8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

ContinuousWaveletTransform
DefinitionoftheContinuousWaveletTransform
LiketheFouriertransform,thecontinuouswavelettransform(CWT)usesinnerproductstomeasurethesimilaritybetween
jt

asignalandananalyzingfunction.IntheFouriertransform,theanalyzingfunctionsarecomplexexponentials, e
.The
resultingtransformisafunctionofasinglevariable,.IntheshorttimeFouriertransform,theanalyzingfunctionsare
jt

windowedcomplexexponentials, w(t)e
,andtheresultinafunctionoftwovariables.TheSTFTcoefficients, F (, ),
representthematchbetweenthesignalandasinusoidwithangularfrequencyinanintervalofaspecifiedlength
centeredat.
IntheCWT,theanalyzingfunctionisawavelet,.TheCWTcomparesthesignaltoshiftedandcompressedorstretched
versionsofawavelet.Stretchingorcompressingafunctioniscollectivelyreferredtoasdilationorscalingandcorresponds
tothephysicalnotionofscale.Bycomparingthesignaltothewaveletatvariousscalesandpositions,youobtaina
functionoftwovariables.Thetwodimensionalrepresentationofaonedimensionalsignalisredundant.Ifthewaveletis
complexvalued,theCWTisacomplexvaluedfunctionofscaleandposition.Ifthesignalisrealvalued,theCWTisareal
valuedfunctionofscaleandposition.Forascaleparameter,a>0,andposition,b,theCWTis:

C(a, b f (t), (t)) =

t b
) dt
a

f (t) G (
a

where denotesthecomplexconjugate.NotonlydothevaluesofscaleandpositionaffecttheCWTcoefficients,the
choiceofwaveletalsoaffectsthevaluesofthecoefficients.
Bycontinuouslyvaryingthevaluesofthescaleparameter,a,andthepositionparameter,b,youobtainthecwtcoefficients
C(a,b).Notethatforconvenience,thedependenceoftheCWTcoefficientsonthefunctionandanalyzingwavelethasbeen
suppressed.
Multiplyingeachcoefficientbytheappropriatelyscaledandshiftedwaveletyieldstheconstituentwaveletsoftheoriginal
signal.

TherearemanydifferentadmissiblewaveletsthatcanbeusedintheCWT.Whileitmayseemconfusingthatthereareso
manychoicesfortheanalyzingwavelet,itisactuallyastrengthofwaveletanalysis.Dependingonwhatsignalfeatures
youaretryingtodetect,youarefreetoselectawaveletthatfacilitatesyourdetectionofthatfeature.Forexample,ifyouare
tryingtodetectabruptdiscontinuitiesinyoursignal,youmaychooseonewavelet.Ontheotherhand,ifyouareinteresting
infindingoscillationswithsmoothonsetsandoffsets,youarefreetochooseawaveletthatmorecloselymatchesthat
behavior.

Scale
Liketheconceptoffrequency,scaleisanotherusefulpropertyofsignalsandimages.Forexample,youcananalyze
temperaturedataforchangesondifferentscales.Youcanlookatyeartoyearordecadetodecadechanges.Ofcourse,
youcanexaminefiner(daytoday),orcoarserscalechangesaswell.Someprocessesrevealinterestingchangesonlong
time,orspatialscalesthatarenotevidentonsmalltimeorspatialscales.Theoppositesituationalsohappens.Someof
ourperceptualabilitiesexhibitscaleinvariance.Yourecognizepeopleyouknowregardlessofwhetheryoulookatalarge
portrait,orsmallphotograph.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

1/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Togobeyondcolloquialdescriptionssuchas"stretching"or"shrinking"weintroducethescalefactor,oftendenotedbythe
lettera.Thescalefactorisainherentlypositivequantity,a>0.Forsinusoids,theeffectofthescalefactorisveryeasyto
see.

Insin(at),thescaleistheinverseoftheradianfrequency,a.
Thescalefactorworksexactlythesamewithwavelets.Thesmallerthescalefactor,themore"compressed"thewavelet.
Conversely,thelargerthescale,themorestretchedthewavelet.Thefollowingfigureillustratesthisforwaveletsatscales
1,2,and4.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

2/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Thisgeneralinverserelationshipbetweenscaleandfrequencyholdsforsignalsingeneral.SeeCWTasaFiltering
TechniqueandScaleandFrequencyformoreinformationontherelationshipbetweenscaleandfrequency.
Notonlyisatimescalerepresentationadifferentwaytoviewdata,itisaverynaturalwaytoviewdataderivedfroma
greatnumberofnaturalphenomena.
ScaleandFrequency
Thereisclearlyarelationshipbetweenscaleandfrequency.Recallthathigherscalescorrespondtothemost"stretched"
wavelets.Themorestretchedthewavelet,thelongertheportionofthesignalwithwhichitisbeingcompared,and
thereforethecoarserthesignalfeaturesmeasuredbythewaveletcoefficients.

Tosummarize,thegeneralcorrespondencebetweenscaleandfrequencyis:

LowscaleaCompressedwaveletRapidlychangingdetailsHighfrequency.

HighscaleaStretchedwaveletSlowlychanging,coarsefeaturesLowfrequency.

Whilethereisageneralrelationshipbetweenscaleandfrequency,nopreciserelationshipexists.Usersfamiliarwith
Fourieranalysisoftenwanttodefineamappingbetweenawaveletatagivenscalewithaspecifiedsamplingperiodtoa
frequencyinhertz.Youcanonlydothisinageneralsense.Therefore,itisbettertotalkaboutthepseudofrequency
correspondingtoascale.TheWaveletToolboxsoftwareprovidestwofunctionscentfrqandscal2frq,whichenable
youtofindtheseapproximatescalefrequencyrelationshipsforspecifiedwaveletsandscales.
ThebasicapproachidentifiesthepeakpowerintheFouriertransformofthewaveletasitscenterfrequencyanddivides
thatvaluebytheproductofthescaleandthesamplinginterval.Seescal2frqfordetails.Thefollowingexampleshows
thematchbetweentheestimatedcenterfrequencyofthedb8waveletandasinusoidofthesamefrequency.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

3/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

TherelationshipbetweenscaleandfrequencyintheCWTisalsoexploredinCWTasaFilteringTechnique.

Shifting
Shiftingawaveletsimplymeansdelaying(oradvancing)itsonset.Mathematically,delayingafunctionf(t)bykis
representedbyf(tk):

CWTasaWindowedTransform
InShortTimeFourierTransform,theSTFTisdescribedasawindowingofthesignaltocreatealocalfrequencyanalysis.A
shortcomingoftheSTFTapproachisthatthewindowsizeisconstant.Thereisatradeoffinthechoiceofwindowsize.A
longertimewindowimprovesfrequencyresolutionwhileresultinginpoorertimeresolutionbecausetheFouriertransform
losesalltimeresolutionoverthedurationofthewindow.Conversely,ashortertimewindowimprovestimelocalization
whileresultinginpoorerfrequencyresolution.
Waveletanalysisrepresentsthenextlogicalstep:awindowingtechniquewithvariablesizedregions.Waveletanalysis
allowstheuseoflongtimeintervalswhereyouwantmorepreciselowfrequencyinformation,andshorterregionswhere
youwanthighfrequencyinformation.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

4/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Thefollowingfigurecontraststime,frequency,timefrequency,andtimescalerepresentationsofasignal.

CWTasaFilteringTechnique
Thecontinuouswavelettransform(CWT)computestheinnerproductofasignal, f (t) ,withtranslatedanddilatedversions
ofananalyzingwavelet, (t). ThedefinitionoftheCWTis:

C(a, b f (t), (t)) =

t b
) dt
a

f (t) G (
a

YoucanalsointerprettheCWTasafrequencybasedfilteringofthesignalbyrewritingtheCWTasaninverseFourier
transform.

C(a, b f (t), (t)) =

1
2

f ()

a ((a)) e

jb

where f () and ()
aretheFouriertransformsofthesignalandthewavelet.
Fromtheprecedingequations,youcanseethatstretchingawaveletintimecausesitssupportinthefrequencydomainto
shrink.Inadditiontoshrinkingthefrequencysupport,thecenterfrequencyofthewaveletshiftstowardlowerfrequencies.
Thefollowingfiguredemonstratesthiseffectforahypotheticalwaveletandscale(dilation)factorsof1,2,and4.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

5/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

ThisdepictstheCWTasabandpassfilteringoftheinputsignal.CWTcoefficientsatlowerscalesrepresentenergyinthe
inputsignalathigherfrequencies,whileCWTcoefficientsathigherscalesrepresentenergyintheinputsignalatlower
frequencies.However,unlikeFourierbandpassfiltering,thewidthofthebandpassfilterintheCWTisinversely
proportionaltoscale.ThewidthoftheCWTfiltersdecreaseswithincreasingscale.Thisfollowsfromtheuncertainty
relationshipsbetweenthetimeandfrequencysupportofasignal:thebroaderthesupportofasignalintime,thenarrower
itssupportinfrequency.Theconverserelationshipalsoholds.
Inthewavelettransform,thescale,ordilationoperationisdefinedtopreserveenergy.Topreserveenergywhileshrinking
thefrequencysupportrequiresthatthepeakenergylevelincreases.Thequalityfactor,orQfactorofafilteristheratioofits
peakenergytobandwidth.Becauseshrinkingorstretchingthefrequencysupportofawaveletresultsincommensurate
increasesordecreasesinitspeakenergy,waveletsareoftenreferredtoasconstantQfilters.

DFTBasedContinuousWaveletTransform
TheequationintheprecedingsectiondefinedtheCWTastheinverseFouriertransformofaproductofFouriertransforms.

C(a, b f (t), (t)) =

1
2

f ()

a (a)e

jb

ThetimevariableintheinverseFouriertransformisthetranslationparameter,b.
ThissuggeststhatyoucancomputetheCWTwiththeinverseFouriertransform.Becausethereareefficientalgorithmsfor
thecomputationofthediscreteFouriertransformanditsinverse,youcanoftenachieveconsiderablesavingsbyusingfft
andifftwhenpossible.
ToobtainapictureoftheCWTintheFourierdomain,startwiththedefinitionofthewavelettransform:

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

6/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

< f (t), a,b(t) >= G

t b
)dt
a

f (t)(

Ifyoudefine:
1
a(t) = G (t/a)
a

youcanrewritethewavelettransformas

(f a)(b) =

f (t) a(b t)dt

whichexplicitlyexpressestheCWTasaconvolution.
ToimplementthediscretizedverionoftheCWT,assumethattheinputsequenceisalengthNvector,x[n].Thediscrete
versionoftheprecedingconvolutionis:

W a[b] =

N1

n=0

x[n] a[b n]

ToobtaintheCWT,itappearsyouhavetocomputetheconvolutionforeachvalueoftheshiftparameter,b,andrepeatthis
processforeachscale,a.
However,ifthetwosequencesarecircularlyextended(periodizedtolengthN),youcanexpressthecircularconvolution
asaproductofdiscreteFouriertransforms.TheCWTistheinverseFouriertransformoftheproduct

W a(b) =

1
N

2a
t

N1

X (2k/Nt) (a2k/Nt)e

j2kb/N

k=0

wheretisthesamplinginterval(period).
ExpressingtheCWTasaninverseFouriertransformenablesyoutousethecomputationallyefficientfftandifft
algorithmstoreducethecostofcomputingconvolutions.
ThecwtftfunctionimplementstheCWTusinganFFTbasedalgorithm.Seecwtftinfoforinformationpertainingtothe
supportedanalyzingwavelets.

InverseContinuousWaveletTransform
TheicwtftfunctionimplementstheinverseCWT.UsingicwtftrequiresthatyouobtaintheCWTfromcwtft.The
WaveletToolboxdoesnotsupporttheinverseCWTforageneralCWTobtainedusingcwt.
BecausetheCWTisaredundanttransform,thereisnotauniquewaytodefinetheinverse.TheinverseCWTimplemented
intheWaveletToolboxutilizesadiscreteversionofthesingleintegralformuladuetoMorlet.
TheinverseCWTisclassicallypresentedinthedoubleintegralform.AssumeyouhaveawaveletwithaFouriertransform
thatsatisfiestheadmissibilitycondition:

C =

2
|()|
d
||

<

Forwaveletssatisfyingtheadmissibilityconditionandfiniteenergyfunctions,f(t),youcandefinetheinverseCWTas:

f (t) =

1
C

< f (t), a,b(t) > a,b(t) db

da
2
a

Foranalyzingwaveletsandfunctionssatisfyingthefollowingconditions,asingleintegralformulafortheinverseCWT
exists.Theseconditionsare:

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

7/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Theanalyzedfunction,f(t),isrealvaluedandtheanalyzingwavelethasarealvaluedFouriertransform.

Theanalyzedfunction,f(t),isrealvaluedandtheFouriertransformoftheanalyzingwavelethassupportonlyontheset
ofnonnegativefrequencies.Thisisreferredtoasananalyticwavelet.AfunctionwhoseFouriertransformonlyhas
supportonthesetofnonnegativefrequenciesmustbecomplexvalued.

Theprecedingconditionsconstrainthesetofpossibleanalyzingwavelets.Ifyouinspectthelistofwaveletssupportedby
cwtft,eachwaveletiseitheranalyticorhasarealvaluedFouriertransform.Becausethetoolboxonlysupportsthe
analysisofrealvaluedfunctions,therealvaluedconditionontheanalyzedfunctionisalwayssatisfied.
Tomotivatethesingleintegralformula,let 1and 2betwowaveletsthatsatisfythefollowingtwowaveletadmissibility
condition:

() ()
1 2

d <

Definetheconstant:

() ()
1
2 d
C , =

1 2

INotethattheaboveconstantmaybecomplexvalued.Letf(t)andg(t)betwofiniteenergyfunctions.Ifthetwowavelet
admissibilityconditionissatisfied,thefollowingequalityholds:

,
1 2

< f, g >=

< f, 1 >< g, 2 > db

da
2
a

where<,>denotestheinnerproduct,*denotesthecomplexconjugate,andthedependenceof 1and 2onscaleand

positionhasbeensuppressedforconvenience.
ThekeytothesingleintegralformulafortheinverseCWTistorecognizethatthetwowaveletadmissibilityconditioncan
besatisfiedevenifoneofthewaveletsisnotadmissible.Inotherwords,itisnotnecessarythatboth 1and 2be
separatelyadmissible.Youcanalsorelaxtherequirementsfurtherbyallowingoneofthefunctionsandwaveletstobe
distributions.Byfirstlettingg(t)betheDiracdeltafunction(adistribution)andalsoallowing 2tobetheDiracdelta
function,youcanderivethesingleintegralformulafortheinverseCWT

f (t) = 2 Re
C

,
1

< f (t), 1(t) >

whereRe{}denotestherealpart.

da
3/2
a

TheprecedingequationdemonstratesthatyoucanreconstructthesignalbysummingthescaledCWTcoefficientsoverall
scales.
BysummingthescaledCWTcoefficientsfromselectscales,youobtainanapproximationtotheoriginalsignal.Thisis
usefulinsituationswhereyourphenomenonofinterestislocalizedinscale.
icwtftimplementsadiscretizedversionoftheaboveintegral.

ContinuousWaveletTransformAlgorithm
ThefollowingoutlinesthebasicalgorithmfortheCWT:
1. Takeawaveletandcompareittoasectionatthestartoftheoriginalsignal.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

8/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

2. Calculateanumber,C,thatrepresentshowcloselycorrelatedthewaveletiswiththissectionofthesignal.Thelarger
thenumberCisinabsolutevalue,themorethesimilarity.ThisfollowsfromthefacttheCWTcoefficientsarecalculated
withaninnerproduct.SeeInnerProductsformoreinformationonhowinnerproductsmeasuresimilarity.Ifthesignal
energyandthewaveletenergyareequaltoone,Cmaybeinterpretedasacorrelationcoefficient.Notethat,in
general,thesignalenergydoesnotequaloneandtheCWTcoefficientsarenotdirectlyinterpretableascorrelation
coefficients.
AsdescribedinDefinitionoftheContinuousWaveletTransform,theCWTcoefficientsexplicitlydependonthe
analyzingwavelet.Therefore,theCWTcoefficientsaredifferentwhenyoucomputetheCWTforthesamesignalusing
differentwavelets.

3. Shiftthewavelettotherightandrepeatsteps1and2untilyou'vecoveredthewholesignal.

4. Scale(stretch)thewaveletandrepeatsteps1through3.

5. Repeatsteps1through4forallscales.

InterpretingCWTCoefficients
BecausetheCWTisaredundanttransformandtheCWTcoefficientsdependonthewavelet,itcanbechallengingto
interprettheresults.
TohelpyouininterpretingCWTcoefficients,itisbesttostartwithasimplesignaltoanalyzeandananalyzingwaveletwith
asimplestructure.
Asignalfeaturethatwaveletsareverygoodatdetectingisadiscontinuity,orsingularity.Abrupttransitionsinsignalsresult
inwaveletcoefficientswithlargeabsolutevalues.
Forthesignalcreateashiftedimpulse.Theimpulseoccursatpoint500.
x=zeros(1000,1);
x(500)=1;
Forthewavelet,picktheHaarwavelet.
[~,psi,xval]=wavefun('haar',10);
plot(xval,psi);axis([011.51.5]);
title('HaarWavelet');

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

9/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

TocomputetheCWTusingtheHaarwaveletatscales1to128,enter:
CWTcoeffs=cwt(x,1:128,'haar');
CWTcoeffsisa128by1000matrix.EachrowofthematrixcontainstheCWTcoefficientsforonescale.Thereare128
rowsbecausetheSCALESinputtocwtis1:128.Thecolumndimensionofthematrixmatchesthelengthoftheinputsignal.
RecallthattheCWTofa1Dsignalisafunctionofthescaleandpositionparameters.ToproduceaplotoftheCWT
coefficients,plotpositionalongthexaxis,scalealongtheyaxis,andencodethemagnitude,orsizeoftheCWT
coefficientsascolorateachpointinthexy,ortimescaleplane.
Youcanproducethisplotusingcwtwiththeoptionalinputargument'plot'.
cwt(x,1:128,'haar','plot');
colormapjet;colorbar;

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

10/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

TheprecedingfigurewasmodifiedwithtextlabelstoexplicitlyshowwhichcolorsindicatelargeandsmallCWT
coefficients.
YoucanalsoplotthesizeoftheCWTcoefficientsin3Dwith
cwt(x,1:64,'haar','3Dplot');colormapjet;
wherethenumberofscaleshasbeenreducedtoaidinvisualization.
ExaminingtheCWToftheshiftedimpulsesignal,youcanseethatthesetoflargeCWTcoefficientsisconcentratedina
narrowregioninthetimescaleplaneatsmallscalescenteredaroundpoint500.Asthescaleincreases,thesetoflarge
CWTcoefficientsbecomeswider,butremainscenteredaroundpoint500.Ifyoutracetheborderofthisregion,itresembles
thefollowingfigure.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

11/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Thisregionisreferredtoastheconeofinfluenceofthepointt=500fortheHaarwavelet.Foragivenpoint,theconeof
influenceshowsyouwhichCWTcoefficientsareaffectedbythesignalvalueatthatpoint.
Tounderstandtheconeofinfluence,assumethatyouhaveawaveletsupportedon[C,C].Shiftingthewaveletbyband
scalingbyaresultsinawaveletsupportedon[Ca+b,Ca+b].Forthesimplecaseofashiftedimpulse, (t ) ,theCWT
coefficientsareonlynonzeroinanintervalaroundequaltothesupportofthewaveletateachscale.Youcanseethisby
consideringtheformalexpressionoftheCWToftheshiftedimpulse.

C(a, b (t ), (t)) =

t b
)dt
a

(t ) G (
a

b
)
a

= G (
a

Fortheimpulse,theCWTcoefficientsareequaltotheconjugated,timereversed,andscaledwaveletasafunctionofthe
shiftparameter,b.YoucanseethisbyplottingtheCWTcoefficientsforaselectfewscales.
subplot(311)
plot(CWTcoeffs(10,:));title('Scale10');
subplot(312)
plot(CWTcoeffs(50,:));title('Scale50');
subplot(313)
plot(CWTcoeffs(90,:));title('Scale90');

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

12/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Theconeofinfluencedependsonthewavelet.Youcanfindandplottheconeofinfluenceforaspecificwaveletwith
conofinf.
Thenextexamplefeaturesthesuperpositionoftwoshiftedimpulses, (t 300) + (t 500) .Inthiscase,usethe
Daubechies'extremalphasewaveletwithfourvanishingmoments,db4.Thefollowingfigureshowstheconeofinfluence
forthepoints300and500usingthedb4wavelet.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

13/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Lookatpoint400forscale20.Atthatscale,youcanseethatneitherconeofinfluenceoverlapsthepoint400.Therefore,
youcanexpectthattheCWTcoefficientwillbezeroatthatpointandscale.Thesignalisonlynonzeroattwovalues,300
and500,andneitherconeofinfluenceforthosevaluesincludesthepoint400atscale20.Youcanconfirmthisby
entering:
x=zeros(1000,1);
x([300500])=1;
CWTcoeffs=cwt(x,1:128,'db4');
plot(CWTcoeffs(20,:));gridon;

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

14/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Next,lookatthepoint400atscale80.Atscale80,theconesofinfluenceforbothpoints300and500includethepoint
400.Eventhoughthesignaliszeroatpoint400,youobtainanonzeroCWTcoefficientatthatscale.TheCWTcoefficientis
nonzerobecausethesupportofthewavelethasbecomesufficientlylargeatthatscaletoallowsignalvalues100points
aboveandbelowtoaffecttheCWTcoefficient.Youcanconfirmthisbyentering:
plot(CWTcoeffs(80,:));
gridon;

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

15/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Intheprecedingexample,theCWTcoefficientsbecamelargeinthevicinityofanabruptchangeinthesignal.Thisability
todetectdiscontinuitiesisastrengthofthewavelettransform.TheprecedingexamplealsodemonstratedthattheCWT
coefficientslocalizethediscontinuitybestatsmallscales.Atsmallscales,thesmallsupportofthewaveletensuresthatthe
singularityonlyaffectsasmallsetofwaveletcoefficients.
Todemonstratewhythewavelettransformissoadeptatdetectingabruptchangesinthesignal,considerashifted
Heaviside,orunitstepsignal.
x=[zeros(500,1);ones(500,1)];
CWTcoeffs=cwt(x,1:64,'haar','plot');colormapjet;

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

16/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Similartotheshiftedimpulseexample,theabrupttransitionintheshiftedstepfunctionresultsinlargeCWTcoefficientsat
thediscontinuity.Thefollowingfigureillustrateswhythisoccurs.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

17/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Intheprecedingfigure,theredfunctionistheshiftedunitstepfunction.TheblackfunctionslabeledA,B,andCdepictHaar
waveletsatthesamescalebutdifferentpositions.YoucanseethattheCWTcoefficientsaroundpositionAarezero.The
signaliszerointhatneighborhoodandthereforethewavelettransformisalsozerobecauseanywaveletintegratesto
zero.
NotetheHaarwaveletcenteredaroundpositionB.ThenegativepartoftheHaarwaveletoverlapswitharegionofthestep
functionthatisequalto1.TheCWTcoefficientsarenegativebecausetheproductoftheHaarwaveletandtheunitstepis
anegativeconstant.Integratingoverthatareayieldsanegativenumber.
NotetheHaarwaveletcenteredaroundpositionC.HeretheCWTcoefficientsarezero.Thestepfunctionisequaltoone.
Theproductofthewaveletwiththestepfunctionisequaltothewavelet.Integratinganywaveletoveritssupportiszero.
Thisisthezeromomentpropertyofwavelets.
AtpositionB,theHaarwavelethasalreadyshiftedintothenonzeroportionofthestepfunctionby1/2ofitssupport.As
soonasthesupportofthewaveletintersectswiththeunityportionofthestepfunction,theCWTcoefficientsarenonzero.In
fact,thesituationillustratedinthepreviousfigurecoincideswiththeCWTcoefficientsachievingtheirlargestabsolute
value.Thisisbecausetheentirenegativedeflectionofthewaveletoscillationoverlapswiththeunityportionoftheunit
stepwhilenoneofthepositivedeflectionofthewaveletdoes.Oncethewaveletshiftstothepointthatthepositive
deflectionoverlapswiththeunitstep,therewillbesomepositivecontributiontotheintegral.Thewaveletcoefficientsare
stillnegative(thenegativeportionoftheintegralislargerinarea),buttheyaresmallerinabsolutevaluethanthose
obtainedatpositionB.
Thefollowingfigureillustratestwootherpositionswherethewaveletintersectstheunityportionoftheunitstep.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

18/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Inthetopfigure,thewavelethasjustbeguntooverlapwiththeunityportionoftheunitstep.Inthiscase,theCWT
coefficientsarenegative,butnotaslargeinabsolutevalueasthoseobtainedatpositionB.Inthebottomfigure,the
wavelethasshiftedpastpositionBandthepositivedeflectionofthewaveletbeginstocontributetotheintegral.TheCWT
coefficientsarestillnegative,butnotaslargeinabsolutevalueasthoseobtainedatpositionB.
Youcannowvisualizehowthewavelettransformisabletodetectdiscontinuities.Youcanalsovisualizeinthissimple
exampleexactlywhytheCWTcoefficientsarenegativeintheCWToftheshiftedunitstepusingtheHaarwavelet.Note
thatthisbehaviordiffersforotherwavelets.
x=[zeros(500,1);ones(500,1)];
CWTcoeffs=cwt(x,1:64,'haar','plot');colormapjet;
%plotafewscalesforvisualization
subplot(311);
plot(CWTcoeffs(5,:));title('Scale5');
subplot(312);
plot(CWTcoeffs(10,:));title('Scale10');
subplot(313);
plot(CWTcoeffs(50,:));title('Scale50');
NextconsiderhowtheCWTrepresentssmoothsignals.Becausesinusoidaloscillationsareacommonphenomenon,this
sectionexamineshowsinusoidaloscillationsinthesignalaffecttheCWTcoefficients.Tobegin,considerthesym4wavelet
ataspecificscalesuperimposedonasinewave.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

19/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

RecallthattheCWTcoefficientsareobtainedbycomputingtheproductofthesignalwiththeshiftedandscaledanalyzing
waveletandintegratingtheresult.Thefollowingfigureshowstheproductofthewaveletandthesinusoidfromthe
precedingfigure.

YoucanseethatintegratingoverthisproductproducesapositiveCWTcoefficient.Thatresultsbecausetheoscillationin
thewaveletapproximatelymatchesaperiodofthesinewave.Thewaveletisinphasewiththesinewave.Thenegative
deflectionsofthewaveletapproximatelymatchthenegativedeflectionsofthesinewave.Thesameistrueofthepositive
deflectionsofboththewaveletandsinusoid.
Thefollowingfigureshiftsthewavelet1/2oftheperiodofthesinewave.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

20/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Examinetheproductoftheshiftedwaveletandthesinusoid.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

21/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

YoucanseethatintegratingoverthisproductproducesanegativeCWTcoefficient.Thatresultsbecausethewaveletis1/2
cycleoutofphasewiththesinewave.Thenegativedeflectionsofthewaveletapproximatelymatchthepositivedeflections
ofthesinewave.Thepositivedeflectionsofthewaveletapproximatelymatchthenegativedeflectionsofthesinusoid.
Finally,shiftthewaveletapproximatelyonequartercycleofthesinewave.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

22/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Thefollowingfigureshowstheproductoftheshiftedwaveletandthesinusoid.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

23/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

IntegratingoverthisproductproducesaCWTcoefficientmuchsmallerinabsolutevaluethaneitherofthetwoprevious
examples.Thatresultsbecausethenegativedeflectionofthewaveletapproximatelyalignswithapositivedeflectionofthe
sinewave.Also,themainpositivedeflectionofthewaveletapproximatelyalignswithapositivedeflectionofthesinewave.
Theresultingproductlooksmuchmorelikeawaveletthantheothertwoproducts.Ifitlookedexactlylikeawavelet,the
integralwouldbezero.
Atscaleswheretheoscillationinthewaveletoccursoneitheramuchlargerorsmallerscalethantheperiodofthesine
wave,youobtainCWTcoefficientsnearzero.Thefollowingfigureillustratesthecasewherethewaveletoscillatesona
muchsmallerscalethanthesinusoid.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

24/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Theproductshowninthebottompanecloselyresemblestheanalyzingwavelet.IntegratingthisproductresultsinaCWT
coefficientnearzero.
Thefollowingexampleconstructsa60HzsinewaveandobtainstheCWTusingthesym8wavelet.
t=linspace(0,1,1000);
x=cos(2*pi*60*t);
CWTcoeffs=cwt(x,1:64,'sym8','plot');colormapjet;

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

25/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

NotethattheCWTcoefficientsarelargeinabsolutevaluearoundscales9to21.Youcanfindthepseudofrequencies
correspondingtothesescalesusingthecommand:
freq=scal2frq(9:21,'sym8',1/1000);
NotethattheCWTcoefficientsarelargeatscalesnearthefrequencyofthesinewave.Youcanclearlyseethesinusoidal
patternintheCWTcoefficientsatthesescaleswiththefollowingcode.
surf(CWTcoeffs);colormapjet;
shading('interp');view(60,12);

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

26/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Thefinalexampleconstructsasignalconsistingofbothabrupttransitionsandsmoothoscillations.Thesignalisa4Hz
sinusoidwithtwointroduceddiscontinuities.
N=1024;
t=linspace(0,1,1024);
x=4*sin(4*pi*t);
x=xsign(t.3)sign(.72t);
plot(t,x);xlabel('t');ylabel('x');
gridon;

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

27/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

Notethediscontinuitiesneart=0.3andt=0.7.
ObtainandplottheCWTusingthesym4wavelet.
CWTcoeffs=cwt(x,1:180,'sym4');
imagesc(t,1:180,abs(CWTcoeffs));
colormapjet;axisxy;
xlabel('t');ylabel('Scales');

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

28/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

NotethattheCWTdetectsboththeabrupttransitionsandoscillationsinthesignal.TheabrupttransitionsaffecttheCWT
coefficientsatallscalesandclearlyseparatethemselvesfromsmoothersignalfeaturesatsmallscales.Ontheotherhand,
themaximaandminimaofthe2HzsinusoidareevidentintheCWTcoefficientsatlargescalesandnotapparentatsmall
scales.
ThefollowinggeneralprinciplesareimportanttokeepinmindwheninterpretingCWTcoefficients.

ConeofinfluenceDependingonthescale,theCWTcoefficientatapointcanbeaffectedbysignalvaluesatpoints
farremoved.Youhavetotakeintoaccountthesupportofthewaveletatspecificscales.Useconofinftodeterminethe
coneofinfluence.Notallwaveletsareequalintheirsupport.Forexample,theHaarwavelethassmallersupportatall
scalesthanthesym4wavelet.

DetectingabrupttransitionsWaveletsareveryusefulfordetectingabruptchangesinasignal.Abruptchangesina
signalproducerelativelylargewaveletcoefficients(inabsolutevalue)centeredaroundthediscontinuityatallscales.
Becauseofthesupportofthewavelet,thesetofCWTcoefficientsaffectedbythesingularityincreaseswithincreasing
scale.Recallthisisthedefinitionoftheconeofinfluence.Themostpreciselocalizationofthediscontinuitybasedon
theCWTcoefficientsisobtainedatthesmallestscales.

DetectingsmoothsignalfeaturesSmoothsignalfeaturesproducerelativelylargewaveletcoefficientsatscales
wheretheoscillationinthewaveletcorrelatesbestwiththesignalfeature.Forsinusoidaloscillations,theCWT
coefficientsdisplayanoscillatorypatternatscaleswheretheoscillationinthewaveletapproximatestheperiodofthe
sinewave.

RedundancyintheContinuousWaveletTransform
Anysignalprocessingperformedonacomputerusingrealworlddatamustbeperformedonadiscretesignalthatis,on
asignalthathasbeenmeasuredatdiscretetime.Sowhatexactlyis"continuous"abouttheCWT?
http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

29/30

8/28/2015

ContinuousWaveletTransformMATLAB&Simulink

What's"continuous"abouttheCWT,andwhatdistinguishesitfromthediscretewavelettransform(tobediscussedinthe
followingsection),isthesetofscalesandpositionsatwhichitoperates.
Unlikethediscretewavelettransform,theCWTcanoperateateveryscale,fromthatoftheoriginalsignaluptosome
maximumscalethatyoudeterminebytradingoffyourneedfordetailedanalysiswithavailablecomputationalhorsepower.
TheCWTisalsocontinuousintermsofshifting:duringcomputation,theanalyzingwaveletisshiftedsmoothlyoverthefull
domainoftheanalyzedfunction.

http://www.mathworks.com/help/wavelet/gs/continuouswavelettransform.html

30/30