Proactive Relay Selection with Joint Impact of Hardware Impairment and Co-channel Interference
Performance Analysis and Joint Source-Channel Coding
-
Upload
khangminh22 -
Category
Documents
-
view
1 -
download
0
Transcript of Performance Analysis and Joint Source-Channel Coding
SINGLE AND MULTIPLE ANTENNA
COMMUNICATION SYSTEMS: PERFORMANCE
ANALYSIS AND JOINT SOURCE-CHANNEL CODING
by
FirouzBehnamfar
A thesissubmittedto the
Departmentof ElectricalandComputerEngineering
in conformitywith therequirementsfor
thedegreeof Doctorof Philosophy
Queen’sUniversity
Kingston,Ontario,Canada
September2004
Copyright c�
FirouzBehnamfar, 2004
Abstract
Performanceanalysisplaysa centralrole in thedesignof reliablecommunicationsys-
tems. It helpsto reducethe probability of error in the transmissionof digital dataor to
improve the quality of the recoveredsignalswhenanalogsignalsarecommunicated.In
this dissertation,we first addresstheproblemof performanceanalysisby presentingana-
lytical methodsto deriveupperandlowerboundsfor theerrorratesof varioussystems.We
thendesigna joint source-channelcodingsystemfor the transmissionof analogsources
overamultipleantennacommunicationlink.
We begin by consideringa single-antennasystemwith additive white Gaussiannoise
aswell asblock Rayleighfadingchannels.It is shown thatonecanreducethecomplexity
of someof the existing methodsand, at the sametime, obtain tighter results,which is
very favorableparticularlywhenthe complexity of the problem(i.e., the codebooksize)
is prohibitive. Closed-formformulasfor theblock pairwiseerrorprobabilitiesarederived
everywhere.
Wethenconsideramulti-antennacommunicationlink with space-timeorthogonalblock
codingandestablishtight upperandlowerboundson thesymbolandbit errorratesof the
system.Thesignalingschemeandthemappingbetweenthesignalsandbits canbearbi-
trary.
We next considerthe casewherethe input bit-streamhasa non-uniformdistribution,
which allows the useof maximuma posteriori (MAP) decoding. We derive the MAP
decodingrule, calculatea closed-formformula for the symbolpairwiseerrorprobability,
andestablishtight upperand lower boundsfor the symbolandbit error rates. Another
ii
contribution is to demonstratethat systemswith MAP decoding,which is a form of joint
source-channelcoding,canoutperformthosewith separatesourceandchannelcoding(i.e.,
tandemsystems)atall errorratesof interest.
Finally, we proposea joint source-channelcodingsystemfor the transmissionof ana-
log sourcesover multiple-antennalinks. The proposedsetupusessoft-decisiondecod-
ing channel-optimizedvectorquantization(COVQ) to achieve a largeportionof thesoft-
decodinggain with reducedcomplexity. Two COVQs with fixed encodersandadaptive
decodersarealsodesigned.Thesuperiorityof theproposedCOVQ over tandemsystems
is alsodemonstrated.
iii
Acknowledgments
I would like to expressmy sincerethanksto my supervisors,ProfessorFady Alajaji and
ProfessorTamasLinder, for their support,continuousencouragement,andthoughtfulsug-
gestionsandcommentsthroughoutthecompletionof this work. I amalsogratefulto my
committeemembersProfessorSteve Blostein,ProfessorLorneCampbell,ProfessorAmir
Khandani,andProfessorShahramYousefi,for spendingvaluabletimeto readmy thesis,at-
tendingmy defense,andproviding detailedandconstructivecomments.I amalsothankful
to ProfessorMartin Guayfor chairingmy defensesession.
I wouldalsoliketo gratefullyacknowledgethefinancialsupportreceivedfrom theNat-
uralSciencesandEngineeringResearchCouncil(NSERC),theOntarioPremier’sResearch
ExcellenceAward (PREA), the Departmentof ElectricalandComputerEngineering,the
Departmentof MathematicsandStatistics,andtheSchoolof GraduateStudies.
I amgreatlyindebtedto my bestfriend andloving wife, Arezou,whosepatience,un-
derstanding,andsacrificesenabledme to pursuemy studies.I cannotthankher enough.
I have beenfortunateto have wonderfulparents,brother, andsisters;I thankthemall for
their constantloveandstimulation.
iv
Contents
Abstract ii
Acknowledgments iv
List of Figures xi
1 Introduction 1
1.1 TheCommunicationSystemModel . . . . . . . . . . . . . . . . . . . . . 2
1.2 TheSource-ChannelSeparationTheorem . . . . . . . . . . . . . . . . . . 4
1.3 PerformanceEnhancement. . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 Improvementvia JointSource-ChannelCoding . . . . . . . . . . . 6
1.3.2 Improvementvia Diversityin SpaceandTime . . . . . . . . . . . . 7
1.4 Summaryof theThesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.1 ProblemStatement. . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.2 ThesisOrganizationandContributions. . . . . . . . . . . . . . . . 14
v
1.5 LiteratureReview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5.1 ErrorAnalysisof ChannelCodedSISOSystems . . . . . . . . . . 16
1.5.2 JointSource-ChannelCodingMethods . . . . . . . . . . . . . . . 17
1.5.3 Multi-antennaSystems. . . . . . . . . . . . . . . . . . . . . . . . 21
2 Bounds on the Probability of a Finite Union of Events 33
2.1 TheLowerandUpperBounds . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.2 TheHunterUpperBound. . . . . . . . . . . . . . . . . . . . . . . 34
2.1.3 TheKouniasLowerBound . . . . . . . . . . . . . . . . . . . . . . 35
2.1.4 ThedeCaenLowerbound . . . . . . . . . . . . . . . . . . . . . . 35
2.1.5 TheKAT LowerBound . . . . . . . . . . . . . . . . . . . . . . . 36
2.1.6 TheCohenandMerhav Lowerbound . . . . . . . . . . . . . . . . 36
2.2 Algorithmic Implementationof theHunterandKouniasBounds . . . . . . 37
2.2.1 TheHunterUpperBound. . . . . . . . . . . . . . . . . . . . . . . 37
2.2.2 TheKouniasLowerBound . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Final Remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 A Brief Review of Space-Time Orthogonal Block Codes 41
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1.1 Encodingin Alamouti’sSystem . . . . . . . . . . . . . . . . . . . 42
vi
3.1.2 Decodingin Alamouti’sSystem . . . . . . . . . . . . . . . . . . . 43
3.1.3 Extensionto ThreeandFour TransmitAntennas . . . . . . . . . . 45
3.2 GeneralTreatmentof Space-Time OrthogonalBlock Codes. . . . . . . . . 47
3.3 TheSymbolPEPfor GeneralSpace-Time OrthogonalBlock Codes . . . . 48
3.4 TheCodeword PEPof Arbitrary Space-Time Codes. . . . . . . . . . . . . 50
4 Bounds on the Block and Bit Error Rates of Coded AWGN and Block Rayleigh
Fading Channels 53
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 TheErrorRatesin Termsof Probabilitiesof aUnion . . . . . . . . . . . . 54
4.3 TheAWGN Channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.4 TheBlock RayleighFadingChannel . . . . . . . . . . . . . . . . . . . . . 57
4.5 LinearBlock CodesandBPSKSignaling . . . . . . . . . . . . . . . . . . 60
4.5.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.5.2 Evaluationof theAlgorithmic Boundsfor LargeBlocks . . . . . . 61
4.6 NumericalResults. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5 Bounds for the Symbol and Bit Error Rates of Space-Time Orthogonal Block
Codes 76
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2 Boundson the Codeword Pairwise Error Probability of Space-Time Or-
thogonalBlock Codes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
vii
5.3 Boundson theErrorRatesof Space-Time OrthogonalBlock Codes. . . . . 79
5.3.1 TheSymbolErrorRate . . . . . . . . . . . . . . . . . . . . . . . . 79
5.3.2 TheBit ErrorRate . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.4 NumericalResults. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6 Error Rates of STOB Coded Systems under MAP Decoding 94
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.2 TheMAP DecodingRule . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.2.1 DecoupledSymbolDetection . . . . . . . . . . . . . . . . . . . . 96
6.2.2 TheMAP DecodingRule. . . . . . . . . . . . . . . . . . . . . . . 97
6.3 ExactSymbolPairwiseErrorProbabilitywith MAP Detection . . . . . . . 98
6.3.1 TheConditionalPEP . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.3.2 Interpretationof thePEPasaLaplaceTransform . . . . . . . . . . 99
6.3.3 ThePEPin ClosedForm . . . . . . . . . . . . . . . . . . . . . . . 99
6.4 TheOptimalBinaryAntipodalSignaling. . . . . . . . . . . . . . . . . . . 102
6.5 Boundson theSERandBER of STOB CodedChannelsunderMAP De-
coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.5.1 TheSymbolErrorRate . . . . . . . . . . . . . . . . . . . . . . . . 104
6.5.2 TheBit ErrorRate . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.6 NumericalResults. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
viii
6.6.1 BinaryAntipodalSignaling . . . . . . . . . . . . . . . . . . . . . 105
6.6.2 TandemversusJointSource-ChannelCoding . . . . . . . . . . . . 106
6.6.3 SERandBERBoundsfor � -arySignaling. . . . . . . . . . . . . 107
6.6.4 TheEffectof ConstellationMapping . . . . . . . . . . . . . . . . 108
7 Quantization with Soft-decision Decoding for MIMO Channels 125
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.2 SystemComponents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.2.1 TheChannel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.2.2 Soft-DecisionDecodingandtheEquivalentDMC . . . . . . . . . . 130
7.2.3 TheStep-Sizeof theUniform Quantizerat theDecoder. . . . . . . 133
7.3 Quantizationwith Soft-DecisionDecoding. . . . . . . . . . . . . . . . . . 135
7.3.1 Soft-DecisionDecodingCOVQ . . . . . . . . . . . . . . . . . . . 135
7.3.2 Fixed-EncoderAdaptive-DecoderSoft-DecisionDecodingCOVQ . 136
7.3.3 On-lineFEAD Soft-DecisionDecodingCOVQ . . . . . . . . . . . 138
7.4 NumericalResultsandDiscussion . . . . . . . . . . . . . . . . . . . . . . 139
7.4.1 ImplementationIssues . . . . . . . . . . . . . . . . . . . . . . . . 139
7.4.2 COVQ for VariousMIMO Channels. . . . . . . . . . . . . . . . . 140
7.4.3 COVQ versusTandemCoding . . . . . . . . . . . . . . . . . . . . 141
7.4.4 COVQ, FEAD COVQ, andOn-lineFEAD COVQ . . . . . . . . . 142
ix
List of Figures
1.1 Block diagramof adigital communicationsystem. . . . . . . . . . . . . . 2
1.2 Theideaof thesource-channelseparationprinciplefor discretesources. . . 5
1.3 Improvementof channelcapacitywhenmultiple antennasareused. The
CSNRis 20dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Block diagramof acommunicationsystemwith MAP sequenceestimation. 11
1.5 Block diagramof a typical multi input-multi output system,where �������� ������������is encodedto ������� �� ��� ������ ���! , attenuatedby " , receivedas#
, anddecodedto $� whichcorrespondsto $�%� � $��� ������ $����� . . . . . . . . . . 22
2.1 Formingtheoptimalspanningtreeto evaluateaBonferroni-typeupperbound. 38
3.1 Thetransmitterof Alamouti’sspace-timeorthogonalblock codingscheme. 42
4.1 Theideato form aspanningtreeandcomputeanupperboundfor theBlER.&('is theindex of codeword )+*�, for which -.*/,10324-.*50 & �76 �������8 . . . . . . . 66
4.2 Block andbit error ratecurvesfor BCH (15, 5) code,8-PSKmodulation,
andAWGN channel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
xi
4.3 Block error rate curves for BCH (15, 7) code, BPSK modulation,and
AWGN channel.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 Comparisonamongthreelowerboundson theBlER. Golay(23,12) code,
BPSKsignaling,andAWGN channel. . . . . . . . . . . . . . . . . . . . . 70
4.5 Block error rate boundsfor Golay (23, 12) code,BPSK signaling,and
AWGN channel.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.6 Block error rate boundsfor BCH (31, 16) code, BPSK signaling, and
AWGN channel.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.7 Block error rate boundsfor BCH (63, 24) code, BPSK signaling, and
AWGN channel.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.8 Block error rateboundsfor Hamming(7, 4) code,BPSK signaling,and
Rayleighfadingchannel. . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.9 Blockerrorrateboundsfor BCH(15,5)code,BPSKsignaling,andRayleigh
fadingchannel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.1 Our upperboundandthelower boundson -:9;9 ��=< ��?>3@ 9 < *< * < ** >;> , Alam-
outi’scodewith Q-PSKmodulation,AB�DC , and EF�76 . . . . . . . . . . . 85
5.2 Resultsfor Q-PSKsignalingwith AB�DC and EF�76 (2 bits/s/Hz). . . . . . 86
5.3 Resultsfor 8-PSKsignalingwith AG�:C and EH�I6 (3 bits/s/Hz).. . . . . . 87
5.4 Resultsfor 16-QAM signalingwith AJ�:C and EH�:C (4 bits/s/Hz). . . . . 88
5.5 Resultsfor 32-PSKsignalingwith AJ�:C and EH�LK (5 bits/s/Hz). . . . . . 89
5.6 Resultsfor 64-QAM signalingwith AJ�:C and EH�:C (6 bits/s/Hz). . . . . 90
xii
5.7 Resultsfor 8-PSKsignalingwith AG�DM and EH�I6 C and K (1.5bits/s/Hz). 91
5.8 Thestar-QAM signalingschemewith quasi-Graymapping.N is anormal-
izing factorandequals6;OPC�Q 6SR and& � Q T 6 . . . . . . . . . . . . . . . . . 92
5.9 Resultsfor Star8-QAM signalingwith AB�4K and EF�U6 (1.5bits/s/Hz).. . 93
6.1 Resultsfor BPSKsignaling,AB�LC , and EF�U6 . . . . . . . . . . . . . . . . 110
6.2 Resultsfor theoptimumbinaryantipodalsignaling,AJ�:C , and EF�I6 . . . 111
6.3 ComparisonbetweenBPSKandoptimumsignalingschemes,AV�WC , andEF�I6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.4 ComparisonbetweentandemandMAP-decodedschemesfor BPSK sig-
naling, AB�:C , and EF�I6 . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.5 ComparisonbetweentandemandMAP-decodedschemesfor 16-QAM sig-
naling, AB�:C , and EF�I6 . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.6 SERcurvesfor 8-PointStarQAM modulationwith quasi-Graymapping,AB�:C , and EF�I6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.7 BER curvesfor 8-PointStarQAM modulationwith quasi-Graymapping,AB�:C , and EF�I6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.8 SERcurvesfor 16-QAM signaling,AB�LC , and EX�U6 . . . . . . . . . . . . 117
6.9 SERcurvesfor 32-PSKsignaling,AB�LC , and EF�U6 . . . . . . . . . . . . . 118
6.10 BER curvesfor 32-PSKsignaling,AB�LC , and EF�U6 . . . . . . . . . . . . . 119
6.11 SERcurvesfor 64-QAM signalingwith Graymapping,AJ�:C , and EH�I6 . 120
xiii
6.12 Thestar-QAM signalingschemewith M1 mapping. . . . . . . . . . . . . . 121
6.13 Comparisonbetweenthequasi-GrayandtheM1 mappingswith StarQAM
for a systemwith AB�:C , and EF�I6 . . . . . . . . . . . . . . . . . . . . . . 122
6.14 Comparisonbetweenthe M1 andGray mappingsfor 64-QAM signaling,AB�:C , and EF�I6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.15 ComparisonbetweentheBER performanceof theM1 andGraymappings
for 64-QAM signaling,AJ�DC , and EF�I6 . . . . . . . . . . . . . . . . . . . 124
7.1 Systemblock diagram,whereevery Y bits in a Z\[ -bit index8
is transmit-
tedvia a space-timecodeword �U�]�^� �� ������ ���_ (with orthogonalcolumns��* & �`6 �������a ��b�Y ), receivedas# �c�ed �� ������ df�_ , andspace-timesoft-
decodedas g# ��� gd � ������ gdfhi . For simplicity, wehaveassumedthat Yj�:Z\[ . 128
7.2 Linearcombinerwhich follows thespace-timesoftdecoderat thereceiver.
Left: problemstatement,right: solutionto theproblem. . . . . . . . . . . . 144
7.3 DMC capacityversusthestep-sizeof theuniformquantizer, AB�:C , EF�I6 . 146
7.4 SimulatedSDRin dB for azero-meanunit-varianceGauss-Markov source
( kl�mR �on ) vectorquantizedat rate [p�`6 � R bpsandsoft-decisiondecoded
with q bits. Quantizationdimensionis Zr�LC . . . . . . . . . . . . . . . . . 148
7.5 Jointlydesignedversustandemcodingschemesfor azero-meanunit-variance
Gauss-Markov source( ks�UR �tn ). Quantizationdimensionis Zu�vC andthe
overall rateis [w�LM � R bps. AJ�:C and EF�I6 . . . . . . . . . . . . . . . . . 149
xiv
7.6 SimulatedSDR in dB for an i.i.d. xy�eR 6; sourcevectorquantizedat rate[z�DC � R bpsandsoft-decodedwith q bits. Quantizationdimensionis Zr�LC .AB�:C and EF�I6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.7 Comparisonamongthe COVQ, FEAD VQ, andOn-line FEAD VQ for axy�eR 6; Gauss-Markov sourcevectorquantizedat rate [F� 6 � R bpsand
harddecoded.Quantizationdimensionis Zs�DC , AB�DC , and EF�76 . . . . . 151
xv
Chapter 1
Introduction
The next generationof wirelesscommunicationsystemsfacesthe demandfor increased
datarates,highermobility, larger carrierfrequencies,andmorelink reliability. Wireless
channelsarecharacterizedby fading,multipath,limited bandwidth,andfrequency andtime
selectivity which make systemdesigna challenge.It is thereforecrucialto have anunder-
standingof thebehavior of wirelesschannelsin orderto know theirperformancelimits and
to beableto designefficient communicationsystemsfor them.This dissertationconsiders
theanalysisof theperformanceof wirelesssystemswith singleor multipleantennasat the
transmitandthereceive sides.It addressesbothof thecaseswherethetransmitteddatais
or is not protectedagainstchannelnoise(i.e., channelcoded).An applicationof theanal-
ysisresultsis presentedby designinga robustcommunicationsystemfor thetransmission
of analogdatasuchasspeechovera multipleantennacommunicationlink.
This chapterbeginswith a shortdescriptionof conventionalcommunicationssystems.
Two methodsto improve the performanceof the generalsystemarethendescribed.The
1
inputsequence { encoder { modulator |
waveformchannel}demodulator}decoder}reconstructedsequence
discretechannel
Figure1.1: Block diagramof a digital communicationsystem.
two problemsconsideredin this thesis,i.e., performanceanalysisandsystemdesign,are
next introduced.Thechapterendswith a review of therelatedliterature.
1.1 The Communication System Model
A typical communicationsystemis depictedin Figure1.1. We considerthe transmission
of discrete-time(discreteor continuous-amplitude)sequencesover noisywaveformchan-
nels.Theassumptionthattheinput is discretein time is justifiedby thesamplingtheorem
whichstatesthatany band-limitedsignalcanberepresentedby its sampleswithout lossof
informationprovidedthatit is sampledataratewhichis at leasttwicethesignalbandwidth
[143, 162]. In practice,thesignalsto be transmittedby a communicationsystemarefirst
low-passfiltered,andhenceit is reasonableto considerband-limitedsignals.
Theencoderanddecoderaresignalprocessingblocks. Theencoderaimsto represent
the sourcein an efficient form (sourceencoding)and to protectthe informationagainst
channelnoise(channelencoding). The decoderexploits the error detectionstructureto
canceltheeffectof channelnoiseandreconstructtheinputmessage.
2
Theuseof sourcecodingis inevitabledueto restrictionsonchannelbandwidthor stor-
agespace. Sourceencodingaims to remove the redundancy of the input and represent
it with as low a bit rateand inaccuracy aspossible. In effect, the sourceencodermaps
the input sequenceto a string of binary symbols. For analogsources,this mappingin-
corporatesquantization,which causespartial lossof information.Theencoderalsoinserts
controlledredundancy to helperrordetectionandcorrectionatthereceiver, whichis known
aschannelcoding.Themodulatormapsblocksof certainlengthfrom theencoderoutputto
multidimensionalwaveforms,sothatthedatasequencecanbetransmittedover theanalog
channel.We assumethatpulseshapingis alsoincludedin themodulatorblock. By pulse
shaping,we meanusinga modifiedwaveforminsteadof therectangularpulse,suchasthe
raisedcosinepulse,in orderto reduceinter-symbolinterference(ISI) at thereceiver [150].
Amplitudeattenuationandnoiseadditionarethe two maineffectsof thewaveformchan-
nel. While bothof thesephenomenaoccurin channelswith afrequency responsenotequal
to unity, suchastheRayleighfadingchannel,thesoleeffectof theadditivewhiteGaussian
noise(AWGN) channelis theadditionof a constantpower densitynoiseto thewaveform
at thechannelinput. Thedemodulatorestimateswhatwaveformwastransmittedover the
analogchannel.Theoutputof thedemodulatoris a discrete-timestreamwhich is passed
throughthe decoderso that the errorscausedby the channelaredetectedandcorrected.
The decoderalsomapsits input sequenceto symbolsthat are in a subsetof the original
inputalphabetto reconstructthetransmittedmessage.Notethattheinputandoutputof the
concatenationof the modulator, the waveform channel,andthe demodulatorarediscrete
sequencesandhencethesethreeblocksmayberegardedasadiscrete(in termsof timeand
alphabet)channelasindicatedin Figure1.1.
3
1.2 The Source-Channel Separation Theorem
The designof the encoder-decoderpairs can be performedbasedon resultsfrom infor-
mationtheory. The resultsstatethat, for a givendistortion ~ , thereexistsa rate ���e~p –
calledtherate-distortionfunction– below which thesourcecannotbecompressedwith a
distortionlessthan ~ . In thepositivedirection,therate ���e~p is asymptoticallyachievable
at distortionlevel ~ by encodinglarge blocksof the sourcetogether. The resultsfurther
statethat,for everycommunicationchannel,thereexistsaquantitycalledchannelcapacity�, which is themaximumratethat informationcanbe transmittedover thechannelwith
arbitrarily small probability of error. The rate-distortionfunction dependson the source
statistics.Thecapacityof a single-antennawaveformchanneldependson thechannelfre-
quency responseandnoisedistribution. For discretechannels,channelcapacitydepends
on thechannelcrossoverprobabilities.
Letusassumethatasource-channelcodewith transmissionrate[ sourcesymbolsO channel
symbolsis usedat the encoder. From the lossy joint source-channelcoding theorem
[160, 161,63,31], it follows thataslongas���e~p �[=� �holds,communicationwithin distortion ~ is asymptoticallypossible.
Many communicationsystemsaredesignedaccordingto thefollowing interpretationof
thefirst two results:“communicationwith distortion ~ canbeachievedby designingtwo
independentcodes:a sourcecodewith rate [S� anda channelcodewith rate [ � . Theonly
constraintis that ����~p �[ ��O�[S��� � ; thereis noperformancepenaltyassociatedwith thesep-
arateindependentdesign.” Thisresultis calledthesource-channelseparationprinciple.For
4
discretesource { sourceencoder { channelencoder |discretechannel}channeldecoder}sourcedecoder}decodedmessage
error-freechanneluniformi.i.d. source
Figure1.2: Theideaof thesource-channelseparationprinciplefor discretesources.
discretememorylesssources,if we requirelosslessbit errorrateperformance(this results
in the Hammingdistancedistortionmeasure),thenwe replace���e~p by sourceentropy� ���� to require� ���� �[ ��O�[S��� �
. This resulthasgiven rise to the communicationsys-
temmodelof Figure1.2. Accordingto this model,channelcodingwill correctall channel
errors,no matterwhatthesourcestatisticsare,andhencethesourceencoder/decoderpair
shouldbedesignedasif thechannelwaserror-free,no matterwhatthechannelconditions
are.
1.3 Performance Enhancement
In this section,we describetwo methodswhich leadto performancegainsover thecom-
municationsystemof Figure1.2.Thefirst schemereconsiderstheseparationof thesource
andchannelcodesandthesecondmethodchangesthephysicalsetupby deploying multiple
antennasat thetransmitandreceivesidesto obtainperformancegains.
5
1.3.1 Improvement via Joint Source-Channel Coding
Thesource-channelseparationprinciplesimplified thetheoryandtechnologyby dividing
adifficult probleminto two lesschallengingproblemswhichmaybesolvedindependently.
In agreementwith the literature,we usethe term“tandemsystems”for systemsdesigned
basedon this principle. Thereare numeroussuccessfultheoreticalresultsandpractical
systemsavailabletodaywhich arebasedon tandemsystems.In fact, the performanceof
sometandemsystemsis becomingverycloseto thetheoreticallimits for simplechannels,
suchassingle-userGaussiannoiselinearchannels[29, 151]. However, thereareanumber
of factsabout the separationapproachwhich merit attention. The first point is that in
order for this result to hold, the sourceandchannelcodelengthsneedto grow without
bound.As a result,communicationsystemsbuilt from independentlydesignedsourceand
channelcodesmay requiregreatercomputationalresourcesand causemore delay than
jointly designedsystems.Separation-basedcodesmaybefar from theoptimaldesignfor
systemswith critical constraintsondelayandcomplexity, suchaswirelesscommunication
systems.Joint source-channelcodingwasshown in [123] to outperformtandemcoding
for systemshaving delayor complexity below a certainthreshold. It wasfurther proved
in [216] that givena probability of block error, optimal joint source-channelcodingmay
requirehalf of theencoding/decodingdelayof optimaltandemcoding.
Secondly, the separationprinciple ignoresthe imperfectionsin real systems.In par-
ticular, sourcecodesaredesignedassumingthat the channelcodeswill correctall errors
introducedby thechannel.Knowing thatthis is notalwaysthecaseevenfor themostpow-
erful channelcodes,oneis not informedby this theoremhow to modify thesourcecodes
6
to dealwith noisychannels.Similarly, channelcodesareusuallydesignedassumingthat
all bits createdby thesourcecodeareof equalimportance[31]. This assumption,which
resultsfrom theassumptionof a perfectsourcecode,leadsto thedesignof channelcodes
thatprotectall bits equally. In a largenumberof applicationsthesourceoutputbit-stream
is thebinaryexpansionof somedecimalnumbersor it is relatedto more“sensitive” parts
of thedata(asin thecasewheretransformcoefficientsareencoded)andhenceallocating
moreprotectionto the moresignificantbits than the lesssignificantbits is a betterway
to usethebit rate. Indeed,equalerrorprotectioncausesperformancedegradationin such
conditions.
The third point is that therearechannelsfor which the separationprinciple doesnot
hold. An exampleof suchchannelsis the multiuserchannelwhosecapacitydependson
thelengthof thesourcecode[31]. Clearly, joint designof thesourceandchannelcodesis
amoreviableway to designcodesfor suchsystems.
Severalmethodshave beenproposedfor joint source-channelcoding.They arebriefly
describedandreviewedin Subsection1.5.2.
1.3.2 Improvement via Diversity in Space and Time
Diversityprovidesthe receiver with several replicasof the transmittedsignal. Therefore,
it is a powerful techniqueto mitigate fading and interferencefrom other users,thereby
improving link reliability. Dependingon thedomainwherediversity is created,diversity
techniquesmaybe divided into threemain categories;i.e., temporal,frequency, andspa-
tial. Thefirst two techniquesnormally introduceredundancy in thetime and/orfrequency
7
domain,causinglossin transmissionefficiency. Spatialdiversitytechniquesemploy mul-
tiple antennasfor transmissionand/orreceptionwithout necessarilysacrificingbandwidth
or power. Therefore,it is naturalto considerchannelswith multiple input (transmit)and
multipleoutput(receive)antennas,oftenreferredto asMIMO channels,to obtainimproved
reliability and/orincreaseddatarateswhereverpossible.
HigherdataratesandmorereliablecommunicationoverMIMO channelsarepromised
by informationtheoreticresults.For thecoherentcase;i.e., whenthe receiver knows the
channelfadingcoefficientsperfectly, andfor independentRayleighfading,channelcapac-
ity at ahighchannelsignal-to-noiseratio (CSNR)is approximatelyequalto [196]�U�D�r��� �eA E� ����P���\� bits/sec/Hz
where A and E arethenumbersof transmitandreceive antennasand ��� is theCSNR.For
thenon-coherentcase,whichis thescenariothatthechannelis notknown at thetransmitter
or thereceiver, channelcapacityathighCSNRapproximatelyequals[215]�U� A '�� 6 T A '� ��� ���+� �\� bits/sec/Hz
where A ' � �s��� �eA E � ��OPCP and� � is the coherencetime of the channel,that is the
numberof symbolintervalsoverwhich thechannelis static.
Both of theabove resultsarevery significant,becausethey show that thecapacityin-
creaseslinearly in theminimumof thenumberof receive andtransmitantennas(provided
that� � is largeenoughfor thenon-coherentcase)andthereforeit canbemuchlargerthan
thesingle-antennacase.To increasechannelcapacityin thesingle-antennascenario,one
hastwo options: increasingthe bandwidth,for which the capacityis asymptoticallyup-
perboundedby a constant,or increasingthe transmitpower which leadsto a logarithmic
8
growth in capacity. Linearincreasein channelcapacityis thereforeveryattractivefor wire-
lessapplications.
For thecoherentcase,channelcapacityis plottedin Figure1.3for threecases;namely,
for a systemwith a singletransmitantennabut multiple receive antennas,a systemwith a
singlereceive elementbut multiple transmitantennas,anda systemwith thesamenumber
of transmitandreceive antennas.It is observedthat thegainof having multiple antennas
at bothsidesis muchlarger thanhaving a numberof antennasat only oneside. If oneis
constrainedto have a singleantennaat onesideof the channel,thenreceive diversity is
morepromising.Transmitdiversity is thechoicewhenthemobilehaslimited processing
resources.
One signal processingapproachto exploit the larger capacityof MIMO channelsis
spatialmultiplexing. In thisapproach,which is typically known astheBLAST architecture
[56], the input bit-streamis de-multiplexed to a numberof brancheswhereconventional
channelcodingmaybeemployed. Theinformationat eachbranchis thentransmittedin a
certainorderover the transmitantennas.At the receiver side,thesignalsareestimatedin
a step-by-stepfashionaccordingto theCSNRof the received signals.Anotherapproach
which combinescodingandmodulationis known asspace-timecoding.Space-timecodes
introducetemporalandspatialcorrelationinto thesignalstransmittedfrom differentanten-
naswithout increasingthe total transmitpower andoftenwithout extra bandwidth.There
is a diversity gain that resultsfrom multiple pathsbetweenthe transmitand the receive
terminals,andacodinggainthatresultsfrom thecorrelationamongthesymbolsacrossthe
transmitantennas1.
1In this work, we areinterestedin point-to-pointcommunications.Therefore,cooperativediversity[158,
103, 114], in which themobilesrelayeachother’smessagesto obtaindiversitygains,is not considered.
9
1 2 3 4 5 6 7 8 9 100
5
10
15
20
25
30
35
40
Number of antennas
Cha
nnel
Cap
acity
in n
ats
K = LK = 1L = 1
Figure1.3: Improvementof channelcapacitywhenmultipleantennasareused.TheCSNR
is 20 dB.
10
¡£¢�¤¦¥ ��§�¨©¨©¨©§ ¥ ��ª { channelencoder[ � « ¢�¤ ¬ ��§�¨©¨©¨©§ ¬® ª { modulatorC°¯ ± ¢X² ³ � ¤¦´ ��§�¨©¨©¨©§ ´ � ª|channel} µ ¢4¤ ¶ � §�¨©¨©¨©§ ¶ � ªMAP channeldecoder
·± ¢�¤ ·´ � §�¨©¨©¨©§ ·´ � ª·¡3¢�¤ ·¥ � §�¨©¨©¨©§ ·¥ � ª}Figure1.4: Block diagramof acommunicationsystemwith MAP sequenceestimation.
1.4 Summary of the Thesis
1.4.1 Problem Statement
Error Analysis of Communication Systems
Let us considerthe communicationsystemin Figure1.4, which encodesa groupof
databits �u� ����� ������������at a time andmapsthemto a sequenceof basebandsignals.The
above mappingmay includeany kind of channelcodingwith rate [� , suchaslinearblock
codes,for which the codedbits aredenotedby )�� �5¸S�� �������¸ �with ¹º�¼»½O�[ � , or trellis
codedmodulation. Eachof the signal sequenceswill be called a “frame” and denoted
by ��*w� �o¾ * � �������¾ *� �� & �¿6 ÀSÀSÀ_�Á , whereÁ �ÂC � ,
a �»½Oi��[ ��ÃÄ is the frame length
in symbols,and C�¯ is the constellationsize to which¾ *Å belongs. We assumethat
¾ *Å is
selectedfrom a unit-energy constellationandweightedlaterat themodulator. When¾ *Å is
transmittedandfor a transmitpowerof ÆÇ� �È[ �1ÃÉÆËÊ , whereÆËÊ is thepowerperanuncoded
(information)bit, theinput-outputrelationshipof thechannelis representedby� Å ��Ì Æ£� � ¾ *Å�ÍÏÎ Å 11
where� Å and Î Å arethechanneloutputandadditivenoise,respectively, and�
is thechan-
nelgain.Weassumethat Î Å , � , and¾ *Å areindependent.Theadditivenoiseis assumedto be
complex Gaussianwith independentandidenticallydistributed(i.i.d.) realandimaginary
parts,eachhaving distribution x���R ÎzÐ OPC+ . We denotesucha distribution by Ñ�xy�eR ÎzÐ .We hereinconsiderthreecases:
� �V6Ò� ¸�Ó ¹ ¾ Ô�� , which indicatesthe AWGN channel,�beingi.i.d. Ñ�x��eR 6; amongframes,which is theblock Rayleighfadingmodel,and
�following an i.i.d. chi squaredistribution, which modelstheblock Rayleighfadingchan-
nel with diversityreceptionandalsothespace-timeorthogonalblock codedchannels(see
Chapter3). In thefadingcases,weassumecoherentdetection,which is thecasewherethe
receivercanaccuratelyestimatethechannelgain�
.
We are interestedin finding upperand lower boundson the error ratesof the above
system.Moreprecisely, wewantto deriveupperandlowerboundsontheframeerrorrate,
which is theprobabilitythata frameis detectedin error;i.e.,
FER �LÕ Ö.� $�Ø×�:�� where � and $� arethetransmittedanddecodedframes,respectively, aswell astheproba-
bility thata bit is detectedwrongly, which is givenby
BER � 6» ÙÚ ÛÜ � Ã.�^� Û Ú 0ÝÜ�Û ~�Þ3�eß áà �-j� $�=�D��0�â¦���D� Û where ~�Þ£�eß áà is thenumberof bit differencesbetweenthedatabits correspondingto � Ûand ��0 and -j� $�ã����0�â5�ã�y� Û is theconditionalprobability that $�ã�y��0 is detectedgiven
that ���L� Û is transmitted.
Weareinterestedin bothuniformandnon-uniformi.i.d. binaryinputsequences.There-
fore,weconsidermaximuma posteriori(MAP) decoding.
12
Quantization for MIMO Wireless Channels
Unlike the previous systemin which the input is a bit-stream,in many applicationsthe
problemat handis to transmitananalogsource,suchasthespeechor imagesignals,over
adigital communicationsystem.In suchcases,it is requiredto quantizethesource;i.e., to
approximateit with asourcewhich hasafinite alphabet.
The performanceof a quantizeris expressedby its signal-to-distortionratio (SDR).
Therefore,whenan analogsourceis transmitted,it is reasonableto chooseto minimize
theend-to-enddistortionof thesystemasthedesigngoal,asopposedto approacheswhich
opt to minimize the systembit, symbol,codeword, or frameerror rate. We measurethe
inaccuracy by a distortionmeasure,suchasthesquarederrorbetweentheoriginal andthe
decodedsequences.
Let ä and $ä be, respectively, the original and reconstructedsourcevectors,i.e., the
systeminput andoutput. We are interestedin this caseto transmitanalogsourcesover
communicationlinks with multiple transmitandpossiblymultiple receive antennas.Our
objective is to minimizethemean-squareerror;namely, wewantto minimize~m�DÆÈåçæ�ä T $ä�æ �+èsubjectto constraintson the averagetransmitpower andchannelbandwidth. The band-
width constraintcansimplybemetby fixing thequantizationrateanddimension,themod-
ulationscheme,andtheconstellationsizeasdescribedin [150,200].
13
1.4.2 Thesis Organization and Contributions
This dissertationis divided into eight chapters.Relevant upperandlower boundson the
probabilityof a unionof a finite numberof eventsarereviewedin Chapter2. We present
theseboundsin their generalform andemploy themthroughtoutthis thesisto establish
upperandlower boundson theerror ratesof variouscommunicationsystems.In Chapter
3, we briefly review the space-timeorthogonalblock codes. Thesecodesarefrequently
usedin the following chaptersof the thesis. The maximumlikelihood (ML) decoding
rule andthe symbolpairwiseerror probability of space-timeorthogonalblock codesare
alsoreviewed. As a contribution, an alternatederivation of the codeword pairwiseerror
probabilityof generalspace-timecodesis presented.
In Chapter4, we derive upperandlower boundson the error ratesof channelcoded
singleinput-singleouput(SISO)AWGN aswell asblock Rayleighfadingchannelswith
arbitrarysignalingschemesandunderML decoding.In this chapter, we show how to use
the HunterandKouniasupperand lower bounds,which areBonferroni-typebounds,to
boundthe error ratesof suchsystemsandalsoimprove an alreadyexisting lower bound,
known asthe KAT bound. We alsodevise a way to make the computationof the above
boundsfeasiblewhenthecodebooksizeis large. Furthermore,we find a closed-formex-
pressionfor theprobabilityof theintersectionof two pairwiseerroreventsunderRayleigh
fading.
MIMO systemswhich usespace-timeorthogonalblock codesandML decodingare
consideredin Chapter5. As suchMIMO channelsareequivalentto a numberof parallel
SISOchannelswith chi squarefading,a similar approachto that in Chapter4 is usedto
14
derive tight upperand lower boundson the symbolandbit error ratesof the system. A
closed-formexpressionfor theprobabilityof the intersectionof two pairwiseerrorevents
is alsoderived.
We considertheerror ratesfor orthogonalspace-timecodedchannelsunderMAP de-
coding in Chapter6. We derive the MAP decodingrule and show that, similar to ML
decoding,symboldetectionis decoupledin thiscase.Themaincontributionof thischapter
is thatwederiveaclosed-formexpressionfor thesymbolpairwiseerrorprobability. Signal
mappingandcomparisonwith tandemsystems,which performsourceandchannelcoding
independently, arealsopresented.
In Chapter7,wedesignacommunicationsystemfor thetransmissionof analogsources
overMIMO channels.Thecontributionis to show how to usespace-timesoft-decodingand
soft-decisiondecodingchannel-optimizedvectorquantization(COVQ) to achieve perfor-
mancegainsover tandemsystemscomprisingvector quantization,channelcoding, and
space-timecodingandalsoover systemswhich employ COVQ with harddecoding.Two
COVQs which do not assumethe knowledgeof the CSNRat the transmitterarealsode-
signed.
Chapter8 presentsasummaryof thethesis.
1.5 Literature Review
In what follows, we mentionthe major work on the subjectof this thesis;namely, per-
formanceanalysis,joint source-channelcoding,andspace-timecodes.For anoverview on
15
thefundamentalproblemsin informationtheory, onecanreferto Shannon’soriginalpapers
[160, 161], furtherdescriptionsonratedistortiontheoryin [163], andGallager’stutorialon
Shannon’scontributionsin [64].
1.5.1 Error Analysis of Channel Coded SISO Systems
Themostcommonestimateof theerrorrateof acommunicationsystemis theunionbound
whichonly dependsonthepairwiseerrorprobabilities(see,e.g.,[150]). Theunionbound,
which presentsan upperboundto the codeword error rate, is looseparticularly at low
CSNRvalues. Shannonpresenteda lower boundto the codeword error rateof systems
with AWGN channelandÁ
-ary modulationin [164] usinga spherepackingapproach.
Shannon’s lowerboundis tight at low CSNR,but it becomeslooseastheCSNRgrows. An
upperboundto thecodeword errorrateof BPSK-modulatedAWGN channelswasderived
in [149], extendedto PSK-modulatedchannelsin [84], andtightenedfor somelongbinary
blockcodesin [211]. A lowerboundto deCaen’s lowerboundontheprobabilityof aunion
[39] wasderived in [156] andappliedto BPSK-modulatedAWGN channelswith linear
blockcodes.Althoughthelowerboundin [156] is poorevenatmoderatelylow errorrates,
it is significantbecauseit is asymptoticallytight andconvergesto theunionupperbound
athighCSNR[156] andit only dependsontheweightenumerationof thecode.Thelower
boundin [39] wasimprovedin [109], wheretheKAT boundwasintroducedby solvingthe
sameoptimizationproblemasin [39], but usingstrongeranalysis.Sincethelowerbounds
in [39, 109] arestrongerthanthe boundin [156], they alsoconverge to the union upper
boundastheCSNRgrows to infinity whenappliedto thesamecommunicationsystemas
16
in [156]. The lower boundsin [39, 156, 109] areall in termsof sumsof ratios. Another
improvementof theboundin [39] is givenin [30], wherethe lower boundis expressedin
termsof aweightedsumof ratiosandtightenedvia carefulselectionof theweights(aswell
asexhaustivesearch).Thesemethodsaredescribedin detail in Chapter2.
1.5.2 Joint Source-Channel Coding Methods
Wecategorizejoint source-channelcodingapproachesasunequalerrorprotection,channel-
optimizedvectorquantization,optimizationof index assignment,andexploitation of the
(residual)redundancy at thesourceencoderoutputvia MAP decoding.COVQ andMAP
decodingarereviewedherewith moreemphasissincethesetwo methodsareusedin the
following chapters.
Unequalerrorprotection(UEP) tradesoff sourceresolutionandchannelerrorprotec-
tion to allocatetheavailablebit rate.Therateof thechannelcodedependson thechannel
conditions,suchasthebit error rate. As thechannelbecomesmorenoisy, thenumberof
bits usedfor channelcodingincreases.Theaim in UEPis to find theoptimalbalancebe-
tweenthesourceandchannelcoderates.Most UEPsystemsareimplementedusingrate-
compatiblepuncturedconvolutional (RCPC)codesfor error protection[76], often with
a measureof parity checking,suchascyclic redundancy check(CRC) [205], for detec-
tion of errorsin the channel-decodedsequence.Goodexamplesof UEP systemsinclude
thework in [187] – whereentropy-codedsubbandcodingtogetherwith packetizationand
RCPCchannelcodingwereusedfor imagetransmissionover memorylessnoisychannels
– andthework in [165] which follows [187], but insteadof adhocsourcecoding,it uses
17
the strongimagecoderSPIHT introducedin [152]. A differentapproachto UEP canbe
foundin [4], wherelargerenergy levelsandsmallersignalingschemeswereassignedto the
moresensitive transformcoefficientsin imagetransformcodingusingthediscretecosine
transform(DCT). A major sourceof complexity in UEP methodsis the allocationof the
availablebit ratebetweenthesourceandchannelcodes.An adaptive sourceandchannel
codingrateallocationschemefor finite statechannels,includingtheGilbertchannelmodel
[68, 48], wasexaminedin [94]. In [142], thelogarithmof thechannelcodeblockerrorrate
wasassumedto havea linearrelationwith thechannelcoderate.Rateallocationwasnext
doneusingiterativeoptimization.
Channeloptimized scalar/vector quantizationis anotheralternative to joint source-
channelcoding.In thisapproach,channelconditionsin theform of COVQ index crossover
probabilitiesareincorporatedin quantizerdesign.As aresult,thesourceis compressedde-
pendingon thechannelconditions.Theoptimality conditionsof a generaldigital commu-
nicationsystemwith noisychannelsandreal-valuedinputandoutputwerestudiedin [54].
Later, thenecessaryconditionsfor optimality of scalarquantizationfor binarysymmetric
channelsgivena fixedindex assignmentwerederivedin [113]. In [50], it wasshown that
the algorithmin [113] neednot converge anda realizability constrainton the codewords
wasimposedto guaranteeconvergence.ThegeneralizedLloyd algorithm(GLA) [66] for
joint source-channelcodingover noisychannelswasstudiedin [14] for trellis andpredic-
tive codingandit wasshown that theconsideredschemesoutperformedtandemsystems.
Theapproachestowardvectorquantizationfor noisychannelsbeganwith thework in [43]
wheretheoptimality conditionsof theencoderwerefound. Finally, in [111], COVQ op-
timality conditionswereformulated.Amongthevariousalgorithmsthatareproposedfor
18
COVQ design,we cite the modifiedGLA initialized by simulatedannealing[49], noisy
channelrelaxation[60, 61], stochasticrelaxation[214], deterministicannealing[136], and
COVQ designusingfuzzylogic [92]. Theoreticalstudiesonvectorquantizationweredone
in [125, 126, 213]. For example,it wasshown in [126] that, asexpected,the distortion
of a quantizerapproachesto what is predictedby theratedistortionfunctionasthequan-
tizer dimensionincreases.In [125], therateof convergenceof a quantizerin thepresence
of channelnoisewasstudied. In [147], COVQ designfor the Polyachannelwith finite
memory[5] wasaddressedandit wasshown thatsignificantgainscanbeachievedthrough
exploiting the memoryof the channel(insteadof interleaving). The work in [25] used
subbandcoding,all-passfiltering, andCOSQfor subbandimagetransmissionover mem-
orylessnoisy channels.This methodis similar in spirit to the approachin [20, 21, 22],
wherechannelmemoryis exploitedanddiscretewavelettransform(DWT)-basedsubband
coding, which is substantiallylesscomplex due to reducedfilter lengths,is used. Soft
decodingCOVQ wasintroducedin [200,179] andappliedto imagecodingin [177]. Soft-
decodingCOVQ is computationallydemandingdueto the needfor matrix inversionand
evaluationof trigonometricfunctions.In aneffort to reducethecomputationalcomplexity
of the decoder, COVQ with soft-decisiondecoderwasdevelopedin [6] for non-coherent
Rayleighfadingchannelsandusedin [146] for channelswith additiveGaussiannoisewith
inter-symbolinterference.
Anotherapproachto joint source-channelcodingassignsindicesto thequantizercode-
vectorsaccordingto channelconditions.Thesourceis codedfor a noiselesschanneland
noexplicit channelcodingis applied.Then,takingthechannelcrossoverprobabilitiesinto
account,indicesareassignedto source(scalaror vector)samplessuchthattheend-to-end
19
distortionis minimized.Theindex assignmentapproachto joint source-channelcodingwas
studiedin [212], whereclosercodevectorsin theEuclideandistancesensewereassigned
to closerindicesin theHammingdistancesensein aniterativemanner. Thesimulatedan-
nealingalgorithmwasappliedin [49] to find thebestassignmentof indicesvia perturbing
theindex assignmentandacceptingthelabelingwhichcouldincreasedistortionaccording
to aprobabilisticmeasure.In [106], thecodevectorswerewrittenasa linearsummationof
theindex bits andtheHadamardtransformwasusedto find thebestindex assignment.
The fourth methodof joint source-channelcodingconsiderssourcea priori probabil-
ities in channeldecoding. Ideally, a sourcecoderwould compressany discrete-alphabet
sequenceinto ani.i.d. bit-streamhaving a rateequalto thesourceentropy rate[31]. Nev-
ertheless,no sourceencoderis ideal.Theoutputof thesourceencoderwould thencontain
redundancy in theform of memoryand/ornon-uniformdistribution. A MAP detectormay
exploit this residualredundancy to gain performanceimprovementsover ML detection.
MAP decodingwasfirst usedin [154] for scalarquantization. This work was later ex-
tendedto the vectorcasein [148]. In [145], the characteristicsof the arithmeticsource
coderandthesourceweremodeledby a Markov sourceof orderoneandMAP decoding
wasemployed using the model statistics. A similar approachwas taken in [52], where
MELP-compressedspeechwasmodeledasa first-degreeMarkov chain. SequenceMAP
decodingfor varioustrellis codingsystemsandchannelconditions,includingtheRayleigh
fadingchannel,wasstudiedin [13, 75]. MAP decodingwasusedfor channelswith memory
in [7, 178] and,later, in [182], wheretheGilbert channelmodelwasused.Robustspeech
transmissionvia MAP decodingwasdevelopedin [8] andapplicationof MAP decodingto
imagecommunicationovernoisychannelswasstudiedin [210].
20
1.5.3 Multi-antenna Systems
Communicationsystemswith receive antennadiversity have attracteda lot of attention
sincerelatively longago.In fact,in currentcellularapplications,receivediversityis already
in usefor improving receptionfrom mobiles.A few generalideason receive diversityare
explainedin [150]. A moreadvancedstudy canbe found, for example,in [16]. In the
following, we cite themajor referenceson thestudieson channelswith multiple transmit
andmultiple receive antennas.This subsectionis divided into four partswhich address
the informationtheoreticissues,codedesign,performanceanalysis/evaluation,andother
issueswhichareimportantin MIMO channelsbut arenotdirectly relatedto this thesis.
Let us considera multiple-antennasystemwhich employs A transmitand E receive
antennasasdepictedin Figure1.5. The encoderconvertsa block of » input bits � into
matrix codewordsof size A é a symbolsdenotedby �c�¿�^� �� ��� ������ ���! , where � Å �� ¾+�áê Å �¾ � ê Å �������¾Sëìê Å áí with¾ * ê Å beingthe signalassignedto transmitantenna
&at time indexÔ
and í denotingtransposition.At time indexÔ, the entriesof � Å are transmittedsimul-
taneouslyfrom the transmitantennas.At the receiver, the E receive antennascollect the
incomingsignalsattenuatedby fadingandperturbedby additive noise. Thereceivedsig-
nalsarethendemodulatedanddecoded.
In this thesis,we considerRayleighflat fading, so that the complex path gain from
transmitantenna&
to receive antennaà, denotedby
� 0 ê * , hasa zero-meanunit-variance
complex Gaussiandistribution, denotedby Ñix���R 6; , with i.i.d. realandimaginaryparts.
In many scenarios,it is assumedthat the channelis quasi-static,which meansthat the
pathgainsremainconstantduring a codeword transmission,but vary in an i.i.d. fashion
21
¡ { encoder î ï 2
ï 1ï ë... ðñ ññ ñ�òó ó 2 ô1 ôh ô ... õ decoder
·î {·¡Figure 1.5: Block diagramof a typical multi input-multi output system,where �`��5��� ������������
is encodedto �J� �^� �� �ö� ������ ���! , attenuatedby " , received as#
, andde-
codedto $� which correspondsto $�÷� � $��� ������ $����� .from onecodeword interval to theother. Theadditivenoiseat receiver
àat time
Ô, Î 0 ê Å , is
assumedto be ÑixW��R 6; distributedwith i.i.d. realandimaginaryparts.Weassumethatthe
input,fadingcoefficients,andchannelnoiseareindependentfrom eachother. Basedonthe
above,for a CSNRof �\� at eachreceive branchandat timeÔ, thesignalat receiveantennaà
canbewrittenas �ø0 ê Å � Ì ùûúëýü ë* Ü � � 0 ê * ¾ * ê Å ÍþÎ 0 ê Å , or in matrix form,d Å � ÿ �\�A "U� Å Í � Å (1.1)
whered Å ���e� �áê Å ������ � h ê Å áí , � Å �W� Î �áê Å ������ Î h ê Å áí , and " � � � 0 ê * � .Information Theoretic Studies
Researchon systemswith multiple transmitantennasbegan in the late 1980sand early
1990s.Thepotentialfor lower error ratesat no extra bandwidthandtransmitpower with
multiple transmitantennaswasdemonstratedvia simulationsin [206, 207,208, 209]using
delayelements,transmit-antennadiversity, andequalization. It wasshown in [196, 57]
that the capacityof fast fading MIMO channelsscaleslinearly in the minimum of the
numberof the transmitandreceive antennasassumingindependentfadingbetweeneach
22
transmit-receive antennapair. This potentialincreasewasshown to hold also for block
Rayleighfadingchannelsin [134]. In [87], it wasshown that temporaldiversity canbe
replacedwith transmitantennadiversityto gainnon-zerocapacityin asinglecoherencein-
terval if thenumberof transmitantennasandthecoherencetime cangrow without bound.
Capacityreductiondueto spatialcorrelationwasinvestigatedthroughMonte-Carlosimu-
lationsin [169] andtheoreticallyin [26], whereit wasshown that thereductionin outage
andShannoncapacityis not significantwhenthecorrelationcoefficient betweenthe fad-
ing coefficients is lessthan0.5. In [27] somedegeneratechannelmodelsto accountfor
spatialcorrelationand keyhole (whereeachentry in the channelmatrix is a productof
two complex Gaussianrandomvariables)wereinvestigatedandvalidatedthroughphysical
measurements.Usingtheideaof [72], tight upperboundsfor thecapacityof spatiallycor-
relatedanddouble-scaterringMIMO channelsweredevelopedin [168] which canbeused
togetherwith the lower boundderivedin [144]. Thecapacityof keyholeMIMO channels
wascalculatedin closedform in [168]. Using randommatrix theory [93] andGaussian
approximationof theunaveragedchannelcapacity, asymptoticbehavior of MIMO outage
capacitywasstudiedin [72, 138, 159]. Finally, in [59], spherepackingboundswerede-
rived for codeword error rateof space-timecodesover block Rayleighfadingchannels.
It wasshown that thecodeword error ratesignificantlyimprovesasthecodewordsspana
largernumberof fadingblocks.
Code Design for Multi-Antenna Systems with Coherent Detection
Thestepnext to informationtheoreticstudiesis to designcodesto achieve a performance
ascloseaspossibleto theinformationtheoreticlimits. In thissubsection,weconsidercode
23
designmethodsfor coherentscenarios;namely, themethodsin which thechannelmatrix
is assumedto beknown at thereceiver. This assumptionis dueto thefact thatthechannel
fadingcoefficientsneedto beestimatedat thereceiver for synchronizationpurposes.
We divide codedesignapproachesinto threegroups:spatialmultiplexing, space-time
coding,andadhocmethods.Spatialmultiplexing containsmostlytheschemeswhich add
little (or no) redundancy at the transmitter, andusesignalprocessingtechniquesfor de-
tection.Spatialmultiplexing wasintroducedin [56], wherethediagonalBell Laboratories
layeredspace-time(D-BLAST) architecturewasdeveloped.In D-BLAST, the input data
bits arefirst de-multiplexed into A sub-sequences,eachof which may be channelcoded
by a conventional(SISO)encoder. The codedbits are thenmodulatedandsentvia an-
tennasin turn; for example,modulator1 sendsits first symbolvia antenna1, its second
symbolvia antenna2, andsoon. At the receiver, thesymbolsaredetectedin an iterative
mannervia “nulling andcancelling”which detectsandremovestheeffect of thesymbols
with thehighestsignal-to-noiseratioateachiteration.TheD-BLAST decodersuffersfrom
beingtoo computationallyintensive. A simplifiedversionof D-BLAST, calledthevertical
BLAST (V-BLAST) wasproposedin [58], wheremodulator&
sendsA symbolsover theA transmitantennasat the& th symbolperiod.Typicalvariationsof BLAST arethework in
[129, 157], whereinterleaving, BLAST, anditerative decodingwereusedto improve the
BLAST performance.Themajordrawbackof BLAST is its complexity asthenumberof
transmitand/orreceiveantennasgrows.
In orderto designspace-timecodes,thecodewordpairwiseerrorprobability(PEP)was
consideredin [194]. The codeword PEPis the probability of the event that when � was
sentandbetween� and $� , $� is detected.Weshalldenotethis probabilityby Õ ÖS�(� @ $�£ .24
UsingtheChernoff bound,it wasshown in [194] thatthecodeword PEPis upperbounded
by Õ Ö;�á� @ $�Ç � �����\��O�K� h�� where [ is the rank of the matrix � T $� and � , alsoknown asthe codingadvantage,is
givenby �þ�mâ©�á� T $�� �(� T $�£ �� â � � where� denotescomplex conjugatetransposeand â�� â for a matrix � is thedeterminantof� . Theproduct E [ is known asthediversityorder. Thecodedesigncriteriaproposedin
[194] is to maximizetheworse-case(minimum) rank [ anddeterminantof thecodeword
differences.The maximumvaluethat [ cantake on is A . Codesfor which [ �JA are
calledfull-rank (or full diversity)codes.
The first coding schemefor multiple transmitantennaswas designedin [9], where
a rate-1full-diversity space-timecodewith orthogonalcolumnswasdesignedfor a dual
transmitantennasystem.Using a generalizationof the theoryof orthogonaldesigns,the
work of [9] wasextendedto any numberof transmitantennasin [192], wherecodesfor
two to eight transmitantennaswere given. The codewordsof suchcodeshad orthogo-
nal columnsandwerehenceknown asspace-timeorthogonalblock (STOB) codes.Such
codesdo not dependon thenumberof receive antennas.Thekey advantageof space-time
orthogonalblock codesis that decodingof the STOB codedsymbolsis decoupled. In
otherwords,onecanperformscalarML decodinginsteadof vectordecodingandhence
have polynomialasopposedto exponentialdecodingcomplexity. An importantresult in
[192] is that rate-1space-timeorthogonalblock codesexist only for two transmitanten-
25
naswith complex signalingschemes.Space-timeorthogonalcodeswerealsostudiedin
[82, 120, 121, 185]. An interestingresultin [121] is that themaximumrateof space-time
orthogonalcodeswith complex signalingand AJ�DC�» or AJ�DC�» T 6 transmitantennasis��» Í 6S ®OPC�» . Space-timeorthogonalblockcodesarereviewedin moredetailin Chapter3.
Space-timeorthogonalblock codesachieve full diversity gain, but they fail to have
a large codinggain. In order to achieve a codinggain (at the price of increaseddecod-
ing – andpossiblyencoding– complexity), onemayusea concatenationof conventional
channelcodingandeitherspace-timeorthogonalcodes(see,for example,[10]) or BLAST
(suchas the work in [132]). Otherchoicesarespace-timetrellis coding [194, 95] (and
their designvia computersearch[18, 73, 189]), usingiterative (turbo)codingideas(with
parallel [34, 98, 98, 184] andserial [23, 70, 124, 127, 183] encoderconcatenation),and
non-orthogonalblock coding.An importantclassof high-ratenon-orthogonalblock codes
wasintroducedin [80] andcalledthe lineardispersion(LD) code. Every entryof anLD
codeword is aweightedsumof thebasebandsignalswith theweightschosensuchthatthe
mutualinformationbetweenthechannelinput andoutputcodewordsis maximizedgiven
the numberof transmitandreceive antennas.Othermajor space-timecodedesignmeth-
odsincludethework in [35, 36], whereconstellationrotationandtheHadamardtransform
wereused,and in [46, 77] and their extensionin [47], wherea connectionbetweenthe
rankcriterionof [194] with binaryfieldsto designspace-timecodesfor PSKconstellations
usingAlgebraictoolswascreated.Amongadhocmethods,we cite thework in [118] that
usedconventionalchannelcodingto obtaincodinggainfollowedby randomconstellation
mappingof thesamedataat every transmitbranchto achievediversitygain.
26
Performance Evaluation/Analysis
Ourmainfocusin this thesis(exceptfor Chapter7) is performanceanalysisof digital com-
municationsystems. Many conventionalmethodsfor analysisof SISO communication
systemsbecomeinfeasiblein multi-antennasystems(e.g.,dueto their large codebooks).
Therefore,it is requiredto find waysto eitherextendtheexistingmethodsto MIMO chan-
nelsor to comeup with new analysisstrategies.
Early paperson space-timecodedesignderivedtheChernoff upperboundon thepair-
wiseerrorprobabilitybetweencodewords[194, 74]. Chernoff upperboundsfor channels
with correlatedRician fadingandcorrelatednoisecanbefound in [41]. Thepairwiseer-
ror probability underfastfadingwasconsideredin [173], wherean integral solutionwas
found,andin [199] wherethesolutionwasin termsof a derivative. In [190] a form of the
momentgeneratingfunctionof theML metricwasderivedwhich wassimplerto evaluate
for block fadingchannels,and the PEPswere found numerically. The exact expression
and a lower boundfor the pairwiseerror probability of arbitrary space-timecodewords
underquasi-staticfadingwerefoundin [131] (we will presentanalternative derivationin
Chapter3). For a land-mobileto satellitelink, thepairwiseerrorprobabilitywasfoundin
[197] for Rician-lognormalfadingandin [198] for Rician-Nakagami-» fading.A moment-
generatingfunctionapproachwasusedin [176], wherethepairwiseerrorprobabilityof the
so-calledsuper-orthogonalspace-timetrellis codeswasfound,andin [222], whereupper
boundson thesymbolerrorrateof D-BLAST systemswithout channelcodingwereeval-
uated. The transferfunction methodcanbe usedto derive upperboundson the bit error
rateof space-timetrellis codesaswasthecasein [173, 223],but theboundsareoftenover
27
anorderof magnitudelargerthansimulationresultsevenat highsignal-to-noiseratiosand
they becometoo complex to evaluatefor largediversityorders.
For standardÁ
-ary PSK and QAM constellationswith Gray signal mapping[150],
thestandardapproachfor performanceanalysisin [175], which is widely adoptedin many
references,is to usethe exact conditionalsymbolerror rateandapproximateconditional
bit error rateformulasof [175] which areconditionedon the channelfadingcoefficients.
For example,theconditionalsymbolerrorrate -ì�°�� � of standardÁ
-ary PSKis givenby- �°�� £â � ���É � 6� ��� Ù < ����� ÙÐ ������� T �\�! ��� � � � O Á "� � ��� �$# %�& # (1.2)
andtheconditionalSERof standardsquareÁ
-ary QAM is givenby- ���� Çâ � ���É � KPq� � �� �Ð ����� � T �\�áb QAM � � ��� �$# % & # T KPq �� � ��('Ð �)�*� � T ���áb QAM � � ��� �+# % & # (1.3)
where qp� 6 T 6;O Q Á and b QAM � M�OPCi� Á T 6; . Similar expressionshold for theBER,
which we do not reportherefor brevity. To derive the error rates,oneneedsto find the
distribution of theequivalentfadingcoefficientsbasedon which thesymbolerrorrateand
approximatebit error maybe evaluated.An exampleof suchan approachis thework in
[53], in which an approximateexpressionfor the bit error rateof space-timeorthogonal
block codedchannelswith correlatedNakagami-» fadingwasderived. In [19] and[166],
the symbolerror rate for space-timeorthogonalcodedchannelswith i.i.d. Rayleighand
keyhole Nakagami-» fading were found, respectively (althoughthe resultsof the latter
referenceare complex to evaluate). An interestingresult in the latter paperis that the
symbolerrorrateof keyholespace-timeorthogonalcodedchannelswith agrowing number
28
of receive antennasapproachesthat of the i.i.d. channelswith a single receive antenna,
which is expectedbecausekeyhole channelshave only onedegreeof freedom(i.e., their
diversitygainin one).Notethattheaboveresultsholdonly for standardsignalingschemes
andGraysignalmapping.
Miscellaneous Issues
We have left out many areaswhich areappliedin MIMO systems,becausethey arenot
directly relatedto this thesis.Many of theseissuesareimportantby themselves. An im-
portantissueis space-timecodedesignfor non-coherentdetection.Dif ferentialandnon-
coherenttechniquesin communicationsdo not requireanexplicit carrierphasereference.
Thesetechniquesarespeciallyneededfor wirelessradio communicationsystems,where
carrierphasesynchronizationis difficult dueto fading,high mobility, severeinterference,
andtheuseof shortdatapackets.Non-coherentdetectionis requiredalsowhentherearea
lot of antennasin thesystemandtrainingsequencesbecometoo largeto bepossible.The
ideasfor communicationovernon-coherentSISOchannelscanbeappliedto MIMO chan-
nelsaswell. Thesesolutionsincludefrequency shift keying (FSK) [117] anddifferential
transmission.
It wasshown in [86] thatfor asymptoticallylargechannelsignal-to-noiseratiosor when
thechannelcoherencetime� � andthenumberof transmitantennasA grow withoutbound
but theratioof� ��O+A is fixed,capacity-achieving codewordsaretheproductof anisotrop-
ically distributedunitary matrix with an independentreal non-negative diagonalmatrix.
A systematicmethodto designunitary constellationswasproposedin [88], but decoding
complexity of suchconstellationsremainedexponential.Designof differentialspace-time
29
codesbasedon the orthogonaldesignswasaddressedin [191, 96], but thesecodesexist
only for certainratesandcertainnumbersof transmitantennas.In orderto designnon-
coherentspace-timecodeswith somestructure(so thatdecodingcomplexity is reduced),
space-timecodeswhich form a groupundermatrix multiplicationwerestudiedanddevel-
opedin [85, 89,90,170, 171,170, 172,99, 81]. As groupcodesdo not exist for arbitrary
rates,otherstrategiessuchascodeconcatenation[12, 15, 155, 188] needto be devised.
An interestingcodedesignmethodproposedin [79] (andits extensionto non-squarecode-
words in [100]), consideredthe designof unitary space-timecodesbasedon the Cayley
transform. The key advantageof the Cayley transformis that it makes it possibleto do
codedesignin alinearvectorspaceovertherealsinsteadof thenon-linearandnon-concave
spaceof complex unitarymatrices,significantlyreducingencodinganddecodingcomplex-
ity.
Many algorithmshave beenproposedto reducethe complexity of encoding/decoding
in multi-antennacommunicationsystems.We categorizethesealgorithmsin threegroups:
algorithmswhichareaimedto reducethedecodingcomplexity throughsub-optimalmeth-
ods(suchas[37, 78]), thosethatusefewer antennasto reducecomputations,andcombi-
nationsof spatialmultiplexing andspace-timecoding. The algorithmin [37] appliedthe
spheredecodingalgorithmof [202] to multi-antennaapplications.The key advantageof
thisalgorithmis thatit doesnotdependon thetransmissionrate.Althoughit hasexponen-
tial complexity [97], it hasa smallerexponentascomparedwith ML decoding.Antenna
selectionfor space-timeorthogonalblock codeswasaddressedin [193] (andextendedin
[104]), where,amongthereceiveantennas,theonewith thehighestchannelsignal-to-noise
ratio was chosen. An optimal algorithm for receive antennaselectionfrom a capacity-
30
maximizingperspectivewasgivenin [69], where,at eachiteration,theantennawhich had
the leasteffect on channelcapacitywasremoveduntil thedesirednumberof antennasre-
mained.In [67], asub-optimalalgorithmwasgivenwhichhadareducedcomplexity by an
orderof magnitudeascomparedwith thealgorithmin [69] by addingtheantennawith the
largestincreasein channelcapacityat eachstep.Thethird approachis to combinespatial
multiplexing andspace-timecoding. In suchmethods,the transmitantennasaredivided
into a numberof groupsamongwhich theincomingbit-streamis de-multiplexed. A form
of space-timecodingis thenappliedto eachgroup; for example,in [38, 45] thediagonal
space-timecodesare usedwith this method,while in [195] space-timetrellis codes,in
[105] space-timeorthogonalblockcodes,andin [203] V-BLAST wereapplied.
An importantissueis theestimationof thechannelfadingcoefficientsat the receiver,
whichis madefor coherentdetection.Channelestimationmethodscanbefound,for exam-
ple, in [133,32, 153, 24]. As thetransmissionbandwidthincreasesbeyondthecoherence
bandwidthof the channel,equalizationbecomesindispensable.An overview of practical
equalizationtechniquesfor space-timecodedsystemscanbefoundin [11]. Whenthechan-
nel is frequency selective, it maybedivided into a numberof sub-channelswhich canbe
assumedfrequency flat andmultiple tonescanbeused.This methodis known asorthog-
onal frequency division multiplexing (OFDM). OFDM for MIMO channelswasstudied,
for example,in [17, 119,139]. Note thatasshown in [219], frequency selectivity canbe
exploited to obtaindiversity gains. If, at least,the channelcovariancematrix canbe fed
backto thetransmitter, optimalpowerallocationvia water-filling [31] andalsobeamform-
ing canbe employed to increasethe throughput.The main result in beamformingis that
theoptimalbeamformershouldhavemultiplebeamspointingalongtheeigenvectorsof the
31
channelcorrelationmatrixwith properpower loadingacrossbeams[218]. Otherissuesin-
cludemultiuserMIMO communication[130,141, 204], broadcastMIMO channels[115],
andMIMO channelmodeling[2, 28, 65,102,122, 137].
32
Chapter 2
Bounds on the Probability of a Finite
Union of Events
2.1 The Lower and Upper Bounds
2.1.1 Introduction
Let ,;N �� N£� ������ N Ù�- beafinite setof eventsandlet -j��N£*� betheprobabilitythattheout-
comeof the experimentis N£* . In this section,we review a numberof upperand lower
boundson -r�(.ì*1NÇ*e , which is the probability of the union of N£* . The upperand lower
boundsthatwe considerin this chapterrequiretheknowledgeof theprobabilityof thein-
dividualevents-r�eN£*� & �76 �������Á , whichwehereinreferto asthefirst orderprobabilities,
andthesecondorderprobabilities-j��N£**/ N 0 & áà ��6 �������Á (which is theprobabilityof
theintersectionof apairof eventsN£* and N 0 ). Theseboundsholdfor any probabilityspace.
33
Weareinterestedin suchboundsbecause,asis demonstratedin thefollowing chapters,the
errorratesof communicationsystemscanbeexpressedin termsof sumsof probabilitiesof
theunionof errorevents.Therefore,establishingupperandlowerboundsontheprobability
of afinite unionhelpsto establishboundson theerrorrates.
We mentiontwo typesof bounds. The first two boundsareBonferroni-typebounds
[62] which areexpressedin termsof sumsof thefirst andsecondorderprobabilities.The
secondgroupof boundsarein termsof ratios.
2.1.2 The Hunter Upper Bound
A Bonferroni-typeupperbound,dueto Hunter, for theprobabilityof theunionof a finite
numberof eventsis givenby [91]-10 Ù2* Ü � NÇ*43 � ÙÚ * Ü � -j��N£*� T �65 �í87"9;: Ú� * ê 0 � 9°í87 -r�eN£*�/ãNø0° (2.1)
where < is the setof all spanningtreesof theÁ
indices,i.e., the treesthat includeall
indicesasnodes.
Thecomputationalcomplexity of findingtheoptimalspanningtreeviaexhaustivesearch
is exponentialmakingit infeasiblefor mostof theapplications.In [91,110], it is notedthat
theproblemof finding theoptimalspanningtreecanberecastinto determiningthemaxi-
mal spanningtreeof a completelyconnectedweightedgraphwhosenodesarethe indices
of theeventsandtheweightof theedgebetweena pair of nodes&
andà
is theprobability
of the joint event NÇ*=/lNø0 . This ideamakesit possibleto useKruskal’s greedyalgorithm
[3] whichfindstheoptimalspanningtreefor aweightedgraph.Thisalgorithmis described
in Subsection2.2.1.
34
2.1.3 The Kounias Lower Bound
For the set of events N �� ������ N ٠, the Kouniaslower bound,which is a Bonferroni-type
bound,is givenby [108]
-10 Ù2* Ü � NÇ*43m2 �65 �> ?@A @B Ú * 9 > -r�eNÇ*� T ÚC�D EGFIHC�J)E -r�eNÇ*�/ Nø0� LK @M@N (2.2)
whereO is a subsetof thesetof indicesPRQ��,�6 C �������Á - . Thecomputationalcomplexity
of directsearchto performtheabovemaximizationis exponential.UnliketheHunterupper
bound,thereis no algorithmicsolution(or otherwise)to find theoptimal index setin this
problem. Nevertheless,a sub-optimalalgorithmwith complexity Su� Á � is proposedin
[110]. Thisalgorithmis describedin Subsection2.2.2.
2.1.4 The de Caen Lower bound
In [39], a lowerboundontheprobabilityof aunionof eventsis derived.This lowerbound,
which is dueto deCaen,is givenby- 0 Ù2* Ü � N£* 3 2 ÙÚ * Ü � -r�eNÇ*� �ü Ù0Ü � -r�eNÇ*�/ Nø0� À (2.3)
Noticethattheaboveboundis differentfrom thetwo previousonesin thatit is in termsof
a ratio. In the following, we review two otherlower boundswhich arealsoexpressedin
termsof asumof ratiosandimproveoverdeCaen’s lowerbound.
35
2.1.5 The KAT Lower Bound
A lower boundon the probability of the union of a finite numberof eventsis derived in
[109]. Thisbound,to whichwereferastheKAT lowerbound,is givenby- 0 Ù2* Ü � N£* 3 2 ÙÚ * Ü � 0 # *�-r�eNÇ*� �ü Ù0Ü � -r�eNÇ*�/½Nø0° Í ��6 T # *e û-r�eNÇ*� Í ��6 T # *e û-r�eNÇ*� �ü Ù0
Ü � -j��N£*�/½N 0 T # *1-j��N£*� 3 (2.4)
where # *!�UT *V * TXW T *V *�Y with Z4[]\ beingthelargestintegersmallerthan [ ,V *!�L-j��N£*1 and T *_� Ú0)^ 0ÝÜ * -j��N£*�/ãNø0° ÀNotethattheaboveboundreducesto deCaen’s lowerboundif weset # *!�DR for all
&. The
KAT boundis tighterthandeCaen’s lowerboundby a factorof at most9/8 [40].
2.1.6 The Cohen and Merhav Lower bound
Anotherlower boundfor theprobabilityof a finite unionis derivedin [30]. Similar to the
KAT bound,this lowerboundis animprovementoverthedeCaenlowerboundalthoughit
requiresmoreinformationthanjust theknowledgeof -j��N£*� and -r�eNÇ*_/ÒN 0 . For a finite
eventset,CohenandMerhav’s lowerboundis givenby-`0 Ù2* Ü � N£*43m2 ÙÚ * Ü � 9 üba 9dc C Ã.��[ �»u*���[ > �ü Ù0Ü � ü a 9;c Cfe c E à ��[ �» �* ��[ (2.5)
36
where»u*���[ ø24R is anarbitraryrealfunctionwhichcanbechosensothatthelowerbound
is tightened.It is notedin [30] thatequalityin (2.5) is achievedwhen »u*���[ ø��6;Ohg � �Ä��[ ,whereg � � ��[ is thenumberof eventsto which [ belongs;in otherwords,g � �Ä��[ iQ�Iâj, &lk P�mn[ k N£* - â �Finally, notethat for »u*á��[ w�c6 , CohenandMerhav’s lower boundreducesto de Caen’s
lower bound.Therefore,by including »u*á��[ ø�y6 in thepossiblechoicesfor »u*���[ , oneis
guaranteedto geta lowerboundwhich is at leastastight asdeCaen’s lowerbound.
2.2 Algorithmic Implementation of the Hunter and Kou-
nias Bounds
2.2.1 The Hunter Upper Bound
A greedyalgorithm, known as Kruskal’s algorithm, is given in [3] to find the optimal
spanningtree that maximizesthe sumon the right handsideof (2.1). The algorithmis
describedasfollows. First,all nodes&and
àareconnectedvia edgesof weight -r�eN£*�/�Nø0°
to form a fully connectedgraph. To form the tree� Ð , the algorithmstartsfrom the edge
with the largestweight � the � & (à with the largest -r�eNÇ*o/ýN 0 û . Then,at eachstep,the
edgewith thelargestweightis addedto� Ð , subjectto theconstraintthatnocycle is formed
in� Ð . This stepis executeduntil all of the nodesare in
� Ð . This algorithmis formally
statedasfollows [110]. Considera fully connectedgraphwithÁ
edges� & áà of weights-r�eNÇ*�/½Nø0° . Constructasetof edges� Ð accordingto thefollowing steps:
37
p p pp
1
2
3
4
q q q q q q q q qq q q q q q q q qq q q q q q q q q rrrrrrrrq q q q q q q q0.2 0.1
0.04
0.02
0.05
0.3
Figure2.1: Formingtheoptimalspanningtreeto evaluateaBonferroni-typeupperbound.
1. Set� ��s .
2. Add theedgewith maximumweightto� Ð .
3. Fromtheremainingedges,addto� Ð theedgewith maximumweightsubjectto the
constraintthat� Ð remainscycle-free.
4. Gobackto step3 until� Ð contains
Á T 6 edges(i.e.,all nodesandacompletetree).
Figure 2.1 shows the result of applying the above algorithm to a set of eventswith-r�eN � / N��� �LR � C , -r�eN � / Nut� �LR � M , -r�eN � / N ' �LR � R�C , -j��N��8/�Nvt� ��LR � 6 , -j��N��8/�N ' ��R � R+K , and -j��Nutw/lN ' � R � Ryx . Thefirst edgeto addto the treeis theonebetween1 and
3, becauseit is hasthe largestweight. The last edgeis the onebetween3 and4. Note
thatalthough-r�eN��o/pNvt� is largerthan -r�eNutz/uN ' , theedgecorrespondingto theformer
probabilityis not addedto thetree,becauseit resultsin a loop.
2.2.2 The Kounias Lower Bound
A stepwisealgorithmis proposedin [110] which findsanindex setto maximizetheright
handside of (2.2). The algorithm in [110] startsfrom two sets O � � s and Oç�ý� P ,
38
andoptimizesO � and Oç� in an iterative (stepwise)mannerby addingan index to O � and
removing anindex from Oç� andvice versa,until they converge. Thealgorithmis detailed
below [110]. In thefollowing, {j��Oø denotesthetermto bemaximizedin (2.2) for asubsetOR|}P . Also, ~���� denotesthesetwhichcontainstheelementsin ~ but not in � .
1. Begin from two setsO � �bs andOç� ��P . Set {u��O � �DR andcompute{u��Oç�� .2. a. If possible,augmentO � by addingto it the bestindex
&�k P��8O � ; that is, the
index for which {j��O � .�, & - is maximizedand {j��O � .�, & - ���{r��O � .b. If possible,shrink Oç� by removing from it the bestindex
&�k O!� ; that is, the
index for which {j��O!�� is maximizedand {u��O���, & - i�R{u��O� .3. Repeatstep2 until thetermontheright-handsideof (2.2)cannolongerbeimproved.
4. a. If possible,shrink O � by removing from it the bestindex&�k O � ; that is, the
index for which {j��O � is maximizedand {u��O � ��, & - i�R{u��O � .b. If possible,augmentO!� by addingto it the bestindex
&�k P��8Oç� ; that is, the
index for which {j��O!�h.�, & - is maximizedand {j��Oç�h.�, & - ���{r��Oç�� .5. Repeatstep4 until thetermontheright-handsideof (2.2)cannolongerbeimproved.
6. Go backto step2 until themetric { canno longerbeimproved. In sucha case,the
outputof thestepwisealgorithmis�65 � ,�{j��O � {j��O� - �Notethatthisalgorithmis sub-optimalanddependsonthewayit is initialized(in theabove
algorithm,the initial setsarethe emptysetandthe global set P ). Onecanusedifferent
39
initial sets(otherthan s and P ) or usemorethantwo initial setsto find a better(yet still
sub-optimal)setof indicesfor O .
2.3 Final Remarks
As is shown in the following chapters,theblock error rate,thesymbolerror rate,andthe
bit errorrateof single-andmulti-antennacommunicationsystemswith arbitrarysignaling
schemesandsymbolmappingscanbeexpressedin termsof sumsof probabilitiesof unions
of errorevents.Therefore,theaboveboundsarepowerful toolsto establishupperandlower
boundsfor theerrorratesof thesystemsunderstudy. Whenspecializedto theanalysisof
the communicationsystems,the overall error rateof the systemcanbe representedby a
finite unionof events,where-r�eNÇ*� is thepairwiseerrorprobability, and -r�eNÇ*y/uNø0� is the
secondorderpairwiseerrorprobability.
40
Chapter 3
A Brief Review of Space-Time
Orthogonal Block Codes
3.1 Introduction
Space-timeorthogonalblock codesarea sub-classof space-timecodeswhosecodewords
haveorthogonalcolumnsandhencethey arefull-rank. Thesecodesareimportantbecause
of their simpleencodingand,specially, becausetheir decodingcomplexity is polynomial
(asopposedto exponential)in the sizeof the signalingscheme.Space-timeorthogonal
block codeswereintroducedby Alamouti in [9] for a systemwith two transmitantennas
andany numberof receive antennas.A theoryto extendthesecodesfor morethantwo
transmitantennaswaspresentedin [192].
In thefollowing,wedescribeencodinganddecodingfor Alamouti’scode,whichis also
known asthe � � code,in detail for a dual-transmitandsingle-receiveantennasystem.ML
41
input bits¡ � § ¡ � { constellationmapper� ���� �� � � ¬ � ¬ ��� { space-timeblock coder� ¬ � ¬ � �o� ��� ¬ ��� ¬ '�¬ � ¬�' ������ ��
Figure3.1: Thetransmitterof Alamouti’sspace-timeorthogonalblock codingscheme.
decodingandthesymbolpairwiseerrorprobability(PEP)of generalspace-timeorthogonal
block codesarenext explained.As a contribution, we presentanalternative derivationof
thecodeword PEPof multi-antennacodeswith arbitrarystructure.This PEPexpressionis
valid for bothspace-timecodesandBLAST.
3.1.1 Encoding in Alamouti’s System
Let us considera dual-transmitsingle-receive antennasystem. The encodingstrategy
of Alamouti’s schemeis as follows (seeFigure 3.1). First, every incoming à bits �4��5��� �������� ¯ � (whereC ¯ is thesizeof thesignalingscheme)aremappedinto abasebandsignal
usingascalarfunction.For example,if � � and �+� aretwo consecutiveblocksof à bitseach,
they aremappedto¸ � � ���á� � and
¸ �Ç� ���á�+�� , respectively, where ��� À is themodulation
functionand¸ �
and¸ � arebasebandsignals.At a givensymbolperiod,two signalsaresi-
multaneouslytransmittedfrom thetwo transmitantennas.At symbolperiodone,we have¾��áê¦� � ¸ �and
¾ � ê¦� � ¸ � . During the next symbolperiod,¾��áê ��� T ¸ '� is transmittedfrom
antennaoneand¾ � ê �Ç� ¸ ' �
is transmittedfrom antennatwo, where'
denotescomplex con-
jugate. Notice thatno explicit channelcodingis performedandbecausetwo consecutive
42
symbolsaresentin two symbolintervals,therateof thischannelcodeis unity. Weassume
thatthefadingconditionsremainthesameduringthetwo consecutivesymbolperiodsused
to transmitsymbols �
and¸ � . Denotingthechannelfadingcoefficientsby
� �and
� � , at
symbolperiodone,we receive � � � � ��¸S� Í � � ¸ � ÍÏÎ �andatsymbolperiodtwo, weget�£�?� T � �û¸ '� Í � � ¸ ' � ÍÏÎ � �3.1.2 Decoding in Alamouti’s System
Defining ¡" Q� ¢£¤ � � � �� '� T � '�d¥;¦§ wenotethat
¡" is anorthogonalmatrix;namely,
¡" � ¡" �W��â � � â � Í â � �+â � I¨�� .For
¡d©Q�v��� � � '� í , )=��� ¸ �B¸ �� í , and
¡� �I� Î � Î '� í , wecanwrite¡du� ¡"U) Í ¡� � (3.1)
At this stage,ML decodingmay be performeddirectly using (3.1); i.e., choosing $)D�� $¸S� $¸ �� í which satisfies
$)=� argminª æ ¡d T ¡"U)çæ � (3.2)
where $[ denotesthe estimateof [ and minimization is doneover all two-tupleswhose
elementsarebasebandsignals.However, decodingmaybefurthersimplifiedby exploiting
43
theorthogonalityof thepathgainsmatrix
¡" . Multiplying (3.1) from the left by
¡" � , we
have gd Q� ¡" � ¡dã� ¡" � ¡"U) Í ¡" � ¡� �W��â � � â � Í â � �+â � ®) Í ¡" � ¡�� ��â � � â � Í â � ��â � û) Í g� � (3.3)
In otherwords, g� � ����â � � â � Í â � �+â � ¸ � Í gÎ � and g���?�I��â � � â � Í â � �+â � ¸ � Í gÎ � whereg� � Q� � '� � � Í � ����� (3.4)g�£��Q� � '� � � T � � �£� (3.5)gÎ � Q� � '� Î � Í � � Î � , and gÎ ��Q� � '� Î � T � � Î � . As theentriesof g� arei.i.d. and� �
and� �
areknown to thereceiver, ML decodingof theentriesof ) canbedoneindependentlyvia$¸ � � argmin� âSg� � T ��â � � â � Í â � �+â � ¸ â � (3.6)$¸ � � argmin� â g�£� T ��â � � â � Í â � �+â � ¸ â � (3.7)
whereminimizationis doneover thesignalalphabet.Sincethemodulationfunction ��� À is
scalar, theestimatedbits wouldsimplybe $� � �«� < � � $¸S� and $�+�ø�b� < � � $¸ �� .For asignalingschemeof size
Á �:C°¯ , while usingtheML detectionruleontheoutput
vector
¡d asoutlinedin (3.2) will requireÁ � ML costfunctioncomputations,thenumber
of thecomputationswill only be C Á if gd , definedin (3.4) and(3.5), is usedfor detection
asoutlinedin (3.6)and(3.7).Thiswill resultin muchshorterdecodingtimes,speciallyfor
largesignalalphabets.In a typical implementationof thesystemwith 16-QAM signaling,
decodingof 5 million bits with detectionrule (3.2) took 62918seconds,while employing
thedetectionrulesin (3.6)and(3.7) took about100seconds.Thetestswererun on a Sun
Ultra 60 computer.
44
3.1.3 Extension to Three and Four Transmit Antennas
Alamouti’s work wasgeneralizedto A transmitantennas,insteadof only 2 antennas,in
[192]. There,it wasstatedthat it is not possibleto constructspace-timeorthogonalblock
codesof rateonewith complex signalingschemeswhentherearemorethantwo transmit
antennas.Also, two examplesof space-timeorthogonalblock codeswereintroducedfor
threeandfour transmitantennasas
� t � ¢£££££¤¸ � T ¸ � T ¸ t T ¸�' ¸ ' � T ¸ '� T ¸ 't T ¸ ''¸ � ¸ � ¸�' T ¸ t ¸ '� ¸ ' � ¸ '' T ¸ 't¸ t T ¸)' ¸ � ¸ � ¸ 't T ¸ '' ¸ ' � ¸ '� ¥ ¦¦¦¦¦§ (3.8)
and
� ' � ¢££££££££¤¸S� T ¸ � T ¸ t T ¸)'¿¸ '� T ¸ '� T ¸ 't T ¸ ''¸ � ¸ � ¸)' T ¸ t ¸ '� ¸ ' � ¸ '' T ¸ 't¸ t T ¸�' ¸S� ¸ � ¸ 't T ¸ '' ¸ ' � ¸ '�¸)' ¸ t T ¸ � ¸ � ¸ '' ¸ 't T ¸ '� ¸ ' �
¥ ¦¦¦¦¦¦¦¦§ �(3.9)
Notethatbothof theabove codewordstransmitfour symbolsin eightsymbolperiodsand
hencethey havea rateof 6;OPC .For encoding,four blocksof à bits eacharemappedinto a block of four signals )l�� ¸S���¸ � �¸ t �¸)' (í . A codeword is thenformedaccordingto (3.8)or (3.9)andtransmittedover
thechannel,with onecolumnof thematrix transmittedat a time.
Thedecodingphaseis similar to thedualinputsystem.It is assumedthatthechannelis
slow fading(quasi-static),with thepathgainsremainingconstantduring thetransmission
of a codeword (eight symbol intervals). Therefore,we canusethe sameapproachas in
45
Subsection3.1.2. In orderto find theML detectionrules,we definetheoutputandnoise
vectorsas¡dã�W�e� �� ��� �ut � ' � '¬ � ' � '® � '¯ í and
¡� �W� Î �� Î � Î t Î '; Î '¬ Î ' Î '® Î '¯ í respectively, so that the input-outputrelationshipbecomesasin (3.1),with thepathgains
matrix
¡" givenby
¡" �¢£££££££££££££££££££££££¤
� � � � � t R� � T � � R T � t� t R T � � � �R � t T � � T � �� '� � '� � 't R� '� T � '� R T � 't� 't R T � '� � '�R � 't T � '� T � '�
¥ ¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦§for thethreetransmitantennasandby
¡" �¢£££££££££££££££££££££££¤
� � � � � t � '� � T � � � ' T � t� t T � ' T � � � �� ' � t T � � T � �� '� � '� � 't � ''� '� T � '� � '' T � 't� 't T � '' T � '� � '�� '' � 't T � '� T � '�
¥ ¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦§46
for systemswith four transmitantennas.Notethat theabove matricesareorthogonal,and
henceaftermultiplying (3.1) from theleft by
¡" � , wehave
gdp�:C�°=) Í g� (3.10)
where °J� ü ë* Ü � â � *�â � . It canbe shown that the elementsof g� arei.i.d. andhenceML
detectioncanbeperformedseparatelyoneachentryof gd .
3.2 General Treatment of Space-Time Orthogonal Block
Codes
Let usnow considerasystemwith A transmitand E receiveantennas.Let )=��� ¸S�� �������¸)± áíbe a vectorof Y consecutive symbolsto encodeand �c� ��� �������� ���_ be the space-time
codewordcorrespondingto it. In thecaseof STOB codes,wehavea �4b�Y , whereb�� a O�Y
is thecodinggainand �3� � �Ib æ°) æ � I ë . As anexample,for thecode � t in (3.8),a �³² ,Yl�yK , and b �yC , andfor Alamouti’s code, b � 6 and
a �IY��yC . It canbeshown that
(1.1)canbere-writtenas[140]¡d 0 � ÿ �\�A¡" 0 ) Í ¡� 0 à �76 ÀSÀSÀ! E (3.11)
where
¡d 0 � � ¡� 0 � ������ ¡� 0 � � í ,
¡� 0 � � ¡Î 0� ������ ¡Î 0� � í , andwe have
¡� 0 Å �]� 0Å and
¡Î 0Å � Î 0Å for6 � Ô � � � , and
¡� 0 Å ��� 0 'Å and
¡Î 0Å � Î 0 'Å for � � � Ô � aand
¡" 0 is derivedfrom theà
th
row of " via negationandcomplex conjugationof someof its entries.It is clearthat
¡Î 0Åarei.i.d. Ñix���R 6; .
47
We will hereafterconsiderthe set2 of orthogonalcodesfor which the corresponding
matrix
¡" 0 hasorthogonalcolumns,i.e.,
¡" 0 � ¡" 0 �Jb*°ö0�¨ ± , where °\0 � ü * â � 0�*�â � áà �6 ������ E . Therefore,when(3.11)is multiplied from theleft by
¡" 0 � we obtain
gd 0 Q� ¡" 0 � ¡d 0 �4b ÿ �\�A °ö0)´ Í g� 0 (3.12)
where g� 0 Q� ¡" 0 � ¡� 0 � � gÎ 0� ������ gÎ 0± � í � Notethateachentryof gd 0 � � g� 0 � ������ g� 0 ± � í is associ-
atedwith only onesymbol. Since,given " , g# � � gd � ������ gdÄh � is an invertible functionof# � � d �� ������ di� � , ML decodingcanbebasedon g# insteadof#
in thefollowing way¸ *_� argmax� -j� ¸ âµ,Ëg�·¶* - h¶Ü � "7
whichcanbeshown to beequivalentto¸ * � argmin� hÚ ¶ Ü � âSg� ¶* T bL¸¹° ¶ ¸ â �° ¶ (3.13)
whereb ¸ �4b Ì ����O+A .
3.3 The Symbol PEP for General Space-Time Orthogonal
Block Codes
Basedon(3.13),it canbeverifiedthatthePEPbetweenapairof symbolstransmittedover
thespace-timeorthogonalblockcodedchannelconditionedon thepathgainsis givenby-r� ¸ * @ ¸ 0Pâ5"U �bº¼»�½�*50 Q °¿¾ (3.14)
2This classof orthogonalcodesincludesmostof suchcodespublishedin theliterature,includingAlam-
outi’scodeandthosein (3.8)and(3.9).
48
where¸ * @ ¸ 0 denotestheevent that
¸ 0 hasa smallerML metric than¸ * (in (3.13))when¸ * is sent(
¸ * and¸ 0 areapair of symbolsinput to theSTOB encoder),ºr� À is theGaussian
tail functiongivenby ºj��[ � 6Q C � �ÁÀa  < Åfà � & Ô�½�*50?� Ì Ä ù ú� ë â ¸ * T ¸ 0�â , and °7� hÚ ¶ Ü � ° ¶ �ëÚ Å Ü � hÚ ¶ Ü � â � ¶ ê Å â � � (3.15)
To find the PEP, oneshouldaverage(3.14) with respectto the distribution of ° , which
canbederivedasfollows. Let usdefinerandomvariable Æ�* as Æ � 0 < ���ëlÇ * �ÉÈÊ, � 0 ê * - andÆ ë h Ç � 0 < ���ëlÇ * �ÌËÍ, � 0�* - for& �76 ������ A and
à �W6 ������ E , where È and Ë aretherealand
imaginaryparts,respectively. Wenotethat Æ Å�Î i.i.d. x���R �� , andwecanwrite ° as°I� Ú * ê 0 â � 0 ê *�â � � Ú * ê 0 ÈÊ, � 0 ê * - � Í Ë·, � 0 ê * - � � Ú * Ü � Æ �* where¹Ò�LAlE . Usingthemomentgeneratingfunctionof normalrandomvariablesyields
theprobabilitydensityfunctionof ° as�ÐÏ ��Ñi � 6�e¹ T 6; )Ò Ñ < �  <LÓ Ñ��ºR (3.16)
Hence,° hasa scaledchi-squareddistribution with C�¹ degreesof freedom.Theexpected
valueof (3.14)is thereforeequalto-r� ¸ * @ ¸ 0� �� 6�e¹ T 6; )Ò � ÀÐ Ñ < �  <LÓ º��Ô½�*50 Q Ñ\ & Ñ � (3.17)
As noticedin [19], thePEPof diversityreceptionsystemswith maximumratio combining
hasa form similar to thatof (3.17)with theexceptionthat ½�*50 is givenby Ì �\�ûOPCÉâ ¸ * T ¸ 0Pâ .49
Therefore,theconditionalPEPfor thosesystemsis thesameas(3.14).MRC systemswere
previously analyzed,for example,in [150, 175], whoseresultscanbe usedto obtainthe
expectedvalueof (3.14)as-j� ¸ * @ ¸ 0° � 6C ¢¤ 6 T ½�*50Õ ½ �*50 Í C ë h < �Ú Å Ü Ð � C+ZZ÷� 6��Ci�½ �*50 Í C+ ® Å ¥§ �(3.18)
3.4 The Codeword PEP of Arbitrary Space-Time Codes
Here we presenta new simple derivation of the codeword PEPfor arbitraryspace-time
codes.Weshow thattheconditionalPEPof any two space-timecodewordsis in theform of
adoublesumof anexpressionsimilar to (3.14),andhenceits expectationcanbeevaluated
with thesameapproachasthatfor theSTOB codes.
For arbitraryspace-timecodingschemesthe receiver computesthe squareddistance
betweenany pair of fadedtransmittedcodedsequences� and $�Ö � � ê�×� � �\�A �Ú Å Ü � hÚ 0 Ü ��ØØØØØëÚ * Ü � � 0 ê * & * ê Å ØØØØØ
� � �\�A hÚ 0 Ü �oÙ �0)Ú í Ù 0 (3.19)
where &nÛ ê ÅzÜ ¾ Û ê Å+ÝßÞ¾ Û ê Å , à ÜÉá &yÛ ê â4ã , Ú Ü àäà � , and Ùoå is thetransposeof the æ th row of ç .
Let è6éëêbì Þêví denotetheprobabilitythatÞê hasa largermetricthan ê when ê is sent.It
is easyto verify that[190]è�éîêRì Þêví Ü�ïñðzòó=ôÔõóäö�÷1øñù úûýüÿþó�� ×ó �����(3.20)
Let � þ� Ü � ���þ�� , where � � is the � th non-zeroeigenvalue of U with multiplicity � � . To
computethe expectationof (3.20),we determinethe pdf of �þ ü þ ó�� ×ó andwrite (3.20)asa
weightedsumof expressionssimilar to (3.14).
50
SinceU is Hermitian(i.e.,U � = U) andnon-negativedefinite,it canbedecomposedas� Ü�� �îà ���whereV is unitary(i.e.,V � V = I � ) andD is anon-negativedefinitediagonal
matrixhaving theeigenvaluesof U on its maindiagonal.V is a unitarymatrix (i.e.,V � V =
I � ) andits columnsaretheunit-normeigenvectorsof U. Substitutingthedecomposed�
in (3.19),wehaveúû ü þ ó�� ×ó Ü ���û � !" å Ü �$# �å � � à � # å Ü ���û%� !" å Ü �'& �å à & å Ü !" å Ü � �" Û Ü � ���û%� � Û)( * å � Û�( þ � (3.21)
where & å Ü+� # å , * å � Û is the , th elementof & å , and � Û Ü à Û � Û . We notethat theentriesof
& å arei.i.d. -/.Éé10 � ú í . Therefore,themomentgeneratingfunctionof �þ ü þ ó2� ×ó canbewritten
as 354ò ðoòó_ôÔõó é Ý76 í Ü �8� Ü � úé ú:9 � þ� 6 í !<; �>= (3.22)
In order to find the pdf of �þ ü þ ó2� ×ó , which equals ?A@ �CB 354ò ðoòó]ôÔõó é ÝD6 í<E (where ? is the
Laplacetransform),weconvert (3.22)into asumandusethelinearityof theLaplacetrans-
form. Letting F � Ü þ���G��� andusingpartial fraction expansion[112], we canwrite (3.22)
as 3 4ò ðoòó]ô õó é ÝD6 í Ü �" � Ü � ! ; � @ �" Û Ü2H I ÛKJ � � �é 6 9 � @ þ� í ÛKJ � � (3.23)
whereI ! ; � @ Û � � Ü ú,)L BNM ÛM 6 ÛPO é 6 9 F � í ! ; �354ò ðoòó*ô õó é Ý76 íRQ E ØØØØ � ÜTS � � , Ü 0 � �U�V� �XW � � Ý ú � (3.24)
TakingtheinverseLaplacetransformof (3.23),wehaveY 4ò ð ò ó]ô õó é[Z$í Ü �" � Ü � !<; � @ �" Û Ü2H I ÛKJ � � �,�L Z Û[\ @ ò^]_ ` �ba � Z�cd0 � (3.25)
51
Thenext stepis simply using(3.25)to evaluate(3.20).Thisyieldsè6éëêRì Þêuí Ü �" � Ü � ! ; �" Û Ü � I Û � �é1, Ý ú í�LfehgH Z Û @ � \ @ ò^]_ ` � a ÷ji�k Zml M Z � (3.26)
The integral in (3.26) is thesameastheonein (3.17)which wassolvedin (3.18). Using
(3.17)and(3.18)in (3.26)andsetting� þ� Ü � �þ�� , weobtaintheexpressionfor thecodeword
pairwiseerrorprobabilityof arbitraryspace-timecodesasè6éëê�ì Þêví Ü �" � Ü � ! ; �" Û Ü �An Û � �û ø ú Ý � �o � þ� 9 û Û @ �" å Ü2Hqp û ææsr úé û � þ� 9ut í å � �(3.27)
wheren Û � � Ü � þ Û� I Û � � . Theaboveresultagreeswith theresultin [131].
52
Chapter 4
Bounds on the Block and Bit Error
Rates of Coded AWGN and Block
Rayleigh Fading Channels
4.1 Introduction
Let usconsiderthecommunicationsystemof Figure1.4. This systemcanuseany typeof
channelcodingandany complex signalingscheme.In this chapter, we assumethe input
to thesystemto beuniform i.i.d. andhencewe considerML decodingat thereceiver. Our
goalis to establishupperandlowerboundsontheblockerrorrate(BlER) andbit errorrate
(BER) of thesystem.To this end,we first write theBlER andBER in termsof a sumof
theprobabilitiesof a unionof errorevents.We will thenusetheKouniasandKAT lower
boundsaswell asHunterupperbound,whosegeneralformsweredescribedin Chapter2,
53
to derive boundson the block and bit error ratesof SISO AWGN and block Rayleigh
fadingchannelsunderML decoding. Theoretically, theseapproachescanbe appliedto
any signalingschemesandchannelcodesof any typeor rate. For large v , where vxw � is
thechannelcoderate,computationof mostof theboundsin theliterature,includingthose
consideredhere,becomesinfeasible. Oneof the contributionsof this chapteris to usea
subsetof the codebookto evaluatethe bounds. In fact, it is shown that the performance
of thelower boundscanbeimprovedin that theboundsbecometightenedby considering
a subsetof the codebook.Anothercontribution of this chapteris deriving a closed-form
formula for the probability of the intersectionof two pairwiseerror eventsfor the block
Rayleighfadingchannel.
4.2 The Error Rates in Terms of Probabilities of a Union
As explainedis Subsection1.4.1,weconsideracommunicationsystemwhichmapsblocks
of inputbits yTz Ûf{ �}|�Ü 0 � �U�V� ��~ Ý ú into blocksof basebandsignalsy ó Ûf{ �}|�Ü 0 � �V�U� �X~ Ý úwhicharetransmittedover thechannel.Theblockerrorrateof thesystemis givenby
BlERÜ�� @ �" Û°Ü2H è�é�� ( ó Û í^F é ó Û í Ü�� @ �" Û°Ü2H F é ó Û í�è Û ø��Û[�Ü�Û � Û Û � �
(4.1)
whereF é ó Û í Ü F éRz Û í is theprobabilitythat z Û is emittedfrom thesource,è Û é = í��Ü è�é = ( ó Û íis the conditionalprobability given that
ó Ûwassent,and � Û Û indicatesthe pairwiseerror
eventbetweenó Û
andó Û ; i.e., theevent thatwhen
ó Ûis sentandbetween
ó Ûand
ó Û , ó Û is
detectedat thedecoder. Similarly, theBERcanbecomputedfrom
BERÜ � @ �" Û°Ü2H F é ó Û íIè Û é1�����[�ëí � (4.2)
54
where è Û é������[�(í Ü úv � @ �" å Ü2HA�q� é æ �)| íIè6é Þó Ü ó å ( ó Ü ó Û íÜ úv � @ �" å Ü2H �q� é æ �)| í ø ú Ý è Û ø'�Û1�Ü å � å Û ��� �(4.3)
whereÞó
is thedecodedsignalsequenceand�5� é æ �)| í is theHammingdistancebetweenthe
databits correspondingtoó å and
ó Û. Notein theaboveequationsthatbecausewe assume
theinput bit-streamto beuniform i.i.d., all symbolsareequallylikely andF é ó Û í Ü ú w ~ .
Equations(4.1)and(4.3)expressboththeBlER andBERin termsof theprobabilityof
a union,which canbe boundedusingtheKAT, Kounias,andCohenandMerhav’s lower
boundsaswell astheHunterupperboundwhicharepresentedin Chapter2.
In thecomputationof theabove bounds,we needto computethefirst andthesecond
orderPEPs(see(2.1)-(2.5)). Clearly, thePEPexpressionsdependon thechannelmodel.
In thefollowing, wewill considertwo channelmodels:theAWGN andtheblockRayleigh
fadingchannel.
4.3 The AWGN Channel
For theAWGN channel,wehave � â Ü�6 â 9u� â �where
� âis the received signalat symbol interval � and � â�� -/.Éé10 � � H í is the additive
whiteGaussiannoiseat thereceiver. Letó Û Ü¼á�6 Û � � =�=�= ��6 Û� ã bethe
| th possiblesequenceof
55
signalsin ablock,andlet M Û � Ûâ Ü�6 Ûâ Ý�6 Ûâ. TheML decodingrule in this caseis givenbyÞó Ü
argmin� ��� Ý ó � þ �where� ܼá � � � �U�V� �
� � ã.
Thepairwiseerrorprobabilitybetweenó Û
andó Û is givenbyè Û é�� Û Û í Ü è6é ó Û ì ó Û í Ü è�é ��� Ý ó Û � þ c ��� Ý ó Û � þ ØØ ó Ü ó Û íÜ è ø k û2� ó Û Ý ó Û �}� �� ó Û Ý ó Û � � H c úk û � ó Û Ý ó Û � � (4.4)Ü ÷ é ü Û � Û í � (4.5)
where��ܼá � � � �V�U� � � � ã
,� Z �)¡2� is thestandardinnerproductof Z and
¡, ü þÛ � Û Ü ���2¢ â ( M Û � Ûâ ( þ ,
and÷ é = í is theGaussiantail function. We alsohave ��� Ü+£ � w � H
for BPSKsignalingand��� Ü+£ � w û � Hfor two dimensionalsignaling.Notethat theargumentof theexpressionon
theleft-handsideof (4.4) is aunit-variancerealGaussianrandomvariableandhence(4.5)
easilyfollows.
Theprobabilityof theintersectionof two pairwiseerroreventsis equaltoè Û é1� Û Û¥¤ � å Û í Ü è6é ó Û ì ó Û � ó Û ì ó å í ܧ¦ é^¨ Û Û å � ü Û � Û � ü å � Û í � (4.6)
where ¦ é1¨ � Z �)¡ í Ü úû<© o ú Ý ¨ þ ehga ehgª \ @C« ò�¬ ò® « `�¯<` òò�° 4 ¬ òR± MT²�M � (4.7)
is thebivariateGaussianfunctionand¨ Û Û å Ü ¢ �â Ü � � M Û � Ûâ � M å � Ûâ �i ¢ �â Ü � ( M Û � Ûâ ( þ l4ò i ¢ �â Ü � ( M å � Ûâ ( þ l
4ò =56
From(4.3),we needto evaluateè Û é�� å Û í and è Û é�� å Û ¤ � å � í for theBER bounds.These
quantitiesaregivenbyè Û é1� å Û í Ü è6é ��� Ý ó å � þ c ��� Ý ó Û � þ ØØ ó Ü ó Û íÜ ÷ p ü þÛ � Û Ý ü þå � Ûü Û � å rand è Û é1� å Û ¤ � å � í Ü�¦ p ¨ å Û � � ü þÛ � Û Ý ü þå � Ûü Û � å � ü þ� � Û Ý ü þå � Ûü � � å r =4.4 The Block Rayleigh Fading Channel
For theblock Rayleighfadingchannel,wehave� â Ü�³´6 â 9u� â �where
³, which is the only parameteraddedto the channelmodel, is a zero-meanunit-
variancecomplex Gaussianrandomvariable:³ � -/.Éé10 � ú í . We assumethat
³remains
unchangedduringthetransmissionof acodewordandchangesin ani.i.d. mannerto anew
valueafterwards.
Similar to theAWGN channelcase,we first needto evaluatethefirst andsecondorder
PEPsto calculatetheBlER andBER upperandlower bounds.Thepairwiseerrorproba-
bilities conditionedon thepathgain³
aresimilar to theAWGN case.More specifically,
wehave è Û é�� Û Û í Ü�£ � y ÷ é ü Û � Û ( ³ ( í { � (4.8)
57
and è Û é�� Û Ûµ¤ � å Û í Ü�£ � y ¦ é^¨ Û Û å � ü Û � Û ( ³ ( � ü å � Û ( ³ ( í { � (4.9)
We notethatequation(4.8) is thesameas(3.14)with ¶ Ü ( ³ ( þ . Hence,it equals(3.18)
with� Ü·W Ü ú . Therefore,thePEPis givenbyè Û é1� Û Û í Ü úûh¸¹ ú Ý ü Û � Ûº û 9 ü þÛ � Û�»¼ =
As for the expectationin (4.9), we can usethe result of [174] to write the¦ é = � = � = í
functionas¦ é^¨ Û Û å � ü Û � Û ( ³ ( � ü å � Û ( ³ ( í Ü úû<© eh½m¾2¿TÀÂÁ¿�à Á� Ä ÁÅÀ Ã�ÆH Ç�È�É BËÊ ü þÌ ÛûÎÍ�ÏÑÐ þÓÒ ( ³ ( þ E M Ò9 úû<© e ½�Ô ¿ à Á¿ ÀÂÁ� Ä ÁÅÀ ÃRÕH Ç�È�É B Ê ü þå Ûû�Í�ÏÑÐ þ Ò ( ³ ( þ E M Ò � (4.10)
where Öié[Z � ¨!íØ×ÚÙXÛ Ð @ �ÝÜ Z o ú Ê ¨ þ w*é ú Ê ¨TZ$í�Þ , and ÙXÛ Ð @ � é[Z$í is re-definedhereas© ÊÙ}Û Ð @ � é Ê Z$í for negative Z . We thenuse[175, Eq. 5A.35] (with ��× ú ) to finally derive the
expressionfor thesecondorderPEPin theblock Rayleighfadingcaseasè Û é1� Ì Ûµ¤ � å Û í �×NßÊé[¨ Û Ì å � ü Ì � Û � ü å � Û íà×·á$é ü Ì Û � ü å Û � ¨ Û Ì å í 9 á$é ü å Û � ü Ì Û � ¨ Û Ì å í � (4.11)
where á$é[Z �}¡m� ¨!íÎ× ú© pãâ Êåä Zk ú:9 Z þ r �with â × Ù}Û Ð @ � ø Z o ú Ê ¨ þ¡ Ê ¨TZ � �
ä × úû B Ù}Û Ð @ � � � 9 © O ú Ê p ú:9 sgné � íû r sgné � í Q E �58
� × û Z k ú:9 Z þ Í)ÏVÐuû â , and� × é úA9 û Z=þ)í/æèç Í+û â Ê ú (with the definition of ÙXÛ Ð @ � é[Z$í
beingthesameasthatin (4.10)).
For theBERbounds,onecansimilarly obtainthefirst andsecondordererrorprobabil-
itiesasfollows è Û é1� Ì å íà× úû ¸¹ ú Ê ü þÌ � Û Ê ü þå � Ûº û ü þÌ � å 9 é ü þÌ � Û Ê ü þå � Û í þ »¼and è Û é�� Ì å ¤ � � å íà×�ß p ¨ å Ì � � ü þÌ � Û Ê ü þå � Ûü Ì � å � ü þ� � Û Ê ü þå � Ûü � � å r �whereßÊé = � = � = í is definedin (4.11)andwhenthesecondandthethird argumentsof
¦ é = � = � = íarenon-negative. If at leastoneof the two argumentsarenegative, the
¦ é = � = � = í function
canbe written assumof a numberof¦ é = � = � = í functionswith non-negative argumentsas
follows
¦ é1¨ � I � n í�×éêêêêêêêêë êêêêêêêêì
¦ é[¨ � I � 0ní 9 ¦ é Ê ¨ � I � 0yí Ê ¦ é Ê ¨ � I � Ê n í I cd0 � nîí 0 �¦ é[¨ � 0 � n í 9 ¦ é Ê ¨ � 0 � n í Ê ¦ é Ê ¨ � Ê I � n í I í 0 � n c·0 �ú Ê ¦ é1¨ � 0 � Ê n í Ê ¦ é Ê ¨ � 0 � Ê n íÊ ¦ é1¨ � Ê I � 0yí Ê ¦ é Ê ¨ � Ê I � 0yí 9 ¦ é[¨ � Ê I � Ê n í I í 0 � nîí 0 �(4.12)
Therefore,theclosed-formexpressionfor theprobabilityof theintersectionof twopairwise
erroreventscanalwaysbecomputedfrom (4.11).
59
4.5 Linear Block Codes and BPSK Signaling
4.5.1 Preliminaries
Sincethe numberof the block PEPsgrows exponentiallyin v , the computationalcom-
plexity of theaboveboundsbecomesprohibitivefor largedatablocks.In importantspecial
cases,suchaswhenlinear block codesandBPSK modulationareusedor for geometri-
cally uniform codes[55] with matchedsignaling[128], it is possibleto consideronly one
codeword insteadof thesumsin (4.1)and(4.2). In otherwords,onecanuse
BlER ×�è Û ø �Ì �Ü�Û � Û Ì � �and BER × úv � @ �" å Ü2H �q� é æ �}| í ø ú Ê è Û ø �Ì �Ü å � å Ì ��� �
(4.13)
where ï Û is an arbitrary codeword. Furthersimplification canbe madeby considering
the all-zerocodeword (| ×ð0 ) whenbinary modulationis used,becausethis will allow
to benefitfrom informationsuchasthe weight distribution of codesto simplify or speed
up analysis.We illustratethis in thefollowing via reviewing Seguin’s lower boundon the
BlER of a codedBPSKmodulatedsystemwith AWGN channelbasedon deCaen’s lower
boundon theprobabilityof aunionwhich is givenin (2.3).
deCaen’s lowerboundcanbeusedto derivealowerboundfor theblockerrorratefrom
BlER ×bè H ø$�Ì �Ü2H � H Ì � c � @ �" Ì Ü � è þ é1� H Ì í¢ � @ �å Ü � è�é�� H Ì ¤ � H å í � (4.14)
where,from [156], è6é1� H Ì íà× ÷ ø ù û £ �Rñóò� Hõô é�ï Ì í � �and è�é�� H Ì ¤ � H å íà× ¦ ø ¨ Ì å � ù û £ � ñóò� H ô é�ï Ì í � ù û £ �Gñóò� H ô éGï å í � �
60
whereô é�ï Ì í is theweight(i.e., thenumberof ones)of codeword ï å and¨ Ì å × ô éGï Ì í 9 ô é�ï å í Ê ô é�ï Ì 9 ï å íû o ô éGï Ì í ô éGï å í =
It is shown in [156] that¦ é[¨ � Z �}¡ í is strictly increasingin thecorrelationcoefficient ¨ . This
allows to computeanupperboundon ¨ Ì å andhenceon è6é1� H Ì ¤ � H å í on thedenominatorof
(4.14)which dependsonly on thecodeword weightsandis thereforevery fastto compute.
Theupperboundon ¨ Ì å is asfollows¨ Ì åDö ô éGï Ì í 9 ô éGï å í Ê ��÷ Ì ;û o ô éGï Ì í ô é�ï å í �where
�ø÷ Ì ; is theminimumweightof thecode.
Finally, wepresentCohenandMerhav’s dot-productlowerbound[30]. Assumingthatï H is sent,this boundreducesto
BlER c \ �Âùûú @ þ ù��Kü�ý ÷ � @ �" Ì Ü � ÷ þ é�þ é I � ï Ì � � H í"í¢ � @ �å Ü � ¦ é1¨ Ì å � þ é I$ÿ � ï Ì � � H í � þ é ICÿ � ï å � � H í�í � (4.15)
whereI × ú:9�� � H w û , I ÿ × ú:9�� � H, n ׳é ú Ê I þ í�w � H
, n ÿ ×¼é ú Ê é ú:9�� � H í þ í�w � H, andþ é I � ï Ì � � H í�× û I £ � ñóò ô é�ï Ì í� H =
Parameter� canbechosento tightenthebound.Thebestvalueof � is foundin [30] through
anexhaustivesearch.
4.5.2 Evaluation of the Algorithmic Bounds for Large Blocks
It followsfrom (2.4)thatevaluatingtheKAT boundonasubsetof thecodebookwill result
in a lowerboundto theoriginalKAT lowerbound.This follows fromè ø ��Ì Ü � � Ì � c�è ø �Ì���� � Ì � c KAT éÔè6é � Ì í � ,���í �61
where �jy ú � �U�V� ��~ {. Theabovecanbefurthertightenedbyè ø���Ì Ü � � Ì � c � Û È� KAT éè�é � Ì í��f,����í �
Theevaluationof thealgorithmicHunterandKouniasboundsmaystill betediousfor
large v . It would be desirableto useonly a subsetof the codebookto computethese
bounds.An examinationof thelower boundin (2.2)directly indicatesthatcomputingthis
boundon a subsetof the codebookwill result in a lower boundfor (2.2). We shall refer
to this boundasthe “Kouniaslower bound,subset”. It is exponentiallyexpensive to find
theoptimalsubsetfor theKouniaslowerbound.To find a sub-optimalsolutionwith a low
complexity, oneoptionis to employ thestep-wisealgorithmusedin Subsection2.1.3to find
a good(yet still sub-optimal)subset.A typical behavior of theapplicationof thestepwise
algorithmto theKAT boundis describedin thefollowing for acodewith asmallcodebook
sizefor clarificationpurposes.Wehaveobservedthesamebehavior for many othercodes,
suchasGolayandBCH codes.Hamming(7, 4) codehas15 non-zerocodewordsandits
generatormatrix is givenby
� × ¸��������¹ú 0T0T0 ú 0 ú0 ú 0T0 ú�únú0T0 ú 0 ú�ú 00T0T0 ú 0 únú
»���������¼ �Table4.1shows theindicesof thecodewordsfor variouscodewordweights.Eachindex is
thedecimalequivalentof thedatabitswhichcorrespondto acodeword(e.g.,databits0001
areencodedinto codeword 0001011,which hasa weight of 3). Table4.2 reportswhich
codewordsarechosenby the stepwisealgorithmto computethe KAT lower bound. The
62
initial index setsare � ×�� and þ ×���× y ú � �V�U� � ú�� { . We noticethatat low£ � w � H
(e.g.,
-5 dB), only the closestcodewordsto codeword 0 areused. As the signal-to-noiseratio
increases(e.g.,at 4 dB), morecodewordsareaddedto the index set. Eventually, at hight£ � w � H(e.g.,at 10 dB), theentirecodebookis usedto calculatetheKAT lowerbound.For
theGolay(23,12)code,usingthecodewordswith weight7 allowsusto computetheKAT
boundusing253codewordsinsteadof using2047codewords.For theBCH (63,24) code,
weonly considercodewordsof weight15. Thereareonly 651codewordswith this weight
andcomputationof theKAT boundtakesa very smallamountof time andmemory, while
usingtheentirecodebookof size16777216codewordsis computationallynot feasible.
If oneevaluatestheexpressionon theright handsideof (2.1) for a subsetof thecode-
book, onewill obtainneitheran uppernor a lower bound. We will refer to this quantity
asthe“Hunter-basedestimate”.Restrictingthemaximizationof thesecondsum(e.g.,via
Kruskal’s algorithm)on the right handsideof (2.1) will leadto a looseupperbound. To
tightenthisbound,morenodesneedto beaddedto the“sub-tree”whichcontainsthenodes
correspondingto the subset. Figure4.1 demonstratesan idea to derive an upperbound
basedonKruskal’salgorithm.First,weconsiderasubsetthatcontainsthecodewordswith
minimumweightandapplyKruskal’s algorithm. This will maximizethesumof thesec-
ondorderPEPsfor thosecodewords. We next considerthe remainingcodewords(whose
weightsare larger than��÷ Ì ; ). Let ï Ì � , × ú � �U�V� � á be the subsetof the codewordswith
minimum weight and let ï å be a codeword not in this subset. For each æ , we computeè Ì å �× è H é1� H Ì ¤ � H å í for , × ú � �V�U� � á . To includeevery nodeæ in thespanningtree,we add
theedgewith maximum è Ì å .63
4.6 Numerical Results
We considera systemwith a uniform i.i.d.binary source,variouschannelcodes,and2-
and1-dimensionalsignaling. A typical plot for a two dimensionalsignalingschemeis
shown in Figure4.2, wherethe incomingbits areBCH (15, 5) coded,8-PSKmodulated,
andsentover theAWGN channel.Theentirecodebookis usedto calculatetheboundsfor
thisfigure.WeconsidertheHunterupperboundandtheKounias,KAT anddeCaenlower
boundsandobserve thattheKouniaslower boundis thetightestamongthelower bounds.
A similar behavior is observed for othercodesandcoderates,suchasthat in Figure4.3,
which shows theboundsfor a systemwith BCH (15,7) code,AWGN channel,andBPSK
modulation.
Figure 4.4 comparesthe result of computingthe KAT lower boundusing the entire
codebookwith thatof usingonly thecodewordswith theminimumHammingweight for
theGolay(23,12) code.For thesubset,weonly usetheclosestcodewordsto codeword ï Hin theHammingdistancesensebeingmotivatedby theresultof Table4.2,which suggests
thatat low£ � w � H
, it is enoughto considerthecodewordswith theminimumweight. Inter-
estingly, wenoticethattheKAT boundcomputedonthesubset,givesa tighter lowerbound
at low to moderate£ � w � H
ascomparedwith the KAT boundcomputedon the codebook
andalsothebestlower boundin [30]. As predictedby Table4.2,at high£ � w � H
, theKAT
boundusedon theentirecodebookis slightly tighter; but this tightnessis achievedat the
priceof having amuchhighercomplexity andtheimprovementis almostnegligible.
Figures4.5,4.6,and4.7comparethetightnessof theBlER boundsfor thecasewhere
only a subsetof thecodebookis usedto computetheboundsfor theGolay(23,12),BCH
64
(31, 16), and BCH (63, 24) codes,respectively. BPSK modulationis used,hencethe
lowerboundof [156] canalsobecomputed.TheKouniasandKAT lowerbounds(usinga
subsetof theclosestcodewords)arestill tight, but theHunterupperboundis looseat low£ � w � H. Poltyrev’s upperbound[149] is superiorto theHunterupperboundoptimizedon
thesubset,althoughthey areboththesameastheunionboundat BlER valuesof interest
(BlER í ú 0 @�� ). Notefirst thattheHunterupperboundis easierto computethanPoltyrev’s;
therefore,it is thechoiceat BlER valuesof interest.Second,theKAT andKouniaslower
boundsaretighter thanSeguin’s bound;hencethey arealsotighter thanShannon’s lower
bound[164] at leastat mediumto high£ � w � H
. Also noticethat theKAT boundis always
tighterthantheCohenandMerhav (dot-product)lowerboundof [30].
Figures4.8and4.9demonstratetheperformanceof theunion,Hunter, Kounias,KAT,
andSeguinboundsfor theRayleighfadingchannelfor Hamming(7, 4) andBCH (15, 5)
codes.TheKouniaslowerboundis still tight, but theHunterupperboundis notastight for
theRayleighfadingchannel,suggestingthatthehigherorderprobabilitiesaresignificantin
theblock fadingcase.As for theBER,weobserve in ourcalculationsthattheupperbound
(resultingfrom the Kouniaslower bound)is tight, but the lower bound(derived from the
Hunterupperbound)is too looseto beuseful.
65
��
!"# $
%'&)( *+-,�./102436587:9<;>=@?BADC65FEHGIGJ;K5�L'M)N OP>QSRTPVUXWZY>W\[]@^_PKUXP`UFa6PVU_bBP>P
Figure4.1: Theideato form aspanningtreeandcomputeanupperboundfor theBlER. ,dcis theindex of codeword ï Ìfe for which è Ìfe å c�è Ì å � , × ú � �U�V� � á .
66
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10−2
10−1
Eb/N
o in dB
Pro
babi
lity
of E
rror
Block Error Rate
Bit Error Rate
union boundHunter upper boundsimulationKounias lower boundKAT lower boundde Caen’s lower bound
Figure4.2: Block andbit error ratecurvesfor BCH (15, 5) code,8-PSKmodulation,and
AWGN channel.
67
−2 −1 0 1 2 3 4 5 6 6.510
−4
10−3
10−2
10−1
100
Eb/N
o in dB
Blo
ck E
rror
Rat
e
union boundHunter upper bound, subsetHunter−based estimatesimulationKAT lower boundKounias lower bound, subsetSeguin lower bound
Figure4.3: Block errorratecurvesfor BCH (15, 7) code,BPSKmodulation,andAWGN
channel.
68
codeword weight databits (in decimal)
3 1, 2, 5, 6, 8, 11,12
4 3, 4, 7, 9, 10,13,14
7 15
Table4.1: Thedatabits (in decimal)correspondingto variouscodewordweightsfor Ham-
ming (7, 4) code.
£ � w � Hcodeword indices
-5 dB 1, 2, 5, 6, 8, 11,12
4 dB 1, 2, 5, 6, 8, 9, 10,11,12,14
10dB 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,11,12,13,14,15
Table4.2: Codewordsusedin the optimizationof the KAT boundvia the stepwisealgo-
rithm. Only theclosestcodewordsareselectedat low£ � w � H
, while thewholecodebookis
usedat high£ � w � H
.
69
0 1 2 3 4 5 610
−5
10−4
10−3
10−2
10−1
Eb/N
o in dB
Blo
ck E
rror
Rat
e
KAT lower bound, subsetCohen and Merhav’s lower boundKAT lower bound
Figure 4.4: Comparisonamongthreelower boundson the BlER. Golay (23, 12) code,
BPSKsignaling,andAWGN channel.
70
−1 0 1 2 3 4 5 5.510
−4
10−3
10−2
10−1
100
Eb/N
o in dB
Blo
ck E
rror
Rat
e
union boundHunter upper bound, subsetHunter−based estimatePoltyrev upper boundsimulationKAT lower bound, subsetCohen and Merhav lower boundKounias lower bound, subsetSeguin lower bound
Figure4.5: Block errorrateboundsfor Golay(23,12) code,BPSKsignaling,andAWGN
channel.
71
−1 0 1 2 3 4 5 5.510
−4
10−3
10−2
10−1
100
Eb/N
o in dB
Blo
ck E
rror
Rat
e
union boundHunter upper bound, subsetHunter−based estimatePoltyrev upper boundKAT lower bound, subsetKounias lower bound, subsetSeguin lower bound
Figure4.6: Block error rateboundsfor BCH (31, 16) code,BPSKsignaling,andAWGN
channel.
72
−1 0 1 2 3 410
−4
10−3
10−2
10−1
100
Eb/N
o in dB
Blo
ck E
rror
Rat
e
union boundPoltyrev upper boundHunter−based estimateKAT lower bound, subsetKounias lower bound, subsetCohen and Merhav lower bound
Figure4.7: Block error rateboundsfor BCH (63, 24) code,BPSKsignaling,andAWGN
channel.
73
10 15 20 25 30 35 40 45 50 55 6010
−6
10−5
10−4
10−3
10−2
10−1
Eb/N
o in dB
Blo
ck E
rror
Rat
e
union boundHunter upper boundKounias lower boundKAT lower boundSeguin lower bound
Figure 4.8: Block error rate boundsfor Hamming (7, 4) code, BPSK signaling, and
Rayleighfadingchannel.
74
10 15 20 25 30 35 40 45 50 55 6010
−6
10−5
10−4
10−3
10−2
10−1
Eb/N
o in dB
Blo
ck E
rror
Rat
e
union boundHunter upper boundKounias lower boundKAT lower boundSeguin lower bound
Figure4.9: Block errorrateboundsfor BCH (15, 5) code,BPSKsignaling,andRayleigh
fadingchannel.
75
Chapter 5
Bounds for the Symbol and Bit Error
Rates of Space-Time Orthogonal Block
Codes
5.1 Introduction
The original paperson space-timetrellis codes[194] and space-timeorthogonalblock
codes[192] adopt the Chernoff upperboundto estimatethe pairwiseerror probability
of codewordsandto establishcodedesigncriteria. Although the Chernoff boundyields
successfulcodeconstructions,it is quitelooseevenathighvaluesof thechannelsignal-to-
noiseratio. Furthermore,it is commonpracticeto usetheunionboundto approximatethe
symbolerror rateor bit error rate. However, theunionboundis loose,particularlyat low
CSNRs.Therefore,usingtheChernoff boundtogetherwith theunionboundmayresultin
76
poorapproximationsto systemperformance(see,e.g.,Section5.4).
With respectto theSER/BERanalysisof space-timeorthogonalblock codes,onecan
usethe closed-formconditionalSERformulasin [175] (see(1.2) and(1.3)) andaverage
themover the fadingcoefficientsdistribution to derive exact SERformulasfor standard
PSK andsquareû þ ÷ -ary QAM, where v is a positive integer; this approachis taken in
[167]. However, the resultingSERformulasdo not hold for arbitraryconstellationssuch
asstar-QAM andit is not clearhow to extendthis approachto evaluatetheBER,specially
if Graymappingis not used.ThesymbolPEPscanbeusedto derive anupperboundon
theBER; however, aswill beshown in thesequel,theresultingBER upperboundis very
looseevenat mediumCSNRs.
In this chapter, we show how to usethe symbolPEPresultsto derive boundsfor the
SERandBERof STOB codesvia establishingtheHunterandKouniasbounds.Oneimpor-
tantfeatureof theboundsis thatthey hold for arbitraryconstellationsandsignalmappings
sincethey donotdependonthegeometryof thesystemathand.Numericalresultsindicate
that the boundsfor STOB codesoften coincidewith the true error probabilitiesobtained
via simulationsevenat low CSNRs.Furthermore,thecomputationalcomplexity of these
algorithmicboundsis very modestevenfor largeconstellationsandlargernumberof an-
tennas.
Theapproachpresentedin thepreviouschaptercanalsobe appliedto studytheerror
rateof space-timeorthogonalblockcodedMIMO channelsusedin conjunctionwith error-
controlblockcodes;wehoweverdo notconsiderthis for thesakeof simplicity.
77
5.2 Bounds on the Codeword Pairwise Error Probability
of Space-Time Orthogonal Block Codes
Let usdenotetwo STOB codewordsby ê and gê , anddefinematrix h ×ji M Ì � kml with M Ì � k ×6 Ì � k Ê g6 Ì � k andmatrix� ×nhoh � . For STOB codes,
�hasonly oneeigenvalue � with
multiplicity�
. To verify this point, we notethata STOB codeword hasthepropertythatê·ê � × � ê � þ I � , hencewehave� ×phqh � ×¼éëê Ê gê�í;éîê Ê gêví � ׳é � ê � þ 9 � gê � þ í I � �which indicatesthat � � × =�=�= ק� � �×h�Ø× � ê � þ 9 � gê � þ �It is shown in [131] thatthePEPbetweenê and gê is givenbyè�éîêRì gêuíÎ× ú© e r òH p Í�ÏÑÐ þÓÒ�s ��� � 9 Í)ÏÑÐ þÓÒ r ; M Ò � (5.1)
A lowerboundis givenin [131] which is derivedby replacingÍ)ÏVÐ þ Ò at thedenominatorof
(5.1)by 1, resultingin è6éëêRì gêuí<t þ û p ú:9 út ��� � r @ ; �where þ × t @ ; i þ ;; l . A simpleupperboundcanbederivedby minimizing thedenominator
of (5.1),i.e.,by replacingÍ�ÏÑÐ þ Ò by zero,to getè6éëê�ì gêví í þ û p út ��� � r @ ; � (5.2)
Note that as the CSNRgrows, the above upperand lower boundsconverge. Figure5.1
shows a typical behavior of the above boundsfor the codewords ê × i �� @ �� l and gê ×78
i @ Ì@ Ì @ ÌÌ l for Alamouti’sschemewith onereceiveantennaandquadraturephase-shiftkeying
(Q-PSK)signaling.Thetightnessof thenew upperboundandthelower boundin [131] is
obvious.Ourupperboundis closerto theactualPEPatmediumto highCSNRandcanbe
usedfor space-timecodedesignasin [131].
5.3 Bounds on the Error Rates of Space-Time Orthogonal
Block Codes
In thissection,weshow how theapproachof thepreviouschapterfor theHunterupperand
Kouniaslowerboundscanbeusedfor space-timeorthogonalblockcodes.
5.3.1 The Symbol Error Rate
Following theapproachin (4.1),for aconstellationyvu Ì � , × ú � �V�U� ��~ {, wherev ×xwVçzy þ é ~ í
is apositive integer, anda uniform i.i.d. bit-stream,theSERis givenby
SER × ú~ �" ÛÜ � è Û ø �Ì �Ü�Û � Û Ì �d�(5.3)è Û é�� Û Ì í is thePEPandis given in (3.18). In orderto find lower andupperboundson the
probabilityof eachunionin (5.3)via (2.2)and(2.1),respectively, we alsoneedto find the
probability of the intersectionof � Û Ì and � Û { , the secondorderPEPof symbols u Ì and u {with u Û . Thisprobabilitycanbederivedfromè Û é�� Ì Û¥¤ � { Û íÎ× ï|q}'¦ Ü ¨ Ì~{ Û � � Ì Û k ¶ � � { Û k ¶ Þ<� �
(5.4)
79
where ¨ Ì~{ Û × � u Ì Ê u Û � u { Ê u Û �( u Ì Ê u Û (Ñ( u { Ê u Û ( �� Z �)¡2� ×�� y�Z { � y ¡ { 9�� y�Z { � y ¡ { ( �sy = { and � y = { arethe real andimaginaryparts,re-
spectively), ¶ is definedin (3.15),and¦ é1¨ � â Ì é ¡ í � â { é ¡ í"í is thebivariateGaussianfunction
(givenin (4.7))withâ Ì é ¡ íµ×+� Ì Û k ¡
. Asâ Ì é ¡ í and
â { é ¡ í arenon-negative,we canusethe
resultof [174] to write¦ é1¨ � â Ì é ¡ í � â { é ¡ í�í as¦ é[¨ � � Ì Û k ¡Ó� � { Û k ¡ í × úû<© e ½ ¾ � À�Á� à Á
� Ä ÆH Ç�È�É p Ê � þÌ ÛûÎÍ�ÏÑÐ þÓÒ ¡ r M Ò9 úûb© e ½ Ô � à Á� À�Á� Ä ÕH Ç�È�É p Ê � þ{ Ûû�Í�ÏÑÐ þ Ò ¡ r�M Ò � (5.5)
where M Ì Û × ( u Ì Ê u Û ( , Öié^Z � ¨!íØ×ðÙ}Û Ð @ �:Ü Z o ú Ê ¨ þ w_é ú Ê ¨TZ$í�Þ , and ÙXÛ Ð @ � é[Z$í is as re-
definedbelow (4.10). Usingthepdf of ¶ in (3.16),we find theexpectedvalueof eachof
theintegralsin (5.5)asfollows.£�| O e ½H \ @�� �_��� | M Ò Q × úé[� Ê ú í�L ehgH ¡ ; @ � \ @ ª e ½H \ @�� �8��� ª M Ò M ¡× úé[� Ê ú í�LTe ½H e gH ¡ ; @ � \ @ � � J � �8���U� ª M ¡ M Ò (5.6)× e ½H p úú:9�� é Ò í r ; M Ò � (5.7)
where � é Ò í is a givennon-negative functionof Ò . Thestepfrom (5.6) to (5.7) follows by
writing theinnerintegralin (5.6)in termsof the é1� Ê ú í st derivativeof theLaplacetransform
of the unit stepfunction. Using (5.7) with � é Ò í�× � þ w û�Í)ÏÑÐ þ Ò and[175, Eq. 5A.35], we
80
have á$éGÖ � ��í �× úû<© eh½H p û�Í�ÏÑÐ þ Ò� þ 9 û�Í)ÏÑÐ þ Ò r ; M Ò× Öû<© Ê nû<© �k û 9 � þ ; @ �" � Ü2H p û �� r úé û é û 9 � þ í�í �Ê �© k û 9 � þ ; @ �" � Ü2H � @ �"÷ Ü2H p û �v r é Ê ú í ÷ J �é û é û 9 � þ í"í � Í)ÏVÐ i û n éG� Ê v í lû éG� Ê v í �(5.8)
where
n × úû Ù}Û Ð @ � �� 9 © û O ú Ê p ú:9 sgné � íû r sgné � í Q �with
� × � k û 9 � þ Í)ÏVÐ�û Ö ,� ×Ué úµ9 � þ í/æèç Í=û Ö Ê ú , andsgné[Z$í �× ( Z ( w<Z if Zx�× 0 and 0
otherwise.From(5.4) and(5.5),we obtainthe following expressionfor thesecondorder
PEP è Û é�� Ì Û¥¤ � { Û íÎ×�á p Ö p � Ì Û� { Û � ¨ Ì~{ Û r � � Ì Û r 9 á p Ö p � { Û� Ì Û � ¨ Ì~{ Û r � � { Û r �(5.9)
whereáýé = � = í is definedin (5.8).
5.3.2 The Bit Error Rate
Thesamealgorithmcanbeusedasoutlinedin thepreviouschapterto estimatetheBER.
Wehave
BER × ú~ �" Û°Ü � è Û é�� � �[� í �whereè Û é�� � �[� í is thebit errorprobabilitywhen u Û is sentandis givenbyè Û é1� � �[� íà× úv �" { Ü � �q� é æ �}| íIè6é guµ×pu { ( u�×xu Û íà× úv �" { Ü � �5� é æ �)| í ø ú Ê è Û ø �Ì �Ü { � Ì~{ �>� �
(5.10)
81
where�q� éfæ �}| í is theHammingdistancebetweenthebit assignmentsof u { and u Û . From
theabove, it is clearthatfinding upperandlower boundson theBER requiresevaluating
lower and upperboundson the probability of the union in (5.10), respectively. These
bounds,in turn, requirethecomputationofè Û é1� Ì~{ í × £�| B ÷ p �ÐþÌ Û Ê � þ{ Û� Ì~{ k ¶ r E× úû ø ú Ê � þÌ Û Ê � þ{ Ûé�� þÌ Û Ê � þ{ Û í þ Ê û � þÌ~{ � � ! @ �" � Ü2H p û ��Pr p � þÌ~{û é�é�� þÌ Û Ê � þ{ Û í þ 9 û � þÌ�{ íbr � �aswell as è Û é�� Ì~{ ¤ � � { íð× £�| B ¦ p ¨ Ì � { � �ÐþÌ Û Ê � þ{ Û� Ì~{ k ¶ � �Ðþ� Û Ê � þ{ Û� � { k ¶ r E× á p Ö p � � { é1� þÌ Û Ê � þ{ Û í� Ì�{ é1� þ� Û Ê � þ{ Û í � ¨ Ì � { r � � þÌ Û Ê � þ{ Û� Ì~{ r9 á p Ö p � Ì~{ é��Ðþ� Û Ê � þ{ Û í� � { é1� þÌ Û Ê � þ{ Û í � ¨ Ì � { r � � þ� Û Ê �Ðþ{ Û� � { r �
(5.11)
whereáýé = � = í is definedin (5.8). Equation(5.11)holdswhenthesecondandthethird argu-
mentsof¦ é = � = � = í arenon-negative. Negativeargumentsshouldbedealtwith asexplained
below (4.12),showing that,in any case,anexpressionexistsfor thesecondorderPEP.
5.4 Numerical Results
For this part, the lengthof the input bit-sequenceis max(100000,100 � BER@ � ). The
computationtime of the algorithmicBonferroniboundsis negligible for our codethat is
written in C andrunson aSUN Ultra 60 machine.TheChernoff andtheunionboundsare
alsogivenfor theSER.
82
The SERandBER versusCSNRcurvesarepresentedin Figures5.5 through5.9 for
various�
,W
, space-timecodes,and~
-ary PSKandQAM constellationswith Graymap-
ping. Figures5.5 and 5.6 show the performanceof Alamouti’s � þ code[9], while the
performanceof thecodes� � and � s of [192] arepresentedin Figures5.7and5.9,respec-
tively. TheHunterandKouniasboundsfor bothSERandBER curvesarevery tight and
canhardlybedistinguishedfrom eachotherandfrom theperformancecurve obtainedvia
simulation. It canbe seenthatasthe constellationsizegrows, the union boundbecomes
lessreliableandtheChernoff boundgetsfartherfrom theothercurves,while thelowerand
upperboundsremainverytight. Figure5.7showstheBERupperandlowerboundsaswell
assimulationresultsfor a systemwith threetransmitandW × ú � û � and t receiveantennas
for a wide rangeof CSNRs.Theboundsaretight evenat negativeCSNRvaluesandhigh
diversityorders.NoticethattheStar8-QAM constellationof Figure5.9(with signalpoints
andmappingsasdepictedin Figure5.8) is not a regular squareQAM nor 8-PSK,but the
boundsarestill very tight.
An upperboundon theBER canbecomputedby upperboundingè6é gu ×�u { ( u ×�u Û í in
(5.10)by thePEPto get
BER ö ú~ v "PÛË" { �q� é |$� æ!íIè6é�u Û ì u { í �Thisboundis alsoplottedin all figures.It is clearlyobservedthatourupperboundis much
tighter, particularlyfor largerconstellations.
AsGraymappingisusedin theabovesystems,it isexpectedthatathighenoughCSNRs
theBER of PSK-modulatedsystemsconvergesto [150] BER � �÷ SER�where
û ÷is the
constellationsize. Thecurveslabeledwith “SER/m” in Figure5.5 show this approxima-
83
tion. We alsoobserve that asthe constellationsizeshrinksor the numberof the receive
antennasgrows, this estimatebecomestighter. Therefore,at high enoughCSNRvalues,it
would beenoughto find only oneof theSERboundsto obtaina goodestimateof system
SERandBER.For example,for 16-PSKsignalingandataCSNRof 15dB, thecoinciding
SERlowerandupperboundsareequalto 0.195682.TheaboveapproximationyieldsBER� 0.048921,while thelower andupperboundson theBER coincideandequal0.051838.
It is alsoimportantto notethatalthoughwe have presentedtheBER resultsfor theGray
signalmapping(exceptfor Figure5.9), our boundsalsoapply to any othermappingand
show thesamebehavior.
84
0 2 4 6 8 10 12 14 16 18 20 2210
−4
10−3
10−2
10−1
100
CSNR in dB
Cod
ewor
d P
airw
ise
Err
or P
roba
bilit
y
upper bound in (5.2)upper bound of Tarokh et al. [194]lower bound of Lu et al. [131]
Figure5.1: Our upperboundandthelowerboundson � i<i �� @ �� l�� i @ Ì@ Ì @ ÌÌ lbl , Alamouti’s
codewith Q-PSKmodulation,� × û
, andW × ú .
85
−5 0 5 10 15 2010
−3
10−2
10−1
100
SER
BER
CSNR in dB
Pro
babi
lity
of E
rror
Chernoff−based union boundunion boundupper boundsimulationlower boundPEP−based BER estimateSER/2
Figure5.2: Resultsfor Q-PSKsignalingwith � �� andW �� (2 bits/s/Hz).
86
−5 0 5 10 15 20 2510
−3
10−2
10−1
100
CSNR in dB
Pro
babi
lity
of E
rror
SER
BER
Chernoff−based union boundunion boundupper boundsimulationlower boundPEP−based BER estimateSER/3
Figure5.3: Resultsfor 8-PSKsignalingwith ����� andW ��� (3 bits/s/Hz).
87
4 6 8 10 12 14 16 1810
−3
10−2
10−1
100
CSNR in dB
Pro
babi
lity
of E
rror
SER
BER
Chernoff−based union boundunion boundupper boundsimulationlower boundPEP−based BER estimate
Figure5.4: Resultsfor 16-QAM signalingwith ���x� andW �x� (4 bits/s/Hz).
88
−5 0 5 10 15 20 2510
−3
10−2
10−1
100
CSNR in dB
Pro
babi
lity
of E
rror
SER
BER
Chernoff−based union boundunion boundupper boundsimulationlower boundPEP−based BER estimateSER/5
Figure5.5: Resultsfor 32-PSKsignalingwith ����� andW �x� (5 bits/s/Hz).
89
8 10 12 14 16 18 20 22 24 26 2810
−4
10−3
10−2
10−1
100
CSNR in dB
Pro
babi
lity
of E
rror
SERBER
Chernoff−based union boundunion boundupper boundsimulationlower boundPEP−based BER estimate
Figure5.6: Resultsfor 64-QAM signalingwith ���x� andW �x� (6 bits/s/Hz).
90
−5 0 5 10 15 2010
−4
10−3
10−2
10−1
100
CSNR in dB
Bit
Err
or E
rror
L = 1
L = 2
L = 4
upper boundsimulationlower bound
Figure5.7: Resultsfor 8-PSKsignalingwith ���x� andW ���z�\��� and � (1.5bits/s/Hz).
91
�
�
� �¡[011]
� £¢6¡[111]
�¡[000]
�¢6¡[100]
� ¡¥¤[001]
� ¢6¡£¤[101]
� £¡£¤[010]
� £¢6¡¥¤[110]
Figure5.8: Thestar-QAM signalingschemewith quasi-Graymapping.¦ is anormalizing
factorandequals�@§z�©¨ �6ª and «¬� ¨ Ê � .
92
2 4 6 8 10 12 14 16 18 20 22 2410
−4
10−3
10−2
10−1
100
CSNR in dB
Pro
babi
lity
of E
rror
SER
BER
Chernoff−based union boundunion boundupper boundsimulationlower boundPEP−based BER estimate
Figure5.9: Resultsfor Star8-QAM signalingwith ���x� andW ��� (1.5bits/s/Hz).
93
Chapter 6
Error Rates of STOB Coded Systems
under MAP Decoding
6.1 Introduction
Whentheinput to thecommunicationsystemis uniformi.i.d., theoptimaldecodingrule in
theminimumcodeworderrorratesense,is theML decodingrule. Thiscasewasconsidered
in the previous chapter, wherewe derived tight upperand lower boundson the symbol
and bit error ratesof space-timeorthogonalblock codedchannelswith arbitrary signal
geometryandsignalmappings.
In this chapter, we study how exploiting the non-uniformsourcedistribution at the
transmitterand/orthe receiver can improve the performanceof MIMO systems.In par-
ticular, the effects of symbol MAP decodingas well as signal designand mappingare
addressedfor systemswhich usespace-timeorthogonalblock coding. As will beseenin
94
thesequel,obtainingthesymbolpairwiseerrorprobabilityof theMAP decodedsymbols
requiresfinding theexpectedvalueof �®B¯±°�²³�´ , where µ is a realnumber. This is more
involvedthantherelatedderivation in theML decodingcase,wheretheproblemis com-
puting theexpectedvalueof �®B¯ ´ . TheMAP decodingrule, which we alsoderive here,
andthepairwiseerrorprobabilityexpressionunderMAP decodingshow thattherecanbe
a largegainin performingMAP decodingascomparedwith ML decoding.
MAP decodingfor sourceswith redundancy (due to non-uniformdistribution and/or
memory)is a form of joint source-channelcoding.It would thenbeinterestingto compare
theperformanceof MAP-detectedschemeswith thatof separate,independentsourceand
channelcoded(tandem)systems.For BPSK signaling,it is observed that at low CSNR
values,MAP decodingis superiorto tandemcoding. As thesizeof thesignalingscheme
grows, this resultholdsfor all practicalerrorratesof interest.
Anothercontribution of this chapteris the establishmentof tight Hunter (upper)and
Kounias(lower) boundson the SERandBER of STOB codedchannelsunderMAP de-
coding.Similar to thepreviouschapter, wedeterminetheprobabilityof theintersectionof
two pairwiseerroreventsanduseit togetherwith our closed-formPEPformula to derive
HunterandKouniasbounds. We demonstratethat the boundsprovide an excellentesti-
mateof the error rateseven at low (negative) CSNRvalues. For moderatevaluesof the
CSNR,the upperandlower boundsareoften the sameup to four significantdigits. The
SERboundsaremuchtighterthantheunionbound,andtheBER boundsaresignificantly
betterthantwo computationallylessexpensiveestimates.
95
6.2 The MAP Decoding Rule
6.2.1 Decoupled Symbol Detection
As outlinedin Chapter3, theinput-outputrelationfor a MIMO channelwhichusesspace-
timeorthogonalblockcodingcanbeconvertedto theformgivenby (3.12),whichwerepeat
herefor simplicity ¶·�¸º¹� »¼ ¸4½ »·�¸ � ¾'¿ ÀÂÁ� à ¸TÄ ° ¶Å£¸@ÆWe recall in theabove equationthateachentryof
¶·�¸ is associatedwith only onesymbol.
Therefore,in thenew case(i.e.,underMAP decoding),if weshow thatthenoisevector
¶Å£¸is composedof i.i.d. randomvariables,thenwecandetectsymbol « by only consideringthe« th entryof thevectors
¶· ¸ , �ÈÇÊÉ�ÇÌË , whichmeansthatdetectionof symbolsis decoupled
underMAP decoding.
In order to find the distribution of the noisevector
¶Å£¸ , we considertwo noisesam-
ples
¶ÍÏÎÐÒÑ and
¶ÍÏÓÐ)Ô at two arbitrary symbol intervals ÕvÖ and ÕS× , and for two arbitrary re-
ceive antennasØ and Ù . Since
¶Í ÎÐ Ñ and
¶Í ÓÐ Ô are weightedsumsof independentnormal
randomvariables,they haveGaussiandistribution. Also, it is straightforwardto verify thatÚÜÛ ¶ÍÏÎÐÒÑ\Ý � ÚÞÛ ¶ÍoÓÐßÔTÝ �pª . Hence,thecorrelationof thesenoisesamplesisàâá ¶Í ÎÐ Ñ ¶Í ÓdãÐ Ôåä � Úçæéèxêë Ì Ü Ö »ì Î�ãÌ�í Ð Ñ »Í ÎÌïî èpêë ¸ Ü Ö »ì ÓÒ㸠í Ð Ô »Í Ó¸ î ãñð� ë Ì ë ¸ »ì Î\ãÌ�í Ð Ñ »ì Ó¸ í Ð Ô Úxò »Í ÎÌ »Í ÓÒã¸�óSince »Í ¸ô �6�ÜÇ ÉÏÇ�Ë , arezero-meani.i.d., theabove sumis zerounlessØq��Ù and «£�xÉ ,
96
in which caseit equalsêë Ì Ü Ö »ì Î\ãÌõí ÐKÑ »ì ÎÌ�í Ð)Ô Ú ò'ö »Í ÎÌ ö × ó � ë Ì »ì Î\ãÌ�í ÐÒÑ »ì ÎÌ�í Ð)Ô (6.1)
� ÷øù øúû Ì ö »ì ÎÌ�í ÐÒÑ ö × �̾ à Πif Õ�Ö£� Õ�ת otherwise,
(6.2)
where(6.1) follows from the fact that »Í ¸ô is unit-varianceand(6.2) follows from the or-
thogonalityof »¼ . Thus,weobtainthat¶Í ÎÌýü i.i.d. þïÿ ®)ªÂ�>¾ à Π´ Æ (6.3)
As we mentionedin thefirst paragraphof this section,the above result,i.e., the fact that
theentriesof thenoisevector
¶Å Î � � ¶Í ÎÖ � ÆõÆfÆ � ¶Í ÎÐ���� arei.i.d., provesthatsymboldetection
is decoupled,similar to theML decodingcase.
6.2.2 The MAP Decoding Rule
For MAP decoding,
¶� � � ¶· Ö � ÆõÆfÆ � ¶·�� � canbeusedinsteadof� � � · ÖD� ÆfÆfÆ � · ê � because
¶�is
generatedfrom�
throughinvertibleoperations.Thedetectionrule is givenby Ì � argmax ��® ö Û ¶�� Ì Ý � Ü Ö � ¼ ´� argmax � ® Û ¶�� Ì Ý � Ü Ö ö � ¼ ´���� ® ´� argmax �� Ü Ö ������� ® Û ¶�� Ì�� ¾�� Ã Ý � Ü Ö ´�� � ® ´ (6.4)
� argmax æ"!$# Û � ® ´ Ý � �ë Ü Ö ö¶� Ì%� ¾ � à ö ×¾ à ð � (6.5)
where¾ � � ¾'& ( )* , and(6.4)and(6.5)holdbecause
¶Í ÎÌ arei.i.d.andGaussian,respectively,
asindicatedin (6.3).
97
6.3 Exact Symbol Pairwise Error Probability with MAP
Detection
6.3.1 The Conditional PEP
Without loss of generality, we considerMAP decodingfor the Ø th symbol period. The
errorprobabilitiesmaybedeterminedusingtheMAP detectionmetricgivenin (6.5). The
receivershouldevaluatethismetricfor symbols Ì and ¸ giventhat Ì is transmitted(hence¶� Î �̾ � Ã Ì ° ¶Í Î ) anddecidein favor of theonewhichyieldsa largermetric.
From(6.5), theprobability that ¸ is preferredover Ì when Ì is sent, +-,@® Ì-. ¸ ´ , is
theprobabilityof theevent!$#`Û � ® Ì ´ Ý � /0 ¾ �ë Ü Ö ö¶� Î � ¾ � Ã Ì ö ×Ã Ç !1# Û � ® ¸ ´ Ý � /0 ¾ �ë Ü Ö ö
¶� Î � ¾ � à ¸ ö ×à �which is equivalentto�ë Ü Ö ¨ 0�2 ¸ � Ì � ¶Í Î43ö ¸ � Ì ö 5 /ö ¸ � Ì ö 60 À�Á !$# � ® Ì ´� ® ¸ ´ ° ¾ ö ¸ � Ì ö ¿ ÀÂÁ0 6 �ë Ü Ö Ã Æ (6.6)
From(6.3),it followsthat 7 × 8 :9�;< � í �� �=?>@ :9�;< � @ arei.i.d. þïÿ ®)ªÂ�>¾ à ´ . Hencethesumontheleft hand
sideof (6.6) is þïÿ ®)ªÂ�4¾ û � Ü Ö Ã ´ . Therefore,theprobabilityof theeventin (6.6),which is
thepairwiseerrorprobabilityconditionedon thepathgains,is givenby�ý® ÌA. ¸ ö ¼ ´ �p è ¿ ¾ À�Á0 6 ö Ì � ¸ ö ¨ à ° 60 ¾ À�Á /ö Ì � ¸ ö !$# � ® Ì ´� ® ¸ ´ /¨ à î � (6.7)
where,asmentionedin theChapter3,
à � �ë Ü Ö Ã � *ë Î Ü Ö �ë Ü Ö ö ì Î ö × Æ98
6.3.2 Interpretation of the PEP as a Laplace Transform
Basedon theprobabilitydensityfunctionof à , which wasderived in (3.16),andusinga
changeof variables,theaverageof (6.7)canbewrittenas�ý® ÌA. ¸ ´ � /®CB � / ´EDGF ×IHÌ ¸ JLKM N H ; ÖPO ;�Q RIS Ô� 9 UT ¨ N ° µ Ì ¸¨ N�VXW N � (6.8)
where F Ì ¸ � & Y (?)× * ö Ì � ¸ ö and µ Ì ¸ � Ö× !$#[Z]\ �1^Z]\ :9 ^ . We notethat the integral in (6.8) is the
Laplacetransformof N H ; Ö `_ ¨ N ° ² � 97 Qba evaluatedat cÜ� F ; ×Ì ¸ . We know that if � ®ed ´ andf ®gc ´ are Laplacetransformpairs (f ®hc ´ �ji Û � ®kd ´ Ý ), so are d H � ®ed ´ and ® � / ´ H�lPml Á m f ®gc ´ .
Therefore,weneedto find the B � / st derivativeof i á _ ¨ N ° ² � 97 Qba ä . Usingintegration
by parts,wehavederivedtheLaplacetransformof `_ ¨ N ° ² � 97 Qba asf�n�o�p ®gc ´ � irq T ¨ N ° µ Ì ¸¨ N Vts� / � sgn®)µ Ì ¸ ´0 c � /0 T /c ¨ 0 c£° / ° sgn®)µ Ì ¸ ´c V O ; \ ² � 9 u @ ² � 9 @ 7 × Á u Ö ^ Æ (6.9)
We now needto evaluatethe B th derivativeof (6.9)which is calculatedin thenext subsec-
tion.
6.3.3 The PEP in Closed Form
Herewederivethe B th derivativeof (6.9)with respectto c . First,wepresentalemmawhich
canbeprovedeasilyvia induction.
Lemma–Thefollowing hold
a) lvml Á mxw ÖÁ]y � \ ; Ö ^ m H{zÁ m | Ñb) l ml Á m ® 0 c<° / ´]} Ô � ® 0 c<° / ´]} Ô ; H H ; Ö� Ì Ü M ® � � 0 « ´
99
c) l ml Q m O ; \�~ u���Q ^ ��® � / ´ H���H O ; \G~ u���Q ^ ÆThederivativeof thefirst termin (6.9) maybefoundvia (a). As for theproductterm,
weusetheLeibniz’s formula[1] to treatthetwo termsseparately. Theformulais
W HW�� HA�%� � Hë Ì�� M T B « V W Ì �Wb� Ì W H; Ì �W�� H ; Ì � (6.10)
Weapply(6.10)with � � ~ Á ° ÖÁ 7 × Á u Ö and � � O ; ® ~ u�� 7 × Á u Ö ´ . Using(6.10)againto find the« th derivativeof thesecondtermin � with � Ö � ÖÁ and � Ö�� Ö7 × Á u Ö , andapplying(a)and(b)
with � � � / resultsin
W ÌW c Ì T�� c ° /c ¨ 0 c¥° / V � ® � / ´ Ì « D �c Ì u Ö ° Ìë ¸ � M ® � / ´ Ì « D®�« � É ´ED � Ì ; ¸ � Ö ® 0�� � / ´c ¸ u Ö ® 0 c¥° / ´ Ì ; ¸ u ÑÔ� ® � / ´ Ì « D �c Ì u Ö ° ® � / ´ Ì « Dc Ì u Ö Ìë Î � M T 0 ØØ V c Î0 Î ® 0 c<° / ´ Î u ÑÔ � (6.11)
As for theexponentialterm,weuseaformulafor the B th derivativeof acompositefunction
from [71] which statesthatfor � ® � ´ � f ® N ´ , N ���Z® � ´ , wehave
W HW�� H � ® � ´ � Hë Ì�� Ö�� Ì« D f \ Ì ^ ® N ´ � where � Ì � Ì ; Öë Î � M T «Ø V ® � / ´ Î N Î W HW�� H N Ì ; Î Æ (6.12)
Lettingf ® N ´ � O ; \�~ u���Q ^
and N � ¨ 0 c¥° / , weuse(6.12),(b), and(c) to get
W H ; ÌW c H ; Ì O ; ® ~ u�� 7 × Á u Ö ´ � O ; ® ~ u�� 7 × Á u Ö ´ H ; Ìë ¸ � Ö ® 0 c¥° / ´ 9 Ô ; H u Ì � ¸É D ¸ ; Öë Î � M T É Ø V ® � / ´ ¸ u Î H ; Ì ; Ö� � M ®õÉ � Ø � 0�� ´ Æ(6.13)
Using (6.11)and(6.13) in (6.10)with � � µ Ì ¸ and � � ö µ Ì ¸ ö yields the B th derivative of
100
(6.9)as
W HW c H fMAP ®gc ´ � ® / � sgn®ßµ Ì ¸ ´>´ ® � / ´ H B D0 c H u Ö � /0 O ; ® ² � 9vu @ ² � 9 @ 7 × Á u Ö ´� Hë Î � M T B Ø V ® � / ´ Î Ø Dc Î u Ö è
sgn®ßµ Ì ¸ ´ ° Îë� � M T 0��� V c �0 � ® 0 c¥° / ´ � u ÑÔ î� H ; Îë � Ö ö µ Ì ¸ ö � D ® 0 c<° / ´ H ; Î ; �Ô ; Öë Z � M T �� V ® � / ´ u Z H ; Î ; Ö� Ó � M ® � � � � 0 Ù ´ Æ (6.14)
Using(6.14)with (6.8) resultsin the following expressionfor theexactPEPof MAP de-
codedspace-timeorthogonalblockcodes��® ÌA. ¸ ´ � / � sgn®)µ Ì ¸ ´0 � /0 O ;�� ² � 9?u @ ² � 9 @�� × Sv� Ô� 9 u Öe�� H ; Öë Î � M ® � / ´ H u Î ; Ö® 0 ° F ×Ì ¸ ´ H ; Î ; Ö��� sgn®ßµ Ì ¸ ´ ° F Ì ¸� 0 ° F ×Ì ¸ Îë� � M T 0��� V /® 0 F ×Ì ¸ °�� ´ ��� � H ; Î ; Öë � Ö ö µ Ì ¸ ö ® F ×Ì ¸ ° 0 ´ R ×� D¡F Ì ¸ ; Öë Z � M T �� V ® � / ´ u Z H ; Î ; ×�Ó � M ® � � � � 0 Ù ´ � (6.15)
wherewehavesetû£¢Ì¤� �"¥ Ì � / for ËL¦ � .
A specialcaseis ML decodingfor which µ Ì ¸ � Ö× !$# Z]\ �:^Z]\ :9 ^ � ª . Hence,thefirst sumin
(6.15)is non-zeroonly for ØÞ�§B � / andwehave�ý® ÌA. ¸ ´ � /0 �� / � F Ì ¸� 0 ° F ×Ì ¸ H ; Öë Î � M T 0 ØØ V /® 0 F ×Ì ¸ °�� ´ Î � �whichagreeswith theresultin (3.18).
101
6.4 The Optimal Binary Antipodal Signaling
In this subsectionwe considerbinary antipodalsignalingandoptimizeit in the senseof
minimizing theBERgivenby
BER �¨�ý®ª© � × ö � Ö ´�� � ® Ö ´ °«�ý®ª© � Ö ö � × ´���� ® × ´ Æ (6.16)
Normally, oneshouldusetheaveragedPEPin (6.16)with Ö�� � and ×<� � � , differentiate
theresult,andfind theoptimal � and � . However, this canbea tediousjob consideringthe
PEPgivenin (6.15).Therefore,weusethePEPsat thereceiverside,i.e.,given¼
, to find
thesolutionin aneasierway. Theoptimalconstellationderivedin thiswaywill notdepend
on¼
, thusjustifying our approach.
Let usassumethat � ® Ö ´ � � , andthebits 0 and1 aremappedto ׺� � � and Ö<� � ,
respectively. Letting ¬ � � × Y (?)* , ¨ ¦ �¬ \G~ u�� ^× ¨ N � and ® � Ö× !$# Ö ; ZZ , we canwrite the
BERconditionedon à as
BERé� � °T ¨ ¦ � ®¨ ¦ V °x® / � �V´ °T ¨ ¦�° ®¨ ¦ V Æ (6.17)
It is easyto verify that the BER is a strictly decreasingfunction of ¦ (regardlessof ® ).
Hence,given À�Á and� , in orderto minimizetheBER,onehasto maximize¦ . Notethat ¦is a scaleddistancebetweenthesignalpoints,therefore,signalingschemeswith thesame
distancebetweentheir signalshave identicalperformance.It is clearthattheconstellation
with constantaveragesignalenergy ( )* which maximizes¦ is thezero-meanconstellation,
becausea constellationwith a non-zeromeancansimply be shiftedto reduceits energy
withoutperformanceloss.
102
Fromthezero-meancondition,wehave� � �/ � � � �andtheaverageenergy conditionrequiresthat� � × °x® / � �V´ � × � À�Á6 �Theabovetwo equalitiesresultin
® � � � � ´ � ¿ ÀÂÁ6 T � ¿ �/ � � � ¿ / � �� V �which is thereforethe optimal binary antipodalconstellation.The above constellationis
identicalto theantipodalsignalingresultin [107] for thecaseof theAWGN channel.
6.5 Bounds on the SER and BER of STOB Coded Chan-
nels under MAP Decoding
In this section,we show how theapproachof thepreviouschapterin deriving Bonferroni-
typeupperandlower boundscanbe appliedto space-timeorthogonalblock codesunder
MAP decoding.Two lessexpensive upperanda lower boundson the BER arealsopre-
sentedin Section6.6andcomparedwith theBonferronibounds.
103
6.5.1 The Symbol Error Rate
As seenin thepreviouschapter, for a constellationof size ± � 0{�, where
�is a positive
integer, andani.i.d. bit-stream,theSERis givenby
SER �¨² ; Öë ³ � M �ý®h´ ö ³ ´k� ® ³ ´ �¨² ; Öë ³ � M � ® ³ ´ � ³ è-µÌC¶� ³ ´ Ì ³ î Æ (6.18)
Notethat � ³ ®h´ ³ Ì ´ is thePEPandis givenin (6.15).We now needto find theprobabilityof
theintersectionof ´ ³ Ì and ´ ³ ¸ which is givenby� ³ ®C´ ³ Ì�· ´ ³ ¸ ´ � à ¯ 湸 èAº Ì ¸ ³ � F ³ Ì ¨ à ° µ ³ ÌF ³ Ì ¨ à � F ³ ¸ ¨ à ° µ ³ ¸F ³ ¸ ¨ à î ð � (6.19)
whereº Ì ¸ ³ � 8 � ;<e» í 19 ;<e» >l � » l 9 » , W Ì ³ � ö Ì � ³ ö , Ã is definedin (3.15), and
¸ ® º � � � � ´ is the
bivariateGaussianfunction definedin (4.7). After trying a few numericalmethods,we
choseto compute(6.19)usingacombinationof 96-pointGaussianintegration[1] together
with Donnelly’s algorithm[42] for each à � N . This selectionled to the fastestandthe
mostaccurateresults.
6.5.2 The Bit Error Rate
Similar to thepreviouschapter, wewrite theBER as
BER �¨² ; Öë ³ � M � ® ³ ´ � ³ ®C´ \ � ^ ´ � (6.20)
where � ³ ®h´ \ � ^ ´ is givenin (5.10). Now, we needto computethepairwiseerrorprobability
between Ì and ¸ giventhat ³ is sent.i.e.,� ³ ®h´ Ì ¸ ´ � Ú ¯ æ è F ׳ Ì%� F ׳ ¸F Ì ¸ ¨ à ° µ Ì ¸F Ì ¸ ¨ à î ð �104
whoseclosed-formsolutionis givenin (6.15). We alsoneedto find theprobabilityof the
intersectionof ´ ¸ Ì and ´ ¸ Î giventhat ³ is sent,which is givenby� ³ ®h´ Ì ¸ · ´ Î ¸ ´ � Ú ¯ 漸 èAº Ì Î ¸ � F ×Ì ³ � F ׸ ³F Ì ¸ ¨ à ° µ Ì ¸F Ì ¸ ¨ à � F ×γ � F ׸ ³F Î ¸ ¨ à ° µ Î ¸F Î ¸ ¨ à î ð �
Therefore,theresultsof Section6.5.1canbeusedhereto computelowerandupperbounds
on theBER.
6.6 Numerical Results
6.6.1 Binary Antipodal Signaling
We first study the BER of binary antipodalsignalingto show the exactnessof our PEP
formulas.We simulatethetransmissionof an i.i.d. bit sequenceover theMIMO channel.
Thelengthof thebit-sequenceis max ® / ª�½6� / ªzª � BER; Ö ´ bits. Wealsolet �ý®H¯ �xª ´ � � .
We considera systemwith two transmitandonereceive antennasusingthe ¾ × (Alam-
outi) space-timecode in Figure 6.1. It is seenthat the analysisand simulationcurves
coincideeverywhere.This simulationalsoindicateshow MAP decodingcanimprove the
BER performance.TheMAP decodinggainover ML decodingis 0.63,1.57,and5.98dB
for � � ª ÆG¿ ��ª ÆGÀ , and ª ÆGÀbÀ , respectively, at BER = / ª ;<Á . Figure6.2 presentstheresultsof
Section6.4. It showsthata largegaincanbeobtainedthroughmodifying theconstellation
accordingto theprior probabilities.At aBERof / ª ;<Á , thegainof usingtheoptimalbinary
signalingandMAP detectionis respectively 2.7, 6.1, and20.1 dB for � ��ª Æ¡¿ �\ª Æ¡À , andª Æ¡ÀbÀ overanML decodedsystemwith BPSKmodulation.
105
In Figure6.3, we comparefour systems,i.e., two systemswith BPSK signalingand
ML or MAP decoding,andtwo systemswith optimalsignalingandML or MAP decoding.
Thesesystemsareindicatedby ML symmetric,MAP symmetric,ML optimum,andMAP
optimum,respectively. Weseethatif theCSNRis highenough,optimumsignalingandML
detectionoutperformsBPSKandMAP detection.In otherwords,usingadaptivesignaling
is superiorto MAP detection.Whenreceiver noiseis strong(i.e., at low À�Á ), the second
term in the argumentof the ý® �8´ function in (6.7) hasthe dominanteffect; henceMAP
decodingwith BPSKsignalsis moreeffective thanML decodingwith optimumsignals.In
lessnoisychannelconditions(high À�Á ), thefirst termin theargumentof the ý® �8´ function
becomesdominantand henceML decodingwith optimum signalingoutperformsMAP
decodingwith symmetricsignals.As previouslymentioned,MAP decodingwith optimum
signalingis alwaysbetterthanothersystems.
6.6.2 Tandem versus Joint Source-Channel Coding
Figures6.4 and6.5 comparea MAP decodedsystemwith two tandemsystemsfor a dual
transmit-singlereceivechannelwith BPSKand16-QAM signaling,respectively. Theinput
bit-streamis i.i.d. with ��®B¯ �nª ´ � ª ÆG¿�À so that the sourceentropy isì ®H¯ ´ � ª ÆG .
The tandemsystemsarecomposedof 4th orderHuffmancodingfollowedby eithera 16-
stateor a 64-staterate-1/2convolutionalcodingblock. Theconvolutionalcodesarenon-
systematicandchosenfrom [101]. The lengthof the input bit-streamis 10000bits. The
testis repeated1000timesandtheaverageBER is reported.It is observedthatthetandem
systembreaksdown (dueto errorpropagationin Huffmandecoding)forÚ � § Í MÄÃ / � dB
106
for thesystemwith BPSKsignalingandamuchhigherÚ � § Í M of � ª dB for thesystemwith
16-QAM modulation.For thesystemwith 16-QAM signaling,theMAP-decodedsystem
outperformsthe16-state(64-state)tandemcodedsystemfor BER ¦ � � / ª ;ÆÅ ( ¦ 0 � / ª ;ÆÅ ),which is therangeof theerrorratesof interestin mostpracticalapplications.This is avery
interestingresultwhich shows thatMAP decoding,which is far lesscomplex thantandem
coding,shouldbethechoicefor applicationsin whichtheinputbit-streamis notuniformly
distributed (suchas FAX). Notice that as the constellationsize (which translatesto the
systembit rate)grows, theMAP decodedsystembecomesmoresuperior.
6.6.3 SER and BER Bounds for Ç -ary Signaling
Figures6.8 to 6.14show theSERandBER boundsandthesimulationresultsfor various
PSKandQAM signalingschemesfor adualtransmit-singlereceivesystem.They compare
the curvesfor two systems:onewith an equiprobablei.i.d. bit-streamandML decoding
andanotherwith ani.i.d. bit-streamwith ��®B¯��pª ´ �pª Æ¡À andMAP decoding.Thelength
of thebit-sequenceis 2000000bits. For theSER,theunionboundgivenby
Unionbound � ë ³ � ® ³ ´ ë Ì �ý® ³ . Ì ´is alsoplotted. For theBER, we alsoplot SER§ � , where
0 Zis theconstellationsize. This
approximationis goodonly athighCSNRvaluesandwhenGraymappingis usedfor signal
mapping.TheothercurveasaBERestimateis basedon thesymbolPEPs.TheexactBER
is givenin (6.20).At highCSNR,onecanreplace�ý®ª© � ¸ ö � Ì ´ with �ý® ÌA. ¸ ´ . This
will resultin anupperboundontheBER,becauseclearly ��® ÌA. ¸ ´ ¦È�ý®ª© � ¸ ö � Ì ´ .107
Figure6.8demonstratestheSERboundsfor 16-QAM signalingandGraymapping(the
mappingis shown in [186, Figure8]). The upperandlower boundsprovide an excellent
approximationto the SERfor all (even negative) CSNRvalues. It is interestingto note
thatasthesourcebecomesmorebiased,theboundsbecometighter. Thecurvesalsoshow
that the union boundbecomeslooseras the CSNR decreases.Finally, we observe that
theMAP-decodedsystemis considerablysuperiorto theML-decodedsystem.TheMAP
decodinggainat BER � / ª ;<Á is 5.5dB.
TheSERandBERcurvesof 32-PSKsignalingareshown in Figure6.9andFigure6.10,
respectively. We observe from Figure6.10that theBER upperandlower boundsarealso
very tight for the entire CSNR range. This figure also shows that the PEP-basedBER
estimateis oftenvery looseascomparedwith theotherbounds.TheMAP decodinggain
for BER (SER)= / ª ;<Á is 1.85dB (1.68dB). Figure6.11presentstheSERcurvesfor 64-
QAM andGray mapping.The MAP decodinggainat SER= / ª ;<Á is aslarge as6.7 dB.
Theupperboundfor theML-decodedsystemis up to 10%larger thanthe lower boundat
low CSNRvalues,but it becomestight astheCSNRgrows. Both boundsaretight for the
MAP-decodedcase.Theboundsfor smallerconstellationsaretighterandothervaluesof�ý®H¯��xª ´ areat leastastight asthosereportedabove.
6.6.4 The Effect of Constellation Mapping
We next demonstratethat a large gain canbe achieved via signalmappingsdesignedac-
cordingto thesourcenon-uniformdistribution over Grayandquasi-Graymappings.The
M1 mappingwasintroducedin [186] andwasdesignedfor thetransmissionof non-uniform
108
binarysourcesover singleantennaadditive white Gaussiannoisechannels.It minimizes
the SERunion boundfor singleantennaRayleighfadingchannelswith ± -ary PSK and
squareQAM signaling[217].
Basedonthedesigncriteriaproposedin [186],weconstructanM1-typesignalmapping
for thestar-QAM signalingschemeasshown in Figure6.12.TheSERperformanceof the
M1 andquasi-Graymappingsarecomparedin Figure6.13. It is observed that even for
this small signalingscheme,theCSNRgainof M1 mappingover quasi-Graymappingis
nearly2 dB with MAP decodingat SER= / ª ;<Á . Thegaindueto sourceredundancy when
the sourcedistribution variesbetween0.5 to 0.9 is 7 dB with M1 mapping. Figure6.14
comparesthe SER curves for the Gray andM1 mappingsfor 64-QAM signaling. This
figureshowsthattheM1 mapperformsverywell for MIMO channels.Thegainof theM1
mappingoverGraymappingis 3.7dB at SER= / ª ;<Á . Thegaindueto sourceredundancy
is 10.4dB.
Note that the specificM1 mappingsusedin the above scenariosarenot designedto
minimizetheBERat all � ; they outperformtheGraymappingfor � �pª Æ¡À . This is demon-
stratedfor 64-QAM signalingin Figure6.15, wherethe M1 mappingis shown to have
superiorperformanceto the Gray mappingfor � � ª Æ¡À , but it is outperformedby Gray
mappingfor � �pª ÆGÂ .
109
−10 −5 0 5 10 1510
−5
10−4
10−3
10−2
10−1
100
CSNR in dB
Bit
Err
or R
ate
ML, p = 0.5, AnalysisML, p = 0.5, SimulationMAP, p = 0.8, AnalysisMAP, p = 0.8, SimulationMAP, p = 0.9, AnalysisMAP, p = 0.9, SimulationMAP, p = 0.99, AnalysisMAP, p = 0.99, Simulation
Figure6.1: Resultsfor BPSKsignaling,6 � 0, and Ë � / .
110
−10 −5 0 5 10 1510
−5
10−4
10−3
10−2
10−1
100
CSNR in dB
Bit
Err
or R
ate
ML, p = 0.5, AnalysisML, p = 0.5, SimulationOptimum, p = 0.8, AnalysisOptimum, p = 0.8, SimulationOptimum, p = 0.9, AnalysisOptimum, p = 0.9, SimulationOptimum, p = 0.99, AnalysisOptimum, p = 0.99, Simulation
Figure6.2: Resultsfor theoptimumbinaryantipodalsignaling,6 � 0, and Ë � / .
111
−10 −5 0 5 10 15 20 2510
−6
10−5
10−4
10−3
10−2
10−1
100
CSNR in dB
Bit
Err
or R
ate
ML symmetric, p = 0.5ML optimum, p = 0.9MAP symmetric, p = 0.9MAP optimum, p = 0.9ML optimum, p = 0.99MAP symmetric, p = 0.99MAP optimum, p = 0.99
Figure 6.3: ComparisonbetweenBPSK and optimum signalingschemes,6 � 0, andË � / .
112
−5 0 5 10 15 20
10−4
10−3
10−2
10−1
100
CSNR in dB
Bit
Err
or R
ate
Tandem, 16 statesTandem, 64 statesMAP Decoded
Figure6.4: ComparisonbetweentandemandMAP-decodedschemesfor BPSKsignaling,6ÊÉ 0, and Ë É / .
113
0 5 10 15 20 25 3010
−7
10−6
10−5
10−4
10−3
10−2
10−1
100
Eb/N
0 in dB
Bit
Err
or R
ate
Tandem, 16−stateTandem, 64−stateMAP decoded
Figure6.5: ComparisonbetweentandemandMAP-decodedschemesfor 16-QAM signal-
ing, 6ËÉ 0, and Ì É / .
114
0 5 10 15 20 2510
−3
10−2
10−1
100
CSNR in dB
Sym
bol E
rror
Rat
e ML
MAP
union BoundHunter upper boundsimulationKounias lower bound
Figure6.6: SERcurvesfor 8-PointStarQAM modulationwith quasi-Graymapping,Í ÉÎ, and Ì ÉÐÏ .
115
0 5 10 15 20 2510
−3
10−2
10−1
100
CSNR in dB
Bit
Err
or R
ate
ML
MAP
PEP−based estimateupper boundsimulationlower bound
Figure6.7: BERcurvesfor 8-PointStarQAM modulationwith quasi-Graymapping,ÍËÑÎ, and ÌÒÑ Ï .
116
−5 0 5 10 15 20 2510
−3
10−2
10−1
100
CSNR in dB
Sym
bol E
rror
Rat
e
ML
MAP
union boundHunter upper boundsimulationKounias lower bound
Figure6.8: SERcurvesfor 16-QAM signaling,ÍÊÑ Î, and ÌÒÑ Ï .
117
−5 0 5 10 15 20 25 30 3510
−3
10−2
10−1
100
CSNR in dB
Sym
bol E
rror
Rat
e
MLMAP
union boundHunter upper boundsimulationKounias lower bound
Figure6.9: SERcurvesfor 32-PSKsignaling,ÍËÑ Î, and ÌÒÑ Ï .
118
0 5 10 15 20 25 3010
−3
10−2
10−1
CSNR in dB
Bit
Err
or R
ate
ML
MAP
PEP−based estimateupper boundsimulationlower boundSER/5
Figure6.10:BER curvesfor 32-PSKsignaling,ÍËÑ Î, and ÌÒÑ Ï .
119
−5 0 5 10 15 20 25 3010
−3
10−2
10−1
100
CSNR in dB
Sym
bol E
rror
Rat
e
ML
MAP
union boundHunter upper boundsimulationKounias lower bound
Figure6.11:SERcurvesfor 64-QAM signalingwith Graymapping,ÍËÑ Î, and ÌÒÑ Ï .
120
Ó
Ô
ÕÖØ×[100]
ÕÖ�Ù]×[101]
Õ×[000]
ÕÙ]×[111]
Õ ×ÛÚ[001]
Õ Ù]×�Ú[011]
Õ Ö�×�Ú[010]
Õ Ö�Ù]×ÛÚ[110]
Figure6.12:Thestar-QAM signalingschemewith M1 mapping.
121
0 5 10 15 20 25 3010
−4
10−3
10−2
10−1
100
CSNR in dB
Sym
bol E
rror
Rat
e
quasi−Gray mapping, P(X = 0) = 0.5M1 mapping, P(X = 0) = 0.5quasi−Gray mapping, P(X = 0) = 0.9M1 mapping, P(X = 0) = 0.9
Figure6.13: Comparisonbetweenthe quasi-Grayandthe M1 mappingswith StarQAM
for asystemwith ÍÊÑ Î, and ÌÒÑ Ï .
122
0 5 10 15 20 25 30 35 4010
−4
10−3
10−2
10−1
100
CSNR in dB
Sym
bol E
rror
Rat
e
Gray mapping, P(X = 0) = 0.5M1 mapping, P(X = 0) = 0.5Gray mapping, P(X = 0) = 0.9M1 mapping, P(X = 0) = 0.9
Figure 6.14: Comparisonbetweenthe M1 and Gray mappingsfor 64-QAM signaling,ÍÊÑ Î, and ÌÜÑ Ï .
123
−5 0 5 10 15 20 25 3010
−3
10−2
10−1
100
CSNR in dB
Bit
Err
or R
ate
Gray mapping, P(X = 0) = 0.5M1 mapping, P(X = 0) = 0.5Gray mapping, P(X = 0) = 0.9M1 mapping, P(X = 0) = 0.9
Figure6.15:ComparisonbetweentheBERperformanceof theM1 andGraymappingsfor
64-QAM signaling,ÍÊÑ Î, and ÌÜÑ Ï .
124
Chapter 7
Quantization with Soft-decision
Decoding for MIMO Channels
7.1 Introduction
Likemany othererrorprotectionschemesthataredesignedin thespirit of Shannon’s sep-
arationtheorem[160,161],space-timecodesaredesignedto operateon uniform indepen-
dentandidenticallydistributedbit-streams.As explainedin Subsection1.3.1,Shannon’s
separationtheoremdoesnot take into considerationconstraintson systemcomplexity and
delay. Sincereal-world communicationsystemsareconstrained,systemswith independent
sourceandchannelcodes,known as tandemcoding systems,may have inferior perfor-
mancecomparedwith systemswhichperformsourceandchannelcodingjointly.
In this chapter, we considerthe transmissionof continuous-alphabet(analog)sources
over MIMO channels.Our proposedsystemaddsonly two blocks,with modestcompu-
125
tationalneeds,to a conventionalspace-timecodedsystem.As thesystemmayhave mul-
tiple receive antennas,an importanttask is the properprocessingof the received signals
from thedifferentantennas.Weproposeto addressthisproblemby performingspace-time
soft-decodingfollowedby linear combiningat the receiver. The linear combinerhasthe
following key advantages:i) it hasaverysimplestructure,ii) its designcriterionallowsthe
COVQ index transitionprobabilitiesto bedeterminedin closedform, andiii) its outputis
continuous,makingsoft-decisiondecodingpossible.Inspiredby thework in [6, 146], we
usesoft-decisiondecodingasopposedto soft-decodingmethodssuchasthework in [179],
to exploit the soft informationavailableat the outputof the linear combiner. Our choice
is motivatedby two factors.First, soft-decisiondecodingmaybeimplementedvia a Ý -bit
uniformquantizerat thereceiver(not to beconfusedwith theCOVQ blocksatthetransmit-
terandthereceiver)whichmakesthetaskof decodingcomputationallysimple.In contrast,
thefirst versionof soft-decodingin [179] needsthecomputationof trigonometricfunctions
andmatrix multiplicationandthesecondversionalsorequiresmatrix inversion. Second,
asdiscussedin [6], theperformanceof soft-decisiondecodingcanbecloseto thatof soft-
decodingandrequireslesscomputationalcomplexity (althoughits storagecomplexity may
behigher).
We show that the concatenationof the space-timeencoder, the MIMO channel,the
space-timesoft-decoder, thecombiner, andtheuniform scalarquantizeris equivalentto a
binary-input,Î�Þ
-outputdiscretememorylesschannel(DMC) usedßÆà times,where ß and àarethequantizerdimensionandrate,respectively. Thestep-sizeof theuniform quantizer
usedfor soft-decisiondecodingis numericallyselected,sothatthecapacityof theequiva-
lentDMC is maximizedfor eachvalueof thechannelsignal-to-noiseratio(CSNR).This is
126
aheuristic(sub-optimal)criterion,but asthesimulationresultsof [146] demonstrate,there
is a substantialcorrelationbetweenmaximizingchannelcapacityandminimizing distor-
tion. We show thatthetransitionprobabilitiesof this equivalentDMC canbeexpressedin
termsof thesymbolpairwiseerrorprobabilityof theML-decodedspace-timeorthogonal
blockcodedchannel.Hence,theseprobabilitiescanbedeterminedusing(3.18).
We designthree soft-decisiondecodingCOVQs for the equivalent DMC. The first
COVQ is the classicalCOVQ which assumesthat the index transitionprobabilitiesare
known at both the transmitterandthe receiver. The COVQ codebookis determinediter-
atively usingthemodifiedgeneralizedLloyd algorithm[66]. As theCSNRis not always
availableat thetransmitter, we considerthedesignof two fixed-encoderadaptive-decoder
(FEAD) COVQs. In a FEAD COVQ, the encoderis designedfor a fixed CSNRandthe
decoder, which canestimatethechannelfadingcoefficientsandtheCSNR,adaptsitself to
thechannelconditions.Ourfirst FEAD COVQ usesonly theknowledgeof theCSNRatthe
receiver (asin [180]), while thesecondone,whichwe call theOn-lineFEAD COVQ, em-
ploys alsotheknowledgeof thechannelfadingcoefficientsat thereceiver and,asa result,
it outperformstheFEAD COVQ. An importantfeatureof FEAD COVQ is thatits decoder
codebooksarecomputedin termsof thetransmitterparameters,andnot througha training
processasfor classicalCOVQ. Therefore,this methoddoesnot needa large memoryat
thereceiver to storeadifferentcodebookfor eachvalueof theCSNR.Wedemonstratethat
with a properchoiceof the designCSNR,the performancelossof FEAD COVQ canbe
significantlyreducedascomparedwith theclassicalCOVQ.
127
Figure7.1: Systemblock diagram,whereevery á bits in a ßÆà -bit index â is transmitted
via a space-timecodeword ã°ÑËägå�æªç]è:è1è:çEåbéëê (with orthogonalcolumns å�ìIç�í¼Ñ Ï ç]è:è:è1ç�îUÑï á ), receivedas ðñÑÊäCòóæEç4è1è:è1ç ò%éóê , andspace-timesoft-decodedas ôðjÑÊä ôò æ ç]è:è1è:ç ôò%õ�ê . For
simplicity, wehaveassumedthat áöÑ�ß<à .
7.2 System Components
Thesystemblock diagramis shown in Figure7.1. A COVQ encoderformsa vector ÷ of
dimensionß from theincomingscalarsourcesamples.It thenencodes÷ at a rateof à bits
per sample(bps) into a binary index â of length ßÆà bits, whosebits areeachmappedto
a signal in the BPSK signalingschemeconsistingof øúù æhû and øúùGü û . Encodingis specified
by thedecisionregions ý�þAìgÿ�� �ì���� (�� Ñ Î���
), which form a partitionof the ß -dimensional
space,using the rule that â¨Ññí if ÷� þAì . Index â is then sentover the channeland
receivedasindex � . TheCOVQ decodersimply uses��� , the � th elementin thecodebook,
to reconstruct÷ as �÷ Ñ���� . Thegoal in COVQ designis to minimizetheexpectedvalue
of �ª÷����÷�� ü throughfinding theoptimalpartitionandcodebook.
Our objective is to modeltheconcatenationof theblocksbetweentheencoderandthe
decoderof thevectorquantizerby adiscretechannelandthendesignefficientvectorquan-
tization systemsfor this channel.The systemis designedso that the channeljudiciously
128
incorporatesthe soft-informationof the STOB-codedchannelandadmitsa closed-form
expressionfor its transitionprobabilities. This is achieved by designinga soft-decision
decodingspace-timereceiver which consistsof a space-timesoft detector, followedby a
linear combiner(with optimizedweight coefficients), anda simple Ý -bit scalaruniform
quantizer(whosecell sizeis chosenso the capacityof the equivalentdiscretechannelis
maximized). As a result of the simplememorylessstructureof our receiver anddue to
theorthogonalityof thespace-timecode,we obtainthat theequivalentdiscretechannelis
indeeda binary-inputÎ Þ
-outputmemorylesschannelwhosedistribution canbeeasilyde-
terminedanalyticallyin termsof thesystemparameters.In thefollowing, we describethe
systemcomponentsin detail.
7.2.1 The Channel
From(3.12),theoutputof theSTOB soft-decoderis givenby
ôò���Ñ ï���� �! #" ô$ � çwhere ï � Ñ ï&% ')(+* Í . We rememberthateachentry ô, � - of ôò�� is associatedwith only one
transmittedsymbolin . From(6.3),weknow thatô� �-. i.i.d. / ä10Æç ï2� � êEç 3ÄÑ Ï ç]è:è:è1ç54�ç�6-Ñ Ï ç4è1è:è1ç?á�ç (7.1)
andhencesymbol í canbe detectedby only consideringthe í th entry of the vectors ôò�� ,Ï87 3 7 4 . For our application,this will imply that the bits correspondingto a COVQ
index canbedetectedindependently. Thepairwiseerrorprobabilitybetweena pair of ML
129
decodedSTOB codedBPSKsymbolsø ù ì�û and ø ù � û wasgivenin (3.18)as9 ä1:�ê<;Ñ>=tähø]ù ì�û@? ø]ù � û ê Ñ ACBEDGF�HI:KJ �LNMÑ ÏÎPO Ï � :J Î " : ü<Q õ�R æS ���UT Î ßßWV Ïä Î : ü "YX ê 2Z ç (7.2)
where: Ñ % Î ï�'2([* Í .
7.2.2 Soft-Decision Decoding and the Equivalent DMC
Soft-Decision Decoding: Linear Combining and Scalar Quantization
Many communicationsystemsemploy harddecodingin processingthereceivedsignalsof
space-timecodedsystemsandsodonotexploit thesoft informationavailableat thespace-
time soft-decoderoutputs ôò æ ç]è:è:è1ç ôò õ . Methodsthat exploit soft information can provide
significantperformancegains. In our case,in additionto usingthe soft informationeffi-
ciently, thesolutionshouldallow theCOVQ index transitionprobabilitiesto bedetermined
in closedform, sincethis is requiredfor the COVQ designandencodingphases.As we
illustratebelow andin Section7.4,linearcombininghasbothof theaboveproperties.
This designproblemis a variationof the classicalmaximumratio combining(MRC)
setup[150] in whichthesignalsto becombined(thespace-timesoft-decodedsignals)have
differentnoisevariances(see(7.1)). Letting ô\ � ì ;Ñ ô, � ì * ä ï � � � ê , wecanwrite theoutputof the
linearcombineras(seeFigure7.2)
ô\ ì Ñ õS � �Aæ@] � ô, � ìï � � � Ñ õS � �Aæ^] � ähø ì " ô_ �ì êEç í Ñ Ï ç]è:è1è:ç á�ç (7.3)
130
where ø ì is the í th transmittedsymbolandthe ] � ’s areweightingcoefficientsto bedeter-
mined. From (7.1), we know that the distribution of ô_ �ì , the noisecomponentof ô\ � ì , is/ ä10�ç ï�* ä ï�` ü � � ê?ê . Therefore,theSNRat theoutputof thelinearcombineris
SNRCO Ñ HIa õ� �Aæ ] � L üa õ� �Aæ bb `dc B+e ] ü�gf (7.4)
In linear combining,the objective is thento choosethe weights ý ] � ÿ õ� �Aæ so that SNRCO
is maximizedby enforcingthe constraint a õ� �Aæ ] � Ñ Ï . Therefore,the problemis to
minimizethedenominatorof (7.4), i.e.,findinghjilk õS � �Aæ ïï ` üm� � ] ü� çsubjectto a õ� �Aæ ] � Ñ Ï è Minimizing theLagrangiannpo Ñ S � ïï ` üm� � ] ü� ">q O S � ] � � Ï Z çwe obtainthat q Ñr� Î ï)* ä ï ` ü � ê , and ] � Ñ � � *)� . Therefore,theoutputof thecombiner
canbewritten from (7.3)as ô\ ì Ѩø ì " ô_ ìIç (7.5)
where ô_ ì Ñ a õ� �Aæ ] � ô_ �ì . Wenotethat Ats ô_ üìvu Ñ ä ï � � ê R ü a õ� �Aæ A�w ô_ � üìyx , andtherefore,ô_ ì . / T 0Æç Íï�'2(+� V ç í Ñ Ï ç]è1è:è:ç á�è (7.6)
It is easyto verify that when Ý Ñ Ï (i.e., harddecoding),the concatenationof the linear
combininganddecisionblocksis equivalentto ML decoding.
The linear combineroutput, ô\ ì , is next fed into a “uniform” scalarquantizerwhich
actsasthe soft-decisiondecoder. Let us indicatethe decisionlevels of this quantizerby
131
ýIz ÿ��{R æ � R æ andits codepointsby ý î ÿ��|R æ ��� , where� Ñ Î{Þ
is thenumberof thecodewords.
As ô\ ì cantake any realvalue,thequantizershouldhave two unboundeddecisionregions.
Thedecisionregionsof theuniformquantizeraregivenby
z Ñ }~~~~~� ~~~~~�����ç if ß Ñ�� Ïägß " Ï � � * Î ê�� ç if ß Ñ�0�ç f�f�f ç � � Î" ��ç if ß Ñ � � Ï ç
andthequantizationrule ��ä f ê is simply
��ä ô\ ê-Ñ ß�ç if ô\ Òä�z R æ ç�z u ç ßxÑ�0�ç f�f�f ç � � Ï èTransition Probabilities of the Equivalent DMC
For COVQ design,weneedto derivethetransitionprobabilitiesof the���
-inputÎ Þ ��
-output
discretechannelrepresentedby the concatenationof the space-timeencoder, the MIMO
channel,thespace-timesoft-decoder, thelinearcombiner, andtheuniformquantizer. Since
thedetectionof bits which correspondto eachquantizerindex is decoupled,we notethat
thetransmissionanddecodingof COVQ indicesareindependentof oneanother, andalso
thediscretechannelis equivalentto abinary-inputÎ�Þ
-outputDMC usedß<à times.Weshall
referto this discretechannelasthe“equivalentDMC”.
Therequiredsetof thetransitionprobabilitiesare =täCî �� ø]ù æhû ê and =täCî �� øúùGü û ê for all î ,where ø ù æhû and ø ù�ü û are the BPSK constellationpointswhich, without lossof generality,
areassumedto correspondto 1 and0, respectively. Decisionis madein favor of the ß th
codepointif theoutputof thelinearcombinerfalls into the ä�z R æEç�z u interval of length � .
132
Using(7.5)and(7.6)wecanwrite
=täCî �� ø]ù ì¤û çm� ê Ñ =tä�z R æ 7 ø]ù ì¤û " ô_ ì��>z �� � êÑ F H ä�z R æ��rø ù ì�ûhê�: J � L �PF H ä�z �rø ù ì¤ûgê+: J � L è (7.7)
Theexpectationover � of eachof theabove Ftä f ê functionscanbedeterminedusing(7.2)
to yield
=täCî �� ø ù ì�ûCê�Ñ 9Y��� z R æ��Üø ù ì�û���:���� 9Y��� z �rø ù ì�û���:��Øç (7.8)
where9 ä f ê is definedin (7.2).NotethattheDMC transitionprobabilitymatrixis symmetric
in thesenseof [63].
For our ß -dimensionalCOVQ with rate à shown in Figure7.1, we denotethe natural
binaryrepresentationof the index of decisionregion þAì by ý��5�kÿ ���d�Aæ andthatof codevectorî � by ý��#�kÿ ���d�Aæ , where�#� is abinary Ý -tuple.As theDMC is memoryless,theCOVQ index
transitionprobabilitiescaneasilybecomputedas
=���� �úäl3 � íIê�Ñ ��� ���Aæ = � îC��� � ø ùGü R�� ��û � è (7.9)
7.2.3 The Step-Size of the Uniform Quantizer at the Decoder
Thefinal designparameterof thesystemis � , thestep-sizeof theuniformquantizerat the
receiver. For agivenCSNR,wenumericallyselectthe � whichmaximizesthecapacityof
theequivalentDMC. This is a sub-optimalcriterionsinceour ultimategoal is minimizing
theMSE,not maximizingthecapacity, but asthesimulationresultsof [146] demonstrate,
thereis a strongcorrelationbetweenhaving a high channelcapacityandreducedmean-
squarederrordistortion.
133
Foragivensoft-decisionresolutionÝ andCSNR'2( , wedeterminethestep-size� which
maximizesthe capacityof the DMC by maximizingthe mutual informationbetweenthe
DMC input andoutput. Becausethechanneltransitionprobabilitymatrix is symmetric,a
uniform input distributionachieveschannelcapacity[63]. Notethatthestep-sizedoesnot
dependon therateor dimensionof theCOVQ (usedto quantizethesource),andis only a
functionof Ý andtheCSNR.As atypicalsetof results,we list, in Table7.1,thecapacityof
theequivalentDMC versusthe“optimal” step-sizeof theuniformquantizerfor Alamouti’s
[9] dual transmitsinglereceive setup. Similar resultscanbe derived for systemswith a
differentnumberof transmitantennas,receive antennas,or space-timecodes.As shown
in Figure7.3,whenthestep-sizeis very small (closeto zero)or very large,soft-decision
decodingdoesnotsignificantlyincreasechannelcapacity. Wealsonotethatif theCSNRis
high,soft-decisiondecodingis not verybeneficialin termsof improving channelcapacity.
However, for moderateandlow CSNRs,andwith the optimal choiceof � , soft-decision
decodingsignificantly increaseschannelcapacity. For example,at CSNR Ñ � ÎdB in
Table7.1, thereis a 15% benefitin usingsoft-decisiondecodingwith Ý Ñ� bits. Also
notethat thechannelcapacityincreaseslessthan1% from ÝxÑU¡ to ÝxÑ� evenfor severe
channelconditions. This shows that typically Ý Ñ¢¡ achievesmostof the capacitygain
offeredby soft-decisiondecoding.
134
7.3 Quantization with Soft-Decision Decoding
7.3.1 Soft-Decision Decoding COVQ
Thetransitionprobabilitygivenin (7.9)canbeusedin themodifiedgeneralizedGLA algo-
rithm [66] to designasoft-decisiondecodingCOVQ for space-timecodedMIMO channels
asexplainedbelow. Every input ß -tuple is encodedat a rateof à bits persample.There-
fore, the input spaceis partitionedinto�� Ñ Î���
subsets.As we useBPSK modulation,
a vectorof ß<à real-valuedsignalsis received for every transmittedindex. This vector is
soft-decisiondecodedat a rate of Ý bits per dimension. Therefore,each ß -dimensional
sourcevectoris decodedto oneof the�£ Ñ Î Þ ��
codevectors.Theinputspacepartitioning
andthe codebookareoptimizedbasedon two necessaryconditionsfor optimality using
trainingdata ý{÷��eç5¤ëÑ¥0�ç]è:è1è:çm¦ � Ï ÿ asfollows.
§ Thenearestneighborcondition: for a fixed codebookand í�Ѩ0�ç f�f�f ç �� � Ï , the
optimalpartition ©Wª Ñ ý�þ|ªì ÿ isþ ªì Ѭ«�÷> �¯®5R æS � ��� =���� � ä°3 � í ê+±�äh÷"ç�� � ê 7 �¯®5R æS � ��� =���� � ä°3 � ¤Cê�±�äh÷"ç�� � êEç+²N¤m³ (7.10)
where ´ is a training vector, ýI� � çµ3 Ѷ0�ç]è:è:è1ç �·£ �¹¸�ÿ is the codebook,±�ä1´"ç���ê is
thesquaredEuclideandistancebetween and � , andtiesarebrokenaccordingto a
presetrule.
135
§ Thecentroid condition: givena partition ©ËÑ ý�þAìgç�í Ѻ0�ç]è:è1è:ç �� �»¸�ÿ , theoptimal
codebook¼ ª Ñ ýI� ª� ÿ is
� ª� Ñ � � R æS ì���� =���� �{ä°3 � í ê S��½ ¾ �À¿vÁà´��� � R æS ì���� =���� �úäl3 � íIê � þAì � ç 3 Ñ�0�ç f�f�f ç �·£ �t¸bè (7.11)
where � þAì � is thenumberof thetrainingvectorsin þAì .Trainingconsistsof usingtheaboveconditionsiteratively to updatethecodebookuntil the
averagetrainingdistortiongivenbyn Ñ ¸ß�¦ Ä R æS ����� � ® R æS � ��� =���� � ä°3 � Å ä1´|�1ê ê�±�ä�´��Cç�� � êconverges,whereÅ ä1´�ê is theindex of thepartitioncell to which ´ belongs.
7.3.2 Fixed-Encoder Adaptive-Decoder Soft-Decision Decoding COVQ
Equations(7.2), (7.7)-(7.9)show that the CSNRshouldbe known at both the transmit-
ter andthe receiver to computetheCOVQ index transitionprobabilities.TheCSNRcan
be estimatedat the receiver and then fed back to the transmitter. As the feedbackpath
may not alwaysbe available,it is of particularinterestto considerthe casewhereno in-
formationaboutthechannelstateis availableto thetransmitter. In [180], a fixed-encoder
adaptive-decoder(FEAD) COVQ is proposedwhich addressesthis issuefor a different
setupinvolving hybrid digital-analogSISO transmissionsystems. In the following, we
show how to designa FEAD COVQ for thesoft-decisiondecodedSTOB codedchannel.
In additionto having multiple antennas,our FEAD COVQ differs from the onein [180]
136
in that we have a digital channel(insteadof a hybrid digital-analogchannel).The block
diagramof the FEAD COVQ looks the sameasin Figure7.1. The key differenceis that
heretheencoderpartitionmatchesa “designCSNR”andthereceivercodebookis adapted
to theactualCSNR.Theencoderusesadesigncodebooký�ÆAìhÿ � � R æì���� to find its partitionviaþ ªì Ñ « ´> � � R æS � ��� =���� � ä°3 � í ê+±Aä�´"ç5Æ � ê 7 � � R æS � ��� =���� � äl3 � ¤hê+±�ä1´"ç5Æ � êªç[²N¤ ³ èBecauseweassumecompletelack of informationat thetransmitter, theencodercodebooký�ÆAìhÿ is designedassumingthechannelis harddecoded(i.e, that Ý�ÑU¸ ).
Theproblemis thento adaptthedecodercodebookaccordingto theactualCSNRvalue,
while theencodercodebookremainsfixed. Let usdenotetheencoderindex by â andthe
decoderindex by � . Theaveragedistortionperdimensionis givenbyn Ñ ¸ß AÇwÈ�!´É�P�Ê� ü x Ñ ¸ß ��®mR æS � ��� A�wË�!´É�P�Ê� ü � ��Ñ>3 x =���äl3<êÑ ¸ß � ® R æS � ��� A w �!´É�Ì� � � ü � ��ÑÍ3 x =���ä°3<êªè (7.12)
The goal in quantizerdesignis to derive the � � which minimize (7.12). This gives the
optimal � � in theminimumMSE senseas� ª� Ñ AÎsÏ´ � ��Ñ>3 u Ñ � � R æS ì���� AÎsÏ´ � â Ѩí?çÐ��Ñ>3 u =��!� �ÆäCí � 3<êÑ � � R æS ì����ÒÑ ìÓ=��m� ��äCí � 3<êªç 3ÄÑÔ0�ç f�f�f ç �·£ �t¸�è (7.13)
whereÑ ì Ñ�Õ Áà´@Ö�ä1´Øê+±�´ denotesthemeanof thesamplesin þAì and(7.13)followsbecauseAts×´ � âÄѨí çÐ� ÑY3 u Ñ AÎsÏ´ � âxÑ í u Ñ Ñ ì . Note that the decodercan simply determine=��!� �ÆäCí � 3<ê as = �m� � äCí � 3<ê-Ñ = ��� � ä°3 � í ê+=�� äeí êa � � R æ�d��� = ��� � ä°3 � ¤Cê�=��úä1¤Cê ç (7.14)
137
where =��{äCíIê ÑØ=öäeâ Ñ�í êÛÑU=tä1´Ù þAìkê . Also notethat =�� äeí ê and Ñ ì canbeapproximated
in thetrainingphaseof theencoderasfollows
=��{äeí ê{Ú � þAì �¦ ç Ñ ì@Ú ¸� þAì � S��½ ¾ �l¿vÁà´|�eè (7.15)
From (7.13), we observe that unlike COVQ, given the encodermeans ý Ñ ìhÿ � � R æì���� , the
FEAD-COVQ doesnot requirea training phaseto determineýI� � ÿ . In otherwords, the
decodercodebookis computedfrom the channeltransitionprobabilitiesandthe encoder
meansvia (7.13)and(7.14).
7.3.3 On-line FEAD Soft-Decision Decoding COVQ
Oneof our assumptionsis that the receiver hasperfectknowledgeof the channelfading
coefficientswithout error. In light of this assumption,we observe in equation(7.5) that
theoutputof the linearcombinerhasan identicalform to theoutputof anadditive white
Gaussiannoisechannelwith the varianceof the additive noise ô_ ì known at the receiver
andgivenby (7.6). This motivatesusto applysoft-decisiondecodingdirectly on ô\ ì using
thestep-size� derived in [146, SectionII]) for AWGN channels.Thechanneltransition
probabilities,giventhepathgainsmatrix � , aregivenby
=öäeî �� ø ù � û çm� ê Ñ =tä�z R æ 7 ø ù � û " ô_ ì��Ùz �� � êÑ F H äÛz R æ��rø ù � ûCê+: J � L �EF H ä�z �rø ù � ûCê�: J � L çNote that the above derivation is valid if the channelfadingcoefficientsremainconstant
duringthetransmissiontimeof anindex.
138
7.4 Numerical Results and Discussion
7.4.1 Implementation Issues
Weconsiderthetransmissionof zero-meanunit-variancei.i.d. GaussianandGauss-Markov
sourcesover MIMO channels.500,000trainingvectorsand850,000testvectorsareem-
ployed. Eachtestis performed5 timesandtheaveragesignal-to-distortionratio (SDR)in
dB is reported.MIMO systemswith Ü transmitand 4 receive antennasarereferredto asä�Ü - 4�ê systems.Alamouti’scode[9] is usedfor thedualtransmit-antennasystems.Thereal
(rate1) codeof [192] is employedfor thequad-transmitsystembecauseour constellation
is real.
Several training strategieswere examined,and the bestone in termsof having con-
sistentresultsandhigh training SDR wasusedasfollows. For any given COVQ rate à ,
dimensionß , andnumberof soft-decisiondecodingbits Ý , we first train a ß -dimensional
rate-Ý�à VQ with theSplit algorithm[66]. We next usetheSimulatedAnnealingalgorithm
[49] which aimsto minimize the averageend-to-enddistortionfor a givenVQ codebook
throughoptimizingtheassignmentof indicesof theVQ codevectors.Theaverageoverall
distortioncanbewrittenasn Ñ ¸ß AÎs×±�ä1´"ç���ê u Ñ ¸ß � � R æS ì���� ��®5R æS � ��� AÎsϱ�ä1´"ç���ê � â ѧí çÐ��Ñ>3 u =���Ý ��äCâÄѨí çÃ� ÑY3<êÑ ¸ß S ì =��{äCí ê S � =���� � äl3 � í ê¯Þ Áà�!´É�ß�&à ù � û�� ü Ö�ä1´�ê+±�´Ñ ¸ß A�wÈ�!´�� ü x " ¸ß S ì =��{äeí ê S � =���� �úäl3 � íIêÐá1�@à ù � ûPçúä��@à ù � ûâ�Yã Ñ ìkê�ä ç (7.16)
where áÈå[ç�æ�äLÑ a ì ] ìÓç�ì is the standardinner productand èé ý�0�ç f�f�f ç �£ �ê¸�ÿ ?139
ý�0�ç f�f�f ç �£ �»¸�ÿ is the one-to-onemappingfunction to be optimized. It follows that the
costfunctionto beminimizedequals� � R æS ì���� =��{äeí ê � ® R æS � ��� =���� �{äl3 � íIêÐá1�ëà ù � ûPç ä��Nà ù � ûN�Eã Ñ ìeê�ä èSimulatedAnnealingis usedonly at thehighestCSNR.We thenuseanapproachsimilar
to the one in [50]; namely, we usethe modified generalizedLloyd algorithm to derive
theCOVQ codebooksstartingfrom thehighestCSNRto the lowestandvice-versa.This
methodis referredto asthe “decrease-increase”(DI) method.Anotherway to obtainthe
codebookscouldbestartingfrom thehighestCSNRdown to thelowest,whichwereferto
asthe“decreasing”method.
Table7.2 comparesthe resultsof the DI anddescendingmethodsfor variousSTOB
codedsystems.It is observedthattheDI methodis mostlybeneficialat low CSNRvalues.
This is becauseat low CSNRsomeencodercells areempty. Empirical resultsshow that
thesecellsareoptimizedmoreefficiently throughtheCSNR-increasingloop.
7.4.2 COVQ for Various MIMO Channels
Figure7.4 plots the SDR curvesof variousCOVQ-basedspace-timecodedsystemsasa
function of the CSNR.Even at the low COVQ dimensionand rateconsideredhere,the
gainof usingMIMO channelsover theSISOchannelis obvious. For example,at SDR=
5 dB, the (2-1) systemoutperformsthe SISOsystemby 6 dB (for hard-decoding)andis
outperformedby the(2-2)systemby 4.3dB.Thisfigurealsodemonstratestheeffectiveness
of our linearcombiner. Notethatasthesignalpowercollectedby the(2-2)systemis twice
140
thatof the(4-1)system,theformerhasabetterperformance,althoughthediversitygainof
bothof thesystemsis thesameandequalsÜ84ÒÑ X .We observe in Table 7.1 that soft-decisiondecodingbecomeslessbeneficialas the
CSNRgrows. Increasingthe numberof transmitor receive antennasresultsin a CSNR
gain dueto spacediversity. Therefore,we expectthat increasingthe numberof transmit
or receive antennaswould leave little roomfor furtherenhancementthroughsoft-decision
decoding. It follows from (7.6) that betweentwo systemswith the samediversity gain,
codinggain ï , andCSNR '2( , theonewith fewer transmitantennashasa higherSNRat its
linearcombineroutput.In otherwords,giventwo systemswith thesamediversitygain,the
onewith moretransmitantennasobtainsa largersoft-decisiondecodinggain. This result
canbe statedmore intuitively: systemswith more receive branchescollect more signal
power. Hence,the SNR at their linear combinerwould be higher, makingsoft-decision
decodinglesseffective. This observationis supportedby thesimulationsof Figure7.4: at
CSNR= 4 dB, thesoft-decisiondecodinggainin SDRis 0.29dB for bothof the(2-1)and
(4-1) systems;this gainreducesto 0.12dB for the(2-2) system.
7.4.3 COVQ versus Tandem Coding
Wenext compareour COVQ-basedsystemwith traditionaltandemcodingschemeswhich
useseparatesourcecodingandchannelcodingblockswith VQ andconvolutionalcoding
(CC), respectively. We consider, in Figure7.5, a (2-1) systemandquantizationwith di-
mensionß Ñìã . The overall rateis 3.0 bpsandhencetherearesix choicesfor the (VQ,
CC) coderates;namely, (0.5,1* 6), (1.0 , 1* 3), (1.5,1* 2), (2.0,2* 3), (2.5 , 5* 6), and(3.0,
141
0). Thefirst four convolutionalcodeshave 64 statesandarenon-systematicwith freedis-
tancesof 27 [116], 14, 10, and5 [101]. For rate5* 6, we usea rate-compatiblepunctured
convolutional (RCPC)codewith a rate-1* 2 mothercode. The generatorpolynomialsof
therate-1* 6, 1* 3, 1* 2, and2* 3 convolutionalcodesaregivenby (754,644,564,564,714,
574),(574,664,744),(634,564),and(3, 4, 5; 4 3 7). They arethestrongestcodesgiven
in [101, 150] for thegivennumberof states.Thegeneratorpolynomialsaredefinedasin
[101]. Thegeneratorpolynomialsof themothercodefor therate-5* 6 codeare(554,744)
andits puncturingmatrix is givenby� æ æ�� æ��æ�� æ�� æ � [150]. Figure7.5 shows a typical behavior:
the jointly designedCOVQ outperformsthe substantiallymorecomplex tandemsystems
almosteverywhere. Furthertestswith i.i.d. sourcesyield even more supportive results
towardsCOVQ.
Notethatif oneaimsto designanunequalerrorprotection(UEP)joint source-channel
coderwith the above separatecoders(i.e., selectthe besttandemcoderat eachCSNR),
oneneedsto designan algorithmto allocatethe sourceandchannelcoderatesfor each
givenCSNR,thusincreasingthecomplexity of theUEPsystem.COVQ doesnothave this
problemsinceerrorprotectionin COVQ is built-in.
7.4.4 COVQ, FEAD COVQ, and On-line FEAD COVQ
Figure 7.6 demonstratesthe performanceof a (2-1) systemquantizingan i.i.d. / ä10�çv¸úêsourcewith variousrate-2bpsFEAD COVQs with dimensionß Ñrã . A FEADí COVQ
is onewhosedesignCSNRis ' dB. TheFEAD VQ assumesthatthechannelis noiseless;
henceit hasalowercomputationalcomplexity attheencoder. Thefigureshowsthatsuchan
142
assumptionwill leadto a significantSDRlossat low to mediumCSNRs.FEAD� COVQ
alsosuffers from a considerableperformancedegradationat high CSNRs. It seemsthat
assumingamid-rangeCSNRof 8 dB (for thegivenMIMO system)will leadto reasonable
performanceeverywhere.
Thethreequantizerspresentedin thischapterarecomparedin Figure7.7,whereaunit-
varianceGauss-Markov sourceis quantizedwith dimension2 andrate1 bpsandsentover
a(2-1)system.TheFEAD VQsaredesignedassuminganoiselesschannel,hencetheSDR
of the threesystemsbecomecloserastheCSNRgrows becausethechannelmismatchof
the VQs decreases.The On-line FEAD VQ maintainsits gain over the FEAD VQ when
soft-decisiondecodingis employed. It is alsoobserved that althoughthe On-line FEAD
VQ encoderassumesthe channelis noiseless(CSNR ? � ), it slightly outperformsthe
COVQ, which is designedfor theexactCSNR,at highCSNRs.
143
îï æì Ó�ðÏñ òIólô ��õ æ+ö R æ Óî÷ æì ðøñ ò5ù æîï üì Ó�ð ñ ò ólô ��õ ü5ö R æ Óî÷ üì ð ñ ò ù ü...îï õì Ó�ð ñ ò ólô ��ú õ ö R æ Óî÷ õì ð ñ ò ù õ
ð û òÓ Ô î÷ ìýüEþ õ� �Aæ ù � óGÿ ì û î� �ì ö Óîï æì
òîï üì Ó...îï õì Ôð û Ó�ð ñ òólô � õ ö R æî÷ ì¯ü ÿ ì û î� ì Ó
Figure 7.2: Linear combinerwhich follows the space-timesoft decoderat the receiver.
Left: problemstatement,right: solutionto theproblem.
144
CSNR � ü�� � ü�� � ü Ù � ü�� � ü�(dB) � � � �
16.0 0.9945 0.9969 0.414 0.9973 0.219 0.9974 0.183 0.9974 0.108
14.0 0.9881 0.9929 0.403 0.9937 0.214 0.9939 0.109 0.9940 0.109
12.0 0.9752 0.9842 0.394 0.9858 0.209 0.9862 0.107 0.9864 0.111
10.0 0.9506 0.9660 0.390 0.9695 0.207 0.9702 0.107 0.9705 0.113
8.0 0.9070 0.9333 0.393 0.9381 0.208 0.9392 0.108 0.9397 0.116
6.0 0.8374 0.8760 0.405 0.8833 0.214 0.8849 0.111 0.8855 0.120
4.0 0.7386 0.7891 0.430 0.7987 0.226 0.8009 0.118 0.8015 0.128
2.0 0.6164 0.6741 0.472 0.6852 0.248 0.6877 0.129 0.6884 0.139
0.0 0.4849 0.5427 0.536 0.5540 0.280 0.5565 0.146 0.5572 0.157
-2.0 0.3608 0.4142 0.627 0.4223 0.327 0.4246 0.170 0.4252 0.183
-4.0 0.2560 0.2974 0.751 0.3057 0.390 0.3076 0.203 0.3081 0.219
Table 7.1: Capacity(in bits/channeluse)of the DMC derived from � -bit soft-decision
decodingof BPSK-modulatedspace-timecodedMIMO channelwith Ü �ã and �� ¸ .� is thestep-sizewhichmaximizesthecapacityand � is numberof soft-decisiondecoding
bits.
145
0 0.1 0.2 0.3 0.4 0.6 0.8 1 1.40.4
0.5
0.6
0.7
0.8
0.9
1
∆
Cap
acity
(bi
ts/c
hann
el u
se)
CSNR = 0 dB
CSNR = 4 dB
CSNR = 10 dB
q = 4q = 3q = 2q = 1
Figure7.3: DMC capacityversusthestep-sizeof theuniformquantizer, �� �� , � �� .
146
CSNR (2-1) (4-1) (2-2)
in dB DI decreasing DI decreasing DI decreasing
10 7.385 7.368 7.809 7.795 7.899 7.876
8 6.886 6.845 7.551 7.536 7.867 7.839
6 6.119 6.026 6.949 6.908 7.712 7.691
4 5.167 4.968 5.948 5.832 7.288 7.272
2 4.178 3.830 4.781 4.508 6.477 6.416
0 3.315 2.783 3.714 3.245 5.353 5.179
-1 2.916 2.329 3.266 2.700 4.772 4.508
-2 2.546 1.930 2.830 2.224 4.217 3.853
-3 2.202 1.578 2.436 1.818 3.713 3.245
Table7.2: ComparisonbetweenthetrainingSDR(in dB) of two COVQ trainingmethods
for a unit-varianceGauss-Markov source( �� ������ ) channel-optimizedvectorquantizedat
rate �� �!��� bps.Quantizationdimensionis "# $� . ThreeMIMO systemsareconsidered
with %&� - �(') *%&� - �+' , %-, - �.' , and %&� - �!' .
147
−3 −2 0 2 4 6 8 10
1
2
3
4
5
6
7
8
CSNR in dB
SD
R in
dB
K=2, L=2
K=4, L=1K=2, L=1
K=1, L=1
COVQ, q = 3COVQ, q = 2COVQ, q = 1
Figure 7.4: SimulatedSDR in dB for a zero-meanunit-varianceGauss-Markov source
( �� ����� ) vector quantizedat rate �/ �!��� bps and soft-decisiondecodedwith � bits.
Quantizationdimensionis "0 �� .
148
−2 0 2 4 6 8 10 12 14 16 18 20−5
0
5
10
15
20
CSNR in dB
SD
R in
dB
COVQVQ onlyTandem, 5/6Tandem, 2/3Tandem, 1/2Tandem 1/3Tandem 1/6
Figure7.5: Jointly designedversustandemcodingschemesfor a zero-meanunit-variance
Gauss-Markov source( �# 1���2� ). Quantizationdimensionis "3 � andtheoverall rateis
�4 65��2� bps. �7 �� and � �� .
149
−3 −2 0 2 4 6 8 10
1
2
3
4
5
6
7
8
9
CSNR in dB
SD
R in
dB
FEAD0 COVQ
FEAD VQ
FEAD8 COVQ
FEAD (CO)VQ, q = 3FEAD (CO)VQ, q = 2FEAD (CO)VQ, q = 1
Figure 7.6: SimulatedSDR in dB for an i.i.d. 81%9��:;�.' sourcevector quantizedat rate
�< �=��� bpsandsoft-decodedwith � bits. Quantizationdimensionis "> � . � � and
� �� .
150
−3 −2 −1 0 1 2 3 4 5 6 7 8
2
3
4
5
6
7
CSNR in dB
SD
R in
dB
COVQ, q = 3On−line FEAD VQ, q = 3FEAD VQ, q = 3On−line FEAD VQ (q = 1)FEAD VQ (q = 1)
Figure 7.7: Comparisonamongthe COVQ, FEAD VQ, and On-line FEAD VQ for a
8*%&��:?�.' Gauss-Markov sourcevector quantizedat rate �@ �!�2� bps and hard decoded.
Quantizationdimensionis "0 �� , �A �� , and � $� .
151
Chapter 8
Summary
In thisthesis,westudiedtheanalysisof theerrorratesof codedSISOchannelsanduncoded
MIMO channelswith both uniform andnon-uniformsources.We alsodesigneda soft-
decisiondecodingCOVQ for the transmissionof analogsourcesover MIMO channels.
Thecontributionsof this thesisaresummarizedin thefollowing.
B Wepresentedanalternativederivationof thepairwiseerrorprobabilityof space-time
codesof arbitrarystructureunderquasi-staticfading.Ourmethod,which is basedon
a momentgeneratingfunctionapproach,is very simpleandcanbeappliedto many
scenariosin singleor multiple antennasystems.It canalsobespecializedto exploit
thestructureof certaincodesto simplify thederivation,aswasdonein [176].
B We establishedupperandlower Bonferroni-typeboundson the block andbit error
ratesof single-antennacommunicationsystems. The KAT lower boundwas also
studied.TheAWGN aswell astheblockRayleighfadingchannelswereconsidered.
152
We derivedclosed-formformulasfor thepairwiseerrorprobabilityandalsofor the
probability of the intersectionof two pairwiseerror events. Thesequantitiesare
requiredfor thecomputationof theBonferroniandKAT bounds.We alsoproposed
to usea subsetof thecodebookandsuggesteda goodinitial setfor thecomputation
of theboundsfor thecasewherethenumberof bits inputto thechannelcoderis large
(thatis, for large " , where�;CD 6"FEHG is thechannelcoderate).Theboundswereoften
identicalup to four significantdigits for theAWGN channel.
B For MIMO channelswith auniformi.i.d. binaryinput,weconsideredthecasewhere
space-timeorthogonalblock coding is usedand establishedtight Bonferroni-type
upperandlowerboundson thesymbolandbit errorratesof thesystem.We derived
theprobabilityof the intersectionof two pairwiseerroreventsin closedform. The
bounds,whichcanbeappliedto any kind of signalinggeometryandarbitrarysymbol
mapping,areoftenthesameupto six significantdigits. In thecaseof PSKsignaling
(with any size),theboundsareidenticalup to tensignificantdigits evenat negative
channelsignal-to-noiseratios.
B We developeda new approachto derive the symbol pairwiseerror probability of
space-timeorthogonalblock codesunderMAP detectionfor non-uniforminputs.
With the pairwiseerror probability in hand,we computedtight upperand lower
boundson the symbol and bit error ratesof the system. It was shown that gains
up to 10 dB canbeobtainedvia MAP decodingandusingsymbolmappingsbetter
thanGraycoding.Furthermore,wederivedtheoptimalbinarysignalingschemeand
demonstratedthat very large gainsexist in usingthis schemeinsteadof BPSK sig-
153
naling. Moreover, it wasshown thatMAP decodingcansignificantlyoutperforma
typical tandemcodingsystemwhich is muchmorecomplex. In fact, it wasshown
thatMAP decodingis superiorto tandemcodingfor thebit errorratesof interest.
B In additionto MAP decoding,which is a form of joint source-channelcoding,we
consideredthe casewherethe input to the MIMO systemis a continuous-alphabet
(analog)signal. We designeda soft-decisiondecodingCOVQ for robust compres-
sionandcommunicationof theanalogsource.Wealsoconsideredthecasewherethe
transmitterdoesnothave thechannelsignal-to-noiseratioanddesignedtwo COVQs
for the systemoneof which exploits the knowledgeof the channelgainsat the re-
ceiver to gain betterperformance.Comparisonswith typical tandemsystemswere
alsopresentedand it wasshown that the COVQ cannot only outperformtandem
systems,but also it can beatunequalerror protectionsystemswhich useanother
approachto joint source-channelcoding.
154
Bibliography
[1] M. Abramowitz andI. Stegun,Handbookof MathematicalFunctionswith Formulas,
Graphs,and MathematicalTables. New York, NY: Dover Publications,fifth ed.,
1966.
[2] A. Abdi and M. Kaveh, “A space-timecorrelationmodel for multielementan-
tennasystemsin mobilefadingchannels,” IEEE J. Select.AreasCommun., vol. 20,
pp.550-560,Apr. 2002.
[3] R. K. Ahuja,T. L. Magnanti,andJ.B. Orlin, NetworkFlows.EnglewoodClif fs, NJ:
PrenticeHall, 1993.
[4] F. Alajaji, S.A. Al-Semari,andP. Burlina,“Visualcommunicationvia trellis coding
andtransmissionenergy allocation,” IEEETrans.Commun., vol. 47,pp.1722-1728,
Nov. 1999.
[5] F. Alajaji and T. Fuja, “A communicationchannelmodeledon contagion,” IEEE
Trans.Inform.Theory, vol. 40,pp.2035-2041,Nov. 1994.
155
[6] F. Alajaji andN. C. Phamdo,“Soft-decisionCOVQ for Rayleigh-fadingchannels,”
IEEE Commun.Let., vol. 2, pp.162-164,June1998.
[7] F. Alajaji, N. Phamdo,N. Farvardin, and T. Fuja, “Detection of binary Markov
sourcesover channelswith additive Markov noise,” IEEE Trans. Inform. Theory,
vol. 42,pp.230-239,Jan.1996.
[8] F. Alajaji, N. PhamdoandT. Fuja, “Channelcodesthatexploit the residualredun-
dancy in CELP-encodedspeech,” IEEE Trans.Speech Audio Processing, vol. 4,
pp.325-336,Sept.1996.
[9] S. M. Alamouti, “A simple transmitdiversity techniquefor wirelesscommunica-
tions,” IEEE J. Select.AreasCommun., vol. 16,pp.1451-1458,Oct.1998.
[10] S. M. Alamouti, V. Tarokh, and P. Poon,“Trellis-codedmodulationand transmit
diversity: designcriteria and performanceevaluation,” in Proc. IEEE Int’l Conf.
UniversalPersonalCommun., Florence,Italy, Oct.5-9,1998,pp.703-707.
[11] N. Al-Dhahir, “Overview andcomparisonof equalizationschemesfor space-time-
codedsignalswith applicationto EDGE,” IEEE Trans.SignalProcessing, vol. 50,
pp.2477-2488,Oct.2002.
[12] N. Al-Dhahir, “A new high-ratedifferentialspace-timeblockcodingscheme,” IEEE
Commun.Let., vol. 7, pp.540-542,Nov. 2003.
[13] S. Al-Semari,F. Alajaji, andT. Fuja,“SequenceMAP decodingof trellis codesfor
Gaussianand Rayleighchannels,” IEEE Trans.Veh. Technol., vol. 48, pp. 1130-
1140,July1999.
156
[14] E. AyanogluandR. M. Gray, “The designof joint sourceandchanneltrellis wave-
form coders,” IEEE Trans.Inform.Theory, vol. 33,pp.855-865,Nov. 1987.
[15] I. Bahceci,andT. M. Duman,“Combinedturbocodingandunitaryspace-timemod-
ulation,” IEEE Trans.Commun., vol. 50,pp.1244-1249,Aug. 2002.
[16] P. BalabanandJ. Salz, “Optimum diversity combiningandequalizationin digital
datatransmissionwith applicationto cellularmobileradio,” IEEETrans.Veh.Tech-
nol., vol. 40,pp.342-354,May. 1991.
[17] I. Barhumi,G. Leus,andM. Moonen,“Optimal trainingdesignfor MIMO OFDM
systemsin mobile wirelesschannels,” IEEE Trans. Signal Processing, vol. 51,
pp.1615-1624,June2003.
[18] S.Baro,G. Bauch,andA. Hansmann,“Improvedcodesfor space-timetrellis-coded
modulation,” IEEE Commun.Let., vol. 4, pp.20-22,Jan.2000.
[19] G. Bauch,J. Hagenauer, andN. Seshadri,“Turbo processingin transmitantenna
diversitysystems,” AnnalsTelecommun., vol. 56,pp.455-471,Aug. 2001.
[20] F. Behnamfar, F. Alajaji and T. Linder, “Progressive imagecommunicationover
binarychannelswith additiveburstynoise,” in Proc.IEEEDataCompressionConf.,
Snowbird, UT, Apr. 2002.
[21] F. Behnamfar, F. Alajaji andT. Linder, “Imagecodingfor binaryburstynoisechan-
nels,” in Proc.21st BiennialSymp.Commun., Kingston,ON, June2002.
157
[22] F. Behnamfar, F. Alajaji, andT. Linder, “Imagetransmissionover thePolyachannel
via channel-optimizedquantization,” IEEE Trans.SignalProcessing, to appear.
[23] R. S. Blum, “Some analytical tools for the designof space-timeconvolutional
codes,” IEEE Trans.Commun., vol. 50,pp.1593-1599,Oct.2002.
[24] C. Budianu and L. Tong, “Channel estimationfor space-timeorthogonalblock
codes,” IEEE Trans.SignalProcessing, vol. 50,pp.2515-2528,Oct.2002.
[25] Q. ChenandT. R. Fisher, “Imagecodingusingrobustquantizationfor noisydigital
transmission,” IEEE Trans.ImageProcessing, vol. 7, pp.496-505,Apr. 1998.
[26] M. Chiani, M. Z. Win, and A. Zanella, “On the capacityof spatially correlated
MIMO Rayleigh-fadingchannels,” IEEE Trans.Inform. Theory, vol. 49, pp. 2363-
2371,Oct.2003.
[27] D. Chizhik, G. J. Foschini,M. J. Gans,andR. A. Valenzuela,“K eyholes,correla-
tions, andcapacitiesof multielementtransmitandreceive antennas,” IEEE Trans.
WirelessCommun., vol. 1, pp.361-368,Apr. 2002.
[28] D. Chizhik, J. Ling, P. W. Wolniansky, R. A. Valenzuela,N. Costa,andK. Huber,
“Multiple-input-multiple-outputmeasurementsandmodelingin Manhattan,” IEEE
J. Select.AreasCommun., vol. 21,pp.321-331,Apr. 2003.
[29] S.-Y. Chung,J.G.D. Forney, T. Richardson,andR.Urbanke,“On thedesignof low-
densityparity-checkcodeswithin 0.0045dB of theShannonlimit,” IEEE Commun.
Lett., vol. 5, pp.58-60,Feb. 2001.
158
[30] A. CohenandN. Merhav, “Lower boundson the error probability of block codes
basedon improvementson de Caen’s inequality,” IEEE Trans. Inform. Theory,
vol. 50,pp.290-310,Feb. 2004.
[31] T. M. Cover andJ. A. Thomas,Elementsof InformationTheory. New York: Wi-
ley, 1991.
[32] C. CozzoandB. L. Hughes,“Joint channelestimationanddatadetectionin space-
timecommunications,” IEEE Trans.Commun., vol. 51,pp.1266-1270,Aug. 2003.
[33] J.W. Craig,“A new, simple,andexactresultfor calculatingtheprobabilityof error
for two-dimensionalsignalconstellations,” in Proc. IEEE Military Commun.Conf.
(MilCom), McLean,VA, Nov. 1991,pp.571-575.
[34] D. Cui andA. M. Haimovich, “A new bandwidthefficient antennadiversityscheme
using turbo codes,” in Proc. 34th Annual Conf. Inform. Sci. Syst., Princeton,NJ,
Mar. 2000,vol. 1, pp.24-29.
[35] M. O. Damen,K. Abed-Meraim,andJ.C. Belfiore,“Diagonalalgebraicspace-time
block codes,” IEEETrans.Inform.Theory, vol. 48,pp.628-636,Mar. 2002.
[36] M. O. DamenandN. C. Beaulieu,“On two high-ratealgebraicspace-timecodes,”
IEEE Trans.Inform.Theory, vol. 49,pp.1059-1063,Apr. 2003.
[37] M. O. Damen,A. Chkeif, andJ. C. Belfiore, “Lattice codedecoderfor space-time
codes,” IEEE Commun.Let., vol. 4, pp.161-163,May 2000.
159
[38] M. O. Damen,H. El Gamal,andN. C. Beaulieu,“Linear threadedalgebraicspace-
timeconstellations,” IEEETrans.Inform.Theory, vol. 49,pp.2372-2388,Oct.2003.
[39] D. deCaen,“A lower boundon theprobabilityof a union,” Discr. Math., vol. 169,
pp.217-220,1997.
[40] A. Dembo,unpublishednotes,communicatedby I. Sason,2000.
[41] A. Dogandzic, “Chernoff boundson pairwise error probabilities of space-time
codes,” IEEE Trans.Inform.Theory, vol. 49,pp.1327-1336,May 2003.
[42] T. G. Donnelly, “Bi variatenormaldistribution,” Commun.ACM, vol. 16, pp. 636,
1973,Algorithm 462.
[43] J. G. DunhamandR. M. Gray, “Joint sourceandnoisy channelencoding,” IEEE
Trans.Inform.Theory, vol. 27,pp.516-519,July1981.
[44] H. El GamalandM. O.Damen,“Universalspace-timecoding,” IEEETrans.Inform.
Theory, vol. 49,pp.1097-1119,May 2003.
[45] H. El Gamaland A. R. Hammons,“A new approachto layeredspace-timecod-
ing and signal processing,” IEEE Trans. Inform. Theory, vol. 47, pp. 2321-2334,
Sept.2001.
[46] H. El GamalandA. R. Hammons,“On the designandperformanceof algebraic
space-timecodesfor BPSKandQPSKmodulation,” IEEETrans.Commun., vol. 50,
pp.907-913,June2002.
160
[47] H. El GamalandA. R. Hammons,“On thedesignof algebraicspace-timecodesfor
MIMO block-fadingchannels,” IEEE Trans.Inform. Theory, vol. 49, pp. 151-163,
Jan.2003.
[48] E. O. Elliott, “Estimatesof errorratesfor codeson burst-noisechannels,” Bell Syst.
Tech. J., vol. 42,pp.1977-1997,Sept.1963.
[49] N. Farvardin, “A studyof vectorquantizationfor noisychannels,” IEEE Trans.In-
form.Theory, vol. 36,pp.799-809,July1990.
[50] N. Farvardin andV. Vaishampayan,“Optimal quantizerdesignfor noisy channels:
an approachto combinedsource-channelcoding,” IEEE Trans. Inform. Theory,
vol. 33,pp.827-838,Nov. 1987.
[51] N. FarvardinandV. Vaishampayan,“On theperformanceandcomplexity of channel
optimizedvector quantizers,” IEEE Trans. Inform. Theory, vol. 37, pp. 155-160,
Jan.1991.
[52] T. FazelandT. Fuja, “Robust transmissionof MELP-compressedspeech:an illus-
trative exampleof joint source-channeldecoding,” IEEE Trans.Commun., vol. 51,
pp.973-982,June2003.
[53] G. Femenias,“BER performanceof linear STBC from orthogonaldesignsover
MIMO correlatedNakagami-mfadingchannels,” IEEETrans.Veh.Technol., vol. 53,
pp.307-317,Mar. 2004.
[54] T. Fine, “Propertiesof an optimal digital systemand applications,” IEEE Trans.
Inform.Theory, vol. 10,pp.443-457,Oct.1964.
161
[55] G. D. Forney, Jr., “Geometricallyuniform codes,” IEEE Trans. Inform. Theory,
vol. 37,pp.1241-1260,Sept.1991.
[56] G. J.Foschini,“Layeredspace-timearchitecturefor wirelesscommunicationovera
fadingenvironmentwhenusingmultipleantennas,” Bell LabsTech.J., vol. 1, pp.41-
59,Autumn1996.
[57] G. J. Foschini and M. J. Gans,“On limits of wirelesscommunicationsin a fad-
ing environmentwhenusingmultiple antennas,” WirelessPers. Commun., vol. 6,
pp.311-335,Mar. 1998.
[58] G. J. Foschini,G. D. Golden,R. A. Valenzuela,andP. W. Wolniansky, “Simplified
processingfor high spectralefficiency wirelesscommunicationemploying multi-
elementarrays,” IEEEJ. Select.AreasCommun., vol. 17,pp.1841-1852,Nov. 1999.
[59] M. Fozunbal,S. W. McLaughlin, and R. W. Schafer, “On performancelimits of
space-timecodes:a sphere-packingboundapproach,” IEEE Trans.Inform.Theory,
vol. 49,pp.2681-2687,Oct.2003.
[60] S. GadkariandK. Rose,“Robust vectorquantizerdesignby noisy channelrelax-
ation,” IEEETrans.Commun., vol. 47,pp.1113-1116,Aug. 1999.
[61] S. GadkariandK. Rose,“Correctionsto “robust vectorquantizerdesignby noisy
channelrelaxation”,” IEEETrans.Commun., vol. 48,pp.176-176,Jan.2000.
[62] J. GalambosandI. Simonelli,Bonferroni-TypeInequalitieswith Appications.New
York: Springer-Verlag,1996.
162
[63] R.G.Gallager, InformationTheoryandReliableCommunication.New York: Wiley,
1968.
[64] R. G. Gallager, “ClaudeE. Shannon:a retrospectiveon his life, work, andimpact,”
IEEE Trans.Inform.Theory, vol. 47,pp.2681-2695,Nov. 2001.
[65] D. Gesbert,H. Bolcskei, D. A. Gore, and A. J. Paulraj, “Outdoor MIMO wire-
lesschannels:modelsandperformanceprediction,” IEEETrans.Commun., vol. 50,
pp.1926-1934,Dec.2002.
[66] A. Gersho,andR. M. Gray, Vector Quantizationand signal compression. Kluwer
AcademicPublishers,1992.
[67] M. Gharavi-Alkhansari and A. B. Gershman,“Fast antennasubsetselectionin
MIMO systems,” IEEE Trans.SignalProcessing, vol. 52,pp.339-347,Feb. 2004.
[68] E. N. Gilbert, “Capacity of a burst-noisechannel,” Bell Syst.Tech. J., vol. 39,
pp.1253-1265,Sept.1960.
[69] A. Gorokhov, D. Gore, and A. J. Paulraj, “Receive antennaselectionfor MIMO
flat-fadingchannels:theoryandalgorithms,” IEEE Trans.Inform. Theory, vol. 49,
pp.2687-2696,Oct.2003.
[70] L. GouletandH. Leib, “Serially concatenatedspace-timecodeswith iterativedecod-
ing andperformancelimits of block-fadingchannels,” IEEE J. Select.AreasCom-
mun., vol. 21,pp.765-773,June2003.
163
[71] I. GradsteinandI. Ryshik,Tablesof Integrals,Series,andProducts. SanDiego,CA:
AcademicPress,fifth ed.,1994.
[72] A. Grant,“Rayleighfadingmultiple-antennachannels,” EURASIPJ. AppliedSignal
Processing, vol. 2002,pp.316-322,Mar. 2002.
[73] J.Grimm,M. P. Fitz, andJ.V. Krogmeier, “Furtherresultsin space-timecodingfor
Rayleighfading,” IEEETrans.Commun., vol. 51,pp.1093-1101,July2003.
[74] J. C. Guey, M. P. Fitz, M. R. Bell, andW. Y. Kuo, “Signal designfor transmitter
diversity wirelesscommunicationsystemsover Rayleighfading channels,” IEEE
Trans.Commun., vol. 47,pp.527-537,Apr. 1999.
[75] A. Guyader, E. Fabre,C. Guillemot, and M. Robert,“Joint source-channelturbo
decodingof entropy-coded sources,” IEEE J. Select.Areas Commun., vol. 19,
pp.1680-1696,Sept.2001.
[76] J. Hagenauer, “Rate-compatiblepuncturedconvolutional codes(RCPCcodes)and
their applications,” IEEE Trans.Commun., vol. 36,pp.389-400,Apr. 1988.
[77] A. R. Hammonsand H. El Gamal, “On the theory of space-timecodesfor PSK
modulation,” IEEE Trans.Inform.Theory, vol. 46,pp.524-542,Mar. 2000.
[78] B. Hassibi,“A fastsquare-rootimplementationfor BLAST,” in Proc.34th Asilomar
Conf. SignalsSyst.Computers, Pacific Grove,CA, Oct.-Nov. 2000,pp.1255-1259.
[79] B. Hassibiand B. M. Hochwald, “Cayley differential unitary space-timecodes,”
IEEE Trans.Inform.Theory, vol. 48,pp.1485-1503,June2002.
164
[80] B. HassibiandB. M. Hochwald,“High-ratecodesthatarelinearin spaceandtime,”
IEEE Trans.Inform.Theory, vol. 48,pp.1804-1824,June2002.
[81] B. HassibiandM. Khorrami, “Fully-diversemultiple-antennasignalconstellations
andfixed-point-freeLie groups,” Proc. IEEE Int’l Symp.Inform. Theory, Washing-
ton,DC, June2001,p. 199.
[82] L. HeandH. Ge,“A new full-rate full-diversityorthogonalspace-timeblockcoding
scheme,” IEEE Commun.Let., vol. 7, pp.590-592,Dec.2003.
[83] R.W. Heath,Jr. andA. J.Paulraj,“Linear dispersioncodesfor MIMO systemsbased
onframetheory,” IEEETrans.SignalProcessing, vol. 50,pp.2429-2441,Oct.2002.
[84] H. Herzberg and G. Poltyrev, “The error probability of I -ary PSK block coded
modulationschemes,” IEEETrans.Commun., vol. 44,pp.427-433,Apr. 1996.
[85] B. M. Hochwald andW. Sweldens,“Dif ferentialunitary space-timemodulation,”
IEEE Trans.Commun., vol. 48,pp.2041-2052,Dec.2000.
[86] B. M. Hochwald andT. L. Marzetta,“Unitary space-timemodulationfor multiple-
antennacommunicationsin Rayleigh flat fading,” IEEE Trans. Inform. Theory,
vol. 46,pp.543-564,Mar. 2000.
[87] B. M. Hochwald, T. L. Marzetta,andB. Hassibi,“Space-timeautocoding,” IEEE
Trans.Inform.Theory, vol. 48,pp.2761-2781,Nov. 2001.
165
[88] B. M. Hochwald, T. L. Marzetta,T. J. Richardson,W. Sweldens,andR. Urbanke,
“Systematicdesignof unitaryspace-timeconstellations,” IEEE Trans.Inform.The-
ory, vol. 46,pp.1962-1973,sept.2000.
[89] B. L. Hughes,“Dif ferentialspace-timemodulation,” IEEE Trans.Inform. Theory,
vol. 46,pp.2567-2578,Nov. 2000.
[90] B. L. Hughes,“Optimal space-timeconstellationsfrom groups,” IEEE Trans. In-
form.Theory, vol. 49,pp.401-410,Feb. 2003.
[91] D. Hunter, “An upperboundfor theprobabilityof aunion,” J. Appl.Probab., vol. 13,
pp.597-603,1976.
[92] W. J. Hwang,F. J. Lin, andC. T. Lin, “Fuzzy channel-optimizedvectorquantiza-
tion,” IEEE Commun.Let., vol. 4, pp.408-410,Dec.2000.
[93] T. M. Ivrlac,W. Utschick,andJ.A. Nossek,“FadingCorrelationsin WirelessMIMO
CommunicationSystems,” IEEE Trans. WirelessCommun., vol. 2, pp. 819-824,
May 2003.
[94] H. Jafarkhani, P. Ligdas and N. Farvardin, “Adaptive rate allocation in a joint
source/channelcodingframework for wirelesschannels,” in Proc. IEEE Veh.Tech-
nol. Conf., Atlanta,GA, pp.492-296,1996.
[95] H. JafarkhaniandN. Seshadri,“Super-orthogonalspace-timetrellis codes,” IEEE
Trans.Inform.Theory, vol. 49,pp.937-950,Apr. 2003.
166
[96] H. JafarkhaniandV. Tarokh,“Multiple transmitantennadifferentialdetectionfrom
generalizedorthogonaldesigns,” IEEE Trans. Inform. Theory, vol. 47, pp. 2626-
2631,Sept.2001.
[97] J.JaldenandB. Ottersten,“An exponentiallowerboundontheexpectedcomplexity
of spheredecoding,” in Proc.IEEEInt. Conf. Acoustics,Speech andSignalProcess-
ing (ICASSP), Montreal,QC,May 2004.
[98] S.JiangandR.Kohno,“A new space-timemultipletrellis codedmodulationscheme
for flat fadingchannels,” Inst.Electron.Inf. Commun.Eng. (IEICE) Trans., vol. E87-
A, pp.640-647,Mar. 2004.
[99] Y. Jing andB. Hassibi,“Fully-diverse JLK (2) codedesign,” Proc. IEEE Int’l Symp.
Inform.Theory, Yokohama,Japan,June-July2003,p. 299.
[100] Y. JingandB. Hassibi,“Unitary space-timemodulationvia Cayley transform,” IEEE
Trans.SignalProcessing, vol. 51,pp.2891-2904,Nov. 2003.
[101] R. JohannessonandK. Zigangirov, Fundamentalsof Convolutional Coding. New
York, NY: IEEE Press,1999.
[102] J. P. Kermoal,L. Schumacher, K. I. Pedersen,P. E. Mogensen,andF. Frederiksen,
“A stochasticMIMO radio channelmodelwith experimentalvalidation,” IEEE J.
Select.AreasCommun., vol. 20,pp.1211-1226,Aug. 2002.
[103] M. A. Khojastepour, A. Sabharwal, and B. Aazhang,“Improved achievable rates
for usercooperationandrelaychannels,” in Proc. IEEE Int’l Symp.Inform. Theory
(ISIT), Chicago,IL, June-July2004.
167
[104] Y. J.Kim andH. S.Lee,“Generalizedprincipalratiocombiningfor space-timecodes
in slowly fadingchannels,” IEEE Commun.Let., vol. 4, pp.343-345,Nov. 2000.
[105] I. M. Kim andV. Tarokh,“Variable-ratespace-timeblock codesin M-ary PSKsys-
tems,” IEEEJ. Select.AreasCommun., vol. 21,pp.362-373,Apr. 2003.
[106] P. KnagenhjelmandE. Agrell, “The Hadamardtransform– a tool for index assign-
ment,” IEEETrans.Inform.Theory., vol. 42,pp.1139-1151,July1996.
[107] I. Korn,J.P. Fonseka,andS.Xing, “Optimal binarycommunicationwith nonequal
probabilities,” IEEETrans.Commun., vol. 51,pp.1435-1438,Sept.2003.
[108] E.G.Kounias,“Boundsontheprobabilityof aunion,with applications,” Ann.Math.
Statist., vol. 39,pp.2154-2158,1968.
[109] H. Kuai, F. Alajaji, andG. Takahara,“A lower boundon theprobabilityof a finite
unionof events,” Discr. Math., vol. 215,pp.147-158,2000.
[110] H. Kuai, F. Alajaji, andG. Takahara,“Tight error boundsfor nonuniformsignal-
ing over AWGN channels,” IEEE Trans.Inform. Theory, vol. 46, pp. 2712-2718,
Nov. 2000.
[111] H. Kumazawa,M. Kasahara,andT. Namekawa,“A constructionof vectorquantizers
for noisychannels,” ElectronicsEng. in Japan, vol. 67-B,pp.39-47,Jan.1984.
[112] B. C.Kuo,AutomaticControl Systems. EnglewoodClif fs,NJ:PrenticeHall, seventh
ed.,1995.
168
[113] A. KurtenbachandP. Wintz, “Quantizingfor noisychannels,” IEEETrans.Commun.
Technol., vol. 17,pp.291-302,Apr. 1969.
[114] J. N. Laneman,G. W. Wornell, andD. N. C. Tse,“An efficient protocolfor realiz-
ing cooperative diversity in wirelessnetworks,” in Proc. IEEE Int’l Symp.Inform.
Theory(ISIT), Washington,DC, June2001.
[115] E. G. Larsson,“Unitary nonuniform space-timeconstellationsfor the broadcast
channel,” IEEE Commun.Let., vol. 7, pp.21-23,Jan.2003.
[116] L. H. C. Lee, Convolutional Coding: Fundamentalsand Applications, Norwood:
ArtechHouse,Inc., 1997.
[117] G. Leus,W. Zhao,G. B. Giannakis,andH. Delic, “Space-timefrequency-shift key-
ing,” IEEETrans.Commun., vol. 52,pp.346-349,Mar. 2004.
[118] Y. Li, C. N. Georghiades,andG.Huang,“Transmitdiversityoverquasi-staticfading
channelsusingmultiple antennasandrandomsignalmapping,” IEEE Trans.Com-
mun., vol. 51,pp.1918-1926,Nov. 2003.
[119] Y. Li, J.H. Winter, andN. R. Sollenberger, “MIMO-OFDM for wirelesscommuni-
cations:signaldetectionandenhancedchannelestimation,” IEEE Trans.Commun.,
vol. 50,pp.1471-1477,Sept.2002.
[120] X. B. Liang, “A high-rateorthogonalspace-timeblock code,” IEEE Commun.Let.,
vol. 7, pp.222-223,May 2003.
169
[121] X. B. Liang,“Orthogonaldesignswith maximalrates,” IEEETrans.Inform.Theory,
vol. 49,pp.2468-2503,Oct.2003.
[122] M. Lienard, P. Degauque,J. Baudet,and D. Degardin, “Investigationon MIMO
channelsin subway tunnels,” IEEE J. Select.AreasCommun., vol. 21,pp.332-339,
Apr. 2003.
[123] J.Lim andD. L. Neuhoff, “Joint andtandemsource-channelcodingwith complexity
anddelayconstraints,” IEEETrans.Commun., vol. 51,pp.757-766,May 2003.
[124] X. Lin andR. S. Blum, “Improved space-timecodesusingserial concatenation,”
IEEE Commun.Let., vol. 4, pp.221-223,July2000.
[125] T. Linder, G. Lugosi, and K. Zeger, “Ratesof convergencein the sourcecoding
theorem,in empiricalquantizerdesign,andin universallossysourcecoding,” IEEE
Trans.Inform.Theory, vol. 40,pp.1728-1740,Nov 1994.
[126] T. Linder, G. Lugosi,andK. Zeger, “Empirical quantizerdesignin thepresenceof
sourcenoiseor channelnoise,” IEEE Trans.Inform. Theory, vol. 43, pp. 612-623,
Mar. 1997.
[127] Y. Liu, M. P. Fitz, andO. Y. Takeshita,“Full ratespace-timeturbocodes,” IEEE J.
Select.AreasCommun., vol. 19,pp.969-980,May 2001.
[128] H. A. Loeliger, “Signal setsmatchedto groups,” IEEE Trans. Inform. Theory,
vol. 37,pp.1675-1682,Nov. 1991.
170
[129] A. LozanoandC. Papadias,“Layeredspace-timereceiversfor frequency-selective
wirelesschannels,” IEEE Trans.Commun., vol. 50,pp.65-73,Jan.2002.
[130] B. Lu andX. Wang,“Iterative receiversfor multiuserspace-timecodingsystems,”
IEEE J. Select.AreasCommun., vol. 18,pp.2322-2335,Nov. 2000.
[131] H. F. Lu, Y. Wang,P. V. Kumar, andK. M. Chugg,“Remarkson space-timecodes
includinga new lower boundandanimprovedcode,” IEEE Trans.Inform. Theory,
vol. 49,pp.2752-2757,Oct.2003.
[132] X. Ma andG. B. Giannakis,“Full-diversityfull-rate complex-field space-timecod-
ing,” IEEETrans.SignalProcessing, vol. 51,pp.2917-2930,Nov. 2003.
[133] A. ManikasandM. Sethi,“A space-timechannelestimatorandsingle-userreceiver
for code-reuseDS-CDMA systems,” IEEETrans.SignalProcessing, vol. 51,pp.39-
51,Jan.2003.
[134] T. L. MarzettaandB. M. Hochwald, “Capacityof a mobilemultiple-antennacom-
munication link in Rayleigh flat fading,” IEEE Trans. Inform. Theory, vol. 45,
pp.139-157,Jan.1999.
[135] J. L. Massey, “Joint sourceand channelcoding,” in Commun.Syst.and random
processtheory, pp.279-293,TheNetherlands:Noordhoff, 1978.
[136] D. Miller and K. Rose, “Combined source-channelvector quantization us-
ing deterministic annealing,” IEEE Trans. Commun., vol. 42, pp. 347-356,
Feb./Mar./Apr. 1994.
171
[137] A. F. Molisch,“A genericmodelfor MIMO wirelesspropagationchannelsin macro-
andmicrocells,” IEEE Trans.SignalProcessing, vol. 52,pp.61-71,Jan.2004.
[138] A. L. Moustakas,S. H. Simon,andA. M. Sengupta,“MIMO capacitythroughcor-
relatedchannelsin thepresenceof correlatedinterferersandnoise:a (not so) large
Manalysis,” IEEE Trans.Inform.Theory, vol. 49,pp.2545-2561,Oct.2003.
[139] S.H. Muller-Weinfurtner, “Coding approachesfor multipleantennatransmissionin
fast fadingand OFDM,” IEEE Trans.Signal Processing, vol. 50, pp. 2442-2450,
Oct.2002.
[140] A. Naguib,N. Seshadri,andA. Calderbank,“Increasingdatarateoverwirelesschan-
nels,” IEEE SignalProcessingMagazine, vol. 46,pp.76-92,May 2000.
[141] B. K. Ng andE. S. Sousa,“On bandwidth-efficient multiuser-space-timesignalde-
signanddetection,” IEEEJ. Select.AreasCommun., vol. 20,pp.320-329,Feb. 2002.
[142] A. Nosratinia,J. Lu, and B. Aazhang,“Source-channelrate allocation for pro-
gressive transmissionof images,” IEEE Trans. Commun., vol. 51, pp. 186-196,
Feb. 2003.
[143] H. Nyquist,“Certaintopicsin telegraphtransmissiontheory,” Trans.AIEE, vol. 47,
p. 617,1928.
[144] O. Oyman,R. U. Nabar, H. Bolcskei, and A. J. Paulraj, “Characterizingthe sta-
tistical propertiesof mutual information in MIMO channels,” IEEE Trans.Signal
Processing, vol. 51,pp.2784-2795,Nov. 2003.
172
[145] B. D. Pettijohn,M. W. Hoffman,andK. Sayood,“Joint source/channelcodingusing
arithmeticcodes,” IEEE Trans.Commun., vol. 49,pp.826-836,May 2001.
[146] N. Phamdoand F. Alajaji, “Soft-decisiondemodulationdesignfor COVQ over
white, colored, and ISI Gaussianchannels,” IEEE Trans. Commun., vol. 48,
pp.1499-1506,Sept.2000.
[147] N. Phamdo,F. Alajaji, andN. Farvardin, “Quantizationof memorylessandGauss-
Markov sourcesover binary Markov channels,” IEEE Trans. Commun., vol. 45,
pp.668-675,June1997.
[148] N. PhamdoandN. Farvardin, “Optimal detectionof discreteMarkov sourcesover
discretememorylesschannels- Applicationsto combinedsource-channelcoding,”
IEEE Trans.Inform.Theory, vol. 40,pp.186-193,Jan.1994.
[149] G. Poltyrev, “Boundson the decodingerror probability of binary linear codesvia
their spectra,” IEEE Trans.Inform.Theory, vol. 40,pp.1284-1292,July1994.
[150] J.G. Proakis,Digital Communications. New York: McGraw-Hill, 4thed.,2001.
[151] T. J. Richardson,M. A. Shokrollahi, and R. L. Urbanke, “Design of capacity-
approachingirregularlow-densityparity-checkcodes,” IEEETrans.Inform.Theory,
vol. 47,pp.619-637,Feb. 2001.
[152] A. SaidandW. A. Pearlman,“A new, fastandefficient imagecodecbasedon set
partitioningin hierarchicaltrees,” IEEE Trans.CircuitsSyst.VideoTechnol., vol. 6,
pp.243-250,June1996.
173
[153] D. SamardzijaandN. Mandayam,“Pilot-assistedestimationof MIMO fadingchan-
nel responseand achievable datarates,” IEEE Trans.Signal Processing, vol. 51,
pp.2882-2890,Nov. 2003.
[154] K. SayoodandJ.C.Borkenhagen,“Useof residualredundancy in thedesignof joint
source/channelcoders,” IEEE Trans.Commun., vol. 39,pp.838-846,June1991.
[155] C. Schlegel andA. Grant, “Dif ferential space-timeturbo codes,” IEEE Trans.In-
form.Theory, vol. 49,pp.2298-2306,Sept.2003.
[156] G. Seguin, “A lower boundon the error probability for signalsin white Gaussian
noise,” IEEETrans.Inform.Theory, vol. 44,pp.3168-3175,Nov. 1998.
[157] M. SellathuraiandS. Haykin, “Turbo-BLAST for wirelesscommunications:the-
ory and experiments,” IEEE Trans. Signal Processing, vol. 50, pp. 2538-2546,
Oct.2002.
[158] A. Sendonaris,E. Erkip, andB. Aazhang,“Increasinguplink capacityvia userco-
operationdiversity,” in Proc. IEEE Int’l Symp.Inform. Theory(ISIT), Camberidge,
MA, Aug. 1998.
[159] M. ShafiandP. J. Smith, “On a Gaussianapproximationto the capacityof wire-
lessMIMO systems,” in Proc. IEEE Int’l Conf. Commun., New York, NY, Apr.-
May 2002,pp.406-410.
[160] C. E. Shannon,“A mathematicaltheoryof communication(Part1),” Bell Syst.Tech.
J., vol. 27,pp.379-423,1948.
174
[161] C. E. Shannon,“A mathematicaltheoryof communication(Part2),” Bell Syst.Tech.
J., vol. 27,pp.623-656,1948.
[162] C. E. Shannon,“Communicationin the presenceof noise,” in Proc. IRE, vol. 37,
pp.10-21,1949.
[163] C. E. Shannon,“Coding theoremsfor a discretesourcewith a fidelity criterion,” in
IRE Nat.Conv. Rec., pp.142-163,1959.
[164] C. E. Shannon,“Probability of error for optimalcodesin a Gaussianchannel,” Bell
Syst.Tech. J., vol. 38,pp.611-656,May 1959.
[165] P. G. SherwoodandK. Zeger, “Error protectionfor progressive imagetransmission
over memorylessandfadingchannels,” IEEE Trans.Commun., vol. 46, pp. 1555-
1559,Dec.1998.
[166] H. ShinandJ.H. Lee,“Performanceanalysisof space-timeblockcodesoverkeyhole
Nakagami-N fading channels,” IEEE Trans.Veh. Technol., vol. 53, pp. 351-362,
Mar. 2004.
[167] H. Shin andJ. H. Lee, “Exact symbolerror probability of orthogonalspace-time
block codes,” Proc. IEEEGlobal Telecom.Conf., Taipei,Taiwan,Nov. 2002,vol. 2,
pp.1197-1201.
[168] H. ShinandJ.H. Lee,“Capacityof multiple-antennafadingchannels:spatialfading
correlation,doublescattering,andkeyhole,” IEEE Trans.Inform. Theory, vol. 49,
pp.2636-2647,Oct.2003.
175
[169] D. S. Shiu, G .J. Foschini, M. J. Gans,andJ. M. Kahn, “Fading correlationand
its effect on thecapacityof multielementantennasystems,” IEEE Trans.Commun.,
vol. 48,pp.502-513,Mar. 2000.
[170] A. Shokrollahi, “Computing the performanceof unitary space-timegroup codes
from their charactertable,” IEEE Trans. Inform. Theory, vol. 48, pp. 1355-1371,
June2002.
[171] A. Shokrollahi,B. Hassibi,B. M. Hochwald, and W. Sweldens,“Representation
theory for high-ratemultiple-antennacodedesign,” IEEE Trans. Inform. Theory,
vol. 47,pp.2335-2367,Sept.2001.
[172] M. A. Shokrollahi, “Design of unitary space-timecodesfrom representationsof
J(O (2),” in Proc. IEEE Int’l Symp.Inform. Theory, Washington,DC, May 2001,
p. 241.
[173] M. K. Simon,“Evaluationof theaveragebit errorprobabilityfor space-timecoding
basedonasimplerexactevaluationof pairwiseerrorprobabillity,” J. Commun.Net.,
vol. 3, pp.257-264,Sept.2001.
[174] M. K. Simon,“A simplerform of theCraigrepresentationfor thetwo-dimensional
joint GaussianQ-Function,” IEEECommun.Let., vol. 6, pp.49-51,Feb. 2002.
[175] M. K. SimonandM. S.Alouini, Digital Communicationsover FadingChannels:a
UnifiedApproach. New York, NY: JohnWiley, 2000.
176
[176] M. K. SimonandH. Jafarkhani,“Performanceevaluationof super-orthogonalspace-
timetrellis codesusingamomentgeneratingfunction-basedapproach,” IEEETrans.
SignalProcessing, vol. 51,pp.2739-2751,Nov. 2003.
[177] M. Skoglund, “A soft decodervector quantizerfor a Rayleigh fading channel-
Applicationto imagetransmission,” in Proc. IEEE Int. Conf. Acoustics,Speech and
SignalProcessing(ICASSP), pp.2507-2510,May 1995.
[178] M. Skoglund,“Soft decodingfor vectorquantizationovernoisychannelswith mem-
ory,” IEEETrans.Inform.Theory, vol. 45,pp.1293-1307,May 1999.
[179] M. SkoglundandP. Hedelin,“Hadamard-basedsoftdecodingfor vectorquantization
overnoisychannels,” IEEETrans.Inform.Theory, vol. 45,pp.515-532,Mar. 1999.
[180] M. Skoglund,N. Phamdo,andF. Alajaji, “Designandperformanceof VQ-basedhy-
brid digital-analogjoint source-channelcodes,” IEEETrans.Inform.Theory, vol. 48,
pp.708-720,Mar. 2002.
[181] F. K. SoongandB. H. Juang,“Line spectrumpair (LSP)andspeechdatacompres-
sion,” in Proc. IEEE Int’l Conf. Acoustics,Speech, and Signal Processing, 1984,
pp.1.10.1-1.10.4.
[182] B. Srinivas, R. Ladner, M. Azizoglu and E. Riskin, “Progressive transmissionof
imagesusingMAP detectionoverchannelswith memory,” IEEE Trans.Image Pro-
cessing, vol. 8, pp.462-475,Apr. 1999.
[183] A. Stefanov andT. M. Duman,“Turbo-codedmodulationfor systemswith transmit
andreceive antennadiversityover block fadingchannels:systemmodel,decoding
177
approaches,andpracticalconsiderations,” IEEE J. Select.AreasCommun., vol. 19,
pp.958-968,May 2001.
[184] H. J. Su andE. Geraniotis,“Space-timeturbo codeswith full antennadiversity,”
IEEE Trans.Commun., vol. 49,pp.47-57,Jan.2001.
[185] W. SuandX. G. Xia, “Two generalizedcomplex orthogonalspace-timeblockcodes
of rates7/11 and3/5 for 5 and6 transmitantennas,” IEEE Trans.Inform. Theory,
vol. 49,pp.313-316,Jan.2003.
[186] G. Takahara,F. Alajaji, N. C. Beaulieu,andH. Kuai, “Constellationmappingsfor
two-dimensionalsignalingof nonuniformsources,” IEEE Trans.Commun., vol. 51,
pp.400-408,Mar. 2003.
[187] N. TanabeandN. Farvardin, “Subbandimagecodingusingentropy-codedquanti-
zationover noisychannels,” IEEE J. Select.AreasCommun., vol. 10, pp. 926-943,
June1992.
[188] M. TaoandR. S. Cheng,“Trellis-codeddifferentialunitaryspace-timemodulation
overflat fadingchannels,” IEEETrans.Commun., vol. 51,pp.587-596,Apr. 2003.
[189] M. TaoandR. S.Cheng,“Improveddesigncriteriaandnew trellis codesfor space-
time codedmodulationin slow flat fadingchannels,” IEEE Commun.Let., vol. 5,
pp.313-315,July2001.
[190] G. TariccoandE. Biglieri, “Exact pairwiseerror probability of space-timecodes,”
IEEE Trans.Inform.Theory, vol. 48,pp.510-513,Feb. 2002.
178
[191] V. TarokhandH. Jafarkhani,“A differentialdetectionschemefor transmitdiversity,”
IEEE J. Select.AreasCommun., vol. 18,pp.1169-1174,July2000.
[192] V. Tarokh,H. Jafarkhani,andA. R. Calderbank,“Space-timeblock codesfrom or-
thogonaldesigns,” IEEE Trans.Inform.Theory, vol. 45,pp.1456-1467,July1999.
[193] V. TarokhandT. K. Y. Lo, “Principal ratiocombiningfor fixedwirelessapplications
whentransmitterdiversity is employed,” IEEE Commun.Let., vol. 2, pp. 223-225,
Aug. 1998.
[194] V. Tarokh,N. Seshadri,andA. R. Calderbank,“Space-timecodesfor high datarate
wirelesscommunication:performancecriterionandcodeconstruction,” IEEETrans.
Inform.Theory, vol. 44,pp.744-765,Mar. 1998.
[195] V. Tarokh, N. Seshadri,and A. R. Calderbank,“Combinedarray processingand
space-timecoding,” IEEETrans.Inform.Theory, vol. 45,pp.1121-1128,May 1999.
[196] E. Telatar, “Capacityof multi-antennaGaussianchannels,” EuropeanTrans.Tele-
com., vol. 10,pp.585-595,Nov.-Dec.1999.
[197] M. Uysal, “Effect of shadowing on the performanceof space-timetrellis codes,”
IEEE Trans.WirelessCommun., to appear.
[198] M. Uysal,“Pairwiseerrorprobabilityof space-timecodesin Rician-Nakagamichan-
nels,” IEEE Commun.Let., vol. 8, pp.132-134,Mar. 2004.
179
[199] M. Uysal and C. N. Georghiades, “Error performanceanalysis of space-time
codesover Rayleighfadingchannels,” J. Commun.Networks, vol. 2, pp. 351-356,
Dec.2000.
[200] V. Vaishampayan,“Combined source-channelcoding for bandlimitedwaveform
channels,” Ph.D.Thesis,Universityof Maryland,CollegePark,MD, 1989.
[201] V. VaishampayanandN. Farvardin,“Joint designof blocksourcecodesandmodu-
lationsignalsets,” IEEETrans.Inform.Theory, vol. 36,pp.1230-1248,July1992.
[202] E. Viterbo and J. Boutros,“A universallattice codedecoderfor fading channel,”
IEEE Trans.Inform.Theory, vol. 45,pp.1639-1642,July1999.
[203] U. Wachsmann,J. Thielecke, and H. Schotten,“Capacity-achieving coding for
MIMO channels: multi-stratum space-timecoding,” IEEE Veh. Technol. Conf.,
Rhodes,Greece,May 2001,vol. 1, pp.199-203.
[204] X. WangandH. V. Poor, “Space-timemultiuserdetectionin multipathCDMA chan-
nels,” IEEE Trans.SignalProcessing, vol. 47,pp.2356-2374,Sept.1999.
[205] S.Wicker, Error Control Systemsfor Digital CommunicationandStorage. Prentice
Hall, 1995.
[206] J. H. Winters,“On the capacityof radio communicationsystemswith diversity in
Rayleighfadingenvironment,” IEEE J. Select.AreasCommun., vol. 5, pp.871-878,
June1987.
180
[207] J. H. Winters, “The diversity gain of transmitdiversity in wirelesssystemswith
Rayleighfading,” IEEETrans.Veh.Technol., vol. 47,pp.119-123,Feb. 1998.
[208] J. H. Winters, J. Salz, and R. D. Gitlin, “The impact of antennadiversity on
the capacityof wirelesscommunicationsystems,” IEEE Trans.Commun., vol. 42,
pp.1740-1751,Feb.-Apr. 1994.
[209] A. Wittneben,“A new bandwidthefficient transmitantennamodulationdiversity
schemefor linear digital modulation,” Proc. IEEE Int’l Conf. Commun., vol. 3,
pp.1630-1634,May 1993.
[210] W. Xu, J.HagenauerandJ.Hollmann,“Joint source-channelcodingusingtheresid-
ual redundancy in compressedimages,” in Proc.IEEEInt’l COnf. Commun., Dallas,
TX., June1996.
[211] S. Yousefi,“Bounds on the performanceof maximum-likelihooddecodedbinary
block codesin AWGN interference,” Ph.D.Dissertation,Departmentof Electrical
andComputerEngineering,Universityof Waterloo,Ontario,Canada.
[212] K. Zeger andA. Gersho,“Pseudo-Graycoding,” IEEE Trans.Commun., vol. 38,
pp.2147-2158,Dec.1990.
[213] K. Zeger and V. Manzella“Asymptotic boundson optimal noisy channelquanti-
zation via randomcoding,” IEEE Trans. Inform. Theory, vol. 40, pp. 1926-1938,
Nov. 1994.
181
[214] K. Zeger, J. Vaisey, and A. Gersho,“Globally optimal vector quantizerdesign
by stochasticrelaxation,” IEEE Trans. Signal Processing, vol. 40, pp. 310-322,
Feb. 1992.
[215] L. ZhengandD. N. C. Tse,“Communicationon the Grassmannmanifold: a geo-
metricapproachto thenoncoherentmultiple-antennachannel,” IEEETrans.Inform.
Theory, vol. 48,pp.359-383,Feb. 2002.
[216] Y. Zhong,F. Alajaji, andL. Campbell,“When is joint source-channelcodingworth-
while: aninformationtheoreticperspective,” in Proc.22nd BiennialSymp.Commun.,
Kingston,ON, June2004.
[217] L. Zhong,F. Alajaji, andG.Takahara,“Error analysisfor nonuniformsignalingover
Rayleighfadingchannels,” IEEE Trans.Commun., accepted.
[218] S. ZhouandG. B. Giannakis,“Optimal transmittereigen-beamformingandspace-
time block codingbasedon channelmeanfeadback,” IEEE Trans.SignalProcess-
ing, vol. 50,pp.2599-2613,Oct.2002.
[219] S. Zhou andG. B. Giannakis,“Space-timecodingwith maximumdiversity gains
over frequency-selective fading channels,” IEEE Signal ProcessingLet., vol. 8,
pp.269-272,Oct.2001.
[220] G. C. Zhu andF. Alajaji, “Turbo codesfor nonuniformmemorylesssourcesover
noisychannels,” IEEE Commun.Let., vol. 6, pp.64-66,Feb. 2002.
182
[221] G.-C.Zhu,F. Alajaji, J.BajcsyandP. Mitran, “Transmissionof non-uniformmem-
orylesssourcesvia non-systematicturbo codes,” IEEE Trans.Commun., vol. 52,
Aug. 2004.
[222] X. ZhuandR.D. Murch,“Performanceanalysisof maximumlikelihooddetectionin
aMIMO antennasystem,” IEEE Trans.Commun., vol. 50,pp.187-191,Feb. 2002.
[223] S.A. ZummoandS.A. Al-Semari,“A tight boundontheerrorprobabilityof space-
time codesfor rapid fadingchannels,” Proc. IEEE WirelessCommun.Networking
Conf., Chicago,IL, Sept.2000,vol. 3, pp.1086-1089.
183