This is a repository copy of On bounded continuous solutions of the archetypal equation with rescaling. White Rose Research Online URL for this paper: http://eprints.whiterose.ac.uk/88014/ Version: Accepted Version Article: Bogachev, LV, Derfel, G and Molchanov, SA (2015) On bounded continuous solutions of the archetypal equation with rescaling. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 471 (2180). 20150351. ISSN 1364-5021 https://doi.org/10.1098/rspa.2015.0351 Reuse Unless indicated otherwise, fulltext items are protected by copyright with all rights reserved. The copyright exception in section 29 of the Copyright, Designs and Patents Act 1988 allows the making of a single copy solely for the purpose of non-commercial research or private study within the limits of fair dealing. The publisher or other rights-holder may allow further reproduction and re-use of this version - refer to the White Rose Research Online record for this item. Where records identify the publisher as the copyright holder, users can verify any specific terms of use on the publisher’s website. Takedown If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing [email protected] including the URL of the record and the reason for the withdrawal request. [email protected] https://eprints.whiterose.ac.uk/ On bounded continuous solutions of the archetypal rspa.royalsocietypublishing.org equation with rescaling 1 2 Leonid V. Bogachev , Gregory Derfel and Research 3 Stanislav A. Molchanov 1 Articlesubmittedtojournal DepartmentofStatistics,SchoolofMathematics, UniversityofLeeds,LeedsLS29JT,UK 2 DepartmentofMathematics,BenGurionUniversityof SubjectAreas: theNegev,Be’erSheva84105,Israel mathematicalanalysis,functional 3 DepartmentofMathematics,UniversityofNorth equations,probabilitytheory CarolinaatCharlotte,CharlotteNC28223,USA Keywords: functional&functional-differential The ‘archetypal’ equation with rescaling is given equations,pantographequation, by y(x)= R2y(a(x−b))µ(da,db) (x∈R), where µ is a prRoRbability measure; equivalently, y(x)= Markovchain,harmonicfunction, E{y(α(x−β))}, with random α,β and E denoting martingale,stoppingtime expectation.Examplesinclude:(i)functionalequation y(x)= ipiy(ai(x−bi)); (ii) functional-differential Authorforcorrespondence: (‘pantoPgraph’)equationy′(x)+y(x)= ipiy(ai(x− LeonidV.Bogachev ci)) (pi>0, ipi=1). Interpreting sPolutions y(x) as harmonicPfunctions of the associated Markov e-mail:[email protected] chain(Xn),weobtainLiouville-typeresultsasserting that any bounded continuous solution is constant. In particular, in the ‘critical’ case E{ln|α|}=0 such a theorem holds subject to uniform continuity of y(x); the latter is guaranteed under mild regularity assumptions on β, satisfied e.g. for the pantograph equation (ii). For equation (i) with ai=qmi (mi∈ Z, ipimi=0), the result can be proved without thePuniformcontinuityassumption.Theproofsutilize the iterated equation y(x)=E{y(Xτ)|X0=x} (with a suitable stopping time τ) due to Doob’s optional stoppingtheoremappliedtothemartingaley(Xn). 1. Introduction (a) The archetypal equation Thispaperconcernstheequationwithrescaling(referred toas‘archetypal’)oftheform y(x)= y(a(x−b))µ(da,db), x∈R, (1.1) ZZ R2 where µ(da,db) is a probability measure on R2. The integral in (1.1) has the meaning of expectation with (cid:13)c TheAuthor(s)PublishedbytheRoyalSociety.Allrightsreserved. respecttoarandompair(α,β)withdistributionP{(α,β)∈da×db}=µ(da,db);thus,equation 2 (1.1)canbewrittenincompactformas y(x)=E{y(α(x−β))}, x∈R. (1.2) ....rspa AshasbeenobservedbyDerfel[1](andwillbeillustratedbelowin§1(b)),thisequationisa .....roy richsourceofvariousfunctionalandfunctional-differentialequationswithrescaling,specifiedby .a ..ls asuitablechoiceoftheprobabilitymeasureµ(i.e.thedistributionof(α,β)).Itisforthatreason .o .c thatweproposetocall(1.2)(aswellasitsintegralcounterpart(1.1))thearchetypalequation(AE). ..ie ..ty Thestudyofthisequationallowsonetoenhanceandunifyearlierresultsforparticularsubclasses .p ofequationswithrescaling,whilemakingtheanalysismoretransparentandefficient(cf.[2]). ...ub ..lis Notingthatanyfunctiony(x)≡constsatisfiestheAE(1.2),itisnaturaltoinvestigatewhether .h ..in tnhaetrueraalrleyaanriysensonin-trthiveiaclo(ni.tee.xntoonf-cfuonncsttiaonnta)lbaonudndfuedncctoionntainl-udoifufser(ebn.cti.)alsoeqluutaiotinosn.sSwucihthareqsuceasltiinogn, ......g.org where the possible existence of bounded solutions (e.g. periodic, almost periodic, compactly ..P supported,etc.)isofmajorinterestinphysicalandotherapplications(seee.g.[3–6]).Solutions ....roc understudymayalsobeboundedbynature,e.g.representingaruinprobabilityasafunctionof ..R . ainsitaiafilrcsatpstiteaplt[o7,w8]a.rOdnsathfeuloltdheesrchriapntdio,ncoonffithneinagsyomnepsteolftitcobbeohuanvdioeudrsooflusotilountisomnsa.ybeconsidered ......SocA Thus,thegoalofthepresentpaperistogiveconditionsonthedistributionµoftherandom ..0 coefficients(α,β),underwhichanyb.c.-solutionofequation(1.2)isconstantonR.Forshorthand, ....000 cwoemrpelfeexrtaonasltyasteismaenndtshoafrmthoisnikcinfudnactsioLniotuhveiloler-yt,ybpeeatrhineogrienmmsibnydatnhaaltoyg(yx)wiinth(1s.i2m)iislaarwreesuiglhtsteidn .....000 averageofothervalues,thusresemblingtheusualharmonicfunction.Moredetailshighlighting thepertinenceof‘harmonicity’inthecontextofequation(1.2)areprovidedin§1(d). Remark 1.1. Continuity of y(x) (or some other regularity assumption) is needed to avoid pathologicalsolutions,asiswellknowninthetheoryoffunctionalequations(cf.[9,Ch.2]).For example,allb.c.-solutionsoftheequationy(x)=py 1(x+1) +(1−p)y 1(x−1) (0≤p≤1) 2 2 areconstantbyTheorem1.1(a)statedbelow,butift(cid:0)hecontin(cid:1)uityrequirem(cid:0) entisdr(cid:1)oppedthen one can easily construct other bounded solutions, e.g. the Dirichlet function y(x)=1Q(x) (i.e. y(x)=1ifxisrationalandy(x)=0otherwise),whichiseverywherediscontinuous. (b) Some subclasses of the archetypal equation; historical remarks Beforeoutliningourresults,weillustratetheremarkablecapacityofequation(1.2)justifying thename‘archetypal’.Generalsurveysoffunctionalandfunctional-differentialequationswith rescalingarefoundinDerfel[10]andBaron&Jarczyk[11],bothwithextensivebibliographies. (i)Functionalequationsandself-similarmeasures Tostartwith,inthesimplestcaseα≡1equations(1.1),(1.2)arereducedto y(x)=Z y(x−t)µβ(dt) ⇐⇒ y(x)=E{y(x−β)}, (1.3) R where µβ(dt):=P(β∈dt). This equation (sometimes called the integrated Cauchy functional equation [12]) plays a central role in potential theory and harmonic analysis on groups [13,14], andisalsoprominentinprobabilitytheoryinrelationtorenewaltheorems[15,§XI.9],Markov chains[16,Ch.5],queues[17,§III.6],characterizationofprobabilitydistributions[12,Ch.2],etc. ALiouville-typeresultinthiscaseisrenderedbythecelebratedChoquet–Denytheorem[18](see also[12]andreferencestherein). Notethatequation(1.3)canbewrittenintheconvolution1formy=y⋆µβ.Moregenerally,if αhasadiscretedistribution(withatomsaiandmassespi)then,denotingbyµβi theconditional 1Theconvolutionbetweenfunctiony(x)andmeasureσinRisdefinedasy⋆σ(x):=RRy(x−t)σ(dt). distributionofβgivenα=ai,theAE(1.2)isconvenientlyexpressedinconvolutions, 3 y(x)= p y(a x)⋆µi. i i β i Forapurelydiscretemeasureµ,withatomsX(ai,bi)andmassespi=P(α=ai,β=bi),theAE(1.2) ....rspa specializesto .....roy y(x)=Xipiy(ai(x−bi)). (1.4) ....also Ifallai>1then(1.4)isanexampleofHutchinson’sequation[19]forthedistributionfunctionof ...cie aself-similarprobabilitymeasurewhichisinvariantunderafamilyofcontractions(here,affine ..ty .p transformationsx7→bi+x/ai).Animportantsubclassof(1.4),withai≡a>1,isexemplifiedby ...ub y(x)=12y(a(x+1))+ 12y(a(x−1)). .....lishin Twhhiesreeqtuhaetsioignndseasrcercibheossethnein(sdeelfp-esinmdielnartl)ydwisittrhibpurtoiobnabfiulintycti21o.nCohfatrhaectrearnizdaotmionseorfiethsiPsd∞nis=t0ri±buat−ionn, .......g.org forRdeiftfuerrneinntgat>oe1qiusatthioento(p1i.4ca)lwBietrhnaouil≡liaco>nv1o,luthtieondsenprsoitbylezm(x[2)0:=]. y′(x)(ifitexists)satisfies .....Proc ..R . z(x)=aXipiz(a(x−bi)), (1.5) ....Soc oftencalledthetwo-scaledifferenceequationorrefinementequation[21].Constructionofcompactly ..A supportedcontinuoussolutionsof(1.5)playsacrucialroleinwavelettheory[22,23]andalsoin ...00 subdivisionschemesandcurvedesign[4,24],whichisarapidlygrowingbranchofapproximation ...00 .0 theory.Aspecialversionof(1.5)knownasSchilling’sequation ...00 . z(x)=α 1z(αx+1)+ 1z(αx)+ 1z(αx−1) 4 2 4 (cid:2) (cid:3) arises in solid state physics in relation to spatially chaotic structures in amorphous materials [5, p.230], where the existence of compactly supported continuous solutions is again of major interest;see[25]forafullcharacterizationofthisproblemintermsofarithmeticalpropertiesofα. (ii)Functional-differentialequations Let us now turn to the situation where the distribution of β conditioned on α is absolutely continuous(i.e.hasadensity).ItappearsthatforcertainsimpledensitiestheAE(1.2)produces some well-known functional-differential equations. An important example is the celebrated pantograph equation, introduced by Ockendon & Tayler [26] as a mathematical model of the overheadcurrentcollectionsystemonanelectriclocomotive.2 Initsclassical(one-dimensional) formthepantographequationreads y′(x)=c0y(x)+c1y(αx). (1.6) Thisequationanditsramificationshaveemergedinastrikingrangeofapplications,including number theory [7], astrophysics [28], queues & risk theory [29], stochastic games [8], quantum theory [6], and population dynamics [30]. The common feature of all such examples is some self-similarity of the system under study. Thorough asymptotic analysis of equation (1.6) was given by Kato & McLeod [31]. A more general first-order pantograph equation (with matrix coefficients, and also allowing for a term with a rescaled derivative) was studied by Iserles [27], where a fine geometric structure of almost-periodic solutions was also described. Further developmentsincludeanalysisinthecomplexdomain[32],higher-orderequations[27,33],and stochasticversions[34].AmongrecentimportantanalyticresultsisaproofbydaCostaetal.[35] oftheunimodalityofsolutionswhichplaysasignificantroleinmedicalimagingoftumours[36]. Abalancedversionofthepantographequationisgivenby(see[1,2]) y′(x)+y(x)= piy(ai(x−ci)), pi>0, pi=1. (1.7) i i X X Asexplainedin§3(c),equation(1.7)isessentiallyequivalenttotheAE(1.2)wherebyαisdiscrete, withP(α=ai)=pi,andβconditionedonα=aihastheunitexponentialdistributionon(ci,∞), 2Theterm‘pantographequation’wascoinedbyIserles[27]. with the density function eci−t1 (t). The discreteness of α is not significant here, and a (ci,∞) 4 similarconnectionwiththeAEholdsformoregeneralintegro-differentialequations(cf.[37]) where γ is a randyo′(mx)v+aryia(bxl)e=anEd{yµ(αα,(γx(d−a,γd)}c)≡=ZZPR(2αy∈(ad(ax,−γ∈c)d)cµ)αi,sγ(thdae,ddics)t,ribution of (α(,1γ.8)). .........rspa.roya Higher-orderpantographequationscanalsobededucedfromtheAE,e.g. ..ls .o .c −y′′(x)+y(x)= p y(a (x−c )), p >0, p =1, ..ie i i i i i i i ..ty X X .p andmoregenerally(cf.[2]) ...ub C2y′′(x)+C1y′(x)+y(x)=E{y(α(x−γ)} (C1,C2∈R, C12−4C2≥0). .....lishin onF[−or1a,n1]e,xthamenpelequoaftaiodnif(f1e.r2e)nitskreinddu,cteadketoα≡2andassumethatβhastheunformdistribution ......g.org 2 2 ..P y(x)=Zx+1/2y(2u)du. (1.9) ......rocR x−1/2 . .S Differentiating(1.9),forz(x):=y′(x)weobtainRvachev’sequation[3] ...oc ..A z′(x)=2 z(2x+1)−z(2x−1) . (1.10) ...00 (cid:2) (cid:3) ...00 Acompactlysupportedsolutionof(1.10)(calledthe‘up-function’)anditsgeneralizations(unified .0 underthenameatomicfunctions)haveextensiveapplicationsinapproximationtheory(see[3,24] ...00 . andreferencestherein);allsuchfunctionscanbeobtainedassolutionsofsuitableversionsofthe AE(1.2)(see[1]). (c) Main results Letussummarizeourresults.First,certaindegeneratecaseswarrantaseparateanalysisbut needtobeexcludedingeneraltheory,namely:(i)α=0withpositiveprobability;(ii)|α|≡1;and (iii) α(c−β)≡c for some c∈R (resonance). Note that (ii) includes the case α≡1 settled in the Choquet–Denytheoremmentionedin§1(b)-i;in§2(b)wegeneralizethisresult(Theorem2.3).As forcases(i)and(iii),aLiouvilletheoremholdshereunconditionally,whichiseasytoprovefor (i),analyticallyandprobabilisticallyalike(seeTheorem2.1).Intheresonancecase(iii),theproof ismoreinvolvedrelyingheavilyontheChoquet–Denytheorem(see§2(c)),buttheresultitselfis quitelucidandappealing. Inthenon-degeneratesituation,existenceofnon-trivialb.c.-solutionsisessentiallygoverned bythesignofK:= ln|a|µ(da×db)=E{ln|α|}.Moreprecisely,onecanprove(see[38])the R2 followingdichotomyRRbetweenthesubcritical(K<0)andsupercritical(K>0)regimes. Theorem1.1. SupposethatK=E{ln|α|}isfiniteandE{lnmax(|β|,1)}<∞. (a) IfK<0thenanyb.c.-solutionoftheAE(1.2)isconstant. (b) If K>0 and α>0, then there is a b.c.-solution of (1.2) given by the distribution function FΥ(x):=P(Υ ≤x), where Υ := ∞n=1βn jn=−11αj−1 and {(αn,βn)} is a sequence of independentidenticallydistributed(Pi.i.d.)randQompairswiththesamedistributionµas(α,β). Remark1.2. Almostsure(a.s.)convergenceoftherandomseriesΥ (forα6=0)andcontinuityof FΥ(x)onRwereprovedbyGrintsevichyus[39]. Remark 1.3. The result of Theorem 1.1 was obtained by Derfel [1] under a stronger moment condition E{|β|}<∞ and only for α>0 (which is essential in (b) but not in (a)); however, his argumentsholdinthegeneralcasewithminorchanges. Remark1.4. IncontrastwiththesubcriticalcaseK<0,whichisinsensitivetothesignofα(see Theorem1.1(a)),thesupercriticalcaseK>0ismoredelicate:ifP(α<0)>0theny=FΥ(x)isno longerasolutionoftheAE(1.2);e.g.ifα<0(a.s.)thenthisfunctionsatisfiestheequationy(x)= 5 1−E{y(α(x−β))}(cf.[39,Eq.(5)]).Moreover,onecanprove[38]thatanyboundedsolutionof (1.2) with limits at ±∞ is constant; thus, any non-trivial solution must be oscillating, which is drasticallydifferentfromthecaseα>0(a.s.). ....rspa The critical case K=0 is much more challenging, and it has remained open since [1]. More .....roy .a recently, for a pantograph equation (1.8) without shift (i.e. γ≡0) and some second-order ..ls .o extensions, aLiouville theoremin thecaseK=0 wasestablished byBogachev et al.[2]. In the ...cie presentpaper,weprovethefollowinggeneralresult(cf.Theorem3.1below). ..ty .p ...ub Theorem1.2. AssumethatP(|α|=6 1}>0andK=E{ln|α|}=0.If y(x)isaboundedsolutionofthe ...lish AE(1.2)whichisuniformlycontinuousonR,thenitisconstant. ........ing.org . Although an unwanted restriction, the uniform continuity assumption can be shown to be .P satisfiedprovidedthereexiststheprobabilitydensityofβ conditionedonα(Theorem3.3).An ....roc alternativecriteriontailoredtothemodelβ=γ+ξ withξ independentof(α,γ)(Theorem3.4) ..R . .S is applicable to a large class of examples including the pantograph equation (1.8) and its ...oc generalizations (§3(c)). As a consequence, we obtain a Liouville theorem for the general ..A (balanced)pantographequationinthecriticalcase(cf.Theorem3.8),significantlyextendingthe ..0 .0 resultofBogachevetal.[2].Inparticular,forthefirst-orderpantographequation(1.8)wehave ...00 .0 ...00 Theorem1.3. If P(|α|=6 1}>0andK=E{ln|α|}=0,thenanyboundedsolutionof (1.8)isconstant. . Remark1.5. Asexplainedin§3(d)below,Theorem1.3extendstothecase|α|=1(a.s.)byvirtue ofthegeneralizedChoquet–Denytheoremprovedin§2(b). On the other hand, for a subclass of functional equations (1.4) with multiplicatively commensurable coefficients {ai} (i.e. ai=qmi, q>1, mi∈Z), where the random shift β is discrete and therefore the criteria of Theorems 3.3, 3.4 do not apply, the Liouville theorem in thecriticalcaseK=0(i.e. ipimi=0)canbeprovedbyadifferentmethodthatcircumvents thehypothesisofuniformcPontinuity(Theorem3.9). (d) The method: associated Markov chain and iterations Inthissubsection,wedescribetheprobabilisticapproachtotheAEbasedonMarkovchains andmartingales,andintroducesomebasicnotationanddefinitions. ConsideraMarkovchain(Xn)onRdefinedrecursivelyby Xn=αn(Xn−1−βn) (n∈N), X0=x, (1.11) where {(αn,βn)}n∈N is a sequence of i.i.d. random pairs with the same distribution as (α,β) (see(1.2)).NotethattheAE(1.2)canthenbeexpressedas y(x)=Ex{y(X1)}, x∈R, (1.12) whereindexxintheexpectationreferstotheinitialconditionin(1.11).Thatistosay,anysolution oftheAE(1.2)isaharmonicfunctionoftheMarkovchain(Xn)(cf.[16,p.40]). Remark1.6. BoundedharmonicfunctionsplayaparamountroleinthegeneraltheoryofMarkov chains(seemoredetailsandsomereferencesinappendixB. Remark 1.7. Stochastic recursion (1.11) is well known in the literature as the random difference equation(seee.g.[40–43]andfurtherreferencestherein). Notethatequation(1.12)propagatesalongtheMarkovchain(Xn),i.e.foranyn∈N 6 y(x)=Ex{y(Xn)}, x∈R. (1.13) Equivalently,anintegralformofequation(1.13)canbeobtainedbyiteratingforwardtheAE(1.1). ....rspa Therecursion(1.11)canalsobeiteratedtogiveexplicitly .....roy .a Xn=Anx−Dn (n∈N), (1.14) ....lsoc ..ie where ..ty .p n n n ...ub An:= αk, Dn:= βk αj. (1.15) ..lis kY=1 kX=1 jY=k ...hin Recall that K=E{ln|α|}. In the subcritical case (K<0), for the proof of Theorem 1.1(a) it ......g.org sufficestoconsideriterations(1.13)asn→∞.Indeed,Kolmogorov’sstronglawoflargenumbers . .P impliesthatSn:= nk=1ln|αk|→−∞andhence|An|=exp(Sn)→0a.s.;inviewof(1.14),this ....roc indicatesthattheriPght-handsideof(1.13)eventuallybecomesx-free(seemoredetailsin[38]).In ..R . (tha.es.c)r;ihtiecnaclec,aasets(Kom=e0r)a,ntdheomrantidmoemswτǫawlke(hSanv)eis|Areτcǫu|r=reenxtpb(uStτnǫ)on<eǫth(feolreassnlyimǫ>inf0n),→w∞hiSchn=can−b∞e ......SocA usedtoinferthaty(x)≡constinasimilarfashionasbefore. ..0 .0 Expanding this idea, our approach to the analysis of equation (1.2), first probed in [2], is ...00 basedonreplacingafixednintheiteratedequation(1.13)byasuitablestoppingtimeτ,defined .0 as a random (integer-valued) variable such that for any n∈N the event {τ≤n} is determined ....00 by(α1,β1),...,(αn,βn).Itsufficesforourpurposestoworkwith‘hittingtimes’τB:=inf{n≥ 1: An∈B}≤∞,3whereAn=α1···αn(see(1.15))andB⊂Risanintervalorasinglepoint. Weshallroutinelyusethefollowingcentrallemma(wherecontinuityofy(x)isnotrequired). Lemma1.4. Let(Xn)betheassociatedMarkovchain(1.11),andτ astoppingtimesuchthatτ<∞a.s. If y(x)isaboundedsolutionoftheAE(1.2)thenitsatisfiesthe‘stopped’equation y(x)=Ex{y(Xτ)}, x∈R. (1.16) Thecrucialfactisthaty(Xn)isamartingale(cf.[16,p.43,Proposition1.8]);indeed,by(1.11) Ex{y(Xn)|(αk,βk), k<n}=Ex{y(αn(Xn−1−βn))|Xn−1}=y(Xn−1) (a.s.), (1.17) which verifies the martingale property [44, §10.3, p.94]. The lemma then readily follows by Doob’s optional stopping theorem (e.g. [44, p.100, Theorem 10.10(b)]). For the sake of a more self-containedexposition,adirectproofofLemma1.4isincludedasappendixA. Layout The rest of the paper is organized as follows. In §2 we work out the degenerate cases mentionedatthebeginningof§1(c),namely:P(α=0)>0(§2(a)),|α|≡1(§2(b)),andα(c−β)≡c (§2(c)). In §3(a) we prove our main result for the critical case (Theorem 3.1; cf. Theorem 1.2), backed up in §3(b) by simple sufficient conditions for the uniform continuity of solutions (Theorems3.3and3.4).In§3(c)weexplainindetailtheremarkablelinkbetweenthepantograph equations and the AE, which enables us to prove a Liouville theorem for the general (integro- differential)pantographequationofanyorder(Theorem3.8).Thisiscomplementedin§3(d)by a Liouville theorem for the functional equation (1.4) with ai=qmi (Theorem 3.9). Appendix A providesanelementaryproofofLemma1.4,andappendixBgivesabriefcompendiumofbasic factsilluminatingthefundamentalroleofboundedharmonicfunctionsinthegeneraltheoryof Markovchains. 3Hereandbelow,weadopttheconventionthatinf∅:=∞. 2. Three degenerate cases 7 BeforeembarkingonageneraldiscussionoftheAE,weneedtostudytheproblemofbounded csoolnustiidoenrsatiinoncethrtearienafstpeer.cRiaelccaallstehseonfoptaotsisoinblAenv:a=luQesnko=f1ααkan(sdeeβ(,1.w15h)i)c.h will be excluded from ........rspa.roy .a (a) Vanishing of the scaling coefficient ...lso .c ..ie Let us consider the case where the scaling coefficient α may take the value zero. Note that ..ty .p continuityofsolutionsy(x)isnotassumedapriori. ...ub ..lis .h TheWoerefimrs2t.g1.ivSeuapnpoesleemp0e:n=taPr(yα‘a=na0l)y>tic0’.pTrhoeonfaonfytbhoiusnsdimedpsloeluthtieoonroefmthaenAdEt(h1e.2n)pisrecsoennsttaanntootnhRer. .........ing.org .P prooftoillustratethemethodbasedonLemma1.4. ....roc ..R Proofof Theorem2.1. Withy0(x):=y(x)−y(0),equation(1.1)maybewrittenintheform ..S ...oc y0(x)=(1−p0)ZZR2\{a=0}y0(a(x−b))µ˜(da,db), x∈R, (2.1) .....A00 wnohremreoµ˜n:R=,(fr1o−mp(02).1−)1wµ,esoobttahiant µ˜(R2\{a=0})=1. Denoting by kfk:=supx∈R|f(x)| the sup- .......00000 . |y0(x)|≤(1−p0)ky0k (x∈R) =⇒ ky0k≤(1−p0)ky0k. (2.2) Since1−p0<1,thesecondinequalityin(2.2)immediatelyimpliesthatky0k=0,andthenthe firstinequalitygivesy0(x)≡0,i.e.y(x)≡y(0),asclaimed. Alternativeproofof Theorem2.1. Considerthestoppingtimeτ0:=inf{n≥1: An=0}.Notethat P(τ0>n)=P(An6=0)=P(α16=0,...,αn6=0)=(1−p0)n→0 (n→∞), henceτ0<∞a.s.Now,usingtheiterationformulas(1.14),(1.15),andnotingthatAτ0=0a.s.,by Lemma1.4weobtain y(x)=Ex{y(Xτ0)}=E{y(xAτ0 −Dτ0)}=E{y(−Dτ0)}, x∈R. (2.3) Sincetheright-handsideof(2.3)doesnotdependonx,itfollowsthaty(x)=const. (b) An extension of the Choquet–Deny theorem (i)Theclassicalcaseα≡1 Asmentionedin§1(b),theAE(1.2)withα≡1isreducedto y(x)=E{y(x−β)}, x∈R. (2.4) ThefamousChoquet–Denytheorem[18](cf.[15,p.382,Corollary]or[16,p.161,Theorem1.3]) assertsthatanyb.c.-solutionof(2.4)isconstantprovidedthat(thedistributionof)theshiftβ is non-arithmetic,i.e.notsupportedonanysetλZ={λk, k∈Z}(withspanλ∈R). Remark 2.1. In connection with the uniform continuity condition in Theorem 1.2, it may be of interest to note that some proofs of the Choquet–Deny theorem (e.g. [15, p.382]) deploy the convolutiony˜(x)=y⋆ϕ0,σ2(x)ofab.c.-solutiony(x)withthedensityfunctionϕ0,σ2(x)ofthe normal distribution with zero mean and variance σ2, whereby the function y˜(x) is uniformly continuousandstillsatisfiesequation(2.4).Onceithasbeenprovedthaty˜(x)isconstant,thisis extendedtotheoriginalsolutiony(x)bytakingthelimitasσ→0. A discrete version of the Choquet–Deny theorem [18] (see also [45, §XIII.11] or [46, p.276, 8 Theorem T1]) refers to the case where equation (2.4) is considered on Z and β is integer- valued.Namely,assumethatthesmallestadditivegroupcontainingtheset{x∈Z: P(β=x)>0} coincideswithZ;thenthetheoremassertsthaty(x)≡constforallx∈Z. ....rspa of bIn.c.t-hsoelcuotinotnesxtinoftheequaartiitohnm(e2t.i4c)coanseth(eexwcluhdoilneglinthee, tdheisgeennearbalteescoanseetβo≡gi0v)e. Tahfeulnledxetsrcersiputlitoins ......roya essentiallywellknowninfolklore;wegiveitsproofforthesakeofcompleteness. ...lso .c ..ie Theorem2.2. Assumethatthedistributionofβ isarithmetic,i.e.itssupportΛiscontainedintheset ..ty .p λZwithmaximalspanλ>0.Thenthegeneralb.c.-solutionofequation(2.4)isgivenbyy(x)=g(x/λ), ...ub whereg(·)isanycontinuousperiodicfunctionofperiod1. ..lis .h ..in wPriotohf.λWZ.eInstdaeretdb,yfosrhnow∈iNngletthdantt∈heλNsmbaelltehsetgardedaitteisvtecsoumbmgroonupdiGvi⊂soλrZofgtehnee(rfianteitde)bsyetΛΛcno:i=nc{idse∈s ......g.org Λin:te|gse|r≤sλmni},⊂anGd.BityfoBlélozwoust’tshiadtednntit∈yG(s(ene∈e.Ng.).[4S7i,n§c1e.2th])ewseeqhuaevnecedndn=/Pλ∈siN∈Λinsmnoisni-iwncitrheassoinmge, ......Proc thereexiststhelimitk∗:=limn→∞dn/λ=dn∗/λ∈NandsoΛ⊂k∗λZ.Butλ>0isthemaximal ...R spaNn,ohwe,nitceiske∗a=sy1toansdeeththuastλeq=udatni∗o∈nG(2,.4w)hsipclhitisminptloiesseGpa=raλtZe,daissccrleatiemeeqdu.ationsz˜(k)=E{z˜(k− ......SocA β/λ)} on every coset x0+λZ (x0∈[0,λ)), where z˜(k):=z(x0+kλ) (k∈Z). The discrete ...00 Choquet–Denytheoremshowsthatz˜(k)isconstantonZ;inotherwords,anyboundedsolution ...00 of(2.13)onRisλ-periodic,andtheclaimofthetheoremeasilyfollows. ....000 . Remark2.2. Itisevidentthatanyfunctiony(x)=g(x/λ)satisfiesequation(2.4);themainpointof Theorem2.2isthattherearenootherb.c.-solutions. Remark2.3. Laczkovich[48]givesafullcharacterizationofnon-negativemeasurablesolutionsof theequationy(x)= ℓi=1Ciy(x−bi)(witharbitrarycoefficientsC1,...,Cℓ>0)intermsofthe realrootsofthecharPacteristicequation ℓi=1Cie−bis=1. P Remark 2.4. The original proof by Choquet & Deny [18] (as well as many subsequent proofs and extensions) is based on a reduction to a uniformly continuous solution (cf. Remark 2.1) andestablishingthatthelattermustreachitsmaximumatafinitepointx0∈R.Inviewofthe martingaletechniquesusedinthepresentpaperforageneralAE,itisofinteresttopointoutan elegantmartingaleprooffoundbySzékely&Zeng[49](cf.Rao&Shanbhag[12,Ch.3]). (ii)Case|α|≡1 WeprovehereanextensionoftheChoquet–Denytheoremforαtakingthevalues±1;tothebest ofourknowledge,sucharesulthasnotyetbeenmentionedintheliterature. Theorem2.3. Supposethat|α|≡1andP(α=1)<1.Letβ+,β−havethedistributionof βconditioned onα=1andα=−1,respectively(incaseα≡−1,setβ+≡0). (a) If β+isnon-arithmetictheneveryb.c.-solutionofequation(1.2)isconstant. (b) Letβ+bearithmeticwithspanλ6=0. (b-i) If the distribution of β−is not supported on any set λ0+λZ (λ0∈R), then every b.c.-solutionofequation(1.2)isconstant. (b-ii) Otherwise,thegeneralb.c.-solutionofequation(1.2)isoftheformy(x)=g(x/λ),where g(·)isacontinuous1-periodicfunctionsymmetricaboutpointx0=21λ0/λ,i.e.g(x0− x)=g(x0+x) (x∈R). Remark 2.5. In part (b-ii), all functions with the required symmetry property may be represented (though not uniquely) as g(x)=g0(x−x0)+g0(−x+x0), where g0(·) is an arbitrary (continuous) 1-periodic function. It is straightforward to check that so constructed functions y(x)=g(x/λ) satisfy equation (1.2) (with |α|≡1), but part (b-ii) asserts that all b.c.-solutionsarecontainedinthisclass. Proof of Theorem2.3. Considerthestoppingtimeτ1:=inf{n≥1: An=1}.Withp1:=P(α=1)< 9 1andq1:=P(α=−1)=1−p1>0,thedistributionofτ1isgivenby4 P(τ1=1)=p1, P(τ1=n)=q12p1n−2 (n≥2), (2.5) ....rspa henceτ1<∞a.s.SinceAτ1=1(a.s.),from(1.14)wehaveXτ1=x−Dτ1,andbyLemma1.4 .....roy .a y(x)=E{y(x−Dτ1)}, x∈R. (2.6) ...lso .c Now, in view of the Choquet–Deny theorem, we need to investigate whether the random shift ..ie aDsτβ1L+eista(nnβodn+n-)βaa−rni,tdrhem(sβpen−etic)ct,ibvie.eetl.wyP.o(CDsoeτnq1du∈ietλinoZcne)isn<ogf1oif.nio.dτr1.arnaanynddλou∈msiRnv.gar(i1a.1b5le)saenadch(2,.w5)i,thwtehoebstaaminedistribution ...........typublishin P(Dτ1∈λZ)=p1P(β1+∈λZ)+ ∞ q12p1n−2P(β1−−β2+−···−βn+−1−βn−∈λZ) ......g.org nX=3 ..P ≤p1+q12 ∞ p1n−2=1. (2.7) ......rocR nX=2 ..S Ifp1>0then(2.7)impliesthatP(Dτ1∈λZ)<1unlessβ1+∈λZ(a.s.)andforalln≥2 .....ocA β1−−β2+−···−βn+−1−βn−∈λZ (a.s.). (2.8) ......0000 Snin≥c2e;ahllenβci+e,acroeni.di.idti.o,nthse(2fi.8rs)taorfetrheedsueciendclutosioβn−s−imβp−l∈iesλZtha(at.sβ.2)+. +In·t·u·r+n,βtn+h−e1la∈stλZcoan.ds.itfioornailsl .....000 1 2 equivalenttoβ−∈λ0+λZforsomeλ0∈R.Indeed,applyingLebesgue’sdecompositiontheorem [15,p.142]toeachofthei.i.d.randomvariablesβ−andβ−,itisevidentthatthecontinuouspart 1 2 oftheircommondistributionmustvanish,sothatthisdistributionispurelydiscrete;furthermore, its(countable)support{bi}satisfiestheconditionbi−bj∈λZforalli,j,andtheclaimfollows. Asimilarargumentisalsovalidforp1=0,wherebyβ+=0, β−=β,and(2.7)simplifiesto P(Dτ1∈λZ)=P(β1−β2∈λZ)≤1. Thiscompletestheproofofparts(a)and(b-i). Finally,weprovepart(b-ii),wherebyβ+∈λZ,β−∈λ0+λZandDτ1∈λZ(a.s.).ByTheorem 2.2,anyb.c.-solutionofequation(2.6)mustbeoftheformy(x)=g(x/λ),withsome1-periodic functiong(·).Substitutingthisintotheoriginalequation(2.6)(with|α|≡1)weget g(x/λ)=p1E{g((x−β+)/λ)}+q1E{g((−x+β−)/λ)} =p1g(x/λ)+q1g((−x+λ0)/λ), (2.9) andsinceq16=0itfollowsthatg(·)satisfiesthefunctionalequationg(x)=g(−x+λ0/λ) (x∈R), whichisequivalenttothesymmetryconditionstatedinthetheorem. Remark2.6. Ifα≡1thenβ+≡β, β−≡0,andparts(a)and(b-ii)ofTheorem2.3formallyrecover theChoquet–Denytheorem.Noextrarequirementong(·)arisesfrom(2.9),sinceq1=0. Remark 2.7. Note that the values α=1 and α=−1 (and the corresponding conditional distributions of β represented by β+ and β−, respectively) feature in Theorem 2.3 in a non- symmetricway:e.g.ifβ+hasanon-arithmeticdistributionthen,accordingtopart(a),thereareno b.c.-solutionsexceptconstants,irrespectivelyofβ−;however,ifβ+isarithmeticwithspanλ6=0 whilst β−is non-arithmetic but supported on a set λ0+λZ (i.e. with λ06=0 incommensurable withλ)thenthereexistnon-trivialb.c.-solutions,accordingtopart(b-ii). Example2.1. Theorem2.3withα≡−1isexemplifiedbytheequations (a) y(x)= ∞y(t−x)e−tdt (equivalenttothepantographequationy′(x)+y(x)=y(−x), 0 cf.(1.7)R),whichbypart(b-i)hasonlyconstantb.c.-solutions; 4Theseformulasincludethecasep1=0(undertheconvention00:=1),wherebyτ1=2a.s.
Description: