known a priori, we write the conditional probability
density of the Rayleigh channel in the form:
(
+∞
p(y |xi ) = p(α)p(y |xi , α )dα
−∞
(
+∞ - . 2

= √ 1
2πσ2
2α exp −α2 exp − |y−αx
2σ 2
i|
dα
0
One development of this expression means that we can explicitly write this
conditional probability density that turns out to be independent of α:
|y|2
|y|2 − 2 2
xi ye |xi | +2σ
−
2 σe 2σ2
p(y |xi ) = π |xi |2 +2σ2 +2 2 2 3/2
! " (|xi | +2σ ) #$
xi y
× 1 − 12 erfc
σ 2(|xi |2 +2σ2 )
which is sufficient to enable the capacity to be evaluated by using (3.14).

In the case where the attenuation is known, the probability density for a partic-
ular realization of α can be written:
" #
2
1 |y − αxi |
p (y |xi , α) = √ exp −
2πσ2 2σ2
The instantaneous capacity Cα for this particular realization of α is first calcu-

lated and then we have to average Cα over the set of realizations of α in order
to obtain the capacity of the channel:
M +∞
1 p (y |xi , α )
Cα = p (y |xi , α ) log2 dy
M i=1 p (y |α )
−∞

+∞
C= Cα p (α) dα = E [Cα ]
0
3.3 Practical limits to performance

In the sections above, we obtained the theoretical limits for performance which
are subject to certain hypotheses that are not realistic in practice, in particular
the transmission of infinite length data blocks. In the great majority of commu-
nication systems today, it is a sequence of data blocks that is transmitted, these
blocks being of very variable size depending on the system implemented. Logi-
cally, limited size block transmission leads to a loss of performance compared to
infinite size block transmission, because the quantity of redundant information
contained in the codewords is lower.
Another parameter used to specify the performance of real transmission sys-
tems is the packet error rate (PER), which corresponds to the proportion of
blocks of wrong data (containing at least one binary error after decoding).
What follows contains some results on the Gaussian channel, for two cases:
the binary input and continuous output Gaussian channel, and the continuous
input and output Gaussian channel. The case of the continuous input can be
assimilated to that of a modulation with an infinite number of states M . The
fewer states we have to describe the input, the less efficient the communication.
Consequently, a binary input channel gives a lower bound on the practical per-
formance of the set of modulations, whereas a continuous input channel gives
its higher limit.
3.3.1 Gaussian binary input channel

Initial work on this channel was done by Gallager [3.2]. We denote again p(y|x)
the probability of transition on the channel, and we consider information mes-
sages of size k. Assuming that a message, chosen arbitrarily and equiprobably

among 2k , is encoded then transmitted through the channel, and assuming that
we use maximum likelihood decoding, then the coding theorem provides a bound
on the mean error probability of decoding the correct codeword. In [3.2], it is
shown that it is possible to limit the PER in the following way, for whatever
value of variable ρ, 0 ≤ ρ ≤ 1:
! $1+ρ
1/1+ρ
PER ≤ (2 − 1)
k ρ
Pr(x)p(y|x) (3.17)
y x
In the case of a channel with equiprobable binary inputs, the probability of each
of the inputs is 1/2 and the vectors x and y can be treated independently in
x and y scalar values. Considering that (3.17) is valid for any ρ, in order to
obtain the closest upper bound to the PER, we must minimize the right-hand
side of (3.17) as a function of ρ. Introducing the rate R = k/n, it therefore
means minimizing for 0 ≤ ρ ≤ 1, the expression :
⎧ +∞
1+ρ ! " # " #$ ⎫k
⎨ 1 1
1
(y − 1)
2
(y + 1)
2 ⎬
2ρR √ × exp − 2 + exp − 2 dy
⎩ 2 σ 2π 2σ (1 + ρ) 2σ (1 + ρ) ⎭
−∞
The explicit value of σ is known for binary inputs (2-PSK and 4-PSK modu-
−1/2
lations): σ = (2REb /N0 ) . An exploitable expression of Gallager’s upper
bound on the PER of a binary input channel is then:
⎧ +∞ " " + ##1+ρ ⎫k
E
⎨ exp−y 2
y 4RE /N ⎬
−k Nb b 0
e 0 min 2ρR+1 √ cosh dy (3.18)
0ρ1 ⎩ π 1+ρ ⎭
0
This expression links the PER, the rate, the size k of the messages and the signal
to noise ratio Eb /N0 , for a Gaussian binary input channel. It gives an upper
bound of the PER and not an equality. This equation is not very well adapted
to all cases. In particular, simulations show that for a rate close to 1, the bound
is far too lax and does not give really useful results.
If we want to determine the penalty associated with a given packet size, we
can compare the result obtained by evaluating (3.18) with the result obtained
by computing the capacity that considers infinite size packets
3.3.2 Gaussian continuous input channel

In the case of a continuous input channel, we consider the case contrasting with
that of the binary input channel, that is, we will obtain an upper bound on
the practical limits of performance (all the modulations show performance lower
than a continuous input channel). Any modulation used will give performance
lower bounded by the limit obtained by a binary input and upper bounded by
a continuous input.
The first results were given by Shannon [3.4] and by the so-called sphere-
packing bound method which provides a lower bound on the error probability
of random codes on a Gaussian channel. We again assume maximum likelihood
decoding. A codeword of length n is a sequence of n whole numbers. Geometri-
cally, this codeword can be assimilated to a point in an n-dimensional Euclidean
space and the noise can be seen as a displacement of this point towards a neigh-
bouring point following a Gaussian distribution (see Section 3.1.4). Denoting P
the power of the emitted
√ signal, all the codewords are situated on the surface of
a sphere of radius nP .
k
√ Observing that we have a code with 2 points (codewords), each at a distance
nP from the origin in n-dimensional space, any two points are equidistant from
the origin, and consequently, the bisector of these two points (a hyperplane of
dimension n − 1) passes through the origin. Considering the set of 2k points
making up the code, all the hyperplanes pass through the origin and form pyra-
mids with the origin as the summit. The error probability, after decoding, is
2k
Pr(e) = 21k Pr(ei ), where Pr(ei ) is the probability that the point associated
i=1
with the codeword i is moved by the noise outside the corresponding pyramid.
The principle of Shannon’s sphere-packing bound involves this geometrical
vision of coding. However, it is very complex to keep the ’pyramid’ approach
and the solid angle pyramid Ωi , around the codeword i, is replaced by a cone
with the same summit and the same solid angle Ωi (Figure 3.4).
Figure 3.4 – Assimilation of a pyramid with one cone in Shannon’s so-called sphere-
packing approach.
It can be shown that the probability that the signal remains in the cone is
higher than the probability that it remains in the same solid angle pyramid.
Consequently, the error probability can be lower-bounded in the following way:
k
1 ∗
2
Pr (e) ≥ k Q (Ωi ) (3.19)
2 i=1
denoting Q∗ (Ωi ) the probability that the noise moves point i out of the solid
angle cone Ωi (therefore a decoding error is made on this point). We also observe
that, if we consider the set of codewords equally distributed on the surface of
√
the sphere of radius nP , the decoding pyramids form a partition of this same
sphere, and therefore the solid angle of this sphere Ω0 is the sum of all the solid
angles of the Ωi pyramids. We can thus replace the solid angles Ωi by the mean
solid angle Ω0 /2k .
This progression, which leads to a lower bound on the error probability for an
optimal decoding of random codes on the Gaussian channel, is called the sphere-
packing bound because it involves restricting the coding to an n-dimensional
sphere and the effects of the noise to movements on this sphere.
Mathematical simplifications give an exploitable form of the lower bound on
the packet error rate (PER):
- .
ln (PER) ≥ R k
ln (G (θi , A) sin θi ) − 12 A2 − AG (θi , A) cos θi
- .
θi ≈ arcsin 2−R
- √ . (3.20)
G (θi , A) ≈ A cos θi + A2 cos2 θi + 4 /2
+
A= 2REb /N0
These expressions link the size k of the messages, the signal to noise ratio Eb /N0
and the coding rate R. For high values of R and for block sizes k lower than a
few tens of bits, the lower bound is very far from the real PER.
Asymptotically, for block sizes tending towards infinity, the bound obtained
by (3.20) tends towards the Shannon limit for a continuous input and output
channel such as presented in Section 3.2. In the same way as for the binary
input channel, if we wish to quantify the loss caused by the transmission of finite
length packets, we must normalize the values obtained by evaluating (3.20) by
removing the Shannon limit (3.15) from them, the penalty having to be null
when the packet sizes tend towards infinity. The losses due to the transmission
of finite length packets in comparison with the transmission of a continuous flow
of data are less in the case of a continuous input channel than in the case of a
binary input channel.
3.3.3 Some examples of limits

Figure 3.5 below gives an example of penalties caused by the transmission of
blocks of size k lower than 10000 bits, in the case of continuous input and in the
case of binary input. These penalty values should be combined with the values
of capacities presented in Figure 3.3, in order to obtain the absolute limits. As
we have already mentioned, this figure is to be considered with caution for small
values of k and high PER.
The results obtained concern the Gaussian channel. It is theoretically pos-
sible to consider the case of fading channels (Rayleigh, for example) but the
computations become complicated and the results very approximate.
Figure 3.5 – Penalties in Eb/N 0 for the transmission of finite length packets for the
continuous input channel and the binary input channel as a function of size k (infor-
mation bits), for a coding rate 5/6 and different PER.
3.4 Minimum distances required

So far, we have highlighted the theoretical limits and they have been calculated
for the Gaussian channel. These limits determine boundaries, expressed in signal
to noise ratio, between transmission channels at the output of which it is possible
to correct the errors and channels for which this correction cannot be envisaged.
Assuming that codes exist whose decoding can be performed close to these limits,
the question now arises about how we can know which minimum Hamming
distances (MHD) these codes should have in order to satisfy a given objective
of error rates.
Here we present some results for the Gaussian channel and modulations
currently used: 4-PSK, 8-PSK and 16-QAM.
3.4.1 MHD required with 4-PSK modulation

With maximum likelihood decoding after transmission on a Gaussian channel,
the PER has a known upper bound, called the union bound:
" #
1 Eb
PER ≤ N (d)erfc dR (3.21)
2 N0
d≥dmin
where erfc(x) denotes the complementary error function defined by erfc (x) =
(∞ - .
√2
π
exp −t2 dt. dmin is the minimum Hamming distance of the code associ-
x
ated with the modulation considered, 2-PSK or 4-PSK in the present case. N (d)
represents the multiplicities of the code (see Section 1.5). In certain cases, these
multiplicities can be determined precisely (like for example simple convolutional
codes, Reed-Solomon codes, BCH codes, etc. . . . ), and (3.21) can easily be eval-
uated. For other codes, in particular turbo codes, it is not possible to determine
these multiplicities easily and we have to consider some realistic hypotheses in
order to get round the problem. The hypotheses that we adopt for turbo codes
and for LDPC codes are the following [3.1]:
• Hypothesis 1: Uniformity. There exists at least one codeword of weight1
dmin having an information bit di equal to "1", for any place i of the
systematic part (1 ≤ i ≤ k).
• Hypothesis 2: Unicity. There is only one codeword of weight dmin such
that di ="1".
• Hypothesis 3: Non-overlapping. The k codewords of weight dmin associated
with the k bits of information are distinct.
Using these hypotheses and limiting ourselves to the first term of the sum in
(3.21), the upper bound becomes an asymptotic approximation (low PERs):
" #
k Eb
PER ≈ erfc dmin R (3.22)
2 N0
The three hypotheses, taken separately, are more or less realistic. Hypotheses 1
and 3 are somewhat pessimistic as to the quantity of codewords at the minimum
distance. As for hypothesis 2, it is slightly optimistic. The three hypotheses to-
gether are suitable for an acceptable approximation of the multiplicity, especially
since imprecision about the value of this multiplicity does not affect the quality
of the final result. Indeed, the targeted minimum distance that we wish to deter-
mine from (3.22) appears in an exponential argument, whereas the multiplicity
is a multiplying coefficient.
It is then possible to combine (3.22) with the results obtained in Section 3.3
which provide the signal to noise ratio limits. Giving Eb /N0 the limit value
beyond which using a code is not worthwhile, we can extract from (3.22) the
MHD sufficient to reach a PER at that limit value. Given, on the one hand,
that (3.22) assumes ideal (maximum likelihood) decoding and, on the other
hand, that the theoretical limit is not reached in practice, the targeted MHD
can be slightly lower than the result of this extraction.
Figure 3.6 presents some results obtained using this method.
1 The codes being linear, distance and weight have the same meaning.
Figure 3.6 – Minimum distances required for 4-PSK modulation and a Gaussian
channel as a function of packet size, for some coding rates and P ER = 10−7 and
P ER = 10−11 .
3.4.2 MHD required with 8-PSK modulation

Here we consider an 8-PSK modulation on a Gaussian channel implemented us-
ing the principle of the "pragmatic" approach, as presented in Figure 3.7. This
approach first involves encoding the data flow in packets to produce codewords
that are then randomly permuted by the interleaver Π, with a permutation law
drawn randomly. The contents of the permuted codewords are then organized
in groups of 3 bits using a Gray coding, before being modulated in 8-PSK. The
demodulator provides the received symbols from which we extract the log like-
lihood ratios (LLRs) for all the bits of the packets. Finally, inverse interleaving
and decoding complete the transmission chain.
Figure 3.7 – Transmission scheme using the pragmatic approach.
The error probability Pe is the probability of deciding about an incorrect

codeword instead of the codeword emitted. Let Ns be the number of modulated
symbols that differ between the incorrectly decoded codeword and the codeword
emitted. Also let {φi } and {φi } (1 ≤ i ≤ Ns ) be the transmitted phase sequences
for these symbols that differ. It is possible to express Pe as a function of these
phases and of the signal to noise ratio:
⎡ ⎤

1 Es ⎣ ϕi − ϕi ⎦
Pe = erfc sin2 (3.23)
2 N0 2
i=1,Ns
where Es is the energy per symbol emitted and N0 the monolateral noise power
spectral density. It is however not possible to exploit (3.23) in the general case.
We require an additional hypothesis, which is then added to the three hypotheses
formulated in the previous section, and assume that NS is much lower than the
size of the interleaved codewords:
• Hypothesis 4: A symbol does not contain more than one opposite bit in
the correct codeword and in the wrong codeword.
This hypothesis allows the following probabilities to be expressed:
Pr {ϕi − ϕi = π/4} = 2/3; Pr {ϕi − ϕi = 3π/4} = 1/3
which means that two times out ofthree on average, the Euclidean distance
between the concurrent symbols is 2 ETs sin(π/8) and, one time out of three, is

raised to 2 ETs sin(3π/8) (Figure 3.8).
Considering the asymptotic case, that is, putting Ns = dmin , yields:
PER8-PSK,Π random ≈

- 2 .dmin d
min
dmin - 1 .j+1 (3.24)
k 3 2 erfc Es
N0 j sin
2 3π
8 + (dmin −j) sin2 π
8
j=0 j
Figure 3.8 – 8-PSK constellation with Gray coding. Es and T are the energy and the
duration of a symbol, respectively.
This relation therefore makes it possible to establish a relation between the

signal to noise ratio, the size of the information blocks and the PER. In the
same way as in Section 3.4, we can combine this result with the limits on the
signal to noise ratio to obtain the MHD targeted for a 8-PSK coded modulation
using the pragmatic approach. Figure 3.9 presents some results obtained with
this method.
3.4.3 MHD required with 16-QAM modulation

The same method as above, based on the same four hypotheses, can be applied
to the case of 16-QAM modulation with pragmatic encoding. The constellation
is a standard 16-state Gray constellation. For +
75% of the bits making up the
symbols, the minimum+Euclidean distance is 2Es /5 and for the remaining
25%, this distance is 3 2Es /5. Estimating the PER gives:
dmin d
min j
3 dmin 1 1 Es
PER ≈ k erfc (8j + dmin ) (3.25)
4 j 3 2 10N0
j=0
Like for 4-PSK and 8-PSK modulations, this relation used jointly with signal to
noise ratio limits makes it possible to obtain targeted MHD values for 16-QAM
modulation (Figure 3.10).
Some observations can be made from the results obtained in Section 3.4.
For example, in the particular case of 4-PSK modulation, for a rate R = 1/2,
size k = 4000 bits and PER of 10−11 , Figure 3.6 provides a targeted MHD of
50. From the evaluation that we can make from the Gilbert-Varshamov bound
Figure 3.9 – Minimum distances required for 8-PSK modulation and a Gaussian
channel as a function of packet size, for some coding rates and P ER = 10−7 and
P ER = 10−11 .
(relation (3.9)), random codes have a minimum distance of about 1000. There
is therefore a great difference between what ideal (random) coding can offer and
what we really need.
A second aspect concerns the dependency of the required MHD upon the
modulation used, a dependency that turns out to be minimum. Thus, a code
Figure 3.10 – Minimum distances required for 16-QAM modulation and a Gaussian
channel as a function of the packet size, for some coding rates and P ER = 10−7 and
P ER = 10−11 .
having a minimum distance sufficient to reach the channel capacity with 4-

PSK modulation will also satisfy specifications with the other modulations, for
a certain size of message (larger than 1000 bits for R = 1/2) or for longer
messages (over 5000 bits) if the rate is higher.
Bibliography
[3.1] C. Berrou, E. Maury, and H. Gonzalez. Which minimum hamming distance
do we really need? In Proceedings of the 3rd International Symposium on
Turbo Codes & related topics (ISTC 2003), pages 141–148, Brest, France,
Sept. 2003.
[3.2] R.E. Gallager. Information Theory and Reliable Communications. John
Wiley & Sons, 1968.
[3.3] C.E. Shannon. A mathematical theory of communication. Bell Systems

Technical Journal, 27, July-Oct. 1948.
[3.4] C.E. Shannon. Probability of error for optimal codes in a gaussian channel.
Bell Systems Technical Journal, 38, 1959.
Chapter 4
Block codes
Block coding involves associating with a data block d of k symbols coming from
the information source, a block c, called the codeword, of n symbols with n ≥ k.
The (n − k) is the amount of redundancy introduced by the code. Knowledge of
the coding rule at reception enables errors to be detected and corrected, under
certain conditions. The ratio k/n is called the coding rate of the code.
The message symbols of the information d and of the codeword c take their
values in a finite field Fq with q elements, called a Galois field, whose main
properties are given in the appendix to this chapter. We shall see that for most
codes, the symbols are binary and take their value in the field F2 with two
elements (0 and 1). This field is the smallest Galois field.
The elementary addition and multiplication operations in field F2 are re-
sumed in Table 4.1.
a b a + b ab
0 0 0 0
0 1 1 0
1 0 1 0
1 1 0 1
Table 4.1 – Addition and multiplication in the Galois field F2
A block code of length n is an application g of the set Fkq towards the set
Fnq that associates a codeword c with any block of data d.
g : Fkq → Fnq
d → c = g(d)
The set of q k codewords generally constitutes a very reduced subset of Fnq .

A block code with parameters (n, k), that we denote C(n, k), is linear if the
codewords are a vector subspace of Fnq , that is, if g is a linear application. A
direct consequence of linearity is that the sum of two codewords is a codeword,
and that the null word made up of n symbols at zero is always a codeword.
We will now consider linear block codes with binary symbols. Linear block
codes with non binary symbols will be addressed later.
4.1 Block codes with binary symbols

In the case of a binary block code, the elements of d and c have values in F2 .
As g is a linear application, we will be able to describe the coding operation
simply as the result of the multiplication of a vector of k symbols representing
the data to be coded by a matrix representative of the code considered, called
a code generator matrix.
4.1.1 Generator matrix of a binary block code

Let us denote d = [d0 · · · dj · · · dk−1 ] and c = [c0 · · · cj · · · cn−1 ] the data-
word and the associated codeword. Expressing the vector d from a base
(e0 , . . ., ej , . . ., ek−1 ) of Fk2 , we can write:

k−1
d= dj ej (4.1)
j=0
Taking into account the fact that application g is linear, the word c associated
with d is equal to:

k−1
c = g(d) = dj g(ej ) (4.2)
j=0
Expressing the vector g(ej ) from a base (e 0 , · · · , e l , · · · , e n−1 ) of Fn2 , we ob-
tain:

n−1
g(ej ) = gjl e l (4.3)
l=0
The vectors g(ej ) = gj = (gj0 · · · gjl · · · gj,n−1 ), 0 ≤ j ≤ k − 1 represent the k

rows of matrix G associated with the linear application g.
⎡ ⎤ ⎡ ⎤
g0 g0,0 ··· g0,l ··· g0,n−1
⎢ .. ⎥ ⎢ .. .. .. .. .. ⎥
⎢ . ⎥ ⎢ . . . . . ⎥
⎢ ⎥ ⎢ ⎥
G=⎢ g ⎥ = ⎢ gj,0 · · · g · · · g ⎥ (4.4)
⎢ j ⎥ ⎢ j,l j,n−1 ⎥
⎢ . ⎥ ⎢ . . . . . ⎥
⎣ .. ⎦ ⎣ .. .. .. .. .. ⎦
gk−1 gk−1,0 · · · gk−1,l · · · gk−1,n−1
4. Block codes 111
Matrix G with k rows and n columns, having its elements gjl ∈ F2 is called
a generator matrix of the code C(n, k). It associates the codeword c with the
block of data d by the matrix relation:
c = dG (4.5)
The generator matrix of a block code is not unique. Indeed, by permuting

the vectors of the base (e 0 , . . . , e l , . . . , e n−1 ) or of the base (e0 , . . . , ej , . . . ,
ek−1 ), we obtain a new generator matrix G whose columns or rows have also
been permuted. Of course, the permutation of the columns or the rows of the
generator matrix always produces the same set of codewords; what changes is
the association between the codewords and the k-uplets of data.
Note that the rows of the generator matrix of a linear block code are in-
dependent codewords, and that they make up a base of the vector subspace
generated by the code. The generator matrix of a linear block code is therefore
of rank k. A direct consequence is that any family made up of k independent
codewords can be used to define a generator matrix of the code considered.
Example 4.1
Let us consider a linear block code called the parity check code denoted
C(n, k), with k = 2 and n = k + 1 = 3 (for a parity check code, the sum of the
symbols of a codeword is equal to zero). We have four codewords:
Dataword Codeword
00 000
01 011
10 101
11 110
To write a generator matrix of this code, let us consider, for example, the canon-
ical base of F22 :
e 0 = 1 0 , e1 = 0 1
and the canonical base of F32 :

e 0 = 1 0 0 , e 1 = 0 1 0 , e 2 = 0 0 1
We can write:
g (e0 ) = [101] = 1.e 0 + 0.e 1 + 1.e 2
g (e1 ) = [011] = 0.e 0 + 1.e 1 + 1.e 2
A generator matrix of the parity check code is therefore equal to :

1 0 1
G=
0 1 1
By permuting the first two vectors of the canonical base of F32 , we obtain a new
generator matrix of the same parity check code:

0 1 1
G =
1 0 1
In this example, we have just seen that the generator matrix of a block code is
not unique. By permuting the rows or the columns of a generator matrix or by
adding one or several other rows to a row, which means considering a new base
in Fn2 , it is always possible to write a generator matrix of a block code in the
following form:
⎡ ⎤
1 0 ··· 0 p0,1 ··· p0,l ··· p0,n−k
⎢⎢ 1
0 ··· 0 p1,1 ··· p1,l ··· p1,n−k ⎥
⎥
G = Ik P = ⎢. . .. .. .. .. ⎥ (4.6)
⎣ .. .. ··· . . ··· . ··· . ⎦
0 0 ... 1 pk−1,1 ... pk−1,l ... pk−1,n−k
where Ik is the identity matrix k × k and P a matrix k × (n− k) used to calculate

the (n − k) redundancy symbols.
Written thus, the generator matrix G is in a reduced form and produces code-
words of the form:

c = d dP (4.7)
The code is therefore systematic. Following 4.7, the code is said to be systematic
when there exist k indices i0 , i1 , . . . , ik−1 , such that for any data word d, the
associated codeword c satisfies the relation:
ciq = dq , q = 0, 1, · · · , k − 1.
4.1.2 Dual code and parity check matrix

Before tackling the notion of dual code, let us define the orthogonality between
two vectors made up of n symbols. Two vectors x = [x0 · · · xj · · · xn−1 ] and
y = [y0 · · · yj · · · yn−1 ] are orthogonal (x⊥y) if their scalar product denoted
x, y is null.

n−1
x⊥y ⇔ x, y = xj yj = 0
j=0
With each linear block code C(n, k), we can associate a dual linear block
code that verifies that any word of the dual code is orthogonal to any word of
the code C(n, k). The dual of code C(n, k) is therefore a vector subspace of Fn2
made up of 2n−k codewords of n symbols. This vector subspace is the orthogonal
4. Block codes 113
of the vector subspace made up of 2k words of the code C(n, k). It results that
any word c of code C(n, k) is orthogonal to the rows of the generator matrix H
of its dual code
cHT = 0 (4.8)
where T indicates the transposition.
A vector y belonging to Fn2 is therefore a codeword of code C(n, k) if, and
only if, it is orthogonal to the codewords of its dual code, that is, if:
yHT = 0
The decoder of a code C(n, k) can use this property to verify that the word
received is a codeword and thus to detect the presence of errors. That is why
matrix H is called the parity check matrix of code C(n, k).
It is easy to see that the matrices G and H are orthogonal (GHT = 0).
Hence, when the code is systematic and its generator matrix is of the form
G = [Ik P], we have:
H = [PT In−k ] (4.9)
4.1.3 Minimum distance

Before recalling what the minimum distance of a linear block code is, let return
to the notion of Hamming distance that measures the difference between two
codewords. The Hamming distance, denoted dH , is equal to the number of
places where the two codewords have different symbols.
We can also define the Hamming weight, denoted wH , of a codeword as the
number of non-null symbols of this codeword. Thus, the Hamming distance
between two codewords is also equal to the weight of their sum.
Example 4.2
Let there be two words u = [1101001] and v = [0101101]. The Hamming

distance between u and v is 2. Their sum u + v = [1000100] has a Hamming
weight 2.
The minimum distance dmin of a block code is equal to the smallest Hamming
distance between its codewords.
dmin = min dH (c, c ), ∀c, c ∈ C(n, k) (4.10)

c=c
Taking into account the fact that the distance between two codewords is equal
to the weight of their sum, the minimum distance of a block code is also equal
to the minimum weight of its non-null codewords.
dmin = min wH (c) (4.11)

c=0,c∈C(n,k)
When the number of codewords is very high, searching for the minimum
distance can be laborious. A first solution to get round this difficulty is to
determine the minimum distance from the parity check matrix.
We have seen that dmin is equal to the minimum Hamming weight of the
non-null codewords. Let us consider a codeword of weight dmin . The orthog-
onality property cHT = 0 implies that the sum of dmin columns of the parity
check matrix is null. Thus dmin corresponds to the minimum number of linearly
dependent columns of the parity check matrix.
A second solution to evaluate dmin is to use higher bounds of the minimum
distance. A first bound can be expressed as a function of the k and n parameters
of the code. For a linear block code whose generator matrix is written in the
systematic form G = [Ik P], the (n − k) columns of the matrix In−k of the
parity check matrix (H = PT In−k ) being linearly independent, any column
of PT can be expressed as at most a combination of these (n − k) columns. The
minimum distance is therefore upper bounded by:
dmin ≤ n − k + 1 (4.12)
Another bound of the minimum distance, called the Plotkin bound, can be
obtained by noting that the minimum distance is necessarily lower than the
average weight of the non-null codewords. If we consider the set of codewords,
it is easy to see that there are as many symbols at 0 as symbols at 1. Thus
the sum of the weights of all the codewords is equal to n2k−1 . The number of
non-null codewords being 2k − 1, the minimum distance can be upper bounded
by:
n2k−1
dmin ≤ k (4.13)
2 −1
4.1.4 Extended codes and shortened codes

From a block code C(n, k) with minimum distance dmin we can build a linear
code C(n + 1, k) by adding to the end of each codeword a symbol equal to 1
(respectively to 0) if the codeword includes an odd (respectively even) number
of 1s. This code is called an extended code and its minimum distance is equal
to dmin + 1 if dmin is an odd number.
The parity check matrix He of an extended code is of the form:
⎡ ⎤
0
⎢ .. ⎥
⎢ H . ⎥
He = ⎢ ⎥
⎣ 0 ⎦
1 ··· 1 1
where H is the parity check matrix of code C(n, k).

4. Block codes 115
A systematic block code C(n, k) with minimum distance dmin can be short-
ened by setting s < k data symbols to zero. We thus obtain a systematic linear
code C(n − s, k − s). Of course the s symbols set to zero are not transmitted,
but they are retained in order to calculate the (n − k) redundancy symbols. The
minimum distance of a shortened code is always higher than or equal to the
distance of code C(n, k).
4.1.5 Product codes

A product code is a code with several dimensions built from elementary codes.
To illustrate these codes, let us consider a product code built from two systematic
block codes C1 (n1 , k1 ) and C2 (n2 , k2 ).
Let there be a table with n2 rows and n1 columns. The k2 first rows are
filled with codewords of length n1 generated by the code C1 (n1 , k1 ). The re-
maining (n2 − k2 ) rows are filled by the redundancy symbols generated by the
code C2 (n2 , k2 ); the k2 symbols of each of the n1 columns being the informa-
tion bits of the code C2 (n2 , k2 ). We can show that the (n2 − k2 ) rows of the
table are codewords of code C1 (n1 , k1 ). It results that all the rows of the table
are codewords of C1 (n1 , k1 ) and all the columns of the table are codewords of
C2 (n2 , k2 ).
The parameters of the two-dimensional product code C(n, k) with minimum
distance dmin are equal to the product of the parameters of the elementary codes.
n = n 1 n2 k = k1 k2 dmin = d1min d2min

where d1min and d2min are the minimum distances of codes C1 (n1 , k1 ) and
C2 (n2 , k2 ) respectively.
A two-dimensional product code can be seen as a double serial concatenation
of two elementary codes (see Chapter 6). An encoder C1 is fed with k2 data
blocks of length k1 and it produces k2 codewords of length n1 that are written
in rows in a matrix. The matrix is read column-wise and produces n1 blocks
of symbols of length k2 that feed an encoder C2 . The latter in turn produces
n1 codewords of length n2 . Figure 4.1 illustrates the implementation of a two-
dimensional product code built from two systematic block codes.
4.1.6 Examples of binary block codes

Parity check code
This code uses a redundancy binary symbol (n = k + 1) determined in such a
way as to ensure the nullity of the modulo 2 addition of the symbols of each
codeword.

k−1
c= d0 d1 ··· dk−2 dk−1 cn−1 with cn−1 = dj
j=0
Figure 4.1 – Product code resulting from the serial concatenation of two systematic
block codes.

where d = d0 d1 · · · dk−1 represents the dataword. The minimum
distance of this code is 2.
Example 4.3
A generator matrix G of this code for n = 5, k = 4 is equal to:

⎡ ⎤
1 0 0 0 1
⎢ 0 1 0 0 1 ⎥
G=⎢ ⎣ 0 0
⎥ = I4 P
⎦
1 0 1
0 0 0 1 1
and the parity check matrix H is reduced to one vector.

H = 1 1 1 1 1 = PT I1
Repetition code
For this code with parameters k = 1 and n = 2m + 1, each bit coming from the
information source is repeated an odd number of times. The minimum distance
of this code is 2m + 1. The repetition code C(2m + 1, 1) is the dual code of the
parity check code C(2m + 1, 2m).
4. Block codes 117
Example 4.4
The generator matrix and the parity check matrix of this code, for k = 1,
n = 5, can be the following:

G= 1 1 1 1 1 = I1 P
⎡ ⎤
1 1 0 0 0
⎢ 1 0 1 0 0 ⎥
H=⎢ ⎣ 1
⎥ = PT I4
0 0 1 0 ⎦
1 0 0 0 1
Hamming code
For a Hamming code, the columns of the parity check matrix are the binary
representations of the numbers from 1 to n. Each column being made up of
m = (n − k) binary symbols, the parameters of the Hamming code are therefore:
n = 2m − 1 k = 2m − m − 1
The columns of the parity check matrix being made up of all the possible com-
binations of (n − k) binary symbols except (00 · · · 0), the sum of two columns is
equal to one column. The minimum number of linearly dependent columns is
3. The minimum distance of a Hamming code is therefore equal to 3, whatever
the value of parameters n and k.
Example 4.5
Let there be a Hamming code with parameter m = 3. The codewords and

the datawords are then made up of n = 7 and k = 4 binary symbols respectively.
The parity check matrix can be the following:
⎡ ⎤
1 1 1 0 1 0 0
H = ⎣ 1 1 0 1 0 1 0 ⎦ = PT I3
1 0 1 1 0 0 1
and the corresponding generator matrix is equal to:

⎡ ⎤
1 0 0 0 1 1 1
⎢ 0 1 0 0 1 1 0 ⎥
G=⎢ ⎣ 0 0 1 0 1
⎥ = I4 P
0 1 ⎦
0 0 0 1 0 1 1
Maximum length code

The columns of the generator matrix of a maximum length code are the binary
representations of the numbers from 1 to n. The parameters of this code are
therefore n = 2m − 1, k = m and we can show that its minimum distance is 2k−1 .
The maximum length code with parameters n = 2m − 1, k = m is the dual code
of the Hamming code with parameters n = 2m − 1, k = 2m − m − 1, that is, the
generator matrix of the one is the parity check matrix of the other.
Hadamard code
The codewords of a Hadamard code are made up of the rows of a Hadamard
matrix and of its complementary matrix. A Hadamard matrix has n rows and
n columns (n even) whose elements are 1s and 0s. Each row differs from the
other rows at n/2 positions. The first row of the matrix is made up only of 0,
the other rows having n/2 0 and n/2 1.
For n = 2, the Hadamard matrix is of the form:

0 0
M2 =
0 1
From a Mn matrix we can generate a M2n matrix.

Mn Mn
M2n =
Mn Mn
where Mn is the complementary matrix of Mn , that is, where each element

at 1 (respectively at 0) of Mn becomes an element at 0 (respectively at 1) for Mn .
Example 4.6
If n = 4 M4 and M4 have the form:

⎡ ⎤ ⎡ ⎤
0 0 0 0 1 1 1 1
⎢ 0 1 0 1 ⎥ ⎢ 1 0 1 0 ⎥
M4 = ⎢
⎣ 0
⎥ M4 = ⎢ ⎥
0 1 1 ⎦ ⎣ 1 1 0 0 ⎦
0 1 1 0 1 0 0 1
The rows of M4 and M4 are the codewords of a Hadamard code with parameters
n = 4, k = 3 and with minimum distance equal to 2. In this particular case, the
Hadamard code is a parity check code.
More generally, the rows of matrices Mn and Mn are the codewords of a
Hadamard code with parameters n = 2m , k = m+ 1 and with minimum distance
dmin = 2m−1 .
4. Block codes 119
Reed-Muller codes
A Reed-Muller code (RM) of order r and with parameter m, denoted RMr,m ,
has codewords of length n = 2m and the datawords are made up of k symbols
with:

m m N N!
k =1+ + ···+ , with =
1 r q q! (N − q)!
where r < m. The minimum distance of an RM code is dmin = 2m−r .
The generator matrix of an RM code of order r is built from the generator matrix
of an RM code of order r − 1 and if G(r,m) represents the generator matrix of
the Reed-Muller code of order r and with parameter m, it can be obtained from
G(r−1,m) by the relation:
(r−1,m)
(r,m) G
G =
Qr

m
where Qr is a matrix with dimensions × n.
r
By construction, G(0,m) is a row vector of length n whose elements are equal
to 1. The matrix G(1,m) is obtained by writing on each column the binary
representation of the index of the columns (from 0 to n − 1). For example, for
m = 4, the matrix G(1,m) is given by:
⎡ ⎤
0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1
⎢ 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1 ⎥
G(1,4) = ⎢
⎣ 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 ⎦.
⎥
0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1
Matrix Qr is obtained simply by considering all the combinations of r rows of
G(1,m) and by obtaining the product of these vectors, component by component.
The result of this multiplication constitutes a row of Qr . For example, for
the combination having the rows of G(1,m) with indices i1 , i2 , . . . , ir , the j-
(1,m) (1,m) (1,m)
th coefficient of the row thus obtained is equal to Gi1 ,j Gi2 ,j · · · Gir ,j , the
multiplication being carried out in the field F2 . For example, for r = 2, we
obtain:
⎡ ⎤
0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1
⎢ 0 0 0 0 0 0 0 0 0 0 1 1 0 0 1 1 ⎥
⎢ ⎥
⎢ 0 0 0 0 0 0 0 0 0 1 0 1 0 1 0 1 ⎥
Q2 = ⎢⎢ ⎥
⎥
⎢ 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 1 ⎥
⎣ 0 0 0 0 0 1 0 1 0 0 0 0 0 1 0 1 ⎦
0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1
We can show that the code RMm−r−1,m is the dual code of the code RMr,m ,
that is, the generator matrix of code RMm−r−1,m is the parity check matrix of
code RMr,m . For some values of r and m, the generator matrix of code RMr,m
is also its parity check matrix. We then say that code RMr,m is self dual. Code
RM1,3 , for example, is a self dual code.
4.1.7 Cyclic codes

Cyclic codes are the largest class of linear block codes. Their relatively easy
implementation, from shift registers and logical operators, has made them at-
tractive and widely-used codes.
Definition and polynomial representation
A linear block code C(n, k) is cyclic if, for any codeword c =
c0 c1 · · · cn−1 , c1 = cn−1 c0 · · · cn−2 , obtained by circu-
lar shift to the right of a symbol of c, is also a codeword. This definition
of cyclic codes means that any circular shift to the right of j symbols of a
codeword, gives another codeword.
For cyclic codes, we use a polynomial representation of the codewords and of the
datawords. Thus, with codeword c we associate the polynomial c(x) of degree
n − 1.
c(x) = c0 + c1 x + · · · + cj xj + · · · + cn−1 xn−1
and with dataword d the polynomial d(x) of degree k − 1.
d(x) = d0 + d1 x + · · · + dj xj + · · · + dk−1 xk−1
where dj and cj take their values in F2 .

Multiplying c(x) by x,
xc(x) = c0 x + c1 x2 + · · · + cj xj+1 + · · · + cn−1 xn
then dividing xc(x) by xn + 1, we obtain:
xc(x) = (xn + 1)cn−1 + c1 (x)
where c1 (x) is the remainder of the division of xc(x) by xn + 1 with:
c1 (x) = cn−1 + c0 x + · · · cj xj+1 + · · · cn−2 xn−1
We can note that c1 (x) corresponds to the codeword c1 = (cn−1 c0 . . . cj . . . cn−2 ).

Using the same method as above, we obtain:
xj c(x) = (xn + 1)q(x) + cj (x) (4.14)
where cj (x) is also a codeword obtained by j circular shifts to the right of the
symbols of c(x).
4. Block codes 121
The codewords of a cyclic code are multiples of a normalized polynomial g(x)

of degree (n − k) called a generator polynomial .
g(x) = g0 + g1 x + · · · + gj xj + · · · + xn−k
where gj takes its values in F2 . The generator polynomial of a cyclic code is a
divisor of xn + 1. There exists a polynomial h(x) of degree k such that equation
(4.15) is satisfied.
g(x)h(x) = xn + 1 (4.15)
The product d(x)g(x) is a polynomial of degree lower than or equal to n − 1,
so it can represent a codeword. The polynomial d(x) having 2k realizations,
d(x)g(x) enables 2k codewords to be generated. Let us denote dl (x) the l-th
realization of d(x) and cl (x) the polynomial representation of the associated
codeword. We can write:
cl (x) = dl (x)g(x) (4.16)
We will now show that the codewords satisfying relation (4.16) satisfy the prop-
erties of cyclic codes. To do so, we re-write relation (4.14) in the form:
cj (x) = (xn + 1)q(x) + xj c(x) (4.17)
Since c(x) represents a codeword, there exists a polynomial d(x) of degree at
most k − 1, such that c(x) = d(x)g(x). Using (4.15), we can therefore express
(4.17) in another way:
cj (x) = g(x)[h(x)q(x) + xj d(x)] (4.18)
The codewords cj (x) are therefore multiples of the generator polynomial, and
they can be generated from the dj (x) by applying relation (4.16).
• Generator polynomial of the dual code of C(n, k)

The dual code of a cyclic block code is also cyclic. Polynomial h(x) of degree
k can be used to build the dual code of C(n, k). The reciprocal polynomial h̃(x)
of h(x) is defined as follows:
h̃(x) = xk h(x−1 ) = 1 + hk−1 x + hk−2 x2 + · · · + h1 xk−1 + xk
We can write (4.15) differently:
[xn−k g(x−1 )][xk h(x−1 )] = xn + 1 (4.19)
The polynomial h̃(x) is also a divisor of xn + 1; it is the generator polynomial
of a C ⊥ = C(n, n − k) code that is the dual code of C(n, k).
Note: the code of generator polynomial h(x) is equivalent to dual code C ⊥ .
The vector representation of the codewords generated by h(x) corresponds to
the reversed vector representation of the codewords of C ⊥ .
C ⊥ , generated by h̃(x) ↔ generated by h(x)
Code
c̃ = c0 c1 · · · cn−1 ↔ c = cn−1 · · · c1 c0
• Generator matrix of a cyclic code

From the generator polynomial g(x) it is possible to build a generator matrix
G of code C(n, k). We recall that the k rows of the matrix G are made up of k
linearly independent codewords. These k codewords can be obtained from a set
of k independent polynomials of the form:
xj g(x) j = k − 1, k − 2, . . . , 1, 0.
Let d(x) be the polynomial representation of any dataword. The k codewords
generated by the polynomials xj g(x) have the expression:
cj (x) = xj g(x)d(x) j = k − 1, k − 2, · · · , 1, 0
and the k rows of the matrix G have for their elements the binary coefficients
of the monomials of cj (x).
Example 4.7
Let C(7, 4) be the generator polynomial code g(x) = 1 + x2 + x3 . Let us take

d(x) = 1 for the dataword. The 4 rows of the generator matrix G are obtained
from the 4 codewords cj (x).
c3 (x) = x3 + x5 + x6
c2 (x) = x2 + x4 + x5
c1 (x) = x + x3 + x4
c0 (x) = 1 + x2 + x3
A generator matrix of the code C(7, 4) is equal to:
⎡ ⎤
0 0 0 1 0 1 1
⎢ 0 0 1 0 1 1 0 ⎥
G=⎢ ⎣ 0 1 0 1 1
⎥
0 0 ⎦
1 0 1 1 0 0 0
Cyclic code in systematic form

When the codewords are in systematic form, the data coming from the infor-
mation source are separated from the redundancy symbols. The codeword c(x)
associated with dataword d(x) is then of the form:
c(x) = xn−k d(x) + v(x) (4.20)
where v(x) is the polynomial of degree at most equal to n − k − 1 associated

with the redundancy symbols.
4. Block codes 123
Taking into account the fact that c(x) is a multiple of the generator poly-
nomial and that the addition and the subtraction can be merged in F2 , we can
then write:
xn−k d(x) = q(x)g(x) + v(x)
v(x) is therefore the remainder of the division of xn−k d(x) by the generator
polynomial g(x). The codeword associated with dataword d(x) is equal to
xn−k d(x) increased by the remainder of the division of xn−k d(x) by the gener-
ator polynomial.
Figure 4.2 – Schematic diagram of a circuit divisor by g(x).
Example 4.8
To illustrate the computation of a codeword written in systematic form, let us

take the example of a C(7,4) code of generator polynomial g(x) = 1 + x + x3 and
let us determine the codeword c(x) associated with message d(x) = 1 + x2 + x3 ,
that is:
c(x) = x3 d(x) + v(x)

The remainder of the division of x3 d(x) by g(x) = 1 + x + x3 being equal to 1,
codeword c(x) associated with dataword d(x) is:
c(x) = 1 + x3 + x5 + x6
Thus, with data block d, made up of 4 binary information symbols, is associated

codeword c with:

d= 1 0 1 1 →c= 1 0 0 1 0 1 1
To obtain the generator matrix, it suffices to encode
d(x) = 1, x, x2 , x3 .
We obtain:
d(x) c(x)
1 1 + x + x3
x x + x2 + x4
x2 1 + x + x2 + x5
x3 1 + x2 + x6
and thus the generator matrix in a systematic form:
⎡ ⎤
1 1 0 1 0 0 0
⎢ 0 1 1 0 1 0 0 ⎥
G=⎢ ⎣ 1
⎥
1 1 0 0 1 0 ⎦
1 0 1 0 0 0 1
We can verify that for

d= 1 0 1 1 ,
the matrix product dG does give

c= 1 0 0 1 0 1 1 .
Implementation of an encoder
We have just seen that the encoder must carry out the division of xn−k d(x)
by the generator polynomial g(x) then add the remainder v(x) of this division
to xn−k d(x). This operation can be done using only shift registers and adders
in field F2 . As the most difficult operation to carry out is the division of
xn−k d(x) by g(x), let us first examine the schematic diagram of a divisor by
g(x) shown in Figure 4.2. The circuit divisor is realized from a shift register
with (n − k) memories denoted Ri and the same number of adders. The shift
register is initialized to zero and the k coefficients of the polynomial xn−k d(x)
are introduced sequentially into the circuit divisor. After k clock pulses, we
can verify that the result of the division is available at the output of the cir-
cuit divisor, as well as the remainder v(x) which is in the shift register memories.
The schematic diagram of the encoder shown in Figure 4.3, uses the circuit
divisor of Figure 4.2. The multiplication of d(x) by xn−k , corresponding to a
simple shift, is realized by introducing polynomial d(x) at the output of the shift
register of the divisor.
The k data coming from the information source are introduced sequentially
into the encoder (switch I in position 1) that carries out the division of xn−k d(x)
by g(x). Simultaneously, the k data coming from the information source are also
transmitted. Once this operation is finished, the remainder v(x) of the division
is in the (n − k) shift register memories. Switch I then moves to position 2, and
the (n − k) redundancy symbols are sent to the output of the encoder.
4. Block codes 125
Figure 4.3 – Schematic diagram of an encoder for a cyclic code.
BCH codes
Bose-Chaudhuri-Hocquenghem codes, called BCH codes, enable cyclic codes to
be built systematically correcting at least t errors in a block of n symbols, that
is, codes whose minimum distance dmin is at least equal to 2t + 1.
To build a BCH code, we set t or equivalently d, called the constructed
distance of the code and we determine its generator polynomial g(x). The code
obtained has a minimum distance dmin that is always higher than or equal to
the constructed distance.
Primitive BCH code
The generator polynomial g(x) of a primitive BCH code constructed over a

Galois field Fq with q = 2m elements, with a constructed distance d has (d − 1)
roots of the form: αl , · · · , αl+j , · · · , αl+d−2 , where 2t + 1 is a primitive element
of Galois field Fq and l an integer. The BCH code is said to be primitive since
the roots of its generator polynomial are powers of α, a primitive element of Fq .
We will see later that it is possible to build non-primitive BCH codes.
Generally, parameter l is set to 0 or 1 and we show that for a primitive
BCH code exponent (l + d − 2) of root αj+d−2 must be even. When l = 0, the
constructed distance is therefore necessarily even, that is, equal to 2t + 2 for a
code correcting t errors. When l = 1, the constructed distance is odd, that is,
equal to 2t + 1 for a code correcting t errors.
• Primitive BCH code with l = 1

The generator polynomial of a primitive BCH code correcting at least t
errors (constructed distance 2t + 1) therefore has α, · · · , αj , · · · , α2t as
roots. We show that the generator polynomial g(x) of a primitive BCH
code is equal to:
g(x) = S.C.M. (mα (x), · · · , mαi (x), · · · , mα2t (x))

where mαi (x) is the minimal polynomial with coefficients in field F2 asso-
ciated with αj , and S.C.M. is the Smallest Common Multiple.
It is shown in the appendix that a polynomial with coefficients in F2 having
αj as its root also has α2j as its root. Thus, the minimal polynomials
mαi (x) and mα2i (x) have the same roots. This remark enables us to
simplify the writing of generator polynomial g(x).
g(x) = S.C.M. (mα (x), mα3 (x), · · · , mα2t−1 (x)) (4.21)
The degree of a minimal polynomial being lower than or equal to m, degree

(n − k) of the generator polynomial of a primitive BCH code correcting
at least t errors, is therefore lower than or equal to mt. Indeed, g(x) is at
most equal to the product of t polynomials of degree lower than or equal
to m.
The parameters of a primitive BCH code constructed over a Galois field
Fq with a constructed distance d = 2t + 1 are therefore the following:
n = 2m − 1; k ≥ 2m − 1 − mt; dmin ≥ 2t + 1
When t = 1 a primitive BCH code is a Hamming code. The generator

polynomial of a Hamming code, equal to mα (x), is therefore a primitive
polynomial.
Example 4.9
Let us determine the generator polynomial of a BCH code having param-

eters m = 4 and n = 15, t = 2 and l = 1. To do this, we will use a Galois
field with q = 24 elements built from a primitive polynomial of degree
m = 4(α4 + α + 1). The elements of this field are given in the appendix.
We must first determine the minimal polynomials mα (x) and mα3 (x) as-
sociated with elements α and α3 respectively of field F16 .
We have seen in the appendix that if α is a root of polynomial mα (x) then
α2 , α4 , α8 are also roots of this polynomial (raising α to the powers of 16,
32 etc. gives, modulo α4 + α + 1, elements α, α2 , α4 , α8 ). We can therefore
write:
mα (x) = (x + α)(x + α2 )(x + α4 )(x + α8 )
Developing the expression of mα (x) we obtain:
mα (x) = [x2 + x(α2 + α) + α3 ][x2 + x(α8 + α4 ) + α12 ]
Using the binary representations of the elements of field F16 , we can show
that α2 +α = α5 and that α4 +α8 = α5 (we recall that the binary additions
4. Block codes 127
are done modulo 2 in the Galois field). We then continue the development
of mα (x) and finally we have:
mα (x) = x4 + x + 1
For the computation of mα3 (x), the roots to take into account are
α3 , α6 , α12 , α24 = α9 (α15 = 1), and the other powers of α3 (α48 , α96 , · · · )
give the previous roots again. The minimal polynomial mα3 (x) is therefore
equal to:
mα3 (x) = (x + α3 )(x + α6 )(x + α12 )(x + α9 )
which after development and simplification gives:
mα3 (x) = x4 + x3 + x2 + x + 1
The S.C.M. of polynomials mα (x) and mα3 (x) is obviously equal to the
product of these two polynomials since they are irreducible and thus, the
polynomial generator is equal to:
g(x) = (x4 + x + 1)(x4 + x3 + x2 + x + 1)
Developing this, we obtain:
g(x) = x8 + x7 + x6 + x4 + 1
Finally the parameters of this BCH code are:
m = 4; n = 15; n − k = 8; k = 7; t = 2
The numerical values of parameters (n, k, t) of the main BCH codes and
the associated generator polynomials have been put table form and can be
found in [4.2]. As an example, we give in Table 4.2 the parameters and
the generator polynomials, expressed in octals, of some BCH codes with
error correction capability t = 1 (Hamming codes).
Note : g(x) = 13 in octals gives 1011 in binary, that is, g(x) = x3 + x + 1
• Primitive BCH code with l = 0

The generator polynomial of a primitive BCH code correcting at least t
errors (constructed distance d = 2t + 2) has (2t + 1) roots of the form:
α0 , α1 , · · · , αj , · · · , α2t ; that is, one root more (α0 ) than when l = 1.
Taking into account the fact that the minimal polynomials mαj (x) and
mα2j (x) have the same roots, generator polynomial g(x) is equal to:
g(x) = S.C.M.(mα0 (x), mα1 (x), mα3 (x), · · · , mα2t−1 (x))

n k t g(x)
7 4 1 13
15 11 1 23
31 26 1 45
63 57 1 103
127 120 1 211
255 247 1 435
511 502 1 1021
1023 1013 1 2011
2047 2036 1 4005
4095 4083 1 10123
Table 4.2 – Parameters of some Hamming codes.
1. Parity check code

Let us consider a BCH code with l = 0 and t = 0. Its generator
polynomial, g(x) = (x + 1) has only one root α0 = 1. This code uses
only one redundancy symbol and the c(x) words of this code satisfy
the condition:
c(α0 ) = c(1) = 0
This code, which is cyclic since (x + 1) divides (xn + 1), is a parity

check code with parameters n = k + 1, k, t = 0. Thus, every time we
build a BCH code by selecting l = 0, we introduce into the genera-
tor polynomial a term in (x+1) and the codewords are of even weight.
2. Cyclic Redundancy Code (CRC)

Another example of a BCH code for which l = 0, is the CRC used
for detecting errors. A CRC has a constructed distance of 4 (t = 1)
and its generator polynomial, from above, is therefore equal to:
g(x) = (x + 1)mα (x)
α being a primitive element, mα (x) is a primitive polynomial and thus

the generator polynomial of a CRC is a code equal to the product of
(x + 1) by the generator polynomial of a Hamming code.
gCRC (x) = (x + 1)gHamming (x)
The parameters of a CRC are therefore:
n = 2m − 1; (n − k) = m + 1; k = 2m − m − 2
4. Block codes 129
Code m g(x)
CRC-12 12 14017
CRC-16 16 300005
CRC-CCITT 16 210041
CRC-32 32 40460216667
Table 4.3 – generator polynomials of some codes CRC.
The most widely-used CRC codes have the parameters m = 12, 16,
32 and their generator polynomials are given, in octals, in Table 4.3.
Note: g(x) = 14017 in octals corresponds to 1 100 000 001 111 in
binary, that is:
g(x) = x12 + x11 + x3 + x2 + x + 1
Non-primitive BCH code
The generator polynomial of a non-primitive BCH code (with l = 1) correct-

ing at least t errors (constructed distance d = 2t + 1) has 2t roots of the form:
β, β 2 , β 3 , . . . , β 2t where β is a non-primitive element of a Galois field Fq . Taking
into account the fact that the minimal polynomials mβ j (x) and mβ 2j (x) have
the same roots, the generator polynomial of a non-primitive BCH code is equal
to:
g(x) = S.C.M.(mβ (x), mβ 3 (x).....mβ 2t−1 (x))
We can show that length n of the words of a non-primitive BCH code is no
longer of the form 2m − 1 but is equal to p, where p is the exponent of β such
that β p = 1 (p is the order of β). A Galois field Fq has non-primitive elements
if 2m − 1 is not prime. The non-primitive elements are then of the form β = αλ
where λ is a divisor of 2m − 1 and α is a primitive element of the field.
Example 4.10
Let there be a Galois field Fq with m = 6 and q = 64. The quantity

2m − 1 = 63 is not equal to a prime number; it is divisible by 3, 7, 9, 21 and 63.
The non-primitive elements of this field are therefore α3 , α7 , α9 , α21 , α63 = 1.
Let us build, for example, a non-primitive BCH code having an error correction
capability at least equal to t = 2 on field F64 and let us take β = α3 as the
non-primitive element. We have two minimal polynomials to calculate mβ (x)
and mβ 3 (x). Taking into account the fact that β 21 = α63 = 1, the roots of these
polynomials are:
mβ (x) : roots β, β 2 , β 4 , β 8 , β 16 , β 32 = β 11
mβ 3 (x) : roots β 3 , β 6 , β 12
The generator polynomial of this code is equal to:
g(x) = mβ (x)mβ 3 (x)
which, after development and simplification, gives:
g(x) = x9 + x8 + x7 + x5 + x4 + x + 1
The parameters of this non-primitive BCH code are:
n = 21; (n − k) = 9; k = 12
• Golay code
Among non-primitive BCH codes, the most well-known is certainly the
Golay code constructed over a Galois field Fq with m = 11, q = 2048.
Noting that 2m − 1 = 2047 = 23 × 89, the non-primitive element used
to build a Golay code is β = α89 . The computation of the generator
polynomial of this code constructed on field F2048 leads to the following
expression:
g(x) = x11 + x9 + x7 + x6 + x5 + x + 1
We can show that the minimum distance dmin of a Golay code is 7 and
thus, its correction capability is 3 errors in a block of 23 binary symbols
(β 23 = α2047 = 1). The parameters of a Golay code are therefore:
n = 23; (n − k) = 11; k = 12; t = 3
Note that the reciprocal polynomial of g(x), equal to g̃(x) = x11 g(x−1 )
also enables a Golay code to be produced.
g̃(x = x11 + x10 + x6 + x5 + x4 + x2 + 1
4.2 Block codes with non-binary symbols

4.2.1 Reed-Solomon codes
Reed-Solomon or RS codes are the most well-known and the most widely-used
codes having non-binary symbols. For codes with non-binary symbols the coef-
ficients cj of the codewords and dj of the datawords take their value in a Galois
field Fq with q = 2m elements. Thus, each symbol of these codes can be en-
coded on m binary symbols. Reed-Solomon codes being cyclic codes, they are
generated by a generator polynomial g(x) divisor of xn + 1 whose coefficients
gj j = 0, 1, · · · , n − k − 1 also take their value in the Galois field Fq .
4. Block codes 131
The generator polynomial of a Reed-Solomon code, with a constructed dis-

tance d has (d− 1) roots αl , · · · , αl+j , · · · , αl+d−2 where α is a primitive element
of Galois field Fq . It therefore has the expression:
g(x) = S.C.M.(mαl (x), · · · , mαl+j (x), · · · , mαl+d−2 (x))
where mαl+j is the minimal polynomial associated with the αl+j element of field
Fq .
Using the results of the appendix on the minimal polynomials with coeffi-
cients in Fq , the minimal polynomial mαl+j has only one root αl+j .
mαj+i (x) = (x + αj+i )
The generator polynomial of a Reed-Solomon code is therefore of the form:
g(x) = (x + αj )(x + αj+1 )...(x + αj+i )...(x + αj+d−2 )
In general, parameter j is set to 0 or 1 like for binary BCH codes. The
generator polynomial of a Reed-Solomon code, of degree (n − k), has (d − 1)
roots, that is n − k = d − 1. Its constructed distance is therefore equal to:
d=n−k+1
For a block code k C(n, k) the minimum distance dmin being lower than or equal
to n − k + 1, the minimum distance of a Reed-Solomon code is therefore always
equal to its constructed distance. A code whose minimum distance is equal to
n − k + 1 is called a maximum distance code.
The parameters of a Reed-Solomon code correcting t errors in a block of n q-ary
symbols are therefore:
n = q − 1; n − k = dmin − 1 = 2t; k = n − 2t
Example 4.11
Let us determine the generator polynomial of a Reed-Solomon code built

from a Galois field with 16 elements having a correction capability of t = 2
errors. The minimum distance of this code is therefore dmin = 5. Taking for
example l = 1, the generator polynomial of this code is therefore of the form:
g(x) = (x + α)(x + α2 )(x + α3 )(x + α4 )
Developing the expression above, we obtain:
g(x) = x2 + x(α + α2 ) + α3 x2 + x(α3 + α4 ) + α7
Using the binary representations of the elements of field F16 (Appendix), the
polynomial g(x) after development and simplification is equal to:
g(x) = x4 + α3 x3 + α6 x2 + α3 x + α10
4.2.2 Implementing the encoder

The schematic diagram of an encoder for Reed-Solomon codes is quite similar
to that of an encoder for cyclic codes with binary symbols, but the encoder
must now carry out multiplications between q-ary symbols and memorize q-ary
symbols.
As an example, we have shown in Figure 4.4 the schematic diagram of the
encoder for the Reed-Solomon code treated in the example above.
Figure 4.4 – Schematic diagram of the encoder for the RS code (15,11).
4.3 Decoding and performance of codes with bi-

nary symbols
4.3.1 Error detection
Considering a binary symmetric transmission channel, the decoder receives bi-
nary symbols assumed to be perfectly synchronized with the encoder. This
means that the splitting into words having n symbols at the input of the de-
coder corresponds to the splitting used by the encoder. Thus, in the absence of
errors, the decoder sees codewords at its input.
Let us assume that codeword c is transmitted by the encoder and let r be
the word of n symbols received at the input of the decoder. Word r can always
be written in the form:
r=c+e
where e is a word whose non-null symbols represent the errors. A non-null
symbol of e indicates the presence of an error in the corresponding position of
c.
Errors are detected by using the orthogonality property of the parity check
matrix with the codewords and calculating the quantity s called the error syn-
drome.
s = rHT = (c + e)HT = eHT
4. Block codes 133
Syndrome s is null if, and only if, r is a codeword. A non-null syndrome implies
the presence of errors. However, it should be noted that a null syndrome does
not necessarily mean absence of errors since r can belong to the set of codewords
even though it is different from c. For this to occur, it suffices for word e to be a
codeword. Indeed, for a linear block code, the sum of two codewords is another
codeword.
Finally, let us note that for any linear block code, there are configurations
of non-detectable errors.
Detection capability
Let cj be the transmitted codeword and cl its nearest neighbour. We have the
following inequality:
dH (cj , cl ) dmin
Introducing the received word r, we can write:
dmin ≤ dH (cj , cl ) ≤ dH (cj , r) + dH (cl , r)
and thus all the errors can be detected if the Hamming distance between r and
cl is higher than or equal to 1, that is, if r is not merged with cl .
The detection capability of a C(n, k) code with minimum distance dmin is
therefore equal to dmin − 1.
Probability of non-detection of errors

Considering a block code C(n, k) and a binary symmetric channel with error
probability p, the probability of non-detection of the errors Pnd is equal to:

n
Pnd = Aj pj (1 − p)n−j (4.22)
j=dmin
where Aj is the number of codewords with weight j.

Examining the hypothesis of a completely degraded transmission, that is, of
an error probability of p = 1/2 on the channel, and taking into account the fact
that for any block code we have:

n
Aj = 2k − 1
j=dmin
(the −1 in the above expression corresponds to the null codeword), probability

Pnd is equal to:
2k − 1 ∼ −(n−k)
Pnd = =2
2n
The detection of errors therefore remains efficient whatever the error proba-
bility on the transmission channel if the number of redundancy symbols (n − k)
is large enough. The detection of errors is therefore not very sensitive to error
statistics.
When erroneous symbols are detected, the receiver generally asks the source
to send them again. To transmit this re-transmission request, it is then necessary
to have a receiver source link, called a return channel. The data rate on the
return channel being low (a priori, requests for retransmission are short and few
in number), we can always arrange it so that the error probability on this channel
is much lower than the error probability on the transmission channel. Thus, the
performance of a transmission system using error detection and repetition does
not greatly depend on the return channel.
In case of error detection, the emission of the source can be interrupted
to enable the retransmission of the corrupted information. The data rate is
therefore not constant, which can present problems in some cases.
4.3.2 Error correction

Error correction involves looking for the transmitted codeword c given the re-
ceived word r. Two strategies are possible. The first one corresponds to a
received word r at the input of the decoder made up of binary symbols (the
case of a binary symmetric channel) and the second, to a received word r made
up of analogue symbols (the case of a Gaussian channel). In the first case, we
speak of hard input decoding whereas in the second case we speak of soft input
decoding. We will now examine these two types of decoding, already mentioned
in Chapter 1.
Hard decoding
• Maximum a posteriori likelihood decoding
For hard decoding the received word r is of the form:
r=c+e
where c and e are words with binary symbols.

Maximum a posteriori likelihood decoding involves looking for the codeword
ĉ such that:
Pr {ĉ |r } > Pr {ci |r } ∀ ci = ĉ ∈ C(n, k)
Using Bayes’ rule and assuming that all the codewords are equiprobable, the
above decision rule can also be written:
ĉ = ci ⇔ Pr (r |c = ci ) > Pr (r |c = cj ) , ∀cj = ci ∈ C (n, k)
Again taking the example of a binary symmetric channel with error probability
p and denoting dH (r, ĉ) the Hamming distance between r and ĉ, the decision
4. Block codes 135
rule is:
n−dH (ci ,r) n−dH (cj ,r)
ĉ = ci ⇔ pdH (ci ,r) (1 − p) > pdH (cj ,r) (1 − p) , ∀cj = ci
Taking the logarithm of the two parts of the above inequality and considering
p < 0.5, the decision rule of the maximum a posteriori likelihood can finally be
written:
ĉ = ci ⇔ dH (r, ci ) ≤ dH (r, cj ), ∀ cj = ci ∈ C(n, k)
If two or several codewords are the same distance from r, the codeword ĉ is
chosen arbitrarily among the codewords equidistant from r.
This decoding procedure which is optimal, that is, which minimizes the
probability of erroneous decoding, becomes difficult to implement when the
number of codewords becomes large, which is often the case for the widely-used
block codes.
• Decoding from the syndrome

To get around this difficulty, it is possible to perform the decoding using
syndrome s. We recall that the syndrome is a vector of dimension (n − k) that
depends solely on the error configuration e. For a binary symbol block code,
the syndrome has 2n−k configurations, which is generally much lower than the
2k codewords.
To decode from the syndrome, we use a table with n rows and two columns.
We write respectively in each row of the first column the null syndrome s (all
the symbols are at zero, no errors) then the syndromes s corresponding to the
configuration of an error then two errors etc. until the n rows are filled. All the
configurations of the syndromes of the first column must be different. In the
second column, we write the error configuration associated with each syndrome
of the first column.
For a received word r we calculate the syndrome s then, using the table, we
deduce the error word e. Finally, we add word e to r and we obtain the most
likely codeword.
Example 4.12
Let us consider a code C(7, 4) with a parity check matrix H with:

⎡ ⎤
1 1 1 0 1 0 0
H=⎣ 1 1 0 1 0 1 0 ⎦
1 0 1 1 0 0 1
This code has 16 codewords but only 8 configurations for the syndrome as indi-
cated in Table 4.4.
Syndrome s Error word e

000 0000000
001 0000001
010 0000100
011 0001000
100 0000100
101 0010000
110 0100000
111 1000000
Table 4.4: Syndromes and corresponding error words for a C(7, 4) code.
Let us assume that the codeword transmitted is c = [0101101] and that the
received word r = [0111101] has an error in position 3. The syndrome is then
equal to s = [101] and, according to the table, e = [0010000]. The decoded
codeword is ĉ = r + e = [0101101] and the error is corrected.
If the number of configurations of the syndrome is still too high to apply
this decoding procedure, we use decoding algorithms specific to certain classes
of codes but that, unfortunately, do not always exploit the whole correction
capability of the code. These algorithms will be presented below.
• Correction power
Let cj be the codeword transmitted and cl its nearest neighbour. We have
the following inequality:
dH (cj , cl ) dmin
Introducing the received word r and assuming that the minimum distance
dmin is equal to 2t + 1 (integer t), we can write:
2t + 1 ≤ dH (cj , cl ) ≤ dH (cj , r) + dH (cl , r)
We see that if the number of errors is lower than or equal to t, cj is the most
likely codeword since it is nearer to r than to cl and thus the t errors can be
corrected. If the minimum distance is now (2t + 2), using the same reasoning,
we arrive at the same error correction capability. In conclusion, the correction
capability of a linear block code with minimum distance dmin with hard decoding
is equal to:
dmin − 1
t= (4.23)
2
where x is the whole part of x rounded down (for example 2.5 = 2).
4. Block codes 137
• Probability of erroneous decoding of a codeword

For a linear block code C(n, k) of error correction capability t, the codeword
transmitted will be wrongly decoded if there are t + j errors, j = 1, 2, · · · , n −
t, in the received word r. For a binary symmetric channel of probability p,
the probability Pe,word of performing an erroneous decoding of the transmitted
codeword is upper bounded by:
n
n
Pe,word < pj (1 − p)n−j (4.24)
j
j=t+1
We can also determine the binary error probability Pe,bit on the information data
after decoding. In presence of erroneous decoding, the maximum a posteriori
likelihood decoder adds at most t errors by choosing the codeword with the
minimum distance from the received word. The error probability is therefore
bounded by:

1
n
n
Pe,bit < (j + t) pj (1 − p)n−j (4.25)
n j
j=t+1
If the transmission is performed with binary phase modulation (2-PSK, 4-PSK),

probability p is equal to:
1 REb
p = erfc
2 N0
where R is the coding rate, Eb the energy received per transmitted information
bit and N0 the unilateral power spectral density of the noise. Figure 4.5 shows
the binary error probability and word error probability after algebraic decoding
for the (15,7) BCH code. The modulation is 4-PSK and the channel is Gaussian.
The higher bounds expressed by (4.24) and (4.25) respectively are also plotted.
Soft decoding
Considering a channel with additive white Gaussian noise and binary phase mod-
ulation transmission (2-PSK or 4-PSK), the components rj , j = 0, 1, · · · , n− 1
of the received word r have the form:
+
rj = Es c̃j + bj , c̃j = 2cj − 1
where cj = 0, 1 is the symbol in position j of codeword c, c̃j is the binary
symbol associated with cj , Es is the energy received per transmitted symbol
and bj is white Gaussian noise, with zero mean and variance equal to σb2 .
• Maximum a posteriori likelihood decoding

Decoding using the maximum a posteriori likelihood criterion means search-
ing for codeword ĉ such that:
ĉ = c ⇔ Pr {c |r } > Pr {c |r } , ∀c = c ∈ C(n, k)
Figure 4.5 – Performance of the algebraic decoding of the (15,7) BCH code. 4-PSK
transmission on a Gaussian channel.
Using the Bayes’ rule and assuming all the codewords equiprobable, the above
inequality can also be written:
ĉ = c if p(r |c ) > p(r |c ), ∀ c = c ∈ C(n, k) (4.26)
where p(r |c ) is the probability density function of observation r conditionally
to codeword c.
For a Gaussian channel, probability density function p(r |c ) is equal to:
⎛ ⎞
n
1 1
n−1 +
p(r |c ) = √ exp ⎝− 2 (rj − Es c̃j )2 ⎠
2πσb 2σb j=0
where σb2 is the variance of the noise.

Replacing the two probability density functions by their respective expres-
sions in inequality (4.26) and after some basic computation, we obtain:

n−1
n−1
ĉ = c ⇔ rj cj > rj c j , ∀c = c ∈ C(n, k)
j=0 j=0
The decoded codeword is the one that maximizes the scalar product r, c. We
could also show that the decoded
B codeword
B2 is the one that minimizes the square
√
of the Euclidean distance Br − Es c̃B .
4. Block codes 139
This decoding procedure is applicable when the number of codewords is not

too high. In the presence of a large number of codewords we can use a Chase
algorithm whose principle is to apply the above decoding procedure and restrict
the search space to a subset of codewords.
• Chase algorithm
The Chase algorithm is a sub-optimal decoding procedure that uses the max-
imum a posteriori likelihood criterion but considers a very reduced subset of
codewords. To determine this subset of codewords, the Chase algorithm works
in the following way.
• Step 1: The received word r, made up of analogue symbols, is transformed

into a word with binary symbols z0 = (z00 · · · z0j · · · z0n−1 ) by threshold-
ing,
z0j = sgn(rj )
with the following convention:
sgn(x) = 1 if x 0
= 0 if x < 0
The binary word z0 is then decoded by a hard decision algorithm other

than the maximum a posteriori likelihood algorithm (we will present algo-
rithms for decoding block codes later). Let c0 be the codeword obtained.
• Step 2: Let j1 , j2 , · · · , jt be the positions of the least reliable symbols, that

is, such that the |rj | amplitudes are the smallest.
2t − 1 words ei are built by forming all the non-null binary combina-
tions possible on positions j1 , j2 , · · · , jt . On the positions other than
j1 , j2 , · · · , jt , the symbols of ei are set to zero. Recall that t is the correc-
tion capability of the code.
• Step 3: Each of the 2t − 1 words ei is used to define the words zi with:
zi = z0 + ei
A hard decoder processes the words zi to obtain at most 2t − 1 codewords

ci . Note that the word at the output of the algebraic decoder is not
always a codeword and only codewords will be considered when applying
the decision criterion.
• Step 4: The maximum a posteriori likelihood rule is applied to the subset

of the codewords ci created in the previous step.
Example 4.13
Let there be a code C(n, k) with correction capability t = 3. The subset of

the codewords is made up of 8 codewords, 7 of which are elaborated from the
words ei . Words ei are words of length n whose components are null except
possibly those with indices j1 , j2 and j3 (see the table below).
i ei,j1 ei,j2 ei,j3

1 001
2 010
3 011
4 100
5 101
6 110
7 111
• Probability of erroneous decoding of a codeword

Let us assume that the transmitted codeword is c0 = (c01 · · · c0j · · · c0n−1 )
and let r0 = (r0 · · · rj · · · rn−1 ) be the received word with:
+
rj = Es c̃0j + bj
Codeword c0 will be wrongly decoded if:

n−1
n−1
rj c0,j < rj cl,j ∀ cl = c0 ∈ C(n, k)
j=0 j=0
The code being linear, we can, without loss of generality, assume that the code-
word transmitted is the null word, that is, c0,j = 0 for j = 0, 1, · · · , n − 1.
The probability of erroneous decoding Pe,word of a codeword is then equal
to: ⎛ ⎞

n−1
n−1
Pe,word = Pr ⎝ rj c1,j > 0 or . . . rj cl,j > 0 or . . .⎠
j=0 j=0
Probability Pe,word can be upper bounded by a sum of probabilities and,

after some standard computation, it can be written in the form:
2k
1 Es
Pe,word ≤ erfc wj
2 j=2 N0
where wj is the Hamming weight of the j-th codeword.

4. Block codes 141
Assuming that code C(n, k) has Aw codewords of weight w, probability

Pe,word can again be written in the form:

1
n
Es
Pe,word < Aw erfc w (4.27)
2 N0
w=dmin
Introducing the energy Eb received per bit of information transmitted, proba-

bility Pe,word can finally be upper bounded by:

1
n
REb
Pe,word < Aw erfc w (4.28)
2 N0
w=dmin
where R is the coding rate.

We can also establish an upper bound of the binary error probability on the
information symbols after decoding.

1 w
n
REb
Pe,bit < Aw erfc w (4.29)
2 n N0
w=dmin
To calculate probabilities Pe,word and Pe,bit we must know the number Aw

of codewords of weight w. For extended BCH codes the quantities Aw are given
in [4.1].
As an example, Table 4.5 gives the Aw quantities for three extended Ham-
ming codes.
n k dmin A4 A6 A8 A10 A12 A14 A16

8 4 4 14 - 1 - - - -
16 11 4 140 448 870 448 140 - 1
32 26 4 1240 27776 330460 2011776 7063784 14721280 18796230
Table 4.5 – Aw for three extended Hamming codes.
For the code (32,26) the missing Aw quantities are obtained from the relation
Aw = An−w for 0 ≤ w ≤ n/2, n/2 even.
The Aw quantities for non-extended Hamming codes can be deduced from
those of extended codes by resolving the following system of equations:
(n + 1)Aw−1 = wAwextended
wAw = (n + 1 − w)Aw−1
where n is the length of the words of the non-extended code.
For the Hamming code (7,4), for example, the Aw quantities are:
8A3 = 4A4extended A3 = 7
4A4 = 4A3 A4 = 7
8A7 = 8A8extended A7 = 1
Weight Aw (23,12) Aw (24,12)

0 1 1
7 253 0
8 506 759
11 1288 0
12 1288 2576
15 506 0
16 253 759
23 1 0
24 0 1
Table 4.6 – Aw for extended Golay and Golay codes.
For Golay and extended Golay codes, the Aw quantities are given in Table 4.6.
With a high signal to noise ratio, error probability Pe,word is well approxi-
mated by the first term of the series:

∼ Ad erfc Rdmin Eb if Eb >> 1
Pe,word =
1
(4.30)
2 min N0 N0
The same goes for error probability Pe,bit on the information symbols.
dmin Eb
Pe,bit ∼
= Pe,word if >> 1 (4.31)
n N0
In the absence of coding, the error probability on the binary symbols is equal
to:
1 Eb
p = erfc
2 N0
As seen in Section 1.5, comparing the two expressions of the binary error proba-
bility with and without coding, we observe that the signal to noise ratio Eb /N0
is multiplied by Rdmin in the presence of coding. If this multiplying coefficient is
higher than 1, the coding acts as an amplifier of the signal to noise ratio whose
asymptotic gain is approximated by
Ga = 10 log(Rdmin )(dB)
To illustrate these bounds, let us again take the example of the (15,7) BCH
code transmitted on a Gaussian channel with 4-PSK modulation. In Figure 4.6,
we show the evolution of the binary error probability and word error probability
obtained by simulation from the sub-optimal Chase algorithm (4 non-reliable
positions). We also show the first two terms of the sums appearing in the
bounds given by (4.28) and (4.29). As a reference, we have also plotted the
binary error probability curve of a 4-PSK modulation without coding.
4. Block codes 143
Figure 4.6 – Performance of the soft input decoding of the (15,7) BCH code. 4-PSK
transmission on a Gaussian channel.
4.4 Decoding and performance of codes with non-

binary symbols
4.4.1 Hard input decoding of Reed-Solomon codes
Hard input decoding algorithms make it possible to decode Reed-Solomon (RS)
codes and BCH codes with binary symbols. We begin by presenting the principle
of decoding RS codes then we treat the case of BCH codes using binary symbols
as a particular case of decoding RS codes.
Assuming that c(x) is the transmitted codeword, then for a channel with
discrete input and output, the received word can always be written in the form:
r(x) = c(x) + e(x)
with:
e(x) = e0 + e1 x + · · · + ej xj + · · · + en−1 xn−1 , e j ∈ Fq ∀j
When ej = 0 there is an error in position j.

It was seen above that the generator polynomial of an RS code or of a BCH
code (with l = 1) correcting t errors had the roots α, · · ·, αj , · · ·, α2t and that the
codewords were multiples of the generator polynomial. Thus, for any codeword,
we can write:
c(αi ) = 0; ∀ i = 1, 2, · · · , 2t
Decoding RS codes and binary BCH codes can be performed from a vector with
2t components S = [S1 · · · Sj · · · S2t ], called a syndrome.
Sj = r(αj ) = e(αj ), j = 1, 2, · · · , 2t (4.32)
When the components of vector S are all null, there are no errors or, at least, no
detectable errors. When some components of the vector S are non-null, errors
are present that, in certain conditions, can be corrected.
In the presence of t transmission errors, the error polynomial e(x) is of the
form:
e(x) = en1 xn1 + en2 xn2 + · · · + ent xnt
where the enl are non-null coefficients taking their value in the field Fq .
The components Sj of syndrome S are equal to:
Sj = en1 (αj )n1 + · · · + enl (αj )nl + · · · + ent (αj )nt
Putting Zl = αnl and, to simplify the notations enl = el , the component Sj
of the syndrome is again equal to:
Sj = e1 Z1j + · · · + el Zlj + · · · + et Ztj (4.33)
To determine the position of the transmission errors it is therefore sufficient
to know the value of quantities Zl ; j = 1, 2, · · · , t then, in order to correct the
errors, to evaluate coefficients el ; l = 1, 2, · · · , t.
The main difficulty in decoding RS codes or binary BCH codes is determining
the position of the errors. Two methods are mainly used to decode RS codes
or binary BCH codes: Peterson’s direct method and the iterative method using
the Berlekamp-Massey algorithm or Euclid algorithm .
4.4.2 Peterson’s direct method

Description of the algorithm for codes with non-binary symbols
This method is well adapted for decoding RS codes or binary BCH codes cor-
recting a low number of errors, typically 1 to 3. Indeed, the complexity of this
method increases as the square of the correction capability of the code, whereas
for the iterative method, the complexity increases only linearly with the correc-
tion capability of the code.
To determine the position of the errors let us introduce a polynomial σd (x)
called the error locator polynomial whose roots are exactly the quantities Zl .
5
t
σd (x) = (x + Zl )
l=1
4. Block codes 145
Developing this expression, the polynomial σd (x) is again equal to:
σd (x) = xt + σ1 xt−1 + · · · + σj xt−j + · · · + σt
where the coefficients σj are functions of the quantities Zl .

From the expression of Sj we can build a non-linear system of 2t equations.

t
Sj = ei Zij , j = 1, 2, · · · , 2t
i=1
The quantities Zl , l = 1, · · · , t being the roots of the error locator polynomial

σd (x), we can write:

t
σd (Zl ) = Zlt + σj Zlt−j = 0, l = 1, 2, · · · , t (4.34)
j=1
Multiplying the two parts of this expression by the same term el Zlq , we
obtain:

t
el Zlt+q + σj el Zlt+q−j = 0, l = 1, 2, · · · , t (4.35)
j=1
Summing relations (4.35) for l from 1 to t and taking into account the defi-
nition of component Sj of syndrome S, we can write:
St+q + σ1 St+q−1 + · · · + σj St+q−j + · · · + σt Sq = 0, ∀q (4.36)
For an RS code correcting one error (t = 1) in a block of n symbols, syndrome

S has two components S1 and S2 . Coefficient σ1 of the error locator polynomial
is determined from relation (4.36) by making t = 1 and q = 1.
S2
S2 + σ1 S1 = 0 → σ1 = (4.37)
S1
In the same way, for an RS code correcting two errors (t = 2) in a block of

n symbols, the syndrome has four components S1 , S2 , S3 , S4 . Using relation
(4.36) with t = 2 and q = 1, 2 we obtain the following system with two equations:
σ1 S2 + σ2 S1 = S3
σ1 S3 + σ2 S2 = S4
Resolving this system of two equations enables us to determine coefficients σ1

and σ2 of the error locator polynomial.
1
σ1 = Δ2 [S
1 S4 + S22S3 ]
1 (4.38)
σ2 = Δ2 S2 S4 + S3
where Δ2 is the determinant of the system with two equations.
Δ2 = S22 + S1 S3
Finally, for an RS code correcting three errors (t = 3), the relation (4.36) with
t = 3 and q = 1, 2, 3 leads to the following system of three equations:
σ1 S3 + σ2 S2 + σ3 S1 = S4
σ1 S4 + σ2 S3 + σ3 S2 = S5
σ1 S5 + σ2 S4 + σ3 S3 = S6
The resolution of this system enables us to determine coefficients σ1 , σ2 and σ3

of the error locator polynomial.

σ1 = Δ13 S1 S3 S6 + S1 S4 S5 + S22 S6 + S2 S3 S5 + S2 S42 + S32 S4
σ2 = Δ13 S1 S4 S6 + S1 S52 + S2 S3 S6 + S 2 S4 S5 + S32 S5 + S3 S42 (4.39)
1 2 2 3
σ3 = Δ3 S2 S4 S6 + S2 S5 + S3 S6 + S4
where Δ3 is the determinant of the system with three equations.
Δ3 = S1 S3 S5 + S1 S42 + S22 S5 + S33
Implementation of Peterson’s decoder for an RS code with parameter t = 3
1. Calculate the 2t syndromes Sj : Sj = r(αj )
2. Determine the number of errors:
• Case (a) Sj = 0, ∀ j: no detectable error.

• Case (b) Δ3 = 0: presence of three errors.
• Case (c) Δ3 = 0 and Δ2 = 0: presence of two errors.
• Case (d) Δ3 = Δ2 = 0 and S1 = 0: presence of one error.
3. Calculate the error locator polynomial σd (x)
• Case (b) Use (4.39)
• Case (c) Use (4.38)
• Case (d) Use (4.37)
4. Look for the roots of σd (x) in field Fq
5. Calculate the error coefficients ei

4. Block codes 147
• Case (b)
1

ei = Δ S1 (Zk2 Zp3 + Zk3 Zp2 ) + S2 (Zk3 Zp + Zk Zp3 ) + S3 (Zk2 Zp + Zk Zp2 ) ,
k = p = i, (i, k, p) ∈ {1, 2, 3}3

Δ= Z1i1 Z2i2 Z3i3
1 ≤ i 1 , i2 , i3 ≤ 3
i1 + i2 + i3 = 6
i1 = i2 = i3
• Case (c)
S1 Zp + S2
ei = , p = i, (i, p) ∈ {1, 2}2
Zi (Z1 + Z2 )
• Case (d)
S12
e1 =
S2
6. Correct the errors: ĉ(x) = r(x) + e(x)
Example 4.14
To illustrate the decoding of an RS code using the direct method, we now

present an example considering an RS code correcting up to three errors (t = 3)
and having the following parameters:
m = 4 q = 16 n = 15 n − k = 6
Let us assume, for example, that the transmitted codeword is c(x) = 0 and that
the received word has two errors.
r(x) = α7 x3 + α3 x6
1. Calculate the components of the syndrome
S1 = α10 + α9 = α13 S4 = α19 + α27 = α6

S2 = α13 + α15 = α6 S5 = α22 + α33 = α4
S3 = α16 + α21 = α11 S6 = α25 + α39 = α13
2. Determine the number of errors
Δ3 = 0 Δ2 = α8
Δ3 being null and Δ2 = 0, we have two errors.

3. Calculate coefficients σ1 and σ2 of the error locator polynomial.

19 17
1
σ1 = [S1 S4 + S2 S3 ] = α α+α
8 = α11 + α9 = α2
Δ2
1
12 22
σ2 = Δ2 S2 S4 + S32 = α α+α8 = α4 + α14 = α9
The error locator polynomial is therefore equal to:
σd (x) = x2 + α2 x + α9
4. Look for the two roots of the error locator polynomial.

Looking through the elements of field F16 we find that α3 and α6 cancel
the polynomial Λ(x). The errors therefore concern the terms in x3 and in
x6 of word r(x).
5. Calculate the error coefficients e1 and e2 .

S1 Z2 +S2 α19 +α6 α12
e1 = Z1 Z2 +Z12
= α9 +α6 = α5 = α7
S1 Z1 +S2 α16 +α6 α11
e2 = Z1 Z2 +Z22
= α9 +α12
= α8
= α3
6. Correct the errors
c(x) = (α7 x3 + α3 x6 ) + (α7 x3 + α3 x6 ) = 0
The transmitted codeword is the null word; the two errors have therefore
been corrected.
Simplification of Peterson’s algorithm for binary codes

For BCH codes with binary symbols it is not necessary to calculate the coeffi-
cients ej . Indeed, as these coefficients are binary, they are necessarily equal to 1
in the presence of an error in position j. The computation of coefficients σj can
also be simplified by taking into account the fact that for a code with binary
symbols we have:
2
S2j = e(α2j ) = e(αj ) = Sj2
For a BCH code with binary symbols correcting up to t = 3 errors, taking into
account the previous remark and using the expressions of the three coefficients
σj of the error locator polynomial, we obtain:
σ1 = S1
S12 S3 +S5
σ2 = S13 +S3 (4.40)
S 2 S3 +S5
σ3 = (S13 + S3 ) + S1 S1 3 +S3
1
4. Block codes 149
For a BCH code with binary symbols correcting up to t = 2 errors, also taking
into account the previous remark and using the expressions of the two coefficients
σj of the error locator polynomial, we obtain:
σ1 = S1
S3 +S13 (4.41)
σ2 = S1
Finally, in the presence of an error σ2 = σ3 = 0 and σ1 = S1 .
Example 4.15
Let us consider a BCH code correcting two errors (t = 2) in a block of n = 15

symbols of a generator polynomial equal to:
g(x) = x8 + x7 + x6 + x4 + 1
Let us assume that the transmitted codeword is c(x) = 0 and that the received
word r(x) has two errors.
r(x) = x8 + x3
There are three steps to the decoding: calculate syndrome S, determine the
coefficients σl of the error locator polynomial and search for its roots in field
F16 .
1. Calculate syndrome S: we only need to calculate the odd index components

S1 and S3 of syndrome S. Using the binary representations of the elements
of field F16 given in the appendix, and taking into account the fact that
α15 = 1, we have:
S1 = r(α) = α8 + α3 = α13
S3 = r(α3 ) = α24 + α9 = α9 + α9 = 0
2. Determine the coefficients σ1 and σ2 of the error locator polynomial. Using

the expressions of coefficients σ1 and σ2 , we obtain:
σ1 = S1 = α13
S +S 3
σ2 = 3S1 1 = S12 = α26 = α11 (α15 = 1)
and the error locator polynomial is equal to:
σd (x) = x2 + α13 x + α11
3. Search for the roots of the error locator polynomial in field F16 . By trying
all the elements of field F16 , we can verify that the roots of the error
locator polynomial are α3 and α8 . Indeed, we have
σ(α3 ) = α6 + α16 + α11 = α6 + α + α11 = 1100 + 0010 + 1110 = 0000

σ(α8 ) = α16 + α21 + α11 = α + α6 + α11 = 0010 + 1100 + 1110 = 0000
The transmission errors concern the terms x8 and x3 of received word r(x).
The transmitted codeword is therefore c(x) = 0 and the two errors have
been corrected.
The reader can verify that in the presence of a single error, r(x) = xj ;
0 ≤ j ≤ (n − 1), the correction is still performed correctly since:
S1 = αj ; S3 = α3j ; σ1 = αj ; σ2 = 0; σd (x) = x(x + σ1 )
and the error locator polynomial has one sole root σ1 = αj .
Chien algorithm
To search for the error locator polynomial roots in the case of codes with binary
symbols, we can avoid going through all the elements of field Fq by using Chien’s
iterative algorithm.
Dividing polynomial σd (x) by xt , we obtain:
σd (x)
σ̃d (x) = = 1 + σ1 x−1 + · · · + σj x−j + · · · + σt x−t
xt
The roots of polynomial σd (x) that are also the roots of σ̃d (x) have the form
αn−j where j = 1, 2, . . ., n − 1 and n = q − 1.
Thus αn−j is a root of σ̃d (x) if:
σ1 α−n+j + · · · + σp x−np+jp + · · · + σt x−nt+jt = 1
Taking into account the fact that αn = 1, the condition to satisfy in order for
αn−j to be a root of the error locator polynomial is:

t
σp αjp = 1; j = 1, 2, · · · , (n − 1) (4.42)
p=1
Chien’s algorithm has just tested whether condition (4.42) is satisfied using the
circuit shown in Figure 4.7.
This circuit has a register with t memories initialized with the t coefficients
σj of the error locator polynomial and a register with n memories that stocks
symbols rj ; j = 0, 1, · · · , (n − 1) of word r(x). At the first clock pulse, the
circuit performs the computation of the left-hand part of expression (4.42) for
j = 1. If the result of this computation is equal to 1, αn−1 is a root of the error
locator polynomial and the error that concerned symbol rn−1 is then corrected.
If the result of this computation is equal to 0, no correction is performed. At
the end of this first phase, the σj coefficients contained in the t memories of
the register are replaced by σj αj . At the second clock pulse the circuit again
4. Block codes 151
Figure 4.7 – Schematic diagram of the circuit implementing the Chien algorithm.
performs the computation of the left-hand part of expression (4.42) for j = 2. If

the result of this computation is equal to 1, αn−2 is a root of the error locator
polynomial and the error that concerned symbol rn−2 is then corrected. The
algorithm continues in the same way for the following clock pulses.
4.4.3 Iterative method

Decoding RS codes or binary BCH codes with the iterative method uses two
polynomials, error locator polynomial Λ(x) and error evaluator polynomial Γ(x).
These two polynomials are defined respectively by:
5
t
Λ(x) = (1 + Zj x) (4.43)
j=1

t
Λ(x)
Γ(x) = ei Zi x (4.44)
i=1
1 + Zi x
The error locator polynomial whose roots are Zj−1 enables the position of the
errors to be determined and the error evaluator polynomial enables the value
of the error ej to be determined. Indeed, taking into account the fact that
Λ(Zj−1 ) = 0, the polynomial Γ(x) taken in Zj−1 is equal to:
'
Γ(Zj−1 ) = ej (1 + Zp Zj−1 )
p=j
= ej Zj−1 Λ (Zj−1 )
where Λ (x) = dΛ
dx (x).
The value of error ej is then given by the Forney algorithm:
Γ(Zj−1 )
e j = Zj (4.45)
Λ (Zj−1 )
Introducing the polynomial S(x) defined by:

2t
S(x) = Sj xj (4.46)
j=1
we can show that:

Λ(x)S(x) ≡ Γ(x) modulo x2t+1 (4.47)
This relation is called the key equation for decoding a cyclic code.
To determine polynomials Λ(x) and Γ(x) two iterative algorithms are mainly
used, the Berlekamp-Massey algorithm and Euclid’s algorithm.
Berlekamp-Massey algorithm for codes with non-binary symbols

Computation of polynomials Λ(x) and Γ(x) using the Berlekamp-Massey al-
gorithm is performed iteratively. It requires two intermediate polynomials
denoted Θ(x) and Ω(x). The algorithm has 2t iterations. Once the algorithm
has terminated, the Chien algorithm must be implemented to determine the
roots Zj−1 of Λ(x) and consequently the position of the errors. Next, the Forney
algorithm expressed by (4.45) enables the value of the errors ej to be calculated.
Initial conditions:
L0 = 0
Λ(0) (x) = 1 Θ(0) (x) = 1
Γ(0) (x) = 0 Ω(0) (x) = 1
Recursion: 1 ≤ p ≤ 2t
(p−1)
Δp = Λj Sp−j
j
δp = 1 if Δp = 0 and 2Lp−1 ≤ p − 1
= 0 otherwise
Lp = δp (p − Lp−1 ) + (1 − δp )Lp−1

Λ(p) Γ(p) 1 Δp x Λ(p−1) Γ(p−1)
=
Θ(p) Ω(p) Δ−1
p δp (1 − δp )x Θ(p−1) Ω(p−1)
Termination :
Λ(x) = Λ(2t) (x)
Γ(x) = Γ(2t) (x)
4. Block codes 153
Example 4.16
To illustrate the decoding of an RS code using the Berlekamp-Massey algo-

rithm, let us consider an RS code correcting up to two errors (t = 2) and having
the following parameters:
m = 4; q = 16; n = 15; n−k =4
Let us assume, for example, that the transmitted codeword is c(x) = 0 and that
the received word has two errors.
r(x) = α7 x3 + α3 x6
The set of calculations performed to decode this RS code will be done in field
F16 whose elements are given in the appendix.
1. Calculate syndrome S = (S1 , S2 , S3 , S4 )
S1 = α10 + α9 = α13 S3 = α16 + α21 = α11

S2 = α13 + α15 = α6 S4 = α19 + α27 = α6
The polynomial S(x) is therefore equal to:
S(x) = α13 x + α6 x2 + α11 x3 + α6 x4
2. Calculate polynomials Λ(x) and Γ(x) from the Berlekamp-Massey algo-

rithm
p Δp δp Lp Λp (x) Θp (x) Γp (x) Ωp (x)

0 0 1 1 0 1
1 α13 1 1 1 + α13 x α2 α13 x 0
2 α 0 1 1 + α8 x α2 x α13 x 0
3 α10 1 2 1 + α8 x + α12 x2 α5 + α13 x α13 x α3 x
4 α10 0 2 1 + α2 x + α9 x2 α5 x + α13 x2 α13 x + α13 x2 α3 x2
In the table above, all the calculations are done in field F16 and take into
account the fact that α15 = 1.
The error locator and error evaluator polynomials are:
Λ(x) = 1 + α2 x + α9 x2
Γ(x) = α13 x + α13 x2
We can verify that the key equation for the decoding has been satisfied.
Indeed, we do have:
Λ(x)S(x) = α13 x + α13 x2 + α4 x5 + x6 ≡ α13 x + α13 x2 = Γ(x) modulo x5

3. Search for the roots of the error locator polynomial

By looking through all the elements of field F16 we find that α12 and
α9 are roots of polynomial Λ(x). The errors are therefore in position
x3 (α−12 = α3 ) and x6 (α−9 = α6 ) and error polynomial e(x) is equal to:
e(x) = e3 x3 + e6 x6
4. Calculate error coefficients ej (4.45).

6
α
e3 = α3 α 2 = α
7
14
α
e6 = α6 α2 = α3
Error polynomial e(x) is therefore equal to:
e(x) = α7 x3 + α3 x6
and the estimated codeword is ĉ(x) = r(x) + e(x) = 0. The two transmis-
sion errors are corrected.
Euclid’s algorithm
Euclid’s algorithm enables us to solve the key equation for decoding, that is, to
determine polynomials Λ(x) and Γ(x).
Initial conditions:
R−1 (x) = x2t ; R0 (x) = S(x); U−1 (x) = 0; U0 (x) = 1

Recursion:
calculate Qj (x), Rj+1 (x) and Uj+1 (x) from the two following expressions:
Rj−1 (x) R (x)
Rj (x)= Qj (x) + Rj+1
j (x)
Uj+1 (x) = Qj (x)Uj (x) + Uj−1 (x)
When deg(Uj ) ≤ t and deg(Rj ) ≤ t then:
Λ(x) = Uj+1 (x)

Γ(x) = Rj+1 (x)
Example 4.17
Let us again take the RS code used to illustrate the Berlekamp-Massey al-
gorithm. Assuming that the received word is always r(x) = α7 x3 + α3 x6 when
the transmitted codeword is c(x) = 0, the decoding algorithm is the following:
4. Block codes 155
1. Calculate syndrome S = (S1 , S2 , S3 , S4 )
S1 = α10 + α9 = α13 S3 = α16 + α21 = α11

S2 = α13 + α15 = α6 S4 = α19 + α27 = α6
Polynomial S(x) is therefore equal to:
S(x) = α13 x + α6 x2 + α11 x3 + α6 x4
2. Calculate polynomials Λ(x) and Γ(x) from Euclid’s algorithm (the calcula-
tions are performed in field F16 whose elements are given in the appendix).
j = 0 j = 1
R−1 (x) = x5 R0 (x) = S(x)
R0 (x) = S(x) R1 (x) = α5 x3 + α13 x2 + α12 x
Q0 (x) = α9 x + α14 Q1 (x) = αx + α5
R1 (x) = α5 x3 + α13 x2 + α12 x R2 (x) = α14 x2 + α14 x
U1 (x) = α9 x + α14 U2 (x) = α10 x2 + α3 x + α
We can verify that deg(U2 (x)) = 2 is lower than or equal to t (t = 2)

and that deg(R2 (x)) = 2 is lower than or equal to t. The algorithm is
therefore terminated and polynomials Λ(x) and Γ(x) respectively have the
expression:
Λ(x) = U2 (x) = α + α3 x + α10 x2 = α(1 + α2 x + α9 x2 )

Γ(x) = R2 (x) = α14 x + α14 x2 = α(α13 x + α13 x2 )
We can verify that the key equation for the decoding is satisfied and that
the two polynomials obtained are identical, to within one coefficient α, to
those determined using the Berlekamp-Massey algorithm.
The roots of the polynomial Λ(x) are therefore 1/α3 and 1/α6 , and error
polynomial e(x) is equal to:
e(x) = α7 x3 + α3 x6
Calculating coefficients ej by a transform

It is possible to calculate the coefficients ej ; j = 0, 1, · · · , (n − 1) of error poly-
nomial e(x) without determining the roots of the error locator polynomial Λ(x).
To do this, we introduce the extended syndrome S ∗ (x) defined by:
1 + xn n
∗
S (x) = Γ(x) = Sj xj (4.48)
Λ(x) j=1
Coefficient ej is null (no errors) if α−j is not a root of error locator polynomial
Λ(x). In this case, we have S ∗ (α−j ) = 0 since α−jn = 1 (recall that n = q − 1
and αq−1 = 1).
A contrario if α−j is a root of the locator polynomial, coefficient ej is non-null
(presence of an error) and S ∗ (α−j ) is of the form 0/0. This indetermination can
be removed by calculating the derivation of the numerator and the denominator
of expression (4.48).
nα−j(n−1)
S ∗ (α−i ) = Γ(α−i ) −j
Λ (α )
Using Equation (4.45) and taking into account the fact that α−j(n−1) = αj and
that na = a for n odd in a Galois field, coefficient ej is equal to:
ej = S ∗ (α−j ) (4.49)
The extended syndrome can be computed from polynomials Λ(x) and Γ(x) using
the following relation deduced from expression (4.48).
Λ(x)S ∗ (x) = Γ(x)(1 + xn ) (4.50)
Coefficients Sj of the extended syndrome are identical to those of syndrome

S(x) for j from 1 to 2t and are determined by cancelling the coefficients of the
xj terms in the product Λ(x)S ∗ (x), for j from 2t + 1 to n.
Example 4.18
Again taking the example of the RS code (q = 16; n = 15; k = 11; t = 2)

used to illustrate the Berlekamp-Massey algorithm let us determine the extended
syndrome.

15
S ∗ (x) = Sj xj
j=1
with:
S1 = α13 S3 = α11
S2 = α6 S4 = α6
Equation (4.50) provides us with the following relation:
S(x) + α2 xS(x) + α9 x2 S(x) = α13 (x + x2 + x16 + x17 )
S1 x + (α2 S1 + S2 )x2

15
+ (α9 Sk−2 + α2 Sk−1 + Sk )xk
k=3
+(α2 S15 + α9 S14 )x16 + α9 S15 x17 = α13 (x + x2 + x16 + x17 )
4. Block codes 157
From this there results the recurrence relation:

Sk = α2 Sk−1 + α9 Sk−2 , k = 3, 4, · · · , 15
We thus obtain the coefficients of the extended syndrome:
S5 = α4 , S6 = α13 , S7 = α6 , S8 = α11 , S9 = α6 , S10 = α4
S11 = α13 , S12 = α6 , S13 = α11 , S14 = α6 , S15 = α4
Another way to obtain the extended polynomial involves dividing Γ(x)(1 + xn )
by Λ(x) by increasing power orders.
The errors being with monomials x3 and x6 , let us calculate coefficients e3 and
e6 .
e3 = S ∗ (α12 ) = α2 + α4 + α10 + α7 = α7
e6 = S ∗ (α9 ) = α4 + α7 = α3
The values found for coefficients e3 and e6 are obviously identical to those ob-
tained in example 4.16. We can verify that the other ej coefficients are all null.
Berlekamp-Massey algorithm for binary cyclic codes

For binary BCH codes the Berlekamp-Massey algorithm can be simplified since
it is no longer necessary to determine the error evaluator polynomial, and since
it is possible to show that the Δj terms are null for j even. This implies:
δ2p = 0
L2p = L2p−1
Λ(2p) (x) = Λ(2p−1) (x)
Θ(2p) (x) = xΘ(2p−1) (x)
Hence the algorithm in t iterations:
Initial conditions:
L−1 = 0
Λ(−1) (x) = 1 Θ(−1) (x) = x−1
Recursion: 0 ≤ p ≤ t − 1
(2p−1)
Δ2p+1 = Λj S2p+1−j
j
δ2p+1 = 1 if Δ2p+1 = 0 and L2p−1 ≤ p
= 0 if not
L2p+1 = δ2p+1 (2p + 1 − L2p−1 ) + (1 − δ2p+1 )L2p−1

Λ(2p+1) 1 Δ2p+1 x2 Λ(2p−1)
=
Θ(2p+1) Δ−1
2p+1 δ2p+1 (1 − δ2p+1 )x2 Θ(2p−1)
Termination:
Λ(x) = Λ(2t−1) (x)
Example 4.19
Again taking the BCH code that was used to illustrate the computation of
the error locator polynomial with the direct method, let us assume that the
received word is r(x) = x8 + x3 when the transmitted codeword is c(x) = 0.
1. Syndrome S has four components.
S1 = r(α) = α8 + α3 = α13
S3 = r(α3 ) = α24 + α9 = 0
S2 = S12 = α26 = α11
S4 = S22 = α22 = α7
Polynomial S(x) is equal to:
S(x) = α13 x + α11 x2 + α7 x4
2. Calculate polynomial Λ(x) from the Berlekamp-Massey algorithm
p Δ2p+1 δ2p+1 L2p+1 Λ2p+1 (x) Θ2p+1 (x)

-1 0 1 x−1
0 α13 1 1 1 + α13 x α2
1 α9 1 2 1 + α13 x + α11 x2 α6 + α4 x
Note that polynomial Λ(x) obtained is identical to that determined using

the direct method. The roots of Λ(x) are 1/α3 and 1/α8 , and the errors
therefore concern terms x3 and x8 . The estimated codeword is ĉ(x) = 0.
Euclid’s algorithm for binary codes

Example 4.20
Let us again take the decoding of the (15,7) BCH code. The received word
is r(x) = x8 + x3 .
j = 0 j = 1
R−1 (x) = x5 R0 (x) = S(x)
R0 (x) = S(x) R1 (x) = α4 x3 + α6 x2
Q0 (x) = α8 x Q1 (x) = α3 x + α5
R1 (x) = α x + α 6 x2
4 3
R2 (x) = α13 x
U1 (x) = α8 x U2 (x) = α x + α13 x + 1
11 2
4. Block codes 159
We can verify that deg(U2 (x)) = 2 is lower than or equal to t (t = 2) and

that the degree of R2 is lower than or equal to t. The algorithm is therefore
terminated and polynomial Λ(x) has the expression:
Λ(x) = U2 (x) = 1 + α13 x + α11 x2

Γ(x) = R2 (x) = α13 x
For a binary BCH code it is not necessary to use the error evaluator polynomial
to determine the value of coefficients e3 and e8 . However, we can verify that:
−3
e3 = α3 ΛΓ(α )
(α−3 ) = 1
−8
e8 = α8 ΛΓ(α )
(α−8 ) = 1
The decoded word is therefore ĉ(x) = r(x) + e(x) = 0 and the two errors have
been corrected.
4.4.4 Hard input decoding performance of Reed-Solomon

codes
Recall that for a Reed-Solomon code, the blocks of information to encode and the
codewords are made up of k and n = q − 1 (q = 2m ) q-ary symbols respectively.
The probability Pe,word of having a wrong codeword after hard decoding can be
upper bounded by:

n
n
Pe,word ≤ pjs (1 − ps )n−j (4.51)
j
j=t+1
where ps is the error probability per q-ary symbol on the transmission channel
and t is the code correction capability in number of q-ary symbols.
When a codeword is wrongly decoded, the error probability per corresponding
Pe,symbol symbol after decoding is upper bounded by:

1
n
n
Pe,symbol ≤ (j + t) pjs (1 − ps )n−j (4.52)
n j=t+1 j
The binary error probability after decoding is obtained from the error probability
per symbol, taking into account that a symbol is represented by m bits:
1
Pe,bit = 1 − (1 − Pe,symbol ) m
At high signal to noise ratio, we can approximate the binary error probability
after decoding:
∼ 1 Pe,symbol Eb >> 1 .
Pe,bit =
m N0
Bibliography
[4.1] R.H. Morelos-Zaragoza. The Art of Error Correcting Coding. John Wiley
& sons, 2005.
[4.2] J. G. Proakis. Digital Communications. McGraw-Hill, New-York, 4th

edition, 2000.
Appendix
Notions about Galois fields

and minimal polynomials
Definition
A Galois field with q = 2m elements denoted Fq , where m is a positive integer is
defined as a polynomial extension of the field with two elements (0, 1) denoted
F2 . The polynomial ϕ(x) used to build field Fq must be
• irreducible, that is, non factorizable in F2 (in other words, 0 and 1 are not
roots of ϕ(x)),
• of degree m,
• and with coefficients in F2 .
The elements of a Galois field Fq are defined modulo ϕ(x) and thus, each element
of this field can be represented by a polynomial with degree at most equal to
(m − 1) and with coefficients in F2 .
Example 1
Consider an irreducible polynomial ϕ(x) in the field F2 of degree m = 2.
ϕ(x) = x2 + x + 1
This polynomial enables a Galois field to be built with 4 elements. The elements
of this field F4 are of the form:
aα + b where a, b ∈ F2
that is:
F4 : {0, 1, α, α + 1}
We can see that if we raise element α to successive powers 0, 1 and 2 we obtain

all the elements of field F4 with the exception of element 0. Indeed, α2 is still
equal to (α + 1) modulo ϕ(α). Element α is called the primitive element of field
F4 .
The elements of field F4 can also be represented in binary form:
F4 : {00, 01, 10, 11}
The binary couples correspond to the four values taken by coefficients a and b.
Primitive element of a Galois field

We call the primitive element of a Galois field Fq , an element of this field that,
when it is raised to successive powers 0, 1, 2, · · · , (q − 2); q = 2m , makes it
possible to retrieve all the elements of the field except element 0. Every Galois
field has at least one primitive element. If α is a primitive element of field Fq
then, the elements of this field are:

Fq = 0, α0 , α1 , · · · , αq−2 with αq−1 = 1
Note that in such a Galois field the "-" sign is equivalent to the "+" sign, that
is:
−αj = αj ∀j ∈ {0, 1, · · · , (q − 2)}
Observing that 2αj = 0 modulo 2, we can always add the zero quantity 2αj to
−αj and we thus obtain the above equality.
For example, for field F4 let us give the rules that govern the addition and
multiplication operations. All the operations are done modulo 2 and modulo
α2 + α + 1.
+ 0 1 α α2
0 0 1 α α2
1 1 0 1 + α = α2 1 + α2 = α
α α 1 + α = α2 0 α + α2 = 1
α2 α2 1 + α2 = α α + α2 = 1 0
Table 4.7 – Addition in field F4 .
Minimal polynomial with coefficients in F2 associ-

ated with an element of a Galois field Fq
The minimal polynomial mβ (x) with coefficients in F2 associated with any
element β of a Galois field Fq , is a polynomial of degree at most equal to
4. Block codes 163
× 0 1 α α2
0 0 0 0 0
1 0 1 α α2
α 0 α α2 α3 = 1
α2 0 α2 3
α =1 α4 = α
Table 4.8 – Multiplication in field F4 .
m = log2 (q), having β as a root. This polynomial is unique and irreducible

in F2 . If β is a primitive element of Galois field Fq then polynomial mβ (x) is
exactly of degree m. Note that a polynomial with coefficients in F2 satisfies the
following property:
2 2p p
[f (x)] = f (x2 ) ⇒ [f (x)] = f (x2 )
So, if β is a root of polynomial f (x) then β 2 , β 4 , · · · are also roots of this poly-
nomial. The minimal polynomial with coefficients in F2 having β as a root can
then be written in the form:
mβ (x) = (x + β)(x + β 2 )(x + β 4 ) · · ·
If β is a primitive element of Fq , the minimal polynomial with coefficients in F2

being of degree m, it can also be written:
m−1
mβ (x) = (x + β)(x + β 2 )(x + β 4 ) · · · (x + β 2 )
Example 2
Let us calculate the minimal polynomial associated with the primitive element
α of Galois field F4 .
F4 : 0, 1, α, α2
The minimal polynomial associated with element α therefore has α and α2
(m = 2) as roots, and can be expressed:
mα (x) = (x + α)(x + α2 ) = x2 + x(α + α2 ) + α3
Taking into account the fact that α3 = 1 and that α + α2 = 1 in field F4 , the
polynomial mα (x) is thus equal to:
mα (x) = x2 + x + 1
Minimal polynomial with coefficients in Fq associ-

ated with an element in a Galois field Fq
The minimal polynomial mβ (x), with coefficients in the Galois field Fq associ-
ated with an element β = αj (α a primitive element of field Fq ) of this field, is
the lowest degree polynomial having β as a root.
Recalling that for a polynomial with coefficients in Fq , we can write:
p p
[f (x)]q = f (xq ) ⇒ [f (x)]q = f (xq )
2
Then if β is a root of polynomial f (x), β q , β q , · · · are also roots of this polyno-
mial.
p p
Since in field Fq αq−1 = 1, then β q = (αj )q = αj = β and, thus, minimal
polynomial mβ (x) is simply equal to:
mβ (x) = x + β
These results on minimal polynomials are used to determine the generator poly-
nomials of particular cyclic codes (BCH and Reed-Solomon).
Primitive polynomials
A polynomial with coefficients in F2 is primitive if it is the minimal polynomial
associated with a primitive element of a Galois field. A primitive polynomial
is thus irreducible in F2 and consequently can be used to build a Galois field.
When a primitive polynomial is used to build a Galois field, all the elements of
the field are obtained by raising the primitive element, the root of the primitive
polynomial, to successively increasing powers. As the main primitive polynomi-
als are listed in the literature, the construction of a Galois field with q = 2m
elements can then be done simply by using a primitive polynomial of degree m.
Table 4.9 gives some primitive polynomials.
To end this introduction to Galois fields and minimal polynomials, let us
give an example of a Galois field with q = 16 (m = 4) elements built from the
primitive polynomial x4 +x+1. This field is used to build generator polynomials
of BCH and Reed-Solomon codes and to decode them. The elements of this field
are:
F16 = 0, 1, α, α2 , α3 · · · α14
where α is a primitive element of F16 . With these 16 elements, we can also asso-
ciate a polynomial representation and a binary representation. The polynomial
representation of an element of this field is of the form:
aα3 + bα2 + cα + d
where a, b, c and d are binary coefficients belonging to F2 .

4. Block codes 165
Degree of the polynomial Primitive polynomial

2 α2 + α + 1
3 α3 + α + 1
4 α4 + α + 1
5 α5 + α2 + 1
6 α6 + α + 1
7 α7 + α3 + 1
8 α + α4 + α3 + α2 + 1
8
9 α9 + α4 + 1
10 α10 + α3 + 1
Table 4.9 – Examples of primitive polynomials
Galois field F16 being made up of 16 elements, the binary representation of

an element of this field is done with the help of 4 binary symbols belonging to
F2 . These 4 symbols are equal to the values taken by coefficients a, b, c and d
respectively.
Elements Polynomial Binary representation

of the representa-
field tion
0 0 0 0 0 0
1 1 0 0 0 1
α α 0 0 1 0
α2 α2 0 1 0 0
3
α α3 1 0 0 0
4
α α+1 0 0 1 1
α5 α2 + α 0 1 1 0
α6 α3 + α2 1 1 0 0
7
α α3 + α + 1 1 0 1 1
8
α α2 + 1 0 1 0 1
9
α α3 + α 1 0 1 0
10 2
α α +α+1 0 1 1 1
α11 α3 + α2 + α 1 1 1 0
α12 α3 + α2 + α + 1 1 1 1 1
13
α α3 + α 2 + 1 1 1 0 1
14
α α3 + 1 1 0 0 1
Table 4.10 – Different representations of the elements of Galois field F16

Example 3
Some calculations in field F16 are given in table 4.11 for addition, in table 4.12
for multiplication and in table 4.13 for division.
+ α2 α4
α8 0100 + 0101 = 0001 = 1 0011 + 0101 = 0110 = α5
α10 0100 + 0111 = 0011 = α4 0011 + 0111 = 0100 = α2
Table 4.11 – Addition in F16
× α2 α6
α8 α10 α14
α14 α16 = α as α15 = 1 α20 = α as α15 = 1
5
Table 4.12 – Multiplication in F16
÷ α2 α12
−6
α8 α = α as α15 = 1
9
α4
α14 α−12 = α3 as α15 = 1 α−2 = α as α15 = 1
13
Table 4.13 – Division in F16

Chapter 5
Convolutional codes and their

decoding
5.1 History
It was in 1955 that Peter Elias introduced the notion of convolutional code
[5.5]. The example of an encoder described is illustrated in Figure 5.1. It
is a systematic encoder, that is, the coded message contains the message to
be transmitted, to which redundant information is added. The message is of
infinite length, which at first sight limits the field of application of this type of
code. It is however easy to adapt it for packet transmissions thanks to tail-biting
techniques.
di
di d i 1 d i 2 d i 3 d i ,r i
D D D Mux
ri
Figure 5.1 – Example of a convolutional encoder.
The encoder presented in Figure 5.1 is designed around a shift register with
three memory elements. The redundancy bit at instant i, denoted ri is con-
structed with the help of a modulo 2 sum of the information at instant i, di and
the data present at instants i − 1 and i − 3 (di−1 and di−3 ). A multiplexer plays
the role of a parallel to serial converter and provides the result of the encoding
at a rate twice that of the rate at the input. The coding rate of this encoder is
1/2 since, at each instant i, it receives data di and delivers two elements at the
output: di (systematic part) and ri (redundant part).
It was not until 1957 that the first algorithm capable of decoding such codes
appeared. Invented by Wozencraft [5.15], this algorithm, called sequential de-
coding, was then improved by Fano [5.6] in 1963. Four years later, Viterbi
introduced a new algorithm that was particularly interesting when the length of
the shift register of the encoder is not too large [5.14]. Indeed, the complexity
of the Viterbi algorithm increases exponentially with the size of this register
whereas the complexity of the Fano algorithm is almost independent of it.
In 1974, Bahl, Cocke, Jelinek and Raviv presented a new algorithm [5.1]
capable of associating a probability with the binary decision. This property is
very widely used in the decoding of concatenated codes and more particularly
turbocodes, which have brought this algorithm back into favour. It is now
referred to in the literature in one of these three ways: BCJR (initials of the
inventors), MAP (Maximum A Posteriori) or APP (A Posteriori Probability).
The MAP algorithm is rather complex to implement in its initial version,
and it exists in simplified versions, the most common ones being presented in
Chapter 7.
In parallel with these advances in decoding algorithms, a number of works
have treated the construction of convolutional encoders. The aim of these stud-
ies has not been to decrease the complexity of the encoder, since its implantation
is trivial. The challenge is to find codes with the highest possible error correction
capability. In 1970, Forney wrote a reference paper on the algebra of convolu-
tional codes [5.7]. It showed that a good convolutional code is not necessarily
systematic and suggested a construction different from that of Figure 5.1. For a
short time, that paper took systematic convolutional codes away from the field
of research on channel coding.
Figure 5.2 gives an example of a non-systematic convolutional encoder. Un-
like the encoder in Figure 5.1, the data are not present at the output of the
encoder and are replaced by a modulo 2 sum of the data at instant i, di , and
of the data present at instants i − 2 and i − 3 (di−2 and di−3 ). The rate of
the encoder remains unchanged at 1/2 since the encoder always provides two
(1) (2)
elements at the output: ri and ri , at instant i.
When Berrou et al. presented their work on turbocodes [5.4], they rehabil-
itated systematic convolutional codes by using them in a recursive form. The
interest of recursive codes is presented in Sections 5.2 and 5.3. Figure 5.3 gives
an example of an encoder for recursive systematic convolutional codes. The
original message being transmitted (di ), the code is therefore truly systematic.
A feedback loop appears, the structure of the encoder now being similar to that
of pseudo-random sequence generators.
This brief overview has allowed us to present the three most commonly used
families of convolutional codes: systematic, non-systematic, and recursive sys-
tematic codes. The next two sections tackle the representation and performance
5. Convolutional codes and their decoding 169
r (1)
i
di d i 1 r (1)
, r (2)
i i
D D D Mux
d i2 d i 3
r (2)
i
Figure 5.2 – Example of an encoder for non-systematic convolutional codes.
of convolutional codes. They give us the opportunity to compare the properties

of these three families. The decoding algorithms most commonly used in cur-
rent systems are presented in Section 5.4. Finally, Section 5.5 tackles the main
tail-biting and puncturing techniques.
di d i ,r i
ri Mux
di s (1)
i
D D D
s (0)
i
s (2)
i s (3)
i
Figure 5.3 – Example of a encoder for recursive systematic convolutional codes.
5.2 Representations of convolutional codes

This chapter makes no claim to tackle the topic of convolutional codes exhaus-
tively. Non-binary or non-linear codes are not treated, nor are encoders with
several registers. Only the most commonly used codes, in particular for the
construction of turbocodes, are introduced. The reader wishing to go further
into the topic can, for example, refer to [5.11].
5.2.1 Generic representation of a convolutional encoder

Figure 5.4 gives a sufficiently general model for us to represent all the convo-
lutional codes studied in this chapter. At each instant i, it receives a vector
di of m bits at its input. The code thus generated is a binary code. However,
d (1)
i c
d (1)
i
d (2)
i c
d (2)
i
(m)
d i c
d (m)
i
a(1)
1
a(2)
1
a(m)
1
a(1)
2
a(2)
2
a2(m) a(1) a(2) a(m)
D D D
s (1)
i s (2)
i s (i )
b1 b2
g(1)1 g(1)2 g(1)

r (1)
i
(1)
g0
g(2)1 g(2)2 g(2)

r (2)
i
(2)
g0
g(n)1 g(n)2 g(n)

r (in)
(n)
g0
Figure 5.4 – Generic representation of an encoder for convolutional codes.
to simplify the writing, we will call it an m-binary and double-binary code if

m = 2 (as in the example presented in Figure 5.5). When c = 1, the code gen-
erated is systematic, since di is transmitted at the output of the encoder. The
code is thus made up of the systematic part di on m bits and the redundancy
ri on n bits. The coding rate is then R = m/(m + n). If c = 0, the code is
non-systematic and the rate becomes R = m/n.
The non-systematic part is constructed with the help of a shift register made
up of ν flip-flops and of binary adders, in other words, of XOR gates. We then
define an important characteristic of convolutional codes: the constraint length,
here equal to ν + 1 (some authors denote it ν, which implies that the register is
then made up of ν − 1 flip-flops). The register at instant i is characterized by the
(1) (2)
ν bits si , si , . . ., sνi memorized: they define its state, that we can thus code
(1) (2)
on ν bits and represent in the form of a vector si = (si , si , · · · , sνi ). This
ν
type of convolutional encoder thus has 2 possible state values, that we often
denote in natural binary or binary decimal form. Thus the state of an encoder
made up of three flip-flops can take 23 = 8 values. If s1 = 1, s2 = 1 and s3 = 0,
the encoder is in state 110 in natural binary, that is, 6 in decimal.
(l)
Using the coefficients aj , each of the m components of the vector di is
selected or not as the term of an addition with the content of a previous flip-flop
(except in the case of the first flip-flop) to provide the value to be stored in the
following flip-flop. The new content of a flip-flop thus depends on the current
input and on the content of the previous flip-flop. The case of the first flip-flop
has to be considered differently. If all the bj coefficients are null, the input is the
result of the sum of the only components selected of di . In the opposite case,
the contents of the flip-flops selected by the non-null bj coefficients are added to
the sum of the components selected of di . The code thus generated is recursive.
Thus, the succession of states of the register depends on the departure state
and on the succession of data at the input. The components of redundancy ri
are finally produced by summing the content of the flip-flops selected by the
coefficients g.
Let us consider some examples.
— The encoder represented in Figure 5.1 is systematic binary, therefore m = 1
(l) (1)
and c = 1. Moreover, all the aj coefficients are null except a1 = 1.
This encoder is not recursive since all the coefficients bj are null. The
(1) (1) (1)
redundancy (or parity) bit is defined by g0 = 1, g1 = 1, g2 = 0 and
(1)
g3 = 1.
— In the case of the non-systematic non-recursive binary (here called "clas-
(l) (1)
sical") encoder in Figure 5.2, m = 1, c = 0 ; among the aj , only a1 = 1
is non-null and bj = 0 ∀j. Two parity bits come from the encoder and
(1) (1) (1) (1) (2) (2)
are defined by g0 = 1, g1 = 1, g2 = 0, g3 = 1 and g0 = 1, g1 = 0,
(2) (2)
g2 = 1, g3 = 1.
— Figure 5.3 presents a recursive systematic binary encoder (m = 1, c = 1

(1)
and a1 = 1). The coefficients of the recursivity loop are then b1 = 1,
(1) (1) (1)
b2 = 0, b3 = 1 and those of the redundancy are g0 = 1, g1 = 0, g2 = 1,
(1)
g3 = 1.
— Figure 5.5 represents a recursive systematic double-binary encoder. The

(l)
only coefficients that differ from the previous case are the aj : coefficients
(1) (2) (2) (2) (l)
a1 , a1 , a2 and a3 are equal to 1, the other aj are null.
To define an encoder, it is not however necessary to make a graphic repre-
sentation since knowledge of the parameters presented in Figure 5.4 is sufficient.
A condensed representation of these parameters is known as generator polyno-
mials. This notation is presented in the following paragraph.
d (1)
i
d (1), d (2),ri
d (2)
i Mux
i i
ri
D D D
Figure 5.5 – Example of a recursive systematic double-binary convolutional encoder.
5.2.2 Polynomial representation

Let us first consider a classical (non-systematic and non-recursive) binary code:
k=1..n
(k)
c = 0, all the coefficients bj are null and m = 1. The knowledge of gj
j=0..ν
(k)
then suffices to describe the code. The coefficients gj therefore define
j=0..ν
n generator polynomials G(k) in D (Delay) algebra:
(k)
G(k) (D) = gj D j (5.1)
j=0...ν
(1)
Let us take the case of the encoder defined in Figure 5.2. The outputs ri and
(2)
ri are expressed as functions of the successive data d as follows:
(1)
ri = di + di−2 + di−3 (5.2)
which can also be written, via the transform in D:
r(1) (D) = G(1) (D) d(D) (5.3)
with G(1) (D) = 1 + D2 + D 3 , the first generator polynomial of the code and
d(D) the transform in D of the message to be encoded. Likewise, the second
generator polynomial is G2 (D) = 1 + D + D 3 .
These generator polynomials can also be resumed by the series of their coeffi-
cients, (1011) and (1101) respectively, generally denoted in octal representation,
(13)octal and (15)octal respectively. In the case of a non-recursive systematic
code , like the example in Figure 5.1, the generator polynomials are expressed
according to the same principle. In this example, the encoder has generator
polynomials G(1) (D) = 1 and G(2) (D) = 1 + D + D3 .
To define the generator polynomials of a recursive systematic code is not
straightforward. Let us consider the example of Figure 5.3. The first generator
polynomial is trivial since the code is systematic. To identify the second, we

must note that
(0) (1) (3) (0) (0)
si = di + si + si = di + si−1 + si−3 (5.4)
and that:
(0) (2) (3) (0) (0) (0)
ri = si + si + si = si + si−2 + si−3 (5.5)
which is equivalent to:
(0) (0) (0)
di = si + si−1 + si−3
(0) (0) (0) (5.6)
ri = si + si−2 + si−3
This result can be reformulated by introducing the transform in D:
d(D) = G(2) (D)s(D)

(5.7)
r(D) = G(1) (D)s(D)
where G(1) (D) and G(2) (D) are the generator polynomials of the code shown in
Figure 5.2, which leads to
d(D)
s(D) = G(2) (D)
(5.8)
G(1) (D)
r(D) = G(2) (D)
d(D)
Thus, a recursive systematic code can easily be derived from a non-systematic

non-recursive code. The codes generated by such encoders can be represented
graphically according to three models: the tree, the trellis and the state machine.
5.2.3 Tree of a code

The first graphic representation of a code and certainly the least pertinent for
the rest of the chapter is the tree representation. It enables all the sequences
of possible states to be presented. The root is associated with the initial state
of the encoder. From the root, we derive all the possible successive states as
a function of input di of the encoder. The branch linking a father state to a
son state is labelled with the value of the outputs of the encoder during the
associated transition. This principle is iterated for each of the strata and so
forth. The tree diagram associated with the systematic encoder of Figure 5.1 is
illustrated in Figure 5.6. This type of diagram will not be used in what follows
and the only use that is made of it concerns a sequential decoding algorithm
(Fano’s algorithm) not treated in this book.
5.2.4 Trellis of a code

The most common representation of a convolutional code is the trellis diagram.
It is of major importance both for defining the properties of a code and for
decoding it, as we shall see in the next part of Chapter 5.
Figure 5.6 – Tree diagram of the code of polynomials [1, 1 + D + D3 ]. The binary pairs
indicate the outputs of the encoder and the values in brackets are the future states.
00 00 00 00 00
(000)
11 11 11
01 01
(001)
10 10
11 11
00 00 00
(010)
11 11 11
01 01
(011) 01
10 10
01
01 01
(100)
10 10 10
10 00 00
(101) 11 11
01
01 01
(110) 10 10 10
00 00
(111) 11 11
di ri di=0
di ri di=1
Figure 5.7 – Trellis diagram of a code with generator polynomials [1, 1 + D + D3 ].
At an instant i, the state of a convolutional encoder can take 2ν values.

Each of these possible values is represented by a node. With each instant i is
associated a states-nodes column and according to input di , the encoder transits
from a state si to si+1 while delivering the coded bits. This transition between
two states is represented by an arc between the two associated nodes and labelled
with the outputs of the encoder. In the case of a binary code, the transition on an
input at 0 (resp. 1) is represented by a dotted (resp. solid) line. The succession
of si states up to instant t is represented by the different paths between the
initial state and the different possible states at instant t.
Let us show this with the example of the systematic encoder of Figure 5.1.
Hypothesizing that the initial state s0 is state (000) :
• If d1 = 0 then the following state, s1 , is also (000). The transition is
represented by a dotted line and labelled in this first case 00, the value of
the encoder outputs;
• If d1 = 1, then the following state, s1 , is (100). The transition is repre-
sented by a solid line and here labelled 11.
• We must next envisage the four possible transitions: s1 = (000) if d2 = 0
or d2 = 1 and s1 = (100) if d2 = 0 or d2 = 1.
Iterating this construction, we reach the representation, in Figure 5.7, of all

the possible successions of states from the initial state to instant 5, without the
unlimited increase of the tree diagram.
A complete section of the trellis suffices to characterize the code. The trel-
lis section of the previous code is thus shown in Figure 5.8(a). Likewise, the
encoders presented in Figures 5.2 and 5.3 are associated with trellis sections,
illustrated in Figures 5.8(b) and 5.8(c) respectively.
(000) 00 (000) 00 (000) 00

11 11 11
01 11 11
(001) (001) (001)
10 00 00
00 10 01
(010) (010) (010)
11 01 10
01 01 10
(011) (011) (011)
10 10 01
01 01 10
(100) (100) (100)
10 10 01
00 10 01
(101) 11 (101) 01 (101) 10
01 11 11
(110) 10 (110) 00 (110) 00
00 00 00
(111) 11 (111) 11 (111) 11
di ri d i=0 ri(1) ri(2) di=0 di ri di=0

di ri d i=1 ri(1) ri(2) d i=1 di ri di=1
Figure 5.8 – Trellis sections of the codes with generator polynomials [1, 1 + D + D3 ]
(a), [1 + D2 + D3 , 1 + D + D3 ] (b) and [1, (1 + D2 + D3 )/(1 + D + D3 )] (c)
Such a representation shows the basic pattern of these trellises: the butterfly,
so called because of its shape. Each of the sections of Figures 5.8 is thus made up
of 4 butterflies (the transitions of states 0 and 1 towards states 0 and 4 make up
one). The butterfly structure of the three trellises illustrated is identical but the
sequences coded differ. It should be noted in particular that all the transitions
arriving at the same node of a trellis of a non-recursive code are due to the same
value at the input of the encoder. Thus, among the two non-recursive examples
treated (Figures 5.8(a) and 5.8(b)), a transition associated with a 0 at the input
necessarily arrives at one of the states between 0 and 3 and a transition with a
1 arrives at one of the states between 4 and 7. It is different in the case of a
recursive code (like the one presented in Figure 5.8(c)): each state allows one
incident transition associated with an input with a 0, and another one associated
with 1. We shall see the consequences of this in Section 5.3.
5.2.5 State machine of a code

To represent the different transitions between the states of an encoder, there is
a final representation, that of a state machine. The convention for defining the
transition branches is identical to that used in the previous section. Only 2ν
nodes are represented, independently of instant i, the previous representation
being translated into a trellis section. The encoders of Figures 5.1, 5.2 and
5.3 thus allow a representation in the form of a state machine illustrated in
Figures 5.9, 5.10 and 5.11 respectively.
10
11 (100) 01 11
(110) 10
11
00 (000) 10 (010) 00 (101) 01 (111) 11
01 00 10 00
(001) 01 (011)
di ri di =0
di ri di =1
Figure 5.9 – State machine for a code with generator polynomials [1, 1 + D + D3 ].
This representation is particularly useful for determining the transfer func-

tion and the distance spectrum of a convolutional code (see Section 5.3). It is
now possible to see a notable difference between a state machine of a recursive
code and that of a non-recursive code: the existence and the number of cycles1
on an all-zero sequence at the input.
In the case of two non-recursive state machines, there is a single cycle on a null
sequence at the input: the loop on state 0.
1 A cycle is a succession of states such that the initial state is also the final state.
10
11 (100) (110) 00
01 01
01
00 (000) 00 (010) 10 (101) 11 (111) 11
11 10 10 00
(001) 01 (011)
ri(1) ri(2) di=0
ri(1) ri(2) di=1
Figure 5.10 – State machine for a code with generator polynomials [1 + D2 + D3 , 1 +

D + D3 ].
Figure 5.11 – State machine for a code with generator polynomials [1, (1+D2 +D3 )/(1+
D + D3 )].
However, the recursive state machine allows another cycle on a null sequence at
the input: state 4 → state 6 → state 7 → state 3 → state 5 → state 2 → state
1 → state 4.
Moreover, this cycle is linked to the loop on state 0 by two transitions associated
with inputs at 1 (transitions 0 → 4 and 1 → 0). There therefore exists an infinite
number of input sequences with Hamming weight 2 equal to 2 producing a cycle
on state 0. This weight 2 is the minimum weight of any sequence that makes
the recursive encoder leave state 0 and return to zero. Because of the linearity
of the code (see Chapter 1), this value of 2 is also the smallest distance that can
separate two sequences with different inputs that make the encoder leave the
same state and return to the same state.
In the case of non-recursive codes, the Hamming weight of the input se-
quences allowing a cycle on state 0 can only be 1 (state 0 → state 4 → state 2
→ state 1 → state 0). This distinction is essential for understanding the interest
of recursive codes used alone (see Section 5.3) or in a turbocode structure (see
Chapter 7).
2 The Hamming weight of a binary sequence is equal to the number of bits equal to 1.
5.3 Code distances and performance

5.3.1 Choosing a good code
As a code is exploited for its correcting capacities, we have to be able to esti-
mate these in order to make a judicious choice of one code rather than another,
according to the application targeted. Among the bad possible choices, catas-
trophic codes are such that a finite number of errors at the input of the decoder
can produce an infinite number of errors at the output of the decoder, which
explains their name. One main property of these codes is that there exists at
least one input sequence of infinite weight that generates a coded sequence of
finite weight: systematic codes therefore cannot be catastrophic. These codes
can be identified very simply if they have a rate of the form R = 1/N . We can
then show that the code is catastrophic if the largest common divisor (L.C.D.) of
its generator polynomials is different from unity. Thus, the code with generator
polynomials G(1) (D) = 1 + D + D2 + D 3 and G(2) (D) = 1 + D3 is catastrophic
since the L.C.D. is 1 + D.
However, choosing a convolutional code cannot be limited to the question "Is
it catastrophic?". By exploiting the graphic representations introduced above,
the properties and performance of codes can be compared.
5.3.2 RTZ sequences

Since convolutional codes are linear, to determine the distances between the
different coded sequences amounts to determining the distances between the
non-null coded sequences and the "all zero" sequence. Therefore, it suffices to
calculate the Hamming weight of all the coded sequences that leave from state
0 and that return to it. These sequences are called Return To Zero (RTZ) se-
quences. The smallest Hamming weight thus obtained is called the free distance
of the code. The minimum Hamming distance of a convolutional code is equal
to its free distance from a certain length of coded sequence. In addition, the
number of RTZ sequences that have the same weight is called the multiplicity
of this weight.
Let us consider the codes that have been used as examples so far. Each RTZ
sequence of minimum weight is shown in bold in Figures 5.12, 5.13 and 5.14.
The non-recursive systematic code has an RTZ sequence with minimum Ham-
ming weight equal to 4. The free distance of this code is therefore equal to 4.
On the other hand, as the classical code and the recursive systematic code each
possess two RTZ sequences with minimum weight 6, their free distance is there-
fore 6. The correction capacity of non-recursive non-systematic and recursive
systematic codes is therefore better than that of the non-recursive systematic
code.
It is interesting, in addition, to compare the weights of the sequences at the
input associated with the RTZ sequences with minimum weight. In the case of
00 00 00 00 00
(000)
11 11 11
01 01
(001)
10 10
11 11
00 00 00
(010)
11 11 11
01 01
(011) 01
10 10
01
01 01
(100)
10 10 10
10 00 00
(101) 11 11
01
01 01
(110) 10 10 10
00 00
(111) 11 11
ri(1) ri(2) di =0
ri(1) ri(2) di =1
Figure 5.12 – RTZ sequence (in bold) defining the free distance of the code with
generator polynomials [1, 1 + D + D3 ].
00 00 00 00 00 00
(000)
11 11 11 11
11 11 11
(001)
00 00 00
11 11
10 10 10 10
(010)
01 01 01 01
01 01 01
(011) 01
10 10 10
01
01 01 01
(100)
10 10 10 10
10 10 10 10
(101) 01 01 01
11
11 11 11
(110) 00 00 00 00
00 00 00
(111) 11 11 11
ri(1) ri(2) di =0
ri(1) ri(2) di =1
Figure 5.13 – RTZ sequences (in bold) defining the free distance of the code with
generator polynomials [1 + D2 + D3 , 1 + D + D3 ].
00 00 00 00 00 00
(000)
11 11 11 11
11 11 11
(001)
00 00 00
11 11
01 01 01 01
(010)
10 10 10 10
10 10 10
(011) 10
01 01 01
10
10 10 10
(100)
01 01 01 01
01 01 01 01
(101) 10 10 10
11
11 11 11
(110) 00 00 00 00
00 00 00
(111) 11 11 11
di ri(2) di =0
di ri(2) di =1
Figure 5.14 – RTZ sequences (in bold) defining the free distance of the code with
generator polynomials [1, (1 + D2 + D3 )/(1 + D + D3 )].
the non-recursive systematic code, the only sequence of this type has a weight
equal to 1, which means that if the RTZ sequence is decided instead of the
transmitted "all zero" sequence, only one bit is erroneous. In the case of the
classical code, one sequence at the input has a weight of 1 and another a weight
of 3: one or three bits are therefore wrong if such an RTZ sequence is decoded.
In the case of the recursive systematic code, the RTZ sequences with minimum
weight have an input weight of 3.
Knowledge of the minimum Hamming distance and of the input weight as-
sociated with it is not sufficient to closely evaluate the error probability at the
output of the decoder of a simple convolutional code. It is necessary to com-
pute the distances, beyond the minimum Hamming distance, and their weight in
order to make this evaluation. This computation is called the distance spectrum.
5.3.3 Transfer function and distance spectrum

The error correction capability of a code depends on all the RTZ sequences,
which we will consider in the increasing order of their weight. Rather than
computing them by reading the graphs, it is possible to establish the transfer
function of the code. The latter is obtained from the state transition diagram
in which the initial state (000) is cut into two states ae and as , which are no
other than the initial state and the arrival state of any RTZ sequence.
Let us illustrate the computation of the transfer function with the example
of the systematic code of Figure 5.1, whose state transition diagram is again
represented in Figure 5.15.
(000)
ae
O²I
OI
(100) O O²I
(110) OI
e g
O²I
OI (010) 1 (101) O (111) O²I
c f h
1 OI 1
(001) O (011)
b d
O
(000)
as
Figure 5.15 – Machine state of the code [1, 1 + D + D3 ], modified for the computation
of the associated transfer function.
Each transition has a label Oi I j , where i is the weight of the sequence coded
and j that of the sequence at the input of the encoder. In our example, j
can take the value 0 or 1 according to the level of the bit at the input of the
encoder at each transition and i varies between 0 and 2, since 4 coded symbols
are possible (00, 01, 10, 11), with weights between 0 and 2.
The transfer function of the code T (O, I) is then defined by:
as
T (O, I) = (5.9)
ae
To establish this function, we have to solve the system of equations coming from
the relations between the 9 states (ae, b, c... h and as):
b = c + Od
c = Oe + f
d = h + Dg
e = O2 Iae + OIb
(5.10)
f = O2 Ic + OId
g = OIe + O2 If
h = O2 Ih + OIg
as = Ob
Using a formal computation tool, it is easy to arrive at the following result:
−I 4 O12 + (3I 4 + I 3 )O10 + (−3I 4 − 3I 3 )O8 + (I 4 + 2I 3 )O6 + IO4

T (O, I) =
I 4 O10+ (−3I 4 − I 3 )O8 + (3I 4 + 4I 3 )O6 + (−I 4 − 3I 3 )O4 − 3IO2 + 1
T (O, I) can then be developed as a series:
T (O, I) = IO4
+(I 4 + 2I 3 + 3I 2 )O6
+(4I 5 + 6I 4 + 6I 3 )O8
(5.11)
+(I 8 + 5I 7 + 21I 6 + 24I 5 + 17I 4 + I 3 )O10
+(7I 9 + 30I 8 + 77I 7 + 73I 6 + 42I 5 + 3I 4 )O12
+···
This enables us to observe that an RTZ sequence with weight 4 is produced

by an input sequence with weight 1, that the RTZ sequences with weight 6 are
produced by a sequence with weight 4, by two sequences with weight 3 and by
three sequences with weight 2, etc.
In the case of the classical code mentioned above, the transfer function is:
T (O, I) = (I 3 + I)O6
+(2I 6 + 5I 4 + 3I 2 )O8
+(4I 9 + 16I 7 + 21I 5 + 8I 3 )O10
(5.12)
+(8I 12 + 44I 10 + 90I 8 + 77I 6 + 22I 4 )O12
+(16I 15 + 112I 13 + 312I 11 + 420I 9 + 265I 7 + 60I 5 )O14
+···
Likewise, the recursive systematic code already studied has as its transfer func-
tion:
T (O, I) = 2I 3 O6
+(I 6 + 8I 4 + I 2 )O8
+(8I 7 + 33I 5 + 8I 3 )O10
(5.13)
+(I 10 + 47I 8 + 145I 6 + 47I 4 + I 2 )O12
+(14I 11 + 254I 9 + 649I 7 + 254I 5 + 14I 3 )O14
+···
Comparing the transfer functions from the point of view of the monomial with
the smallest degree allows us to appreciate the error correction capability at
very high signal to noise ratio (asymptotic behaviour). Thus, the non-recursive
systematic code is weaker than its rivals since it has a lower minimum distance.
A classical code and its equivalent recursive systematic code have the same free
distance, but their monomials of minimal degree differ. The first is in (I 3 + I)O6
and the second in 2I 3 O6 . This means that with the classical code an input
sequence with weight 3 and another with weight 1 produce an RTZ sequence
with weight 6 whereas with the recursive systematic code two sequences with
weight 3 produce an RTZ sequence with weight 6. Thus, if an RTZ sequence
with minimum weight is introduced by the noise, the classical code will introduce
one or three errors, whereas its recursive systematic code will introduce three
or three other errors. In conclusion, the probability of a binary error on such
a sequence is lower with a classical code than with a recursive systematic code,
which explains that the former will be slightly better at high signal to noise ratio.
Things are generally different when the codes are punctured (see Section 5.5) in
order to have higher rates [5.13].
To compare the performance of codes with low signal to noise ratio, we
must consider all the monomials. Let us take the example of the monomial in
O12 for the non-recursive systematic code, the classical code and the recursive
systematic code, respectively:
(7I 9 + 30I 8 + 77I 7 + 73I 6 + 42I 5 + 3I 4 )O12

(8I 12 + 44I 10 + 90I 8 + 77I 6 + 22I 4 )O12
(I 10 + 47I 8 + 145I 6 + 47I 4 + I 2 )O12
If 12 errors are introduced by the noise on the channel, 232 RTZ sequences
are "available" as errors for the first code, 241 for the second and 241 again
for the third. It is therefore (a little) less probable that an RTZ sequence will
appear if the code used is the non-recursive systematic code. Moreover, the
error expectancy per RTZ sequence of the three codes is 6.47, 7.49 and 6.00,
respectively: the recursive systematic code therefore introduces, on average,
fewer decoding errors than the classical code on RTZ sequences with 12 errors
on the frame coded. This is also true for higher degree monomials. Recursive
and non-recursive systematic codes are therefore more efficient at low signal to
d 6 8 10 12 14 ...
ω(d) 6 40 245 1446 8295 ...
Table 5.1 – First terms of the spectrum of the recursive systematic code with generator
polynomials [1, (1 + D2 + D3 )/(1 + D + D3 )].
noise ratio than the classical code. Moreover, we find the monomials I 2 O8+4c ,
where c is an integer, in the transfer function of the recursive code. The infinite
number of monomials of this type is due to the existence of the cycle on a
null input sequence different from the loop on state 0. Moreover, such a code
does not provide any monomials of the form IOc , unlike non-recursive codes.
These conclusions concur with those drawn from the study of state machines in
Section 5.2.
This notion of transfer function is therefore efficient for studying the per-
formance of a convolutional code. A derived version is moreover essential for
the classification of codes according to their performance. This is the distance
spectrum ω(d) whose definition is as follows:
∞

∂T (O, I)
( )I=1 = ω(d)Od (5.14)
∂I
d=df
For example, the first terms of the spectrum of the recursive systematic code,
obtained from (5.13), are presented in Table 5.1. This spectrum is essential for
estimating the performance of codes in terms of calculating their error proba-
bility, as illustrated in the vast literature on this subject [5.9].
The codes used in the above examples have a rate of 1/2. By increasing the
number of redundancy bits n the rate becomes lower. In this case, the powers
of O associated with the branches of the state machines will be higher than or
equal to those of the figures above. This leads to higher transfer functions with
powers of O, that is, to RTZ sequences with a greater Hamming weight. The
codes with lower rates therefore have a higher error correction capability.
5.3.4 Performance
The performance of a code is defined by the decoding error probability after
transmission on a noisy channel. The previous section allows us to intuitively
compare non-recursive non-systematic, non-recursive systematic and recursive
systematic codes with the same constraint length. However, to estimate the
absolute performance of a code, we must be able to estimate the decoding error
probability as a function of the noise, or at least to limit it. The literature,
for example [5.9], thus defines many bounds that are not described here and we
will limit ourselves to comparing the three categories of convolutional codes. To
do this, a transmission on a Gaussian channel of blocks of 53 then 200 bytes
Figure 5.16 – Comparison of simulated performance (Binary Error Rate and Packet
Error Rate) of three categories of convolutional codes after transmission of packets of
53 bytes on a Gaussian channel (decoding using the MAP algorithm).
Figure 5.17 – Comparison of simulated performance of three categories of convolutional

codes after transmission of blocks of 200 bytes on a Gaussian channel.
coded according to different schemes was simulated (Figures 5.16 and 5.17):
classical (non-recursive non-systematic), non-recursive systematic and recursive
systematic.
The blocks were constructed following the classical trellis termination tech-
nique for non-recursive codes whereas the recursive code is circular tail-biting
(see Section 5.5). The decoding algorithm used is the MAP algorithm.
The BER curves are in perfect agreement with the conclusions drawn during
the analysis of the free distance of codes and of their transfer function: the
systematic code is not as good as the others at high signal to noise ratio and
the classical code is then slightly better than the recursive code. At low signal
to noise ratios, the hierarchy is different: the recursive code and the systematic
code are equivalent and better than the classical code.
Comparing performance as a function of the size of the frame (53 and 200
bytes) shows that the performance hierarchy of the codes is not modified. More-
over, the bit error rates are almost identical. This was predictable as the sizes
of the frames are large enough for the transfer functions of the codes not to be
affected by edge effects. However, the packet error rate is affected by the length
of the blocks since although the bit error probability is constant, the packet
error probability increases with size.
The comparisons above only concern codes with 8 states. It is, however, easy
to see that the performance of a convolutional code is linked with its capacity
to provide information on the succession of data transmitted: the more the
code can integrate successive data into its output symbols, the more it improves
the quality of protection these data. In other words, the greater the number
of states (therefore the size of the register of the encoder), the more efficient
a convolutional code is (within its category). Let us compare three recursive
systematic codes:
• 4 states [1, (1 + D2 )/(1 + D + D 2 )],
• 8 states [1, (1 + D 2 + D3 )/(1 + D + D 3 )]

• and 16 states [1, (1 + D + D2 + D4 )/(1 + D3 + D4 )].
Their performance in terms of BER and PER were simulated on a Gaussian

channel and are presented in Figure 5.18.
The higher the number of states of the code, the lower the residual error rates
are. For a BER of 10−4 , 0.6 dB are thus gained when passing from 4 states to
8 states and 0.5 dB when passing from 8 states to 16 states. This remark is
coherent with the qualitative justification of the interest of a large number of
states. It would therefore seem logical to choose a convolutional code with a large
number of states to ensure the desired protection, especially since such codes
offer the possibility of producing redundancy on far more than two components,
and therefore of providing even higher protection. Thus, the Big Viterbi Decoder
project at NASA’s Jet Propulsion Laboratory used for transmissions with space
Figure 5.18 – Comparison of simulated performance of recursive systematic convolu-

tional codes with 4, 8 and 16 states after transmitting packets of 53 bytes on a Gaussian
channel (decoding according to the MAP algorithm)
probes was designed to process frames encoded with convolutional codes with
2 to 16384 states and rates much lower than 1/2 (16384 states and R = 1/6
for the Cassini probe to Saturn and the Mars Pathfinder probe). Why not
use such codes for terrestrial radio-mobile transmissions for the general public?
Because the complexity of the decoding would become unacceptable for current
terrestrial transmissions using a reasonably-sized terminal operating in real time,
and fitting into a pocket.
5.4 Decoding convolutional codes

There exist several algorithms for decoding convolutional codes. The most fa-
mous is probably the Viterbi algorithm which relies on the trellis representation
of codes [5.14][5.8]. It enables us to find the most probable sequence of states
in the trellises from the received symbol sequence.
The original Viterbi algorithm performs hard output decoding, that is, it
provides a binary estimation of each of the symbols transmitted. It is therefore
not directly adapted to iterative systems that require information about the
trustworthiness of the decisions. Adaptations of the Viterbi algorithm such
as those proposed in [5.2], [5.10] or [5.3] have led to versions with weighted
output called Soft-Output Viterbi Algorithm (SOVA). SOVA algorithms are not
described in this book since, for turbo decoding, we prefer another family of
algorithms relying on the minimization of the error probability of each symbol
transmitted. Thus, the Maximum A Posteriori (MAP) algorithm enables the
calculation of the exact value of the a posteriori probability associated with each
symbol transmitted using the received sequence [5.1]. The MAP algorithm and
its variants are described in Chapter 7.
Figure 5.19 – Model of the transmission chain studied.
5.4.1 Model of the transmission chain and notations

Viterbi and MAP algorithms are used in the transmission chain shown in Fig-
ure 5.19. The convolutional code considered is systematic and, for each data
sequence of information d = dN 1 = {d1 , · · · , dN }, calculates a redundant se-
quence r = rN1 = {r1 , · · · , rN }. The code is m-binary with rate R = m/(m + n):
each vector of data di at the input of the encoder is thus made up of m bits,
(1) (2) (m)
di = (di , di , · · · , di ), and the corresponding redundancy vector at the out-
(1) (2) (n)
put is written ri = (ri , ri , · · · , ri ). The value of di can also be represented

m
(l)
by the scalar integer variable j = 2l−1 di , between 0 and 2m − 1, and we can
l=1
then write di ≡ j.
The systematic d and redundant r data sequences are transmitted after a
binary/antipodal conversion making an antipodal value (-1 or +1) transmitted
towards the channel correspond to each value binary (0 or 1) coming from the
encoder. X and Y represent the noisy systematic and redundant symbol se-
quences received at the input of the decoder and d̂ the decoded sequence that
can denote either a binary sequence in the case of the Viterbi algorithm, or a
sequence of weighted decisions associated with the di at the output of the MAP
algorithm.
5.4.2 The Viterbi algorithm

The Viterbi algorithm is the most widely used method for the maximum-
likelihood (ML) decoding of convolutional codes with low constraint length (typ-
ically ν ≤ 8). Beyond this limit, its complexity of implementation means that
we have to resort to a sequential decoding algorithm, like Fano’s [5.6].
ML decoding is based on a search for the codeword c that is the shortest
distance away from the received word. In the case of a channel with binary deci-
sions (binary symmetric channel), ML decoding relies on the Hamming distance,

whereas in the case of a Gaussian channel, it relies on the Euclidean distance
(see Chapter 2). An exhaustive search for codewords associated with the differ-
ent paths in the trellis leads to 2ν+k paths being taken into account. In practice,
searching for the path with the minimum distance on a working window with a
width l lower than k, limits the search to 2ν+l paths.
The Viterbi algorithm enables a notable reduction in the complexity of the
computation. It is based on the idea that, among the set of paths of the trellis
that converge in a node at a given instant, only the most probable path can be
(1) (2) (ν)
retained for the following search steps. Let us denote si = (si , si , . . ., si )
the state of the encoder at instant i and T (i, si−1 , si ) the branch of the trellis
corresponding to the emission of data di and associated with the transition
between nodes si−1 and si . Applying the Viterbi algorithm involves performing
the set of operations described below.
• At each instant i, for i ranging from 1 to k:
• Calculate for each branch of a branch metric, d (T (i, si−1 , si )). For a binary
output channel, this metric is defined as the Hamming distance between
the symbol carried by the branch of the trellis and the received symbol,
d (T (i, si−1 , si )) = dH (T (i, si−1 , si )).
For a Gaussian channel, the metric is equal to the square of the Euclidean
distance between the branch considered and the observation at the input
of the decoder (see also Section 1.3):
2 2
d (T (i, si−1 , si )) = Xi − xi + Yi − yi
m
2 n
2
(j) (j) (j) (j)
= xi − Xi + yi − Yi
j=1 j=1
• Calculate the accumulated metric associated with each branch T (i, si−1 , si )
defined by:
λ(T (i, si−1 , si )) = μ(i − 1, si−1 ) + d(T (i, si−1 , si ))
where μ(i − 1, si−1 ) is the accumulated metric associated with node si−1 .
• For each node si , select the branch of the trellis corresponding to the
minimum accumulated metric and memorize this branch in memory (in
practice, it is the value of di associated with the branch that is stored).
The path in the trellis made up of the branches successively memorized at
the instants between 0 and i is the survivor path arriving in si . If the two
paths that converge in si have identical accumulated metrics, the survivor
is then chosen arbitrarily between these two paths.
• Calculate the accumulated metric associated with each node si , μ(i, si ).

It is equal to the accumulated metric associated with the survivor path
arriving in si :
μ(i, si ) = min (λ (T (i, si−1 , si ))) .
si−1
• Initialization (instant i =0):

The initialization values of the metrics μ when commencing the algorithm
depend on the initial state s of the encoder: μ(0, s) = +∞ if s = s0 and
μ(0, s0 ) = 0. If this state is not known, all the metrics are initialized to the
same value, typically 0. In this case, the decoding of the beginning of the frame
is less efficient since the accumulated metrics associated with each branch at
instant 1 depend only on the branch itself. The past cannot be taken into
account since it is not known: λ(T (1, s0 , s1 )) = d(T (1, s0 , s1 )).
• Calculating the decisions (instant i=k):

At instant k, if the final state of the encoder is known, the maximum likeli-
hood path is the survivor path coming from the node corresponding to the final
state of the encoder. The decoded sequence is given by the series of values of
di , i ranging from 1 to k, stored in the memory associated with the maximum
likelihood path. This operation is called trellis traceback.
If the final state is not known, the maximum likelihood path is the survivor
path coming from the node with minimum accumulated metric. In this case,
the problem is similar to the one mentioned for initialization: the decoding of
the end of the frame is less efficient.
When the sequence transmitted is long, or even infinite in length, it is not
possible to wait for the whole transmitted binary sequence to be received to
begin the decoding operation. To limit the decoding latency and the size of
memory necessary to memorize the survivor paths, the trellis must be truncated.
Observing the algorithm unfold, we can note that by tracing back sufficiently in
time from instant i, the survivor paths coming from the different nodes of the
trellis nearly always converge towards a same path. In practice, memorizing the
survivors can therefore be limited to a time interval of duration l. It is then
sufficient to do a trellis traceback at each instant i over a length l in order to
take the decision on the data di−l . To decrease the complexity, the survivors
are sometimes memorized on an interval higher than l (for example l+3). The
number of trellis traceback operations is then decreased (divided by 3 in our
example) but each of the tracebacks provides several decisions (di−l , di−l−1 and
di−l−2 , in our example).
The higher the code memory and coding rate are, the greater the value of l
must be. We observe that, for a systematic code, values of l corresponding to
the production by the encoder of a number of redundancy symbols equal to 5
times the constraint length of the code are sufficient. As an example, for coding
rate R = 1/2, we typically take l equal to 5(ν + 1).
From the point of view of complexity, the Viterbi algorithm requires the
calculation of 2ν+1 accumulated metrics at each instant i and its complexity
varies linearly with the length of sequence k or of decoding window l.
Example of applying the Viterbi algorithm

Let us illustrate the different steps of the Viterbi algorithm described above by
applying it to decode the binary recursive systematic convolutional code (7,5)
with 4 states and encoding rate R = 1/2 (m = n = 1). The structure of the
encoder and the trellis are shown in Figure 5.20.
s i 1 si
(1, 1)
(00)
(+1, +1)
(+1, +1)
(01) (1, 1)
(+1, 1)
(10)
(1, +1) (1, +1)
(+1, 1)
(11)
(x i, y i) di=0
(x i, y i) di=1
Figure 5.20 – Structure of the recursive systematic convolutional code (7,5) and asso-
ciated trellis.
Calculate the branch metrics at instant i:

2 2
d(T (i, 0, 0)) = d(T (i, 1, 2)) = (Xi + 1) + (Yi + 1)
2 2
d(T (i, 0, 2)) = d(T (i, 1, 0)) = (Xi − 1) + (Yi − 1)
2 2
d(T (i, 2, 1)) = d(T (i, 3, 3)) = (Xi − 1) + (Yi + 1)
2 2
d(T (i, 2, 3)) = d(T (i, 3, 1)) = (Xi + 1) + (Yi − 1)
Calculate the accumulated metrics of the branches at instant i:

λ(T (i, 0, 0)) = μ(i − 1, 0) + d(T (i, 0, 0)) = μ(i − 1, 0) + (Xi + 1)2 + (Xi + 1)2
λ(T (i, 1, 2)) = μ(i − 1, 1) + d(T (i, 1, 2)) = μ(i − 1, 1) + (Xi + 1)2 + (Xi + 1)2
λ(T (i, 0, 2)) = μ(i − 1, 0) + d(T (i, 0, 2)) = μ(i − 1, 0) + (Xi − 1)2 + (Xi − 1)2
λ(T (i, 1, 0)) = μ(i − 1, 1) + d(T (i, 1, 0)) = μ(i − 1, 1) + (Xi − 1)2 + (Xi − 1)2
λ(T (i, 2, 1)) = μ(i − 1, 2) + d(T (i, 2, 1)) = μ(i − 1, 2) + (Xi − 1)2 + (Xi + 1)2
λ(T (i, 3, 3)) = μ(i − 1, 3) + d(T (i, 3, 3)) = μ(i − 1, 3) + (Xi − 1)2 + (Xi + 1)2
λ(T (i, 2, 3)) = μ(i − 1, 2) + d(T (i, 2, 3)) = μ(i − 1, 2) + (Xi + 1)2 + (Xi − 1)2
λ(T (i, 3, 1)) = μ(i − 1, 3) + d(T (i, 3, 1)) = μ(i − 1, 3) + (Xi + 1)2 + (Xi − 1)2
Calculate the accumulated metrics of the nodes at instant i:

μ(i, 0) = min (λ(T (i, 0, 0), λ(T (i, 1, 0))
μ(i, 1) = min (λ(T (i, 3, 1), λ(T (i, 2, 1))
μ(i, 2) = min (λ(T (i, 0, 2), λ(T (i, 1, 2))
μ(i, 3) = min (λ(T (i, 3, 3), λ(T (i, 2, 3))
For each of the four nodes of the trellis, the value of di corresponding to the
transition of minimum accumulated metric λ is stored in memory.
(0)
(1)
(2)
(i, 2 ) = min ( (i, s ) )
s {0, 1, 2, 3}
(3)
i 15 i 14 i 13 i 2 i 1 i
^d
i 15
Figure 5.21 – Survivor path traceback operation (in bold) in the trellis from instant i
and determining the binary decision at instant i − 15.
After selecting the node with minimum accumulated metric, denoted s (in
the example of Figure 5.21, s = 3), we trace back in the trellis along the survivor
path to a depth l = 15. At instant i − 15, the binary decision dî−15 is equal to
the value of di−15 stored in the memory associated with the survivor path.
The aim of applying the Viterbi algorithm with weighted inputs is to search
for codeword c that is the shortest Euclidean distance between two codewords.
Equivalently (see Chapter 1), this also means looking
m for the codeword that

k (l) (l) n
(l) (l)
maximizes the scalar product x, X + y, Y = xi Xi + yi Yi .
i=1 l=1 l=1
In this case, applying the Viterbi algorithm uses branch metrics of the form

m
(l) (l)
n
(l) (l)
d (T (i, si−1 , si )) = xi X i + yi Yi and the survivor path then corre-
l=1 l=1
sponds to the path with maximum accumulated metric.
Figure 5.22 provides the performance of the two variants, with hard and
weighted inputs, of a decoder using the Viterbi algorithm for the code (7,5)
RSC for a transmission on a channel with additive white Gaussian noise. In
practice, we observe a gain of around 2 dB when we substitute weighted input
decoding for hard input decoding.
Figure 5.22 – Example of correction performance of the Viterbi algorithm with hard
inputs and with weighted inputs on a Gaussian channel. Recursive systematic convo-
lutional code (RSC) with generator polynomials 7 (recursivity) and 5 (redundancy).
Coding rate R = 1/2.
5.4.3 The Maximum A Posteriori algorithm or MAP al-

gorithm
The Viterbi algorithm determines the codeword closest to the received word.
However, it does not necessarily minimize the error probability of the bits or
symbols transmitted. The MAP algorithm enables us to calculate the a posteri-
ori probability of each bit or each symbol transmitted and, at each instant, the
corresponding decoder selects the most probable bit or symbol. This algorithm
was published in 1974 by Bahl, Cocke, Jelinek and Raviv [5.1]. The impact of
this decoding method remained little known until the discovery of turbocodes
since it does not provide any notable improvement in performance compared
to the Viterbi algorithm for decoding convolutional codes and it turned out
to be more complex to implement. However, the situation changed in 1993 as
decoding turbocodes uses elementary decoders with weighted or soft outputs
and the MAP algorithm, unlike the Viterbi algorithm, enables us to associate
a weighting naturally with each decision. The MAP algorithm is presented in
Chapter 7.
5.5 Convolutional block codes

Convolutional codes are naturally adapted to transmission applications where
the message transmitted is of infinite length. However, most telecommunica-
tions systems use independent frame transmissions. Paragraph 5.4.2 showed the
importance of knowing the initial and final states of the encoder during the de-
coding of a frame. In order to know these states, the technique used is usually
called trellis termination. This generally involves forcing the initial and final
states to values known by the decoder (in general zero).
5.5.1 Trellis termination

Classical trellis termination
As the encoder is constructed around a register, it is easy to initialize it by using
reset inputs before beginning the encoding of a frame. This operation has no
consequence on the coding rate. But termination at the end of a frame is not
so simple.
When the k bits of the frame have been coded, the register of the encoder is
in any of the 2ν possible states. The aim of termination is to lead the encoder
towards state zero by following one of the paths in the trellis so that the decoding
algorithm can use this knowledge of the final state. In the case of non-recursive
codes, the final state is forced to zero by injecting ν zero bits at the end of the
frame. It is as if the coded frame were of length k + ν with dk+1 = dk+2 = . . . =
dk+ν = 0. The encoding rate is slightly decreased by the transmission of the
termination bits. However, taking into account the size of the frames generally
transmitted, this degradation in rate is very often negligible. In the case of
recursive codes, it is also possible to inject a zero at the input of the register.
Figure 5.23 shows a simple way to solve this question.
d’i d’i , r i
ri Mux
di s (1)
i
1 D D D s (3)
i
2
s (0)
i
(2)
si
Figure 5.23 – Example of an encoder of recursive systematic convolutional codes al-

lowing termination at state 0 in ν instants.
After initializing the register to zero, switch I is kept in position 1 and data
d1 to dk are coded. At the end of this encoding operation, instants k to k + ν,
switch I is placed in position 2 and di takes the value coming from the feedback
(0)
of the register, that is, a value that forces one register input to zero. Indeed, Si
is the result of a modulo-2 sum of two identical members. As for the encoder,
it continues to produce the associated redundancies ri .
This classical termination has one main drawback: the protection of the data
is not independent of their position in the frame. In particular, this can lead to
edge effects in the construction of a turbocode (see Chapter 7).
Tail-biting
A technique was introduced in the 70s and 80s [5.12] to terminate the trellis of
convolutional codes without edge effects: tail-biting. This involves making the
decoding trellis circular, that is, ensuring that the departure and the final states
of the encoder are identical. This state is then called the circulation state. This
technique is trivial for non-recursive codes as the circulation state is merely the
last ν bits of the sequence to encode. As for RSC codes, tail-biting requires
operations that are described in the following. The trellis of such a code, called
circular recursive systematic codes (CRSC), is shown in Figure 5.24.
(000)
(001)
(010)
(011)
(100)
(101)
(110)
(111)
Figure 5.24 – Trellis of a CRSC code with 8 states.
It is possible to establish a relation between state si+1 of the encoder at an

instant i + 1, its state si and the input data di at the previous time:
si+1 = Asi + Bdi (5.15)
where A is the state matrix and B the input matrix. In the case of the recur-
sive systematic code of generator polynomials [1, (1+D 2 + D 3 )/( 1+D + D3 )]
mentioned above, these matrices are
⎡ ⎤ ⎡ ⎤
1 0 1 1
A = ⎣ 1 0 0 ⎦ and B = ⎣ 0 ⎦.
0 1 0 0
If the encoder is initialized to state 0 (s0 = 0), the final state s0k obtained at the
end of a frame of length k is:

k
s0k = Aj−1 Bdk−j (5.16)
j=1
When it is initialized in any state sc , the final state sk is expressed as follows:

k
sk = Ak sc + Aj−1 Bdk−j (5.17)
j=1
For this state sk to be equal to the departure state sc and for the latter therefore
to become the circulation state, it is necessary and sufficient that:
- .
k
I − Ak sc = Aj−1 Bdk−j (5.18)
j=1
where I is the identity matrix of dimension ν × ν.

Thus, by introducing state s0k of the encoder after initialization to 0
- .−1 0
sc = I − Ak sk (5.19)
This is only possible on condition that the matrix (I − Ak ) is invertible : this is

the condition for the existence of the circulation state.
In conclusion, if the circulation state exists, it can be obtained in two steps.
The first involves coding the frame of k bits at the input after initializing the
encoder to state zero and keeping the termination state. The second is to simply
deduce the circulation state from the previous termination state and from a table
(obtained by inverting I − Ak ).
Take for example the recursive systematic code used above. Since it is a
binary code, addition and subtraction are equivalent: the circulation state exists
if I + Ak is invertible. Matrix A is such that A7 is equal to I. Thus, if k is a
multiple of 7(= 2ν − 1), I + Ak is null and therefore non-invertible: this case
should be avoided. Another consequence is that Ak = Ak mod 7 : it suffices to
calculate once and for the 6 state transformation tables associated with the 6
possible values of (I − Ak )−1 , to store them and to read the right table, after
calculating k mod 7. The table of the CRSC code with 8 states of generator
polynomials [1, (1 + D 2 + D3 )/(1 + D + D 3 )] is given in Table 5.2.
A simple method for encoding according to a circular trellis can be summa-
rized in five steps, after checking the existence of the circulation state:
1. Initialize the encoder to state 0

2. Code the frame to obtain the final state s0k ;
XX
XXXk mod 7
XXX 1 2 3 4 5 6
s0k XX
0 0 0 0 0 0 0
1 6 3 5 4 2 7
2 4 7 3 1 5 6
3 2 4 6 5 7 1
4 7 5 2 6 1 3
5 1 6 7 2 3 4
6 3 2 1 7 4 5
7 5 1 4 3 6 2
Table 5.2 – Table of the CRSC code with generator polynomials [1, (1+D2 +D 3 )/
(1 + D + D3 )] providing the circulation state as a function of k mod 7 (k being
the length of the frame at the input) and of the terminal state s0k obtained after
encoding initialized to state 0.
3. Calculate the circulation state sc from the tables already calculated and
stored;
4. Initialize the encoder to state sc ;
5. Code the frame and transmit the redundancies calculated.
5.5.2 Puncturing
Some applications can only allocate a small space for the redundant part of the
codewords. But, by construction, the natural rate of a systematic convolutional
code is m/(m + n), where m is the number of input bits di of the encoder and n
is the number of output bits. It is therefore maximum when n = 1 and becomes
R = m/(m + 1). High rates can therefore only be obtained with high values
of m. Unfortunately, the number of transitions leaving any one node of the
trellis is 2m . In other words, the complexity of the trellis, and therefore of the
decoding, increases exponentially with the number of input bits of the encoder.
Therefore, this solution is generally not satisfactory. It is often avoided in favour
of a technique with a slightly lower error correction capability, but easier to
implement: puncturing.
The puncturing technique is commonly used to obtain high rates. It involves
using an encoder with a low value of m (1 or 2 for example), to keep a reasonable
decoding complexity, but transmitting only part of the bits coded. An example
is proposed in Figure 5.25. In this example, a 1/2 rate encoder produces outputs
di and ri at each instant i. Only 3 bits out of 4 are transmitted, which leads to
a global rate of 2/3. The pattern in which the bits are punctured is called the
puncturing mask.
d0 r 0 d1 r 1 d2 r 2 d3 r 3 di ri dk2 rk2 dk1 rk 1
Figure 5.25 – Puncturing a systematic code to obtain a rate 2/3.
In the case of systematic codes, it is generally the redundancy that is punc-

tured. Figure 5.26 shows the trellis of code [1, (1 + D 2 + D3 )/(1 + D + D3 )]
resulting from a puncturing operation according to the mask of Figure 5.25. The
"X"s mark the bits that are not transmitted and that therefore cannot be used
for the decoding.
00 0X 00 0X 00
(000)
11 1X 11
1X 11
(001)
0X 00
11 1X
01 0X 01
(010)
10 1X 10
1X 10
(011) 1X
0X 01
10
1X 10
(100)
01 0X 01
0X 0X 01
(101) 1X 10
11
1X 11
(110) 00 0X 00
0X 00
(111) 1X 11
di ri di =0
di ri di =1
Figure 5.26 – Trellis diagram of the punctured recursive code for a rate 2/3.
The most widely used decoding technique involves taking the decoder of
the original code and inserting neutral values in the place of the punctured
elements. The neutral values are values representing information that is a priori
not known. In the usual case of a transmission using antipodal signalling (+1
for the logical ’1’, -1 for the logical ’0’), the null value (analogue 0) is taken as
the neutral value.
The introduction of puncturing increases the coding rate but, of course,
decreases its correction capability. Thus, in the example of Figure 5.26, the free
distance of the code is reduced from 6 to 4 (an associated RTZ sequence is shown
in the figure). Likewise Figure 5.27, in which we present the error rate curves
of code [1, (1 + D2 + D3 )/(1 + D + D3 )] for rates 1/2, 2/3, 3/4 and 6/7, shows
a decrease in error correction capability with the increase in coding rate.
The choice of puncturing mask obviously influences the performance of the
code. It is thus possible to favour one part of the frame, transporting sensitive
data, by slightly puncturing it to the detriment of another part that is more
highly punctured. A regular mask is, however, often chosen as it is simple to
implement.
Figure 5.27 – Simulated performance of a CRSC code with 8 punctured states, as a

function of its rate.
Puncturing is therefore a flexible technique, and easy to implement. As the

encoder and the decoder remain identical whatever the puncturing applied, it
is possible to modify the encoding rate at any moment. Some applications use
this flexibility to adapt the rate, as they go along, to the channel and/or to the
importance of the data transmitted.
Bibliography
[5.1] L. R. Bahl, J. Cocke, F. Jelinek, and J. Raviv. Optimal decoding of
linear codes for minimizing symbol error rate. IEEE Transactions on
Information Theory, IT-20:284–287, March 1974.
[5.2] G. Battail. Weighting of the symbols decoded by the viterbi algorithm.
Annals of Telecommunications, 42(1-2):31–38, Jan.-Feb. 1987.
[5.3] C. Berrou, P. Adde, E. Angui, and S. Faudeuil. A low complexity soft-

output viterbi decoder architecture. In Proceedings of IEEE International
Conference on Communications (ICC’93), pages 737–740, GENEVA, May
1993.
[5.4] C. Berrou, A. Glavieux, and P. Thitimajshima. Near shannon limit error-

correcting coding and decoding: turbo-codes. In Proceedings of IEEE In-
ternational Conference on Communications (ICC’93), pages 1064–1070,

GENEVA, May 1993.
[5.5] P. Elias. Coding for noisy channels. IRE Convention Records, 3(4):37–46,
1955.
[5.6] R. M. Fano. A heuristic discussion of probabilistic decoding. IEEE Trans-
actions on Information Theory, IT-9:64–74, Apr. 1963.
[5.7] G. D. Forney. Convolutional codes i: Algebraic structure. IEEE Trans-
actions on Information Theory, IT-16:720–738, Nov. 1970.
[5.8] G. D. Forney. The viterbi algorithm. Proceedings of the IEEE, 61(3):268–
278, March 1973.
[5.9] A. Glavieux. Codage de canal, des bases théoriques aux turbocodes.
Hermès-Science, 2005.
[5.10] J. Hagenauer and P. Hoeher. A viterbi algorithm with soft-decision out-

puts and its applications. In Proceedings of IEEE Global Communications
Conference (Globecom’89), pages 1680–1686, Dallas, Texas, USA, Nov.
1989.
[5.11] R. Johannesson and K. Sh. Zigangirov. Fundamentals of Convolutional

Coding. Piscataway, IEEE Press, 1999.
[5.12] H. H. Ma and J. K. Wolf. On tail-biting convolutional codes. IEEE

Transactions on Communications, COM-34:104–111, Feb. 1986.
[5.13] P. Thitimajshima. Les codes convolutifs récursifs systématiques et leur ap-

plication à la concaténation parallèle. Phd thesis, Université de Bretagne
Occidentale, Dec. 1993.
[5.14] A. J. Viterbi. Coding for noisy channels. IEEE Transactions on Infor-
mation Theory, IT-13:260–269, Apr. 1967.
[5.15] J. M. Wozencraft. Sequential decoding for reliable communication. IRE
National Convention Records, 5 part 2(2):11–25, 1957.
Chapter 6
Concatenated codes
The previous chapters presented the elementary laws of encoding like BCH,
Reed-Solomon or CRSC codes. Most of these elementary codes are asymptoti-
cally good, in the sense that their minimum Hamming distances (MHD) can be
made as large as we want, by sufficiently increasing the degree of the generator
polynomials. The complexity of the decoders is unfortunately unacceptable for
the degrees of polynomials that would guarantee the MHD required by practical
applications.
A simple means of having codes with a large MHD and nevertheless easily
decodable is to combine several reasonably-sized elementary codes, in such a
way that the resulting global code has a high error correction capability. The
decoding is performed in steps, each of them corresponding to one of the ele-
mentary encoding steps. The first composite encoding scheme was proposed by
Forney during work on his thesis in 1965, called concatenated codes [6.4]. In
this scheme, a first encoder, called the outer encoder, provides a codeword that
is then re-encoded by a second encoder, called the inner encoder. If the two
codes are systematic, the concatenated code is itself systematic. In the rest of
this chapter, only systematic codes will be considered.
Figure 6.1(a) shows a concatenated code, as imagined by Forney, and the
corresponding step decoder. The most judicious choice of constituent code is an
algebraic code, typically a Reed-Solomon code, for the outer code, and a convo-
lutional code for the inner code. The inner decoder is then the Viterbi decoder,
which easily takes advantage of the soft values provided by the demodulator,
and the outer decoder, which works on symbols with several bits (for example,
8 bits), can handle errors in bursts at the output of the first decoder. A per-
mutation or interleaving function inserted between the two encoders, and its
inverse function placed between the two decoders, can greatly increase the ro-
bustness of the concatenated code (Figure 6.1(b)). Such an encoding scheme has
worked very successfully in applications as varied as deep space transmissions
and digital, satellite and terrestrial television broadcasting. In particular, it is
the encoding scheme adopted in many countries for digital terrestrial television
[6.1].
Figure 6.1 – serial concatenated code, (a) without and (b) with permutation. In both
cases, the output of the outer encoder is entirely recoded by the inner encoder.
Nowadays, this first version of concatenated codes is called serial concatena-

tion (SC). Its decoding, presented in Figures 6.1 is not optimal. Indeed, even if,
locally, the two elementary decoders are optimal, the simple sequencing of these
two decodings is not globally optimal as the inner decoder does not take any
advantage of the redundancy produced by the outer code. It is this observation,
that occurred fairly late in the history of information theory, that led to the
development of new decoding principles, beginning with turbo decoding. We
now know how to decode, quasi-optimally, all sorts of concatenated schemes,
with the sole condition that the decoders of elementary codes are of the SISO
(soft-in/soft-out) type. In this sense, we can note that the concept of concate-
nation has greatly evolved in the last few years, moving towards a wider notion
of multi-dimensional encoding. Here, the dimension of a code, which should not
be confused with the length (k) of the information message that we also call
dimension, is the number of elementary codes used in the production of the final
codeword.
Figure 6.2 – Parallel concatenation of systematic encoders.

6. Concatenated codes 203
A new form of concatenation, called parallel concatenation (PC), was intro-

duced at the beginning of the 1990s to elaborate turbo codes [6.3]. Figure 6.2
presents a PC with dimension 2, which is the classical dimension for turbo codes.
In this scheme, the message is coded twice, in its natural order and in a per-
muted order. The redundant part of the codeword is formed by concatenating
the redundant outputs of the two encoders. PCs differ from SCs in several ways,
described in the next section.
6.1 Parallel concatenation and serial concatena-

tion
Limiting ourselves to dimension 2, the PC, which associates two elementary
codes with rates R1 (code C1 ) and R2 (code C2 ), has a global encoding rate:
R1 R2 R1 R2
Rp = = (6.1)
R1 + R2 − R1 R2 1 − (1 − R1 )(1 − R2 )
This rate is higher than the global rate Rs of a serial concatenated code (Rs =
R1 R2 ), for identical values of R1 and R2 , and the lower the encoding rates
the greater the difference. We can deduce from this that with the same error
correction capability of component codes, parallel concatenation offers a better
encoding rate, but this advantage diminishes when the rates considered tend
towards 1. When the dimension of the composite code increases, the gap between
Rp and Rs also increases. For example, three component codes of rate 1/2
form a concatenated code with global rate 1/4 for parallel, and 1/8 for serial
concatenation. That is the reason why it does not seem to be useful to increase
the dimension of a serial concatenated code beyond 2, except for rates very close
to unity.
However, with SC, the redundant part of a word processed by the outer
decoder has benefited from the correction of the decoder(s) that precede(s) it.
Therefore, at first sight, the correction capability of a serial concatenated code
seems to be greater than that of a parallel concatenated code, in which the values
representing the redundant part are never corrected. In other terms, the MHD
of a serial concatenated code must normally be higher than that of a parallel
concatenated code. We therefore find ourselves faced with the dilemma given
in Chapter 1: PC performs better in the convergence zone (near the theoretical
limit) since the encoding rate is more favourable, and the SC behaves better
at low error rates thanks to a larger MHD. Encoding solutions based on the
SC of convolutional codes have been studied [6.3], which can be an interesting
alternative to classical turbo codes, when low error rates are required. Serial
convolutional concatenated codes will not, however, be described in the rest of
this book.
When the redundant parts of the inner and outer codewords both undergo
supplementary encoding, the concatenation is said to be double serial concate-
nation. The most well-known example of this type of encoding structure is
the product code, which implements BCH codes (see Chapters 4 and 8). Mixed
structures, combining parallel and serial concatenations have also been proposed
[6.6]. Moreover, elementary concatenated codes can be of a different nature, for
example a convolutional code and a BCH code [6.2]. We then speak of hybrid
concatenated codes. From the moment elementary decoders accept and produce
weighted values, all sorts of mixed and/or hybrid schemes can be imagined.
Whilst SC can use systematic or non-systematic codes indifferently, parallel
concatenation uses systematic codes. If they are convolutional codes, at least
one of these codes must be recursive, for a fundamental reason to do with the
minimum input weight wmin , which is only 1 for non-recursive codes but is 2 for
recursive codes (see Chapter 5). To show this, see Figure 6.3 which presents two
non-recursive systematic codes, concatenated in parallel. The input sequence is
"all zero" (reference sequence) except in one position. This single "1" perturbs
the output of the encoder C1 for a short length of time, equal to the constraint
length 4 of the encoder. The redundant information Y1 is poor, in relation to
this particular sequence, as it contains only 3 values different from 0. After
permutation, of whatever type, the sequence is still "all zero", except in one
single position. Again, this "1" perturbs the output of the encoder C2 for a
length of time equal to the constraint length, and redundancy Y2 provided by the
second code is as poor in information as redundancy Y1 . In fact, the minimum
distance of this two-dimensional code is not higher than that of a single code,
with the same rate as that of the concatenated code. If we replace at least one
of the two non-recursive encoders by a recursive encoder, the "all zero" sequence
except in one position is no longer a "Return to Zero" (RTZ, see Section 5.3.2)
sequence for this recursive encoder, and the redundancy that it produces is thus
of much higher weight.
What we have explained above about the PC of non-recursive convolutional
codes suggests that the choice of elementary codes for the PC in general is
limited. As another example, let us build a parallel concatenated code from the
extended Hamming code defined by Figure 1.1 and the encoding Table 1.1. The
information message contains 16 bits, arranged in a 4x4 square (Figure 6.4(a)).
Each line and each column is encoded by the elementary Hamming code. The

horizontal and vertical parity bits are denoted ri,j and ri,j , respectively. The
global coding rate is 1/3. Decoding this type of code can be performed using the
principles of turbo decoding (optimal local decoding according to the maximum
likelihood and continuous exchanges of extrinsic information).
The MHD of the code is given by the pattern of errors of input weight 1
(Figure 6.4(b)). Whatever the position of the 1 in the information message,
the weight is 7. The figure of merit Rdmin (see Section 1.5) is therefore equal
to 7x(1/3), compared with the figure of merit of the elementary code 4x(1/2).
Figure 6.3 – The parallel concatenation of non-recursive systematic codes is a poor code
concerning the information sequences of weight 1. In this example, the redundancy
symbols Y1 and Y2 each contain only three 1.
Figure 6.4 – Parallel concatenation of extended Hamming codes (global rate: 1/3). On
the right: a pattern of errors of input weight 1 and total weight 7.
The asymptotic gain has therefore not been extraordinarily increased by means
of the concatenation (0,67 dB precisely), and a great reduction in the coding
rate has occurred. If we wish to keep the same global rate of 1/2, a part of the
redundancy must be punctured. We can choose, for example, not to transmit
the 16 symbols present in the last two columns and the last two lines of the
table of Figure 6.4(a). The MHD then drops to the value 3, that is, less than
the MHD of the elementary code. The PC is therefore of no interest in this case.
Again from the extended Hamming code, a double serial concatenation can
be elaborated in the form of a product code (Figure 6.5(a)). In this scheme,
the redundant parts of the horizontal and vertical codewords are themselves
re-encoded by elementary codes, which produce redundancy symbols denoted
wi,j . One useful algebraic property of this product code is the identity of the
redundancy symbols coming from the second level of encoding, in the horizontal
and vertical directions. The MHD of the code, which has a global rate 1/4, is
again given by the patterns of errors of input weight 1 and is equal to 16, that
is, the square of the MHD of the elementary code (Figure 6.5(b)). The figure
of merit Rdmin = 4 has therefore been greatly increased compared to parallel
concatenation. To attempt to increase the rate of this code by puncturing the
redundancy symbols while keeping a good MHD is bound to fail.
Figure 6.5 – Double serial concatenation (product code) of extended Hamming codes
(global rate: 1/4). On the right: a pattern of errors of input weight 1 and total weight
16.
In conclusion, parallel concatenation cannot be used with just any elementary

code. Today, only convolutional recursive systematic codes are used in this type
of concatenation, with 2 dimensions. Serial concatenation can offer large MHD.
The choice of codes is greater: convolutional codes, recursive or not, BCH codes
or Reed-Solomon codes. However, with the same coding rates of elementary
codes, serial concatenation has a lower global rate than parallel concatenation.
6.2 Parallel concatenation and LDPC codes

LDPC codes, which are described in Chapter 9, are codes where the lines and
columns of the parity check matrix contain few 1s. LDPC codes can be seen
as a multiple concatenation of n − k parity relations containing few variables.
Here it is not a concatenation in the sense that we defined above, since the
parity relations contain several redundancy variables and these variables appear
in several relations. We cannot therefore assimilate LDPC codes to standard
serial or parallel concatenation schemes. However, we can, like MacKay [6.5],
observe that a turbo code is an LDPC code. An RSC code with generator
polynomials GX (D) (recursivity) and GY (D) (redundancy), whose input is X
and redundant output Y , is characterized by the sliding parity relation:
GY (D)X(D) = GX (D)Y (D) (6.2)
Using the tail-biting technique (see CRSC, Section 5.5.1, the parity check
matrix takes a very regular form, such as the one presented in Figure 6.6 for a
coding rate 1/2, and choosing GX (D) = 1 + D + D3 and GY (D) = 1 + D2 + D 3 .
A CRSC code is therefore an LDPC code since the check matrix is sparse. This
is certainly not a good LDPC code, as the check matrix does not respect certain
properties about the positions of the 1s. In particular, the 1s on a same line
are very close to each other, which is not favourable to the belief propagation
decoding method.
Figure 6.6 – Check matrix of a tail-biting convolutional code. A convolutional code,

particularly of the CRSC type, can be seen as an LDPC code since the check matrix
is sparse.
A parallel concatenation of CRSC codes, that is a turbo code, is also an

LDPC code since it associates elementary codes that are of the LDPC type. Of
course, there are more degrees of freedom in the construction of an LDPC code,
as each 1 of the check matrix can be positioned independently of the others. On
the other hand, decoding a convolutional code, via an algorithm based on the
trellis, does not encounter the problem of correlation between successive symbols
that a belief propagation type of decoding would encounter, if it was applied to
a simple convolutional code. A turbo code cannot therefore be decoded like an
LDPC code.
6.3 Permutations
The functions of permutation or interleaving, used between elementary encoders
in a concatenated scheme, have a twofold role. On the one hand, they ensure, at
the output of each component decoder, a time spreading of the errors that can
be produced by it in bursts. These packets of errors then become isolated errors
for the following decoder, with far lower correlation effects. This technique for
the spreading of errors is used in a wider context than that of channel coding.
We can use it profitably, for example, to reduce the effects of more or less long
attenuation in transmissions affected by fading, and more generally in situations
where perturbations can alter consecutive symbols. On the other hand, in close
liaison with the characteristics of constituent codes, the permutation is designed
so that the MHD of the concatenated code is as large as possible. This is a
problem of pure mathematics associating geometry, algebra and combinatory
logic which, in most cases, has not yet found a definitive answer. Sections 7.3.2
and 9.1.6 develop the topic of permutation for turbo codes and graphs for LDPC
codes, respectively.
6.4 Turbo crossword

To end this chapter, here is an example of parallel concatenation that is familiar
to everyone: crosswords. The content of a grid has been altered during its
retranscription, as can be seen in Figure 6.7. Fortunately, we have a correct clue
for each line and for each column and we have at our disposal a dictionary of
synonyms.
Figure 6.7 – Crossword grid with wrong answers but correct clues.
To correct (or decode) this grid, we must operate iteratively by line and by
column. The basic decoding rule is the following: "If there is a word in the
dictionary, a synonym or an equivalent to the definition given that differs from
the word in the grid by at most one letter, then this synonym is adopted".
Figure 6.8 – First iteration of the line - column decoding process.
The horizontal definitions allow us to begin correcting the lines in the grid
(Figure 6.8(a)):
I. There must be more than one wrong letter.

II. THETA is a Greek letter.
III. We can replace the L with a C, getting CANES (sticks).
IV. There must be more than one wrong letter.
V. There must be more than one wrong letter
After decoding this line, two words are correct or have been corrected, and
three are still to be found. Using the vertical definitions, we can now decode the
columns (Figure 6.8(b)):
1. There must be more than one wrong letter. No correction is possible.

2. Replacing the I with an E, we get SHAPE (representation).
3. A TENON is a projection (of wood).
4. Replacing the T with a G, we get AGENT (force).
5. A LASSO is a kind of rope.
After decoding the columns, there are still some unknown words and we have to
perform a second iteration of the line - column decoding process (Figure 6.9).
Line decoding leads to the following result (Figure 6.9(a)):
I. Replacing the S with an I, we get VITAL (animate).

Figure 6.9 – Second iteration of the line - column decoding process.
II. There must be at least 2 wrong letters.

III. CANES is correct
IV. Replacing the P with a G, we get AGONS (conflicts)
V. We can replace the P with an L, getting LENTO (slow).
After this step, there is still one wrong line. It is possible to correct it by
decoding the columns again (Figure 6.9(b)).
1. VOCAL is a synonym of oral.

2. IMAGE is also a kind of representation.
3. TENON is correct.
4. AGENT is correct.
5. LASSO is correct.
After this final decoding step, the wrong word on line II is identified: it is
OMEGA, another Greek letter.
A certain number of remarks can be made after this experience of decoding

a parallel concatenated code.
• To arrive at the right result, two iterations of the line and column decoding
were necessary. It would have been a pity to stop after just one iteration.
But that is what we do in the case of a classical concatenated code such
as that of Figure 6.1.
• A word that was correct at the beginning (THETA is indeed a Greek letter)
turned out to be wrong. Likewise, a correction made during the second
step (SHAPE) turned out to be wrong. So the intermediate results must
be considered with some caution and we must avoid any hasty decisions.
In modern iterative decoders, this caution is measured by a probability
that is never exactly 0 or 1.
Bibliography
[6.1] Dvb-terrestrial. ETSI EN 302 296 V1.1.1 (2004-04).
[6.2] P. Adde, R. Pyndiah, and C. Berrou. Performance of hybrid turbo codes.
Elect. Letters, 32(24):2209–2210, Nov. 1996.
[6.3] S. Benedetto, D. Divsalar, G. Montorsi, and F. Pollara. Serial concate-
nation of interleaved codes: performance analysis, design, and iterative
decoding. IEEE Trans. Info. Theory, 44(3):909–926, May 1998.
[6.4] Jr. G. D. Forney. Performance of concatenated codes. In E. R. Berlekamp,
editor, Key papers in the development of coding theory, pages 90–94. IEEE
Press, 1974.
[6.5] D. J. C. MacKay. Good error-correcting codes based on very sparse ma-
trices. IEEE Transactions on Information Theory, 45(2):399–431, March
1999.
[6.6] K. R. Narayanan and G. L. Stüber. Selective serial concatenation of turbo
codes. IEEE Comm. Letters, 1(5):136–139, Sept. 1997.
Chapter 7
Convolutional turbo codes
The error correction capability of a convolutional code increases when the length
of the encoding register increases. This is shown in Figure 7.1, which provides
the performance of four RSC codes with respective memories ν = 2, 4, 6 and 8,
for rates 1/2, 2/3, 3/4 and 4/5, decoded according to the MAP algorithm. For
each of the rates, the error correction capability improves with the increase in ν,
above a certain signal to noise ratio that we can assimilate almost perfectly with
the theoretical limit calculated in Chapter 3 and identified here by an arrow.
To satisfy the most common applications of channel coding, a memory of the
order of 30 or 40 would be necessary (from a certain length of register and for a
coding rate 1/2, the minimum Hamming distance of a convolutional code with
memory ν is of the order of ν). If we knew how to easily decode a convolutional
code with over a billion states, we would no longer speak much about channel
coding and this book would not exist.
A turbo code is a coding trick, aiming to imitate a convolutional code with
a large memory ν. It is built on the principle of the saying divide and rule, that
is, by associating several small RSC codes whose particular decodings are of rea-
sonable complexity. A judicious exchange of information between the elemen-
tary decoders enables the composite decoder to approximate the performance
of maximum likelihood decoding.
7.1 The history of turbo codes

The invention of turbo codes is not the outcome of a mathematical development.
It is the result of an intuitive experimental approach whose origin can be found
in the work of several European researchers: Gerard Battail, Joachim Hagenauer
and Peter Hoeher who, at the end of the 80s [7.8, 7.7, 7.31, 7.30] highlighted
the interest of probabilistic processing in receivers. Others before them, mainly
in the United States: Peter Elias [7.25], Michael Tanner [7.45], Robert Gallager
Figure 7.1 – Performance of recursive systematic convolutional codes (RSC) for dif-
ferent rates and four values of the memory of code ν. Comparison with Shannon
limits.
[7.26], etc. had earlier imagined procedures for coding and decoding that were
the forerunners of turbo codes.
In a laboratory at École Nationale Supérieure des Télécommunications de
Bretagne (Telecom Bretagne), Claude Berrou and Patrick Adde were attempt-
ing to transcribe the Viterbi algorithm with weighted input (SOVA: Soft-Output
Viterbi Algorithm) proposed in [7.7], into MOS transistors, in the simplest possi-
ble way. A suitable solution [7.10] was found after two years which enabled these
researchers to form an opinion about probabilistic decoding. Claude Berrou,
then Alain Glavieux, pursued the study and observed, after Gerard Battail,
that a decoder with weighted input and output could be considered as a sig-
nal to noise ratio amplifier. This encouraged them to implement the concepts
commonly used in amplifiers, mainly feedback. Perfecting turbo codes involved
many very pragmatic stages and also the introduction of neologisms, like "paral-
lel concatenation" or "extrinsic information", nowadays common in information
theory jargon. The publication in 1993 of the first results [7.14], with a perfor-
mance 0,5 dB from the Shannon limit, shook the coding community. A gain of
almost 3 dB, compared to solutions existing at that time, had been found by
a small team that was not only unknown, but also French (France, a country
known for its mathematical rigour, versus turbo codes, an empirical invention
to say the least). There followed a very distinct evolution in habits, as under-
lined by A. R. Calderbank in [7.20] (p. 2573): "It is interesting to observe that
the search for theoretical understanding of turbo codes has transformed coding
theorists into experimental scientists"
[7.13] presents a chronology describing the successive ideas that appeared in
the search to perfect turbo codes. This new coding and decoding technique was
7. Convolutional turbo codes 215
first baptized turbo-code, with a hyphen to show that it was a code decoded in
a turbo way (by analogy with the turbo engine that uses exhaust gas to increase
its power). As the hyphen is not used much in English, it became turbo code,
that is, a "turbo" code, which does not mean very much. In French today, turbo
code is written as a single word: turbocode.
7.2 Multiple concatenation of RSC codes
Figure 7.2 – Multiple parallel concatenation of circular recursive systematic convo-

lutional (CRSC) codes. Each encoder produces k/N redundancy symbols uniformly
distributed on the circular trellis. Global coding rate: 1/2.
Since the seminal work of Shannon, random codes have always been a ref-
erence for error correction coding (see Section 3.1.5). The systematic random
coding of a block of k information bits, leading to a codeword of length n, can,
as the first step and once for all, involve drawing at random and memorizing k
binary markers containing n − k bits, whose memorization address is denoted
i (0 ≤ i ≤ k − 1). The redundancy associated with any block of information
is then formed by the modulo 2 sum of all the markers whose address i is such
that the i-th information bit equals 1. In other words, the k markers are the
bases of a vector space of dimension k. The codeword is finally made up of the
concatenation of the k information bits and of the n − k redundancy bits. The
rate R of the code is k/n. This very simple construction of the codeword relies
on the linearity property of the addition and leads to high minimum distances
for sufficiently large values of n − k. Because two codewords are different by at
least one information bit and the redundancy is drawn at random, the average
minimum distance is 1 + n−k 2 . However, the minimum distance of this code
being a random variable, its different realizations can be lower than this value.
A simple realistic approximation of the effective minimum distance is n−k 4 .
A way to build an almost random encoder is presented in Figure 7.2. It is
a multiple parallel concatenation of circular recursive systematic convolutional
codes (CRSC, see Chapter 5) [7.12]. The sequence of k binary data is coded N
times by N CRSC encoders, in a different order each time. The permutations Πj
are drawn at random, except the first one that can be the identity permutation.
k
Each elementary encoder produces N redundancy symbols (N being a divisor
of k), the global rate of the concatenated code being 1/2.
The proportion of input sequences of a recursive encoder built from a pseudo-
random generator with memory ν, initially positioned in state 0, which return
the register back to the same state at the end of the coding, is:
p1 = 2−ν (7.1)
since there are 2ν possible return states, with the same probability. These
sequences, called Return To Zero (RTZ) sequences,(see Chapter 5), are linear
combinations of the minimum RTZ sequence, which is given by the recursivity
polynomial of the generator (1 + D + D 3 in the case of Figure 7.2).
The proportion of RTZ sequences for the multi-dimensional encoder is low-
ered to:
pN = 2−N ν (7.2)
since the sequence must, after each permutation, remain RTZ for the N encoders.
The other sequences, with proportion 1 − pN , produce codewords that have
a distance d satisfying:
k
d> (7.3)
2N
This worst case value assumes that a single permuted sequence is not RTZ
and that redundancy Y takes the value 1, every other time on average, on the
corresponding circle. If we take N = 8 and ν = 3 for example, we obtain
p8 ≈ 10−7 and, for sequences to encode of length k = 1024, we have dmin = 64,
which is a sufficient minimum distance if we refer to the curves of Figure 3.6.
Random coding can thus be approximated by using small codes and random
permutations. The decoding can be performed following the turbo principle,
described in Section 7.4 for N = 2. The scheme of Figure 7.2 is, however, not
used in practice, for reasons linked to the performance and complexity of the
decoding. First, the convergence threshold of the turbo decoder, that is, the
signal to noise ratio from which the turbo decoder can begin to correct most of
the errors, degrades when the dimension of the concatenation increases. Indeed,
the very principle of turbo decoding means considering the elementary codes
one after the other, iteratively. As their redundancy rate decreases when the
dimension of the composite code increases, the first steps in the decoding are
penalized compared to a concatenated code with a simple dimension 2. Then,
the complexity and the latency of the decoder are proportional to the number
of elementary encoders.
7.3 Turbo codes

Fortunately, concerning the above, it is not necessary to carry dimension N
to a high value. By replacing the random permutation Π2 by a judiciously
elaborated permutation, good performance can be obtained by limiting ourselves
to a dimension N = 2. That is the principle of turbo codes.
Figure 7.3 – A binary turbo code with memory ν = 3 using identical elementary RSC
encoders (polynomials 15, 13). The natural coding rate of the turbo code, without
puncturing, is 1/3.
Figure 7.3 presents a turbo code in its most classical version [7.14]. The
binary input message, of length k, is encoded in its natural order and in a
permuted order by two RSC encoders called C1 and C2 , which can be terminated
or not. In this example, the two elementary encoders are identical (generator
polynomials 15 for the recursivity and 13 for the construction of the redundancy)
but this is not a necessity. The natural coding rate, without puncturing, is 1/3.
To obtain higher rates, redundancy symbols Y1 and Y2 are punctured. Another
way to have higher rates is to adopt m-binary codes (see 7.5.2).
As the permutation function (Π) concerns a message of finite size k, the turbo
code is by construction a block code. However, to distinguish it from concate-
nated algebraic codes decoded in a "turbo" way, like product codes which were
later called block turbo codes, this turbo coding scheme is called a convolutional
code or, more technically, a Parallel Concatenated Convolutional Code (PCCC).
Arguments in favour of this coding scheme (some of which have already been
introduced in Chapter 6) are the following:
1. A decoder for convolutional codes is vulnerable to errors arriving in pack-
ets. Coding the message twice, following two different orders (before and
after permutation), makes fairly improbable the simultaneous appearance
of error packets at the input of the decoders of C1 and C2 . If there are
errors grouped at the input of the decoder of C1 , the permutation dis-
perses them over time and they become isolated errors that are easy for
the decoder of C2 to correct. This reasoning also holds for error packets at
the input of this second decoder, which correspond, before permutation,
to isolated errors. Thus two-dimensional coding, on at least one of the
two dimensions, greatly reduces the vulnerability of convolutional coding
concerning grouped perturbations. But which of the two decoders should
be relied on to take the final decision? No criterion allows us to be more
confident about one or the other. The answer is given by the "turbo" algo-
rithm that avoids having to make this choice. This algorithm implements
exchanges of probabilities between the two decoders and constrains them
to converge, during these exchanges, towards the same decisions.
2. As we saw in Section 6.1, parallel concatenation leads to a higher coding
rate than that of serial concatenation. Parallel concatenation is therefore
more favourable when signal to noise ratios close to the theoretical limits
are considered, with average error rates targeted. It can be different when
very low error rate are sought for since the MHD of a serial concatenated
code can be larger.
3. Parallel concatenation uses systematic codes and at least one of these codes
must be recursive, for reasons also described in Section 6.1
4. Elementary codes are small codes: codes with 16, 8, or even 4 states. Even
if the decoding implements repeated probabilistic processing, it remains of
reasonable complexity.
Figure 7.4 presents turbo codes used in practice and Table 7.2 lists some
industrial applications. For a detailed overview of the applications of turbo and
LDPC codes see [7.29]. The parameters defining a particular turbo code are the
following:
a– m is the number of bits in the symbols applied to the turbo encoder. The
applications known to this day consider binary (m = 1) or double-binary
(m = 2) symbols.
Figure 7.4 – Turbo codes used in practice.
b– Each of the two elementary encoders C1 and C2 is characterized by
– its code memory ν

– its generator polynomials for recursivity and redundancy
– its rate
The values of ν are in practice lower than or equal to 4. The generator

polynomials are generally those that we use for classical convolutional codes
and that were the subject of much literature in the 1980s and 1990s.
c– The way in which we perform the permutation is important when the tar-
geted binary error rate is lower than around 10−5 . Above this value, per-
Application Turbo Termination Polynomials Rates

code
CCSDS binary, tail bits 23, 33, 25, 37 1/6, 1/4, 1/3, 1/2
(deep space) 16 states
UMTS, binary, 8 tail bits 13, 15, 17 1/4, 1/3, 1/2
CDMA2000 states
(mobile 3G)
DVB-RCS double- circular 15, 13 from 1/3 to 6/7
(Return binary, 8
Channel on states
Satellite)
DVB-RCT double- circular 15, 13 1/2, 3/4
(Return binary, 8
Channel on states
Terrestrial)
Inmarsat binary, none 23, 35 1/2
(M4) 16 states
Eutelsat double- circular 15, 13 4/5, 6/7
(skyplex) binary, 8
states
IEEE 802.16 double- circular 15, 13 from 1/2 to 7/8
(WiMAX) binary, 8
states
Table 7.2 – Standardized applications of convolutional turbo codes.
formance is not very sensitive to permutation, on condition of course that

it respects at least the principle of dispersion (which, for example, can be a
regular permutation). For a low or very low targeted error rate, performance
is dictated by the minimum distance of the code and the latter is highly
dependent on the permutation Π.
d– The puncturing pattern must be the as regular as possible, in the same way
as for classical convolutional codes. However, it can be advantageous to
have a slightly irregular puncturing pattern when we are looking for very
low error rates and when the puncturing period is a divisor of the period of
the generator polynomial of recursivity or parity.
Puncturing is performed classically on the redundancy symbols. It can be

envisaged instead to puncture the information symbols, in order to increase the
minimum distance of the code. This is done to the detriment of the convergence
threshold of the turbo decoder. From this point of view, in fact, puncturing
data shared by the two decoders is more penalizing than puncturing data that
are only useful to one of the decoders.
What must be closely considered when building a turbo code and decoding
it, are the RTZ sequences, whose output weights limit the minimum distance of
the code and fix its asymptotic performance. In what follows it will be assumed
that the error patterns that are not RTZ do not contribute to the MHD of the
turbo code and will therefore not have to be considered.
7.3.1 Termination of constituent codes

For a turbo code, there are two trellises to be terminated and the solutions
presented in Section 5.5.1 can be envisaged:
Doing nothing in particular concerning the terminal states: the data
situated at the end of the block, in either the natural order or in the permuted
order, are thus less well protected. This leads to a decrease in the asymptotic
gain, but this degradation, which is a function of the size of the block, may be
compatible with some applications. It should be noted that non-termination of
the trellis penalizes the PER (Packet Error Rates) more greatly than the BER.
Terminating the trellis of one or both elementary codes using tail
bits: CCSDS [7.4] and UMTS [7.3] standards use this technique. The bits
ensuring the termination of one of the two trellises are not used in the other
encoder. These bits are therefore not turbo encoded, which leads, but to a
lesser degree, to the same drawbacks as those presented in the previous case.
Moreover, the transmission of the tail bits causes a decrease in the coding rate
and therefore in the spectral efficiency.
Using interleaving enables the automatic termination of the trellis:
it is possible to close the trellis of a turbo code automatically, without adding any
tail bits, by slightly transforming the coding scheme (self-concatenation) and by
using interleaving that respects certain periodicity rules. This solution described
in [7.15] does not decrease the spectral efficiency but imposes constraints on the
interleaving which makes it difficult to control performance at low error rates.
Adopting circular encoding: a circular encoder for convolutional codes
guarantees that the initial state and the final state of the register are identical.
The trellis then takes the form of a circle, which, from the point of view of the
decoder, can be considered as a trellis with infinite length [7.32, 7.48]. This
termination process, already known as tail-biting for non-recursive codes, offers
two main advantages:
• Unlike the other techniques, circular termination does not present any edge
effects: all the bits of the message are protected in the same way and are
all doubly encoded by the turbo code. Therefore, during the design of the
permutation, there is no need to give special importance to such and such
a bit, which leads to simpler permutation models.
• The sequences that are not RTZ have an influence on the whole circle: on
average, one parity symbol out of two is modified along the block. For
typical values of k (a few hundred or more), the corresponding output
weight is therefore very high and these error patterns do not contribute
to the MHD of the code, as already mentioned at the end of the previous
section. Without termination or with termination using tail bits, only the
part of the block after the beginning of the non-RTZ sequence has any
effect on the parity symbols.
To these two advantages we can, of course, add the interest of having to trans-
mit no additional information about termination and therefore losing nothing
in spectral efficiency.
The circular termination technique was chosen for the DVB-RCS and DVB-
RCT [7.2, 7.1] standards, for example.
7.3.2 The permutation function

Whether we call it interleaving or permutation, the technique that involves dis-
persing the data over time has always been very useful in digital communications.
For example, we use it profitably to reduce the effects of more or less long at-
tenuations in transmissions affected by fading and, more generally, in situations
where perturbations can alter consecutive symbols. In the case of turbo codes
too, permutation allows us to efficiently combat the appearance of error packets,
on at least one of the dimensions of the composite code. But its role does not
stop there: in close relation with the properties of the constituent codes, it also
determines the minimum distance of the concatenated code.
Let us consider the turbo code presented in Figure 7.3. The worst permu-
tation that we can use is, of course, identity permutation, which minimizes the
diversity of the coding (we then have Y1 = Y2 ). On the other hand, the best
permutation that we could imagine, but that probably does not exist [7.42],
would enable the concatenated code to be equivalent to a sequential machine of
which the number of irreducible states would be 2k+6 . There are indeed k + 6
binary memorization elements in the structure: k for the permutation memory
and 6 for the two convolutional codes. If we could assimilate this sequential
machine to a convolutional encoder and for common values of k, the number of
corresponding states would be very large, in any case large enough to guarantee
a large minimum distance. For example, a convolutional encoder with a code
memory of 60 (1018 states !) shows a free distance of the order of a hundred (for
R = 1/2), which is quite sufficient.
Thus, from the worst to the best permutation, there is a wide choice and
we have not yet discovered any perfect permutation. Having said that, good
permutations have been defined even so in order to elaborate standardized turbo
coding schemes.
There are two ways to specify a permutation, the first by equations linking
addresses before and after permutation, the second by a look-up table providing
the correspondence between addresses. The first is preferable from the point
of view of simplicity in the specification of the turbo code (standardization
committees are sensitive to this aspect) but the second can lead to better results
since the degree of freedom is generally larger when designing the permutation.
Regular permutation
The point of departure when designing interleaving is regular permutation,
which is described in Figure 7.5 in two different forms. The first assumes that the
block containing k bits can be organized as a table of M rows and N columns.
The interleaving then involves writing the data in an ad hoc memory, row by
row, and reading them column by column (Figure 7.5(a)). The second applies
without any hypothesis about the value of k. After writing the data in a linear
memory (address i, 0 ≤ i ≤ k − 1), the block is assimilated to a circle, the two
extremities (i = 0 and i = k − 1) then being adjacent (Figure 7.5(b)). The
binary data are then extracted in such a way that the j-th datum read has been
previously written in position i, with value:
i = Π(j) = P j + i0 mod k (7.4)
where P is a prime integer with k and i0 is the index of departure1 .
Figure 7.5 – Regular permutation in rectangular (a) and circular (b) form.
For circular permutation, let us define the accumulated spatial distance

S(j1 , j2 ) as the sum of the two spatial distances separating two bits, before
and after permutation, whose reading indices are j1 and j2 :
S(j1 , j2 ) = f (j1 , j2 ) + f (Π(j1 ), Π(j2 )) (7.5)

1 The permutation can, of course, be defined in a reciprocal form, that is, j function of i. It
is a convention that is to be adopted once and for all, and the one that we have chosen is
compatible with most standardized turbo codes.
where:
f (u, v) = min {|u − v| , k − |u − v|} (7.6)

Function f is introduced to take into account the circular nature of the
addresses. Finally, we call Smin the smallest of the values of S(j1 , j2 ), for all
the possible pairs j1 and j2 :
Smin = min {S(j1 , j2 )} (7.7)

j1 ,j2
It is proved in [7.19] that an upper bound of Smin is:

√
sup Smin = 2k (7.8)
This upper bound is only reached in the case of a regular permutation and
with conditions:
√
P = P0 = 2k (7.9)
and:
P0
k= mod P0 (7.10)
2
Let us now consider a sequence of any weight that, after permutation, can be
written:

k−1
˜
d(D) = aj D j (7.11)
j=0
where aj can take the binary value 0 (no error) or 1 (one error) and, before
permutation:

k−1
k−1
d(D) = ai D i = aΠ(j) DΠ(j) (7.12)
i=0 j=0
We denote jmin and jmax the j indices corresponding to the first and last non-
˜
null values aj in d(D). Similarly, we define imin and imax for sequence d(D).
Then, the regular permutation satisfying (7.9) and (7.10) guarantees the prop-
erty:
√
(jmax − jmin ) + (imax − imin ) > 2k (7.13)
˜
This is because d(D) and d(D), both considered between min and max indices,
contain at least 2 bits whose
√ accumulated spatial distance, as defined by (7.5),
is maximum and equal to 2k. We must now consider two cases:
˜
• sequences d(D) and d(D) are both of the simple RTZ type, that is, they
begin in state 0 of the encoder and return to it once, at the end. The
parity bits produced by these sequences are statistically 1s, every other
time. Taking into account (7.13), for common values of k (k > 100), the
redundancy weights are high and these RTZ sequences do not contribute
to the MHD of the turbo code.
˜
• at least one of sequences d(D) and d(D) is of the multiple RTZ type, that
is, it corresponds to the encoder passing several times through state 0.
If these passes through state 0 are long, the parity associated with the
sequence may have reduced weight and the associated distance may be
low. Generally, in this type of situation, the sequences before and after
permutation are both multiple RTZ.
The performance of a turbo code, at low error rates, is closely linked with
the presence of multiple RTZ patterns and regular permutation is not a good
solution for eliminating these patterns.
Figure 7.6 – Possible error patterns of weight 2, 3, 6 or 9 with a turbo code whose
elementary encoders have a period 7 and with regular permutation.
Necessity for disorder

Again assuming that the error patterns that are not RTZ have weights high
enough not to have any effect on performance, an ideal permutation for a turbo
code could be defined by the following rule:
If a sequence is of the RTZ type before permutation, then it is no

longer so after permutation and vice-versa.
The rule above is impossible to satisfy in practice and a more realistic ob-
jective is:
If a sequence is of the RTZ type before permutation, then it is
no longer so after permutation or it has become a simple long RTZ
sequence and vice-versa.
The dilemma in designing good permutations for turbo codes lies in the
need to satisfy this objective for two distinct classes of input sequences that
require opposing types of processing: simple RTZ sequences and multiple RTZ
sequences, as defined above. To illustrate this problem, consider a rate 1/3
turbo code, with regular rectangular permutation (writing in M rows, lecture
in N columns) over blocks of k = M N bits (Figure 7.6). Elementary encoders
are encoders with 8 states whose period is 7 (recursivity generator 15).
The first pattern (a) of Figure 7.6 concerns a sequence of possible errors
with input weight w = 2 : 10000001 for code C1 , that we can also call the
horizontal code. This is the RTZ minimum sequence with weight 2 for the
encoder considered. The redundancy produced by this encoder is of weight 6
(exactly: 11001111). The redundancy produced by the vertical encoder C2 , for
which the sequence considered is also RTZ (its length is a multiple of 7), is much
more informative because it is simple RTZ and produced over seven columns.
Assuming that Y2 is equal to 1 every other time on average, the weight of this
redundancy is around w(Y2 ) ≈ 7M . When we make k tend towards infinity via
2 √
the values of M and N (M ≈ N ≈ k), the redundancy produced by one of
the two codes, for this type of pattern, also tends towards infinity. We then say
that the code is good.
The second pattern (b) is that of the minimum RTZ sequence of input weight
3. Here again, the redundancy is poor on the first dimension and much more
informative on the second. The conclusions are the same as above.
The other two diagrams in (c) present examples of multiple RTZ sequences,
made up of short RTZ sequences on each of the two dimensions. The input
weights are 6 and 9. The distances associated with these patterns (respectively
30 and 27 for this rate 1/3 code) are not generally sufficient to ensure good
performance at low error rates. Moreover, these distances are independent of
block size and therefore, in relation to the patterns considered, the code is not
good.
Regular permutation is therefore a good permutation for the class of simple
RTZ error patterns. For multiple RTZ patterns, however, regular permutation is
not appropriate. A good permutation must "break" the regularity of rectangular
composite patterns like those of Figure 7.6(c), by introducing some disorder.
But this must not be done to the detriment of the patterns for which regular
permutation is good. The disorder must therefore be managed well! Therein
lies the whole problem when looking for a permutation that must lead to a high
enough minimum distance. A good permutation cannot be found independently

of the properties of elementary codes, of their RTZ patterns, their periodicities,
etc.
Intra-symbol disorder
When the elementary codes are m-binary codes, we can introduce a certain
disorder into the permutation of a turbo code without however removing its
regular nature! To do this, in addition to intersymbol classical permutation, we
implement intra-symbol permutation, that is, a non-regular modification of the
content of the symbols of m bits, before coding by the second code [7.11]. We
briefly develop this idea with the example of double-binary turbo codes (m = 2).
Figure 7.7 – Possible error patterns with binary (a) and double-binary (b) turbo codes
and regular permutation.
Figure 7.7(a) presents the minimum pattern of errors with weight w = 4,

again using the code of Figure 7.6. It is a square pattern whose side is equal
to the period of the pseudo-random generator with polynomial 15, that is, 7.
It has already been mentioned that some disorder had to be introduced into
the permutation to "break" this kind of error pattern but without altering the
properties of the regular permutation in relation to patterns with weight 2 and
3, which is not easy. If, as an elementary encoder, we replace the binary encoder
by a double-binary encoder, the error patterns to consider are no longer made up
of bits but of couples of bits. Figure 7.7(b) gives an example of a double-binary
encoder and of possible error patterns, when the permutation is regular. The
couples are numbered from 0 to 3, according to the following correspondence:
(0, 0) : 0; (0, 1) : 1; (1, 0) : 2; (1, 1) : 3

The periodicities of the double-binary encoder are resumed in the diagram

of Figure 7.8. There we can find all the combinations of pairs of couples of
the RTZ type. For example, if the encoder, initialized in state 0, is fed by the
successive couples 1 and 3, it immediately returns to state 0. It is the same for
the sequences 201 or 2003 or 3000001, for example.
Figure 7.8 – Periodicities of the double-binary encoder of Figure 7.7(b). The four
input couples (0, 0), (0, 1), (1, 0) and (1, 1) are denoted 0, 1, 2 and 3, respectively. This
diagram gives all the combinations of pairs of couples of the RTZ type.
Figure 7.7(b) gives two examples of rectangular, minimum size error patterns.
First note that the perimeter of these patterns is larger than half the perimeter
of the square of Figure 7.7(a). Now, for a same coding rate, the redundancy of
a double-binary code is twice as dense as that of a binary code. We thus deduce
that the distances of the double-binary error patterns will naturally be larger,
everything else being equal, than those of binary error patterns. Moreover, there
is a simple way to eliminate these elementary patterns.
Figure 7.9 – The couples of the grey boxes are inverted before the second (vertical)
encoding. 1 becomes 2, 2 becomes 1; 0 and 3 remain unchanged. The patterns of
Figure 7.7(b), redrawn in (a), are no longer possible error patterns. But those of (b)
are, with distances 24 and 26 for coding rate 1/2.
Assume, for example, that the couples are inverted (1 becomes 2 and vice-
versa), every other time, before being applied to the vertical encoder. Then the
error patterns presented in Figure 7.9(a) no longer exist; for example, although
30002 does represent an RTZ sequence for the encoder considered, 30001 no
longer does. Thus, many of the error patterns, in particular the smallest, disap-
pear thanks to the disorder introduced inside the symbols. Figure 7.9(b) gives
two examples of patterns that the periodic inversion does not modify. The cor-
responding distances are high enough (24 and 26 for a rate 1/2) not to pose
a problem for small or average block sizes. For long blocks (several thousand
bits), additional intersymbol disorder, of low intensity, can be added to the
intra-symbol non-uniformity, to obtain even higher minimum distances.
Figure 7.10 – Permutation of the DRP type. This is a circular regular permutation to
which local permutations before writing and after reading are added.
Irregular permutations
In this section, we will not describe all the irregular permutations that have
been imagined so far, and that have been the subject of numerous publications
and several book chapters (see [7.40, 7.34] for example). We prefer to present
what seems, for the moment, to be both the simplest and the most efficient type
of permutation. These are almost regular circular permutations, called almost
regular permutation (ARP)[7.17] or dithered relatively prime (DRP ) [7.21]
permutations, depending on their authors. In all cases, the idea is not to stray
too far away from the regular permutation, which is well adapted to simple RTZ
error patterns and to instil some small, controlled disorder to counter multiple
RTZ error patterns.
Figure 7.10 gives an example, taken from [7.21], of what this small disorder
can be. Before the circular regular permutation is performed, the bits undergo
local permutation. This permutation is performed in groups of CW bits. CW ,
which is the writing cycle disorder, is a divisor of length k of the message. Sim-
ilarly, a local CR reading cycle permutation is applied before the final reading.
In practice, CW and CR can be identical values CW = CR = C, typically

4 or 8. This way of introducing disorder, in small local fluctuations, does not
significantly
√ decrease the accumulated spatial distance, whose maximum value
is 2k. However, it enables us to suppress the error patterns comparable to
those of Figures 7.6(b) and 7.7(c) on condition that the heights and widths of
these patterns are not both multiples of C.
Another way to perturb the regular permutation in a controlled way is shown
in Figure 7.11. The permutation is represented here in rectangular form, visually
more accessible, but it can also be very well applied to circular permutation.
One piece of information (bit or symbol) is placed where each row and column
cross. With regular permutation, these data are therefore memorized row by
row and read column by column. In Figure 7.11, the disorder is introduced by
means of four displacement vectors V1 , · · · , V4 that are applied alternately during
reading. These vectors have a small amplitude compared to the dimensions of
the permutation matrix.
Figure 7.11 – Permutation of the ARP type, following [7.17].
The mathematical model associated with this almost regular permutation,

in its circular form, is an extension of (7.4):
i = Π(j) ≡ P j + Q(j) + i0 mod k (7.14)
If we choose
Q(j) = A(j)P + B(j) (7.15)
where the positive integers A(j) and B(j) are periodic, with cycle C (divisor
of k), then these values correspond to the positive shifts applied respectively
before and after regular permutation. That is the difference between the per-
mutation shown in Figure 7.10, in which the writing and reading perturbations
are performed inside small groups of data and not by shifts.
For the permutation to really be a bijection, parameters A(j) and B(j) are
not just any parameters. To ensure the existence of the permutation, there is
one sufficient condition: all the parameters have to be multiples of C. This
condition is not very restricting in relation to the efficiency of the permutation.
(7.15) can then be rewritten in the form:
Q(j) = C(α(j)P + β(j)) (7.16)
where α(j) and β(j) are more often than not small integers, with values 0 to 8.
In addition, since the properties of a circular permutation are not modified by
a simple rotation, one of the Q(j) values can systematically be 0.
Two typical sets of Q values, with cycle 4 and α = 0 or 1, are given below:
if j = 0 mod 4, then Q=0

if j = 1 mod 4, then Q = 4P + 4β1
(7.17)
if j = 2 mod 4, then Q = 4β2
if j = 3 mod 4, then Q = 4P + 4β3
if j =0 mod 4, then Q=0
if j =1 mod 4, then Q = 4β1
(7.18)
if j =2 mod 4, then Q = 4P + 4β2
if j =3 mod 4, then Q = 4P + 4β3
These models require the knowledge of only four parameters (P , β1 , β2 and
β3 ), which can be determined using the procedure described in [7.17]. The
utilization of m-binary codes (see Section 7.5), instead of binary codes, sim-
ply requires k to be replaced by k/m in Equation (7.14). In particular, the
permutations defined for double-binary turbo codes (m = 2) of the DVB-RCS,
DVB-RCT and WiMax standards are inspired by Equations (7.14) and (7.21).
Quadratic Permutation
Recently, Sun and Takeshita [7.41] proposed a new class of deterministic in-
terleavers based on permutation polynomials (PP) over integer rings. The use
of PP reduces the design of interleavers to simply a selection of polynomial
coefficients. Furthermore, PP-based turbo codes have been shown to have a)
good distance properties [7.38] which are desirable for lowering the error floor
and b) a maximum contention-free property [7.43] which is desirable for parallel
processing to allow high-speed hardware implementation of iterative turbo
decoders.
A. Permutation Polynomials
Before addressing the quadratic PP, we will define the general form of a
polynomial and discuss how to verify whether a polynomial is a PP over the
ring of integers modulo N , ZN . Given an integer N ≥ 2, a polynomial
f (x) = a0 + a1 + a2 x2 + . . . + am xm modulo N (7.19)
where the coefficients a0 , a1 , a2 , . . . , am and m are non-negative integers, is said
to be permutation polynomial over ZN when f (x) permutes {0, 1, 2, . . . , N − 1}
[7.41]. Since we have modulo N operation, it is sufficient for the coefficients
a0 , a1 , a2 , . . . , am to be in ZN . Let us recall that the formal derivative of the
polynomial f (x) is given by

f (x) = a1 + 2a2 x + 3a3 x2 + . . . + mam xm−1 modulo N (7.20)
To verify whether a polynomial is a PP over ZN , let us discuss the following
three cases a) the case N = 2n , where n is an element of the positive integers
Z+ , b) the case N = pn where p is any prime number, and c) the case where N
is an arbitrary element of Z+ .
1. Case I (N = 2n ): a theorem in [7.36] states that f (x) is a PP over the
integer ring Zn2 if and only if 1) a1 is odd, 2) a2 + a4 + a6 + . . . is even,
and 3) a3 + a5 + a7 + . . . is even.
Example 1 for N = 23 = 8 : f (x) = 1 + 5x + x2 + x3 + x5 + 3x6 is

a PP over ZN =8 because it maps the sequence {0, 1, 2, 3, 4, 5, 6, 7} to
{1, 4, 7, 2, 5, 0, 3, 6}. Note that a1 = 5 is odd, a2 + a4 + a6 = 1 + 0 + 3 = 4
is even, and a3 + a5 = 1 + 1 = 2 is even.
Example 2 for N = 23 = 8 : f (x) = 1 + 4x + x2 + x3 + x5 + 3x6 is

not PP over ZN =8 because it maps the sequence {0, 1, 2, 3, 4, 5, 6, 7} to
{1, 3, 5, 7, 1, 3, 5, 7}. Note that a1 = 4 is even.
2. Case II (N = pn ): a theorem in [7.33] guarantees that f (x) is a PP

modulo pn if and only if f (x) is a PP modulo p and f (x) = 0 modulo p,
for every integer x ∈ Znp . Note that the Case I is simply a special case of
the Case II because p = 2 is a prime number.
Example 1 for N = 3n (p = 3) : f (x) = 1+2x+3x2 is a PP over Zn3 because

f (0) = 1 modulo 3 = 1, f (1) = 6 modulo 3 = 0, f (2) = 17 modulo 3 = 2,

and f (x) = 2 + 6x = 2 modulo 3 = 2 is a non-zero constant for all xinZn3 .
Example 2 for N = 3n (p = 3) : f (x) = 1 + 6x+ 3x2 is not a PP Zn3 because

f (x) = 6+6x = 0 modulo 3 = 0 for all xinZn3 . For instance, for N = 32 =
9, f (x) maps the sequence {0, 1, 2, 3, 4, 5, 6, 7, 8} to {1, 1, 7, 1, 1, 7, 1, 1, 7}.
3. Case III (arbitrary N ): let P = {2, 3, 5, 7, . . .} be the

' set of prime numbers.
Then, every N ∈ Z+ can be factored as N = p∈P pnN,p , where all p
values are distinct prime numbers, nN,p ≥ 1 for a certain number of p
and nN,p = 0 otherwise. For example, if N = 2500 = 22 × 54 , then
we have n2500,2' = 2 and n2500,5 = 4. A theorem in [7.41] states that
for any N = p∈P pnN,p , f (x) is a PP modulo N if and only if f (x)
is also a PP modulo pnN,p ,∀p such that nN,p ≥ 1. With this theorem,
verifying whether a polynomial is a PP modulo N reduces to verifying the
polynomial modulo each pnN,p factor of N . For p = 2, we use the theorem
reported in Case I, which is a simple test on the polynomial coefficients.
For p = 2, we must use the theorem reported in Case II, which cannot be
done by simply testing the polynomial coefficients. For an arbitrary N , it is
difficult to develop a simple coefficient test to check whether an arbitrary
m-degree polynomial f (x) is a PP modulo N . However, for quadratic
polynomial (m = 2), f (x) = a0 + a1 x + a2 x2 , a simple coefficient test have
been proposed in [7.39]. Next section will address this coefficient test in
details.
B. Quadratic Permutation Polynomials
Since the constant q0 in the quadratic polynomial q(x) = q0 + q1 x + q2 x2

only causes a "cyclic shif" to the permuted values, we define in this section
-without loosing generality- quadratic polynomials as q(x) = q1 x + q2 x2 . Let
us first establish some abbreviations borrowed from [7.43], that we will use
throughout this section. We express the fact that b is divisible by a, or a
is divisor of b, by a |b . We also use anot |b to express the contrary of a |b .
The greatest common divisor of a and b is denoted by gcd(a, b). Remember
that gcd(a, b) = 1 indicates that a and b are relatively prime. As we will
see in Proposition 1 below, we are mainly interested in the factorization of
the coefficient
' q2 , which can be written according to the previous notation as
q2 = p∈P pnq2 ,p . The following Proposition 1 provides a necessary and suf-
ficient condition for verifying whether a quadratic polynomial is a PP modulo N.
' nN,p
Proposition 1: Let N = p∈P p . For a quadratic polynomial
2
q(x) = q1 x + q2 x modulo N to be a PP, the following necessary and suffi-
cient conditions must be satisfied [7.39]
1)'Either 2not |N or 4 |N (i.e., nN,2 = 1) gcd(q1 , N ) = 1 and

q2 = p∈P pnq2 ,p , nq2 ,p ≥ 1, ∀p such that nN,p ≥ 1.
2)'2 |N or 4not |N (i.e., nN,2 = 1), q1 + q2 is odd, gcd(q1 , N2 ) = 1 and

q2 = p∈P pnq2 ,p , nq2 ,p ≥ 1, ∀p such that p =
2 and nN,p ≥ 1.
'
The statement q2 = p∈P pnq2 ,p , nq2 ,p ≥ 1, ∀p such that nN,p ≥ 1 can be
expressed in simple words as follows: each p factor of N must also be a factor
of q2 . It is important to note that this statement still allows q2 to have prime
factors that differ from all p factors of N .
Example 1: if N = 36, then we have case 1) of Proposition 1 because

4 |N . All possible values of q1 are simply the set of numbers that are relatively
prime to N . Consequently, q1 = {1, 5, 7, 11, 13, 17, 19, 23, 25, 29, 31, 35}. Since
36 = 22 ×32 (p1 = 2 and p2 = 3), then 2 and 3 must be a factor of q2 . That
is, q2 = (2 × 3), (22 × 3), (23 × 3), (2 × 32 ), 5 × (2 × 3) = {6, 12, 24, 18, 30}.
As mentioned above, the use of the prime number 5 in 5 × (2 × 3) does not
violate the condition in case 1) of Proposition 1. In total, for N = 36 there are
12 × 5 = 60 possible quadratic PPs (QPP).
'
The statement q2 = p∈P pnq2 ,p , nq2 ,p ≥ 1, ∀p such that p = 2 and nN,p ≥ 1
of case 2) can also be expressed in simple words as follows: each p6 = 2 factor
of N must also be a factor of q2 . It is important to note that this statement
still allows q2 to have the prime factor 2 and all prime factors that differ from
all p factors of N (q2 may or may not have 2 as a factor).
Example 2: if N = 90, then we have case 2) of Proposition 1 because 2 |N

and 4not |N . Since N = 90 = 2 × 32 × 5, all p values that differ from 2 are
p1 = 3 and p2 = 5. Therefore, the potential values for q2 are

(3 × 5), (32 × 5), (3 × 52 ), 2 × (3 × 5), 22 × (3 × 5) = {15, 45, 75, 30, 60}
Under the condition gcd(q1 , N/2 = 45) = 1, the potential values for q1 are have
120 possible QPPs.
{1, 2, 4, 7, 8, 11, 13, 14, 16, 17, 19, 22, 23, 26, 28, 29, 31, 32, 34, 37, 38, 41, 43, 44, 46,
47, 49, 52, 53, 56, 58, 59, 61, 62, 64, 67, 68, 71, 73, 74, 76, 77, 79, 82, 83, 86, 88, 89}
Despite the tight conditions imposed by Proposition 1 on q1 and q2 , the

search space of QPPs is still large, especially for medium to large interleavers.
Thus, it is desirable to reduce the search space further. A solution is to consider
only QPPs that do have a quadratic inverse (for more details on quadratic
inverse for QPP, see [7.39]). This solution for reducing the search space is based
on the interesting finding reported by Rosnes and Takeshita [7.38], namely,
for 32 ≤ N ≤ 512 and N = 1024, the class of QPP-based interleavers with
quadratic inverses are strictly superior (in term of minimum distance) to the
class of QPP-based interleavers with no quadratic inverse. Using exhaustive
computer search, Rosnes and Takeshita provided, for turbo codes that use 8 and
16-state constituent codes, a very useful list for the best (in term of minimum
distance) QPPs for a wide range of N (32 ≤ N ≤ 512 and N = 1024) [7.38].
After discussing a necessary and sufficient condition for verifying whether a
quadratic polynomial is a PP modulo N, and providing some examples, let us
discuss some properties of QPPs. It is well known that a linear polynomial,
l(x) = l0 + l1 x(or simply l(x) = l1 x), is guaranteed to be a PP modulo N if
l1 is chosen relatively prime to N (i.e., gcd(l1 , N ) = 1). Consequently, linear
permutation polynomials (LPP) always exist for any N , but unfortunately this
is not true for QPPs. For example, there are no QPP for N = 11 and for
2 ≤ N ≤ 4096 there are only 1190 values of N that have QPPs (roughly 29%)
[7.44]. A theorem in [7.44] guarantees the existence of QPP for all N = 8i,
i ∈ Z+ (i.e., multiples of a typical computer byte size of 8). It is shown in [7.44]
that some QPP degenerate to LPP (i.e., there exists an LPP that generates the
same permutation over the ring ZN ). A QPP is called reducible if it degenerates
to an LPP; otherwise it is called irreducible. For instance, example 1 in Case I of
sub-section A could be simply reduced to f (x) = l + 3x modulo 8 to obtain the
same permutation. In [7.38], it is shown that some reducible QPPs can achieve
better minimum distances than irreducible QPP for some short to medium
interleavers. However, for large intereavers, the class of irreducible QPPs are
better (in term of minimum distance) than the class of LPP; and if not, that
particular length will not have any good minimum distance [7.38].
7.4 Decoding turbo codes

7.4.1 Turbo decoding
Decoding a binary turbo code is based on the schematic diagram of Figure 7.12.
The loop allows each decoder to take advantage of all the information avail-
able. The values considered at each node of the layout are LLRs, the decoding
operations being performed in the logarithmic domain.
The LLR at the output of a decoder of systematic codes can be seen as
the sum of two terms: the intrinsic information, coming from the transmission
channel, and the extrinsic information, which this decoder adds to the former
to perform its correction operation. As the intrinsic information is used by the
two decoders (at different instants), it is the extrinsic information produced by
each of the decoders that must be transmitted to the other as new information,
to ensure joint convergence. Section 7.4.2 details the operations performed to
calculate the extrinsic information, by implementing the MAP algorithm or its
simplified Max-Log-MAP version.
Because of latency effects, the exchange of extrinsic information, in a digital
processing circuit, must be implemented via an iterative process: first decoding
Figure 7.12 – 8-state turbo encoder and schematic structure of the corresponding
turbo decoder. The two elementary decoders exchange probabilistic information, called
extrinsic information (z)
.
by DEC1 and putting extrinsic information z1 in the memory, second decoding

by DEC2 and putting extrinsic information z2 in the memory (end of the first
iteration), again using DEC1 and putting z1 in the memory, etc. Different
hardware architectures, with more or less great degrees of parallelism, can be
envisaged to accelerate the iterative decoding.
If we wanted to decode the turbo code using a single decoder, which would
take into account all the possible states of the encoder, for each element of the
message decoded, we would obtain one and only one probability of having a
binary value equal to 0 or to 1. As for the composite structure of Figure 7.12,
it uses two decoders working jointly. By analogy with the result that the single
decoder would provide, they therefore need to converge towards the same deci-
sions, with the same probabilities, for each of the data considered. That is the
fundamental principle of "turbo" processing, which justifies the structure of the
decoder, as the following reasoning shows.
The role of a SISO decoder (see Section 7.4.2), is to process the LLRs at its
input to try to make them more reliable, thanks to local redundancy (that is,
y1 for DEC1, y2 for DEC2). The LLR produced by a decoder of binary codes,
relative to data d, can be written simply as
LLRoutput (d) = LLRinput(d)+z(d) (7.21)
where z(d) is the extrinsic information specific to d. The LLR is improved when
z is negative and d is a 0, or when z is positive and d is a 1.
After p iterations, the output of DEC1 is:
LLRpoutput,1 (d) = (x + z2p−1 (d)) + z1p(d)
and the output of DEC2 is
LLRpoutput,2 (d) = (x + z1p−1 (d)) + z2p(d)
If the iterative process converges towards a stable solution, z1p(d) − z1p−1 (d) and
z2p (d) − z2p−1 (d) tend towards zero when p tends towards infinity. Consequently,
the two LLRs relative to d become identical, thus satisfying the basic criterion of
common probability mentioned above. As for proof of the convergence, it is still
being studied further and on this topic we can, for example, consult [7.49, 7.24].
Apart from the permutation and inverse permutation functions, Figure 7.13
details the operations performed during turbo decoding:
Figure 7.13 – Operations shown (clipping, quantization, attenuation of the extrinsic

information) in the turbo decoder of Figure 7.12.
1. Analogue to digital (A/D) conversion transforms the data coming from

the demodulator into samples exploitable by the digital decoder. Two
parameters are involved in this operation: nq , the number of quantization
bits, and Q, the scale factor, that is, the ratio between the average absolute
value of the quantized signal and its maximum absolute value. nq is fixed
to a compromise value between the precision required and the complexity
of the decoder. With nq = 4, the performance of the decoder is very
close to what we obtain with real samples. The value of Q depends on
the modulation, on the coding rate and on the type of channel. It is, for
example, larger for a Rayleigh channel than for a Gaussian channel.
2. The role of SISO decoding is to increase the equivalent signal to noise ratio
of the LLR, that is, to provide more reliable extrinsic information at output
zoutput than at input (zinput ). The convergence of the iterative process
(see Section 7.6) will depend on the transfer function SNR(zoutput ) =
G(SNR(zinput )) of each of the decoders.
When data is not available at the input of the SISO decoder, due to punc-
turing for example, a neutral value (analogue zero) is substituted for this
missing data.
3. When the elementary decoding algorithm is not the optimal MAP algo-
rithm but a sub-optimal simplified version, the extrinsic information has
to undergo some transformations before being used by a decoder:
• multiplying the extrinsic information by factor γ, lower than 1, guar-

antees the stability of the looped structure. γ can vary over the itera-
tions, for example, from 0.7 at the beginning of the iterative process,
to 1 for the last iteration.
• clipping the extrinsic information solves both the issue of limiting the
size of the memories and that of participating in the stability of the
process. A typical value of the maximum dynamics of the extrinsic
information is twice the input dynamics of the decoder.
4. Binary decision taking is performed by thresholding at value 0.

The number of iterations required by turbo decoding depends on the size
of the block and on the coding rate. Generally, the larger the decoded
block and the slower the convergence, the higher the MHD of the code.
The same occurs when the coding rates are low. In practice, we limit the
number of iterations to a value between 4 and 10, according to the speed,
latency and consumption constraints imposed by the application.
Figure 7.14 gives an example of the performance of a binary turbo code,

taken from the UMTS standard [7.3]. We can observe a decrease in packet error
rates (PER), just close to the theoretical limit (that is, around 0,5 dB, taking
into account the size of the block), but also a fairly pronounced change in slope,
due to an MHD that is not very high (dmin = 26) for a rate of 1/3.
7.4.2 SISO decoding and extrinsic information

Here we present processing performed in practice in a SISO decoder using the
MAP algorithm [7.6] or its simplified version, the Max-log-MAP algorithm, also
called the SubMAP algorithm [7.37], to decode RSC m-binary codes and imple-
ment iterative decoding. For binary codes and turbo codes, all these equations
can be simplified by taking m = 1.
Figure 7.14 – Performance in packet error rates (PER) of the UMTS standard turbo
code for k = 640 and R = 1/3 on a Gaussian channel with 4-PSK modulation. Decod-
ing using the Max-Log-MAP algorithm with 6 iterations.
Notations
A sequence of data d is defined by d ≡ dk−1 0 = (d0 · · · di · · · dk−1 ), where di
is the vector of m-binary data applied at the input of the encoder at instant i:
di = (di,1 · · · di,l · · · di,m ). The value of di can also be represented by the scalar
m
integer value j = 2l−1 di,l , ranging between 0 and 2m − 1 and we can then
l=1
write di ≡ j.
In the case of two or four-phase PSK modulation (2-PSK, 4-PSK), the en-
coded modulated sequence u ≡ uk−1 0 = (u0 · · · ui · · · uk−1 ) is made up of vec-
tors ui of size m + m : ui = (ui,1 · · · ui,l · · · ui,m+m ), where ui,l = ±1 for l =
1 · · · m + m and m is the number of redundancy bits added to the m bits of
information. The symbol ui,l is therefore representative of a systematic bit for
l ≤ m and of a redundancy bit for l > m.
The sequence observed at the output of the demodulator is denoted v ≡
v0k−1 = (v0 · · · vi · · · vk−1 ), with vi = (vi,1 · · · vi,l · · · vi,m+m ). The series of
the states of the encoder between instants 0 and k is denoted S = Sk0 =
(S0 · · · Si · · · Sk ). The following is based on the results presented in the chapter
on convolutional codes.
Decoding following the Maximum A Posteriori (MAP) criterion

At each instant i, the weighted (probabilistic) estimates provided by the
MAP decoder are the 2m a posteriori probabilities (APP) Pr(di ≡ j | v ),
j = 0 · · · 2m − 1. The corresponding hard decision, d̂i , is the binary repre-
sentation of value j that maximizes the APP.
Each APP can be expressed as a function of the joint likelihoods p(di ≡ j, v):
p(di ≡ j, v) p(di ≡ j, v)
Pr(di ≡ j | v ) = = 2m −1 (7.22)
p(v)
p(di ≡ l, v)
l=0
In practice, we calculate the joint likelihoods p(di ≡ j, v) for j = 0 · · · 2m − 1

then each APP is obtained by normalization.
The trellis representative of a code with memory ν has 2ν states, taking
their scalar value s in (0, 2ν − 1). The joint likelihoods are calculated from
the recurrent forward αi (s) and backward probabilities βi (s) and the branch
likelihoods gi (s , s):

p(di ≡ j, v) = βi+1 (s) αi (s ) gi (s , s) (7.23)
(s ,s)/di (s ,s)≡j
where (s , s)/di (s , s) ≡ j denotes the set of transitions from state to state s → s
associated with the m-binary j. This set is, of course, always the same in a trellis
that is invariant over time.
The value gi (s , s) is expressed as:
gi (s , s) = Pra (di ≡ j, di (s , s) ≡ j).p(vi | ui ) (7.24)
where ui is the set of systematic and redundant information symbols associated

with transition s → s of the trellis at instant i and Pra (di ≡ j, di (s , s) ≡ j)
is the a priori probability of transmitting the m-tuple of information and that
this would correspond to transition s → s at instant i. If transition s → s does
not exist in the trellis for di ≡ j , then Pra (di ≡ j, di (s , s) ≡ j) = 0, otherwise
the transition is given by the source statistics (usually uniform, in practice).
In the case of a Gaussian channel with binary inputs, value p(vi | ui ) can
be written:
5
m+m
1 (vi,l − ui,l )2
p(vi | ui ) = √ exp − (7.25)
σ 2π 2σ2
l=1
where σ2 is the variance of the additive white Gaussian noise. In practice, we

keep only the terms that are specific to the transition considered and that are
not eliminated by division in the expression (7.22):

⎛
⎞

m+m
⎜ vi,l · u i,l ⎟
⎜ ⎟
p (vi | ui ) = exp ⎜ l=1 2 ⎟ (7.26)
⎝ σ ⎠
The forward and backward recurrent probabilities are calculated as follows:

ν
2 −1
αi (s) = αi−1 (s ) gi−1 (s , s) for i = 1 · · · k (7.27)
s =0
and: ν
2 −1
βi (s) = βi+1 (s ) gi (s, s ) for i = k − 1 · · · 0 (7.28)
s =0
To avoid problems of precision or of overflow in the representation of these

values, they have to be normalized regularly. The initialization of the recursions
depends on the knowledge or not of the state of the encoder at the beginning
and at the end of encoding. If the initial state S0 of the encoder is known,
then α0 (S0 ) = 1 and α0 (s) = 0 for any other state, otherwise all the α0 (s) are
initialized to the same value. The same rule is applied for the final state Sk . For
circular codes, initialization is performed automatically after the prologue step,
which starts from identical values for all the states of the trellis.
In the context of iterative decoding, the composite decoder uses two ele-
mentary decoders exchanging extrinsic probabilities. Consequently, the basic
decoding brick described above must be reconsidered in order to:

1. take into account an extrinsic probability, Prex (di ≡ j v ), in expres-
sion (7.24), calculated by the other elementary decoder of the composite

decoder, from its own input sequence v ,

2. produce its own extrinsic probability Prex (di ≡ j v ) that will be used
by the other elementary decoder.
In practice, for each value of j, j = 0 · · · 2m − 1:
1. in expression (7.24), the a priori probability Pra (di ≡ j, di (s , s) ≡ j) is

replaced by the modified a priori probability Pr@ (di ≡ j, di (s , s) ≡ j),
having for its expression, to within one normalization factor:
Pr@ (di ≡ j, di (s , s) ≡ j) = Pra (di ≡ j, di (s , s) ≡ j).Prex (di ≡ j | v ) (7.29)

1. Prex (di ≡ j | v ) is given by:

βi+1 (s) αi (s ) gi∗ (s , s)
(s ,s)/di (s ,s)≡j
Prex (di ≡ j | v ) = (7.30)
βi+1 (s) αi (s ) gi∗ (s , s)
(s ,s)
The terms gi∗ (s , s) are non-zero if s → s corresponds to a transition of the

trellis and are then inferred from the expression of p(vi | ui ) by eliminating the
systematic part of the information. In the case of a transmission over a Gaussian
channel with binary inputs and starting from the simplified expression (7.26) of
p (vi | ui ) , we have:
⎛ m+m ⎞

⎜ vi,l u i,l ⎟
⎜ l=m+1 ⎟
gi∗ (s , s) = exp ⎜ ⎟ (7.31)
⎝ σ2 ⎠
The simplified Max-Log-MAP or SubMAP algorithm

Decoding following the MAP criterion requires a large number of operations,
including calculating exponentials and multiplications. Re-writing the decoding
algorithm in the logarithmic domain simplifies the processing. The weighted es-
timations provided by the decoder are then values proportional to the logarithms
of the APPs, called Log-APP logarithms, denoted L:
σ2
Li (j) = − ln Pr(di ≡ j | v) , j = 0 · · · 2m − 1 (7.32)
2
We define Miα (s) and Miβ (s) the forward and backward metrics relative to node
s at instant i, and Mi (s , s), the branch metric relative to the s → s transition
of the trellis at instant i by:
Miα (s) = −σ 2 ln αi (s)

Miβ (s) = −σ 2 ln βi (s) (7.33)
Mi (s , s) = −σ 2 ln gi (s , s)
Introduce values Ai (j) and Bi calculated as:

⎡ ⎤

Ai (j) = −σ 2 ln ⎣ βi+1 (s)αi (s )gi (s , s)⎦ (7.34)
(s ,s)/di (s ,s)≡j
⎡ ⎤

Bi = −σ 2 ln ⎣ βi+1 (s)αi (s )gi (s , s)⎦ (7.35)
(s ,s)
Li (j) can then be written, by reference to (7.22) and (7.23), as follows:

1
Li (j) = (Ai (j) − Bi ) (7.36)
2
Expressions (7.34) and (7.35) can be simplified by applying the so-called Max-
Log approximation:
ln(exp(a) + exp(b)) ≈ max(a, b) (7.37)
For Ai (j) we get:

Ai (j) ≈ min β
Mi+1 (s) + Miα (s ) + Mi (s , s) (7.38)
(s ,s)/di (s ,s)≡j
and for Bi :

β α
Bi ≈ min M i+1 (s) + M i (s ) + M i (s , s) = min Ai (l) (7.39)
(s ,s) l=0···2m −1
and finally we get:

1
Li (j) = Ai (j) − minm Ai (l) (7.40)
2 l=0···2 −1
Note that these values are always positive or equal to zero.

Introduce the values La proportional to the logarithms of the a priori prob-
abilities Pra :
σ2
Lai (j) = − ln Pra (di ≡ j) (7.41)
2
Branch metrics Mi (s , s) can be written, according to (7.24) and (7.33):
Mi (s , s) = 2Lai (d(s , s)) − σ 2 ln p(vi | ui ) (7.42)
If the statistic of the a priori transmission of the m-tuples di is uniform,

term 2Lai (d(s , s)) can be omitted from the above relation since it is the same
value that is used in all the branch metrics. In the case of a transmission over
a Gaussian channel with binary inputs, we have according to (7.26):

m+m

Mi (s , s) = 2Lai (d(s , s)) − vi,l · ui,l (7.43)
l=1
The forward and backward metrics are then calculated from the following
recurrence relations:
⎛
⎞

m+m
Miα (s) = minν ⎝Mi−1 α
(s ) − vi−1,l · ui−1,l + 2Lai−1 (d(s , s))⎠ (7.44)
s =0···2 −1
l=1
⎛
⎞

m+m
Miβ (s) = min ⎝Mi+1
β
(s ) − vi,l · ui,l + 2Lai (d(s, s )⎠ (7.45)
s =0···2ν −1
l=1
Applying the Max-Log-MAP logarithm in fact amounts to performing two

Viterbi decodings, in the forward and backward directions. That is the rea-
son why it is also called the dual Viterbi algorithm.
If the initial state of the encoder, S0 , is known, then M0α (S0 ) = 0 and
M0 (s) = +∞ for any other state, otherwise all the M0α (s) are initialized to the
α
same value. The same rule is applied for the final state. For circular codes, all
the metrics are initialized to the same value at the beginning of the prologue.
Finally, taking into account (7.38) and replacing Mi (s , s) by its expression
(7.43), we obtain:
⎛
⎞

m+m
Ai (j) = min ⎝Mi+1
β
(s) + Miα (s ) − vi,l · ui,l ⎠ + 2Lai(j) (7.46)
(s ,s)/di (s ,s)≡j
l=1
The hard decision taken by the decoder is the value of j, j = 0 · · · 2m − 1, which

minimizes Ai (j). Let us denote this value j0 . According to (7.40), Li (j) can be
written:
1
Li (j) = [Ai (j) − Ai (j0 )] pour j = 0 · · · 2m − 1 (7.47)
2
We note that the presence of coefficient σ2 in definition (7.32) of Li (j) allows us
to ignore the knowledge of this parameter for computing the metrics and hence
for all the decoding. This is an important advantage of the Max-Log-MAP
method over the MAP method.
In the context of iterative decoding, term Lai (j) is modified in order to take
into account extrinsic information L∗i (j) coming from the other elementary de-
coder:
∗
L@ a
i (j) = Li (j) + Li (j) (7.48)
On the other hand, the extrinsic information produced by the decoder is obtained
by eliminating in Li (j) the terms containing the direct information about di ,
that is, the intrinsic and a priori information:
! "
#
∗ β α

m+m
Li (j) = 21
min Mi+1 (s) + Mi (s ) − vi,l · ui,l
(s ,s)/di (s ,s)≡j
" l=m+1

#$ (7.49)
β α

m+m
− min Mi+1 (s) + Mi (s ) − vi,l · ui,l
(s ,s)/di (s ,s)≡j0 l=m+1
The expression of Li (j) can then be formulated as follows:
1
m

Li (j) = L∗i (j) + vi,l · [ ui,l |di ≡j − ui,l |di ≡j0 ] + L@
i (j) − Li (j0 )
@
(7.50)
2
l=1
This expression shows that extrinsic information L∗i (j) can, in practice, be
deduced from Li (j) by simple subtraction. Factor 12 in definition (7.32) of Li (j)
allows us to obtain a weighted decision and extrinsic information L∗i (j) on the
same scale as the noisy samples vi,l .
7.4.3 Practical considerations

The simplest way to perform turbo decoding is totally sequential and uses the fol-
lowing operations, here founded on the Max-Log-MAP algorithm and repeated
as many times as necessary:
1. Backward recursion for code C2 (Figure 7.12), calculation and memoriza-

tion of metrics Miβ (s), i = k − 1, ..., 0 and s = 0, ..., 2ν − 1,
2. Forward recursion for code C2 , calculation of metrics Miα (s), i = 0, ...,

k − 1 and s = 0, ..., 2ν − 1. Calculation and memorization of the extrinsic
information,
3. Backward recursion for code C1 , calculation and memorization of metrics
Miβ (s), i = k − 1, ..., 0 and s = 0, ..., 2ν − 1,
4. Forward recursion for code C1 , calculation of metrics Miα (s), i = 0, ...,

k − 1 and s = 0, ..., 2ν − 1. Calculation and memorization of the extrinsic
information. Binary decisions (at the last iteration).
The first practical problem lies in the memory necessary to store metrics
Miβ (s). Processing the coded messages of k = 1000 bits, for example, with 8-
state decoders and quantization of the metrics on 6 bits, at first sight requires
a storage capacity of 48000 bits for each decoder. In sequential operation (al-
ternate processing of C1 and C2 ), this memory can, of course, be used by the
two decoders in turn. The technique used to greatly reduce this memory is
that of the sliding window. It involves (Figure 7.15) replacing all the backward
processing, from i = k − 1 to 0, by a succession of partial forward processings,
from i = iF to 0, then from i = 2iF to iF , from i = 3iF to 2iF etc., where iF
is an interval of some tens of trellis sections. Each partial backward processing
includes a "prologue" (dotted line), that is, a step without memorization whose
aim is to estimate as correctly as possible the accumulated backward metrics in
positions iF , 2iF , 3iF , etc. The parts shown by a solid line correspond to the
phases during which these metrics are memorized. The same memory can be
used for all the partial backward recursions. The forward recursion is performed
without any discontinuity.
The process greatly reduces the storage capacity necessary which, in addition,
becomes independent of the length of the messages. The drawback lies in the
necessity to perform the additional operations – the prologues – that can increase
the total calculation complexity by 10 to 20 %. However, these prologues can be
avoided after the first iteration if the estimates of the metrics at the boundary
indices are put into memory to be used as departure points for the calculations
of the following iteration.
Figure 7.15 – Operation of the forward and backward recursions when implementing
the MAP algorithm with a sliding window.
The second practical problem is that of the speed and latency of decoding.
The extent of the problem depends of course on the application and on the ra-
tio between the decoding circuit clock and the data rate. If the latter is very
high, the operations can be performed by a single machine, in the sequential
order presented above. In specialized processors of the DSP (digital signal pro-
cessor ) type, cabled co-processors may be available to accelerate the decoding.
In dedicated circuits of the ASIC (application-specific integrated circuit ) type,
acceleration of the decoding is obtained by using parallelism, that is, multiplying
the number of arithmetical operators, if possible without increasing the capacity
of the memories required to the same extent. Then, problems of access to these
memories are generally posed.
Note first that only knowledge of permutation i = Π(j) is necessary for
implementation of the iterative decoding and not that of inverse permutation
Π−1 , as could be wrongly assumed from the schematic diagrams of Figures 7.12
and 7.13. Consider, for example, two SISO decoders working in parallel to
decode the two elementary codes of the turbo code and based on two dual-
port memories for the extrinsic information (Figure 7.16). The DEC1 decoder
associated with the first code produces and receives the extrinsic information in
the natural order i. The DEC2 decoder associated with the second code works
according to index j but writes and recovers its data at addresses i = Π(j).
Knowledge of Π−1 , which could pose a problem depending on the permutation
model selected, is therefore not required.
Figure 7.16 – Implementing turbo decoding does not require explicit knowledge of Π−1
In addition, the two extrinsic information memories can be merged into a

single one, observing that extrinsic information that has just been read and
exploited by a decoder no longer has to be retained. It can thus be replaced
immediately afterwards by another datum, which can be the extrinsic informa-
tion output from the same decoder. Figure 7.17 illustrates this process, which
imposes a slight hypothesis: working indices i and j have the same parity and
permutation i = Π(j) inverses the parity. For example, with the permutation
defined by (7.4), this hypothesis is satisfied if the departure index i0 is odd and
the length of the message k is even.
Figure 7.17 – In practice, the storage of the extrinsic information uses only a single
memory.
The extrinsic information memory is divided in two pages corresponding to

the sub-sets of the even and odd addresses. Access to these two pages, in a dual-
port memory, is alternated regularly. In Figure 7.17(a), in the even cycles, DEC1
reads at an even address and writes the extrinsic information produced during
the previous cycle, via a buffer memory with unit delay, to an odd address.
Meanwhile, DEC2 reads at an even address and writes to an odd address. In
Figure 7.17(b), during the odd cycles, the accesses to the reading-writing pages
are exchanged.
To further increase the degree of parallelism in the iterative decoder, the
forward and backward recursion operations can also be tackled inside each of the
two decoders (DEC1 and DEC2). This can be easily implemented by considering
the diagram of Figure 7.15.
Finally, depending on the permutation model used, the number of elemen-
tary decoders can be increased beyond two. Consider for example the circular
permutation defined by (7.14) and (7.16), with cycle C = 4 and k a multiple
of 4.
The congruences of j and Π(j) modulo 4, are periodic. Parallelism with
degree 4 is then possible following the principle described in Figure 7.18 [7.17].
For each forward or backward recursion (these also can be done in parallel),
four processors are used. At the same instant, these processors process data
whose addresses have different congruences modulo 4. In the example in the
figure, the forward recursion is considered and we assume that k/4 is also a
multiple of 4. Then, we have first processor begin at address 0, the second at
address k/4 + 1, the third at address k/2 + 2 and finally the fourth at address
3k/4 + 3. At each instant, as the processors advance by one place each time, the
congruences modulo 4 of the addresses are always different. Addressing conflicts
are avoided via a router that directs the four processors towards four memory
pages corresponding to the four possible congruences. If k/4 is not a multiple
of 4, the departure addresses are no longer exactly 0, k/4 + 1, k/2 + 2, 3k/4 + 3
but the process is still applicable.
Figure 7.18 – The forward recursion circle is divided into 4 quadrants.

Whatever the value of cycle C, higher degrees of parallelism of value pC, can
be implemented. Indeed, any multiple of C, the basic cycle in the permutation,
is also a cycle in the permutation, on condition that pC is a divisor of k. That
is, j modulo pC and Π(j) modulo pC are periodic on the circle of length k,
which can then be cut into pC fractions of equal length. For example, a degree
64 parallelism is possible for a value of k equal to 2048.
However, whatever the degree of parallelism, a minimum latency is unavoid-
able: the time required for receiving a packet and putting it into the buffer
memory. While this packet is being put into memory, the decoder works on the
information contained in the previous packet. If this decoding is performed in
a time at least equal to the memorization time, then the total decoding latency
is at maximum twice this memorization time. The level of parallelism in the
decoder is adjusted according to this objective, which may be a constraint in
certain cases.
For further information about the implementation of turbo decoders, of all
the publications on this topic, [7.47] is a good resource.
7.5 m-binary turbo codes

m−binary turbo codes are built from recursive systematic convolutional (RSC)
codes with m binary inputs (m ≥ 2). There are at least two ways to build an
m−binary convolutional code: either from the Galois field F2m , or from the
Cartesian product (F2 )m . Here, we shall only deal with the latter, which is
more convenient. Indeed, a code elaborated in F2m , with a memory ν, has 2νm
possible states, whereas the number of states of the code defined in (F2 )m , with
the same memory, can be limited to 2ν .
The advantages of the m-binary construction compared to the classical (m =
1) turbo code scheme, are varied: better convergence of the iterative process,
larger minimum distances, less puncturing, lower latency, robustness towards the
sub-optimality of the decoding algorithm, in particular when the MAP algorithm
is simplified into its Max-Log-MAP version [7.23].
The case m = 2 has already been adopted in the European standards for
the return path in digital video broadcasting via the satellite network and in
the terrestrial network [7.2, 7.1] as well as in the IEEE 802.16 standard [7.5].
Combined with the circular trellis technique, these 8-state turbo codes, called
double-binary turbo codes, offer good average performance and great flexibility
in adapting to different block sizes and different rates, whilst retaining reasonable
decoding complexity.
7.5.1 m−binary RSC encoders

Figure 7.19 presents the general structure of an m−binary RSC encoder. It uses
a pseudo-random generator with code memory ν and generator matrix G (size
ν × ν). The input vector d with m components is connected to the different

possible nodes thanks to a grid of interconnections whose binary matrix, size
ν × m, is denoted C. The vector T applied to the ν possible taps of the register
at instant i, is given by:
Ti = C.di (7.51)
with di = (d1,i . . . dm,i ).
Figure 7.19 – General structure of an m-binary RSC encoder with code memory ν.
The time index is not shown.
If we wish to avoid parallel transitions in the trellis of the code, condition

m ≤ ν must be respected and matrix C must be full rank. Except for very
particular cases, this encoder is not equivalent to an encoder with a single input
on which we would successively present d1 , d2 , · · · , dm . An m−binary encoder
is therefore not decomposable generally.
The redundant output of the machine (not shown in the figure) is calculated
at instant i according to the expression:

yi = dj,i + RT Si (7.52)
j=1...m
where Si is the state vector at instant i and RT is the transposed redundancy

vector. The p-th component of R equals 1 if the p-th component of Si is used in
the construction of yi and 0 otherwise. We can show that yi can also be written
as:
yi = dj,i + RT G−1 Si+1 (7.53)
j=1...m
on condition that :
RT G−1 C ≡ 0 (7.54)
Expression (7.52) ensures, first, that the Hamming weight of vector
(d1,i , d2,i , · · · , dm,i , yi ) is at least equal to 2 when we leave the reference path
("all zero" path), in the trellis. Indeed, inverting any component of di modifies
the value of yi . Second, expression (7.53) indicates that the Hamming weight
of the same vector is also at least equal to 2 when we return to the reference
path. In conclusion, relations (7.52) and (7.53) together guarantee that the
free distance of the code, whose rate is R = m/(m + 1), is at least equal to 4,
whatever m.
7.5.2 m-binary turbo codes
Figure 7.20 – m-binary turbo encoder.
We consider a parallel concatenation of two m−binary RSC enoders associ-

ated with a permutation as a function of N words of m bits (k = mN ) (Fig-
ure 7.20). The blocks are encoded twice by this two-dimensional code, whose
rate is m/(m + 2). The circular trellis principle is adopted to enable encoding
of the blocks without a termination sequence and without edge effects.
The advantages of this construction compared to classical turbo codes are
the following :
• Better convergence. This advantage, observed first in [7.9], commented

in [7.16] and in a different way in [7.23], can be explained by a lower
density of errors on each of the two dimensions of the iterative process.
Take relation (7.8) that provides the upper bound of the accumulated
spatial distance for a binary code and adapt it to an m-binary code:

2k
sup Smin = (7.55)
m
For a coding rate R, the number of parity bits produced by the sequence
with accumulated length sup Smin is:

1−R m 1−R mk
nparity (sup Smin ) = sup Smin = (7.56)
R 2 R 2
Thus, replacing an (m = 1) binary turbo code by an (m = 2) double-binary

code,
√ the number of parity bits in the sequence considered is multiplied
by 2, although the accumulated spatial distance has been reduced by
the same ratio. Because the parity bits are local information for the two
elementary decoders (and are therefore not a source of correlation between
them), increasing the number of the former improves convergence. To
increase m beyond 2 slightly improves behaviour concerning correlation
but the effects are less visible than when passing from m = 1 to m = 2.
• Larger minimum distances. As explained above, the number of par-
ity bits produced by the RTZ sequences of input weight 2 is increased by
using m-binary codes. The same is true for all the simple RTZ sequences
defined in Section 7.3.2. The number of parity bits for these sequences is
at least equal to nparity (sup Smin ). The corresponding Hamming distances
are therefore even higher than those obtained with binary codes and con-
tribute even less to the MHD of the turbo code. As for the distances
associated with multiple RTZ patterns, which are generally those that fix
the MHD, they depend on the quality of the permutation implemented
(see Section 7.3.2).
• Less puncturing. To obtain coding rates greater than m/(m + 1) from
the encoder of Figure 7.20, it is not necessary to suppress as many redun-
dancy symbols as with a binary encoder. The performance of elementary
codes is improved by this, as Figure 7.21 shows. This figure compares the
correction capability of convolutional codes of rates 2/3 and 6/7, in the
binary (m = 1) and double-binary (m = 2) versions.
• Reduced latency. From the point of view of encoding as well as de-
coding, the latency (that is, the number of clock cycles necessary for the
processing) is divided by m since the data are processed in groups of m
bits. However, it may happen that the critical path of the decoding circuit
is increased compared to the case m = 1 as more data are to be considered
in a clock cycle. Parallelism solutions, such as those proposed in [7.35],
can help to increase the frequency of the circuit.
• Robustness of the decoder. For binary turbo codes, the difference in

performance between the MAP algorithm and its simplified versions or
between the MAP algorithm and the SOVA algorithm, vary from 0.2 to
0.6 dB, depending on the size of the blocks and the coding rates. This
difference is divided by two when we use double-binary turbo codes and
can be even lower for m > 2. This favourable (and slightly surprising)
property can be explained as follows: for a block of a given size (k bits),
the lower the number of steps in the trellis, the closer the decoder is to
the Maximum Likelihood (ML) decoder, whatever the algorithm on which
Figure 7.21 – Performance of simple binary (m = 1) and double-binary (m = 2)

convolutional codes, for R = 2/3 and R = 6/7.
it is based. Ultimately, a trellis reduced to a single step and therefore

containing all the possible codewords is equivalent to an ML decoder.
8-state double-binary turbo code

Figure 7.22(a) gives some examples of performance obtained with the turbo code
of [7.2], for a rate 2/3. The parameters of the constituent encoders are:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0 1 1 1 1
G = ⎣ 1 0 0 ⎦,C = ⎣ 0 1 ⎦,R = ⎣ 1 ⎦
0 1 0 0 1 0
The permutation function uses both inter- and intrasymbol interference. In

particular, we can observe:
• good average performance for this code whose decoding complexity re-
mains very reasonable (around 18,000 gates per iteration plus the mem-
ory);
• a certain coherence concerning the variation of performance with block
size (in agreement with the curves of Figures 3.6, 3.9, 3.10). The same
coherence could also be observed for the variation of performance with
coding rate;
• quasi-optimality of decoding with low error rates. The theoretical asymp-
totic curve for 188 bytes has been calculated from the sole knowledge of
the minimum distance of the code (that is, 13 with a relative multiplicity
of 0.5) and not from the total spectrum of the distances. In spite of this,
the difference between the asymptotic curve and the curve obtained by
simulation is only 0.2 dB for a PER of 10−7 .
16-state double-binary turbo code

The extension of the previous scheme to 16-state elementary encoders allows the
minimum distances to be greatly increased. We can, for example, choose:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
0 0 1 1 1 1 1
⎢ 1 0 0 0 ⎥ ⎢ 0 1 ⎥ ⎢ 1 ⎥
G=⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ 0 1 0 0 ⎦,C = ⎣ 0 0 ⎦,R = ⎣ 1 ⎦
0 0 1 0 0 1 0
For the rate 2/3 turbo code, again with blocks of 188 bytes, the minimum
distance obtained is equal to 18 (relative multiplicity of 0.75) instead of 13 for
the 8-state code. Figure 7.22(b) shows the gain obtained for low error rates:
around 1 dB for a PER of 10−7 and 1.4 dB asymptotically, considering the
respective minimum distances. We can note that the convergence threshold is
almost the same for 8-state and 16-state decoders, the curves being practically
identical for a PER greater than 10−4. The theoretical limit (TL), for R = 2/3
and for a blocksize of 188 bytes, is 1.7 dB. The performance of the decoder in
this case is: TL + 0.9 dB for a PER of 10−4 and TL + 1.3 dB for a PER of
10−7 . These intervals are typical of what we obtain in most rate and blocksize
configurations.
Replacing 4-PSK modulation by 8-PSK modulation, in the so-called prag-
matic approach, gives the results shown in Figure 7.22(b), for blocks of 188
and 376 bytes. Here again, good performance of the double-binary code can
be observed, with losses compared to the theoretical limits (that are around
3.5 and 3.3 dB, respectively) close to those obtained with 4-PSK modulation.
Associating turbo codes with different modulations is described in Chapter 10.
For a particular system, the choice between an 8-state or 16-state turbo code
depends, apart from the complexity desired for the decoder, on the target error
Figure 7.22 – (a) PER performance of a double-binary turbo code with 8 states for
blocks of 12, 14, 16, 53 and 188 bytes. 4-PSK, AWGN noise and rate 2/3. Max-Log-
MAP decoding with input samples of 4 bits and 8 iterations. (b) PER performance of
a double-binary turbo code with 16 states for blocks of 188 bytes (4-PSK and 8-PSK)
and 376 bytes (8-PSK), AWGN noise and rate 2/3. Max-Log-MAP decoding with
input samples of 4 bits (4-PSK) or 5 bits (8-PSK) and 8 iterations.
rates. To simplify, let us say that an 8-state turbo code suffices for PERs greater
than 10−4 . This is generally the case for transmissions having the possibility
of repetitions (ARQ: Automatic Repeat reQuest). For lower PERs, typical of
broadcasting or of mass memory applications, the 16-state code is highly prefer-
able.
7.6 Analysis tools

7.6.1 Theoretical performance
Figure 1.6 shows two essential parameters allowing the performance of an error
correcting code and its decoder to be evaluated:
• the asymptotic gain measuring the behaviour of the coded system at low
error rates. This is mainly dictated by the MHD of the code (see Sec-
tion 1.5). A low value of the MHD leads to a great change in the slope
(flattening) in the error rate curve. When the asymptotic gain is reached,
the BER(Eb /N0 ) curve with coding becomes parallel to the curve without
coding.
• the convergence threshold defined as the signal to noise ratio from which
the coded system becomes more efficient than the non-coded transmission
system;
In the case of turbo codes and the iterative process of their decoding, it is
not always easy to estimate the performance either of the asymptotic gain or of
the convergence. Methods for estimating or determining the minimum distance
proposed by Berrou et al. [7.18], Garello et al. [7.27] and Crozier et al. [7.22]
are presented in the rest of this chapter. The EXIT diagram method proposed
by ten Brink [7.46] to estimate the convergence threshold is also introduced.
7.6.2 Asymptotic behaviour

Determining the performance of error correcting codes with low error rates by
simulation requires high calculation power. It is, however, possible to estimate
this performance when the MHD dmin and the multiplicity are known (see Sec-
tion 1.5). Thus, the packet error rate with high signal to noise ratio Eb /N0 is
given by the first term of the union bound (UB). The expression of the UB is de-
scribed by relation (3.21), and estimation of the PER, given by Equation (1.16),
is shown again here:
" #
1 Eb
PER ≈ N (dmin ) erfc Rdmin (7.57)
2 N0
where N (dmin ), the multiplicity, represents the number of codewords at the

minimum distance.
The minimum distance of a code is not, in the general case, simple to de-
termine except if the number of codewords is low enough for us to make an
exhaustive list of them, or if particular properties of the code enable us to es-
tablish an analytical expression of this value (for example, the minimum distance
of a product code is equal to the product of the minimum distances of the con-
stituent codes). In the case of convolutional turbo codes, the minimum distance
is not obtained analytically; the only methods proposed are based on the total
or partial [7.28] enumeration of codewords whose input weight is lower than or
equal to the minimum distance. These methods are applicable in practice only
for small sizes blocksizes and small minimum distances, which is why they will
not be described here.
Error impulse method

This method, proposed by Berrou et al. [7.18], is not based on the analysis of
the properties of the code but on the correction capacity of the decoder. Its
principle, illustrated in Figure 7.23, involves superposing on the input sequence
of the decoder an error impulse whose amplitude Ai is increased until the decoder
no longer knows how to correct it.
Figure 7.23 – Schematic diagram of the error impulse method.
The code considered being linear, the sequence transmitted is assumed to be

the "all zero" sequence. The coding operation then produces codewords that
also contain only zeros. These are next converted into real values equal to -
1. If this succession of symbols was directly applied at the decoder, the latter
would not encounter any difficulty in retrieving the original message since the
transmission channel is perfect.
The proposed method involves adding an error impulse to the i-th symbol
(0 ≤ i ≤ k − 1) of the information sequence (systematic part), that is, trans-
forming a "−1" symbol into a symbol having a positive value equal to −1 + Ai .
If amplitude Ai is high enough, the decoder does not converge towards the "all
zero" word. Let us denote A∗i the maximum amplitude of the impulse in position
i such that the decoded codeword is the "all zero" word. It is shown in [7.18]
that, if the decoder performs maximum likelihood decoding, impulse distance
dimp = min (A∗i ) is also the minimum distance dmin from the code.
i=0,··· ,k−1
It is generally not necessary to test all the positions of the sequence. For a
shift invariant code (which is the case of convolutional codes), it suffices to apply
the error impulse to just one position of the datablock. For a code presenting a
periodicity of period P , it is necessary to test P positions. This method is appli-
Figure 7.24 – Measured and estimated PER (UB) of the DVB-RCS turbo code for
the transmission of MPEG (188 bytes) blocks with coding rates 2/3 and 4/5. 4-PSK
modulation and Gaussian channel.
cable to any linear code, for any blocksize and any coding rate, and it requires
only a few seconds to a few minutes calculation on an ordinary computer, the
calculation time being a linear function of the blocksize or of its period P .
When the decoding is not maximum likelihood, this method is no longer rig-
orous and produces only an estimation of the minimum distance. In addition,
the multiplicity of the codewords at distance dmin is not provided and Equa-
tion (7.57) cannot be applied without particular hypotheses about the prop-
erties of the code. In the case of turbo codes, two realistic hypotheses can be
formulated to estimate multiplicity: a single codeword at distance A∗i has its i-th
information bit at 1 (unicity), and the A∗i values corresponding to all positions
i come from distinct codewords (non-overlapping).
An estimation of the PER is then given by:

1
k−1
Eb
PER ≈ erfc( RA∗i ) (7.58)
2 i=0 N0
The first hypothesis (unicity) under-evaluates the value of the error rate, unlike
the second (non-overlapping) that over-evaluates it and, overall, the two effects
compensate each other. As an example, Figure 7.24 compares the measured per-
formance of the DVB-RCS turbo code, for two coding rates, with their estimate
deduced from (7.58). The parameters obtained by the error impulse method
are:
• dmin = 13 and n(dmin ) = 752 for R = 2/3

• dmin = 8 and n(dmin ) = 940 for R = 4/5
For packet error rates of 10−7, less than 0.2 dB separates the measured and
estimated curves.
Modified error impulse method

The approach of Garello et al. [7.27] is similar to the error impulse method pre-
sented above. It involves placing an impulse in row i in the "all zero" codeword.
This time, the amplitude of the impulse is high enough for the decoder not to
converge towards the "all zero" codeword but towards another sequence that
contains a 1 in position i. In addition, Gaussian noise is added to the input
sequence of the decoder, which tends to help the latter converge towards the
concurrent word having the lowest weight. This is what often happens when the
level of noise is well adjusted. In all cases, the weight of the codeword provided
by the decoder is an upper limit of the minimum distance of all the codewords
containing a 1 in row i. The minimum distance and the multiplicity are esti-
mated by sweeping all the positions. This algorithm works very well for small
and average distances.
Double error impulse method

The method proposed by Crozier et al. [7.22] is an improvement of the previous
method, at the expense of higher computation time. It involves placing a first
high level impulse at row i and a second at row j to the right of i and such that
j − i < r. The upper limit of r is 2D where D is an upper bound of the distance
to be evaluated. Then, decoding is applied similar to that described above but
with a stronger probability of obtaining a codeword at the minimum distance.
The calculation time is increased by a ratio r.
7.6.3 Convergence
A SISO decoder can be seen as a processor that transforms one of its input
values, the LLR of the extrinsic information used as a priori information, into
an output extrinsic LLR. In iterative decoding, the characteristics of the extrinsic
information provided by decoder 1 depend on the extrinsic information provided
by decoder 2 and vice-versa. The degree of dependency between the input and
output extrinsic information can be measured by the mutual information (MI).
The idea implemented by ten Brink [7.46] is to follow the exchange of ex-
trinsic information through the SISO decoders working in parallel on a diagram,
called an EXtrinsic Information Transfer (EXIT) chart. To elaborate the EXIT

chart, it is necessary to know the transfer characteristics of the extrinsic infor-
mation of each SISO decoder used in the decoding. This section shows how
to establish the transfer function of the extrinsic information for a SISO de-
coder, then construct the EXIT chart, and finally analyse the convergence of
the iterative decoder.
Figure 7.25 – Variation of variance σz as a function of the the mutual information IA
Transfer function for a SISO decoder of extrinsic information

a. Definition of the mutual information (MI)
If the weighted extrinsic information z on the coded binary element x ∈ { −

1, +1} follows a conditional probability density f (z|x), the MI I(z, x) measures
the quantity of information provided on average by z on x and equals

+∞
1 2f ( z| x)
I (z, x) = f ( z| x) × log2 dz (7.59)
2 x=−1,+1 f ( z| − 1) + f ( z| + 1)
−∞
Figure 7.26 – Algorithm for determining the transfer function IE = T (IA, Eb/N0 )
b. Definition of the a priori mutual information
Hypotheses:
• Hyp. 1: when the interleaving is large enough, the distribution of the input
extrinsic information can be approximated by a Gaussian distribution after
a few iterations.
• Hyp. 2: probability density f (z|x) satisfies the exponential symmetry
condition, that is, f (z|x) = f (−z|x)exp(−z).
The first hypothesis allows the a priori LLR ZA of a SISO decoder to be mod-
elled by a variable having independent Gaussian noise nz , with variance σz and
expectation μz , applied to the transmitted information symbol x according to
the expression
ZA = μz x + nz
The second hypothesis imposes σz2 = 2μz . The amplitude of the extrinsic infor-
mation is therefore modelled by the following distribution:
! $
2
1 (λ − μz x)
f ( λ| x) = √ exp − (7.60)
4πμz 4μz
From (7.59) and (7.60), observing that f ( z| 1) = f ( −z| − 1), we deduce the a
priori mutual information:

+∞ ! $
2
1 (λ − μz ) 2
IA = √ exp − × log2 dλ
4πμz 4μz 1 + exp (−λ)
−∞
or again

+∞ ! - C .2 $
1 λ − σz2 2
IA = 1 − √ exp − × log2 [1 + exp (−λ)] dλ (7.61)
2πσz 2σz2
−∞
We can note that lim IA = 0 (the extrinsic information does not provide any
σz →0
information about datum x) and that lim IA = 1 (the extrinsic information
σz →+∞
perfectly determines datum x).
IA is an increasing monotonous function of σz ; it is therefore invertible. Function
σz = f (IA) is shown in Figure 7.25.
Figure 7.27 – (a) Transfer characteristic of the extrinsic information for a 16-state
binary encoder, with rate 2/3 and MAP decoding with different Eb/N0.
(b) EXIT chart and trajectory for the corresponding turbo code, with pseudo-random
interleaving on 20,000 bits, for Eb /N0 = 2dB.
c. Calculation of the output mutual information
Relation (7.59) allows the calculation of the mutual information IS linked

with the extrinsic information produced by the SISO decoder:

+∞
1 2fs ( z| x)
IS = fs ( z| x) × log2 dz (7.62)
2 x=−1,+1 fs ( z| − 1) + fs ( z| + 1)
−∞
We can note that IS ∈ [0, 1].

The distribution fs is not Gaussian. It is therefore necessary to use a digital

calculation tool to determine it, which is the great drawback of this method.
If we view the MI of output IS as a function of IA and of the signal to noise
ratio Eb /N0 , the transfer function of the extrinsic information is defined by:
IE = T (IA, Eb /N0 ) (7.63)

d. Practical method to obtain the transfer function of the extrinsic informa-
tion
Figure 7.26 shows the path taken to establish the transfer characteristic of
the extrinsic information of a SISO decoder.
• step 1: Generation of the pseudo random message d to be transmitted; at

least 10000 bits are necessary for the statistical properties to be represen-
tative.
• step 2: Encoding the data with rate R then 2-PSK modulation of the
signal; the systematic and redundancy data both belong to the alphabet
{-1,+1}.
• step 3: Application of a Gaussian noise with signal to noise ratio Eb /N0
(in dB), with variance

Eb
1 10−0,1× /N0
σ= ·
2 R
• step 4: Application to the data transmitted (stored in a file) of normal

law N (μz , σz ) corresponding to the mutual information IA desired (see
Figure 7.25) to obtain the distribution of a priori extrinsic information.
• step 5: Initialization of the SISO decoder with the a priori LLRs (it might
be necessary, depending on the decoding algorithm chosen, to transform
the LLRs into probabilities).
• step 6: Recovering the LLRs at the output of the SISO decoder (corre-
sponding to one half-iteration of the decoding process), in a file.
• step 7: Utilization of digital calculation software to evaluate IS (relation
(7.62)).
• Trace the histograms of the LLR distributions output as a function

of the bit transmitted (hence the necessity to store this information
in two files).
• Evaluate the integral by the trapeze method.
• The result is the MI of output IS corresponding to the MI of input IA.
e. An example
The simulations were performed on a 16-state binary turbo code with rate
2/3, with a pseudo-random interleaving of 20,000 bits. The decoding algorithm
is the MAP algorithm. Figure 7.27(a) shows the relation between IS and IA as
a function of the signal to noise ratio of the Gaussian channel.
Figure 7.28 – EXIT charts for different Eb /N0 in the case of binary turbo codes,
rate 2/3, pseudo-random interleaving of 20,000 bits, (a) 16-state and (b) 8-state MAP
decoding.
EXIT chart
The extrinsic information transfer characteristic is now known for a SISO de-
coder. In the case of iterative decoding, the output of decoder 1 becomes
the input of decoder 2 and vice versa. Curves IS1 = f (IA1 = IS2) and
IS2 = f (IA2 = IS1), identical to one symmetry if the SISO decoders are the
same, are placed on the same graph as shown in Figure 7.27(b). In the case
of a high enough signal to noise ratio (here 2 dB), the two curves do not have
any intersection outside the point of coordinates (1,1) which materializes the
knowledge of the received message. Starting from null mutual information, it
is then possible to follow the exchange of extrinsic information along the iter-
ations. In the example of Figure 7.27(b), arrival at point (1,1) is performed in

3.5 iterations.
Figure 7.29 – Binary error rates of a 16-state (a) and 8-state (b) binary turbo code
with rate 2/3, with pseudo-random interleaving of 20000 bits. MAP decoding with 1,
3, 6, 10, 15 and 20 iterations and comparison with the convergence threshold estimated
by the EXIT method.
When the signal to noise ratio is too low, as in case Eb /N0 = 1.4 dB in
Figure 7.28(b), the curves have intersection points other than point (1, 1). The
iterative process starting from null MI at the input will therefore not be able to
lead to a perfectly determined message. The minimum signal to noise ratio for
which there is no intersection other than point (1,1) is the convergence threshold
of the turbo encoder. In the simulated example, this convergence can be esti-
mated at around 1.4 dB for 16-state (Figure 7.28(a)) and 8-state (Figure 7.28(b))
binary turbo codes.
Figure 7.29 shows the performance of 16-state and 8-state binary turbo codes
as a function of the number of iterations, and compares them with the conver-
gence threshold estimated by the EXIT chart method.
Bibliography
[7.1] Interaction channel for digital terrestrial television. DVB, ETSI EN 301
958, V1.1.1, pp. 28-30, Aug. 2001.
[7.2] Interaction channel for satellite distribution systems. DVB, ETSI EN 301
790, V1.2.2, pp. 21-24, Dec. 2000.
[7.3] Multiplexing and channel coding (fdd). 3GPP Technical Specification
Group, TS 25.212 v2.0.0, June 1999.
[7.4] Recommendations for space data systems. telemetry channel coding. Con-
sultative Committee for Space Data Systems, BLUE BOOK, May 1998.
[7.5] Ieee standard for local and metropolitan area networks, IEEE Std 802.16a,
2003. Available at
http://standards.ieee.org/getieee802/download/802.16a-2003.pdf.
[7.7] G. Battail. Weighting of the symbols decoded by the viterbi algorithm.
Annals of Telecommunications, 42(1-2):31–38, Jan.-Feb. 1987.
[7.8] G. Battail. Coding for the gaussian channel: the promise of weighted-
output decoding. International Journal of Satellite Communications,
7:183–192, 1989.
[7.9] C. Berrou. Some clinical aspects of turbo codes. In Proceedings of 3rd
International Symposium on turbo codes & Related Topics, pages 26–31,
Brest, France, Sept. 1997.
[7.10] C. Berrou, P. Adde, E. Angui, and S. Faudeuil. A low complexity soft-
output viterbi decoder architecture. In Proceedings of IEEE International
Conference on Communications (ICC’93), pages 737–740, GENEVA, May
1993.
[7.11] C. Berrou, C. Douillard, and M. Jézéquel. Designing turbo codes for
low error rates. In IEE colloquium : turbo codes in digital broadcasting –
Could it double capacity?, pages 1–7, London, Nov. 1999.
[7.12] C. Berrou, C. Douillard, and M. Jézéquel. Multiple parallel concatenation

of circular recursive convolutional (crsc) codes. Annals of Telecommuni-
cations, 54(3-4):166–172, March-Apr. 1999.
[7.13] C. Berrou and A. Glavieux. Reflections on the prize pa-

per: "near optimum error correcting coding and decoding: turbo
codes". IEEE IT Society Newsletter, 48(2):136–139, June 1998.
www.ieeeits.org/publications/nltr/98_jun/reflections.html.
GENEVA, May 1993.
[7.15] C. Berrou and M. Jézéquel. Frame-oriented convolutional turbo-codes.
Electronics letters, 32(15):1362–1364, July 1996.
[7.16] C. Berrou and M. Jézéquel. Non binary convolutional codes for turbo
coding. Electronics Letters, 35(1):39–40, Jan. 1999.
[7.17] C. Berrou, Y. Saouter, C. Douillard, S. Kerouédan, and M. Jézéquel.
Designing good permutations for turbo codes: towards a single model.
In Proceedings of IEEE International Conference on Communications
(ICC’04), Paris, France, June 2004.
[7.18] C. Berrou, S. Vaton, M. Jézéquel, and C. Douillard. Computing the min-

imum distance of linear codes by the error impulse method. In Proceed-
ings of IEEE Global Communication Conference (Globecom’2002), pages
1017–1020, Taipei, Taiwan, Nov. 2002.
[7.19] E. Boutillon and D. Gnaedig. Maximum spread of d-dimensional multiple
turbo codes. IEEE Transanction on Communications, 53(8):1237–1242,
Aug. 2005.
[7.20] A. R. Calderbank. The art of signaling: fifty years of coding theory. IEEE
Transactions on Information Theory, 44(6):2561–2595, Oct. 1998.
[7.21] S. Crozier and P. Guinand. Distance upper bounds and true minimum
distance results for turbo-codes with drp interleavers. In Proceedings of
3rd International Symposium on turbo codes & Related Topics, pages 169–
172, Brest, France, Sept. 2003.
[7.22] S. Crozier, P. Guinand, and A. Hunt. Computing the minimum distance
of turbo-codes using iterative decoding techniques. In Proceedings of the
22nd Biennial Symposium on Communications, pages 306–308, Kingston,
Ontario, Canada, 2004,May 31-June3.
[7.23] C. Douillard and C. Berrou. Turbo codes with rate-m/(m+1) constituent

convolutional codes. IEEE Trans. Commun., 53(10):1630–1638, Oct.
2005.
[7.24] L. Duan and B. Rimoldi. The iterative turbo decoding algorithm has fixed
points. IEEE Transactions on Information Theory, 47(7):2993–2995, Nov.
2001.
[7.25] P. Elias. Error-free coding. IEEE Transactions on Information Theory,
4(4):29–39, Sept. 1954.
[7.26] R. G. Gallager. Low-density parity-check codes. IRE Transactions on
Information Theory, IT-8:21–28, Jan. 1962.
[7.27] R. Garello and A. Vila Casado. The all-zero iterative decoding algo-
rithm for turbo code minimum distance computation. In Proceedings of
IEEE International Conference on Communications (ICC’2004), pages
361–364, Paris, France, 2004.
[7.28] R. Garello, P. Pierloni, and S. Benedetto. Computing the free distance of
turbo codes and serially concatened codes with interleavers: Algorithms
and applications. IEEE Journal on Selected Areas in Communications,
May 2001.
[7.29] K. Gracie and M.-H. Hamon. Turbo and turbo-like codes: Principles
and applications in telecommunications. In Proceedings of the IEEE, vol-
ume 95, pages 1228–1254, June 2007.
[7.30] J. Hagenauer and P. Hoeher. Concatenated viterbi-decoding. In Proceed-
ings of International Workshop on Information Theory, pages 136–139,
Gotland, Sweden, Aug.-Sept. 1989.
[7.31] J. Hagenauer and P. Hoeher. A viterbi algorithm with soft-decision out-

puts and its applications. In Proceedings of IEEE Global Communications
Conference (Globecom’89), pages 1680–1686, Dallas, Texas, USA, Nov.
1989.
[7.32] J. Hagenauer and P. Hoeher. Turbo decoding with tail-biting trellises. In
Proceedings of Signals, Systems and Electronics, URSI Int’l Symposium,
pages 343–348, Sept.-Oct. 1998.
[7.33] G. H. Hardy and E. M. Wright. An Introduction to the Theory of Numbers.

Oxford Univ. Press, Oxford, UK, 1979.
[7.34] C. Heegard and S. B. Wicker. Chapter 3. In Turbo coding. Kluwer Aca-

demic Publishers, 1999.
[7.35] H. Lin and D. G. Messerschmitt. Algorithms and architectures for concur-

rent viterbi decoding. In Proceedings of IEEE International Conference
on Communications (ICC’89), pages 836–840, Boston, June 1989.
[7.36] R. L. Rivest. Permutation polynomials modulo 2w. Finite Fields Their
Applications, 7:287–292, Apr. 2001.
[7.37] P. Robertson, P. Hoeher, and E. Villebrun. Optimal and suboptimal
maximum a posteriori algorithms suitable for turbo decoding. European
Transactions on Telecommunication, 8:119–125, March-Apr. 1997.
[7.38] E. Rosnes and O. Y. Takeshita. Optimum distance quadratic permutation
polynomial-based interleavers for turbo codes. In Proceedings of IEEE
International Symposyum on Information Theory (ISIT’07), pages 1988–
1992, Seattle, Washington, July 2006.
[7.39] J. Ryu and O. Y. Takeshita. On quadratic inverses for quadratic permu-
tation polynomials over integer rings. IEEE Transactions On Information
Theory, 52:1254–1260, March 2006.
[7.40] H. R. Sadjadpour, N. J. A. Sloane, M. Salehi, and G. Nebe. Interleaver
design for turbo codes. IEEE Journal on Selected Areas in Commununi-
cations, 19(5):831–837, May 2001.
[7.41] J. Sun and O. Y. Takeshita. Interleavers for turbo codes using permuta-
tion polynomials over integer rings. IEEE Transactions On Information
Theory, 51:101–119, Jan. 2005.
[7.42] Y. V. Svirid. Weight distributions and bounds for turbo-codes. European
Transactions on Telecommunication, 6(5):543–55, Sept.-Oct. 1995.
[7.43] O. Y. Takeshita. On maximum contention-free interleavers and permuta-
tion polynomials over integer rings. IEEE Transactions On Information
Theory, 52:1249–1253, March 2006.
[7.44] O. Y. Takeshita. Permutation polynomial interleavers: an algebraic-
geometric perspective. IEEE Transactions On Information Theory,
53:2116–2132, June 2007.
[7.45] R. M. Tanner. A recursive approach to low complexity codes. IEEE
Transactions on Information Theory, IT-271:533–547, Sept. 1981.
[7.46] S. ten Brink. Convergence behavior of iteratively decoded parallel con-
catenated codes. IEEE Transactions On Communications, 49(10), Oct.
2001.
[7.47] Z. Wang, Z. Chi, and K. K. Parhi. Area-efficient high-speed decoding
schemes for turbo decoders. IEEE Trans. VLSI Systems, 10(6):902–912,
Dec. 2002.
[7.48] C. Weiss, C. Bettstetter, and S. Riedel. Code constuction and decoding of

parallel concatenated tail-biting codes. IEEE Comm. Letters, 47(1):366–
386, Jan. 2001.
[7.49] Y. Weiss and W. T. Freeman. On the optimality of solutions of the max-

product belief-propagation algorithm in arbitrary graphs. IEEE Trans-
actions on Information Theory, 47(2):736–744, Feb. 2001.
Chapter 8
Turbo product codes
8.1 History
Because of the Gilbert-Varshamov bound, it is necessary to have long codes in
order to obtain block codes with a large minimum Hamming distance (MHD) and
therefore high error correction capability. But, without a particular structure,
it is almost impossible to decode these codes.
The invention of product codes, due to Elias [8.4], can be seen in this con-
text: it means finding a simple way to obtain codes with high error correction
capability that are easily decodable from simple elementary codes. These prod-
uct codes can be seen as a particular realization of the concatenation principle
(Chapter 6).
The first decoding algorithm results directly from the construction of these
codes. This algorithm alternates the hard decision decoding of elementary codes
on the rows and columns. Unfortunately, this algorithm does not allow us to
reach the maximum error correction capability of these codes. The Reddy-
Robinson algorithm [8.15] does allow us to reach it. But no doubt due to its
complexity, it has never been implemented in practical applications.
The aim of this chapter is to give a fairly complete presentation of algorithms
for decoding product codes, whether they be algorithms for hard data or soft
data.
8.2 Product codes

With conventional constructions, it is theoretically possible to build codes
having a high MHD. However, the decoding complexity becomes prohibitive,
even for codes having an algebraic structure, like Reed-Solomon codes or BCH
(see Chapter 4) codes. For example, for Reed-Solomon codes on F256 , the
most complex decoder to have been implemented on a circuit has an error
correction capability limited to 11 error symbols, which is insufficient for most

applications today. The construction of product codes allows this problem to be
circumvented: by using simple codes with low correction capability, but whose
decoding is not too costly, it is possible to assemble them to obtain a longer
code with higher correction capability.
Definition
Let C1 (resp. C2 ) be a linear code of length n1 (resp. n2 ) and with dimension1
k1 (resp. k2 ). The product code C = C1 ⊗ C2 is the set of matrices M of size
n1 × n2 such that:
• Each row is a codeword of C1 ,

• Each column is a codeword of C2 .
This code is a linear code of length n1 × n2 and with dimension k1 × k2 .
Example 8.1
Let H be the Hamming code of length 7 and P be the parity code of length 3.
The dimension of H is 4 and the dimension of P is 2. The code C = H ⊗ P
is therefore of length 21 = 7 × 3 and dimension 8 = 4 × 2. Let the following
information word be coded:

0 1 1 0
I= .
1 0 1 0
Each row of a codeword of C must be a codeword of H. Therefore to code

I, we begin by multiplying each row of I by the generating matrix of code H:
⎡ ⎤
1 0 0 0 1 1 1
⎢0 1 0 0 1 1 0⎥
0 1 1 0 ·⎢
⎣0
⎥= 0 1
⎦ 1 0 0 1 1
0 1 0 1 0 1
0 0 0 1 0 1 1
⎡ ⎤
1 0 0 0 1 1 1
⎢0 1 0 0 1 1 0⎥
1 0 1 0 ·⎢
⎣0
⎥= 1 0 1 0 0 1 0
0 1 0 1 0 1⎦
0 0 0 1 0 1 1
1 or length of message
8. Turbo product codes 273
After encoding the rows, the provisional codeword is therefore:

0 1 1 0 0 1 1
.
1 0 1 0 0 1 0
Each column of the final codeword must now be a codeword with parity P .
The final codeword is therefore obtained by adding a third row made up of the
parity bits of each column. The complete codeword is:
⎡ ⎤
0 1 1 0 0 1 1
⎣1 0 1 0 0 1 0⎦.
1 1 0 0 0 0 1
For the codeword to be valid, it must then be verified that the third row of the
word is indeed a codeword of H. This row vector must therefore be multiplied
by the parity control matrix of H:
⎡ ⎤
1
⎢1⎥
⎡ ⎤ ⎢ ⎥ ⎡ ⎤
1 1 1 0 1 0 0 ⎢ ⎥ 0
⎢0⎥
⎣1 1 0 1 0 1 0⎦ · ⎢0⎥ = ⎣0⎦
⎢ ⎥
1 0 1 1 0 0 1 ⎢ ⎥ 0
⎢0⎥
⎣0⎦
1
In fact, it is not worthwhile doing this verification: it is ensured by con-
struction since codes H and P are linear. In addition, the encoding order is not
important: if we first code by columns then by rows, the codeword obtained is
the same.
8.3 Hard input decoding of product codes

8.3.1 Row-column decoding
The first decoding algorithm results directly from the construction of the code:
we successively alternate decoding the rows by a decoder of code C1 and decod-
ing the columns by a decoder of code C2 . Let d1 (resp. d2 ) be the minimum
distance of code C1 (resp. C2 ). Then, syndrome decoding of C1 (resp. C2 ) is
t1 -correcting (resp. t2 -correcting) with t1 = d1 /2 (resp. t2 = d2 /2).
Property
Row-column decoding is limited by a correction capability of (t1 + 1) · (t2 + 1)
errors. In other words, row-column decoding decodes any word having at least
(t1 + 1) · (t2 + 1) errors (even if it might decode certain patterns having more
errors) and there are words having exactly (t1 + 1) · (t2 + 1) errors that will not
be decoded.
Indeed, assume that we have a pattern with a number of errors strictly lower
than (t1 + 1) · (t2 + 1). Since the algorithm decoding in rows corrects up to t1
errors, after the first decoding step, any row with errors contains at least t1 + 1
errors. There are therefore at least t2 rows with errors after row decoding. Each
column therefore contains at least t2 errors and column decoding then eliminates
all the errors.
There are undecodable patterns having exactly (t1 + 1) · (t2 + 1) errors. Take
a codeword of C1 ⊗ C2 for which we choose (t2 + 1) rows and (t1 + 1) columns at
random. At each intersection between a row and a column, we insert an error in
the initial codeword. By construction, for the word thus obtained, there exists
a codeword for the product code at a distance of (t1 + 1) · (t2 + 1) errors, but
row decoding and column decoding fail.
We can note that row-column decoding is less powerful than syndrome decod-
ing for a product code. Indeed, a product code is a linear code whose minimum
distance is d1 d2 . Therefore syndrome decoding allows us to correct all the words
having at least t errors with t = (d1 d2 )/2. But row-column decoding allows
only all the words having less than (t1 + 1) · (t2 + 1) = (d1 /2 + 1)(d2 /2 + 1)
errors to be corrected. We therefore lose around a factor 2 in error correction
capability.
Example 8.2
We assume that we have a product code whose row code and column code both
have a minimum distance equal to 5. They are therefore both 2-correcting. The
row-column decoding of the product code, according to the above, can thus
correct any word having at most 8 errors. Figure 8.1 illustrates a word having
10 errors (shown as points) but that can be corrected by row-column decoding.
Figure 8.2 shows a pattern having the same number of errors but not correctable.
8.3.2 The Reddy-Robinson algorithm

The Reddy-Robinson algorithm [8.15] is a more sophisticated iterative algorithm
which, now, assumes that for codes C1 and C2 we have decoders with errors and
erasures. An erasure is an unreliable position in the frame received, whose
symbol we think might be erroneous. The difference with a standard error is
that the position of the erasure is known at the moment of decoding. For an
MHD code equal to d, syndrome decoding can be adapted to take into account
the erasures and then it is possible to decode any frame with t errors and e
erasures as long as we have
2t + e < d (8.1)
Figure 8.1 – Erroneous word that can be corrected by row-column decoding.
Figure 8.2 – Erroneous word that cannot be corrected by row-column decoding.
The algebraic algorithms of BCH and Reed-Solomon codes can also be

adapted to treat erasures.
The Reddy-Robinson algorithm is thus the following:

• Step 1: Find the errors in each row i by applying the decoder of code C1
without erasures and assign them a weight li equal to the number of errors
detected. If the decoding fails, the weight assigned is li = d1 /2. The rows
are passed without correction to step 2.
• Step 2: Decode the columns by the decoder of code C2 with erasures.
For each column, we successively perform all the decodings by erasing the
least reliable symbols (i.e. those that have the highest row weight). We
therefore have at least d1 /2 decodings per row. At each decoding we
n
assign a weight W such that W = wi where wi = li if the symbol of
i=1
row i is unchanged by the decoding and wi = d2 /2 otherwise.
• Step 3: At the final decoding, for each column we choose the decoded word
giving the smallest weight W .
Reddy-Robinson decoding allows any pattern of errors to be corrected whose
weight is strictly lower than d1 d2 [8.15].
Example 8.3
Let us again take the above example with the word of Figure 8.2. During the
first step, those rows with 3 errors will not be able to be corrected by the row
code since the latter can correct only a maximum of 2 errors (MHD 5). They
will therefore be assigned an equal weight 2.5. The row with one error will have
a weight equal to 1, while all the remaining rows will have a weight equal to 0.
The configuration is then as shown in Figure 8.3.
At the second step, the correction becomes effective. Only three columns
have errors. Concerning the most left-hand column with errors, according to
the weights provided by step 1, three symbols in this column have a weight
equal to 2.5, one symbol with a weight equal to 1 and all the others have a
null weight. The second step of the algorithm for this column will therefore
involve three successive decodings: decoding without erasure, decoding with
three erasures (the symbols having a weight equal to 2.5) and decoding with
four erasures (the three previous symbols plus the symbol having a weight of 1).
The first decoding fails since the code is only 2-correcting. The second decoding
succeeds. Indeed, the column code has a minimum distance of 5. It can thus
correct t errors and e erasures when 2t + e < 5. Now, for this decoding, we have
e = 3 (since 3 erasures are placed) and t = 0 (since there are no additional errors
in the column). The weight associated with the second decoding is the sum of the
weights of the symbols erased, that is, 7.5. Likewise, the third decoding (with
4 erasures) succeeds and the weight associated with the word decoded is thus
8.5. The algorithm then chooses from among the decodings having succeeded
the one whose weight is the lowest, that is, the second decoding in this case.
Figure 8.3 – Calculation of the weight of row coming from the first step of the Reddy-
Robinson algorithm.
The two other columns with errors are also decoded. However for the most
right-hand column, the second decoding fails (since there is one error in the
non-erased symbols) and the word decoded for this column is therefore that of
the third decoding. Finally, all the errors are corrected.
In this algorithm the role of the rows and of the columns is not symmetric.
Thus, if the whole decoding fails in the initial order, it is possible to try again
by inverting the role of the rows and the columns.
The Reddy-Robinson decoding algorithm is not iterative in its initial version.
We can make it iterative, for example, by starting another decoding with the
final word of step 2, if one of the decodings of step 2 succeeded. There are also
more sophisticated iterative versions [8.16].
8.4 Soft input decoding of product codes

8.4.1 The Chase algorithm with weighted input
The Chase algorithm, in the case of a block code, enables the values received
on the transmission channel to be used to decode in the maximum likelihood
sense or, at least approximate its performance. In its basic version, it produces
hard decoding. Following the idea of convolutional turbo code decoding [8.2],
Pyndiah’s improved version [8.14] of the algorithm allows a soft value of the
decoded bit to be obtained at the output. With the help of iterative decoding,
it is then possible for the row and column decoders of a product code to exchange
extrinsic information about the bits. We can thus decode the product codes in
a turbo manner.
Let therefore r = (r1 , ..., rn ) be the received word after encoding and trans-
mission on a Gaussian channel. The Chase-Pyndiah algorithm with t places is
decomposed as follows:
• Step 1: Select the t places Pk in the frame containing the least reliable
symbols in the frame (i.e. the t places j for which the rj values are the
smallest in absolute value).
• Step 2: Generate the vector of the hard decisions h0 = (h01 , ..., hon ) such
that h0j = 1 if rj > 0 and 0 otherwise. Generate the vectors h1 , ..., h2t −1
such that hij = h0j if j ∈ / {Pk } and hiPk = h0Pk ⊕ N um(i, k) where
N um(i, k) is the k-th bit in the binary writing of i.
• Step 3: Decode the words h0 , ..., h2t −1 with the hard decoder of the linear
code. We thus obtain the concurrent words c0 , ..., c2t −1 .
• Step 4: Calculate the metrics of the concurrent words

Mi = rj (1 − 2cij )
1≤j≤n
• Step 5: Determine the index pp such that
Mpp = min {Mi }
Codeword cpp is then the most probable codeword.

• Step 6: For each bit j in the frame, calculate the reliability
Fj = 1/4(min {Mi , cij = cpp,j } − Mpp )
If there are no concurrent words for which the j-th bit is different from
cpp,j , then the reliability Fj is at a fixed value β.
• Step 7: Calculate the extrinsic value for each bit j,
Ej = (2 × cpp,j − 1) × Fj − rj
The extrinsic values are then exchanged between row decoders and column
decoders in an iterative process. The value β, as well as the values of the feed-
backs are here more sensitive than in the case of decoding convolutional turbo
codes. Inadequate values can greatly degrade the error correction capability.
However, it is possible to determine them incrementally. First, for the first iter-
ation we search to see which values give the best performance (for example by
dichotomy). Then, these values being fixed, we perform a similar search for the
second iteration, and so forth.
Example 8.4
Let r = (0.5; 0.7; −0.9; 0.2; −0.3; 0.1; 0.6) be a received sample of a Hamming
codeword (7, 4, 3) and the number of places of the Chase algorithm is equal to
t = 3. We choose β = 0.6. The above algorithm thus gives:
• Step 1: P1 = 6, P2 = 4, P3 = 5.
• Step 2:
I hi
0 (1;1;0;1;0;1;1)
1 (1;1;0;1;0;0;1)
2 (1;1;0;0;0;1;1)
3 (1;1;0;0;0;0;1)
4 (1;1;0;1;1;1;1)
5 (1;1;0;1;1;0;1)
6 (1;1;0;0;1;1;1)
7 (1;1;0;0;1;0;1)
The bits underlined correspond to the inversions performed by the Chase
algorithm.
• Step 3:
I hi ci
0 (1;1;0;1;0;1;1) (1;1;0;1;0;1;0*)
1 (1;1;0;1;0;0;1) (1;1;0;0*;0;0;1)
2 (1;1;0;0;0;1;1) (1;1;0;0;0;0*;1)
3 (1;1;0;0;0;0;1) (1;1;0;0;0;0;1)
4 (1;1;0;1;1;1;1) (1;1;1*;1;1;1;1)
5 (1;1;0;1;1;0;1) (0*;1;0;1;1;0;1)
6 (1;1;0;0;1;1;1) (1;0*;0;0;1;1;1)
7 (1;1;0;0;1;0;1) (1;1;0;0;0*;0;1)
The bits with a star in the column of concurrent words ci correspond to
the places corrected by the hard decoder in word hi .
• Step 4:
I ci Mi
0 (1;1;0;1;0;1;0*) -(0.5)-(0.7)+(-0.9)-(0.2)+(-0.3)-(0.1)+(0.6)=-2.1
1 (1;1;0;0*;0;0;1) -(0.5)-(0.7)+(-0.9)+(0.2)+(-0.3)+(0.1)-(0.6)=-2.7
2 (1;1;0;0;0;0*;1) -(0.5)-(0.7)+(-0.9)+(0.2)+(-0.3)+(0.1)-(0.6)=-2.7
3 (1;1;0;0;0;0;1) -(0.5)-(0.7)+(-0.9)+(0.2)+(-0.3)+(0.1)-(0.6)=-2.7
4 (1;1;1*;1;1;1;1) -(0.5)-(0.7)-(-0.9)-(0.2)-(-0.3)-(0.1)-(0.6)=-0.9
5 (0*;1;0;1;1;0;1) +(0.5)-(0.7)-(-0.9)-(0.2)-(-0.3)+(0.1)-(0.6)=0.3
6 (1;0*;0;0;1;1;1) -(0.5)+(0.7)+(-0.9)+(0.2)-(-0.3)-(0.1)-(0.6)=-0.9
7 (1;1;0;0;0*;0;1) -(0.5)-(0.7)+(-0.9)+(0.2)+(-0.3)+(0.1)-(0.6)=-2.7
• Step 5: We have pp = 1. The word decoded is (1; 1; 0; 0; 0; 0; 1). Note that

we encounter it several times in the list of concurrent words.
• Steps 6 and 7:
j Fj Ej
1 ((0.3)-(-2.7))/4=0.75 0.75-0.5=0.25
2 ((-0.9)-(-2.7))/4=0.475 0.475-0.7=-0.225
3 ((-0.9)-(-2.7))/4=0.475 -0.475-(-0.9)=0.525
4 ((-0.9)-(-2.7))/4=0.475 -0.475-0.2=-0.675
5 ((-2.1)-(-2.7))/4=0.15 -0.15-(-0.3)=0.15
6 ((-0.9)-(-2.7))/4=0.475 -0.475-0.1=-0.575
7 ((-2.1)-(-2.7))/4=0.15 0.15-0.6=-0.4
8.4.2 Performance of the Chase-Pyndiah algorithm

The Chase-Pyndiah algorithm is the most widespread of weighted decoding al-
gorithms for block codes. Figure 8.4 gives the performance obtained by this
algorithm in the context of turbo decoding for different product codes. The
curves are circuit oriented, i.e. the data are quantified on q = 5 bits. The
simulations are done on a Gaussian channel.
Figure 8.5 shows the evolution of the binary error rates during the turbo
decoding of the extended BCH code, BCH(64,57,4)2 , with the Chase-Pyndiah
algorithm. For this code, we can see that most of the gain in correction is
obtained during the first 4 iterations.
8.4.3 The Fang-Battail algorithm

Like the Chase-Pyndiah algorithm, the Fang-Battail algorithm enables the re-
liability of the decoded bits to be calculated from soft data sampled at the
output of the transmission channel. We can see this algorithm as a variant of
the Chase-Pyndiah algorithm for which only the least reliable bits are modi-
fied. Let r = (r1 , ..., rn ) be the received word. The code used is a linear code of
length n and dimension k of a generating matrix G = {gij , 0 ≤ i < k, 0 ≤ j < n}
and parity control matrix H = {hij , 0 ≤ i < n − k, 0 ≤ j < n}. The algorithm
proceeds in the following steps:
• Step 1: Sort the values ri in the order of decreasing absolute values. We
thus obtain a permutation P on all the indices [1..n]. We put rl = rP (l)
and we therefore have |r1 | ≥ |r2 | ≥ ... ≥ |rn |.
• Step 2: Let H∗ be the matrix obtained by permuting the columns of H by

P . Systematize the k most right-hand columns of H∗ by Gauss reduction.
Let H be the matrix obtained.
Figure 8.4 – Performance of the Chase-Pyndiah algorithm in the turbo decoding of

different 1-correcting codes (2-PSK on a Gaussian channel, 8 iterations, q = 5). (Source
P. Adde, Electronics Department, ENST Bretagne).
• Step 3: Let s be the vector of the hard decisions of r . Let M1 be the

accumulated metrics associated with the n − k first values of r . That is,

s = H s1 s2 ... sn−k 0 0 ... 0
Let M2 be the accumulated metrics associated with s in the k last values
of r . The total accumulated metric is then M = M1 + M2 .
Figure 8.5 – Evolution of the binary error rate during Chase-Pyndiah turbo decoding
for 1, 2, 3, 4 and 6 iterations (BCH code(64,51,6), 2-PSK on a Gaussian channel).
• Step 4: List the concurrent words in the order of increasing metrics M1

(we use a combination generation algorithm with exclusion). Stop as soon
as the total metric M begins to increase.
• Step 5: After the inverse permutation of the metrics, calculate the extrinsic
values from the concurrent words by using the same equations as in the
Chase-Pyndiah algorithm.
Example 8.5
We consider the same example as for the Chase algorithm.
Let r = (0.5; 0.7; −0.9; 0.2; −0.3; 0.1; 0.6) be a received sample. The parity
control matrix of the Hamming code of dimension 4 and length 7 is, as we have
already seen, ⎡ ⎤
1 1 1 0 1 0 0
H = ⎣1 1 0 1 0 1 0⎦
1 0 1 1 0 0 1
• Step 1: We have r = (−0.9; 0.7; 0.6; 0.5; −0.3; 0.2; 0.1) and P =
[3, 2, 7, 1, 5, 4, 6].
• Step 2: ⎡ ⎤
1 1 0 1 1 0 0
H∗ = ⎣0 1 0 1 0 1 1⎦
1 0 1 1 0 1 0
After systematization on the last three columns, we obtain

⎡ ⎤
1 1 0 1 1 0 0
H = ⎣1 0 1 1 0 1 0⎦
1 1 1 0 0 0 1
• Step 3: The hard decision on the n − k = 4 highest received values gives

s = (0, 1, 1, 1). Matrix H without the k last columns makes it possible
to find the missing redundancy values:
⎡ ⎤
⎡ ⎤ 0 ⎡ ⎤
1 1 0 1 ⎢ ⎥ 0
1
⎣1 0 1 1⎦ · ⎢
⎣1⎦
⎥ = ⎣0⎦
1 1 1 0 0
1
The initial decoded vector is therefore (0, 1, 1, 1, 0, 0, 0) which, after inverse
permutation, gives the vector (1, 1, 0, 0, 0, 0, 1). The initial metric is M =
M1 + M2 = −2.7 as M1 = (−0.9) − 0.7 − 0.6 − 0.5 = −2.7 and M2 =
(−0.3) + 0.2 + 0, 1 = 0.0.
• Step 4: To list the concurrent words, we apply inversion masks to the
n − k first bits (we can have from 1 to 4 inversions maximum). Each
inversion mask will increase the M1 part of the metric. For an inversion,
the bonus is at minimum 1.0 and at maximum 1.8. The minimum bonus
for two inversions is at minimum 2 × (0.6 + 0.5) = 2.2. The first concurrent
words to consider are therefore all those corresponding to a single inversion.
Moreover, modifications on the M2 part of the metric could decrease it.
However, the decrease cannot exceed 2 × (0.2 + 0.1) = 0.6 compared to
the initial metric. We therefore already know that the found previously
word is the most likely word. However, if we decide to list them anyway,
we find the following concurrent words which, unlike the Chase-Pyndiah
algorithm, are all different from each other:
si Ci M
(0,1,1,1) (0,1,1,1,0,0,0) -2.7
(0,1,1,0) (0,1,1,0,1,1,0) -2.7+1.0+0.2=-1.5
(0,1,0,1) (0,1,0,1,0,1,1) -2.7+1.2-0.6=-2.1
(0,0,1,1) (0,0,1,1,1,0,1) -2.7+1.4+0.2=-1.1
(1,1,1,1) (1,1,1,1,1,1,1) -2.7+1.8+0.0=-0.9
(0,1,0,0) (0,1,0,0,1,0,1) -2.7+2.2+0.2=-0.3
(0,0,1,0) (0,0,1,0,0,1,1) -2.7+2.4-0.6=-0.9
(0,0,0,1) (0,0,0,1,1,1,0) -2.7+2.6+0.2=0.1
(1,1,1,0) (1,1,1,0,0,0,1) -2.7+2.8-0.2=-0.1
(1,1,0,1) (1,1,0,1,1,0,0) -2.7+3.0+0.6=0.9
(1,0,1,1) (1,0,1,1,0,1,0) -2.7+3.2-0.2=0.3
(0,0,0,0) (0,0,0,0,0,0,0) -2.7+3.6+0.0=0.9
(1,1,0,0) (1,1,0,0,0,1,0) -2.7+4.0-0.2=1.1
(1,0,1,0) (1,0,1,0,1,0,0) -2.7+4.2+0.6=2.1
(1,0,0,1) (1,0,0,1,0,0,1) -2.7+4.4-0.2=1.5
(1,0,0,0) (1,0,0,0,1,1,1) -2.7+5.4+0.0=2.7
• Step 5:
j FP (j)
1 ((-0.9)-(-2.7))/4=0.475
2 ((-1.1)-(-2.7))/4=0.4
3 ((-2.1)-(-2.7))/4=0.15
4 ((-1.5)-(-2.7))/4=0.3
5 ((-1.5)-(-2.7))/4=0.3
6 ((-1.5)-(-2.7))/4=0.3
7 ((-2.1)-(-2.7))/4=0.15
J EJ
1 0.3-0.5=-0.2
2 0.4-0.7=-0.3
3 -0.475-(-0.9)=0.525
4 -0.3-0.2=-0.5
5 -0.3-(-0.3)=0.0
6 -0.15-0.1=-0.25
7 0.15-0.6=-0.4
The Fang-Battail algorithm is theoretically maximum likelihood. However,

that will not necessarily suffice for the algorithm applied for decoding product
codes to be maximum likelihood itself.
8.4.4 The Hartmann-Nazarov algorithm

Hartmann et al. [8.8] used the properties of duality to describe a maximum
likelihood decoding method for linear codes. Initially, only a hard decision was
provided by the algorithm. Hagenauer [8.7] then took up these ideas in order
to make available the extrinsic values necessary for turbo decoding. Finally,
Nazarov and Smolyaninov [8.13] showed how to reduce the complexity cost by
using the fast Hadamard transform. This paragraph summarizes these three
reference articles.
Let r = (r1 , ..., rn ) be the received word. The code used is a linear code of
length n and dimension k of a generating matrix G = {gij , 0 ≤ i < k, 0 ≤ j < n}
and parity control matrix H = {hij , 0 ≤ i < n − k, 0 ≤ j < n}. We assume that
the transmission has been done on a channel perturbed by a Gaussian noise with
null mean and variance equal to σ 2 .
Putting ρl = − tanh(rl /σ2 ), we show that the LLR maximum likelihood of
the l-th bit in the frame is:

1 − C(l)
LLRl = ln (8.2)
1 + C(l)
with:

ρl FD (hl ) 1 FD (hl )
C(l) = 1+ + 1−
2 FD (0) 2ρl FD (0)
where hl is the l-th column of the parity control matrix and function FD is such
that
% &
−1
2n−k
n−k−1
FD (hl ) = Dν (l) exp jπ νm hml
ν=0 m=0

n−k−1
νm is the m-th bit of the binary representation of integer ν (ν = νm 2m ),
m=0
and:
5
n−1
Dν (l) = (ρl )tν (l)
l=0
tν (l) being the l-th bit of the ν-th vector of the dual of the code, that is, therefore:
"n−k−1 #

tν (l) = νm hml mod 2 = ν, hl mod 2
m=0
The calculation of Dν (l) normally requires n multiplications. But using the

dual code and applying a fast Hadamard transform, it is possible to lower this
cost to one term of the order of n − k multiplications. If the coding rate is
high, then n is much higher than n − k and the gain in terms of computation
complexity is high.
To do this, we re-write the general term of Dν (l) in the form:
(ρl )tν (l) = exp {tν (l) ln |ρl |} exp {jπql tν (l)}
where ql is such that ρl = (−1)ql |ρl |.
We then have: %n−1 &

Dν (l) = exp (tν (l) ln |ρl |) + (jπql tν (l))
l=0
%n−1 & %n−1 &

= exp (tν (l) ln |ρl |) exp (jπql tν (l))
l=0 l=0
Put: % &

n−1
n−k−1
Fρ (w) = ln (|ρl |) exp jπ wm hml
l=0 m=0
for any integer 0 ≤ w ≤ 2n−k − 1. We have, therefore:

n−1
Fρ (w) = ln (|ρl |) exp {jπtw (l)}
l=0

n−1
with, in particular, Fρ (0) = ln (|ρl |).
l=0
1−exp{jπt}
On the other hand, if t = 0 or 1, then 2 = t and
Fρ (0) − Fρ (ν)
n−1
= (tν (l) ln |ρl |) .
2
l=0
/ n−k−1 0

n−1
Likewise, if we put Fq (w) = ql exp jπ wm hml , we have:
l=0 m=0
Fq (0) − Fq (ν)
n−1
= (ql ln |ρl |)
2
l=0
and therefore:

1 1
Dν (l) = exp (Fρ (0) − Fρ (ν)) exp jπ (Fq (0) − Fq (ν))
2 2
The two terms Fρ (ν) and Fq (ν) have a common expression of the form:
% n−k−1 &

n−1
F (w) = fl exp jπ wm hml
l=0 m=0
where fl are real numbres that depend

only on l.
We define function g on the set 0, · · · , 2n−k − 1 with by:
⎧
⎨
n−k−1
fl if ∃l, p = hml 2m
g(p) =
⎩ m=0
0 otherwise
Function g is well defined since the columns of H are linearly independent and
therefore a fortiori two-by-two distinct. The Hadamard
of G is then
transform
a function with real values defined on the interval 0..2n−k − 1 by:
−1
2n−k
G(w) = g(p)(−1)<p,w>
p=0

n−k−1
Now, function g is null except for points pl = hml 2m for l ∈ [0, · · · , n − 1].
l=0
Thus, we have:

" #

n−k−1
n−1 n−1 n−k−1
< hml 2m ,w>
G(w) = fl (−1) m=0 = fl exp jπ wm hml = F (w)
l=0 l=0 m=0
The two terms Fρ (ν)and Fq (ν) are therefore expressed as Hadamard transforms
and can be calculated by means of the fast Hadamard transform.
Fast Hadamard transform
Let R = [R0 , R1 , ..., R2n −1 ] be a vector with real components. The vector
obtained from R by the Hadamard transform is vector R̂ = [R̂0 , R̂1 , · · · , R̂2n −1 ]
such that
n
2 −1
R̂j = Ri (−1)<i,j> (8.3)
i=0
The scalar product < i, j > is, as above, the bit-by-bit scalar product of the
binary developments of i and j. We also write in vector form, R̂ = RH2n where
H2n is the Hadamard matrix of order 2n whose coefficient (H2n )i,j = (−1)<i,j> .
Let A be a matrix size a1 ×a2 and B a matrix size b1 ×b2 with real coefficients.
Then the Kronecker product of A by B, denoted A⊗B, is a matrix size (a1 b1 )×
(a2 b2 ) such that (A ⊗ B)i,j = Aq1 q2 Br1 r2 where i = b1 q1 + r1 and j = b2 q2 + r2
with 0 ≤ r1 < b1 and 0 ≤ r2 < b2 .
If N = 2n , we show that ([8.11]):
HN = (H2 ⊗ IN/2)(I2 ⊗ HN/2 )

where Ik is the unit matrix of order k.

Developing recursively, we obtain:
5
n
- .
HN = I2i−1 ⊗ H2 ⊗ IN/2i
i=1
The fast Hadamard transform is in fact the use of this factorization to calculate
R̂.
Example 8.6
Let us calculate the Hadamard transform of vector
R = [0.2; 0.5; −0.7; 1.3; 0.1; −1.1; 0.8; −0.3]
We have R̂ = RH8 with

⎡ ⎤
1 1 1 1 1 1 1 1
⎢ 1 −1
⎢ 1 −1 1 −1 1 −1 ⎥⎥
⎢ 1 1
⎢ −1 −1 1 1 −1 −1 ⎥
⎥
⎢ 1 −1 −1 1 1 −1 −1 1 ⎥⎥
H8 = ⎢
⎢ 1 1
⎢ 1 1 −1 −1 −1 −1 ⎥
⎥
⎢ 1 −1
⎢ 1 −1 −1 1 −1 1 ⎥⎥
⎣ 1 1 −1 −1 −1 −1 1 1 ⎦
1 −1 −1 1 −1 1 1 −1
The direct calculation gives:
R = [0.8; 0.0; −1.4; 1.8; 1.8; −4.6; 1.6; 1.6]
Now, according to the above, H8 = G1 G2 G3 where Gi = Ii ⊗ H2 ⊗ I8/2i ,

we have:
⎡ ⎤
1 0 0 0 1 0 0 0
⎢ 0 1 0 0 0 1 0 0 ⎥
⎡ ⎤ ⎢ ⎥
1 0 0 0 ⎢ 0 0 1 0 0 0 1 0 ⎥
⎢ ⎥
1 1 ⎢ 0 1 0 0 ⎥ ⎢ 0 0 0 1 0 0 0 1 ⎥
G1=[1] ⊗ ⊗⎢ ⎥=⎢ ⎥
1 −1 ⎣ 0 0 1 0 ⎦ ⎢ 1 0 0 0 −1 0 0 0 ⎥
⎢ ⎥
0 0 0 1 ⎢ 0 1 0 0 0 −1 0 0 ⎥
⎢ ⎥
⎣ 0 0 1 0 0 0 −1 0 ⎦
0 0 0 1 0 0 0 −1
⎡ ⎤
1 0 1 0
1 0 1 1 1 0 1 0 ⎢ 0 1 0 1 ⎥
G2 = ⊗ ⊗ = ⊗⎢
⎣ 1
⎥
0 1 1 −1 0 1 0 1 0 −1 0 ⎦
0 1 0 −1
⎡ ⎤
1 0 1 0 0 0 0 0
⎢ 0 1 0 1 0 0 0 0 ⎥
⎢ ⎥
⎢ 1 0 −1 0 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 1 0 −1 0 0 0 0 ⎥
=⎢ ⎥
⎢ 0 0 0 0 1 0 1 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 0 1 0 1 ⎥
⎢ ⎥
⎣ 0 0 0 0 1 0 −1 0 ⎦
0 0 0 0 0 1 0 −1
⎡ ⎤
1 1 0 0 0 0 0 0
⎢ 1 −1 0 0 0 0 0 0 ⎥
⎡ ⎤ ⎢ ⎥
1 0 0 0 ⎢ 0 0 1 1 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 1 0 0 ⎥ 1 1 ⎢ 0 0 1 −1 0 0 0 0 ⎥
G3 = ⎢ ⎥⊗ =⎢ ⎥
⎣ 0 0 1 0 ⎦ 1 −1 ⎢ 0 0 0 0 1 1 0 0 ⎥
⎢ ⎥
0 0 0 1 ⎢ 0 0 0 0 1 −1 0 0 ⎥
⎢ ⎥
⎣ 0 0 0 0 0 0 1 1 ⎦
0 0 0 0 0 0 1 −1
We can then calculate R̂ by three matrix multiplications:
R̂ = [0.8; 0.0; −1.4; 1.8; −1.0; −1.8; 0.0; 3.2] · G1 · G2 · G3

= [0.8; 0.0; −1.4; 1.8; 1.8; −4.6; 1.6; 1.6]
The matrices Gi are sparse matrices having only two non-null elements per
column. Moreover, factorization of the Hadamard matrix HN has a length
proportional to log(N ). The total computation cost is therefore in N log(N ) by
the fast transform, instead of N 2 by the direct method. Figure 8.6 presents the
graph of the calculations for the fast Hadamard transform in the case N = 8.
In terms of error correcting performance, article [8.6] shows that we obtain
the same performance as the Chase-Pyndiah algorithm for two times fewer iter-
ations.
8.4.5 Other soft input decoding algorithms

There are still other algorithms for decoding block codes with weighted input and
output. One of the oldest is Farrell et al.’s algorithm [8.5] for decoding by local
Figure 8.6 – Graph of the computation flow for the fast Hadamard transform N = 8.
optimization. This technique involves posing the problem of searching for the
most probable word as a problem of minimizing a global cost function, having as
its variables the information bits of the word. This optimization must be done
theoretically in integers, which makes the problem very difficult. The technique
used is to replace these integer variables by real variables. The problem is then
no longer strictly equivalent to the initial problem but it becomes possible to use
classical non-linear optimization techniques, like the gradient method, to search
for the minimum of the cost function.
Another approach introduced by Kschischang and Sorokine [8.17] involves
using the fact that most block codes used in turbo product codes are in fact trellis
codes. It is then possible to use classical(MAP, Max-Log-MAP, . . . ) algorithms
for decoding trellis codes to obtain the soft decisions of the decoder. This type
of algorithm is an application in the case of the turbo product codes of the
maximum likelihood decoding algorithms of Wolf [8.19].
Another more recently invented algorithm is Kötter and Vardy’s [8.10]. This
algorithm is only applicable to very specific codes, mainly including Reed-
Solomon codes and algebraic codes. It is based on the Sudan algorithm [8.18]
which is capable of determining the codewords in a neighbourhood close to the
received word. It is then possible to use conventional weighting techniques. Al-
though the initial version of this algorithm is relatively computation-costly, theo-
retical improvements have been made and this algorithm is even more promising
since Reed-Solomon codes are widely used in the domain of telecommunica-
tions. There are also decoding algorithms based on sub-codes [8.12]. Although
slightly complex at implementation level, these algorithms provide excellent per-
formance. Finally, recent studies have shown that decoding algorithms based on
belief propagation can be applied to linear codes in general [8.9].
8.5 Implantation of the Chase-Pyndiah algo-

rithm
The Fang-Battail algorithm requires the systematization of the parity matrix
which is a very costly operation. The Hartmann-Nazarov algorithm has been
implemented on DSP [8.6], but the precision necessary for the computation of
the metrics is too great for it to be envisaged to implement the decoder on a
reasonably sized ASIC. This is the reason why the Chase-Pyndiah algorithm is
the most commonly used decoding algorithm with weighted output for imple-
menting dedicated circuits [8.1], since by a judicious use of memories, it allows
turbo product codes architectures to be elaborated that are adapted to high-rate
transmissions [8.3].
Turbo decoding using the Chase-Pyndiah algorithm alternates weighted de-
coding of the rows and columns. This iterative process leads to the architecture
of Figure 8.7. Between each half-iteration is inserted a phase for reconstructing
the matrix in order to obtain the decoded word.
Figure 8.7 – Global architecture of a turbo product decoder.
Each half-iteration processor takes at its input the channel data as well as
the extrinsic values produced by the previous half-iteration. At the output,
the processor transfers the channel data (to ensure a pipeline operation of the
decoding chain), the hard decisions of the most probable codeword and the ex-
trinsic values calculated by the Chase-Pyndiah algorithm. The architecture of
the processor is illustrated in Figure 8.8. The FIFOs (First-In/First-Out) are
used to synchronize the input and output data of the SISO decoder which has
a certain latency. Small sized FIFOs are generally implemented by rows of D
flip-flops. When size increases, for reasons of hardware space and consumption,
it becomes fairly quickly worthwhile using a RAM memory and two pointers
(one for writing and one for reading), incremented at each new datum, with the
addressing managed circularly. We can also make a remark about the multipli-
cation of the extrinsic values by α. In a conventional implementation, values
Wk are generally integers of reduced size (5 or 6 bits) but the hardware cost
of a real multiplier is prohibitive, and we generally prefer to substitute it by a
simple table.
Figure 8.8 – Internal architecture of the half-iteration processor.
The SISO decoder, described by Figure 8.9, performs the steps of the Chase-
Pyndiah algorithm. The decoder is made up of five parts:
– The module for sequential processing of the data calculates in parallel the
input codeword syndrome and the least reliable positions in the frame.
– The algebraic decoding module performs the algebraic decoding of the

words built from input Rk data and knowledge of the least reliable places.
– The selection module determines the most probable codeword as well as
the closest concurrent word or words.
– The module for calculating the weightings determines the reliability of the
decoded bits.
– The memory module stores the total input weightings that are used to
calculate the weightings.
The module for processing the data receives the sample bits, one after the
other. If the code is cyclic (BCH, for example) calculating the syndrome is
then very simply done by using the factorization of the generator polynomial
following the Hörner scheme. Determining the least reliable positions is often
done by sequentially managing the list of not so reliable positions in a small
local RAM. There are also other solutions that are more economical in size, like
Leiserson’s systolic array, but the gain obtained is small.
Figure 8.9 – Internal architecture of the SISO decoder.
The algebraic decoding module uses the value of the syndrome to determine
the erroneous places in the concurrent vectors. In the case of BCH codes, it is
possible to use the Berlekamp-Massey algorithm or the extended Euclid algo-
rithm to make the correction. It should, however, be noted that this solution
is really only economical for block codes with high correcting power. For codes
with low error correction capability, it is less costly to store the bits to be cor-
rected in a local ROM, for each possible value of the syndrome.
The selection module must sort among the words generated by the algebraic
decoding module to determine the most probable ones. It must therefore calcu-
late the metric of each of these words (which it does sequentially by additions)
and determine, by computation of the minimum value, the most probable among
them (their number is generally limited, for the sake of space).
Finally, the module for calculating the weightings uses the list of concurrent
words chosen above to generate the weightings from the equation of step 6 of
the Chase-Pyndiah algorithm. This module has low complexity since the calcu-
lations to be done are relatively simple and, for each bit, it must keep only two
values sequentially (the smallest metric among the concurrent words having 0 as
the corresponding value for this bit in its binary development, and the smallest
metric for the candidate words having 1 as their binary value). This module also
contains value β. In the case of an FPGA (Field Programmable Gate Array) im-
plantation, all the iterations are generally executed on the same hardware which
is re-used from half-iteration to half-iteration. We must therefore anticipate a
procedure for loading value β.
Bibliography
[8.1] P. Adde, R. Pyndiah, and O. Raoul. Performance and complexity of
block turbo decoder circuits. In Proceedings of Third International Con-
ference on Electronics, Circuits and Systems, (ICECS’96), pages 172–175,
Rhodes, Greece, 1996.

GENEVA, May 1993.
[8.3] J. Cuevas, P. Adde, and S. Kerouedan. Information theory turbo decoding
of product codes for gigabit per second applications and beyond. European
Transactions on Telecommunications, 17(1):345–55, Jan. 2006.
[8.4] P. Elias. Error-free coding. IEEE Transactions on Information Theory,
4(4):29–39, Sept. 1954.
[8.5] K. Farrell, L. Rudolph, C. Hartmann, and L. Nielsen. Decoding by local
optimization. IEEE Transactions on Information Theory, 29(5):740–743,
Sept. 1983.
[8.6] A. Goalic, K. Cavalec-Amis, and V. Kerbaol. Real-time turbo decod-

ing of block turbo codes using the hartmann-nazarov algorithm on the
dsp texas tms320c6201. In Proceedings of International Conference on
Communications (ICC’2002), New-York, 2002.
[8.7] J. Hagenauer, E. Offer, and L. Papke. Iterative decoding of binary block
and convolutional codes. IEEE Transactions on Information Theory, IT-
42:429–445, 1996.
[8.8] C.R.P. Hartmann and L.D. Rudolph. An optimum symbol-by-symbol

decoding rule for linear codes. IEEE Transactions on Inforamtion Theory,
IT-22:514–517, 1974.
[8.9] A. Kothiyal and O.Y. Takeshita. A comparison of adaptative belief prop-
agation and the best graph algorithm for the decoding of linear block
codes. In Proceedings of IEEE International Symposium on Information
Theory, pages 724–728, 2005.
[8.10] R. Kotter and A. Vardy. Algebraic soft-decision decoding of reed-solomon
codes. In Proceedings og IEEE International Symposium on Information
Theory, page 61, 2000.
[8.11] M. Lee and M. Kaveh. Fast hadamard transform based on a simple ma-
trix factorization. IEEE Transactions on Acoustics, Speech, and Signal
Processing, 34:1666–1667, 1986.
[8.12] C.Y. Liu and S. Lin. Turbo encoding and decoding of reed-solomon codes
through binary decomposition and self-concatenation. IEEE Transactions
on Communications, 52(9):1484–1493, Sept. 2004.
[8.13] L.E. Nazarov and V.M. Smolyaninov. Use of fast walsh-hadamard trans-
formation for optimal symbol-by-symbol binary block-code decoding.
Electronics Letters, 34:261–262, 1998.
[8.14] R. Pyndiah, A. Glavieux, A. Picart, and S. Jacq. Near optimum decod-

ing of product codes. In Proceedings of IEEE Global Communications
Conference (GLOBECOM’94), San Francisco, 1994.
[8.15] S. M. Reddy and J. P. Robinson. Random error and burst corrections
by iterated codes. IEEE Transactions on Information Theory, 18(1):182–
185, 1972.
[8.16] N. Sendrier. Codes correcteurs à haut pouvoir de correction. PhD thesis,

Université Paris VI, 1991.
[8.17] V. Sorokine, F.R. Kschischang, and V. Durand. Trellis based decoding of
binary linear block codes. Lecture Notes in Computer Science, 793:270–
286, 1994.
[8.18] M. Sudan. Decoding of reed-solomon codes beyond the error correction
bound. Journal of Complexity, 12:180–193, 1997.
[8.19] J. Wolf. Efficient maximum likelihood decoding of linear block codes using
a trellis. IEEE Transactions on Information Theory, 24:76–80, 1978.
Chapter 9
LDPC codes
Low Density Parity Check (LDPC) codes make up a class of block codes that
are characterized by a sparse parity check matrix. They were first described
in Gallager’s thesis at the beginning of the 60s [9.21]. Apart from the hard
input decoding of LDPC codes, this thesis proposed iterative decoding based
on belief propagation (BP). This work was forgotten for 30 years. Only a few
rare studies referred to it during this dormant period, in particular, Tanner’s
which proposed a generalization of the Gallager codes and a bipartite graph
[9.53] representation.
After the invention of turbo codes, LDPC codes were rediscovered in the
middle of the 90s by MacKay et al. [9.39], Wilberg [9.64] and Sipser et al.
[9.52]. Since then, considerable progress concerning the rules for building good
LDPC codes, and coding and decoding techniques, have enabled LDPC codes
to be used, like turbo codes, in practical applications.
This chapter gives an overview of the encoding and decoding of LDPC codes,
and some considerations about hardware implementations.
9.1 Principle of LDPC codes

LDPC codes are codes built from the simplest elementary code: the single parity
check code. We therefore begin this chapter by describing the single parity check
code and its soft in soft out decoding before dealing with the construction of
LDPC codes.
9.1.1 Parity check code

Definition
A parity equation, represented graphically by Figure 9.1, is an equation linking n
binary data to each other by the exclusive or, denoted ⊕ operator. It is satisfied
if the total number of 1s in the equation is even or null.
Figure 9.1 – Graphic representation of a parity equation.
The circles represent the binary data ci , also called variables. The rectangle con-
taining the exclusive or operator represents the parity equation (also called the
parity constraint, or parity). The links between the variables and the operator
indicate the variables involved in the parity equation.
Parity code with three bits

We consider that the binary variables c1 , c2 and c3 are linked by the parity
constraint c1 ⊕ c2 ⊕ c3 = 0, and that they make up the codeword (c1 , c2 , c3 ). We
assume that we know the log likelihood ratio (LLR) L(c1 ) and L(c2 ) of variables
c1 and c2 : what can we then say about the LLR L(c3 ) of variable c3 ?
We recall that L(cj ) is defined by the equation:

Pr (cj = 1)
L (cj ) = ln (9.1)
Pr (cj = 0)
There are two codewords in which bit c3 is equal to 0: codewords (0,0,0) and
(1,1,0). Similarly, there are two codewords in which bit c3 is equal to 1: code-
words (1,0,1) and (0,1,1). We deduce from this the following two equations in
the probability domain:
/
Pr(c3 = 1) = Pr(c1 = 1) × Pr(c2 = 0) + Pr(c1 = 0) × Pr(c2 = 1)
(9.2)
Pr(c3 = 0) = Pr(c1 = 0) × Pr(c2 = 0) + Pr(c1 = 1) × Pr(c2 = 1)
Using the expression of each probability according to the likelihood ratio func-
tion, deduced from Equation (9.1):
⎧
⎪ exp(L(cj ))
⎨ Pr(cj = 1) = 1+exp(L(cj ))
⎪
⎩ Pr(cj = 0) = 1 − Pr(cj = 1) = 1
1+exp(L(cj ))
9. LDPC codes 299
we have:

1 + exp(L(c2 ) + L(c1 ))
L(c3 ) = ln L(c1 ) ⊕ L(c2 ) (9.3)
exp(L(c2 )) + exp(L(c1 ))
Equation (9.3) enables us to define the switching operator ⊕ between the two
LLRs of the variables c1 and c2 .
exp(x)−1
Applying function tanh (x/2) = exp(x)+1 to Equation (9.3), the latter be-
comes:
tanh L(c2 3 ) = exp(L(c 0 ))−1

exp(L(c0 ))+1
× exp(L(c 1 ))−1
exp(L(c1 ))+1
'
1
(9.4)
L(cj )
= tanh 2
j=0
It is practical (and frequent) to separate the processing of the sign and the
magnitude in Equation (9.4) which can then be replaced by the following two
equations:
5
1
sgn (L(c3 )) = sgn (L(cj )) (9.5)
j=0
5
1
|L(c3 )| |L(cj )|
tanh = tanh (9.6)
2 j=0
2
where the sign function sgn(x) is such that sgn(x) = +1 if x ≥ 0 and sgn(x) = −1
otherwise.
Processing the magnitude given by Equation (9.6) can be simplified by taking
the inverse of the logarithm of each of the terms of the equation, which gives:
⎛ ⎞

|L(c3 )| = f −1 ⎝ f (|L(cj )|)⎠ (9.7)
j=1,2
where function f, satisfying f −1 (x) = f (x), is defined by:
f (x) = ln (tanh (x/2)) (9.8)
These different aspects of the computation of function ⊕ will be developed in

the architecture part of this chapter.
Expression (9.6) in fact corresponds to the computation of (9.3) in the Fourier
domain. Finally, there is also a third writing of the LLR of variable c2 [9.65, 9.23]:
L(c3 ) = sign (L(c1 )) sign (L(c2 )) min (|L(c1 )| , |L(c2 )|)

− ln (1 + exp (− |L(c1 ) − L(c2 )|)) (9.9)
+ ln (1 + exp (− |L(c1 ) + L(c2 )|))
This other expression of operator ⊕ can easily be processed by using a look-up

table for function g defined by:
g(x) = ln (1 + exp (− |x|)) (9.10)
Practical example
Let us assume that Pr(c1 = 1) = 0.8 and Pr(c2 = 1) = 0.1. We then have,
Pr(c1 = 0) = 0.2 et Pr(c2 = 0) = 0.9
It is therefore more probable that c1 = 1 and c2 = 0. A direct application of
Equation (9.2) then gives
Pr(c3 = 0) = 0.26 and Pr(c3 = 1) = 0.74
c3 is therefore more probably equal to 1, which intuitively is justified since
the number of 1s belonging to the parity check equation must be even. Using
Equation (9.3) gives
L(c3 ) = L(c1 ) ⊕ L(c2 ) = (−1.386) ⊕ (2.197) = −1.045
that is, Pr(c3 = 0) = 0.26 again. We find the same result again using (9.7) and
(9.9).
Parity check code with n bits

We can now proceed to the parity check equation with n bits. We consider that
the binary variables c1 , · · · , cn are linked by the parity constraint c1 ⊕· · ·⊕cn = 0
and that they make up the codeword (c1 , · · · , cn ). The LLR of the variables
{cj }j=1..n,j=i is assumed to be known and we search for the LLR of variable ci .
It is then simple to generalize the equations obtained for the parity code with 3
bits. Thus, taking the operator defined by (9.2) again:
E
L(ci ) = L(c1 ) ⊕ L(c2 ) ⊕ · · · ⊕ L(cj=i ) ⊕ · · · ⊕ L(cn ) = L(cj ) (9.11)
j=i
Similarly, the hyperbolic tangent rule is expressed by:

5
L(ci ) L(cj )
tanh = tanh (9.12)
2 2
j=i
or, separating the sign and the magnitude:

5
sgn (L(ci )) = sgn (L(cj )) (9.13)
j=i
⎛ ⎞

|L(ci )| = f −1 ⎝ f (|L(cj )|)⎠ (9.14)
j=i
where f is defined by Equation (9.8).

9. LDPC codes 301
9.1.2 Definition of an LDPC code

Linear block codes
Linear block codes (see Chapter 4) can be defined by a parity check matrix H
of size m × n, where m = n − k. This matrix can be seen as a linear system of
m parity check equations. The words c of the code defined by H are the binary
words whose n bits simultaneously satisfy the m parity check equations. This
system of linear equations is represented graphically in Figure 9.2 for the case
of the Hamming binary block code of length n = 7.
Figure 9.2 – Graphic representation of a block code: example of a Hamming code of

length 7.
Such a representation is called the bipartite graph of the code. In this graph,
branches link two different classes of nodes to each other:
• The first class of nodes called variable nodes, correspond to the bits of the
codewords (cj , j ∈ {1, · · · , n}), and therefore to the columns of H.
• The second class of nodes, called parity check nodes, correspond to the
parity check equations (ep , p ∈ {1, · · · , m}), and therefore to the rows of
H.
Thus, to each branch linking a variable node cj to a parity check node ep corre-
sponds the 1 that is situated at the intersection of the j-th column and the p-th
row of the parity check matrix.
By convention, we denote P (j) (respectively J(p)) all the indices of the
parity nodes (respectively variable nodes) connected to the variable with index
j (respectively to the parity with index p). We denote by P (j)\p (respectively
J(p)\j) all the P (j) not having index p (respectively, all the J(p) not having
index j). Thus, in the example of Figure 9.2, we have
P (5) = {1, 3} and J(1)\5 = {4, 6, 7}
A cycle on a bipartite graph is a path on the graph which makes it possible to

leave a node and to return to this same node without passing twice through the
same branch. The size of a cycle is given by the number of branches contained
in the cycle. The graph being bipartite, the size of the cycles is even. The
size of the shortest cycle in a graph is called the girth. The presence of cycles
in the graph may degrade the decoding performance by a phenomenon of self-
confirmation during the propagation of the messages. Figure 9.3 illustrates two
cycles of sizes 4 and 6.
Figure 9.3 – Cycles in a bipartite graph.
Low density parity check codes

LDPC codes are linear block codes, the term low density coming from the fact
that parity check matrix H contains a low number of non null values: it is
a sparse matrix. In the particular case of binary LDPC codes studied here,
the parity check matrix contains a small number of 1s. In other words, the
associated bipartite graph contains a small number of branches. The adjective
"low" mathematically means that when the length n of a codeword increases,
the number of 1s in the matrix increases in O(n) (compared to an increase in
O(n2 ) of the number of elements of the matrix if the rate remains fixed).
The class of LDPC codes generates a very large number of codes. It is
convenient to divide them into two sub-classes:
• regular LDPC codes

• irregular LDPC codes
An LDPC code is said to be regular in the particular case where parity

check matrix H contains a constant number dc of 1s in each row, and a constant
number dv of 1s in each column. We then say that the variables are of degree
dv and that the parities are of degree dc . The code is denoted a (dv ,dc ) regular
LDPC code. For example, the parity check matrix of a regular LDPC code
(3,6) contains only 3 non zero values in each column, and 6 non zero values in
each row. Figure 9.4 presents an example of a regular (3,6) code of size n = 256
obtained by drawing randomly the non zero positions. Out of the 256x128 inputs
of the matrix, only 3x256 are non zero, that is, around 2.3%. This percentage
tends towards 0 if the size of the code, for a fixed rate, tends towards infinity.
The irregularity profile of
the variables of an irregular LDPC code is defined
by the polynomial λ(x) = λj xj−1 where coefficient λj is equal to the ratio
between the accumulated number of 1s in the columns (or variable) of degree j
9. LDPC codes 303
120
100
80
60
40
20
50 100 150 200 250
Figure 9.4 – Parity check matrix of a regular (3,6) LDPC code of size n = 256 and
rate R = 0, 5.
and the total number E of 1s in matrix H. For example, λ(x) = 0, 2x4 + 0, 8x3
indicates a code where 20% of the 1s are associated with variables ofdegree 5
and 80% with variables of degree 4. Note that, by definition, λ(1) = λj = 1.
Moreover, the proportion of variables of degree j in the matrix is given by
λj /j
λ̄j =
λk /k
k
Symmetrically, the
irregularity profile of the parities is represented by the poly-
nomial ρ(x) = ρp xp−1 , coefficient ρp being equal to the ratio between the
accumulated number of 1s in the rows (or parity) ofdegree p and the total num-
ber of 1s denoted E. Similarly, we obtain ρ(1) = ρj = 1. The proportion ρ̄p
of columns of degree p in matrix H is given by
ρp /p
ρ̄p =
ρk /k
k
Irregular codes have more degrees of freedom than regular codes and it is
thus possible to optimize them more efficiently: their asymptotic performance
is better than that of regular codes.
Coding rate
Consider a parity equation of degree dc . It is possible to arbitrarily fix the
values of the dc − 1 first bits; only the last bit is constrained and corresponds
to the redundancy. Thus, in a parity matrix H of size (m,n), each of the m
rows corresponds to 1 redundancy bit. If the m rows of H are independent, the
code then has m redundancy bits. The total number of bits of the code being
n, the number of information bits is then k = n − m and the coding rate is

R = (n − m)/n = 1 − m/n. Note that in the case where the m rows are not
independent (for example, two identical rows), the number of constrained bits
is lower than m. We then have R > 1 − m/n.
In the case of a regular LDPC code (dv ,dc ), each of the m rows has dc non zero
values, that is, a total of E = mdc non zero values in matrix H. Symetrically,
each of the n columns contains dv non zero values. We deduce from this that E
satisfies E = ndv = mdc , that is, m/n = dv /dc . The rate of such a code then
satisfies R (1 − dv /dc ).
In the case of an irregular code, the expression of the rate is generalized,
taking into account each degree weighted by:

p ρp /p
R1− (9.15)
j λj /j
Equality is reached if matrix H is of rank m.
9.1.3 Encoding
Encoding an LDPC code can turn out to be relatively complex if matrix H does
not have a particular structure. There exist generic encoding solutions, including
an algorithm with complexity in O(n), requiring complex preprocessing on the
matrix H. Another solution involves directly building matrix H so as to obtain
a systematic code very simple to encode. It is this solution, in particular, that
was adopted for the standard DVB-S2 code for digital television transmission
by satellite.
Generic encoding
Encoding with a generator matrix
LDPC codes being linear block codes, the coding can be done via the gen-
erator matrix G of size k × n of the code, such as defined in Chapter 4. As
we have seen, LDPC codes are defined from their parity check matrix H, which
is generally not systematic. A transformation of H into a systematic matrix
Hsys is possible, for example with the Gaussian elimination algorithm. This
relatively simple technique, however, has a major drawback: the generator ma-
trix Gsys of the systematic
- . code is generally not sparse. The coding complexity
increases rapidly in O n2 , which makes this operation too complex for usual
length codes.
Coding with linear complexity
Richardson et al. [9.49] proposed a solution enabling quasi-linear complexity
encoding, as well as greedy algorithms making it possible to preprocess parity
check matrix H. The aim of the preprocessing is to put H as close as possible
to a lower triangular form, as illustrated in Figure 9.5, using only permutations
9. LDPC codes 305
Figure 9.5 – Representation in the lower pseudo-triangular form of the parity check
matrix H.
of rows or columns. This matrix is made up of 6 sparse sub-matrices, denoted

A,B,C,D,E and T , the latter being a lower triangular sub-matrix. Once the
preprocessing of H is finished, the encoding principle is based on solving the
system represented by the following matrix equation:
cHT = 0 (9.16)
The codeword searched for is decomposed into three parts: c = (d, r1 , r2 ),

where d is the systematic part that is known and where the redundancy bits
searched for are split into two vectors r1 and r2 , of respective −
size g and n
I 0
k − g. After multiplication on the right-hand side by matrix ,
−ET −1 I
Equation (9.16) becomes:
Adt + Brt1 + T rt2 = 0 (9.17)

- . - .
−ET −1A + C dt + −ET −1B + D rt1 = 0 (9.18)
−1
Equation (9.18) enables r1 to be found by inverting Φ = −ET B + D. Then
Equation (9.17) enables r2 to be found by inverting T . Many time-consuming
operations can be done once and for all during preprocessing. All the operations
repeated- during the encoding
. - in O.(n) except the multiplica-
have a complexity
tion of −ET −1A + C dt by square matrix −Φ−1 of .size g × g which after
-
inversion is no longer sparse, hence a complexity in O g 2 . It is shown in [9.49]
that we can obtain a value of g equal - to. a small fraction of n: g = αn where α
is a sufficiently low coefficient for O g 2 << O (n) for values of n up to 105 .
Specific constructions
Coding with a sparse generator matrix
One idea proposed by Oenning et al. [9.45] involves directly building a sparse
systematic generator matrix, so the coding is performed by simple multiplication
and the parity check matrix remains sparse. These codes are called Low-Density
Generator-Matrix (LDGM) codes. Their performance is however poor [9.36],
even if it is possible to optimize their construction [9.22] and lower the error
floor.
Encoding by solving the system cH T = 0 obtained by substitution
Mackay et al. [9.40] propose to constrain the parity matrix so that it is
composed of the three sub-matrices A, B and C arranged as in Figure 9.6.
Figure 9.6 – Specific construction of parity check matrix H facilitating the encoding.
Systematic encoding is performed by solving Equation (9.16), which means

solving (n − k − δ) equations by substitution. Each row of the parity matrix
containing a low number of 1s, this operation is linear with n. The remaining
δ equations are solved by inversion of matrix C defined in Figure 9.4. That
leads
- .to multiplication by a non-sparse matrix and therefore a complexity in
O δ2 . Bond et al. [9.7] and Hu et al. [9.29] have proposed building parity
check matrices with δ = 0. In [9.25] Haley et al. define a class of codes enabling
equation 9.12 to be solved by an iterative algorithm similar to the one used for
the decoding.
Cyclic coding
The classes of LDPC codes defined by finite geometry or by projective ge-
ometry [9.29, 9.46, 9.34, 9.1, 9.58] enable cyclic or pseudo-cyclic codes to be
obtained (see Section ?? for further details about this type of construction).The
codes thus obtained can be encoded efficiently by using shift registers. In ad-
dition, they offer good properties in terms of the distribution of cycle length
( 6). The main drawback is that the cardinal of these classes of code is rela-
9. LDPC codes 307
tively small. These classes therefore offer only a very limited number of possible
size – rate – irregularity profile combinations.
Summary
Table 9.1 summarizes the different possible types of coding encountered in the
literature. In practice, the conventional encoding of block codes by a generator
matrix is not used for LDPC codes due to the large the length of the codewords.
The codes obtained by projective or finite geometry cannot be optimized (opti-
mal design of the irregularity profiles). There therefore remain only codes built
to facilitate the encoding by solving the equation cH T = 0 by substitution, such
as the one chosen for the DVB-S2 standard.
Type of encoding Complexity Remarks

Generator ∼ O(n2 ) Not used in practice
matrix
Generic
Pseudo-linear ∼ O(n) A lot of preprocessing
encoding
Solving ∼ O(n) (si Possible loss of
cHT = 0 by δ = 0) performance
Ad hoc substitution
construction Cyclic or ∼ O(n) Limited number of
pseudo-cyclic possible combinations of
the different parameters
Table 9.1 – Summary of the different possible encodings.
Analogy between an LDPC code and a turbo code

Figure 9.7 presents a turbo code in the form of a bipartite graph proposed by
Tanner [9.53], thus showing the very close relation which links the family of
turbo codes and that of LDPC codes.
In the case of a turbo code, the constraints are greater than in the case of
an LDPC code since elementary codes are convolutional codes. But in the same
way as for LDPC codes, a word is a codeword if and only if the two constraints
of the graph are respected. Note here that the degree of the bits of a turbo code
is two for the information bits and one for the redundancy bits.
Similarly, a product code can also √
be represented by a bipartite graph. The
number of constraint nodes is then 2 n (compared with 2 for the turbo code
and n/2 for an LDPC code with rate 0,5) and the latter have an intermediate
complexity between that of the turbo code and that of the LDPC code. Turbo
codes and LDPC codes are thus the two extremities of the spectrum of "com-
posite" codes. The former contain only two very complex constraint nodes, the
latter, very many constraint nodes, each node being made up of the simplest
linear code possible (parity code). Note that from this representation of the
bipartite graph, an infinite number of more or less exotic codes can be built.
It is interesting to note that encoding LDPC codes is tending to be performed
more and more like the encoding of turbo codes (in a serial concatenation).
The precursors were, without doubt, the Repeat Accumulate codes proposed in
[9.16] whose encoding is composed of a repetition code, an interleaver and an
accumulator. These codes are then decoded by an algorithm of the LDPC type
with an adapted schedule. In the literature, we can now find many variants
of this type of encoding, which involves combining elementary encoders and
interleavers.
Figure 9.7 – Representation of a turbo code in the form of a bipartite graph.
The similarity between turbo codes and LDPC codes is even greater than can
be assumed from the representations in the form of bipartite graphs. Indeed,
it is shown in [9.36], [9.44] and in Section 6.2 that it is possible to represent
a turbo code in the form of an LDPC matrix. The resemblance stops there.
Parity check matrix H of a turbo code contains many rectangular patterns (four
1s making a rectangle in matrix H), that is, many cycles of length 4, which make
the algorithms for decoding LDPC codes, to be described below, inefficient.
9.1.4 Decoding LDPC codes

Decoding an LDPC code is done using the same principle as decoding a turbo
code by an iterative algorithm called a belief propagation algorithm. Each vari-
able node sends to the parity nodes with which it is associated a message about
the estimated value of the variable (a priori information). The set of a priori
messages received enables the parity constraint to compute then return the ex-
trinsic information. The successive processing of the variable then parity nodes
make up one iteration. At each iteration, there is therefore a bilateral exchange
of messages between the parity nodes and variable nodes, on the arcs of the
bipartite graph representing the LDPC code. At the level of the receiver, the
9. LDPC codes 309
method for quantifying the sequence received, r, determines the choice of de-
coding algorithm.
Hard input algorithm

Quantization on one bit involves processing only the sign of the samples received.
Hard input decoding algorithms are based on the one proposed by Gallager under
the name of algorithm A [9.21]. These decoders of course offer lower performance
than those of soft input decoders. They are only implemented for very particular
applications like optical communications, for example [9.17]. These algorithms
will not be considered in the remainder of this chapter.
Belief propagation algorithm

When quantization is done on more than one bit, the decoding may use soft
inputs: the a priori probability of the received symbols. In the case of binary
codes and in the logarithm domain, we use the a priori log likelihood ratio
(LLR) of samples rj :

p (rj |cj = 0 )
L (rj |cj ) = ln (9.19)
p (rj |cj = 1 )
where cj is the j-th bit of the codeword and rj = 2cj−1 + bj . In the case of
the additive white Gaussian noise channel, the noise samples bj follow a centred
Gaussian law with variance σ2 , that is:
! $
2
1 (rj − (2cj − 1))
p (rj |cj ) = √ exp (9.20)
2πσ2 2σ 2
Combining (9.19) and (9.20), intrinsic information Ij can be defined :
Δ 2rj
Ij = L (rj |cj ) = − (9.21)
σ2
Each iteration of the BP algorithm is decomposed into two steps:
1. Processing the parities:

⎡ ⎤
5 Lj ,p ⎦
Zj,p = 2 tanh−1 ⎣ tanh (9.22)
2
j ∈J(p)/j
2. Processing the variables:

Lj,p = Ij + Zj ,p (9.23)
p ∈P (j)/p
The iterations are repeated until the maximum number of iterations Nit is
reached. It is possible to stop the iterations before Nit when all the parity
equations are satisfied. This enables either a gain in mean throughput, or a
limit in consumption.
We call Lj the total information or the LLR of bit j. This is the sum of
the intrinsic information Ii and the total extrinsic information Zj which is by
definition the sum of the extrinsic information of branches Zj,p :
Δ
Zj = Zj,p (9.24)
p∈P (j)
We therefore have Lj = Ij + Zj and Equation (9.23) can then be written:
Lj,p = Lj − Zj,p = Ij + Zj − Zj,p (9.25)
The BP algorithm is optimal in the case where the graph of the code does not
contain any cycle: all the schedules1 give the same result. As LDPC codes
involve cycles, their decoding by the BP algorithm can lead to phenomena of
self-confirmation of the messages which degrade the convergence and make the
BP algorithm distinctly sub-optimal. However, these phenomena can be limited
if the cycles are large enough.
The first schedule proposed is called the "flooding schedule" [9.35]. It
involves successively processing all the parities then all the variables.
Flooding schedule algorithm
Initialization :
(0)
1- nit = 0, Zj,p = 0 ∀p ∀j ∈ J(p), Ij = 2yj /σ 2 ∀j
Repeat until nit = Nit or until the system has converged towards a code-
word :
2- nit = nit + 1
3- ∀ j ∈ {1, · · · , n} do: {Computation of the variable towards parity
messages}
(n ) (n −1) (n ) (n )
4- Zj it = Zj,pit and Lj it = Ij + Zj it
p∈P (j)
5- ∀ p ∈ P (j):
(n )
(n −1) (nit ) (n −1)
Lj,pit = Ij + Zj,pit = Ij + Zj − Zj,pit
p ∈P (j)/p
1 By schedules, we mean the order in which each parity and each variable is processed.
9. LDPC codes 311
6- ∀ p ∈ {1, · · · , m} faire : {Computation of the parity towards

variable message}
(n ) F (n )
7- ∀ j ∈ J(p): Zj,pit = Lj,pit
j ∈J(p)/j
(n )
The bits decoded are then estimated by sgn −Lj it .
It is interesting to note that it is possible to modify the algorithm by "order-
ing" the flooding schedule depending on the parity nodes. The latter are then
processed serially, and the algorithm becomes:
(nit +1)
3’- ∀j ∈ {1, · · · , n}: Zj =0
4’- ∀p ∈ {1, · · · , m} do:
5’- Computation of the input messages

(n ) (nit ) (n −1)
∀j ∈ J(p) Lj,pit = Ij + Zj − Zj,pit
6’- Computation of the extrinsic information

(n )
E (n )
∀ j ∈ J(p) Zj,pit = Lj,pit
j ∈J(p)/j
7’- Update for the following iteration

(nit +1) (nit +1) (n )
∀ j ∈ J(p) Zj = Zj + Zj,pit
A similar organization of computations for the variable nodes will be called

"distributed computation" since the computations linked with a variable node
will be distributed during one iteration. In Section 9.2, the different types of
schedule will be detailed then generalized.
It must also be noted that the notion of iteration (the computation of all
the messages of the graph once and only once) is not strict. Thus, Mao et al.
[9.42] proposed a variant of the flooding schedule in order to limit the impact
of the effect of the cycles on the convergence. This variant, called "probabilis-
tic scheduling", involves not processing some variables at each iteration. The
choice of these variables is random and depends on the size of the smallest cycle
associated with this variable: the smaller the cycle, the lower the probability of
processing the variables involved in this cycle. This method limits phenomena
of self-confirmation introduced by short cycles. It enables convergence to be ob-
tained more rapidly than with the flooding schedule. The architectures linked
with this schedule will not be discussed in this chapter.
9.1.5 Random construction of LDPC codes

The construction of an LDPC code (or of a family of LDPC codes) must nat-
urally be done so as to optimize the performance of the code while minimizing
the hardware complexity of the associated decoder.
Building a code remains a delicate problem so we refer the reader wishing to
explore the subject further to the references given in this chapter. The problem
of building an LDPC code adapted to decoding hardware will be dealt with in
Section 9.2.
Optimizing an LDPC code is carried out in three steps:
• a priori optimization of the irregularity profiles of the parity and variable
nodes;
• construction of matrices H of an adequate size respecting the irregularity
profile and maximizing the length of the cycles;
• if necessary, selection or rejection of the codes using the minimum distance

criterion or the performance computed by simulation.
Optimization of irregularity profiles

We make the hypothesis of codes with infinite size and an infinite number of
iterations. Indeed, this enables optimization of their asymptotic characteristics
(irregularity profile, rate) as a function of the channel targeted. Two techniques
exist: the density evolution algorithm and its Gaussian approximation, and the
extrinsic information transfer chart.
The density evolution algorithm was proposed by Richardson [9.48]. This
algorithm calculates the probability density of messages Lj,p and Zj,p after each
new iteration. The algorithm is initialized with the probability density of the
input samples, which depends on the level of noise σ2 of the channel. Using
this algorithm enables us to know the maximum value of σ 2 below which the
algorithm converges, that is, such that the error probability is lower than a given
threshold. It is also possible to determine by linear programming an irregularity
profile that gives the lowest possible threshold.
A simplification of the density evolution algorithm proposed by Chung et
al. [9.14, 9.13], is obtained by approximating the real probability densities by
Gaussian densities. The interest of Gaussian density approximation is that it
suffices to calculate the evolution of a single parameter by making the hypothesis
that these Gaussian densities are consistent, that is, the variance is equal to
two times the mean. Indeed, assuming that the "all zero"word was sent, at
initialization we have (nit = 0):
(0) 2rj - .
Lj,p = − with rj ∼ N −1, σ 2 (9.26)
σ2
9. LDPC codes 313

(0) 2 4
therefore Lj,p ∼ N ,
σ2 σ2
We denote:
(0)
mj = σ22 the average of the consistent Gaussian probability density of
variable cj of degree dv sent to parities ep of degree dc which are connected
to it,
(nit ) (n )
μp the mean of messages Zj,pit .
(nit )
To follow the evolution of the average mj during the iterations nit , it then
suffices to take the mathematical expectation of Equations (9.22) and (9.23)
relative to the variable and parity processing , which gives:
(n )
(n )
dc −1
with Ψ (x) = E [tanh (x/2)] , x ∼ N (m, 2m) (9.27)
it it
Ψ μp = Ψ mj
(nit +1)2
mj + (dv − 1) μp(nit )
= (9.28)
σ2
Thus for a regular (dv , dc ) LDPC code and for a given noise with variance
σ 2 , Equations (9.27) and (9.28) enable us, by an iterative computation, to know
if the mean of the messages tends towards infinity or not. If such is the case,
it is possible to decode without errors with a codeword of infinite size and an
infinite number of iterations. In the case of an irregular code, it suffices to make
the weighted mean on the different degrees of Equations (9.27) and (9.28).
The maximum value of σ for which the mean tends towards infinity, and
therefore for which the error probability tends towards 0, is the threshold of
the code. For example, the threshold of a regular code (3,6), obtained with the
density evolution algorithm, is σmax = 0.8809 [9.13], which corresponds to a
Eb
minimum signal to noise ratio of N 0 min
= 1.1dB.
Another technique derived from extrinsic information transfer (EXIT)
charts2 proposed by ten Brink [9.55, 9.56] enables the irregularity profiles to
be optimized. Whereas the density evolution algorithm is interested in the evo-
lution of the probability densities of the messages during the iterations, these
charts are interested in the transfer of mutual information between the input
and the output of the decoders of the constituent codes [9.56]. The principle of
these charts has also been used with parameters other than mutual information,
like the signal to noise ratio or error probability [9.3, 9.2]. It has also been
applied to other types of channels [9.15].
2 The principle of building EXIT charts is described in Section 7.6.3
Optimization of cycle size

Optimization of the irregularity profiles being asymptotic, we now have to build
a parity check matrix of finite size. This phase can be performed randomly:
we draw the non zero inputs of the parity check matrix at random, respecting
as far as possible the irregularity profile of the nodes. It is also possible to
build codes by randomly drawing permutations of an elementary matrix which
are then concatenated. Another way to build LDPC codes is the deterministic
construction of a matrix (finite and projective geometry).
In all cases, we must pay attention to the cycles present in the graph of
the code. The belief propagation decoding algorithm assumes that the cycles
that would deteriorate the independence of the messages entering a node do not
exist. In practice, the presence of cycles in the graph is inevitable, but if they are
large enough, the independence of the messages remains a good approximation.
Building good LDPC codes must therefore ensure the absence of smaller cycles,
those of size 4. Very many solutions are proposed in the literature to build
LDPC codes. For example, Campello et al. [9.9] propose optimizing the size
of the minimum cycle for a given rate. Hu et al. [9.65] suggest building the
graph branch by branch in order to avoid at maximum the lowest cycle sizes
(Progressive Edge Geometry or PEG). Zhang et al. [9.67] build LDPC codes
whose smallest cycles are size 12, 16 or 20, but the variables are only degree 2.
Tian et al. [9.57] use the fact that not all the small size cycles have the same
influence and omit only those that are the most penalizing.
Selecting the code by the impulse method

The decoding performance using the belief propagation algorithm is improved
by avoiding small sized cycles. But it is also important to have "good" error
correcting codes, that is to say, those that have a large minimum distance. The
impulse method was first proposed by Berrou et al. [9.4, 9.5] to evaluate the
minimum Hamming distance of a turbo code. It was then adapted to the case
of LDPC codes by Hu et al. [9.26]. It thus enables us to simply verify that
the minimum distance of the code designed is sufficient to reach the error rate
targeted for the application required.
Selecting the code by simulation

Two codes of the same size and the same rate, built with the same irregular-
ity profiles, not having a short cycle and having the same minimum distance
can nevertheless have a fairly different performance. These differences can be
explained by two phenomena: the existence of "parasitic" fixed points intro-
duced by the sub-optimality of the iterative decoding algorithm which increase
the binary error rate in relation to the theoretical value [9.33]. The number of
codewords with minimum distance also influences the performance of the code.
9. LDPC codes 315
Figure 9.8 shows the performance of an LDPC code for different sizes and differ-
ent rates in the case of a DVB-S2 decoder implemented on an Altera Stratix80
FPGA.
LDPC codes therefore have an excellent theoretical performance. This must
however be translated by simplicity in their hardware implementation to enable
these codes to be used in practice. That is why particular attention must be
paid to LDPC decoder architectures and implementations.
R=1/2
R=3/4
R=4/5
0.1 R=5/6
R=8/9
R=9/10
0.01
0.001
FER
0.0001
1e05
1e06
1e07
1 2 3 4 5
Eb/No
Figure 9.8 – Packet error rate (or Frame error rate, FER) obtained for codeword sizes
of 64 kbits and different rates of the DVB-S2 standard (50 iterations, fixed point).
With the permission of TurboConcept S.A.S, France.
9.1.6 Some geometrical constructions of LDPC codes

To complete the random constructions presented above, we list below some
deterministic constructions leaving much less room, if any, for random ones.
Cayley / Ramanujan constructions

Margulis [9.43] was the first to propose algebraic LDPC codes. Then Rosenthal
and Votonbel [9.50, 9.51] extended these results to obtain high expansion factor
graphs with high girths, using Ramanujan graphs instead of Cayley graphs.
Some drawbacks were raised by MacKay and Postol [9.38] about these codes
(error floor and low minimum distance for some sets of parameters).
Figure 9.9 – Simple illustration of finite geometry
Kou, Lin and Fossorier’s Euclidian / Projective Geometry LDPC

These LDPC codes are built on finite geometry [9.34]. Finite geometry is based
on n points and m rows, such that each row contains k points and each point
is on j lines. Two lines are either parallel or they have only one common point.
A parity check matrix H = (hij ) can be built by assuming that hij = 1 if and
only if the i-th row contains the j-th point. Of course, j and k have to be small
compared to m and n in order to build an LDPC code that is called type I. An
example of simple finite geometry is illustrated in Figure 9.9.
These codes are cyclic. When considering the transposed version of H, we
can obtain quasi-cyclic LDPC codes which are referred to as type II codes.
Two kinds of finite geometry are used: Euclidean geometry (EG) and projective
geometry (PG).
Constructions based on permutation matrices

Tanner et al. have proposed an algebraic construction of LDPC parity-check
matrices based on an idea of Tanner [9.54]. A regular (j, k) code of length
n = kp and m = jp parity checks can be obtained, assuming that p is a prime
number and j, k are chosen among the prime factors of p − 1. The structure of
the parity check matrix is:
⎛ ⎞
I1 Ia Ia2 ··· Iak−1
⎜ Ib Iab Ia2 b ··· Iak−1 b ⎟
⎜ ⎟
H=⎜ . . . . .. ⎟
⎝ .. .. .. .. . ⎠
Ibj−1 Iabj−1 Ia2 bj−1 · · · Iak−1 bj−1
where Ix is the identity matrix whose columns have been right-shifted x times,
and where a and b are two non-zero elements in Fp of order j and k respectively.
9. LDPC codes 317
Matrices based on Pseudo random generators

An important drawback of random constructions is that the parity check ma-
trix has to be saved in memory, which takes up a lot of room for long codes.
Prabhakar and Narayanan [9.47] found an interesting solution to circumvent this
issue by using linear congruential sequences to design the parity-check matrix.
Hence, after a non-complex computation, the generator outputs the address
of the non-zero entries of the matrix. This solution has been implemented by
Verdier et al. [9.62]. Girths of at least 6 can be obtained by correctly choosing
the parameters of the generator.
Array-based LDPC
Array codes are two-dimensional codes that have been proposed for detecting
and correcting burst errors [9.6]. When viewed as binary codes, the parity check
matrix of array codes exhibit sparseness, which can be exploited for decoding
them as LDPC codes using the BP algorithm [9.19]. Therefore, array codes
provide the framework for defining a family of LDPC codes that lend themselves
to deterministic constructions [9.18]. The parity check matrix of an array-based
LDPC code is:
⎛ ⎞
I I I ··· I
⎜ I α α2 ··· αk−1 ⎟
⎜ ⎟
H = ⎜ .. .. .. . . ⎟
⎝ . . . . . .
. ⎠
I α j−1
α 2(j−1)
··· α (j−1)(k−1)
where I is the p × p identity matrix, p being an odd prime number, α is the I

matrix whose rows have been shifted once, and j, k ≥ p.
BIBDs, Latin rectangles

A block design is an incidence system [9.63] (v, k, λ, r, b) in which a set X of v
points is partitioned into a family A of b subsets (blocks) in such a way that
any two points determine λ blocks with k points in each block, and each point
is contained in r different blocks. It is also generally required that k < v, which
leads to a balanced incomplete block design (BIBD) of the LDPC code. The
five parameters are not independent, but satisfy the two relations
vr = bk
λ(v − 1) = r(k − 1)
A BIBD is therefore commonly simply written as (v, k, λ), since b and r are
given in terms of v, k, and λ by
v(v − 1)λ λ(v − 1)
b= r=
k(k − 1) k−1
A BIBD is said to be symmetric if b = v (or, equivalently, r = k).

These constructions have been widely studied in the literature. For example,
MacKay and Davey [9.37] first proposed the use of Steiner triple systems to
design short LDPC codes. Johnson and Weller [9.30, 9.31], and also B. Vasic
[9.60], presented a family of LDPC codes based on Kirkman triple systems.
A design based on anti-Pasch Steiner systems is also presented in [9.59], and
mutually orthogonal Latin rectangles (MOLR) are used in [9.61].
9.2 Architecture for decoding LDPC codes for

the Gaussian channel
When the belief propagation algorithm is implemented, the general architecture
of decoders for LDPC codes can be performed with the help of generic node pro-
cessors (GNP) modelling either the parity processing, or the variable processing.
This section describes the different possible implementations of these processors
after analysing the decoding complexity of LDPC codes. The different possibili-
ties for controlling this GNP-based architecture enables us to define three classes
of schedule for the belief propagation algorithm: a two-pass schedule, a "verti-
cal" schedule and a "horizontal" schedule. This original, unified presentation of
the architectures of decoders for LDPC codes enables us to cover many existing
architectures published so far, and to synthesize innovatory architectures.
9.2.1 Analysis of the complexity

The decoding complexity of LDPC codes is directly linked with the number of
branches in the bipartite graph of the code, or with the number of 1s in the
parity check matrix. The iterative decoding belief propagation algorithm has
two steps. At each step, we have to calculate information Lj,p or Zj,p which is
associated with the branch linking variable j to parity p. Let us denote B the
number of branch in the bipartite graph of the LDPC code. For example, in the
case of a regular code (dv , dc ) of size n, the number of branches B is given by:
B = dv n = dc m (9.29)
The computing power Pc necessary to decode LDPC codes is then defined

as the number of branches to process per clock cycle. This parameter depends
on:
• the number k of information bits to transmit per codeword,
• the number of branches B,
• the data rate D of information desired,
• the maximum number of iterations Nit ,
9. LDPC codes 319
• the clock frequency fclk .
In one second, the number of codewords to process in order to obtain an infor-

mation data rate D is equal to D/k (words/second). In the worst case, decoding
a codeword requires computing B × Nit branches. To guarantee a data rate D,
an architecture must provide the power to compute D × B × Nit /k branches
per second. The minimum computing power Pc to provide per clock cycle is
therefore:
BNit D
Pc = (branches/cycle) (9.30)
kfclk
Note that, for a fully parallel architecture in which each node of the graph is
associated with a processor, all the branches of the graph are processed in one
clock cycle. The computing power is then (Pc )max = B. There is no practical
interest in trying to go beyond this power since the critical path then becomes
longer.
9.2.2 Architecture of a generic node processor (GNP)

The computations performed in a variable node processor (VNP) and in a parity
node processor (PNP) have an identical dependency between the inputs and the
outputs. Indeed, for both the PNP and VNP, the d outputs are calculated from
the d inputs, with the i-th output depending on all the inputs less the i-th input.
It is thus possible to represent the different processor architectures abstractly
by a generic node processor. The latter will then be specialized according to the
decoding algorithm used. The GNP therefore receives at its input d messages
(ei )i=1..d and produces at its output d messages (sj )j=1..d defined by:
sj = ⊗ e i (9.31)
i=j
Operator ⊗ is a generic associative-commutative operator for computations,

whose implementation will be specified below. The condensed expression 9.31
means that the operator is applied to all the variables ei for i = j.
Figure 9.10 illustrates the three main versions of GNP parallel architectures:
• Direct architecture: the computations of the d output messages are per-
formed independently (Figure 9.10(a)). The computations of the different
outputs can also be factorized. The number of ⊗ components traversed is
of the order of log2 (d).
• Trellis architecture (Forward-Backward type): This architecture corre-

sponds to a particular factorized form of parallel architecture which has
great regularity, but whose number of ⊗ operators is linear with d.
• Total sum architecture: this architecture is possible only if the generic op-
erator ⊗ allows an inverse denoted inv⊗ . In this case, the generic operator
Figure 9.10 – The different "compact mode" architectures for implementing the generic
operator ⊗.
is applied at all the inputs (total sum) then each output is calculated by
eliminating the contribution of the corresponding input, with the help of
the inverse operator.
It is possible to modify these architectures in order to introduce intermediate
pipeline registers enabling the critical path to be reduced. There are also archi-
tectures of the serial type (Figure 9.10(c)).
In what follows, the degree of parallelism of a GNP will be denoted αg . This
is the number of cycles necessary to process a node (without considering latency
due to the pipeline processing). Thus, for a parallel architecture capable of
9. LDPC codes 321
processing one node at each clock cycle, αg = 1, whereas for a serial architecture,
αg = d.
Note that in all the GNP architectures presented, we implicitly made the
hypothesis that all the inputs were available and that all the outputs had to
be generated either simultaneously (parallel architecture), or grouped in time
(serial architecture). This kind of GNP control mode is called the "compact
mode".
Figure 9.11 – Principle of the distributed mode (delayed update).
It is possible to imagine different execution modes, like the "distributed

mode", in which the GNP inputs and outputs are distributed throughout the
decoding iteration.
Figure 9.11 shows the distributed mode operation of a processor :
• During the current iteration nit , we consider that the input variables ei
come from the previous iteration nit − 1 whereas the output variables
belong to the current iteration.
• At the end of an iteration, we assume that the d input variables

it −1)
(e(n
i
)i=1..d are memorized in a memory (internal or external to the
it −1)
GNP) as well as the value of E (nit ) = ⊗ e(n
i
.
i=1..d
• The GNP can therefore, at the request of the system, calculate the i-th
(n ) (n −1)
output si it = E (nit ) ⊗ inv⊗ (ei it ).
• This output is sent, via the interleaver, to the opposite node which, once
(n )
the computation is over, returns ei it .
(n −1)
• This new value then replaces ei it in the memory and is also accu-
mulated to obtain the value of E (nit +1) at the end of the iteration. Two
accumulation modes are possible:
1. The first mode – delayed update (Figure 9.11) – involves using an

accumulation register initialized to zero at each new iteration. This
register enables E (nit +1) = ⊗ e(ni
it )
to be calculated directly. This
i=1..d
architecture therefore has d + 2 memory words, d for the inputs
it −1)
(e(n
i
)i=1..d , a word for E (nit ) and a word for the accumulation
of E (nit +1) .
2. The second mode – immediate update – involves replacing the con-
(n −1) (n )
tribution of ei it in E (nit ) by that of ei it as soon as a new input
(n )
ei it arrives, that is:

(n ) (n −1)
E (nit ) = E (nit ) ⊗ ei it ⊗ inv⊗ ei it (9.32)
At the end of the iteration, we thus have E (nit +1) = E (nit ) . This
solution offers two advantages in relation to delayed updating:
• one less memory word;
• an acceleration in the convergence of the algorithm as the new
values of the inputs are taken into account sooner.
Choice of a generic operator

Figure 9.12 gives a "cross-section" view of the belief propagation algorithm on
the bipartite graph of the LDPC code. We assume that each branch is split into
two to differentiate the variable towards parity messages from the parity towards
variable messages. This view shows the great resemblance between processing
variables and processing the parities and enables us to imagine other positions
of the interconnection networks in the computation cycle. Each position of the
interconnection graph is thus translated by a different processing of the parity
nodes and the variable nodes. Table 9.2 gives the different computations to be
carried out according to the position of the interconnection network.
When the interconnection network is in position 1 (Table 9.2), we again have
the classical separation between a variable processor and a parity processor. The
latter can then either be performed in the frequency domain (as indicated in
Figure 9.11), or directly in the domain of the LLRs via the ⊕ operator defined
in equation (9.3).
9.2.3 Generic architecture for message propagation

Presentation of the model
PNPs and GNPs are characterized by their architecture and their generic opera-
tor, depending on the position of the interconnection network. The architecture
presented in Figure 9.13 enables the exchange of messages between these differ-
ent processors.
9. LDPC codes 323
Figure 9.12 – Different positions of the interconnection network obtained by processing

the parity nodes in the Fourier domain. These positions separate those parts of the
iteration to be performed in the VNP from those performed in the PNP.
Position of the network 1 2 3 4

VNP Σ f ◦Σ Σ ◦ f −1 f ◦ Σ ◦ f −1
PNP Fourier f −1 F
◦Σ◦f f −1 ◦ Σ Σ◦f Σ
module Direct
PNP sign Product of the signs
Table 9.2 – Value of the generic operator associated with the variable processors
(VNPs) and parity processors (PNPs) as a function of the position of the intercon-
nection network.
This architecture is composed of P PNPs which each generate d = dp mes-

sages in αp clock cycles. These processors therefore have dc = dc /αp inputs
and outputs. They are connected to an interleaving network that is direct - and.
inverse. On the other side of this interleaving network are placed P dp /dv
VNPs. These processors similarly generate d = dv messages in αv clock cycles,
and therefore have dv = dv /αv inputs and outputs.
The degree of parallelism of the architecture is defined by the three parame-
ters P , αp and αv . It is possible to obtain all the degrees of parallelism possible,
ranging from the completely parallel architecture where P = m, αp = 1 and
αv = 1 to the completely serial architecture where P = 1, αc = dc and αv = dv .
Note that such an architecture has a computing power (equation (9.30)) Pc =
P × dc .
Direct inverse interleaving networks enable the messages associated with the
different VNPs to be routed towards different PNPs and vice-versa. This kind
of network generally makes it possible to perform several permutations, called
space permutations. Another type of permutation, called a time permutation,
Figure 9.13 – Generic serial-parallel architecture.
makes it possible to randomly access the nodes associated with a same processor
of nodes, by memory addressing, for example. The combination of these two
types of permutations enables a random interconnection such as that existing
between the variable nodes and the parity nodes of the LDPC code.
Example of an implementation
To help clarify ideas about the way to organize the computations and the propa-
gation of the messages in the decoder, and to truly understand the link between
the organization of the propagation of the messages and the structure of the
LDPC code, Figure 9.14 shows a simple example of decoding an LDPC code of
length n = 12 and rate R = 0, 5 (therefore m = 6), with P = 2, dc = 3, α = 1
and β = dv . There are therefore P = 2 parity node processors and n/P = 6
variable node processors. One iteration is performed in m/P = 3 steps:
• At the first cycle, reading the information relative to the bits is done in
each of the VNPs, each of them containing n/P = 2 bits of the codeword
(in practice, n/P can be much higher). These bits are shaded in grey in
each VNP: this is space permutation.
9. LDPC codes 325
• This information is then sent to the PNPs via the permutation network,
whose address was generated from the cycle number (read into a memory
for example): this is time permutation.
• Combining the two describes the random interleaving between the vari-
ables and the first two parities of the bipartite graph shown on the right-
hand part of the figure.
In a single cycle, the first two parities will therefore be able to be processed.
The following two will be processed at the second cycle and so on and so forth,
until all the parities of the code have been processed. Note that this technique
where the PNP information arrives simultaneously prevents two bits contained
in the same VNP being involved in the same parity. Thus, for example, bits
1 and 2 cannot be involved in the same parity otherwise that would lead to a
memory conflict. This solution therefore imposes constraints on matrix H, if we
want it to be decodable by this structure. One solution to relax the constraints
involves, for example, entering the data serially into the parities.
9.2.4 Combining parameters of the architecture

A certain number of parameters characterizing LDPC decoder architectures have
been defined above:
• Node processors:
– 3 possible architectures (direct, trellis, total sum)
– 4 possible positions of the interconnection network (see Figure 9.12)
– 3 input-output control modes (compact, distributed with delayed or
immediate update)
• Message propagation architecture:
– 3 parameters characterizing the level of parallelism (P, α, β)
All the combinations of these different parameters are possible to describe or
create an LDPC decoder architecture. Of course, some of these combinations are
more or less of interest, depending on the specifications required. For example,
the combinations of the control modes between VNPs and PNPs, showing the
different possible decoding schedules, are given in Table 9.3.
In the case where the controls on the two processors are of the compact
flow of inputs-outputs type, the schedule performed is of the flooding type:
all the PNPs are processed then all the VNPs. This schedule can easily be
used with completely parallel (P = m) architectures. For mixed (P < m)
architectures, the parities cannot all be processed completely before processing
the variables. The control of the VNPs in distributed mode with delayed update
allows the processing to be done since it guarantees that the new outputs will
Figure 9.14 – Example of a message propagation architecture: link between the de-
coder’s addressing and code structure.
only be computed when all the parities have been processed. This control mode
implements a flooding schedule according to the parities. Symmetrically, we
make a flooding schedule appear according to the variables, when the PNPs are
in distributed mode and the VNPs are in compact mode. These three types of
9. LDPC codes 327
VNP
distributed
Compact Delayed Immediate
update update
Flooding Flooding Interleaving
Compact
(parity) (horizontal)
Delayed Flooding
PNP
update (variable)
distributed Branches
Immediate interleaving
update (vertical)
Table 9.3 – Schedules associated with the different combinations of the node processor
controls.
schedules converge towards the same values: They do not change the information
propagation operation.
When one of the two types of processor is controlled in compact mode and
the other in distributed mode with immediate update, we implement a schedule
of the horizontal or vertical interleaving (shuffle) type. The order in which
the processors are activated is similar to the flooding schedule according to the
variables or the parities. Only the update of information changes since it is
performed as soon as a new input has arrived, thus accelerating the convergence
of the code.
The case where the two VNP and PNP processors are controlled in dis-
tributed mode is not of great interest. It would in fact correspond to controlling
the decoding, branch by branch.
The memory required to implement these different combinations is given in
Table 9.4.
VNP
Distributed
Compact Delayed Immediate
update update
Compact B+n 3n + g(B, dc ) 2n + g(B, dc )
Delayed
B + n + 2m 3n + 2m + B 2n + 2m + B
update
PNP Distributed
Immediate
B+n+m 3n + m + B 2n + m + B
update
Table 9.4 – Quantity of memory necessary as a function of the combinations of the

different node processor controls.
Parameter B designates, like above, the number of branches in the graph.

Each extrinsic branch information must be memorized, whatever the schedule
used. In all cases the intrinsic variable information must also be memorized, that
is, n values. When the parity check mode is the compact one, the accumulation
of the messages of the n variables in each VNP must be memorized, that is,
n memories if we update them immediately, and 2n in the opposite case. The
reasoning is the same if the VNPs are in compact mode and the PNPs are in
distributed mode, but in this case, it is the accumulations of messages in the m
parities that must be memorized.
It is sometimes possible, as we shall see later, to memorize the Zj,p messages
in a compressed way. The number of messages to memorize then passes from B
to g(B, dc ), with g representing a compression function (g(B, dc ) < B).
9.2.5 Example of synthesis of an LDPC decoder architec-

ture
The two examples described in this part allow us to show two LDPC decoder
architectures using two different schedules. For each of these examples, we will
give the values of the parameters characterizing these architectures.
Parameters Values
Message propagation architecture (αp = 1, αv = dv = 3, P = 3)
Position of the interconnection network 1
Control Distributed, delayed update
VNP
Data path Total sum, serial
Control Compact
PNP
Data path Trellis, parallel
Table 9.5 – Parameters characterizing the flooding schedule architecture.
Flooding schedule (according to parities)

The architecture described here to illustrate the flooding schedule is based on the
one proposed initially by Boutillon et al. [9.8]. It is schematized in Figure 9.15.
In this example, P = 3 PNP operate simultaneously in compact mode. As
αp = 1 and αv = 3, there are 12 VNPs that operate simultaneously but in
distributed mode (only one VNP and one PNP are shown in the figure). Note
that the computing power of such an architecture is 12 branches per cycle.
The architecture of the PNPs is of the trellis type, with parallel implemen-
tation. The chronogram at the bottom of Figure 9.15 indicates that at time T1 ,
dv = 4 messages Lj,p (T1 ) are produced at each PNP. After a latency of T2 − T1
clock cycles, the Zj,p (T2 ) messages leaving are sent to the VNPs. This operation
is reproduced m/P times to carry out one complete iteration. The data path of
the VNPs is of the total sum type, with a serial implementation. The delayed
update is shown by using two memory blocks, one for the extrinsic information
during accumulation (Lacc), and another for the total extrinsic information of
9. LDPC codes 329
Figure 9.15 – Example of architecture for a flooding schedule (according to parities).
the previous iteration (Lold). At the end of each iteration, the role of these two
memories is exchanged. In this architecture, the extrinsic branch information
Zj,p can be saved either on the VNP side (solid line in the figure) or on the PNP
side (dotted line in the figure), like in the Chen et al. [9.12] and Guilloud et al.
[9.24] architectures.
Horizontal interleaving schedule

This second type of architecture illustrates the horizontal interleaving schedule,
proposed by Mansour et al. [9.41] in a particular case of the turbo decoding
of LDPC codes. In this example, illustrated in Figure 9.16, there are P = 3
PNPs that operate simultaneously in compact mode. As αp = 4 and αv = 3,
there are 3 VNPs that operate simultaneously in distributed mode, which gives
a computing power of 3 branches per cycle.
The data paths of the VNPs and PNPs are both of the total sum type, with
a serial implementation. From time T1 , dc = 4 messages Lj,p enter serially into
the PNP. After a computation latency of T2 − T1 , the messages Zj,p calculated
are sent back, again serially, to the VNPs which are controlled in distributed
mode. But in this case, the update of the information is immediate. This is
translated by using a single block of Lacc memory. Thus, the sum of the extrinsic
information of the j bits is updated as soon as a new input Zj,p arrives.
Parameters Values
Message propagation architecture (αp = dc = 4, αv = dv = 3, P = 3)
Position of the interconnection network 4
Control Compact
VNP
Control Distributed, immediate update
PNP
Table 9.6 – Values of the parameters characterizing vertical interleaving architecture.
9.2.6 Sub-optimal decoding algorithm

In order to reduce the complexity of the LDPC decoder, many "sub-optimal" de-
coding algorithms have been proposed. These algorithms are based on the same
principle: reduction in complexity and in memory of the (parity or variable)
node processors, by replacing the individual computation of the d output mes-
sages (with d the degree of the node) by the computation of Δ (Δ < d) distinct
values. Of course, using a sub-optimal algorithm generally degrades the perfor-
mance of the code. A compromise thus has to be found between performance
and complexity.
Single message decoding algorithm (Δ = 1)

This is the simplest algorithm since all the outputs of the (variable or parity)
node processor are assigned a same single value at each step in the iterative
process.
VNP with Δ = 1
In this technique, the VNP simply returns Lj to the parity constraints to
which it is connected. Thus, it is no longer necessary to memorize the Zj,p mes-
sages since the latter are no longer used by the VNP. There results a significant
economy in memory. This algorithm, APP algorithm, was first proposed by
Fossorier et al. in [9.20] , and taken up again by E. Yeo et al. in [9.66].
Note that the hypothesis of independence between the messages leaving
and entering a parity node is absolutely not verified. That is why the iterative
9. LDPC codes 331
Figure 9.16 – Example of architecture for a vertical interleaving schedule. The serial
implementation of the PNP is not detailed in this figure.
decoding algorithm diverges very rapidly as it is subject to the self-confirmation

phenomenon: the propagation of the information occurs as if cycles with length
2 existed in the graph.
PNP with Δ = 1
This is the algorithm symmetric to the previous one: the PNP returns a
unique value. This technique, which is very efficient in terms of complexity,
enables the algorithm to reach its correction capacity in very few iterations,
typically 5. Although its correction capacity is very low in relation to the BP
algorithm, it is interesting to note that for 5 iterations, such an algorithm is more
efficient than the BP algorithm after this same number of iterations. Thus, such
procedures can be successfully applied for high data-rate applications where only
a reduced number of iterations can be performed.
Sub-optimal PNP algorithms (Δ > 1)

In the state of the art there are three algorithms for Δ >1 concerning the PNP.
Min-sum or BP-Based algorithm (Δ = 2)

This algorithm proposed by Fossorier et al. [9.20] requires no computations in
the PNP. Indeed, the authors suggest approximating parity processing algorithm
(9.22) by: ⎧
⎨ |Zj,p | = j ∈J(p)/j
Min (|Lj ,p |)
' (9.33)
⎩ sign (Zj,p ) = sign (Lj ,p )
j ∈J(p)/j
Only the computation of the magnitude changes: it is approximated by excess by

the minimum of the magnitudes of the messages entering the PNP. Processing
in the PNP therefore involves only computing the sign and sorting the two
lowest magnitudes of the input messages. Note that this approximation makes
the iterative decoding processing independent of the knowledge of the level of
noise σ 2 of the channel. The loss in performance is of the order of around 1 dB
compared to the BP algorithm.
This approximation by excess of the Min-Sum algorithm can however be
compensated by simple methods. It is thus possible to reduce the value of
|Zj,p | by assigning it a multiplicative factor A strictly lower than 1. It is also
possible to subtract from it an offset B (B>0), taking the precaution, however,
of saturating the result to zero if the result of |Zj,p | − B is negative. The value
c
of |Zj,p | corrected |Zj,p | is therefore:
% c
|Zj,p | = A × max(|Z
' j,p | − B, 0)
sign (Zj,p ) = sign (Lj ,p ) (9.34)
j ∈J(p)/j
These two variants of the Min-sum algorithm are called Offset BP-based and
Normalized BP-Based [9.10] respectively. The optimization of coefficients A
and B enables decoders to be differentiated. They can be constant or variable
according to the signal to noise ratio, the degree of the parity constraint, or the
processed iteration number, etc.
λ − min algorithm(Δ = λ + 1)
This algorithm was presented initially by Hu et al. [9.27, 9.28] then re-
formulated independently by Guilloud et al. [9.24]. Function f , defined by
Equation (9.35), is such that f (x) is large for low x, and low when x is large.
Thus, the sum in (9.35) can be approximated by its λ highest values, that is to
say, by the λ lowest values of |Lj,p |. Once the set denoted Jλ (p) of minima λ is
obtained, the PNP will calculate Δ = λ + 1 distinct magnitudes:
⎛ ⎞
x
|Zj,p | = f ⎝ f |Lj ,p |⎠ with f (x) = ln tanh (9.35)

2
j ∈Jλ (p)/j
9. LDPC codes 333
Indeed, if j is the index of a bit having sent one of the values of the set of min-
ima, the magnitude is calculated on all the λ − 1 other minima (λ computations
on λ − 1 values). However, for all the bits, the same magnitude is returned
(a computation on λ values). It must be noted that the performance of the
λ − min algorithm can be improved by adding a correction factor A and B as
defined in equation (9.34).
A − min ∗algorithm(Δ = 2)
The last sub-optimal algorithm published so far is called the "A − min ∗"
algorithm and was proposed by Jones et al. [9.32]. Here also, the first step is
to find index j0 of the bit having the message with the lowest module: j0 =
Arg Minj∈J(p) (|Lj,p |). Then, two distinct messages are calculated:
⎛ ⎞

If j = j0 : |Zj0 ,p | = f ⎝ f |Lj,p |⎠ (9.36)
j∈J(p)/j0
⎛ ⎞

If not j = j0 : |Zj,p | = f ⎝ f |Lj,p |⎠ (9.37)
j∈J(p)
A comparison of performance in terms of binary error rate and packet error

rate between the sub-optimal algorithms and the BP algorithm is presented in
Figure 9.17. The code simulated is an irregular code with size n = 1008 and
rate R = 0, 5, built by Hu (IBM Zurich Research Labs) with the help of the
PEG (Progressive Edge Growth) technique [9.29]. The degrees of the parities
are thus almost all equal to 8. The degrees of the variables vary from 2 to 15. We
see that the sub-optimal algorithm is not necessarily less efficient than the BP
algorithm (typically for the A − min ∗ algorithm). No compensation of extrinsic
information was used for these simulations. It is possible to greatly improve the
performance of these algorithms by adding this compensation to them, like for
the offset and normalized BP-based algorithms cited above.
9.2.7 Influence of quantization

Quantization is likely to create a degradation in performance by essentially af-
fecting two points: the computation of intrinsic information Ii defined by (9.21)
and the computation of branch (9.22) and total (9.24) extrinsic information.
We distinguish two quantization parameters: the number of quantization bits
nq and the quantization dynamic δ. The representable values thus lie between
−δ and +δ. This means that the position of the decimal point is not necessarily
defined in the nq bits or, to put it another way, that the coding dynamic is not
necessarily a power of two. Thus, the bit with the lowest weight will be equal
to:
δ
qLSB = nq −1 (9.38)
2
1.00E+00
LDPC (50 it)
BP
Amin*
1.00E01 lambda=4
lambda=3
lambda=2
BPBased
1.00E02
Error probability
1.00E03
1.00E04
1.00E05
1.00E06
1 1.5 2 2.5 3 3.5 4 4.5 5
Eb/No
Figure 9.17 – Comparison of performance between the 3−min and A−min ∗ algorithms
in the case of decoding an irregular code C3 .
and we pass from a quantified scalar aq to a non-quantified scalar a by the

relations: % n −1
q
aq = trunc a 2 δ + 0.5 ,
(9.39)
a = aq 2nqδ−1
where trunc designates the truncature operation.
These two parameters can influence the decoding performance. Too low a dy-
namic allows an error floor to appear on the error rate curves. This floor appears
much earlier than that associated with the minimum distance of the code.
However it is important to note that increasing the dynamic without increas-
ing the number of quantization bits increases the value of the bit with the lowest
weight, and consequently decreases the precision of the computations done in
the PNP. Increasing the dynamic without increasing the number of bits degrades
the decoding performance in the convergence zone.
9. LDPC codes 335
The decoding performance obtained in practice is very close to that obtained

with floating decimal points for quantizations on 4 to 6 bits (see the table in
Figure 9.18). The influence of the parameters can be studied by the density
evolution algorithm [9.15, 9.11].
9.2.8 State of the art of published LDPC decoder archi-

tectures
The table in Figure 9.18 groups the main characteristics of LDPC decoder cir-
cuits published in the literature so far. The inputs of the table are as follows:
• Circuit: description of the type of circuit used (ASIC or FPGA).

• Authors: reference to authors and articles concerning the platform.
• Architecture:
– Decoder: indication of the type of decoder (serial, parallel or mixed)
with parameters (P , αp , αv ) associated with the message propagation
architecture.
– Data path: indication for each node processor (variable and con-
straint) of type of architecture used (direct, trellis or total sum).
– Control: indication for each node processor (variable and constraint)
of type of control used, compact or distributed and, if applicable, of
type of update.
– Position of the interconnection network (between 1 and 4).
• Characteristics of the LDPC code: size, rate and regularity of the LDPC
code.
• quantization format: number of bits used to represent the data in the

decoder.
• Clock frequency of the chip in MHz.
• Data rate (in bits per second). The information binary rate is obtained
by multiplying by the coding rate.
• Maximum number of iterations.

So far, no architecture has been published that describes in detail a decoder
using a sub-optimal algorithm. However, we can mention that of Jones et al.
[9.32], but which contains too little information to be classified here.
Figure 9.18 – State of the art of the different platforms published.

9. LDPC codes 337
Bibliography
[9.1] PB. Ammar, B. Honary, Y. Kou, and S. Lin. Construction of low density
parity check codes: a combinatoric design approach. In Proceedings of
IEEE International Symposium on Information Theory (ISIT’02), July
2002.
[9.2] M. Ardakani, T.H. Chan, and F.R. Kschischang. Properties of the exit
chart for one-dimensional ldpc decoding schemes. In Proceedings of Cana-
dian Workshop on Information Theory, May 2003.
[9.3] M. Ardakani and F.R. Kschischang. Designing irregular lpdc codes using
exit charts based on message error rate. In Proceedings of International
Symposium on Information Theory (ISIT’02), July 2002.
[9.4] C. Berrou and S. Vaton. Computing the minimum distance of linear
codes by the error impulse method. In Proceedings of IEEE International
Symposium on Information Theory, July 2002.
[9.5] C. Berrou, S. Vaton, M. Jézéquel, and C. Douillard. Computing the min-
imum distance of linear codes by the error impulse method. In Proceed-
ings of IEEE Global Communication Conference (Globecom’2002), pages
1017–1020, Taipei, Taiwan, Nov. 2002.
[9.6] M. Blaum, P. Farrel, and H. Van Tilborg. Chapter 22: Array codes. In
Handbook of Coding Theory. Elsevier, 1998.
[9.7] J.W. Bond, S. Hui, and H. Schmidt. Constructing low-density parity-

check codes. EUROCOMM 2000. Information Systems for Enhanced
Public Safety and Security, IEEE AFCEA, May 2000.
[9.8] E. Boutillon, J. Castura, and F.R. Kschischang. Decoder-first code design.
In Proceedings of 2nd International Symposium on Turbo Codes & Related
Topics, pages 459–462, Brest, France, 2000.
[9.9] J. Campello and D.S. Modha. Extended bit-filling and ldpc code design.
Proceedings of IEEE Global Telecommunications Conference (GLOBE-
COM’01), pages 985–989, Nov. 2001.
[9.10] J. Chen and M. Fossorier. Near optimum universal belief propagation
based decoding of low-density parity check codes. IEEE Transactions on
Communications, 50:406–414, March 2002.
[9.11] J. Chen and M.P.C. Fossorier. Density evolution for two improved bp-
based decoding algorithms of ldpc codes. IEEE Communications Letters,
6:208–210, May 2002.
[9.12] Y. Chen and D. Hocevar. A fpga and asic implementation of rate 1/2,
8088-b irregular low density parity check decoder. In Proceedings of IEEE
Global Telecommunications Conference (GLOBECOM’03), 1-5 Dec. 2003.
[9.13] S.-Y. Chung. On the Construction of some Capacity-Approaching Coding

Schemes. PhD thesis, MIT, Cambridge, MA, 2000.
[9.14] S.-Y. Chung, T.J. Richardson, and R.L. Urbanke. Analysis of sum-
product decoding of low-density parity-check codes using a gaussian ap-
proximation. IEEE Transactions on Information Theory, 47, Feb. 2001.
[9.15] D. Declercq. Optimisation et performances des codes ldpc pour des
canaux non standards. Master’s thesis, Université de Cergy Pontoise,
Dec. 2003.
[9.16] D. Divsalar, H. Jin, and R. J. McEliece. Coding theorems for turbo-like
codes. In Proceedings of 36th Allerton Conference on Communication,
Control, and Computing, pages 201–210, Sept. 1998.
[9.17] I.B. Djordjevic, S. Sankaranarayanan, and B.V. Vasic. Projective-plane
iteratively decodable block codes for wdm high-speed long-haul transmis-
sion systems. Journal of Lightwave Technology, 22, March 2004.
[9.18] E. Eleftheriou and S. Olcer. Low-density parity-check codes for digital
subscriber lines. In Proceedings of International Conference on Commu-
nications, pages 1752–1757, 28 Apr.-2 May 2002.
[9.19] J. L. Fan. Array codes as low-density parity-check codes. In Proceedings

of 2nd Symposium on Turbo Codes, pages 543–546, Brest, France, Sept.
2000.
[9.20] M.P.C. Fossorier, M. Mihaljevic, and I. Imai. Reduced complexity itera-
tive decoding of low-density parity-check codes based on belief propaga-
tion. IEEE Transactions on Commununications, 47:673–680, May 1999.
[9.21] R. G. Gallager. Low-Density Parity-Check Codes. MIT Press, Cambridge,
MA, 1963.
[9.22] J. Garcia-Frias and Wei Zhong. Approaching shannon performance by
iterative decoding of linear codes with low-density generator matrix. IEEE
Communications Letters, 7:266–268, June 2003.
[9.23] F. Guilloud. Generic Architecture for LDPC Codes Decoding. PhD thesis,
ENST Paris, July 2004.
[9.24] F. Guilloud, E. Boutillon, and J.-L. Danger. Lambda-min decoding algo-

rithm of regular and irregular ldpc codes. In Proceedings of 3rd Interna-
tional Symposium on Turbo Codes & Related Topics, 1-5 Sept. 2003.
9. LDPC codes 339
[9.25] D. Haley, A. Grant, and J. Buetefuer. Iterative encoding of low-density

parity-check codes. In Proceedings of IEEE Global Telecommunications
Conference (GLOBECOM’02), Nov. 2002.
[9.26] X.-Y. Hu, M.P.C. Fossorier, and E. Eleftheriou. On the computation of
the minimum distance of low-density parity-check codes. In Proceedings
of IEEE International Conference on Communications (ICC’04), 2004.
[9.27] X.-Y. Hu and R. Mittelholzer. An ordered-statistics-based approxima-
tion of the sum-product algorithm. In Proceedings of IEEE International
Telecommunications Symposium, Natal, Brazil, 8-12 Sept. 2002.
[9.28] X.-Y. Hu and R. Mittelholzer. A sorting-based approximation of the sum-
product algorithm. Journal of the Brazilian Telecommunications Society,
18:54–60, June 2003.
[9.29] X.Y. Hu, E. Eleftheriou, and D.-M. Arnold. Progressive edge-growth
tanner graphs. In Proceedings of IEEE Global Telecommunications Con-
ference (GLOBECOM’01), Nov. 2001.
[9.30] S.J. Johnson and S.R. Weller. Construction of low-density parity-check
codes from kirkman triple systems. In Proceedings of IEEE Global
Telecommunication Conference (GLOBECOM’01), volume 2, pages 970–
974, 25-29 Nov. 2001.
[9.31] S.J. Johnson and S.R. Weller. Regular low-density parity-check codes from
combinatorial designs. In Proceedings of Information Theory Workshop,
pages 90–92, Sept. 2001.
[9.32] C. Jones, E. Vallés, M. Smith, and J. Villasenor. Approximate min*
constraint node updating for ldpc code decoding. In Proceedings of IEEE
Military Communications Conference (MILCOM’03), 13-16 Oct. 2003.
[9.33] R. Koetter and P.O. Vontobel. Graph-covers and the iterative decoding
of finite length codes. In Proceedings of 3rd International Symposium on
turboCodes & Related Topics, Sept. 2003.
[9.34] Y. Kou, S. Lin, and M.P.C. Fossorier. Low-density parity-check codes
based on finite geometries: A rediscovery and new results. IEEE Trans-
actions on Information Theory, 47:2711–2736, Nov. 2001.
[9.35] F.R. Kschischang and B.J. Frey. Iterative decoding of compound codes by
probability propagation in graphical models. IEEE Journal on Selected
Areas in Commununications, 16:219–230, 1998.
[9.36] D. J. C. MacKay. Good error-correcting codes based on very sparse ma-
trices. IEEE Transactions on Information Theory, 45(2):399–431, March
1999.
[9.37] D. J. C. MacKay and M. C. Davey. Evaluation of gallager codes for short

block length and high rate applications. In B. Marcus and J. Rosenthal,
editors, Codes, Systems and Graphical Models, volume 123 of IMA Vol-
umes in Mathematics and its Applications,, pages 113–130. Springer, New
York, 2000.
[9.38] David J.C. MacKay and Michael S. Postol. Weaknesses of margulis and
ramanujan-margulis low-density parity-check codes. Electronic Notes in
Theoretical Computer Science, 74, 2003.
[9.39] D.J.C MacKay and R.M. Neal. Good codes based on very sparse matri-
ces. In Proceedings of 5th IMA Conference on CryproGraphy and Coding,
Berlin, Germany, 1995.
[9.40] D.J.C MacKay, S.T. Wilson, and M.C. Davey. Comparison of construc-
tions of irregular gallager codes. IEEE Transactions on Communications,
47:1449–1454, Oct. 1999.
[9.41] M.M. Mansour and N.R. Shanbhag. Turbo decoder architectures for low-
density parity-check codes. In Proceedings of IEEE Global Telecommuni-
cations Conference (GLOBECOM’02), 17-21 Nov. 2002.
[9.42] Y. Mao and A.H. Banihashemi. Decoding low-density parity-check codes
with probabilistic scheduling. IEEE Communications Letters, 5:414–416,
Oct. 2001.
[9.43] G. A. Margulis. Explicit construction of graphs without short cycles and
low density codes. Combinatorica, 2(1):71–78, 1982.
[9.44] R. J. McEliece, D. J. C. MacKay, and J.-F. Cheng. Turbo decoding as
an instance of pearle’s belief propagation algorithm. IEEE Journal on
Selected Areas in Commununications, 16:140–152, Feb. 1998.
[9.45] T.R. Oenning and Jaekyun Moon. A low-density generator matrix inter-
pretation of parallel concatenated single bit parity codes. IEEE Transac-
tions on Magnetics, 37:737– 741, 2001.
[9.46] T. Okamura. Designing ldpc codes using cyclic shifts. In Proceedings of
IEEE International Symposium on Information Theory (ISIT’03), July
2003.
[9.47] A. Prabhakar and K. Narayanan. Pseudorandom construction of low den-
sity parity-check codes using linear congruential sequences. IEEE Trans-
actions on Communications, 50:1389–1396, Sept. 2002.
[9.48] T.J. Richardson, M.A. Shokrollahi, and R.L. Urbanke. Design of capacity-
approaching irregular low-density parity-check codes. IEEE Transactions
on Information Theory, 47:619–637, Feb. 2001.
9. LDPC codes 341
[9.49] T.J. Richardson and R.L Urbanke. Efficient encoding of low-density

parity-check codes. IEEE Transactions on Information Theory, 47:638–
656, Feb. 2001.
[9.50] J. Rosenthal and P. Vontobel. Constructions of ldpc codes using ramanu-

jan graphs and ideas from margulis. In Proceedings of the 38-th Annual
Allerton Conference on Communication, Control, and Computing, pages
248–257, 2000.
[9.51] J. Rosenthal and P. Vontobel. Constructions of regular and irregular ldpc
codes using ramanujan graphs and ideas from margulis. In Proceedings of
IEEE International Symposium on Information Theory (ISIT’01), page 4,
June 2001.
[9.52] M. Sipser and D.A. Spielman. Expander codes. IEEE Transactions on
Information Theory, 42:1710–1722, Nov. 1996.
[9.53] R. M. Tanner. A recursive approach to low complexity codes. IEEE

Transactions on Information Theory, IT-271:533–547, Sept. 1981.
[9.54] R.M. Tanner. A [155, 64, 20] sparse graph (ldpc) code. In Proceedings of
IEEE International Symposium on Information Theory, Sorrento, Italy,
June 2000.
[9.55] S. ten Brink. Convergence of iterative decoding. IEE Electronics Letters,
35:806–808, May 1999.
[9.56] S. ten Brink. Iterative decoding trajectories of parallel concatenated
codes. In Proceedings of 3rd IEEE ITG Conference on Source and Chan-
nel Coding, pages 75–80, Munich, Germany, Jan. 2000.
[9.57] T. Tian, C. Jones, J.D. Villasenor, and R.D. Wesel. Construction of irreg-
ular ldpc codes with low error floors. In Proceedings of IEEE International
Conference on Communications (ICC’03), 2003.
[9.58] B. Vasic. Combinatorial constructions of low-density parity check codes
for iterative decoding. In Proceedings of IEEE International Symposium
on Information Theory (ISIT’02), July 2002.
[9.59] B. Vasic. High-rate low-density parity check codes based on anti-pasch
affine geometries. In Proceedings of IEEE International Conference on
Communications (ICC’02), volume 3, pages 1332–1336, 2002.
[9.60] B. Vasic, E.M. Kurtas, and A.V. Kuznetsov. Kirkman systems and their
application in perpendicular magnetic recording. IEEE Transactions on
Magnetics, 38:1705–1710, July 2002.
[9.61] B. Vasic, E.M. Kurtas, and A.V. Kuznetsov. Ldpc codes based on mu-
tually orthogonal latin rectangles and their application in perpendicu-
lar magnetic recording. IEEE Transactions on Magnetics, 38:2346–2348,
Sept. 2002.
[9.62] F. Verdier and D. Declercq. A ldpc parity check matrix construction for
parallel hardware decoding. In Proceedings of 3rd International Sympo-
sium on Turbo Codes & related topics, 1-5 Sept. 2003.
[9.63] Eric W. Weisstein. Weisstein. from Mathworld,
http://mathworld.wolfram.com/BlockDesign.html.
[9.64] N. Wiberg. Codes and Decoding on General Graphs. PhD thesis,
Linköping University, 1996.
[9.65] X.-Y.Hu, E. Eleftheriou, D.-M. Arnold, and A. Dholakia. Efficient im-

plementations of the sum-product algorithm for decoding ldpc codes. In
Proceedings of IEEE Global Telecommunications Conference (GLOBE-
COM’01), pages 1036–1036, Nov. 2001.
[9.66] E. Yeo, B. Nikolic, and V. Anantharam. High throughput low-density

parity-check decoder architectures. In Proceedings of IEEE Global
Telecommunications Conference (GLOBECOM’01), San Antonio, 25-29
Nov. 2001.
[9.67] H. Zhang and J.M.F. Moura. The design of structured regular ldpc codes
with large girth. In Proceedings of IEEE Global Telecommunications Con-
ference (GLOBECOM’03), Dec. 2003.
Chapter 10
Turbo codes and large

spectral efficiency
transmissions
Transporting information in telecommunication systems is carried out at higher

and higher data rates and in narrower and narrower frequency bands. Conse-
quently, we wish to maximize the ratio of the useful data rate to bandwidth,
that is to say, the spectral efficiency of the transmissions. To do this, it seems
natural to couple digital modulations having large constellations with powerful
high-rate error correcting codes like turbo codes.
The studies undertaken in this domain are essentially based on two ap-
proaches: turbo trellis coded modulation and pragmatic turbo coded modulation.
10.1 Turbo trellis coded modulation (TTCM)

Turbo trellis coded modulation or TTCM was introduced by Robertson and
Wörz in 1995 [10.5, 10.6]. It uses the notion of parallel concatenation, which
is at the origin of turbo coding, applied to two trellis coded modulations, or
TCMs.
TCM, introduced by Ungerboeck at the beginning of the 80s [10.7] is based
on the joint optimization of error correction coding and modulation. The coding
is performed directly in the signal space so that the error correcting code and
the bit to signal mapping of the modulation can be represented jointly using
a single trellis. The criterion for optimizing a TCM thus involves maximizing
the minimum Euclidean distance between two coded sequences. To do this,
Ungerboeck proposed a two-step approach: partitioning the constellation of the
modulation into sub-constellations presenting increasing minimum Euclidean
distances, then assigning to each branch of the trellis a signal belonging to the
constellation, respecting a set of rules such as those described in [10.7].
The TTCM scheme presented by Robertson and Wörz is shown in Fig-
ure 10.1. Each TCM encoder is made up of a recursive systematic convolutional
encoder, or RSC encoder, with rate q/(q+1), and a modulation without memory
of order Q = 2q+1 . The binary symbols coming from the source are grouped
into symbols of q bits. These symbols are encoded by the first TCM in the order
in which are produced by the source and by the second TCM after interleaving.
Figure 10.1 – Diagram of the principle of turbo trellis coded modulation (TTCM),
according to Robertson and Wörz [10.5, 10.6]. Spectral efficiency η = q bit/s/Hz.
Each q-tuple coming from the source being encoded two times, a selection
operator alternatively transmits the output of one of the two TCM encoders,
in order to avoid the double transmission of information, which would lead to
a spectral efficiency of the system of q/2 bit/s/Hz. This in fact amounts to
puncturing half of the redundancy sequence for each convolutional code.
At reception, the TTCM decoder is similar to a turbo decoder, except that
the former directly processes the (q + 1)-ary symbols coming from the demod-
ulator. Thus, the calculation of the transition probabilities at each step of
the MAP algorithm (see Section 7.4) uses the Euclidean distance between the
received symbol and the symbol carried by each branch of the trellis. If the
decoding algorithm operates in the logarithmic domain (Log-MAP, Max-Log-
MAP), it is the branch metrics that are taken equal to the Euclidean distances.
Computing an estimate of the bits carried by each demodulated symbol, before
decoding, would indeed be a sub-optimal implementation of the receiver.
Similarly, for efficient implementation of turbo decoding, the extrinsic infor-
mation exchanged by the elementary decoders must directly concern the q-tuples
of information transmitted and not the binary elements that they are made up
10. Turbo codes and large spectral efficiency transmissions 345
of. At each decoding instant, the elementary decoders thus exchange 2q values
of extrinsic information.
Figure 10.2 provides two examples of elementary RSC codes used in [10.5,
10.6] to build an 8-PSK TTCM with 8 states of spectral efficiency η = 2 bit/s/Hz
and a 16-QAM TTCM with spectral efficiency η = 3 bit/s/Hz.
(a)
(b)
Figure 10.2 – Examples of elementary RSC codes used in [10.5, 10.6] for the construc-
tion of a 8-PSK turbo trellis (a) and a 16-QAM turbo trellis (b) coded modulations.
Figures 10.3 and 10.4 show the performance of these two TTCMs in terms of
binary error rates (BER) as a function of the signal to noise ratio for transmission
over a Gaussian channel. At high and average error rates, these schemes show
correction performance close to capacity: a BER of 10−4 is reached at around
0.65 dB from Shannon’s theoretical limit for the transmission of packets of 5,000
coded modulated symbols. On the other hand, as the interleaving function of
the TTCM has not been the object of any particular optimization in [10.5, 10.6],
the error rates curves presented reveal early changes in slope (BER∼ 10−5 ) that
are very pronounced.
A variant of this technique, proposed by Benedetto et al. [10.1] made it pos-
sible to improve its asymptotic performance. An alternative method to build a
TTCM with spectral efficiency q bit/s/Hz involves using two RSC codes with
rate q/(q + 1) and for each of them to puncture q/2 information bits (q is as-
sumed to be even). For each elementary code we thus transmit only half the
information bits and all the redundancy bits. The bits at the output of each
encoder are associated with a modulation with 2(q/2)+1 points. The same oper-
ation is performed for the two RSC codes, taking care that each systematic bit
is transmitted once and only once, so that the resulting turbo code is system-
Figure 10.3 – Binary error rate (BER) as a function of the signal to noise ratio Eb /N0
of the 8-PSK TTCM with 8 states using the RSC code of Figure 10.2(a). Transmis-
sion over a Gaussian channel. Spectral efficiency η = 2 bit/s/Hz. Blocks of 10,000
information bits, 5,000 modulated symbols. MAP decoding algorithm. Curves taken
from [10.6].
atic. On the other hand, this technique uses interleaving at bit level, and not at
symbol level like in the previous approach.
The criterion for optimizing the TTCM proposed in [10.1] is based on maxi-
mizing the effective Euclidean distance, defined as the minimum Euclidean dis-
tance between two encoded sequences whose information sequences have a Ham-
ming weight equal to 2. Figures 10.5 and 10.6 show two examples of TTCMs
built on this principle.
The correction performance of these two TTCMs over a Gaussian channel
are presented in Figures 10.7 and 10.8. At high and average error rates, they are
close to those given by the scheme of Robertson and Wörz; on the other hand,
using interleavers operating on the bits rather than on the symbols has made it
possible to significantly improve the behaviour at low error rates.
TTCMs lead to excellent correction performance over a Gaussian channel,
since they are an ad hoc approach to turbo coded modulation. However, they
have the main drawback of very limited flexibility: a new code must be defined
for each coding rate and each modulation considered. This drawback is cumber-
some in any practical system requiring a certain degree of adaptability. On the
other hand, although they are a quasi-optimal solution to the problem of coded
modulations for the Gaussian channel, their behaviour over fading channels like
Rayleigh channels leads to mediocre performance [10.9].
Figure 10.4 – BER as a function of the signal to noise ratio Eb /N0 of the 16-QAM
TTCM with 8 states using the RSC code of Figure 10.2(b). Transmission over a
Gaussian channel. Spectral efficiency η = 3 bit/s/Hz. Blocks of 15,000 information
bits, 5,000 modulated symbols. MAP decoding algorithm. Curves taken from [10.6].
10.2 Pragmatic turbo coded modulation

The so-called pragmatic approach was chronologically the first implementation.
It was introduced by Le Goff et al. [10.4] in 1994. This technique takes its
name from its similarities with the technique of associating a convolutional
code and modulation proposed by Viterbi [10.2] as an alternative solution to
Ungerboeck’s TCMs. The coding and modulation functions are processed in-
dependently, without joint optimization. It uses a "good" turbo code, a bit to
signal mapping which minimizes the probability of binary error at the output
of the corresponding demapper (Gray coding) and associates the two functions
via puncturing and multiplexing to adapt the whole scheme to the spectral ef-
ficiency targeted. Figures 10.9 and 10.10 present the general diagram for the
principle of the transmitter and the receiver for the pragmatic association of a
turbo code and modulation with Q = 2q states.
With this pragmatic approach to turbo coded modulation, the encoder and
the decoder used are standard turbo encoders and decoders, identical for all
coding rates and modulations considered. If the size of the blocks of data trans-
mitted is variable, simple parametering of the code’s permutation function must
allow it to adapt to different sizes.
When the targeted coding rate is higher than the natural rate of the turbo
code, the puncturing operation enables it to erase, that is to say, not transmit,
(a) Elementary encoder
(b) TTCM encoder
Figure 10.5 – Construction of a 16-QAM TTCM according to the method described

in [10.1]. Spectral efficiency η = 2 bit/s/Hz.
certain coded bits. In practice, for practical reasons of hardware implementa-

tion, the puncturing pattern is periodic or quasi-periodic. If possible, only the
parity bits are punctured. Indeed, puncturing the systematic bitsleads to a rapid
degradation in the decoding convergence threshold, as these bits take part in
the process of decoding the two codes, unlike the redundancy bits. When the
coding rate is high, a slight puncturing of the data can nevertheless improve the
asymptotic behaviour of the system.

The presence of interleaving functions Π and Π −1 is justified by the need
to decorrelate the data at the input of the turbo decoder. In fact, it is shown in
[10.4] that inserting this interleaving has no significant effect on the error rates
at the output of the decoder in the case of a transmission on a Gaussian channel.
However, in the case of fading channels, the interleaving is necessary as we must
prevent bits coming from the same coding instant from belonging to a same
symbol transmitted over the channel, so that they will not be affected simulta-
neously by fading. The studies carried out in this domain [10.9, 10.4, 10.8, 10.2]
have shown that in the case of fading channels, the best performance is ob-
tained by using independent interleavers at bit level. This technique is called
Bit-Interleaved Coded Modulation (BICM). When the interleaver is placed at
modulation symbol level, the order of diversity of the coded modulation is equal
to the minimum number of different symbols between two coded modulated se-
quences. With independent interleavers placed at each output of the encoder, it
(a) Elementary encoder
(b) TTCM encoder
Figure 10.6 – Construction of an 8-PSK TTCM according to the method described in

[10.1]. Spectral efficiency η = 2 bit/s/Hz
.
can ideally reach the Hamming distanceof the code. Consequently, transmission
schemes using the BICM principle in practice have better performance on fading
channels than TTCMs have.
The code and the modulation not being jointly optimized, unlike a TTCM
scheme, we choose binary mapping of the constellation points which minimizes
the mean binary error rates at the input of the decoder. When it can be envis-
aged, Gray encoding satisfies this condition. For simplicity in implementing the
modulator and demodulator, in the case of square QAM (q even), the in-phase
and in-quadrature axes, I and Q, are mapped independently.
In Figure 10.9, the role of the "Multiplexing / symbol composition" block is
to distribute the encoded bits, after interleaving for fading channels, into modu-
lation symbols. This block, the meeting point between the code and the modu-
lation, enables a certain level of adjustment of the coded modulation according
to the performance targeted. This adjustment is possible since the code and the
modulation do not play the same role in relation to all the bits transmitted.
On the one hand, we can distinguish two distinct families of encoded bits
at the output of the encoder: systematic bits and redundancy bits. These two
families of bits play a different role in the decoding process: the systematic
bits, coming directly from the source, are used by the two elementary decoders
Figure 10.7 – BER as a function of the signal to noise ratio Eb /N0 of the 16-QAM
TTCM with 16 states using the RSC code of Figure 10.5. Transmission over a Gaussian
channel. Spectral efficiency η = 2 bit/s/Hz. 2×16,384 information bits. MAP decoding
algorithm. Curves taken from [10.1].
at reception whereas the redundancy bits, coming from the two elementary
encoders, are used by only one of the two decoders.
On the other hand, the binary elements contained in a modulated symbol
are not, in general, all protected identically by the modulation. For example,
in the case of PSK or QAM modulation with Gray encoding, only modulations
with two or four points offer the same level of protection to all the bits of a same
symbol. For higher order modulations, certain bits are better protected than
others.
As an illustration, consider a 16-QAM modulation, mapped independently
and in an analogue manner on the in-phase and in-quadrature axes by Gray
encoding. The projection of this modulation on each of the paths is amplitude
shift keying (ASK) with 4 symbols (see Figure 10.11).
We can show that, for a transmission over a Gaussian channel, the error
probabilities on the binary positions s1 and s2 , given the transmitted symbol
(±3 or ±1), are expressed in the form:
Figure 10.8 – BER as a function of the signal to noise ratio Eb /N0 of the 8-PSK
TTCM with 16 states using the RSC code of Figure 10.6. Transmission over a Gaussian
channel. Spectral efficiency η = 2 bit/s/Hz. 4×4,096 information bits. MAP decoding
algorithm. Curves taken from [10.3].

1 3
Peb (s2 | ±3 ) = erfc √
2 σ 2

1 1
Peb (s2 | ±1 ) = erfc √
2 σ 2
(10.1)

1 1 1 5 1 1
Peb (s1 | ±3 ) = erfc √ − erfc √ ≈ erfc √
2 σ 2 2 σ 2 2 σ 2

1 3 1 1 1 1
Peb (s1 | ±1 ) = erfc √ + erfc √ ≈ erfc √
2 σ 2 2 σ 2 2 σ 2
where erfc represents the complementary error function and σ 2 designates the
noise variance on the channel. We observe that binary position s2 is on average
better protected by the modulation than position s1 .
Consequently, it is possible to define several strategies for building modula-
tion symbols by associating as a matter of priority the systematic bits or redun-
Figure 10.9 – Diagram of the principle of the transmitter in the case of the pragmatic
association of a turbo code and modulation with Q = 2q states.
Figure 10.10 – Diagram of the principle of the receiver for the pragmatic turbo coded
modulation scheme of Figure 10.9.
Figure 10.11 – Diagram of the signals of 4-ASK modulation with Gray encoding.
dant bits with the positions that are the best protected by the modulation. Two
extreme strategies can thus be defined in all cases:
• so-called "systematic" scheme: the bits best protected by the modulation

are associated as a matter of priority with the systematic bits;
• so-called "redundant" scheme: the bits best protected by the modulation
are associated as a matter of priority with the redundancy bits.
Modulations of orders higher than 16-QAM offer more than two levels of
protection for the different binary positions. 64-QAM modulation, for example,
gives three different levels of protection, if the in-phase and in-quadrature axes
are mapped independently and in an analogue manner by using a Gray code.
In this case, other schemes, falling in between "systematic" and "redundant"

schemes, can be defined.
The reception scheme corresponding to the transmitter of Figure 10.9 is
described in Figure 10.10. A standard turbo decoder is used, which requires
calculating a weighted estimation of each of the bits contained in the symbols
at the output of the demodulator before carrying out the decoding.
The weighted estimation of each bit si is obtained by calculating the log
likelihood ratio (LLR) defined by:

σ2 Pr( si = 1 | r) σ2 Pr( si = 1 | I, Q)
ŝi = Λ(si ) = ln = ln (10.2)
2 Pr( si = 0 | r) 2 Pr( si = 0 | I, Q)
The sign of the LLR provides the binary decision on si (0 if Λ(si ) ≤ 0,

1 otherwise) and its absolute value represents the weight, that is to say, the
reliability, associated with the decision. In the case of a transmission over a
Gaussian channel, ŝi can be written:
⎛ " #⎞
d2r,s
⎜ exp − 2 ⎟
σ2 ⎜
2σ ⎟
ŝi =
⎜ s∈Si,1
ln ⎜ " #⎟
⎟ (10.3)
2 ⎜ d 2 ⎟
⎝ exp − 2 ⎠
r,s
2σ
s∈Si,0
where Si,1 and Si,0 represent the sets of points s of the constellation such that
the ith bit si is equal to 1 or 0, and dr,s is the Euclidean distance between the
received symbol r and the constellation point considered s.
In practice, the Max-Log approximation is commonly used to simplify the
calculation of the LLRs:
ln(exp(a) + exp(b)) ≈ max(a, b) (10.4)

and the weighted estimations are calculated as:

1
ŝi = min (d ) − min(d ) .
2 2
(10.5)
4 s∈Si,0 r,s s∈Si,1 r,s
We note that it is not necessary to know the noise variance on the channel, when
this simplification is used.
The other blocks of the receiver perform the inverse operations of the blocks
of Figure 10.9. The depuncturing operation corresponds to the insertion of an
LLR equal to 0, this is to say, a zero reliability decision, at the input of the
decoder for all the non-transmitted bits.
The pragmatic turbo coded modulation approach enables performance very
close to that of TTCMs to be obtained. Figures 10.12 and 10.13 show the
performance of two pragmatic turbo coded modulations using the double-binary
turbo code with 16 states presented in Section 7.5 for transmission conditions
similar to those leading to the curves obtained in Figures 10.3 and 10.4. We
observe that after 8 decoding iterations, the performance of the two turbo coded
modulation families are equivalent down to BERs from 10−4 to 10−5 . The better
behaviour of the pragmatic solution at lower error rates is due, on the one hand,
to the use of 16-state elementary codes and, on the other hand, to the careful
design of the turbo code interleaver.
Figure 10.12 – BER as a function of the signal to noise ratio Eb /N0 of pragmatic
turbo-coded 8-PSK using a 16-state double-binary code. Transmission over a Gaussian
channel. Spectral efficiency η = 2 bit/s/Hz. Blocks of 10,000 information bits, 5,000
modulated symbols. MAP decoding algorithm. "Systematic" scheme.
The curves of Figure 10.14 show the influence of the strategy of construct-
ing symbols on the performance of turbo coded modulation. They show the
behaviour of the association of a 16-state double-binary turbo code and a 16-
QAM mapped independently on the in-phase and in-quadrature axes using the
Gray code. The two extreme strategies for building the symbols described above
were simulated. The size of the blocks, 54 bytes, and the simulated rates 1/2 and
3/4, are representative of concrete applications in the wireless technology sector
(IEEE 802.16 standard, Wireless Metropolitan Area Network ). Figure 10.14 also
shows the theoretical limits of the transmission studied. These limits take into
Figure 10.13 – BER as a function of the signal to noise ratio Eb /N0 of pragmatic turbo-
coded 16-QAM using a 16-state double-binary code. Transmission over a Gaussian
channel. Spectral efficiency η = 3 bit/s/Hz. Blocks of 15,000 information bits, 5,000
modulated symbols. MAP decoding algorithm. "systematic" scheme.
account the size of the blocks transmitted as well as the packet error rates (PER)
targeted. They are obtained from the value of the capacity of the channel, to
which we add a correcting term obtained with the help of the so-called "sphere
packing" bound, (see Section 3.3).
We observe that at high or average error rates, the convergence of the itera-
tive decoding process is favoured by a better protection of the systematic bits.
This result can be explained by the fact that, in the decoding process, each sys-
tematic data is used at the input of the two decoders. Consequently, an error on
a systematic bit at the output of the channel causes an error at the input of the
two elementary decoders, whereas erroneous redundancy only affects the input
of one of the two elementary decoders. Consequently, reinforcing the protection
of the systematic bits benefits the two elementary decoders simultaneously.
The higher the proportion of redundancy bits transmitted, that is to say,
the lower the coding rate, the greater the gap in performance between the two
schemes. As an example, at a binary error rate of 10−4, we observe a gap of

0.9 dB for a coding rate R = 1/2, and 0.2 dB for R = 3/4.
Figure 10.14 – Performance in binary error rate (BER) and packet error rate (PER) of
the pragmatic association of a 16-QAM and a 16-state double-binary turbo code, for
the transmission of blocks of 54 bytes over a Gaussian channel. Coding rates 1/2 and
3/4. Max-Log-MAP decoding, inputs of the decoder quantized on 6 bits, 8 decoding
iterations.
For low and very low error rates, the scheme favouring the protection of the
redundancy gives the best performance. This behaviour is difficult to prove by
simulation for the lowest rates, as the assumed crossing point of the curves is
situated at an error rate that is difficult to obtain by simulation (P ER ≈ 10−8
for R = 1/2). The interpretation of this result requires analysis of the erroneous
paths in trellises with a high signal to noise ratio. We have observed that, in
the majority of cases, the erroneous sequences contain a fairly low number of
erroneous systematic bits and a rather high number of erroneous redundancy
bits. In other words, the erroneous sequences generally have a low input weight.
In particular, the erroneous paths in question mainly correspond to rectangular
patterns of errors (see Section 7.3.2). The result, from the point of view of
the asymptotic behaviour of turbo coded modulation, is that it is preferable to
ensure better protection of the parity bits.
The curves shown in Figure 10.14 were obtained with the help of the simpli-
fied Max-Log-MAP decoding algorithm, using data quantized on 6 bits at the
input of the decoder. These conditions correspond to a hardware implementa-

tion of the decoder. In spite of these constraints, the performance obtained is
fairly close to the theoretical limits: only 1 dB at a PER of 10−4 and 1.5 dB at
a PER of 10−6.
Bibliography
[10.1] S. Benedetto, D. Divsalar, G. Montorsi, and F. Pollara. Parallel concate-
nated trellis codes modulation. In Proceeedings of International Confer-
ence on Communications (ICC’96), pages 974–978, Dallas, USA, 1996.
[10.2] G. Caire, G. Taricco, and E. Biglieri. Bit-interleaved coded modulation.
IEEE Transactions on Information Theory, 44(3):927–946, May 1998.
[10.3] C. Fragouli and R. Wesel. Bit vs. symbol interleaving for parallel con-
catenated trellis coded modulation. In Proceedings of Global Telecommu-
nications Conference (Globecom’01), pages 931–935, San Antonio, USA,
Nov. 2001.
[10.4] S. Le Goff, A. Glavieux, and C. Berrou. Turbo codes and high spec-
tral efficiency modulation. In Proceedings of International Conference on
Communications (ICC’94), pages 645–649, New Orleans, USA, May 1994.
[10.5] P. Robertson and T. Wörz. Coded modulation scheme employing turbo

codes. Electronics Letters, 31(2):1546–1547, Aug. 1995.
[10.6] P. Robertson and T. Wörz. Bandwidth-efficient turbo trellis-coded modu-
lation using punctured component codes. IEEE Journal on Selected Areas
in Communications, 16(2):206–218, Feb. 1998.
[10.7] G. Ungerboeck. Channel coding with mutilevel/phase signals. IEEE

Trans. Info. Theory., IT-28(1):55–67, Jan. 1982.
[10.8] A. J. Viterbi, J. K. Wolf, E. Zehavi, and R. Padovani. A pragmatic
approach to trellis-coded modulation. IEEE Communications Magazine,
27(7):11–19, July 1989.
[10.9] E. Zehavi. 8-psk trellis codes for a rayleigh channel. IEEE Transactions
on Communications, 40(5):873–884, May 1992.
Chapter 11
The turbo principle applied to

equalization and detection
The invention of turbo codes at the beginning of the 90s totally revolutionized
the field of error correcting coding. Codes relatively simple to build and decode,
making it possible to approach Shannon’s theoretical limit very closely, were at
last available. However, the impact of this discovery was not limited to one sin-
gle coding domain. More generally, it gave birth to a new paradigm for designing
digital transmission systems, today commonly known as the "turbo principle".
To solve certain very complex a priori signal processing problems, we can envis-
age dividing these problems into a cascade of elementary processing operations,
simpler to implement. However, today we know that the one-directional succes-
sion of these processing operations leads to a loss of information. To overcome
this sub-optimality, the turbo principle advocates establishing an exchange of
probabilistic information, "in the two directions", between these different pro-
cessing operations. All of the information available is thus taken into account
in solving the global problem and a consensus can be found between all the
elementary processing operations in order to elaborate the final decision.
The application of the turbo principle to a certain number of classical prob-
lems in digital transmission has provided impressive gains in performance in
comparison to traditional systems. Therefore its use rapidly became popular
within the scientific community. This chapter presents the first two systems
having historically benefited from the application of the turbo principle to a
context other than error correction coding. The first system, called turbo equal-
ization, iterates between the equalization function and a decoding function to
improve the processing of the intersymbol interference for data transmission over
multipath channels. The second, commonly called turbo CDMA, exploits the
turbo principle to improve the discrimination between users in the case of a
radio-mobile communication between several users based on the Code Division

Multiple Access technique.
11.1 Turbo equalization

Multipath channels have the particularity of transforming a transmitted signal
into a linear superposition of several different copies (or echoes) of this signal.
Turbo equalization is a digital reception technique that makes it possible to
detect data deteriorated by multipath transmission channels. It combines the
work of an equalizer and a channel decoder using the turbo principle. Schemat-
ically, this digital reception system involves a repetition of the equalization-
interleaving-decoding processing chain. First, the equalization performs an ini-
tial estimation of the transmitted data. Second, the estimation is transmitted
to the decoding module which updates this information. Then the information
updated by the decoder is sent to the equalization module. Thus, over the itera-
tions, the equalization and decoding processing operations exchange information
in order to reach the performance of a transmission on a channel with a single
path.
The purpose of this section is to present the turbo equalization principle
and its implementation in two versions: turbo equalization according to the
Maximum A Posteriori (MAP) criterion and turbo equalization according to
the Minimum Mean Square Error (MMSE) criterion. We will describe the algo-
rithms associated with these two techniques, as well as their respective complex-
ity. This will lead us to present the possible architectures and give examples of
implementation. Finally, potential and existing applications for these techniques
will be shown.
11.1.1 Multipath channels and intersymbol interference

This section is dedicated to transmissions on multipath channels whose partic-
ularity is to generate one or several echoes of the signal transmitted. Physically
these echoes can, for example, correspond to reflections off a building. The
echoes thus produced come and superpose themselves on the signal initially
transmitted and thus degrade the reception. The equivalent discrete channel
model allows a mathematically simple representation of these physical phenom-
ena in the form of a linear filtering of the transmitted discrete-time symbol
sequence. Let xi be the symbol transmitted at discrete instant i, and yi be the
received symbol at this same instant. The channel output is then given by

L−1
yi = hk (i)xi−k + wi (11.1)
k=0
where hk (i) represents the action of the channel (echo) at instant i on a symbol
transmitted at instant i − k. The impulse response of the channel at instant i
11. The turbo principle applied to equalization and detection 361
is then written in the following way in the form of a z-transform:

L−1
h(z) = hk (i)z −k (11.2)
k=0
The impulse response of the channel is assumed to have finite duration (L coef-
ficients), which is a realistic hypothesis in practice in most scenarios.
Equation (11.1) shows that generally, received symbol yi is a function of the
symbols transmitted before, or after (if the channel introduces a propagation de-
lay) information symbol xi considered at instant i. In accordance with what was
introduced in Chapter 2, we then say that the received signal is spoiled by inter-
symbol interference (ISI). If we now assume that the transmission channel does
not vary (or very little) on the duration of a transmitted block of information,
model(11.1) can be simplified as follows:

L−1
yi = hk xi−k + wi (11.3)
k=0
where we have suppressed the time dependency from the coefficients of the equiv-
alent discrete channel. The representation of the equivalent discrete channel in
the form of a digital filter with finite impulse response presented in Figure 11.1
comes directly from (11.3). The coefficients of the filter are precisely those of
the impulse response of the channel.
Figure 11.1 – Representation of the equivalent discrete channel in the form of a digital
filter.
ISI can be a major obstacle for establishing a good quality digital trans-
mission, even in the presence of very low noise. As an illustration, we have
shown in Figure 11.2 the constellation of the symbols received at the output of
a channel highly perturbed by ISI, for a signal to noise ratio of 20 dB1 , given
that we have transmitted a sequence of discrete symbols with four phase states
(QPSK modulation). We thus observe that when the ISI is not processed by an
1 We recall that a signal to noise ratio of 20 dB corresponds to a power of the transmitted
signal 100 times higher than the power of the additive noise on the link.
adequate device, it can lead to great degradation in the error rate at reception,
and therefore in the general quality of the transmission.
(a) (b)
Figure 11.2 – Illustration of the phenomenon of ISI in the case of a 5-path highly
frequency-selective channel, for a signal to noise ratio of 20 dB.
We now study the characteristics of a multipath channel in the frequency do-

main. We show in Figure 11.3 the frequency response of the channel generating
the constellation presented in Figure 11.2. The latter is highly perturbed by ISI.
We note that the frequencies of the signal will not be attenuated and delayed in
the same way over the whole frequency band. Thus, a signal having a band W
between 0 and 3 kHz will be distorted by the channel. We then speak of a fre-
quency selective channel in opposition to a flat non-frequency selective channel,
for which all the frequencies undergo the same distortion. To resume, when a
multipath channel generates intersymbol interference in the time domain, it is
then frequency selective in the frequency domain.
We mainly have three different techniques to combat the frequency selectiv-
ity of transmission channels: multi-carrier transmissions, spread spectrum and
equalization. In this chapter, we deal only with the third solution, applied here
to transmissions on a single carrier frequency ("single-carrier" transmissions).
Note, however, that some of the concepts tackled here can be transposed rela-
tively easily to systems of the multi-carrier type (Orthogonal Frequency Division
Multiplex, or OFDM systems).
11.1.2 The equalization function

In its most general form, the purpose of the equalization function is to give an
estimation of the transmitted sequence of symbols from the sequence observed
at the output of the channel, the latter being perturbed both by intersymbol
interference and additive noise, assumed to be Gaussian. We distinguish different
equalization strategies. Here we limit ourselves to a succinct overview of the
Figure 11.3 – Frequency response of the 5-path discrete-time channel.
main techniques usually implemented in systems. The interested reader can

find additional information in Chapters 10 and 11 of [11.44], in articles [11.45]
and [11.54] or in book [11.11], for example.
A first solution, called Maximum Likelihood Sequence Detection, or MLSD,
involves searching for the most probable sequence transmitted relatively to the
observation received at the output of the channel. We can show that this cri-
terion amounts to choosing the candidate sequence at the minimum Euclidean
distance from the observation, and that it thus minimizes the error probability
per sequence, that is to say, the probability of choosing a candidate sequence
other than the sequence transmitted. A naive implementation of this criterion
involves listing the set of admissible sequences in such a way as to calculate the
distance between each sequence and the observation received, then to select the
sequence closest to this observation. However, the complexity of this approach
increases exponentially with the size of the message transmitted, which turns
out to be unacceptable for a practical implementation.
In a famous article dating from 1972 [11.27], Forney noted that a frequency
selective channel presents a memory effect whose content characterizes its state
at a given instant. More precisely, state s of the channel at instant i is perfectly
defined by the knowledge of the L − 1 previous symbols, which we denote s =
(xi , . . ., xi−L+2 ). This fact is based on the representation of the channel in the
form of a shift register (see Figure 11.1). The evolution of the state of the channel
over time can then be represented by trellis diagram having M L−1 states, where
M denotes the number of discrete symbols in the modulation alphabet. As
an illustration, we have represented in Figure 11.4 the trellis associated with
a channel having L = 3 coefficients h = (h0 , h1 , h2 ) in the case of a binary

phase-shift-keying (BPSK) modulation.
Figure 11.4 – Trellis representation for BPSK transmission on a frequency-selective

discrete-time channel with L = 3 paths.
Each candidate sequence takes a single path in the trellis. Searching for the
sequence with the minimum Euclidean distance from the observation can then
be performed recursively, with a linear computation cost depending on the size
of the message, by applying the Viterbi algorithm on the trellis of the channel.
The MLSD equalizer offers very good performance. However, the complexity
of its implementation increases proportionally with the number of states in the
trellis, and therefore exponentially with duration L of the impulse response of
the channel and size M of the modulation alphabet. Its practical utilization is
consequently limited to transmissions using modulations with a small number of
states (2, 4, or 8) on channels with few echoes. On the other hand, it should be
noted that this equalizer requires prior estimation of the impulse response of the
channel in order to build the trellis. The MLSD solution has been adopted by
many manufacturers to perform the equalization operation in mobile telephones
for the worldwide GSM (Global System Mobile) standard.
In the presence of modulations with a large number of states or on channels
whose impulse response length is large, the MLSD equalizer has an unaccept-
able computation time for real-time applications. An alternative strategy then
involves combating the ISI with the help of equalizers having less complexity,
implementing digital filters.
In this perspective, the simplest solution involves applying a linear transverse
filter at the output of the channel. This filter is optimized so as to compensate
("equalize") the irregularities of the frequency response of the channel, with
the aim of converting the frequency selective channel into an equivalent ide-
ally ISI-free (or frequency-flat) channel, perturbed only by additive noise. The
transmitted message is then estimated thanks to a simple operation of symbol
by symbol decision (threshold detector) at the output of the equalizer, optimal
on an additive white Gaussian noise (AWGN) channel. This equalizer, shown

in Figure 11.5, is called a linear equalizer or LE.
Figure 11.5 – Linear equalizer.
We distinguish several optimization criteria to define the coefficients of the

transverse filter. The optimal criterion involves minimizing the symbol error
probability at the output of the filter, but its application leads to a system of
equations difficult to solve. In practice, we prefer criteria sub-optimal in terms
of performance, but leading to solutions easily implementable, like the Minimum
Mean Square Error or MMSE criterion [11.44]. The linear MMSE equalizer is an
attractive solution due to its simplicity. However, this equalizer suffers from the
problem of amplification of the noise level on highly selective channels having
strong attenuations at certain points in the frequency spectrum.
Figure 11.6 – Decision-feedback equalizer.
Examining the diagram of the principle of the linear equalizer, we can note
that when we take a decision on symbol xi at instant i, we have an estimation
on the previous symbols x̂i−1 , x̂i−2 , . . . We can therefore envisage rebuilding
the (causal) interference caused by these data and therefore cancel it, in order to
improve the decision. The equalizer which results from this reasoning is called a
Decision-Feedback Equalizer or DFE. The diagram of the principle of the device
is illustrated in Figure 11.6. It is made up of a forward filter, in charge of
converting the impulse response of the channel into a purely causal response,
followed by a decision device and a feedback filter, in charge of estimating the
residual interference at the output of the feedback filter in order to cancel it via
a feedback loop.
As a general rule, the DFE provides performance higher than that of the lin-
ear equalizer, particularly on channels that are highly frequency selective. How-
ever, this equalizer is non-linear in nature, due to the presence of the decision
device in the feedback loop, which can give rise to an error propagation phe-
nomenon (particularly at low signal to noise ratio) when some of the estimated
data are incorrect. In practice, the filter coefficients are generally optimized
following the MMSE criterion, by assuming that the estimated data are equal
to the transmitted data, in order to simplify the computations (see Chapter 10

in [11.44], for example).
Figure 11.7 – Interference canceller.
If we assume now that we have a estimation x̂i+l on the transmitted data

both before (l < 0) and after (l > 0) the symbol considered at instant i, we
can then envisage removing the whole of the ISI at the output of the channel.
The equalization structure obtained is called an interference canceller, or IC [11-
6,11-7]. It is detailed in Figure 11.7. This structure is made up of two digital
transverse filters, with finite impulse response: a forward filter (matched to the
channel) whose aim is to maximize the signal to noise ratio before the decision,
and a canceller filter, in charge of rebuilding the ISI present at the output of the
matched filter. Note that by construction, the central coefficient of the canceller
filter is necessarily null in order to avoid subtracting the useful signal. With the
reserve that the estimated data x̂i+l be equal to the transmitted data, we can
show that this equalizer eliminates all the ISI, without any increase in noise level.
We thus reach the matched-filter bound, which represents what we can best do
with an equalizer on a frequency selective channel. Of course, we never know a
priori the transmitted data in practice. The difficulty then lies in building an
estimation of the data that is sufficiently reliable to keep performance close to
optimal.
None of the equalizer structures presented so far take into account the pres-
ence of a possible error correcting code on transmission. We shall now see how
we can best combine the equalization and decoding functions to improve the
global performance of the receiver.
11.1.3 Combining equalization and decoding

Most single-carrier digital transmission systems operating on frequency selective
channels incorporate an error correction coding function before the actual mod-
ulation step at transmission. The error correcting code is traditionally inserted
to combat the errors caused by the additive noise on the link. However, coupled
with a carefully built interleaving function, the encoder also offers an additional
degree of protection faced with power fading caused by the channel, when the
characteristics of the latter vary over time. We saw in the previous section that
independently of the nature of the equalizer used, the ISI systematically leads
to a loss in performance compared with a non-selective AWGN channel. The
presence of the encoder can then be exploited to reduce this gap in performance,
by benefiting from the coding gain at reception.
In the following part of this section, we are going to examine a transmission
system shown in Figure 11.8. The modulation and transmission operations on
the channel are here represented in equivalent baseband, in order not to have to
consider a carrier frequency.
Figure 11.8 – Baseband model of the transmission system considered.
The source sends a sequence of Q information bits, d = (d0 , . . ., dQ−1 ). This

message is protected by a convolutional code of rate R, to provide a sequence
c = (c0 , . . ., cP −1 ) of P = Q/R coded bits. The coded bits are then interleaved
following a permutation Π, then finally converted in groups of m successive bits
into discrete complex symbols chosen in an alphabet with M = 2m elements,
that we will denote {X1 , . . ., XM }. This is the mapping operation. We thus
obtain a sequence of N = P/m complex symbols: x = (x0 , . . ., xN −1 ). Later,
the vector of m coded and interleaved binary elements associated with symbol xi
at instant i will be denoted (xi,1 , . . . , xi,j , . . . , xi,m ). The transmitted symbols
are discrete random variables with zero mean and variance σx2 = E{|xi |2 }.
This transmission scheme is called Bit Interleaved Coded Modulation or
BICM. We still find it today in many systems: the mobile telephony standard
GSM, for example. For voice transmission at 13 kbits/s (TCH/FS channel), the
specifications of the radio interface indicate that at the output of the speech
encoder, the most sensitive bits (class 1A and 1B bits) are protected by a convo-
lutional code with rate R = 1/2 and generator polynomials (33,23) in octal [11.1].
The coded message is then interleaved on 8 consecutive packets (or bursts), to
benefit from time-diversity in the presence of fading, then finally modulated
following a waveform of the Gaussian Minimum Shift Keying (GMSK) type.
If we now return to the scenario of Figure 11.8, sequence x is transmit-
ted within a frequency selective channel with L coefficients, with discrete im-
pulse response h = (h0 , . . ., hL−1 ). The resulting signal is then perturbed by a
vector w with complex centred AWGN samples and variance σw 2
= E{|wi |2 }.
The noisy sequence observed at the output of the channel is finally denoted
y = (y0 , . . ., yN −1 ), the expression of sample yi at instant i being given by
Equation (11.3).
In this context, the problem that faces the designer of the receiver is the
following: how can we best combine equalization and decoding, so that each
processing benefits from the result of the other processing?
In reply to this question, estimation theory tells us that to minimize the

error probability in this case, the equalization and decoding operations must
be performed jointly, following the maximum likelihood criterion. Conceptu-
ally, implementing the optimal receiver then amounts to applying the Viterbi
algorithm, for example, on a "super-trellis" simultaneously taking into account
the constraints imposed by the code, the channel and the interleaver. However,
the "super-trellis" has a number of states that increases exponentially with the
size of the interleaver, which excludes a practical implementation of the optimal
receiver. It is therefore legitimate to question the feasibility of such a receiver
in the absence of an interleaver. Historically, this question has been asked in
particular in the context of data transmission over twisted-pair telephone ca-
bles (voice-band modems). These systems implement error correction coding
in Euclidean space (trellis coded modulations), without interleaving, and the
telephone channel is a typical example of a frequency-selective, time-invariant
channel. Assuming an encoder with S states, a constellation of M points and
a discrete channel with L coefficients, the studies undertaken in this context
have shown that the corresponding "super-trellis" has exactly S(M/2)L−1 states
[11.13]. It is then easy to verify that in spite of the absence of an interleaver, the
complexity of the optimal receiver again rapidly becomes prohibitive, whenever
we wish to transmit a high rate of information (with modulations having a large
number of states) or when we are confronted with a channel having large delays.
To counter the unaffordable complexity of the optimal receiver, the solution
commonly adopted in practice involves performing the equalization and decoding
operations disjointly, sequentially in time. If we again take the example of GSM,
the received data are thus first processed by an MLSD equalizer. The estimated
sequence provided by the equalizer is then transmitted, after deinterleaving, to
a Viterbi decoder. The permutation function then plays a twofold role in this
context: not only combating slow fading on the channel, but also dispersing
error packets at the input of the decoder. This strategy presents the advantage
of simplicity of implementation, since the total complexity is then given by the
sum of the individual complexities of the equalizer and the decoder. However, it
necessarily leads to loss in performance compared with the optimal receiver since
the equalization operation does not exploit all the available information. To be
more precise, the estimation sent by the equalizer will not necessarily correspond
to a valid coded sequence since the equalizer does not take into account the
constraints imposed by the code. The performance of the disjoint solution can be
improved when we introduce the passing of weighted (probabilistic) information
instead of an exchange of binary data between the equalizer and the decoder. By
propagating a reliability measure on the decisions of the equalizer, the decoder
thus benefits from additional information to produce its own estimation of the
message, and we benefit from a correction gain generally of the order of several
dB (see for example [11.28, 11.23] or Chapter 3 in [11.15]). Despite this, the
drawback of this solution remains: only one-way communication between the

equalizer and the decoder.
Therefore, does a strategy exist that can somehow produce the best of both
worlds, capable of reconciling both good performance of the optimal joint re-
ceiver and simplicity in implementation of the sub-optimal disjoint receiver?
Today, it is possible to reply in the affirmative, thanks to what we have called
"turbo equalization".
11.1.4 Principle of turbo equalization

The concept of turbo equalization first saw the light of day in the laboratories
of ENST Bretagne at the beginning of the 90s, under the impulsion of the spec-
tacular results obtained with turbo codes. It was the outcome of a very simple
realization: the transmission scheme in Figure 11.8 can be seen as the serial
concatenation of two codes (Chapter 6), separated by an interleaver, the second
code being formed by cascading the mapping operation with the channel2 . Seen
from this angle, it would then seem natural to apply a decoding strategy of the
"turbo" type at reception, that is, a reciprocal, iterative exchange of probabilis-
tic information (extrinsic information) between the equalizer and the decoder.
The first turbo equalization scheme was proposed in 1995 by Douillard et al.
[11.12]. This scheme implements a weighted input and output (Soft Input Soft
Output, or SISO) Viterbi equalizer according to the Soft Output Viterbi Algo-
rithm (SOVA). The principle was then used in 1997 by Bauch et al., substituting
the SOVA equalizer by a SISO equalizer that was optimal in the sense of the
MAP criterion, using the algorithm developed by Bahl et al. (BCJR algorithm
[11.7]) .
The simulation results quickly showed that the turbo equalizer was capable
of totally removing ISI, under certain conditions. Retrospectively, this excellent
performance can be explained by the fact that this transmission scheme brings
together two key ingredients which are the force of the turbo principle:
1. The implementation of iterative decoding at reception, introducing an

exchange of probabilistic information between the processing operations,
about which we today know that, when the signal to noise ratio exceeds a
certain "convergence threshold", it converges towards the performance of
the optimal joint receiver after a certain number of iterations.
2. The presence of an interleaver at transmission, whose role here mainly
involves breaking up the error packets at the output of the equalizer (to
avoid the phenomenon of error propagation), and decorrelating as far as
2 Note that, strictly speaking, transmission on a selective channel does not represent a coding
operation in itself, despite its convolutional character, as it does not provide any gain. In-
deed, it only degrades performance. Nevertheless, this analogy makes sense from the iterative
decoding point of view.
possible the probabilistic data exchanged between the equalizer and the
channel decoder. The turbo equalizer is then capable of totally compensat-
ing the degradation caused by the ISI, with the reserve that the interleaver
be large enough and carefully constructed.
Figure 11.9 – Turbo equalizer for BICM transmission systems.
Generally, the turbo equalizer corresponding to the transmission scenario

in Figure 11.8 takes the form shown in Figure 11.9. It is first made up of a
SISO equalizer, which at the input takes both vector y of the data observed
at the output of the channel, and a priori probabilistic information on the
set of coded, interleaved bits xi,j , here formally denoted La (x) = {La (xi,j )}.
The probabilistic information is propagated in the form of log likelihood ratios
(LLRs), the definition of which we recall here for a binary random variable d
with values in {0, 1}:
Pr(d = 1)
L(d) = ln (11.4)
Pr(d = 0)
The notion of LLR provides twofold information since the sign of the quantity
L(d) gives the hard decision on d, while its absolute value |L(d)| measures the
reliability this decision can be given.
From the two pieces of information y and La (x), the SISO equalizer produces
extrinsic information denoted Le (x) = {Le(xi,j )} on the coded, interleaved bi-
nary message. This vector Le (x) is then deinterleaved to give a new sequence
La (c), which is the a priori information on the sequence coded for the SISO
decoder. The latter then deduces two pieces of information from this: a hard
decision on the information message transmitted, here denoted d̂, and some new
extrinsic information on the coded message, denoted Le (c). This information is
then re-interleaved and sent back to the SISO equalizer where it is exploited as
a priori information for a new equalization step at the following iteration.
The turbo equalization scheme that we have presented above corresponds to
BICM transmitters. It is however important to note that the turbo equalization
principle also applies in the case of a system implementing traditional coded
modulation, that is to say, a system where the coding and modulation operations
are jointly optimized, on condition however that the symbols to be transmitted
are interleaved before being modulated and sent on the channel (Figure 11.10).
The main difference with the previous scheme thus lies in the implementation of
the SISO equalizer and SISO decoder. Indeed, these latter no longer exchange
probabilistic information at binary level but at symbol level, whether in LLR
form or directly in probability form. The interested reader can find further
details on this subject in [11.8], for example.
Figure 11.10 – Baseband model of traditional coded modulation systems.
As a general rule, the channel code is a convolutional code and the chan-
nel decoder uses a soft-input soft-output decoding algorithm of the MAP type
(or its derivatives in the logarithmic domain: Log-MAP and Max-Log-MAP).
Again, we will not consider the hardware implementation of the decoder since
this subject is dealt with in Chapter 7. Note, however, that unlike classical
turbo decoding schemes, the channel decoder here does not provide extrinsic
information on the information bits, but instead on the coded bits.
On the other hand, we distinguish different optimization criteria to imple-
ment the SISO equalizer, leading to distinct families of turbo equalizers. The
first, sometimes called "turbo detection" and what we call MAP turbo equal-
ization here, uses an equalizer that is optimal in the Maximum A Posteriori
sense. The SISO equalizer is then typically performed thanks to the BCJR-
MAP algorithm. As we shall see in the following section, this approach leads
to excellent performance, but like the classical MLSD equalizer, it has a very
high computation cost which excludes any practical implementation in the case
of modulations with a large number of states and for transmissions on channels
having large time delays. We must then turn towards alternative solutions, with
less complexity but that will necessarily be sub-optimal in nature. One strategy
that can be envisaged in this context involves reducing the number of branches
to examine at each instant in the trellis. This approach is commonly called
"reduced complexity MAP turbo equalization". We know different methods to
reach this result, which will be briefly presented in the following section. An-
other solution is inspired by classical equalization methods and implements an
optimized SISO equalizer following the minimum mean square error (MMSE)
criterion. We thus obtain an MMSE (filtering-based) turbo equalizer, a scheme
described in Section 11.1.6 and that appears as a very promising solution today
for high data rates transmissions on highly frequency-selective channels.
11.1.5 MAP turbo equalization

MAP turbo equalization corresponds to the turbo equalization scheme originally
introduced by Douillard et al. [11.12]. In this section, we first present the
equations for implementing the SISO equalizer. The good performance of the
MAP turbo equalizer is illustrated by simulation. We also introduce solutions
of less complexity derived from the MAP criterion. Finally, we examine the
problems encountered during a circuit implementation of the turbo equalizer, as
well as potential applications of this reception technique.
Implementation of the BCJR-MAP equalizer

The MAP equalizer shown in Figure 11.11 takes at its input vector y of the
discrete symbols observed at the output of the channel, as well as a priori
information denoted La (x) on the coded interleaved bits. This information
comes from the channel decoder and is produced at the previous iteration. In
the particular case of the first iteration, we do not generally have any a priori
information other than the hypothesis of equiprobability on the bits transmitted,
which leads us to put La (xi,j ) = 0.
Figure 11.11 – Block diagram of the MAP equalizer.
The purpose of the MAP equalizer is to evaluate the a posteriori LLR L(xi,j )
on each coded interleaved bit xi,j , defined as follows:

Pr(xi,j = 1 |y )
L(xi,j ) = ln (11.5)
Pr(xi,j = 0 |y )
Using conventional results in detection theory, we can show that this equal-
izer is optimal in the sense of the minimization of the symbol error probability.
To calculate the a posteriori LLR L(xi,j ), we will use the trellis representa-
tion associated with transmission on the frequency selective channel. Applying
Bayes’ relation, the previous relation can also be written:

Pr(xi,j = 1, y)
L(xi,j ) = ln (11.6)
Pr(xi,j = 0, y)
Among the set of possible sequences transmitted, each candidate sequence

traces a single path in the trellis. The joint probability Pr(xi,j = 0 or 1, y) can
then be evaluated by summing the probability Pr(s , s, y) of passing through a
particular transition in the trellis linking a state s at instant i − 1 to a state s
at instant i, on all of the transitions between instants i − 1 and i for which the
j-th bit making up the symbol associated with these transitions equals 0 or 1.
Thus, ⎛ ⎞
Pr(s , s, y)
⎜ (s ,s)/x

i,j =1 ⎟
L(xi,j ) = ln ⎝ ⎠ (11.7)
Pr(s , s, y)
(s ,s)/xi,j =0
Adopting a similar approach now to the one presented in the original article
by Bahl et al. [11.3], we can show that the joint probability Pr(s , s, y) associated
with each transition considered can be decomposed into a product of 3 terms:
Pr(s , s, y) = αi−1 (s )γi−1 (s , s)βi (s) (11.8)
Figure 11.12 shows the conventions of notation used here.
Figure 11.12 – Conventions of notation used to describe the MAP equalizer.
Forward and backward state probabilities αi−1 (s ) and βi (s) can be calcu-
lated recursively for each state and at each instant in the trellis, by applying the
following update equations:

αi (s) = αi−1 (s )γi−1 (s , s) (11.9)
(s ,s)

βi (s ) = γi (s , s)βi+1 (s) (11.10)
(s ,s)
These two steps are called forward recursion and backward recursion, respec-
tively. Summations are performed over all the couples of states with indices (s ,
s) for which there is a valid transition between two consecutive instants in the
trellis. Forward recursion uses the following initial condition:
α0 (0) = 1, α0 (s) = 1 for s = 0 (11.11)
This condition translates the fact that the initial state in the trellis (with
index 0, by convention) is perfectly known. Concerning the backward recursion,
we usually assign the same weight to each state at the end of the trellis since
the arrival state is generally not known a priori:
1
βN (s) = ∀s (11.12)
M L−1
In practice, we see that the dynamic of values αi−1 (s ) and βi (s) increases
during the progression in the trellis. Consequently, these values must be normal-
ized at regular intervals in order to avoid overflow problems in the computations.
One natural solution involves dividing these metrics at each instant by constants
Kα and Kβ chosen in such a way as to satisfy the following normalization con-
dition:
1 1
αi (s) = 1 and βi (s) = 1 (11.13)
Kα s Kβ s
the sums here concerning all possible states s of the trellis at instant i.
To complete the description of the algorithm, it remains for us to develop
the expression of the term γi−1 (s , s). This term can be assimilated to a branch
metric. We can show that it is decomposed into a product with two terms:
γi−1 (s , s) = Pr(s|s )P (yi |s , s) (11.14)
Pr(s|s ) represents the a priori probability of going through the transi-

tion between state s and state s , that is to say, the a priori probability
Pa (Xl ) = Pr(xi = Xl ) of having transmitted at time instant i the constel-
lation symbol Xl labeling the branch considered in the trellis. Owing to the
presence of the interleaver at transmission, bits xi,j composing symbol xi can
be assumed statistically independent. Consequently, probability Pa (Xl ) has the
following decomposition:
5
m
Pa (Xl ) = Pr(xi = Xl ) = Pa (Xl,j ) (11.15)
j=1
where we have written Pa (Xl,j ) = Pr(xi,j = Xl,j ), binary element Xl,j taking the
value 0 or 1 according to the symbol Xl considered and the mapping rule. Within
the turbo equalization iterative process, the a priori probabilities Pa (Xl,j ) are
deduced from the a priori information available at the input of the equalizer.
From the initial definition (11.4) of the LLR, we can in particular show that
probability Pa (Xl,j ) and corresponding a priori LLR La(xi,j ) are linked by the
following expression:
Pa (Xl,j ) = K exp (Xl,j La (xi,j )) with Xl,j ∈ {0, 1} (11.16)
The term K is a normalization constant that we can omit in the following

computations without compromising the final result in any way.
Conditional probability Pr(s|s ) is therefore finally given by:

⎛ ⎞
m
Pr(s |s ) = exp ⎝ Xl,j La (xi,j )⎠ (11.17)
j=1
As for the second term P (yi |s , s), it quite simply represents the likelihood
P (yi |zi ) of observation yi relative to branch label zi associated with the tran-
sition considered. The latter corresponds to the symbol that we would have
observed at the output of the channel in the absence of noise:

L−1
zi = hk xi−k (11.18)
k=0
The sequence of symbols (xi , xi−1 , . . ., xi−L+1 ) occurring in the computation

of zi is deduced from the knowledge of initial state s and of information symbol
Xl associated with transition s → s. In the presence of zero-mean circularly-
2
symmetric complex additive white Gaussian noise with total variance σw , we
obtain: " #
2
1 |yi − zi |
P (yi |s , s ) = P (yi |zi ) = 2
exp − 2
(11.19)
πσw σw
2
Factor 1/πσw is common to all the branch metrics and can therefore be
omitted in the computations. On the other hand, we see here that calculating
branch metrics γi−1 (s’,s) requires both knowledge of the impulse response of the
equivalent discrete channel and knowledge of the noise variance. In other words,
in the context of a practical implementation of the system, the MAP equalizer
will have to be preceded by a channel estimation procedure.
To summarize, after computing branch metrics γi−1 (s , s) then performing
the forward and backward recursions, the a posteriori LLR L(xi,j ) is finally
given by:
αi−1 (s )γi−1 (s , s)βi (s)
(s ,s)/xi,j =1
L(xi,j ) = ln (11.20)
αi−1 (s )γi−1 (s , s)βi (s)
(s ,s)/x i,j =0
In reality and in accordance with the turbo principle, it is not this a posteriori
information that is propagated to the SISO decoder, but rather the extrinsic
information. Here, the latter measures the equalizer’s own contribution in the
global decision process, excluding the information relating to the bit considered
coming from the decoder at the previous iteration, that is to say, the a priori
LLR La (xi,j ). If we develop the expression of branch metric γi−1 (s , s) in the
computation of L(xi,j ), we obtain:

⎡ ⎤
|yi −zi |2
m

α
⎢ (s ,s)/x i−1 (s ) exp − 2
σw + X L (x
l,k a i,k ) β i (s) ⎥
⎢ i,j =1

k=1
⎥
L(xi,j ) = ln ⎢ ⎥ (11.21)
⎣ |y −z | 2
m ⎦
αi−1 (s ) exp − σ2 + i i
Xl,k La (xi,k ) βi (s)
w
(s ,s)/xi,j =0 k=1
We can then factorize the a priori information term in relation to the bit xi,j
considered, both in numerator (Xl,j = 1) and in denominator (Xl,j = 0), which
gives:
L(xi,j ) = La (xi,j ) " #

⎡ ⎤
2
⎢

αi−1 (s ) exp − |yiσ−z
2
i|
+ Xl,k La (xi,k ) βi (s) ⎥
⎢ (s ,s)/xi,j =1 w
k=j ⎥
⎢
+ ln ⎢ " # ⎥
⎥ (11.22)
⎣ 2 ⎦
αi−1 (s ) exp − |yiσ−z
2
i|
+ Xl,k La (xi,k ) βi (s)
w
(s ,s)/xi,j =0 k=j
1 23 4
Le (xi,j )
Finally, we see that the extrinsic information is obtained quite simply by sub-
tracting the a priori information from the a posteriori LLR calculated by the
equalizer:
Le (xi,j ) = L(xi,j ) − La (xi,j ) (11.23)
This remark concludes the description of the MAP equalizer. As we have
presented it, this algorithm proves to be difficult to implement on a circuit due
to the presence of numerous multiplication operations. In order to simplify
the computations, we can then envisage transposing the whole algorithm into
the logarithmic domain (Log-MAP algorithm), the advantage being that the
multiplications are then converted into additions, which are simpler to do. If we
wish to further reduce the processing complexity, we can also use a simplified
(but sub-optimal) version, the Max-Log-MAP (or Sub-MAP) algorithm. These
two variants were presented in the context of turbo codes in Chapter 7. The
derivation is quite similar in the case of the MAP equalizer. Reference [11.5]
presents a comparison in performance between these different algorithms in a
MAP turbo equalization scenario. In particular, it turns out that the Max-Log-
MAP equalizer offers the best performance/complexity compromise when the
estimation of the channel is imperfect.
Example of performance
In order to illustrate the good performance offered by MAP turbo equalization,
we chose to simulate the following transmission scenario: a binary source gener-
ates messages of 16382 bits of information, which are then protected by a rate
R = 1/2 non-recursive non-systematic convolutional code with 4 states, and

with generator polynomials (5,7) in octal. Two null bits (tail-bits) are inserted
at the end of the message in order to force the termination of the trellis in state
0. Thus we obtain a sequence of 32768 coded bits, which are then randomly
interleaved and mapped into a sequence of BPSK symbols. These symbols are
transmitted on a 5-path discrete-time channel with impulse response:
h = (0.227, 0.460, 0.688, 0.460, 0.227)
Figure 11.13 – Performance of the MAP turbo equalizer for BPSK transmission on
the Proakis C channel, using a rate R = 1/2 4-state non-recursive non-systematic
convolutional code and a pseudo-random interleaver of size 32768 bits.
This channel model, called Proakis C, taken from Chapter 10 in [11.44], is

relatively difficult to equalize. At reception, we implement 10 iterations of the
MAP turbo equalizer described above. The SISO decoder is performed using the
BCJR-MAP algorithm. Figure 11.13 presents the bit error rate after decoding,
measured at each iteration, as a function of the normalized signal to noise ratio
Eb /N0 on the channel. For reference, we have also shown the performance
obtained after decoding on a non-frequency selective AWGN channel. This curve
shows the optimal performance of the system. We see that beyond a signal to
noise ratio of 5 dB, the turbo equalizer suppresses all the ISI after 5 iterations,
and we reach the ideal performance of the AWGN channel. Furthermore, for a
target bit error rate of 10−5 , the iterative process provides a gain of the order of
6.2 dB compared with the performance of the conventional receiver performing
the equalization and decoding disjointly, given by the curve at the 1st iteration.
This performance is very similar to that presented in reference [11.7].
These results give rise to a certain number of remarks, since the example
considered here presents the characteristic behaviour of turbo systems. In par-
ticular, we see that the gain provided by the iterative process only appears
beyond a certain signal to noise ratio (convergence threshold, equal to 3 dB
here). Beyond this threshold, we observe a rapid convergence of the turbo
equalizer towards the asymptotic performance of the system, given by the error
probability after decoding on a non-selective AWGN channel. To improve the
global performance of the system, we can envisage using a more powerful error
correcting code. Experience shows that we then come up against the necessity
of finding a compromise in choosing the code, between rapid convergence of the
iterative process and good asymptotic performance of the system (at high signal
to noise ratios). The greater the correction capacity of the code, the higher the
convergence threshold. On this topic, we point out that today there exist semi-
analytical tools such as EXIT (EXtrinsic Information Transfer ) charts [11.49],
enabling the value of the convergence threshold to be predicted precisely, as
well as the error rate after decoding for a given transmission scenario, under
the hypothesis of ideal interleaving (infinite size). A second solution involves
introducing a feedback effect in front of the equivalent discrete-time channel,
by inserting an adequate precoding scheme at transmission. Cascading the pre-
encoder with the channel produces a new channel model, recursive in nature,
leading to a performance gain that is greater, the larger the dimension of the
interleaver considered. This phenomenon is known as "interleaving gain" in the
literature dedicated to serial turbo codes. Subject to carefully choosing the pre-
encoder, we can then exceed the performance of classical non-recursive turbo
equalization schemes as has been shown in [11.35] and [11.39].
Complexity of the MAP turbo equalizer and alternative solutions

The complexity of the MAP turbo equalizer is mainly dictated by the complexity
of the MAP equalizer. Now, the latter increases proportionally with the number
of branches to examine at each instant in the trellis. Considering a modulation
with M states and a discrete channel with L coefficients, the total number of
branches per section of the trellis rises to M × M L−1 = M L . We therefore see
that the processing cost associated with the MAP equalizer increases exponen-
tially with the number of states of the modulation and the length of the impulse
response of the channel. As an illustration, EDGE (Enhanced Data Rate for
GSM Evolution) introduces the use of 8-PSK modulation on channels with 6
coefficients maximum, that is, slightly more than 262000 branches to examine
at each instant! MAP turbo equalization is therefore an attractive solution for
modulations with a low number of states (typically BPSK and QPSK) on chan-
nels having ISI limited to a few symbol periods. Beyond that, we must turn to
less complex, but less efficient, solutions.
There are several ways to deal with this problem. If we limit ourselves to us-
ing equalizers derived from the MAP criterion, one idea is to reduce the number
of paths examined by the algorithm in the trellis. A first approach performs a
truncation of the channel impulse response in order to keep only the J < L first
coefficients. The number of states in the trellis will then be decreased. The ISI
terms ignored in the definition of the states are then taken into account when
calculating the branch metrics, using past decisions obtained from the knowl-
edge of the survivor path in each state. This strategy is called Delayed Decision
Feedback Sequence Estimation (DDFSE). It offers good performance provided
most of the channel’s energy be concentrated in its first coefficients which, in
practice, requires the implementation of a minimum-phase pre-filtering oper-
ation. Applying this technique to turbo equalization has, for example, been
studied in [11.2]. A refinement of this algorithm involves grouping some states
of the trellis together, in accordance with the set-partitioning rules defined by
Ungerboeck [11.52] for designing trellis coded modulations. This improvement,
called Reduced State Sequence Estimation (RSSE), includes DDFSE as a par-
ticular case [11.19]. In a similar way, we can also envisage retaining more than
one survivor path in each state to improve the robustness of the equalizer and if
necessary to omit the use of pre-filtering [11.42]. Rather than reduce the number
of states of the trellis by truncation, it can also be envisaged to examine only
a non-exhaustive list of the most likely paths at each instant. The resulting
algorithm is called the "M algorithm", and its extension to SISO equalization
was studied in [11.17]. Whatever the case, the search for efficient equalizers with
reduced complexity regularly continues to give rise to new contributions.
All the strategies that we have mentioned above enter into the category of
MAP turbo equalizers with reduced complexity. Generally, these solutions are
interesting when the number of states of the modulation is not too high. On
the other hand, in the case of high data rate transmissions on channels with
long delay spreads, it is preferable to envisage filter-based turbo equalizers of
the MMSE type.
Architectures and applications

When systems based on MAP turbo equalization require real time processing
with relatively high data rates (of the order of several Mbits/s), a software im-
plementation cannot be envisaged. In this case, we must resort to specific ASIC
circuits. The circuit implementation of a MAP turbo equalizer poses problems
similar to those encountered in the context of the hardware implementation of
a turbo decoder. Two architectural solutions can be envisaged:
• The first uses an implementation of the turbo decoder in the form of

a pipeline by cascading several elementary modules, each module imple-
menting one detection and one decoding iteration.
• The second uses a single hardware module, implementing the successive

iterations sequentially, by looping back on itself.
The first architecture presents a smaller latency, so is better adapted to applica-

tions requiring high data rates. On the other hand, the second solution enables
an economy in the number of transistors and therefore in the silicon surface. In
order to further reduce the surface used, some authors have proposed sophisti-
cated architectures enabling part of the SISO algorithm to be shared between
the equalizer and the decoder, despite the different structure of the trellises con-
cerned [11.36]. This approach also enables a reduction in length of the critical
path, and therefore in the global latency of the system. This last factor can be a
major obstacle to the practical implementation of turbo equalization (and turbo
systems more generally) since not all applications may tolerate an increase in
the processing delay at reception. Resorting to analogue electronics will perhaps
soon enable this obstacle to be overcome. An analogue implementation of a sim-
plified MAP turbo equalizer has thus been reported in [11.24], with promising
results.
From the algorithmic point of view, the application of MAP turbo equaliza-
tion to the GSM system has been the subject of several studies [11.15, 11.43,
11.6, 11.18]. The traditional turbo equalization scheme must thus be revised in
order to take into account the specificities of the standard (inter-frame interleav-
ing, different levels of protection applied to the bits at the output of the speech
encoder, GMSK modulation, . . . ). Simulation results show generally moderate
gains in performance, in return for a large increase in the complexity of the
receiver. This can be partly explained by the fact that the conventional GSM
system faces only a limited level of ISI on the majority of the test channels
defined in the standard. On the other hand, the introduction of 8-PSK modu-
lation in the context of EDGE greatly increases the level of interference. This
scenario therefore seems more appropriate for the use of turbo equalization, and
has given rise to several contributions. In particular, the authors of [11.40]3 have
studied the implementation of a complete turbo equalization system relying on
a SISO equalizer of the DDFSE type with pre-filtering, coupled to a channel
estimator. They have obtained gains of the order of several dB, depending on
the modulation and coding scheme considered, compared with the performance
of the classical receiver. Furthermore, they have also proved the fact that the
equalization and iterative decoding principle could be carefully exploited in the
context of the ARQ retransmission protocol defined in EDGE (the Incremental
Redundancy scheme) to improve the global quality of service at reception.
3 See also the references cited in this article.
11.1.6 MMSE turbo equalization

The increase in data rates, in response to current multimedia service require-
ments, combined with the infatuation with mobility and wireless infrastructures,
present receivers with severe propagation conditions. Thus, if we take the ex-
ample of the radio interface of the Wireless MAN (Metropolitan Area Network )
802.16a standard normalized by IEEE during 2003 and operating in the 2-11
GHz band, the ISI encountered is likely to recover up to 50 symbol durations,
or even more. Underwater acoustic communications is another example. The
application of turbo equalization to such scenarios involves using low complex-
ity SISO equalizers. MMSE turbo equalization is an attractive solution in this
context.
In contrast with the approaches described in the previous section, MMSE
turbo equalization mainly involves substituting for the MAP equalizer an equal-
izer structure based on digital filters, optimized according to the minimum mean
square error criterion4 . This solution presents a certain number of advantages.
First of all, simulations show that the MMSE turbo equalizer gives very good
performance on average, sometimes very close to the performance offered by
MAP turbo equalization. On the other hand, the complexity of the MMSE
equalizer increases linearly (and not exponentially) with the length of the chan-
nel impulse response, independently of the order of the modulation. Finally,
as we shall see in what follows, this approach naturally lends itself well to an
implementation in adaptive form, appropriate for tracking the time variations
of the channel.
Historically, the first MMSE turbo equalization scheme was proposed by
Glavieux et al. in 1997 [11.20, 11.32, 11.34]. This original contribution laid
down the bases of MMSE turbo equalization, particularly for the design of a
filter-based soft-input soft-output equalizer. Indeed, classical equalizers based
on digital filters do not naturally lend themselves to handling probabilistic infor-
mation. This difficulty was overcome by inserting a binary to M -ary conversion
operation at the input of the equalizer, in charge of rebuilding a soft estimation
of the symbols transmitted using the a priori information sent by the decoder.
In addition, a SISO demapping module placed at the output of the equalizer
converts the equalized data (complex symbols) into extrinsic LLR on the coded
bits, which are then sent to the decoder. This initial scheme relied on the imple-
mentation of an equalization structure of the interference canceller type, whose
coefficients were updated adaptively thanks to the Least Mean Square (LMS)
algorithm.Least Mean Square Remarkable progress was then achieved with the
work of Wang and Poor [11.56], taken up by Reynolds and Wang [11.47] then
by Tüchler et al. [11.51, 11.50]. These contributions have made it possible to
4 Equalizers optimized according to the Zero Forcing criterion could also be envisaged. How-
ever these equalizers usually introduce significant noise enhancement on channels with deep
nulls in their frequency response, and thus generally turn out to be less efficient than MMSE
equalizers.
establish a theoretical expression for the coefficients of the equalizer, explicitly

taking into account the presence of a priori information on the transmitted
data. This progress has proved to be particularly interesting for packet mode
transmissions, in which the coefficients of the equalizer are precalculated once
from an estimation of the impulse response of the channel, and applied to the
whole received block.
MMSE turbo equalization relies on a soft-input soft-output linear equaliza-
tion scheme optimized according to the MMSE criterion. This type of equalizer
is also sometimes known as a "linear MMSE equalizer with a priori information"
in the literature. This section describes the principle of this equalizer, assuming
that we know the parameters of the channel, which enables the filter coefficients
to be calculated directly. Its implementation in adaptive form is also discussed.
We next present some examples of MMSE turbo equalizer performance, and we
describe the (Digital Signal Processor, or DSP) implementation of this solution.
This part ends with a reflection on the potential applications of MMSE turbo
equalization.
Principle of soft-input soft-ouput linear MMSE equalization

Generally, the linear soft-input soft-output MMSE equalizer can formally be
decomposed into three main functions (Figure 11.14).
Figure 11.14 – Soft-input soft-output linear equalizer optimized according to the

MMSE criterion.
1. The first operation, the SISO mapping, calculates a soft estimate for the
transmitted symbols, denoted x̄ = (x̄0 , . . . , x̄N −1 ), from the a priori in-
formation La (x) coming from the decoder at the previous iteration.
2. The linear equalizer then uses estimated data x̄i to rebuild and cancel the
ISI affecting the received signal. The resulting signal is filtered in order
to eliminate residual interference. The filter coefficients are optimized so
as to minimize the mean square error between the equalized data and
the corresponding transmitted data. However, unlike the classical linear

MMSE equalizer, the reliability information coming from the decoder is
here explicitly taken into account when calculating the coefficients.
3. The equalizer is finally followed by a soft-input soft-output demapping

module whose role is to convert the equalized symbols into extrinsic LLRs
on the (interleaved) coded bits.
We now examine in greater detail the implementation of each of these three

functions.
• SISO mapping
This operation involves calculating the soft estimate x̄i , defined as the math-
ematical expectation of symbol xi transmitted at instant i:

M
x̄i = Ea {xi } = Xl × Pa (Xl ) (11.24)
l=1
The sum here concerns all of the discrete symbols in the constellation. The
term Pa (Xl ) denotes the a priori probability Pr(xi = Xl ) of symbol Xl being
transmitted at instant i. We have put index a at the level of the expectation
term to highlight the fact that these probabilities are deduced from the a priori
information at the input of the equalizer. Indeed, provided the m bits making
up symbol xi are statistically independent, it is possible to write:
5
m
Pa (Xl ) = Pa (Xl,j ) (11.25)
j=1
where binary element Xl,j takes the value 0 or 1 according to the symbol Xl
considered and the mapping rule. On the other hand, starting from the general
definition (11.4) of the LLR, we can show that the a priori probability and the
a priori LLR are linked by the following relation:

1 La (xi,j )
Pa (Xl,j ) = 1 + (2Xl,j − 1) tanh with Xl,j ∈ {0, 1} (11.26)
2 2
In the particular case of a BPSK modulation, the above computations are

greatly simplified. We then obtain the following expression for the soft esti-
mate x̄i :
La (xi )
x̄i = tanh (11.27)
2
In the classical situation where we make the hypothesis of equiprobability
on the transmitted symbols, we have La (xi,j ) = 0 and x̄i = 0. On the other
hand, in the ideal case of perfect a priori information, La (xi,j ) → ±∞ and
the soft estimate x̄i is then strictly equal to the transmitted symbol xi (perfect
estimate). To summarize, the value of the soft estimate x̄i evolves as a function
of the reliability of the a priori information provided by the decoder. This
explains the name of "soft" (or probabilistic) estimate for x̄i . By construction,
the estimated data x̄i are random variables. In particular, we can show (see
[11.33] for example) that they satisfy the following statistical properties:
E {x̄i } = 0 (11.28)

E x̄i x∗j = E x̄i x̄∗j = σx̄2 δi−j (11.29)
The parameter σx̄2
here denotes the variance of estimated data x̄i . In practice,
this quantity can be estimated using the sample variance estimator on a frame
of N symbols as follows:
N −1
1 2
σx̄2 = |x̄i | (11.30)
N i=0
We easily verify that under the hypothesis of equiprobable a priori symbols,
σx̄2 = 0. Conversely, we obtain σx̄2 = σx2 in the case of perfect a priori informa-
tion on the transmitted symbols. To summarize, the variance of the estimated
data offers a measure of the reliability of the estimated data. This parameter
plays a major role in the behaviour of the equalizer.
• Calculating the linear equalizer coefficients

As explained above, the equalization step can be seen as the cascad-
ing of an interference cancellation operation followed by a filtering operation.
The filter coefficients are optimized so as to minimize the mean square error
E{|zi − xi−Δ |2 } between the equalized symbol zi at instant i and symbol xi−Δ
transmitted at instant i − Δ. The introduction of a delay Δ enables the anti-
causality of the solution to be taken into account. Here we will use a matrix
formalism to derive the optimal form of the equalizer coefficients. Indeed, dig-
ital filters always have a finite number of coefficients in practice. The matrix
formalism takes this aspect into account and thus enables us to establish the
optimal coefficients under the constraint of a finite-length implementation.
Here we consider a filter with F coefficients: f = (f0 , . . ., fF −1 ). The channel
impulse response and the noise variance are assumed to be known, which requires
prior estimation of these parameters in practice. Starting from the expression
(11.3) and grouping the F last samples observed at the output of the channel
up until instant i in the form of a vector column yi , we can write:
yi = Hxi + wi (11.31)
T T
with yi = (yi , . . . , yi−F +1 ) , xi = (xi , . . . , xi−F −L+2 ) and wi = (wi , . . . ,
wi−F +1 )T . Matrix H, of dimensions F × (F + L − 1), is a Toeplitz matrix5
5 The coefficients of the matrix are constant along each of the diagonals.
describing the convolution by the channel:

⎛ ⎞
h0 · · · hL−1 0 ··· 0
⎜ .. ⎟
⎜ 0 h0 hL−1 . ⎟
H=⎜ ⎜ .
⎟
⎟ (11.32)
⎝ . . . . .. ⎠
. . 0
0 ··· 0 h0 ··· hL−1
With these notations, the interference cancellation step from the estimated signal
x̄ can then be written formally:
ỹi = yi − Hx̃i (11.33)
where the vector x̃i = (x̄i , . . . , x̄i−Δ+1 , 0, x̄i−Δ−1 , . . . , x̄i−F −L+2 )T is of dimen-
sion F + L−1. The component related to symbol xi−Δ is set to zero in order to
cancel only the ISI and not the signal of interest. At the output of the forward
filter, the expression of the equalized sample at instant i is given by:
zi = f T ỹi = f T [yi − Hx̃i ] (11.34)
It remains to determine the theoretical expression of the coefficients of the

filter f minimizing the mean square error E{|zi − xi−Δ |2 }. In the most general
case, these coefficients vary in time. The corresponding solution, developed in
detail by Tüchler et al. [11.51, 11.50], leads to an equalizer whose coefficients
must be recalculated for each received symbol. This equalizer represents what
can best be done currently for MMSE equalization in the presence of a priori
information. On the other hand, the computation load associated with updating
the coefficients symbol by symbol increases quadratically with the number F of
coefficients, which again turns out to be too complex for real time implemen-
tations. The equalizer that we present here can be seen as a simplified, and
therefore sub-optimal, version of the solution cited above. The coefficients of
filter f are calculated only once per frame (at each iteration) and then applied
to the whole block, which considerably decreases the implementation cost. On
the other hand, and despite this reduction in complexity, this equalizer retains
performance close to the optimal one6 , which makes it an excellent candidate
for practical realizations. This solution was derived independently by several
authors, including [11.51] and [11.33].
With these hypotheses, the optimal form of the set of coefficients f is obtained
using the projection theorem, which stipulates that the estimation error must
be orthogonal to the observations7 :

E (zi − xi−Δ )ỹiH = 0 (11.35)
6 The degradation measured experimentally in comparison with Tüchler’s original time-
varying solution is at most 1 dB, depending on the channel model considered.
7 We recall here that the notation AH denotes the Hermitian (conjugate) transpose of ma-
trix A.
We then obtain the following solution:

−1
f ∗ = E{ỹi ỹiH } E{x∗i−Δ ỹi } (11.36)
Using the statistical properties of the estimated data x̄i , we note that:
E{x∗i−Δ ỹi } = E{x∗i−Δ H(xi − x̃i )} = HeΔ σx2 (11.37)
where we have introduced the unit vector eΔ with dimension F + L − 1 that

has a 1 in coordinate Δ and 0 elsewhere. Denoting by hΔ the Δ-th column Δ
of matrix H, the previous expression can also be written:

E x∗i−Δ ỹi = hΔ σx2 (11.38)
In addition,

E ỹi ỹiH = HE (xi − x̃i )(xi − x̃i )H HH + σw
2
I
(11.39)
= (σx2 − σx̄2 )HHH + σx̄2 hΔ hH 2
Δ + σw I
To summarize, the optimal form of the equalizer coefficients can finally be writ-
ten: −1
f ∗ = (σx2 − σx̄2 )HHH + σx̄2 hΔ hH 2
Δ + σw I hΔ σx2 (11.40)
By bringing into play a simplified form of the matrix inversion lemma8 , the
previous solution can then be written:
σx2
f∗ = f̃ ∗ (11.41)
1 + βσx̄2
where we have introduced vector f̃ and scalar quantity β defined as follows:

−1
f̃ ∗ = (σx2 − σx̄2 )HHH + σw
2
I hΔ and β = f̃ T hΔ (11.42)
By means of this new expression, we note that the computation of the

coefficients of the equalizer is mainly based on the inversion of the matrix
(σx2 − σx̄2 )HHH + σw2
I, with dimensions F × F . This matrix has a rich structure
since it is a Toeplitz matrix with Hermitian symmetry. Consequently, matrix in-
version can be performed efficiently with the help of dedicated algorithms, with
a computation cost in O(F 2 ) (see Chapter 4 in [11.21], for example). In order to
reduce even further the complexity of determining the coefficients, the authors
of [11.33] have proposed a sub-optimal, but nevertheless efficient, method using
the Fast Fourier Transform, (or FFT), with a cost in O(F log2 (F )). However,
the number of coefficients F must be a power of 2.
8
−1 A−1 uuH A−1
A + uuH = A−1 − 1+uH A−1 u
It is particularly instructive to study the limiting form taken by the equalizer

in the classical case where the transmitted symbols are assumed to be equiprob-
able (which corresponds to the 1st iteration of the turbo equalizer). In this case,
σx̄2 = 0 and the equalizer coefficients can be written:
−1
f ∗ = σx2 HHH + σw
2
I hΔ σx2 (11.43)
Here we can recognize the form of a classical linear MMSE equalizer with finite
length. Inversely, under the hypothesis of perfect a priori information on the
transmitted symbols, we have σx̄2 = σx2 . The equalizer then takes the following
form:
σx2
L−1
h∗Δ
2 2
f= 2 with h = hH
Δ hΔ = |hk | (11.44)
σx2 h + σw
2
k=0
and the equalized signal zi can be written:
σx2 h
2
- .
zi = 2 xi−Δ + hH
Δ wi (11.45)
σx2 h + σw
2
We recognize here the output of a classical MMSE interference canceller, fed

by a perfect estimation of the transmitted data. The equalized signal can be
decomposed as the sum of the useful signal xi−Δ , up to a scale factor that is char-
acteristic of the MMSE criterion, and a coloured noise term. In other words, the
equalizer suppresses all the ISI without raising the noise level and thus reaches
the theoretical matched-filter bound corresponding to ISI-free transmission.
To summarize, we see that the SISO MMSE linear equalizer adapts the
equalization strategy according to the reliability of the estimated data, measured
here by parameter σx̄2 .
To conclude this description of the equalizer, we point out that the interfer-
ence cancellation operation defined formally by Equation (11.33) has no physical
reality in the sense that it cannot be performed directly in this way using trans-
verse linear filters. In practice, we prefer to use one of the two architectures
presented in Figure 11.15, strictly equivalent from a theoretical point of view.
The coefficient gΔ appearing in implementation (1) is the central coefficient
gΔ = f T hΔ of the global filter formed by the cascade of the channel with filter
f . In the case of implementation (2), we again find the classical structure of an
interference canceller type equalizer, operating here on the estimated signal x̄.
Filter g = f T H is given by the convolution of filter f with the impulse response
of the channel, the central coefficient gΔ then being forced to zero.
•SISO demapping
The role of this module is to convert the equalized data zi into extrinsic LLRs
on the interleaved coded bits, which will be then transmitted to the channel
Figure 11.15 – Practical implementation of the equalizer using transverse filters.
decoder. Generally, we can always decompose the expression of zi as the sum of

two quantities:
zi = gΔ xi−Δ + νi (11.46)
The term gΔ xi−Δ represents the useful signal up to a constant factor gΔ . We
recall that this factor quite simply corresponds to the central coefficient of the
cascading of the channel with the equalizer. The term νi accounts for both
residual interference and noise at the output of the equalizer. In order to perform
the demapping operation, we make the hypothesis9 that interference term νi
follows a complex Gaussian distribution, with zero mean and total variance σν2 .
Parameters gΔ and σν2 are easy to deduce from the knowledge of the set of
equalizer coefficients. We can thus show ([11.51, 11.33]) that we have:
) *
2
gΔ = f T hΔ and σν2 = E |zi − gΔ xi−Δ | = σx2 gΔ (1 − gΔ ) (11.47)
Starting from these hypotheses, the demapping module calculates the a poste-
riori LLR on the coded interleaved bits, denoted L(xi,j ) and defined as follows:

Pr(xi,j = 1 |zi )
L(xi,j ) = ln (11.48)
Pr(xi,j = 0 |zi )
The values present in the numerator and denominator can be evaluated by sum-
ming the a posteriori probability Pr(xi = Xl |zi ) of having transmitted a par-
ticular symbol Xl of the constellation on all the symbols for which the j-th bit
making up this symbol takes the value 0 or 1 respectively. Thus, we can write:
9 This hypothesis rigorously only holds on condition that the equalizer suppresses all the
ISI, which assumes perfect knowledge of the transmitted data. Nevertheless, it is a good
approximation in practice, particularly in a turbo equalization context where the reliability
of the decisions at the output of the decoder increases along the iterations which, in its turn,
improves the equalization operation.
"
Pr(xi =Xl |zi )
#
Xl /Xl,j =1
L(xi,j ) = ln
Pr(xi =Xl |zi )
Xl /Xl,j =0
(11.49)
"
P (zi |xi =Xl )Pa (Xl )
#
Xl /Xl,j =1
= ln
P (zn |xi =Xl )Pa (Xl )
Xl /Xl,j =0
The second equality results from applying Bayes’ relation. It shows the a priori
probability Pa (Xl ) = Pr(xi = Xl ) of having transmitted a given symbol Xl of
the modulation alphabet. This probability is calculated from the a priori infor-
mation available at the input of the equalizer (relations (11.25) and (11.26)). By
exploiting the above hypotheses, the likelihood of observation zi conditionally to
the hypothesis of having transmitted the symbol Xl at instant i can be written:
" #
2
1 |zi − gΔ Xl |
P (zi |xi = Xl ) = exp − (11.50)
πσν2 σν2
After simplification, the a posteriori LLR calculated by the demapping operation
becomes:
⎛ ⎞
|zi −gΔ Xl |2 m
⎜ X /X =1 exp − σν2
+ X L (x
l,k a i,k ) ⎟
⎜ l l,j ⎟
k=1
L(xi,j ) = ln ⎜ ⎟ (11.51)
⎝ |zi −gΔ Xl |2 m ⎠
exp − σ2
+ Xl,k La (xi,k )
ν
Xl /Xl,j =0 k=1
Like in the case of the BCJR-MAP equalizer, we can factorize in the numerator
and denominator the a priori information term in relation to the considered bit,
in order to obtain the extrinsic information that is then provided to the decoder:
⎛ " #⎞
|zi −gΔ Xl |2
⎜ exp − σν2
+ Xl,k La (xi,k ) ⎟
⎜ Xl /Xl,j =1 k=j ⎟
L(xi,j ) = La (xi,j ) + ln ⎜
⎜ " #⎟
⎟
⎝ 2 ⎠
exp − |zi −gσΔ2 Xl | + Xl,k La (xi,k )
ν
Xj /Xj,i =0 k=j
1 23 4
Le (xi,j )
(11.52)
Finally, the extrinsic information is obtained quite simply by subtracting the a
priori information from the a posteriori LLR calculated by the equalizer:
Le (xi,j ) = L(xi,j ) − La (xi,j ) (11.53)
In the particular case of BPSK modulation, the SISO demapping equations are
simplified to give the following expression of the extrinsic LLR:
4
L(xi ) = Re {zi } (11.54)
1 − gΔ
When Gray mapping rules are used, experience shows that we can reduce the
complexity of the demapping by ignoring the a priori information coming from
the decoder in the equations above10 , without really affecting the performance
of the device. On the other hand, this simplification no longer applies when
we consider other mapping rules, like the Set Partitioning rule used in coded
modulation schemes. This point has been particularly well highlighted in [11.14]
and [11.30].
This completes the description of the soft-input soft-output linear MMSE
equalizer. Finally, we can note that unlike the BCJR-MAP equalizer, the com-
plexity of the SISO mapping and demapping operations increases linearly (and
not exponentially) as a function of size M of the constellation and of the number
L of taps in the impulse response of the discrete-time equivalent channel model.
Adaptive implementation of the equalizer

Historically, the first MMSE turbo equalizer was proposed in 1997, directly in
adaptive form [11.20, 11.32]. The closed-form expression (11.40) enabling the
computation of the equalizer coefficients from the knowledge of the channel im-
pulse response was not known at that time. The chosen solution thus involved
determining the filter coefficients in an adaptive manner, using stochastic gra-
dient descent algorithms aiming at minimizing the mean square error between
the transmitted data and the equalizer output. As we shall see in the following,
when evaluating performance, the adaptive MMSE turbo equalizer remains a
very interesting solution for time-invariant or slowly time-varying channels. The
purpose of this section is to show that, for such channels, the adaptive MMSE
equalizer and the MMSE equalizer proposed in (11.40) have similar performance
and characteristics.
The structure of the considered equalizer is shown in Figure 11.15 (imple-
mentation (2)). An adaptive procedure is used to obtain the filters’ coefficients.
This adaptive algorithm is composed of two distinct phases: the training phase
and the tracking phase. The training phase makes use of sequences known by the
receiver (training sequences) to initialize the equalizer coefficients. Next, during
the tracking period, the coefficients are continuously updated in a decision-
directed manner, based on the receiver estimate of the transmitted sequence.
Adaptive algorithms involve determining, for each symbol entering the equal-
izer, output zi from the following relation:
zi = fiT yi − giT x̃i (11.55)

T
where yi = (yi+F , . . . , yi−F ) is the vector of channel output samples and
T
x̃i = (x̄i+G , . . . , x̄i−Δ+1 , 0, x̄i−Δ−1 , . . . , x̄i−G ) is the vector of estimated sym-
bols, with respective lengths 2F +1 and 2G+1. Note that the coordinate relative
10 This amounts to assuming the transmitted symbols to be equiprobable, i.e. to putting
Pa (Xl ) = 1/M whatever the symbol and the iteration considered.
to the soft estimate x̄i−Δ in x̃i is set to zero in order not to cancel the signal
of interest. Vectors fi = (fi,F , . . . , fi,−F )T and gi = (gi,G , . . . , gi,−G )T represent
the coefficients of filters f and g, respectively. Both vectors are a function of
time since they are updated at each new received symbol.
The relations used to update the vectors of the coefficients can be obtained
from a least-mean square (LMS) gradient algorithm:
fi+1 = fi − μ (zi − xi−Δ ) yi∗

(11.56)
gi+1 = gi − μ (zi − xi−Δ ) x̃∗i
where μ is a small, positive, step-size that controls the convergence properties

of the algorithm.
During the first iteration of the turbo equalizer, x̃i is a vector all the compo-
nents of which are null; the result is that the coefficients vector gi is also null.
The MMSE equalizer then converges adaptively towards an MMSE transversal
equalizer. When the estimated data are very reliable and close to the transmit-
ted data, the MMSE equalizer converges towards an ideal (genie) interference
canceller, then having the performance of a transmission without intersymbol
interference. The limiting forms of the adaptive equalizer are therefore totally
identical to those obtained in (11.43) and (11.44), on condition of course that the
adaptive algorithm can converge towards a local minimum close to the optimal
solution.
Note, however, that for intermediate iterations where the estimated infor-
mation symbols x̄i are neither null nor perfect, filter gi must not be fed directly
with the transmitted symbols otherwise the equalizer will converge towards the
solution of the genie interference canceller, which is not the aim searched for. To
enable the equalizer to converge towards the targeted solution, the idea here in-
volves providing filter gi with soft estimates built from the transmitted symbols,
during the training periods:

(x̄i )its = σx̄ xi + 1 − σx̄2 ηi (11.57)
where σx̄2 corresponds to the variance of the soft estimates x̄i obtained from
(11.24) and ηi a zero-mean complex circularly-symmetric additive white Gaus-
sian noise with unit variance.
In the tracking period and in order to enable the equalizer to follow the
variations of the channel, it is possible to replace the transmitted symbols xi in
relations (11.56) by the decisions x̂i at the output of the equalizer, or by the
decisions on the estimated symbols x̄i .
When the SISO MMSE equalizer is realized in adaptive form, we do not
explicitly have access to the channel impulse response, and the updating relation
of gi does not enable gΔ to be obtained since component gi,Δ is constrained to
be zero. To perform the SISO demapping operation, we must however estimate
both bias gΔ on the data zn provided by the equalizer and variance σν2 of the
residual interference at the output of the equalizer. As we will see, these two
parameters can be estimated from the output of the equalizer. From relation
(11.46) again, we can show the general following result:

2
E |zi | = gΔ σx2 (11.58)
Assuming that the variance of the transmitted data is normalized to unity, an

estimate of gΔ is given by:
N −1
1
ĝΔ = |zi |2 (11.59)
N i=0
Once we have estimated gΔ , we immediately deduce the value of σν2 thanks to

relation (11.47):
σν2 = σx2 ĝΔ (1 − ĝΔ ) (11.60)
One particularity of adaptive MMSE turbo equalization concerns the deter-
mination of the estimated symbols. Indeed, in accordance with the remarks
of [11.20] and [11.55], using a posteriori information instead of extrinsic infor-
mation at the output of the channel decoder in (11.27) can yield significant
performance improvement.
We have therefore defined an adaptive MMSE turbo equalizer whose coeffi-
cients are obtained from a low complexity stochastic gradient descent algorithm,
making it possible to track the slow time variations of the transmission channel.
A drawback of this technique lies in the necessity to transmit training sequences,
which lower the spectral efficiency. The size of training sequences can be signif-
icantly reduced by considering self-learning or blind algorithms. In particular,
the equalizer in the first iteration can be advantageously replaced by a self-
learning equalizer called Self Adaptive Decision-Feedback Equalizer (SADFE)
[11.32] that requires a very small transmission overhead. The work of Hélard
et al. [11.26] has shown that such a turbo equalizer can reach performance
virtually identical to that of the adaptive MMSE turbo equalizer with learning
sequence, while operating at a higher spectral efficiency. On the other hand, a
higher number of iterations is then required.
Examples of performance
For comparison purposes, the performance of the MMSE turbo equalizer has
been simulated by considering the same transmission scenario as for the turbo
MAP equalizer.
First, the parameters of the channel are assumed to be perfectly estimated.
The coefficients are calculated once per frame by matrix inversion, by considering
a digital filter with F = 15 coefficients and a designed delay Δ = 9. The
simulation results, obtained after 10 iterations, are presented in Figure 11.16.
Figure 11.16 – Performance of the MMSE turbo equalizer for BPSK transmission on
the Proakis C channel, with a 4-state rate R = 1/2 non-recursive non-systematic
convolutional code and a 16384 bit pseudo-random interleaver.
Convergence of the iterative process occurs here at threshold signal to noise

ratio of 4 dB, and the turbo equalizer suppresses all the ISI beyond a signal to
noise ratio of 6 dB (after 10 iterations). Compared to the results obtained with
the MAP turbo equalizer (Figure 11.13), we can therefore make the following
remarks:
1. The convergence occurs later with MMSE turbo equalization (of the order
of 1 dB here, compared to MAP turbo equalization).
2. The MMSE turbo equalizer requires more iterations than the MAP turbo
equalizer to reach comparable error rates.
However, the MMSE turbo equalizer here shows its capacity to suppress all the
ISI when the signal to noise ratio is high enough, even on a channel that is known
to be difficult to equalize. It is therefore a serious alternative solution to the
MAP turbo equalizer when the latter cannot be used for reasons of complexity.
Second, the hypothesis of perfect knowledge of the channel parameters has
been removed and the turbo equalizer is simulated in the adaptive form, keep-
ing the same transmission parameters. The communication begins with the
transmission of an initial training sequence of 16384 symbols assumed to be per-
fectly known by the receiver. Then, frames composed of 1000 training symbols
followed by 16384 information symbols are periodically sent into the channel.
Figure 11.17 – Performance of the adaptive MMSE turbo equalizer for BPSK trans-
mission on the Proakis C channel, with a 4-state rate R = 1/2 non-recursive non-
systematic convolutional code and a 16384 bit pseudo-random interleaver.
During the processing of the 16384 information symbols, the turbo equalizer
operates in a decision-directed manner. The equalizer filters each have 21 coef-
ficients (F = G = 10). The coefficients are updated using the LMS algorithm.
The step size is set to μ = 0, 0005 during the training period, and then to
μ = 0, 000005 during the tracking period. Simulation results are given in Fig-
ure 11.17, considering 10 iterations at reception. We observe a degradation of
the order of only 1 dB compared to the ideal situation where the channel is as-
sumed to be perfectly known. We note that when the channel is estimated and
used for the direct computation of the coefficients of the MMSE equalizer, losses
in performance will also appear, which reduces the degradation in comparison
to the ideal situation of Figure 11.16. Note also that, to track the performance
of Figure 11.17, we have not taken into account the loss in the signal to noise
ratio caused by the use of training sequences.
In the light of these results, we note that the major difference between adap-
tive MMSE turbo equalization and that which uses direct computation of the
coefficients from the estimate of the channel lies in the way the filter coeffi-
cients are determined, since the structure and the optimization criterion of the
equalizers are identical.
To finish, we point out that, in the same way as for the turbo MAP equalizer,
we can use EXIT charts to predict the theoretical convergence threshold of the
MMSE turbo equalizer, under the hypothesis of ideal interleaving. The reader
will find further information on this subject in [11.8] or [11.50], for example.
Example of implementation and applications

The implementation of an MMSE turbo equalizer on a signal processor was
reported in [11.9]. The target was the TMS320VC5509 processor by Texas
Instruments. This is a 16-bit fixed-point DSP with low power consumption,
which makes it an ideal candidate for mobile receivers. The considered trans-
mission scheme included a 4-state rate 1/2 convolutional encoder and a 1024
bit interleaver followed by a QPSK modulator. The whole turbo equalizer was
implemented in C language on the DSP, with the exception of some processing
optimized in assembly (filtering and FFT) provided by a specialized library. The
equalizer included 32 coefficients. The decoding was performed using the Max-
Log-MAP algorithm. The simulation results showed that, subject to carefully
choosing the representation in fixed decimal points of the data handled (within
the limit of 16 bits maximum granted by the DSP), data quantization did not
cause any loss in performance in comparison with the corresponding floating-
point receiver. The final data rate obtained was of the order of 42 kbits/s after 5
iterations, which shows the feasibility of such receivers using current technology.
The challenge now involves defining appropriate circuit architectures, capable of
operating at several Mbits/s, in order to respond to emerging demands for high
data rate services.
MMSE turbo equalization is a relatively recent technology. Therefore, at
the moment of writing this book, there have been few studies on the poten-
tial applications of this technique at reception. Generally, resorting to MMSE
turbo equalization is an effective solution in the context of high data rate trans-
missions on highly frequency-selective channels. In particular, this system has
shown excellent performance on the ionospheric channel typically used in the
context of HF military communications. Indeed, the long echoes produced by
this channel prevent the use of MAP equalizers. On the other hand, conven-
tional linear equalization schemes do not make it possible to reach a transmission
quality acceptable when high-order modulations (e.g. 8-PSK or 16-QAM) are
considered. MMSE turbo equalization is thus an attractive solution to the prob-
lem of increasing the data rate of military transmissions. In the context of HF
communications, the interest of MMSE turbo equalization for high spectral effi-
ciency modulations has been validated by the work of Langlais [11.29] and Otnes
[11.41], which shows that this technique can offer gains up to 5 dB compared
to conventional receivers. To our knowledge, MMSE turbo equalization has not
yet been implemented in standardized modems. However, it is important to
note that this reception technique enables the transmission performance to be
notably improved while keeping standardized transmitters.
11.2 Multi-user turbo detection and its applica-

tion to CDMA systems
11.2.1 Introduction and some notations
In a Code Division Multiple Access (CDMA) system, such as the one shown
in Figure 11.18, user k (1 k K) transmits a sequence of binary elements
{dk } = ±1 with an amplitude Ak . For each of the users, a channel encoder (CCk )
is used, followed by an external interleaver (πk ) before the spreading operation
(multiplication by size N , normalized spreading code sk ,) which provides binary
symbols called chips. This code can vary at each symbol time.
Figure 11.18 – CDMA transmitter.
The received signal, r, can be written in matrix form by:
r = SAb + n (11.61)
where:
• S is the N × K matrix formed by the normalized codes of each user (the

k-th column represents the k-th code sk whose norm is equal to unity),
• A is a diagonal K × K matrix made up of the amplitudes Ak of each user.

• b is the vector of dimension K made up of the elements transmitted after
coding channel by the K users.
• n is the N -dimensional centred Gaussian vector with covariance matrix

σ2 IN .
The source data rates of the different users can be different. The size of the
spreading code is such that the chip data rate (after spreading) is the same for all
users. The received signal r is given by the contribution of all the K users plus
a centred AWGN with variance σ 2 . From observation r, we wish to recover the
information bits dk of each user. Figure 11.19 gives the diagram of the receiver
using a turbo CDMA type technique to jointly process the multi-detection and
the channel decoding:
Figure 11.19 – Turbo CDMA receiver.
11.2.2 Multi-user detection

This section presents the main multi-user detection methods. In order to simplify
the description of these methods, only the case of synchronous transmissions over
Gaussian channels is considered.
Standard receiver
The simplest (conventional or standard) detector is the one which operates as
if each user was alone on the channel. The receiver is quite simply made up of
the filter adapted to the signature of the user concerned (this operation is also
called despreading), see Figure 11.20.
At the output of the adapted filter bank the signal can be written in the
form:
y = ST r = RAb + ST n (11.62)
We note that the vector of the additive noise at the output of the adapted
filter bank, is made up of correlated components. Its covariance matrix depends
directly on the intercorrelation matrix of the spreading sequences used, R =
ST S. We have ST n ∼ N (0, σ 2 R).
Figure 11.20 – Standard detector.
We can show that the error probability (before channel decoding) for the k-th
user can be written in the form:
⎛ ⎞

1 Ak Aj
Pe,k = P b̂k = bk = K−1 Q⎝ + bj ρjk ⎠ (11.63)
2 σ σ
b−k ∈{−1,+1}
K−1 j=k
where ρjk = sTj sk measures the intercorrelation between the codes of users j
and k, with b−k = (b1 , b2 , · · · , bk−1 , bk+1 , · · · , bK ).
Assuming that the spreading codes used are such that the intercorrelation
coefficients are constant and equal to 0.2, Figure 11.21 gives the performance of
the standard receiver, in terms of error probability of the first user as a function
of the signal to noise ratio, for a number of users varying from 1 to 6. The
messages of all the users are assumed to be received with the same power. We
note of course that the higher the number of users, the worse the performance.
This error probability can even tend towards 1/2, while the signal to noise ratio
increases if the following condition (Near Far Effect ) is not satisfied:

Ak > Aj |ρjk |
j=k
Optimal joint detection

Optimal joint detection involves maximizing the a posteriori probability (prob-
ability of vector b conditionally to observation y). If we assume that the binary
elements transmitted are equiprobable and given that y = ST r = RAb + ST n
with ST n ∼ N (0, σ 2 R), we can deduce that the optimal joint detection is given
Figure 11.21 – Error probability of the first user as a function of the signal to noise
ratio Eb/N 0 for constant intercorrelation coefficients ρ = 0.2, for K=6 users sharing
the resource.
by the equivalences:
- .
- .
2
Maxb fyb (y) ⇔ Minb y−RAbR−1 ⇔ Maxb 2bTAy−bTARAb (11.64)
An exhaustive search for the optimal solution is relatively complex.
Decorrelator detector
Decorrelator detectors involves multiplying observation y by the inverse of the
intercorrelation matrix of the codes: R−1 y = Ab + R−1 ST n. This equation
shows that the decorrelator enables the multiple access interference to be can-
celled completely, which makes it robust in relation to the Near Far Effect. On
the other hand, the resulting additive Gaussian noise has greater variance. In-
deed, we have: R−1 ST n ∼ N (0, σ 2 R−1 ). The error probability of the k-th user
can then be written in the form:
" #

Ak
Pe,k = P b̂k = bk = Q + (11.65)
σ (R−1 )kk
Linear MMSE detector

The MMSE detector involves finding the transformation

M that minimizes the
2
mean squared error: MinM∈RK×K E b − My . This transformation is no
other than: - .−1

M = A−1 R + σ2 A−2 (11.66)
and consequently, the error probability of the k-th user can be written in the
form:
⎛ ⎛ ⎞⎞
1 Ak (MRk,k ) ⎝ (MRk,j )Aj bj ⎠⎠
Pe,k = K−1 Q⎝ + 1+
2 σ (MRM)k,k (MRk,k )Ak
b−k ∈{−1,+1}
K−1 j=k
(11.67)
Figure 11.22 – Comparison between the decorrelator method and MMSE.
In order to compare the two techniques, the decorrelator detector and

MMSE, the two receivers were simulated with 2 users whose spreading codes
are highly correlated (intercorrelation coefficients equal to 0,75). Figure 11.22
shows the curves of the BER for the first user and parametred by the power of
the 2nd user (or the user’s amplitude). The performance of the MMSE receiver
is always better than that of the decorrelator. For low power of the 2nd user,
performance is close to that of the single-user. However, for a high power of the
2nd user, the MMSE performance will tend towards that of the decorrelator.
Iterative detector
Decorrelator receivers or MMSE receivers can be implemented with iterative ma-
trix inversion (Jacobi, Gauss-Siedel, or relaxation) methods. The Jacobi method
leads to the Parallel Interference Cancellation(PIC) method. The Gauss-Siedel

method leads to the Successive Interference Cancellation (SIC) method. Fig-
ure 11.23 gives the diagram for implementing the SIC (with K = 4 users and
M = 3 iterations) where ICUm,k is the interference cancellation unit (or ICU)
for the k-th user at iteration m (see Figure 11.24). The binary elements are
initialized to zero: at iteration m = 0 b0,k = 0 for k = 1, · · · , K.
Figure 11.23 – Iterative SIC (Successive Interference Cancellation) detector, with K =

4 users and M = 3 iterations.
Figure 11.24 – Interference cancellation unit ICUm,k for user k, at iteration m.
Function F Ak (respectively F A−1 k ) is the despreading (respectively spread-

ing) of the k-th user. Function φ(.) can be chosen as a non-linear function or
quite simply as being equal to the identity function (see also the choice of φ(.)
in the case of turbo CDMA). If we choose the identity function, the ICU unit of
Figure 11.24 can of course be simplified and is easy to define. In this case, we
can verify that, for user k and at iteration m, the output of the receiver can be
written as the result of linear filtering:
5
k−1
-
. m−1
bm,k = sTk I − sj sTj ΦpK r = gm,k
T
r (11.68)
j=1 p=0
with:
5
K
- .
ΦK = I − sj sTj (11.69)
j=1
We can show that the error probability for the k-th user at iteration m can be
written in the following form, where S is the matrix of the codes and A is the
diagonal matrix of the amplitudes:
⎛ ⎞
1 g T
SAb
Q⎝ ⎠
m,k
Pe (m, k) = K−1 (11.70)
2 σ gT g
b/bk =+1 m,k m,k
Figure 11.25 gives an example of simulations of the SIC method with 5 users, a
spreading factor of 20 and an intercorrelation matrix given by:
⎛ ⎞
1 0, 3 0 0 0
⎜ 0, 3 1
⎜ 0, 3 0, 3 0, 1 ⎟ ⎟
R=⎜ ⎜ 0 0, 3 1 0 −0, 2 ⎟ ⎟
⎝ 0 0, 3 0 1 0 ⎠
0 0, 1 −0, 2 0 1
Figure 11.25 – Simulation of a SIC receiver.
We note that after 3 iterations, the SIC converges towards the result obtained
with the decorrelator (we can prove mathematically that the SIC converges
towards this result when the number of iterations M tends towards infinity).
11.2.3 Turbo CDMA

Several turbo CDMA type techniques have been proposed to jointly process
multi-detection and channel decoding:
• Varanasi and Guess [11.53] have proposed (hard estimation) decoding and
immediately recoding each user before subtracting this contribution from
the received signal. The same operation is performed on the residual signal
to decode the information of the second user, and so on, until the final user.
• Reed and Alexander [11.46] have proposed to use an adapted filter bank
followed (in parallel) by different decoders before subtracting, for each
user, the multiple access interference linked to the K − 1 other users.
• Wang and Poor [11.56] have proposed a multi-user detector that involves
implementing in parallel the MMSE filters associated with each user, fol-
lowed by the corresponding channel decoders. These two elements ex-
change their extrinsic information iteratively.
• Tarable et al. [11.48] have proposed a simplification of the method pre-
sented in [11.56]. For the first iterations, an MMSE type multi-user detec-
tor is used, followed by channel decoders placed in parallel. For the final
iterations, the MMSE filter is replaced by an adapted filter bank.
Turbo SIC detector

In this section, channel decoding is introduced into a new successive interference
cancellation (SIC) structure. Figure 11.23 remains valid, only units ICUm,k
change. Each interference cancellation unit ICUm,k , relative to the k-th user
and at iteration m, is given in Figure 11.26. The originality lies in the way in
which this unit is designed: the residual error signal em,k is despread (by sk ) then
deinterleaved (πk−1 ) before adding the weighted estimation of the bm−1,k data
of the same user calculated at the previous iteration. The signal thus obtained,
ym,k , passes through the channel decoder that provides the a posteriori log
likelihood ratio, conditionally to the whole observation, of all the binary elements
(both for the information bits and the parity bits):

P [bk = +1/ym,k ]
LLR(bk /ym,k ) = log (11.71)
P [bk = −1/ym,k ]
This ratio is then transformed into a weighted estimation of the binary ele-
ments:
1
b̃m,k = E [bk /ym,k ] = tanh LLR (bk /ym,k ) (11.72)
2
The soft estimation of user k at iteration m is given by bm,k = Ak b̃m,k . The
difference (bm,k − bm−1,k ) is interleaved by πk before spreading by sk . The result
Figure 11.26 – Interference cancellation unit for the turbo SIC decoder in CDMA for
the k-th user and at iteration m.
thus obtained Δem,k is subtracted from residual signal em,k to obtain the new
residual signal em,k+1 of the following user (if k < K) or to obtain the new
residual signal em+1,1 for the first user at the following iteration (em,K+1 =
em+1,1 ). Here, ym,k is written in the form ym,k = Ak bk + νm,k where νm,k
(residual multiple access interference plus the additive noise) is approximated
by a centred Gaussian random variable whose variance is given by:

var (νm,k ) = A2i ρ2i,k 1 − b̃2m,i + A2i ρ2i,k 1 − b̃2m−1,i + σ2 (11.73)

i<k i>k
We show that the extrinsic information of user k at iteration m is given by:

P [ym,k /bk = +1] 2ym,k Ak
λm (bk ) = log = (11.74)
P [ym,k /bk = −1] var(νm,k )
This extrinsic information serves as the input at the decoder associated with the
k-th user.
Some simulations
To give an idea of the performance of the turbo SIC decoder, Gold sequences
of size 31 are generated. The channel turbo encoder (rate R = 1/3) normalized
for UMTS [11.1] is used. We consider frames of 640 bits per user. The external
interleavers of the different users are produced randomly. The BER and PERs
are averaged over all the users. For the channel turbo encoder, the Max-Log-
MAP algorithm is used, 8 being the number of iterations internal to the turbo
decoder. Figure 11.27(a) gives the performance of the turbo SIC decoder for one,
two and three iterations with K = 31 users (that is, 100% load rate) having the
same power. The performance of the single-user detector and of the conventional
detector are also indicated. Figure 11.27(b) shows performance in terms of PER.
Turbo SIC/RAKE detector

In the case where the propagation channel of the k-th user has an impulse
response with multiple paths ck (t), it suffices to replace the despreading function
Figure 11.27 – Performance of the turbo SIC decoder: (a) mean Binary Error Rates
(BER) (b) mean Packet Error Rates (PER). K = 31 users, spreading factor of 31, with
frame size 640 bits.
by a RAKE filter (filter adapted to the spreading sequence convolved with the
transfer function ck (t) in the ICUm,k ) unit, and to replace the spreading function
by the spreading function convolved by ck (t). This new structure is called a
turbo SIC/RAKE decoder.
The turbo SIC /RAKE decoder is used particularly in the context of the
uplink in the UMTS-FDD system.
11.3 Conclusions
In this chapter, we have presented the first two systems to have benefited from
applying the turbo principle to a context other than error correction coding. In
the first part, we have described the principle of turbo equalization, which relies
on an iterative exchange of probabilistic information between a SISO equalizer
and a SISO decoder. The SISO equalizer can take different forms according
to the chosen optimization criterion. We have presented two types of SISO
equalizers: the BCJR-MAP equalizer, operating on the trellis representation of
the ISI channel, and the MMSE equalizer, which uses linear filtering. The MAP
turbo equalizer leads to excellent performance compared to the conventional
receiver. However, this approach is often avoided in practice since it leads to a
very high computation cost. We have discussed several solutions for reducing
the complexity of the BCJR-MAP equalizer. As for the MMSE turbo equalizer,
it offers a good compromise between performance and complexity. For many
transmission configurations it leads to performance close to that offered by the
BCJR-MAP turbo equalizer, with reasonable complexity. In addition, unlike
the BCJR-MAP turbo equalizer, the MMSE turbo equalizer can be realized
in adaptive form, thereby jointly performing equalization and tracking of the
channel time variations.
In the second part, we have dealt with the application of the turbo princi-
ple to the domain of multi-user communications in code-division multiple access
systems. We have presented a survey of conventional multi-user detection tech-
niques. In particular, the PIC and SIC methods for cancellation of multi-user
interference have been described. Their particular structures lead to a relatively
simple exploitation of the turbo principle in a multi-user transmission context.
Like for turbo equalization, different detectors can be implemented based on
MMSE filters or matched-filter banks, for example.
In this chapter, we have deliberately limited ourselves to the presentation of
two particular systems exploiting the turbo principle. However, more generally,
any problem of detection or parameter estimation may benefit from the turbo
principle. Thus, the range of solutions dealing with interference caused by a
multi-antenna system at transmission and at reception (MIMO) has been en-
riched by iterative techniques such as the turbo BLAST (Bell Labs layered space
time) [11.25]. The challenge involves proposing SISO detectors of reasonable
complexity, without sacrificing data rates and/or the high performance of such
systems.
We can also mention the efforts dedicated to receiver synchronization. In-
deed, the gains in power provided by the turbo principle lead to moving the
systems’ operation point towards low signal to noise ratios. Now, conventional
synchronization devices were not initially intended to operate in such difficult
conditions [11.31]. One possible solution is to integrate the synchronization into
the turbo process. A state of the art of turbo methods for timing synchronization
was presented in [11.4]. More generally, when the choice of turbo processing at
reception is performed, it seems interesting, or even necessary, to add a system
to the receiver to iteratively estimate the transmission parameters, like channel
turbo estimation or turbo synchronization.
Among other applications, the uplink of future radio-mobile communications
systems will require higher and higher data rates, with an ever-increasing number
of users. This is the one of the favourite applications of the turbo principle, the
generalization of which will be essential in order to respond to the never-ending
technological challenge posed by the evolution of telecommunications.
Understanding the turbo principle has led to the introduction of novel theo-
retical tools and concepts, like EXIT charts or factor graphs. While the former
enable accurate prediction of the convergence threshold of iterative decoding
schemes, the latter offer a graphical framework for representing complex detec-
tion/estimation problems and then deriving efficient turbo-like iterative algo-

rithms for solving them. The interested reader will find good overviews of factor
graphs and their applications in [11.37] and [11.38]. The use of factor graphs in a
turbo equalization context has been considered in particular in [11.22] et[11.16]
and an in-depth study of multi-user detection from a factor graph perspective
has been presented in [11.10].
Bibliography
[11.1] Etsi digital cellular telecommunication system (phase 2+). GSM 05
Series, Rel. 1999.
[11.2] B. S. Ünal A. O. Berthet and R. Visoz. Iterative decoding of convolu-
tionally encoded signals over multipath rayleigh fading channels. IEEE
Journal of Selected Areas in Communications, 19(9):1729–1743, Sept.
2001.
[11.4] J. R. Barry, A. Kavcic, S. W. McLaughlin, A. Nayak, and W. Zeng.

Iterative timing recovery. IEEE Signal Processing Magazine, 21(1):89–
102, Jan. 2004.
[11.5] G. Bauch and V. Franz. A comparison of soft-in/soft-out algorithms for
turbo detection. Proceedings of International Conference on Telecom-
munications (ICT’98), pages 259–263, June 1998.
[11.6] G. Bauch and V. Franz. Iterative equalization and decoding for the
gsm system. In Proceedings of IEEE Vehicular Technology Conference
(VTC’98), pages 2262–2266, Ottawa, Canada, May 1998.
[11.7] G. Bauch, H. Khorram, and J. Hagenauer. Iterative equalization and
decoding in mobile communication systems. In Proceedings of 2nd Eu-
ropean Personal Mobile Commununications Conference (EPMCC’97),
pages 307–312, Bonn, Germany, Sept.-Oct. 1997.
[11.8] R. Le Bidan. Turbo Equalization for Bandwidth-Efficient Digital Com-
munication over Frequency-Selective Channels. PhD thesis, INSA de
Rennes, Nov. 2003.
[11.9] R. Le Bidan, C. Laot, and D. Leroux. Real-time mmse turbo equalization

on the tms320c5509 fixed-point dsp. In Proceedings of IEEE Interna-
tional Conference on Accoustics, Speech and Signal Processing ICCASP
2004, volume 5, pages 325–328, Montreal, Canada, May 2004.
[11.10] J. Boutros and G. Caire. Iterative multiuser joint decoding: Unified

framework and asymptotic analysis. IEEE Transactions on Information
Theory, 48(7):1772–1793, July 2002.
[11.11] J.-M. Brossier. Signal et Communication Numérique – Equalisation et

Synchronisation. Collection Traitement du Signal. Hermès, Paris, 1997.
[11.12] C. Berrou C. Douillard, M. Jézequel, A. Picart, P. Didier, and

A. Glavieux. Iterative correction of intersymbol interference: Turbo
equalization. European Transactions on Telecommununication, 6(5):507–
511, Sept.-Oct. 1995.
[11.13] P. R. Chevillat and E. Eleftheriou. Decoding of trellis-encoded signals in

the presence of intersymbol interference and noise. IEEE Transactions
on Commununications, 37(7):669–676, July 1989.
[11.14] A. Dejonghe and L. Vandendorpe. Turbo equalization for multilevel

modulation: An efficient low-complexity scheme. In Proceedings of IEEE
International Conference on Communications (ICC’2002), volume 3,
pages 1863–1867, New-York City, NY, 28 Apr.-2 May 2002.
[11.15] P. Didier. La Turbo-Egalisation et son Application aux Communications
Radiomobiles. PhD thesis, Université de Bretagne Occidentale, Déc.
1996.
[11.16] R. J. Drost and A. C. Singer. Factor-graph algorithms for equalization.

IEEE Transactions on Signal Processing, 55(5):2052–2065, May 2007.
[11.17] C. Fragouli, N. Al-Dhahir, S. N. Diggavi, and W. Turin. Prefiltered
space-time m-bcjr equalizer for frequency-selective channels. IEEE
Transactions on Commununications, 50(5):742–753, May 2002.
[11.18] V. Franz. Turbo Detection for GSM Systems – Channel Estimation,

Equalization and Decoding. PhD thesis, Lehrstuhl für Nachrichten Tech-
nik, Nov. 2000.
[11.19] G. Ferrari G. Colavolpe and R. Raheli. Reduced-state bcjr-type algo-
rithms. IEEE Journal of Selected Areas in Communications, 19(5):849–
859, May 2001.
[11.20] A. Glavieux, C. Laot, and J. Labat. Turbo equalization over a frequency

selective channel. In Proceedings of International Symposium on Turbo
Codes & Related Topics, pages 96–102, Brets, France, Sept. 1997.
[11.21] G. H. Golub and C. F. Van Loan. Matrix Computations. The Johns

Hopkins University Press, Baltimore, 3rd edition, 1996.
[11.22] Q. Guo, L. Ping, and H.-A. Loeliger. Turbo equalization based on factor
graphs. In Proceedings of IEEE International Symposium on Information
Theory (ISIT’05), pages 2021–2025, Sept. 2005.
[11.23] J. Hagenauer. Soft-in / soft-out – the benefits of using soft decisions in all
stages of digital receivers. In Proceedings of 3rd International Symposium
on DSP Techniques applied to Space Communications, Noordwijk, The
Netherlands, Sept. 1992.
[11.24] J. Hagenauer, E. Offer, C. Measson, and M. Mörz. Decoding and
equalization with analog non-linear networks. European Transactions
on Telecommunications, pages 107–128, Oct. 1999.
[11.25] S. Haykin, M. Sellathurai, Y. de Jong, and T. Willink. Turbo mimo
for wireless communications. IEEE Communications Magazine, pages
48–53, Oct. 2004.
[11.26] M. Hélard, P.J. Bouvet, C. Langlais, Y.M. Morgan, and I. Siaud. On the
performance of a turbo equalizer including blind equalizer over time and
frequency selective channel. comparison with an ofdm system. In Pro-
ceedings of International Symposium on Turbo Codes & Related Topics,
pages 419–422, Brest, France, Sept. 2003.
[11.27] G. D. Forney Jr. Maximum-likelihood sequence estimation of digital
sequences in the presence of intersymbol interference. IEEE Transactions
on Information Theory, IT-18(3):363–378, May 1972.
[11.28] W. Koch and A. Baier. Optimum and sub-optimum detection of coded
data disturbed by time-varying intersymbol interference. In Proceedings
of IEEE Global Telecommununication Conference (GLOBECOM’90),
volume 3, pages 1679–1684, San Diego, CA, 2-5 Dec. 1990.
[11.29] C. Langlais. Etude et amélioration d’une technique de réception

numérique itérative: Turbo Egalisation. PhD thesis, INSA de Rennes,
Nov. 2002.
[11.30] C. Langlais and M. Hélard. Mapping optimization for turbo equalization
improved by iterative demapping. Electronics Letters, 38(2), Oct. 2002.
[11.31] C. Langlais, M. Hélard, and M. Lanoiselée. Synchronization in the carrier
recovery of a satellite link using turbo codes with the help of tentative
decisions. In Proceedings of IEE Colloqium on Turbo Codes in Digital
Broadcasting, Nov. 1999.
[11.32] C. Laot. Egalisation Auto-didacte et Turbo-Egalisation – Application

aux Canaux Sélectifs en Fréquence. PhD thesis, Université de Rennes I,
Rennes, France, July 1997.
[11.33] C. Laot, R. Le Bidan, and D. Leroux. Low-complexity mmse turbo equal-

ization: A possible solution for edge. IEEE Transactions on Wireless
Communications, 4(3), May 2005.
[11.34] C. Laot, A. Glavieux, and J. Labat. Turbo equalization: Adaptive equal-

ization and channel decoding jointly optimized. IEEE Journal of Selected
Areas in communications, 19(9):1744–1752, Sept. 2001.
[11.35] I. Lee. The effects of a precoder on serially concatenated coding systems
with an isi channel. IEEE Transactions on Communications, 49(7):1168–
1175, July 2001.
[11.36] S.-J. Lee, N. R. Shanbhag, and A. C. Singer. Area-efficient high-

throughput vlsi architecture for map-based turbo equalizer. In Proceed-
ings of IEEE Workshop on Signal Processing Systems SIPS 2003, pages
87–92, Seoul, Korea, Aug. 2003.
[11.37] H.-A. Loeliger. An introduction to factor graphs. IEEE Signal Processing
Magazine, 21(1):28–41, Jan. 2004.
[11.38] H.-A. Loeliger, J. Dauwels, J. Hu, S. Korl, L. Ping, and F. R. Kschis-
chang. The factor graph approach to model-based signal processing.
Proceedings of IEEE, 95(6):1295–1322, June 2007.
[11.39] K. Narayanan. Effects of precoding on the convergence of turbo equal-
ization for partial response channels. IEEE Journal of Selected Areas in
Communications, 19(4):686–698, Apr. 2001.
[11.40] N. Nefedov, M. Pukkila, R. Visoz, and A. O. Berthet. Iterative re-
ceiver concept for tdma packet data systems. European Transactions on
Telecommunications, 14(5):457–469, Sept.-Oct. 2003.
[11.41] R. Otnes. Improved receivers for digital High Frequency communica-

tions: iterative channel estimation, equalization, and decoding (adaptive
turbo equalization). PhD thesis, Norwegian University of Science and
Technology, 2002.
[11.42] B. Penther, D. Castelain, and H. Kubo. A modified turbo detector for
long delay spread channels. In Proceedings of International Symposium
on Turbo Codes & Related Topics, pages 295–298, Brest, France, Sept.
2000.
[11.43] A. Picart, P. Didier, and A. Glavieux. Turbo detection: A new approach
to combat channel frequency selectivity. In Proceedings of IEEE In-
Montreal, Canada, June 1997.
[11.44] J. G. Proakis. Digital Communications. McGraw-Hill, New-York, 4th

edition, 2000.
[11.45] S. U. H. Qureshi. Adaptive equalization. Proceedings of the IEEE,
73(9):1349–1387, Sept. 1985.
[11.46] M.C. Reed and P.D. Alexander. Iterative multiuser detection using an-
tenna arrays and fec on multipath channels. IEEE Journal of Selected
Areas in communications, 17(12):2082–2089, Dec. 1999.
[11.47] D. Reynolds and X. Wang. Low-complexity turbo equalization for di-
versity channels. Signal Processing, 81(5):989–995, May 2001.
[11.48] A. Tarable, G. Montorsi, and S. Benedetto. A linear front end for
iterative soft interference cancellation and decoding in coded cdma.
In Proceedings of IEEE International Conference on Communications
(ICC’01), 11-14 June 2001.
[11.49] S. ten Brink. Designing iterative decoding schemes with the extrinsic
information transfer chart. AEÜ International Journal of Electronics
Communications, 54(6):389–398, Nov. 2000.
[11.50] M. Tüchler, R. Kötter, and A. C. Singer. Turbo equalization: Principles
and new results. IEEE Transactions on Communications, 50(5):754–767,
May 2002.
[11.51] M. Tüchler, A. C. Singer, and R. Kötte. Minimum mean-squared error
equalization using a priori information. IEEE Transactions on Signal
Processing, 50(3):673–683, March 2002.
[11.52] G. Ungerboeck. Channel coding with mutilevel/phase signals. IEEE
Trans. Info. Theory., IT-28(1):55–67, Jan. 1982.
[11.53] M.K. Varanasi and T. Guess. Optimum decision feedback multiuser
equalization with successive decoding achieves the total capacity of the
gaussian multiple-access channel. In Conference Record of the Thirty-
First Asilomar Conference on Signals, Systems & Computers, volume 2,
pages 1405–1409, 2-5 Nov. 1997.
[11.54] G. M. Vitetta, B. D. Hart, A. Mämmelä, and D. P. Taylor. Equalization
techniques for single-carrier unspread digital modulations. In A. F. Mol-
ish, editor, Wideband Wireless Digital Communications, pages 155–308.
Prentice-Hall, Upper Saddle River, NJ, 2001.
[11.55] F. Volgelbruch and S. Haar. Improved soft isi cancellation for turbo
equalization using full soft output channel decoder’s information. In
Proceedings of IEEE Global Telecommununication Conference (Globe-
com’2003), pages 1736–1740, San Francisco, USA, Dec. 2003.
[11.56] X. Wang and H. V. Poor. Iterative (turbo) soft interference cancellation

and decoding for coded cdma. IEEE Transactions on Communications,
47(7):1046–1061, July 1999.
bit-interleaved coded modulation, 348,
367
branch metric, 57
butterfly, 176
capacity, 331
Index CDMA, see code division multiple ac-

cess
Chase algorithm, 139
Chien algorithm, 150
circular recursive systematic codes,
16-QAM, 352 194, 207, 216
2-PSK, 137 code division multiple access, 396
4-PSK, 137 code memory, 189, 222
codeword, 109, 319
a posteriori probability, 168, 240, 330 coding rate, 109, 303
accumulated spatial distance, 230 coherence bandwidth, 70
additive white gaussian noise, 9, 37, coherence time, 70
191, 240, 309, 318, 365 constellation, 20, 343
all zero sequence, 5, 178, 204, 257, 312 constraint length, 170, 204
almost regular permutation , 229 convergence, 249, 311
APP, see a posteriori probability convolutional code, 167, 307
ARP, see almost regular permutation CRC, see cyclic redundancy code
ARQ, see automatic repeat request CRSC, see circular recursive system-
asymptotic gain, 14, 142, 256 atic codes
automatic repeat request, 86, 255, 380 cyclic code, 122
AWGN, see additive white gaussian cyclic redundancy code, 128
noise
decorrelator, 399
Bahl-Cocke-Jelinek-Raviv (BCJR) al- density evolution, 312
gorithm, 168, 369 digital signal processor, 10, 246, 291,
band-limited channel, 59, 93 382
BCH code, 125 disorder, 225
BCJR, see Bahl-Cocke-Jelinek-Raviv distance spectrum, 183
algorithm dithered relatively prime, 229
belief propagation, 291, 297 diversity, 2, 74, 222, 348, 367
Berlekamp-Massey algorithm, 144, 293 double-binary, 170, 227
Bernoulli distribution, 85 DRP, see dithered relatively prime
BICM, see bit-interleaved coded mod- DSP, see digital signal processor
ulation dual code, 112
binary symmetric channel, 7, 84, 132,
188 EDGE (Enhanced Data Rate for GSM
bipartite graph, 297 Evolution), 378
bit error rate(BER), 12 entropy, 86
equalization, 81 log likelihood ratio, 91, 299, 309, 353,

error detection, 132 403
Error impulse method, 257 Low Density Parity Check codes, see
Euclid algorithm, 144, 293 LDPC codes
Euclidean distance, 138 LRC (L-ary raised cosine), 34
EXIT chart, 260, 313
extended code, 114 M-ASK (amplitude-shift keying with
extrinsic information, 235, 308 M states), 22
eye diagram, 61 M-FSK (frequency modulation with M
states), 29
fading channel, 69, 346 M-QAM, see Quadrature Amplitude
Fang-Battail algorithm, 280 Modulation
free distance, 178 MAP (maximum a posteriori), 168,
frequency band, 2, 343, 362 213, 344
MAP equalizer, 372
Galois field, 109 Max-Log approximation, 235, 290, 344,
generator matrix, 110, 304, 306 371
generator polynomial, 121 maximum likelihood, 134
Gilbert-Varshamov bound, 91 m-binary code, 217
GMSK (Gaussian Minimum Shift Key- MHD, see minimum Hamming dis-
ing), 35 tance
Golay code, 130 minimum Hamming distance, 6, 100,
Gray coding, 23 113, 178, 201, 256, 271, 314,
GSM, 364 349
minimum mean square error (MMSE),
Hadamard code, 118
360
Hamming code, 117
MLSD equalizer, 364
Hamming weight, 113
MMSE equalizer, 365
hard input, 134, 143, 309
MSK (Minimum Shift Keying), 32
hard output, 11, 186
multi-user detection, 397
Hartmann-Nazarov algorithm, 285
mutual information, 93, 259, 313
ICU, see interference cancellation unit
Nyquist criterion, 64
interconnection networks, 322
interference cancellation unit, 401 OFDM (orthogonal frequency division
interleaver, 102, 231, 308, 346, 368 multiplexing), 79, 362
intersymbol interference, 61, 361 OOK (on off keying), 22
irregularity profile, 302
ISI, see intersymbol interference packet error rate (PER), 12
parallel concatenation, 203, 214
LDPC codes, 297, 302
parallel interference cancellation, 401
linear code, 115, 301
parity check code, 115, 298
linear equalizer, 365
parity check matrix, 112, 301
LLR, see log likelihood ratio
parity equation, 298
LMS, see Least Mean Square
Index 415
permutation, 5, 201, 216, 222, 231, 283, soft-in/soft-out, 11, 202, 238, 291, 297,
304, 347, 368 370
Peterson algorithm, 144 SOVA, see Soft-Output Viterbi Algo-
Phase Shift Keying, 25, 31, 239, 350, rithm
364 sphere-packing bound, 98
PIC, see parallel interference cancella- spread spectrum, 362
tion state machine, 173
PNP (parity node processor), 319, 331 successive interference cancellation,
polynomial representation, 120 401, 403
power spectral density, 9, 24, 34, 35, survivor path, 189
37, 59, 79, 103, 137 syndrome, 132, 144
product code, 115 systematic bits, 348
psd, see power spectral density
PSK, see Phase Shift Keying tail bits, 221
puncturing, 169, 196, 206, 220, 347 tail-biting, 169, 194
transfer function, 180, 260
Quadrature Amplitude Modulation, transmission channel, 6, 83, 361
26, 95, 352 tree of a code, 173
quantization, 8, 237, 333, 395 trellis, 17, 56, 173, 207, 240, 290, 319,
343, 363
Rayleigh channel, 73, 95, 237 TTCM, see turbo trellis coded modu-
Reddy-Robinson algorithm, 274 lation
redundancy, 109, 183, 216, 303, 345 turbo CDMA, 359
Reed-Muller code, 119 turbo code, 11, 83, 203, 213, 307
Reed-Solomon code, 130, 143 turbo detection, 371
repetition code, 116 turbo equalization, 359
Return To Zero, 178, 204, 216 turbo estimation, 406
RTZ, see Return To Zero turbo product code, 290
turbo synchronization, 406
scalar product, 9 turbo trellis coded modulation, 343
sequential decoding, 168
serial concatenation, 115, 202 Viterbi algorithm, 56
Shannon limit, 2, 93, 214, 345, 359 VNP (variable node processor), 319,
shortened code, 114 330
SIC, see successive interference cancel-
lation
SIC detector, 403
SISO, see soft-in/soft-out
SISO demapping, 387
SISO equalizer, 370
soft input, 134, 309
soft output, 11, 92
Soft-Output Viterbi Algorithm, 186,
214

Codes and Turbo Codes - C. Berrou (Springer, 2010) BBS

Uploaded by

Copyright:

Available Formats

Codes and Turbo Codes - C. Berrou (Springer, 2010) BBS

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Codes and Turbo Codes - C. Berrou (Springer, 2010) BBS

Uploaded by

Copyright:

Available Formats

Codes and Turbo Codes

ISBN : 978-2-8178-0038-7 Springer Paris Berlin Heidelberg New York

© Springer-Verlag France, Paris, 2010

First edition in French © Springer-Verlag France 2007

Cover design: Jean-François MONTMARCHÉ

The following have contributed to this work:

Translation: Janet Ormrod (Télécom Bretagne).

Cover illustration: Jean-Noël Jaﬀry (Télécom Bretagne).

Pythagoras, ﬁfth century BC

To our late lamented colleagues and friends,

however, remain incomparably more powerful, due to its extraordinary network

munications systems, especially coding and equalization, in order to counter the

[1] A. J. Viterbi, "Error Bounds for Convolutional Codes and an Asymptoti-

4 Block codes 109

5 Convolutional codes and their decoding 167

6 Concatenated codes 201

7 Convolutional turbo codes 213

7.6.2 Asymptotic behaviour . . . . . . . . . . . . . . . . . . . . 256

8 Turbo product codes 271

9 LDPC codes 297

10 Turbo codes and large spectral eﬃciency transmissions 343

11 The turbo principle applied to equalization and detection 359

More elaborate coding involves transmitting the original message no longer

1.1 Digital messages

1.2 A ﬁrst code

Figure 1.1 – Extended Hamming encoder: a simple, easily decodable code.

The encoder contains four identical operators performing the exclusive-or

It is therefore quite simply 0 if the number of logical 1s appearing in the input

A decoder capable of implementing such a process is called a maximum likelihood

1.3 Hard input decoding and soft input decoding

Figure 1.2 – Binary symmetric channel with error probability p.

Figure 1.3 gives an example of a transmission during which the transmitted

Figure 1.3 – Transmission channel with additive noise of an analogue type.

Minimizing the Euclidean distance between two codewords c and c’ therefore

In a Digital Signal Processor (DSP) or in an Application Speciﬁc Integrated

(−1)(0.2) + (−1)(0.4) + (−1)(0.1) + (−1)(−1)

(+1)(0.2) + (+1)(0.4) + (+1)(0.1) + (−1)(−1)

for the competitor word (1, 1, 1, 0, 0, 0, 0, 1).

1.4 Hard output decoding and soft output decod-

1.5 The performance measure

The comparison of diﬀerent coding and decoding processes or the variation in

rate, the asymptotic performance of Pe is approximated by:

To obtain the error binary probability Pe,bit , it suﬃces to multiply Pe,word

1.6 What is a good code?

Figure 1.6 – Possible behaviours for a coding/decoding scheme on a Gaussian channel

1.7 Families of codes

• algebraic codes are appropriate for protecting blocks of data independent

in the associated decoders and by an exchange of information of a probabilistic

2.1 Digital Modulations

s(t) = a cos(2πf0 t + ϕ) (2.1)

σ(t) = a exp[j(2πf0 t + ϕ)] (2.2)

σ(t) = α exp(2πjf0 t), (2.3)

where only the ﬁrst factor

be either analogue or digital, with memory or memoryless. In all that follows

2.1.2 Linear Memoryless Modulations

Figure 2.1 – On Oﬀ Keying (OOK).

sj (t) = Aj h(t) cos(2πf0 t + ϕ0 ) (2.5)

with Aj = (2j − 1 − M )A j = 1, 2, · · · , M where A is constant (2.6)

and h(t) = 1 for t ∈ [0, T [