Você está na página 1de 27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

ApplyingtheHaarWaveletTransformto
TimeSeriesInformation
Contents
Introduction(/misl/misl_tech/wavelets/haar.html#Introduction)
Whatalong,strangetripsitsbeen(/misl/misl_tech/wavelets/haar.html#StrangeTrip)
Thewavelettechniqueforanalyzingasignalortimeseries
(/misl/misl_tech/wavelets/haar.html#WaveletTechnique)
TheLanguageofWavelets(/misl/misl_tech/wavelets/haar.html#language)
ApplyingWaveletsandJavaSourceCode
(/misl/misl_tech/wavelets/haar.html#ApplyingWavelets)
FinancialTimeSeries(/misl/misl_tech/wavelets/haar.html#FinancialTimeSeries)
WhyHaarWavelets?(/misl/misl_tech/wavelets/haar.html#WhyHaar)
HaarWavelets(/misl/misl_tech/wavelets/haar.html#HaarWavelets)
FilteringSpectrum(/misl/misl_tech/wavelets/haar.html#FilteringSpectrum)
NoiseFilters(/misl/misl_tech/wavelets/haar.html#NoiseFilters)
Waveletsvs.SimpleFilters(/misl/misl_tech/wavelets/haar.html#SimpleFilters)
LimitationsoftheHaarWaveletTransform(/misl/misl_tech/wavelets/haar.html#limitations)
WaveletsandParallelism(/misl/misl_tech/wavelets/haar.html#parallelism)

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

1/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

JavaSourceCodeDownload(/misl/misl_tech/wavelets/haar.html#download)
Resources(/misl/misl_tech/wavelets/haar.html#Resources)

References,Books(/misl/misl_tech/wavelets/haar.html#ReferencesBooks)
References,Webpublished(/misl/misl_tech/wavelets/haar.html#ReferencesWeb)

Linkstosubpages
DaubechiesWavelets(/software/java/wavelets/daubechies/index.html)
FilteringUsingHaarWavelets(/misl/misl_tech/wavelets/spectrum_plots/index.html)
TheHaarWaveletTransformandNoiseFilters
(/misl/misl_tech/wavelets/close_images/index.html)
WaveletNoiseThresholding(/misl/misl_tech/wavelets/noise.html)
Waveletsvs.SimpleFilters(/misl/misl_tech/wavelets/simple_filters/index.html)
WaveletsinJava(includeshistogrammingandsimplestatisticalalgorithms)
(/software/java/wavelets/index.html)

Introduction
ThiswasthefirstwebpageIwroteonWavelets.Fromthisseedgrewotherwebpageswhich
discussavarietyofwaveletrelatedtopics.Fora"tableofcontents"seeWaveletsandSignal
Processing(/misl/misl_tech/wavelets/index.html).Thiswebpageappliesthewavelettransformto
atimeseriescomposedofstockmarketcloseprices.Laterwebpagesexpandonthisworkina
varietyofareas(e.g.,compression,spectralanalysisandforecasting).

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

2/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

WhenIstartedoutIthoughtthatIwouldimplementtheHaarwaveletandthatsomeofmy
colleaguesmightfindituseful.Ididnotexpectsignalprocessingtobesuchaninterestingtopic.
NordidIunderstandwhomanydifferentareasofcomputerscience,mathematics,and
quantitativefinancewouldbetouchedbywavelets.Ikeptfindingthat"onethingleadtoanother",
makingitdifficulttofindalogicalstoppingplace.Thiswanderingpathofdiscoveryonmypartalso
accountsforthesomewhatorganicgrowthofthesewebpages.Ihavetriedtotamethisgrowth
andorganizeit,butIfearthatitstillreflectsthefactthatIdidnotknowwhereIwasgoingwhenI
started.
TheJavacodepublishedalongwiththiswebpagereflectthefirstworkIdidonwavelets.More
sophisticated,liftingschemebased,algorithms,implementedinJavacanbefoundonotherweb
pages.Thewaveletliftingschemecode,publishedonotherwebpages,issimplerandeasierto
understand.Thewaveletliftingschemealsoprovidesanelegantandpowerfulframeworkfor
implementingarangeofwaveletalgorithms.
Inimplementingwaveletpacketalgorithms,IswitchedfromJavatoC++.Thewaveletpacket
algorithmIusedissimplerandmoreelegantusingC++'soperatoroverloadingfeatures.C++also
supportsgenericdatastructures(templates),whichallowedmetoimplementagenericclass
hierarchyforwavelets.Thiscodeincludesseveraldifferentwaveletalgoriths,includingHaar,
linearinterpolationandDaubechiesD4.
Likethewaveletalgorithms,thefinancialmodelingdonehererepresentsveryearlywork.WhenI
startedworkingonthesewebpagesIhadnoexperiencewithmodelingfinancialtimeseries.The
workdescribedonthiswebpageleadtomoreintensiveexperimentswithwaveletfiltersin
financialmodels,whichIcontinuetoworkon.OnthiswebpageIusestockmarketcloseprices.
Infinancialmodelingoneusuallyusesreturns,sincewhatyouaretryingtopredictisfuturereturn.
Ibecameinterestedinwaveletsbyaccident.Iwasworkingonsoftwareinvolvedwithfinancial
timeseries(e.g.,equityopenandcloseprice),soIsupposethatitwasanaccidentwaitingto
happen.IwasreadingtheFebruary2001issueofWIREDmagazinewhenIsawthegraph
includedbelow.EverymonthWIREDrunsvariousgraphicvisualizationsoffinancialdataandthis
wasoneofthem.

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

3/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

Ifstockpricesdoindeedfactorinallknowableinformation,acompositeprice
graphshouldproceedinanorderlyfashon,asnewinformationnudgesperceived
valueagainstthepullofestablishedtendencies.Waveletanalysis,widelyusedin
communicationstoseparatesignal(patternedmotion)fromnoise(random
activity),suggestsotherwise.
ThisimageshowstheresultsofrunningaHaartransformthefundamental
waveletformulaonthedailycloseoftheDowandNASDQsince1993.The
bluemountainsconstitutesignal.Theembeddedredspikesrepresentnoise,of
whichtheyellowlinefollowsa50daymovingaverage.
Noise,whichcanberegardedasinvestorignorance,hasrisenalongwiththe
valueofbothindices.ButwhilenoiseintheDowhasgrown500percenton
average,NASDAQnoisehasballooned3,000percent,faroutstripping
NASDAQ'sspectacular500percentgrowthduringthesameperiod.Mostofthis
increasehasoccurredsince1997,withanextraordinarysurgesinceJanuary
2000.PerhapstherewasaY2Kglichafterallonethatderailednotoperating
systemsandCPUs,but>investorpsychology.ClemChambers
(clemc@advfn.com).
GraphandquotefromWIREDMagazine,February2001,page176
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

4/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

IamaPlatonist.Ibelievethat,intheabstract,thereistruth,butthatwecanneveractuallyreach
it.Wecanonlyreachanapproximation,orashadowoftruth.Modernscienceexpressesthisas
Heisenberguncertainty.
APlatonistviewofafinancialtimeseriesisthatthereisa"true"timeseriesthatisobscuredto
someextentbynoise.Forexample,aclosepriceorbid/asktimeseriesforastockmovesonthe
basisofthesupplyanddemandforshares.Inthecaseofabid/asktimeseries,the
supply/demandcurvewillbesurroundedbythenoisecreatedbyrandomorderarrival.If,
somehow,thenoisecouldbefilteredout,wewouldseethe"true"supply/demandcurve.Software
whichusesthisinformationmightbeabletodoabetterjobbecauseitwouldnotbeconfusedby
falsemovementscreatedbynoise.
TheWIREDgraphabovesuggeststhatwaveletanalysiscanbeusedtofilterafinancialtime
seriestoremovetheassociatednoise.Ofcoursethereisavastareathatisnotaddressedbythe
WIREDquote.What,forexample,constitutesnoise?WhatarewaveletsandHaarwavelets?
Whyarewaveletsusefulinanalyzingfinancialtimeseries?WhenIsawthisgraphIknew
answerstononeofthesequestions.
TheanalysisprovidedinthebriefWIREDparagraphisshallowaswell.Noiseinthetimeseries
increaseswithtradingvolume.Inordertoclaimthatnoisehasincreased,thenoiseshouldbe
normalizedfortradingvolume.

Whatalong,strangetripitsbeen
Readingisadangerousthing.Itcanlaunchyouoffintostrangedirections.ImovedfromCalifornia
toSantaFe,NewMexicobecauseIreadabook
(http://www.bearcave.com/bookrev/predictors.html).ThatonegraphinWIREDmagazine
launchedmedownapaththatIspentmanymonthsfollowing.Likeanyadventure,I'mnotsureifI
wouldhaveembarkedonthisoneifIhadknownhowlongand,attimes,difficult,thejourney
wouldbe.
Yearsago,whenitfirstcameout,IboughtacopyofthebookTheWorldAccordingtoWavelets
byBarbaraHubbard,onthebasisofareviewIreadinthemagazineScience.Thebooksaton
myshelfunreaduntilIsawtheWIREDgraph.
Waveletshavebeensomewhatofafad,abuzzwordthatpeoplehavethrownaround.Barbara
HubbardstartedwritingTheWorldAccordingtoWaveletswhenthewaveletfadwasstartingto
catchfire.Sheprovidesaninterestinghistoryofhowwaveletsdevelopedinthemathematicaland
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

5/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

engineeringworlds.Shealsomakesavaliantattempttoprovideanexplanationofwhatthe
wavelettechniqueis.Ms.Hubbardisasciencewriter,notamathematician,butshemastereda
fairamountofbasiccalculusandsignalprocessingtheory(whichIadmireherfor).Whenshe
wroteTheWorldAccordingtoWaveletstherewerefewbooksonwaveletsandnointroductory
material.AlthoughIadmireBarbaraHubbard'sheroiceffort,Ihadonlyasurfaceunderstandingof
waveletsafterreadingTheWorldAccordingtoWavelets.
Thereisavastliteratureonwaveletsandtheirapplications.Fromthepointofviewofasoftware
engineer(withonlyayearofcollegecalculus),theproblemwiththewaveletliteratureisthatithas
largelybeenwrittenbymathematicians,eitherforothermathematiciansorforstudentsin
mathematics.I'mnotamemberofeithergroup,soperhapsmyproblemisthatIdon'thavea
fluentgraspofthelanguageofmathematics.IcertianlyfeelthiswheneverIreadjournalarticles
onwavelets.However,Ihavetriedtoconcentrateonbooksandarticlesthatareexplicitly
introductoryandtutorial.Eventhesehaveproventobedifficult.
ThefirstchapterofthebookWaveletsMadeEasybyYvesNievergeltstartsoutwithan
explainationofHaarwavelets(thesearethewaveletsusedtogeneratethegraphpublishedin
WIRED).ThischapterhasnumerousexamplesandIwasabletounderstandandimplementHaar
waveletsfromthismaterial(linkstomyJavacodeforHaarwaveletscanbefoundbelow).Alater
chapterdiscussestheDaubechieswavelettransform.Unfortunately,thischapterofWavelets
MadeEasydoesnotseemtobeasgoodasthematerialonHaarwavelets.Thereappeartobea
numberoferrorsinthischapterandimplementingthealgorithmdescribedbyNievergeltdoesnot
resultinacorrectwavelettransform.Amongotherthings,thewaveletcoefficientsforthe
Daubechieswaveletsseemtobewrong.MywebpageontheDaubechieswavelettransformcan
befoundhere(/software/java/wavelets/daubechies/index.html).ThebookRipplesinMathematics
(seethereferencesattheendofthewebpage)isabetterreference.

Thewavelettechniqueforanalyzingasignalortime
series
Thereisavastliteratureonwavelets.Thisincludesthousandsofjournalarticlesandmany
books.ThebooksonwaveletsrangefromrelativelyintroductoryworkslikeNievergelt'sWavelets
MadeEasy(whichisstillnotlightreading)tobooksthatareaccessableonlytograduatestudents
inmathematics.ThereisalsoagreatdealofwaveletmaterialontheWeb.Thisincludesanumber
oftutorials(seeWebbasedreference(/misl/misl_tech/wavelets/haar.html#ReferencesWeb),
below).
Giventhevastliteratureonwavelets,thereisnoneedforyetanothertutorial.Butitmightbe
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

6/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

worthwhiletosummarizemyviewofwaveletsastheyareappliedto1Dsignalsortimeseries
(animageis2Ddata).Atimeseriesissimplyasampleofasignalorarecordofsomething,like
temperature,waterlevelormarketdata(likeequitycloseprice).
Waveletsallowatimeseriestobeviewedinmultipleresolutions.Eachresolutionreflectsa
differentfrequency.Thewavelettechniquetakesaveragesanddifferencesofasignal,breaking
thesignaldownintospectrum.AllthewaveletalgorithmsthatI'mfamiliarwithworkontimeseries
apoweroftwovalues(e.g.,64,128,256...).Eachstepofthewavelettransformproducestwo
setsofvalues:asetofaveragesandasetofdifferences(thedifferencesarereferredtoas
waveletcoefficients).Eachstepproducesasetofaveragesandcoefficientsthatishalfthesize
oftheinputdata.Forexample,ifthetimeseriescontains256elements,thefirststepwillproduce
128averagesand128coefficients.Theaveragesthenbecometheinputforthenextstep(e.g.,
128averagesresultinginanewsetof64averagesand64coefficients).Thiscontinuesuntilone
averageandonecoefficient(e.g.,20)iscalculated.
Theaverageanddifferenceofthetimeseriesismadeacrossawindowofvalues.Mostwavelet
algorithmscalculateeachnewaverageanddifferencebyshiftingthiswindowovertheinputdata.
Forexample,iftheinputtimeseriescontains256values,thewindowwillbeshiftedbytwo
elements,128times,incalculatingtheaveragesanddifferences.Thenextstepofthecalculation
usestheprevioussetofaverages,alsoshiftingthewindowbytwoelements.Thishastheeffect
ofaveragingacrossafourelementwindow.Logically,thewindowincreasesbyafactoroftwo
eachtime.
Inthewaveletliteraturethistreestructuredrecursivealgorithmisreferredtoasapyramidal
algorithm.
Thepoweroftwocoefficient(difference)spectrumgeneratedbyawaveletcalculationreflect
changeinthetimeseriesatvariousresolutions.Thefirstcoefficientbandgeneratedreflectsthe
highestfrequencychanges.Eachlaterbandreflectschangesatlowerandlowerfrequencies.
Thereareaninfinitenumberofwaveletbasisfunctions.Themorecomplexfunctions(likethe
Daubechieswavelets)produceoverlappingaveragesanddifferencesthatprovideabetter
averagethantheHaarwaveletatlowerresolutions.However,thesealgorithmsaremore
complicated.

TheLanguageofWavelets
Everyfieldofspecialtydevelopsitsownsublanguage.Thisiscertainlytrueofwavelets.I've
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

7/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

listedafewdefinitionsherewhich,ifIhadunderstoodtheirmeaningwouldhavehelpedmeinmy
wanderingsthroughthewaveletliterature.
Wavelet
Afunctionthatresultsinasetofhighfrequencydifferences,orwaveletcoefficients.Inlifting
scheme(/misl/misl_tech/wavelets/lifting/index.html)termsthewaveletcalculatesthe
differencebetweenapredictionandanactualvalue.
Ifwehaveadatasamples i,s i+1,s i+2...theHaarwaveletequationsis

Wherec iisthewaveletcoefficient.
ThewaveletLiftingScheme(/misl/misl_tech/wavelets/lifting/index.html)usesaslightly
differentexpressionfortheHaarwavelet:

ScalingFunction
Thescalingfunctionproducesasmootherversionofthedataset,whichishalfthesizeofthe
inputdataset.Waveletalgorithmsarerecursiveandthesmootheddatabecomestheinputfor
thenextstepofthewavelettransform.TheHaarwaveletscalingfunctionis

whereaiisasmoothedvalue.
TheHaartransformpreservestheaverageinthesmoothedvalues.Thisisnottrueofall
wavelettransforms.
Highpassfilter

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

8/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

Indigitalsignalprocessing(DSP)terms,thewaveletfunctionisahighpassfilter.Ahighpass
filterallowsthehighfrequencycomponentsofasignalthroughwhilesuppressingthelow
frequencycomponents.Forexample,thedifferencesthatarecapturedbytheHaarwavelet
functionrepresenthighfrequencychangebetweenanoddandanevenvalue.
Lowpassfilter
Indigitalsignalprocessing(DSP)terms,thescalingfunctionisalowpassfilter.Alowpass
filtersuppressesthehighfrequencycomponentsofasignalandallowsthelowfrequency
componentsthrough.TheHaarscalingfunctioncalculatestheaverageofanevenandanodd
element,whichresultsinasmoother,lowpasssignal.
Orthogonal(orOrthonormal)Transform
Thedefinitionoforthonormal(a.k.a.orthogonal)tranformsinWaveletMethodsforTimeSeries
AnalysisbyPercivalandWalden,CambridgeUniversityPress,2000,Chaper3,section3.1,
isoneofthebestI'veseen.I'vequotedthisbelow:

Orthonormaltransformsareofinterstbecausetheycanbeusedtore
expressatimeseriesinsuchawaythatwecaneasilyreconstructtheseries
fromitstransform.Inaloosesense,the"information"inthetransformisthus
equivalenttothe"information"istheoriginalseriestoputitanotherway,the
seriesanditstransformcanbeconsideredtobetworepresentationsofthe
samemathematicalentity.
Intermsofwavelettransformsthismeansthattheoriginaltimeseriescanbeexactly
reconstructedfromthetimeseriesaverageandcoefficientsgeneratedbyanorthogonal
(orthonormal)wavelettransform.
Signalestimation
Thisisalsoreferredtoas"denoising".Signalestimationalgorithmsattempttocharacterize
portionsofthetimeseriesandremovethosethatfallintoaparticularmodelofnoise.

ApplyingWaveletsandJavaSourceCode
TheseWebpagespublishsomeheavilydocumentedJavasourcecodefortheHaarwavelet
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

9/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

transform.BookslikeWaveletsMadeEasyexplainsomeofthemathematicsbehindthewavelet
transform.Ihavefound,however,thattheimplemationofthiscodecanbeatleastasdifficultas
understandingthewaveletequations.Forexample,theinplaceHaarwavelettransformproduces
waveletcoefficientsinabutterflypatternintheoriginaldataarray.TheJavasourcepublished
hereincludescodetoreorderthebutterflyintocoefficientspectrumswhicharemoreusefulwhen
itcomestoanalyzingthedata.Althoughthiscodeisnotlarge,ittookmemostofaSaturdayto
implementthecodetoreorderthebutterflydatapattern.
ThewaveletLiftingScheme,developedbyWimSweldensandothersprovidesasimplerwayto
lookasmanywaveletalgorithms.IstartedtoworkonLiftingSchemewaveletimplementations
afterIhadwrittenthiswebpageanddevelopedthesoftware.TheHaarwaveletcodeismuch
simplerwhenexpressedintheliftingscheme.SeemywebpageTheWaveletLiftingScheme
(/misl/misl_tech/wavelets/lifting/index.html).
ThelinktotheJavasourcedownloadWebpageisbelow.

FinancialTimeSeries
Thereareavarietyofwaveletanalysisalgorithms.Differentwaveletalgorithmsareappplied
dependingonthenatureofthedataanalyzed.TheHaarwavelet,whichisusedhereisveryfast
andworkswellforthefinancialtimeseries(e.g.,theclosepriceforastock).Financialtimeseries
arenonstationary(touseasignalprocessingterm).Thismeansthatevenwithinawindow,
financialtimeseriescannotbedescribedwellbyacombinationofsinandcosterms.Norare
financialtimeseriescyclicalinapredictablefashion(unlessyoubelieveinElliotwaves
(http://www.crbindex.com/techtip/tipv2n19.htm)).FinancialtimeserieslendthemselvestoHaar
waveletanalysissincegraphsoffinancialtimeseriestendtojagged,withoutalotofsmooth
detail.Forexample,thegraphbelowshowsthedailyclosepriceforAppliedMaterialsovera
periodofabouttwoyears.

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

10/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

DailyclosepriceforAppliedMaterials(symbol:AMAT),12/18/97to12/30/99.
TheHaarwaveletalgorithmsIhaveimplementedworkondatathatconsistsofsamplesthatarea
poweroftwo.Inthiscasethereare512samples.

WhyHaarWavelets?
Thereareawidevarietyofpopularwaveletalgorithms,includingDaubechies
(/software/java/wavelets/daubechies/index.html)wavelets,MexicanHatwaveletsandMorlet
wavelets.Thesewaveletalgorithmshavetheadvantageofbetterresolutionforsmoothly
changingtimeseries.Buttheyhavethedisadvantageofbeingmoreexpensivetocalculatethan
theHaarwavelets.Thehigerresolutionprovidedbythesewavletsisnotworththecostfor
financialtimeseries,whicharecharacterizedbyjaggedtransitions.

HaarWavelets
TheHaarwaveletalgorithmspublishedhereareappliedtotimeserieswherethenumberof
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

11/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

samplesisapoweroftwo(e.g.,2,4,8,16,32,64...)TheHaarwaveletusesarectangular
windowtosamplethetimeseries.Thefirstpassoverthetimeseriesusesawindowwidthoftwo.
Thewindowwidthisdoubledateachstepuntilthewindowencompassestheentiretimeseries.
Eachpassoverthetimeseriesgeneratesanewtimeseriesandasetofcoefficients.Thenew
timeseriesistheaverageoftheprevioustimeseriesoverthesamplingwindow.Thecoefficients
representtheaveragechangeinthesamplewindow.Forexample,ifwehaveatimeseries
consistingofthevaluesv0,v1,...vn,anewtimeseries,withhalfasmanypointsiscalculatedby
averagingthepointsinthewindow.Ifitisthefirstpassoverthetimeseries,thewindowwidthwill
betwo,sotwopointswillbeaveraged:
for(i=0;i<n;i=i+2)
si=(vi+vi+1)/2;
The3Dsurfacebelowgraphsninewaveletspectrumsgeneratedfromthe512pointAMATclose
pricetimeseries.Thexaxisshowsthesamplenumber,theyaxisshowstheaveragevalueat
thatpointandthezaxisshowslog2ofthewindowwidth.

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

12/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

Thewaveletcoefficientsarecalcalculatedalongwiththenewaveragetimeseriesvalues.The
coefficientsrepresenttheaveragechangeoverthewindow.Ifthewindowswidthistwothiswould
be:
for(i=0;i<n;i=i+2)
ci=(vivi+1)/2;
Thegraphbelowshowsthecoefficientspectrums.Asbeforethezaxisrepresentsthelog2ofthe
windowwidth.Theyaxisrepresentsthetimeserieschangeoverthewindowwidth.Somewhat
counterintutitively,thenegativevaluesmeanthatthetimeseriesismovingupward
viislessthanvi+1
so
vivi+1willbelessthanzero
Positivevaluesmeanthethetimeseriesisgoingdown,sincev iisgreaterthanv i+1.Notethatthe
highfrequencycoefficientspectrum(log2(windowWidth)=1)reflectsthenoisiestpartofthetime
series.Herethechangebetweenvaluesfluctuatesaroundzero.

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

13/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

PlotoftheHaarcoefficientspectrum.Thesurfaceplotsthehighestfrequencyspectruminthe
frontandthelowestfrequencyspectrumintheback.Notethatthehighestfrequencyspectrum
containsmostofthenoise.

FilteringSpectrum
Thewavelettransformallowssomeorallofagivenspectrumtoberemovedbysettingthe
coefficientstozero.Thesignalcanthenberebuiltusingtheinversewavelettransform.Plotsof
theAMATclosepricetimeserieswithvariousspectrumfilteredoutareshownhere.
(/misl/misl_tech/wavelets/spectrum_plots/index.html)

NoiseFilters
Eachspectrumthatmakesupatimeseriescanbeexaminedindependently.Anoisefiltercanbe
appliedtoeachspectrumremovingthecoefficientsthatareclassifiedasnoisebysettingthe
coefficientstozero.
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

14/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

Thiswebpage(/misl/misl_tech/wavelets/close_images/index.html)showsahistogramanalysis
ofthethreehighestfrequencyspectrumoftheAMATcloseprice.Theresultofafilterthat
removesthepointsthatfallwithinagaussiancurveineachspectrumisalsoshown.The
gaussiancurvehasameanandstandarddeviationofthecoefficientsinthatspectrum.
Anotherwaytoremovenoiseistousethresholding.Mywebpageoutliningonethresholding
algorithmcanbefoundhere(/misl/misl_tech/wavelets/noise.html).

Waveletsvs.SimpleFilters
HowdoHaarwaveletfilterscomparetosimplefilters,likewindowedmeanandmedianfilters?A
plotoftheAMATtimeseries,filteredwithamedianfilter(whichinthiscaseisvirtuallyidenticalto
ameanfilter)isshownherehere(/misl/misl_tech/wavelets/simple_filters/index.html).These
filterscanbecomparedtothespectrumfilters(whereagivenwaveletcoefficientspectrumis
fileredout)here.(/misl/misl_tech/wavelets/spectrum_plots/index.html).
Whetherawaveletfilterisbetterthanawindowedmeanfilterdependsontheapplication.The
waveletfilterallowsspecificpartsofthespectrumtobefiltered.Forexample,theentirehigh
frequencyspectrumcanberemoved.Orselectedpartsofthespectrumcanberemoved,asis
donewiththegaussiannoisefilter.ThepowerofHaarwaveletfiltersisthattheycanbeefficiently
calculatedandtheyprovidealotofflexibility.Theycanpotentiallyleavemoredetailinthetime
series,comparedtothemeanormedianfilter.Totheextentthatthisdetailisusefulforan
application,thewaveletfilterisabetterchoice.

LimitationsoftheHaarWaveletTransform
TheHaarwavelettransformhasanumberofadvantages:
Itisconceptuallysimple.
Itisfast.
Itismemoryefficient,sinceitcanbecalculatedinplacewithoutatemporaryarray.
Itisexactlyreversiblewithouttheedgeeffectsthatareaproblemwithotherwavelet
trasforms.
TheHaartransformalsohaslimitations,whichcanbeaproblemforsomeapplications.
Ingeneratingeachsetofaveragesforthenextlevelandeachsetofcoefficients,theHaar
transformperformsanaverageanddifferenceonapairofvalues.Thenthealgorithmshiftsover
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

15/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

bytwovaluesandcalculatesanotheraverageanddifferenceonthenextpair.
Thehighfrequencycoefficientspectrumshouldreflectallhighfrequencychanges.TheHaar
windowisonlytwoelementswide.Ifabigchangetakesplacefromanevenvaluetoanodd
value,thechangewillnotbereflectedinthehighfrequencycoefficients.
Forexample,inthe64elementtimeseriesgraphedbelow,thereisalargedropbetweenelements
16and17,andelements44and45.

Sincethesearehighfrequencychanges,wemightexpecttoseethemreflectedinthehigh
frequencycoefficients.However,inthecaseoftheHaarwavelettransformthehighfrequency
coefficientsmissthesechanges,sincetheyareoneventooddelements.
Thesurfacebelowshowsthreecoefficientspectrum:32,16and8(wherethe32element
coefficientspectrumisthehighestfrequency).Thehighfrequencyspectrumisplottedonthe
leadingedgeofthesurface.thelowestfrequencyspectrum(8)isthefaredgeofthesurface.

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

16/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

Notethatbothlargemagnitudechangesaremissingfromthehighfrequencyspectrum(32).The
firstchangeispickedupinthenextspectrum(16)andthesecondchangeispickedupinthelast
spectruminthegraph(8).
Manyotherwaveletalgorithms,liketheDaubechieswaveletalgorithm,useoverlappingwindows,
sothehighfrequencyspectrumreflectsallchangesinthetimeseries.LiketheHaaralgorithm,
Daubechiesshiftsbytwoelementsateachstep.However,theaverageanddifferenceare
calculatedoverfourelements,sothereareno"holes".
Thegraphbelowshowsthehighfrequencycoefficientspectrumcalculatedfromthesame64
elementtimeseries,butwiththeDaubechiesD4waveletalgorithm.Becauseoftheoverlapping
averagesanddifferencesthechangeisreflectedinthisspectrum.

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

17/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

The32,16and8coefficientspectrums,calculatedwiththeDaubechiesD4waveletalgorithm,are
shownbelowasasurface.Notethatthechangeinthetimeseriesisreflectedinallthree
coefficientspectrum.

WaveletsandParallelism
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

18/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

Waveletalgorithmsarenaturallyparallel.Forexample,ifenoughprocessingelementsexist,the
wavelettransformforaparticularspectrumcanbecalculatedinonestepbyassigninga
processorforeverytwopoints.Theparallelisminthewaveletalgorithmmakesitattractivefor
hardwareimplementation.

JavaSourceCodeDownload
TheWebpagefordownloadingtheHaarwaveletsourcecodecanbefoundhere

(/software/java/wavelets/index.html).ThisJavacodeisextensivelydocumentedandthisweb
pageincludesalinktotheJavadocgenerateddocumentation.
AsimplerversionoftheHaarwaveletalgorithmcanbefoundviamywebpageTheWavelet
LiftingScheme(/misl/misl_tech/wavelets/lifting/index.html).

Resources
GnuPlot
Theplotsabovearegeneratedwithgnuplot(http://www.gnuplot.org)forWindowsNT.Seemy
webpageofGnuplotlinkshere(/misl/misl_tech/plotting.html).Iamonlymarginallystatisified
withgnuplot.ThesoftwareiseasytouseandtheWindowsNTversioncomeswithanice
GUIandanicehelpsystem.However,whenitcomesto3Dplots,thesoftwareleavessome
thingstobedesired.Thehiddenlineremovalconsumesvastamountsofvirtualmemory.
WhenItriedtoplotoneofthecoefficientssurfaceswiththexandzaxesswitched,itranout
ofmemoryonaWindowsNTsystemwith256Kofvirtualmemory.Also,thesurfacewouldbe
mucheasiertounderstandifitcouldbecoloredwithaspectrum.Ifyouknowofabetter3D
plottingpackagethatrunsonWindowsNT,pleasedropmeanote.
Ihavealsohadahardtimegettinggnuplottogenerate2Dplotswithmultiplelinesthathave
differentcolors.Ihavesucceededindoingthisonlywhenthedataforeachlinewasina
separatefile,whichcanbeawkward.
ROOT:AnObjectOrientedDataAnalysisFramework(http://root.cern.ch/)
IwassentthereferencetoRootbyaphysicist,CostasA.Rootisadataanalysisframework
thatistargetedatthemassiveamountsofdatageneratedbyhighenergyphysics
experimentsatCERNandelsewhere.

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

19/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

AlthoughRootleansheavilytowardphysics,itlookstomelikeRootwouldbeusefulinother
areas.Someofthestatisticaltechniquesthatareusedtoanalyzeresultsinexperimental
physicsisalsousedinquantitivefinance,forexample.
RoothasdifferentgoalsthangnuPlot.Itistargetedatamuchmorechallengingdataanalysis
enviroment(terabytesofdata).ButithasalargelearningcurveandI'mskepticalifitcanbe
easilyusedbythosewhodonothaveasophisticatedcommandofC++.IncontrastgnuPlot
isasimpleplottingenvironment.Somysearchforabetterplottingenvironmentcontinues.I
knowthatsuchenvironmentsaresupportedbyMatlabandMathematics,butthesepackages
aretooexpensiveformylimitedsoftwarebudget.

References
Books
RipplesinMathematics:theDiscreteWaveletTransformbyJensenandlaCourHarbo,2001
SofarthisisthebestbookI'vefoundonwavelets.IreadthisbookafterIhadspentmonths
readingmanyofthereferencesthatfollow,soI'mnotsurehoweasythisbookwouldbefor
someonewithnopreviousexposuretowavelets.ButIhaveyettofindany"easy"reference.
RipplesinMathematicscoversLiftingSchemewaveletswhichareeasiertoimplementand
understand.Thebookiswrittenatarelativelyintroductorylevelandisaimedatengineers.
Theauthorsprovideimplementationsforanumberofwaveletalgorithms.Ripplesalsocovers
theproblemofapplyingwaveletalgorithmslikeDaubechiesD4tofinitedatasets(e.g.,they
coversomesolutionsfortheedgeproblemsencounteredforDaubechieswavelets).
WaveletsandFilterBanksbyGilbertStrangandTruongNguyen,WellesleyCambridgePr,
1996
Acolleaguerecommendthisbook,althoughhecouldnotloadittomesinceitispackedaway
inabox.Sadlythisbookishardtofind.Iboughtmycopyviaabebooks.com,used,froma
bookdealerinAustralia.WhileIwaswaitingforthebookIreadafewofGilbertStrang's
journalarticles.GilbertStrangisoneofthebestwritersI'veencounteredinmathematics.I
haveonlyjuststartedworkingthroughthisbook,butitlookslikeanexcellent,although
mathematical,bookonwavelets.
WaveletsMadeEasybyYvesNievergelt,Birkhauser,1999
ThisbookshastwoexcellentchaptersonHaarwavelets(Chapter1covers1DHaar
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

20/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

waveletsandChapter2covers2Dwavelets).AtleastinhiscoverageofHaarwavelts,Prof.
Nievergeltwritesclearlyandincludesplentyofexamples.ThecoverageofHaarwavelets
usesonlybasicmathematics(e.g.,algebra).
FollowingthechapteronHaarwaveletsthereisachapteronDaubechieswavelets.
Daubechieswaveletsarederivedfromageneralclassofwavelettransforms,whichincludes
Haarwavelets.Daubechieswaveletsarebetterforsmoothlychangingtimeseries,butare
probablyoverkillforfinancialtimeseries.AsWaveletsMadeEasyprogresses,itgetsless
easy.FollowingthechapteronDaubechieswaveletsisadiscussionofFouriertransforms.
Thelaterchaptersdelveintothemathematicsbehindwavelets.Prof.Nievergeltprettymuch
leftmebehindatthechapteronFouriertransforms.Foranapproachablediscussionof
Fouriertransforms,seeUnderstandingDigitalSignalProcessingbyRichardG.Lyons
(below).
AsWaveletsMadeEasyprogresses,itbecomeslessandlessusefulforwaveletalgorithm
implementation.Infact,whilethemathematicsNievergeltusestodescribeDaubechies
waveletsiscorrect,thealgorithmhedescribestoimplementtheDaubechiestransformand
inversetransformseemstobewrong.
WaveletsMadeEasydoesnotliveuptothe"easy"partofitstitle.Giventhisandtheapparent
errorsintheDaubechiescoverage,IamsorrytosaythatIcan'trecommendthisbook.Save
yourmoneyandbuyacopyofRipplesinMathematics.
DiscoveringWaveletsbyEdwardAboufadelandStevenSchlicker
At125pages,thisisoneofthemostexpensivewaveletbooksI'vepurchased,onaperpage
basis.ItsellsonAmazon(http://www.amazon.com)for$64.95US.Iboughtitusedfor$42.50.
IfDiscoveringWaveletsprovidedashort,cleardescriptionofwavelets,thelengthwouldbea
virtue,notafault.Sadlythisisnotthecase.DiscoveringWaveletsseemstobeabookwritten
forcollegestudentswhohavecompletedcalculusandlinearalgebra.Thebookisheavyon
theorms(whichareincompletelyexplained)andverysortonusefulexplaination.Ifoundthe
descriptionofwaveletsunnecessarilyobscure.Forexample,Haarwaveletsaredescribedin
termsoflinearalgebra.Theycanbemuchmoresimplydescribedintermsofsums,
differencesandthesocalledpyramidalalgorithm.
WhileDiscoveringWaveletscoverssomeimportantmaterial,itscoverageissoobscureand
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

21/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

cursorythatIfoundthebookuseless.Thebookresemblesasetoflecturenotesandisof
littleusewithoutthelecture(fortheirstudent'ssakeIhopethatAboufadelandSchlickerare
betterteachersthanwriters).ThisisabookthatIwishIhadnotpurchased.
WaveletMethodsforTimeSeriesAnalysisbyDonaldB.PercivalandAndrewT.Walden,
CambridgeUniversityPress,2000
I'mnotamathematicianandIdon'tplayoneontelevision.Sothisbookisheavygoingforme.
Nevertheless,thisisagoodbook.Forsomeonewithabettermathematicalbackgroundthis
mightbeanexcellentbook.Theauthorsprovideacleardiscussionofwaveletsandavariety
oftimeseriesanalsysistechniques.Unlikesomemathematicians,PercivalandWalden
actuallycodedupthewaveletalgorithmsandunderstandthedifficultiesofimplementation.
Theycomparevariouswaveletfamiliesforvariousapplicationsandchosethesimplestone
(Haar)insomecases.
OneofthegreatbenifitsofWaveletMethodsforTimeSeriesAnalysisisthatitprovidesa
clearsummaryofagreatdealoftherecentresearch.ButPercivalandWaldenputthe
researchinanappliedcontext.ForexampleDonohoandJohnstonepublishedanequationfor
waveletnoisereduction.IhavebeenunabletofindalloftheirpapersontheWebandIhave
neverunderstoodhowtocalculatesomeofthetermsintheequationinpractice.Ifoundthis
definitioninWaveletMethods.
TheWorldAccordingtoWavelets:TheStoryofaMathematicalTechniqueintheMakingby
BarbaraBurkeHubbard,A.K.Peters,1996
Thisbookprovidesaninterestinghistoryofthedevelopmentofwavelets.Thisincludes
sketchesofmanyofthepeopleinvolvedinpioneeringtheapplicationandmathematicaltheory
behindwavelets.AlthoughMs.Hubbardmakesaheroiceffort,Ifoundtheexplainationof
waveletsdifficulttofollow.
TheCartoonGuideToStatisticsbyLarryGonicandWoollcottSmith,HarperCollins
Iworkwithanumberofmathematicians,soit'sabitembarrassingtohavethisbookonmy
disk.Inevertookstatistics.IncollegeeveryoneIknewwhotookstatisticsdidn'tlikeit.Since
itwasnotrequiredformymajor(ascalculuswas),Ididnottakestatistics.I'vecometo
understandhowusefulstatisticsis.IwantedtofilteroutGaussiannoise,soIneededto
understandnormalcurves.Althoughthetitleisabitembarrassing,TheCartoonGuideto
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

22/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

Statisticsprovidedaveryrapidandreadableintroductiontostatistics.
UnderstandingDigitalSignalProcessingbyRichardG.Lyons.
Thisbookisfantastic.Perhapsthebestintroductorybookeverwrittenondigitalsignal
processing.Itisthebookonsignalprocessingforsoftwareengineerslikemyselfwithtepid
mathematicalbackgrounds.ItprovidesthebestcoverageI'veeverseenonDFTsandFFTs.
Infact,thisbookhasinspiredmetotryFFTsonfinancialtimeseries
(http://www.bearcave.com/misl/misl_tech/signal/nonstat/index.html)(aninteresting
experiment,butwaveletsproducebetterresultsandFouriertransformsonnonstationarytime
series).
SeemywebpageANotebookCompiledWhileReadingUnderstandingDigitalSignal
ProcessingbyLyons(/misl/misl_tech/signal/index.html)

Webbasedreferences
Mywebpage(/misl/misl_tech/wavelets/lifting/index.html)onthewaveletLiftingScheme.The
HaarwaveletalgorithmexpressedusingthewaveletLiftingSchemeisconsiderablysimpler
thanthealgorithmreferencedabove.TheLiftingSchemealsoallowsHaarwavelettobe
extendedintoawaveletalgorithmsthathaveperfectreconstructionandhavebettermultiscale
resolutionthanHaarwavelets.
HaarWaveletTransform(http://dmr.ath.cx/gfx/haar/)byEmilMikulic
EmilMikulichaspublishedasimpleexplainationoftheHaartransform,forboth1Dand2D
data.Forthosewhofindmyexplainationobscure,thismightbeagoodresource.
TheWaveletTutorial(http://engineering.rowan.edu/~polikar/WAVELETS/WTtutorial.html):The
Engineer'sUltimateGuidetoWaveletAnalysis,byRobiPolikar.
The"ultimateguide"towaveletanalysishasyettobewritten,atleastformypurposes.But
Prof.Polikar'sWaveletTutorialisexcellent.WhenitcomestoexplainingWaveletsand
Fouriertransforms,thisisoneofthebestoverviewsI'veseen.Prof.Polikarputagreatdealof
workintothistutorialandIamgreatefulforhiseffort.However,therewasnotsufficientdetail
inthistutorialtoallowmetocreatemyownwaveletandinversewavelettranformsoftware.
AReallyFriendlyGuidetoWavelets
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

23/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

(http://perso.wanadoo.fr/polyvalens/clemens/wavelets/wavelets.html)
ThisWebpage(whichisalsoavailableinPDF)providesaniceoverviewofthetheorybehind
wavelets.ButaswithRobiPolikar'swebpage,itsabigstepfromthismaterialtoasoftware
implementation.WhetherthisWebpageis"reallyfriendly"dependsonwhoyourfriendsare.If
youfriendsarecalculusandtaylorseries,thenthispaperisforyou.Afterworkingmyway
throughagoodpartofWaveletsMadeEasythispaperfilledinsomeholeforme.ButIwould
nothaveunderstooditifIhadreaditbeforeWaveletsMadeEasy.
BellLabsWaveletsGroupHomePage(http://cm.belllabs.com/who/jelena/Wavelet/)
WimSweldens,whohaspublishedalotofmaterialontheWeb(heistheeditorofWavelet
Digest(http://www.wavelet.org))andelsewhereonWaveletsisamemberofthisgroup.An
interestingsitewithlotsofgreatlinkstootherwebresources.
SeealsoWimSwelden'sWaveletCascadeJavaApplet(http://netlib.bell
labs.com/cm/ms/who/wim/cascade/index.html)
LiftingSchemeWavelets
WinSweldensandIngridDaubechiesinventedanewwavelettechniqueknownasthelifting
scheme.GabrielFernandezhaspublishedanexcellentbibliographyontheliftingscheme
waveletswhichcanbefoundhere(http://www.cse.sc.edu/~fernande/liftpack/liftbibl.html).This
bibliographyhasapointertoWimSweldens'andPeterSchroder'sliftingschemetutorial
BuildingYourOwnWaveletsatHome.
ClemensValenshaswrittenatutorial
(http://perso.wanadoo.fr/polyvalens/clemens/lifting/lifting.html)onthefastliftingwavelet
transform.Thisisarathermathematicallyorientedtutorial.Formany,WimSweldens'paper
BuildingYourOwnhWavletsatHomemaybeeasiertounderstand(althoughIstillfoundthis
paperheavygoing).
GabrielFernandezhasdevelopedLiftPack.TheLiftPackHomePage
(http://www.cse.sc.edu/~fernande/liftpack/index.html)publishestheLiftPacksoftware.The
bibliographyisasubpageoftheLiftPackHomepage.
SeealsomywebpageonTheWaveletLiftingScheme
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

24/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

(/misl/misl_tech/wavelets/lifting/index.html).
WaveletsinComputerGraphis
OneofthepapersreferencedinGabrielFernandez'sliftingschemebibliographyisWim
SweldensandPeterSchroder'spaperBuildingYourOwnWaveletsatHome.Thisispartofa
courseonWaveletsinComputerGraphics
(http://www.multires.caltech.edu/teaching/courses/waveletcourse/)givenatSigGraph1994,
1995and1996.ThesigGraphcoursecoverdanamazingamountofmaterial.BuildingYour
OwnWaveletsatHomewasapparentlycoveredinamorning.Therearealotof
mathematicallygiftedpeopleincomputergraphics.Butevenforthesepeople,thislookslike
toughgoingforamorning.I'vespenthoursreadingandrereadingthistutorialbeforeI
understooditenoughtoimplementthepolynomialinterpolationwaveletsthatitdiscusses.
WaveletShrinkage(denoising)()
D.Donoho(http://wwwstat.stanford.edu/~donoho/)WaveletShrinkageandW.V.D.A
TenMinuteTour(http://wwwstat.stanford.edu/~donoho/Reports/1993/toulouse.ps.Z),
(figures)(http://wwwstat.stanford.edu/~donoho/Reports/1993/toulouse_figs.ps.Z),1993.
D.Donoho(http://wwwstat.stanford.edu/~donoho/)DeNoisingBySoftThresholding
(http://wwwstat.stanford.edu/~donoho/Reports/1992/denoiserelease3.ps.Z),IEEETrans.
onInformationTheory,Vol41,No.3,pp.613627,1995.
D.Donoho(http://wwwstat.stanford.edu/~donoho/)AdaptingtoUnknownSmoothness
viaWaveletShrinkage(http://spib.rice.edu/spib/papers/dsp/1994.02/009.ps.Z),JASA,
1995.
CalTechMultiResolutionModelingGroupPublications
TheWaveletsinComputerGraphicspage,referencedabove,isoneofthelinksfromthe
CalTechMultiresolutionModelingGroupPublications(http://www.multires.caltech.edu/pubs/)
webpage.Thewaveletpublicationsreferencedonthispageconcentrateonwavelet
applicationsforcomputergraphics.
TutorialonContinuousWaveletAnalysisofExperiementalData
(http://www.mame.syr.edu/faculty/lewalle/tutor/tutor.html)byJacquesLewalle,Syracuse
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

25/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

University.
Thisisyetanother"introductory"tutorialbyamathematician.Itgivesafeelingforwhatyou
candowithwavelets,butthereisnotenoughdetailtounderstandthedetailsofimplementing
waveletcode.
Amara'sWaveletPage(http://www.amara.com/current/wavelet.html)
AmaraGraps'webpageprovidessomegoodbasicintroductorymaterialonwaveletsand
someexcellentlinkstootherWebresources.Thereisalsoalinktotheauthor's(Amara)
IEEEComputationalSciencesandEngineeringarticleonwavelets.
Wave++fromRyersonPolytechnicUniversityComputationalSignalsAnalysisGroup
Wave++(http://www.scs.ryerson.ca/~lkolasa/CppWavelets.html)isaC++classlibraryfor
waveletandsignalanalysis.Thislibraryisprovidedinsourceform.Ihavenotexamineditin
detailyet.
Waveletandsignalprocessingalgorithmsareusuallyfairlysimple(theyconsistofarelatively
smallamountofcode).Myexperiencehasbeenthattheimplementationofthealgorithmsis
notastimeconsumingasunderstandingthealgorithmsandhowtheycanbeapplied.Since
oneofthebestwaystounderstandthealgorithmsistoimplementandapplythem,I'mnot
surehowmuchleverageWave++providesunlessyoualreadyunderstandwavelet
algorithms.
WaveletCompressionArrives(http://www.seyboldreports.com/SRIP/wavelet/)byPeter
Dyson,SeyboldReports,April1998.
Thisisanincreasinglydateddiscussiononwaveletcompressionproducts,especiallyfor
images.Thedescriptionofthecompressionproductsstrengthsandweaknessesisgood,but
thedescriptionofwaveletsispoor.
WebpageofZbigniewR.Struzik(http://www.cwi.nl/~zbyszek/)
Prof.ZbigniewR.StruzikofCentrumvoorWiskundeenInformaticaintheNetherlandshas
donesomeveryinterestingworkwithwaveletsinavarietyofareas,includingdataminingin
finance.ThiswebpagehasalinktoProf.Struzik'spublications(atthebottomoftheWeb
http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

26/27

7/13/2015

ApplyingtheHaarWaveletTransformtoTimeSeriesInformation

page).Prof.Struzik'sworkalsoshowssomeinterestingconnectionsbetweenfractalsand
wavelets.

Disclaimer
Thiswebpagewaswrittenonnightsandweekends,usingmycomputerresources.ThisWeb
pagedoesnotnecessarilyreflecttheviewsofmyemployer(atthetimethiswebpagewas
written).Nothingpublishedhereshouldbeinterpretedasareflectiononanytechniquesusedby
myemployer(atthattime).
IanKaplan,July2001
Revised:February2004
(mailto:iank@bearcave.com)

http://www.bearcave.com/misl/misl_tech/wavelets/haar.html

27/27

Você também pode gostar