ONTHESTRONGLAWOFLARGENUMBERSFOR SUMSOFRANDOMELEMENTSINBANACHSPACES By AMYM.CANTRELL ADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOL OFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENT OFTHEREQUIREMENTSFORTHEDEGREEOF DOCTOROFPHILOSOPHY UNIVERSITYOFFLORIDA 2001 ACKNOWLEDGMENTS Iwouldliketothankmyadvisor,AndrewRosalsky,forhisinspirationalatti- tudeandunderstanding. Hisdirectionandmotivationhavemadethisworkpossible. Iwouldalso liketothankthemembersofmy dissertationcommittee, Dr. Malay Ghosh,Dr.RamonLittell,Dr.JamesHobert,andDr.IreneHueter. IthankDr.Ron Randlesforhissupportandencouragementbeforeandthroughoutmystudyatthe UniversityofFlorida. IthankmyparentsforalwaysbelievingthatIcouldaccomplishmygoals,and fortheirconstantloveandsupport. Ithankmyhusband,Emory,forhislove,en- couragementandunderstanding. Igivemythankstoalloftheteachersthatinspiredandmotivatedmydesireto learn,withspecialthankstoRitaCater,RonGoolsby,JamesBentley,andThomas Polaski. I wishtogivespecialthankstoMrs. AnneClosefor hermanyyears of support,encouragementandfriendship. 11 TABLEOFCONTENTS ACKNOWLEDGMENTS ii ABSTRACT iv CHAPTERS 1 INTRODUCTION 1 2 STRONGLAWSOFLARGENUMBERSFORSUMSOFRANDOM ELEMENTS 11 2.1 Introduction 11 2.2 NecessaryConditionsforaSLLN 14 2.3 SufficientConditionsforaSLLN 21 2.3.1 SLLNsfor SumsofIndependent RandomElementsinRade- macherTypepBanachSpaces 21 2.3.2 SLLNsforCompactlyUniformlyIntegrableSequencesofInde- pendentRandomElements 36 2.3.3 SLLNsforBanachSpaceValuedSummandsIrrespectiveof TheirJointDistributions 44 2.4 SomeInterestingExamples/Counterexamples 51 3 STRONGLAWSOFLARGENUMBERSFORSUMSOFRANDOM VARIABLES 65 3.1 Introduction 65 3.2 NecessaryConditionsforaSLLN 65 3.3 SufficientConditionsforaSLLN 66 3.3.1 SLLNsforSumsofIndependentRandomVariables 66 3.3.2 SLLNsforRandomVariableSummandsIrrespectiveofTheir JointDistributions 77 3.4 SomeInterestingExamples/Counterexamples 80 4 SUMMARYANDIDEASFORFUTURERESEARCH 97 4.1 Summary 97 4.2 IdeasForFutureResearch 98 REFERENCES 101 BIOGRAPHICALSKETCH 105 iii AbstractofDissertationPresentedtotheGraduateSchool oftheUniversityofFloridainPartialFulfillmentofthe RequirementsfortheDegreeofDoctorofPhilosophy ONTHESTRONGLAWOFLARGENUMBERSFOR SUMSOFRANDOMELEMENTSINBANACHSPACES By AmyM.Cantrell May2001 Chairman: AndrewRosalsky MajorDepartment:Statistics Forasequenceofrandomelements{Vn, n>1}inarealseparableBanachspace A,sufficientconditionsareprovidedforthestronglawoflargenumbersX^Li(Vj~ Ci)/bn —> 0 almostcertainlytoholdwhere {c„,ra > 1} and {bn > 0,n > 1} are suitablesequencesofcenteringelementsinXandnormingconstants,respectively. Inthecaseofindependentrandomelements,separatenecessaryconditionsare alsoprovidedforastronglawoflargenumbers. Thenecessityresultextendsareal lineresultofMartikainentoaBanachspacesetting.Thesufficiencyresultsinthecase ofindependentsummandsarenewevenwhenXisthereallineandaredividedinto twocategories. ThefirstassumesthatXisofRademachertypep(1<p<2). The resultisgeneralenoughtoincludeasspecialcasesastronglawofAdler,Rosalsky, andTaylorforsumsofindependentandidenticallydistributedrandomelementsand astronglawofHeydeforsumsofindependent(real-valued)randomvariables. The secondimposesnoconditionsontheunderlyingBanachspace; instead,itassumes IV thatthesequenceofrandomelementsiscompactlyuniformlyintegrable. Theresult includesasspecialcasesastronglawofAdler,Rosalsky,andTaylorforcompactly uniformlyintegrablesequencesandastronglawofTaylorandWeiforuniformlytight sequences. Strong laws are also provided where no conditions are imposedon thejoint distributionsoftherandomelementsorontheunderlyingBanachspace,andthese resultsarenewevenwhentheBanachspaceistherealline. Illustrativeexamplesareprovidedwhichcomparetheresultsorwhichshowhow theresultsimproveuponoraredifferentfromotherresultsintheliterature.Examples arealsoprovidedwhichshowthattheresultsaresharp. v > CHAPTER 1 INTRODUCTION Thehistoryand literatureofinvestigationon laws oflarge numbers arevast andrich,asthisconceptiscrucialinprobabilityandstatisticaltheoryandintheir application. Thereisnothingmorefundamentaltotheveryfoundationofstatistical sciencethanthelawsoflargenumbers. Itisthelawsoflargenumberswhichelucidate thenotionthatprobabilityisalimitingrelativefrequencyandhencethelawsoflarge numbersprovidejustificationforKolmogorov’s(1933)axiomatictheoryofprobability beingaphysicallyrealisticsubject. Indeed,thelawsoflargenumbersprovidearig- orousformulationandjustificationforthenotionthat“thesamplemeanapproaches thepopulationmeanasthesamplesizeapproachesinfinity”(i.e.,thesamplemeanis aconsistentestimatorofthepopulationmean). Itistheareaofconsistencywherein thelawsoflargenumbershavemanyapplications. Thefirsttheoremonthelawoflargenumbers,duetoBernoulliinthelate1600’s, wastheweaklawoflargenumbers (WLLN)forBernoullitrialswhichstatesthatif Snisthenumberofsuccessesobservedinnindependentidenticaltrialswithsuccess probabilitypineachtrial, then Sn/n —»pinprobabilityasn — oo. Bernoulli’s resultwastoutedbyKolmogorovin1986asthebeginningofprobabilityproper(see Bingham(1989)). AfterLebesgue’sworkonmeasuretheoryintheearly1900’s,the (first) stronglawoflargenumbers (SLLN) couldbeformulated. Thefirstofsuch SLLNsiscreditedto Borel (1909) and isthe extension ofBernoulli’sresultfrom convergence in probabilitytoconvergencewithprobabilityone (oralmostcertain (a.c.) convergence). Itisinterestingtoobservethattherewasoveratwohundred 1 2 yeargapbetweenthisSLLNanditsearlierWLLNcounterpart. Cantelli(1917) is creditedwiththefirstSLLNregardingthea.c.convergenceofthesamplemeanto thepopulationmean(seeSeneta(1992)). Thelawsoflargenumbersprovidetheconsistencyofmanyestimatorsincluding numerouscommonstatistics (suchas, ofcourse, thesamplemean), aswellasfor estimatesfoundbyMonteCarlosimulation.Manyoftheseconceptsandapplications canbeextended. InterestingapplicationsofSLLNsalsooccurinphysicsandcom- puterscience. Thefieldofergodictheory,whichliesattheinterfaceofprobability theoryandstatisticalphysics,hasasitsfoundationtheSLLNtyperesultprovidedby theBirkoff-Khintchine-vonNeumannpointwiseergodictheorem(see,e.g., Breiman (1968,Chapter6)orStout(1974,Section3.5)). Thisresultgivesconditionsforan “ensembleaverage” tobeestimatedbya “timeaverage.” ASLLNwasderivedby Wehr(1997)withapplicationstorandomresistornetworksandthedurabilityofcom- positefibers. OtherapplicationsoftheSLLNtoresistancecanbefoundinEssohand Bellissard(1989). Anotherinterestingand“non-statistical”applicationoccursinthe fieldofnumbertheoryintheinvestigationofnormalnumbersx€ [0,1] (seeRevesz (1968,pp.151-157)). The primaryobjectiveofthecurrentwork isto investigate conditions under which the SLLN for independent Banach space valued random elements obtains, althoughsomeresultswillbepresentedforwhichtheassumptionofindependenceis notneeded. Beforediscussionofthisobjectivesomenotationanddefinitionsmust beintroduced. LetA"bealinearspaceoverM;thatis,A”isavectorspaceoverM. Let||•||: X-*[0,oo)beafunctionsatisfyingthefollowingthreeproperties: (i)Forv6X,||v||=0ifandonlyifv—0. 3 (ii)||av||=|a||M|forallvGAandaGM. (Hi)||«i+W2II<IKII+11^211forallvuv2GA. Thefunction ||•||iscalledanormonA,andXisthensaidtobearealnormed linearspace(withnorm|| ||). Asequence{vn,n>1}inAissaidtoconvergetovGXiflim^oo||un—v||=0. Arealnormedlinearspaceissaidtobe complete ifeveryCauchysequencein X convergestoamemberofX. (ACauchysequenceinXis, ofcourse,asequence {vn,n>1}inXsuchthat mnl——i>>m0000|K-i>m||=0.) ArealnormedlinearspaceXwhichiscompleteissaidtobearealBanachspace. A realBanachspaceissaidtobeseparableifitcontainsacountabledensesubset. Let(f2,P,P)beaprobabilityspace,letAbearealseparableBanachspacewith norm ||•||,andletXbeequippedwithitsBorelcr-algebraB; thatis,Bisthea- algebrageneratedbytheclassofopensubsetsofXdeterminedby viathemetric 11 11 d(vi,v2) = |K—v2\\,Vi,v2 G X. A randomelement AinAisanJF-measurable transformationfrom12tothemeasurablespace(A,B). Theexpectedvalueormean ofA,denotedEA,isdefinedtobethePettisintegralprovideditexists;thatis,Ahas expectedvalueEVinAifL(EV)=E(L(A))foreveryLinX*whereA*denotes the(dual) spaceofallcontinuouslinearfunctionalsonA. Werecallthatalinear functionalisafunctionL:A-4EsatisfyingL(av+bw) =aL(v)+bL(w)forall v,wGAandalla,6GM. Ofcourse,alinearfunctionalLiscontinuousifwhenever {vn, n>1}isasequenceinAwithvn-»vGAwehaveL(v„)-»L(v). Asufficient conditionforEVtoexististhat£||A|| < 00 (see,e.g., Taylor(1978, p.40)). A completecharacterizationofwhenthePettisintegralexistswasprovidedbyBrooks ) 4 (1969). SeeHilleand Phillips (1957, pp. 76-85) forfurtherdiscussionanddetails regardingthepropertiesofthePettisintegral. An exampleillustratingexpectationin Banach spacesfollows: Let 1 < p < ooandX =CP(M.)where,asusual,thisdenotestheclassofLebesguemeasurable functions/: R —>• Rsuch that fR|/(x)|pdx < oo where thenorm isdefinedby ||/|| = Ifp\f(x)\pdx) . Thisclass isaBanach spaceaccordingtothefamous Riesz-Fischertheorem(see,forexample,Royden(1988,p. 125)). ForfixedfoEX andintegrablerandomvariablesXandY,definetherandomelementV=A”/o+Y. WeclaimthatEV = E(X)fo+E(Y). Toverifythis,notethatforanyL E X*, wehaveL(E(X)f0+E(Y))=E(X)L(f0)+E(Y)L(1)=E(XL(f0))+E(YL(1))= E(XL(f0)+YL{1))=E(L{Xf0+Y)) =E{L{V))andthusthecandidateEV= E(X)fo+E(Y)satisfiesthedefiningpropertyL(EV)=E(L(V)).Weremarkthatthe dualspaceX*canbecompletelycharacterizedforthisexample. ForeverygE£g(M) whereq—^(1<q<oo),thefunctionalLdefinedby L{f)=J[rf{x)g(x)dx (/€ Cp) (1.1) isinX* (see,e.g.,Royden(1988,p. 131)). Conversely,thefamousRieszRepresen- tationTheorem(see,e.g.,Royden(1988,p.132))assertsthateveryLEX*isofthe form(1.1)forsomefunctiongE£,(M)whereq= (1<q<oo). Thefunctiong isuniqueuptosetsofLebesguemeasure0. Let n > 1} beasymmetric Bernoullisequence; that is, {F„, n > 1} is asequenceofindependent andidenticallydistributed (i.i.d.) randomvariables withP{Y{ =1}=P{YX = -1} = 1/2. LetX°° =XxXxXx •• anddefine C(X) ={(vi,v2,... E X°° :Y^=iYnvnconvergesinprobability}.Let 1 <p< 2. ThenXissaidtobeofRademachertypepifthereexistsaconstant0<C<oo 1 5 suchthat OO P E IKIIPfora11(«i,«*,-••)€ C(X). n=l n-l Itshouldbepointedoutthattheconditionthattheseries Y„vnconverges inprobabilityisequivalenttotheconditionthattheseries Ynvnconvergesa.c. Indeed, ItoandNisio (1968) provedthataseriesofindependent randomelements convergesinprobabilityifandonlyifitconvergesa.c.therebyextendingthecele- bratedtheoremofLevy(see,e.g.,ChowandTeicher(1997,p.72))fromtherealline toBanachspaces. Hoffmann-JprgensenandPisier(1976)provedfor1<p<2thatarealseparable BanachspaceisofRademachertypepifandonlyifthereexistsaconstant0<C<oo suchthat E (1-2) «=i i- foreveryfinitecollection{Pi,-••,Vn}ofindependentrandomelementswithEV{= 0, 1<f<n. OthercharacterizationsofaBanachspacebeingofRademachertype pareprovidedbyWoyczyriski(1978). IfarealseparableBanachspaceisofRademachertypepforsome1 <p<2, thenitisofRademachertypeqforall1<q<p.EveryrealseparableBanachspaceis ofRademachertype(atleast)1whilethe£p-spacesand^-spacesareofRademacher typemin{2,p} forp > 1. (SeeWoyczynski (1978, pp. 343-353) fordetails and a thoroughdiscussion.) EveryrealseparableHilbertspaceandrealseparablefinite- dimensionalBanachspaceisofRademachertype2(seePisier(1986)). Inparticular, thereallineIRisofRademachertype2.