signal to noise ratio estimation using the expectation
TRANSCRIPT
Rochester Institute of TechnologyRIT Scholar Works
Theses Thesis/Dissertation Collections
2012
Signal to noise ratio estimation using theExpectation Maximization AlgorithmAhmed Almradi
Follow this and additional works at: http://scholarworks.rit.edu/theses
This Thesis is brought to you for free and open access by the Thesis/Dissertation Collections at RIT Scholar Works. It has been accepted for inclusionin Theses by an authorized administrator of RIT Scholar Works. For more information, please contact [email protected].
Recommended CitationAlmradi, Ahmed, "Signal to noise ratio estimation using the Expectation Maximization Algorithm" (2012). Thesis. Rochester Instituteof Technology. Accessed from
SignaltoNoiseRatioEstimationusingtheExpectationMaximizationAlgorithm
By
AHMEDMALMRADI
AThesisissubmittedinpartialfulfillmentoftherequirementsforthedegreeof
MASTEROFSCIENCEinElectricalEngineering
APPROVEDBY:
…………………………………………………………
Dr.SohailDianat,Thesisadvisor
…………………………………………………………
Dr.ErnestFokoué,ThesisCommitteemember
………………………………………………………….
Dr.ChanceM.Glenn,ThesisCommitteemember
………………………………………………………….
Dr.VincentAmuso,ThesisCommitteemember
………………………………………………………….
Dr.SohailDianat,Departmenthead
DEPARTMENTOFELECTRICALANDMICROELECTRONICENGINEERING
KATEGLEASONCOLLEGEOFENGINEERING
ROCHESTERINSTITUTEOFTECHNOLOGY
ROCHESTER,NEWYORK
MARCH2012
SignaltoNoiseRatioEstimationusingtheExpectationMaximizationalgorithm
By:
AHMEDALMRADI
Thesisadvisor:
Dr.SOHAILDIANAT
Abstract
Signal to Noise Ratio (SNR) estimationwhen the transmitted symbols are unknown is acommonprobleminmanycommunicationsystems,especiallythosewhichrequireanaccurateSNRestimation. For instance, modern wireless communication systems usually require accurateestimate of SNR without knowledge of the transmitted symbols. In addition, SNR estimation isrequiredinordertoperformefficientsignaldetection,powercontrol,andadaptivemodulationInthis study, Non data Aided (NDA) SNR estimation for Binary Phase Shift Keying (PBSK) andQuadrature Phase Shift Keying (QPSK) using the Expectation Maximization (EM) algorithm isdeveloped. The assumption here is that the received data samples are drawn from amixture ofGaussiansdistributionandtheyareindependentandidenticallydistributed i. i. d. .Thequalityofthe proposed estimator is examined via the Cramer‐Rao Lower Bound (CRLB) of NDA SNRestimator.Itisalsoassumedthatthechannelgainisconstantduringeachsymbolinterval,andthenoiseisAdditiveWhiteGaussian(AWGN).MaximumLikelihoodestimatorisbeingusedifwehaveaccesstothecompletedata,inthiscasetheproblemwouldbemucheasiersincewegettheexactclosed form solution, but when the observed data are incomplete or partially available, the EMalgorithmwillbeused.Thisapproachisaniterativemethodtogetanapproximatedresultwhichiseitheranapproximatedglobalmaximumorlocalmaximum.However,intheNDASNRestimation,we only have a global maximum since our assumption is that the distribution is a mixture ofGaussians.This isbeing investigated forthecasesofSingle InputSingleOutput(SISO)andSingleInputMultipleOutput (SIMO).Themainconcernabout thereceivediversity is thecost, size,andpower,thatiswhyweresorttothetransmitdiversitysuchasMultipleInputsingleOutput(MISO)with space time block codes (STBC). The base station usually serves hundreds to thousands ofremoteunitswhich is thesolereasonofusingtransmitdiversityat thebasestation insteadofateveryremoteunitcoveredbythebasestation.Itismoreeconomicalinthiscasetoaddequipmenttothebasestationinsteadoftheremoteunits.Alamoutiusedasimpletransmitdiversitytechniqueandassumedinhispaperthatthereceiverhasperfectknowledgeofthechanneltransitionmatrix.However,thisassumptionmayseemhighlyunrealistic.Oneofourcontributionsistoestimatethechannelinformation,aswellasthenoisevariancewhichwouldbeusedinestimatingtheSNRandderiving the CRLB for both DA and NDA case. The performance of our estimator would beempiricallyassessedusingMonte‐Carlosimulations,withCRLBasaperformancemetric.
ii
TABLEODCONTENTS
Listoffigures......................................iv
Listofsymbolsandabbreviations............................v
Acknowledgments....................................viii
Abstract.........................................ix
1. Introduction.....................................1
1.1Estimation....................................2
1.2MultivariateGaussiandistribution.........................5
1.3MinimumVarianceUnbiasedEstimator(MVUE)..................7
1.4CramerRaoLowerBound(CRLB)...........................8
2. MixturesofGaussians(MoG).............................10
2.1TheBasicdefinitionofMoG...........................10
2.2TheLikelihoodfunction..............................12
3.TheEMalgorithm....................................15
3.1Introduction....................................15
3.2Jensen’sEquality..................................16
3.3Kullback‐Leibler(KL)Divergence..........................17
3.4TheEMDerivation.................................18
3.5 MixtureofGaussiansandtheEMalgorithm.....................
22
4.SISONDASNRestimationusingtheEMalgorithmintheLiterature...............27
4.1TheBPSKcase...................................27
4.1.1Thesystemmodel...............................27
4.1.2TheCRLBforBPSK.............................28
4.1.3SISONDASNRestimationusingtheEMalgorithmfortheBPSKcase.....31
iii
4.2TheQPSKcase...................................38
4.2.1TheCRLBforQPSK..............................41
4.2.2 SISONDASNRestimationusingtheEMalgorithmfortheQPSKcase....43
5.MISOwithSTBCNDASNRestimationusingtheExpectationMaximizationalgorithm.....51
5.1ConventionalReceivediversity(SIMO)model......................51
5.2Transmitdiversity(MISO)model............................52
5.3MIMOmodel.......................................53
5.4ThetransmitdiversityforBPSKcase...........................54
5.5TheCRLBforthetransmitdiversityintheBPSKcase..................58
5.6STBCNDASNRestimationusingtheEMalgorithm....................65
6.Conclusions..........................................70
7.Futurework......................................70
8.References.........................................71
iv
LISTOFFIGURES
Figure1.1Normaldistributionwith 0and 1............................3
Figure2.1MixturesoftwoGaussians......................................11
Figure2.2Samplesandtheircontourofthepreviousdistribution...................11
Figure3.1mixturesoffourGaussiansdataandtheirfitwhenC 0.35I................26
Figure3.2mixturesoffourGaussiansdataandtheirfitwhenC 0.05I................26
Figure4.1TheMSE&CRLBindB2vs.SNRindBfortheBPSKwhenN=128..............35
Figure4.2TheMSE&CRLBindB2vs.SNRindBfortheBPSKwhenN=512..............36
Figure4.3TheMSE&CRLBindB2vs.SNRindBfortheBPSKwhenN=1024.............37
Figure4.4TheErrorindB2vs.SNRindBforBPSKatdifferentNs....................37
Figure4.5TheMSEindB2vs.N..........................................38
Figure4.6TheMSE&CRLBindB2vs.SNRindBfortheQPSKwhenN=128.............47
Figure4.7TheMSE&CRLBindB2vs.SNRindBfortheQPSKwhenN=512..............48
Figure4.8TheMSE&CRLBindB2vs.SNRindBfortheQPSKwhenN=1024.............49
Figure4.9TheErrorindB2vs.SNRindBforQPSKatdifferentNs....................49
Figure5.1SIMO/Receivediversitysystemmodel..............................51
Figure5.1MISO/Transmitdiversitysystemmodel.............................52
Figure5.3MIMOsystemmodel.........................................53
Figure5.4.TheMSE&CRLBindB2vs.SNRindBforSTBCwhenN=128................67
Figure5.5.TheMSE&CRLBindB2vs.SNRindBforSTBCwhenN=512................67
Figure5.6.TheMSE&CRLBindB2vs.SNRindBforSTBCwhenN=1024...............68
Figure5.7.TheMISOwithSTBCBPSKcase...........................69
v
LISTOFSYMBOLSANDABBREVIATIONS
Symbols
SignaltoNoiseRatioindB
EstimatedSignaltoNoiseRatioindB
NonDataAided
DataAided
BinaryPhaseShiftKeying
QuadraturePhaseShiftKeying
ExpectationMaximization
. . .IndependentandIdenticallyDistributed
CramerRaoLowerBound
AdditiveWhiteGaussianNoise
MaximumLikelihood
SingleInputSingleOutput
SingleInputMultipleOutput
MultipleInputSingleOutput
MultipleInputMultipleOutput
Space‐TimeBlockCodes
MinimumVarianceUnbiasedEstimator
MixturesofGaussians
GaussianMixtureModel
MaximumLikelihoodEstimate
vi
MonteCarlo
ProbabilityDensityFunction
ӨParameterswhichneedtobeestimated
ӨTheEstimatedparameterValues
μ, ∑ RealGaussiandistributionwithmeanμandCovariance∑
μ, ∑ ComplexGaussiandistributionwithmeanμandCovariance∑
Ө Likelihoodfunction
Ө LogLikelihoodfunction
Gradientwithrespecttotheparameters
Realof
Imaginaryof
μMean
Variance
μEstimatedmean
Estimatedvariance
TheExpectedvalue
Ө FisherInformationMatrix
Ө ElementijoftheFisherInformationMatrix
MeanSquareError
Ө Bias
Mixingproportion
Ө , Ө Auxiliaryfunction
vii
Kullback‐Leibler
SignaltoNoiseratio
⊗KronckerProduct
viii
Acknowledgments
I introduce my deep gratitude to my advisor, Dr. Sohail Dianat, for his constant guidance and
supporttillwemadethisworkvisibleandvaluable.Hisvisionandideasmakeminegrowgradually
during my work in finishing this project. I am thankful for the opportunity in being under his
guidanceandsupport.
I also would like to send my thanks to Dr. Ernest Fokoue, from the center for quality and
appliedstatistics,forhisappreciatedandvaluablediscussionwhichhelpedmeshapemyprojectin
a betterway.His knowledge filled in the gapwhich I had in some statistical concepts regarding
densityestimationwhenincompletedataispresent.
Finally, special thanks go tomy family;mymother (Mabrooka) andmy father (Mohammed)
whowerealwayssupportiveandhelpful.Ialsothankmywife(Afaf)andmydaughter(Shahad)for
theirpatiencewhileIambusyworkingonthisproject.Theirsupportandsmilehashelpedmein
completingthisworkproperly.
ix
Abstract
Signal toNoiseRatio (SNR)estimationwhenthe transmittedsymbolsareunknown isacommon
problem in many communication systems, especially those which require an accurate SNR
estimation. In this study, Non data Aided (NDA) SNR estimation for Binary Phase Shift Keying
(PBSK) and Quadrature Phase Shift Keying (QPSK) using the Expectation Maximization (EM)
algorithmisdeveloped.Theassumptionhere is that thereceiveddatasamplesaredrawn froma
mixtureofGaussiansdistributionandtheyareindependentandidenticallydistributed . . . .The
quality of the proposed estimator is examined via the Cramer‐Rao LowerBound (CRLB) ofNDA
SNRestimator.Itisalsoassumedthatthechannelgainisconstantduringeachsymbolinterval,and
thenoise isAdditiveWhiteGaussian(AWGN).MaximumLikelihoodestimator isbeingused ifwe
haveaccesstothecompletedata,inthiscasetheproblemwouldbemucheasiersincewegetthe
exact closed formsolution,butwhen theobserveddataare incompleteorpartiallyavailable, the
EM algorithmwill be used. This approach is an iterativemethod to get an approximated result
which is either an approximated globalmaximum or localmaximum.However, in theNDA SNR
estimation, we only have a global maximum since our assumption is that the distribution is a
mixtureofGaussians.ThisisbeinginvestigatedforthecasesofSingleInputSingleOutput(SISO);
[3],[4],[10],andSingleInputMultipleOutput(SIMO);[5],[9].Themainconcernaboutthereceive
diversityisthecost,size,andpower,thatiswhyweresorttothetransmitdiversitysuchasMultiple
Input single Output(MISO)with space time block codes (STBC). The base station usually serves
hundredstothousandsofremoteunitswhichisthesolereasonofusingtransmitdiversityatthe
basestation insteadofateveryremoteunitcoveredbythebasestation.It ismoreeconomical in
thiscasetoaddequipmentstothebasestation insteadof theremoteunits.Alamouti [12]useda
simple transmit diversity technique and assumed in his paper that the receiver has perfect
knowledgeofthechanneltransitionmatrix.However,thisassumptionmayseemhighlyunrealistic.
Oneofourcontributionsistoestimatethechannelinformation,aswellasthenoisevariancewhich
would be used in estimating the SNR and deriving the CRLB for both DA and NDA case. The
performanceofourestimatorwouldbeempiricallyassessedusingMonte‐Carlosimulations,with
CRLBasaperformancemetric.
1
1Introduction
ManymodernwirelesscommunicationsystemsrequiretheSignaltoNoiseratio(SNR)estimation
without the knowledge of the transmitted data symbols in Non‐Data Aided (NDA)manner. SNR
estimatorsmaybedivided to twocategories,DataAided (DA)andNonDataAided (NDA). InDA
estimators,thetransmittedsymbolsareusedorknownatthereceiverandusedintheestimation
process, in NDA estimators; the estimation process is being donewithout the knowledge of the
transmittedsymbols,justbasedonthereceivedsamples.TheCramer‐RaoLowerBound(CRLB)for
NDA SNR estimation,which gives theminimumvariance of unbiased estimators,will be used in
evaluating our estimator [5]. The CRLB for NDA SNR estimation for the BPSK andQPSKwill be
derived and plotted against the EM‐ML NDA SNR estimator. Here, the transmitted symbols are
treatedasunknownparameters,butnuisanceparameter;unknownandunwanted.Thisestimation
procedurewouldbeclose to theCRLB for theSNRof interest;atsufficientlyhighSNR.Themain
purposeofthisstudyistoderiveaclosedformapproximationfortheNDACRLBintheBPSKand
QPSK case. Besides, the derivations of the EM algorithm for NDA SNR estimation which will
iterativelymaximizethelikelihoodfunctiontillreachapproximatelytheglobalmaximum.
The structure of the rest of this research is as follows, in this first chapter, we introduce the
problem of estimation especially for a Gaussian distributed data, followed by the Minimum
Variance Unbiased Estimator (MVUE) which will then lead us to the Cramer‐Rao Lower Bound
(CRLB).Inchaptertwo,MixturesofGaussians(MoG)willbeintroducedfollowedbytheMaximum
Likelihood estimate (MLE) of the MoG model. In chapter three, we introduce the Expectation
Maximization (EM) algorithm, and then we solve one of its applications which is mixtures of
Gaussiansparameterestimation.Inchapterfour,weintroduceSISONDASNRestimationusingthe
EMalgorithmintheliterature,besidessolvingthismodelfortheBPSKandQPSKcase.AlsoCRLB
forbothBPSKand/orQPSKwillbederivedandplotted.Inchapterfive,MISOwithspaceandtime
block codes (STBC) NDA SNR estimation using the EM algorithm will be addressed. NDA SNR
estimatorwill be derived and plotted in both cases and for different sample sizeN usingMonte
Carlo simulations. Improvements in SNR estimation accuracy will be proven when exploiting
diversity method due to the use of mutual information. The last chapter will talk about the
conclusionandsummaryoftheresearch.
2
1.1 Estimation
Theproblemofparameterestimationisacommonproblemthatiswidelyfacedinmanyareassuch
as Radar, speech, image analysis, biomedicine, communications, control, and much more. The
problemwearefacedwithisestimatingsomeparametersbasedonsomeseenorobservedsamples
due to the use of digital communication systems or digital computers [2]. Mathematically, we
observeNdatapointsatthereceiverside 1 , 2 , . . . . . , N whichsampledfromanunknown
probability density function ; Ө parameterized by an unknown parameter Ө, which is the
parameter of interest. Unfortunately, we don’t really get to see the distribution of the received
sampleY.Weonlygetthatsample,Y,whichwethenusetoestimateourparameterӨ.Asaresultof
that, our estimatorwon’t give us the exact Ө, but an estimated version of it. Our estimatorwill
dependhighlythroughsomefunctionofthereceivedsampleYas,
Ө 1 , 2 , . . . . . , N 1.1
Thisestimatormightbeusedtoestimatethecarrierfrequency,channelphase,signalpower,noise
power,….etc.Sincethereceiveddataarerandom,wecandescribeitsbehavioraccordingtoitsPDF
function, 1 , 2 , . . . . . , N ; Ө .ThesemicolonisusedtoindicatethatthedistributionofPDF
isparameterizedbytheparameterӨ.Andit isconstant.Thiskindofestimationiscalledclassical
parameter estimation because the parameters of interest are assumed to be deterministic and
needs to be estimated. Since we are assuming the noise which corrupted our received signal is
additivewhiteGaussiannoise,AWGN, thedistributionof thePDFof thereceivedsampleswillbe
Gaussiandistribution.Forexample,ifwereceivedonesampleandtheparameterwhichneedstobe
estimatedisthemeanandvariance.Figure1.1showsanormaldistributionofzeromeanandunite
variance. This is just an illustrative example which shows us the parameters which we will be
estimating.AssumethatwehaveasignalwhichhappentobeconstantcorruptedbyAWGN,w(n),
ourmodelwillbeasfollows,
μ
Theparameterswhichneedtobeestimatedherearetwo, theDCcomponentoraverage,andthe
varianceofourGaussiannoise.
Ifonesamplehasbeenobserved,
1 μ 1
3
Figure1.1Normaldistributionwithmu=0,sigma=1
Ө μ
then,
1 ; Ө1
√2
1 μ2
1.2
Theassumptionisthatthenoise 1 ~ 0, .Oringeneral,eachsampleof hasthesame
PDF as 1 . Since these samples are uncorrelated with each other, uncorrelated means
independentifthedistributionisGaussian,wecannowgeneralizethePDFofthereceivedsamples
YofsizeNasfollows,
1 , 2 , . . . . . , N ; Ө ; Ө ; Ө Ө/
-1.5 -1 -0.5 0 0.5 1 1.5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normal distribution
Y
P(Y
)data 1
mu
sigma
4
1
√2
n μ2
1
2
12
n μ 1.3.1
Equation(1.3)iscalledthelikelihoodfunctionofthereceiveddata,whichwouldbeusedtofindthe
MaximumLikelihoodestimate(MLE)of themeanandvariance incaseofourreceiveddatawere
generatedbysingleGaussian.Ijustwantedtomakethisclearbecauseinthenextchapterwewill
betalkingaboutreceivedsamplewhichhasbeengeneratedbymorethanoneGaussian;Mixtureof
Gaussians. To find the MLE of equation (1.3), we need to find the values of μ and which
maximizesthelikelihoodfunction Ө .Wecaninsteadmaximizetheloglikelihoodfunctionsince
thelogarithmisamonotonicfunction.MLEhastheasymptoticpropertyofbeingunbiased.Itisalso
provedthatMLEachievestheCRLBforhighSNR.
Ө2
212
n μ 1.3.2
ӨӨ
Ө
whichcanbefoundbysettingthegradientof Ө withrespecttoӨequaltozero;
Ө Ө 0.
TheMLsolutionisgivenas,
μ1
1.4
1μ 1.5
Tochecktheunbiasedness,wefindtheexpectedvalueoftheestimatedparametersandcompare
themwiththetrueparameters.Itcanbeeasilyshownas,
5
μ1
μ μ 1.6
1μ
1 2 1
1 1.7
So,weseethattheMLEofthevarianceisbiased,whiletheMLEofthemeanisunbiased.TheMLE
distributionmaybeapproximatedas,
Ө Ө, Ө~ 1.8
Where, ~ asymptotically distributed according to, Ө is the fisher informationwhich is
givenby,
Ө; Ө
Ө 1.9
AcomputersimulationcouldbeperformedtodefinehowlargethedatalengthNhadtobeinorder
togetthisasymptoticresult.MonteCarlomethodcouldbeusedforvariousrealizations(sayM)of
ourestimateӨ,Themeanandvarianceoftheestimatorcouldthenbecalculatedas,
Ө1
Ө 1.10
Ө1
Ө Ө 1.11
Thechoiceofabetterestimatorcouldbedoneas;amongalltheestimatorswehave,wechoosethe
onewhichgivesacloserestimate(onaverage) to theglobal truevalue,andhashigherprecision
(lowervariance)[2].
6
1.2 MultivariateGaussiandistribution
SincethefocusonthisresearchwillbeonGaussiandistribution,itisworthintroducingthegeneral
multivariateGaussiandistribution,itsbehaviorandproperties.
SupposewehavearandomvariablethatfollowsamultivariateGaussiandistribution.Thisrandom
variablemaybesignalcorruptedbynoise.
…
; μ …
Itscovariancematrixis; ∑ Y μ Y μ
Then,thedistributionofthisrandomvariablewillbegivenby,
1
|2 ∑ |
12
μ ∑ μ 1.12
IftheCovariancematrix∑ ,independentrandomvariableswithequalvariances,eachofthe
received samples dimension in equation (1.12)will return to equation (1.2). Let usmention an
importantpropertywhichistheeffectoflineartransformationonGaussiandistribution,let,
where,Aandbareconstantsmatrixandvectorrespectively,andXisavectorofrandomvariable.
, μ μ
, ∑ ∑
TheMLE in themultivariate casewould be as follows, themodel parameterswhich need to be
estimatedis,
Ө μ ∑ ,assumingi.i.dsamples,
; Ө1
|2 ∑ |
12
μ ∑ μ
|2 ∑ | 12
μ ∑ μ
7
Ө ; Ө2
|2 ∑ |12
μ ∑ μ
TofindtheMLEofthisparameter,weneedtogothroughthesameprocessofdifferentiatingthe
likelihoodofequation(1.12)andequatingittozerowegetthefollowing,
ӨӨ
, firstwedifferntiatewithrespecttoμ asӨμ
, weget,
μ1
1.13
anddifferntiatingwithrespecttothecovariancematrixӨ∑
, weget,
∑1
μ μ 1.14
The parameter estimate in this case is straight forward computed in a closed form since the
assumption thatY’ssamplesarebeingdrawn fromonlyoneGaussiandistribution.However, this
won’tbethecasewhenwehaveamixtureofGaussians.Theproposedalgorithmwillbeamixture
of multivariate Gaussian distribution since we are dealing with PBSK and QPSK constellation
systemsandthetransmitteddatasymbolsareunknown.
1.3 MinimumVarianceUnbiasedEstimator(MVUE)
Thesearchforgoodestimatorsofunknowndeterministicparametersisn’talwaysaneasytask.Our
focuswillbeonestimatorswhichonaverageyield the trueparametervalue,and thenamongall
thoseestimators,welookfortheonewhichhastheleastvariance.Foranestimatortobeunbiased,
Ө Ө 1.15
Thebiasofanestimatormaybedefinedas,
Ө Ө Ө
Tofindanoptimalestimator,weneedtofindacriterionlikethemeansquareerror(MSE),
8
Ө Ө Ө 1.16
Ө Ө Ө 1.17
GoodestimatorshavesmallMSEorsmallvarianceif thebias iszero.Equations(1.16)and(1.17)
areunrealizableestimatorssincetheyarefunctionofthetrueparametervalue,sowecannotassess
theperformanceoftheestimatorinthiscase.
MSEiscomposedtoerrorduetothebiasandvariance.MVUEdoesnotalwaysexist,andif itdid,
sometimesitisnoteasytobefound[2].OnewayoffindingitisusingtheCramerRaoLowerBound
(CRLB).WhenCRLBsatisfieswithequalityforallӨ,the .
1.4 CramerRaoLowerBound(CRLB)
Finding a lower bound on the variance of an unbiased estimatorwill be extremely important in
practice.Thiswillserveasabenchmarkincomparingthegoodnessofanestimator.CRLBallowsus
toassurethatanestimatoristheminimumvarianceunbiasedestimator(MVUE),whichwillbethe
casewhentheestimatorattainstheCRLBforallvaluesoftheunknownparameter.
If ; Ө satisfiesregularityconditions,thevarianceofunbiasedestimatorisdefinedas,
Ө1 ; ӨӨ
1.18
AnestimatormaybefoundthatattainstheboundforallӨifandonlyifthescorefunctioncanbe
writtenas,
; ӨӨ
Ө Ө
ForsomefunctionsgandI,thatestimatorwhichistheMVUEisgivenby,
Ө
Andtheminimumvarianceis
Ө1Ө
9
Thenthisestimatorissaidtobeefficient.
LetusextendtheCRLBtoavectorparameter,
Ө Ө Ө Ө …Ө
Ө Ө 1.19
where Ө isthe Fisherinformationmatrix.
Ө ; ӨӨ Ө
1.20
Forexampleif
Ө μ
Then,
Ө
; Өμ
; Өμ
; Өμ
; Ө
Thismatrixissymmetricandpositivedefinite.
Ifweareestimatingafunctionoftheparameterasintheresearchstudy,if
Ө
Then,
ӨӨ
ӨӨӨ
10
2 MixturesofGaussians(MoG)
WhenitisnotpossibleforoneGaussiandistributiontorepresentourmodel,weresortthemixture
ofGaussiansmethod.Itisacombined(mixture)ofsomeknowndistributionsthatarecombinedto
formourmodel.Inourestimationproblem,ourassumptionisthatwearegivenareceivedsamples
YoflengthN,whicharedrawnfromamixtureofGaussiansdistribution.ThenumberofGaussians
isdeterminedby theconstellationorderweuse.Forexample, inPBSK,wehave twomixturesof
Gaussiansperantennawithmeans{‐1,+1},andinQPSK,wehavefourmixturesofGaussiansper
antenna with means {√,
√}. Gaussian Mixture Model (GMM) can be used in modeling
complicated probability density functions (PDFs)which have complicated shapes that cannot be
modeledusingasingleGaussian.
2.1 ThebasicdefinitionofMoG
ThegeneralformofaGaussianmixturemodelsis
/Ө , /Ө
/ , Ө
; μ , ∑ 2.1
where isthecomponentpriorofeachGaussiancomponent.Thereisanecessaryconditionfor
to be a valid probability density function that, ∑ 1 and 0. The parameters
whichneedtobeestimatedare , ,...., ,Ө ,Ө ,....,Ө .Increasingthenumberof
componentswouldresult inmoreparameterswhichneedtobeestimated,whichleadstoamore
accurateclassconditionalprobabilitydistributionfunction.Weneedtobecarefulaboutfindingthe
correct number of mixture components to avoid singularities, when using theML estimation to
estimate the parameters [1]. Figure 2.1 below shows the probability distribution function of a
mixture of two Gaussians, and figure 2.2 shows the contours of those two Gaussians with the
samplesdrawnfromthesamedistribution. In fact thisexactdistributionwillnotbegiven inour
caseofstudy.WhatisgivenisjustareceivedsampleofsizeNwhichisdrawninfigure2.2.
11
Figure2.1.MixturesoftwoGaussians.
Figure2.2.SamplesandContoursofthepreviousdistribution.
-10-5
05
10
-10
-5
0
5
100
0.01
0.02
0.03
0.04
0.05
x
Mixture of two Gaussians
y
p(x,
y)
-10 -5 0 5-10
-5
0
5
x
y
12
2.2 TheLikelihoodfunction
Thelikelihoodfunctionisafunctionoftheparametersofastatisticalmodelgiventheobservation
data. The maximum likelihood solution selects values for the model parameters that give a
distributionwhichgivestheobserveddatathehighestprobability.LetusintroducetheMaximum
Likelihoodestimate(MLE)ofaGaussianMixtureModel(GMM).Maximizingthelikelihoodfunction
isthesameasmaximizingtheLogLikelihoodfunctionsincethelikelihoodfunctionisamonotonic
function.This,inthecaseofoneGaussianorwhenthereisnohiddenvariableswouldsimplifythe
matha lotsincethelikelihoodfunctionisexponential function,sothelogfunctionwillcancelthe
exponential andwewould be leftwith the exponent. The Log likelihood function for the data is
givenby,
Ө /Ө
/ , Ө
; μ , ∑ 2.2
Where, it is clear here that the likelihood function depends on the PDF at a given mixture
component. Intheestimationproblemwe’veassumedthatthenoiseisAWGN,so∑ ,but
∑ in general may take any covariance matrix form. Therefore the log Likelihood function
becomes,
Ө 2
‖ μ ‖
2 2.3
ӨӨ
Ө 2.4
ThismaybesolvedbydifferentiatingthelogLikelihoodfunctionandequateittozeroasfollows,
Ө Ө ,
The maximum likelihood solution for this likelihood function is given by setting the partial
derivativeoftheloglikelihoodfunction, Ө withrespecttotheparameterofinteresttozeroand
solvingfortheparameterofinterest.Consideringaparticularparameter,Ө ,whichisrelatedto
themthmixturecomponent.
13
ӨӨ /Ө
/ , ӨӨ
0
Where,Ө μ , ∑ intheGaussiandistributioncase,thedistributionofinterest,
LetusfirstletӨ μ then,
Өμ
/μ , ∑∑ /μ , ∑
∑ μ 0
, / , Ө/μ , ∑
∑ /μ , ∑
,μ∑ / , Ө∑ / , Ө
2.5
Similarly,letӨ ∑ then,
Ө∑ ∑ /μ , ∑
/μ , ∑∑
0
Ө∑ ∑ /μ , ∑
12
/μ , ∑ ∑ μ ∑ μ ∑
, ∑∑ / , Ө μ μ
∑ / , Ө 2.6
We conclude that μ , ∑ estimates will only be correct when using the correct posterior
distribution / , Ө , but the posterior distribution / , Ө depends on them both. So,
there isno compact formsolution for theseparameters.We then turn to an iterative solution to
findingthemaximumlikelihoodestimatewhichwillbeourtargetinthisresearchwherewewillbe
using the ExpectationMaximization as an iterative approach in finding theMaximumLikelihood
solution.
Letusnowmaximizetheloglikelihoodfunctionwithrespecttothemixingcoefficients, .Thisisa
constraintoptimizationproblem,soLagrangemultiplierwillbeusedtosolvethisproblem.
14
Ө, /Ө 1
Thennowdifferentiatethisfunction, Ө, ,withrespecttoboth,themixingcoefficients, ,and
theLagrangemultiplier, ,wegettheestimatedmixturecomponentasbelow,
1N
/ , Ө 2.7
Finally,we realize thatwithout the knowledge of the class or component of each drawn sample
/ , Ө ,wecannot findthemaximumlikelihoodestimateoftheparametersof interest.The
estimatesμ and∑ willbeoneiterationinthedirectionofthemaximumlikelihoodestimateof
theEMalgorithmwhichwouldbeclearinthenextchapter.
15
3 TheExpectationMaximization(EM)Algorithm
3.1 Introduction
The EM algorithm was invented and implemented by several researchers till Dempster [6]
collectedtheirideastogether,assuredconvergence,andstatedtheterm“EMalgorithm.“which
hasbeenusedsincethen.OneofthemainapplicationareasoftheEMalgorithmisestimating
parameters ofmixture distributionwhich is the aim in this research. The EM algorithm has
broadareasofapplicationssomeofwhichisingenetics,econometric,clinical,andsociological
studies.InsignalprocessingareaslikeMaximumLikelihoodtomographicimagereconstruction,
trainingofhiddenMarkovmodels inspeechrecognition[7]. Theadventof theEMalgorithm
hascometosolvingtheproblemoflatent(hidden)(unobserved)variableswhichMLEcouldn’t
afford. Ifwe introducea jointdistributionoverboth theobservedandhiddenvariables, then
the corresponding distribution (the distribution of the observed variables alone) is given by
marginalization.Thiswouldhelpputtingcomplexdistributionsoverobservedvariablestobein
amoretractableformofdistributionformedbybothobservedandlatentvariables.Themain
focusinthisstudywillbeonEMalgorithminMixturesofGaussians(MoG).Thelatentvariable
inourcasewouldbetheclasslabel.Wewouldbegivenabunchofdatawhicharenotlabeled,
and need to estimate some parameters out of it. Herewe have two problems, the first is to
estimate the latent ormissing data, and thenwe come to our goal of estimating the desired
parameterswhichwouldnotbeestimatedwithoutestimatingthemissing(latent)classes.The
latent variable will be discrete and constant. For example, in the case of Mixture of two
Gaussians,wehavetwolatentvariables.Itwilleventuallytendtoaclassificationproblem.The
classwill thenbebelonging to either classoneor twodependingon theobserveddata.This
wouldgivewhichoftheGaussianmixturecomponentsisassociatedwitheachvector(sample)
intheobserveddata.TheEMalgorithmwouldbethemethodofchoicewhendirectmaximum
likelihood (ML) parameter estimation is not possible without the knowledge of the latent
variables. ThegoaloftheEMalgorithmistofindtheMaximumLikelihood(ML)solutionsfor
modelswhichhavelatentvariables.LetusdenotethesetofobserveddatabyY,anddenotethe
latentvariablesbyZ.Thesetofallmodelparameter isdenotedbyӨ,and then the likelihood
functionwouldbegivenby,
/Ө , /Ө 3.1
16
Thisisfordiscreterandomvariables,thesameistruewhenworkingwithcontinuousrandom
variables except switching the summation in equation (3.1) in to integration over the latent
variableZ.Thebadnewshereinequation(3.1)isthatthesummationoverthelatentvariableZ
is insidethelogarithmwhichmadeitclearthatthesolutionofthisequationwon’tbeaneasy
task.Wecanseethatsincethelogarithmandexponentialswon’tbecanceledwitheachotheras
we’ve seen in the previous chapter. In the previous chapter the summationwas outside the
logarithmwhichmadeiteasytobecancelledwiththeexponentialoftheGaussiandistribution.
LetusassumethatforeveryobservedsampleinY,weknowthecorrespondinglatentvariable
Z. Then we can call , as the complete data set. We will refer to Y as incomplete data.
Maximizationofthecompletedataloglikelihoodfunction , isstraightforwarddone
byMLE.However,wedon’treallygettoseethelatentvariablesZ,butonlytheincompletedata
Y.weknowZbytheposteriordistribution / , Ө .Sincewecannotusethecompletedatalog
likelihood, we instead take its expected value under the posterior distribution of the latent
variable.This stepcorresponds to theEstepof theEMalgorithm.TheMstep thenwouldbe
maximizingthisexpectationwithrespecttotheparametersofinterest.Ifwedenotethecurrent
estimateasӨ ,thentheEandMstepswouldresultinabetterestimateӨ .Thesetwosteps
will then be repeated till we get some precision degree [1]. We start our EM algorithm by
choosinganinitialpointӨ fortheparameterstostartwith.IntheEstep,weusethecurrent
parametervalueӨ tofindtheposteriordistributionofthelatentvariables / , Ө .Wethen
use this posterior distribution in finding the expectation of the complete data log likelihood
computedatӨ ,wedenotethisexpectationas Ө , Ө ,whichcalledauxiliaryfunction,
isgivenby,
Ө , Ө / , Ө , /Ө 3.2
Ө isthencomputedintheMstepas,
ӨӨ
Ө , Ө 3.3
LetusprovethoseequationsalongsidewiththeconvergenceoftheEMalgorithmregardlessof
anyinitialstartingpoint.Todothat,weneedtointroducesomedefinitions.
3.2 Jensen’sInequality
Thisinequalitywillbeusefulinthederivationoftheupdateformulasinthemixturemodels.Itsays
that,
17
3.4
Wherefisanyconcavefunctionand∑ 1,and0 1.
Also,
or,
1 1
3.3 Kullback‐Leibler(KL)Divergence
Thishelpsus findthedivergencebetweentwoprobabilitydistribution functions.Letustaketwo
distributions,p(y)andq(y).ThentheKLdivergencebetweenthemisgivenby,
,
Usingthefactthat 1,wethenget,
1
0
Thenweseethat , 0.
Thisgivesthefollowing,
Orindiscreteversion,
3.5
18
LetustrytowritetheKLdivergenceinthecaseoftwoGaussians,
Let, ; μ , ∑ and, ; μ , ∑ then,
,12
∑ ∑ μ μ ∑ μ μ |∑ ||∑ |
3.6
3.4 TheEMDerivation
If we consider a mixture of Gaussians distribution, the parameters values which need to be
estimated are the means, , , . . , covariances, ∑ , ∑ ,… , ∑ and component priors,
, , … , . Thosevalueswill change from iteration to iteration,Ө toӨ till they reach the
optimumvalueorstabilizeatsomevalues.OncetheӨ goestoӨ ,thePDFwillalsoiteratefrom
/Ө to /Ө .Theincreaseintheloglikelihoodwillbe,
Ө Ө /Ө /Ө
/Ө/Ө
3.7
In the case of mixture model, we need to introduce the latent variables as being one of the M
mixtures.
Ө Ө1/Ө
, /Ө
1/Ө
/ , Ө , /Ө/ , Ө
Sincethelogfunctionisstrictlyconcave,wecanapplyJensen’sInequalityasfollows,
Ө Ө / , Ө , /Ө
/Ө / , Ө 3.8
We’veused / , Ө inJensen’sinequality.
19
Also,we’veseenfrom(3.2)thattheauxiliaryfunctionmaybewrittenas,
Ө , Ө / , Ө , /Ө
Wecanthenrewriteequation(3.7)as,
Ө Ө Ө , Ө Ө , Ө
Thedifference in theauxiliary functiongivesa lowerboundonthe increase in the log likelihood.
Wecanseethat Ө , Ө dependsonthecurrentparameters,soifwejustmaximizetheauxiliary
function Ө , Ө , the likelihood functionwillalsobemaximizedasa resultof that.Nowour
aimistomaximizethefunction Ө , Ө togettotheparametersofinterest.Wecandothatby
differentiating the auxiliary function Ө , Ө with respect to the new parameters Ө and
equatetheresulttozero.Weneedtomakesurethatwhenmaximizingthefunctionwithrespectto
themixingproportion,weneedtouseLagrangemultiplier.
Ө Ө , Ө 0.
WhatweneedtoguaranteeisthatasweiteratefromӨ toӨ ,weneedtomakesurethatweare
in fact increasing the log likelihood function or elsewe are not improving or going towards the
optimumsolution.Weneed,
Ө Ө or,
Ө Ө 0.
Letusintroducetheposteriordistributionofthelatentvariable / , Ө ,
/Ө /Ө / , Ө /Ө /Ө
Since,
/ , Ө 1
Fromthedefinitionofconditionalprobability,
20
/Ө, /Ө/ , Ө
Wecanthenwritethefollowing,
/ , Ө /Ө / , Ө , /Ө/ , Ө
Similarlyforthesecondterm,
/ , Ө /Ө / , Ө , /Ө/ , Ө
Combiningthelasttwoequationsweget,
Ө Ө / , Ө , /Ө / , Ө / , Ө
/ , Ө , /Ө / , Ө / , Ө
FromthedefinitionoftheKLdivergence,
/ , Ө , / , Ө / , Ө / , Ө/ , Ө
0
UsingthisKLdivergenceequationinthelastequationweget,
Ө Ө / , Ө , /Ө / , Ө , /Ө
Where,thedifferencebetweentheleftandrighthandsidesistheKLdivergence.Ifwecanassure
thattherighthandsideispositive,thenthelefthandsidewillbealsopositive.
Sufficestoshowthat,
/ , Ө , /Ө / , Ө , /Ө
Whichwillassurethat,
Ө Ө .
21
Letusrecalltheauxiliaryfunction,
Ө , Ө / , Ө , /Ө
Iftheauxiliaryfunctionincreases,thenthelikelihoodwillalsoincrease.If,
Ө , Ө Ө , Ө
Then,
Ө Ө .
We can see the difference between the likelihoods and the auxiliary functions would be the KL
divergence,asfollows,
Ө Ө Ө , Ө Ө , Ө / , Ө , / , Ө 3.9
Theincreaseintheauxiliaryfunctionwouldbealowerboundontheincreaseoftheloglikelihood.
We noted thatmaximizing the auxiliary function once, doesn’t result in the optimumparameter
values, we need to iterate till convergence is being achieved. We can now summarize the EM
algorithmintwosteps:
⤇ The Expectation step which is done after guessing a starting point Ө and calculating the
posterior distribution of the latent variable, / , Ө , by taking the expected value of the log
likelihood of the complete data in terms of the new parameter values Ө , , /
Ө , Ө with respect to the posterior distribution of the latent variables , / , Ө . As
follows,
Ө , Ө / ,Ө , /Ө , Ө 3.10
⤇TheMaximizationstepwhichmaximizestheauxiliaryfunction, Ө , Ө ,withrespecttothe
newparametersӨ .
ӨӨ
Ө , Ө
The major problem with the initial parameter values that if there is at least one local maxima
besidestheglobalmaxima,thentheEMalgorithmmightonlygetalocalmaximainsteadofgetting
22
the globalmaxima.Oneway of trying to overcome this problem is by starting atmanydifferent
initialpointsandthenchoosestheonewhichgivesthehighestlikelihood.
3.5 MixtureofGaussiansandtheEMalgorithm
One of the applications of the EM algorithm is to find the MLE of the Mixture of Gaussians
parametersintheexistingofunobservedvariablesordata.We’veseenthattheparameterscouldn’t
be foundby a directMLE.An iterative procedureneeds to be done in order for us to get to the
optimum results for the parameters. Themajor problemwhichwe encounter here is thatwhich
Gaussiancomponentwasresponsibleforgeneratingthatspecificsample.Weknowthatthelatent
variables are discrete which determines the number of mixtures which have been used in
generating the observed data. If we have prior knowledge about the components which were
responsible for generating the data, then theproblemwould become easier and could be solved
usingadirectMLE.Sonowourmajorproblemisestimatingthecomponentwhichwasresponsible
forgeneratingeachdatapointinmyobserveddata.Onceweknowtheposteriordistributionofthe
hidden variables / , Ө , the problem then become like a directMLE problem in the auxiliary
functionsense.
Assumethat, 1iftheobserveddata wasgeneratedbycomponent0otherwise
Let us have a look at a single received sample which we suppose that it was generated by
component .Wecanwrite,
, /Ө / , Ө
/ , Ө 3.11
InadditionofYbeing . . .samples,Zalsoare . . .samples,
, /Ө , /Ө
In the Expectation step of the EM algorithm we need to compute iteratively the posterior
distribution of the latent variable / , Ө in order to find the auxiliary function. Then we
maximizetheauxiliaryfunction.
23
Ө , Ө / , Ө , /Ө
/ , Ө , /Ө
/ , Ө , /Ө
/ , Ө , /Ө
/ , Ө / , Ө
/ , Ө / , Ө
/ , Ө
Ө , Ө / , Ө / , Ө
/ , Ө 3.12
Weknowthattheloglikelihoodfunctionforacomponentmixtureis,
; μ , ∑12
2 |∑ | μ ∑ μ
Thisisthensubstitutedintheauxiliaryfunctionasfollows,
24
Ө , Ө / , Ө12
μ ∑ μ
/ , Ө12
2 ∑
/ , Ө 3.13
We now need to estimate the parameters of interest of component at iteration 1 by
differentiatingtheauxiliaryfunctionwithrespecttooneoftheparametersandequatingtozero,
Ө Ө , Ө 0
Inthecaseofestimatingthemeanofthe mixturecomponent,
Ө , Ө 0
Thiswillresultintheupdatingmeanformula,
μ∑ / , Ө∑ / , Ө
3.14
∑ Ө , Ө 0
ThiswillresultintheupdatingCovarianceformula,
∑∑ / , Ө μ μ
∑ / , Ө 3.15
Equations(3.13)to(3.15)willbeiteratedtillconvergenceisachieved.
Tosumup,theEMalgorithmforMixtureofGaussiansmaybesummarizedasfollows,
1. Initialize the parameters mean μ , Covariance ∑ , and mixing coefficient
,thencalculatetheinitialvaluefortheloglikelihoodfunction.
2. IntheEstep,weneedtocalculatetheresponsibilities, / , Ө ,
/ , Ө/μ , ∑
∑ /μ , ∑
3. IntheMstep,weneedtoreestimatetheparameters,usingthecurrentresponsibility,
25
∑ / , Ө∑ / , Ө
∑∑ / , Ө
∑ / , Ө
∑ / , Ө
4. Evaluate the log likelihood checking for increasing the log likelihood and stopping the
iterationprocessifagivenprecisionhasbeenachieved.
/ , , ∑ /μ , ∑
Iftheprecisionhasn’tbeenachieved,wegobacktopointnumber2anditerateonceagain.
Wedothesameproceduretillweachievetheprecisioncondition.
For instance, let us consider an example showing a mixture of four Gaussians as an
illustrationexample.Iassumedthatthemeansasthevaluesofthefoursentsymbolsinthe
QPSKcase, ∈ 1, 1 ,andhaveequalcovarianceof∑ 0.35 ,whereIisa2×2matrix.
Figure 3.1 shown shows these generated data with its Gaussian fit using the previous
algorithm,myassumptionmodelwaslike,
Where isthereceiveddatawhichcorruptedbyadditivewhiteGaussiannoiseAWGN, of
covariance,∑,bothofsizeN,and 1, 1 .Thisisdonebygenerating800samples
drawn from eachmixture of four Gaussians and then pretending likewe don’t know the
parametersfromwhichthosesamplesweredrawnandtryingtoestimatethoseparameters
byfindingthebestcontourmatchandestimatingtheparametersusingtheEMalgorithm.
Aswecansee fromthecoming figures, themorewe increasethevariances, themorethe
inference become difficult and we might not get the true parameters as a result. As we
decreasethevariance,themixtureswouldbecomemorecircularandseparatedandeasyto
be classified. Figure 3.2 shows the mixture of four Gaussians exactly as in the previous
figure except that we decreased the variances to 0.05. We clearly see the difference in
estimatingthecontoursofthosemixtures.Inthisfigureclassificationcouldbedoneeasily
without encountering toomuch error. K‐means plot is also drawn to see that it is just a
specialcaseoftheEMalgorithmwhenthecovarianceapproacheszero.
26
Figure3.1.mixturesoffourGaussiansdataandtheirfitwhen 0.35
Figure3.2.mixturesoffourGaussiansdataandtheirfitwhen 0.05
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
y1(n)
y 2(n)
Scatter plot and its PDF Contours estimate using the EM algorithm
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
y1(n)
y 2(n)
Scatter plot with noise Covariance 0.35I
-3 -2 -1 0 1 2 3-4
-3
-2
-1
0
1
2
3
4
y1(n)
y 2(n)
Cluster 1
Cluster 2Cluster 3
Cluster 4
Com
pone
nt 1
Pos
terio
r P
roba
bilit
y
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
-3 -2 -1 0 1 2 3-3
-2
-1
0
1
2
3
y1(n)
y 2(n)
partitioned data using the K-means
Cluster 1
Cluster 2
Cluster 3Cluster 4
Centroids
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
y1(n)
y 2(n)
Cluster 1
Cluster 2Cluster 3
Cluster 4
Com
pone
nt 1
Pos
terio
r P
roba
bilit
y
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
y1(n)
y 2(n)
partitioned data using the k-means
Cluster 1
Cluster 2
Cluster 3Cluster 4
Centroids
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
y1(n)
y 2(n)
Scatter plot with noise Covariance 0.05I
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
y1(n)
y 2(n)
Scatter plot and PDF Contours estimate using the EM algorithm
27
4 SISONDASNRestimationusingtheEMalgorithmintheLiterature
4.1 BPSKcase
In the binaryphase shift keying,wehave amixtureof two classes, ∈ 1, 1 .Those are the
transmittedsymbols.
4.1.1 Thesystemmodel
We are assuming that there is no error in timing, phase deviation, or frequency deviation. The
matchedfilteroutputmaybegivenby,
, 1, … . . , 4.1
where, isthereceivedsample, isthetransmittedsymbol, isthechannelgain,and isa
zeromeanAWGNwithvariance .
Wecanwriteequation(4.1)inavectorformas,
4.2
where, . . . . , . . . . , . . . . . The transmitted
symbols may be modeled as realizations of . . . binary random variables. The probability
distributionfunctionmaybewrittenas,
/ , Ө1
√2
2 4.3
Assumingequallylikelysymbols,wehave,
; Ө / , Ө /Ө
1
√2
12 2
12
2
where,Ө .NowforNreceived . . .sampleswehave,
; Ө ; Ө
28
; Ө1
√2
12 2
2
4.4
This is the likelihood functionof thereceiveddata.Nowgiven theobservationdata,estimate the
parametervaluesӨ andthensubstitutetofindtheSNR,
10
4.1.2 TheCRLBforBPSK
TheCramer‐RaoLowerBound(CRLB) isa lowerbound for theMeanSquareError (MSE)of any
unbiasedestimatorwhichsatisfiestheregularitycondition[2].InourcaseofSNRestimation,the
CRLBcanbewrittenas,
ӨӨ
Ө 4.5
Where Ө isthefisherinformationmatrixgivenby,
Ө
; Ө ; Ө
; Ө ; Ө 4.6
And,
Ө20
10
1010
4.7
FortheNDAcase,
Theloglikelihoodfunctionisgivenas,
Ө /Ө
2
212
4.8
29
After differentiating this function with respect to the channel gain and the variance we get the
following,
; Ө
; Ө
; Ө ; Ө
; Ө2
1 2
Ө
1
12
4.9
where,
2√2
2
4.10
Wecannowsubstitutethisresultintoequation(4.5)andgetthefollowing,
20 10
10
10
1
12
20 10
1010
20010
21
1 2 4.11
30
FortheDAcase,
Ө / , Ө
12
22
where isgiventobeeither+1or‐1.
After differentiating this function with respect to the channel gain and the variance we get the
following,
; Ө
; Ө
; Ө ; Ө
; Ө2
1
Ө0
02
4.12
Wemaynowsubstitutethisresultintoequation(4.5)andgetthefollowing,
20 10
10
10
0
02
20 10
1010
20010
2
1 4.13
31
4.1.3 NDASNRestimationusingtheEMalgorithm
WenowintroducethemethodofEMalgorithminestimatingtheSNR.Ourassumptionsarethatthe
data was drawn from a mixture of two Gaussians, and are uncorrelated with each other. The
samplesaredrawninan . . .manner.Theparameterswhichweneedtoestimatenoware,
Ө
Ө
; Ө
10 4.14
TheEMalgorithmwillbeusedtoiterativelyestimatethechannelgainandthenoisevariancethen
usethemtoestimatethesignaltonoiseratio.Thereceiveddatawillbeseenasincompletedataset
sincewedon’tknowitsclass(transmittedsymbol).Theauxiliaryfunctioninthiscasemaybegiven
as,
Ө , Ө / ,Ө , /Ө , Ө
Ө , Ө / , Ө / , Ө
/ , Ө 4.15
ӨӨ
Ө , Ө 4.16
Now solving those two equations will give us the E & M steps. Since we know the mixing
proportion ,wedon’tneedtoestimateit.
Theauxiliaryfunctionmaybegivenas,
Ө , Ө 1 / , Ө / , Ө
32
Ө , Ө / , Ө12
∑
/ , Ө12
2 ∑ / , Ө
/ , Ө12
22
Ө , Ө / , Ө
20
/ , Ө 0
/ , Ө 0
/ , Ө
/ , Ө / , Ө
, / , Ө / , Ө 1
/ , Ө 1 / , Ө
2 / , Ө 1
12 / , Ө 1
33
2 / , Ө 1 2
2
2
2
1
2
2
2
2
2
2
2
2
2
2
1 4.17
Solvingforthevariance,wehave,
Ө , Ө
/ , Ө / , Ө / , Ө / , Ө 0
/ , Ө1
2 2/ , Ө
1
2 20
/ , Ө1
2 21 / , Ө
1
2 2
0
/ , Ө2
1
2 2/ , Ө
2
0
34
2 / , Ө1
2 20
4 / , Ө 2 0
2 2 / , Ө 1 0
2 2
0
2 0
1 4.18
10
1 4.19
1 4.20
Wedon’tneedtoworryaboutthemixingcoefficientsinceweknowthattheyareequallylikely.The
hyperbolictangentinupdatingthechannelgainwillserveasasoftthreshold.Whenthehyperbolic
tangent argument is large, this thresholdwould becomemore like a signum function threshold.
After Starting with initial values for the channel gain and noise variance, equations (4.17) and
(4.18)willbeusedtoiterativelyfindtheestimatestotheparameterswhichwouldbethenusedto
35
estimate the signal to noise ratio, , as given above. The will be found for different
estimatedvaluesofSNR.ThenthisMSEwillbeassessedusingtheNDACRLBinunitsofdB2.The
estimationprocesswillbedoneforarangeofSNRfrom1dBto20dB,andfordifferentvaluesof
receivedsamplesize,N,128,512and1024.Figure4.1shownexplainsthreethingswhenBPSKis
usedas the transmittedsymbols.Theredcurve is theCRLB forNDAsignal, thebluecurve is the
CRLBforDAsignal,andtheblackcurveshowstheMSEestimatedusingtheEMalgorithm.TheEM
algorithm estimate is repeated 10000 times using Monte Carlo simulations to get 10000
realizationsof theSNRestimates, thenaveragingthemtogettheestimatedSNR,thenfindingthe
variance to get the MSE for that specific SNR estimate value. We then repeat this process for
different values of SNR, say from 0 dB to 20 dB and their resultant MSE. Those estimates then
assessedwiththeCRLBasshownbelow.WeseethatforlowSNRestimatevalues,theMSEislarge
comparedwiththeNDACRLB.ThehighertheSNRbecome,thelowertheMSEwegettillthey(the
CRLB andMSE) bothmatch. This was for sample size of N=128. The next step we would do is
increaseNandseewhatisgoingtohappentoourestimationaccuracy.
Figure4.1TheMSE&CRLBindB2vs.SNRindBwhenN=128fortheBPSKcase
0 2 4 6 8 10 12 14 16 18 200
0.5
1
1.5
2
2.5
3
SNR (dB)
MS
E &
CR
LB (
dB)2
N=128
CRLBNDA
CRLBDAMSE-EM
36
Figure4.2TheMSE&CRLBindB2vs.SNRindBwhenN=512fortheBPSKcase
Figure4.2showstherelationbetweenMSEandCRLBvs.SNRforN=512.Whatwerealizehereis
that theMSE&CRLBaredecreased comparedwith the casewhenN=128.CRLB itselfdecreased
from0.3dB2forhighSNRwhenNwas128,tolessthan0.1dB2forhighSNRwhenNis512.Figure
4.3showsthesamethingbutforN=1024.Figure4.3showsthattheprecisionoftheMSEandCRLB
isintheorderoflessthan0.05dB2forSNRsgreaterthan8dB.Figure4.4showstheerrorwhichis
the difference between theMSE and CRLB in dB2 and the SNRs for the previous three cases of
differentNs,128,512,and1024.ThisfigureprovesthatasNincreases,theerrordecreaseswhich
meansthattheMSEbecomecloserandclosertotheCRLBuntiltheybothmatchatsomeSNRvalue
and beyond. From the SNR equationwe see that the SNR is inversely proportional to the noise
variance,whichmeansasweincreasethenoisevariance,theSNRwilldecreaseasaresultofthat,
andthetwomixturesthenwillintervenewitheachotherwhichmakesitdifficultforclassification.
Ontheotherhandhowever,aswedecreasethenoisevariance,theSNRwillincreaseasaresultof
thatandthetwomixturesthenwillbefarawayfromeachotherwhichmakesclassificationbecome
easierandresultsinalmostnoerror..
0 2 4 6 8 10 12 14 16 18 200
0.2
0.4
0.6
0.8
1
1.2
SNR (dB)
MS
E &
CR
LB (
dB)2
N=512
CRLBNDA
CRLBDAMSE-EM
37
Figure4.3TheMSE&CRLBindB2vs.SNRindBwhenN=1024fortheBPSKcase
Figure4.4TheErrorindB2vs.SNRindBfordifferentNs.
0 2 4 6 8 10 12 14 16 18 200
0.1
0.2
0.3
0.4
0.5
SNR (dB)
MS
E &
CR
LB (
dB)2
N=1024
CRLBNDA
CRLBDAMSE-EM
0 2 4 6 8 10 12 14 16 18 200
0.5
1
1.5
2
2.5
SNR (dB)
Err
or in
(dB
)2
N=128
N=512
N=1024
38
Figure4.5TheMSEindB2vs.N.
The last figure4.5which shows the changes in theMSEaswe increaseN. this is easily couldbe
inferred from the previous figures, but we want it to be clear that as we increase N, MSE will
decreaseasaresult.WevaryNfrom128to4096andseetheresultantMSEwhenSNRhasbeen
chosenconstantinthiscasetobe4dB.TheMSEdecreasedfrom0.45dB2whenNwas128toabout
0.025dB2whenNis4096.OurexpectationisthatasNtendstoinfinity,MSEwilltendtozeroasa
result
4.2 TheQPSKcase
In the case ofQPSK case, it is assumed that the received samplesweredrawn independently an
identicallyfromamixtureoffourcircularsymmetriccomplexGaussiandistribution.Ourmodelin
thiscasewouldbeexactlythesameexceptforthetransmittedsymbolsnowfourinsteadoftwo.We
areassumingthatthereisnoerrorintiming,phasedeviation,orfrequencydeviation.Thematched
filteroutputmaybegivenas,
, 1, … . . , 4.21
500 1000 1500 2000 2500 3000 3500 40000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
N
MS
E in
dB
2
MSE as we increase N for SNR=4dB
39
where, isthereceivedsample, ∈√ √
isthetransmittedsymbol, isthechannelgain,
and isazeromeanAWGNwithvariance .
Wecanwriteequation(4.21)inavectorformas,
4.22
where, . . . . , . . . . , . . . . . The transmitted
symbolsmaybemodeledasrealizationsof . . .circularsymmetriccomplexGaussiandistribution.
Theprobabilitydistributionfunctionmaybewrittenas,
/ , Ө1
exp1| | 4.23
Assumingequallylikelysymbols,
; Ө / , Ө /Ө
1 1
41 1
√2
1
√2
14
1 1
√2
1
√2
14
1 1
√2
1
√2
14
1 1
√2
1
√2
1 1
41
√21
1
√21
1
√21
1
√21
| | ∗
| | | | 2 ∗
40
; Ө1 1
41
4
1
4
1
4
1
4
; Ө1 1
4
| |
2
4
2
42
4
2
4
; Ө1 1
4
| | 2
24
22
4
; Ө1 1
2
| |
24
24
; Ө1
| |
√2
√2
TheprobabilitydistributionfunctionofN . . .receivedsamplesof willbeasfollows,
; Ө1
| |
√2
√2
4.24
TheLogLikelihoodfunctionisthendefinedasfollows,
Ө /Ө
41
Ө 1
| |
√2
√2
1 | | √2
√2
4.25
ThisistheLogLikelihoodfunctionofthereceiveddataof ,nowtheproblemofestimationwould
beasfollows,giventheobservationvector, ,estimatetheparameters,Ө .Thenusethem
tofindtheSNRestimate,
10
4.2.1 TheCRLBforQPSK
TheCramer‐RaoLowerBound(CRLB) isa lowerbound for theMeanSquareError (MSE)of any
unbiasedestimatorwhichsatisfiestheregularitycondition[2].InourcaseofSNRestimation,the
CRLBcanbegivenas,
ӨӨ
Ө
Where Ө isthefisherinformationmatrixgivenby,
Ө
; Ө ; Ө
; Ө ; Ө
And,
Ө20
10
1010
FortheNDAcase,
Theloglikelihoodfunctionisgivenby,
42
Ө1 | | √2
√2
ThisprocedureissimilartotheonewhichbeingdonebyAlagha[8],
Ө
2 1 2 4
4 1 4
4.26
where,
2√
2 4.27
Wecannowsubstitutethisresultinequation(4.5)andgetthefollowing,
20 10
10
10
2 1 2 4
4 1 4
20 10
1010
10010
12
2
1 2 4 4.28
FortheDAcase,
Theloglikelihoodfunctionisgivenby,
Ө / , Ө
Ө 1
exp| |
| |
Thensubstitutingtofindthefisherinformationmatrix,
43
Ө
20
0
20 10
10
10
20
0
20 10
1010
10010
21 4.29
4.2.2 NDASNRestimationusingtheEMalgorithm
Weuse theEMalgorithmnow inestimating theSNR, itsMSEandcompareourestimatedvalues
with the CRLB values. Our assumptions are that the data was drawn from a mixture of four
Gaussians,andareuncorrelatedwitheachother.Thesamplesaredrawninan . . .manner.The
parameterswhichweneedtoestimatenoware,
Ө
Ө
; Ө
10 4.30
TheEMalgorithmwillbeusedtoiterativelyestimatethechannelgainandthenoisevariancethen
usethemtoestimatethesignaltonoiseratio.Thereceiveddatawillbeseenasincompletedataset
sincewedon’tknowtheclassofeachtransmittedsymbol.Theauxiliaryfunctioninthiscasemaybe
givenby,
Ө , Ө / ,Ө , /Ө , Ө
44
Ө , Ө / , Ө / , Ө
/ , Ө
Ө , Ө / , Ө12
∑
/ , Ө12
2 ∑
/ , Ө
/ , Ө 1
Ө , Ө / , Ө
ӨӨ
Ө , Ө 4.31
NowsolvingthosetwoequationswillgiveustheE&Msteps.
Whensolvingforthechannelgainweget,
Ө , Ө / , Ө 0
Ө , Ө / , Ө 0
/ , Ө 2 2 ∗ 0
/ , Ө / , Ө ∗
45
/ , Ө exp 2 1 /4
/ , Ө exp /4 / , Ө exp 3 /4
/ , Ө exp 5 /4 / , Ө exp 7 /4
/ , Ө 1 / , Ө / , Ө / , Ө
/ , Ө exp /4 / , Ө exp 3 /4
/ , Ө exp 5 /4
1 / , Ө / , Ө / , Ө exp 7 /4
/ , Ө exp /4 exp 7 /4
/ , Ө exp 3 /4 exp 7 /4
/ , Ө exp 5 /4 exp 7 /4 exp 7 /4
√2 / , Ө / , Ө 1 / , Ө
12
1
√2 / , Ө / , Ө
12
/ , Ө / , Ө12
4.32
Whensolvingforthevarianceweget,
Ө , Ө / , Ө 0
46
/ , Ө1
0
/ , Ө
/ , Ө / , Ө
/ , Ө
1 / , Ө / , Ө / , Ө
/ , Ө 2 ∗ 2 ∗
/ , Ө 2 ∗ 2 ∗
/ , Ө 2 ∗ 2 ∗
2 / , Ө exp /4 exp 7 /4
2 / , Ө exp 3 /4 exp 7 /4
2 / , Ө exp 5 /4 exp 7 /4
2 / , Ө √2 2 / , Ө 2 1
2 / , Ө √2
2√2 / , Ө / , Ө 1
/ , Ө1
2√2 7 /4
47
2√2 / , Ө / , Ө
12
/ , Ө / , Ө12
1
2√2| |
1
2√2
1| | 4.33
Wedon’t need toworry about themixing coefficient sincewe know that they are equally likely.
After Starting with initial values for the channel gain and noise variance, equations (4.27) and
(4.28)willbeusedtoiterativelyfindtheestimatestotheparameterswhichwouldbethenusedto
estimatethe asgivenabove.The willbefoundfordifferentestimatedvaluesofdifferent
SNRvaluesandwillbeassessedsingtheCRLB.
Figure4.6TheMSE&CRLBindB2vs.SNRindBwhenN=128
ThenthisMSEperformancewillbecomparedtothefundamentallimitoftheCRLBinunitsofdB2at
eachestimatedSNRvalue.TheestimationprocesswillbedoneforarangeofSNRfrom0dBto20
0 2 4 6 8 10 12 14 16 18 200
0.5
1
1.5
2
2.5
3
SNR (dB)
MS
E &
CR
LB (
dB)2
N=128
CRLBNDA
CRLBDAMSE-EM
48
dB, and for different values of received sample size N, 128, 512, and 1024. Figure 4.6 shown
explainsthreethingswhenQPSKisusedasthetransmittedsymbols.TheredcurveistheCRLBfor
NDA signal, theblue curve is theCRLB forDA signal, and theblack curvewhich shows theMSE
estimatedusingtheEMalgorithm.TheEMalgorithmSNRestimateisrepeated10000timesusing
MonteCarlosimulationstoget10000realizationsoftheSNRestimates,thenaveragingthemtoget
theestimatedSNR,andfindtheirvariancetogettheMSEforthatspecificSNRestimatevalue.
Figure4.7TheMSE&CRLBindB2vs.SNRindBwhenN=512
WethenrepeatthisprocessfordifferentvaluesofSNR,sayfrom0dBto20dBandtheirresultant
MSE. Those estimates then comparedwith the CRLB as shown above.We see that for low SNR
estimatevalues; theMSE is largecomparedwiththeNDACRLB.Thehigher theSNRbecome, the
lowertheMSEwegettillthey(theCRLBandMSE)bothmatch.ThiswasforsamplesizeofN=128
infigure4.6inthepreviouspage.ThenextstepwewoulddoisincreaseNandseewhatisgoingto
happentoourestimationaccuracy.Figure4.7showstherelationbetweenMSEandCRLBvs.SNR
forN=512.WhatwerealizehereisthattheMSEisdecreased,alsoCRLBisdecreasedfromwhatit
wasbeforewhenN=128.CRLBitselfdecreasedfrom2whenNwas128andSNRof0dB,tolessthan
0.5whenNis512atthesameSNR.
0 2 4 6 8 10 12 14 16 18 200
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
SNR (dB)
MS
E &
CR
LB (
dB)2
N=512
CRLBNDA
CRLBDAMSE-EM
49
Figure4.8.TheMSE&CRLBindB2vs.SNRindBwhenN=1024
Figure4.9.TheErrorindB2vs.SNRindBfordifferentNs.
0 2 4 6 8 10 12 14 16 18 200
0.05
0.1
0.15
0.2
0.25
SNR (dB)
MS
E &
CR
LB (
dB)2
N=1024
CRLBNDA
CRLBDAMSE-EM
0 2 4 6 8 10 12 14 16 18 200
0.5
1
1.5
SNR (dB)
Err
or in
(dB
)2
N=128
N=512
N=1024
50
Figure4.8showsthesamethingFigure4.7shows,butforN=1024.TheprecisionoftheMSEand
CRLBisintheorderoflessthan0.025dB2forSNRsgreaterthan6dB.Figure4.9showstheerror
whichisthedifferencebetweentheMSEandNDACRLBindB2atdifferentvaluesofSNRsshown
for the previous three cases of different Ns, 128, 512, and 1024. This figure proves that as N
increases, the error decreases whichmeans that theMSE become closer and closer to the NDA
CRLBuntiltheybothmatchatsomeSNRvalueandbeyond.FromtheSNRequationweseethatthe
SNRisinverselyproportionaltothenoisevariance,whichmeansasweincreasethenoisevariance,
theSNRwilldecreaseasaresultofthatandthefourmixturesthenwillintervenewitheachother
whichmakesitdifficultforustoclassifyeachclass.Ontheotherhandhowever,aswedecreasethe
noisevariance,theSNRwillincreaseasaresultofthatandthefourmixturesthenwillbefaraway
fromeachother(classesareseparatedfromeachother)whichmakesclassificationbecomeeasier
andresultsinalmostnoerror.
51
5. MISOwithSTBCNDASNRestimationusingtheExpectationMaximizationalgorithm
Modern communication systems require accurate estimation of the Signal to Noise Ratio
SNR for optimal usage of radio resources [5], [9]. The knowledge of the SNR is a
requirementinmostapplicationsinordertomakeoptimalsignaldetection,powercontrol.
MIMO systems gives a significant enhancement to the accuracy of estimating the SNR
besidesmanyotherimprovementsindatarate,channelcapacity,andbeingabletoreduce
Multipath fading at the remote units. Different diversity modes are being used, time
diversity, frequency diversity, spatial diversity. A MIMO system usually consists of M
transmitandNreceiveantennas.Attentionwillbegiventothetransmitdiversityforsome
reasons, which then will be compared with SISO and SIMO. Transmit diversity or MISO
wouldimproveoursystem’sperformancewithoutthecost,size,andpowerthatSIMOwill
have inorder to improve thequalityof thereceivedsignal [12].For that reason,antenna
diversitytechniquesareutilizedattheBasestation.Asimpletransmitdiversitytechnique
willbeexplainedherewhichbeendonebyAlamoutiin[12].Space‐timecodesareusedto
generate a redundant signal. The signal copy is transmitted not only from a different
antenna,butatadifferenttime.Thesignals1ands2aresentinthefirstsymbol,thenasignal
replicationof–s2*ands1*isaddedtocreatetheAlamoutispace‐timeblockcode.
5.1ConventionalReceivediversity(SIMO)model:
Figure5.1SIMO/ReceiveDiversitysystemmodel
h
h2
hNR
TX#1
RX#1
RX#2
RX#NR
Combining schem
e
52
h1
h2
hNT
wherei=1,2,.....,NR.andn=1,2,.....,N.
/ , Ө1
exp
/ , Ө1
exp
5.2Transmitdiversity(MISO)model:
Figure5.2MISO/Transmitdiversitysystemmodel
wherei=1,2,.....,NT.andn=1,2,.....,N.
/ , Ө1
exp∑ 1
/ , Ө1
exp∑ 1
TX#1
TX#2
TX#N
RX#1
53
5.3MIMOmodel:
Figure5.3MIMOsystemmodel
wherei=1,2, . . . . . ,NT, j=1,2, . . . . . ,NR ,andn=1,2, . . . . . ,N,andhij isthechannelgain
betweenthetransmitantennaiandthereceiveantennaj.
/ , Ө1
exp
/ , Ө1
exp
TX#1
TX#2
TX#N
RX#1
RX#2
RX#NR
h11
hNT NR
54
5.4 ThetransmitdiversityforBPSKcase:
ThisapproachisbeingdonebyAlamouti[12]wheretwotransmitantennasandonereceive
antennahasbeenused.Atagiventime,n,twosignalsaresimultaneouslysentfromthetwo
antennas.Denotingthesignalsentbyantennaoneby,S1,andthesignalsentbyantennatwo
by, S2. Then we repeat sending – S2* from the first antenna, then S1* from the second
antenna,thisorderingwillbedonetoallofourdata.Thisiscalledspaceandtimecoding.
Ourmodelinthiscasewilllooklike,
∗ ∗
Wethenmaywriteitinamatrixformas,
∗
∗ ∗ , 1,2, … . . , /2. 5.1
⊗
whereR(n)isthereceivedsamplesfrombothantennasattimen,Histhecomplexchannels
gains matrix, S(n) are the sent symbols at time n from the two antennas, W(n) are the
CircularlysymmetriccomplexGaussiannoisesamplesattimen,Risthevectorofreceived
samples,Iis IdentitymatrixwhereNisthetotalnumberofreceivedsamples,Sisthe
sentvectorofsymbols,Wisthenoisevectoratthereceiver,⊗isthekronckerproduct,and
Hisasfollows;
∗ ∗
The probability distribution function for the received data at a given time n, given the
transmitteddata,givenby,
/ , Ө1
exp‖ ‖
5.2
; Ө / , Ө /Ө
Supposeanequallylikelydatasymbols.
; Ө1
4exp
‖ ‖
55
; Ө1
4exp
| | | |
FortheBPSKcase, 1, 1 , 1, 1 .
where ∈ 11, 1
1, 1
1, 1
1.
Wecanthenrewrite ; Ө asfollows,
; Ө1
4exp
| | | |
exp| | | |
exp| | | |
exp| | | |
; Ө
14
exp| | | | 2 ∗ 2 ∗ | | | |
exp| | | | 2 ∗ 2 ∗ | | | |
exp| | | | 2 ∗ 2 ∗ | | | |
exp| | | | 2 ∗ 2 ∗ | | | |
; Ө1
4exp
| | | | | | | |∗
56
exp2 ∗ 2 ∗
exp2 ∗ 2 ∗
exp2 ∗ 2 ∗
exp2 ∗ 2 ∗
; Ө1
4exp
| | | | | | | |
∗ 2 cosh2 ∗ 2 ∗
2cosh2 ∗ 2 ∗
Weknowthat,
cosh cosh 2cosh2
cosh2
; Ө1
2exp
| | | | | | | |
∗ 2 cosh2 ∗ 2 ∗
cosh2 ∗ 2 ∗
; Ө1
exp| | | | 2| | 2| |
∗ cosh2 ∗ 2 ∗
cosh2 ∗ 2 ∗
5.3
ForN . . .receivedsamplesofR(n),theJointPDFdistributionwillbewrittenasfollows,
; Ө1
exp| | | | 2| | 2| |
/
∗ cosh2 ∗ 2 ∗
cosh2 ∗ 2 ∗
5.4
TheLogLikelihoodfunctionwillthenbegivenby,
57
Ө ; Ө
/
Ө 1
exp| | | | 2| | 2| |
/
∗ cosh2 ∗ 2 ∗
cosh2 ∗ 2 ∗
Ө
/| | | | 2| | 2| |
cosh2 ∗ 2 ∗
cosh2 ∗ 2 ∗
5.5
Now given the received sample R, estimate the parameters h1, h2, and . However, h1, h2 are
complexvariables,soweneedtoestimatetheirrealandimaginarypartsseparately.
Let
Nowourparametersareasfollows,
Ө ̂
Ө
; Ө
10 5.6
58
5.5 TheCRLBforthetransmitdiversityintheBPSKcase.
FortheDAcase:
InthiscaseweareassumingthatweknowthetransmitteddataSandwouldliketofinda lower
boundonthevarianceinthiscase.WeknowthatCRLBisgivenby,
ӨӨ
Ө
where Ө isthefisherinformationmatrixgivenby,
Ө ; ӨӨ Ө
And,
Ө20
10
2010
20
10
2010
10
10
where | | | | .
/ , Ө1
exp‖ ‖
/ , Ө1
exp‖ ‖
/
Ө
/‖ ‖
/| | | |
/
| | | |
59
/
| | | ∗ |
/
∗ ∗
Ө 1/
2 2 ∗
Ө 1/
2 2
1
/
2 2
1
Ө 1/
42
ӨӨ
ӨӨ
ӨӨ
Ө
Ө 2, Ө
Ө
Ө Ө
Ө Ө0
60
Ө Ө
20000
02000
00200
00020
00001
ӨӨ
Ө
2010
20
10
2010
20
10
1010
20000
02
000
002
00
0002
0
0000
2010
2010
2010
2010
1010
100
log 102
1 5.7
FortheNDAcase:
Ө
/| | | | 2| | 2| |
cosh2 ∗ 2 ∗
cosh2 ∗ 2 ∗
WeneedtofindthefisherinformationmatrixforthisLikelihoodfunctioninordertofindtheCRLB
fortheNDAcase.
Ө Ө
Ө Ө, and Ө Ө for ,
61
Ө Ө
, Ө Ө
, Ө Ө
,
Ө Ө
, Ө Ө
, Ө Ө
,
Ө Ө
, Ө Ө
, Ө Ө
,
ӨӨ
, Ө Ө
, Ө Ө
Ө Ө
, Ө Ө
, Ө Ө
.
Ө 4/
2 ∗ 2 ∗ 2 ∗
2 ∗ 2 ∗ 2
Ө/
2 ∗ 2 ∗ 2 ∗ 2 ∗
2 ∗ 2 ∗ 2 2
Ө/
2 ∗ 2 ∗ 2 ∗ 2
2 ∗ 2 ∗ 2 2 ∗
Ө/
2 ∗ 2 ∗ 2 ∗ 2
2 ∗ 2 ∗ 2 2 ∗
62
Ө 4 2 ∗ 2 ∗/
2 ∗
2 ∗ 2 ∗ 2 ∗ 2 ∗ 2 ∗
2 ∗ 2 ∗ 2
2 ∗ 2 ∗ 2 2 ∗ 2 ∗
Ө 4/
2 ∗ 2 ∗ 2 ∗
2 ∗ 2 ∗ 2
Ө/
2 ∗ 2 ∗ 2 ∗ 2
2 ∗ 2 ∗ 2 2 ∗
Ө/
2 ∗ 2 ∗ 2 ∗ 2
2 ∗ 2 ∗ 2 2 ∗
Ө 4 2 ∗ 2 ∗/
2 ∗
2 ∗ 2 ∗ 2 ∗ 2 ∗ 2 ∗
2 ∗ 2 ∗ 2
2 ∗ 2 ∗ 2 2 ∗ 2 ∗
63
Ө 4/
2 ∗ 2 ∗ 2
2 ∗ 2 ∗ 2 ∗
Ө/
2 ∗ 2 ∗ 2 2
2 ∗ 2 ∗ 2 ∗ 2 ∗
Ө 4 2 ∗ 2 ∗/
2
2 ∗ 2 ∗ 2 2 ∗ 2 ∗
2 ∗ 2 ∗ 2 ∗
2 ∗ 2 ∗ 2 ∗ 2 ∗ 2 ∗
Ө 4/
2 ∗ 2 ∗ 2
2 ∗ 2 ∗ 2 ∗
Ө 4 2 ∗ 2 ∗/
2
2 ∗ 2 ∗ 2 2 ∗ 2 ∗
2 ∗ 2 ∗ 2 ∗
2 ∗ 2 ∗ 2 ∗ 2 ∗ 2 ∗
64
Ө 2/
2| | | ∗ | 2| | 2| |
2 ∗ 2 ∗
22 ∗ 2 ∗
2 ∗ 2 ∗ 2 ∗ 2 ∗
2 ∗ 2 ∗
22 ∗ 2 ∗
2 ∗ 2 ∗ 2 ∗ 2 ∗
Numericalintegrationisusedtofindthestatisticalexpectationofeachentryinthefollowingfisher
informationmatrix.
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
Ө
ThisisthenusedtofindtheCRLBfortheSNRasfollows,
ӨӨ
Ө
5.6STBCNDASNRestimationusingtheEMalgorithm.
InthecomingSNRestimation,BPSKdatawillbeused,butonemaychooseanysignalingscheme.
TheEMalgorithmwillbeusedtoestimatetheSNR.TheEMalgorithmwilliterativelyestimatethe
channelsandthenoisevariancewhichthenbeusedtoestimatetheSNR.Theproblemis,giventhe
received samples R without the knowledge of the transmitted symbols, estimate the unknown
parameters.Ateachtime,wereceiveasample,wedon’tknowtowhichofthefourmixturesdoes
thissamplebelongsto.Ourreceiveddatawillbeincompletewhichwon’tbeeasyfortheMLEtobe
65
usedtoestimateitsparameters;weresorttoaniterativemethodliketheEMalgorithm.Westartby
writingtheauxiliaryfunction,
Ө , Ө / ,Ө , /Ө , Ө
Ө , Ө / , Ө
/
/ , Ө
/ , Ө
/
Ө , Ө / , Ө ∑
/
/ , Ө 2 ∑
/
/ , Ө
/
/ , Ө 1
Ө , Ө / , Ө
/
ӨӨ
Ө , Ө
Ө , Ө / , Ө
/
1 / , Ө
/
∗ 5.8
Ө , Ө / , Ө
/
66
1 / , Ө
/
∗ 5.9
Ө , Ө / , Ө
/
1 / , Ө
/
∗ 5.10
Ө , Ө / , Ө
/
1 / , Ө
/
∗ 5.11
Ө , Ө / , Ө
/
1 / , Ө
/
5.12
whereM=4 in our PBSK case, and N is the received sample size. Also, themixtures classes are
assumed equally likely. Equations (5.8)‐(5.12)will be used to update the parameters iteratively
starting from initial values for those parameters. Luckily that ourmodel has one global optimal
parameters,whichkeepusawayfromthinkingthatthevalueswhichweestimatedmightbelocal
maxima.TheMSE&CRLBwillbedrawnfordifferentvaluesofN.
67
Figure5.4.TheMSE&CRLBindB2vs.SNRindBwhenN=128.
Figure5.5.TheMSE&CRLBindB2vs.SNRindBwhenN=512.
0 2 4 6 8 10 12 14 16 18 200.1
0.2
0.3
0.4
0.5
0.6
SNR (dB)
MS
E &
CR
LB (
dB)2
Alamouti for N=128
MSE-EM
CRLBDACRLBNDA
0 2 4 6 8 10 12 14 16 18 200.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
SNR (dB)
MS
E &
CR
LB (
dB)2
Alamouti for N=512
CRLBDA
MSE-EMCRLBNDA
68
Figure5.6.TheMSE&CRLBindB2vs.SNRindBwhenN=1024.
Convergencetime,τoftheEMAlgorithm.(SNR=8dB):
Datasize,N Convergencetime,τinsecondsforonerun
Convergencetime,τ insecondsfor2000MonteCarlosimulations
Convergencetime,τinsecondsfor10000MonteCarlosimulations
τ1 τ2000 τ10,000
N=128 0.131456 6.881052 34.264368
N=512 0.140788 27.668970 139.303882
N=1024 0.173774 91.254040 456.952999
ThesimulationswereperformedusingMATLABversion7.13.0.564(R2011b),withprocessorIntel(R)Core(TM)i7CPU860 @ 2.80GHzwithRAM4GB.
0 2 4 6 8 10 12 14 16 18 200.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
SNR (dB)
MS
E &
CR
LB (
dB)2
Alamouti for N=1024
CRLBDA
MSE-EMCRLBNDA
69
Fig.5.7.TheMISOwithSTBCBPSKcase.
Fig.5.8.TheSISOBPSKcase.
0 2 4 6 8 10 12 14 16 18 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
SNR (dB)
MS
E &
CR
LBs
in (
dB)2
NDA CRLB, N=128
DA CRLB, N=128MSE-EM, N=128
NDA CRLB, N=512
DA CRLB, N=512
MSE-EM, N=512
NDA CRLB, N=1024DA CRLB, N=1024
MSE-EM, N=1024
0 2 4 6 8 10 12 14 16 18 200
0.1
0.2
0.3
0.4
0.5
0.6
SNR (dB)
MS
E &
CR
LBs
in (
dB)2
NDA CRLB, N=128
DA CRLB, N=128MSE-EM, N=128
NDA CRLB, N=512
DA CRLB, N=512
MSE-EM, N=512
NDA CRLB, N=1024DA CRLB, N=1024
MSE-EM, N=1024
70
6. Conclusions
Inthisresearch,transmitdiversityisusedasawayofimprovingtheSNRestimationaccuracy.The
SNRwasestimatedusingtheEMalgorithminaNDAmannerusingspace‐timeblockcoding(STBC)
which was done by Alamouti [12]. Also, the CRLB for the DA and NDA case were derived and
plottedwhichthenbeusedtoassesstheperformanceofourestimator.ThiswasdonefortheBPSK
constellationsystem.Itisbeingdonefordifferentsamplesizes,N;128,512,and1024.We’veseen
thattheCRLBandMSEareinverselyproportionaltothesamplesize,N.Inaddition,asweincrease
theSNR,CRLBandMSEbecomecloserandclosertoeachothertill theymatchatsomehighSNR
point.Ingeneral,thedetectionproblemdependshighlyontheSNRandN.Itisatradeoffbetween
thesetwofactors.Forexample,iftheSNRislow,thenweneedtoincreaseNinordertodecrease
the errorprobability and theMSEestimate.On theotherhandhowever, if the SNR is high, then
smallNwillwork forus. This techniquewas thencomparedwith theSISO[3], [4], [8], [10]and
SIMO[5],[9]models.SIMOsystemswouldhavesomedrawbacksespeciallywhenbeingappliedfor
mobile phones like, size and cost of the phone, and power consumption. Instead of having the
diversitydoneateachremoteunit,instead,itisbeingdoneatthebasestation.Forexample,having
twoantennasatthebasestationcoveringhundredsofremoteunits(havingonereceiveantenna)
wouldbemoreeconomicalthanhavinghundredsofremoteunits(eachhavingtwoantennas)and
one transmit antenna at the base station. This is the main reason of choosing MISO with STBC
systemsoverSIMOsystems.
7. Futurework
1. SNRestimationusingtheEMalgorithmwithotherlinearlymodulatedsignalsotherthan
BPSKlikeM‐QAMandM‐PSK.
2. BayesianSNRestimation.
3. MIMOSNRestimationusingtheEMalgorithm.
4. SNRestimationinnon‐Gaussiannoise.
71
8.References:
[1]C.M.Bishop,Patternrecognitionandmachinelearning.Springer2006.
[2]S.M.Kay,Fundamentalsofstatisticalsignalprocessing.PrenticeHall,1993.
[3] A. Wiesel, J. Goldberg, and H. Messer, “Non‐data‐aided signal‐to‐noise‐ratio estimation,”
Communications,2002.ICC2002.IEEEInternationalConferenceon,vol.1,pp.197‐201,2002.
[4] A. Das, “NDA SNR estimation: CRLBs and EM based estimators,” in proc. TENCON 2008,
Hyderabad,India,Nov.2008,pp.1‐6.
[5]M. A. Boujelben, F. Bellili, S. Affes, and A. Stephenne, “EM algorithm for non‐data‐aided SNR
estimation of linearly‐modulated signals over SIMO channels,” in proc. Communications IEEE
GLOBCOM2009,pp.1‐6.
[6]A.P.Dempester,N.M.Laird,andD.B.Rubin,“Maximumlikelihoodfromincompletedataviathe
EMalgorithm,”JournaloftheRoyalStatisticalSociety.SeriesB,no.1,pp.1‐38,1977.
[7]T.K.Moon,“Theexpectation‐maximizationalgorithm,”IEEESignalprocess.Mag.,vol.13,pp.47‐
60,Nov.1996.
[8]N.S.Alagha,“Cramer‐raoboundsofSNRestimatesforBPSKandQPSKmodulatedsignals,”IEEE
Communicationsletters,vol.5,no.1,pp.10‐12,January2001.
[9]M.A.Boujelben,F.Bellili,S.Affes,andA.Stephenne,“SNRestimationoverSIMOchannelsfrom
linearlymodulatedsignals,”inproc.SignalprocessingIEEEBritishCrown,vol.58,n0.12,December
2010,pp.6017‐6028.
[10]A.DasandM.Miller,“Remotenon‐linearitydetectionviaburstpowerditheringandEMbased
SNR estimation,” International workshop on Satellite and space communications. IWSSC, sept.
2007.
[11] F. Rice, B. Cowley, and M. Rice, “Cramer‐rao lower bounds for QAM phase and frequency
estimation,”IEEETrans.Communications,vol.49,pp.1582‐1591,sept.2001.
[12] S. M. Alamouti, “A simple transmit diversity technique for wireless communications,” IEEE
Trans.Communications,vol.16,pp.1451‐1458,sept.1998.
72
[13 ]N. Celandroni, E. Ferro, andF. Potorti, “Quality estimation of PSKmodulated signals,” IEEE
Comm.Mag.,pp.50‐55.July1997.
[14] K. Balachandran, S. R. Kadaba, and S. Nanda, “Channel Quality estimation and rate
adaptationforcellularmobileradio,”IEEEJ.SelectAreasComm.,vol.17,pp.1244‐1256,July1999.
[15]F.Bellili,S.Affes,andA.Stephenne,“Cramer‐RaoBoundforNDASNRestimatesofSquareQAM
ModulatedSignals,”IEEEinWCNC2009,pp.1‐5.
[16] F. Bellili, S. Affes, andA. Stephenne, “Cramer‐RaoBounds forNDA SNR estimates of Square
QAMModulatedTransmissions,”IEEEinComm.2010,pp.3211‐3218.
[17]S.Caiyao,Y.Hongyi,andZ.Tianzhong,“Cramer=RaoboundsofSNREstimationforBPSKand
QPSKsignalsoverSIMOChannels,”IEEEICSPS2010,pp.647‐650.
[18]W.Gappmair,R. Lopez‐Valcare, andC.Mosquera, “MLandEMalgorithm fornon‐data‐aided
SNRestimationoflinearlymodulatedsignals,”inProc.IEEEsymp.Commun.Syst.,Graz,Austria,Jul.
2008,pp.530‐534.
[19]R.Lopez‐Valcare,C.Mosquera,andW.Gappmair,“IterativeenvelopbasedSNRestimationfor
nonconstantmodulusconstellation,”inProc.IEEESPAWC’07,Helsinki,Finland,Jun.2007,pp.1‐5.
[20]W.Gappmair,R.Lopez‐Valcare,andC.Mosquera,“Cramer‐RaolowerboundandEMalgorithm
forenvelope‐basedSNRestimationofnonconstantmodulusconstellation ,” IEEETrans.Commun.
Vol.57,no.6,June2009,pp.1622‐1627.