Series

SeriesISSN:
Series ISSN:1932-1244
ISSN: 1932-1244
1932-1244

M
M

DINIZ
DINIZ •• MARTINS
DINIZ
SSYNTHESIS
YNTHESISL
YNTHESIS LECTURES ONC
ECTURESON
ECTURES ON COMMUNICATIONS
OMMUNICATIONS
OMMUNICATIONS
&C
& C Mor
Mor gan&
Morgan
gan &Cl
Claypool
Cl aypool Publishers
aypool Publishers
Publishers

• MARTINS •• LIMA
MARTINS • LIMA
Block
BlockTransceivers
Transceivers
Series
SeriesEditor:
Series Editor:William
Editor: WilliamTranter,
William Tranter,Virginia
Tranter, VirginiaTech
Virginia Tech
Tech

LIMA
Block
BlockTransceivers:OFDM
Block Transceivers:OFDMand
Transceivers:OFDM andBeyond
and Beyond
Beyond
Paulo
PauloS.R.
Paulo
Wallace
S.R.Diniz,
S.R.
WallaceA.
Wallace
Diniz,Universidade
Diniz,
A.Martins,
A.
UniversidadeFederal
Universidade
Martins,Universidade
Martins,
Federaldo
Federal
UniversidadeFederal
Universidade
doRio
do
Federaldo
Federal
Riodede
Rio
doRio
do
deJaneiro
Janeiro
Janeiro
Riodede
Rio deJaneiro
Janeiroand
Janeiro andCEFET/RJ
and CEFET/RJ
CEFET/RJ
OFDM
OFDM and
and Beyond
Beyond
Markus
MarkusV.S.
Markus V.S.Lima,
V.S. Lima,Universidade
Lima, UniversidadeFederal
Universidade Federaldo
Federal doRio
do Riodede
Rio deJaneiro
Janeiro
Janeiro
The
Thedemand
The demandfor
demand fordata
for datatraffic
data trafficover
traffic overmobile
over mobilecommunication
mobile communicationnetworks
communication networkshas
networks hassubstantially
has substantiallyincreased
substantially increasedduring
increased duringthe
during thelast
the last
last
decade.
decade.
decade.AsAs a result,
As aa result, these
result,these mobile
these mobile broadband
mobile broadband devices
broadband devices
devices spendspend the
spend the available
the available spectrum
available spectrum fiercely,
spectrum fiercely, requiring
fiercely,requiring
requiring thethe search
the search
search
for
fornew
for newtechnologies.
new technologies.InIn
technologies. Intransmissions
transmissionswhere
transmissions wherethe
where thechannel
the channelpresents
channel presentsaaafrequency-selective
presents frequency-selectivebehavior,
frequency-selective behavior,multicarrier
behavior, multicarrier
multicarrier
modulation
modulation(MCM)
modulation (MCM)schemes
(MCM) schemeshave
schemes haveproven
have proventoto
proven tobebemore
be moreefficient,
more efficient,inin
efficient, interms
termsofof
terms ofspectral
spectralusage,
spectral usage,than
usage, thanconventional
than conventional
conventional
modulations
modulations
modulations and and spread
and spread spectrum
spread spectrum techniques.
spectrum techniques.
techniques.

BLOCK
BLOCK TRANSCEIVERS
BLOCK
The
Theorthogonal
The orthogonalfrequency-division
orthogonal frequency-divisionmultiplexing
frequency-division multiplexing(OFDM)
multiplexing (OFDM)isisisthe
(OFDM) themost
the mostpopular
most popularMCM
popular MCMmethod,
MCM method,since
method, sinceitititnot
since not
not
only
onlyincreases
only increasesspectral
increases spectralefficiency
spectral efficiencybut
efficiency butalso
but alsoyields
also yieldssimple
yields simpletransceivers.
simple transceivers.All
transceivers. AllOFDM-based
All OFDM-basedsystems,
OFDM-based systems,including
systems, includingthe
including the
the

TRANSCEIVERS
single-carrier
single-carrierwith withfrequency-division
frequency-divisionequalization
equalization(SC-FD), (SC-FD),transmit transmitredundancy
redundancyinin inorder
ordertoto tocope
copewith withthe

TRANSCEIVERS
single-carrier with frequency-division equalization (SC-FD), transmit redundancy order cope with the
the
problem
problemofof
problem ofinterference
interferenceamong
interference amongsymbols.
among symbols.This
symbols. Thisbook
This bookpresents
book presentsOFDM-inspired
presents OFDM-inspiredsystems
OFDM-inspired systemsthat
systems thatare
that areable
are ableto,
able to,atat
to, atmost,
most,
most,
halve
halvethe
halve
comparable.
theamount
the
comparable.Such
comparable.
amountofof
amount ofredundancy
Suchsystems,
Such
redundancyused
redundancy
systems,herein
systems,
usedby
used
hereincalled
herein
byOFDM
by OFDMsystems
OFDM
calledmemoryless
called memorylesslinear
memoryless
systemswhile
systems whilekeeping
while
lineartime-invariant
linear
keepingthe
keeping
time-invariant(LTI)
time-invariant
thecomputational
the
(LTI)
computationalcomplexity
computational
(LTI)transceivers
transceiverswith
transceivers
complexity
complexity
withreduced
with reduced
reduced
Paulo
Paulo S.R.
S.R.Diniz
Diniz
redundancy,
redundancy,require
redundancy,
block
block transmitter
block transmitter
requirelow-complexity
require
transmitter and and
low-complexityarithmetical
low-complexity
re-ceiver
and re-ceiver have
re-ceiver have
arithmeticaloperations
arithmetical
memory
have memory and/or
memory and/or
operationsand
operations
and/or are are linear
are linear
andfast
and fastalgorithms.
fast
time-varying
linear time-varying
algorithms.In
algorithms. Inaddition,
In
(LTV),it
time-varying (LTV),it
addition,whenever
addition,
is possible
(LTV),it isis possible
wheneverthe
whenever
possible toto reduce
to reduce
the
the
reduce
Wallace
Wallace A.
A.Martins
Martins
the
theredundancy
the
memory
redundancyinin
redundancy
memoryitititisisispossible
memory
inthethetransmission
the
possibletoto
possible
transmissioneven
transmission
toeliminate
eliminatethe
eliminate
evenfurther,
even further,asas
further,
theredundancy
the redundancyatat
redundancy atthe
asalso
alsodiscussed
also
thecost
the
discussedinin
discussed
costofof
cost ofmaking
makingthe
making
inthis
thisbook.
this book.For
book.
thechannel
the
Forthe
For
channelequalization
channel
thetransceivers
the transceiverswith
transceivers
equalizationmore
equalization moredifficult.
more
with
with
difficult.
difficult.
Markus
MarkusV.S.
V.S.Lima
Lima
Moreover,
Moreover,when
Moreover, whentime-varying
when time-varyingblock
time-varying blocktransceivers
block transceiversare
transceivers arealso
are alsoemployed,
also employed,then
employed, thenthe
then theamount
the amountofof
amount ofredundancy
redundancycan
redundancy canbe
can be
be
asas
aslow
lowasas
low asaaasingle
singlesymbol
single symbolper
symbol perblock,
per block,regardless
block, regardlessofof
regardless ofthe thesize
the sizeofof
size ofthethechannel
the channelmemory.
channel memory.
memory.
With
Withthe
With thetechniques
the techniquespresented
techniques presentedinin
presented inthe
thebook
the bookitititisisispossible
book possibletoto
possible toaddress
addresswhat
address whatlies
what liesbeyond
lies beyondthe
beyond theuse
the useofof
use ofOFDM-
OFDM-
OFDM-
related
relatedsolutions
related solutionsinin
solutions inbroadband
broadbandtransmissions.
broadband transmissions.
transmissions.

About
AboutSYNTHESIs
About SYNTHESIs
SYNTHESIs
This
Thisvolume
This volumeisisisaaaprinted
volume printedversion
printed versionofof
version ofaaawork
workthat
work thatappears
that appearsinin
appears inthe
theSynthesis
the Synthesis
Synthesis

MOR
MOR GAN
MOR
Digital
DigitalLibrary
Digital Libraryofof
Library ofEngineering
Engineeringand
Engineering andComputer
and ComputerScience.
Computer Science.Synthesis
Science. SynthesisLectures
Synthesis Lectures
Lectures
provide
provideconcise,
provide concise,original
concise, originalpresentations
original presentationsofof
presentations ofimportant
importantresearch
important researchand
research anddevelopment
and development
development

GAN &
GAN &
topics,
topics,published
topics, publishedquickly,
published quickly,inin
quickly, indigital
digitaland
digital andprint
and printformats.
print formats.For
formats. Formore
For moreinformation
more information
information
visit
visitwww.morganclaypool.com
visit www.morganclaypool.com
www.morganclaypool.com
SSYNTHESIS
YNTHESISL
LECTURES ONCCOMMUNICATIONS

& CL
YNTHESIS ECTURESON
ECTURES ON OMMUNICATIONS
OMMUNICATIONS

CL AYPOOL
CL AYPOOL
ISBN:
ISBN:978-1-60845-829-5
&
ISBN: 978-1-60845-829-5
978-1-60845-829-5
Mor
Morgan
Mor gan Cl
gan Claypool
Cl &
&
aypoolPublishers
aypool Publishers
Publishers

AYPOOL
90000
90000
90000
wwwwwwwww. .mmooor rgr ggaaannncccl lal aayyypppooooool l.l .c. ccooomm
.m m
999781608
781608458295
781608 458295
458295
William
WilliamTranter,
William Tranter,Series
Tranter, SeriesEditor
Series Editor
Editor
Block Transceivers
OFDM and Beyond
 Synthesis Lectures on
Communications
Editor
William Tranter, Virginia Tech

Block Transceivers: OFDM and Beyond

Paulo S.R. Diniz, Wallace A. Martins, and Markus V.S. Lima
2012

Basic Simulation Models of Phase Tracking Devices Using MATLAB

William Tranter, Ratchaneekorn Thamvichai, and Tamal Bose
2010

Joint Source Channel Coding Using Arithmetic Codes

Dongsheng Bi, Michael W. Hoffman, and Khalid Sayood
2009

Fundamentals of Spread Spectrum Modulation

Rodger E. Ziemer
2007

Code Division Multiple Access(CDMA)

R. Michael Buehrer
2006

Game Theory for Wireless Engineers

Allen B. MacKenzie and Luiz A. DaSilva
2006
Copyright © 2012 by Morgan & Claypool

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means—electronic, mechanical, photocopy, recording, or any other except for brief quotations in
printed reviews, without the prior permission of the publisher.

Block Transceivers: OFDM and Beyond

Paulo S.R. Diniz, Wallace A. Martins, and Markus V.S. Lima
www.morganclaypool.com

ISBN: 9781608458295 paperback

ISBN: 9781608458301 ebook

DOI 10.2200/S00424ED1V01Y201206COM007

A Publication in the Morgan & Claypool Publishers series

SYNTHESIS LECTURES ON COMMUNICATIONS

Lecture #7
Series Editor: William Tranter, Virginia Tech
Series ISSN
Synthesis Lectures on Communications
Print 1932-1244 Electronic 1932-1708
Block Transceivers
OFDM and Beyond

Paulo S.R. Diniz, Wallace A. Martins, and Markus V.S. Lima

Universidade Federal do Rio de Janeiro

SYNTHESIS LECTURES ON COMMUNICATIONS #7

M
&C Morgan & cLaypool publishers
ABSTRACT
The demand for data traffic over mobile communication networks has substantially increased during
the last decade. As a result, these mobile broadband devices spend the available spectrum fiercely,
requiring the search for new technologies. In transmissions where the channel presents a frequency-
selective behavior, multicarrier modulation (MCM) schemes have proven to be more efficient, in
terms of spectral usage, than conventional modulations and spread spectrum techniques.
The orthogonal frequency-division multiplexing (OFDM) is the most popular MCM method,
since it not only increases spectral efficiency but also yields simple transceivers. All OFDM-based
systems, including the single-carrier with frequency-division equalization (SC-FD), transmit re-
dundancy in order to cope with the problem of interference among symbols. This book presents
OFDM-inspired systems that are able to, at most, halve the amount of redundancy used by OFDM
systems while keeping the computational complexity comparable. Such systems, herein called mem-
oryless linear time-invariant (LTI) transceivers with reduced redundancy, require low-complexity
arithmetical operations and fast algorithms. In addition, whenever the block transmitter and re-
ceiver have memory and/or are linear time-varying (LTV), it is possible to reduce the redundancy in
the transmission even further, as also discussed in this book. For the transceivers with memory it is
possible to eliminate the redundancy at the cost of making the channel equalization more difficult.
Moreover, when time-varying block transceivers are also employed, then the amount of redundancy
can be as low as a single symbol per block, regardless of the size of the channel memory.
With the techniques presented in the book it is possible to address what lies beyond the use
of OFDM-related solutions in broadband transmissions.

KEYWORDS
block transceivers, multicarrier modulation (MCM), orthogonal frequency-division
multiplexing (OFDM), reduced-redundancy transceivers, broadband digital commu-
nications
 vii

Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv

List of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii

List of Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi

1 The Big Picture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Digital Communications Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Orthogonal Frequency-Division Multiplexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 Wired Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Wireless Systems and Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.3 Basics of OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Cellular Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Multiple Access Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5.1 TDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5.2 FDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5.3 CDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5.4 OFDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6 Duplex Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6.1 TDD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6.2 FDD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.7 Wireless Channels: Fading and Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.7.1 Fading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.7.2 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.8 Block Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.9 Multicarrier Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.10 OFDM as MIMO System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.11 Multiple Antenna Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.12 Mitigating Interference and Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.13 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
viii
2 Transmultiplexers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Multirate Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Filter-Bank Transceivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.1 Time-Domain Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.2 Polyphase Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 Memoryless Block-Based Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4.1 CP-OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.4.2 ZP-OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.4.3 CP-SC-FD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.4.4 ZP-SC-FD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.4.5 ZP-ZJ Transceivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3 OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2 Analog OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2.1 From TDM to FDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2.2 Orthogonality Among Subcarriers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.3 Orthogonality at Receiver: The Role of Guard Interval . . . . . . . . . . . . . . . . 62
3.2.4 Spectral Efficiency, PAPR, CFO, and I/Q Imbalance . . . . . . . . . . . . . . . . . 69
3.2.5 Implementation Sketch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.3 Discrete-Time OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3.1 Discretization of The OFDM Symbol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3.2 Discretization at Receiver: The CP-OFDM . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.3.3 Discrete-Time Multipath Channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.4 Block-Based Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.4 Other OFDM-Based Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.4.1 SC-FD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.4.2 ZP-Based Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.4.3 Coded OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.4.4 DMT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

4 Memoryless LTI Transceivers with Reduced Redundancy . . . . . . . . . . . . . . . . . . . 105

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2 Reduced-Redundancy Systems: The ZP-ZJ Model Revisited . . . . . . . . . . . . . . . 107
 ix
4.3 Structured Matrix Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.3.1 Displacement-Rank Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.3.2 Toeplitz, Vandermonde, Cauchy, and Bezoutian Matrices . . . . . . . . . . . . . 118
4.3.3 Properties of Displacement Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.4 DFT-Based Representations of Bezoutian Matrices . . . . . . . . . . . . . . . . . . . . . . . 128
4.4.1 Representations of Cauchy Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.4.2 Transformations of Bezoutian Matrices into Cauchy Matrices . . . . . . . . . 132
4.4.3 Efficient Bezoutian Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.5 Reduced-Redundancy Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.5.1 Complexity Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.5.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

5 FIR LTV Transceivers with Reduced Redundancy . . . . . . . . . . . . . . . . . . . . . . . . . . 149

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
5.2 Time-Varying Reduced-Redundancy Systems with Memory . . . . . . . . . . . . . . . . 151
5.2.1 FIR MIMO Matrices of LTI Transceivers . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.2.2 FIR MIMO Matrices of LTV Transceivers . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.3 Conditions for Achieving ZF Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5.3.1 The ZF Constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5.3.2 Lower Bound on The Receiver Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.3.3 Lower Bound on The Amount of Redundancy . . . . . . . . . . . . . . . . . . . . . . 163
5.3.4 Achieving the Lower Bound of Redundancy . . . . . . . . . . . . . . . . . . . . . . . . 164
5.3.5 Role of The Time-Variance Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.4 Transceivers with No Redundancy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
5.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

Authors’ Biographies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

 xi

Preface
The widespread use of mobile devices with high processing capabilities, like smartphones and tablets,
as well as the increasing number of users and growing demand for higher data rates are some of the
main reasons why data traffic over mobile communication networks has increased so much during
the last decade. As a result, these mobile broadband devices spend the available spectrum fiercely,
requiring the search for new technologies.
From the physical-layer viewpoint, the first step to address this problem is to choose a mod-
ulation scheme which is more adequate for the type of channel through which the signal (wave)
propagates. Indeed, the spectral efficiency of communications systems can significantly increase
by properly choosing the modulation scheme. For example, in broadband transmissions in which
the channel presents a frequency-selective behavior, multicarrier modulation (MCM) schemes have
proven to be more efficient, in terms of spectral usage, than conventional modulations and spread
spectrum techniques.The multicarrier transmission illuminates the physical channel utilizing several
non-overlapping narrowband subchannels, where each subchannel appears to be flat, thus turning
the equalization process simpler.
Among the existing MCM schemes, the orthogonal frequency-division multiplexing
(OFDM) is the most notorious, since it not only increases spectral efficiency but also yields simple
transceivers. OFDM is capable of eliminating the intersymbol interference (ISI) with very simple
transmitter and receiver, by performing low-complexity computations such as insertion and removal
of a prefix, and by using fast algorithms such as the fast Fourier transform. It is worth noting that
ISI is one of the most harmful effects in broadband transmissions. Simple transceivers are attractive
since they lead to lower latency and require less power consumption. Therefore, it is no surprise
that OFDM has been adopted by many wired and wireless broadband communication technologies.
For instance, the long-term evolution (LTE) is a wireless communication standard whose down-
link connection is based on OFDM, whereas its uplink connection is based on the single-carrier
with frequency-division equalization (SC-FD), which is similar to OFDM in many aspects and is
composed of the same building blocks. The enhancements introduced by LTE physical layer are so
significant that LTE achieves much higher data rates, as compared to 3rd generation (3G) systems,
and is already being considered a 4th generation (4G) system.
At this moment, given the desirable features of OFDM and SC-FD schemes and their
widespread use in both wired and wireless communication standards, one could ask the following
questions: Is this the best we can do in terms of spectral efficiency? When LTE spectrum gets
overloaded, what comes next? Can we further improve these schemes?
This book tries to provide some directions to address these questions. Both OFDM and SC-
FD systems transmit redundancy (the prefix) in order to cope with the problem of ISI. Indeed, a
xii PREFACE
portion of each data block is reserved for the prefix, whose size must be larger than the channel
memory for OFDM and SC-FD systems. Thus, spectral efficiency may be increased if less redun-
dancy could be used. This book presents OFDM-inspired systems that are able to, at most, halve
the amount of redundancy used by OFDM systems while keeping the computational complexity
comparable. Such systems, herein called memoryless linear time-invariant (LTI) transceivers with
reduced redundancy, require low-complexity arithmetical operations and fast algorithms. In addition,
whenever the block transmitter and receiver have memory and/or are linear time-varying (LTV), it
is possible to reduce the redundancy in the transmission even further, as discussed in the last chapter
of this book. For the transceivers with memory it is possible to eliminate the redundancy at the
cost of making the channel equalization a more difficult task. Moreover, when time-varying block
transceivers are also employed, then the amount of redundancy can be as low as a single symbol per
block, regardless of the size of the channel memory. An example is the code-division multiple access
(CDMA) system with long spreading codes which is always able to achieve ISI elimination as long
as the system is not at full capacity, i.e., at least one spreading/de-spreading code is unused.
The approach followed in this book is to present both OFDM and SC-FD systems as particular
cases of the so-called transmultiplexer (TMUX). In fact, these two systems belong to the category
of memoryless TMUXes. Special attention is given to OFDM and its desirable properties, since
OFDM is being employed in many standards. Then, the TMUX is used as the main framework to
derive both LTI and LTV transceivers with reduced redundancy.
With the techniques presented in the book it is possible to address what lies beyond the use of
OFDM-related solutions in broadband transmissions. In summary, this book presents solutions to
reduce the redundancy in transmission aiming at increasing data throughput. However, it is worth
mentioning that reducing redundancy might increase mean square error (MSE) and bit-error rate
(BER), as well as turn the design of the block transceiver more challenging. The optimum solutions
are environment-dependent and its proper sensing leads to much more efficient spectral usage.

ORGANIZATION OF THE BOOK

Chapter 1 aims at providing a big picture of digital and wireless communications. This chapter
differs from the others in the sense that the material is presented in a pictorial manner, avoiding the
mathematics whenever possible. The reasons for this choice are: (i) to provide a quick overview of the
field without wasting time explaining concepts that are not central to this book, most of which are
thoroughly explained in digital communication and wireless communication textbooks; and (ii) to
concentrate on ideas rather than mathematics. Therefore, at the end of this chapter the reader should
have recollected topics such as: digital modulation, channel encoder, cellular systems, multiple access
methods, frequency-selective channels, multicarrier modulation schemes, and OFDM.
Chapter 2 briefly presents multirate signal processing fundamentals that are of major impor-
tance to fully understand the TMUX framework, which is employed throughout the rest of the book.
In this chapter, it is shown that TMUXes are general structures that can be used to represent/model
several communications systems. In particular, OFDM and SC-FD systems can be interpreted as
 PREFACE xiii
memoryless TMUXes, whose implementations are based on memoryless block-based transceivers.
This chapter also introduces some initial results related to what is beyond OFDM-based systems.
Chapter 3 introduces OFDM from its original analog conception to its actual discrete-time
practical usage. The chapter starts with the analog OFDM, exploring the role played by the guard
interval in maintaining the orthogonality among OFDM subcarriers, and explaining the choices for
some parameters such as the OFDM symbol duration and distance between adjacent subcarriers.
Then, the discrete-time OFDM is studied in connection with its analog version. In addition, many
other topics related to OFDM are covered, such as the coded OFDM (C-OFDM), issues of OFDM
transmissions like the peak-to-average power ratio (PAPR), discrete multitone (DMT) systems, and
optimal power allocation.
Chapter 4 presents multicarrier and single-carrier memoryless LTI transceivers that use a
reduced amount of redundancy, as compared to OFDM and SC-FD systems. That is, this chap-
ter describes how transceivers with reduced redundancy can be implemented employing superfast
algorithms based on the concepts of structured matrix representations. Thus, part of this chapter de-
scribes structured matrices and the displacement theory that allows the derivation of these superfast
algorithms.
The focus of Chapter 5 is on the fundamental limits of some parameters related to LTV
transceivers. In particular, we consider the memory of the multiple-input multiple-output (MIMO)
receiver matrix and the number of transmitted redundant elements, which are inherent to finite
impulse response (FIR) LTV transceivers satisfying the zero-forcing (ZF) constraint. In addition,
it is shown that ZF equalizers cannot be achieved when no redundancy is used, and as alternative
pure MMSE-based solutions are presented.

PREREQUISITES
We attempted to make this book as self contained as possible. Although basic knowledge of wireless
communications, digital transmission, and multirate signal processing is highly desirable, it is not
necessary since the first two chapters revisit the main concepts which are used throughout the book.
Thus, the main prerequisites to follow this book are: digital communications, basic concepts of
stochastic processes—involving expected values, means, and variances of random variables—and
linear algebra—involving operations with vectors and matrices, ranks, determinants, null and range
spaces. More advanced concepts, such as structured matrices and displacement theory, are explained.

Paulo S.R. Diniz, Wallace A. Martins, and Markus V.S. Lima

Rio de Janeiro
June 2012
 xv

Acknowledgments
The authors are grateful to Joel Claypool for kindly pushing us to finish this project. They are also
thankful to Professors M.L.R. de Campos, E.A.B. da Silva, L.W.P. Biscainho, and S.L. Netto of
UFRJ, and Professor T.N. Ferreira of UFF for their incentive and always being available to answer our
questions. They would like to thank Professors R. Sampaio Neto of PUC-RJ and V.H. Nascimento
of USP who influenced some parts of this book. Wallace thanks his colleagues at the Federal Center
for Technological Education Celso Suckow da Fonseca (CEFET/RJ–UnED-NI), in particular at
the Department of Control and Automation Industrial Engineering.
We also would like to thank our families for their patience and support during this challenging
process of writing a book. Paulo would like to thank his parents, his wife Mariza, and his daughters
Paula and Luiza for illuminating his life. Wallace thanks his fiancee Claudia and his parents, Renê
and Perpétua. Markus thanks his parents, Luiz Álvaro and Aracy, and his girlfriend Bruna.

Paulo S.R. Diniz

Wallace A. Martins
Markus V.S. Lima
Rio de Janeiro, Brazil
June 2012
 xvii

List of Abbreviations
2G 2nd Generation
3G 3rd Generation
3GPP 3rd Generation Partnership Project
4G 4th Generation
ADSL Asymmetric Digital Subscriber Line
BER Bit Error Rate
CDMA Code-Division Multiple Access
CFO Carrier-Frequency Offset
CP-OFDM Cyclic-Prefix Orthogonal Frequency-Division Multiplexing
CP-SC-FD Cyclic-Prefix Single-Carrier with Frequency-Division equalization
CSI Channel-State Information
DAB Digital Audio Broadcasting
DFT Discrete Fourier Transform
DHT Discrete Hartley Transform
DMT Discrete MultiTone
DS-CDMA Direct Sequence CDMA
DSP Digital Signal Processing
ETSI European Telecommunications Standards Institute
FDD Frequency-Division Duplex
FDM Frequency-Division Multiplexing
FDMA Frequency-Division Multiple Access
FFT Fast Fourier Transform
FH-CDMA Frequency-Hopping CDMA
FIR Finite Impulse Response
GSM Global System for Mobile communications
IBI InterBlock Interference
ICI InterCarrier Interference
IDFT Inverse Discrete Fourier Transform
IEEE Institute of Electrical and Electronics Engineers
IIR Infinite Impulse Response
ISI InterSymbol Interference
LAN Local Area Network
LTE Long Term Evolution
LTI Linear Time-Invariant
xviii LIST OF ABBREVIATIONS
MA Multiple Access
MAN Metropolitan Area Network
Mbps Megabits per second
MC-MRBT MultiCarrier Minimum-Redundancy Block Transceiver
MC-RRBT MultiCarrier Reduced-Redundancy Block Transceiver
MIMO Multiple-Input Multiple-Output
MISO Multiple-Input Single-Output
MMSE Minimum Mean Square Error
MSC Mobile Switching Center
MSE Mean Square Error
MUI MultiUser Interference
OFDM Orthogonal Frequency-Division Multiplexing
OFDMA Orthogonal Frequency-Division Multiple Access
OLA Overlap-And-Add
PAM Pulse-Amplitude Modulation
PAN Personal Area Network
PAPR Peak-to-Average Power Ratio
PSD Power Spectrum Density
PSK Phase-Shift Keying
QAM Quadrature Amplitude Modulation
QPSK Quadrature PSK
SC Single Carrier
SC-FD Single-Carrier with Frequency-Domain equalization
SC-FDMA Single-Carrier Frequency-Division Multiple Access
SC-MRBT Single-Carrier Minimum-Redundancy Block Transceiver
SC-RRBT Single-Carrier Reduced-Redundancy Block Transceiver
SGSN Serving GPRS Support Node
SIMO Single-Input Multiple-Output
SISO Single-Input Single-Output
SNR Signal-to-Noise Ratio
SVD Singular Value Decomposition
VDSL Very high-speed Digital Subscriber Line
TDD Time-Division Duplex
TDM Time-Division Multiplexing
TDMA Time-Division Multiple Access
TMUX Transmultiplexer
UMTS Universal Mobile Telecommunications System
 LIST OF ABBREVIATIONS xix
WAN Wide Area Network
Wi-Fi Wireless Fidelity
WiMAX Worldwide interoperability for Microwave ACCess
WLAN Wireless Local Area Network
WPAN Wireless Personal Area Network
WSS Wide-Sense Stationary
xDSL high-speed Digital Subscriber Line
ZF Zero-Forcing
ZP Zero-Padding
ZP-OFDM Zero-Padding OFDM
ZP-OFDM-OLA ZP-OFDM OverLap-and-Add
ZP-SC-FD Zero-Padding SC-FD
ZP-SC-FD-OLA ZP-SC-FD OverLap-and-Add
ZP-ZJ Zero-Padding Zero-Jamming
 xxi

List of Notations
Scalars Lowercase letters, e.g., x
Vectors Lowercase boldface letters, e.g., x
Matrices Uppercase boldface letters, e.g., X
 Definition
N Set of natural numbers, which is defined as N  {1, 2, 3, . . .}
Z Set of integer numbers
R Set of real numbers
C Set of complex numbers
t Real-valued variable representing continuous time
n Integer number representing discrete time
j Imaginary unit, j2 = −1
ω Angular frequency
δ(t) Dirac impulse
δ[n] Kronecker delta
WM Unitary DFT matrix of size M × M
IM Identity matrix of size M × M
em Canonical vector, e.g., e0 = [1 0 . . . 0]T
(·)T Transpose of matrix (·)
(·)H Hermitian (conjugate) transpose of matrix (·)
0M×N M × N matrix with all entries equal to 0
E[(·)] Expected value of (·)
Z {(·)} Z -transform applied to (·)
Z −1 {(·)} Inverse Z -transform applied to (·)
tr{(·)} Trace of matrix (·)
rank{(·)} Rank of matrix (·)
diag{(·)} Diagonal matrix whose entries in its diagonal are (·)
ker{(·)} Kernel (Null space) of matrix (·)
R{(·)} Range (Column space) of matrix (·)
F {(·)} Fourier transform of (·)
[(·)]↓N Decimation operator by N
[(·)]↑N Interpolation operator by N
[(·)]ml Entry of matrix (·) in the mth row and lth column
DFT{(·)} Discrete Fourier transform of sequence (·)
 1

CHAPTER 1

The Big Picture

1.1 INTRODUCTION
In communications the ultimate goal is the transport of as much information as possible through a
propagation medium. Currently, most transmissions are performed by propagating electromagnetic
energy through the air or wired channels. The majority of our current communications systems
transmit digital data to benefit from the widely available digital technology which has become
relatively cheap. The digital technology is also reliable, amenable to error detection and correction,
and reproducible. All these features reduce the cost of transmission and make available new services
to the end users.
The wired channels involve a physical connection between fixed communication terminals,
usually consisting of guided electromagnetic channels such as twisted-pair wirelines and coaxial
cables. In addition, optical fibers are becoming increasingly popular due to their channel bandwidth.
Indeed, the bandwidth of optical fiber is in general some orders of magnitude larger than in coaxial
cables. Even though there is a trend for replacing wireline by optical-fiber channels, the low costs and
improvements in modem designs have extended the lifetime of many wireline connections which
were already deployed.
In wireless communications, the channel is the medium through which the electromagnetic
energy propagates. Examples of such mediums are the air and the water. In this book, the focus is on
wireless communications through the air, which we will refer to only as wireless communications.1 In
such communications systems, the electromagnetic energy is radiated to the propagation medium
via an antenna.
Unlike wired transmissions, wireless transmissions require the use of radio spectrum, which
in turn should be carefully managed by government regulators. The current trend of increasing the
demand for radio transmissions shows no sign of settling. The amount of wireless data services is
more than doubling each year leading to spectrum shortage as a sure event in the years to come. As
a consequence, all efforts to increase the efficiency of spectrum usage are highly justifiable at this
point. In response this book addresses the issue of how to increase the spectral efficiency of radio
links by properly designing the transceivers, especially for multicarrier systems, which include the
popular orthogonal frequency-division multiplexing (OFDM) as a special case.
This chapter starts with a brief description of digital communications systems in Section 1.2.
Section 1.3 motivates the study of OFDM by giving examples of several communications standards
that use it, with emphasis to the WiMAX and LTE systems. The next two sections contain key
1 Communications through the water are usually referred to as underwater communications.
2 1. THE BIG PICTURE
concepts of mobile and multiuser communications. Indeed, Section 1.4 introduces the cellular-
division paradigm, while Section 1.5 briefly describes the most used multiple access schemes and
also explains how multiple access can be performed in OFDM-based systems, leading to the so-
called OFDMA (orthogonal frequency-division multiple access). After that, Section 1.6 describes
duplex methods.Then, Section 1.7 introduces the main problems present in wireless communications
systems, emphasizing the multipath effect. In Section 1.8, a central topic of this book, namely block
transceivers, is introduced. In this section, the use of a guard time between blocks is exemplified
as a naive solution to avoid interblock interference. Section 1.9 presents the fundamental idea of
multicarrier systems, which is the division of the channel spectrum in narrowband and approximately
flat fading subchannels, that allows simple equalization schemes at the receiver, as it will be discussed
in the following chapters. Section 1.10 shows for the first time the simple, yet powerful, mathematical
model for representing OFDM as a MIMO (multiple-input multiple-output) system. The target
is to illustrate that the low complexity of the OFDM is due to parameter decoupling, which is
achieved by judicious design of both the precoder (transmitter) and postcoder (receiver). Finally,
in Sections 1.11 and 1.12 we briefly discuss multiple antenna systems and the interference issue,
respectively.

1.2 DIGITAL COMMUNICATIONS SYSTEMS

Using a simplified building-block representation, the main elements of a digital communication
system are depicted in Figure 1.1 (see [4, 7, 22, 26, 47, 57, 65, 66, 75] for further details).
Firstly, typical transmitted input signals, like data, speech, audio, image, and video, are com-
pressed at the source encoder. By exploring the redundancy2 of the input signal, the source encoder is
capable of representing it in a more compact form, i.e., using less bits. Depending on the input signal
nature, the compression process can be either lossless, which means that the original signal can be
exactly recovered from the compressed signal, or lossy, which does not allow exactly restoration of the
input signal, but yields a much higher compression rate. Lossy compression schemes use perceptive
criteria to discard information that is not perceived by the end users. For instance, an audio signal at
the receiver end should keep the perceptive quality as close as possible to the original input signal.
The next building block is the channel encoder whose primary task is to protect the compressed
input information against the physical-channel impairments. The channel encoder allows for error
detection and correction by adding some bits to the compressed signal. This building block acts as
a wrap to encase and secure the information to be transported [38, 90].
Once our package of information is ready, it must be represented in a proper format to cross
the channel.This task is performed by the digital modulator which is responsible for mapping bits into
waveforms. These waveforms are mathematically represented by complex numbers and are usually
called symbols. After crossing the channel the received waveform (symbol) is usually a distorted
2 Here, the term redundancy means predictability, which is a common characteristic of natural signals, and is related to the concept
of entropy [14]. In the rest of the book the term redundancy is related to the prefix or suffix used in block transmissions, a topic
that will be covered in detail in the following chapters.
 1.3. ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING 3
Transmitter

Input Source Channel Digital

Signal Encoder Encoder Modulator

Channel

Output Source Channel Digital

Signal Decoder Decoder Demodulator

Receiver

Figure 1.1: Simplified representation of digital communication systems.

version of the transmitted waveform since the latter suffers attenuation and other wave propagation
effects, and interferences caused by environmental noise and by other signals being transmitted
through the same channel.
At the receiver end, all the strategies utilized at the transmitter to improve the channel usage
are undone. The received waveform is converted to bits by the digital demodulator, the channel decoder
is responsible for correcting some bits that were erroneously detected, and then the source decoder
undoes the compression process, which is not a perfect reversion process in cases of using lossy
compression schemes (as explained before), generating an output signal as close as possible to the
input signal.
The next section addresses the OFDM, a transmission scheme that enables efficient trans-
mission of symbols (waveforms) through the channel.

1.3 ORTHOGONAL FREQUENCY-DIVISION

MULTIPLEXING
Orthogonal frequency-division multiplexing (OFDM) is a transmission technique that is currently
used in a number of practical systems, such as:

• Digital radio broadcasting or digital audio broadcasting (DAB);

• Wireless local area network (WLAN);

• Wireless broadband links;

• High-speed digital subscriber line (xDSL);

• 4th generation (4G) cellular communications;

• Digital broadcasting TV.

4 1. THE BIG PICTURE
Because of its great importance and widespread use, Chapter 3 is dedicated to explain OFDM
and some of its variants. By now, it is important to notice that OFDM is applicable to both wired
and wireless links, as exemplified in the previous list of applications.

1.3.1 WIRED SYSTEMS

An important player in wired communications is the xDSL, a general term for all broadband access
technologies based on digital subscriber line. The xDSL systems provide customers with high data
rates using the already existing copper pairs inherited from the fixed telephony. The main xDSL
systems are the following:

• Asymmetric DSL (ADSL) – the asymmetric term means that upstream (i.e., from costumer to
network) and downstream (i.e., from network to costumer) data rates are different. Nowadays,
evolved versions of the ADSL, such as the ADSL2+M, can achieve upstream rates up to
3.5 Mbps (megabits per second) and downstream rates up to 24 Mbps.

• Very high-speed DSL (VDSL) – the first VDSL systems, known as VDSL1, provided data
rates higher than the ADSL. The drawback of VDSL1 is that its data rate decreases too fast
as the distance from the subscriber premises to the network increases, limiting its usage to
short local loops. Nowadays, VDSL2 systems can achieve upstream rates up to 10 Mbps and
downstream rates up to 50 Mbps, and perform quite similar to the ADSL2+ when transmitting
over long distances, unlike VDSL1.

Both ADSL and VDSL use discrete multitone (DMT), which is essentially a sophistication of
OFDM.

1.3.2 WIRELESS SYSTEMS AND NETWORKS

A set of complementary wireless standards are available. These standards are mainly divided accord-
ing to their coverage area and their main classes are:3

• Personal area network (PAN);

• Local area network (LAN);

• Metropolitan area network (MAN);

• Wide area network (WAN).

Figure 1.2 depicts these main classes according to their coverage area, and also shows examples
of wireless standards. Note that some technologies may appear in more than one class of wireless
networks, since they may receive upgrades that extend their original coverage area. For instance,
3 Some of these classes also exist for wired systems. Indeed, PAN and LAN can be used for both wired and wireless systems. In
cases when one wants to refer just to the wireless part of these systems, the terms wireless PAN (WPAN) and wireless LAN
(WLAN) can be used.
 1.3. ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING 5

WAN

IEEE 802.20 ETSI GSM (2G) UMTS (3G)
(Mobile-Fi) 3GPP
MAN LTE-Advanced (4G)

IEEE 802.16 ETSI

(WiMAX) HIPERMAN
LAN
IEEE 802.11 ETSI
(Wi-Fi) HIPERLAN
PAN
IEEE 802.15
(Bluetooth)

Figure 1.2: Wireless networks and corresponding technologies.

the WiMAX (worldwide interoperability for microwave access) is usually considered an example of
MAN, but it can also be seen as a WAN system.
Many wireless network standards use OFDM in the air interface. Two examples are the LTE
(long-term evolution) and WiMAX standards. LTE was designed to fully replace the 3rd generation
(3G) networks for mobile communications. The WiMAX, although originally conceived to provide
wireless broadband services to homes, has been upgraded to be employed by mobile phones as access
method in recent years, competing with LTE.

(A) WiMAX

The main target of IEEE 802.16 standard, known as WiMAX, is to deliver wireless high-speed
Internet access over longer distances than the ones supported by the IEEE 802.11 standard,
commonly known as Wi-Fi (wireless fidelity). Indeed, WiMAX provides a MAN with wireless
broadband service in an area of 50 km (about 30 miles) of radius. The WiMAX is guided by an
association called WiMAX Forum and its data rates can reach up to 40 Mbps for low mobility
access and up to 15 Mbps for mobile access. The capabilities and coverage area of WiMAX
systems, especially due to the amendment e of IEEE 802.16 standard (IEEE 802.16e, also
called Mobile WiMAX), increased in such a way that the IEEE 802.20 standard was put to
hibernation.

WiMAX employs an adaptive modulation scheme as illustrated in Figure 1.3 in which the
digital modulator is adjusted according to the signal-to-noise ratio (SNR). Through low-SNR
channels, usually consisting of channels where the user is far from the base station, a sparser
modulation scheme, the quadrature phase-shift keying (QPSK), is used. On the other hand,
users near the base station are likely to have high-SNR channels and, therefore, transmissions
with higher-order modulation schemes, such as the 16- or 64-quadrature amplitude modulation
6 1. THE BIG PICTURE
High SNR

Medium SNR

Low SNR

64-QAM
16-QAM
QPSK

Figure 1.3: Adaptive modulation in WiMAX.

(QAM), are allowed. In addition, the multiple access scheme used in WiMAX is the orthogonal
frequency division multiple access (OFDMA).
Other goals of WiMAX are:

• Broadband on demand – the deployment of LAN hotspots can be accelerated by MAN,

especially at locations not served by xDSL or other cable-based technologies, and the
phone companies cannot provide broadband services at short notice.
• Cellular operator backhaul – these operators usually use wired and leased connections
from third-party service providers and they could be replaced by MAN.
• Residential broadband – in many areas it is difficult to provide wired broadband services.
• Wireless services to rural and scarcely populated areas.
• Wireless everywhere – as the number of hotspots increases, the demand for wireless services
in areas not covered by LAN increases as well.

(B) LTE
LTE is a standard for wireless data communication that is capable of overcoming some limi-
tations of GSM/UMTS (global system for mobile communications/universal mobile telecom-
munications system) standards. It was developed by the 3rd Generation Partnership Project
(3GPP) and its first version was established in December 2008, when 3GPP Release 8 was
frozen. The main motivations for LTE are the user demands for higher data rates and quality
of service, the necessity of optimizing the packet-switched system, and the demands for cost
and complexity reduction.
 1.3. ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING 7

Δf

Figure 1.4: Basics of OFDM: overlapping but not interfering subcarriers.

Similar to the WiMAX, LTE uses OFDMA in the downlink connection. However, in the
uplink connection, LTE opted for using the single-carrier frequency-division multiple access
(SC-FDMA). In addition, LTE has very low latency and can operate on different bandwidths:
1.4, 3, 5, 10, 15, and 20 MHz.
3GPP Releases are continuously provided such as the most recent version called Release 11. It is
worth mentioning that from Release 10 and beyond, the LTE is usually called LTE-Advanced,
since it became IMT-Advanced compliant. Some of the key features of IMT-Advanced com-
pliant standards are:

• Worldwide functionality and roaming;

• Compatibility of services;
• Interworking with other radio access systems;
• Enhanced peak data rates to support advanced services and applications (100 Mbps for
high and 1 Gbps for low mobility).

1.3.3 BASICS OF OFDM

In the basic application of OFDM, a data stream is divided into blocks and each entry of the data
block modulates a subcarrier with overlapping but not interfering frequency spectrum. As illustrated
in Figure 1.4, the OFDM subcarriers are overlapped in frequency and the distance between the
mainlobes of adjacent subcarriers is equal to f Hertz. Once a bandwidth is defined for provision of
a certain service, it can be occupied by a set of OFDM subcarriers. The summation of bandlimited
8 1. THE BIG PICTURE

Figure 1.5: Coverage area: cellular division and frequency reuse.

OFDM subcarriers should occupy most of the available bandwidth so that the amount of symbols
transmitted is maximized.
In the next section we introduce the concept of cellular division, which was crucial to the
success (widespread use) of mobile communications.

1.4 CELLULAR DIVISION

The mobile communication capitalizes on the significant attenuation a transmitted signal faces in
the wireless channel. This allows the wireless-service providers to reuse the available frequency range
at distinct locations distant enough from each other. During wireless service deployment, the service
provider defines a coverage region that consists of cell clusters. Within these clusters each cell is
assigned with a different channel group.
Figure 1.5 illustrates a coverage area in which each cell is represented by a hexagon and each
collection of hexagons including all colors is a cluster. Hexagons of the same color represent cells
sharing the same radio resources, i.e., transmitting on the same frequencies. Each cluster utilizes the
whole radio resources that can be reused by all other clusters. In actual wireless networks there are
equipments, such as the mobile switching center (MSC) for GSM systems and the serving GPRS
support node (SGSN) for UMTS, which dynamically distribute the radio resources among the users.
These equipments can also manage an exchange of wireless channels among cells and among clusters.
 1.5. MULTIPLE ACCESS METHODS 9
When the mobile moves toward the frontier of two cells, the signal connection with the
original cell becomes weak. Before it reaches the minimum level of acceptance in quality, a handoff
to the neighboring cell should be made without interruption.
Note that the cellular division enables different users to transmit simultaneously provided
they are in different cells. But, if the users are in the same cell, then a multiple access scheme must
be used in order to allow simultaneous transmission for these users.

1.5 MULTIPLE ACCESS METHODS

Multiple access (MA) methods enable multiple users to transmit over the same channel. Indeed, the
choice of the MA method determines how the radio resources, sometimes called channel resources,
are shared among users. The fundamental idea underlying MA methods is the concept of separability
among users, which means that the signals transmitted by different users, although sharing the same
transmission medium, should be completely separable at the receiver. Mathematically, this implies
that there must exist some domain (e.g., time, frequency, space, or code) in which the waveforms
corresponding to different users are orthogonal to each other. The appropriate choice of the access
method is key to achieve high data-rate transmissions, thus increasing the capacity of the networks
to provide multiple services to several users.
The MA methods can be classified according to the domain in which the users are separable.
The main MA methods under use are:

• Time-division multiple access (TDMA);

• Frequency-division multiple access (FDMA);

• Code-division multiple access (CDMA);

• Orthogonal frequency-division multiple access (OFDMA).

These MA methods are briefly described in the following subsections. Note, however, that
they are not mutually exclusive. Indeed, the most successful 2G system, the GSM, is a hybrid of
TDMA and FDMA. In GSM, each 200 kHz of channel bandwidth is shared among 8 users using
TDMA.

1.5.1 TDMA
In TDMA, the separability among users occurs in the time domain, i.e., each user receives a time
slot (or a set of them) for his transmission. Figure 1.6 illustrates two users sharing the same medium
using a TDMA scheme. The duration of each slot is T seconds, and each user receives two of them.
During the slots that User 1 is transmitting, User 2 is in silence, and vice versa. The signal that arrives
at the receiver, assuming the channel does not introduce any distortion to the transmitted signals, is
the superposition of the signals sent by the two users. Therefore, in TDMA the transmitters must
be tightly synchronized in order to avoid interference among them. It is clear that the receiver must
10 1. THE BIG PICTURE
be synchronized with the transmitters as well, since it must know how to properly chop the received
signal in order to isolate the signals sent by the different users.
T T T T

T T T T
User 1
t

User 2
t Received Signal
(Without Distortion)

Figure 1.6: Example of TDMA involving two users.

1.5.2 FDMA
Unlike TDMA, in FDMA users may transmit all the time. The separability among them occurs in
the frequency domain, i.e., for each user it is assigned a different frequency band. Figure 1.7 depicts
two users employing an FDMA scheme to share the channel resources. Both users transmit over
the channel using a bandwidth of f Hz, but the central frequency of User 1 is f1 , whereas it is f2
for User 2. Clearly, f1 and f2 must be distant enough from each other in order to avoid interference
between the two users. At the receiver side, the signals belonging to the two users can be separated
through bandpass filters that dramatically attenuate all frequencies, except the ones that fall within
the desired band.
As the number and types of services available to users increase, the fixed channel assignment
inherent to FDMA and TDMA becomes less efficient. As a result, more spectrum would be required
with such fixed assignment. Even if some kind of flexible channel assignment is incorporated in the
current FDMA and TDMA schemes, there is always a fixed upper bound in the number of users
that can be served.

1.5.3 CDMA
The CDMA, sometimes called spread-spectrum, appears in two distinct forms: the direct sequence
CDMA (DS-CDMA) and the frequency-hopping CDMA (FH-CDMA).
In DS-CDMA, the spectrum of the baseband signal is spread to occupy a wider bandwidth,
as depicted in Figure 1.8. In this figure, the power spectrum density (PSD) is represented as a
function of the frequency. Before transmitting the waveform conveying information, this waveform
 1.5. MULTIPLE ACCESS METHODS 11
Δf Δf

Δf Δf
User 1
f1 f2 f

f1 f2 f

User 2
f1 f2 f Received Signal
(Without Distortion)

Figure 1.7: Example of FDMA involving two users.

is modified in order to spread it over a wider frequency band. At the receiver, the reverse process
is performed and the spread-spectrum waveform is converted back to the original waveform from
which the information can be extracted.

PSD PSD
Original
Signal Spreading
Spread
Original
Signal
Despreading

Frequency Frequency

Figure 1.8: Spreading the spectrum in CDMA.

One of the main benefits from using spread-spectrum is reducing narrowband interference.
For instance, consider Figure 1.9 in which the spread original signal suffers interference due to
two narrowband interferers. In this case, the de-spreading process performed at the receiver de-
spreads the original signal and also spreads the interferers, as depicted in Figure 1.10. Consequently,
a significant portion of the signal power corresponding to the interferers is spread over frequency
bands different from the ones used by the original signal, thus reducing the inband interference.
Another main advantage of CDMA is that it allows users to share all the available bandwidth
simultaneously. Hence, since the signal transmitted by each user is wideband, CDMA must resolve
such type of wideband interference in order to guarantee the separability among users. Indeed,
CDMA introduces a new domain, called code domain, in which multiple users can be fully separated.
12 1. THE BIG PICTURE

PSD Narrowband
Interferers

Spread
Original
Signal

Frequency

Figure 1.9: Narrowband interference in CDMA.

Original
PSD Signal

Spread
Narrowband
Interferers

Frequency

Figure 1.10: Mitigating narrowband interference in CDMA de-spreading process.

Figure 1.11 illustrates a DS-CDMA passband-transmission scheme. In this case, the mth
user, m ∈ N, intends to transmit a sequence of symbols bm (nTb ), in which Tb denotes the symbol
duration and the sequence of symbols is indexed by n ∈ Z. Before transmitting, the symbols bm (nTb )
are multiplied by a code sequence cm (kTc ), where k ∈ Z represents the sequence index and Tc
corresponds to the duration of each element of the code sequence. In addition, cm (kTc ) is a very simple
sequence whose elements are 1 or −1. The main difference between the sequences bm (nTb ) and
cm (kTc ) is that the sampling rate of the latter is much higher, i.e., Tc  Tb . Hence, the signal resulting
from the multiplication of bm (nTb ) by cm (kTc ) is a spread-spectrum signal whose bandwidth is
increased by a factor equal to Tb /Tc , also known as spreading factor or processing gain. Finally, this
resulting signal modulates the carrier cos (ωc t), where ωc is the carrier frequency, resulting in a
passband signal ready to be transmitted. The key point that guarantees the separability among users
is that these code sequences are unique, i.e., a different code sequence is assigned to each user, and
they are also orthogonal to each other.4

4 In fact, due to some transmission issues, such as synchronization, orthogonal code sequences are sometimes exchanged by (or
used together with) pseudonoise sequences.
 1.5. MULTIPLE ACCESS METHODS 13

bm (nTb )

cm (kTc ) cos (ωc t)

Figure 1.11: DS-CDMA passband-transmission scheme.

Note that while TDMA and FDMA have a maximum number of simultaneous users, CDMA
does not have such a constraint. Nevertheless, in CDMA, if the number of users increases the service
smoothly degrades.
In cellular systems, CDMA allows soft handover between neighboring cells. Close to the
frontier, where the signal is weaker, the user communicates with two base stations simultaneously so
that the diversity helps compensating for signal degradation at cell edges. In TDMA and FDMA
neighboring cells must use different frequencies because they control interference based on spatial
attenuation of the signals (frequency reuse). In CDMA, all cells use the same frequency range, elim-
inating the necessity for frequency-use planning, whereas TDMA and FDMA may use adaptive
frequency reallocation. In addition, CDMA requires strict power control and base station synchro-
nization, and allows intercell interference to be suppressed at the receiver.
Another type of spread-spectrum technique is the FH-CDMA, in which each user employs
a different frequency band within a given time frame. During transmission, the user frequency
band hops to different bands according to a prescribed hopping pattern (code). The receiver hops
synchronously with the transmitter with the knowledge of the code. The hopping can be slow, where
hopping occurs at the symbol rate, or fast where more than one hop occurs during symbol duration.
The latter case is more difficult to implement.
As a rule, wideband MA schemes can operate in the frequency range of existing narrowband
services and allow flexibility in the number of users and services provided to each user.They also allow
improved interference rejection, originated from multiuser, multipath, and narrowband interferences.
On the other hand, wideband MA systems require more advanced technology for implementation.
Figure 1.12 summarizes how the radio resources are assigned to one user for TDMA, FDMA,
and CDMA schemes. As we have already discussed, the separability among the multiple users occurs
in one of the following domains: time, frequency, or code domains. Therefore, in TDMA the user
receives a time interval for his transmission, which may use the whole channel bandwidth. In FDMA,
a frequency band is dedicated for the user’s transmission. In CDMA, a single user may transmit all
the time and using the entire channel bandwidth, but with a unique code.

1.5.4 OFDMA
In some standards, sets of OFDM subchannels can be assigned to distinct users leading to an MA
scheme known as orthogonal frequency-division multiple access (OFDMA). In the simplest case,
14 1. THE BIG PICTURE

Frequency Frequency
Slot

Slot

Time Time

Code Code
(a) TDMA. (b) FDMA.

Frequency

Slot

Time

Code
(c) CDMA.

Figure 1.12: Channel sharing for: (a) TDMA; (b) FDMA; (c) CDMA.

multiple access can be implemented in a TDMA format, where at a given time slot a specific user
is allowed to employ all subchannels for his transmission, as illustrated in Figure 1.13. As depicted
in Figure 1.14, it is also possible to assign distinct frequency bands for different users provided the
users know in which subchannels (bands) they can transmit at a given time slot.
 1.5. MULTIPLE ACCESS METHODS 15
Frequency

·
User 1

··
User 2

User 3

···

Time

Figure 1.13: OFDMA system: TDMA case.

Frequency

·
User 1

··
User 2

User 3

···

Time

Figure 1.14: OFDMA system: FDMA case.

The most efficient way to assign the subchannels to multiple users is through random assign-
ment, which guarantees that all users enjoy approximately the same quality of service. Figure 1.15
illustrates an OFDMA scheme with random assignment of subchannels. As depicted in this fig-
ure, the subchannels used by each user may change at each time slot and, therefore, this kind of
OFDMA scheme avoids users to get stuck in low-quality subchannels. Note, however, that the
OFDMA schemes that are usually employed in standards are the ones depicted in Figures 1.13
and 1.14. This is justified by some issues and limitations concerning the OFDMA with random
assignment. For example, some multiple-antenna transmission schemes, especially those transmis-
sions with diversity, cannot be employed in OFDMA with random assignment, since they usually
require the transmission over adjacent subchannels.
16 1. THE BIG PICTURE
Frequency

·
User 1

··
User 2

User 3

···

Time

Figure 1.15: OFDMA system: random assignment.

In this section we have seen how different users can transmit over the same medium. But
what if the users want to be capable of both transmitting and receiving data? In this case, duplex
methods must be used.

1.6 DUPLEX METHODS

In cellular systems, duplex methods are used to separate the signal sent by the mobile to the base station,
called uplink connection, from the signal sent by the base station to the mobile station, known as
downlink connection. In fact, in any bidirectional communication system the duplex method has to
be specified in order to determine how these connections can coexist without interfering with each
other.5 There are mainly two duplex methods, the time-division duplex (TDD) and the frequency-
division duplex (FDD), which are briefly explained in the following.

1.6.1 TDD
As illustrated in Figure 1.16, TDD schemes assign different time intervals to uplink and downlink
connections. Both of these connections can use the whole channel bandwidth during their trans-
missions. It is common practice to separate the uplink and the downlink connections in TDD by
a time interval known as guard time, which avoids interference between these two connections that
might be caused by propagation effects such as multipath, which are addressed in the next section.

5That is, bidirectional communications systems also have uplink and downlink connections like cellular systems. The difference
can be on the elements that are at the ends of these connections, which can be computers and Internet service providers instead
of mobile stations and base stations.
 1.7. WIRELESS CHANNELS: FADING AND MODELING 17
1.6.2 FDD
Figure 1.17 illustrates the FDD method in which different frequency bands are assigned to the
uplink and downlink connections.

1.7 WIRELESS CHANNELS: FADING AND MODELING

The signals (waveforms) transmitted by wireless communications systems are ordinary electromag-
netic waves and, as such, they suffer from wave-propagation effects. A detailed description of wireless
channel effects over the transmitted signal as well as channel modeling can be found in [67, 80]. In
this section, we briefly explain some of the main problems introduced by wireless channels, empha-
sizing the multipath fading, in order to motivate the mathematical model used throughout this book.
These main impairments are: path loss, shadowing, and multipath fading. The problem of multiuser
interference (MUI) associated with multiple access schemes is not addressed here.

1.7.1 FADING
The main effects on transmitted signals inherent to wireless channels can be summarized in one
word: fading. Fading is a phenomenon concerning the time-variation of the channel strengths. If
these variations are due to transmissions over long distances (ranging from hundreds to thousand
meters), then they are known as large-scale fading, whereas the term small-scale fading is used for
channel variations due to relative movements of transmitter/receiver over short distances (of the
order of the carrier wavelength).

Frequency

ϕ2
Uplink Downlink
ϕ1

t1 t2 t3 t4 Time
  
Guard Time

Figure 1.16: Time-division duplex (TDD).

18 1. THE BIG PICTURE

Frequency

ϕ4
Uplink
ϕ3

ϕ2
Downlink
ϕ1

t1 t2 Time

Figure 1.17: Frequency-division duplex (FDD).

Some examples of wave-propagation effects that fall into the category of large-scale fading
are the path loss and the shadowing effects. Path loss, also known as path attenuation, is the power
reduction of an electromagnetic signal as it propagates through the medium, i.e., longer distances
between transmitter and receiver leads to lower power of the received signal, assuming that the power
of the transmitted signal does not vary. The path loss Ploss is defined as
Pt
Ploss  , (1.1)
Pr
where Pt and Pr stand for the power of the electromagnetic signal at the transmitter and receiver,
respectively. The path loss, measured in decibels, has a linear dependence on log10 (d), where d is
the distance that the signal travels. The other effect mentioned above, namely shadowing, is a fading
caused by the obstruction of the line-of-sight between transmitter and receiver. The shadowing
generates signal variations that are usually modeled by log-normal distribution.
There are two main types of small-scale fading: the frequency-selective and the flat fading. If
different frequency components of the electromagnetic signal are affected differently by the channel,
such effect corresponds to frequency-selective fading. Otherwise, in cases the variation induced on the
signal by the channel does not depend on/vary with the frequency of the electromagnetic wave, then
such effect falls into the category of flat fading.
The multipath fading is the most common type of small-scale fading that is present in mobile
communications. It originates from the transmitted-signal reflections in local buildings, hills, or
structures around a few hundred wavelengths from the mobile. Therefore, the multipath fading
can be understood as variations on the signal caused by interferences from attenuated and delayed
versions of the same signal (the reflections). Figure 1.18 illustrates an example in which two reflected
 1.7. WIRELESS CHANNELS: FADING AND MODELING 19
signals are interfering with the signal received from the line-of-sight. This figure depicts the three
waveforms independently, but it is clear that the received signal y(t) is a summation of the three
signals. Note also that in the time interval [0, Ts ], in which Ts stands for the duration of one symbol
generated by the digital modulator, the interference among the three signals will be constructive,
whereas some destructive interferences occur within other intervals, such as [Ts , 2Ts ].

Received Signals

... ...

0 Ts 2Ts time

Figure 1.18: Example of multipath fading.

Figure 1.19 depicts an example of power loss in wireless environment. For a given transmitted
power Pt , such figure illustrates how the received power Pr varies as a function of the transmitter-
receiver distance d. These power variations occur due to the path loss, shadowing, and multipath
fading phenomena.

1.7.2 MODELING
In rough environments, such as urban areas, frequency-selective fading occurs due to the reflections
of the transmitted signal that arrive with distinct delays at the receiver (multipath). In these cases,
the propagation medium (channel) is usually modeled as a linear system with memory, which can
be characterized by its impulse responses. Indeed, since wireless channels are time-varying, their
impulse responses h(t, τ ) can change along the time t. Thus, the signal arriving at the receiver at a
time instant t ∈ R can be written as
 t
y(t) = s(τ )h(t, τ )dτ, (1.2)
0
where h(t, τ ) corresponds to the channel response to an impulse applied at the instant t and s(τ )
represents the transmitted signal (concatenation of waveforms representing symbols). In addition,
20 1. THE BIG PICTURE
Pr
(dB) Mulipath - Fading
Pt
Path-loss + Shadowing
Path-loss

d (log)

Figure 1.19: Example of power loss in wireless environment.

it is common practice to consider that the memory of h(t, τ ) is finite since, due to path loss, not all
existing reflections will have enough power to be sensed at the receiver.
If one is interested only in the discrete-time representation of the signals, the signal arriving
at the receiver at instant k ∈ Z can be written as

k
y(k) = s(i)h(k, i), (1.3)
i=0

where h(k, i) is the channel response to an impulse applied at instant k, and s(i) represents the
transmitted signal as a sequence of symbols.
Note that, in practical communications systems, the sequence of symbols s(i) are usually
divided into blocks, and then each block is transmitted through the channel, as will be shown in the
next section.

1.8 BLOCK TRANSMISSION

Block transmission schemes are employed in some modern digital communications systems. Fig-
ure 1.20 illustrates the single-input single-output (SISO) model of a block transmission. In this
figure, each data block s(n), where n ∈ Z represents the block index, is comprised of eight symbols
coming from a digital modulator. Each data block s(n) is processed and prepared for transmission at
the SISO transmitter, generating the block u(n) that is propagated through the SISO channel. If the
SISO channel has memory, as usually happens in wireless channels, it will give rise to intersymbol
 1.8. BLOCK TRANSMISSION 21

s(n) SISO u(n) SISO y(n) SISO ŝ(n)

Transmitter Channel Receiver

     y(n − 1)
s(n) s(n − 1)   

  
y(n)

Figure 1.20: SISO model for block transmission.

interference (ISI) as well as interblock interference (IBI) at the receiver end. In this case, the currently
received block y(n) and the previously received block y(n − 1), corresponding to the transmitted
blocks u(n) and u(n − 1), respectively, will have an overlap as depicted in Figure 1.20. Then, the
SISO receiver is responsible for eliminating both ISI and IBI and detecting the transmitted symbols,
yielding a data block ŝ(n) that must be as close as possible to the originally transmitted block s(n).
In the case of wideband transmission, the ISI can be severe enough to make the SISO receiver
very complex to implement. Roughly speaking, each transmitted symbol would spread over the time
slot(s) of the neighboring symbol(s), turning their correct detection more challenging.
A naive, but widely used, solution to avoid interference among symbols of different blocks
(i.e., IBI) is to allow a guard period between each block transmission. In Figure 1.20, this solution
corresponds to separate each block u(n) by an amount of time that is sufficiently large to guarantee
that overlaps do not occur between the received blocks y(n), n ∈ Z. The drawback of such a solution
is the reduction of the data-transmission rate. This reduction can be significant in cases the blocks
u(n) are not much larger than the guard period.
A more general block-transmission framework is illustrated in Figure 1.21. This MIMO
model for block transmission encompasses many block-transmission schemes, including the SISO
model, multicarrier schemes, and multiple antenna configurations. In the MIMO model for block
transmission, a given data block s is modified by a MIMO Transmitter, yielding the block u whose
length is greater than the length of s. This larger length is due to many different reasons, such as the
replication of previously transmitted blocks, in case the transmitter has memory, or just by the use
of a guard period. The MIMO channel model can be described through a MIMO transfer function
possibly with memory, generating ISI and IBI. It is up to the MIMO receiver to process the received
signal block y in order to generate a reliable estimate of the transmitted signal block ŝ.
22 1. THE BIG PICTURE

s MIMO
u MIMO
y MIMO
ŝ

Transmitter Channel Receiver

Figure 1.21: MIMO model for block transmission.

The MIMO model might represent a wide range of signal-processing tasks. In this book the
main type of MIMO processing is the multicarrier transmission, which consists of transmitting each
symbol in a block through a narrowband subcarrier. The benefits from using this technique are the
following:
• Each subcarrier illuminates a narrow range of channel frequencies, so that the equivalent
subchannel appears to be flat. This turns the equalization for each subcarrier much simpler.
• Since it consists of a block transmission, the time support for transmission of a symbol modu-
lated by a subcarrier is roughly multiplied by the number of symbols in the block. That means
there is much more time to decode the information conveyed by each subcarrier, reducing or
even avoiding ISI within a block.
• If a guard period is inserted to avoid IBI, the time overhead is relatively low as long as each
block carries several symbols. The guard time is a function of the length of the channel impulse
response (time-delay spread).
As we have seen, if ISI and IBI are not tackled, they can deteriorate the performance of
communications systems. We have also seen that IBI can be avoided just by using a simple guard
period between the blocks. In the following section, we will introduce systems that tackle the ISI
problem in a simple and efficient manner.

1.9 MULTICARRIER SYSTEMS

In currently deployed communications systems, multicarrier transmissions seem to be the standard
choice. Multicarrier-modulation methods play a key role in modern data transmissions to deal with
channels with moderate to severe intersymbol interference.
 1.9. MULTICARRIER SYSTEMS 23
  jω 
H e 

  jω 
H e 
Zoom

Δω

Δω
ω

Figure 1.22: Multicarrier system dividing the channel bandwidth into non-overlapping flat subchannels.

The basic idea of multicarrier systems, whose most popular implementation is the OFDM, is
the transport of information through a wideband channel by energizing it with several narrowband
subcarriers simultaneously.The success of this technique relies on the partition of the physical channel
into non-overlapping narrowband subchannels through a transmultiplexer, as will be explained in
the next chapter. If the subchannels are narrow enough, the associated channel response in each
subchannel-frequency range appears to be flat, thus avoiding the use of sophisticated equalizers.
Figure 1.22 illustrates the effect of splitting a wideband channel in flatter subchannels.
Figure 1.23 depicts a transmultiplexer implementing a 4-band multicarrier system, i.e., the
channel is divided into 4 subchannels. At the transmitter end, a set of symbols (represented by colors)
is prepared for transmission through distinct subcarriers, where each subcarrier is represented by a
finite-impulse response (FIR) filter whose transfer function is denoted by Fi (z), i ∈ {0, 1, 2, 3}. At
the receiver side there are related FIR filters Gi (z). The FIR filters can be thought as narrowband
filters with distinct central frequencies, so that the symbols sent at different subcarriers travel through
different subchannels, thus not interfering with each other. In addition, Figure 1.23 assumes that
perfect transmission (reconstruction) is possible, i.e., the symbol transmitted at each subcarrier is
perfectly recovered at the receiver.
24 1. THE BIG PICTURE

F0 (z) G0 (z)

Noise
F1 (z) G1 (z)

Channel

F2 (z) G2 (z)

F3 (z) G3 (z)

Figure 1.23: Example of a 4-band multicarrier system.

In practical systems, perfect reconstruction is usually not achievable due to degradations caused
by physical-channel and noise effects, as well as power limitations. In this case, the subchannel
division allows, whenever possible, the exploitation of the signal-to-noise ratio (SNR) in the distinct
subbands by managing their data load in each subchannel. Indeed, if the transmitter has knowledge
about the SNR at the channel output for each subcarrier, then some loading scheme could be applied,
as illustrated in Figure 1.24. As can be observed in this figure, at the subcarriers with higher SNR it is
possible to transmit symbols belonging to higher-order modulations, such as an 8-PSK modulation,
whereas low SNR ones use lower-order modulation schemes, such as binary-PSK (BPSK). For very
low SNR subcarriers, it can be even decided not to transmit any symbol at all.
In a general setup some redundancy is required at the transmission in order to keep the
equalization as simple as possible. This is an important issue that will be addressed in the following
chapters. In addition, several methods for jointly optimizing the transmitter and receiver of FIR
MIMO systems can be employed to combat near-end crosstalk and additive-noise sources.

1.10 OFDM AS MIMO SYSTEM

In a noiseless environment, an OFDM system can be described using the MIMO framework de-
picted in Figure 1.21. For this case, the estimated signal vector ŝ can be described as a function of
 1.10. OFDM AS MIMO SYSTEM 25

F0 (z) G0 (z)

Noise
F1 (z) G1 (z)

Channel

F2 (z) G2 (z)

F3 (z) G3 (z)

Figure 1.24: Example of a loading scheme applied to a 4-band multicarrier system.

the input-signal vector s as

ŝ = GHF s, (1.4)

where F represents the precoder matrix applied at the transmitter, G represents the postcoder matrix
applied at the receiver, and H is the MIMO channel matrix. In this simplified description, all matrices
are considered memoryless so that each input-signal vector is processed independently. In addition,
allowing the existence of additive noise v at the channel output and assuming that F modifies and
inserts a prefix on s, the block transmission can be modeled as

F s + 
ŝ = GH  Gv , (1.5)
u v̄

as represented in Figure 1.25.

For channels with memory, in which IBI exists, the OFDM system adds some redundancy to
the input-signal vector in order to be able to eliminate the IBI at the receiver. As will be explained
later, if the redundancy consists of a cyclic prefix and the transmitter and receiver matrices are based
on discrete Fourier transform (DFT), the detection of the symbols at the receiver are decoupled
from each other meaning that ISI within each block is also eliminated.
26 1. THE BIG PICTURE

ŝ = GHFs

+ Gv

u v̄
Prefix

ŝ G H u G v

0
ŝ s v̄
0

Figure 1.25: Parameter decoupling in OFDM using a MIMO model.

1.11 MULTIPLE ANTENNA CONFIGURATIONS

Although the main topic of this book is multicarrier systems, in many current applications the
MIMO formulation allows the incorporation of multiple antennas at the transmitter and receiver,
on top of the usual precoder and postcoder blocks inherent to these systems, as shown in Figure 1.26.
These multiple-antenna building blocks introduce another degree of freedom, the space, that enables
an efficient use of the radio resources. For instance, this new degree of freedom can be exploited
to increase the system throughput by employing a spatial multiplexing scheme, or to enhance a
transmission (i.e., decrease bit-error rate) by using a transmission with diversity scheme.
In a general multiple-antenna setup we can consider the transmission of several blocks of data,
belonging to one or multiple users, where all the pre-processing at the transmitter is incorporated
in a single matrix building block Tx , and transmitted through an array of antennas. At the receiver
there is also an array of antennas whose output signals feed a single post processing building block
Rx that is responsible for separating and detecting each transmitted signal block, as illustrated in
Figure 1.27.
There are several ways to compose the input- and output-signal vectors as well as the channel
matrix in a digital communication setup. In any case, by properly stacking the transmitted and
received information, the representation given by Equation (1.5) is quite powerful and accommodates
several transceiver configurations.
The capacity gains of multiple-antenna systems with respect to the conventional single-
antenna systems depend on:

• number of antennas at the transmitter;

 1.12. MITIGATING INTERFERENCE AND NOISE 27

s1 ŝ1

···
···
Tx Rx

···
···

sN ŝM

Figure 1.26: General setup of MIMO precoding with multiple antennas.

Space-Time Prefilter RF De-interleaving

s Coding & Demodulation & ŝ
···

···

···

···
···

···

& RF Modulation & Space

Interleaving Post Filter Receiver

Figure 1.27: General setup of multiple block MIMO precoding with multiple antennas.

• number of antennas at the receiver;

• number of paths in the channel.

1.12 MITIGATING INTERFERENCE AND NOISE

Practical communications systems must be able to deal with interference and noise in an efficient
manner. As we have already seen, mobile communications suffer ISI and IBI due to multipath fading,
and they also suffer interference caused by other users sharing the same radio resources, which is
usually called multiuser interference (MUI) or co-channel interference.
In this context, the signal-to-interference-plus-noise ratio (SINR) plays a key role in assessing
the quality of a transmission. A high SINR, indicating a high-quality transmission, can be achieved
by mitigating interference and/or noise, promoting enhancements in the performance of the physical
layer.
28 1. THE BIG PICTURE
There are several strategies and techniques to increase SINR, such as: designing equalizers for
MIMO systems, employing transmission with diversity in multiple-antenna systems, optimizing
multicarrier systems, using subspace methods to mitigate noise, etc. In the next chapters we will
discuss how multicarrier systems can combat ISI, IBI, and noise.

1.13 CONCLUDING REMARKS

In this chapter, many aspects of digital communications and transmissions were briefly introduced.
In the following chapters, some of these aspects will be used or carefully revisited. Among the
material covered in this chapter, block transmissions and multicarrier systems, in which OFDM is
the most notorious case, are the central topics of this book. Indeed, the rest of this book is dedicated
to thoroughly explain the fundamental ideas of block transmissions and multicarrier systems, to
present the OFDM and SC-FD systems, and to introduce block transceivers that are capable of
increasing the system throughput by reducing the amount of redundancy necessary to remove IBI.
In addition, we tried to motivate and expose in an intuitive manner the importance and necessity of
such topics to current communications systems. Therefore, the approach followed here consists of
presenting the material in a pictorial way, leaving the mathematical details to the following chapters.
 29

CHAPTER 2

Transmultiplexers
2.1 INTRODUCTION
The proposal of new techniques for channel and source coding, along with the development of
integrated circuits and the use of digital signal processing (DSP) for communications have allowed
the deployment of several communications systems to meet the demands for transmissions with
high data-rates. Typical DSP tools, such as digital filtering, are key to retrieving, at the receiver end,
reliable estimates of signals associated with one or several users who share the same physical channel.
There are various classes of digital filters. Those employed in communications systems can
be either fixed or adaptive, linear or nonlinear, with finite impulse response (FIR) or with infinite
impulse response (IIR), just to mention a few. Among such classes of systems, fixed, linear, and
FIR filters are rather common in practice because of their simpler implementation, good stability
properties, and lower costs as compared to other alternatives.
Nonetheless, modern communications systems require more sophisticated techniques, thus
calling for more features than fixed, linear, and FIR filters can offer. In this context, multirate signal
processing adds some degrees of freedom to the standard linear time-invariant (LTI) signal process-
ing through the inclusion of decimators and interpolators. These degrees of freedom are crucial to
develop some interesting representations of communications systems based on filter banks, especially
multicarrier transceivers. A filter bank is a set of filters (usually LTI FIR filters) sharing the same
input-output pair and internally employing decimators and interpolators.
Filter-bank representations are widely used in source coding and spectral analysis. In com-
munications, the transmultiplexer (TMUX) configuration can be employed to represent multicarrier
or single-carrier transceivers, and can be considered a system dual to the filter-bank configuration
in the sense that the signal processing which takes place at the input of a filter bank actually appears
at the output of a TMUX, and vice versa. Indeed, several practical systems can be modeled using
TMUXes.
Unlike filter banks that usually require sharp frequency-selective subfilters, practical multicar-
rier transceivers can be modeled as TMUXes which use short-length subfilters with poor frequency
selectivity. In the majority of practical cases, these transceivers are implemented as memoryless block-
based transceivers. The most commonly used memoryless block-based transceivers are the orthogonal
frequency-division multiplexing (OFDM) and the single-carrier with frequency-domain equaliza-
tion (SC-FD) systems.
The main feature of OFDM-based transceivers is the elimination of intersymbol interfer-
ence (ISI) with low computational complexity, i.e., using just a small amount of numerical opera-
30 2. TRANSMULTIPLEXERS
tions to undo the harmful effects induced by frequency-selective channels. A competing alternative
to OFDM is the SC-FD transceiver, which presents lower peak-to-average power ratio (PAPR)
and lower sensitivity to carrier-frequency offset (CFO), as explained in [63, 87]. In addition, for
frequency-selective channels, the bit-error rate (BER) of SC-FD can be lower than for its OFDM
counterpart, particularly for the cases in which the channel has high attenuation at some subchannels.
In this chapter some key multirate signal-processing tools are revised (Section 2.2) aiming at
their use in the modeling of communications systems (Section 2.3). These tools will be particularly
utilized to represent OFDM and SC-FD systems, as well as to introduce some initial results related
to what is beyond OFDM-based systems, namely: the memoryless LTI block-based transceivers
using reduced redundancy (Section 2.4).

2.2 MULTIRATE SIGNAL PROCESSING

In many signal-processing applications, it is quite common that signals with distinct sampling
rates coexist. In general, multirate signal-processing systems include as building blocks both the
interpolator and the decimator. The interpolation consists of increasing the sampling rate of a given
signal, whereas the decimation entails a sampling-rate reduction of its input signal. The loss of data
inherent to decimation may give rise to aliasing in the decimated signal spectrum.
The interpolation by a factor N ∈ N consists of including N − 1 zeros between each pair of
adjacent samples, generating a signal whose sampling rate is N times larger than the sampling rate
of the original signal. Indeed, given a complex-valued signal s(n), in which the integer number n
denotes the time index at the original sampling rate, the interpolated signal sint (k) is given by

s(n), whenever k = nN
sint (k)  , (2.1)
0, otherwise

where the integer number k denotes the time index at the new sampling rate. In the frequency
domain, the effect of interpolation can be described as (see, for example, [17])

Sint (ejω ) = S(ejωN ), (2.2)

in which

X(ejω )  F {x(n)}

= x(n)e−jωn (2.3)
n∈Z

is the discrete-time Fourier transform of the sequence x(n), with ω ∈ R denoting the frequency
variable.1
1 It is assumed that the discrete-time Fourier transform of the sequence x(n) exists, i.e., the series in expression (2.3) is convergent
for all real-valued scalar ω. For instance, an absolutely summable (i.e., an 1 -signal) x(n) is sufficient to guarantee the convergence
of the series.
 2.2. MULTIRATE SIGNAL PROCESSING 31
The decimation by a factor N consists of discarding N − 1 samples from each non-overlapping
block containing N samples of the input signal. The resulting signal has a sampling rate N times
lower than the sampling rate of the original signal. Indeed, given the signal s(n), the decimated
signal sdec (k) is defined by

sdec (k)  s(kN), (2.4)

for all integer number k. In the frequency domain, it is possible to show that the decimated signal
is represented by (see, for example, [17])
1  j ω−2π n

Sdec (ejω ) = S e N , (2.5)

N
n∈N

where N  {0, 1, · · · , N − 1}. Unlike the interpolation, the decimation is a periodically time-
varying operation.

|S(ejω )| |Sint(ejω )|

ω ω
−2π −π 0 π 2π −2π −π 0 π 2π

N
s(n) sint(k)

n k
−6 −4 −2 0 2 4 6 −6 −4 −2 0 2 4 6

Figure 2.1: Interpolation (N = 2).

It is worth mentioning that a more appropriate nomenclature for the interpolation and dec-
imation processes just described should be upsampling and downsampling, reserving the nouns in-
terpolation and decimation for the cases in which a filtering process is also present. However, it is
rather common in the literature and in practice to use interchangeably the nomenclatures upsam-
pling/interpolation and downsampling/decimation. We will follow this practice, but the reader will
be able to identify easily when a filtering process takes place or not.
Figures 2.1 and 2.2 depict the respective effects of interpolation and decimation by a factor
N = 2 in both time and frequency domains. These signals are only for illustration purposes and
they do not represent true time-frequency pairs. By examining Figures 2.1 and 2.2 it is possible to
32 2. TRANSMULTIPLEXERS
|S(ejω )| |Sdec(ejω )|

ω ω
−2π −π 0 π 2π −2π −π 0 π 2π

N
s(n) sdec(k)

n k
−6 −4 −2 0 2 4 6 −6 −4 −2 0 2 4 6

Figure 2.2: Decimation (N = 2).

verify that, in order to avoid aliasing due to decimation and to eliminate the spectrum repetition
due to interpolation, a digital filtering operation is required before the decimation and after the
interpolation. The decimation filter narrows the spectrum of the input signal in order to avoid
that aliasing corrupts the spectrum of the resulting decimated signal. For a lowpass real signal,
for instance,
 we have
 to maintain the input signal information only at the low frequencies within
the range − N π π
, N , so that the spectrum at this range is not corrupted after decimation. The
interpolation filter smooths the interpolated signal sint (k), eliminating abrupt transitions between
non-zero and zero samples, which is the source of the spectrum repetitions (also known as spectral
images). The central frequencies of the spectrum repetitions are located at ± 2π N n, with n ∈ N .
Figure 2.3 illustrates how the decimation and interpolation operations are implemented in practice.

s(n) N f (k) sint(k) s(n) g(n) N sdec(k)

Figure 2.3: General interpolation and decimation operations in time domain.

Example 2.1 (Decimation & Interpolation) Let h(n) be a signal defined as


2n , whenever n ∈ {0, 1, · · · , 7}
h(n)  . (2.6)
0, otherwise

Determine Hint (z) = [H (z)]↑3 and Hdec (z) = [H (z)]↓3 , for all non-zero complex number z, in
which H (z)  Z {x(n)} is the Z -transform of the sequence x(n). In addition, the notations [(·)]↑N
and [(·)]↓N denote the interpolation and decimation by N applied to (·), respectively.
 2.2. MULTIRATE SIGNAL PROCESSING 33
Solution. We know that

H (z) = x(n)z−n
n∈Z
= 1 + 2z−1 + 4z−2 + 8z−3 + 16z−4 + 32z−5 + 64z−6 + 128z−7 , (2.7)

for all z  = 0. The interpolation by a factor of 3 is equivalent to insert 2 zero-valued samples between
adjacent samples of x(n). Hence, we have

Hint (z) =1 + 0.z−1 + 0.z−2 + 2z−3 + 0.z−4 + 0.z−5 + 4z−6 + 0.z−6 + 0.z−8 + 8z−9
+ 0.z−10 + 0.z−11 + 16z−12 + 0.z−13 + 0.z−14 + 32z−15 + 0.z−16 + 0.z−17
+ 64z−18 + 0.z−19 + 0.z−20 + 128z−21
=1 + 2z−3 + 4z−6 + 8z−9 + 16z−12 + 32z−15 + 64z−18 + 128z−21
=H (z3 ), (2.8)

for all z  = 0. The decimation by a factor of 3 will generate a discrete-time signal hdec (k) = h(3k) =
23k = 8k , if k ∈ {0, 1, 2}, or hdec (k) = 0, otherwise. Hence, we have

Hdec (z) = 1 + 8z−1 + 64z−2 , (2.9)

for all z  = 0. 2

In multirate systems, there are very useful manners to manipulate the interpolation and dec-
imation building blocks. We are particularly interested in ways to commute the decimation and
interpolation operations with linear time-invariant filters. Some forms of commuting are based on
the so-called noble identities.

S(z) F (z) N U (z) S(z) N F (z N ) U (z)

Y (z) N G(z) Ŝ(z) Y (z) G(z N ) N Ŝ(z)

Figure 2.4: Noble identities in Z-domain.

Figure 2.4 illustrates the building-block representations of the noble identities. In the inter-
polation process, instead of first upsampling the input signal and then filtering it, one can first filter
the input signal in a lower sampling rate and then upsample the resulting signal. This strategy allows
one to reduce the number of operations required by the overall interpolation process. As for the
decimation process, the decimator followed by a filter is equivalent to filter the input signal by the
34 2. TRANSMULTIPLEXERS
interpolated filter followed by the downsampling. These operations can be mathematically described
as

[F (z)S(z)]↑N  U (z) = F (zN ) [S(z)]↑N , (2.10)


G(z) [Y (z)]↓N  Ŝ(z) = G(zN )Y (z) . (2.11)
↓N

Analysis Bank Synthesis Bank

g0 (k) N N f0 (k)

g1 (k) N N f1 (k)

gM −1 (k) N N fM −1 (k)

Figure 2.5: Analysis and synthesis filter banks in time domain.

A widespread application of multirate systems is the filter-bank design. A filter bank consists
of a set of filters with the same input signal, or a set of filters whose outputs are added to form the
overall output signal, as depicted in Figure 2.5. The set of M ∈ N filters represented by the family
of impulse responses {gm (k)}m∈M , in which M  {0, 1, · · · , M − 1}, is the so-called analysis filter
bank, whereas the set of filters represented by the family of impulse responses {fm (k)}m∈M is the
synthesis filter bank. It is possible to verify that the analysis filter bank divides the input signal in
subbands, generating narrowband signals which can be further decimated.The subband signals can be
employed for analyses and manipulations according to the particular application. For reconstruction,
the subband signals are interpolated and combined by the synthesis filter bank.
Transmultiplexers, also known as filter-bank transceivers, are considered systems dual to the
filter-bank configurations since the roles of analysis and synthesis filter banks are interchanged in
transmultiplexers. Indeed, the inputs of a transmultiplexer are first combined by the synthesis bank
and, after some further processing stages, the outputs are obtained as a result from the analysis bank,
as shown in Figure 2.6.
It is worth mentioning that this section is based on [17, 81], which contain a thorough
treatment of this subject.
 synthesis filter bank
2.3. FILTER-BANK TRANSCEIVERS 35
v(k)

s0 (n) u(k) x(k) y(k) ŝ0 (n)

N f0 (k) h(k) g0 (k) N

s1 (n) ŝ1 (n)

N f1 (k) g1 (k) N

sM −1 (n) ŝM −1 (n)

N fM −1 (k) gM −1 (k) N

Figure 2.6: TMUX system in time domain. analysis filter bank

2.3 FILTER-BANK TRANSCEIVERS

Consider the transceiver model described in Figure 2.6, where a communication system is modeled
as a multiple-input multiple output (MIMO) system. The data samples of each sequence sm (n)
belong to a particular digital constellation C ⊂ C, such as PAM, QAM, or PSK.2 The sequence
sm (n) represents the mth transceiver input, where m ∈ M and n ∈ Z represents the time index.
The corresponding transceiver output is denoted as ŝm (n) ∈ C, which should be a reliable estimate
of sm (n − δ), where δ ∈ N represents the delay introduced by the overall transmission/reception
process.
A communication system can be properly designed by carefully choosing the set of causal
transmitter filters with impulse responses represented by {fm (k)}m∈M , and the set of causal receiver
filters represented by {gm (k)}m∈M . These filters operate at a sampling rate N times larger than the
why ?
sampling rate of the sequences sm (n). Note that the index n represents the sample index at the
input and output of the transceiver, whereas k ∈ Z is employed to represent the sample index of the
subfilters and of the internal signals between the interpolators and decimators. The transmitter and
receiver subfilters are time invariant in our discussions.
The input signals sm (n), for each m ∈ M, are processed by the transmitter and receiver sub-
filters aiming at reducing the channel distortion, so that the output signals ŝm (n) may represent good
estimates of the corresponding transmitted signals. The usual objective in a communication system
is to produce estimates of sm (n − δ) achieving low BER and/or maximizing the data throughput.
The channel model can be represented by an FIR filter of order L ∈ N whose impulse response
is h(k) ∈ C. The FIR transfer function accounts for the frequency-selective behavior of the physical

2 Pulse-amplitude modulation, quadrature-amplitude modulation, or phase-shift keying, respectively.

36 2. TRANSMULTIPLEXERS
channel. The additive noise v(k) ∈ C accounts for the thermal noise from the environment and
possibly for the multi-user interference (MUI).

2.3.1 TIME-DOMAIN REPRESENTATION

According to Figure 2.6 the channel input signal is given by


u(k)  sm (i)fm (k − iN). (2.12)
(i,m)∈Z×M

The relation between input and output of the channel is described as


y(k)  h(j )u(k − j ) + v(k). (2.13)
j ∈Z

At the receiver end, the signal y(k) is processed in order to generate estimates of the transmitted
data according to


ŝm (n)  gm (l)y(nN − l). (2.14)
l∈Z

By employing Equations (2.12), (2.13), and (2.14) it is possible to describe the relation between
the input signal sm (n) and its estimate ŝm (n), as given by

 
ŝm (n) = gm (l)h(j )sm (i)fm (nN − l − j − iN) + gm (l)v(nN − l). (2.15)
(i,j,l,m)∈Z3 ×M l∈Z

The description above is not the easiest one to analyze the system and draw conclusions. For
example, it is possible to employ a time-domain approach using matrix description, as described
in [70, 72]. Another approach is to apply polyphase decomposition in a Z -domain formulation as
described as follows.

2.3.2 POLYPHASE REPRESENTATION

As long as the interpolation and decimation factors are equal to N, it is convenient to describe
the transmitter and receiver filters by their polyphase decompositions of order N, according to the
 2.3. FILTER-BANK TRANSCEIVERS 37
expressions

Fm (z)  Z {fm (k)}


= fm (k)z−k
k∈Z 

= fm (j N)z−j N + fm (j N + 1)z−(j N +1) + · · · + fm (j N + N − 1)z−(j N +N−1)
j ∈Z
 
= fm (j N + i)z−(j N +i)
j ∈Z i∈N

= fm (j N + i)z−j N z−i
i∈N j ∈Z
 
−i
= z fm (j N + i)z−j N
i∈N j ∈Z
  
Fi,m (zN )

= z−i Fi,m (z ) , N
(2.16)
i∈N

and

Gm (z)  Z {gm (k)}


= gm (k)z−k
k∈Z 

= gm (j N)z−j N + gm (j N − 1)z−(j N −1) + · · · + gm (j N − N + 1)z−(j N −N+1)
j ∈Z
 
= gm (j N − i)z−(j N −i)
j ∈Z i∈N

= gm (j N − i)z−j N zi
i∈N j ∈Z
 
= i
z gm (j N − i)z−j N
i∈N j ∈Z
  
Gm,i (zN )

= zi Gm,i (zN ) , (2.17)
i∈N

where m ∈ M, and Fm (z) and Gm (z) are the Z -transforms of fm (k) and gm (k), respectively. The
transfer functions Fi,m (z) are the Type-I polyphase components of order N associated with Fm (z),
whereas the transfer functions Gm,i (z) are the Type-II polyphase components of order N associated
with Gm (z).
38 2. TRANSMULTIPLEXERS
Example 2.2 (Polyphase Decomposition) Let us consider the signal h(n) defined in Example 2.1.
Determine the Type-I polyphase decomposition of order 3 associated with the transfer function
H (z).

Solution. Consider that i ∈ {0, 1, 2}. Thus, we have


Hi (z) = h(3j + i)z−j , (2.18)
j ∈Z

yielding

H0 (z) = 1 + 8z−1 + 64z−2 , (2.19)

H1 (z) = 2 + 16z−1 + 128z−2 , (2.20)
H2 (z) = 4 + 32z−1 . (2.21)

Observe that

H (z) = H0 (z3 ) + z−1 H1 (z3 ) + z−2 H2 (z3 )

= 1 + 8z−3 + 64z−6 + z−1 2 + 16z−3 + 128z−6 + z−2 4 + 32z−3

= 1 + 2z−1 + 4z−2 + 8z−3 + 16z−4 + 32z−5 + 64z−6 + 128z−7 . (2.22)

Comparing
  with the solution of Example 2.1, the reader should also notice that Hdec (z) = H0 (z) =
H0 (z3 ) ↓3 . This is a useful property that will be further exploited. 2

By using a matrix approach, we can rewrite Equations (2.16) and (2.17) as follows:
⎡ ⎤
F0,0 (zN ) ··· F0,M−1 (zN )
   ⎢ .. .. ⎥
F0 (z) · · · FM−1 (z) = 1 z−1 · · · z−(N −1) ⎣ .
..
. . ⎦
  
dT (z) FN −1,0 (z ) · · · FN −1,M−1 (z )
N N
  
F(zN )
= d (z)F(z ),
T N
(2.23)
⎡ ⎤ ⎡ ⎤⎡ ⎤
G0 (z) G0,0 (zN ) ··· G0,N−1 (zN ) 1
⎢ .. ⎥ ⎢ .. .. .. ⎥ ⎢ .. ⎥
⎣ . ⎦=⎣ . . . ⎦⎣ . ⎦
GM−1 (z) GM−1,0 (z ) · · · GM−1,N−1 (z )
N N z(N −1)
     
G(zN ) d(z−1 )
−1
= G(z )d(z
N
). (2.24)
 2.3. FILTER-BANK TRANSCEIVERS 39
V (z)

S0 (z) U (z) X(z) Y (z) Ŝ0 (z)

N H(z) N

z −1 z
S1 (z) Ŝ1 (z)
N N

F(z N ) z −1 z G(z N )

z −1 z
SM −1 (z) ŜM −1 (z)
N N

Figure 2.7: Polyphase representation of TMUX systems.

Now, by defining Sm (z)  Z {sm (n)}, U (z)  Z {u(k)}, X(z)  Z {x(k)}, V (z)  Z {v(k)},
Y (z)  Z {y(k)}, and Ŝm (z)  Z {ŝm (n)}, then one can write

⎡ ⎤
S0 (zN )
⎢ .. ⎥
U (z) = dT (z)F(zN ) ⎣ . ⎦, (2.25)
SM−1 (zN )
  
s(z)
X(z) = H (z)U (z), (2.26)
Y (z) = X(z) + V (z), (2.27)
⎡ ⎤
Ŝ0 (z) 
⎢ .. ⎥ −1
⎦ = G(z )d(z )Y (z)
N
⎣ . . (2.28)
↓N
ŜM−1 (z)
  
ŝ(z)

Figure 2.7 illustrates the transceiver model utilizing the polyphase decompositions of the
transmitter and receiver subfilters. By employing the noble identities described in Section 2.2, it is
possible to transform the transceiver of Figure 2.7 into the equivalent transceiver of Figure 2.8.
The highlighted area of Figure 2.8 that includes delays, forward delays, decimators, interpola-
tors, and the SISO channel model can be represented by a pseudo-circulant matrix H(z) of dimension
40 2. TRANSMULTIPLEXERS
V (z)

S0 (z) U0 U (z) X(z) Y (z) Y0 Ŝ0 (z)

N H(z) N

z −1 z
S1 (z) U1 Y1 Ŝ1 (z)
N N

F(z) z −1 z G(z)

z −1 z
SM −1 (z) UN −1 YN −1 ŜM −1 (z)
N Pseudo-Circulant Channel N
H(z)

Figure 2.8: Equivalent representation of TMUX systems employing polyphase decompositions.

N × N, given by
⎡ ⎤
z−1 HN −1 (z) z−1 HN −2 (z) · · · z−1 H1 (z)
H0 (z)
⎢ H1 (z)
H0 (z) z−1 HN −1 (z) · · · z−1 H2 (z)⎥
⎢ ⎥
H(z)  ⎢ .. .. .. .. .. ⎥, (2.29)
⎣ . . . . . ⎦
HN −1 (z) HN −2 (z) HN −3 (z) ··· H0 (z)

in which
 
H (z)  Hi (zN )z−i and Hi (z)  h(j N + i)z−j . (2.30)
i∈N j ∈Z
0≤j N +i≤L

Indeed, given the indexes m and l within the set N , the (m, l)th element of the matrix H(z),
denoted as [H(z)]ml , represents the transfer function from the lth input element of the highlighted
area shown in Figure 2.8 to the mth output element of this area. Hence, by assuming that v(k) = 0
for all integer number k, if Ul (z) is the lth input at the transmitter end of the highlighted area in
Figure 2.8 and Ym (z) is the mth output of this area at the receiver end, then

Ym (z) = z−l H (z)Ul (zN )zm
 ↓N
= zm−l H (z)Ul (zN )
 ↓N
= Ul (z) z H (z)
m−l
, (2.31)
↓N
 2.3. FILTER-BANK TRANSCEIVERS 41
in which we have applied the noble identity described in Equation (2.11) and we also have considered
that the only non-zero input of the highlighted area in Figure 2.8 is Ul (z).
Therefore, based on Equation (2.31) and on the first type of polyphase representation of the
channel-transfer function, we can write
Ym (z)
[H(z)]ml =
Ul (z) 
= zm−l H (z)
 ↓N 

= zm−l Hi (zN )z−i
i∈N ↓N
 

= Hi (zN )zm−l−i
i∈N ↓N 
= H0 (z )z
N m−l
+ H1 (zN )zm−l−1 + · · · + HN −1 (zN )zm−l−N +1 . (2.32)
↓N

We know that the decimation operation retains the first coefficient out of N coefficients within
a block, starting from the 0th element. In the Z -domain, this means that the decimation operation
keeps only the coefficients which multiply a power of zN . Thus, the j th coefficient of the decimated
signal corresponds to the (j N )th coefficient of the signal before the decimation. Another way of
interpreting this fact is that, given an index i0 ∈ N such that (m − l − i0 ) is a multiple of N, the
decimation operation which appears in expression (2.32) retains the i0 th term Hi0 (zN )zm−l−i0 and
decimates it, as illustrated in Example 2.2.
We also know that −(N − 1) ≤ m − l ≤ N − 1, since m and l are within the set N . Hence,
if m − l ≥ 0, then expression (2.32) yields

[H(z)]ml = H(m−l) (zN ) = H(m−l) (z). (2.33)
↓N

On the other hand, if m − l < 0, then


[H(z)]ml = z−N H(N −l+m) (zN ) = z−1 H(N −l+m) (z), (2.34)
↓N

confirming the relations described in Equations (2.29) and (2.30).

Figure 2.9 describes the transceiver through the polyphase decomposition of appropriate
matrices, including the pseudo-circulant representation of the channel matrix. It is worth noting
that the descriptions of Figures 2.6 and 2.9 are equivalent.
As Figure 2.9 illustrates, the transmitted and received vectors are denoted as

s(n)  [ s0 (n) s1 (n) · · · sM−1 (n) ]T , (2.35)

ŝ(n)  [ ŝ0 (n) ŝ1 (n) · · · ŝM−1 (n) ]T . (2.36)
42 2. TRANSMULTIPLEXERS

v(n)

s(n) u(n) y(n) ŝ(n)

F(z) H(z) G(z)

Figure 2.9: Block-based transceivers in Z-domain employing polyphase decompositions.

From Figure 2.9, it is also possible to infer that the transfer matrix T(z) of the transceiver can
be expressed as

T(z)  G(z)H(z)F(z), (2.37)

where we considered the particular case in which v(k) = 0 for all integer number k, inspired by the
zero-forcing (ZF) design. A transceiver is zero forcing whenever T(z) = αz−d IM , for some α ∈ C
and d ∈ N. Notice that, if there is no noise, a zero-forcing solution is able to retrieve a scaled and
delayed version of all transmitted signals.
An important observation about Figure 2.9 is that, in order to be able to recover a block with
M transmitted symbols, one must send through the channel at least M elements in a data block,
i.e., we must necessarily have N ≥ M (this fact explains the shapes of the boxes in Figure 2.9).
Nonetheless, if N = M (no redundancy is included), then the matrices F(z), H(z), and G(z) are
square matrices and, therefore, a zero-forcing solution would not be achieved using only FIR filters,
considering that the channel model is not a simple delay, as explained in [44, 45]. Hence, some
redundancy must be introduced in order to work with FIR transceivers.3
Now, let us assume that we choose N > L, i.e., the interpolation/decimation factor is greater
than or equal to the channel order L, a common situation in practice.4 Based on Equation (2.30),
we have that the only integer number j which satisfies the inequality constraint 0 ≤ j N + i ≤ L
is j = 0, which lead us to conclude that Hi (z) = h(i), for N
i ≤ L. On the other hand, if there
exists i > L within the set N , then Hi (z) = 0 since there is no term to be added in order to form
Hi (z). In other words, we can say that, for N > L, each element of the matrix H(z) will consist of
filters with a single (possibly null) coefficient. In this case, the pseudo-circulant channel matrix in

3 Employing IIR filters may bring about many drawbacks, such as instability issues. This is the reason why FIR transceivers are the
prevalent choice.
4 Usually, practical block transceivers use N = M + K, where K is an integer number larger than or equal to L.
 2.3. FILTER-BANK TRANSCEIVERS 43
Equation (2.29) is represented by a first-order FIR matrix described as
⎡ ⎤ ⎡ ⎤
h(0) 0 0 ··· 0 0 · · · 0 h(L) · · · h(1)
⎢ h(1) h(0) 0 ··· 0 ⎥ ⎢ .. .. ⎥
⎢ ⎥ ⎢0 0 · · · 0 . . ⎥
⎢ .. .. .. .. .. ⎥ ⎢ ⎥
⎢ . . . . ⎥
. ⎥ ⎢ . . . . .. h(L)⎥
.
⎢ ⎢ .. .. .. .. ⎥
⎢ ⎥ −1 ⎢ ⎥
H(z) = ⎢h(L) h(L − 1) . . . · · · 0 ⎥ +z ⎢ 0 0 0 · · · 0 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0 ··· ··· 0 ⎥ ⎥ ⎢ ··· 0 ⎥
⎢ h(L) ⎢0 0 0 0 ⎥
⎢ . .. .. .. .. ⎥ ⎢. .. .. .. .. .. ⎥
⎣ .. . . . . ⎦ ⎣ .
. . . . . . ⎦
0 0 h(L) · · · h(0) 0 0 0 0 ··· 0
     
HISI HIBI
−1
= HISI + z HIBI . (2.38)

Notice that Equation (2.38) implies the following relation in the time domain:

x(k) = HISI u(k) + HIBI u(k − 1)

 
  u(k − 1)
= HIBI HISI , (2.39)
u(k)

where

u(k)  [ u(kN − N + 1) u(kN − N + 2) · · · u(kN ) ]T , (2.40)

x(k)  [ x(kN − N + 1) x(kN − N + 2) · · · x(kN) ]T . (2.41)

The relationship described in Equation (2.39) makes clear the roles of the matrices HISI and
HIBI . Indeed, matrix HISI mixes the symbols transmitted in the current data block, i.e., such a matrix
introduces interferences among the current data-block symbols, while matrix HIBI mixes some of
the symbols transmitted in the past block. The channel output vector x(k) is the result of adding
both effects: ISI and IBI.
Another way to derive Equation (2.39) is by analyzing what happens in the time domain when
a signal u(k) passes through an FIR channel h(k) of order L. In this case, we know that the channel
output x(k) is the linear convolution between the signals u(k) and h(k), that is, x(k) = (h ∗ u)(k).
Hence, if we look at a block of size N > L containing the channel output signals (in other words,
if we examine the elements of the vector x(k)), then we can verify that the first L elements of this
block are affected by the last L elements of the previous block, due to the channel memory and the
way the linear convolution is computed.
It is worth pointing out that generalizations of standard multicarrier communications systems
may call for sophisticated transmultiplexer designs in which the transmitted signal is filtered by a
precoder with memory consisting of a MIMO FIR filter.The inherent memory at the transmitter can
be viewed as a kind of redundancy since a given signal block is transmitted more than once along with
neighboring blocks. Sophisticated transmitters may require more complex receivers, but they might
44 2. TRANSMULTIPLEXERS
allow a reduction in the amount of redundant signals necessary to attain zero-forcing solution, for
example. All of these facts indicate that communication engineers should master the TMUX-related
tools in order to pursue new advances in communications systems, especially regarding multicarrier
transceivers. The case of transceivers with memory will be addressed in Chapter 5. In this chapter,
we shall consider the widespread used memoryless systems.

2.4 MEMORYLESS BLOCK-BASED SYSTEMS

The particular and very important case where the transceivers are memoryless, that is, F(z) =
F and G(z) = G, is addressed in this section. This case encompasses the memoryless block-based
transceivers, since these systems do not use data from previous or future blocks in the transmission
and reception processing of the current data block. That is, only the current block takes part in the
transceiver computations.The traditional OFDM and SC-FD transceivers are well-known examples
of memoryless block-based systems.
The non-overlapping behavior associated with memoryless transceivers is only possible if the
lengths of the FIR causal subfilters {fm (k)}m∈M and {gm (k)}m∈M are less than or equal to N .
Indeed, from Equations (2.16) and (2.17), we know that


[F(z)]im = fm (j N + i)z−j , (2.42)
j ∈Z

[G(z)]mi = gm (j N − i)z−j , (2.43)
j ∈Z

for all pairs of numbers (i, m) within the set N × M. Hence, the matrix F(z) will have memory
(i.e., will depend on z) if, and only if, there exists both a non-zero natural5 number j0 and a pair of
numbers (i0 , m0 ) ∈ N × M, such that fm0 (j0 N + i0 )  = 0, which occurs if, and only if, fm0 (k) is
a causal impulse response with length larger than N, since j0 N + i0 ≥ N. The same conclusion can
be drawn for the matrix G(z).
We shall briefly describe now the main memoryless LTI block transceivers which will be
considered throughout this book. Further details will be given in Chapters 3 and 4.

5The index j cannot be negative because we are only interested in causal subfilters.
0
 2.4. MEMORYLESS BLOCK-BASED SYSTEMS 45
2.4.1 CP-OFDM
The OFDM transceiver employing cyclic prefix as redundancy (also known as cyclic-prefix OFDM,
or just CP-OFDM) is described by the following transmitter and receiver matrices, respectively:
 
0K×(M−K) IK
F H
WM , (2.44)
IM
  
ACP ∈CN×M
 
G  EWM 0M×K IM , (2.45)
  
RCP ∈CM×N

where the integer number K denotes the amount of redundant elements, WM is the unitary M × M
discrete Fourier transform (DFT) matrix, that is,

W ml
[WM ]ml  √M , (2.46)
M

with (m, l) ∈ M2  M × M and

2π
WM  e−j M . (2.47)

In addition, IM is the M × M identity matrix, 0M×N is an M × N matrix whose entries are zero,
and E is an M × M diagonal equalizer matrix placed after the removal of the cyclic prefix and the
application of the DFT matrix.
As can be noted, the data block to be transmitted has length M, however, due to the prefix,
the transceiver actually transmits a block of length N = M + K, in which K must be larger than or
equal to the channel order L, i.e., one must necessarily have M ≥ K ≥ L so that the CP-OFDM
system works properly. The first K elements are repetitions of the last K elements of the inverse
discrete Fourier transform (IDFT) output in order to implement the cyclic prefix.
Matrix ACP adds and matrix RCP removes the related cyclic prefix. Note that, based on
Equation (2.38), the product RCP H(z)ACP  Hc ∈ CM×M is given by
⎡ ⎤
h(0) 0 ··· 0 h(L) · · · h(1)
⎢ .. .. ⎥
⎢ h(1) h(0) ··· 0 0 . . ⎥
⎢ ⎥
⎢ .. .. . ⎥
⎢ . . . . h(L) ⎥
⎢ ⎥
⎢ .. .. ⎥
Hc = ⎢ h(L) h(L − 1) . . 0 ⎥ , (2.48)
⎢ ⎥
⎢ .. .. ⎥
⎢ 0 h(L) . . ⎥
⎢ ⎥
⎢ .. .. .. .. ⎥
⎣ . . . . 0 ⎦
0 ··· 0 h(L) ··· h(0)
46 2. TRANSMULTIPLEXERS
where we can observe that RCP removes the IBI (there is no dependency on z anymore), whereas
matrix ACP right-multiplies the resulting memoryless matrix RCP H(z) ∈ CM×N so that the overall
matrix product is a circulant matrix of dimension M × M. Indeed, one can observe that each row of
matrix Hc can be obtained by circular-shifting the related previous row.
After inclusion and removal of the cyclic prefix, the resulting circulant matrix can be diago-
nalized by its right-multiplication by the IDFT and left-multiplication by the DFT matrices, where
these matrices are placed at the transmitter and receiver sides, respectively. Indeed, we have
  H
H
WM Hc WM = WM h0 h1 · · · hM−1 WM
  H
= WM h0 WM h1 · · · WM hM−1 WM

=  h0  h1 · · ·  H
hM−1 WM , (2.49)

where hm is the mth column of matrix Hc and  hm  WM hm is its DFT. Note that one can in-
terpret the elements of vector hm as a periodic discrete-time signal hm (k) whose period is M,
which respects the relation hm (k) = h0 (k − m), where h0 (0) = h(0), h0 (1) = h(1), · · · , h0 (L) =
h(L), h0 (L + 1) = 0, · · · , h0 (M − 1) = 0, h0 (M) = h(0), and so forth.Thus, by remembering the
circular-shifting property of the DFT stating that, given

H0 (l)  DFT{h0 (k)}

1 
=√ lk
h0 (k)WM , (2.50)
M k∈M

for all l ∈ M, then one has

DFT{h0 (k − m)} = WM
ml
H0 (l). (2.51)

Hence, by applying this result, we have

Hm (l) = DFT{hm (k)}

= DFT{h0 (k − m)}
= WM
ml
H0 (l), (2.52)

yielding


hm = diag{WMml
}l∈M h0

m
= diag{WMl
}l∈M  h0
m
= D h0 , (2.53)
 2.4. MEMORYLESS BLOCK-BASED SYSTEMS 47
in which D  l }
diag{WM denotes an M × M
l∈M
l ,
diagonal matrix whose the (l, l)th element is WM
for each l ∈ M. We can therefore rewrite Equation (2.49) as
  
H
WM Hc WM = h0 D
h0 D2
h0 · · · DM−1 H
h0 WM
lm lm ⎡ ⎤
−m.0
WM
⎢
⎢ WM
⎥
−m.1
⎥ √
= l0
H0 (l)WM l1
H0 (l)WM ···
l(M−1)
H0 (l)WM ⎢ ⎥/ M
⎢ ..⎥
⎣ .⎦
−m(M−1)
WM
H0 (l)  li −im
= √ WM WM
M i∈M
H0 (l)  i(l−m)
= √ WM . (2.54)
M i∈M

i(l−m)
If l = m, then WM = 1 for all i ∈ M, implying
 H0 (l) √
H
WM Hc WM = √ × M = MH0 (l), (2.55)
ll M
(l−m)
while, if l  = m, then WM  = 1, implying
 
 H0 (l)
M(l−m)
WM −1
H
WM Hc WM = √ (l−m)
lm M
 WM − 1
H0 (l) 1−1
= √ (l−m)
M WM −1
= 0. (2.56)

Therefore, we can conclude that

 WM Hc WM H

= WM RCP H(z)ACP WM H
√ 
= diag Mh0
√ 
= diag MWM h0
  
√ h
= diag MWM
0(M−L−1)×1
 diag{λm }m∈M , (2.57)

in which h  [ h(0) h(1) · · · h(L) ]T .

48 2. TRANSMULTIPLEXERS
Matrix includes at its diagonal the distortion imposed by the channel on each symbol of
the data block. Hence, the model of a CP-OFDM transceiver is described by

s + Ev ,
ŝ  E (2.58)

with v  WM RCP v and, for the sake of simplicity, the time dependency of the expressions was
omitted. As can be observed, the estimates of the transmitted symbols are uncoupled, that is, each
symbol can be estimated independently of any other symbol within the related block, avoiding
intersymbol interference. One can interpret this fact as if each symbol were transmitted through a
flat-fading subchannel.
From a signal processing perspective, the model described in Equation (2.58) has a simple
interpretation. Indeed, the addition and removal of the cyclic prefix turns the linear convolution
described in Equation (2.39) into a circular convolution. In this case, the CP-OFDM system loads
each subcarrier in the frequency domain with a constellation symbol and, after that, performs the
inverse discrete Fourier transformation, generating a vector in the time domain. The elements of
this vector can be thought as a periodic signal which is processed by the channel through a linear
convolution. After that, the signal is brought back to the original frequency domain. A basic fact
of digital signal processing is that the circular convolution of two signals can be implemented in
the frequency domain by performing the product of the DFTs of the related signals. Therefore,
the CP-OFDM system can be further simplified if we take this fact into account. All we have to
do is to perform the entire processing in the frequency domain. The symbols which are loaded at
each subcarrier can be directly mapped to the received signals at each subcarrier by performing the
product with the frequency response of the channel (DFT of the zero-padded impulse response, as
in Equation (2.57)).
The equalizer E for this transceiver can be defined in several ways, where the most popular are
the zero-forcing (ZF) and the minimum mean square error (MMSE) equalizers. In the ZF solution,
it is aimed to undo the distortions introduced by the channel. Indeed, when there is no noise, the
ZF solution is able to perfectly recover the transmitted vector. It is assumed that matrix can be
inverted, thus yielding

EZF  −1 . (2.59)

As for the MMSE solution, there is no requirement that matrix be invertible since this
latter operation is not needed. In fact, the linear MMSE equalizer matrix is the solution to the
following optimization problem:


EMMSE  arg min J (E) , (2.60)
∀E∈CM×M
 2.4. MEMORYLESS BLOCK-BASED SYSTEMS 49
where J is a real-valued function of a complex-valued matrix argument defined as

J (E)  E s − E( s + v )22
 H  
= E s − E s − Ev s − Ev
s − E
   H 
= tr E s − E s − Ev s − E s − Ev
 
= tr σs2 IM + σs2 E H EH − σs2 E − σs2 H EH + σv2 EEH , (2.61)

where E[·] and tr{·} are the expected value and trace operators, respectively. The derivation above
assumes that the transmitted symbols and environment noise within a block are independent and
identically distributed (i.i.d.), originating from a wide-sense stationary (WSS) white random se-
quences with zero means and uncorrelated.These assumptions imply that E[sv H ] = E[s]E[v ]H =
0M×M = E[v ]E[s]H = E[v sH ] and that E[ssH ] = σs2 IM and E[v v H ] = σv2 IM , where the pos-
itive real numbers σs2 and σv2 are the variances of the related WSS random sequences.6
Now, by using the following derivatives of scalar functions of complex matrices [83]:
! "
∂tr ZAZH
= ZA, (2.62)
! ∗ H"
∂Z
∂tr AZ
= A, (2.63)
∂Z∗
then we have
∂ J (E)
= σs2 E
H − σs2 H + σv2 E. (2.64)
∂E∗
∂ J (EMMSE )
We know that the optimal solution EMMSE is such that ∂E∗ = 0M×M , which implies
that7
# $−1
σv2
EMMSE = H
H
+ 2 IM
⎧ σs ⎫
⎨ λ∗m ⎬
= diag . (2.65)
⎩ |λ |2 + σv2 ⎭
m σs2 m∈M

It is worth highlighting that the CP-OFDM transceiver is the most popular type of OFDM-
based techniques which are employed in practical applications.

2.4.2 ZP-OFDM
An alternative OFDM system inserts zeros as redundancy and is called zero-padding OFDM (ZP-
OFDM). There are many variants of ZP-OFDM. One possible choice is the ZP-OFDM-OLA
6 In this book, we shall not employ distinct notations for deterministic and random variables.
7 We encourage the reader to justify why this is actually the minimum solution of the objective function.
50 2. TRANSMULTIPLEXERS
(overlap-and-add) whose transmitter and receiver matrices are implemented as
 
IM
F WM H
, (2.66)
0K×M
  
AZP ∈CN×M
 
IK
G  EWM IM , (2.67)
0(M−K)×K
  
RZP ∈CM×N

where, as in the CP-OFDM case, K ≥ L elements are inserted as redundancy, and N = M + K.

The name OLA stems from the way the received signals are processed by the matrix RZP .
Matrices AZP and RZP perform the addition and removal of the guard period of zero redundancy,
respectively. The matrix product RZP H(z)AZP ∈ CM×M is given by
⎡ ⎤
h(0) 0 ··· 0 h(L) · · · h(1)
⎢ .. .. ⎥
⎢ h(1) h(0) ··· 0 0 . . ⎥
⎢ ⎥
⎢ .. .. .. ⎥
⎢ . . . h(L) ⎥
⎢ ⎥
⎢ .. .. ⎥
RZP H(z)AZP = ⎢ h(L) h(L − 1) . . 0 ⎥ = RCP H(z)ACP .
⎢ ⎥
⎢ . . ⎥
⎢ 0 h(L) . . . ⎥
.
⎢ ⎥
⎢ .. .. .. .. ⎥
⎣ . . . . 0 ⎦
0 ··· 0 h(L) · · · h(0)
(2.68)

As can be observed, matrix AZP removes the interblock interference, whereas matrix RZP left-
multiplies the resulting memoryless Toeplitz matrix 8 H(z)AZP ∈ CN ×M so that the overall product
becomes a circulant matrix of dimension M × M. The reader should note that RZP H(z)AZP =
RCP H(z)ACP = Hc .
The ZP-OFDM-OLA transceiver discussed here is a simplified version of a more general
transceiver proposed in [55].9 In fact, the general transceiver allows the recovery of the transmitted
symbols using zero-forcing equalizers independently of the locations of the channel zeros, unlike
the ZP-OFDM-OLA or CP-OFDM that might have zero eigenvalues under certain channel con-
ditions. Unfortunately, from the computational point of view, this transceiver implementation is not
as simple as, for instance, the CP-OFDM, since the equivalent channel matrix is not circulant, turn-
ing impossible its diagonalization through fast transforms, such as fast Fourier transform (FFT).10
Furthermore, even for the design of a simple ZF equalizer, the general ZP-OFDM transceiver
8 See Subsection 4.3.2 for a formal definition of a Toeplitz matrix.
9There are other variants of ZP-OFDM, such as the ZP-OFDM-FAST.
10 Actually, it is possible to implement the general ZP-OFDM system using FFTs, but without diagonalizing the equivalent channel
matrix (see Chapter 4).
 2.4. MEMORYLESS BLOCK-BASED SYSTEMS 51
would require the inverse of a Toeplitz matrix, being therefore more complex than the inversion of
a circulant matrix required by a ZP-OFDM-OLA system.

2.4.3 CP-SC-FD
The cyclic-prefix single-carrier with frequency-domain (CP-SC-FD) equalization transceiver em-
ploys cyclic prefix as redundancy and it is closely related to the CP-OFDM transceiver. The CP-
SC-FD system is described by the following transmitter and receiver matrices:
 
0K×(M−K) IK
F , (2.69)
IM
 
G  WMH
EWM 0M×K IM . (2.70)

2.4.4 ZP-SC-FD
The zero-padding single-carrier with frequency-domain (ZP-SC-FD) equalization transceiver in-
serts redundant zeros to the block to be transmitted, as in the ZP-OFDM transceiver. The ZP-SC-
FD-OLA version may be modeled through the following transmitter and receiver matrices:
 
IM
F , (2.71)
0K×M
 
IK
G  WM
H
EWM IM . (2.72)
0(M−K)×K

2.4.5 ZP-ZJ TRANSCEIVERS

Lin and Phoong [39, 40, 44] showed that the amount of redundancy (guard samples) K  N −
M ∈ N required to eliminate IBI and ISI in memoryless block-based transceivers must satisfy the
inequality 2K ≥ L. They proposed a family of memoryless discrete multitone transceivers with
reduced redundancy. A particular transceiver of interest to our studies here is the zero-padding zero-
jamming (ZP-ZJ) system, which is characterized by the following transmitter and receiver matrices:
 
F
F , (2.73)
0K×M N ×M
 
G  0M×(L−K) G M×N
, (2.74)

in which F ∈ CM×M and G ∈ CM×(M+2K−L) .

The transfer matrix related to this transceiver is given by

T(z) = GH(z)F = G H F = T, (2.75)

52 2. TRANSMULTIPLEXERS
where, after removing the redundancy, the effective channel matrix is defined as
⎡ ⎤
h(L − K) · · · h(0) 0 0 ··· 0
⎢ .. .. .. ⎥
⎢ . . . ⎥
⎢ ⎥
⎢ . ⎥
⎢ h(K) .. 0 ⎥
⎢ ⎥
⎢ .. . . ⎥
⎢ . . . . . h(0) ⎥
H⎢ ⎢ ⎥ ∈ C(M+2K−L)×M . (2.76)
.. ⎥
⎢ h(L) . ⎥
⎢ ⎥
⎢ .. ⎥
⎢ 0 . h(L − K)⎥
⎢ ⎥
⎢ .. .. ⎥
⎣ . . ⎦
0 ··· 0 0 h(L) · · · h(K)

Considering v(k) = 0, for all k ∈ Z, we have

ŝ(n) = G H Fs(n) = Ts(n). (2.77)

Observe that the requirement of having 2K − L ≥ 0 makes sense when we analyze the
above expression. Indeed, in order to recover the M transmitted symbols, the memoryless transfer
matrix T of dimension M × M must be full-rank. This means that min{M, M + 2K − L} ≥ M,
i.e., 2K − L ≥ 0 ⇔ K ≥  L2 .
For this transceiver there are some constraints to be imposed upon the channel impulse
response model so that a zero-forcing solution exists. These constraints are related to the con-
cept of congruous zeros.11 The congruous zeros of a transfer function H (z) are the distinct zeros
z0 , z1 , · · · , zμ−1 ∈ C, with μ ∈ N, which meet the following condition: ziN = zjN , with H (zi ) =
H (zj ) = 0, for all i, j ∈ {0, 1, · · · , μ − 1}. Note that μ is a function of N. As shown in [44], the
channel model must satisfy the constraint μ(N ) ≤ K, where μ(N ) denotes the cardinality (number
of elements) of the largest set of congruous zeros with respect to N .
Therefore, assuming the existence of minimum-redundancy solutions for a given channel, i.e.,
considering that μ(N) ≤ L/2 ∈ N, then the ZF solution is such that its associated receiver matrix
is given by
−1 −1
G  (H F)−1 = F H , (2.78)

where H ∈ CM×M is given and F is predefined.

This solution is computationally intensive since it requires the inversions of M × M matrices,
entailing O(M 3 ) arithmetic operations. The conventional OFDM and SC-FD transceivers need
O(M log M) operations for the design of ZF and MMSE equalizers. The equalization process
associated with the minimum-redundancy solution consists of multiplying the received vector by
the receiver matrix, entailing O(M 2 ) operations. This complexity is high as compared to that of
11 We shall address this topic in Chapter 5 in a more detailed manner.
 2.5. CONCLUDING REMARKS 53
O(M log M) required by traditional OFDM and SC-FD transceivers. This efficient equalization
originates from the use of DFT matrices as well as the multiplication by memoryless diagonal
matrices, as explained in this chapter.

2.5 CONCLUDING REMARKS

This chapter has briefly reviewed the modeling of communications systems using the transmulti-
plexer framework.The LTI memoryless transceivers were the main focus of our presentation. Among
these transceivers we particularly addressed the CP-OFDM, ZP-OFDM-OLA, CP-SC-FD, and
ZP-SC-FD-OLA transceivers, highlighting their corresponding ZF and MMSE designs. Some re-
sults taken from the open literature related to transceivers with reduced redundancy (ZP-ZJ systems)
were also discussed.
A lesson learned from this chapter is that the conventional OFDM and SC-FD transceivers are
very efficient since the receiver and the equalizer have very simple implementations. These systems
capitalize on the circulant structure of the effective channel matrix whenever a cyclic prefix of length
at least L is inserted, where L is the channel order. The circulant matrices can be diagonalized
using a pair of DFT and IDFT transformations. Chapter 3 contains an in-depth description of
OFDM and SC-FD techniques, including details about the effects of employing different types of
prefixes/suffixes.
A further query is if it is possible to derive transceivers similar to the OFDM and SC-FD
while employing reduced redundancy and whose implementations rely on fast transforms as well.
The answer to such a query is yes, as will be clarified later on in Chapter 4. Another relevant question
is if it is possible to reduce even more the transmitted redundancy by working with time-varying
transmultiplexers with memory. Once again, the answer is yes, as described in Chapter 5.
 55

CHAPTER 3

OFDM
3.1 INTRODUCTION
As discussed in the previous chapters, the orthogonal frequency-division multiplexing (OFDM) is
a transmission technique that is currently used in a number of wired and wireless systems. This
chapter describes OFDM in more detail, starting from its original conception in the continuous-time
domain (herein called, analog OFDM ) and arriving at its current implementation in the discrete-time
domain. In fact, the discrete-time description of OFDM has already been addressed in Section 2.4
of Chapter 2. However, that description is solely based on the useful mathematical properties related
to circulant matrices, without necessarily calling for physical intuition of actual transmissions. The
focus of the present chapter, on the other hand, is to motivate the construction of the OFDM system
by analyzing its very insightful analog version and to derive the discrete-time implementation from
this physically meaningful continuous-time system. Indeed, it was only with the widespread use of
digital integrated circuit technology that the discrete-time OFDM transmission technique became
popular, especially due to the existence of fast Fourier transform (FFT) algorithms, which enable
efficient computations of the discrete Fourier transform (DFT) employed for modulation.
From a historical perspective, the origins of frequency-division multiplexing (FDM) date back
to the late nineteen century, according to the review article by S. B. Weinstein [88]. The analog
version of OFDM was first proposed by R. W. Chang in 1966 [10], who filed a patent that was
granted in 1970 [11]. A major breakthrough was the perception that the use of analog subcarrier
oscillators and their corresponding coherent demodulators could be avoided by replacing them by
DFT-based transceivers. In this context, S. B. Weinstein and P. M. Ebert [89] were the originators
of the DFT-based modulation and demodulation schemes. Another key result related to the digital
OFDM implementation was conceived by A. Peled and A. Ruiz [64] who advanced the use of cyclic
prefix as solution for maintaining orthogonality among subcarriers at the receiver side. Although the
analog and digital versions of OFDM systems are closely related, they are not always fully equivalent
as discussed in [43]. OFDM has become widely adopted in commercial applications, thus explaining
why there are so many works addressing its history [13, 88].
This chapter is organized as follows. Section 3.2 describes the origins of OFDM in its analog
version. Such topic is particularly interesting for understanding the choices of some important
parameters, such as the OFDM symbol duration, sampling period, and guard period. In addition,
Section 3.2 also introduces the importance of orthogonality in OFDM. Section 3.3 describes the
discrete-time implementation of OFDM systems. The idea of Section 3.3 is to connect what we
have seen in Section 2.4 of Chapter 2 with the theory of analog OFDM. Section 3.4 describes
56 3. OFDM
some variants of OFDM-based systems, including single-carrier with frequency-domain equalization
(SC-FD), zero-padding (ZP) schemes, coded OFDM (C-OFDM), and discrete multitone (DMT)
systems. Finally, some conclusions are drawn in Section 3.5.

3.2 ANALOG OFDM

Digital communications require the conversion of a discrete-time signal to a continuous-time
signal that is actually transmitted. Such an operation is performed by a digital-to-analog con-
verter (DAC). If we assume that sm denotes the mth element of a discrete-time signal, with
m ∈ M  {0, 1, . . . , M − 1} and M ∈ N, then the conversion to its related continuous-time base-
band signal sDAC (t) can be (theoretically) implemented by first multiplying each element sm by a
continuous-time Dirac impulse δ(t − mT ), and then passing the resulting signal through a linear
time-invariant (LTI) analog filter with impulse response p(t). In this context, the positive real-valued
parameter T denotes the sampling period of the DAC. Mathematically, we have
 

sDAC (t)  sm δ(t − mT ) ∗ p(t)
m∈M

= sm [δ(t − mT ) ∗ p(t)]

m∈ M
= sm p(t − mT ), (3.1)
m∈M

where ∗ represents linear convolution.

In other words, the usual digital-to-analog conversion, which is always present in digital
communications, can be regarded as a time-division multiplexing (TDM) operation of the elements
which compose a discrete-time signal. A natural question arises at this point: is there anything we can
do to perform this conversion in a frequency-division multiplexing (FDM) manner? The answer is
yes, and we will show that such an FDM-based representation is a natural starting point to conceive
the so-called analog OFDM.

3.2.1 FROM TDM TO FDM

In general, the continuous-time signal s(t) associated with a discrete-time signal sm can be described
as

s(t)  sm pm (t), (3.2)
m∈M

where pm (t) is a continuous-time pulse signal. The choice of the pulse signal determines how the
elements of the discrete-time signal are distributed over the time-frequency plane.
For example, by choosing pm (t) in a TDM fashion, so that pm (t)  p(t − mT ) whose time
support is the real interval [mT , (m + 1)T ), we generate the following continuous-time signal (see
 3.2. ANALOG OFDM 57
Equation (3.1) as well):

sTDM (t)  sm p(t − mT ). (3.3)
m∈M

The former equation implies that sTDM (t) is a concatenation of pulses p(t), each of them starting at
time t = mT with duration of T seconds, modulated by their corresponding symbol sm originating
from a digital modulator. From Fourier analysis we know that the Fourier transform (FT) of sTDM (t)
is

STDM (ω) = sm P (ω)e−jωmT , (3.4)
m∈M

where the FT of p(t) is represented by P (ω), whose bandwidth is . Since in TDM schemes
each symbol is transmitted in a time slot with T seconds of duration, then the transmission of M
symbols lasts MT seconds. In frequency domain each of these symbols occupies the entire available
bandwidth .
On the other hand, in FDM schemes we utilize a dual strategy for signal transmission. Indeed,
in FDM each symbol occupies a portion of the whole channel bandwidth . The frequency response
of the transmitted signal is

SFDM (ω)  sm P (ω − m ), (3.5)
m∈M

in which the support of P (ω − m ) is the real interval [m , (m + 1) ), where represents a

fraction of the channel bandwidth , that is, = M . A signal with such representation in the
frequency domain can be written in time domain as

sFDM (t) = sm p (t)ejm t
M
m∈ 

jm t
= sm e p (t). (3.6)
m∈M

Thus, FDM transmission is obtained when we choose pm (t)  p (t)ejm t , which implies that the
symbols sm are all superposed in time domain.
Equation (3.6) will be key to the forthcoming description of analog OFDM since it reveals
clearly the central role that complex exponentials play in FDM-based transmissions. Note that
the existence of such complex exponentials is a natural consequence of the FDM characteristic.
In addition, Equation (3.5) is related to the ideal concept of multicarrier systems, which focus on
dividing the available channel into many narrowband subchannels so that the channel frequency
response can be considered constant in each individual subchannel. In the remaining of this section
it will be shown how analog OFDM exploits FDM transmissions in an efficient manner.
58 3. OFDM
3.2.2 ORTHOGONALITY AMONG SUBCARRIERS
The starting point of analog OFDM is the term between parenthesis in expression (3.6), where each
entry (symbol) of a data block modulates a subcarrier, which can be interpreted as a tone. In order
for these symbols to be easily recovered at the receiver, the subcarriers should be orthogonal. The
concept of orthogonality among subcarriers will be explored in the following discussion.
Let us consider that the OFDM subcarriers consist of equally spaced tones in frequency
domain. Indeed, if we define fm , fm+1 ∈ R as the central frequencies corresponding to the mth and
(m + 1)th subcarriers, respectively, where

m ∈ M  {0, 1, . . . , M − 1} (3.7)

is the subcarrier index, and M is a positive integer number representing the number of subcarriers,
then the frequency separation between two consecutive subcarriers is

1
 fm+1 − fm , (3.8)

for all m ∈ M \ {M − 1}. Note that, by assuming that a subcarrier is comprised of a single tone, let
us say fm , the time-domain representation of such subcarrier consists of a complex exponential at
that frequency,1 that is, ej2πfm t .
The transmission of a block with M symbols belonging to a given constellation C ⊂ C, in
which each symbol is denoted as sm (n) ∈ C , is performed by transmitting these symbols using
subcarriers with distinct central frequencies. In this context, n is an integer number that identifies the
block with M constellation symbols. Such association between symbols and subcarriers is exemplified
below as

s0 (n)  → f0  0,
1
s1 (n)  → f1 = ,

..
.
M −1
sM−1 (n)  → fM−1 = . (3.9)

Hence, the mth symbol is associated with the subcarrier whose central frequency is fm =  m
,
for each m within M. The nth data block to be transmitted, usually called OFDM symbol, is 2

a complex signal denoted by ûn (t), in which t is a real variable representing time. The OFDM
symbol ûn (t) is generated as the superposition of the subcarriers, each of them modulated by its
1 Indeed, a single-tone signal whose tone is f
m is a signal whose frequency-domain representation consists of an impulse centered
at frequency fm . From Fourier analysis, we know that the inverse Fourier transform of an impulse at frequency fm corresponds
to ej2πfm t .
2The reader should not confuse the terms symbol and OFDM symbol. While the former is a complex number generated at the
output of a digital modulator (see Section 1.2), the latter is associated with a collection of constellation symbols.
 3.2. ANALOG OFDM 59
corresponding symbol sm (n), yielding
1 
ûn (t)  + sm (n)ej2πfm (t−nT̂ )
T̂ m∈M
1  2π
=+ sm (n)ej  m(t−nT̂ ) , for nT̂ ≤ t < (n + 1)T̂ , (3.10)
T̂ m∈M

where T̂ is a positive real number representing the duration of an OFDM symbol.

In the time domain, if we assume that each symbol sm (n) represents an entry of a serial
T̂
data, then the time support of each symbol is M , since each OFDM symbol is comprised of M
constellation symbols, as described in Equation (3.10). The OFDM symbol duration T̂ must be
long enough to keep the subcarriers orthogonal to each other so that the individual data symbols
can be extracted from the OFDM symbol.
2π 2π
Indeed, for any two modulated subcarriers, for example sm (n)ej  mt and sm+l (n)ej  (m+l)t ,
where l is an integer number such that m + l ∈ M, their temporal cross-correlation computed over
the OFDM symbol duration is given by
 (n+1)T̂
1 2π
∗ 2π
sm (n)ej  mt sm+l (n)e−j  (m+l)t dt
T̂ nT̂
 (n+1)T̂
∗ 1 2π 2π
= sm (n)sm+l (n) ej  mt e−j  (m+l)t dt
T̂ nT̂

∗ 1 (n+1)T̂ 2π
= sm (n)sm+l (n) e−j  lt dt
 T̂ nT̂
|sm (n)|2 if l = 0
= , (3.11)
0 if l  = 0

provided T̂ is a multiple of , that is, T̂ = κ, with κ being a positive integer. Indeed, the last
2π
equality follows easily by considering that e−j  lt can be rewritten as
# $ # $
2π 2π 2π
e−j  lt = cos − lt + j sin − lt , (3.12)
 
and remembering that both sine and cosine functions integrate to zero in intervals corresponding to
multiples of their fundamental period, which in this case is given by l . Therefore, we must choose
T̂ in Equation (3.11) in such a way that the cross-correlation is equal to zero for all l ∈ M \ {0}.
This implies that the choice of T̂ must be based on the slowest complex exponential,3 which occurs
when l = 1, which in turn shows that T̂ must be a multiple of 1 = .
Therefore, as pointed out in Equation (3.11), the orthogonality among subcarriers plays a key
role in the choice of T̂ and its relation with . Indeed, this orthogonality can be obtained by choosing
3The one that takes more time to complete a cycle. In addition, note that the fundamental period of the slowest complex exponential
is already a multiple of the periods of the other exponentials.
60 3. OFDM
an OFDM symbol duration T̂ = κ. Note that, as κ increases, the OFDM symbol duration also
increases, but the amount of transmitted data is exactly the same that would be transmitted if κ = 1,
that is, κ > 1 reduces the system throughput. That is why κ = 1 is the natural choice. In addition,
note that the orthogonality does not depend on the symbols that modulate the subcarriers.
This implies that the OFDM symbol ûn (t) can be redefined as
1  j 2π m(t−nT̂ )
ûn (t)  + sm (n)e T̂ , for nT̂ ≤ t < (n + 1)T̂ ,
T̂ m∈M
1  j 2π mt
=+ sm (n)e T̂ e−j2πmn  , for nT̂ ≤ t < (n + 1)T̂ ,
T̂ m∈M =1
 j 2π mt
= sm (n) p̂(t − nT̂ )e T̂
  
m∈M
ϕ̂m (t−nT̂ )

= sm (n)ϕ̂m (t − nT̂ ), (3.13)
m∈M

where function ϕ̂m (t) represents the mth subcarrier and the pulse signal is
⎧
⎨ +1 , for 0 ≤ t < T̂ ,
p̂(t)  (3.14)
⎩ T̂
0, otherwise.
Let us interpret Equation (3.13) pictorially. Each subcarrier ϕ̂m (t) is a complex exponential
multiplied by p̂(t), a rectangular window of duration T̂ . Figure 3.1(a) depicts a given pulse p̂(t),
whose Fourier transform is the well-known sinc function. The square of the sinc represents the
subcarrier spectrum, as depicted in Figure 3.1(b). Figure 3.1(c) illustrates many OFDM subcarriers,
placed at their correct positions,4 in order to show the distance between the central frequencies
of neighboring subcarriers and to emphasize that, at each subcarrier central frequency, all other
subcarriers have amplitude equal to zero, as illustrated by the dotted lines.
Figure 3.1(c) illustrates several subcarriers belonging to a single OFDM symbol as if they were
isolated (i.e., we have not added the curves associated with each subcarrier). However, as given in
Equation (3.13), an OFDM symbol is formed by the summation of all M subcarriers modulated by
their corresponding symbol. The result of such summation is represented in Figure 3.2. This figure
depicts a frequency-domain representation of an OFDM symbol comprising three subcarriers. Note
that the support of such a representation is the entire real axis.
At this point, it is worth mentioning that simple TDM- and FDM-based transmissions can
also yield orthogonal signals to the input of the communication channel, as illustrated in the dis-
cussions of Section 1.5 of Chapter 1 within the framework of multiple-access schemes. Indeed,
4 Remember that the Fourier transform of the product of two signals is equivalent to the convolution between the Fourier transforms
j 2π mt
of each individual signal. Thus, each subcarrier p̂(t)e T̂ can be represented in the frequency domain by the convolution of
a sinc function (Fourier transform of p̂(t)) with a Dirac impulse at frequency m (Fourier transform of the exponential), which
T̂
results in the sinc function centered at frequency m .
T̂
 3.2. ANALOG OFDM 61

0 “ ” 12
Subcarrier p sin πf T̂
@ T̂ “ ” A
Spectrum πf T̂

Pulse Shape
1
p̂(t)
T̂
t
T̂ − T̂4 − T̂2 0 2
T̂
4
T̂
f

(a) Time domain. (b) Frequency domain.

1
T̂

(c) Subcarriers.

Figure 3.1: Representation of OFDM subcarriers: (a) time-domain representation of p̂(t); (b) frequency-
domain representation of p̂(t); and (c) a set of non-interfering subcarriers.

Equations (3.3) and (3.5) are examples of (theoretically) orthogonal TDM and FDM. However, in
the case of FDM transmissions, one must necessarily let empty spectral regions for separating the
frequency content associated with each subcarrier, otherwise the filters employed in such separa-
tion would be for certain non-causal filters due to the required sharp transitions in the frequency
domain. The aforementioned analog OFDM avoids this waste of spectrum that generally occurs
in FDM-based systems by allowing spectrum superposition of the subcarriers. In standard TDM
transmissions, when the transmitted signal crosses a frequency-selective channel, the original time-
domain orthogonality is lost. A possible solution is adding guard intervals between the transmission
62 3. OFDM
Subcarrier A
Subcarrier B
Subcarrier C
A+B+C

··· ···

Figure 3.2: Example of frequency representation of an OFDM symbol comprised of three subcarriers.

of each constellation symbol, which also represents a waste of resources. Actually, TDM-based solu-
tions employ time-domain equalizers to decrease the interference among symbols due to the loss of
orthogonality. But, even in this case, if the interference level is too high, then the order of the time-
domain equalizer can turn its implementation impractical. Actual analog OFDM transmissions, on
the other hand, are able to circumvent the interferences introduced by frequency-selective channels
by using subcarriers that are orthogonal to each other at the receiver end, thus justifying the name
orthogonal FDM (OFDM). The key feature present in analog OFDM is the introduction of guard
intervals between each OFDM symbol, as explained in Subsection 3.2.3.

3.2.3 ORTHOGONALITY AT RECEIVER: THE ROLE OF GUARD INTERVAL

So far we have seen how to design the OFDM transmitter in such a way that its subcarriers are
orthogonal to each other, which is an important feature that allows easy extraction of the symbols
within an OFDM symbol. However, in practical communications systems this extraction is actually
performed at the receiver end.Thus, we are interested in transmitting OFDM symbols that maintain
their orthogonality among subcarriers when they reach the receiver end.
The task of keeping the subcarriers orthogonal at the receiver is very challenging, especially
when the OFDM symbol faces multipath fading channels. As described in Section 1.7, this kind of
channel has memory, which means that delayed and attenuated versions of the transmitted signal
arrive at the receiver. This generates interference among transmitted constellation symbols and
OFDM symbols (assuming the transmission of more than one OFDM symbol). An instinctive
solution to maintain the orthogonality and combat interference is to extend the time support of the
subcarriers.
Let us consider the simplest case in which a subcarrier is a simple sinusoid, as depicted in
Figure 3.3, where the solid line represents the original subcarrier with a finite time support and the
 3.2. ANALOG OFDM 63

t
     

Cyclic Prefix Sinusoidal Carrier

Figure 3.3: Extending an OFDM carrier with a cyclic prefix.

initial dashed version represents its time extension. This OFDM subcarrier extension is known as
cyclic prefix (CP). Assuming this carrier energizes a frequency-selective channel, whose time-delay
spread spans up to τmem ∈ R seconds, at the receiver end the first τmem seconds will be corrupted
by the previous OFDM symbol, thus generating the so-called interblock interference (IBI). After the
period τmem the IBI is over and if we keep the subcarriers illuminating the channel for T seconds,
then the subcarriers will be orthogonal to each other as long as the time period T − τmem is long
enough and as long as we are able to eliminate the interference among constellation symbols within
a given data block, i.e., the so-called intersymbol interference (ISI).
Figure 3.4 illustrates how OFDM symbols are concatenated taking into consideration a guard
interval5 of duration τ ∈ R seconds. The useful symbol time T̂ (see Figure 3.4) corresponds to the
original duration of the OFDM symbol whose subcarriers are orthogonal to each other only at
the transmitter side. Thus, the extended OFDM symbol has duration T  T̂ + τ , where the guard
interval τ is longer than the longest multipath delay (τ > τmem ) and is used to avoid the harmful
interferences introduced by frequency-selective channels.

τ
T

guard useful
period symbol time

Figure 3.4: Using a guard period to avoid interblock interference.

5 A guard interval or guard period is a more general concept which includes the cyclic prefix as a special case.
64 3. OFDM
1
Mathematically, if we define M subcarriers separated in frequency by T −τ Hz as
2π
ϕm (t)  p(t)ej T −τ mt , (3.15)

for m ∈ M, where the subcarrier index m is within the set M and the pulse signal is6
⎧
⎨√ 1 , for − τ ≤ t < T − τ,
p(t)  T −τ (3.16)
⎩
0, otherwise,

then the transmitted signal, which is the concatenation of extended OFDM symbols, can be written
as

u(t)  un (t)
n∈Z 

= sm (n)ϕm (t − nT ), (3.17)
n∈Z m∈M
  
=un (t)

where sm (n) is the mth symbol within the nth block representing an extended OFDM symbol.
An extended OFDM symbol obtained by the extension of each subcarrier-time support is
equivalent to a time-domain signal with a cyclic prefix of length τ , i.e., in which the first τ seconds of
the data block coincide with the last τ seconds. Indeed, consider the nth extended OFDM symbol,
whose time support is [nT − τ, nT + T − τ ). Let t be an arbitrary real number within the first
τ seconds of that interval, i.e., t ∈ [nT − τ, nT ). In addition, let t be a real number defined as
t  t + (T − τ ), which denotes a time instant within the last τ seconds of the referred block. Thus,
we have
 2π
un (t) = sm (n)p(t − nT )ej T −τ m(t−nT )
m∈M
 1 2π
= sm (n) √ ej T −τ m[t −(T −τ )−nT ]
m∈M
T −τ
 1 2π
= sm (n) √ ej T −τ m(t −nT ) · e−j2πm
 
T −τ
m∈M    =1
=p(t −nT )
= un (t ). (3.18)

A detailed representation of an extended OFDM symbol7 using cyclic prefix is depicted in

Figure 3.5. The reader should note the similarities and differences between Equations (3.13) and
(3.17). The signal ûn (t) in Equation (3.13) is simply an OFDM symbol, i.e., no guard period is
6 We considered the interval [−τ, T − τ ), instead of [0, T ), since this choice simplifies the forthcoming notations used in Section 3.3.
7 In some texts, especially in standards, the extended OFDM symbol is simply referred to as OFDM symbol.
 3.2. ANALOG OFDM 65

Cyclic Prefix
     

τ
T

guard useful
period symbol time

Figure 3.5: Cyclic prefix in an OFDM symbol.

being used and, as a consequence, the received signal corresponding to this ûn (t) after crossing a
multipath fading channel would not yield orthogonal subcarriers. On the other hand, the signal u(t)
in Equation (3.17) is the concatenation of infinitely many extended OFDM symbols using cyclic
prefix in order to ensure orthogonality among subcarriers at the receiver side.
Indeed, the orthogonality among subcarriers is maintained since the interference between
OFDM symbols (i.e., the IBI) can be eliminated by discarding the first τ seconds out of T seconds
of each received data block. As for the remaining interference (ISI) due to constellation-symbol
superpositions within the resulting block of duration T − τ , we can eliminate it using the subcarrier
orthogonality. In order to verify mathematically these facts, let us analyze the received signal y(t)
assuming a noiseless baseband channel model whose time-delay spread is τmem , as follows:

y(t)  H {u(t)}
, -
 

=H sm (n )ϕm (t − n T )
∈Z m∈M
 n ! "
= sm (n )H ϕm (t − n T )
∈Z m∈M
n  ! "
= sm (n )H ϕm (t − n T ) , (3.19)
m∈M n ∈Z

where H {·} represents the linear system that models the referred baseband channel. Now, by con-
sidering that the channel model
! remains" constant during the interval of an OFDM symbol, we can
compute the quantity H ϕm (t − n T ) through a convolution integral of ϕm (t − n T ) with the
66 3. OFDM
channel impulse response hn (t) associated with the n th symbol. Therefore, we have

  ∞
j T2π 2π
y(t) =
sm (n )e −τ m(t−n T ) hn (τ )p(t − n T − τ )e−j T −τ mτ dτ
m∈M n ∈Z −∞
τmem
  2π 2π
= sm (n )ej T −τ m(t−n T ) hn (τ )p(t − n T − τ )e−j T −τ mτ dτ , (3.20)
m∈M n ∈Z 0

where hn (τ ) = 0 for all τ  ∈ [0, τmem ] and8

⎧
⎨√ 1 , for t + τ − (n + 1)T ≤ τ < t + τ − n T ,

p(t − n T − τ ) = T −τ (3.21)
⎩
0, otherwise.

As we are studying both interblock and intersymbol interferences associated with the trans-
mission of blocks un (t), with n ∈ Z, of length T , it is convenient to separate the received signal into
blocks with the same length T . Thus, we can write

y(t) = y(nT + t )
  
n∈Z
yn (t )


= yn (t ), (3.22)
n∈Z

where the time instant t is in the interval [−τ, T − τ ). This way, it follows from Equation (3.20)
that the nth received block can be expressed by
  2π
yn (t ) = sm (n )ej T −τ m[(n−n )T +t ] In,n (t ), (3.23)
m∈M n ∈Z

with
τmem

2π

In,n (t )  hn (τ )p (n − n )T + t − τ e−j T −τ mτ dτ . (3.24)
0

Observe that, for n ≥ n + 1, we have p((n − n )T + t − τ )  = 0 only for some values of

τ outside the integration interval [0, τmem ]. The same occurs for n ≤ n − 2. Indeed, based on
Equation (3.21), one has that the pulse signal p((n − n )T + t − τ ) assumes non-zero values only
for values of τ such that [n − (n + 1)]T + t + τ < τ ≤ (n − n )T + t + τ . Therefore, if n ≥
n + 1, then the upper bound for τ will be such that (n − n )T + t + τ ≤ t + τ − T = t − (T −
τ ) < 0, since t ∈ [−τ, T − τ ), which means that τ cannot be within the interval [0, τmem ] for this
8 Actually, it would be more appropriate to state that h (τ ) is approximately zero for all τ  ∈ [0, τ
n mem ].
 3.2. ANALOG OFDM 67
n
condition. Similarly, if ≤ n − 2 then the lower bound for will be such that τ + 1)]T + [n − (n

t + τ ≥ t + τ + T > τ > τmem , which means that τ cannot be within the interval [0, τmem ] in
this case as well. Hence, the integrand which appears in the above integral In,n (t ) is always zero,
except for n = n and n = n − 1. When n = n, then the pulse signal p((n − n )T + t − τ ) is
non-zero whenever τ is such that t + τ − T < 0 < τ ≤ t + τ . Similarly, when n = n − 1 the
interval is 0 < t + τ < τ ≤ t + τ + T > τmem . Therefore, we have
min{t +τ,τmem }
δ[n − n] 2π
In,n (t ) = √ hn (τ )e−j T −τ mτ dτ
T −τ
0
τmem
δ[n − n + 1] 2π
+ √ hn (τ )e−j T −τ mτ dτ , (3.25)
T −τ
min{t +τ,τmem }

where δ[n] denotes the Kronecker delta, which is defined as


1, if n = 0,
δ[n]  (3.26)
0, otherwise.

Now, if we consider only time instants t within the interval [0, T − τ ), with τ > τmem , then
min{t + τ, τmem } = τmem , and, therefore, the second integral which appears at the right-hand side
of Equation (3.25) will be zero. In other words, when we discard the first τ > τmem seconds of
each received block yn (t ), then there is no interference between the nth and (n − 1)th transmitted
OFDM symbols, i.e., for any t ∈ [0, T − τ ), we have
τmem
δ[n − n] 2π

In,n (t ) = √ hn (τ )e−j T −τ mτ dτ
T −τ
0
 


2π m
Hn T −τ
# $
δ[n − n] 2π m
= √ H n , (3.27)
T −τ T −τ

so that (see Equation (3.23) as well)

 # $
1 2π m 2π
yn (t ) = √ Hn sm (n)ej T −τ mt , (3.28)
T − τ m∈M T − τ

in which, once again, we highlight that t must be in the interval [0, T − τ ).

From Equation (3.28), it is clear that, for each fixed time interval [nT − τ, nT + T − τ ), if
one discards the first τ > τmem seconds of the received signal in that interval, then one ends up
with a continuous-time signal that is composed of a sum of M complex exponentials modulated by
68 3. OFDM

sm (n)Hn T2π−τ
m
complex numbers √ . Hence, if we compute the temporal cross-correlation over a useful
T −τ
symbol duration T̂ = T − τ for any of those two modulated subcarriers, we would reach a similar
result to Equation (3.11). This means that the orthogonality between subcarriers is achieved at the
receiver side. This is the main feature of analog OFDM, since the IBI is eliminated and, at the
same time, the ISI can be eliminated by using the resulting orthogonality among subcarriers at the
receiver.

1
Δf =
T −τ

t
T

Figure 3.6: Time-frequency map of analog OFDM signals.

Therefore, analog OFDM transmissions illuminate the channel at each T = T̂ + τ period

of time, with τ > τmem , using M subcarriers whose central frequencies of neighboring subcarriers
are separated by f  1 = T −τ 1
, as illustrated in Figure 3.6. Each subcarrier will be responsible
T̂
for transporting a single symbol during a time slot, taking into consideration the longest path
propagation time τmem .
Note that the extended OFDM symbols do not have orthogonal subcarriers at the transmitter
due to the insertion of the guard interval. This can be verified through different ways. For example,
if one computes the temporal cross-correlation between two distinct subcarriers described in Equa-
tion (3.15), one would end up with a non-zero temporal cross-correlation in general. This occurs
because we do not integrate over an integer multiple of the period T − τ . Another way of verifying
this fact is to observe Figure 3.7, in which the time-domain and frequency-domain representations
of the pulse p(t) in extended OFDM symbols are depicted. As the subcarrier central frequencies
1
are proportional to T −τ and the zeros of the subcarrier spectra are proportional to T1 , then there
 3.2. ANALOG OFDM 69

„ «2
Subcarrier √ sin (πf T )
T
Spectrum (πf T )

Pulse Shape
1
p(t) √
T
t
T − T4 − T2 0 2
T
4
T
f

(a) Time domain. (b) Frequency domain.

Figure 3.7: Representation of extended OFDM subcarriers: (a) time-domain representation of p(t);
and (b) frequency-domain representation of p(t).

will exist interferences from adjacent subcarriers at those central frequencies. This means that the
subcarriers are not orthogonal at the transmitter side. However, this is not an issue since the infor-
mation extraction occurs at the receiver. Thus, the orthogonality among subcarriers at the receiver
end allows a proper extraction of the symbols associated with different subcarriers within a received
extended OFDM symbol, even when the received signal has been severely distorted by the channel.
That is why we exchange the original orthogonality present in OFDM symbols by an orthogonality
of the received extended OFDM symbols with cyclic prefix (after discarding the first τ seconds).

3.2.4 SPECTRAL EFFICIENCY, PAPR, CFO, AND I/Q IMBALANCE

Before we move on to describe an implementation sketch of analog OFDM, let us briefly comment
on its spectral efficiency and some of the OFDM issues, namely peak-to-average power ratio (PAPR),
carrier-frequency offset (CFO), and in-phase/quadrature-phase (I/Q) imbalance.
Let us start with the spectral efficiency. We know that, during a period of T seconds, OFDM
transmits M symbols from a given constellation with 2b points, where b is a natural number denoting
the number of bits required to represent a single symbol. Thus, the OFDM bit rate (BR) is

Mb
BR = (3.29)
T

in bits per second (bps). If we add the spectra of all subcarriers and consider that side lobes below
20 dB from the main lobe are negligible, corresponding to the second side lobes on each side of the
70 3. OFDM
main lobe, then the total bandwidth is9

M −1 3
BW = +2 . (3.30)
T −τ T
The ratio
BR Mb
= (M−1)T
BW
T −τ +6
b
=
(3.31)
M−1
M
T
T −τ + 6
M

is the so-called spectral efficiency, which tends to b(1 − Tτ ) for large M. That means the OFDM is
an optimal modulation in terms of spectral efficiency as long as Tτ  1. However, when τ  T (very
dispersive environment), then the spectral efficiency of OFDM transmissions is quite small. Other
multicarrier and single-carrier transmissions which address this drawback of spectral efficiency will
be described in Chapters 4 and 5.
Figure 3.8 depicts the instantaneous power of a given OFDM symbol un (t). A dotted line is
used to represent the average power of the OFDM symbol. In this figure, we can observe that there
exist some peaks in the power of un (t) that are much higher than the average power, i.e., they are
well above the dotted line. Indeed, it is well known that the peak-to-average power ratio (PAPR) of
OFDM transmissions is higher than the PAPR of single-carrier transmissions (see Subsection 3.4.1
for further details in the discrete-time domain). High PAPR is undesirable because it implies a wide
dynamic range of the signal to be transmitted, which in turn requires power amplifiers with linear
response over a wide range, increasing the cost of such devices. This is one of the main reasons why in
LTE the use of OFDMA in the uplink was avoided. Therefore, PAPR is an important impairment
related to OFDM transmissions.
In cases where the carrier frequency of the received signal does not match the carrier frequency
of the transmitted signal, we have the so-called carrier-frequency offset (CFO). Thus, CFO is the
offset (difference) between two numbers representing carrier frequencies, one at the transmitter and
the other at the receiver end. Ideally, CFO should be close to zero, but there are many practical cases
in which non-negligible CFO occurs. For example, when the transmitter and/or receiver are moving,
which usually happens in mobile communications, the Doppler effect acts as a source of CFO. Note
that, from our previous discussion about the importance of orthogonality in OFDM systems, it is
rather intuitive that CFO has the potential to severely degrade the quality of OFDM transmissions.
Most of the solutions to the CFO issue rely on blind estimation of the frequency offset, and are a bit
complex and/or applicable to very particular cases. Some of the low-complexity solutions to CFO
are presented in [46, 92].
9The bandwidth must be computed by considering that the central frequencies of the M subcarriers are separated by 1 Hz,
T −τ
and each subcarrier is a sinc whose second side lobes decay more than 20 dB at T3 Hz, as can be observed in Figure 3.7(b).
 3.2. ANALOG OFDM 71
 
un(t)2

Figure 3.8: Instantaneous power of a single OFDM symbol.

Moreover, digital transmissions usually employ two branches: an in-phase (I) and a quadrature-
phase (Q) branch. These branches are associated with the real and imaginary parts of the transmitted
signal, respectively. I/Q imbalance occurs when there is phase and/or amplitude mismatches between
I and Q branches. Such mismatches are usually due to the imperfections in the process of the
radio-frequency signal down-conversion to baseband signal and are, therefore, unavoidable in the
analog front-end [74]. In most cases, I/Q imbalance can only deteriorate the bit-error rate (BER)
performance of OFDM systems when they are employing high-order modulation schemes, such as
64-QAM (quadrature amplitude modulation). When I/Q imbalance is a major issue, one can use
digital signal processing techniques to compensate such mismatch. Indeed, there already exists several
techniques to compensate for the I/Q imbalance without increasing significantly the computational
burden. For more details, see [6, 78, 85] and references therein.

3.2.5 IMPLEMENTATION SKETCH

Figure 3.9 depicts the implementation sketch of an analog OFDM transmission scheme. First the
symbols sm (n) modulate their corresponding time-extended subcarriers ϕm (t − nT ). This operation
is represented by the filtering of the continuous-time signal sm (n)δ(t − nT ) through the analog filter
ϕm (t). The signals resulting from each of these modulations are then added, forming an extended
OFDM symbol, which crosses an analog channel and is then corrupted by the environment noise
v(t). Here, we assume that the length of the cyclic prefix is greater than the channel time-delay
spread. At the receiver end, there are filters ψm (t) which are responsible for discarding the cyclic
prefix as well as extracting the symbol sm (n) from the corrupted OFDM symbol. Indeed, by defining
the receiver filters as
j2π
ψm (t)  q(t)e T −τ mt , (3.32)
72 3. OFDM
for m ∈ M, where
⎧
⎨√ 1 , for 0 ≤ t < T − τ,
q(t)  T −τ (3.33)
⎩
0, otherwise,

the output of the filter ψm (t) will be equivalent to the temporal cross-correlation between the
OFDM symbol and the mth receiver filter. Following the same steps performed in Equation (3.11)
and remembering that the fundamental period of the slowest subcarrier is  = T − τ , it is easy to
verify that ψm (t) will remove all subcarriers except ϕm (t) due to orthogonality. In fact, the pair of
functions ϕm (t) and ψm (t) are biorthogonal [17]. In addition, note that the basis function at the
receiver ψm (t) has a time support shorter than the basis function at the transmitter ϕm (t).

s0 (n)δ(t − nT ) ϕ0 (t) v(t) ψ0 (t)

Analog
s1 (n)δ(t − nT ) ϕ1 (t) ψ1 (t)
Channel
···

···
sM −1 (n)δ(t − nT ) ϕM −1 (t) ψM −1 (t)

Each
T

Figure 3.9: Analog OFDM implementation sketch.

Even though analog OFDM can be derived from a very insightful view of the digital-to-
analog conversion implemented in an orthogonal FDM-based fashion, the resulting implementation
sketch depicted in Figure 3.9 also summarizes its main drawback: in general, practical solutions
entail the use of a large number of orthogonal subcarriers, thus hindering the applicability of this
structure in practice. Indeed, if M is large, then we would have to implement a large amount of
different oscillators/modulators, which may not be practical. This is one of the main reasons why
this analog version of OFDM was not employed in commercial applications after its proposal by
R. W. Chang [10, 11]. However, many of its properties and interpretations are still useful, as S. B.
Weinstein and P. M. Ebert [89], as well as A. Peled and A. Ruiz [64] noticed when they realized
that OFDM could be efficiently implemented in the discrete-time domain. This implementation
will be addressed in the next section.
 3.3. DISCRETE-TIME OFDM 73
3.3 DISCRETE-TIME OFDM
Digital signal processing (DSP) has emerged as a powerful and efficient tool in a growing number
of applications. Indeed, there are many situations where the use of DSP-based techniques has either
greatly simplified the implementation of practical systems or simply enhanced their performance.
The discrete-time implementation of OFDM systems is an example of how DSP can even enable
the practical usage of a given technique, which could be quite hard to implement otherwise.
In this section we will apply sampling to the OFDM symbols and show a discrete-time
implementation of OFDM systems. As will be shown, the discrete Fourier transform (DFT) plays
a central role in this process, enabling efficient implementations of OFDM by means of fast Fourier
transform (FFT) algorithms.

3.3.1 DISCRETIZATION OF THE OFDM SYMBOL

As in the continuous-time case, let us start with OFDM symbols without guard intervals described
in Subsection 3.2.2. The first challenge that one faces while trying to derive a discrete-time imple-
mentation of OFDM is that the OFDM symbols are not bandlimited. Indeed, due to the use of a
time-domain rectangular window p̂(t) as illustrated in Figure 3.1(a), each subcarrier in an OFDM
symbol and the entire OFDM symbol itself contain spectral components at infinitely large frequen-
cies, as depicted in Figures 3.1(b) and 3.1(c). Therefore, a uniform sampling of such continuous-time
signal would invariably result in aliasing effects, regardless of the particular sampling frequency em-
ployed.
In other words, no discrete-time implementation of OFDM obtained by sampling the analog
OFDM can be used to recover the same original continuous-time signal associated with analog
OFDM (which employs rectangular windows). In general, the recovering process is implemented as
a digital-to-analog conversion, such as the one described in expression (3.1), with the constellation
symbols being replaced by the elements that compose an OFDM symbol. Such impossibility is also
pointed out in [43].
In principle, the presence of aliasing effects could hinder any interpretation based on the
intuitive results of analog OFDM presented in Section 3.2. Fortunately, the aliasing effects are not
determinant here and the reason is quite simple to understand: the original signal to be transmitted,
sm (n), has a discrete nature, which means that we are not interested in the particular form of the
continuous-time OFDM symbol as long as we can detect the original discrete-time signal.10 It turns
out that each element of this discrete signal is the scaling factor of the subcarrier spectra at their
central frequencies. Hence, if the spectral components at the subcarrier central frequencies do not
suffer from aliasing effects, then we could recover the discrete-time signal sm (n). Once again, the
orthogonality of the subcarriers in an OFDM symbol plays an important role here (see Figure 3.1(c)).

10This is true only for detection purposes. If one is interested in studying the spectral roll-off related to the output of OFDM
systems or the effect of CFO (just to mention a few examples), then one should work directly with the discrete-time OFDM
model after the digital-to-analog conversion [43].
74 3. OFDM
Indeed, when we sample an OFDM symbol,11 the spectral repetitions due to the sampling
process are spaced apart by fs Hz, where fs  T1s denotes the sampling frequency and Ts ∈ R
denotes the sampling period. First of all, fs must be larger than M−1 , otherwise we would have
T̂
at least one central frequency being shared by two distinct subcarriers, constituting a harmful type
of aliasing in this case. But if we consider any integer multiple of 1 larger than M − 1, then we
T̂
would not have any kind of interference at the subcarrier central frequencies. For simplicity reasons,
it is better to use the smallest sampling frequency that does not cause interference at the subcarrier
central frequencies, i.e., fs = M . Thus, we have
T̂

T̂
Ts  . (3.34)
M
Without loss of generality, since the OFDM symbols are non-overlapping in time, we can
analyze each block of symbols separately. The nth OFDM symbol is non-zero only for time instants
within the interval [nT̂ , (n + 1)T̂ ).This means that, for all t ∈ [0, T̂ ), we have from Equation (3.13)
that

ûn (t + nT̂ ) = sm (n)ϕ̂m (t + nT̂ − nT̂ )

m∈ M
= sm (n)ϕ̂m (t ). (3.35)
m∈M

Now, let ûn,k be the discrete-time signal stemming from sampling the continuous-time sig-
nal ûn (t + nT̂ ) described in the above equation at each time instant t = kTs , where, due to the
definition of Ts in Equation (3.34) and as t ∈ [0, T̂ ), we must have k within the set M. Hence, the
resulting discrete-time representation of an OFDM symbol is

ûn,k  ûn (kTs + nT̂ )


= sm (n) ϕ̂m (kTs )
  
m∈M
ϕ̂m,k

= sm (n)ϕ̂m,k , (3.36)
m∈M

where, based on Equation (3.13), we have that the discrete-time version of the mth subcarrier in an
OFDM symbol is given as

1 j 2π m( kT̂ )
ϕ̂m,k = + e T̂ M
T̂
1 j 2π mk
=+ eM . (3.37)
T̂
11 Here, sampling means to multiply the continuous-time function by a train of Dirac impulses.
 3.3. DISCRETE-TIME OFDM 75
Thus, we can rewrite the discrete OFDM symbol in a more convenient manner as follows:

1  2π
ûn,k = + sm (n)ej M mk
T̂ m∈M
  
√
 MIDFT{sm (n)}k
.
M
= IDFT {sm (n)}k , (3.38)
T̂

for each k ∈ M, where the inverse DFT (IDFT) of the discrete-time signal sm (n), with m ∈ M,
is also a sequence with length M. Hence, the following relation also holds:

.
T̂ ! "
sm (n) = DFT ûn,k m , (3.39)
M

in which the DFT of the sequence ûn,k is also a sequence whose mth element is defined as

! " 1  2π
DFT ûn,k m  √ ûn,k e−j M mk , (3.40)
M k∈M

thus implying that the symbols sm (n) can be actually written as

T̂  ∗
sm (n) = ûn,k ϕ̂m,k . (3.41)
M
k∈M

In summary, the discrete-time version of OFDM symbols without guard intervals is easily
computed through an IDFT of the transmitted discrete-time signal, which can be efficiently im-
plemented by using an FFT algorithm. Thus, assuming the channel introduces no distortion on the
transmitted signal, the sequence of symbols sm (n) could be recovered at the receiver end by taking
the DFT of the sequence ûn,k and scaling the result, as shown in Equation (3.39). This is a rather
obvious conclusion since the DFT and IDFT are inverse operations.
Note that the discrete-time OFDM symbols ûn,k in expression (3.36) can be thought as the
superposition of M subcarriers ϕ̂m,k modulated by the symbols sm (n). It is worth mentioning that the
original orthogonality present in the analog OFDM symbol is preserved in the discrete-time case.
Indeed, we can see that the following relation between any two subcarriers with indexes i, j ∈ M
76 3. OFDM
is valid:
 1  2π
∗
ϕ̂j,k ϕ̂i,k = ej M k(j −i)
k∈M T̂
⎧ k∈#M 2π $
⎨ 1 ej M M(j −i) −1 , if i  = j,
= T̂ e
j 2π
M (j −i) −1
⎩ M
, otherwise.
, T̂
0, if i  = j,
= M
, otherwise.
T̂
M
= δ[i − j ]. (3.42)
T̂
Equation (3.42) is nothing but the orthogonality between subcarriers which are synchronized
and have the same duration T̂ .This expression stems from the projection of a transmitted signal onto
a subcarrier for detection purposes at the receiver end, as exemplified in Equation (3.41). For analog
OFDM, we have shown that the subcarrier orthogonality at the transmitter side is not sufficient
to allow the detection of the transmitted symbols in practical situations, i.e., when the data faces a
frequency-selective channel. This occurs since frequency-selective fading channels extend the time
support of the transmitted signals generating both IBI and ISI, as explained in Subsection 1.7. In
Subsection 3.2.3, we ensure subcarrier orthogonality at the receiver by inserting a cyclic prefix before
transmission. In the following subsection, we shall generate a discrete-time version of the results
related to extended analog OFDM symbols using cyclic prefix.

3.3.2 DISCRETIZATION AT RECEIVER: THE CP-OFDM

Our starting point in this subsection is the expression (3.28) of the received block after removing
the first τ seconds. As pointed out before, τ is chosen in such a manner that it is larger than the
delay spread of the channel τmem . Let L be a positive integer number defined as
/ 0
τmem
L , (3.43)
Ts

where Ts = M T̂
is the sampling rate associated with the discretization process. In addition, let K be
a positive integer number such that K ≥ L. Thus, by choosing the length of the cyclic prefix as

τ  KTs , (3.44)

then we have that τ > τmem and, therefore, expression (3.28) holds for any t ∈ [0, T − τ ).
Now, by remembering from Subsection 3.2.3 that the useful symbol time T̂ is given by T − τ ,
then the sampling rate can be written as
T −τ
Ts = , (3.45)
M
 3.3. DISCRETE-TIME OFDM 77
which means that [0, T − τ ) = [0, MTs ).
Let yn,k be the discrete-time signal originating from sampling the continuous-time signal
yn (t ) described in Equation (3.28) at each time instant t = kTs , with k ∈ M. As explained before,
this continuous-time signal is the nth received block after removing the first τ = KTs seconds.
Thus, the resulting discrete-time representation of the nth received extended OFDM symbol after
removing the first K elements12 is

yn,k  yn (kTs )
 # $
1 2π m 2π T −τ
=√ Hn sm (n)ej T −τ mk M
T − τ m∈M T −τ
 # $
1 2π m 2π
=√ Hn sm (n)ej M mk
T − τ m∈M T −τ
 # $ 2π
2π m ej M mk
= Hn sm (n) √
T −τ T −τ
m∈M      
λm (n) =ϕ̂m,k

= λm (n)sm (n)ϕ̂m,k . (3.46)
m∈M

Equation (3.46) means that the received signal, after removing the first K elements associated
with the guard interval introduced at the transmitter side, is composed of M modulated subcarriers
ϕ̂m,k . Even though the kth element of such signal is affected by all transmitted symbols sm (n), with
m ∈ M, one can use the subcarrier orthogonality expressed in Equation (3.42) in order to recover a
scaled version of the transmitted symbols without ISI. Therefore, by projecting the received signal
onto the mth subcarrier, one gets
T −τ  ∗
λm (n)sm (n) = yn,k ϕ̂m,k . (3.47)
M
k∈M

The former projection process can be implemented in a much more efficient way. Indeed, based
on Equation (3.46) and following a similar approach which was employed in Subsection 3.3.1, we
can rewrite the received discrete OFDM symbol in a more convenient manner as follows:
1
M
yn,k = IDFT {λm (n)sm (n)}k , (3.48)
T −τ
for each k ∈ M. Therefore, the following relation also holds:
1
T −τ ! "
λm (n)sm (n) = DFT yn,k m . (3.49)
M
12 If we are ignoring the first τ = KT seconds of the continuous-time received block, then we are ignoring the first K elements of
s
the related discrete-time signal with corresponding sampling rate of T1 Hz.
s
78 3. OFDM
Hence, in order to recover sm (n), all we need is to multiply the mth element of the DFT of the
received OFDM symbol by λm1(n) , assuming λm (n)  = 0, for all m ∈ M. This is the so-called zero-
forcing (ZF) equalizer. In fact, there are many other ways to perform equalization in order to estimate
sm (n). The choice of the equalizer depends on the types of distortion faced by the transmitted signal.
For instance, in our previous discussion we have neglected the existence of additive noise. In the
presence of such type of noise, an equalizer that minimizes the mean square error (MSE) would be
more appropriate than an equalizer that eliminates only the ISI, such as the ZF equalizer. In this
case, the equalization would consist of multiplying the mth element of the DFT of the received
OFDM symbol by

λ∗m (n)
, (3.50)
σv2 (n)
|λm (n)|2 + σs2 (n)

where σs2 (n) and σv2 (n) represent the variance of symbols and noise, respectively.13 This type of
equalizer is known as minimum MSE (MMSE) equalizer, as previously discussed in Subsection 2.4.1.
As explained in Subsection 3.2.3, the subcarrier orthogonality at the receiver end is obtained
by including a cyclic prefix of length τ at the transmitter end, generating the extended OFDM
symbols. Thus, by following the same steps employed in Subsection 3.3.1, but now considering the
nth extended OFDM symbol un (t) of Equation (3.17), we have

un (t + nT ) = sm (n)ϕm (t + nT − nT )

m∈ M
= sm (n)ϕm (t ), (3.51)
m∈M

for all t ∈ [−τ, T − τ ) = [−KTs , MTs ). Hence, for each k ∈ {−K, · · · , −1, 0, · · · , M − 1}, we
can define the discrete-time representation of an extended OFDM symbol as (see Equation (3.15))

un,k  un (kTs + nT )
1  2π
=√ sm (n)ej M mk , (3.52)
T − τ m∈M

where, for k ∈ M, one has

 2π √
sm (n)ej M mk = MIDFT {sm (n)}k , (3.53)
m∈M

13 It is assumed here that all constellation symbols within the nth OFDM symbol have the same variance σ 2 (n). In addition, it was
s
assumed that all noise components have the same variance σv2 (n) as well. Those assumptions are not necessary, but they simplify
the notation. See Subsection 3.4.4 for the case where we do not consider those equal-power assumptions.
 3.3. DISCRETE-TIME OFDM 79
whereas, for k ∈ {−K, · · · , −1}, one has
 2π  2π 2π
sm (n)ej M mk = sm (n)ej M mk ej MmM
m∈M m∈M =1
 2π
= sm (n)ej M (M+k)m
√M
m∈
= MIDFT {sm (n)}(M+k) , (3.54)

with M + k ∈ {M − K, · · · , M − 1} ⊂ M, assuming that M ≥ K, i.e., the useful symbol duration

is larger than or equal to the duration of the guard interval.
Therefore, we can use these results in Equation (3.52) so that, for each k ∈
{−K, · · · , −1, 0, · · · , M − 1}, we have
1
M
un,k = IDFT {sm (n)}k , (3.55)
T −τ
in which the above notation considers the inherent periodicity property of the IDFT, i.e.,
IDFT {sm (n)}(−k) = IDFT {sm (n)}(M−k) , for k ∈ {1, · · · , K}. Therefore, the first K elements of
the signal un,k are equal to its last K elements, thus characterizing the type of guard interval as cyclic
prefix (CP-OFDM). In the discrete-time domain, this guard interval is also called redundancy since
it corresponds to entries that do not carry additional information.
So far, we described the discrete-time versions of the transmitter and receiver. We shall address
the channel model in Subsection 3.3.3.

3.3.3 DISCRETE-TIME MULTIPATH CHANNEL

The term λm (n) which appears in expression (3.46) is associated with the frequency response of the
analog channel evaluated at the central frequency T m−τ Hz, for each subcarrier index m ∈ M and
for each block n ∈ Z. The aim of this subsection is to show the relation between λm (n) and the
discrete-time model of the analog baseband channel.
In actual discrete-time OFDM implementations, one must always associate the discrete
OFDM signal un,k with a related continuous-time signal, let us say u(t). As already mentioned,
due to aliasing effects this continuous-time version is not the analog OFDM symbol presented in
Section 3.2. The standard way to perform such conversion is through a DAC process, as briefly
described in Subsection 3.2.1, so that

 M−1

u(t)  un,k pT (t − (k + nN )Ts ) , (3.56)
n∈Z k=−K

where

N M +K (3.57)
80 3. OFDM
denotes the amount of transmitted elements per extended OFDM symbol, Ts is the sampling period
of the DAC, and pT (t) is the transmitting pulse that may be chosen based on several distinct criteria,
such as best spectral roll-off.
After applying u(t) to a continuous-time baseband channel represented by the linear operator
H {·}, we have the following resulting signal:

r(t)  H {u(t)}
 M−1

= un,k H {pT (t − (k + nN )Ts )} , (3.58)
n∈Z k=−K

where H {pT (t − (k + nN )Ts )} is the output of the linear channel corresponding to nth transmitted
block, which can be computed as

H {pT (t − (k + nN )Ts )} = (hn ∗ pT ) (t − (k + nN )Ts ) . (3.59)

The function hn (t) represents the impulse response of the analog baseband channel associated with
the nth transmitted block.14 Hence, one has

 M−1

r(t) = un,k (hn ∗ pT ) (t − (k + nN )Ts ) . (3.60)
n∈Z k=−K

At the receiver end, an analog-to-digital conversion is also implemented by first filtering the
received waveform and then sampling it at a given rate (e.g., the same sampling rate employed at
the transmitter). By considering that the receiving filter impulse response is pR (t), then the signal
y(t) resulting from this first filtering stage at the receiver front-end is given by

y(t)  (pR ∗ r)(t)


 M−1
= un,k (pR ∗ hn ∗ pT ) (t − (k + nN )Ts )
  
n∈Z k=−K
hn (t−(k+nN)Ts )
 M−1

= un,k hn (t − (k + nN )Ts ) , (3.61)
n ∈Z k=−K

where the equivalent analog baseband channel model hn (t) is given by

hn (t) = (pR ∗ hn ∗ pT )(t). (3.62)

14 It is assumed that the channel model does not vary with k during the transmission of the nth block.
 3.3. DISCRETE-TIME OFDM 81
We can now define the discrete-time version of y(t) as

yn ,k  y(k Ts + n NTs )
 M−1    
= un,k hn k − k Ts + (n − n)N Ts
k=−K
  
n∈Z
hdn ((k −k)+(n −n)N )
 
M−1
 
= un,k hdn (k − k) + (n − n)N , (3.63)
n∈Z k=−K

in which the impulse response

hdn (l)  hn (lTs )

= (pR ∗ hn ∗ pT ) (lTs ) (3.64)

is the discrete-time equivalent channel model associated with the nth transmitted block.
Now, in order to determine the relation between λm (n) and hdn (l), consider that hn (t) is the
impulse response of a multipath fading channel with delay spread τmem ≤ LTs . This way, it can be
approximated as15


hn (t) = hdn (l)δ(t − lTs ), (3.65)
l∈L

in which the set of channel-tap indexes L is defined as

L  {0, 1, · · · , L}. (3.66)

15 We use the word “approximated,” since the delay associated with the lth tap is not necessarily equal to lT . However, if T is small
s s
enough, then this expression is a reasonable approximation for an equivalent multipath baseband channel whose corresponding
transmitting and receiving filters were properly designed.
82 3. OFDM
−τ
Taking this fact into account, assuming that Ts = TM , and considering Equations (3.27)
and (3.46), we have
# $
2π
λm (n) = Hn m
T −τ
τmem
2π
= hn (τ )e−j T −τ mτ dτ
0
τmem
 2π
= hdn (l) δ(τ − lTs )e−j T −τ mτ dτ
l∈L
 0
d −j T2π
= hn (l)e −τ mlTs
l∈L
 2π
= hdn (l)e−j M ml + 0
l∈L
 2π  2π
= hdn (l)e−j M ml + 0 × e−j M mk
l∈L k∈M\L
√ 2π

= MHnd ej M m
√
 MHnd (m) , (3.67)

in which Hnd (m) denotes √ the mth bin of the M-length DFT of a discrete-time equivalent channel
hdn (l). The scaling factor M appeared since we are working with the normalized version of the
DFT. In general, we assume that the number of subcarriers M employed in CP-OFDM systems
is larger than or equal to the length L + 1 of the channel model, i.e., we assume that M > L. This
means that we may have to insert some zeros at the end of the channel model hdn (l) before computing
the M-length DFT Hnd (m). Such insertion is also known as zero-padding.16
It is worth mentioning that the notation here is “heavier” than the notation in Chapter 2
since in this chapter we are working with time-varying continuous and discrete variables. So, we
shall keep in this chapter the dependency on the index of the OFDM symbol denoted by n and we
shall use the superscript “d” for the discrete-time model of the channel hdn (l), instead of the simpler
notation h(l).
Using the results presented in Subsections 3.3.1, 3.3.2, and 3.3.3, we can now clearly connect
the discrete-time implementation of CP-OFDM described in Chapter 2, Subsection 2.4.1 and the
CP-OFDM derived through a discretization of the analog OFDM.

3.3.4 BLOCK-BASED MODEL

This section describes a block-based model of the CP-OFDM transceiver, based on the results
derived in the previous subsections. Such a model is exactly equal to the model previously presented
16This does not have connections with the zero-padding (ZP) type of guard interval that will be explained in Subsection 3.4.2.
 3.3. DISCRETE-TIME OFDM 83
in Equations (2.44) and (2.45) of Chapter 2, except for a scaling factor which appears due to the
definition of analog OFDM. In that chapter we have proved that those choices of transmitter
and receiver matrices are quite useful since they transform the channel matrix into a circulant one,
which in turn can be diagonalized using DFT matrices. Such approach is very insightful from the
mathematical viewpoint, but it may lack from physical meaning.
Our objective here is to show the structure of the transmitter and receiver matrices, high-
lighting that such simple structures are natural consequences of the sampling process implemented
based on a physically meaningful analog OFDM version.
Firstly, consider the transmitter side. From what we have shown in Subsection 3.3.2, we can
form a block of elements u(n) ∈ CN ×1 representing the nth extended OFDM symbol, as follows:
 T
u(n)  un,(−K) · · · un,(−1) un,0 · · · un,(M−1)
 T
= un,(M−K) · · · un,(M−1) un,0 · · · un,(M−1)
⎡ ⎤
  un,0
0 I ⎢ .. ⎥
= K×(M−K) K ⎣ . ⎦
IM
   un,(M−1)
ACP
⎡ ⎤
1 IDFT {sm (n)}0
M ⎢ .. ⎥
= ACP ⎣ . ⎦
T −τ
IDFT {sm (n)}(M−1)
⎡ ⎤
1 s0 (n)
M H ⎢ .. ⎥
= ACP WM ⎣ . ⎦
T −τ
sM−1
  
s(n)
1
M
= H
ACP WM s(n), (3.68)
T −τ
where WM is the unitary M × M DFT matrix defined in Equation (2.46). There is a clear resem-
blance between the former equation and Equation (2.44) of Chapter 2. Note that this structure
appeared in a rather natural way as a consequence of the stacking of the related variables.
Secondly, consider the receiver side. If we define the matrix containing the channel-attenuation
factors (n) ∈ CM×M and the vector comprised of the received elements y(n) ∈ CN ×1 , with N =
M + K, as
(n)  diag{λm (n)}m∈M , (3.69)
 T
y(n)  yn,(−K) · · · yn,(−1) yn,0 · · · yn,(M−1) , (3.70)
then
 T
(n)s(n) = λ0 (n)s0 (n) · · · λM−1 (n)sM−1 (n) , (3.71)
84 3. OFDM
and, therefore, we can rewrite Equation (3.49) as follows:
1
T −τ ! " ! " T
(n)s(n) = DFT yn,k 0 · · · DFT yn,k M−1
M

⎡ ⎤
1 yn,0
T −τ ⎢ .. ⎥
= WM ⎣ . ⎦
M
yn,(M−1)
1
T −τ  
= WM 0M×K IM y(n)
M   
RCP
1
T −τ
= WM RCP y(n). (3.72)
M

Hence, we can define an estimate for s(n) in the following manner:

1
T −τ
ŝ(n) = E(n)WM RCP y(n), (3.73)
M

where the equalizer matrix E(n) is an M × M diagonal matrix containing the scaling factors corre-
sponding to each one of the M subcarriers. Once again, it is clear the similarities between the former
expression and Equation (2.45) of Chapter 2. In addition, as pointed out in Subsection 3.3.2, one
can choose

1
[E(n)]mm = , (3.74)
λm (n)

for the ZF solution, or

λ∗m (n)
[E(n)]mm = , (3.75)
σv2 (n)
|λm (n)|2 + σs2 (n)

for the MMSE solution (considering the presence of an additive noise signal at the receiver front-
end).
As mentioned before, the OFDM discussed here is the so-called cyclic prefix OFDM (CP-
OFDM), whose block diagram is depicted in Figure 3.10. For the CP-OFDM the equalizer consists
of a scalar correction performed at each DFT output to compensate for the channel distortion
λm (n) at the mth frequency bin. The digital channel appearing in Figure 3.10 can also be modeled
within the block-based framework by simply rewriting Equation (3.63) as follows (compare with
 3.3. DISCRETE-TIME OFDM 85
Transmitter

s0(n)
s1(n) I
s2(n) D P Digital
···

···
F S Cyclic Channel
T Prefix
sM −2(n)
sM −1(n)

Equalizer

ŝ0(n)
ŝ1(n)
S D ŝ2(n)
F
···

···
Remove
Prefix
P T
ŝM −2(n)
ŝM −1(n)

Receiver

Figure 3.10: OFDM digital transceiver with CP.

Equations (2.38) and (2.39) of Chapter 2):

⎡ ⎤
hdn (0) 0 0 ··· 0
⎢ hdn (1) h d (0) 0 ··· 0 ⎥
⎢ n ⎥
⎢ .. .. .. .. .. ⎥
⎢ . . . . . ⎥
⎢ ⎥
⎢ d . ⎥
y(n) = ⎢h (L) hd (L − 1) .. ··· 0 ⎥ u(n)
⎢ n n ⎥
⎢ 0 d
hn (L) ··· ··· 0 ⎥
⎢ ⎥
⎢ . . .. .. .. ⎥
⎣ .. .. . . . ⎦
0 0 hdn (L) · · · hdn (0)
  
HISI (n)
86 3. OFDM
⎡ ⎤
0 ··· 0 hdn−1 (L) · · · hdn−1 (1)
⎢ .. .. ⎥
⎢0 0 ··· 0 . . ⎥
⎢ ⎥
⎢ .. .. .. .. .. ⎥
⎢. . . . . hn−1 (L)⎥
d
⎢ ⎥
+ ⎢0 0 0 ··· 0 0 ⎥ u(n − 1)
⎢ ⎥
⎢0 0 0 ··· 0 0 ⎥ ⎥
⎢
⎢. .. .. .. .. .. ⎥
⎣ .. . . . . . ⎦
0 0 0 0 ··· 0
  
HIBI (n)
= HISI (n)u(n) + HIBI (n)u(n − 1), (3.76)

where we use the fact that hdn (l) is a causal system with order L, which implies that hdn (l) = 0
whenever l ∈ Z \ L.
As pointed out in Subsection 2.4.1 of Chapter 2, the effect of matrices ACP and RCP in the
former block-based description is to eliminate the IBI while turning the resulting effective channel
matrix into a circulant matrix. Due to the use of IDFT and DFT at the transmitter and receiver
sides, respectively, the overall block-based system reduces to a set of M uncoupled subchannels whose
mth received signal is given as (see also Equation (2.58) of Chapter 2)

λm (n)sm (n) + vm (n), (3.77)

in which we include the effect of additive noise modeled by vm (n). The parameter λm (n) is the so-
called channel tap associated with the mth subchannel of the nth transmitted block. This uncoupled
subchannel model will be very useful in the description of DMT in Subsection 3.4.4.

Example 3.1 (Block-Based Model) Design a CP-OFDM communication system with four
subcarriers and a fixed noiseless channel with transfer function

Hnd (z) = 1 + 0.1z−1 . (3.78)

For simplicity reasons, consider that T − τ = M.

Solution. Firstly, observe that L = 1 and M = 4. We can therefore choose K = 1, yielding N =

M + K = 5. Hence, the received signals are given as
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤⎡ ⎤
yn,(−1) 1 0 0 0 0 un,(−1) 0 0 0 0 0.1 u(n−1),(−1)
⎢ yn,0 ⎥ ⎢0.1 1 0 0 0⎥ ⎢ un,0 ⎥ ⎢0 0 0 0 0⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥ ⎢ u(n−1),0 ⎥
⎢ yn,1 ⎥ = ⎢ 0 0.1 1 0⎥ ⎢ ⎥ ⎢ 0 0⎥ ⎢ ⎥
⎢ ⎥ ⎢ 0 ⎥ ⎢ un,1 ⎥ + ⎢0 0 0 ⎥ ⎢ u(n−1),1 ⎥
⎣ yn,2 ⎦ ⎣ 0 0 0.1 1 0 ⎦ ⎣ un,2 ⎦ ⎣0 0 0 0 0 ⎦ ⎣ u(n−1),2 ⎦
yn,3 0 0 0 0.1 1 un,3 0 0 0 0 0 u(n−1),3
 3.3. DISCRETE-TIME OFDM 87
⎡ ⎤
1 0 0 0 0 ⎡ ⎤
⎢0.1 1 ⎥ un,0
0 0 0⎥  
⎢ 01×3 1 ⎢ ⎥
=⎢ 0 0.1 1 0 0⎥ ⎢un,1 ⎥
⎢ ⎥ I4 ⎣ un,2 ⎦
⎣0 0 0.1 1 0⎦
un,3
0 0 0 0.1 1
⎡ ⎤
0 0 0 0 0.1 ⎡ ⎤
⎢0 ⎥ u(n−1),0
⎢ 0 0 0 0 ⎥ 
01×3 1 ⎢⎢u(n−1),1 ⎥
⎥
+⎢
⎢0 0 0 0 0⎥ ⎥ ⎣
⎣0 I4 u(n−1),2 ⎦
0 0 0 0⎦
u(n−1),3
0 0 0 0 0
⎡ ⎤ ⎡ ⎤
0 0 0 1 ⎡ ⎤ 0 0 0 0.1 ⎡ ⎤
⎢1 u
⎥ n,0 ⎢0 ⎥ u(n−1),0
⎢ 0 0 0.1 ⎥ ⎢un,1 ⎥ ⎢ 0 0 0 ⎥⎢ ⎥
=⎢ 0⎥ ⎢ ⎥ ⎢ 0 0⎥ ⎢u(n−1),1 ⎥
⎢0.1 1 0 ⎥ ⎣un,2 ⎦ + ⎢0 0 ⎥ ⎣u(n−1),2 ⎦ . (3.79)
⎣ 0 0.1 1 0 ⎦ ⎣0 0 0 0⎦
un,3 u(n−1),3
0 0 0.1 1 0 0 0 0

The first operation at the receiver is the removal of the redundancy through the multiplication
by RCP = [04×1 I4 ]. This operation is equivalent to discarding the first element, yn,(−1) , of the
received data block, thus removing the effect of the IBI and yielding the following description:

⎡ ⎤ ⎡ ⎤⎡ ⎤
yn,0 1 0 0 0.1 un,0
⎢yn,1 ⎥ ⎢0.1 1 0 0⎥ ⎢ ⎥
⎢ ⎥ =⎢ ⎥ ⎢un,1 ⎥
⎣yn,2 ⎦ ⎣ 0 0.1 1 0 ⎦ ⎣un,2 ⎦
yn,3 0 0 0.1 1 un,3
⎡ ⎤ ⎡ ⎤
1 0 0 0.1 s0 (n)
⎢0.1 1 0 0⎥ ⎢ ⎥
=⎢ ⎥ W H ⎢s1 (n)⎥
⎣ 0 0.1 1 0 ⎦ 4 ⎣ s2 (n)⎦
0 0 0.1 1 s3 (n)
⎡ ⎤⎡ ⎤⎡ ⎤
1 0 0 0.1 1 1 1 1 s0 (n)
1 ⎢0.1 1 0 0⎥ ⎢ ⎥⎢
⎥ ⎢1 j −1 −j ⎥ ⎢s1 (n)⎥ .
⎥
= ⎢⎣ ⎦ ⎣ ⎦ ⎣ (3.80)
2 0 0.1 1 0 1 −1 1 −1 s2 (n)⎦
0 0 0.1 1 1 −j −1 j s3 (n)
88 3. OFDM
Now, by taking the DFT of the resulting signal, one has
⎡ ⎤ ⎡ ⎤⎡ ⎤⎡ ⎤⎡ ⎤
yn,0 1 1 1 1 1 0 0 0.1 1 1 1 1 s0 (n)
⎢yn,1 ⎥ 1 ⎢1 −j −1 j ⎥ ⎢0.1 1 0⎥ ⎢ ⎥⎢ ⎥
W4 ⎢ ⎥ ⎢ ⎥⎢ 0 ⎥ ⎢1 j −1 −j ⎥ ⎢s1 (n)⎥
⎣yn,2 ⎦ = 4 ⎣1 −1 1 −1⎦ ⎣ 0 0.1 1 0 ⎦ ⎣ 1 −1 1 −1 ⎦ ⎣ s2 (n)⎦
yn,3 1 j −1 −j 0 0 0.1 1 1 −j −1 j s3 (n)
⎡ ⎤⎡ ⎤
1.1 0 0 0 s0 (n)
⎢ 0 1 − 0.1j 0 0 ⎥ ⎢s1 (n)⎥
=⎢
⎣0
⎥⎢ ⎥, (3.81)
0 0.9 0 ⎦ ⎣s2 (n)⎦
0 0 0 1 + 0.1j s3 (n)
leading to an uncoupled model with four subchannels. 2

3.4 OTHER OFDM-BASED SYSTEMS

In practice, there are several distinct ways of configuring the block transceivers that are closely related
to the CP-OFDM discussed in Section 3.3. In this section, we shall present some of them, starting
from the single-carrier with frequency-domain equalization (SC-FD) in Subsection 3.4.1. This
type of transceiver is able to significantly reduce the PAPR as compared to the CP-OFDM, a very
desirable feature from the implementation viewpoint. In addition, we will also describe the so-called
ZP-OFDM in Subsection 3.4.2, which employs zero-valued elements in the guard interval, instead
of a cyclic prefix. This saves transmitter power, a valuable resource in many applications. Another
very useful system in practice is the so-called coded OFDM (C-OFDM), which combines the good
features inherent to OFDM system with the additional protection enabled by error-correcting codes.
In this context, Subsection 3.4.3 gives a brief introduction to C-OFDM systems. The section ends
with Subsection 3.4.4, which describes a standard manner of taking into account the knowledge
of the channel state at the transmitter side in order to enhance the related data rates. The discrete
multitone (DMT) systems are able to take into account such side information about the channel by
using a technique called water-filling.

3.4.1 SC-FD
As previously discussed in Subsection 3.2.4, the OFDM symbol presents high PAPR imposing
a high linearity constraint in the transmitter power amplifier. In the case where the power of the
transmitted signal has high dynamic range, it is possible that the nonlinear behavior of the transmitter
amplifier introduces distortions to the transmitted OFDM symbol, leading to loss of orthogonality
among the subcarriers. The deterministic PAPR of a discrete-time OFDM symbol can be defined as
! "
max |un,k |2
k
PAPR OFDM (n)  1 2 , (3.82)
M |un,k |2
k∈M
 3.4. OTHER OFDM-BASED SYSTEMS 89
! "
in which max |un,k denotes the peak power level associated with the nth OFDM symbol, whereas
|2
1 2
k
M |un,k | denotes the average power of this nth block. In the case of CP-OFDM, the insertion
2
k∈M
of the guard interval does not change the maximum achievable power (peak) of the elements that
compose an OFDM symbol. In addition, it is very unlikely that the guard interval could change
the average power as well. Therefore, we can consider only the elements after the application of the
IDFT, without considering the cyclic prefix which is inserted. Thus, based on Equation (3.55), one
has
1  2π
un,k = √ sm (n)ej M mk , (3.83)
T − τ m∈M
yielding (see also Equation (3.42))
1  1 
|un,k |2 = un,k u∗n,k
M M
k∈M k∈M  
1  1  
∗ j 2π )k
= sm (n)sm (n)e M (m−m
M T −τ
k∈M m∈M m ∈M  
1   1  2π
∗
= sm (n)sm (n) ej M (m−m )k
T −τ M
m∈M m ∈M k∈M
1  
= sm (n) sm (n)δ[m − m ]
∗
T −τ
m∈M m ∈M
1 
= |sm (n)|2
T −τ
m∈M
 
M 1 
= |sm (n)| ,2
(3.84)
T −τ M
m∈M
and
3 3
3 3
1 3 2π 3
|un,k | = √ 3 sm (n)ej M mk 3
3
T − τ m∈M 3
 33 3
1 2π 3
≤√ 3sm (n)ej M mk 3
T − τ m∈M

1 
=√ |sm (n)|
T − τ m∈M
1 
≤√ max {|sm (n)|}
T − τ m∈M m
M
=√ max {|sm (n)|} , (3.85)
T −τ m
90 3. OFDM
thus implying that

M2  
max |un,k |2 ≤ max |sm (n)|2 . (3.86)
k T −τ m
The results of Equations (3.84) and (3.86) give us that
⎛ ! "⎞
max |sm (n)| 2
⎜ m ⎟
PAPR OFDM (n) ≤ M ⎝ 1 2 ⎠
M |sm (n)| 2
m∈M
  
PAPR SC (n)
= MPAPR SC (n), (3.87)

where SC stands for single-carrier transmissions, i.e., no pre-processing is implemented at the

transmitter side. The result of Equation (3.87) means that the upper bound for the PAPR related
to OFDM systems is M times larger than the corresponding PAPR of a single-carrier system.
In practice, it is observed that this is a tight upper bound, which means that the actual value of
PAPR OFDM (n) is close to MPAPR SC (n). In many applications, M is made large in order to cope
with the loss of spectral efficiency due to the redundancy insertion (guard interval), which might
turn the PAPR problem even more critical [45].17
A possible solution to the PAPR issue is to perform the block transmission in a single-carrier
fashion while performing the equalization in the frequency domain, as illustrated in Figure 3.11.
In this configuration the cyclic prefix is inserted at the transmitter end, whereas the IDFT is im-
plemented at the receiver in order to allow a decoupled equalization, which is represented by the
multipliers between the IDFT and DFT blocks in Figure 3.11. This block transmission setup is
called single-carrier with frequency-domain equalization (SC-FD) transceiver. Since it uses a cyclic
prefix, we will call it cyclic prefix SC-FD (CP-SC-FD), as also described in Subsection 2.4.3.

3.4.2 ZP-BASED SCHEMES

The zero-padding (ZP) represents an alternative to the use of CP as guard time. ZP is a quite effective
way to eliminate the IBI that pervades block-based transmissions. Indeed, in several different setups,
ZP systems are optimal solutions in the MSE sense [83]. This optimality characteristic leads to
better performance of ZP-based transceivers, as compared to CP-based systems in a number of
situations [55].
In comparison with CP-based schemes, one of the main advantages of ZP-based transceivers
is their lower required power due to their trailing zeros. On the other hand, ZP-based transceivers
are usually able to eliminate only IBI effects, whereas the ISI effects should be mitigated by a
17 Another definition for the PAPR considers the OFDM symbol as a random sequence and takes the expected value of that sequence
instead of the deterministic average value. However, the results and conclusions are the same of those presented here. See, for
instance, [45] for further details.
 3.4. OTHER OFDM-BASED SYSTEMS 91
Transmitter

s0(n)
s1(n)
s2(n) P Digital

···
S Cyclic Channel
Prefix
sM −2(n)
sM −1(n)

Equalizer

ŝ0(n)
ŝ1(n)
D I
S D ŝ2(n)
F

···
···

···
Remove F
Prefix
P T T ŝM −2(n)
ŝM −1(n)

Receiver

Figure 3.11: SC-FD digital transceiver with CP.

properly designed equalizer. Thus, in general, ZP-based transceivers do not yield so simple (one-
tap) equalizers as CP-based systems do. These observations are valid for the general ZP versions
of the OFDM and SC-FD, herein called ZP-OFDM and ZP-SC-FD, respectively. In fact, there
are some tools related to structured matrix representations which can work around these difficulties,
yielding general ZP-transceivers which are still based on DFTs and one-tap equalizers (see Figure 4.3
of Chapter 4). However, there are particular versions of ZP-OFDM and ZP-SC-FD that have the
same simple equalizers as the ones used in CP-OFDM and CP-SC-FD systems. These particular
versions, known as ZP-OFDM-OLA and ZP-SC-FD-OLA, perform overlap-and-add (OLA)
operations at the receiver side and they are depicted in Figures 3.12 and 3.13. The mathematical
details regarding these particular versions of ZP-OFDM and ZP-SC-FD were already explained in
Subsections 2.4.2 and 2.4.3, respectively.
The following topics are research results concerning the differences between the general ZP-
OFDM and ZP-SC-FD and their cyclic-prefix counterparts [55, 87]:

• ZP-OFDM introduces more nonlinear distortion than CP-OFDM transceivers.

92 3. OFDM
Transmitter

s0(n)
s1(n) I
s2(n) D P Digital

···

···
F S Zero Channel
T Padding
sM −2(n)
sM −1(n)

Equalizer

ŝ0(n)
ŝ1(n)
Post S D ŝ2(n)
··· F

···
Processing
P T
ŝM −2(n)
ŝM −1(n)

Receiver

Figure 3.12: OFDM digital transceiver with ZP.

• ZP-OFDM has better performance in terms of BER or MSE than CP-OFDM, for a given
average-bit-energy-to-noise power ratio, Eb /N0 .

• The ZP-SC-FD has lower PAPR, presents robustness to CFO, and has also better uncoded
performance. However, the equalization is a bit more complex to implement.

• In the case some kind of channel coding is included (C-OFDM), the coded version is better
when code rate is low and the error correcting coding capability is enhanced. In the coded case
ZP-SC-FD is better than C-OFDM for high code rate.

• ZP-SC-FD only has clear performance advantages over uncoded OFDM.

• Uncoded CP-OFDM is inferior to zero-forcing equalized CP-SC-FD transceiver.

• CP-OFDM with equal-gain power allocation has the same performance as zero-forcing equal-
ized CP-SC-FD transceiver.
 3.4. OTHER OFDM-BASED SYSTEMS 93
Transmitter

s0(n)
s1(n)
s2(n) P Digital

···
S Zero Channel
Padding
sM −2(n)
sM −1(n)

Equalizer

ŝ0(n)
ŝ1(n)
I
Post S D
D ŝ2(n)
F

···
···

···
Processing
P F
T T ŝM −2 (n)
ŝM −1 (n)

Receiver

Figure 3.13: SC-FD digital transceiver with ZP.

• CP-OFDM with AMBER (an approximately minimum BER) power allocation is better than
zero-forcing equalized CP-SC-FD.

3.4.3 CODED OFDM

The usual complete coded OFDM (C-OFDM) system utilized in some broadcast systems is depicted
in Figure 3.14 where two levels of coding and an interleaving are employed. These building blocks
are required to protect the transmitted information against distortions and deep fadings in some
subcarriers [13, 20]. C-OFDM includes a few building blocks as described in [13].
The outer encoder (coder 1) is meant to insert redundancy in the data stream, e.g., using
Reed-Solomon codes [38, 90]. This will increase the required bandwidth for transmission, but on
the other hand, allows more reliable reception. The aim is either to obtain block codes with increased
Hamming distances or to encode the data in a continuous way such that a correlation between the
data is induced to help detection at the receiver end.
The inner encoder (coder 2) is meant to increase the Euclidean distance among the symbols of
the constellation, usually employing a trellis coded modulation. The idea is to increase the number of
possible symbols in comparison with the number of points in the given constellation and subdivide
94 3. OFDM
OFDM Transmitter
bitstream Outer Inner OFDM
Interleaver
Encoder Encoder Modulator

Channel

Outer Inner OFDM

Deinterleaver
bitstream Decoder Decoder Demodulator

OFDM Receiver

Figure 3.14: Coded OFDM (C-OFDM) system.

them in subsets with greater Euclidean distances, while maintaining the transmission energy. The
subsets and the points in the subsets are tied using a trellis diagram derived with the convolutional
code. If a Viterbi decoder [38] is used at the receiver some coding gain is expected [13].
Error bursts could occur if the inner decoder (decoder 2) chooses a wrong decoding path
which could then be corrected by the combination of the interleaver and the outer decoder (decoder
1).
In the final analysis the reason for coding is to provide a link among the symbols transmitted on
different subcarriers such that a symbol transmitted in a strongly faded subcarrier (i.e., λm (n) ≈ 0)
can be recovered at the receiver by estimating the symbols transmitted by other subcarriers. The
bottom line is that we are distributing and mixing the transmitted information to increase the
chance of proper detection at the receiver end.

3.4.4 DMT
A discrete multitone (DMT) transceiver is essentially an OFDM system comprised of three particular
features: (i) there is no passband conversion to a higher carrier frequency, which means that the actual
transmitted signals are baseband signals; (ii) since any actual transmission employs real-valued signals,
then the baseband transmitted signals must be real-valued. This means that the input constellation
∗
symbols sm (n) must have the conjugate symmetric property, i.e., sm (n) = s(M−m) (n), for all m ∈ M;
and (iii) there is some kind of channel-state information (CSI) at the transmitter side, so that the
transceiver can use some smart techniques in order to cope with possible channel impairments in
advance.
The third DMT property above is indeed its key feature since it enables transmissions with
higher data rates. Nevertheless, it is usually applicable in wired connections, in which the channel
state does not change too often. As described in Subsection 1.3.1 of Chapter 1, the DMT system is
currently employed in many digital subscriber line (xDSL) applications. The aim of this subsection
is to describe how DMT-based systems take into account the availability of information about the
channel at the transmitter side in order to enhance the overall transmission performance.
As described in Subsection 3.3.4, OFDM systems can be thought as M parallel uncou-
pled subchannels, whose mth received signal of the nth transmitted OFDM symbol is given by
 3.4. OTHER OFDM-BASED SYSTEMS 95
λm (n)sm (n) + vm (n), in which we consider that sm (n) is a random sequence with zero-mean and
2 (n), whereas v (n) is a random sequence with zero-mean and variance σ 2 (n), where
variance σs,m m v,m
m ∈ M and n ∈ Z. In this case, we have the following signal-to-noise ratio (SNR) associated with
the mth subchannel:
|λm (n)|2 σs,m
2 (n)
SNR m (n)  2 (n)
. (3.88)
σv,m
If vm (n) is a Gaussian random sequence, then it follows from basic knowledge on digital
communications that, for each fixed pair of numbers (m, n) ∈ M × Z, one has the following error
probability of symbols [22, 26, 45, 66, 83]:
+

Pe,m (n) ≈ cQ A SNR m (n) , (3.89)

in which Q(·) is the area under a Gaussian tail, being a decreasing function of its argument defined
as
 ∞
1
e−w /2 dw,
2
R
x → Q(x)  √ (3.90)
2π x
whereas c and A depend upon the number of bits bm (n) ∈ N employed to represent sm (n). Large
number of bits leads to both relatively large values of c and small values of A. In the particular case
of a 2bm (n) -QAM constellation, for instance, one has [83]
bm (n)
c = 4(1 − 2− 2 ), (3.91)
1
3
A= . (3.92)
2bm (n) − 1
This means that, for sm (n) pertaining to a 2bm (n) -QAM constellation, one can rewrite Equa-
tion (3.89) as follows:
# $
1 −1 Pe,m (n)
A≈ √ Q , (3.93)
SNR m (n) c
yielding (see Equations (3.91) and (3.92))
⎛ ⎞
⎜ SNR m (n) ⎟
bm (n) ≈ log2 ⎝1 +
2 ⎠
P (n)
Q−1 e,mc /3
# $
SNR m (n)
= log2 1 + , (3.94)
QAM
where the number QAM is the so-called SNR gap associated with the QAM constellation, being
defined as

2
Q−1 Pe,mc (n)
QAM  . (3.95)
3
96 3. OFDM
Strictly speaking, Equation (3.94) is not a closed-form expression for bm (n) since the definition
of c in Equation (3.92) depends on bm (n), thus implying that the SNR gap is a function of the number
of bits itself. Nevertheless, the dependency of QAM on bm (n) is quite weak in the sense that QAM
varies slowly when the number of bits changes. For instance, considering an error probability of
symbols of Pe,m (n) = 10−5 , then QAM ∈ [6.5, 6.9] when bm (n) varies from 2 to 8. Besides, when
  2
bm (n) is large, then QAM ≈ Q−1 Pe,m (n)/4 /3, i.e., there is no dependence on bm (n).
The important conclusion that we can draw from Equation (3.94) is that power and bit alloca-
tions are closely related problems. Indeed, for a predefined target error probability of symbols, large
values of SNR allow increasing the number of bits following the rule expressed in Equation (3.94).
Increasing the number of transmitted bits without sacrificing the error-probability performance is a
rather desirable feature of a transmission scheme. On the other hand, we cannot increase the SNR
as we wish since this may entail an unrealistic transmission power.
So far, we are working only with one predefined subchannel index m ∈ M. When all sub-
channels are considered, then we can work with the average number of bits b(n) ∈ R in an OFDM
symbol, which is the sum of the number of bits transmitted on each subcarrier divided by the total
number of subcarriers M, i.e.,
# $
1  SNR m (n)
b(n) ≈ log2 1 +
M QAM
m∈M
 $ M1
: # SNR m (n)
= log2 1+
QAM
#m∈M $
SNR(n)
= log2 1 + , (3.96)
QAM
where SNR(n) is defined as
:  1
SNR(n)  QAM + SNR m (n) M − QAM , (3.97)
m∈M

which means that SNR(n) is basically the geometric mean of the sum of the SNR gap with the mth
subchannel SNR, subtracting the SNR gap from the final result.
Thus, the parallel and independent subchannels together behave as a single white Gaussian
noise channel with SNR corresponding to SNR(n), which is essentially equal to the geometric mean
of the subcarrier SNRs increased by the SNR gap. This geometric mean is a good approximation
for moderate to high SNRs on all subcarriers. As the geometric mean is much more sensitive to
small values than the arithmetic mean,18 then the existence of poor subchannels (i.e., subchannels
with low SNRs) can significantly degrade the system performance in terms of error probability of
symbols. An alternative view is that, for a given error-probability performance, the presence of poor
subcarriers can significantly decrease the transmission data rates.
18 In fact, the geometric mean of positive real numbers is always smaller than or equal to the arithmetic mean of these numbers.
 3.4. OTHER OFDM-BASED SYSTEMS 97
The former reasoning clarifies how the quality of the subchannels affects the overall system
performance, since the SNR of the mth subchannel is affected directly by the channel gain λm (n)
(see Equation (3.88)). Standard uncoded OFDM-based systems suffer critically from performance
degradations caused by deep fadings in some subchannels. In order to appreciate the advantages
of the DMT over standard OFDM-based systems, the SNR(n) must be compared with unbiased
SNR of equalized baseband systems. In practice, SNR(n) can be improved considerably when the
available energy is distributed non-uniformly over all, or a subset of parallel carriers, giving rise
to high-performance multicarrier systems. This energy allocation is implemented by the so-called
loading algorithms, whose most notorious example is the water-filling approach. Let us formulate
this approach from now on.
From information theory, it is well known that the ultimate transmission data rate is dictated
by the channel capacity [14]. Indeed, the channel capacity is the upper bound on the possible data
rates which any communication system is able to achieve, assuring that error probability of symbols
is zero. It is possible to show that the channel capacity (in bits/transmission, i.e., channel capacity
normalized by the channel bandwidth) of the mth Gaussian subchannel of a CP-OFDM system is
given by (in the case of complex-valued variables) [14]

Cm (n) = log2 (1 + SNR m (n)) , (3.98)

implying that the average channel capacity, C(n), considering all M subchannels of the nth data
block, is

1 
C(n) = log2 (1 + SNR m (n)) . (3.99)
M
m∈M

This expression is quite similar to the average amount of transmitted bits expressed in Equa-
tion (3.96).
Therefore, in order to increase the transmission data rates one could try to compensate for
the channel taps λm (n) leading to low the subchannel SNRs, and an intuitive solution to this
problem is maximizing the average channel capacity in expression (3.99) subject to a predefined
power constraint. It is worth mentioning at this point that, although a transmission scheme that
achieves the related capacity could be theoretically implemented by using codes with a long block
length [14], practical communications systems avoid high complexity codes, thus implying that the
capacity may not be achievable. In other words, for a given small average error probability of symbols
Pe (n), the channel code might require infinite complexity to achieve the capacity. In practice, at a
given fixed error probability of symbols, one can characterize the modulation/coding by a gap  ∈ R+
(for example,  = QAM in Equation (3.95)) that quantifies the effective loss in SNR with respect
to capacity. Extremely powerful codes (such as turbo code and concatenated code) can reduce the
gap to up 1 or 2 dB.
98 3. OFDM
Let us formulate the problem of maximizing the average channel capacity (or minimizing
−C(n) expressed using the natural logarithm) as the following optimization problem:
,  -
1  |λm (n)|2 σs,m
2 (n)
min − ln 1 + 2 (n)
, (3.100)
σs,m (n)∈R M σv,m
m∈M
m∈M
subject to:

2
σs,m (n) = p0 (n), (3.101)
m∈M

in which the positive real number p0 (n) denotes the total transmission power of the nth block. In
2 (n) and λ (n) are fixed and known numbers, for
this optimization problem, it is assumed that σv,m m
all m ∈ M.
By applying the Lagrange-multiplier method (with Lagrange multiplier λ ∈ R), we have the
following cost-function:
 
  1  |λm (n)|2 σs,m
2 (n)
J σs,0 (n), · · · , σs,(M−1) (n)  − ln 1 + 2 (n)
M σv,m
m∈ M 

+λ σs,m (n) − p0 (n) ,
2
(3.102)
m∈M

which can be optimized by finding its associated extreme points, as follows:

 
∂J σs,0 (n), · · · , σs,(M−1) (n) 2 σs,m (n)
=− 2 (n)
+ 2λσs,m (n) . (3.103)
∂σs,m (n) M σv,m + σ 2 (n)|λm (n)|2 s,m

Thus, for each m ∈ M, we have

 
∂J σs,0 (n), · · · , σs,m
o (n), · · · , σ
s,(M−1) (n)
=0 (3.104)
∂σs,m (n)
o (n) is such that σ o (n) = 0 or (discarding the negative root)
if and only if the optimal value σs,m s,m
.
1 2 (n)
σv,m
o
σs,m (n) = − . (3.105)
Mλ |λm (n)|2

Note that, based on Equation (3.103), when σs,mo (n)  = 0, one has λ  = 0, which means that 1 is
λ
well defined.19
o (n), with m ∈ M, are
Now, we can consider, without loss of generality, that the numbers σs,m
ordered as follows:
o o o
σs,0 (n) ≥ σs,1 (n) · · · ≥ σs,(M−1) (n). (3.106)
19 We shall omit the proof that those optimal values correspond to solutions that minimize the original objective function.
 3.4. OTHER OFDM-BASED SYSTEMS 99
In addition, let us assume that we will transmit using M
subchannels, where M
− 1 ∈ M is the

largest index of the smallest value σs,m (n), with m ∈ M. Note that M ≤ M and that
o

o o
σs,0 (n) ≥ · · · ≥ σs,(M −1) (n) > 0. (3.107)

Thus, by defining M  {0, · · · , M − 1}, the constraint in Equation (3.101) can be used to find
the Lagrange multiplier, as follows:
 

2  1
2 (n)
σv,m
o
σs,m (n) = −

Mλ |λm (n)|2
m ∈M m ∈M
M 1  σv,m
2 (n)

= −
M λ |λm (n)|2
m ∈M
= p0 (n), (3.108)

yielding
 
1 1  σv,m
2 (n)

= p0 (n) + . (3.109)
Mλ M |λm (n)|2
m ∈M

Therefore, according to Equation (3.105), the variances of the first M ≤ M constellation

symbols of the nth block are given by
 
 o 2 1  σv,m 2 (n)

2 (n)
σv,m
σs,m (n) = p0 (n) + − , (3.110)
M
|λm (n)|2 |λm (n)|2
m ∈M
 o 2
with m ∈ M , whereas the remaining M − M symbols do not transmit data, i.e., σs,m (n) = 0
for m ∈ M \ M . In other words, we can state that
⎧ 2 (n)

⎪ 2 σv,m
 o 2 ⎨ p0 (n)+m ∈M |λm (n)|2 σ 2 (n)
σs,m (n) = − |λv,m(n)|2 , if this quantity is larger than zero, (3.111)
⎪
⎩ M m
0, otherwise.

σ 2 (n)  o 2
Note that large values of |λv,m(n)|2 leads to low values of σs,m (n) , which means that less power
m
is allocated to poor subchannels. The solution in expression (3.111) is the so-called water-filling. It
essentially says that the power allocation among the subcarriers is implemented in such a way that
 o 2 σ 2 (n)
σs,m (n) + |λv,m(n)|2 is a constant and that occasionally it will not be distributed any energy to some
m
of the last subcarriers due to the power constraint. The name water-filling is motivated by the way
it can be interpreted, namely, as communicating vessels which are filled with a homogeneous liquid
that achieves the same level in all containers, regardless of their volume and/or shape.
100 3. OFDM

Power
1
4λ
 o 2
σs,1 (n)
 o
2
σs,0 (n)

2 2
σv,2 (n) σv,3 (n)
|λ2 (n)|2 |λ3 (n)|2
2
σv,1 (n)
|λ1 (n)|2 2
σv,0 (n)
|λ0 (n)|2

1 2 0 3 Channel

Figure 3.15: Discrete water-filling for four subchannels.

Figure 3.15 illustrates pictorially a water-filling solution for the case of four subchannels (i.e.,

2
2
M = 4) where the total energy distributed is p0 (n) = σs,0 o (n) + σs,1o (n) . The subchannels

with indexes 3 and 4 did not receive any power (i.e., M = 2) since the values of their corresponding
2 (n)
σv,0 σ 2 (n)
2 (n)
σv,2 2 (n)
σv,3 p0 (n)+ 2 + v,1 2
|λ0 (n)| |λ1 (n)|
factors and
|λ2 (n)|2
were both higher than 4λ
|λ3 (n)|2
1
= 2 . Example 3.2 is a toy
example which help us verify the gains introduced by the water-filling approach.

Example 3.2 (Water-Filling) Given a time-invariant channel model with transfer function

Hnd (z) = 1 − z−2 , (3.112)

2π 2π
whose zeros are ej 8 0 = 1 and ej 8 4 = −1. Assume that the noise variance is given by σv,m
2 (n) = 0.08

for all m ∈ {0, 1, · · · , 7} and n ∈ Z. Considering eight subchannels, compute the channel capacity
for each subchannel, Cm (n), and average channel capacity, C(n), before and after applying the water-
filling approach. Assume two possible values of total transmission power, namely p0 (n) = 0.04 and
p0 (n) = 0.40. In addition, determine the power distribution for each case.
Solution. In order to compute the initial SNR (i.e., the SNR before applying the water-filling
2 (n),
optimization approach), one has to determine three quantities for each subchannel, namely σv,m
 3.4. OTHER OFDM-BASED SYSTEMS 101
2 (n), and |λ (n)|2 , as described in Equation (3.88). The noise variance is fixed at 0.08, while the
σs,m m
average power of the symbols is σs,m2 (n) = p (n)/8, i.e., σ 2 (n) = 0.04/8 = 0.005 or σ 2 (n) =
0 s,m s,m
0.40/8 = 0.05. As for |λm (n)| , one has (see Equation (3.67) or Equation (2.57) of Chapter 2)
2

3  d  332
3√ 
3 hn 3
|λm (n)|2 = 3 M WM 3 , (3.113)
3 05×1 m 3

where hnd  [1 0 − 1]T .

0.12 0.12

0.1 0.1

0.08 0.08
Power

Power

0.06 0.06

0.04 0.04

0.02 0.02

0 0
6 2 0 3 7 4 1 5 6 2 0 3 7 4 1 5
Subchannel Index Subchannel Index

(a) p − 0 = 0.04. (b) p − 0 = 0.40.

Figure 3.16: Water-filling example.

 o 2
The power distribution, the initial and optimized SNRs (the optimized SNR uses σs,m (n)
given in expression (3.111)), the initial and optimized capacity (see Equation (3.98)) for each subcar-
rier, and the average capacity (see Equation (3.99)) are summarized in the following table, considering
both transmission powers:
102 3. OFDM
Subchannel index 6 2 0 3 7 4 1 5
|λm (n)|2 0.0 2.0 4.0 2.0 0.0 2.0 4.0 2.0
2 (n)/|λ (n)|2
σv,m ∞ 0.04 0.02 0.04 ∞ 0.04 0.02 0.04
m
p0 (n) 0.04
2 (n)
Initial σs,m 0.005 0.005 0.005 0.005 0.005 0.005 0.005 0.005
Initial SNRm (n) 0.00 0.12 0.25 0.12 0.00 0.12 0.25 0.12
Initial Cm (n) 0.00 0.17 0.32 0.17 0.00 0.17 0.32 0.17
Initial C(n) 0.1654
Optimized σs,m2 (n) 0.00 0.00 0.02 0.00 0.00 0.00 0.02 0.00
Optimized SNRm (n) 0.00 0.00 1.00 0.00 0.00 0.00 1.00 0.00
Optimized Cm (n) 0.00 0.00 1.00 0.00 0.00 0.00 1.00 0.00
Optimized C(n) 0.2500
p0 (n) 0.40
2 (n)
Initial σs,m 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05
Initial SNRm (n) 0.00 1.25 2.50 1.25 0.00 1.25 2.50 1.25
Initial Cm (n) 0.00 1.17 1.81 1.17 0.00 1.17 1.81 1.17
Initial C(n) 1.0368
Optimized σs,m2 (n) 0.00 0.06 0.08 0.06 0.00 0.06 0.08 0.06
Optimized SNRm (n) 0.00 1.50 4.00 1.50 0.00 1.50 4.00 1.50
Optimized Cm (n) 0.00 1.32 2.32 1.32 0.00 1.32 2.32 1.32
Optimized C(n) 1.2414

A NOTE ABOUT OFDM SYSTEMS AND BEYOND

Before closing this chapter, let us comment on the OFDM-based schemes that we have just pre-
sented. There is a growing demand for transmission resources which shows no sign of settling. In
the case of wireless data services, for instance, it is possible to predict that spectrum shortage is a sure
event in the near future. As we mentioned before, the first step to address this problem is to choose
a modulation scheme that is efficient in terms of channel capacity in bits/transmission, particularly
in broadband transmissions where the channel presents a frequency-selective model. In such cases,
multicarrier communication systems is a very smart solution given that this modulation scheme is
efficient for data transmission through channels with moderate and severe ISI.
The most widely used and simplest multicarrier system is the OFDM (comprising its variants
presented in this chapter). Whenever the subchannels are narrow enough, it is possible to consider
that each subchannel-frequency range is flat, avoiding the use of sophisticated equalizers. In appli-
cations where there is a return channel, it is also possible to exploit the SNR in different subchannels
in order to select the subchannel-modulation order, so that subchannels with high SNR utilize
high-order modulation, whereas low-order modulation should be employed in subchannels with
moderate SNR. This loading strategy enables more effective usage of frequency-selective channels.
 3.5. CONCLUDING REMARKS 103
In OFDM, a serial data stream is divided into blocks where each block is appended with redundancy
(guard interval) in order to avoid IBI. The special features of the OFDM system are the elimination
of IBI, ISI, and its inherent low computational complexity. OFDM system utilizes IDFT at the
transmitter and DFT at the receiver, enabling the use of computationally efficient FFT algorithms.
The key strategy to avoid IBI in transmissions through frequency-selective channels is to include
some redundancy at the transmitter, thus reducing the data rate.
The SC-FD scheme has also emerged as an alternative solution to overcome some drawbacks
inherent to OFDM-based systems, such as PAPR and CFO. Also for some frequency-selective
channels, the BER of an SC-FD system can be lower than for the OFDM, particularly if some
subchannels have high attenuation. In this solution, the data stream is inserted with redundancy to
avoid IBI and at the receiver the equalization is performed in the frequency domain, keeping the
efficient equalization scheme inherited from OFDM.
A way to postpone the spectrum shortage is to increase the data throughput for a given
bandwidth by utilizing some smart technology. Assuming we are employing block transmission it is
worth discussing how to reduce the amount of redundancy in multicarrier system, while constraining
the transceiver to employ fast algorithms. A possible solution is to employ reduced-redundancy
transceivers utilizing DFTs and diagonal matrices. The reduced-redundancy transceiver allows for
higher data throughput than OFDM in a number of practical situations. A computationally efficient
solution for reduced-redundancy transceiver is presented in Chapter 4.

3.5 CONCLUDING REMARKS

In this chapter we addressed distinct aspects of OFDM systems, for example, continuous-time
(analog) OFDM, discrete-time OFDM, coded OFDM (C-OFDM), and DMT. We derived the
discrete-time OFDM as the sampled version of the analog OFDM and discussed the issue of ensur-
ing orthogonality among subcarriers in frequency-selective channels.The C-OFDM was introduced
as a discrete-time OFDM in which we include some channel encoding elements (inner and outer
coders, interleavers and/or scramblers) in order to protect the information to be transmitted. In
addition, the DMT can be seen as a discrete-time OFDM system in which the transmitter has
channel state information, and, therefore, it is capable of performing optimal power allocation, thus
increasing the channel capacity given a power constraint.
All these systems require the use of a guard period whose time extension is greater or equal
to the channel memory (or time delay spread, for the analog OFDM). Therefore, the guard period
τ increases with the channel memory τmem . As a consequence, for an OFDM symbol of a given
duration T , the useful symbol time T − τ decreases, thus reducing the system capacity. In other
words, when the durations of the OFDM symbol and the channel impulse response are of the same
order, the throughput of OFDM systems decreases. In order to address this issue, the next chapter
describes systems whose computational complexity are comparable to the OFDM complexity, with
the advantage of employing a reduced number of redundant elements, that is, their guard time is
shorter.
 105

CHAPTER 4

Memoryless LTI Transceivers

with Reduced Redundancy
4.1 INTRODUCTION
Multicarrier-modulation methods are of paramount importance to many data-transmission systems
whose channels induce severe or moderate intersymbol interference (ISI). As previously discussed,
the key idea behind the success of multicarrier techniques is the partition of the physical channel
into (ideally) non-interfering narrowband subchannels. If the bandwidths associated with these
subchannels are narrow enough, then the related channel-frequency response of each subchannel
appears to be flat, thus avoiding the use of sophisticated equalizers. In addition, the subchannel
division allows, whenever possible, the exploitation of signal-to-noise ratios (SNRs) in the different
subchannels by managing the data load in each individual subchannel.
The orthogonal frequency-division multiplexing (OFDM) described in Chapter 3 is the most
popular multicarrier-modulation technique. OFDM-based systems feature lots of good properties
regarding their performance and implementation simplicity. However, the OFDM has some draw-
backs, such as high peak-to-average power ratio (PAPR), high sensitivity to carrier-frequency offset
(CFO), and (possibly) significant loss in spectral efficiency due to the redundancy insertion re-
quired to eliminate the interblock interference (IBI). The single-carrier with frequency-domain
equalization (SC-FD) is an efficient transmission technique which reduces both PAPR and CFO
as compared to the OFDM system. These advantages are attained without changing drastically the
overall complexity of the transceiver, as shown in Chapter 3.
Alternatively, one could consider the general transmultiplexer (TMUX) framework described
in Chapter 2 in order to conceive new multicarrier-modulation techniques which are able to cir-
cumvent some of the OFDM limitations. For example, intuition might suggest that we can reduce
intercarrier interference (ICI) caused by the loss of orthogonality between the subcarriers as long
as we are able to design TMUXes containing a large number of subchannels with sharp transitions.
Therefore, it would be possible to achieve better solutions (in terms of bit-error rate, for example)
for filter-bank transceivers to be employed in both future generations of wireless systems and current
developments of local/broadband wireless networks. This strategy, however, remains to be proved as
a viable solution in practice. In fact, one should always take into account the fundamental trade-off
between performance gains and cost effectiveness from a practical perspective. The computational
complexity1 is among the factors that directly affects the cost effectiveness of new advances in com-
1 Number of arithmetical operations employed in the related processing.
106 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
munications. TMUXes with sharp transitions must necessarily have long memory, which might
hinder their use in practice. This explains why memoryless linear time-invariant (LTI) TMUXes are
still preferred in many practical applications.
The price paid for using memoryless LTI TMUXes is that some redundancy must be nec-
essarily included at the transmitter end in order to allow the elimination of the ISI induced by the
channel. Redundancy plays a central role in communications systems. Channel-coding schemes are
good examples of how to apply redundancy in order to achieve reliable transmissions. In addition,
redundancy is also employed in many block-based transceivers in practice, such as OFDM and
SC-FD systems, in order to eliminate the inherent IBI and to yield simple equalizer structures.
Regarding the spectral-resource usage, the amount of redundancy employed in both OFDM
and SC-FD systems depends on the delay spread of the channel, implying that both transceivers waste
the same bandwidth on redundant data. Nevertheless, there are many ways to increase the spectral
efficiency of communications systems, such as by decreasing the overall symbol-error probability in
the physical layer, so that less redundancy needs to be inserted in upper-layers by means of channel
coding. In general, this approach increases the costs in the physical layer, since it leads to more
computationally complex transceivers, hindering its implementation in some practical applications.
Other means to improve spectral efficiency are, therefore, highly desirable. Reducing the
amount of transmitted redundancy inserted in the physical layer is a possible solution.2 Although
reliability and simplicity are rather important in practical applications, the amount of redundancy
should be reduced to potentially increase the spectral efficiency. In the context of fixed and memo-
ryless TMUXes, it is possible to show that the minimum redundancy required to eliminate IBI and
still allow the design of zero-forcing (ZF) solutions is only half the amount of redundancy used by
standard OFDM and SC-FD systems.
In this context, an important question arises: can we design memoryless LTI transceivers
with reduced redundancy whose computational complexity is comparable to OFDM and SC-FD
systems? That is, these transceivers should be amenable to superfast implementations in order to
keep their computational complexities competitive with practical OFDM-based systems. If this
is possible, the resulting transceivers would probably allow higher data throughputs in broadband
channels.
This chapter describes how transceivers with reduced redundancy (Section 4.2) can be imple-
mented employing superfast algorithms based on the concepts of structured matrix representations
(Section 4.3). To achieve this objective, we describe some mathematical decompositions of a special
class of structured matrix: the so-called Bezoutian matrices (Section 4.4). The resulting structures of
multicarrier and single-carrier reduced-redundancy systems are then presented and analyzed through
some examples (Section 4.5).

2The bottom line here is that many distinct and interesting ways of designing multicarrier systems are available. We will focus on a
particular type of solution (reduced-redundancy system) that allows us to present to the reader a set of tools related to structured
matrix representations, which can be eventually employed in many other contexts.
 4.2. REDUCED-REDUNDANCY SYSTEMS: THE ZP-ZJ MODEL REVISITED 107
4.2 REDUCED-REDUNDANCY SYSTEMS: THE ZP-ZJ
MODEL REVISITED
As briefly described in Chapter 2, Subsection 2.4.5, Lin and Phoong proposed a family of memoryless
discrete multitone transceivers which employ a reduced amount of redundancy to eliminate both
IBI and ISI [44]. A useful particular type of reduced-redundancy system is the so-called zero-padded
zero-jammed (ZP-ZJ) system. Let us revisit now the ZP-ZJ model in a more detailed manner.
Consider a complex-valued finite impulse-response (FIR) channel model h(l), where the
channel-tap index l can assume any value within the set L  {0, 1, · · · , L}. The integer number L
is the so-called channel order, which in turn is associated with the delay spread of the channel [80]: a
measure of how much a signal is spread over the time when passing through the channel.
In ZP-ZJ systems, the transmitter is responsible for linearly processing the input vector
s(n) ∈ C M×1 , where the integer number n is the time index, the natural number M denotes the
amount of symbols to be transmitted, and C ⊂ C is a given digital constellation. Such a processing
is implemented through the multiplication by the matrix
 
F
F ∈ CN ×M , (4.1)
0K×M

where N ∈ N is the total length of the transmitted message, whereas F ∈ CM×M is the actual
transmitter matrix, or just the transmitter matrix, which is in charge of preparing the vector for the
transmission through the FIR channel. The matrix 0K×M is responsible for adding the redundant
zero elements (zero-padding process). The amount of redundancy inserted is the natural number
K  N − M. (4.2)
If one assumes that the channel order L is less than the total length of the transmitted message
N , then the received block y(n) ∈ CN ×1 will suffer from ISI and IBI effects according to the channel
model described in Chapter 2, Equation (2.39), which can be rewritten as
y(n) = HISI Fs(n) + HIBI Fs(n − 1) + v(n), (4.3)
where v(n) ∈ CN×1 accounts for an additive noise at the receiver front-end, as depicted, for example,
in Figure 2.9 of Chapter 2.
In order to generate an estimate ŝ(n) ∈ CM×1 of the transmitted data vector,3 the receiver
applies another linear transformation to the incoming signal vector through the matrix
 
G  0M×(L−K) G ∈ CM×N , (4.4)

where G ∈ CM×(N+K−L) is the actual receiver matrix, or just the receiver matrix. The matrix
0M×(L−K) is responsible for removing the remaining IBI (zero-jamming process) that was not com-
pletely eliminated by the transmitted redundancy.
3 Observe that the estimated elements do not have to pertain to the digital constellation C. Only after a hard-decision-detection
process the estimated elements are mapped into constellation symbols belonging to C.
108 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
Now, by using the definitions of HISI and HIBI given in Equation (2.38) of Chapter 2, the
previous ZP-ZJ model description yields the following estimate:

ŝ(n)  Gy(n)
= GHISI Fs(n) + GHIBI F s(n − 1) + Gv(n)
     
=0M×M v(n)
= GHISI Fs(n) + v(n)
 
  F
= 0M×(L−K) G HISI s(n) + v(n)
0K×M
=G H F s(n) + v(n)
 
T
= Ts(n) + v(n), (4.5)

where the equivalent channel matrix H is defined as

⎡ ⎤
h(L − K) · · · h(0) 0 0 ··· 0
⎢ .. .. ..⎥
⎢ . . .⎥
⎢ ⎥
⎢ . ⎥
⎢ h(K) .. 0 ⎥
⎢ ⎥
⎢ .. .. .. ⎥
⎢ . . . h(0) ⎥
H⎢
⎢ ..
⎥ ∈ C(N +K−L)×M .
⎥ (4.6)
⎢ h(L) . ⎥
⎢ ⎥
⎢ .. ⎥
⎢ 0 . h(L − K)⎥
⎢ ⎥
⎢ .. .. ⎥
⎣ . . ⎦
0 ··· 0 0 h(L) · · · h(K)

Equation (4.5) shows that the estimated vector ŝ(n) does not depend on the transmitted vector
s(n − 1), which means that the overall transmission/reception process is performed on a block-by-
block basis. We can therefore simplify the notation by omitting the time-dependency of the related
variables. Thus, Equation (4.5) can be rewritten as

ŝ = Ts + v. (4.7)

Figure 4.1 illustrates the ZP-ZJ model omitting the time-dependency of the related variables.
Note that the choices for the dimensions of the all-zero matrices which implement the zero-padding
zero-jamming processes make sense. Indeed, by remembering the definition of matrix HIBI in
Equation (2.38), we can verify that only the first L elements of a received data block are affected by
the last L elements of the previous transmitted data block. This interblock interference occurs due to
the memory of the channel. Hence, the IBI effect can be eliminated by discarding the first L out of
N elements from all received data blocks. However, if one desires to transmit N = M + K elements,
 4.2. REDUCED-REDUNDANCY SYSTEMS: THE ZP-ZJ MODEL REVISITED 109

v
⎡ ⎤ ⎡ ⎤
s0

T M ŝ0
⎢ s1 ⎥ M N HISI N N ⎢ ŝ1 ⎥
F 0(L−K)×M
s=⎢
⎣ ... ⎦
⎥ + T ⎢ . ⎥ = ŝ
⎣ .. ⎦
0K×M −1
G
sM −1 z HIBI ŝM −1
N =M +K

Figure 4.1: ZP-ZJ transceiver model.

with K redundant zeros included at the end of each transmitted block, then instead of discarding the
first L elements from the received data block, one can discard only the first L − K elements, since the
immediately following L − (L − K) = K elements are “affected” by the zero-valued transmitted
elements appended at the end of the previous transmitted block (see Example 4.1 below).
In general, K can assume any non-negative integer value in order to eliminate the IBI effect.
In fact, if K ≥ L, then we must adjust the notation employed to define matrix G in Equation (4.4),
since L − K ≤ 0. In such a case, we could redefine G without zero jamming, i.e., G = G ∈ CM×N .
By doing so, we see that ZP-ZJ systems encompass general ZP-OFDM and ZP-SC-FD systems
which transmit K ≥ L redundant zeros.

Example 4.1 (Zero-Padding Zero-Jamming Process) Consider an FIR channel with transfer
function

H (z) = h(0) + h(1)z−1 + h(2)z−2 + h(3)z−3 . (4.8)

In this case, we have L = 3. Now, assume that M = 4 and K = 1, implying that N = M + K = 5.

Thus, we have (see Equation (2.38))
⎡ ⎤ ⎡ ⎤
h(0) 0 0 0 0 0 0 h(3) h(2) h(1)
⎢h(1) 0 0 0 ⎥ ⎥ ⎢ h(3) h(2)⎥
⎢ h(0) ⎢0 0 0 ⎥
H(z) = ⎢
⎢h(2) h(1) h(0) 0 0 ⎥ ⎥ +z −1 ⎢
⎢0 0 0 0 h(3)⎥
⎥. (4.9)
⎣h(3) h(2) h(1) h(0) 0 ⎦ ⎣0 0 0 0 0 ⎦
0 h(3) h(2) h(1) h(0) 0 0 0 0 0
     
=HISI =HIBI

Observe that, if we use the following matrices (see Equations (4.1) and (4.4))
 
I4
F= F, (4.10)
0
 1×4 
G = 04×2 G , (4.11)
110 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
in which F ∈ C4×4 and G ∈ C4×3 , then the IBI effect can be eliminated since

⎡ ⎤
0 0 h(3) h(2) h(1)
⎢0 0 0 h(3) h(2)⎥  
 ⎢ ⎥ I4
GHIBI F = 04×2 G ⎢
⎢0 0 0 0 h(3)⎥⎥ 01×4 F
⎣0 0 0 0 0 ⎦
0 0 0 0 0
⎛ ⎞
⎜⎡ ⎤ ⎡ ⎤ ⎟
⎜ 0 0 h(3) h(2) h(1) ⎟
⎜ ⎟
⎜⎢ h(3)⎥ ⎢h(2)⎥ ⎟
  ⎜⎢0 0 0 ⎥ ⎢ ⎥ ⎟
= 04×2 G ⎜ ⎢
⎜⎢0 0 0 ⎥ I + ⎢ ⎥
0 ⎥ 4 ⎢h(3)⎥ 1×4 ⎟ 0 ⎟F
⎜⎣0 0 0 0 ⎦ ⎣ 0 ⎦ ⎟
⎜ ⎟
⎜ 0 ⎟
⎝ 0 0 0 0 ⎠
  
=05×4
⎡ ⎤
0 0 h(3) h(2)
⎢0 0 0 h(3)⎥
 ⎢ ⎥
= 04×2 G ⎢
⎢0 0 0 0 ⎥ ⎥F
⎣0 0 0 0 ⎦
0 0 0 0
⎛ ⎞
⎜   ⎟
⎜ 0 0 h(3) h(2) ⎟
=⎜ 0
⎜ 4×2 0 0 + G 0 ⎟
3×4 ⎟ F
⎝ 0 h(3)   ⎠
  =04×4
=04×4
= 04×4 . (4.12)

Moreover, it is also possible to verify that the structure of the equivalent channel matrix H
described in Equation (4.6) holds in this example. Indeed, we have

⎡ ⎤
h(0) 0 0 0 0
⎢h(1) h(0) 0 0 0 ⎥  
 ⎢ ⎥ I4
GHISI F = 04×2 G ⎢
⎢h(2) h(1) h(0) 0 0 ⎥ ⎥ 01×4 F
⎣h(3) h(2) h(1) h(0) 0 ⎦
0 h(3) h(2) h(1) h(0)
 4.2. REDUCED-REDUNDANCY SYSTEMS: THE ZP-ZJ MODEL REVISITED 111

⎛ ⎞
⎜⎡ ⎤ ⎡ ⎤ ⎟
⎜ h(0) 0 0 0 0 ⎟
⎜ ⎟
⎜⎢ 0 0 ⎥ ⎥ ⎢ ⎥ ⎟
  ⎜⎢h(1) h(0) ⎢ 0 ⎥ ⎟
= 04×2 G ⎜ ⎢
⎜⎢h(2) h(1) h(0) 0 ⎥ ⎢ ⎥ ⎟
⎥ I4 + ⎢ 0 ⎥ 01×4 ⎟ F
⎜⎣h(3) ⎦ ⎣ 0 ⎦ ⎟
⎜ h(2) h(1) h(0) ⎟
⎜ 0 ⎟
⎝ h(3) h(2) h(1) h(0) ⎠
  
=05×4
⎡ ⎤
h(0) 0 0 0
⎢h(1) h(0) 0 0 ⎥
 ⎢ ⎥
= 04×2 G ⎢
⎢h(2) h(1) h(0) 0 ⎥ ⎥F
⎣h(3) h(2) h(1) h(0)⎦
0 h(3) h(2) h(1)
⎛ ⎞
⎜ ⎡ ⎤⎟
⎜   h(2) h(1) h(0) 0 ⎟
⎜ h(0) 0 0 0 ⎟
=⎜
⎜ 0 4×2 +G ⎣h(3) h(2) h(1) h(0)⎦⎟ F
⎟
⎜ h(1) h(0) 0 0
⎝   0 h(3) h(2) h(1) ⎟
=04×4   ⎠
=H∈C3×4
= G H F. (4.13)

As we have seen in Example 4.1, the redundancy is padded at the transmitter and jammed
at the receiver end in such a way that the IBI effect is completely eliminated. Nevertheless, the
amount of redundancy should be such that the ISI effect might also be eliminated, i.e., the amount
of redundant zeros should not be too small, otherwise it would prevent us from finding ZF solutions
(or, in other words, ISI elimination). Indeed, the definition of the M × M equivalent transfer matrix
T in Equation (4.5) makes clear that a zero-forcing solution is achievable if, and only if, the matrix
T = G H F is full-rank (i.e., rank{T} = M). Since the dimensions of matrices G, H, and F are
M × (M + 2K − L), (M + 2K − L) × M, and M × M, respectively, and since the rank of the
product of matrices is not larger than the rank of each individual matrix, then we must necessarily
have

M ≤ min {M, M + 2K − L} , (4.14)

implying M ≤ M + 2K − L ⇔ 0 ≤ 2K − L. This result means that

/ 0
L
K≥ , (4.15)
2
112 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
where the notation x denotes the ceiling operator, which gives us the unique integer number within
the interval [x, x + 1) ⊂ R, for any real number x. Notice that the ZF solution is only achieved if
matrices G, H, and F have full-rank.4
Consider that we do insert at least L/2 zeros before the transmission and assume that
channel-state information (CSI) is available at the receiver end, i.e., h(l) is known for all l ∈ L. In
addition, assume that the transmitter employs a predefined channel-independent unitary precoder F,5
so that the task of defining the rectangular matrix G is left to the designer. In this context, the usual
objective is to minimize either the ISI or the mean square error (MSE) at the receiver. Such goals
can be achieved by the ZF and the minimum MSE (MMSE) receivers, respectively. These solutions
are

GZF  (H F)†
= [(H F)H (H F)]−1 (H F)H
H H H H
= [F H H F]−1 F H
−1 H
H
−1 H H
= F (H H)−1 F F H
−1 H H
=F (H H)−1 H
  
†
H
H †
=F H , (4.16)
 −1
σ2
GMMSE  (H F)H (H F) + v2 IM (H F)H
σs
# $−1
H H σv2 H H H
= F H HF + 2F F F H
σs
# $−1

−1 H σ2 H −1 H H
=F H H + v2 IM F F H
σs
# $−1
H H σv2 H
=F H H + 2 IM H , (4.17)
σs
where (·)† denotes the Moore-Penrose pseudo-inverse of (·). The optimal MMSE solution is de-
rived following the same procedure as in Subsection 2.4.1. In addition, the vectors s and v are
considered uncorrelated wide-sense stationary (WSS) random sequences with zero-mean,6 yielding
E[sv]H = E[s]E[v]H = 0M×M = E[v]E[s]H = E[vsH ]. We also assume that E[ssH ] = σs2 IM
and E[v v H ] = σv2 IM , in which σv2 , σs2 ∈ R+ denote the variances of the random sequences (they
do not depend on time because the random sequences are WSS). The given expression for GZF is
4 In Example 4.1, we see that, even though IBI is eliminated, it is not possible to achieve ISI elimination since matrix G ∈ C4×3
< =
and H ∈ C3×4 have ranks at most 3, instead of 4 = M. This happens since we adopted K = 1 < L 2 = 2, thus violating the
minimum required redundancy.
5The unitary property means F −1 = F H .
6 Remember that the time index was omitted for the sake of conciseness.
 4.2. REDUCED-REDUNDANCY SYSTEMS: THE ZP-ZJ MODEL REVISITED 113
valid in the case H has full-rank. As a result, we are assuming that the FIR channel model leads to
a rank-M matrix H.
The ZF design does not require any estimate of the environment-noise variance and it is
mainly discussed for theoretical purposes. This lack of information about the noise variance leads
to poor performance when the SNR is low at the receiver end. On the other hand, the practical
MMSE design is computationally more demanding than the ZF solution, but leads to more effective
equalization.
The next step is to discuss how complex is performing both the equalization of a received
data block and the design of the related receiver. The equalization consists of processing the received
vector after the zero-jamming process through its multiplication by matrix G. This operation entails
O(M 2 ) complex-valued arithmetical operations for general unstructured matrices. The equalization
process requires the knowledge or the estimation of H and possibly its regularized pseudo-inverse.
This estimation is made during the receiver design. As a result, the computational complexity of the
receiver design is O(M 3 ) complex-valued arithmetical operations for general unstructured matrices.
The cyclic-prefix OFDM (CP-OFDM) and the CP-SC-FD transceivers have the advantage of
performing the equalization as well as the receiver design employing only O(M log2 M) complex-
valued operations due to their structural simplicity.
The use of discrete Fourier transform (DFT) and inverse DFT (IDFT) in order to decouple
the estimation of the symbols at the receiver end are paramount to the success of CP-OFDM-based
systems. Unfortunately, we cannot decouple so easily the estimation of the symbols in a ZP-ZJ
system with reduced redundancy. Indeed, such decoupling process would require the computation of
singular-value decompositions (SVDs), hindering its implementation in several practical problems.
Despite this potential drawback, we shall describe some tools that allow one to implement
low complexity ZP-ZJ systems with reduced redundancy. As a motivating example, let us consider
how a zero-forcing CP-SC-FD system is implemented. The insertion of the cyclic prefix turns the
linear convolution into a circular convolution between the transmitted data symbols and the channel-
impulse response. Using the vector notation for a noiseless channel, we can write y  Hc s, where
Hc is a circulant matrix which contains the channel coefficients. From linear algebra, we know that
all circulant matrices may be diagonalized by using the same set of orthonormal eigenvectors. These
eigenvectors are the columns of the unitary DFT matrix, W. In addition, the eigenvalues of circulant
channel matrices are easily computed by means of the DFT of the first column of the circulant matrix
(see the discussions in Subsection 2.4.1). Thus, we have y = Hc s = W H Ws, which implies that
we can recover s by performing s = H−1 H −1
c y = W Wy, considering that
−1 is computable,

i.e., all eigenvalues of Hc are non-zero. Hence, the ZF-SC-FD system that employs cyclic prefix
decomposes the inverse of the equivalent channel matrix using DFT and diagonal matrices. In fact, this
decomposition is quite special since it is a diagonalization of the inverse of the equivalent channel
matrix.
Our aim is to follow a similar approach, that is, to look for an efficient decomposition for the
“inverse” of the equivalent channel matrix associated with ZP-ZJ systems with reduced redundancy.
114 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
In such systems, the equivalent channel matrix H is no longer circulant; rather, it is an (M + 2K −
L) × M Toeplitz matrix, as described in Equation (4.6). Nevertheless, we could take into account
the Toeplitz structure in order to decompose the generalized inverse of H, maybe using only DFT
and diagonal matrices. Such an approach employs the same basic ideas present in CP-OFDM-based
systems, except for two main features present only in OFDM-based systems: (i) the inverse of the
equivalent channel matrix has exactly the same structure as the equivalent channel matrix itself
(circulant structure); and (ii) the efficient decomposition of the inverse of the equivalent channel
matrix corresponds to its diagonalization.

4.3 STRUCTURED MATRIX REPRESENTATIONS

A matrix is considered structured if its entries follow some predefined pattern. The origin of this
pattern can be elementary mathematical relations among the matrix coefficients or simply the way
certain coefficients appear repeatedly as entries of the given matrix. Some of these patterns happen
naturally in the matrices describing the behavior of many practical systems (e.g., linear and circular
discrete-time convolutions can be modeled by structured matrices) or can be easily induced by
applying simple matrix transformations.
In any case, a structured matrix can be described by using a reduced set of distinct parameters,
that is, the number of parameters required to describe the matrix is much smaller than its number
of entries. There are several examples of structured matrices which are usually found in signal-
processing and communication applications, such as: diagonal, circulant, pseudo-circulant, Toeplitz,
among others.
Such structural patterns may bring about efficient means for exploiting features of the related
problems. Besides, computations involving structured matrices can be further simplified by taking
into account these structural patterns. Consider, for instance, the sum of two M2 × M1 Toeplitz
matrices. If one ignores the structural patterns present in such matrices, then this operation will
require M2 M1 additions, since there are M2 M1 entries in each matrix. However, these matrices
are completely defined by up to M2 + M1 − 1 elements, since the first row (M1 elements) along
with the first column (M2 elements, in which the first element pertains to the first row as well)
are enough to define a given Toeplitz matrix. This way, it would be quite reasonable to expect that
matrix operations may be performed faster by using a reduced amount of parameters. Indeed, if one
considers the structure of the matrices, then this operation will require only M2 + M1 − 1 additions
corresponding to the sum of the first row and the first column of each matrix. The resulting Toeplitz
matrix can be built by rearranging the elements of the resulting vectors accordingly.
As previously mentioned, the widespread use of OFDM and SC-FD transceivers relies on
their key feature of transforming the original description of the Toeplitz channel-convolution matrix
into a circulant matrix, for the case where the channel is linear and time-invariant.7 Since a circulant
matrix has eigenvectors comprising the columns of the unitary DFT matrix, the diagonalization
7The time-invariance assumption only needs to hold during the transmission of one data block.
 4.3. STRUCTURED MATRIX REPRESENTATIONS 115
process is quite efficient due to the existence of fast Fourier transform (FFT) algorithms, turning
the equalization process very simple.
The effective channel matrix associated with ZP-ZJ systems is a rectangular Toeplitz matrix
(see Equation (4.6)). It is therefore natural to expect that linear equalizers, such as linear MMSE or
ZF equalizers, can take advantage of the structure of this channel matrix like OFDM-based systems
do. In this context, three questions arise: (i) How to recognize a structured matrix by using analytical
tools? (ii) How to represent the linear optimal solutions (either MMSE or ZF) by employing such
analytical tools? and (iii) How to effectively take advantage of such representations?
This section answers those questions in the context of rectangular structured matrices. A
useful tool for exploiting the structure of a matrix is the displacement approach proposed by Kailath
et al. [34]. This section addresses the representation of structured matrices using the displacement
structure. Indeed, we apply the concept of displacement to some structured matrices that appear in
reduced-redundancy transceivers. An important result is the decomposition of Bezoutian matrices,
because they play a key role in the design of transceivers using reduced redundancy.

4.3.1 DISPLACEMENT-RANK APPROACH

There are several ways to measure how much a matrix is structured. One of the most widely used
is the displacement operator. The structural property of a matrix can be determined by analyzing the
matrix resulting from the application of the displacement operation on the original matrix. Such
analysis can be as simple as verifying the rank of the resulting matrix after the application of the
displacement operator.
While studying techniques to solve Toeplitz-like problems, Kailath et al. [34] proposed a
procedure to measure how much a given matrix deviates from a Toeplitz matrix based on the com-
putation of the so-called displacement rank. The Toeplitz structure was chosen due to the inherent
computational simplicity associated with the algorithms employed to solve Toeplitz-like problems,
as described thoroughly in [35, 61].
Let us assume that A ∈ CM1 ×M1 and B ∈ CM2 ×M2 are two given square matrices, where M1
and M2 are positive integers. From now on, we shall refer to those matrices as operator matrices.
Consider the linear transformations

∇A,B : CM1 ×M2 → CM1 ×M2

C  → ∇A,B (C)  AC − CB, (4.18)
A,B : CM1 ×M2 → CM1 ×M2
C  → A,B (C)  C − ACB, (4.19)

which associate each M1 × M2 complex-valued rectangular matrix C with other M1 × M2 rectan-

gular matrices ∇A,B (C) and A,B (C), respectively. These linear mappings are the so-called Sylvester
(∇A,B ) and Stein (A,B ) displacement operators.
The displacement ranks related to these linear operators when applied to C are the ranks
of the resulting displacement matrices ∇A,B (C) and A,B (C). In general, matrix C is represented
116 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
through M1 M2 coefficients. However, if R  rank{∇A,B (C)} or R  rank{A,B (C)} is such that
R  min{M1 , M2 }, then the original matrix can be represented using up to R(M1 + M2 )  M1 M2
independent coefficients,8 thus reducing the amount of elements required to express matrix C, as
will be further clarified.
The operator matrices A and B define completely the displacement operator that will be
utilized. The choices of these matrices should rely on which matrices lead to lower displacement
ranks. The most common types of operator matrices are the λ-circulant and the diagonal matrices.
The λ-circulant matrix, denoted as Zλ , and the diagonal matrix, denoted as Dν , are M × M square
matrices, respectively, given by
⎡ ⎤
0 λ
⎢ ⎥
⎢1 . . . ⎥
Zλ  ⎢ ⎢
⎥
⎥
⎣ .. .. ⎦
. .
1 0
 
= e1 · · · eM−1 λe0 , (4.20)

Dν  diag{νν }
⎡ ⎤
ν0
⎢ ν1 ⎥
⎢ ⎥
=⎢ .. ⎥
⎣ . ⎦
νM−1
 
= ν0 e0 ν1 e1 · · · νM−1 eM−1 , (4.21)

where λ is a complex-valued scalar, em is an M × 1 vector whose mth entry is equal to 1, whereas

the remaining elements are equal to 0, and ν  [ ν0 ν1 · · · νM−1 ]T is an M × 1 complex-valued
vector. The reader is encouraged to verify that

Zλ−1 = ZT1
λ
= ZH1 , (4.22)
λ∗

for all non-zero scalar λ and that

 T
Dν−1 = diag 1
ν0
1
ν1 ··· 1
νM−1 , (4.23)

as long as νm is a non-zero scalar, for each m within the set M defined as

M  {0, 1, · · · , M − 1}. (4.24)

8The reader should remember that a rank-1 matrix can be expressed as the outer product between two vectors. In general, a rank-R
matrix can be expressed as the sum of R outer products between vectors.
 4.3. STRUCTURED MATRIX REPRESENTATIONS 117
Example 4.2 (λ-Circulant Operator Matrices) This example illustrates the effects of right- and
left-multiplying a given matrix by λ-circulant matrices. Consider a 3 × 3 matrix C given by
⎡ ⎤
1 2 3
C = ⎣4 5 6⎦ . (4.25)
7 8 9

In addition, assume that λ = −1 and

⎡ ⎤
0 0 −1
Z−1 = ⎣1 0 0 ⎦ . (4.26)
0 1 0

Hence, we have
⎡ ⎤
2 3 −1
CZ−1 = ⎣5 6 −4⎦ , (4.27)
8 9 −7
⎡ ⎤
−7 −8 −9
Z−1 C = ⎣ 1 2 3 ⎦. (4.28)
4 5 6

Therefore, right-multiplication by a λ-circulant matrix shifts all columns to the left, where the first
original column multiplied by λ is moved to the last column of the resulting matrix. On the other
hand, left-multiplying by a λ-circulant matrix shifts down all rows, where the last original row
multiplied by λ is moved to the first row of the resulting matrix. 

Usually, structured matrices can be associated with some linear displacement operator. These
operators might reveal if a given structured matrix can be represented by a reduced number of
parameters. This representation is the key feature that allows the derivation of superfast algorithms 9
for inverting as well as performing matrix-to-vector multiplication involving the related structured
matrix.
The fast implementation of the reduced-redundancy transceivers rely on the displacement
rank of the matrices involved. The procedure entails the following steps.

1. Compression: If the rank of the displacement matrix of a given M1 × M2 structured matrix C

is lower than the dimensions of C, then it is possible to represent this matrix with a reduced
number of coefficients. Indeed, the displacement operator applied to the original matrix can
be compressed and represented by the so-called displacement-generator pair of matrices (P, Q )
with the following features: assuming we are dealing with a Sylvester displacement operator
9That is, algorithms that require O(M logd M) numerical operations, where d ≤ 3 [61].
118 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
of rank R ∈ N, it is true that

∇A,B (C) = pr qrH
r∈R
= PQ H
, (4.29)

in which R  {0, 1, · · · , R − 1} and the generator matrices are given by

 
P  p0 p1 · · · pR−1 ∈ CM1 ×R , (4.30)
 
Q  q0 q1 · · · qR−1 ∈ CM2 ×R , (4.31)

where, for each r within the set R, we have

 T
pr  p0,r p1,r · · · p(M1 −1),r , (4.32)
 T
qr  q0,r q1,r · · · q(M2 −1),r . (4.33)

2. Operation: The idea behind the displacement approach is that the compressed form of a
structured matrix contains all the information about the original matrix, but with a reduced
amount of elements. Therefore, instead of performing operations with the original matrix, it
is worth using the appropriate displacement-generator pair to perform such operations.

3. Decompression: Once the required operations are performed, the processed version of the orig-
inal matrix can be recovered through a displacement decompression operation, as long as the
operator matrices satisfy some mild constraints.

In order to illustrate the power of the displacement-rank tool, let us define some types of
structured matrices in a more formal manner in Subsection 4.3.2.

4.3.2 TOEPLITZ, VANDERMONDE, CAUCHY, AND BEZOUTIAN MATRICES

Toeplitz matrices have the property that their entries do not vary along the diagonals, as can be seen
in Equation (4.36). Indeed, an M2 × M1 complex-valued matrix T is a Toeplitz matrix when, for
each pair of indexes (m2 , m1 ) within the set M2 × M1 , in which

M1  {0, 1, · · · , M1 − 1}, (4.34)

M2  {0, 1, · · · , M2 − 1}, (4.35)

the (m2 , m1 )th element of T , denoted as [T T ]m2 m1  tm2 ,m1 ∈ C, satisfies the equality tm2 ,m1 =
t(m2 +1),(m1 +1) , as long as the pair of indexes (m2 + 1, m1 + 1) is also within the set M2 × M1 . In
 4.3. STRUCTURED MATRIX REPRESENTATIONS 119
this case, matrix T can be rewritten in a more convenient form, as follows:

⎡ ⎤
t0 t−1 t−2 ··· t1−M1
⎢ t1 t0 t−1 ··· t2−M1 ⎥
⎢ ⎥
⎢ ··· ⎥
T =⎢ t2 t1 t0 t3−M1 ⎥
⎢ .. .. .. .. .. ⎥
⎣ . . . . . ⎦
tM2 −1 t tM2 −3 · · · tM2 −M1
  M2 −2
 t(m2 −m1 ) (m2 ,m1 )∈M2 ×M1
. (4.36)

In order to exemplify the compression capability of the displacement operators when dealing
with rectangular Toeplitz matrices, let us consider the application of the Sylvester displacement op-
erator ∇Zη ,Zξ , in which Zη ∈ CM2 ×M2 and Zξ ∈ CM1 ×M1 , on an M2 × M1 complex-valued Toeplitz
matrix T :

∇Zη ,Zξ (T
T ) = ZηT − T Zξ

⎡ ⎤
ηtM2 −1 ηtM2 −2 · · · ηtM2 −M1 +1 ηtM2 −M1
⎢ t0 t−1 ··· t2−M1 t1−M1 ⎥
⎢ ⎥
⎢ ··· ⎥
= ⎢ t1 t0 t3−M1 t2−M1 ⎥
⎢ . .. .. .. .. ⎥
⎣ .. . . . . ⎦
tM2 −2 tM2 −3 ··· tM2 −M1 tM2 −M1 −1

⎡ ⎤
t−1 t−2 ··· t1−M1 ξ t0
⎢ t0 t−1 ··· t2−M1 ξ t1 ⎥
⎢ ⎥
⎢ .. .. .. .. .. ⎥
−⎢ . . . . . ⎥
⎢ ⎥
⎣tM tM2 −4 · · · tM2 −M1 −1 ξ tM 2 ⎦
2 −3 −2
tM2 −2 tM2 −3 · · · tM2 −M1 ξ tM2 −1

⎡ ⎤
ηtM2 −1 − t−1 ηtM2 −2 − t−2 · · · ηtM2 −M1 +1 − t1−M1 ηtM2 −M1 − ξ t0
⎢ 0 0 ··· 0 t1−M1 − ξ t1 ⎥
⎢ ⎥
⎢ .. .. .. .. .. ⎥
=⎢ . . . . . ⎥
⎢ ⎥
⎣ 0 0 ··· 0 tM2 −M1 −2 − ξ tM2 −2
⎦
0 0 ··· 0 tM2 −M1 −1 − ξ tM2 −1
120 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
⎡ ⎤
1
⎢0 ⎥
⎢ ⎥ 
⎢ ⎥
= ⎢ ... ⎥ ηtM2 −1 − t−1 ηtM2 −2 − t−2 · · · ηtM2 −M1 +1 − t1−M1 ηtM2 −M1
⎢ ⎥  
⎣0 ⎦
q̂1H
0

p̂1
⎡ ⎤
−ξ t0
⎢ t1−M1 − ξ t1 ⎥
⎢ ⎥ 
⎢ .. ⎥
+⎢ . ⎥ 0 0 ··· 0 1
⎢ ⎥  
⎣tM − ξ tM 2 ⎦
2 −M1 −2 −2 q̂2
H
tM −M −1 − ξ tM2 −1
 2 1  
p̂2

= p̂1 q̂1H + p̂2 q̂2H

 
  q̂1H
= p̂1 p̂2
   q̂2H
  
P̂
Q̂ H

= P̂Q̂ H
. (4.37)

Note that the resulting displacement matrix ∇Zη ,Zξ (T T ) can be represented by the
displacement-generator pair of matrices (P̂, Q̂ ) ∈ C M 2 ×2 × C ×2 . Thus, if one assumes that M2
M 1

and M1 are integer numbers much larger than 2, then the former example shows that rectangular
Toeplitz matrices can always be compressed, since matrix ∇Zη ,Zξ (T T ) has rank at most 2. In fact,
since the vectors p̂1 and q̂2 do not depend on T , then at most M2 + M1 coefficients (corresponding
to the elements of the vectors p̂2 and q̂1 ) are really required to represent the displacement of T . This
reduced amount of coefficients is rather close to the total number of possibly different entries of
matrix T , which is M2 + M1 − 1.
Now, let us define another important class of structured matrix: the Vandermonde matrix,
which is a matrix whose entries in each row are the terms of a geometric progression, as can be seen
in Equation (4.38) below. Indeed, given an M × 1 complex-valued vector ν  [ ν0 ν1 · · · νM−1 ]T ,
an M × M matrix V ν is a Vandermonde matrix when, for each pair of indexes (m2 , m1 ) within the
set M2  M × M, the (m2 , m1 )th element of V ν , denoted as [V V ν ]m2 m1  νm2 ,m1 ∈ C, satisfies
 4.3. STRUCTURED MATRIX REPRESENTATIONS 121
m1
the equality νm2 ,m1 = νm 2 . In this case, V ν presents the following form:

⎡ ⎤
ν00 ν01 · · · ν0M−2 ν0M−1
⎢ ν10 ν11 · · · ν1M−2 ν1M−1 ⎥
⎢ ⎥
⎢ .. ⎥
Vν = ⎢
⎢
..
.
..
.
..
.
..
.
⎥
. ⎥
⎢ 0 M−1 ⎥
⎣νM−2 νM−2 1 · · · νM−2
M−2
νM−2 ⎦
0
νM−1 νM−1 1 · · · νM−1
M−2 M−1
νM−1
 [νm
m1
]
2 (m2 ,m1 )∈M
2. (4.38)

As in the Toeplitz case (see Equation (4.37)), a similar kind of compression can be applied
to Vandermonde matrices as well. Let us consider the application of the Sylvester displacement
operator ∇Dν ,Z0 , in which Dν = diag{νν } ∈ CM×M , on a given M × M Vandermonde matrix V ν .
In this case, we have

V ν ) = Dν V ν − V ν Z0
∇Dν ,Z0 (V
⎡ ⎤ ⎡ ⎤
ν0 · · · ν0M−1 ν0M ν0 · · · ν0M−1 0
⎢ ν1 · · · ν1M−1 ν1M ⎥ ⎢ ν1 · · · ν1M−1 0⎥
⎢ ⎥ ⎢ ⎥
=⎢ . .. .. ⎥−⎢ .. .. .. ⎥
⎣ .. . . ⎦ ⎣ . . .⎦
νM−1 · · · νM−1
M−1 M
νM−1 νM−1 · · · νM−1
M−1
0
⎡ ⎤
0 · · · 0 ν0M
⎢0 · · · 0 ν M ⎥
⎢ 1 ⎥
= ⎢. .. .. ⎥ , (4.39)
⎣ .. . . ⎦
0 · · · 0 νM−1
M

which consists of a rank-1 matrix with M degrees of freedom. Note that, even though the original
Vandermonde matrix V ν is comprised of M 2 entries, the M elements which compose the vector ν are
enough to completely define V ν . This compression example, therefore, shows that the displacement
approach is able to reveal analytically this reduced number of degrees of freedom.
In addition, there is a close relation between Vandermonde and DFT matrices. In order to
derive such a useful relation, let us first remember that the Mth roots of a given complex number

ξ consist of M distinct complex numbers ξm , with m ∈ M, such that ξmM = ξ = |ξ |ej ξ , where
j = −1 and  ξ ∈ (−π, π ] ⊂ R represents the (principal) phase of the complex number ξ when
2 10

10 Remember that, if  ξ is a phase of a given complex number ξ , then  ξ + 2iπ is also a phase of ξ , for any integer number i. The
principal phase is the unique phase of ξ within the interval (−π, π ].
122 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
expressed in its polar form; that is,

1  ξ 2π
ξm  |ξ | M ej M e−j M m
  
ξ0
⎛ ⎞m
⎜ 2π ⎟
= ξ0 ⎝e−j
 ⎠
M

WM
= ξ0 WM
m
. (4.40)

Therefore, by defining the vector ξ as

 T
ξ  ξ0 ξ1 · · · ξM−1 , (4.41)

its corresponding Vandermonde matrix is given by

⎡ ⎤
ξ00 ξ01 · · · ξ0M−1
⎢ ξ10 ξ11 · · · ξ1M−1 ⎥
⎢ ⎥
Vξ = ⎢ .. .. .. .. ⎥
⎣ . . . . ⎦
0
ξM−1 1
ξM−1 · · · ξM−1M−1
⎡ ⎤
ξ00 WM0 1
ξ0 WM 0 ··· ξ0M−1 WM0
⎢ξ 0 W 0 ξ01 WM 1 · · · ξ0M−1 WM M−1 ⎥
⎢ 0 M ⎥
=⎢
⎢ .. . .. .. .. ⎥
⎥
⎣ . . . ⎦
M−1 M−1 (M−1)2
0
ξ W 0 ξ W1 · · · ξ0 WM
⎡ 0 0 M 00 M ⎤⎡ ⎤
WM WM ··· WM 0
ξ00 0 ··· 0
⎢W 0 1 · · · WM M−1 ⎥ ⎢ ⎥
⎢ M WM ⎥ ⎢ 0 ξ01 ··· 0 ⎥
=⎢
⎢ ... .
.. .
.. .
..
⎥⎢ .
⎥⎣ . .. .. .. ⎥
⎣ ⎦ . . . . ⎦
2
W 0 WMM−1
· · · WM
(M−1) 0 0 · · · ξ0M−1
√ M
= MWM diag{ξ0m }m∈M , (4.42)

where WM is the M × M unitary DFT matrix previously defined in Equation (2.46). Such a
relation between Vandermonde and DFT matrices will be key to developing efficient factorizations
of structured matrices which appear in the ZP-ZJ model.
Another important structured matrix is the so-called Cauchy matrix, which may not be
 T
familiar to some readers. Given two vectors, ν  ν0 ν1 · · · νM1 −1 ∈ CM1 ×1 and λ 
 T
λ0 λ1 · · · λM2 −1 ∈ CM2 ×1 , an M1 × M2 complex-valued matrix C ν ,λλ is a Cauchy matrix
 pairof indexes (m1 , m2 ) within the set M1 × M2 , the (m1 , m2 )th element of C ν ,λλ ,
when, for any
denoted as C ν ,λλ m m  cm1 ,m2 ∈ C, satisfies the equality cm 1,m = 1 − νm1 λm2 , assuming that
1 2 1 2
 4.3. STRUCTURED MATRIX REPRESENTATIONS 123
νm1 λm2  = 1. Thus, C ν ,λλ can be represented as follows:
⎡ ⎤
1 1
1−ν0 λ0 1−ν0 λ1 ··· 1
1−ν0 λ(M2 −1)
⎢ ⎥
⎢ 1 1
··· 1 ⎥
⎢ 1−ν1 λ0 1−ν1 λ1 1−ν1 λ(M2 −1) ⎥
C ν ,λλ =⎢ .. .. .. ⎥
⎢ ..
. ⎥
⎣ . . . ⎦
1
1−ν(M1 −1) λ0
1
··· 1
  1−ν(M1 −1) λ1 1−ν(M1 −1) λ(M2 −1)
1
 . (4.43)
1 − ν m 1 λm 2 (m1 ,m2 )∈M1 ×M2

A Cauchy matrix C ν ,λλ can be easily compressed by using the Stein displacement operator
Dν ,Dλ . Indeed, one has

C ν ,λλ ) = C ν ,λλ − Dν C ν ,λλ Dλ

Dν ,Dλ (C
   
1 νm1 λm2
= −
1 − νm1 λm2 (m1 ,m2 )∈M1 ×M2 1 − νm1 λm2 (m1 ,m2 )∈M1 ×M2
 
1 − νm1 λm2
=
1 − νm1 λm2 (m1 ,m2 )∈M1 ×M2
⎡ ⎤
1 ··· 1
⎢ ⎥
= ⎣ ... . . . ... ⎦ , (4.44)
1 ··· 1

which is also a rank-1 matrix. In fact, the resulting displacement matrix in Equation (4.44) enjoys
the most simple form that we have found up to now. This feature will be important for us in
g
Subsection 4.4.1. Moreover, a rectangular matrix C ν ,λλ is defined as a generalized Cauchy matrix, as
g
long as the Stein displacement matrix Dν ,Dλ (C C ν ,λλ ) has rank much smaller than the number of
g
rows and columns of C ν ,λλ .11
The last class of structured matrix of our interest is the Bezoutian matrix. As opposed to most
of the previous structured matrices, there is no simple closed-form expression for the entries of a
Bezoutian matrix. In fact, the easiest way to define a Bezoutian matrix is based on its displacement
rank, as in the case of generalized Cauchy matrices. Indeed, an M1 × M2 Bezoutian matrix B is such
that

B ) = PQ
∇Zν ,Zλ (B H
(4.45)

for some complex-valued scalars ν and λ, where (P, Q ) ∈ CM1 ×R × CM2 ×R , and with M2  R
and M1  R.
It is possible to show (see Subsection 4.3.3) that the inverse of a non-singular square Toeplitz
matrix is a particular type of Bezoutian matrix, also known as T-Bezoutian matrix.
11There are some other classes of Cauchy matrices which will not be of particular importance in this book. The interested reader
should refer to [27, 28] for detailed information on this topic.
124 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
So far, we have only illustrated the compression capabilities of the displacement operators.
Let us analyze now the operation stage of the displacement approach.

4.3.3 PROPERTIES OF DISPLACEMENT OPERATORS

As we already mentioned, the compression stage is key to representing a given structured matrix using
a reduced amount of elements. Nevertheless, in order to benefit from the reduced amount of elements
in practice, it is necessary to conceive some ways of translating the operations on the compressed
form into related operations over the original matrix entries. The operation stage associated with
the displacement approach is in charge of such a translation. Thus, this subsection addresses some
properties which describe how standard matrix operations, namely linear combination, product, and
inversion, are mapped into their corresponding displacement matrices.
A first key result related to the displacement operators is the conversion between the Sylvester
and Stein operators. Such a conversion is possible as long as at least one of the operator matrices,
either A or B (see expressions (4.18) and (4.19)), is invertible. For example, if the operator matrix
A is a square non-singular matrix, then

∇A,B (C) = AC − CB

= A C − A −1 CB
= AA −1 ,B (C), (4.46)

in which it is assumed that all matrices have compatible dimensions.12 Similarly, if B is invertible,
then

∇A,B (C) = −A,B−1 (C)B. (4.47)

These relations between the Sylvester and Stein displacement operators show that, given two operator
matrices A and B, the displacement rank of a third matrix C does not depend on the particular
choice of the displacement operator (whether Sylvester or Stein), as long as at least one of the
operator matrices has full-rank and, in addition, some minor adjustments on the operator matrices
are performed, as for example, replacing A by A −1 or B by B−1 .
Let us now discuss two important results dealing with linear combinations of matrices and
matrix products, which basically show how these operations affect the displacement-generator
pairs of the original matrices. With respect to the generator of a linear combination, assume that
∇A,B (C) = PC QCH and that ∇A,B (D) = PD QDH , where matrices C and D have the same dimen-

12That is, the number of rows and columns are such that one can perform operations with the involved matrices, like addition or
product, without worrying about the dimensions of the related matrices.
 4.3. STRUCTURED MATRIX REPRESENTATIONS 125
sions. Thus, for any complex-valued scalar α, one always has

∇A,B (C + αD) = ∇A,B (C) + α∇A,B (D)

= PC QCH + αPD QDH
 
  QH
= PC αPD C
   QDH
  
P
Q H

= PQ H
. (4.48)

As for the generator of a product, consider that ∇A,B (C) = PC QCH and that ∇B,D (E) = PE QEH ,
where all matrices have compatible dimensions. Hence, we have

∇A,D (CE) = A(CE)

 − (CE)D
= ACE − CBE
 + (CBE − CED)
= AC − CB E + C (BE − ED)

= PC QCH E + C PE QEH
 
  Q HE
= PC CPE C
   QEH
  
P
Q H

= PQ H
. (4.49)

Thus, the knowledge about the displacement-generator pairs of the individual matrices before the
algebraic operation (whether linear combination or product) helps one determine the displacement-
generator pair of the resulting matrix after the related operation.
Now, we will analyze the inversion operation. The importance of this result is the fact that the
compression of an inverse matrix can be easily generated from the compressed version of the original
matrix. Indeed, given an invertible square complex-valued matrix C and the operator matrices A
and B, one has

∇B,A (C−1 ) = BC−1 − C−1 A

= B − C−1 AC C−1
 
= C−1 CB − AC C−1
= −C−1 ∇A,B (C)C−1 . (4.50)

In other words, if the displacement-generator pair of ∇A,B (C) is (PC , QC ), then the displacement-
generator pair of ∇B,A (C−1 ) is (−C−1 PC , C−H QC ), in which C−H denotes the Hermitian of the
inverse of C.
An immediate application of the inversion property is the verification of the following fact:
for a given non-singular square Toeplitz matrix T ∈ CM×M whose displacement generator pair
126 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
associated with the Sylvester displacement operator ∇Zη ,Zξ is (P̂, Q̂ ), one has ∇Zξ ,Zη (T T −1 ) =
(−T T −1 P̂)(Q̂ H T −1 ), whose rank is at most 2 due to Equation (4.37). Therefore, T −1 is a Bezoutian
matrix whenever M  2.
Actually, any regularized pseudo-inverse of an M2 × M1 complex-valued Toeplitz matrix T
is a Bezoutian matrix, as long as min{M2 , M1 }  4. Indeed, given a strictly positive real number ρ,
the regularized pseudo-inverse of T is defined as

−1
T †ρ  T H T + ρIM1 T H. (4.51)

 −1
Note that, by applying the matrix inversion lemma 13 to the expression T H T + ρIM1 that appears
in Equation (4.51), one has

−1 # $−1
1 1 H 1 1
T T + ρIM1
H
= IM1 − T T T + IM 2
H
T
ρ ρ ρ ρ
1
−1 
= IM1 − T H
T T + ρIM2
H
T , (4.52)
ρ

and, by right-multiplying both sides of Equation (4.52) by matrix T H , we get


−1 
1
T †ρ = T −T
H H
T T + ρIM2
H
TT H
ρ  
1 H
−1
= T IM2 − T T + ρIM2
H
TT H
ρ
1
−1

= T H T T H + ρIM2 T T H + ρIM2 − T T H
ρ
1
−1  
= T H T T H + ρIM2 ρIM2
ρ

−1
= T H T T H + ρIM2 . (4.53)

−1
−1
13The following identity holds: A + BCDH = A −1 − A −1 B C−1 + DH A −1 B DH A −1 , assuming that all operations
are valid (see, for example, [83]).
 4.3. STRUCTURED MATRIX REPRESENTATIONS 127
Therefore, the Sylvester displacement of the regularized pseudo-inverse described in Equa-
tions (4.51) and (4.53) is given by

−1
−1
∇Zξ ,Zη T †ρ = Zξ T H T T H + ρIM2 − T H T + ρIM1 T H Zη

−1

= T H T + ρIM1 T H T + ρIM1 Zξ T H


−1
− T H Zη T T H + ρIM2 T T H + ρIM2

−1

= T H T + ρIM1 T H T Z ξ T H − T H Zη T T H


−1
+ ρ Zξ T H − T H Zη T T H + ρIM2

−1
= T H T + ρIM1 −TT H ∇Zη ,Zξ (T
T )T H


−1
+ ρ∇Zξ ,Zη T H T T H + ρIM2 . (4.54)

 
Hence, if we define ∇Zη ,Zξ (T
T ) as P̂Q̂ H and ∇Zξ ,Zη T H as P̌Q̌ H , then we have

−1
−1
∇Zξ ,Zη T †ρ = −TT †ρ P̂Q̂ H T †ρ + T H T + ρIM1 P̌Q̌ H ρ T T H + ρIM2
  −1 
   H ρ T T H + ρI
−1 Q̌ M
= T H T + ρIM1 P̌ −T
T †ρ P̂ 2

   Q̂ H T †ρ
M ×4
  
P∈C 1
Q H ∈C4×M2

= PQ H
, (4.55)

which proves that T †ρ is a Bezoutian matrix if one assumes that min{M2 , M1 }  4.

The above derivation considers that ρ > 0. In fact, if we assume that M2 ≥ M1 and that
T } = M1 , then the M2 × M2 matrix T T H turns out to be rank deficient. In such a case, the
rank{T
previous derivation would not be valid for ρ → 0, due to the use of operations like the inversion of
 H   −1 H
T T + ρIM2 . Nevertheless, it is still possible to verify that T †  T H T T is also a Bezoutian
 H −1
matrix. Indeed, by applying the matrix inversion lemma to the expression ρ T T + ρIM2 which
appears in Equation (4.55), we have
 # $−1 

−1 1 1 1 H H 1
ρ T T H + ρIM2 =ρ IM − T T T + I M1 T
ρ 2 ρ ρ ρ

−1
= IM2 − T T H T + ρIM1 TH
= IM2 − T T †ρ . (4.56)
128 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
Now, by substituting Equation (4.56) into Equation (4.55), we end up with an expression
which does not depend upon the matrix T T H . In this case, one has

∇Zξ ,Zη T † = lim ∇Zξ ,Zη T †ρ

ρ→0

−1

= −TT † P̂Q̂ H T † + T H T P̌Q̌ H IM2 − T T †

 
   Q̌ H I − T T † 
−1 M
= T HT P̌ −T T † P̂ 2

   Q̂ H T †
M ×4
  
P∈C 1
Q H ∈C4×M2

= PQ H
. (4.57)

The reader should notice the similarity between the above (regularized) pseudo-inverses and
the optimal ZF and MMSE linear solutions defined in Equations (4.16) and (4.17), apart from the
transmitter-matrix inverse which appears only in the definitions of the optimal receiver matrices.14
This means that the optimal receivers associated with ZP-ZJ transceivers are, essentially, Bezoutian
matrices multiplied by a predefined unitary matrix F. Hence, inspired by the fact that the standard
CP-OFDM and CP-SC-FD transceivers decompose the inverse of circulant channel matrices using
DFT, IDFT, and diagonal matrices, it seems to be a good idea to look for decompositions of general
Bezoutian matrices entailing the same type of matrices, namely DFT, IDFT, and diagonal.

4.4 DFT-BASED REPRESENTATIONS OF BEZOUTIAN

MATRICES
Since the formal definitions and results in connection with the displacement operations are now
available, we can proceed to investigate their use in the structured matrices related to reduced-
redundancy transceivers. Thus, this section discusses the main mathematical results required to
design block transceivers with reduced redundancy which are based on DFT.
From what we have described in Subsection 4.3.2, we know that Cauchy matrices are able to
yield really simple Stein displacement matrices, as exemplified in Equation (4.44).This indicates that
Cauchy matrices may have efficient representations, which could be exploited as long as we are able
to transform Bezoutian matrices into Cauchy ones. In fact, this transformation is possible and many
works have taken advantage of this possibility, as explained, for instance, in [27, 28, 49] and references
therein. In other words, we are ultimately interested in transforming a Bezoutian-like matrix G (see
Equations (4.16) and (4.17)) into a new Cauchy-like matrix in order to benefit from the simplicity
of Cauchy matrices. Figure 4.2 illustrates the strategy that we shall follow to decompose generalized
(and possibly regularized) inverses of the effective channel matrix H described in Equation (4.6).
The entire process starts from a Toeplitz matrix and arrives at an efficient decomposition for the
14 Remember that H is a rectangular Toeplitz matrix (see Equation (4.6)).
 4.4. DFT-BASED REPRESENTATIONS OF BEZOUTIAN MATRICES 129
generalized inverse of such a matrix.The crucial stages of this strategy are the efficient decompositions
of Cauchy matrices and the efficient transformations of Bezoutian matrices into Cauchy matrices.

Toeplitz Operation Bezoutian Efficient

Matrix (·)† Matrix Transformation

Cauchy
Matrix
Factorized Efficient Factorized Efficient
Bezoutian Transformation Cauchy Decomposition
Matrix Matrix

Figure 4.2: Strategy to decompose generalized inverses of Toeplitz matrices.

This section describes how one can decompose a particular type of Cauchy matrix using
DFTs (Subsection 4.4.1) and, after that, how one can transform Bezoutian matrices into Cauchy
matrices in a simple and effective manner (Subsection 4.4.2). As a result, one arrives at an efficient
representation of Bezoutian matrices (Subsection 4.4.3).

4.4.1 REPRESENTATIONS OF CAUCHY MATRICES

The Cauchy matrices are known for their efficient decompositions. Let us consider a particular
type of rectangular Cauchy matrix Cν,λ with dimension M1 × M2 , in which the vectors ν ∈ CM1 ×1
and λ ∈ CM2 ×1 are comprised of the distinct M1 th roots of ν ∈ C and the M2 th roots of λ ∈ C,
respectively. That is,

 T
ν  ν0 ν1 · · · νM1 −1 ,
 T
λ  λ0 λ1 · · · λM2 −1 , (4.58)

m1 m2
where νm1  ν0 WM 1
, for all m1 within the set M1 , whereas λm2  λ0 WM 2
, for all m2 within the
set M2 (see Equation (4.40)). Let us also assume that M2 ≥ M1 . Thus, from expression (4.43), we
130 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
have
  1
Cν,λ m =
1 m2 1 − ν m 1 λm 2
1
=

m1 m2
1 − ν 0 WM λ0 W M2
 1
m1 m2 M 1 
1 − (ν0 λ0 WM W M2 ) 1
= m1 m2 M 1 ×
1

1 − (ν0 λ0 WM WM ) 1 − W m1
ν λ W m2
1 2 M1 0 0 M 2
⎡ ⎤
m1 m2 M 1
1 1 − (W ν λ W
0 0 M2 )
= m1 m2 M 1
⎣ M1
⎦
1 − (ν0 λ0 WM 1
W M 2
) 1 − W m1
ν λ
0 0 W m2
M1 M2

1 
M 1 −1
m
m1 m2
= m1 M1 m2 M 1 WM ν0 λ 0 W M
1 − (ν0 WM 1
) (λ0 WM 2
) m=0
1 2


M 1 −1

1 m1 m mm2
= WM (ν0 λ0 )m WM , (4.59)
1 − νλM
m2
1
m=0
1 2

in which we considered that ν and λ are such that

1 − νλm
M1
2
 = 0, ∀m2 ∈ M2 . (4.60)
mm2
Now, let us observe that, for each m within the set M1 , we have that WM is the mth entry
√ m1 m m
2
of the m2 th column of the matrix M2 WM2 , while WM1 (ν0 λ0 ) is the mth entry of the m1 th
√  
row of the matrix M1 WM1 diag{(ν0 λ0 )m1 }m1 ∈M1 0M1 ×(M2 −M1 ) . Therefore, we can rewrite
Equation (4.59) as
√
  M2 M1    
Cν,λ m m = M1
WM1 diag{(ν0 λ0 )m1 }m1 ∈M1 0M1 ×(M2 −M1 ) WM2 m m , (4.61)
1 2
1 − νλm2 1 2

yielding
,√ -
  M2 M1
Cν,λ = WM1 diag{(ν0 λ0 )m1 }m1 ∈M1 0M1 ×(M2 −M1 ) WM2 diag . (4.62)
1 − νλM
m2
1
m2 ∈M2

Equation (4.62) contains an efficient matrix factorization which is also useful to express
generalized Cauchy matrices. Indeed, as described in Subsection 4.3.2, we know that an M1 × M2
generalized Cauchy matrix Cν,λ g is such that
g 
C ν,λ ) =
Dν ,Dλ (C pr qrH , (4.63)
r∈R
 4.4. DFT-BASED REPRESENTATIONS OF BEZOUTIAN MATRICES 131
where R  {0, 1, · · · , R − 1}, pr ∈ CM1 ×1 and qr ∈ CM2 ×1 are vectors, and (in fact, this is the
main condition) R  M1 ≤ M2 . Observe that, by the definition of Stein displacement expressed in
Equation (4.19), we have
 
g g g
Dν ,Dλ (CC ν,λ ) = C ν,λ − Dν C ν,λ Dλ
m1 m 2  m1 m2 
g g
= C ν,λ − Dν C ν,λ Dλ
m1 m2   m1 m2  
g g
= C ν,λ − Dν m m C ν,λ Dλ m m
m1 m 2  1 1 m1 m2 2 2
g
= (1 − νm1 λm2 ) C ν,λ , (4.64)
m1 m2

which, along with Equation (4.63), implies

 
2 H
pr qr
 r∈R
g m1 m2
C ν,λ =
m1 m2
2 (1 − ν m1 m2 )
λ
∗
pm1 ,r qm 2 ,r
r∈R
=
(1 − νm1 λm2 )
 1
= pm1 ,r q∗
(1 − νm1 λm2 ) m2 ,r
r∈R 
   
= Dpr m m C ν,λ m m D∗qr
1 1 1 2 m2 m2
R
r∈ 

= Dpr C ν,λ D∗qr , (4.65)
r∈R m1 m2

in which, for each r ∈ R, we define the following diagonal matrices:

Dpr  diag{pr } ∈ CM1 ×M1 ,
Dqr  diag{qr } ∈ CM2 ×M2 . (4.66)
Therefore, based on Equations (4.62) and (4.65), we conclude that the generalized Cauchy
g
matrix C ν,λ admits the following efficient decomposition:
g   
C ν,λ = Dpr WM1 diag{(ν0 λ0 )m1 }m1 ∈M1 0M1 ×(M2 −M1 )
r∈R ,√ -
M2 M1
× WM2 D∗qr diag M1
. (4.67)
1 − νλm 2 m2 ∈M2

Equation (4.67) is a quite efficient way of decomposing a given generalized Cauchy matrix,
since the decomposition only employs DFT and diagonal matrices. What remains is to find a way
of transforming a given rectangular Bezoutian matrix into such a type of Cauchy matrix.
132 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
4.4.2 TRANSFORMATIONS OF BEZOUTIAN MATRICES INTO CAUCHY
MATRICES
This subsection describes a computationally efficient way to transform rectangular Bezoutian ma-
trices into rectangular Cauchy matrices in order to benefit from the fact that Cauchy matrices enjoy
efficient decompositions. The transformation which we shall describe is based on the displacement
operators of Sylvester and Stein.15
From a displacement-approach point of view, we know that Bezoutian and Cauchy matrices
have an inherent feature in common: their resulting displacement matrices have small rank, as
compared to the dimensions of the original matrices. At a first look, it seems that the similarities
end here, since Bezoutian matrices employ Sylvester displacements, whereas Cauchy matrices use
Stein displacements. In addition, the operator matrices utilized to compress both classes of matrices
are also of different types. These facts might erroneously lead us to conclude that the transformation
from Bezoutian to Cauchy matrices is not possible.
Fortunately, this is not the case, i.e., it is possible to transform Bezoutian into Cauchy matrices.
Indeed, the difference between the Sylvester and Stein displacements is not a problem since we know
how to convert one into another, as explained in Equations (4.46) and (4.47). With regard to the
difference between the operator-matrix types, a Bezoutian matrix uses ν-circulant operator matrices,
such as Zν , whereas a Cauchy matrix uses diagonal operator matrices, let us say Dν . Even though
these operator matrices are quite different, there is a simple relationship between them which will
help us link Bezoutian to Cauchy matrices.
Consider the particular case in which vector ν is comprised of all distinct Mth roots of ν ∈ C,
i.e.,
 T
ν  ν0 ν1 · · · νM−1 , (4.68)

1  ν
where ν0 = |ν| M ej M and νm = ν0 WM m , for all m within the set M. Now, by remembering the

effect of left-multiplying a Vandermonde matrix Vν by a diagonal matrix Dν , as described in Equa-

M = ν, for any m in M), we see that the resulting matrix D V is equal
tion (4.39) (recall that νm ν ν
to the matrix Vν Zν , thus relating the diagonal operator matrix to the ν-circulant operator matrix
through the use of a Vandermonde matrix.
Such a fact can also be derived in a more formal way as follows: first consider that m1 ∈
  m (m +1)
M \ {M − 1}. Thus, Dν V ν m2 m1 = νm2 WM m2 m 1 m1
ν0 = ν0 WM WM ν0 = WM 2 1 ν0m1 +1 =
m2 m2 m 1 m1
 
V ν ]m2 (m1 +1) . Now consider that m1 = M − 1. In this case, we have that Dν V ν m (M−1) =
[V
2
m (M−1) M−1 m2 M M
νm2 WM 2 ν0 = WM ν0 = ν = ν [V
V ν ]m2 0 . We can therefore conclude that

V ν Zν = Dν V ν (4.69)

15The results derived here are heavily based on the outstanding works developed by Heinig and Rost, for example, in [27]. These
authors, however, employed a polynomial formulation to develop most of their results.
 4.4. DFT-BASED REPRESENTATIONS OF BEZOUTIAN MATRICES 133
and, for the useful case where ν  = 0, one has

Zν = V ν−1 Dν V ν . (4.70)

We can now proceed to derive the expected transformation from Bezoutian to Cauchy ma-
trices. We know that, given two natural numbers M1 and M2 , with M1 ≤ M2 , and given two non-
zero complex numbers ξ and η, an M1 × M2 complex-valued Bezoutian matrix B is such that
∇Zξ ,Zη (B
B ) = PQ H , where the operator matrices have compatible dimensions and the resulting
displacement rank is much smaller than M1 . Our first task is to change from the Sylvester to the
Stein displacement and, after that, to convert the circulant operator matrices into diagonal ones.
Thus, by defining η  1/η and by using the relations in Equations (4.22), (4.47), and (4.70), we
have

PQ H
= ∇Zξ ,Zη (B
B)
= Zξ B − B Z η
= −Zξ ,ZT (B B )Zη
1/η

= − B − Zξ B ZηT Zη

= − B − V ξ−1 Dξ V ξ B V ηT Dη V η−T Zη



V ξ−1 V ξ B V ηT − Dξ V ξ B V ηT Dη V η−T Zη
= −V
V ξ−1 Dξ ,Dη (V
= −V V η−T Zη ,
V ξ B V ηT )V (4.71)

implying that

V ξ B V ηT ) = −V
Dξ ,Dη (V V ξ PQ H ZηH∗ V ηT
 
H
= −V V ξ P V η∗ Zη∗ Q
     
P Q H
H
=PQ
  H
= p 0 p 1 · · · p R−1 q0 q1
· · · qR−1 , (4.72)

in which R is a natural number such that M1  R ≥ rank{∇Zξ ,Zη (B B )}.

Equation (4.72) means that matrix V ξ B V ηT is a generalized Cauchy matrix, since its related
Stein displacement matrix is such that

V ξ B V ηT )} = rank{∇Zξ ,Zη (B
rank{Dξ ,Dη (V B )} ≤ R  M1 . (4.73)

4.4.3 EFFICIENT BEZOUTIAN DECOMPOSITIONS

We are now in a position to describe how to decompose a given rectangular Bezoutian matrix by
employing computationally efficient matrices like DFT, IDFT, and diagonal matrices. Indeed, given
134 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
the expression in Equation (4.72), we can apply the result to Equation (4.67), replacing ν by ξ , λ by
g
η = 1/η, pr by p r , qr by qr , and C ν,λ by V ξ B V ηT . The resulting expression is
  
V ξ B V ηT = Dp r WM1 diag{(ξ0 η0 )m1 }m1 ∈M1 0M1 ×(M2 −M1 )
r∈R ,√ -
M2 M1
× WM2 D∗q diag , (4.74)
r
1 − ξ ηM
m2
1
m2 ∈M2

which yields

  
V ξ−1
B =V Dp r WM1 diag{(ξ0 η0 )m1 }m1 ∈M1 0M1 ×(M2 −M1 )
r∈R ,√ - 
M2 M1
×WM2 D∗q diag V η−T . (4.75)
r
1 − ξ ηM
m2
1
m2 ∈M2

Equation (4.75) represents the last step of the displacement approach, namely the decom-
pression operation. It is worth mentioning that such expression is a rather efficient decomposition
of a general rectangular Bezoutian matrix, since it employs only DFT and diagonal matrices, and
considering that a given Vandermonde matrix can be described as a product of a DFT and a diagonal
matrix, as described in Equation (4.42).

Example 4.3 (Bezoutian Factorization) Let us analyze a numerical toy example which will help
us better understand the computations involved in Bezoutian factorizations. Consider the following
4 × 3 Toeplitz matrix
⎡ ⎤
2 1 0
⎢3 2 1⎥
T =⎢ ⎣0 3 2⎦ ,
⎥ (4.76)
0 0 3
 −1 H
whose associated Moore-Penrose pseudo-inverse T † = T H T T is given by16
⎡ ⎤
0.1633 0.2245 −0.2041 0.0612
T † ≈ ⎣−0.0028 0.0019 0.3330 −0.2226⎦ . (4.77)
−0.0334 0.0223 −0.0037 0.3284

As discussed in Subsection 4.3.3, matrix T † is a particular type of Bezoutian matrix. We could

therefore apply the result described in Equation (4.75), with M1 = 3 and M2 = 4.17 Assume, for
16The symbol “≈” is employed rather than “=” since the entries of T † are in fact real numbers with more than four decimal digits
of precision.
17 Note that neither M nor M is much larger than the displacement rank of T † . Nevertheless, the obtained factorization does not
1 2
depend on having R  min{M1 , M2 }. Thus, even though it is a bit misleading to say that T † is a Bezoutian matrix due to its
small dimensions, we still can apply the result in Equation (4.75).
 4.4. DFT-BASED REPRESENTATIONS OF BEZOUTIAN MATRICES 135
instance, that ξ = 2 and η = 1 = η. With such assumptions, most of the√matrices appearing in
the decomposition in Equation (4.75) can be readily defined, where ξ0 = 3 2 (note that  ξ = 0),
√ 2π 3π
η0 = 4 1 = 1 ( η = 0 as well), and ηM m2 M1
m2 = (η0 WM2 )
1
= (e−j 4 m2 )3 = e−j 2 m2 = jm2 , for each
m2 ∈ {0, 1, 2, 3}. Thus, by using Equation (4.42), we have
√

m1
m1 =2 −1
V ξ−1 = 3W3 diag{2 3 }m 1 =0
1 m
− 31 m1 =2

= √ diag{2 }m1 =0 W3H

3
⎡ ⎤⎡ ⎤
1 0 0 1 1 1
1⎢ 1 ⎥⎢ 2π 4π ⎥
= ⎣0 2 − 3 0 ⎦ ⎣1 e j 3 ej 3 ⎦ , (4.78)
3 1 4π 2π
0 0 4− 3 1 ej 3 ej 3
√

m2
m2 =3 −T
V η−T = 4W4 diag{1 4 }m 2 =0
1 H
= W4
2⎡ ⎤
1 1 1 1
1 ⎢1 j −1 −j ⎥
= ⎢ ⎥. (4.79)
4 ⎣1 −1 1 −1⎦
1 −j −1 j

In addition, we have
⎡ ⎤
1 0 0 0
  ⎢ 1 ⎥
diag{(ξ0 η0 )m1 }m1 ∈M1 0M1 ×(M2 −M1 ) = ⎣0 2 3 0 0⎦ , (4.80)
1
0 0 43 0

and
⎡ ⎤
,√ - −1 0 0 0
M2 M1 √ ⎢
⎢0
1
1−2j 0 0 ⎥
⎥
diag = 2 3⎢ 1 ⎥. (4.81)
1 − ξ ηM
m2
1 ⎣0 0 3 0 ⎦
m2 ∈M2 1
0 0 0 1+2j

The reader should note that we have just defined almost all matrices to be employed in the
factorization, but we have not used any information about matrix T yet. Indeed, all information of T is
contained in the entries of the diagonal matrices Dp r and Dqr , with r ∈ {0, 1, · · · , R − 1}. Thus, we
have to determine vectors p r and qr based on Equation (4.72). But, before that, we have to determine
matrices P and Q based on Equation (4.57). These latter matrices, on the other  hand, depend on
the knowledge of the displacement matrices ∇Zη ,Zξ (T T ) = P̂Q̂ H and ∇Zξ ,Zη T H = P̌Q̌ H , which
136 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
can be straightforwardly acquired using Equation (4.37), as follows:
⎡ ⎤
1 −4
⎢0 −6⎥
P̂ = ⎢ ⎥
⎣0 1 ⎦ , (4.82)
0 2
⎡ ⎤
−1 0
Q̂ = ⎣ 0 0⎦ , (4.83)
3 1

and
⎡ ⎤
1 −2
P̌ = ⎣0 −1⎦ , (4.84)
0 0
⎡ ⎤
−3 0
⎢ 2 0⎥
Q̌ = ⎢ ⎥
⎣ 4 0⎦ . (4.85)
6 1

Now, Equation (4.57) gives us

⎡ ⎤
0.1224 −0.1633 −0.1633 2.0816
P ≈ ⎣−0.0816 0.0028 0.0028 0.1122 ⎦ , (4.86)
0.0204 0.0334 0.0334 −0.6531
⎡ ⎤
−2.0288 0.1002 −0.2635 −0.0334
⎢ 1.3525 −0.0668 −0.1577 0.0223 ⎥
Q ≈⎢ ⎣−0.2254 0.0111
⎥. (4.87)
0.1929 −0.0037⎦
−0.3006 0.0148 0.9239 0.3284

Thus, using Equation (4.72), we have R = 4 and

⎡ ⎤
−0.0520 0.1067 0.1067 −1.1864
P ≈ ⎣−0.1577 − 0.1171j 0.1915 − 0.0429j 0.1915 − 0.0429j −2.5293 + 1.0203j⎦ ,
−0.1577 + 0.1171j 0.1915 + 0.0429j 0.1915 + 0.0429j −2.5293 − 1.0203j

(4.88)
⎡ ⎤
−1.2022 0.0594 0.6957 0.3135
⎢−1.6531 − 1.8033j 0.0816 + 0.0891j 1.0816 − 0.4564j 0.3061 − 0.0297j⎥
Q ≈⎢
⎣
⎥.
⎦
3.3061 −0.1633 0.8367 0.3878
−1.6531 + 1.8033j 0.0816 − 0.0891j 1.0816 + 0.4564j 0.3061 + 0.0297j
(4.89)
 4.5. REDUCED-REDUNDANCY SYSTEMS 137
By using matrices P
and Q we can define the diagonal matrices Dp r and Dqr , with r ∈ {0, 1, 2, 3}.
For example, we have
⎡ ⎤
0.0594 0 0 0
⎢ 0 0.0816 + 0.0891j 0 0 ⎥
Dq1 ≈ ⎢⎣ 0
⎥.
⎦ (4.90)
0 −0.1633 0
0 0 0 0.0816 − 0.0891j

It is possible to verify that Equation (4.75) holds for this particular case using the numerical values
in Equations (4.78), (4.79), (4.80), (4.81), (4.88), and (4.89). 

4.5 REDUCED-REDUNDANCY SYSTEMS

This section addresses the issue of designing memoryless LTI transceivers with reduced redun-
dancy which are amenable to superfast implementations at the transmitter and receiver ends. The
transceivers are generated by exploiting the displacement structure of the Toeplitz equivalent channel
matrix, thus yielding systems whose corresponding asymptotic computational complexity is com-
parable to the complexity of the conventional OFDM and SC-FD transceivers with respect to the
equalization process. The reduced-redundancy transceivers may lead to higher data throughputs
when compared to OFDM and SC-FD systems, due to the reduced amount of redundancy which
is employed.
As pointed out before, the equivalent channel matrix H associated with reduced-redundancy
ZP-ZJ systems is a rectangular Toeplitz matrix (see Equation (4.6)). In addition, Equations (4.55)
and (4.57) show that any (regularized) Moore-Penrose pseudo-inverse of a given full-rank Toeplitz
matrix is indeed a Bezoutian matrix. Moreover, if we define the transmitter-independent receiver
matrix K as

K  F G ∈ CM×(M+2K−L) , (4.91)

with 2K ≥ L, then one can easily verify based on Equations (4.16) and (4.17) that
†
KZF = H , (4.92)
# $−1
H σv2 H
KMMSE = H H + 2 IM H (4.93)
σs

are (possibly regularized) pseudo-inverses of the tall channel Toeplitz matrix H, thus implying that
the transmitter-independent receiver matrix is also a particular type of rectangular Bezoutian matrix.
Hence, we can tailor the efficient decompositions of Bezoutian matrices previously described
in Subsection 4.4.3 to the particular cases of MMSE and ZF transmitter-independent receiver
matrices. As a result, a family of superfast multicarrier and single-carrier linear transceivers are
derived with their respective structures.
138 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
Firstly, we must carefully choose both parameters ξ and η which appear in the defini-
tion of the operator matrices employed in the compression stage of both the Toeplitz chan-
π
nel matrix and the receiver Bezoutian matrix. Let us assume that ξ = 1 and η = ej M . In this
case, by using Equation (4.37) one can readily compute the displacement-generator pair of ma-
trices (P̂, Q̂ ) ∈ C(M+2K−L)×2 × CM×2 and (P̌, Q̌
) ∈ CM×2 × C(M+2K−L)×2 Hassociated

with
H
matrices H and H , respectively, where ∇Zη ,Zξ H = P̂Q̂ and ∇Zξ ,Zη H
H = P̌Q̌ . By
H

using such displacement-generator pairs, one can precompute the displacement-generator pair
(P, Q ) ∈ CM×4 × C(M+2K−L)×4 associated with the transmitter-independent receiver matrix,
where ∇Zξ ,Zη (K) = PQ H . Indeed, for the MMSE solution, Equation (4.55) gives us

−1 
H σv2
P = H H + 2 IM P̌ −KMMSE P̂ , (4.94)
σs

−H 
2 H 2
Q = σv2 H H + σv2 I(M+2K−L) H
Q̌ KMMSE Q̂ , (4.95)
σs σs

whereas, for the ZF solution, Equation (4.57) gives us


−1 
H
P= H H P̌ −KZF P̂ , (4.96)

H 
Q = I(M+2K−L) − HKZF Q̌ H Q̂
KZF . (4.97)

Thus, by considering the previous compressed forms of the Bezoutian matrix K for either the
MMSE or the ZF solutions, one can employ the factorization in Equation (4.75), which depends
on the Vandermonde matrices Vξ and Vη . Since ξ = 1 and η = 1/η = e−j M , then
π

 ξ
ξ0 = |ξ |1/M ej M
= 1 × e0
=1 (4.98)
and
 η
η0 = |η|1/(M+2K−L) ej (M+2K−L)
−jπ
= 1 × e M(M+2K−L)
−jπ
= e M(M+2K−L) . (4.99)
These facts, along with Equation (4.42), imply that
√
Vξ = MWM , (4.100)
whereas
√ −jπ m
(M+2K−L−1)
Vη = M + 2K − L × W(M+2K−L) diag{e M(M+2K−L) }m=0 . (4.101)
 4.5. REDUCED-REDUNDANCY SYSTEMS 139
We can therefore apply the decomposition presented in Equation (4.75) to obtain the following
result:
 3 
  
K = WMH
Dpr WM DM 0M×(2K−L) W(M+2K−L) D∗q W(M+2K−L) H
DH(M+2K−L) ,
r
r=0
(4.102)
in which the diagonal matrices DM and D(M+2K−L) do not depend on K and are, respectively,
defined as
 M−1
−jπ m
DM  diag e M(M+2K−L) , (4.103)
 m=0
M+2K−L−1
−jπ m
D(M+2K−L)  diag e M(M+2K−L) . (4.104)
m=0

In addition, based on Equations (4.72) and (4.75), the pair of matrices (P, Q ) ∈ CM×4 ×
C(M+2K−L)×4 are defined as
 
P  p0 p1 p2 p3
= −V V P, (4.105)
 ξ 
Q  q0 q1 q2 q3
  
M+2K−L−1
1
= diag V η∗ Zη∗ Q , (4.106)
1 − ξ ∗ (η∗m )M m=0

where P and Q are defined in Equations (4.94) and (4.95) for the MMSE solution, and in Equa-
tions (4.96) and (4.97) for the ZF solution, whereas ξ = 1 and η = e−j M .
π

Note that the choices of ξ and η were quite arbitrary. We have chosen ξ = 1, since, in the
case of multicarrier systems, we would like to cancel out the last IDFT matrix appearing in the
decomposition of matrix K described in Equation (4.102). Indeed, in the multicarrier systems, the
receiver matrix is G = WM K. If ξ = 1, one would not be able to cancel out the DFT matrix with
the last IDFT matrix presented in the decomposition of K, due to the presence of an additional
diagonal matrix. After fixing ξ = 1, we have chosen η in such a way that 1 − ξ ηM m = 1 − ηm  = 0,
M

for all m within the set {0, 1, · · · , M + 2K − L − 1}, where η = 1/η. There are infinite possible
π
choices for η and we have arbitrarily chosen η = ej M . Note that, for this choice of η, one has

m = (η0 W(M+2K−L) )
ηM m M
−jπ −j2π mM
= e (M+2K−L) e (M+2K−L)
−jπ(2mM+1)
= e (M+2K−L)
 = 1, (4.107)

for all m within the set {0, 1, · · · , M + 2K − L − 1}, since 2mM+1

M+2K−L is not an even number.
140 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
for all m within the set {0, 1, · · · , M + 2K − L − 1}, since 2mM+1
M+2K−L is not an even number.
−1
Thus, a multicarrier system can be designed by setting F = WM
H and G = F K = WM K,
yielding
 3 
  
G= Dpr WM DM 0M×(2K−L) W(M+2K−L) D∗q H
W(M+2K−L) DH
(M+2K−L) , (4.109)
r
r=0

where the definitions of vectors pr and qr depend on whether the ZF or the MMSE solution is
chosen. In any case, the resulting multicarrier structure is depicted in Figure 4.3.

Noise
Add Remove
Data Guard Phase
IDFT Channel Guard
Block Period IDFT
Period Shift
P/S S/P

One-Tap Phase
DFT One-Tap
Equalizer Shift DFT
Equalizer
Ignore

One-Tap Phase
DFT One-Tap
Equalizer Shift DFT
Data Equalizer
Block Ignore
Estimate
One-Tap Phase
DFT One-Tap
Equalizer Shift DFT
Equalizer
Ignore

One-Tap Phase
DFT One-Tap
Equalizer Shift DFT
Equalizer
Ignore

Figure 4.3: Multicarrier reduced-redundancy block transceiver (MC-RRBT).

Figure 4.3 illustrates that, after removing the guard period, there are M + 2K − L phase shifts
to be performed in parallel, where the mth phase rotation is implemented through the multiplication
π
by ej M(M+2K−L) m . The first equalization step is performed after applying the IDFT to the data vector.
The resulting data vector is simultaneously processed by four different branches at the receiver. The
equalizers at this point consist of single-tap equalizers. The coefficients of the equalizers are the
entries of vectors q∗r , with r ∈ {0, 1, 2, 3}. After the DFT application, the last 2K − L elements
are discarded and new phase shifts are performed on the remaining M elements, where in this case
−j π m
the mth rotation consists of a multiplication by e M 2 . The final stage of the equalization consists
 4.5. REDUCED-REDUNDANCY SYSTEMS 141
of applying DFTs in parallel to the output signal of the rotators, followed by another single-tap
equalization stage whose coefficients are the entries of vectors pr , with r ∈ {0, 1, 2, 3}.
−1
Also, a single-carrier system can be designed by setting F = IM and G = F K = K, yielding
 3 
   ∗
G = WM H
Dpr WM DM 0M×(2K−L) W(M+2K−L) Dq W(M+2K−L) H
DH
(M+2K−L) ,
r
r=0
(4.110)

in which, once again, the definitions of vectors pr and qr depend on whether the ZF or the MMSE
solution is chosen.
The superfast multicarrier and single-carrier transceivers of this chapter yield an additional
degree of freedom in the ZP-ZJ-based transmissions, since the amount of redundancy can vary from
the minimum value, L/2, to the most commonly used value, L.18 Nonetheless, one must deal with
two distinct DFT sizes, M and M + 2K − L. When M is a power of 2, then M + 2K − L is not
necessarily a power of two. Thus, a radix-2 FFT algorithm could only be applied to implement those
DFTs with size M. As for the DFTs with size M + 2K − L, one could implement the operations
using a radix-2 FFT of size 2M (which is assumed to be larger than M + 2K − L), along with zero-
padding of the related signals. Another possibility is to choose the amount of redundant elements
in such a way that M + 2K − L can be decomposed as a product of small prime numbers, leading
to fast implementations as well. This is an open research topic.

4.5.1 COMPLEXITY COMPARISONS

This subsection describes some possible complexity comparisons between OFDM-based systems
2 log2 M − 2 + 2
3M
and superfast ZP-ZJ systems. Firstly, assume that an FFT algorithm requires M
complex-valued multiplications for processing size-M data blocks, as described in [17]. In addition,
consider that L = M 4 , as performed in [55]. In this case, it is possible to show that the number of
complex-valued multiplications required by the superfast multicarrier reduced-redundancy system,
by the overlap-and-add (OLA), and by fast proposals of zero-padded OFDM systems described
in [55] is given by:

• ZP-OFDM-OLA: M log2 M − 2M + 4.

• ZP-OFDM-FAST: 5M
4 log2 M − 5M + 20.

• MC-RRBT: 15M
2 log2 M − 9M
2 + 20 + 5(2K − L).

In the MC-RRBT, it is possible to implement part of the receiver side using parallel processing
(see Figure 4.3). In this case, if we consider that the time that is required to perform a generic complex-
valued multiplication is T ∈ R+ s, then the MC-RRBT requires T (3M log2 M + 2(2K − L) + 8) s,
18 It is worth highlighting that Figure 4.3 is a way to implement general ZP-OFDM systems (K = L) which allows for superfast
equalization of received data blocks.
142 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
whereas the ZP-OFDM-OLA requires
T (M log2 M − 2M + 4) s, and the ZP-OFDM-FAST
requires T 4 log2 M − 5M + 20 s.
5M

It is worth noting that, in the minimum-redundancy case where K = L/2 (assuming a channel
with even order), the ZF solution can be further simplified. Indeed, as 2K − L = 0, then KZF =
−1
H , implying that Equation (4.97) can be rewritten as Q = [0M×2 KZF H Q̂ ]. As the product

PQ H defines completely the displacement matrix ∇Zξ ,Zη (KZF ), then we can also discard the first
two columns of P, since they will be multiplied by zero anyway. Thus, for ZF minimum-redundancy
systems, we can rewrite Equations (4.96) and (4.97) as follows:

P = −KZF P̂, (4.111)

Q = KZF
H
Q̂ . (4.112)

Such a result could be alternatively obtained by using Equation (4.50). An important consequence
of this result is that all DFT matrices are M × M now and that the number of equalizer branches
at the receiver end changes from four to two branches.
We assumed that the pair of generator matrices (P, Q ) is known. In fact, these matrices
completely define the reduced-redundancy equalizers, since they are the only ones that contain
information about the channel. These matrices, however, must be previously computed in the so-
called receiver-design stage, which could be performed using up to O(M log22 M) operations, as
described, for example, in [51]. Besides, there are many applications in which the receiver-design
problem is not frequently solved. In wireline communications systems, the channel model is not
updated so often. In this case, the main computational burden refers to the equalization itself.
Reducing the amount of operations in the receiver-design stage is still an interesting and
rather important research direction in the context of superfast reduced-redundancy systems based
on the ZP-ZJ model.

4.5.2 EXAMPLES
Up to this point, we showed efficient structures for reduced-redundancy systems allowing the equal-
ization of received data blocks in a superfast manner. Nevertheless, we should note that ZP-ZJ
systems do not necessarily outperform OFDM-based transceivers. In fact, the authors in [54] have
proved mathematically that the MSE performance of reduced-redundancy transceivers degrades as
the number of transmitted redundant elements reduces. That is, transmissions using larger amounts
of redundancy leads to lower average MSE of symbols in ZP-ZJ systems.
If on one hand, we want to reduce the transmitted redundancy in order to save bandwidth;
on the other hand, we need to use as much redundancy as possible in order to have a good MSE
performance. The throughput is a good figure of merit to study the tradeoff between bandwidth
 4.5. REDUCED-REDUNDANCY SYSTEMS 143
usage and error performance. Thus, let us analyze the throughput performance of superfast ZP-ZJ
systems through some simulation examples.

Example 4.4 (Minimum-Redundancy Transceivers) The figure of merit adopted here is the
throughput, defined as

M
Throughput = brc (1 − BLER)fs bps, (4.113)
M +K
in which b denotes the number of bits that is required to represent one constellation symbol, rc
denotes the code rate considering the protection of channel coding, K denotes the amount of
redundancy, fs denotes the sampling frequency, where symbol and channel models use the same
sampling frequency, and BLER stands for block-error rate, assuming that a data block is discarded
when at least one of its original bits is incorrectly decoded at the receiver end. In addition, the
definition of the SNR used throughout the simulations is the ratio between the mean energy of the
transmitted symbols at the input of the multipath channel and the power-spectral density of the
additive noise at the receiver front-end. Besides, we also consider that both synchronization and
channel estimation are perfectly performed at the receiver end.
In this example, we transmit 200 blocks, each one containing M = 32 BPSK data symbols
(without taking redundancy into account), and compute the throughput by using a Monte-Carlo
averaging process with 10,000 simulations. Consider these symbols are sampled at a frequency
fs = 1.0 GHz and that they are transmitted through a channel with a model operating at the same
frequency as the symbols and with a long impulse response of order L = 30.19 All the channel
 taps
have the same variance, and the channel model is always normalized, that is, E h2 = 1. Both
2

the imaginary and real parts of the channel are independently drawn from a white and Gaussian
random sequence (random Rayleigh channel). For each simulation a new channel is generated.
Furthermore, since the ZP-ZJ transceivers use zeros as redundant elements, the adopted
OFDM and SC-FD systems in the simulations are the ZP-OFDM-OLA and ZP-SC-FD-
OLA [55], where ZP and OLA stand for zero-padding and overlap-and-add, respectively, (see
Subsections 2.4.2 and 2.4.4). Like the traditional cyclic-prefix-based systems, these ZP-based
transceivers also induce a circulant channel matrix. We chose these transceivers as benchmarks
since they are superfast transceivers that transmit L redundant zeros for each M data symbols. In
summary, from now on we shall consider that OFDM means ZP-OFDM-OLA and SC-FD means
ZP-SC-FD-OLA in all results here.
Figures 4.4 and 4.5 show the throughput curves for the OFDM, the SC-FD, the multi-
carrier minimum-redundancy block transceiver (MC-MRBT), and the single-carrier minimum-
redundancy block transceiver (SC-MRBT), using both ZF and MMSE designs. By observing these
figures it is possible to verify that the minimum-redundancy superfast transceivers outperform the
traditional OFDM and SC-FD transceivers in this particular setup, except for SNRs lower than
19 We consider that L = 30 corresponds to a long impulse response since L is very close to M in this case.
144 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY

350

300

250
Throughput [Mbps]

200

150

100

ZF−OFDM
50 MMSE−OFDM
ZF−MC−MRBT
MMSE−MC−MRBT
0
0 5 10 15 20 25 30
SNR [dB]

Figure 4.4: Throughput [Mbps] as a function of SNR [dB] for random Rayleigh channels, considering
multicarrier transmissions.

12 dB in the ZF solutions. In this example, we use a convolutional code with constraint length 7,
rc = 1/2, and generators g0 = [133] (octal) and g1 = [165] (octal). This configuration is adapted
from the LTE (long-term evolution) specifications [94]. In addition, for the BLER computation,
we consider that a block (16 bits) is lost if, at least, one of its received bits is incorrect. We have
employed a MATLAB implementation of a hard-decision Viterbi decoder. Note that such favor-
able result stems from the choices for M and L, representing delay constrained applications in very
dispersive environment. These types of applications are suitable for the ZP-ZJ transceivers. In the
cases where M  L, the traditional OFDM and SC-FD solutions are more adequate. 

Example 4.5 (Reduced-Redundancy Transceivers) In Example 4.4, we have shown that

minimum-redundancy systems may significantly improve the throughput performance of multi-
carrier and single-carrier transmissions. Nevertheless, minimum-redundancy transceivers may also
incur in high noise enhancements induced by the “inversion” of the Toeplitz effective channel matrix
in the equalization process. In this example, we first chose a fourth-order channel model (see [83],
 4.5. REDUCED-REDUNDANCY SYSTEMS 145
pp. 306–307)

HA (z)  0.1659 + 0.3045z−1 − 0.1159z−2 − 0.0733z−3 − 0.0015z−4 (4.114)

for which the throughput performance of the minimum-redundancy systems is poor. For this channel
(Channel A), we transmit 50,000 data blocks carrying M = 16 symbols of a 64-QAM constellation
(b = 6 bits per symbol). In fact, each data block stems from 48 data bits that, after channel coding,
yield 96 bits to be baseband modulated. The channel coding is the same as in Example 4.4 and we
assume that the sampling frequency is fs = 100 MHz.
Figure 4.6 depicts the obtained throughput results. We compare four different transceivers:
the ZP-OFDM-OLA and the three possible multicarrier reduced-redundancy block transceivers
(MC-RRBT).There are three possible MC-RRBT systems since the amount of redundant elements
respects the inequality L2 ≤ K ≤ L (i.e., K ∈ {2, 3, 4}). From Figure 4.6, one can observe that
the minimum-redundancy multicarrier system (MC-RRBT for K = 2) that employs an MMSE
equalizer is not able to produce a reliable estimate for the transmitted bits. However, if just one

350

300

250
Throughput [Mbps]

200

150

100

ZF−SC−FD
50 MMSE−SC−FD
ZF−SC−MRBT
MMSE−SC−MRBT
0
0 5 10 15 20 25 30
SNR [dB]

Figure 4.5: Throughput [Mbps] as a function of SNR [dB] for random Rayleigh channels, considering
single-carrier transmissions.
146 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY

300

250

200
Throughput [Mbps]

MMSE−OFDM
MMSE−MC−RRBT (K = 2)
150
MMSE−MC−RRBT (K = 3)
MMSE−MC−RRBT (K = 4)

100

50

0
15 20 25 30 35
SNR [dB]

Figure 4.6: Throughput [Mbps] as a function of SNR [dB] for Channel A, considering multicarrier
transmissions.

additional redundant element is included in the transmission, the resulting MC-RRBT system (K =
3) is enough to outperform the MMSE-OFDM. One should bear in mind that such throughput
gains are attained without increasing substantially the computational complexity related to OFDM-
based systems. Moreover, the MC-RRBT system with K = 3 also outperforms the MC-RRBT
system with K = 4 in terms of throughput, especially for large SNR values, i.e., adding another
redundant element in the transmission (MC-RRBT for K = 4) does not contribute to improving
the throughput performance in this case.
Now, we will consider an FIR-channel model (Channel B) whose zeros are
0.999, −0.999, 0.7j, −0.7j, and −0.4j. This channel has zeros very close to the unit circle. We there-
fore expect that the performance of the traditional SC-FD system should be rather poor. Apart from
the channel model, all simulation parameters are the same of the previous experiment. Figure 4.7
depicts the throughput results. One can observe that the SC-RRBT systems always outperform the
traditional SC-FD system. Another important fact is that the throughput performance does not
necessarily improves as the number of transmitted redundant elements is increased. For example, for
low SNR values, it is better to use a reduced-redundancy system that transmits with a large number of
 4.6. CONCLUDING REMARKS 147

200

180

160

140
Throughput [Mbps]

120

100

80

60
MMSE−SC−FD
MMSE−SC−RRBT (K = 3)
40 MMSE−SC−RRBT (K = 4)
MMSE−SC−RRBT (K = 5)
20
15 20 25 30 35
SNR [dB]

Figure 4.7: Throughput [Mbps] as a function of SNR [dB] for Channel B, considering single-carrier
transmissions.

redundant elements (K = 5), whereas for large SNR values, it is better to use a reduced-redundancy
system that transmits with a small number of redundant elements (K = 3). Once again, it is im-
portant to highlight that the superfast ZP-ZJ systems described in this chapter are just examples of
how to transmit with a small number of redundant elements while using superfast transforms and
single-tap equalizers. 

4.6 CONCLUDING REMARKS

This chapter described how to design memoryless LTI single-carrier and multicarrier transceivers
with reduced redundancy for both ZF and MMSE optimal receivers. The block transceivers pre-
sented here are computationally efficient. We also introduced the mathematical tools which allow
structured matrices to be represented through displacement operators. In particular, we emphasized
the representation of Bezoutian matrices employing DFT, IDFT, and diagonal matrices. It is worth
mentioning that similar decomposition is possible by using transforms with real entries such as those
148 4. MEMORYLESS LTI TRANSCEIVERS WITH REDUCED REDUNDANCY
described in [48, 52] based on discrete Hartley transform (DHT), diagonal, and antidiagonal ma-
trices. In addition, the complexity of the efficient transceivers described here can be further reduced
by employing suboptimal solutions, as shown in [50].
As described in previous chapters, OFDM-based transceivers are rather efficient systems
incorporating two desired features in practical systems, namely good/fair performance and compu-
tational simplicity. Nevertheless, there is always room for improvements. We showed in this chapter
how to increase the transmission data-rates in block transceivers, while keeping the computational
complexity close to the complexity of OFDM-based systems. We believe that the strategy followed
in this chapter is even more important than the described transceivers themselves, since it illus-
trates how one can deduce new systems using a set of efficient tools related to structured matrix
representations.
An important question still remains: what happens when we move from memoryless LTI to
time-varying FIR systems? Is it possible to reduce even further the amount of transmitted redun-
dancy? These and other questions related to what is beyond OFDM-based systems will be addressed
in Chapter 5.
 149

CHAPTER 5

FIR LTV Transceivers with

Reduced Redundancy
5.1 INTRODUCTION
Nowadays, many practical communications systems employ the orthogonal frequency-division mul-
tiplexing (OFDM) as their core physical-layer modulation, as previously highlighted in Section 1.3.
Such widespread adoption is due to many good properties that OFDM-based systems enjoy, as
thoroughly explained in Chapter 3. However, one of the main drawbacks of OFDM is its related
loss of spectral efficiency caused by the insertion of redundant elements in the transmission. This is
an important issue considering the current trend of increasing the demand for data transmissions
which shows no sign of settling. For example, the amount of wireless data services is more than
doubling each year [37] leading to spectrum shortage as a sure event soon. It turns out that all efforts
to maximize the spectrum usage are therefore highly justifiable at this point.
The memoryless linear time-invariant (LTI) transceivers with reduced redundancy described
in Chapter 4 are a possible way of tackling the problem of increasing the spectral efficiency. Indeed,
we show in Chapter 4 that, as compared to OFDM-based systems, reduced-redundancy LTI systems
may decrease the amount of transmitted redundant elements in up to fifty percent, thus allowing
a better use of the available spectrum for transmissions. However, this amount of reduction may
still not be enough, especially in delay-constrained applications in very dispersive environments, in
which the size of the transmitted data block cannot be too large and, in addition, the channel model
is lengthy. Hence, we should ask ourselves if the memoryless LTI-based solutions could be further
improved.
The transceivers described in Chapters 2, 3, and 4 are concrete examples of memoryless,
linear, and time-invariant systems. Linearity plays a central role in the overall system design since it
yields simpler transceivers, a very desirable feature in practice. A step ahead would be to investigate
how to design simple time-varying transceivers with memory implemented through time-varying
finite-impulse response (FIR) filters.1 The memory and time-varying properties would introduce
additional degrees of freedom in the design of systems that could help us circumvent some limitations
inherent to memoryless LTI transceivers.
In fact, the time-varying FIR transceiver increases the transmission diversity, since more than
one version of the transmitted message is received due to the memory of the system, and, in addition,
1 In this case, we consider that the orders of the related filters are larger than or equal to 1. A memoryless system can be regarded
as an FIR system whose order is equal to 0.
150 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
these versions are more likely to differ from each other due to the time-varying characteristic. If the
receiver can take these facts into account, then it could trade off the amount of transmitted redundant
elements with the amount of memory and degree of time-variance in the systems. Moreover, FIR
linear time-varying (LTV) transceivers allow the design of high-selective transmitter filters and the
interpretation of code-division multiple access (CDMA) schemes with long codes as a particular
type of LTV transmultiplexer (TMUX), as will be explained later on.
It is worth recalling that intersymbol interference (ISI) is one of the most harmful effects
inherent to broadband communications. That is why communication engineers often constrain
their designs to eliminate ISI, especially when noise and other types of interference are negligible,
leading to the so-called zero-forcing (ZF) transceivers.This chapter will describe some conditions that
FIR LTV systems must satisfy in order to achieve ZF solutions. It turns out that some redundant
elements must always be introduced in the ZF designs. On the other hand, ZF systems can be
regarded as suboptimal solutions in the mean square error (MSE) sense. In this case, we can give up
imposing the ZF constraint upon the transceiver, thus allowing the transmission without adding any
redundant element, as long as pure minimum MSE-based solutions are employed. By pure minimum
MSE (MMSE) we mean that there is no ZF constraint imposed upon the transceiver design.
Hence, the way to reduce even further the amount of redundancy inserted in the physical
layer is either to design LTV transceivers and/or to allow the transmitter and receiver multiple-input
multiple-output (MIMO) filters to have memory. In such cases, it is possible to derive ZF solutions
using only one redundant element. Although time-varying transceivers with memory bring about a
reduction in redundancy to allow for ZF solutions, their fast implementations are not known. In
addition, the numerical accuracy of transceivers with reduced redundancy is not fully exploited in
the open literature and for certain is a crucial issue to be addressed before any attempt to include
these solutions in practical implementations or standards.
This chapter adopts a rather different approach as compared to Chapters 2, 3, and 4. The
previous chapters focus on the derivation/description of either structures or performances of OFDM
and beyond-OFDM systems, whereas the main focus of this chapter is on the limits of some
parameters related to FIR LTV transceivers. We start by describing carefully how to use our previous
knowledge about FIR LTI transceivers developed in Chapter 2 in order to model FIR LTV systems
(Section 5.2). Then, we study the fundamental limits concerning some parameters inherent to FIR
LTV transceivers that satisfy the ZF constraint, namely memory of receiver MIMO matrix and
number of transmitted redundant elements (Section 5.3). After showing that when no redundancy
is employed in the transmission ZF solutions cannot be achieved, then we present pure MMSE-
based solutions (Section 5.4). Some examples are also given in this chapter (Section 5.5), including
the interpretation of the ZF conditions within the framework of CDMA systems with long codes.
 5.2. TIME-VARYING REDUCED-REDUNDANCY SYSTEMS WITH MEMORY 151
5.2 TIME-VARYING REDUCED-REDUNDANCY SYSTEMS
WITH MEMORY
The time-varying property of FIR LTV systems hinders the direct application of the Z -domain tools
described in Chapter 2. Hence, in order to derive an appropriate model for the LTV transceivers, we
will revisit the FIR LTI TMUX model introduced in Chapter 2 and derive its matrix representation
in Subsection 5.2.1. Such matrix representation will be the key tool to describe the time-varying
transceivers in Subsection 5.2.2.

5.2.1 FIR MIMO MATRICES OF LTI TRANSCEIVERS

Let us start with Figure 2.9 of Chapter 2, which depicts the Z -domain description of block-based
LTI transceivers. In this figure, F(z) represents an FIR MIMO transmitter matrix whose entries are
polynomials in the complex variable z−1 , as given in Equation (2.42). If we assume that, among all
entries of F(z), the highest polynomial order is T − 1, with T being a given natural number,2 then
we can decompose F(z) as

F(z) = Fτ z−τ , (5.1)
τ ∈T

where T  {0, · · · , T − 1} is a set of integer indexes and Fτ is an N × M matrix with complex-

valued entries, rather than polynomial entries. The natural numbers M and N are, respectively, the
number of symbols within a data block and the actual number of transmitted elements per block,
which includes both data symbols and redundant elements. The entries of Fτ , for all τ ∈ T , can be
determined directly from F(z) using Equation (5.1). However, these coefficients are not important
in our discussions here.
A transmitter matrix with memory (T > 1) enables one to design high-selective transmitter
filters, which may be very useful to avoid spectrum leakage (out-of-band transmissions), to reduce in-
tercarrier interference (ICI) that occurs specially when carrier-frequency offset (CFO) is significant,
and to compensate for the channel distortions when channel-state information (CSI) is available at
the transmitter side. Besides, having memory at the transmitter increases the transmitter diversity,
which could enhance the performance of the related transceiver.
Similarly to what we did with F(z), by assuming that the order of the FIR MIMO receiver
matrix G(z) in Figure 2.9 is Q − 1, where Q is a natural number, we can decompose G(z) as

G(z) = Gq z−q , (5.2)
q∈Q

where Gq is an M × N complex-valued matrix, for each integer index q within the set Q 
{0, · · · , Q − 1}. Once again, the entries of matrix Gq are not important in our discussions here, but
2 From now on, we will define the order of a matrix as the highest polynomial order among all of its entries. For instance, the order
of F(z) is T − 1.
152 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
these entries are defined by G(z) which in turn is defined in Equation (2.43). It is worth mention-
ing that, if the transmitted redundant elements are not able to completely eliminate the interblock
interference (IBI), then matrix G(z) must have memory, i.e., it is mandatory to have Q > 1 in some
applications, as will be clearer soon.
Once again we will perform the same type of decomposition, but now for the FIR MIMO
channel matrix H(z) in Figure 2.9. Thus, we are interested in representing H(z) as

H(z) = Hb z−b , (5.3)
b∈B

where Hb is an N × N complex-valued Toeplitz matrix, for each integer index b within the set
B  {0, 1, · · · , B}, where B stands for the order of the channel matrix H(z). The entries of Hb
can be determined by the entries of H(z), which in turn is a pseudo-circulant matrix defined by
Equations (2.29) and (2.30) of Chapter 2. Following these steps, it is possible to show that the
entries of Hb are related to the channel impulse response h(l) of order L by means of the following
equality:
 
Hb ml
 h(bN + m − l), (5.4)

for any indexes m and l within the set N  {0, 1, · · · , N − 1}. In addition, since we assumed that
the order of the causal channel is L, then h(bN + m − l) = 0 whenever bN + m − l is larger than
L or smaller than 0.

Example 5.1 (Determining Hb from H(z)) Suppose the order L of the FIR channel impulse
response h(l) is such that N < L < 2N, with N being the length of the transmitted block. Thus,
we are assuming that the channel memory L is longer than N. By using Equations (2.29) and (2.30)
of Chapter 2, we will decompose H(z) according to Equation (5.3), as follows:
⎡ ⎤
h(0) 0 0 ··· ··· ··· 0
⎢ h(1) h(0) 0 ··· ···⎥ ··· 0
⎢ ⎥
⎢ h(2) h(1) h(0) 0 ··· ⎥ ··· 0
⎢ ⎥
⎢ .. .. .. .. .. ⎥ .. ..
⎢ . .⎥
H(z) = ⎢ . . . ⎥ . .
⎢ .. .. .. .. .. ⎥ ..
⎢h(N −3) . . . . .⎥ .
⎢ ⎥
⎢ .. ⎥
⎣h(N −2) h(N −3) · · · · · · h(0)
. 0 ⎦
h(N −1) h(N −2) h(N −3) · · · · · · · · · h(0)
  
H0 ∈CN×N
 5.2. TIME-VARYING REDUCED-REDUNDANCY SYSTEMS WITH MEMORY 153
⎡ ⎤
h(N) h(N −1) · · · h(4) h(3) h(2) h(1)
⎢ . . ⎥
⎢h(N +1) h(N ) .. . . h(4) h(3) h(2) ⎥
⎢ ⎥
⎢ .. .. .. .. .. ⎥
⎢ . h(N + 1) . . . . h(3) ⎥
⎢ ⎥
⎢ .. .. ⎥
+ z−1 ⎢ . .
..
.
..
.
..
.
..
. h(4) ⎥
⎢ ⎥
⎢ .. .. .. .. .. .. .. ⎥
⎢ . . . . . . . ⎥
⎢ ⎥
⎢ .. .. . . . . ⎥
⎣ . . . . . . . . . . h(N −1)⎦
··· ··· ··· ··· · · · h(N + 1) h(N )
  
H1 ∈CN×N
⎡ ⎤
0 ··· 0 h(L) · · · h(N + 1)
⎢ .. .. ⎥
⎢0 0 ··· 0 . . ⎥
⎢ ⎥
⎢ .. .. .. .. .. ⎥
⎢. . . . . h(L) ⎥
⎢ ⎥
+ z−2 ⎢0 0 0 ··· 0 0 ⎥
⎢ ⎥
⎢0 0 0 ··· 0 0 ⎥
⎢ ⎥
⎢. .. .. .. .. .. ⎥
⎣ .. . . . . . ⎦
0 0 0 0 ··· 0
  
H2 ∈CN×N

=H0 + z−1 H1 + z−2 H2


= Hb z−b , (5.5)
b∈B

with B = 2. In addition, note that the Toeplitz matrices Hb are such that H0 is a lower triangular
matrix, whereas HB = H2 is an upper triangular matrix. 

Equation (5.4) together with Example 5.1 suggest that B, the order of H(z), can be related
to the order of the channel impulse response L and the length of the transmitted block N . Now, let
us analyze such relation.
When L < N , Equation (5.3) coincides with Equation (2.38) of Chapter 2, with H0 = HISI
and H1 = HIBI , thus we have B = 1. When L = N, the decomposition in Equation (5.3) still
requires only two matrices, i.e., we still have B = 1, but in this case H1 = HIBI includes a non-
zero diagonal entry h(L) = h(N ). Physically, this means that the entire transmitted block suffers
interference from the previous transmitted block. As L becomes larger than N, other previously
transmitted blocks start interfering with the current transmitted block and, therefore, additional
matrices Hb appear in the decomposition of H(z). Example 5.1 illustrates the case when N < L <
2N, which leads to B = 2. In addition, when 2N < L ≤ 3N we have B = 3. Hence, we can define
154 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
B as
/ 0
L
B . (5.6)
N
The definition of B in expression (5.6) implies that BN ≥ L. This observation will be particularly
useful in Section 5.3.
As previously observed, when the length of the transmitted data block N is smaller than the
channel order L, the resulting interblock interference (IBI) is caused by more than one block that have
been previously transmitted. This case (B > 1) is representative of delay-constrained applications
whose related channel-impulse responses are very long, thus hindering the transmission of data
blocks with length N ≥ L (B = 1). Another observation is that, since the channel order is L and
we are considering causal impulse responses, then h(L) must be necessarily different from zero. In
addition, we will assume that the channel model has a non-zero first term in its impulse response,
i.e., we assume h(0) = 0, which means that the channel transfer function has no zeros at the origin.
With those previous definitions, we can now derive a matrix description for the LTI transceiver
of Figure 2.9. To do that, let us first assume that the Z -transform of a generic time-domain vector
signal x(n) is denoted as
x(z)  Z {x(n)} , (5.7)
where x(n) has size Nx ∈ N and can be expressed as
x(n)  [ x(nNx − Nx + 1) x(nNx − Nx + 2) · · · x(nNx ) ]T . (5.8)
Thus, in the time domain, considering a noiseless channel, i.e., v(n) = 0N ×1 for all time
index n ∈ Z, we have the following expressions for the transmitted, received, and estimated vectors,
respectively:
u(n)  Z −1 {F(z)s(z)}
 ! "
= Fτ Z −1 z−τ s(z)
∈T
τ
= Fτ s(n − τ ), (5.9)
τ ∈T
−1
y(n)  Z {H(z)u(z)}
  
= Hb Z −1 z−b u(z)

b∈ B
= Hb u(n − b), (5.10)
b∈B ! "
−1
ŝ(n)  Z G(z)y(z)
 ! "
= Gq Z −1 z−q y(z)
q∈Q

= Gq y(n − q). (5.11)
q∈Q
 5.2. TIME-VARYING REDUCED-REDUNDANCY SYSTEMS WITH MEMORY 155
We can write the input-output relationship of FIR LTI transceivers in a much more compact
form by conveniently stacking the variables which appear in Equations (5.9), (5.10), and (5.11).
Indeed, note that Equation (5.11) can be rewritten as follows:
⎡ ⎤
y (n − (Q − 1))
 ⎢ .. ⎥
ŝ(n) = GQ−1 · · · G0 ⎣ . ⎦
  
G ∈CM×QN
y(n)
  
y(n)∈CQN ×1
= G y(n). (5.12)
Equation (5.12) depends on vectors y (n − Q + 1) , · · · , y(n), which can be determined using
Equation (5.10) as follows:
⎡ ⎤⎡ ⎤
HB · · · H0 0 ··· 0 u (n − (B + Q − 1))
⎢ 0 HB · · · H0 · · · 0 ⎥ ⎢ .. ⎥
⎢ ⎥⎢ . ⎥
y(n) = ⎢ . . . . . ⎥ ⎢ ⎥
⎣ .. .
. . . . . . ⎦⎣
. u(n − 1) ⎦
0 0 · · · HB · · · H0 u(n)
    
H ∈CQN ×(B+Q)N u(n)∈C(B+Q)N ×1
= H u(n). (5.13)
Equation (5.13), on its turn, depends on u (n − B − Q + 1) , · · · , u(n), which can be com-
puted using Equation (5.9) as follows:
⎡ ⎤⎡ ⎤
FT −1 ··· F0 0 ··· 0 s (n − (B + Q + T − 2))
⎢ 0 FT −1 · · · F0 ··· 0 ⎥ ⎢ .. ⎥
⎢ ⎥⎢ . ⎥
u(n) = ⎢ . . . . . ⎥ ⎢ ⎥
⎣ .. .. .. . . .. ⎦ ⎣ s(n − 1) ⎦
0 0 · · · FT −1 · · · F0 s(n)
    
F ∈C(B+Q)N ×[B+Q+(T −1)]M s(n)∈C[B+Q+(T −1)]M×1
= F s(n). (5.14)
Using Equations (5.12), (5.13), and (5.14), we have the following overall input-output de-
scription of the block-based LTI transceiver of Figure 2.9:
ŝ(n) = G HF s(n), (5.15)
which is the time-domain version of equation
ŝ(z) = G(z)H(z)F(z)s(z) (5.16)
in the Z -domain. This transceiver is time-invariant since the global transmitter and receiver matrices,
F and G , respectively, are constant matrices that do not depend on the time index n. The above
matrix description is rather useful for the time-varying case, since we cannot apply the Z -domain
description of the TMUX of Chapter 2 for this case.
156 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
5.2.2 FIR MIMO MATRICES OF LTV TRANSCEIVERS
As previously mentioned, time-variance brings about extra degrees of freedom that play the role of
introducing additional diversity to the system. In order to generalize the previous LTI model to the
LTV case, we could consider that each transmitter matrix Fτ , with τ ∈ T , is actually a time-varying
matrix Fτ (n), which depends on the time index n. This implies that matrices F0 (n), · · · , FT −1 (n)
are employed to generate the channel input u(n), whereas matrices F0 (n − 1), · · · , FT −1 (n − 1)
are employed to generate the channel input u(n − 1), and so forth. Similarly, we could consider that
the receiver matrix Gq , with q ∈ Q, is in fact a matrix Gq (n), which varies with time as well.
In summary, by using this simple generalization we can rewrite Equations (5.9), (5.10),
and (5.11) as follows:

u(n) = Fτ (n)s(n − τ ), (5.17)
∈T
τ
y(n) = Hb u(n − b), (5.18)

b∈ B
ŝ(n) = Gq (n)y(n − q). (5.19)
q∈Q

With these new equations it is straightforward to define generalized time-varying transmitter

and receiver matrices, F (n) and G (n), as follows:
 
G (n)  GQ−1 (n) · · · G0 (n) , (5.20)
⎡ ⎤
FT −1 (n − Q − B + 1) · · · F0 (n − Q − B + 1) 0 ··· 0
⎢ . . . . .. ⎥ . (5.21)
F (n)  ⎣ .. .. .. .. . ⎦
0 0 ··· FT −1 (n) · · · F0 (n)

Hence, we have the following estimate at the receiver end:

ŝ(n) = G (n)HF (n)s(n), (5.22)

and due to the time dependency of the transceiver matrices,3 there is no Z -domain counterpart for
such equation.
The reader should observe that, due to the definition of s(n) in Equation (5.14), we have
 
s(n) = 0M×(B+Q+T −2)M IM s(n), (5.23)

implying that the ZF solution is achieved, i.e., ŝ(n) = s(n), whenever

 
G (n)HF (n) = 0M×(B+Q+T −2)M IM , (5.24)
3 For simplicity, we are considering that the channel does not vary with time. However, since we are interested in bounds rather
than transceiver designs, we could decompose time-varying channels into two parts, one being time-varying and the other time
invariant, and then incorporate the time-varying part of the channel in the transmitter matrix, as shown in [70]. Therefore, some
conditions/bounds derived in this chapter are also applicable to time-varying channels.
 5.3. CONDITIONS FOR ACHIEVING ZF SOLUTIONS 157
or, in other words, whenever we have
 
0(B+Q+T −2)M×M
F (n)
H
× H H
 × G (n) =
H
. (5.25)
      IM
(Q+B+T −1)M×(Q+B)N (Q+B)N ×QN QN ×M

The specific constraints that the transceiver matrices F (n) and G (n) must satisfy in order to
achieve the ZF solution expressed in (5.25) will be described in the following section.

5.3 CONDITIONS FOR ACHIEVING ZF SOLUTIONS

The aim of this section is to examine the conditions that the global transceiver matrices F (n) and
G (n) must satisfy in order for Equation (5.25) to hold.There are a variety of parameters that can affect
such ZF constraint. Therefore, we shall fix some of them by first assuming that the order of the FIR
MIMO transmitter matrix is already known, being denoted by T − 1. In addition, we shall assume
that the order L of the scalar channel h(l) is also known. The number of symbols to be transmitted
in a block was previously fixed at M. With those assumptions, our ultimate target is to find some
mathematical bounds concerning the order of the FIR MIMO receiver matrix (Subsection 5.3.2),
the number of redundant elements K that might possibly be inserted in the transmission, and, as a
consequence, the length of the transmitted data block N = M + K (Subsections 5.3.3 and 5.3.4),
and the degree of time-variance of the transceiver (Subsection 5.3.5).
Before proceeding we will first rewrite the ZF constraint expressed in (5.25) in a different
manner, as described in Subsection 5.3.1.

5.3.1 THE ZF CONSTRAINT

Let us start by interpreting Equation (5.25) as follows: for any (B + Q + T − 1)M × 1 canon-
ical vector em ,4 with m within the set {(B + Q + T − 1)M − M, (B + Q + T − 1)M − M +
1, · · · , (B + Q + T − 1)M − 1}, one has
     
em ∈ R F H (n)H H G H (n) ⊂ R F H (n)H H ⊂ R F H (n) ⊂ C(B+Q+T −1)M , (5.26)

in which R {·} denotes the column or range space of the argument.5 This means that, for all such
vectors em , Equation (5.25) holds if there always exists a QN × 1 vector β m such that

em = F H (n)H H β m , (5.27)

with m ∈ {(B + Q + T − 1)M − M, (B + Q + T − 1)M − M + 1, · · · , (B + Q + T − 1)M −

1}.
By observing the previous equation, one can see that the solution to such linear system depends
on how matrix H H transforms the vector β m into a new vector within the domain of F H (n). An
4 Here, e is a (B + Q + T − 1)M × 1 vector whose mth entry is equal to 1, whereas the remaining elements are equal to 0.
m
5The range space of a given matrix is the vector subspace spanned by linear combinations of the columns of the related matrix. It
is the image of the linear mapping associated with such a matrix.
158 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
important fact concerning the range space of matrix H H is that, for all (Q + B)N × 1 vector α
within the range space of H H , one has
 
0(BN −L)×1
α= , (5.28)
α
 
where α is within R H H ⊂ CQN +L , in which H H is comprised of the last QN + L rows of
H H . That is, we can construct matrix H H by simply discarding the first BN − L ≥ 0 rows of
H H and keeping the remaining rows unchanged. Here, we consider that, if BN − L = 0, then
Equation (5.28) should be read as α = α .
It remains to answer why Equation (5.28) holds. If BN = L, then there is nothing to be proved.
Thus, assume that BN > L. By observing Equation (5.13), one can see that the only possibility of
having non-zero entries in the first BN − L rows of matrix H H occurs when there are some non-
zero entries in the first BN − L rows of matrix HH B . Based on Equation (5.4), the first BN − L
rows of matrix HH contains the elements h∗ (BN + m − l), where m ∈ N denotes the column
B
index and l ∈ {0, · · · , BN − L − 1} denotes the row index. However, BN + m − l is always larger
than L for such indexes m and l (indeed, the smallest argument is achieved when m = 0 and l =
BN − L − 1, thus yielding BN + m − l = L + 1 > L), which implies that h∗ (BN + m − l) = 0.
As a consequence, all linear combinations of the columns of H H have their first BN − L entries
equal to zero, whereas the remaining entries are corresponding linear combinations of the columns
of H H . This means that Equation (5.28) is always true. Example 5.2 illustrates the above reasoning.

Example 5.2 (Range Space of Matrix H H ) Let us start with the same pseudo-circulant chan-
nel matrix H(z) and its related matrices H0 , H1 , and H2 of Example 5.1, where N < L < 2N ,
thus implying that B = 2. In this example, we will study the range space of matrix H H . From
expression (5.13), we can write
⎡ ⎤H
H2 H1 H0 0 ··· 0
⎢0 H2 H1 H0 ··· 0 ⎥
⎢ ⎥
HH ⎢ . .. .. .. .. ⎥
⎣ .. . . . . ⎦
0 0 · · · H2 H1 H0
⎡ ⎤
HH2 0 ··· 0
⎢HH HH ··· 0 ⎥
⎢ 1 2 ⎥
⎢ H . .. ⎥
⎢H H1 H . . . ⎥
⎢ 0 ⎥
=⎢ .. ⎥, (5.29)
⎢ 0 H0 H . H2 ⎥
H
⎢ ⎥
⎢ . .. .. ⎥
⎣ .. . . H1 ⎦
H

0 0 · · · HH 0
 5.3. CONDITIONS FOR ACHIEVING ZF SOLUTIONS 159
with HH
2 ∈ CN×N being a lower triangular Toeplitz matrix given by
⎡ ⎤
0 ··· 0 0 ··· 0
⎢ .. .. .. .. .. ⎥
⎢ . . . . . 0⎥
⎢ ⎥
⎢ .. .⎥
⎢ ··· . .. ⎥
⎢ ∗0 0 0 ⎥
HH2 ⎢ ⎢ h (L) 0 0 · ·· 0 0⎥ ⎥
. (5.30)
⎢ h∗ (L − 1) ∗
h (L) 0 ··· 0 0⎥ ⎥
⎢
⎢ . . . .. .. .. ⎥
⎣ .. .. .. . . .⎦
h∗ (N + 1) h∗ (N + 2) · · · h∗ (L) · · · 0

Therefore, every entry corresponding to the first BN − L = 2N − L rows of matrix HH 2 is equal

to 0. Consequently, the first BN − L rows of H are also comprised of 0 elements only. Hence,
H

any vector formed by linear combination of the columns of H H will have zeros in its first BN − L
entries, as shown in Equation (5.28). 

Taking Equation (5.28) into account and by defining α m  H H β m , we have that the linear
system of Equation (5.27) can be rewritten as
em = F H (n)α
αm
 
0(BN −L)×1
= F (n)
H
αm
  
α
αm
= F H (n)α
αm , (5.31)

in which F H (n) is a (Q + B + T − 1)M × (QN + L) matrix containing the last QN + L columns

of matrix F H (n) (i.e., its first BN − L columns are discarded).
 we must have at least one solution α m within the column space of matrix H , i.e.,
Therefore, H

α m ∈ R H H ⊂ C(QN +L)×1 . From linear algebra, we know that if α m is a vector within the column
space of matrix H H , then α m must be orthogonal
 to all vectors belonging to the kernel (nullspace)

of matrix H , since the subspace ker H is the orthogonal complement of the subspace R H H ,
in which ker {·} denotes the kernel of the argument.6 Hence, we can restate the ZF constraint as
satisfying the linear system in expression (5.31), while also respecting the constraint
z H α m = 0, (5.32)
 
for all vector z within the subspace ker H , and for each m within the set {(B + Q + T − 1)M −
M, (B + Q + T − 1)M − M + 1, · · · , (B + Q + T − 1)M − 1}.
6The kernel (or null space) of a given matrix is the vector subspace spanned by vectors that have zero image. An important result
in linear algebra is that the dimension of the null space plus the dimension of the range space of a matrix is equal to its number
of columns.
160 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
Now, we can use this interpretation of the ZF constraint in order to derive a lower bound for
Q, the order of the MIMO receiver matrix described in Equation (5.19).

5.3.2 LOWER BOUND ON THE RECEIVER LENGTH

Assume that the (Q + B + T − 1)M × (QN + L) matrix F H (n) has been previously designed in
such a way that it has rank (Q + B + T − 1)M − (BN − L). This might sound a bit artificial, but
we shall illustrate it later on, in a concrete example in Subsection 5.3.4. By now it is sufficient to
know the motivation for such a choice. The main idea is that we must have a rank large enough to
allow us to solve M linear systems such as the one described in Equation (5.31). It turns out that
the number of degrees of freedom induced by a rank of (Q + B + T − 1)M − (BN − L) is more
than enough to solve the related linear systems. Such an assumption requires that

QN + L ≥ (Q + B + T − 1)M − (BN − L). (5.33)

From linear algebra we know that the dimension of the null space of F H (n) is

QN + L − [(Q + B + T − 1)M − (BN − L)] = (Q + B)K − (T − 1)M, (5.34)

in which the integer number

K N −M (5.35)

denotes the amount of redundancy introduced in the transmission.

The rank constraint upon matrix F H (n) implies that there
 is a (QN + L) × [(Q + B)K −
(T − 1)M] matrix (n), whose rows span the subspace ker F H (n) . Thus, for all given vector

em , there always exists at least one particular solution γ̂γ m such that em = F H (n)γ̂γ m . The general
solution to the linear system in Equation (5.31) can therefore be written as

α m =  (n)α̂
α + γ̂γ m , (5.36)

α ∈ C[(Q+B)K−(T −1)M]×1 is a vector which parameterizes the general solution and introduces
where α̂
the additional degrees of freedom that allow α m to satisfy the requirement of being orthogonal to
all vectors within the null space of H (see Equation (5.31)). Note that

F H (n)α
α m = F H (n) α + F H (n)γ̂γ m
(n) α̂
     
=0 =em
= 0 × α̂
α + em
= em . (5.37)
 
Now, consider the restriction of having α m ∈ R H H , which is equivalent to find α m such
 
that z H α m = 0 for all vector z within the subspace ker H . But, instead of considering any vector
 5.3. CONDITIONS FOR ACHIEVING ZF SOLUTIONS 161
   
z ∈ ker H , we can verify if α m is orthogonal to all vectors from a basis of ker H . Indeed,
if the dimension of the kernel ofmatrix H is D ∈ N, then there are D linear independent vectors
z0 , z1 , · · · , zD−1 which span ker H and which respect the identity H zd = 0QN ×1 , for all d within
the set D  {0, 1, · · · , D − 1}. Since any vector z ∈ H can be written as

z = ζ0 z0 + · · · + ζD−1 zD−1 , (5.38)

then

z H α m = ζ0 z0H α m + · · · + ζD−1 zD−1

H
α m. (5.39)

Hence,if zdH α m = 0 for all d within the set D, then z H α m = 0. Thus, our task is to find a basis
for ker H . It turns out that such a basis is associated with the zeros of the scalar channel transfer
function, as formulated below.
Given the FIR channel model, whose transfer function is

H (z)  h(0) + h(1)z−1 + · · · + h(L)z−L , (5.40)

with h(0)  = 0  = h(L), consider that the zeros of H (z) are all distinct, i.e., there are L distinct scalars
z0 , · · · , zL−1 ∈ C such that H (z0 ) = · · · = H (zL−1 ) = 0. Since h(0) = 0  = h(L) and due to the
Toeplitz structure of matrix H described in Equation (5.13), it turns out that this channel matrix is
full-row rank, i.e., rank of H is QN. As H is generated from H by discarding the first BN − L zero
columns, then the rank of H is QN as well. This means that the dimension of the null space of H is

D = QN + L − QN
= L. (5.41)

Moreover, the set composed by vectors of the form

T
QN +L−1
zlT  1 zl zl2 · · · zl , (5.42)

for each l ∈ {0, · · · , L − 1}, is a linearly independent set of vectors within the kernel of H , since

H zl = 0QN ×1 , (5.43)

for all l ∈ {0, · · · , L − 1}. Indeed, the ith row of H , with i ∈ {0, 1, · · · , QN − 1}, is given by

T  
hi  0 · · · 0 h(L) · · · h(0) 0 · · · 0 ∈ C1×(QN +L) , (5.44)
162 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
where there are i zero entries at the beginning of this vector and QN − i − 1 zero entries at its end.
Therefore, we have
T
hi zl = h(L)zli + h(L − 1)zli+1 + · · · + h(0)zlL+i

−(L−1)
= zlL+i h(L)zl−L + h(L − 1)zl + · · · + h(0)
= zlL+i H (zl )
  
=0
= 0, (5.45)

implying that H zl = 0QN ×1 , for any l ∈ {0, · · · , L − 1}.

Thus, the constraint described in Equation (5.32) can be rewritten as

ZH α m = 0L×1 , (5.46)

where the (QN + L) × L full column-rank matrix Z is defined as

 
Z  z0 z1 · · · zL−1 . (5.47)

Hence, we can substitute the general solution described in Equation (5.36) into the constraint
in Equation (5.46), thus giving

α + ZH γ̂γ m = 0L×1 ,
ZH  (n)α̂ (5.48)

yielding

α = −ZH γ̂γ m .
ZH  (n) α̂ (5.49)

A condition7 for achieving the solution is having

H   H −1
α = − ZH  (n)
α̂ ZH  (n) ZH  (n) ZH γ̂γ m , (5.50)

  H
which is achievable only when matrix ZH  (n) ZH  (n) is non-singular. This fact has two
consequences:

1. we must have L ≤ (Q + B)K − (T − 1)M; and

2. the rank of the L × [(Q + B)K − (T − 1)M] matrix ZH  (n) is in fact L.

7 Such a constraint may not be a necessary condition, which means that we may derive some results more restrictive than the
necessary ones.
 5.3. CONDITIONS FOR ACHIEVING ZF SOLUTIONS 163
The first consequence of the aforementioned rank requirement is that the number of columns
  H
of matrix ZH  (n) ZH  (n) cannot be smaller than its number of rows. Therefore, we must
have
(Q + B)K ≥ (T − 1)M + L, (5.51)
yielding
/ 0
(T − 1)M + L
Q≥ − B, (5.52)
K
which is a lower bound on Q that must be used in the design of time-varying receiver matrix
(remember that, by hypothesis, Q ≥ 1).

5.3.3 LOWER BOUND ON THE AMOUNT OF REDUNDANCY

In order to achieve the sufficient solution of Equation (5.50), one must always introduce redundancy,
i.e., one must have K = N − M ≥ 1. In fact, we have already used this result in order to derive
expression (5.52) from expression (5.51). Indeed, by examining inequality (5.51), we know that
(T − 1)M + L ≥ L ≥ 1, by hypothesis. Besides, as Q + B is always larger than or equal to 2, then
K = N − M cannot be smaller than or equal to zero 0. Hence, we must necessarily have
K ≥ 1. (5.53)
Actually, having N < M is not meaningful anyway since in this case the transmitter discards
some data which are supposed to be transmitted, constituting a loss of information. However, when a
design criteria other than the ZF solution is employed, one may consider using N = M (i.e., K = 0)
without conceptual problems (see Subsection 5.4 for further details on this topic).
Example 5.3 (Equalizer-Length Bound) This example is just a numerical illustration of the
bound described in expression (5.52). Let us first assume that the transmitter is memoryless (T =
1), that M ≥ L/2, and that K = L/2 (even order). Using these values we can deduce that B =
 M+L/2
L
 = 1 and, by using inequality (5.52), we get Q ≥ 1, i.e., the receiver can be memoryless as
well. See Chapter 4 for the designs of such memoryless transceivers.
In the case where we have an FIR transmitter matrix with T = 3, and assuming K = 2 and
L = 2N, then B =  2N N  = 2 and Q ≥ N + M − 2, i.e., the time-varying receiver might have long
memory. 

The reader should observe that the equalizer-length bound in expression (5.52) depends on
some known variables, namely L, M, and T , and on the amount of redundancy K which is to be
defined. Given L,
! M, and T" , one might argue if it is always possible to satisfy inequality (5.52) and the
constraint rank ZH  (n) = L (see discussions after expression (5.50)), regardless of the amount of
transmitted redundant elements K. The next subsection shows a particular example of time-varying
transceiver with memory that is able to achieve the lower bound on K in expression (5.53) and
discusses further design constraints to achieve such lower bound.
164 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
5.3.4 ACHIEVING THE LOWER BOUND OF REDUNDANCY
Let us address this topic by first considering for the time being that the transmitter matrices are
memoryless (T = 1), since having more columns in matrix F (n) could only add more diversity
(more redundancy) to the system, so that the memoryless transmitter case is somehow the worst
case scenario for finding a lower bound on K which is actually required to achieve the ZF solution.
Indeed, it is reasonable to expect that the minimum achievable bound8 for K would also be sufficient
when we have memory at the transmitter end.
The remaining of this section will focus on proving that the lower bound on the number of
transmitted redundant elements in expression (5.53) can be reached and we encourage the reader
to keep this target in mind while examining the forthcoming lengthy mathematical computations.
Another important observation is that we are not interested in the performance of the described
transceiver, rather we focus only on the parameter bounds.
In the memoryless transmitter case, we consider that the time-varying transmitter matrix in
its Hermitian transpose version is given by
⎡ ⎤
F0H (n − Q − B + 1) 0 ··· 0
⎢ .. ⎥
⎢ 0 . ··· 0 ⎥
F H (n)  ⎢
⎢ .. .. ..
⎥ ∈ C(Q+B)M×(Q+B)N ,
⎥ (5.54)
⎣ . . F0H (n − 1) . ⎦
0 0 ··· F0H (n)

where we assume that F H (n) has full row rank, which means that the rank of F H (n) is (Q + B)M.
In addition, assume that, for each time index n, we have
 
IM  F(n) ∈ CN ×M ,
F0 (n) =  (5.55)
0K×(BN −L)  (n)

where F(n) is an M × M matrix with full rank, whereas  (n) is a given K × M matrix, with
M  M − (BN − L) ≥ 0.9 Such structure for the transmitter matrices means that the redundant
elements are included at the end of each block to be transmitted and that these redundant elements
are linear combinations of the last M non-redundant elements.
Now, matrix F H (n) can be generated by discarding the first BN − L columns of F H (n).
Based on expressions (5.54) and (5.55), this means that we are removing BN − L linearly indepen-
dent columns of F H (n), thus reducing the rank from (Q + B)M to (Q + B)M − (BN − L) =
(Q + B − 1)M + M − (BN − L) = (Q + B − 1)M + M. This coincides with the proposed di-
mension for F H (n) described in the first paragraph of Subsection 5.3.2 (for T = 1). Therefore, we
8The term “achievable” was employed here since we do not know a priori if, for any channel model and for any choices of M and
 
T , we could use K = 1 and still satisfy the constraint: rank ZH (n) = L.
9This assumption (M ≥ 0) is not a very restrictive constraint, since if B = 1, then BN − L ≤ BN − K = N − K = M (assuming
L ≥ K), whereas if B > 1, then N can be chosen so that BN is close to L, implying that it is easy to satisfy BN − L ≤ M.
 5.3. CONDITIONS FOR ACHIEVING ZF SOLUTIONS 165
have the following (Q + B)M × (QN + L) matrix:
⎡ H ⎤
F (n − Q − B + 1) 0 ··· 0
⎢ .. ⎥
⎢ 0 . ··· 0 ⎥
F (n)  ⎢
H ⎢ ⎥
.. .. .. ⎥
⎣ . . F (n − 1)
H . ⎦
0 0 ··· H
F (n)
⎡  ⎤
0(BN −L)×M 0(BN −L)×K
⎢ 0 ⎥
⎢ IM  H (n − Q − B + 1) ⎥
×⎢    0 ⎥.
⎣ 0(BN −L)×K ⎦
0 diag IM
 H (n − i) i=Q+B−2
(5.56)

From linear algebra, we know that the kernel of matrix F H (n) has dimension QN + L −
[(Q + B)M − (BN − L)] = (Q + B)K. Let  (n) be a (QN + L) × (Q + B)K block-diagonal
matrix defined as
⎡ ⎤
 (n − Q − B + 1) 0 ··· 0
⎢ .. .. .. .. ⎥
⎢ . . . . ⎥
 (n)  ⎢ ⎥, (5.57)
⎣ 0 · · ·  (n − 1) 0 ⎦
0 0 ···  (n)

where
⎡ ⎤
0(BN −L)×K
H (n − i) ⎦ ∈ CN ×K ,
 (n − i)  ⎣ − (5.58)
IK

for all i ∈ {0, · · · , Q + B − 2}, whereas

 
H (n − Q − B + 1)
−
 (n − Q − B + 1)  ∈ C(M+K)×K . (5.59)
IK

With this definition, one always has F H (n)

(n) = 0(Q+B)M×(Q+B)K , which allows us to
write the matrix equation

α m =  (n)α̂
α + γ̂γ m , (5.60)

which is a concrete example of Equation (5.36), where γ̂γ m is a particular solution of the linear system
F H (n)α α m = em . Observe that, in order to determine α m , it is only necessary to determine α̂
α , since
γ̂γ m and  (n) are assumed to be already known, since they depend exclusively on matrix F (n). Let
us also simplify the forthcoming notation by omitting the index m of the related variables. Thus,
α =  (n)α̂ α + γ̂γ must be a solution to the linear system e = F H (n)α α.
166 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
As the solution to such linear system must also satisfy ZH α = 0L×1
 , where Hthe L × (QN +
H
L) matrix Z is defined in Equation (5.47), then we have Z  (n) α̂ H α = −Z γ̂γ . The former
equation can be rewritten in a much more convenient way by first defining vectors z̃l , for each
l ∈ {0, · · · , L − 1}, as

 T
z̃lT  1 zl zl2 · · · zlN −1 , (5.61)

and then by rewriting the L × (Q + B)K matrix ZH  (n) as follows:

⎡ H ⎤
z̃0  (n − Q − B + 1) ··· (z0∗ )(Q+B−2)N z̃0H  (n − 1) (z0∗ )(Q+B−1)N z̃0H  (n)
⎢ z̃ H  (n − Q − B + 1) ··· (z1∗ )(Q+B−2)N z̃1H  (n − 1) (z1∗ )(Q+B−1)N z̃1H  (n) ⎥
⎢ 1 ⎥
ZH  (n)=⎢
⎢ .. .. .. ⎥
⎥
⎣ . . ··· . ⎦
H  (n − Q − B + 1)
z̃L−1 ··· ∗
(zL−1 H  (n − 1)
)(Q+B−2)N z̃L−1 ∗
(zL−1 ) (Q+B−1)N H
z̃L−1 (n)
(5.62)

where  (n − Q − B + 1) is an N × K matrix defined as

 
0(BN −L)×K
 (n − Q − B + 1)  . (5.63)
 (n − Q − B + 1)

α , as in Equation (5.50), is having (Q +

Now, a possible condition for finding a solution for α̂
B)K ≥ L, or

/ 0
L
K≥ . (5.64)
Q+B

It is worth pointing out that, in >the?memoryless receiver case (Q = 1) with N ≥ L (i.e.,

B = 1), condition (5.64) becomes K ≥ L2 , which is the same ZF condition previously deduced
in Chapter 4.
By observing the structure of matrix ZH  (n) one can verify its strong dependency on the
quantities zlN , for l ∈ {0, · · · , L − 1}, where zl is a zero of the scalar channel transfer function H (z).
The number of degrees of freedom associated with that matrix is reduced when there exist distinct
indexes l1 , l2 such that zlN1 = zlN2 . In such cases, the chances of matrix ZH  (n) to have full row rank
is somehow decreased. In order to deal with such worst case setup, let us define the congruous zeros of
H (z) with respect to N as the distinct complex numbers zl1 and zl2 which satisfy zlN1 = zlN2  c ∈ C.
 5.3. CONDITIONS FOR ACHIEVING ZF SOLUTIONS 167
Thus, assume that the first μ ∈ {1, · · · , L} zeros of H (z) are congruous. Then, the first μ
rows of ZH (n) can be rewritten as
⎡ H ⎤
z̃0  (n − Q − B + 1) · · · (c∗ )Q+B−2 z̃0H  (n − 1) (c∗ )Q+B−1 z̃0H  (n)
H (n) = ⎢ .. .. .. ⎥
Z ⎣ . . ··· . ⎦
z̃ H  (n − Q − B + 1) · · · (c∗ )Q+B−2 z̃μ−1
H  (n − 1) (c∗ )Q+B−1 z̃ H  (n)
⎡ μ−1H ⎤ μ−1
z̃0
⎢ . ⎥ 
= ⎣ .. ⎦  (n − Q − B + 1) · · ·  (n − 1)  (n)
H   
z̃μ−1 N ×(Q+B)K
  
μ×N
⎡ ⎤
IK
⎢ c ∗ IK ⎥
⎢ ⎥
×⎢ .. ⎥. (5.65)
⎣ . ⎦
(c∗ )Q+B−1 IK
  
(Q+B)K×(Q+B)K

Consider that the global transmitter matrix was designed in such a way that there exist
N
δ ≤ Q + B matrices, let us say  (n − δ + 1), · · · ,  (n − 1),  (n) ∈ CN ×K , such that
⎧ ⎫
⎪
⎨ ⎪
⎬
rank  (n − δ + 1) · · ·  (n − 1)  (n) = δK, (5.66)
⎪
⎩  ⎪
⎭
N ×δK

where, as a consequence of the above matrix dimensions, we have

δK ≤ N. (5.67)

A necessary condition for ZH  (n) be a full-row rank matrix is that Z H  (n) be full row rank,

which eventually requires that

μ ≤ δK ≤ N, (5.68)

since the rank of matrix Z H  (n) is smaller than or equal to the ranks of each matrix that comprises

the matrix product which defines Z  H  (n), and we already know that (Q + B)K ≥ L ≥ μ.

Therefore, using
<μ=
K≥ , (5.69)
δ
with
 @ A
N
δ ≤ min Q + B, (5.70)
K
168 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY

guarantees that Z H  (n) is full row rank, which indicates that it would be possible to satisfy the

original zero-forcing constraint as well.

The previous conditions are dependent on the congruous zeros of the channel. In order to
develop a condition which is independent of the channel-zero locations, we could concentrate efforts
on the worst case scenario, which occurs when μ = L. Intuitively, this can be thought as the worst
setup since the number of degrees of freedom in matrix ZH  (n) is somehow minimized when
μ = L. This means that for such a case one has more chances of having a row of matrix ZH  (n)
described as a linear combination of its other rows. Similarly, to what we have performed before, we
have
⎡ ⎤
z̃0H  (n − Q − B + 1) ··· (c∗ )Q+B−2 z̃0H  (n − 1) (c∗ )Q+B−1 z̃0H  (n)
 ⎢ . . .. ⎥
Z H  (n) = ⎢ .. .. ⎥
⎣ ··· . ⎦
z̃L−1 (n − Q − B + 1) · · · (c∗ )Q+B−2 z̃L−1
H H  (n − 1) ∗
(c ) Q+B−1 H
z̃L−1 (n)
⎡ ⎤
z̃ H
⎢ 0. ⎥  
=⎢ ⎥
⎣ .. ⎦  (n − Q − B + 1)  · · ·  (n − 1)  (n)

H
z̃L−1 N×(Q+B)K
  
L×N
⎡ ⎤
IK
⎢ c∗ I ⎥
⎢ K ⎥
×⎢ .. ⎥. (5.71)
⎣ . ⎦
(c∗ )Q+B−1 IK
  
(Q+B)K×(Q+B)K

Once again, by assuming that

! "
L ≤ rank  (n − δ + 1) · · ·  (n − 1)  (n) = δK ≤ N, (5.72)
and by using N = M + K ≥ L (i.e., B = 1), or in other words,
M ≥ L − K, (5.73)
we can achieve a ZF solution using K = 1 redundant elements, as long as (see Equation (5.64))
min {Q + 1, N} ≥ δ ≥ L, (5.74)
which can be achieved with a large enough equalizer length (Q ≥ L − 1), thus requiring many
time-variant receiver matrices to estimate a block. Indeed, with such assumptions we can design a

matrix Z H  (n) with rank L, which implies that we can design a matrix ZH  (n) with rank L. This

proves that the lower bound K = 1 on the number of transmitted redundant elements is achievable.

5.3.5 ROLE OF THE TIME-VARIANCE PROPERTY

It is important to point out the role of the time-variance property in the FIR reduced-redundancy
systems. As an example, if we were working with LTI systems, then Equation (5.65) would have
 5.4. TRANSCEIVERS WITH NO REDUNDANCY 169
been written as follows:
⎡ ⎤
z̃0H  · · · (c∗ )Q+B−2 z̃0H  (c∗ )Q+B−1 z̃0H 
 H = ⎢ . .. .. ⎥
Z ⎣ .. . ··· . ⎦
H  · · · (c∗ )Q+B−2 z̃ H  (c∗ )Q+B−1 z̃ H 
z̃μ−1
⎡μ−1 μ−1
⎤
⎡ H ⎤ IK
z̃0
⎢ .. ⎥  ⎢⎢ c∗ IK ⎥
⎥
= ⎣ . ⎦  ···   ⎢ . ⎥. (5.75)
H   ⎣ . . ⎦
z̃μ−1 N ×(Q+B)K ∗ Q+B−1
   (c ) IK
μ×N   
(Q+B)K×(Q+B)K

Note that the rank of matrix [

 · · ·   ] is K. Hence, we would have to guarantee that
K ≥ μ, (5.76)
in order to achieve a ZF solution, implying that N = M + K ≥ μ.This is a much stronger constraint
than the previous one in expression (5.69) associated with LTV systems. For example, if all zeros
of a channel transfer function are on the unitary circle, so that μ = L, then we must use K ≥ L
redundant elements in LTI systems, whereas we may still be able to use K = 1 in LTV systems!

5.4 TRANSCEIVERS WITH NO REDUNDANCY

As mentioned before, the ZF solution is suboptimal in the MSE sense. Indeed, an optimal MMSE-
based solution could be achieved by jointly designing the transmitter and receiver filters. In this
case, it would be possible to transmit without using any redundant elements at all. For example, a
transceiver with memory and with no redundancy allows for the minimization of the MSE by jointly
designing the transmitter and receiver FIR MIMO filters in the presence of, for example, near-end
crosstalk and additive noise. This discussion is representative of other discussions related to some
transceiver optimization, such as the ones described in [30, 83].
In the transceiver design we have to determine the matrices representing the transmitter and
receiver. These matrices appear, for instance, in Equations (5.20) and (5.21), or in a more compact
manner in Equation (5.22).
There are many ways to design the global transmitter and receiver matrices, respectively,
denoted as F (n) (remember that this matrix has a block-Toeplitz structure as described in Equa-
tion (5.21)) and G (n). Here, we shall focus on minimizing the MSE of symbols, EMSE ∈ R+ . The
minimum MSE (MMSE) designs are very common in practical systems [83]. Consider that the
transmitted vector s(n) and the channel-noise vector v(n) are, respectively, drawn from the zero-
mean jointly wide-sense stationary (WSS) random sequences s (n) and v (n). Assuming that the
estimated symbols are given by
 
ŝs (n) = G (n) HF (n)ss (n) + v (n)
= G (n)HF (n)ss (n) + G (n)vv (n), (5.77)
170 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY

Transmiter Receiver
Optimization Optimization

Figure 5.1: Optimization strategy.

then the overall MSE of symbols is given by (see also Equation (5.23))

EMSE  E{ŝs (n) − s (n)22 }

B   B2
= E{BG (n)HF (n)ss (n) + G (n)vv (n) − 0 IM s (n)B2 }
     
= tr (G (n)HF (n) − 0 IM )Rss (n)(G (n)HF (n) − 0 IM )H
 
+ tr G (n)Rvv (n)G H (n) , (5.78)

in which Rss (n) = E{ss (n)ss H (n)} and Rvv (n) = E{vv (n)vv H (n)}. In addition, we have also assumed
that s (n) and v (n) are uncorrelated, i.e., Rsv = E{ss (n)vv H (n)} = E{ss (n)}E{vv (n)}H = 0.
In summary, it assumed that the FIR MIMO channel transfer matrices are known and that
all the input signals to the system are uncorrelated with each other. In addition, we usually consider
that the second order statistics of all the input signals are known. The transceiver is represented by
FIR MIMO transfer matrices with given orders.
Our problem is to minimize MSE, usually subject to power constraints. The procedure to
find a solution is to derive analytical expressions for the MSE criterion and power constraint as
for example presented in [31]. For that we find the Karush-Kuhn-Tucker (KKT) conditions for
optimality and implement an iterative numerical algorithm based on those necessary conditions.
The basic strategy is to design a transmitter assuming a known initial receiver and then design
a receiver assuming the transmitter is the one previously designed, as Figure 5.1 illustrates. The
procedure can be repeated until no further reduction in the MSE is achieved.

5.5 EXAMPLES
This section presents three examples that illustrate some of the concepts developed in this chapter.
We will start with Example 5.4 in order to show a particular situation in which an FIR MIMO
transceiver outperforms a CP-OFDM system, considering the bit-error rate (BER) as figure of merit.
Then, Example 5.5 illustrates the BER performance of ZF- and MMSE-based systems satisfying or
not the zero-forcing conditions described in Section 5.3. After that, we explain how the theoretical
zero-forcing bounds derived here relate to the ZF detection of signals in long-code CDMA systems
in Example 5.6.

Example 5.4 (OFDM & FIR MIMO Transceivers) This example assesses the BER versus
signal-to-noise ratio (SNR) performance evaluated through a Monte Carlo simulation, where M = 8
 5.5. EXAMPLES 171
BPSK symbols are transmitted per block through a channel whose transfer function is

H (z) = 1 − 0.9z−1 1 − 0.7ej2π 0.256 z−1 1 − 0.4ej2π 0.141 z−1 , (5.79)

thus implying that L = 3. The number of redundant elements is always K = L = 3, which means
that N = 11. The uncorrelated additive channel noise samples are complex, circularly symmetric,
and Gaussian distributed. Two distinct systems are considered in this example: the CP-OFDM and
the FIR MIMO transceiver proposed in [31], which optimizes the transmitter and receiver matrices
through an iterative MMSE-based approach, such as the one briefly described in Section 5.4.
Figure 5.2 depicts the obtained results, in which it is clear the striking advantages of using FIR
MIMO systems over OFDM-based systems in this particular setup. Of course, we are not taking
into account implementation issues, such as computational complexity, modularity, just to mention
a few. Indeed, a great advantage of OFDM is the low complexity implementation costs required to
implement the fast Fourier and inverse fast Fourier transforms. 

0
10

−1
10

−2
10
BER

−3
10

−4
10

OFDM
FIR MIMO Transceiver
−5
10
0 2 4 6 8 10 12 14 16 18 20
SNR [dB]

Figure 5.2: BER as a function of SNR for Example 5.4.

172 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY

10 0
ZF
MMSE
ZF (K = 5)
MMSE (K = 5)
ZF (Q = 1)
MMSE (Q = 1)
BER

10 −1

10 −2
−5 0 5 10 15 20 25 30 35 40
SNR [dB]

Figure 5.3: BER as a function of SNR for Example 5.5.

Example 5.5 (BER Performance & ZF Conditions) Consider the transmission of data blocks
containing N = 30 entries, including M symbols of a 16-QAM digital constellation as well as K =
30 − M redundant elements. Assume that the channel-transfer function has the following zeros:
2π 2π 3π 3π 4π 4π
1.8ej 5 , 1.8e−j 5 , 1.8ej 5 , 1.8e−j 5 , 1.8ej 5 , 1.8e−j 5 , 1.1ej 8 , 1.1e−j 8 , 0.9ej 8 , 0.9e−j 8 , ej 10 ,
7π 7π 7π 7π 5π

5π
e−j 10 , 0.8, which means that the channel order is L = 13, whereas the number of congruous zeros
with respect to N is μ = 6, corresponding to the first six zeros with magnitude 1.8.
In a time-invariant transmission, one must have K ≥ μ = 6 in order to achieve the ZF
solution (see Subsection 5.3.5). This means < =that M must be smaller than or equal to 24. In addition,
based on Equation (5.6), one has B = 30 13
= 1, while based on Equation (5.52), and considering
< =
that T = 1 and K = 6, one has Q ≥ 13 6 − 1 = 2. Figure 5.3 depicts some results obtained using
time-invariant ZF- and MMSE-based equalizers. The “ZF” and “MMSE” curves employ K = 6
redundant elements and Q = 2 as the order of the related receiver matrices. The “ZF (K = 5)” and
“MMSE (K = 5)” curves employ K = 5 and Q = 2, which means that the number of redundant
elements is violating the ZF constraint described in Section 5.3. The “ZF (Q = 1)” and “MMSE
 5.5. EXAMPLES 173
(Q = 1)” curves employ K = 6 and Q = 1, which means that the orders of the related receiver
matrices are not satisfying the ZF constraint. The obtained results show that, by not following the
ZF conditions, either by choosing K = 5 or Q = 1, a floor in the BER curves appears, starting from
20 dB of SNR. On the other hand, for the designs following the ZF conditions, the BER tends to
zero as the SNR increases.
This example also illustrates that when alternative design criteria are used, such as the min-
imization of the MSE, the ZF conditions described in Section 5.3 are still useful in order to avoid
performance loss due to errors in the reconstruction of the signal. This eventually means that, even
for very high SNRs, it is still possible to have a “BER floor” due to the nonexistence of the ZF so-
lution. On the other hand, this “BER floor” does not appear by observing the conditions presented
in this chapter. 

Example 5.6 (Long-Code CDMA Systems) As previously illustrated in Figure 1.21 of Chap-
ter 1, MIMO models encompass many block-transmission configurations, ranging from single-user
point-to-point communications employing multiple antennas, to multiple-access schemes in mul-
tiuser systems. This example illustrates an application of the theoretical results related to FIR LTV
transceivers with reduced redundancy within the framework of CDMA systems.
Indeed, CDMA systems can be described using the concept of MIMO transceivers. By using
such a description it is possible to derive some theoretical conditions for designing equalizers that
guarantee the perfect reconstruction of the transmitted signal at the receiver end. For example, we
can apply the theoretical analysis in order to obtain multiuser detection in CDMA systems with
long codes, i.e., codes which last for more than one symbol duration, as described in the following
reasoning.
Consider the TMUX structure of Figure 2.6 of Chapter 2. Assume that sm (n) is the symbol
associated with the mth user at the time instant n. It is possible to imagine that the impulse re-
sponses of the synthesis filters fm (k), with m ∈ M, that appear in that figure can play the role of a
spreading sequence of a CDMA system, whereas the impulse responses of analysis filters gm (k), with
m ∈ M, can be thought as the de-spreading codes, all of them associated with the mth user. If, in
addition, we assume that the synthesis and analysis filters are actually time-varying filters, then we
can consider that each subfilter implements a piece of a given CDMA spreading/de-spreading long
code associated with a user. In other words, CDMA with long codes can be interpreted as CDMA
with time-varying short codes.
As a result, it is possible to verify that CDMA with long codes can be represented by a
time-varying structure, i.e., it is an example of FIR LTV system. Each signal block faces a time-
varying code, which is implemented using time-varying transmit filters.Thus, by adapting the results
concerning the amount of redundancy required to allow a ZF solution, it is possible to establish the
conditions for the existence of zero-forcing multiuser detectors.
In this context, N represents the spreading factor, M denotes the number of codes that are
going to be used, whereas K represents the number of unused codes. The main conclusion of this
174 5. FIR LTV TRANSCEIVERS WITH REDUCED REDUNDANCY
analysis is that ZF equalization is always possible in CDMA systems using long codes as long as the
system is not at full capacity, i.e., one must have at least one unused code (K ≥ 1). In addition, it can
be shown that both the complexity of the receiver and its performance depend directly on the number
of unused codes. The conditions derived serve as useful guidelines for the design of communications
systems, allowing the tradeoff between performance and complexity of the receiver.
In conclusion, the existence of ZF equalizers is guaranteed if the amount of redundancy is
greater than or equal to the number of congruous zeros or if there are enough different transmission
filters at the transmitter. For practical channels, approximately congruous zeros may cause numerical
instability in equalizer design. A CDMA system with long codes may be interpreted as a transmul-
tiplexer with memoryless time-varying filters. In the uplink direction, the system can be modeled
as a transmultiplexer with time-varying filters with memory. In any case, for the CDMA system,
the block length is determined by the spreading factor, implying that the redundancy is equal to the
number of unused codes. 

5.6 CONCLUDING REMARKS

This chapter addressed the problem of further reducing transmission redundancy in block
transceivers by relieving their designs from the memoryless and time-invariance constraints. The
use of time-varying transceivers with memory allows ZF solutions whose amount of required re-
dundancy can be reduced to a single element. We discussed several results related to the amount of
memory at the transmitter and receiver as well as the time variance of the transceiver in order to
understand what are available to achieve ZF solutions with as little amount of redundancy in the
transmission as possible. Despite this attractive feature some performance issues have to be consid-
ered particularly when the environment noise is not negligible. In this case, the MMSE solution is a
practical solution where the ZF constraint can be relieved, leading to solutions where no redundancy
is required. On the other hand, it is known that in many practical situations an increase in the amount
of redundancy leads to reduction in the MSE. In summary, the actual designer challenge is to find
the best trade-off between data throughput and system performance while keeping the transceiver
implementation numerically robust and efficient.
 175

Bibliography
[1] A. N. Akansu, P. Duhamel, X. Lin, and M. de Courville, “Orthogonal transmultiplexers in
communication: a review,” IEEE Trans. on Signal Processing, vol. 46, no. 4, pp. 979–995, April
1998. DOI: 10.1109/78.668551 Cited on page(s)

[2] A. N. Akansu and M. J. Medley, Wavelets, Subband and Block Transforms in Communications
and Multimedia, 2nd Edition, Kluwer Academic Publishers, Norwell, MA, 1999. Cited on
page(s)

[3] N. Al-Dhahir and S. N. Diggavi, “Guard sequence optimization for block transmission over
linear frequency-selective channels,” IEEE Trans. Communications, vol. 50, no. 6, pp. 938–946,
June 2002. DOI: 10.1109/TCOMM.2002.1010613 Cited on page(s)

[4] J. B. Anderson, Digital Transmission Engineering, IEEE Press, New York, NY, 1999. Cited on
page(s) 2

[5] A. Antoniou and W.-S. Lu, Practical Optimization: Algorithms and Engineering Applications.
New York: Springer, 2008. Cited on page(s)

[6] L. Anttila, M. Valkama, and M. Renfors, “Circularity-based I/Q imbalance compensation in

wideband direct- conversion receivers,” IEEE Trans. on Vehicular Technology, vol. 57, no. 4, pp.
2099–2113, July 2008 DOI: 10.1109/TVT.2007.909269 Cited on page(s) 71

[7] J. R. Barry, E. A. Lee, and D. G. Messerschmitt, Digital Communications, 3rd edition, Kluwer
Academic Publishers, Norwell, MA, 2004. Cited on page(s) 2

[8] T. Berger, Rate Distortion Theory, Prentice Hall, Englewood Cliffs, NJ, 1991. Cited on page(s)

[9] B. F.-Boroujeny, “OFDM versus filter bank multicarrier: development of broadband commu-
nication systems,” IEEE Signal Processing Magazine, vol. 28, no. 3, pp. 92–112, May 2011.
DOI: 10.1109/MSP.2011.940267 Cited on page(s)

[10] R. W. Chang, “High-speed multichannel data transmission with bandlimited orthogonal sig-
nals,” Bell. Sys. Tech. Journal, vol. 45, pp. 1775–1796, December 1966. Cited on page(s) 55,
72

[11] R. W. Chang, “Orthogonal frequency multiplex data transmission systems,” US-Patent,

US3488445 (a), January 1970 Cited on page(s) 55, 72
176 BIBLIOGRAPHY
[12] Y.-H. Chung and S.-M. Phoong, “Low complexity zero-padding zero-jamming DMT sys-
tems,” in Proc. of the 14th Eur. Signal Process. Conf. (EUSIPCO), Florence, Italy, September
2006, pp. 1–5. Cited on page(s)
[13] T. de Couasnon, R. Monnier, and J. B. Rault, “OFDM for digital TV broadcasting,” Signal
Processing, vol. 39, pp. 1–32, September 1994. DOI: 10.1016/0165-1684(94)90120-1 Cited
on page(s) 55, 93, 94
[14] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd edition, John Wiley,
Hoboken, NJ, 2006. Cited on page(s) 2, 97
[15] Z. Cvetković, “Modulating waveforms for OFDM,” in Proc. of IEEE Int. Conf. Acoust., Speech,
and Signal Process. (ICASSP), Phoenix, USA, March 1999, pp. 2463–2466.
DOI: 10.1109/ICASSP.1999.760629 Cited on page(s)
[16] P. S. R. Diniz, Adaptive Filtering: Algorithms and Practical Implementation. 4th edition, Springer,
New York, 2013. Cited on page(s)
[17] P. S. R. Diniz, E. A. B. da Silva, and S. L. Netto, Digital Signal Processing: Systems Analysis and
Design, 2nd edition, Cambridge, UK, 2010. Cited on page(s) 30, 31, 34, 72, 141
[18] K-L. Du and M. N. Swamy, Wireless Communication Systems: From RF Subsystems to 4G En-
abling Technologies, Cambridge University Press, Cambridge, UK, 2010. Cited on page(s)
[19] O. Edfors, M. Sandell, J.-J. van de Beek, D. Landström, and F. Sjöberg, “An introduction to
orthogonal frequency-division multiplexing,” Internal Report Lulea University of Technology,
Sweden, September 1996. DOI: 10.1002/9780470031384.ch2 Cited on page(s)
[20] B. L. Floch, M. Alard, and C. Berrou, “Coded orthogonal frequency division multiplex,” Pro-
ceedings of IEEE, vol. 83, no. 6, pp. 982–996, June 1995. DOI: 10.1109/5.387096 Cited on
page(s) 93
[21] M. B. Furtado Jr., P. S. R. Diniz, and S. L. Netto, “Redundant paraunitary FIR transceivers for
single-carrier transmission over frequency selective channels with colored noise,” IEEE Trans.
on Communications, vol. 55, no. 6, pp. 1125–1130, June 2007.
DOI: 10.1109/TCOMM.2007.898826 Cited on page(s)
[22] R. G. Gallager, Principles of Digital Communication, Cambridge University Press, Cambridge,
UK, 2008. Cited on page(s) 2, 95
[23] R. M. Gray, Toeplitz and Circulant Matrix: A Review, Now Publisher New York, NY, 2006.
Cited on page(s)
[24] R. Guan and T. Stromer, “Krylov subspace algorithms and circulant-embedding method for
efficient wideband single-carrier equalization,” IEEE Trans. on Signal Processing, vol. 56, no. 6,
pp. 2483–2495, June 2008. DOI: 10.1109/TSP.2007.912287 Cited on page(s)
 BIBLIOGRAPHY 177
[25] P. R. Halmos, Finite-Dimensional Vector Spaces. Boston: Springer, 1974.
DOI: 10.1007/978-1-4612-6387-6 Cited on page(s)

[26] S. Haykin, Communication Systems, 4th Edition, John Wiley, New York, NY, 2001. Cited on
page(s) 2, 95

[27] G. Heinig and K. Rost, “DFT representations of Toeplitz-plus-Hankel Bezoutians with appli-
cation to fast matrix-vector multiplication,” Linear Algebra Appl., vol. 284, no. 1–3, pp. 157–175,
November 1998. DOI: 10.1016/S0024-3795(98)10076-9 Cited on page(s) 123, 128, 132

[28] G. Heinig and K. Rost, “Hartley transform representations of inverses of real Toeplitz-plus-
Hankel matrices,” Numer. Funct. Anal. and Optimiz., vol. 21, no. 1–2, pp. 175–189, February
2000. Cited on page(s) 123, 128

[29] G. Heinig and K. Rost, “Hartley transform representations of symmetric Toeplitz matrix in-
verses with application to fast matrix-vector multiplication,” SIAM J. Matrix Anal. Appl., vol.
22, no. 1, pp. 86–105, May 2000. DOI: 10.1137/S089547989833961X Cited on page(s)

[30] A. Hjørungnes, P. S. R. Diniz, and M. L. R. de Campos, “Jointly minimum BER FIR MIMO
transmitter and receiver filters for binary signal vectors,” IEEE Trans. on Signal Processing, vol.
52, no. 4, pp. 1021–1036, April 2004. DOI: 10.1109/TSP.2003.822291 Cited on page(s) 169

[31] A. Hjørungnes, M. L. R. de Campos, and P. S. R. and Diniz, “Jointly optimized transmitter

and receiver filters FIR MIMO filters in the presence of near-end crosstalk,” IEEE Trans. on
Signal Processing, vol. 53, no. 1, pp. 346–359, January 2005. DOI: 10.1109/TSP.2004.838973
Cited on page(s) 170, 171

[32] A. Hjørungnes and P. S. R. Diniz,“Minimum BER FIR prefilter transform for communications
systems with binary signaling and known FIR MIMO channel,” IEEE Signal Processing Letters,
vol. 12, no. 3, pp. 234–237, March 2005. DOI: 10.1109/LSP.2004.842272 Cited on page(s)

[33] M. Joham, Optimization of Linear and Nonlinear Transmit Signal Processing, Dr. Ing. Thesis,
Technical University of Munich, Shaker Verlag, 2004. Cited on page(s)

[34] T. Kailath, S.-Y. Kung, and M. Morf, “Displacement ranks of a matrix,” Bulletin of The American
Math. Soc., vol. 1, no. 5, pp. 769–773, September 1979.
DOI: 10.1090/S0273-0979-1979-14659-7 Cited on page(s) 115

[35] T. Kailath and A. H. Sayed, “Displacement structure: theory and applications,” SIAM Review,
vol. 37, no. 3, pp. 297–386, September 1995. DOI: 10.1137/1037082 Cited on page(s) 115

[36] A. Lapidoth, A Foundation in Digital Communication, Cambridge, UK, 2009. Cited on page(s)

[37] M. Lazarus, “The great spectrum famine,” IEEE Spectrum, vol. 47, no. 10, pp. 26–31, October
2010. DOI: 10.1109/MSPEC.2010.5583459 Cited on page(s) 149
178 BIBLIOGRAPHY
[38] S. Lin and D. J. Costello, Error Control Coding, 2nd edition, Prentice Hall, Englewood Cliffs,
NJ, 2004. Cited on page(s) 2, 93, 94

[39] Y.-P. Lin and S.-M. Phoong, “Perfect discrete multitone modulation with optimal
transceivers,” IEEE Trans. on Signal Processing, vol. 48, no. 6, pp. 1702–1711, June 2000.
DOI: 10.1109/78.845928 Cited on page(s) 51

[40] Y.-P. Lin and S.-M. Phoong, “ISI-free FIR filter-bank transceivers for frequency selective
channels,” IEEE Trans. on Signal Processing, vol. 49, no. 11, pp. 2648–2658, November 2001.
DOI: 10.1109/78.960412 Cited on page(s) 51

[41] Y.-P. Lin and S.-M. Phoong, “Optimal ISI-free DMT transceivers for distorted channels with
colored noise,” IEEE Trans. on Signal Processing, vol. 49, no. 11, pp. 2702–2712, November
2001. DOI: 10.1109/78.960417 Cited on page(s)

[42] Y.-P. Lin and S.-M. Phoong, “BER minimized OFDM systems with channel independent
precoders,” IEEE Trans. on Signal Processing, vol. 51, no. 9, pp. 2369–2380, September 2003.
DOI: 10.1109/TSP.2003.815391 Cited on page(s)

[43] Y.-P. Lin and S.-M. Phoong, “OFDM transmitters: analog representation and DFT-based
implementation,” IEEE Trans. on Signal Processing, vol. 51, no. 9, pp. 2450–2453, September
2003. DOI: 10.1109/TSP.2003.815392 Cited on page(s) 55, 73

[44] Y.-P. Lin and S.-M. Phoong, “Minimum redundancy for ISI free FIR filter-bank transceivers,”
IEEE Trans. on Signal Processing, vol. 50, no. 4, pp. 842–853, April 2002.
DOI: 10.1109/78.992130 Cited on page(s) 42, 51, 52, 107

[45] Y.-P. Lin, S.-M. Phoong, and P. P. Vaidyanathan, Filter Bank Transceivers for OFDM and DMT
Systems, Cambridge University Press, Cambridge, UK, 2011. Cited on page(s) 42, 90, 95

[46] X. Ma, “Low-complexity block double-differential design for OFDM with carrier frequency
offset,” IEEE Trans. on Communications vol. 53, no. 12, pp. 2129–2138, December 2005.
DOI: 10.1109/TCOMM.2005.860052 Cited on page(s) 70

[47] U. Madhow, Fundamentals of Digital Communication, Cambridge University Press, Cambridge,

UK, 2008. Cited on page(s) 2

[48] W. A. Martins and P. S. R. Diniz, “Minimum redundancy multicarrier and single-carrier

systems based on Hartley transform,” in Proc. of the 17th Eur. Signal Process. Conf. (EUSIPCO),
Glasgow, Scotland, August 2009, pp. 661–665. Cited on page(s) 148

[49] W. A. Martins and P. S. R. Diniz, “Block-based transceivers with minimum redundancy,” IEEE
Trans. on Signal Processing, vol. 58, no. 3, pp. 1321–1333, March 2010.
DOI: 10.1109/TSP.2009.2033000 Cited on page(s) 128
 BIBLIOGRAPHY 179
[50] W. A. Martins and P. S. R. Diniz, “Suboptimal linear MMSE equalizers with minimum
redundancy,” IEEE Signal Processing Letters, vol. 17, no. 4, pp. 387–390, April 2010.
DOI: 10.1109/LSP.2010.2042515 Cited on page(s) 148
[51] W. A. Martins and P. S. R. Diniz, “Pilot-aided designs of memoryless block equalizers with
minimum redundancy,” in Proc. of the IEEE Int. Symp. Circuit. Syst. (ISCAS), Paris, France,
May 2010, pp. 275–279. DOI: 10.1109/ISCAS.2010.5537975 Cited on page(s) 142
[52] W. A. Martins and P. S. R. Diniz, “Memoryless block transceivers with minimum redundancy
based on Hartley transform,” Signal Processing, vol. 91, no. 2, pp. 240–251, February 2011.
DOI: 10.1016/j.sigpro.2010.06.026 Cited on page(s) 148
[53] W. A. Martins and P. S. R. Diniz, “Analysis of Zero-Padded Optimal Transceivers,” IEEE
Trans. on Signal Processing, vol. 59, no. 11, pp. 5443–5457, November 2011.
DOI: 10.1109/TSP.2011.2162327 Cited on page(s)
[54] W. A. Martins and P. S. R. Diniz, “LTI transceivers with reduced redundancy,” IEEE Trans. on
Signal Processing, vol. 60, no. 2, pp. 766–780, February 2012. DOI: 10.1109/TSP.2011.2174056
Cited on page(s) 142
[55] B. Muquet, Z. Wang, G. B. Giannakis, M. de Courville, and P. Duhamel, “Cyclic prefixing
or zero padding wireless multicarrier transmissions?,” IEEE Trans. on Communications, vol.
50, no. 12, pp. 2136–2148, December 2002. DOI: 10.1109/TCOMM.2002.806518 Cited on
page(s) 50, 90, 91, 141, 143
[56] A. F. Naguib, Adaptive Antennas for CDMA Wireless Networks, Ph.D. Thesis, Stanford Univer-
sity, Stanford, CA, August 1996. Cited on page(s)
[57] H. H. Nguyen and E. Shwedyk, A First Course in Digital Communications, Cambridge Univer-
sity Press, Cambridge, UK, 2009. Cited on page(s) 2
[58] S. Ohno, “Performance of single-carrier block transmissions over multipath fading channels
with linear equalization,” IEEE Trans. on Signal Processing, vol. 54, no. 10, pp. 3678–3687,
October 2006. DOI: 10.1109/TSP.2006.879321 Cited on page(s)
[59] S. Ohno and G. B. Giannakis, “Optimal training and redundant precoding for block transmis-
sions with application to wireless OFDM,” IEEE Trans. Communications, vol. 50, no. 12, pp.
2113–2123, December 2002. DOI: 10.1109/TCOMM.2002.806547 Cited on page(s)
[60] V. Y. Pan, “A unified superfast divide-and-conquer algorithm for structured matrices over
abstract fields,” MSRI Preprint 1999-033, Mathematical Science Research Institute, Berkeley,
CA, 1999. Cited on page(s)
[61] V. Y. Pan, Structured Matrices and Polynomials: Unified Superfast Algorithms, Springer, New
York, NY, 2001. DOI: 10.1007/978-1-4612-0129-8 Cited on page(s) 115, 117
180 BIBLIOGRAPHY
[62] V. Y. Pan and X. Wang, “Inversion of displacement operators,” SIAM J. Matrix Anal. Appl., vol.
24, no. 3, pp. 660–667, 2003. DOI: 10.1137/S089547980238627X Cited on page(s)

[63] F. Pancaldi, G. Vitetta, R. Kalbasi, N. Al-Dhahir, M. Uysal, and H. Mheidat, “Single-carrier

frequency domain equalization,” IEEE Signal Processing Magazine, vol. 25, no. 5, pp. 37–56,
September 2008. DOI: 10.1109/MSP.2008.926657 Cited on page(s) 30

[64] A. Peled and A. Ruiz, “Frequency domain data transmission using reduced computational com-
plexity algorithms,” in Proc. of IEEE Int. Conf. Acoust., Speech, and Signal Process. (ICASSP),
Denver, USA, April 1980, pp. 964–967. DOI: 10.1109/ICASSP.1980.1171076 Cited on
page(s) 55, 72

[65] J. G. Proakis, Digital Communications, 4th edition, McGraw-Hill, New York, NY, 2001. Cited
on page(s) 2

[66] J. G. Proakis and M. Salehi, Communication Systems Engineering, Prentice Hall, Upper Saddle
River, NJ, 1994. Cited on page(s) 2, 95

[67] T. S. Rappaport, Wireless Communications: Principles and Practice, 2nd Edition, Prentice Hall,
Upper Saddle River, NJ, 2002. Cited on page(s) 17

[68] C. B. Ribeiro, M. L. R. de Campos, and P. S. R. Diniz, “FIR equalizers with minimum redun-
dancy,” in Proc. of IEEE Int. Conf. Acoust., Speech, and Signal Process. (ICASSP), Orlando, USA,
May 2002, pp. 2673–2676. DOI: 10.1109/ICASSP.2002.5745198 Cited on page(s)

[69] C. B. Ribeiro, M. L. R. de Campos, and P. S. R. Diniz, “Zero-forcing multiuser detection

in CDMA systems using long codes,” in Proc. of IEEE Global Telecom. Conf. (Globecom), San
Francisco, USA, December 2003, pp. 2463–2467. DOI: 10.1109/GLOCOM.2003.1258679
Cited on page(s)

[70] C. B. Ribeiro, M. L. R. de Campos, P. S. R. Diniz, “Time-varying FIR transmultiplexers with

minimum redundancy,” IEEE Trans. on Signal Processing, vol. 57, no. 3, pp. 1113–1127, March
2009. DOI: 10.1109/TSP.2008.2010007 Cited on page(s) 36, 156

[71] A. Scaglione, S. Barbarossa, and G. B. Giannakis, “Filter-bank transceivers optimizing in-

formation rate in block transmissions over dispersive channels,” IEEE Trans. on Information
Theory, vol. 45, no. 3, pp. 1019–1032, April 1999. DOI: 10.1109/18.761338 Cited on page(s)

[72] A. Scaglione, G. B. Giannakis and S. Barbarossa, “Redundant filter-bank precoders and equal-
izers part I: unification and optimal designs,” IEEE Trans. on Signal Processing, vol. 47, no. 7,
pp. 1988–2006, July 1999. DOI: 10.1109/78.771047 Cited on page(s) 36

[73] A. Scaglione, G. B. Giannakis and S. Barbarossa, “Redundant filterbank precoders and equaliz-
ers part II: blind channel equalization, synchronization, and direct equalization,” IEEE Trans.
 BIBLIOGRAPHY 181
on Signal Processing, vol. 47, no. 7, pp. 2007–2022, July 1999. DOI: 10.1109/78.771048 Cited
on page(s)

[74] A. Schoonen, IQ imbalance in OFDM Wireless LAN systems, M.Sc. Thesis, Eindhoven Univer-
sity of Technology, Netherlands, 2006. Cited on page(s) 71

[75] B. Sklar, Digital Communications: Fundamentals and Applications, 2nd edition, Prentice Hall,
Upper Saddle River, NJ, 2001. Cited on page(s) 2

[76] T. Starr, J. M. Cioffi, and P. J. Silverman, Understanding Digital Subscriber Line Technology,
Prentice Hall, Upper Saddle River, NJ, 1999. Cited on page(s)

[77] G. Strang, Linear Algebra and Its Applications, Harcourt Brace Jovanovich San Diego, CA,
1988. Cited on page(s)

[78] A. Tarighat, R. Bagheri, and A. H. Sayed, “Compensation schemes and performance analysis
of IQ imbalances in OFDM receivers ,” IEEE Trans. on Signal Processing, vol. 53, no. 8, pp.
3257–3268, August 2005. DOI: 10.1109/TSP.2005.851156 Cited on page(s) 71

[79] S. Trautmann and N. J. Fliege, “Perfect equalization for DMT systems without guard interval,”
IEEE J. Sel. Areas in Communications, vol. 20, no. 5, pp. 987–996, June 2002.
DOI: 10.1109/JSAC.2002.1007380 Cited on page(s)

[80] D. Tse and P. Viswanath, Fundamentals of Wireless Communications, Cambridge University

Press, Cambridge, UK, 2005. Cited on page(s) 17, 107

[81] P. P. Vaidyanathan, Multirate Systems and Filter Banks, Prentice Hall, Englewood Cliffs, NJ,
1993. Cited on page(s) 34

[82] P. P. Vaidyanathan, “Filter banks in digital communications,” IEEE Circuits and Systems Mag-
azine, vol. 1, pp. 4–25, 2001. DOI: 10.1109/MCAS.2001.939098 Cited on page(s)

[83] P. P. Vaidyanathan, S.-M. Phoong, and Y.-P. Lin, Signal Processing and Optimization for
Transceiver Systems, Cambridge University Press, Cambridge, UK, 2010.
DOI: 10.1017/CBO9781139042741 Cited on page(s) 49, 90, 95, 126, 144, 169

[84] P. P. Vaidyanathan and B. Vrcelj, “Transmultiplexers as precoders in modern digital commu-

nication: a tutorial review,” in Proc. of the IEEE Int. Symp. Circuit. Syst. (ISCAS), Vancouver,
Canada, May 2004, pp. V-405–V-412. DOI: 10.1109/ISCAS.2004.1329590 Cited on page(s)

[85] M. Valkama, M. Renfors, and V. Koivunen, “Advanced methods for I/Q imbalance com-
pensation in communication receivers,” IEEE Trans. on Signal Processing, vol. 49, no. 10, pp.
2335–2344, October 2001. DOI: 10.1109/78.950789 Cited on page(s) 71
182 BIBLIOGRAPHY
[86] Z. Wang, and G. B. Giannakis, “Wireless multicarrier communications,” IEEE Signal Pro-
cessing Magazine, vol. 17, no. 3, pp. 29–48, May 2000. DOI: 10.1109/79.841722 Cited on
page(s)

[87] Z. Wang, X. Ma, and G. B. Giannakis, “OFDM or single-carrier block transmissions?,” IEEE
Trans. on Communications, vol. 52, no. 3, pp. 380–394, March 2004.
DOI: 10.1109/TCOMM.2004.823586 Cited on page(s) 30, 91

[88] S. B. Weinstein, “The history of orthogonal frequency-division multiplexing,” IEEE Commu-

nications Magazine, vol. 47, no. 11, pp. 26–35, November 2009.
DOI: 10.1109/MCOM.2009.5307460 Cited on page(s) 55

[89] S. B. Weinstein and P. M. Ebert, “Data transmission by frequency-division multiplexing using

the discrete Fourier transform,” IEEE Trans. on Communication Technology, vol. 19, no. 5, pp.
628–634, October 1971. DOI: 10.1109/TCOM.1971.1090705 Cited on page(s) 55, 72

[90] S. B. Wicker, Error Control Systems for Digital Communication and Storage, Prentice Hall, Upper
Saddle River, NJ, 1994. Cited on page(s) 2, 93

[91] X. G. Xia, “New precoding for intersymbol interference cancellation using nonmaximally dec-
imated multirate filterbanks with ideal FIR equalizers,” IEEE Trans. on Signal Processing, vol.
45, no. 10, pp. 2431–2441, October 1997. DOI: 10.1109/78.640709 Cited on page(s)

[92] Y. Yao and G. B. Giannakis, “Blind carrier frequency offset estimation in SISO, MIMO,
and multiuser OFDM systems,” IEEE Trans. on Communications vol. 53, no. 1, pp. 173–183,
January 2005. DOI: 10.1109/TCOMM.2004.840623 Cited on page(s) 70

[93] W. Zhang, X. Ma, B. Gestner, and D. V. Anderson, “Designing low-complexity equalizers for
wireless systems,” IEEE Communications Magazine, vol. 47, no. 1, pp. 56–62, January 2009.
DOI: 10.1109/MCOM.2009.4752677 Cited on page(s)

[94] Evolved Universal Terrestrial Radio Access (E-UTRAN): Multiplexing and Channel Coding,
3GPP TS 36212, ver. 8.7.0, 3rd Generation Partnership Project, May 2009. Cited on page(s)
144
 183

Authors’ Biographies

PAULO S. R. DINIZ
Paulo S. R. Diniz was born in Niterói, Brazil. He received the
Electronics Eng. degree (Cum Laude) from the Federal Uni-
versity of Rio de Janeiro (UFRJ) in 1978, a M.Sc. degree from
COPPE/UFRJ in 1981, and a Ph.D. from Concordia University,
Montreal, P.Q., Canada, in 1984, all in Electrical Engineering.
Since 1979 he has been with the Department of Electronic
Engineering (undergraduate) at UFRJ. He has also been with the
Program of Electrical Engineering (the graduate studies dept.),
COPPE/UFRJ, since 1984, where he is presently a Professor. He
served as Undergraduate Course Coordinator and Chairman of the Graduate Department. He has
received the Rio de Janeiro State Scientist award from the Governor of Rio de Janeiro.
From January 1991 to July 1992, he was a visiting Research Associate in the Department of
Electrical and Computer Engineering of University of Victoria, Victoria, B.C., Canada. He also
held a Docent position at Helsinki University of Technology. From January 2002 to June 2002, he
was a Melchor Chair Professor in the Department of Electrical Engineering of University of Notre
Dame, Notre Dame, IN, USA. His teaching and research interests are in analog and digital signal
processing, adaptive signal processing, digital communications, wireless communications, multirate
systems, stochastic processes, and electronic circuits. He has published several refereed papers in
some of these areas and wrote the books Adaptive Filtering: Algorithms and Practical Implementation,
4th ed., Springer, NY, 2012, and Digital Signal Processing: System Analysis and Design, 2nd ed.,
Cambridge University Press, Cambridge, UK, 2010 (with E.A.B. da Silva and S.L. Netto).
He has served as the Technical Program Chair of the 1995 MWSCAS held in Rio de Janeiro,
Brazil. He was the General co-Chair of the IEEE ISCAS2011, and Technical Program co-Chair of
the IEEE SPAWC2008. He has been on the technical committee of several international conferences
including ISCAS, ICECS, EUSIPCO, and MWSCAS. He has served as Vice President for region
9 of the IEEE Circuits and Systems Society and as Chairman of the DSP technical committee of
the same Society. He is also a Fellow of IEEE. He has served as associate editor for the following
Journals: IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing from
1996–1999, IEEE Transactions on Signal Processing from 1999–2002, and the Circuits, Systems and
Signal Processing Journal from 1998–2002. He was a distinguished lecturer of the IEEE Circuits and
Systems Society from 2000–2001. In 2004, he served as distinguished lecturer of the IEEE Signal
184 AUTHORS’ BIOGRAPHIES
Processing Society and received the 2004 Education Award of the IEEE Circuits and Systems
Society. He also holds some best-paper awards from conferences and from an IEEE journal.

WALLACE A. MARTINS
Wallace A. Martins was born in Brazil in 1983. He received
an Electronics Engineering degree (Cum Laude) from the Fed-
eral University of Rio de Janeiro (UFRJ) in 2007, and M.Sc. and
D.Sc. degrees in Electrical Engineering from COPPE/UFRJ in
2009 and 2011, respectively. He worked as a technical consul-
tant for Nokia Institute of Technology (INDT), Brazil, and for
TechKnowledge Training, Brazil. In 2008, he was a research vis-
itor at the Department of Electrical Engineering, University of
Notre Dame, Notre Dame, IN. Since 2010 he has been with the
Department of Control and Automation Industrial Engineering, Federal Center for Technological
Education Celso Suckow da Fonseca (CEFET/RJ – UnED-NI), where he is presently a Lecturer
of Engineering. His research interests are in the fields of digital communication, microphone ar-
ray signal processing, and adaptive signal processing. Dr. Martins received the Best Student Paper
Award from EURASIP at EUSIPCO-2009, Glasgow, Scotland.

MARKUS V. S. LIMA
Markus V. S. Lima was born in Rio de Janeiro, Brazil in 1984.
He received an Electronics Engineering degree from the Federal
University of Rio de Janeiro (UFRJ) in 2008, an M.Sc. degree in
Electrical Engineering from COPPE/UFRJ in 2009, and is cur-
rently pursuing his D.Sc. degree at COPPE/UFRJ. He has served
as a teaching assistant for the following undergraduate courses
taught at UFRJ: Digital Transmission, Digital Signal Process-
ing, and Linear Systems. His main interests are in adaptive signal
processing, microphone array signal processing, digital commu-
nications, wireless communications, statistical signal processing,
and linear algebra.