Bretherton Winter 2014 Reference Hartmann Atm S 552 notes Chapter 612 111 White noise A common way to statistically assess the signi64257cance of a broad spec tral peak as in the Nino34 example is to compare with a simple noise process White noise h ID: 30463
Download Pdf The PPT/PDF document "Lecture White and red noise Christopher..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Lecture11:Whiteandrednoisec\rChristopherS.BrethertonWinter2014Reference:HartmannAtmS552notes,Chapter6.1-2.11.1WhitenoiseAcommonwaytostatisticallyassessthesignicanceofabroadspectralpeakasintheNino3.4exampleistocomparewithasimplenoiseprocess.Whitenoisehaszeromean,constantvariance,andisuncorrelatedintime.Asitsnamesuggests,whitenoisehasapowerspectrumwhichisuniformlyspreadacrossallallowablefrequencies.InMatlab,w=randn(N)generatesasequenceoflengthNofn(01)`Gaussian'whitenoise(i.e.withanormaldistributionofmean0andstd1).TheuppertwopanelsofFig.1showawhitenoisesequenceoflengthN=128anditsperiodogram,whichshowsthatthepowerspectrumisuniformlyspreadacrossfrequencieswithameanspectralpowerof1=Nperharmonic.11.2RednoiseRednoisehaszeromean,constantvariance,andisseriallycorrelatedintime,suchthatthelag-1autocorrelationbetweentwosuccessivetimesampleshascorrelationcoecient0r1.Aswewillshowshortly,rednoisehasapowerspectrumweightedtowardlowfrequencies,buthasnosinglepreferredperiod.Its`redness'dependsonr,whichcanbetunedtomatchtheobservedtimeseries.FortheNino3.4case,areasonablestatisticalnullhypothesiswouldbethattheobservedpowerspectrumcouldhavebeengeneratedpurelybyrednoise.Tosequentiallygeneratean(01)rednoisesequencexjfromawhitenoisesequencewj,wesetx1=w1xj+1=rxj+(1 r2)12wj+1;j1(11.2.1)Usingpropertiesofnormaldistributions,itiseasilyshownthatxj+1isn(01)(Gaussian)andthatthelag-1correlationcoecientofxj+1andxjisr.Itisalsoeasytoshowbyinductionthatthecorrelationcoecientofxj+pand1 Amath482/582Lecture11Bretherton-Winter20142 0 50 100 -3 -2 -1 0 1 2 3 jwjWhite noise 0 50 100 -3 -2 -1 0 1 2 3 jxjRed noise (r = 0.85) -50 0 50 0 0.01 0.02 0.03 0.04 0.05 MSpectral power -50 0 50 0 0.05 0.1 0.15 0.2 MSpectral power Figure1:Whiteandrednoisetimeseries(left)andtheirperiodograms(right)xjforp1isrp=exp( plogr)=exp( Rpt),whereR= logr=tisthedecorrelationrate.Theautocovariancesequenceofrednoisethusdecaysexponentiallywithlag.Thelagatwhichtheautocorrelationdropsto1=eis=R 1Thefunctionrednoise.m(classwebpage)implementsthisalgorithm,ItwasusedwiththewhitenoisesequenceontheupperleftofFig.1andr=085togeneratetherednoisetimeseriesonthelowerleft.Theperiodogramofthissequence,showninthelowerright,nowhasapredominanceofspectralpowerinlowharmonicsM11.3TheoreticalpowerspectrumofrednoiseThetruepowerspectrumofn(01)rednoiseismosteasilydeducedastheDFTofitsautocovariancesequence.Ratherthangrindingthroughdiscretesums,itismorehelpfultointerprettheDFTasaRiemannsumthatapproximatesthecontinuousintegralforthecomplexFouriercoecientsofthecontinuous Amath482/582Lecture11Bretherton-Winter20143functiona(t)=e Rjtj;L-periodicallyextendedfortL=2):Sm=N 1DFT(a)cMMa(t)]]=L 1ZL=2 L=2e Rjtj i!MtdtSofar,theapproximationisgoodifRt1and! 1Mt1,sothata(t)exp( i!mt)iswellresolvedbythegridofspacingt.IfinadditionRL=21,theintegrandbecomesverysmallfortL=2.Then,withnegligi-bleerrorwecanextendtherangeofintegrationtoinnity:SmL 1Z1 1e Rjtj i!Mtdt=L 1Z0 1et[R i!M]dt+Z10e t[R+i!M]=L 11 R i!M+1 R+i!M=L 12R R2+!2M=! R R2+!2M(!=2 L)(11.3.1)Thepowerspectrumofrednoisehasamaximumvalueforlowfrequencies!MR,anddecreasesathighfrequencies-a'red'spectrum,asclaimed.Ifwesumthepowerspectrumacrosstheharmonics,andthinkofitasaRiemannsumapproximationtoacontinuousintegralNXm=1SmN=2 1XM= N=2! R R2+!2M1 Z(N=2 1)! N!=2d!R R2+!21 Z1 1d!R R2+!2=1 tan 1(!=R)1 1=1ConsistentwithParseval'stheorem,wehavededucedthatthepowerspectrumsumsto1,thevariancethatweconstructedourrednoisetohave.Inthisderiva-tion,extendingthelimitsoftheintegraltoinnityisagoodapproximationifN!=2R.SinceN!=2=N=L==t,thisisequivalenttoRt1,whichwastheassumptionwemadeinderivingthediscreterednoisespectrum(thattherednoiseiswellresolved). Amath482/582Lecture11Bretherton-Winter2014411.4FittingrednoisetodataOnecommonwayofttingtheautocorrelationsequenceisarednoiset,asanexponentiallydecreasingfunctionoflag.Thistisshowninplotaschaindash,usingane-foldingtimeof=61months.Roughlyspeaking,measurementsclosertogetherthanwillbesignicantlycorrelatedandthosefurtherapartwillbeonlyweaklycorrelated.Thiscanbecastintermsofaeectivelag-1autocorrelationr=exp(t=)(=0.85inourcase).Becausetheactualautocorrelationisnotexactlyanexponentiallydecreasingfunctionoflag,risnotexactlythesameasthetruelag-1autocorrelationof0.9.Scriptnino2addsared-noisettotheSSTApowerspectrumbasedon=61monthsandscaledtomatchtheobservedvarianceofSSTA.Therearefourharmonicsinthe0.2-0.4yr 1rangethatclearlystandabovetherednoisespectrum.Totesthowlikelythisistobeachanceoccurrence,wenowlookforalessnoisywaytoestimatethepowerspectrum.