Beruflich Dokumente
Kultur Dokumente
LeastmeansquaresfilterWikipedia,thefreeencyclopedia
Leastmeansquaresfilter
FromWikipedia,thefreeencyclopedia
Leastmeansquares(LMS)algorithmsareaclassofadaptivefilterusedtomimicadesiredfilterbyfindingthefiltercoefficients
thatrelatetoproducingtheleastmeansquaresoftheerrorsignal(differencebetweenthedesiredandtheactualsignal).Itisa
stochasticgradientdescentmethodinthatthefilterisonlyadaptedbasedontheerroratthecurrenttime.Itwasinventedin1960by
StanfordUniversityprofessorBernardWidrowandhisfirstPh.D.student,TedHoff.
Contents
1Problemformulation
1.1Relationshiptotheleastmeansquaresfilter
1.2Definitionofsymbols
2Idea
3Derivation
4Simplifications
5LMSalgorithmsummary
6Convergenceandstabilityinthemean
7Normalisedleastmeansquaresfilter(NLMS)
7.1Optimallearningrate
7.2Proof
8Seealso
9References
10Externallinks
Problemformulation
Relationshiptotheleastmeansquaresfilter
TherealizationofthecausalWienerfilterlooksalotlikethesolutiontotheleastsquaresestimate,exceptinthesignalprocessing
domain.Theleastsquaressolution,forinputmatrix andoutputvector is
http://en.wikipedia.org/wiki/Least_mean_squares_filter
1/6
4/19/2015
LeastmeansquaresfilterWikipedia,thefreeencyclopedia
TheFIRleastmeansquaresfilterisrelatedtotheWienerfilter,butminimizingtheerrorcriterionoftheformerdoesnotrelyon
crosscorrelationsorautocorrelations.ItssolutionconvergestotheWienerfiltersolution.Mostlinearadaptivefilteringproblems
canbeformulatedusingtheblockdiagramabove.Thatis,anunknownsystem
istobeidentifiedandtheadaptivefilter
attemptstoadaptthefilter
tomakeitascloseaspossibleto
,whileusingonlyobservablesignals
,
and
but
,
and
arenotdirectlyobservable.ItssolutioniscloselyrelatedtotheWienerfilter.
Definitionofsymbols
isthenumberofthecurrentinputsample
isthenumberoffiltertaps
(Hermitiantransposeorconjugatetranspose)
estimatedfilterinterpretastheestimationofthefiltercoefficientsafternsamples
Idea
ThebasicideabehindLMSfilteristoapproachtheoptimumfilterweights
,byupdatingthefilterweightsinamannerto
convergetotheoptimumfilterweight.Thealgorithmstartsbyassumingasmallweights(zeroinmostcases),andateachstep,by
findingthegradientofthemeansquareerror,theweightsareupdated.Thatis,iftheMSEgradientispositive,itimplies,theerror
wouldkeepincreasingpositively,ifthesameweightisusedforfurtheriterations,whichmeansweneedtoreducetheweights.In
thesameway,ifthegradientisnegative,weneedtoincreasetheweights.So,thebasicweightupdateequationis:
,
where representsthemeansquareerror.Thenegativesignindicatesthat,weneedtochangetheweightsinadirectionoppositeto
thatofthegradientslope.
Themeansquareerror,asafunctionoffilterweightsisaquadraticfunctionwhichmeansithasonlyoneextrema,thatminimises
themeansquareerror,whichistheoptimalweight.TheLMSthus,approachestowardsthisoptimalweightsby
ascending/descendingdownthemeansquareerrorvsfilterweightcurve.
Derivation
TheideabehindLMSfiltersistousesteepestdescenttofindfilterweights
definingthecostfunctionas
where
istheerroratthecurrentsamplenand
whichminimizeacostfunction.Westartby
denotestheexpectedvalue.
Thiscostfunction(
)isthemeansquareerror,anditisminimizedbytheLMS.ThisiswheretheLMSgetsitsname.Applying
steepestdescentmeanstotakethepartialderivativeswithrespecttotheindividualentriesofthefiltercoefficient(weight)vector
where isthegradientoperator
http://en.wikipedia.org/wiki/Least_mean_squares_filter
2/6
4/19/2015
LeastmeansquaresfilterWikipedia,thefreeencyclopedia
Now,
isavectorwhichpointstowardsthesteepestascentofthecostfunction.Tofindtheminimumofthecostfunction
weneedtotakeastepintheoppositedirectionof
.Toexpressthatinmathematicalterms
where isthestepsize(adaptationconstant).Thatmeanswehavefoundasequentialupdatealgorithmwhichminimizesthecost
function.Unfortunately,thisalgorithmisnotrealizableuntilweknow
Generally,theexpectationaboveisnotcomputed.Instead,toruntheLMSinanonline(updatingaftereachnewsampleisreceived)
environment,weuseaninstantaneousestimateofthatexpectation.Seebelow.
Simplifications
Formostsystemstheexpectationfunction
estimator
where
mustbeapproximated.Thiscanbedonewiththefollowingunbiased
indicatesthenumberofsamplesweuseforthatestimate.Thesimplestcaseis
Forthatsimplecasetheupdatealgorithmfollowsas
IndeedthisconstitutestheupdatealgorithmfortheLMSfilter.
LMSalgorithmsummary
TheLMSalgorithmfora thorderalgorithmcanbesummarizedas
Parameters:
filterorder
stepsize
Initialisation:
Computation: For
Convergenceandstabilityinthemean
AstheLMSalgorithmdoesnotusetheexactvaluesoftheexpectations,theweightswouldneverreachtheoptimalweightsinthe
absolutesense,butaconvergenceispossibleinmean.Thatis,eventhoughtheweightsmaychangebysmallamounts,itchanges
abouttheoptimalweights.However,ifthevariancewithwhichtheweightschange,islarge,convergenceinmeanwouldbe
misleading.Thisproblemmayoccur,ifthevalueofstepsize isnotchosenproperly.
http://en.wikipedia.org/wiki/Least_mean_squares_filter
3/6
4/19/2015
LeastmeansquaresfilterWikipedia,thefreeencyclopedia
If ischosentobelarge,theamountwithwhichtheweightschangedependsheavilyonthegradientestimate,andsotheweights
maychangebyalargevaluesothatgradientwhichwasnegativeatthefirstinstantmaynowbecomepositive.Andatthesecond
instant,theweightmaychangeintheoppositedirectionbyalargeamountbecauseofthenegativegradientandwouldthuskeep
oscillatingwithalargevarianceabouttheoptimalweights.Ontheotherhandif ischosentobetoosmall,timetoconvergetothe
optimalweightswillbetoolarge.
Thus,anupperboundon isneededwhichisgivenas
where
isthegreatesteigenvalueoftheautocorrelationmatrix
algorithmbecomesunstableand
.Ifthisconditionisnotfulfilled,the
diverges.
Maximumconvergencespeedisachievedwhen
where
isthesmallesteigenvalueofR.Giventhat islessthanorequaltothisoptimum,theconvergencespeedisdetermined
by
,withalargervalueyieldingfasterconvergence.Thismeansthatfasterconvergencecanbeachievedwhen
iscloseto
,thatis,themaximumachievableconvergencespeeddependsontheeigenvaluespreadof .
Awhitenoisesignalhasautocorrelationmatrix
where isthevarianceofthesignal.Inthiscasealleigenvaluesare
equal,andtheeigenvaluespreadistheminimumoverallpossiblematrices.Thecommoninterpretationofthisresultisthereforethat
theLMSconvergesquicklyforwhiteinputsignals,andslowlyforcoloredinputsignals,suchasprocesseswithlowpassorhigh
passcharacteristics.
Itisimportanttonotethattheaboveupperboundon onlyenforcesstabilityinthemean,butthecoefficientsof
canstill
growinfinitelylarge,i.e.divergenceofthecoefficientsisstillpossible.Amorepracticalboundis
where
denotesthetraceof
.Thisboundguaranteesthatthecoefficientsof
donotdiverge(inpractice,thevalueof
shouldnotbechosenclosetothisupperbound,sinceitissomewhatoptimisticduetoapproximationsandassumptionsmadeinthe
derivationofthebound).
Normalisedleastmeansquaresfilter(NLMS)
Themaindrawbackofthe"pure"LMSalgorithmisthatitissensitivetothescalingofitsinput
.Thismakesitveryhard(if
notimpossible)tochoosealearningrate thatguaranteesstabilityofthealgorithm(Haykin2002).TheNormalisedleastmean
squaresfilter(NLMS)isavariantoftheLMSalgorithmthatsolvesthisproblembynormalisingwiththepoweroftheinput.The
NLMSalgorithmcanbesummarisedas:
Parameters:
filterorder
stepsize
Initialization:
Computation: For
Optimallearningrate
http://en.wikipedia.org/wiki/Least_mean_squares_filter
4/6
4/19/2015
LeastmeansquaresfilterWikipedia,thefreeencyclopedia
Itcanbeshownthatifthereisnointerference(
),thentheoptimallearningratefortheNLMSalgorithmis
andisindependentoftheinput
andthereal(unknown)impulseresponse
),theoptimallearningrateis
Theresultsaboveassumethatthesignals
and
.Inthegeneralcasewithinterference(
areuncorrelatedtoeachother,whichisgenerallythecaseinpractice.
Proof
Letthefiltermisalignmentbedefinedas
,wecanderivetheexpectedmisalignmentforthenext
sampleas:
Let
and
Assumingindependence,wehave:
Theoptimallearningrateisfoundat
,whichleadsto:
Seealso
Recursiveleastsquares
ForstatisticaltechniquesrelevanttoLMSfilterseeLeastsquares.
SimilaritiesbetweenWienerandLMS
http://en.wikipedia.org/wiki/Least_mean_squares_filter
5/6
4/19/2015
LeastmeansquaresfilterWikipedia,thefreeencyclopedia
Multidelayblockfrequencydomainadaptivefilter
Zeroforcingequalizer
Kerneladaptivefilter
References
MonsonH.Hayes:StatisticalDigitalSignalProcessingandModeling,Wiley,1996,ISBN0471594318
SimonHaykin:AdaptiveFilterTheory,PrenticeHall,2002,ISBN0130484342
SimonS.Haykin,BernardWidrow(Editor):LeastMeanSquareAdaptiveFilters,Wiley,2003,ISBN0471215708
BernardWidrow,SamuelD.Stearns:AdaptiveSignalProcessing,PrenticeHall,1985,ISBN0130040290
WeifengLiu,JosePrincipeandSimonHaykin:KernelAdaptiveFiltering:AComprehensiveIntroduction,JohnWiley,2010,
ISBN0470447532
PauloS.R.Diniz:AdaptiveFiltering:AlgorithmsandPracticalImplementation,KluwerAcademicPublishers,1997,ISBN0
792399129
Externallinks
LMSAlgorithminAdaptiveAntennaArrays(http://www.antennatheory.com/arrays/weights/lms.php)www.antenna
theory.com
LMSNoisecancellationdemo(http://www.advsolned.com/example_ale_nc.html)www.advsolned.com
Retrievedfrom"http://en.wikipedia.org/w/index.php?title=Least_mean_squares_filter&oldid=653139450"
Categories: Digitalsignalprocessing Filtertheory Stochasticalgorithms
Thispagewaslastmodifiedon23March2015,at10:43.
TextisavailableundertheCreativeCommonsAttributionShareAlikeLicenseadditionaltermsmayapply.Byusingthissite,
youagreetotheTermsofUseandPrivacyPolicy.WikipediaisaregisteredtrademarkoftheWikimediaFoundation,Inc.,a
nonprofitorganization.
http://en.wikipedia.org/wiki/Least_mean_squares_filter
6/6