Probab.TheoryRelat.Fields DOI10.1007/s00440-017-0777-x Mass concentration and aging in the parabolic Anderson model with doubly-exponential tails MarekBiskup1,2 · WolfgangKönig3,4 · RenatoS.dosSantos3 Received:4September2016/Revised:19March2017 ©Springer-VerlagBerlinHeidelberg2017 Abstract Westudythenon-negativesolutionu =u(x,t)totheCauchyproblemfor theparabolicequation∂ u =Δu+ξuonZd×[0,∞)withinitialdatau(x,0)=1 (x). t 0 HereΔisthediscreteLaplacianonZdandξ =(ξ(z))z∈Zd isani.i.d.randomfieldwith doubly-exponentialuppertails.We(cid:2)provethat,forlarget andwithlargeprobability, most of the total mass U(t) := u(x,t) of the solution resides in a bounded x neighborhood of a site Z that achieves an optimal compromise between the local t DirichleteigenvalueoftheAndersonHamiltonianΔ+ξandthedistancetotheorigin. Theprocessest (cid:4)→ Z andt (cid:4)→ 1logU(t)areshowntoconvergeindistributionunder t t suitable scaling of space and time. Aging results for Z , as well as for the solution t totheparabolicproblem,arealsoestablished.Theproofusesthecharacterizationof eigenvalue order statistics for Δ+ξ in large sets recently proved by the first two authors. MathematicsSubjectClassification 60H25·82B44 ©2017 MarekBiskup,WolfgangKönigandRenatoS.dosSantos.Reproduction,byanymeans,ofthe entirearticlefornon-commercialpurposesispermittedwithoutcharge. B MarekBiskup [email protected] 1 DepartmentofMathematics,UCLA,LosAngeles,CA,USA 2 CenterforTheoreticalStudy,CharlesUniversity,Prague,CzechRepublic 3 Weierstraß-InstitutfürAngewandteAnalysisundStochastik,Berlin,Germany 4 InstitutfürMathematik,TechnischeUniversitätBerlin,Berlin,Germany 123 M.Biskupetal. Contents 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Mainresults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Assumptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Results:Massconcentration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Results:Scalinglimit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Results:Aging. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5 Results:Limitprofiles. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Connectionsandheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Relationstoearlierwork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Someheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Mainresultsfromkeypropositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Definitionofthelocalizationprocess . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Propertiesofthecostfunctional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Massdecompositionandnegligiblecontributions . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Proofofmassconcentrationresults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6 Proofofagingandlimitprofiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Preparations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Potentialsandeigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Islands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Connectivitypropertiesofthepotentialfield . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Spectralbounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Pathexpansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 Keypropositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Massofthesolutionalongexcursions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Equivalenceclassesofpaths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 ProofofPropositions6.1–6.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Analysisofthecostfunctional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1 Apointprocessapproach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Orderstatistics:ProofofPropositions7.1and4.5andTheorem2.6 . . . . . . . . . . . . . . . 8 Massdecomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1 Lowerboundforthetotalmass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Macroboxtruncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Negligiblecontributions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4 UpperboundforthetotalmassandproofofTheorem2.5 . . . . . . . . . . . . . . . . . . . . 9 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Pathlocalization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.1 Fastapproachtothelocalizationcenter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Localconcentration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Localprofiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Contributionofu(2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Contributionofu(1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Appendix:Atailestimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Appendix:Compactification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 Appendix:Propertiesofthecostfunctional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Introduction RandomSchrödingeroperators—mostnotably,theAndersonHamiltonian H =Δ+ ξ—havebeenasubjectofintenseresearchoverseveraldecades.Mostoftheattention hasbeenpaidtothecharacterofthespectrumandtheensuingphysicalconsequences forthequantumevolution.However,theassociatedparabolicproblem—characterized 123 MassconcentrationandagingintheparabolicAnderson… bythePDE∂ u =Δu+ξu—isofasmuchinterestbothfortheoryandapplications. t Herewestudythelatterfacetofthisproblemforaspecificclassofrandompotentials. OurmainresultistheproofoflocalizationofthesolutiontotheabovePDEforlarge timeinaneighborhoodofaprocessdeterminedsolelybytherandompotential. A standard way to describe the parabolic Anderson model (PAM) is via a non- negativesolutionu: Zd ×[0,∞)→[0,∞)oftheCauchyproblem ∂ u(z,t)=Δu(z,t)+ξ(z)u(z,t), z ∈Zd, t ∈(0,∞), (1.1) t u(z,0)=1 (z), z ∈Zd. (1.2) 0 Hereξ =(ξ(z))z∈Zd isani.i.d.randompotentialtakingvaluesin[−∞,∞),1x isthe indicatorfunctionofapoint x ∈ Zd,∂ abbreviatesthederivativewithrespecttot, t andΔisthediscreteLaplacianactingon f : Zd →Ras (cid:3) (cid:4) (cid:5) Δf(z):= f(y)− f(z) , (1.3) y:|y−z|=1 where|·|denotesthe(cid:5)1normonZd. The interest in(1.1–1.2) for mathematics as wellas applications comes fromthe competingeffectofthetwotermsontheright-handsideof(1.1).Indeed,theLaplacian tendstomakethesolutionsmootherovertime,whilethefieldmakesitrougher.The problem(1.1)appearsinthestudiesofchemicalkinetics[13],hydrodynamics[8],and magneticphenomena[23].Werefertothereviews[8,19]formorebackground,and to[13]forthefundamentalmathematicalpropertiesofthemodel.Arecentcompre- hensivesurveyofmathematicalresultsonthePAMandrelatedmodelscanbefound in[15];therelatedspectralorder-statisticsquestionsarereviewedin[3]. A non-negative solution to the Cauchy problem (1.1–1.2) exists and is unique as soon as the upper tail of [ξ(0)/logξ(0)]d is integrable [13]. Under this condition, thereisalsoarepresentationintermsofthechanged-pathmeasure, (cid:6)(cid:7) (cid:8) 1 t Q(ξ)(dX):= exp ξ(X )ds P (dX), (1.4) t U(t) s 0 0 on nearest-neighbor paths X = (Xs)s≥0 on Zd, where P0 stands for the law of a continuous-time random walk on Zd (with generator Δ) started at zero. Indeed, the Feynman–Kacformulashows (cid:9) (cid:11) (cid:10) u(z,t)=U(t)Q(tξ)(Xt =z)=E0 e 0tξ(Xs)ds1{Xt=z} , (1.5) wherebythenormalizationconstantU(t)obtainsthemeaning (cid:12) (cid:7) (cid:13) (cid:3) t U(t)= u(x,t)=E exp ξ(X )ds . (1.6) 0 s x∈Zd 0 The aforementioned competitio(cid:10)n is now obvious probabilistically: the walk would liketomaximizethe“energy” tξ(X )ds,byspendingitstimeatthesiteswhereξ 0 s islarge,againstthe“entropy”ofsuchtrajectoriesunderthepathmeasureP . 0 123 M.Biskupetal. Analternativeandequallyusefulwaytoview(1.1)isasthedefinitionofasemigroup t (cid:4)→et(Δ+ξ)on(cid:5)2(Zd).Thesolutionto(1.1–1.2)isthengivenby (cid:14) (cid:15) u(x,t)= 1 ,et(Δ+ξ)1 . (1.7) x 0 (cid:5)2(Zd) Thisopensupthepossibilitytocontrolthelarge-t behaviorthroughspectralanalysis of the Anderson Hamiltonian. To this end, it is useful to restrict the problem to a sufficiently large (in t-dependent fashion) finite volume Λ ⊂ Zd (with 0 ∈ Λ) as follows.DenotebyHΛtheAndersonHamiltonianinΛwith(zero)Dirichletboundary conditions,i.e.,forφ ∈RΛ,HΛφ = Hφ˜whereH =Δ+ξandφ˜istheextensionofφ toRZd thatisequaltozeroonΛc.LetuΛbethesolutionto(1.1–1.2)restrictedtoΛ andwiththeright-handsideof(1.1)substitutedbyHΛu.Thentheaboveinterpretation yields (cid:3)|Λ| uΛ(x,t)= etλ(Λk)φΛ(k)(x)φΛ(k)(0), (1.8) k=1 whereλ(Λk) aretheeigenvaluesandφΛ(k) thecorrespondingeigenvectorsof HΛ,which weassumetobeorthonormalin(cid:5)2(Λ).Hereafter,weextendboththesolutionuΛ(·,t) andtheeigenfunctionsof HΛtoZd bysettingthemtobeequalto0onΛc. The competition we described in the context of the changed-path measure (1.4) nowmanifestsitselfasfollows.Theterminthesumin(1.8)thatgrowsthefastestint isthatwiththelargesteigenvalue.However,thereisnoapriorireasonforittobethe dominanttermatafixedtime.Indeed,aneigenvaluewillonlycontributeto(1.8)when itseigenvectorputsnon-trivialmassonboth0andx.Sincetheleadingeigenvectors decay exponentially away from their localization centers (Anderson localization), |φ(k)(0)|willinfactbetypicallyextremelysmall.Itisthusthecombinedeffectofboth Λ etλ(Λk) andφΛ(k)(x)φΛ(k)(0)thatdecideswhichindexk willgivethemaincontributionto thesum. In the present paper, we analyze these competing effects for a class of random potentialswithuppertailsclosetothedoubly-exponentialdistribution,characterized by (cid:16) (cid:17) (cid:18) (cid:19) Prob ξ(0)>r =exp −er/ρ , r ∈R, (1.9) whereρ ∈(0,∞).(PrecisedefinitionswillappearinSect.2.)Forthesepotentialswe showthat,atalllarget,mostofthetotalmassU(t)ofthesolutionresidesinabounded neighborhood of a random point Z determined entirely by ξ. This point marks the t optimal local peak of ξ for the strategy where the random walk in (1.4) traverses to Z intimeo(t),andthereafter“sticksaround” Z inordertoenjoythebenefitsofa t t “strong”localDirichleteigenvalue.Wealsocharacterizethescalinglimitsof Z and t 1logU(t),andobtainagingresultsforboth Z andu(x,t). t t OurresultsbuildonalargebodyofliteratureonthePAMwhosefullaccounthere woulddivertfromthemainmessageofthepaper.Fornowletusjustsaythatweextend resultsfrom[9,17,21,26],dealingwithlocalizationononelatticesite,toabenchmark classofrandompotentialsexemplifiedby(1.9),wherethelocalizationtakesplacein large domains, albeit not growing with t. An important technical input for us is the 123 MassconcentrationandagingintheparabolicAnderson… recentwork[7],whereeigenvalueorderstatisticsfortheAndersonHamiltonianH = Δ + ξ was characterized for this class of ξ. Further connections will be given in Sect.3.1. 2 Mainresults Wenowmovetothestatementsofourmainresults.Throughoutthepaper,lnxdenotes the natural logarithm of x, and ln x := lnlnx, ln x := lnlnlnx, etc denote its 2 3 iterates.Wewilluse“Prob”todenotetheprobabilitylawofthei.i.d.randomfieldξ. 2.1 Assumptions We begin by identifying the class of potentials that we will consider in the sequel. Besidessomeregularity,thefollowingensuresthattheuppertailsofξ(0)areinthe vicinityofthedoubly-exponentialdistribution(1.9). Assumption2.1 (Uppertails)Supposethatesssupξ(0)=∞andlet 1 F(r):=ln , r >essinfξ(0). (2.1) 2 Prob(ξ(0)>r) Weassumethat F isdifferentiableonitsdomainandthat 1 lim F(cid:8)(r)= forsome ρ ∈(0,∞). (2.2) r→∞ ρ Theassumptionabove isexactly asAssumption1.1in[7],and impliesAssump- tion (F) of [14]. While the latter would be enough for most of our needs, the extra requirements of Assumption 2.1 are used in the crucial step, performed in [7], of identifying the max-order class of the local principal eigenvalues of the Anderson Hamiltonian. In order to avoid technical inconveniences, we will also assume the followingconditiononthelowertailofξ. Assumption2.2 (Lowertails)Letξ−(x):=max{0,−ξ(x)}.Weassumethat (cid:7) ∞ (cid:16) (cid:17) Prob ξ−(0)>es d1ds <∞. (2.3) 0 Assumption2.2isonlyusedintheproofofLemma8.1,whichisusedinProposi- tion4.6togivealowerboundforthetotalmassU(t).Notethat(2.3)holdswhenever ln(1+ξ−(0))hasa(d+ε)-thfinitemoment(cf.[18]).Webelievethat,withtheuseof percolationarguments,thisassumptioncanberelaxedtoξ(0) > −∞almostsurely ind ≥2.Ind =1,(2.3)isequivalenttoln(1+ξ−(0))havingthefirstmoment,which isknowninthecaseofboundedpotentialstobe“essentiallynecessary”inthesense that,when|ln(1+ξ−(0))|δ isnotintegrableforsomeδ ∈ (0,1),thesolutionmight scaledifferently.See[6],inparticularRemarks3and4therein. 123 M.Biskupetal. WewillassumethevalidityofAssumptions2.1–2.2throughouttherestofthepaper withoutexplicitlystatingthisineachinstance. 2.2 Results:Massconcentration Recallthat|x|denotesthe(cid:5)1-normof x.Ourfirstresultconcernstheconcentration ofthetotalmassofthesolutiontotheCauchyproblem(1.1–1.2): Theorem2.3 (Mass concentration) There is a Zd-valued càdlàg stochastic process (Zt)t>0 dependingonlyonξ suchthatt (cid:4)→|Zt|isnon-decreasingandsuchthatthe followingholds:Foreachδ >0,thereexistsR ∈Nsuchthat,foranyl >0satisfying t limt→∞ 1tlt =0, ⎛ ⎞ lim Prob ⎝ sup (cid:3) u(x,s) >δ⎠=0. (2.4) t→∞ s∈[t−lt,t+lt] x: |x−Zt|>R U(s) Inwords,(2.4)meansthatthesolutionattimet iswithlargeprobabilityconcentrated nearasinglepointZ ,andthecontrolinfactextendstosublinearly-growingintervals t of time around t. This cannot be extended to linearly growing time-intervals due to the jumps of the process s (cid:4)→ Z (cf. Theorem 2.6 below), but a refinement of our s methodswouldshowthat,inthiscase,twoislandswouldsuffice,i.e.,(2.4)wouldstill holdifthesumistakenoverboxesofradius R centeredaroundtwoprocesses Z(1), s Z(2)[see(4.9)].Wealsobelievethatthealmost-sureversionofthisstatement,dubbed s asa“two-citiestheorem”andprovedin[16]forthecaseofParetopotentials,could beobtainedwithmoreworkbutprefernottopursuethishere. IntermsofthepathmeasureQ(ξ),Theorem2.3canbeinterpretedasconcentration t forthelawofthepositionofthepathattimet.Bylettingtheradius Rgrowslowlyto infinity,thiscanbeimprovedtoincludeamajorityoftherandomwalkpath: Theorem2.4 (Pathlocalization)Foranyεt ∈(0,1)withlimt→∞εtln3t =∞, (cid:24) (cid:25) lim Q(ξ) sup |X −Z |>ε lnt =0 inprobability, (2.5) t→∞ t s∈[εtt,t] s t t where(Zt)t>0 isthestochasticprocessinTheorem2.3. To the best of our knowledge, statements about path localization such as Theo- rem2.4werenotyetavailableintheliteratureoftheParabolicAndersonModel.The scalesabovecomeoutofourmethodsandmaybeartificial;inparticular,wedonot knowiflnt/ln3t isthecorrectscalingforsupεtt≤s≤t|Xs −Zt|. 2.3 Results:Scalinglimit Ournexttheoremidentifiesthelarge-t behaviorofthepairofprocessest (cid:4)→ Z and t t (cid:4)→ 1lnU(t).WhileU(t)iscontinuous, Z isonlycàdlàgandthusitisnaturalto t t 123 MassconcentrationandagingintheparabolicAnderson… usetheSkorohodtopologytodiscussdistributionalconvergence.Tworelevantscales are ρ td ρ t d := and r := t = , (2.6) t t dlnt ln t dlnt ln t 3 3 markingthesizeoffluctuationsof 1lnU(t),andthetypicalsizeof|Z |. t t Todescribethescalinglimit,considerasample{(λ ,z ): i ∈N}fromthePoisson i i pointprocessonR×Rd withintensitymeasuree−λdλ⊗dz.Forθ >0,define |z| ψθ(λ,z):=λ− , (λ,z)∈R×Rd. (2.7) θ Itcanbechecked that,foreveryθ > 0,theset{ψθ(λi,zi): i ∈ N}isbounded and locallyfinite.Moreover,themaximizingpointisuniqueatallbutatmostacountable setofθ’sandwecanthusdefine(Λθ,Zθ)tobethecàdlàgmaximizerofψθ overthe samplepointsoftheprocess(cf.Sect.7.2).Weset Ψθ :=ψθ(Λθ,Zθ). (2.8) Thenwehave: Theorem2.5 (Scalinglimitofthelocalizationprocessandthetotalmass)Thereisa non-decreasingscalefunctiona >0obeying t a lim t =ρ (2.9) t→∞ln2t suchthatthefollowingholds:Thestochasticprocess(Zt)t>0inTheorems2.3and2.4 canbechosensuchthat,foralls ∈(0,∞)andrelativetotheSkorohodtopologyon D([s,∞),R×Rd), (cid:24) (cid:25) θ1t lnU(dθtt)−art, Zrθtt θ∈[s,∞) t−→la→w∞ (cid:16)Ψθ,Zθ(cid:17)θ∈[s,∞). (2.10) Inparticular,foreachθ >0,thepair([θ1t lnU(θt)−art]/dt,Zθt/rt)convergesinlaw tothepair(Ψθ,Zθ)∈R×Rd whosecoordinatesareindependentanddistributedas follows:Ψθ followsaGumbeldistributionwithscale1andlocationdln(2θ),while Zθ has i.i.d. coordinates, each of which is Laplace-distributed with location 0 and scaleθ. Thescalingfunctiona characterizestheleading-orderscaleoftheprincipalDirich- t leteigenvalueoftheAndersonHamiltonianinaboxofradiust,asidentifiedin[7]. See(7.3)belowforaprecisedefinition. 123 M.Biskupetal. 2.4 Results:Aging Thetechniquesusedtoprovetheabovetheoremsalsopermitustoaddressthephe- nomenon of aging in the problem under consideration. The term “aging” usually referstothefactthatcertaindecisivechangesinthesystemoccurattimescalesthat increase proportionally to the age of the system. Our next result addresses aging in theprocess(Zt)t>0: Theorem2.6 (Aging for the localization process) For each s > 0, and for (Zt)t>0 and(Zt)t>0 asinTheorems2.3,2.4and2.5, (cid:16) (cid:17) (cid:16) (cid:17) tl→im∞Prob Zt+θt = Zt ∀θ ∈[0,s] =tl→im∞(cid:16)Prob Zt+st =(cid:17) Zt (2.11) =Prob Z1+s = Z1 =Prob(Θ >s), wheretherandomvariable Θ :=inf{θ >0: Z1+θ (cid:12)= Z1} (2.12) ispositiveandfinitealmostsurely.Moreover, sd dd lim Prob(Θ >s)= . (2.13) s→∞(logs)d d! InlightofTheorem2.5,Theorem2.6canbeseenasareflectionofthefactthatthe functionalconvergencestatedinTheorem2.5isnotachievedthroughalargenumber ofmicroscopicjumps,butratherthroughsporadicmacroscopicjumps. Oursecondagingresultdealswiththejumpsintheprofileofthenormalizedsolution u(·,t)/U(t).Itcomesasaconsequenceofthemassconcentrationofthenormalized solutionaround Z togetherwithTheorem2.6. t Theorem2.7 (Agingforthesolution)Foranyε ∈(0,1),therandomvariable ⎧ ⎫ 1t inf⎨⎩s >0: (cid:3) (cid:29)(cid:29)(cid:29)(cid:29)uU(x(,tt++ss)) − uU(x(,t)t)(cid:29)(cid:29)(cid:29)(cid:29)>ε⎬⎭ (2.14) x∈Zd convergesindistributionast →∞totherandomvariableΘ definedin(2.12). AkeypointtonoteaboutTheorem2.7isthatthelimitingrandomvariabledoesnot dependonε.Thissuggeststhat,infact,thesumin(2.14)jumpsfromvaluesnear0 tovaluesnear1ass variesinatimeintervalofsublinearlengthint. 2.5 Results:Limitprofiles ThelocalizationstatedinTheorem2.3canbegiveninamorepreciseformprovided thatwemakeanadditionaluniquenessassumption.Inordertostatethisassumption, 123 MassconcentrationandagingintheparabolicAnderson… weneedfurtherdefinitions.GivenapotentialV: Zd →R,let (cid:3) V(x) L(V):= e ρ . (2.15) x∈Zd ThefunctionalLplaystheroleofalargedeviationratefunctionforrandompotentialsξ withdoubly-exponentialtails.WheneverL(V)<∞(infact,wheneverV(x)→−∞ as|x|→∞),Δ+V hasacompactresolventasanoperatoron(cid:5)2(Zd),anditslargest eigenvalueλ(1)(V)iswell-definedandsimple.Theconstant (cid:18) (cid:19) χ =χ(ρ):=−sup λ(1)(V): V ∈RZd, L(V)≤1 ∈ [0,2d] (2.16) iskeyintheanalysisoftheasymptoticgrowthofU(t).Thesetofcenteredmaximizers ! " M∗ρ := V ∈RZd: 0∈argmax(V),L(V)≤1 and λ(1)(V)=−χ (2.17) isknowntobenon-empty.Theassumptionbelowdealswithuniqueness: Assumption2.8 (Uniqueness of maximizer) We assume that M∗ρ = {Vρ}, i.e., the variationalproblem(2.16)admitsauniquecenteredsolutionVρ. Theuniquenessofthecenteredminimizerisconjecturedtoholdforallρ >0,but hassofaronlybeenprovedforρ largeenough;see[11].Inthelatterpaperitisalso shownthat,foranyV ∈M∗ρ,thenon-negativeprincipaleigenfunctionoftheoperator Δ+V isstrictlypositiveandliesin(cid:5)1(Zd).UnderAssumption(2.8),wewilldenote henceforthbyvρ theprincipaleigenfunctionofΔ+Vρ,normalizedsothat vρ >0 and (cid:14)vρ(cid:14)(cid:5)1(Zd) =1. (2.18) Thenwehave: Theorem2.9 (Limiting profiles) Suppose Assumption 2.8 and let (Zt)t>0 be the process from Theorems 2.3, 2.4 and 2.5. There exist μ ∈ N and#a > 0 satisfy- t t inglimt→∞μt =∞andlimt→∞#at/(ρln2t)=1suchthat,forallε ∈(0,1), (cid:29) (cid:29) s∈[εstu,εp−1t] x∈Zds:u|px|≤μt(cid:29)ξ(x +Zs)−#at −Vρ(x)(cid:29) t−→→∞ 0 inprobability. (2.19) Moreover,foranylt >0satisfyinglimt→∞ 1tlt =0, (cid:29) (cid:29) sup (cid:3) (cid:29)(cid:29)(cid:29)u(Zt +x,s) −vρ(x)(cid:29)(cid:29)(cid:29) −→ 0 inprobability. (2.20) s∈[t−lt,t+lt]x∈Zd U(s) t→∞ Thescale#a in(2.19)coincides(uptotermsthatvanishast →∞)withthemaximum t ofξinsideaboxofradiust[see(5.1)forthedefinition,andalsoLemma5.1].Moreover, the scales at and#at (with at as in Theorem 2.5) satisfy limt→∞#at −at = χ. The 123 M.Biskupetal. scaleμ providedintheproofofTheorem2.9satisfiesμ (cid:15)(lnt)κforsomearbitrary t t κ <1/d,butitsactualrateofgrowthisnotcontrolledexplicitly. Therestofthepaperisorganizedasfollows.InSect.3belowwediscussconnec- tions to the literature and provide some heuristics. Section 4 contains an extensive overview of our proofs including the definition of the localization process Z . The t technicalcoreofthepaperisformedbySect.5(propertiesofthepotentialandspec- tral bounds), Sect. 6 (path expansions) and Sect. 7 (a point process approach). The bulkoftheproofsrelatedtoourmainresultsiscarriedoutinSects.8–11,concern- ingrespectivelynegligiblecontributionstotheFeynman–Kacformula,localizationof relevanteigenfunctions,pathlocalizationpropertiesandtheanalysisoflocalprofiles. TheproofsofsometechnicalresultsaregiveninAppendices12–14. 3 Connectionsandheuristics Inthissection,wemakeconnectionstoearlierworkonthisproblem,andalsoprovide ashortheuristicargumentmotivatingthedefinitionofthescalesin(2.6). 3.1 Relationstoearlierwork Letusgiveaquicksurveyonearlierworksontheparticularquestionthatweconsider; wereferto[15]foracomprehensiveaccountontheparabolicAndersonmodel,and to[20]forasurveyoncertainaspectscloselyrelatedtothepresentpaper. Since1990,muchoftheeffortwentintodevelopingacharacterizationofthelog- arithmicasymptotics oft (cid:4)→ U(t)anditsmoments,whichareallfiniteifandonly ifallthepositiveexponentialmomentsofξ(0)arefinite.Forthiscase,underamild regularityassumption,[27]identifiedfouruniversalityclassesofasymptoticbehav- iors: potentials with tails heavier than (1.9) (corresponding formally to ρ = ∞), double-exponentialtailsoftheform(1.9),theso-called“almostbounded”potentials (correspondingformallytoρ =0),andboundedpotentials.Thefirsttwocaseswere treatedin[14],andthelasttwoin[27]and[5],respectively.Potentialswithinfinite exponentialmomentswereanalysedin[28](moreprecisely,ParetoandWeibulltails), whereweaklimitsandalmostsureasymptoticsforU(t)wereobtained. Inalloftheclassesmentionedabove,theasymptoticsofU(t)isexpressedinterms of a variational principle for the local time of the path in Q(ξ) and/or the “profile” t of ξ that maximizes a local eigenvalue. The picture that emerges is that a typical path sampled from Q(ξ) for t large will spend an overwhelming majority of time in t a relatively small volume whose location is characterized by a favourable value of thelocalDirichleteigenvalue.Proofsofsuchstatementshavefirstbeenavailablefor a related version of the model using the method of enlargement of obstacles [25] andlateralsoforthedouble-exponentialclassbyprobabilisticpathexpansions[12]. However, neither of these approaches was sharp enough to distinguish among the many “favourable eigenvalues.” In fact, while the expectation was that only a finite numberofsucheigenvaluesneedstobeconsidered,thebestavailableboundontheir numberwasto(1). 123
Description: