TheAnnalsofAppliedProbability 2010,Vol.20,No.1,1–25 DOI:10.1214/09-AAP616 ©InstituteofMathematicalStatistics,2010 DYNAMICS OF THE TIME TO THE MOST RECENT COMMON ANCESTOR IN A LARGE BRANCHING POPULATION BY STEVEN N. EVANS1 AND PETER L. RALPH2 UniversityofCaliforniaatBerkeleyandUniversityofCaliforniaatDavis If we follow an asexually reproducing population through time, then theamountoftimethathaspassedsincethemostrecentcommonancestor (MRCA)ofallcurrentindividualslivedwillchangeastimeprogresses.The resulting“MRCAage”processhasbeenstudiedpreviouslywhenthepopu- lationhasaconstantlargesizeandevolvesviathediffusionlimitofstandard Wright–Fisherdynamics.Foranypopulationmodel,thesamplepathsofthe MRCAageprocessaremadeupofperiodsoflinearupwarddriftwithslope +1punctuatedbydownwardjumps.WebuildotherMarkovprocessesthat havesuchpathsfromPoissonpointprocessesonR++×R++withintensity measuresoftheformλ⊗μwhereλisLebesguemeasure,andμ(the“family lifetimemeasure”)isanarbitrary,absolutelycontinuousmeasuresatisfying μ((0,∞))=∞andμ((x,∞))<∞forallx>0.Specialcasesofthiscon- structiondescribethetimeevolutionoftheMRCAagein(1+β)-stablecon- tinuousstatebranchingprocessesconditionedonnonextinction—aparticular caseofwhich,β=1,isFeller’scontinuousstatebranchingprocesscondi- tioned on nonextinction. As well as the continuous time process, we also consider the discrete time Markov chain that records the value of the con- tinuousprocessjustbeforeandafteritssuccessivejumps.Wefindtransition probabilitiesforboththecontinuousanddiscretetimeprocesses,determine whentheseprocessesaretransientandrecurrentandcomputestationarydis- tributionswhentheyexist.Moreover,weintroduceanewfamilyofMarkov processesthatstandsinarelationwithrespecttothegeneral(1+β)-stable continuousstatebranchingprocessanditsconditionedversionthatissimilar totheonebetweenthefamilyofBessel-squareddiffusionsandtheuncondi- tionedandconditionedFellercontinuousstatebranchingprocess. 1. Introduction. Anyasexuallyreproducingpopulationhasauniquemostre- centcommonancestor,fromwhomtheentirepopulationisdescended.Insexually reproducingspecies,thesameistrueforeachnonrecombiningpieceofDNA.For instance, our “mitochondrial Eve,” from whom all modern-day humans inherited ReceivedDecember2008;revisedApril2009. 1SupportedinpartbyNSFGrantsDMS-04-05778andDMS-09-07630. 2SupportedinpartbyaVIGREgrantawardedtotheDepartmentofStatistics,UniversityofCali- forniaatBerkeley. AMS2000subjectclassifications.92D10,60J80,60G55,60G18. Keywordsandphrases.Genealogy, most recent common ancestor, MRCA, continuous state branching, Poisson point process, Poisson cut-out, transience, recurrence, stationary distribution, duality,Bessel,self-similar,piecewisedeterministic. 1 2 S.N.EVANSANDP.L.RALPH theirmitochondrialDNA,isestimatedtohavelivedaround180,000yearsago[19] whileour“Y-chromosomalAdam”isestimatedtohavelivedaround50,000years ago[33].TherehavealsobeeneffortstoestimatethetimesincetheMRCAlived (whichwewillalsocallthe“ageoftheMRCA”)inpopulationsofotherorganisms, particularly pathogens [32, 34]. These studies, using sophisticated models incor- porating ofdemographic history, arefocused onestimatingtheageoftheMRCA atasinglepointintime(thepresent). As time progresses into the future, eventually the mitochondrial lineages of all but one of the daughters of the current mitochondrial Eve will die out, at which pointthenewmitochondrialEvewillhavelivedsomewhatlaterintime.Theageof theMRCAisthusadynamicallyevolvingprocessthatexhibitsperiodsofupward lineargrowthseparatedbydownwardjumps. Recently, [26] and [29] independently investigated the MRCA age process for thediffusionlimitoftheclassicalWright–Fishermodel.TheWright–Fishermodel isperhapsthemostcommonlyusedmodelinpopulationdynamics:eachindividual in a fixed size population independently gives birth to an identically distributed random number of individuals (with finite variance), and after the new offspring are produced, some are chosen at random to survive so that the total population sizeremainsconstant.Thediffusionlimitarisesbylettingthepopulationsizegoto infinityandtakingthetimebetweengenerationstobeproportionaltothereciprocal ofthepopulationsize. In this paper, we investigate the MRCA age process for a parametric family of population models in a setting in which the population size varies with time, and, by suitable choice of parameters, allows control over the extent to which rareindividuals can have large numbers of offspring that survive to maturity. The model for the dynamics of the population size is based on the critical (1+β)- stablecontinuousstatebranchingprocessfor0<β ≤1.Theseprocessesariseas scalinglimitsofGalton–Watsonbranchingprocessesasfollows. (n) Write Z for the number of individuals alive in a critical continuous time t Galton–Watson branching process with branching rate λ and offspring distribu- tionγ.Thedistributionγ hasmean1(andthus,theprocessis“critical”).Suppose thatifW isarandomvariablewithdistributionγ,thentherandomwalkwithsteps distributed as the random variable (W −1) falls into the domain of attraction of a stable process of index 1+β ∈(1,2]. The case β =1 corresponds to γ having finite variance and the random walk converging to Brownian motion after rescal- ing. Set X(n) =n−1/βZ(n) and suppose that X(n) →x as n→∞. Then, up to t t 0 a time-rescaling depending on λ and the scaling of the stable process above, the processes X(n) converge to a Markov process X that is a critical (1+β)-stable continuous state branching process, and whose distribution is determined by the Laplacetransform (cid:2) (cid:3) θx (1.1) E[e−θXt|X =x]=exp − . 0 (1+θβt)1/β DYNAMICSOFTHEMRCA 3 Ifβ =1,thisisFeller’scriticalcontinuousstatebranchingprocess[17,22].(Note that time here is scaled by a factor of 2 relative to some other authors so that the generatorofour“Fellercontinuousstatebranchingprocess”isx ∂2 .) ∂x2 Let τ =inf{t >0:X =0} denote the extinction time of X (it is not hard to t showthatX =0forallt ≥τ).Takingθ →∞in(1.1)gives t (cid:2) (cid:3) x P{τ >t|X =x}=1−exp − , 0 t1/β so X dies outalmostsurely.However,itispossible tocondition X tolive forever inthefollowingsense: 1 lim E[f(X )|X =x,τ >T]= E[f(X )X |X =x]. t 0 t t 0 T→∞ x (cid:8) (cid:8)(cid:8) Thus if P (x ,dx ) are the transition probabilities of X, then there is a Markov t processY withtransitionprobabilities 1 Q (y(cid:8),dy(cid:8)(cid:8))= P (y(cid:8),y(cid:8)(cid:8))y(cid:8)(cid:8). t y(cid:8) t The process Y is the critical (1+β)-stable continuous state branching process X conditioned on nonextinction. The distribution of Y is determined by its Laplace t transform (cid:2) (cid:3) yθ (1.2) E[exp(−θY )|Y =y]=exp − (1+tθβ)−(β+1)/β t 0 (tθβ +1)1/β (see Section 6). Moreover, it is possible to start the process Y from the initial stateY =0,andtheformula(1.2)continuestoholdfory=0.Thesuper-process 0 generalizationofthisconstructionwasconsideredforβ =1in[11,14,15]andfor generalβ in[16]. For β =1, the conditioned process Y can be described informally as a single “immortalparticle”constantlythrowingoffinfinitesimallysmallmasseswitheach massthenevolvingaccordingtothedynamicsoftheunconditionedprocess.These infinitesimalmassescanbeinterpretedasthesingleprogenitorsoffamilieswhose lineagesplitsfromtheimmortalparticleatthebirthtimeoftheprogenitorandare eventually doomed to extinction. Most such families die immediately, but a rare few live for a noninfinitesimal amount of time. More formally, there is a σ-finite measureν onthespaceofcontinuouspositiveexcursionpaths E0:={u∈C(R+,R+):u0=0&∃γ >0s.t.ut >0⇔0<t <γ}, suchthatif(cid:10)isaPoissonpointprocessonR+×E0 withintensityλ⊗ν whereλ is Lebesgue measure, and (X¯t)t≥0 is an independent copy of X begun at X¯0=y, thentheprocess (cid:2) (cid:3) (cid:4) (1.3) X¯t + u(t−s)∨0 (s,u)∈(cid:10) t≥0 4 S.N.EVANSANDP.L.RALPH has the same distribution as (Yt)t≥0 begun at Y0 =y (see [15]). A point (s,u)∈ (cid:10) corresponds to a family that grows to nonnegligible size; the time s records the moment the family splits off from the immortal particle, and the value u of r the trajectory u gives the size of the family r units of time after it split off. The family becomes extinct after the period of time γ(u):=inf{r >0:u(t)=0,∀t > r}. The σ-finite measure ν may be identified explicitly, but it suffices to remark here that it is Markovian with transition probabilities the same as those of the unconditionedprocessX—inotherwords,ν arisesfromafamilyofentrancelaws ¯ for the semigroup of X. The process (Xt)t≥0 records the population mass due to descendants of individuals other than the immortal particle who are present at time0. AnanalogousdescriptionoftheconditionedprocessY forthecaseβ ∈(0,1)is presentedin[16].Thereisagainasingleimmortallineage,butnowfamiliessplit offfromthatlineagewithanoninfinitesimalinitialsize,reflectingtheheavy-tailed offspring distributions underlying these models. More precisely, a decomposition similar to (1.3) holds, but the Poisson point process (cid:10) is now on R+×E where E ={u∈D(R+,R+):∃γ >0s.t.ut >0⇔0≤t <γ}, the set of càdlàg paths starting above zero that eventually hit zero. The nondecreasing process (Mt)t≥0 where (cid:4) M := u t 0 (s,u)∈(cid:10)∩[0,t]×E isthetotaloftheinitialfamilysizesthatsplitofffromtheimmortalparticleinthe timeinterval[0,t].Itisastablesubordinatorofindexβ. Suppose now that β ∈ (0,1] is arbitrary. Take Y = 0 so that X¯ ≡ 0 in the 0 t decomposition (1.3) and all “individuals” belong to families that split off from the immortal particle at times s ≥0. Extend the definition of γ(u) given above for u ∈ E to u ∈ E in the obvious way. The individuals, besides the immortal 0 particlealiveattime t >0,belongtofamiliesthatcorrespondtothesubsetA := t {(s,u) ∈ (cid:10):0 ≤ t −s < γ(u)} of the random set (cid:10). At time t, the amount of timesincethemostrecentcommonancestoroftheentirepopulationlivedisA := t sup{t −s:(s,u)∈At}. As depicted in Figure 1, the MRCA age process (At)t≥0 hassaw-toothsamplepathsthatdriftupwithslope1untilthecurrentoldestfamily is extinguished, at which time they jump downward to the age of the next-oldest family. It is not necessary to know the Poisson point process (cid:10) in order to construct the MRCA age process (At)t≥0. Clearly, it is enough to know the point process (cid:11) on R+ ×R++ given by (cid:11):={(s,γ(u)):(s,u)∈(cid:10)}. Indeed, if we define the left-leaningwedgewithapexat(t,x)tobetheset (1.4) (cid:15)(t,x):={(u,v)∈R2:u<t &u+v>t +x}, then A =t −inf{s:∃x>0s.t.(s,x)∈(cid:11)∩(cid:15)(t,0)} t DYNAMICSOFTHEMRCA 5 FIG.1. Thepointsof(cid:11)aremarked“x”;thesamplepathoftheprocessAisdrawnwithasolidline; theoldestextantfamilyattimet isrepresentedbythepointin(cid:11)labeled“O”andtheleft-leaning wedge(cid:15)(t,0)isthedarklyshadedregionwithapexat(t,0).AttimeT,thefamilyrepresentedbyO willdieoutandthefamilyrepresentedbyNwillcontaintheMRCAbecauseOisontheboundaryof (thelightlyshaded)(cid:15)(T,0).IfthecoordinatesofO are(u,y)andthecoordinatesofN are(v,z), thenthesizeofthejumpthatAmakesatT isAT−AT−=u−v=(T−v)−(T−u)=(T−v)−y. (seeFigure1). Note that (cid:11) is a Poisson point process with intensity λ⊗μ where μ is the push-forward of ν by γ; that is, μ((t,∞))=ν({u:γ(u)>t}). We will show in Section5thatμ((t,∞))=(1+β)/(βt). With these observations in mind, we see that if (cid:11) is now any Poisson point processonR+×R++ withintensityλ⊗μwhereμisanymeasureonR++ with μ(R++)=∞ and 0<μ((x,∞))<∞ for all x >0, then the construction that built(At)t≥0 fromtheparticularpointprocess(cid:11)consideredabovewillstillapply andproduceanR+-valuedprocesswithsaw-toothsamplepaths.Wearetherefore ledtothefollowinggeneraldefinition. DEFINITION 1.1. Let(cid:11)beaPoissonpointprocessonR+×R++ withinten- sity measure λ⊗μ where λ is Lebesgue measure and μ is a σ-finite measure on R++ withμ(R++)=∞andμ((x,∞))<∞forallx>0.Define(At)t∈R+ by A :=t −inf{s≥0:∃x>0s.t.(s,x)∈(cid:11)∩(cid:15)(t,0)}, t where(cid:15)(t,0)isdefinedby(1.4),andA =0if(cid:11)∩(cid:15)(t,0)isempty. t We will suppose from now on that we are in this general situation unless we specify otherwise. We will continue to use terminology appropriate for the ge- nealogical setting and refer to (At)t≥0 as the MRCA age process and μ as the lifetimemeasure.Wewillassumeforconveniencethatthemeasureμisabsolutely continuous with a density m with respect to Lebesgue measure that is positive Lebesguealmosteverywhere.Itisstraightforwardtoremovetheseassumptions. 6 S.N.EVANSANDP.L.RALPH The strong Markov property of the Poisson point processes (cid:11) implies that (At)t≥0 is a time-homogeneous strong Markov process. In particular, there is a family of probability distributions (Px)x∈R+ on the space of R+-valued càdlàg paths with Px interpreted in the usual way as the “distribution of (At)t≥0 started from A =x.” More concretely, the probability measure Px is the distribution of 0 the process (Axt)t≥0 defined as follows. Let (cid:11)x be a point process on [−x,∞)× R++ that has the distribution of the random point set {(t −x,y):(t,y)∈(cid:11)}∪ {(−x,Z)}whereZ isanindependentrandomvariablethatisdefinedonthesame probability space ((cid:12),F,P) as the point process (cid:11), takes values in the interval (x,∞)andhasdistribution P{Z≤z}=μ((x,z])/μ((x,∞)). Thenwecandefine Ax :=t −inf{s≥−x:∃y>0s.t.(s,y)∈(cid:11)x ∩(cid:15)(t,0)}, t ≥0, t whereweadopttheconventionthatAx :=0ifthesetaboveisempty[seetheproof t ofpart(a)ofTheorem1.1belowforfurtherinformation].Fromnowon,whenwe speakoftheprocess(At)t≥0 wewillbereferringeithertotheprocessconstructed asinDefinition1.1fromthePoissonprocess(cid:11)definedonsomeabstractprobabil- ity space ((cid:12),F,P) or to the canonical process on the space of càdlàg R+-valued pathsequippedwiththefamilyofprobabilitymeasures(Px)x≥0.Thisshouldcause noconfusion. Wenoteinpassingthattheanalogueof(At)t≥0 intheconstantpopulationsize Wright–FishersettingisnotMarkov(seeRemark4.1.3of[26]). Weprovethefollowingpropertiesoftheprocess(At)t≥0 inSection2. THEOREM 1.1. (a) The transition probabilities of the time-homogeneous Markovprocess(At)t≥0 haveanabsolutelycontinuouspart (cid:2) (cid:5) (cid:3) μ((x,x+t]) x+t Px{A ∈dy}= exp − μ((z,∞))dz μ((y,∞))dy, t μ((x,∞)) y fory<x+t andasingleatom μ((x+t,∞)) Px{A =x+t}= . t μ((x,∞)) (b) Thetotalrateatwhichtheprocess(At)t≥0 jumpsfromstatex>0is m(x) , μ((x,∞)) and when the process jumps from state x >0, the distribution of the state to whichitjumpsisabsolutelycontinuouswithdensity (cid:2) (cid:5) (cid:3) x y(cid:17)→exp − μ((z,∞))dz μ((y,∞)), 0<y<x. y DYNAMICSOFTHEMRCA 7 (c) The probability P0{∃t > 0:At = 0} that the process (At)t≥0 returns to the statezeroispositiveifandonlyif (cid:5) (cid:2)(cid:5) (cid:3) 1 1 exp μ((y,∞))dy dx<∞. 0 x (d) If (cid:5) (cid:2) (cid:5) (cid:3) ∞ x exp − μ((y,∞))dy dx=∞, 1 1 then for each x >0 the set {t ≥0:A =x} is Px-almost surely unbounded. t Otherwise,limt→∞At =∞,Px-almostsurely,forallx≥0. (e) Astationarydistributionπ existsfortheprocess(At)t≥0 ifandonlyif (cid:5) ∞ μ((z,∞))dz<∞, 1 inwhichcaseitisunique,and (cid:2) (cid:5) (cid:3) ∞ π(dx)=μ((x,∞))exp − μ((z,∞))dz dx. x (f) If(At)t≥0 hasastationarydistributionπ,then (cid:2) (cid:5) (cid:3) ∞ μ([x,x+t)) d (Px{A ∈·},π)≤1−exp − μ((y,∞))dy × , TV t t+x μ([x,∞)) where d denotes the total variation distance. In particular, the distribution TV ofA underPx convergestoπ intotalvariationast →∞. t Specializing Theorem 1.1 to the case when A is the MRCA age process of the conditioned critical (1+β)-stable continuous state branching process gives parts (a)to(d)ofthefollowingresult.Part(e)followsfromanobservationthataspace– timerescalingofthisMRCAageprocessisatime-homogeneousMarkovprocess that arises from another Poisson process by the general MRCA age construction ofDefinition1.1.TheproofisinSection5. COROLLARY 1.1. Suppose that A is the MRCA age process associated with thecritical(1+β)-stablecontinuousstatebranchingprocess. (a) The transition probabilities of the process A have an absolutely continuous part (1+β)ty1/β Px{A ∈dy}= dy, 0<y<x+t, t β(x+t)2+1/β andasingleatom x Px{A =x+t}= . t x+t 8 S.N.EVANSANDP.L.RALPH (b) The total rate at which the process A jumps from the state x >0 is 1/x, and when it jumps from state x >0, the distribution of the state to which it jumps isabsolutelycontinuouswithdensity y1/β (1+1/β) , 0<y<x. x1+1/β (c) TheprobabilityP0{∃t >0:A =0}thattheprocessAreturnstothestatezero t is0. (d) Foreachx≥0,limt→∞At =∞,Px-almostsurely. (e) Theprocess (e−tAet)t∈R indexed by the whole real line is a time-homogeneous Markov process under Px for any x ≥0, and it is stationary when x =0. Moreover, A /t converges t in distribution to the Beta(1+1/β,1) distribution as t →∞ under Px for any x ≥0, and A /t has the Beta(1+1/β,1) distribution for all t >0 when t x=0. Note that the sample paths of (At)t≥0 have local “peaks” immediately before jumps and local “troughs” immediately after. We investigate the discrete time MarkovchainofsuccessivepairsofpeaksandtroughsinSection4.Wealsocon- siderthejumpheightsandinter-jumpintervalsanddescribeaninterestingduality betweenthesesequencesinSection3. Finally,recallthattheBessel-squaredprocessindimensionγ,whereγ isanar- bitrarynonnegativerealnumber,istheR+-valueddiffusionprocesswithinfinites- imalgenerator 2xd2/dx2+γ d/dx.When γ isapositiveinteger,suchaprocess has the same distribution as the square of the Euclidean norm of a Brownian mo- tion in Rγ. Feller’s critical continuous state branching process is thus the zero- dimensionalBessel-squaredprocess,moduloachoiceofscaleintimeorspace.It wasshowninExample3.5of[28]thatfor 0≤γ <2,theBessel-squaredprocess withdimension γ conditionedonneverhittingzeroistheBessel-squaredprocess with dimension 4−γ. Thus for β = 1, the conditioned process Y is the four- dimensional Bessel-squared process. We introduce a new family of processes in Section 6 that are also indexed by a nonnegative real parameter and play the role oftheBessel-squaredfamilyforvaluesofβ otherthan1.Theseprocesseswillbe studiedfurtherinaforthcomingpaper. We end this introduction by commenting on the connections with previous work. First, we may think of each point (s,x)∈(cid:11) as a “job” that enters a queue with infinitely many servers at time s and requires an amount of time x to com- plete.Wethushaveaclassical M/G/∞ queue[31],except weareassumingthat the total arrival rate of jobs of all kinds is infinite. With this interpretation, the quantity A is how long the oldest job at time t has been in the queue. Properties t DYNAMICSOFTHEMRCA 9 ofsuch M/G/∞ queues withinfinitearrivalrateshavebeenstudied(see,forex- ample,[10]),buttheageoftheoldestjobdoesnotappeartohavebeenstudiedin thiscontext. Second,notethattheprocess(At)t≥0isanexampleofapiecewisedeterministic Markov process: it consists of deterministic flows punctuated by random jumps. Such processes were introduced in [5] and studied further in [6] (see also [20], where the nomenclature jumping Markov processes is used). The general proper- tiesofsuchprocesseshavebeenstudiedfurtherin,forexample,[3,4,7]. Last, piecewise deterministic Markov processes like (At)t≥0 that have periods oflinearincreaseinterspersedwithrandomjumpshavebeenusedtomodelmany phenomena, such as stress in an earthquake zone [2], congestion in a data trans- missionnetwork[8]andgrowth-collapse[1].Theyalsohaveappearedinthestudy oftheadditivecoalescent[12]andR-tree-valuedMarkovprocesses[13]. 2. ProofofTheorem1.1. (a)Supposethat (At)t≥0 isconstructedfrom (cid:11) as in Definition 1.1. For s ≥x consider the conditional distribution of As+t given A = x. The condition A = x is equivalent to the requirements that there is a s s point (s −x,Z) in (cid:11) for some Z >x and that, furthermore, there are no points of (cid:11) in the left-leaning wedge (cid:15)(s −x,x). The conditional probability of the event {Z >z}, given A =x, is μ((z,∞))/μ((x,∞)) for z ≥x. If Z >x +t, s thenAs+t =x+t.Otherwise,As+t <x+t.Thesecondclaimofpart(a)follows immediately. Now consider P{As+t ∈ dy|As = x} for y < x +t. This case is depicted in Figure2.Byconstruction,As+t =y ifandonlyifthereisapoint(s+t−y,W)∈ (cid:11) forsome W >y,andtherearenopointsof (cid:11) in (cid:15)(s+t −y,y).Fromabove, theconditionA =x requirestheretobenopointsof(cid:11)inthewedge(cid:15)(s−x,x) s FIG.2. ThecomputationofP{As+t ∈dy|As=x}. 10 S.N.EVANSANDP.L.RALPH (thelightlyshadedregioninFigure2).Therefore,ifAs =x,thenAs+t =y ifand onlyifZ≤x+t,therearenopointsof(cid:11)inthedarklyshadedregionofFigure2 andthereisapointof(cid:11)oftheform(s+t −y,w)withw>y. Now,theconditionalprobabilityoftheeven{Z≤x+t}givenA =x is s μ((x,x+t]) . μ((x,∞)) Theprobabilitythatnopointsof(cid:11)areinthedarklyshadedregionis (cid:2) (cid:5) (cid:3) x+t exp − μ((u,∞))du . y The probability that (cid:11) has a point in the infinitesimal region [s +t −y,s +t − y +dy]×(y,∞) is μ((y,∞))dy. Multiplying these three probabilities together givesthefirstclaimofpart(a). (b)Bothclaimsfollowreadilybydifferentiatingtheformulaein(a)at t =0.It isalsopossibletoarguedirectlyfromtherepresentationintermsof(cid:11). (c)Supposethat(At)t≥0 isconstructedfrom(cid:11)asinDefinition1.1.Notethat (cid:6) (cid:7) {t >0:At =0}=R++ (t,t +x). (t,x)∈(cid:11) Thequestionofwhensucha“Poissoncutout”randomsetisalmostsurelyempty wasaskedin[23],andthenecessaryandsufficientconditionpresentedinpart(c) issimplytheonefoundin[30](seealso[18]). (d)Forx>0,therandomset{t ≥0:A =x}isadiscreteregenerativesetunder t Px thatisnotalmostsurelyequalto{0}.Hencethissetisarenewalprocesswithan inter-arrival time distribution that possibly places some mass at infinity; in which case, the number of arrivals is almost surely finite with a geometric distribution, andthesetisalmostsurelybounded. Suppose that the set {t ≥0:A =x} is Px-almost surely unbounded for some t x >0. Let 0=T <T <··· be the successive visits to x. It is clear that Px{∃t ∈ 0 1 [0,T ]:A = y} > 0 for any choice of y > 0 and hence, by the strong Markov 1 t property, the set {t ≥ 0:A = y} is also Px-almost surely unbounded. Another t application of the strong Markov property establishes that the set {t ≥0:A =y} t is Py-almost surely unbounded. Thus the set {t ≥0:A =x} is either unbounded t Px-almostsurelyforallx>0orboundedPx-almostsurelyforallx>0. Recall that away from the set {t ≥0:A =0} the sample paths of A are piece- t wiselinearwithslope1.Itfollowsfromthecoareaformula(see,e.g.,Section3.8 of[24])that (cid:5) (cid:5) ∞ ∞ f(A )dt = f(y)#{t >0:A =y}dy t t 0 0 foraBorelfunctionf :R+→R+.HencebyFubini’stheorem, (cid:5) (cid:5) (cid:5) ∞ ∞ ∞ Px{A ∈dy} f(y)Ex[#{t >0:A =y}]dy= f(y) t dtdy t dy 0 0 0
Description: