Interacting Measure-Valued Diffusions and their Long-Term Behavior
by
Hardeep Gill
B.Sc. Pure Mathematics and Statistics, Calgary, 2005 M.Sc. Mathematics, University of British Columbia, 2007
A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHEREQUIREMENTSFORTHEDEGREEOF
Doctor of Philosophy
in
THE FACULTY OF GRADUATE STUDIES (Mathematics)
The University Of British Columbia (Vancouver)
August 2011
c Hardeep Gill, 2011 Abstract
The focus of this dissertation is a class of random processes known as interacting measure-valued stochastic processes. These processes are related to another class of stochastic processes known as superprocesses. Both superprocesses and inter- acting measure-valued stochastic processes arise naturally from branching particle systems as scaling limits. A branching particle system is a collection of particles that propagate randomly through space, and that upon death give birth to a random number of particles (children). Therefore when the populations of the particle sys- tem and branching rate are large one can often use a superprocess to approximate it and carry out calculations that would be very difficult otherwise. There are many branching particle systems which do not satisfy the strong in- dependence assumptions underlying superprocesses and thus are more difficult to study mathematically. This dissertation attempts to address two measure-valued processes with different types of dependencies (interactions) that the associated particles exhibit. In both cases, the method used to carry out this work is called Perkins’ historical stochastic calculus, and has never before been used to investi- gate interacting measure-valued processes of these types. That is, we construct the measure-valued stochastic process associated with an interacting branching parti- cle system directly without taking a scaling limit. The first type of interaction we consider is when all particles share a common chaotic drift from being immersed in the same medium, as well as having other types of individual interactions. The second interaction involves particles that at- tract to or repel from the center of mass of the entire population. For measure- valued processes with this latter interaction, we study the long-term behavior of the process and show that it displays some types of equilibria.
ii Preface
The work provided below was conducted solely by myself, under the supervision of my advisor Professor Edwin A. Perkins. The problems investigated were orig- inally conceived by Professor Perkins. I proved the results (solutions of those problems), wrote the research articles and subsequently submitted them to journals of my choosing. This dissertation is based on two publications that resulted from my time as a doctoral student:
(1) Gill, H.S. A super Ornstein-Uhlenbeck process interacting with its center of mass. To appear.
(2) Gill, H.S. Superprocesses with spatial interactions in a random medium. Stochastic processes and their applications. December, 2009. 119 (12): 3981-4003.
Publication (1) forms the basis of Chapter 3, and (2) forms the basis for Chapter 2.
iii Table of Contents
Abstract ...... ii
Preface ...... iii
Table of Contents ...... iv
Acknowledgments ...... vi
Dedication ...... vii
1 Introduction ...... 1 1.1 Branching particle systems and superprocesses ...... 2 1.2 Superprocesses in a random medium ...... 5 1.3 An interacting super Ornstein-Uhlenbeck process ...... 8
2 Superprocesses with spatial interactions in a random media . . . . 12 2.1 Introduction ...... 12 2.2 Preliminaries ...... 13 2.3 The strong equation ...... 28 2.4 A martingale problem ...... 35 2.5 Additional properties of solutions ...... 40
3 A super Ornstein-Uhlenbeck process interacting with its center of mass ...... 45 3.1 Introduction ...... 45 3.2 Background ...... 48 3.3 Existence and preliminary results ...... 50
iv 3.4 Convergence ...... 66 3.5 The repelling case ...... 88
4 Conclusions ...... 95
Bibliography ...... 98
v Acknowledgments
There were many people who made this dissertation possible. Some of them im- pacted it directly, and some of them kept me in “Chahrdi Kala,” which in Punjabi means in high spirits. My advisor, Professor Edwin A. Perkins, has my deepest gratitude for his pa- tient and thoughtful guidance. I am especially grateful for the research problems he suggested, which provided me with a great deal of enjoyment and an invaluable source for learning. I quickly found that Ed is a veritable fount of ideas for those difficult times I was stuck while working on a problem. I also appreciate his careful readings of this manuscript and the papers which resulted from it. I would like to thank the members of my supervisory committee, particularly Martin Barlow and John Walsh, for their insightful comments and suggestions. I am grateful to Jie Xiong for suggesting additional avenues of research. My career path might have been completely different without Ms. Williams, my high school math teacher. Her creative and challenging class fostered my inter- est in mathematics. Kiran has my gratitude for her encouragement and support during the writing of this manuscript. This was nowhere more evident than by her willingness to undergo an allergic reaction rather than leave my doctoral defense room. Kiran, Lisa, Sidd, and my brothers Gurbeer and Harbir have my thanks for interspersing my studies with pleasant diversions which helped to keep me going. Lastly, I would like to thank the many other people with whom I have interacted during the course of this work, for making my stay at UBC vibrant and enjoyable.
vi Dedication
Dedicated to my parents, for their unconditional love and support. Their perseverance and hard work allowed me the opportunity to pursue my aspirations.
vii Chapter 1
Introduction
Branching processes are a well studied branch of probability and have been of in- terest since Galton and Watson first considered the problem of extinction in family names, in the late part of the nineteenth century. The process that Galton and Wat- son first studied consisted of particles that die after each generation leaving behind an independent and identically distributed number of offspring. If one considers a large Galton-Watson population, then its size can be approximated by what is known as Feller’s continuous state branching process. This in turn makes it much simpler to determine various properties of the original Galton-Watson process. The goal in what follows is to construct and study processes that act as analogues of Feller’s continuous state branching process for more complicated branching parti- cle systems which have interactions between particles. A branching particle system (BPS) can essentially be thought of as a Galton- Watson branching process where each particle alive at a given time is also assigned a randomly changing position in space. There are many examples of such systems in biology: consider a single celled bacteria which has the ability to move using its flagellum and which reproduces through binary fission. If each individual bac- terium is not affected by the location of the other bacteria, this could be considered a non-interacting BPS. A second example is a larger species called phytoplankton which live in the ocean. Each plankton moves on some random path determined by its internal motor and reproduces (branches) randomly. The plankton spatial posi- tion is affected both by individual interactions (e.g. reproductive) with each other
1 as well as by a common drift due to the water current and therefore this system is an example of an interacting BPS. It can be computationally difficult to simulate a BPS when the population is large since this involves simulating the random paths for each of the particles. Fortunately, if for a BPS the number of particles is large and we observe the sys- tem over a large period of time, then it can be approximated by a measure-valued stochastic process. Equivalently, the behavior of a large population, rapidly branch- ing, BPS can be inferred from the properties of the corresponding measure-valued process. A mathematically rigorous result of this type was established indepen- dently by Watanabe [25] and later by Dawson [1] for non-interacting BPS, where the associated measure-valued process is a superprocess. The situation for inter- acting BPS is much more difficult to ascertain since different interactions require different methods of examination. Hence the results in this direction tend to be proven on a case-by-case basis. Chapters 2 and 3 study branching particle systems with two different types of interactions. We construct the scaling limits of these systems explicitly using something that is known as “Perkins historical stochastic calculus,” which is a the- ory pioneered in the early nineties by Edwin Perkins [13] to study many types of interacting measure-valued processes.
1.1 Branching particle systems and superprocesses Before mathematically describing an interacting particle system, we define a non- interacting branching particle system. Suppose the spatial motion of a typical par- ticle follows the stochastic process {Yt : t ≥ 0}, which takes values in a space E d (usually E is R ). Also suppose that when the particle dies it branches into k par- ticles with probability ν(k). Let Λ be the set of all particles in the particle system. We assume that a particle, labeled by α, born at a (possibly random) time τ(α), lives until time d(α) whereupon it dies after giving birth to Mα offspring, where {Mα : α ∈ Λ} are independent and identically distributed random variables each with law ν. Typically, the lifetime of a particle, `(α) ≡ d(α) − τ(α) is taken to be an independent exponential random variable, or to be constant.
2 We give ourselves the family of stochastic processes
{Y α (t),τ(α) ≤ t < d(α) : α ∈ Λ} which will describe the paths of the particles. In particular, for particle α we will 0 insist that Y α (τ(α)) = Y α (d(α0)−), where α0 is the parent of α. This simply says that where the parent α0 died is where the particle α was born. We will also assume that the particles alive at a given time t move independently of each other. Then one can define a process Z which counts the number of particles in a region of space at a given time: For A a subset of E, define for t > 0,
Zt (A) = ∑ δY α (t)∈A, α∼t where α ∼ t denotes that particle α is alive at time t. Z0 is defined by the particles alive at time 0, which we will take to be given. Assume in the above that Y is a nice process (say a continuous Markov pro- cess). Suppose the branching law νN has mean 1+β/N and variance γ > 0, where β and γ are fixed numbers. Assume that each labeled particle α is born at an in- tegral multiple of N−1 and lives for a time `(α) = N−1. Suppose also that we are given a finite measure m(·) on E. Then for A ⊂ E and t > 0, let
N 1 X (A) = α , (1.1) t ∑ δYt ∈A N α∼t
N where X0 (·) is given by a Poisson point process with intensity Nm(·). This gives rise to the aforementioned result: Proposition 1.1.1 (Watanabe, 1968; Dawson, 1975). There is a measure-valued, N Markov process X such that Xt −→ Xt , for all t, in distribution with X0 = m. The process X is called a (Y,β,γ)-Dawson-Watanabe superprocess. This prop- osition says that to estimate the probability of most types of events for a large population BPS, one can simply calculate the probability of such events for the corresponding superprocess (which is typically easier). The value of β determines whether the process is subcritical or supercritical. If β > 0 then it is supercritical and there is a positive probability that the process
3 will exhibit exponential mass growth and survive indefinitely. If β ≤ 0, the process will die out in finite time almost surely. The case β = 0 is called critical and β < 0 is subcritical. One immediate property of X is that it is not an atomic measure, as N the X were. Indeed, if Y is one dimensional Brownian motion, Xt is almost surely absolutely continuous with Lebesgue measure. In higher dimensions, the support of Xt tends to be fractal-like. For these results and much more, Perkins [14] is a wonderful reference.
The superprocess Xt is deficient in the sense that it only keeps track of the mass alive at time t and ignores any other genealogical information present in the associated branching particle system. To remedy this, we can keep track of not just the position of a particle at time t, but instead the entire path of the particle and those of its ancestors until time t: for B ⊂ C([0,∞),E) measurable, define
N 1 H (B) = α . t ∑ δY·∧t ∈B N α∼t
α Here, for t < τ(α), we let Yt be the location of the unique ancestor of α alive at time t. Note: we say y·∧t is the path y until time t and fixed thereafter. Then one can adapt Proposition 1.1.1 for a special case to show that
N Ht −→ Ht for all t, in distribution. We call H the (Y,β,γ)-historical process. It is then easy to project
Ht down at time t to recover the process Xt (where X is the (Y,β,γ)-superprocess): For A ⊂ E, Z Xt (A) = 1(yt ∈ A)Ht (dy).
One of the special features of superprocesses that results from the indepen- dence of particles alive at a given time is a type of log-laplace equation. This is an equation that can be used to infer properties of the superprocess (and therefore the associated BPS) through the solution of a particular non-linear partial differential equation. More precisely, for certain positive functions φ : E → R, one can show that −R φdX −R V φdX E e t = e t 0 (1.2)
4 where Vs ≡ Vsφ is the solution to
∂V γ s = AV + βV − V 2, (1.3) ∂s s s 2 s with initial condition V0 = φ, where A is a second order differential operator called the infinitesimal generator of the process Y. This connection allows one to calculate probabilities and infer properties of X from the behavior of V, and conversely use properties of X to establish results for the non-linear PDE (1.3) (see for example, Le Gall [11]).
1.2 Superprocesses in a random medium In the first chapter below, which was published in [8], we construct a family of interacting measure-valued diffusions that live in a random medium by solving a stochastic equation driven by historical Brownian motion. This work is intended to provide oceanographers with a probabilistic model for phytoplankton populations during an algal bloom (colloquially known as a “red tide”). During a bloom, the populations of plankton in a region increase rapidly to the point where one can actually observe them in the water by its reddish tinge. This can have negative consequences higher in the food chain as some of these plankton, such as dinoflag- ellates, release neurotoxins in the water which can then become concentrated in filter-feeding marine animals. When these animals are consumed by humans it can cause possibly serious illnesses like paralytic shellfish poisoning. As such, it is vital to predict the likelihood of algal blooms taking place at a given location. Plankton are by definition ocean drifters and range in size from less than a micrometer in width to possibly bigger than a centimeter. The position of a typical organism is affected by the current of the ocean on a large scale. On a smaller scale the position of a plankton is determined by its internal motor and by eddies and ripples of water. As the size of the plankton is typically very tiny, we may think of it as Brown originally thought of the motion of a pollen grain in a glass of water - that it is constantly being bombarded by millions of water molecules which affect its position in a random manner. Thus the motion of a small plankton may be modeled by a diffusion.
5 The first probabilistic work that could be used to model large plankton popu- lations was found in the Ph.D. dissertation of Wang, published in [23] and [24], though it was not motivated directly by such applications. Dawson, Li and Wang in [2] carry this work further to include a variable branching mechanism. The con- nection to plankton dynamics was made in the work of Skoulakis and Adler [19]. Young in his survey paper [27] for the oceanographic community, discusses similar models under the guise ‘Brownian bugs.’ Young et. al. in [28] provide more support for the models such as the one above by showing that the observed ‘patchy’ dis- tribution of organisms in the water can be explained by the effect of the branching (reproduction) rather than other physical or chemical cues. The technique used in each of [23], [24], [2] and [19] consists of defining a sequence of interacting branching particle systems and taking a high density limit, α as in Proposition 1.1.1. Under this method the particle α follows a path Yt on its life time, where Y α solves the following stochastic differential equation:
Z t Z t Z t Z α α α α α α Yt = Yτ + σ1(Ys )dBs + b(Ys )ds + σ2(Ys ,ξ)dW(s,ξ) (1.4) τ τ τ where σi and b are coefficients satisfying some reasonable conditions and τ ≡ α τ(α). Here the {B }α∈Λ are an independent family of Brownian motions and W is a space-time white noise term independent of that family. The white noise term is shared across all the particles and therefore acts as a random medium which affects all particles simultaneously (i.e. as the ocean does for plankton) . This also means that the spatial motion of the different particles will depend on each other since they all have a common component driven by the same white noise. The white noise W is an object arrived at by taking a scaling limit of a space- time lattice of random variables. We divide up space-time in to cubes of side- length N−1 and to each of these, we assign an appropriately scaled independent N random (spatial) direction. Then for a set A in Euclidean space, Wt (A) denotes the sum of all the random directions of all the space-time cubes located inside the set [0,t] × A. We then send N to infinity to get the random direction Wt (A). Hence one can think of white noise as a random field generated by the movement of a very large number of infinitesimally small particles moving about according to Brownian motion (such as what happens in a liquid).
6 Skoulakis and Adler in [19] assume the white noise W in (1.4) is really a Brow- nian motion (which is a very specific white noise) whereas Wang and his coauthors all assume that the drift term b = 0. Also note that even in the full setting of (1.4), the only interactions between particles are as a result of the common drift due to the random medium. That is, one would expect that for a plankton population there should be some interactions of the plankton with each other, which are not present here. More is known for the setting where W is a Brownian motion in (1.4) than in the generalized case since it is easier to work with a Brownian motion in place of a white noise. In fact Xiong in [26], by utilizing random duals, shows the existence of a log-laplace equation similar to that of (1.2) where the Vt φ in that equation is replaced by a solution of a non-linear stochastic partial differential equation similar to Equation 1.3. This allows one to perform calculations for the limiting measure-valued process in a manner similar to that used for superprocesses, which is remarkable since as mentioned earlier, superprocesses arise as limits of non- interacting BPS’s, whereas the BPS presented here has interactions. The approach employed in Chapter 2 takes an existing historical process K and an independent white noise W and uses them to construct another measure- valued diffusion much in the same way one takes Brownian motion and constructs diffusions by solving stochastic differential equations. In particular, we solve the following stochastic equation: Let K be a critical historical Brownian motion (the process constructed above with Y a Brownian motion and β = 0), and A measur- able. Then define (SE)
Z t Z t Z t Z (a) Zt = Z0 + σ1(Xs,Zs)dy(s) + b(Xs,Zs)ds + σ2(Xs,Zs,ξ)dW(s,ξ) 0 0 0 Rm Z (b) Xt (A) = 1(Zt (y) ∈ A)Kt (dy)
One can think of (SE) (a) as a stochastic differential equation describing the mo- tion of a given particle driven by a certain path y and a white noise, and as the natural analogue to (1.4). One major difference is that (SE)(a) includes interac- tions between individual particles (for example reproductive interactions between plankton) whereas the models considering (1.4) do not. (SE)(b) gives a measure-
7 valued process by integrating over all paths y with respect to K. Since Kt typically puts mass on those paths y that resemble Brownian sample-paths, (SE)(a) is akin to an SDE where y is replaced by a Brownian motion. Note that the white noise W does not depend on y. The main theorem in this work shows that there exists a unique (strong) solu- tion to (SE):
Theorem 1.2.1. Under some Lipschitz conditions for the coefficients σi and b, there exists a pathwise unique solution (X,Z) to (SE).
With this theorem in hand, additional properties of X such as the compact sup- port property (which shows that a particle cannot travel too far in finite time), strong Markov property and stability of the solutions are established. One can also show without too much difficulty that a sequence of interacting branching particle sys- tems determined by the formula (1.1), with spatial position for particle α given by (1.4), converge in distribution to the solution X of (SE). In the case where there is no spatial dependence of the coefficients, we recover the results of Wang, Skoulakis and Adler.
1.3 An interacting super Ornstein-Uhlenbeck process The second chapter of the thesis (forming the basis for the article [9]) is devoted to the construction and examination of a particular class of interacting measure- valued processes with a singular drift called the super Ornstein-Uhlenbeck (SOU) process interacting with its center of mass (COM). What this means is that the BPS associated with this measure-valued process has particles whose spatial motion is described by a modified Ornstein-Uhlenbeck process. The Ornstein-Uhlenbeck process is one of the simplest stochastic processes – originally introduced in 1930 by Ornstein and Uhlenbeck in [21] to model the velocity V of a particle in a medium with a certain friction coefficient. The process satisfies
dV = −γVds + dBs, where B is a Brownian motion and γ > 0. The thought was that since Brownian motion is not differentiable, one cannot talk about the velocity of a Brownian parti-
8 cle and hence it is better to model the velocity by a diffusion process directly. Two of the key properties are that this process is mean-reverting and has a stationary distribution that is a Gaussian random variable. The interacting particle system corresponding to this process is made up of particles that move in a Brownian fashion, but are attracted to or repelled from the COM of the system. That is, for a particle α in the corresponding particle system, its motion during its lifetime is described by
Z t Z t α α ¯ α α Yt = Yτ(α) + γ (Y(s) −Ys )ds + dBi (s), (1.5) τ(α) τ(α) where ¯ 1 α Yt = ∑ Yt #{α ∼ t} α∼t is the center of mass and {Bα : α ∈ Λ} is an independent family of Brownian mo- tions. This constructs a branching modified-OU system Z as follows:
Z (A) = α . (1.6) t ∑ δYt ∈A α∼t
The particles in the system Z exhibit attraction to its COM when γ > 0 and repul- sion when γ < 0. The motivation for the model came from a paper of Englander¨ [3] where he constructs a branching particle system Z0 where each particle moves according to (1.5). The system starts with a single particle, and all particles live for one unit of time and die at the integral times leaving behind exactly two offspring particles. It is clear that the particle system will have exactly 2btc particles alive at time t. Englander¨ then proves various theorems about the long term behavior of the parti- cle system and shows that it achieves a natural equilibrium. He proceeds by first showing that the center of mass Y¯t converges as t approaches infinity and then that d for φ : R → R nice, Z Z −n 0 Y¯∞ 2 φ(x)Zn(dx) −→ φ(x)P (dx), (1.7)
¯ as n goes to infinity, almost surely, where PY∞ is a Gaussian random variable cen-
9 tered at the limiting value of the COM, Y¯∞. This particle system is particularly simple in that one does not have to worry about the population of particles going extinct at a given time and the singularity that results in the definition of the center of mass, as is the case for the more general
BPS, Zt . Nevertheless, one can make an analogy between this system and a super- critical BPS on its survival set, as in both situations the systems exhibit exponential population growth. Therefore one expects that on the survival set a general super- critical BPS should exhibit similar long-term behavior as in this binary branching case. Given this insight, it is not surprising that if one can construct the scaling limit of such a sequence of BPS’s then it should exhibit some form of equilibrium behavior. To make this mathematically sound we define (similar to (SE) above) the fol- lowing stochastic equation: Let K be a supercritical historical Brownian motion, (i.e. β > 0 in the definition). Then define the process (X,Y) by the solution of the following equation:
∗ (SE) (a) dYs = γ(Y¯s −Ys)ds + dys Z (b) Xt (A) = 1(Yt (y) ∈ A)Kt (dy)
d where A ⊂ R and the COM is defined as R xX (dx) ¯ s Ys = R . 1Xs(dx)
We proceed to show that Y¯t converges almost surely and use this fact to prove the main theorem:
x Theorem 1.3.1. If P∞ is the stationary distribution of an Ornstein-Uhlenbeck pro- cess with attraction to x at rate γ > 0, then on the survival set
Xt (·) a.s. ¯ −→ PY∞ (·), Xt (1) as t → ∞.
The convergence obtained here is in fact even stronger than the one Englander¨
10 showed for (1.7) since here we have convergence in measure. The methods used to get this result differ considerably from those of Englander¨ due to the generalized setting being considered. The novelty of this work is that it is one of few cases where we can explic- itly determine the long-term equilibrium behavior of an interacting measure-valued process. Unlike previous applications of the historical calculus, we must use it here to prove uniqueness in a non-Lipschitz setting.
11 Chapter 2
Superprocesses with spatial interactions in a random media
2.1 Introduction We will consider the following stochastic equation:
Z t (SE)(a) Zt = Z0 + σ1(Xs,Zs)dy(s) 0 Z t Z Z t + σ2(Xs,Zs,ξ)dW(s,ξ) + b(Xs,Zs)ds 0 Rm 0 Z d (b) Xt (A) = 1(Zt (y) ∈ A)Kt (dy) ∀A ∈ B(R ).
d d d×d d d The coefficients are such that σ1 : MF (R ) × R → R , σ2 : MF (R ) × R × m d×d d d d d R → R and b : MF (R ) × R → R , where MF (R ) denotes the space of all d non-negative finite measures on R . K is a historical super-Brownian motion on d m some space, y is a path in R , and W is a d-dimensional white noise (on R+ × R ) that is independent of K. The precise definition of a historical process can be found in [14] and the exact meaning of a stochastic integral with respect to y is given below in Section 2.2. Later we impose some Lipschitz-like conditions on b,σ1 and σ2 to use a Picard iteration argument. Note that Z takes values in the d d space of paths in R , while X takes values in the space of paths in MF (R ).
12 One can think of (SE) (a) as a stochastic differential equation describing the motion of a given particle driven by a certain path y and a white noise, and (SE)(b) as the measure-valued process obtained by integrating over all paths y with respect to a historical super-Brownian motion. As mentioned earlier, since Kt typically puts mass on those paths y that resemble Brownian sample-paths, (SE)(a) is like a stochastic differential equation where y is replaced by a Brownian motion. Note that the white noise W is independent of y. Although Kt is not normally atomic, if it were, Kt would keep track of the number of particles alive at time t, as well as the history of each particle’s driving Brownian motions until time t. Dawson et al [2] use a high density limit of a sequence of interacting particle systems, where each of the particles move according to an SDE similar to (SE) (a), to construct a measure-valued diffusion. Earlier, Wang in [24] and [23] constructed special cases of the model of [2]. Using a similar method, the special case where the white noise W is replaced by a Brownian motion B (i.e. setting m = 0 here) was studied initially by Skoulakis and Adler [19] in relation to plankton dynamics. Their model also did not incorporate the spatial interactions between particles that are built into (SE). Before we discuss solutions for (SE), we will need some background material, as well as a framework in which to work. This is provided in Section 2.2. In Section 2.3, we will set up (SE) in a more rigorous manner and then prove the existence and pathwise uniqueness of solutions for a broad range of coefficients σ1,σ2, and b. In Section 2.4, we show that the solutions of (SE) satisfy a martingale problem– an extension of the martingale problems that appears in [2] and [13]. We conclude with a section about various additional properties of the solutions, including the strong Markov and compact support properties.
2.2 Preliminaries For a good reference for much of what follows in this section, consult Chapters II and V of [14].
Let K be an (Ft )-adapted, critical, historical super-Brownian motion on the space (Ω,F,(Ft )t≥0,P). d t t Let C = C(R+,R ) and C be its Borel σ-field. Also define C = {y : y ∈ C}
13 s t and Ct = σ(y ,s ≤ t,y ∈ C) where y = y·∧t . Define MF (C) to be the space of finite 0 0 0 measures on C. For metric spaces E and E , take Cc(E,E ) and Cb(E,E ) to be the space of compactly supported continuous functions from E to E0 and the space of bounded continuous functions from E to E0, respectively.
Notation. For a measure µ on a space E and a measurable function f : E → R, let µ( f ) = R f dµ.
Note that Kt takes values in MF (C), and so Kt (·) will typically mean integration over the y variable below.
We define a martingale problem for K. Let Bˆt = B(· ∧ t) be the path-valued process associated with B, taking values in Ct . Then for φ ∈ bC (bounded and C s,y measurable), if s ≤ t let Ps,t φ(y) = E (φ(Bˆt )), where the right hand side denotes expectation at time t given that until time s, Bˆ follows the path y. We can now introduce the weak generator, Aˆ, of Bˆ. If φ : R+ ×C → R we say φ ∈ D(Aˆ) if and only if φ is bounded, continuous and (Ct )-predictable, and for some Aˆsφ(y) with the same properties as φ,
Z t φ(t,Bˆ) − φ(s,Bˆ) − Aˆrφ(Bˆ)dr,t ≥ s, s
s is a (Ct )-martingale under Ps,y for all s ≥ 0,y ∈ C . d Then if m ∈ MF (R ), we say K satisfies the historical martingale problem, (HMP)m, if and only if K0 = m a.s. and
Z t ∀φ ∈ D(Aˆ), Mt (φ) ≡ Kt (φt ) − K0(φ0) − Ks(Aˆsφ)ds 0 Z t 2 is a continuous (Ft )-martingale with hM(φ)it = Ks(φs )ds ∀t ≥ 0, a.s. 0
The definition of Mt (·) can be extended to form an orthogonal martingale mea- sure using the method of Walsh [22]. Denote by P, the σ-field of (Ft )-predictable sets in R+ × Ω. If ψ : R+ × Ω ×C → R is P × C-measurable and
Z t 2 Ks(ψs )ds < ∞ ∀ t ≥ 0, (2.1) 0 then there exists a continuous local martingale Mt (ψ) with quadratic variation
14 R t 2 given by hM(ψ)it = 0 Ks(ψs )ds. If the expectation of the term in (2.1) is finite, 2 then Mt (ψ) is an L martingale. In this paper, we will require m(·) ≡ P(K0(·)) to be a finite, positive measure d on R . Recall that for a measure space (E,E), the universal completion of E, denoted E∗, is given by ∗ E = ∩µ E¯ µ , where E¯ µ denotes the completion of E with respect to the measure µ and the inter- section is over all probability measures µ on E.
ˆ ˆ ˆ ˆ ∗ Definition. Let (Ω,F,Ft ) = (Ω ×C,F × C,Ft × Ct ). Let Ft denote the universal completion of Fˆ t . If T is a bounded (Ft )-stopping time (denote the set of all such stopping times by Tb), the normalized Campbell measure associated with T is the ˆ ˆ ˆ measure PT on (Ω,F) given by
ˆ −1 PT (A × B) = P(1AKT (B))m(1) for A ∈ F,B ∈ C.
ˆ ˆ We denote sample points in Ω by (ω,y). Hence, under PT , ω has law KT (1)dP and conditionally on ω, y has law KT (·)/KT (1).
ˆ Note that here PT is a probability measure, since
ˆ −1 −1 PT (1) = P(KT (1))m(1) = P(K0(1))m(1) = 1 as Kt (1) is Feller’s critical branching diffusion (a uniformly integrable martingale). To avoid carrying constants of m(1)−1 from line to line, we will without loss of generality assume m(1) = 1. The following results will be useful later. The first is Proposition V.2.4 and the second, Proposition V.3.1 of [14].
Proposition 2.2.1. Assume that T ∈ Tb and ψ ∈ Fˆ T , bounded. Then
Z t Z Kt (ψ) = KT (ψ) + ψ(y)dM(s,y) ∀t ≥ T P-a.s. T
15 ˆ ˆ Proposition 2.2.2. If T ∈ Tb then under PT , y is an (Ft )-adapted Brownian motion stopped at T.
m Suppose that {W(t,ξ),ξ ∈ R ,t ≥ 0} is an (Ft )-adapted, d-dimensional Brow- nian sheet on (Ω,F,(Ft )t≥0,P) that is independent of K. For a function φ ∈ m n×d R t R C ( , ), let W ( ) = m ( )dW(s, ). This defines the associated white c R R t φ 0 R φ ξ ξ noise process of W, and hence we will identify the two in what follows. The next ˆ ˆ theorem will be useful in proving that W is also an (Ft )-adapted PT Brownian sheet for each T ∈ Tb.
Theorem 2.2.3. Suppose N is a real-valued, square-integrable, (Ft )-adapted mar- tingale that is independent of K. Then for each T ∈ Tb, N is an (Fˆ t )-adapted ˆ j ˆ martingale under PT and hN,y it = 0 ∀t ≥ 0 PT -a.s ∀ j ≤ d.
Proof. Let A ∈ Cs and B ∈ Fs. Then
ˆ PT [(Nt − Ns)1A1B]
= P[(Nt − Ns)KT (A)1B]
= P[1B1T≤sKT (A)E(Nt − Ns|Fs)] + P[(Nt − Ns)1B1T>sKT (A)] Z T Z = P (Nt − Ns)1B1T>s Ks(A) + 1AdM(r,y) s Z T∧t Z = P (Nt − Ns)1B1T>s Ks(A) + 1AdM(r,y) s where in the first equality we use the definition of the Campbell measure and in the second that KT (A)1{T≤s} ∈ Fs. For the third equality, we use the fact that Nt is an (Ft )-martingale and Proposition 2.2.1. Finally we condition on Ft and appeal to the Optional Sampling Theorem for the last expression. By conditioning first on
Fs and then on FT we can reduce the above as
Z T∧t Z ˆ PT [(Nt − Ns)1A1B] = P (Nt − Ns)1B0 1AdM(r,y) s Z T∧t Z = P (NT∧t − Ns)1B0 1AdM(r,y) s
0 where B = B ∩ {T > s} ∈ Fs. Now, if we show that Nt Mt (ψ) is an (Ft )-adapted martingale for any bounded function ψ : R+ × C → R, then we may apply the
16 Optional Stopping Theorem to show that the above expression reduces to zero. Note that,
Z t Z t Nt Mt (ψ) = NsdMs(ψ) + Ms(ψ)dNs + hN,M(ψ)it 0 0 Z t Z t = NsdMs(ψ) + Ms(ψ)dNs 0 0 since hN,M(ψ)it = 0 by independence of N and M. It also follows from the inde- pendence that the first term on the right side is a martingale. For the second term we use independence and the square integrability of N to conclude that it is a mar- ˆ tingale as well. Hence Nt Mt (ψ) is a martingale and PT [(Nt − Ns)1A1B] = 0. This proves the first assertion. From the independence of N and K we have for constant T, a Borel subset A of C(R+,R) and B ∈ C that,
ˆ PT (N ∈ A,y ∈ B) = P(1N∈AKT (1y∈B))
= P(1N∈A)P(KT (1y∈B)) ˆ ˆ = PT (N ∈ A)PT (y ∈ B).
ˆ j ˆ Hence, N and y are independent under PT . This implies that hN,y i = 0,PT -a.s. ˆ In the general case where T ∈ Tb, independence of N and y under PT no longer j ˆ necessarily holds. Let Z ≡ hN,y i. Then under PT , for u > T fixed, ˆ PT sup|Zs| ∧ 1 = P KT sup|Zs| ∧ 1 s≤T s≤T Z u Z = P Ku sup|Zs| ∧ 1 − sup|Zs| ∧ 1dM(r,y) s≤T T s≤T = 0
ˆ In the second line, we have noted that sups≤T |Zs| ∧ 1 is bounded, FT -measurable ˆ and applied Proposition 2.2.1. To get the third line, we use the fact that Z = 0, Pu- a.s. by the special case considered above and that the second term is a martingale. j ˆ Hence hN,y it = 0 ∀t ≥ 0 PT -a.s. ∀T ∈ Tb (note that we have used the property ˆ T that under PT , y = y a.s. implicitly here).
17 m m Let R denote the Borel σ-algebra of R . The following proposition is needed to complete the proof.
m m Proposition 2.2.4. Let M be an orthogonal martingale measure on (R ,R ) as defined in Chapter 2 of Walsh [22]. Suppose further that for each A ∈ R, t 7→ Mt (A) is continuous. Then M is a white noise if and only if its covariance measure is given by Lebesgue measure.
For the proof, see Proposition 2.10 of Walsh [22].
m ˆ Corollary 2.2.5. For each T ∈ Tb, {W(t,ξ),ξ ∈ R ,t ≥ 0} is an (Ft )-adapted ˆ ˆ Brownian sheet under PT . Furthermore, W and y are orthogonal under PT ; that j m d ˆ is, hW(φ),y it = 0 ∀φ ∈ Cc(R ,R ),t ≥ 0 PT -a.s. ∀T ∈ Tb for all j ≤ d.
1 d ˆ Proof. Let Wt (A) = (Wt (A),...,Wt (A)). We simply need to verify that under PT , i i {Wt } are independent white noises. Each Wt is an orthogonal martingale measure ˆ m i i i under PT since for disjoint bounded sets A,B ∈ R , Wt (A) and Wt (A)Wt (B) are ˆ square-integrable (Ft )-martingales with respect to P, hence also (Ft )-martingales ˆ i with respect to PT by Theorem 2.2.3. The map t 7→ Wt (A) is continuous since i ˆ i Wt (A)(ω,y) defined on Ω is equal to Wt (A)(ω) on Ω, for which the analogous ˆ map is continuous and PT |Ω is absolutely continuous with respect to P. i 2 ˆ By noting that Theorem 2.2.3 implies Wt (A) − tν(A) is an (Ft )-martingale, where ν is the m-dimensional Lebesgue measure, we see that the covariance mea- i sure is deterministic. Hence each Wt is a white noise process by Proposition 2.2.4. i j Then by a final use of Theorem 2.2.3, we see that Wt (A)Wt (A) − tν(A)δi, j is an i j (Fˆ t )-martingale where δi, j is Kronecker’s δ. Therefore W (A) and W (A) are inde- pendent Brownian motions (for any A with ν(A) < ∞). Hence Wt is a d-dimensional white noise process. The orthogonality of y,W follows from the previous theorem.
Note that this corollary will still hold if W is a white noise on a space E × R+ with covariance measure ν × ` where ν is an arbitrary σ-finite measure on E and ` is the one dimensional Lebesgue measure. Indeed the results below all extend to this generality as well.
18 n×d 2 2 n For a matrix A ∈ R , let ||A|| = ∑i, j Ai j. Note for a vector x ∈ R , we will use |x|2 instead to denote the same quantity.
ˆ n×d ˆ ∗ Definition. We say f ∈ D1(n,d) iff f : R+ × Ω → R is (Ft )-predictable and
Z t 2 k f (s,ω,y)k ds < ∞ Kt − a.a. y ∀t ≥ 0,P a.s. 0
m ˆ n×d ˆ ∗ We say g ∈ D2(m,n,d) iff g : R+ × R × Ω → R is (Ft )-predictable and
Z t Z 2 kg(s,ξ,ω,y)k dξds < ∞ Kt − a.a. y ∀t ≥ 0,P a.s. 0 Rm ˆ Definition. If X,Y : R+ × Ω → E, we say X = YK-a.e. iff X(s,ω,y) = Y(s,ω,y) for all s ≤ t, Kt -a.e. for all t ≥ 0 P-a.s. If E is a metric space we say that X is continuous K-a.e. iff s 7→ X(s,ω,y) is continuous on [0,t] for Kt -a.a. y for all t ≥ 0 P-a.s. ˆ ˆ Since with respect to each measure PT ,T ∈ Tb, {W(s,ξ),0 ≤ s} is an (Fs)- adapted Brownian sheet, we may define the stochastic integral of a function f with R t R 2 d k f ( ,s, ,y)k d ds < , ˆ -a.s. as in Walsh [22]. However as this integral 0 R ξ ω ξ ∞ PT ˆ R t R depends on T, we denote it by PT - 0 f (s,ξ,ω,y)dW(s,ξ). ˆ R t Similarly, define PT - 0 g(s,ω,y)dy(s) to be the stochastic integral g with re- ˆ spect to y under PT . The following part of Proposition V.3.2 of [14] is useful in relating these different stochastic integrals.
Proposition 2.2.6. For g ∈ D1(n,d), there exists an (Fˆ t )-predictable, K-a.e. con- tinuous process I(g,t,ω,y) such that
Z t ˆ ˆ I(g,t ∧ T,ω,y) = PT - g(s,ω,y)dy(s) for all t ≥ 0,PT -a.s. for all T ∈ Tb. 0
0 ˆ ∗ Moreover, I is unique. That is, if I is an (Ft )-predictable process satisfying the above, then I(g,s,ω,y) = I0(g,s,ω,y) K-a.e.
ˆ R t R We can prove a similar theorem for PT - 0 f (s,ξ,ω,y)dW(s,ξ), but we will need to recognize that it only holds for t ≤ T as the following example demon- strates.
19 u 2 m Example 2.2.7. Consider f (s,ξ,ω,y) = y (s)η(ξ)1(u,v](s) with η ∈Cc (R ). Note ˆ T that under PT , y = y a.s., and hence
Z t Z ˆ T PT - f (s,ξ,ω,y)dW(s,ξ) = y (u)(Wv∧t (η) −Wu∧t (η)). 0
This expression clearly depends on the stopping time T. If T = T1 < u is fixed then our integral is y(T1)(Wv∧t (η) −Wu∧t (η)), else if T = T2 > u is fixed, then the integral is y(u)(Wv∧t (η) −Wu∧t (η)). If t < T1, then the integrals in both cases agree (and are equal to zero).
This shows that even simple functions do not necessarily have unique integrals ˆ when integrated with respect to W under the different measures PT when t > T. The next theorem is the analogue of Proposition 2.2.6 and shows how the stochastic ˆ integrals of W are related under different measures PT .
n ˆ Theorem 2.2.8. (a) If f ∈ D2(m,n,d), there is an R -valued (Ft )-predictable pro- cess J( f ,t,ω,y) such that for all T ∈ Tb
Z t∧T Z ˆ ˆ J( f ,t ∧ T,ω,y) = PT - f (s,ξ,ω,y)dW(s,ξ) ∀t ≥ 0,PT -a.s. 0 Rm
0 ˆ ∗ 0 (b) If J ( f ) is an (Ft )-predictable process satisfying (a), then J( f ,s) = J ( f ,s) K-a.e.
(c) J( f ,t) is continuous in t, K-a.e.
(d) (Dominated Convergence) For any N > 0, if fk, f ∈ D2(m,n,d) satisfy
Z N Z 2 lim P KN k fk(s,ξ) − f (s,ξ)k dξds > ε = 0,∀ε > 0, k→∞ 0 Rm
then, 2 lim P supKt supkJ( fk,s) − J( f ,s)k > ε = 0 ∀ε > 0. k→∞ t≤N s≤t
20 (e) For any S ∈ Tb if fk, f ∈ D2(m,n,d) satisfy
Z S Z 2 lim P KS k fk(s,ξ) − f (s,ξ)k dξds = 0, k→∞ 0 Rm
then, 2 P supKt supkJ( fk,s) − J( f ,s)k → 0 as k → ∞. t≤S s≤t The proof of Proposition 2.2.6 can be easily adapted to prove this theorem. We R t R will at times write 0 f (s,ξ,ω,y)dW(s,ξ) or Wt ( f ) in place of J( f ,t,ω,y) where the latter is well defined.
Corollary 2.2.9. Let T ∈ Tb. If f ∈ D1(n,d), g ∈ D2(m,n,d) and there exists an ˆ ∗ (Ft )-predictable process S( f ,g) satisfying
Z t∧T Z t∧T Z ˆ ˆ S( f ,g,t ∧ T,ω,y) = PT - f (s,y)dy(s) + PT - g(s,ξ)dW(s,ξ) (2.2) 0 0
ˆ ∀t ≥ 0,PT -a.s., then S( f ,g) = I( f ) + J(g),K-a.e.
Proof. Define Z L(t,ω) = supkI( f ,s) + J( f ,s) − S( f ,g,s)k ∧ 1Kt (dy). (2.3) s≤t
Note that L is (Ft )-predictable. The proof of this is essentially given in the last half of the proof for Proposition V.3.2(b) of [14]. Assume T is a bounded, predictable stopping time. Then ˆ P(L(T,ω)) = PT supkI( f ,s) + J( f ,s) − S( f ,g,s)k ∧ 1 s≤T = 0
ˆ R s∧T ˆ R s∧T R since S( f ,g,T ∧ s) = PT - 0 f (s,y)dy(s) + PT - 0 g(s,ξ)dW(s,ξ), ˆ R s∧T I( f ,s ∧ T) = PT - 0 f (s,ω,y)dy(s) by Proposition 2.2.6 and since J(g,s ∧ T) = ˆ R s∧T R PT - 0 g(s,ξ,ω,y)dW(s,ξ) by Theorem 2.2.8(a) above. Then by the Section Theorem, we have that L(t,ω) = 0 ∀t ≥ 0 a.s.
21 n ˆ ˆ Notation. If X(t) = (X1(t),...,Xn(t)) is an R -valued process on (Ω,F) and µ ∈ R MF (C), let µ(Xt ) = (µ(X1(t)),..., µ(Xn(t))) where µ(Xi(t)) = Xi(t,ω,y)µ(dy). Also let
Z t Z Z t Z Z t Z X(s)dM(s,y) = X1(s)dM(s,y),..., Xn(s)dM(s,y) 0 0 0 whenever these integrals are defined. We do the same for stochastic integrals with respect to W.
The next theorem is needed in order to prove a version of Ito’sˆ Lemma.
R 2 Theorem 2.2.10. If f ∈ D2(m,n,d) and sups,ω,y k f (s,ξ,ω,y)k dξ < ∞ then
Kt (J( f ,t)) Z t Z Z t Z = Ws( f (s))dM(s,y) + Ks( f (s))dW(s,ξ) ∀t ≥ 0 P-a.s. (2.4) 0 0
2 and each integral on the right is a continuous L (Ft )-martingale.
Proof. To simplify the notation, assume that n = d = 1. We first show the result for simple functions. Let
f (s,ξ,ω,y) = φ1(ω)φ2(y)φ3(ξ)1(u,v](s) (2.5)
∞ m where φ1 ∈ bFu,φ2 ∈ bCu,φ3 ∈ Cc (R ) and 0 ≤ u < v. Then,
Z t Z Kt f (s,ξ,ω,y)dW(s,ξ) = Kt (φ1(ω)φ2(y)(Wv∧t (φ3) −Wu∧t (φ3))) 0
= φ1(ω)(Wv∧t (φ3) −Wu∧t (φ3))Kt (φ2) Z t Z = φ1(ω)φ3(ξ)1(u,v](s)dW(s,ξ)Kt (φ2) 0 where in the second line we make use of the fact that Wt and φ1 depend only on ω. We can now apply the integration by parts formula, while noting that
22 hW(φ3),K(φ2)it = 0 (by independence), to get
Z t Z Z t Z Kt f (s,ξ)dW(s,ξ) = φ1(ω)φ3(ξ)1(u,v](s)Ks(φ2)dW(s,ξ) 0 0 Z t Z s Z + φ1(ω)φ3(ξ)1(u,v](r)dW(r,ξ)dKs(φ2) u u Z t Z = Ks( f (s,ξ))dW(s,ξ) 0 Z t Z s Z Z + φ1(ω)φ3(ξ)1(u,v](r)dW(r,ξ) φ2(y)dM(s,y) 0 0 Z t Z Z t Z Z s Z = Ks( f (s))dW(s,ξ) + f (r,ξ,ω,y)dW(r,ξ) dM(s,y). 0 0 0
Here in the second line we have used Proposition 2.2.1 with T = u and then the fact that the function φ1(ω)φ3(ξ)1(u,v](s) disappears for s < u. The differential R φ2(y)dM(s,y) is just dMs(φ2). Note that for these simple functions f , we need only look at the square function R t R 2 of 0 Ks( f (s,ξ))dW(s,ξ) to show that it is an L martingale. That is, since φ1φ2 ∞ m is bounded and φ3 ∈ Cc (R ),
Z t Z Z t Z 2 2 2 P Ks( f (s,ξ)) dξds ≤ cP φ3(ξ) Ks(1) dξds . 0 0
This expectation is finite because Kt (1) is Feller’s branching diffusion. R t R 2 To show that 0 J( f ,s)dM(s,y) is an L martingale consider
Z t Z t Z s Z 2! 2 2 P Ks J( f ,s) ds = P Ks φ2 φ1φ31(u,v]dW(r,ξ) ds 0 0 0 Z t Z s Z 2 ≤ c P Ks (1)P (φ1φ31(u,v]) dξdr ds 0 0 < ∞.
2 Hence the sum of these two quantities is also an L martingale. Since Mt (φ2) is R t R R t R continuous, 0 J( f ,s)dM(s,y) is continuous. As 0 Ks( f (s,ξ))dW(s,ξ) is con- R t R tinuous so is Kt 0 f (s,ξ,ω,y)dW(s,ξ) . Suppose the statement of the theorem holds for a sequence of (Fˆ t )-predictable
23 ˆ ∗ processes fk and f is an (Ft )-predictable process such that fk → f pointwise, R 2 R 2 supk k fk(ξ)k dξ ∨ k f (ξ)k dξ < ∞ and
ZZ N Z 2 lim P ( fk(s,ξ) − f (s,ξ)) dξdsKN(dy) = 0 ∀N ∈ N. (2.6) k→∞ 0
ˆ Since J( f ,s) is a White noise integral with respect to Ps, consider for each N ∈ N,
Z N Z 2 P(hM (J( f ,s) − J( fk,s))iN) = P (J( fk,s) − J( f ,s)) Ks(dy)ds 0 Z N ˆ 2 = Ps (J( fk,s) − J( f ,s)) ds 0 Z N Z s Z ˆ 2 = Ps ( fk(r,ξ) − f (r,ξ)) dξdr ds 0 0 Z N Z s Z ˆ 2 = PN ( fk(r,ξ) − f (r,ξ)) dξdr ds 0 0 where in the second line we have used Fubini’s theorem and the definition of the ˆ Campbell measure Ps. In the third line we make use of the Itoˆ isometry and then finally we use Remark V.2.5 (d) of [14] in the last line. The last expres- sion now goes to zero as k → ∞ by (2.6) (and Fubini’s theorem). This shows that R t R 2 0 J( f ,s,y)dM(s,y) is a continuous L martingale. Now by the Dominated Con- vergence Theorem, ∀t as k → ∞,
Z t Z 2 P hJ(Ks( f (s))) − J(Ks( fk(s)))it = P Ks( f (s,ξ) − fk(s,ξ)) dξds 0 → 0,
R t R 2 and so 0 Ks( f (s,ξ))dW(s,ξ) is a continuous L martingale as well.
By using Theorem 2.2.8(e) with S = N and (2.6), we get
P supKt (|J( fk,t) − J( f ,t)|)→ 0 as k → ∞,∀N ∈ N. t≤N Sending k → ∞ now shows that (2.4) holds for f as well.
The rest of the proof proceeds by appealing to a Monotone Class Theorem to pass to the bounded pointwise closure of functions satisfying (2.4). This is exactly
24 what is done in the last part of the proof of Proposition V.3.4 of [14].
ˆ n Theorem 2.2.11 (Ito’sˆ Lemma). Let Z0 be F0-measurable and take values in R . n ˆ ∗ Let f ∈ D1(n,d), g ∈ D2(m,n,d), h an R -valued (Ft )-predictable process and 1,2 n ψ ∈ Cb (R+ × R ). Assume
Z t Z t Z 2 2 Ks(k fsk + |hs|)ds + Ks kg(s,ξ)k dξ ds < ∞, ∀t a.s., (2.7) 0 0 and let
Zt (ω,y) = Z0(ω,y) (2.8) Z t Z t Z Z t + f (s,ω,y)dy(s) + g(s,ξ,ω,y)dW(s,ξ) + h(s,ω,y)ds, 0 0 0 then
Z Z ψ(t,Zt )Kt (dy) = ψ(0,Z0)dK0(y) Z t Z Z t Z + ψ(s,Zs)dM(s,y) + Ks(∇ψ(s,Zs)g(s,ξ))dW(s,ξ) 0 0 ! Z t ∂ψ 1 n n + Ks (s,Zs) + ∇ψ(s,Zs) · hs + ∑ ∑ ψi, j(s,Zs)ai j(s) ds, (2.9) 0 ∂s 2 i=1 j=1 where ∇ψ and ψi j are the gradient and second order partial derivatives in the spatial variables and a ≡ f f ∗ + R gg∗(ξ)dξ. The second term on the right is an L2 martingale, the third is a local martingale and the last term on the right has continuous paths with finite variation over compact intervals a.s. Therefore R ψ(t,Zt )Kt (dy) is a continuous (Ft )-semimartingale.
Proof. For now assume that k f k and |h| are bounded and that R 2 ¯ sups,y kg(s,ξ,ω,y)k dξ < ∞. Let T ∈ Tb and Zt = (t,Zt ). Using the classical ˆ Ito’sˆ Lemma we have, PT -a.s., ∀t ≥ 0
25 ψ(Z¯T∧t ) − ψ(Z¯0) = Z T∧t Z T∧t Z T∧t ¯ ∂ψ ¯ 1 ¯ i j ∇ψ(Zs) · dZs + (Zs)ds + ∑ ψi j(Zs)dhZ ,Z is 0 0 ∂s 2 i, j≤n 0 Z T∧t Z T∧t Z ˆ ˆ = PT - ∇ψ(Z¯s) f (s) · dy(s) + PT - ∇ψ(Z¯s)g(s,ξ) · dW(s,ξ) 0 0 Z T∧t n n ! ¯ ∂ψ ¯ 1 ¯ + ∇ψ(Zs) · h(s) + (Zs) + ∑ ∑ ψi, j(Zs)ai j(s) ds. 0 ∂s 2 i=1 j=1
Here in the second line we have simply plugged in (2.8), used Corollary 2.2.5 to manage the square function and done some linear algebra. Letting b˜(s) denote the term inside the last integral above, we see that
Z t S(t) = ψ(Z¯t ) − ψ(Z¯0) − b˜(s)ds 0
ˆ ∗ is an (Ft )-predictable process satisfying (2.2). Hence we apply Corollary 2.2.9 (with f and g replaced by ∇ψ(Z¯) f and ∇ψ(Z¯)g respectively) to get
Z t ψ(Z¯t ) = ψ(Z¯0) + I(∇ψ(Z¯) f ,t) + J(∇ψ(Z¯)g,t) + b˜(s)ds. (2.10) 0
Since |b˜|,k∇ψ(Z¯) f k and R k∇ψ(Z¯)g(ξ)k2dξ are all bounded and ˆ ∗ (Ft )-predictable, we can apply Theorem 2.2.10 above, Proposition V.3.4 and Re-
26 marks V.2.5 of [14] to get P-a.s. for all t ≥ 0,
Z Z Z t Z ψ(Z¯t )Kt (dy) = ψ(Z¯0)K0(dy) + ψ(Z¯0)dM(s,y) 0 Z t Z Z t Z + I(∇ψ(Z¯) f ,s)dM(s,y) + J(∇ψ(Z¯)g,s)dM(s,y) 0 0 Z t Z Z s Z t Z + b˜(r)dr dM(s,y) + Ks(∇ψ(Z¯s)g(s,ξ))dW(s,ξ) 0 0 0 Z t + Ks(b˜(s))ds 0 Z Z t Z = ψ(Z¯0)K0(dy) + ψ(Z¯s)dM(s,y) 0 Z t Z Z t + Ks(∇ψ(Z¯s)g(s,ξ))dW(s,ξ) + Ks(b˜(s))ds. 0 0 In the second equality, we used (2.10) to simplify. This gives the result in this case. Now assume that f ,g and h satisfy (2.7). We may choose f k,gk and hk bounded ˆ ∗ R k 2 k k and (Ft )-predictable and such that kg (ξ)k dξ is bounded and f → f ,g → g and hk → h pointwise (by, for example, truncating f ,g and h) with k f kk ≤ k f k,kgkk ≤ kgk and |hk| ≤ |h|. Therefore, by the Dominated Convergence Theo- rem,
Z t Z k 2 k k 2 Ks k fs − fsk + |hs − hs| + kgs − gsk dξ ds → 0 ∀t ≥ 0 a.s. (2.11) 0
Using (2.7) choose Sn ∈ Tb, Sn ↑ ∞, a.s. such that
Z Sn Z Sn Z 2 2 Ks k fsk + |hs| ds + Ks kgsk dξ ds ≤ n a.s. 0 0
Let Zk be as in (2.8) except with f ,g,h replaced by f k,gk,hk respectively. Using K (R t |h |ds) < n Lemma V.3.3(b) of [14] we get that supt≤Sn t 0 s ∞ for all , a.s. and hence Zt (ω,y) is well-defined Kt -a.a. y for all t ≥ 0 a.s. The same lemma shows that Z t k P sup Kt |hs − hs|ds →0 as k → ∞ ∀n. (2.12) t≤Sn 0
27 Using Proposition V.3.2(e) and Lemma V.3.3(b) of [14] gives k 2 P sup Kt supkI( f ,s) − I( f ,s)k → 0 as k → ∞ ∀n. (2.13) t≤Sn s≤t
Similarly, using Theorem 2.2.8 above and Lemma V.3.3(b) of [14] gives k 2 P sup Kt supkJ(g ,s) − J(g,s)k → 0 as k → ∞ ∀n. (2.14) t≤Sn s≤t
Together, (2.12), (2.13) and (2.14) give k P supKt sup|Z (s) − Z(s)| →0 as k → ∞ ∀T > 0. (2.15) t≤T s≤t
Since we have that (2.9) holds for (Zk, f k,gk,hk) in place of (Z, f ,g,h), the bound- edness of ψ and its derivatives together with (2.11) and (2.15) lets us send k → ∞ and derive the result using the Dominated Convergence Theorem. Note that the proofs of Theorem 2.2.10 above and Proposition V.3.4 of [14] imply that the second term of (2.9) is an L2 martingale and the third is a local martingale. The condition (2.7) can be used to show that the last term of (2.9) has finite variation on bounded intervals a.s.
Corollary 2.2.12. If f ,g,h and Z are as in Theorem 2.2.11, then
Z Xt (A) = 1(Zt (ω,y) ∈ A)Kt (dy)
n defines an a.s. continuous (Ft )-predictable MF (R )-valued process.
The proof for this corollary is the same as for Corollary V.3.6 of [14].
2.3 The strong equation
In this section, let K be as in Section 2.2 and let ν be the law of K0. Let σ1 : d d d×d d d m d×d d d MF (R ) × R → R , σ2 : MF (R ) × R × R → R , b : MF (R ) × R → d d d R t R and Z0 : R → R be Borel maps. If 0 σ1(Xs,Zs)dy(s) is the integral I and R t R 0 σ2(Xs,Zs,ξ)dW(s,ξ) is the integral J discussed in Section 2.2 above, then the
28 precise interpretation of (SE) is as follows:
Z t (a) Zt (ω,y) = Z0(y0) + σ1(Xs,Zs)dy(s) 0 Z t Z Z t (SE)Z0,K + σ2(Xs,Zs,ξ)dW(s,ξ) + b(Xs,Zs)ds K-a.e. 0 0 Z d (b) Xt (ω)(A) = 1(Zt (ω,y) ∈ A)Kt (dy) ∀A ∈ B(R ),∀t ≥ 0 a.s.
ˆ ∗ d (X,Z) is a solution of (SE)Z0,K iff Z is an (Ft )-predictable R -valued process and d X is an (Ft )-predictable MF (R )-valued process such that (SE)Z0,K holds.
Before we start discussing the existence of a solution of (SE)Z0,K, we must impose some conditions on the problem to make it more tractable. Let Lip1 = {φ : d d d R → R : kφk∞ ≤ 1,|φ(x) − φ(z)| ≤ |x − z| ∀x,z ∈ R } and for µ,ν ∈ MF (R ), d denote the Vasserstein metric on MF (R ) by
d(µ,ν) = sup |µ(φ) − ν(φ)|. φ∈Lip1
Let σ1,σ2 and b be such that there is an increasing function L : R+ → R+ with
0 0 0 0 Lip (a) kσ1(µ,z) − σ1(µ ,z )k + |b(µ,z) − b(µ ,z )| 0 0 0 0 d 0 d ≤ L(µ(1) ∨ µ (1)) d(µ, µ ) + |z − z | ∀µ, µ ∈ MF (R ),z,z ∈ R Z 0 0 2 (b) kσ2(µ,z,ξ) − σ2(µ ,z ,ξ)k dξ
≤ L(µ(1) ∨ µ0(1))[d(µ, µ0)2 + |z − z0|2] Z 2 (c) sup(kσ1(0,z)k + |b(0,z)|) < ∞ and sup kσ2(0,z,ξ)k dξ < ∞. z z
The following Lemma follows easily from the Lipschitz type conditions above.
Lemma 2.3.1. There exists a non-decreasing function α : R+ → R+ such that
Z 2 sup kσ1(µ,z)k + kσ2(µ,z,ξ)k dξ + |b(µ,z)| ≤ α(µ(1)) (2.16) z
Define TN = inf{t : Kt (1) ≥ N} ∧ N and
29 d S1 = {X : R+ × Ω → MF (R ) : X is (Ft )-predictable, a.s. continuous and
Xt (1) ≤ N ∀t < TN ∀N ∈ N} ˆ d ˆ ∗ S2 = {Z : R+ × Ω → R : Z is (Ft )-predictable and continuous K-a.e.}
S = S1 × S2.
Let
Φ2(X,Z)(t,y) ≡ Z˜t (y) Z t Z t Z Z t = Z0(y0) + σ1(Xs,Zs)dys + σ2(Xs,Zs,ξ)dW(s,ξ) + b(Xs,Zs)ds 0 0 0 Z Φ1(X,Z)(t)(·) ≡ X˜t (·) = 1(Z˜t (ω,y) ∈ ·)Kt (dy) and let Φ = (Φ1,Φ2). Then note that Φ : S → S by Corollary 2.2.12.
Lemma 2.3.2. Let (X˜ i,Z˜i) = Φ(Xi,Zi) for i = 1,2. If there exist universal con- stants cN such that for any stopping time T ≤ TN,
Z ˆ ˜1 ˜2 2 ˜1 ˜2 2 PT sup |Z (s) − Z (s)| ∧ 1 ≤ cN |Z0 − Z0 | ∧ 1dm (2.17) s≤t∧T Z t∧T Z t∧T 1 2 2 ˆ ˜1 ˜2 2 + cN P d(Xs ,Xs ) ∧ 1ds + PT |Z (s) − Z (s)| ∧ 1ds , 0 0
∀N ∈ N, then there is a pathwise unique solution (X,Z) to (SE)Z0,K. In particular, X is unique up to P-null sets and Z is unique K-a.e. Also, the map t 7→ Xt is a.s. continuous in t.
Proof. The proof of this lemma is essentially provided in the proof of Theorem V.4.1(a) of [14]. To see this, first set K1 = K2 in that proof, and then note that inequality (2.17) above is the same as inequality V.4.9 there, and from there on follow the proof substituting (2.17) for V.4.9 where it appears. The idea is to perform a contraction argument on the complete metric space
(S,d0), where d0 is an appropriately chosen metric that depends on the sequence
{cN}. One has to prove that Φ is a contraction on S.
Theorem 2.3.3. If (Lip) holds, then there exists a pathwise unique solution (X,Z)
30 to (SE)Z0,K. In particular, X is unique up to P-null sets and Z is unique K-a.e. Also, the map t 7→ Xt is a.s. continuous in t.
Proof. We will simply verify that the hypotheses in Lemma 2.3.2 hold. Let TN be i i as above and T ∈ Tb with T ≤ TN. Let (X˜ ,Z˜ ) be as above, for i = 1,2. Using first Doob’s strong L2 inequality and then successive applications of the Cauchy-
Schwartz inequality and the fact that TN ≤ N, we have ˆ 1 2 2 PT sup |Z˜ (s) − Z˜ (s)| ∧ 1 s≤t∧T Z t∧T Z ˆ 1 2 2 1 1 2 2 2 ≤ cPT |Z0 − Z0 | ∧ 1 + kσ2(Xs ,Zs ,ξ) − σ2(Xs ,Zs ,ξ)k dξds 0 Z t∧T ˆ 1 1 2 2 2 1 1 2 2 2 + cPT kσ1(Xs ,Zs ) − σ1(Xs ,Zs )k + N|b(Xs ,Zs ) − b(Xs ,Zs )| ds 0 ˆ 1 2 2 ≤ cPT (|Z0 − Z0 | ∧ 1) Z t∧T ˆ 1 2 2 1 2 1 2 2 + cNPT L(Xs (1) ∨ Xs (1)) (d(Xs ,Xs ) + |Zs − Zs |) 0 ! 1 2 2 2 ∧ α(Xs (1)) + α(Xs (1)) ds " ˆ 1 2 2 ≤ cPT |Z0 − Z0 | ∧ 1
Z t∧T # 2 1 2 1 2 2 2 + NL(N) (d(Xs ,Xs ) + |Zs − Zs |) ∧ α(N) ds , (2.18) 0 where we have used (Lip) (a), (b) and Lemma 2.3.1 for the third inequality. For the fourth inequality we use the fact that T ≤ N and Corollary 2.2.12 (to conclude i that Xs(1) ≤ N for s ≤ T) and that α and L are non-decreasing. Note that in the above, all the constants (not depending on N) are collected under the term c, and this term then varies from line to line. Also, if T = 0, then KT (1) may exceed
N, but the integral in (2.18) above is then zero. If T > 0 then TN > 0 and so
31 KT (1) ≤ supt≤T Kt (1) ≤ N. Hence we have
ˆ 1 2 2 PT sup |Z˜ (s) − Z˜ (s)| ∧ 1 s≤t∧T Z Z t∧T 1 2 2 2 2 2 1 2 2 ≤ c |Z0 − Z0 | dm + cN L(N) α(N) P d(Xs ,Xs ) ∧ 1ds 0 Z t∧T 2 2 ˆ 1 2 2 + cNL(N) α(N) PT |Zs − Zs |) ∧ 1ds , 0 where we have used Proposition 2.2.1 to handle the first term and then factored 2 2 2 2 out the α(N) term inside the integral in (2.18). Letting cN = N L(N) α(N) and invoking Lemma 2.3.2 completes the proof.
Examples of functions σ1,b satisfying (Lip) (a) can be found in Section V.1 of [14]. The following remark gives some broad classes of coefficients σ2 for which (Lip)(b) holds.
Remark 2.3.4. Suppose there exists a finite constant c0 such that for all ξ,
0 0 0 0 kσ2(µ,z,ξ) − σ2(µ ,z ,ξ)k ≤ c0(d(µ, µ ) + |z − z |) and Z 2 sup kσ2(0,z,ξ)k dξ < ∞. z Additionally, suppose one of the following properties hold:
(i) σ2(µ,z,·) has compact support for each µ,z and satisfying |suppσ2(µ,z,·)| ≤ L(µ(1)) for all z for some increasing function L : R+ → R+ (where |A| m represents the m-dimensional volume of A ⊂ R ).
(ii) σ2(µ,z,ξ) has no measure dependence and has first and second derivatives α in z. Also, there exists M > 0 such that supz k∂ (σ2)(z,·)k2 < M for all multi-indices α up to order 2.
(iii) σ2(µ,z,ξ) has no measure dependence, is continuous, integrable in ξ for each z, and has uniformly bounded first and second derivatives in z. Also assume that there exists M > 0 such that for all i ≤ d, sup k ∂ σ (z,·)k z ∂zi 2 1 < M.
32 R n nd d+nd+m d×d (iv) σ2(µ,z,ξ) = f (z,x,ξ)dµ (x) where x ∈ R and f : R → R such that f is bounded, twice differentiable in x and z and there exists M > 0 with α supz,x k∂ f (z,x,·)k2 < M for all multi-indices α with |α| ≤ 2. Additionally 2 m assume that there exists an L function c : R → R+ such that for all z, the function f (z,·,ξ) is Lipschitz with constant c(ξ).
n+d+m d×d (v) σ2(µ,z,ξ) = f (µ(φ1),..., µ(φn),z,ξ) where f : R → R and such that f ((x,z),·) satisfies the conditions on σ2 in either (ii) or (iii) above and
each φi bounded and measurable.
Then (Lip) (b) holds.
Proof. The proof of (Lip) (b) under case (i) is not hard (just use the Lipschitz property followed by the compact support property). Assume, for case (ii) that ξ 2 ξ g (z) = kσ2(z,ξ) − σ2(z0,ξ)k . Since σ2 is twice differentiable, so is g . Note ξ ξ that for each y, g has a minimum when z = z0 and that g (z0) = 0. Hence, by looking at the Taylor expansion we get
Z 1 ξ 2 α α ξ g (z) = ∑ (z − z0) (1 −t)∂ (g )(z0 +t(z − z0))dt |α|=2 α! 0 and, after letting z(t) = z0 +t(z − z0) we have
Z ZZ 1 ξ 2 α ξ g (z)dξ ≤ c|z − z0| ∑ (1 −t)k∂ (g )(z(t))kdtdξ α=2 0 Z 1 Z 2 2 α i j i j ≤ c1|z − z0| ∑ ∑ |∂ σ2 (z(t),ξ) − σ2 (z0,ξ) |dξdt α=2 0 i, j≤d 2 2 ≤ c2M |z − z0|
2 by the uniform L bound on the derivatives up to order 2 of σ2. This proves the lemma under this case. R 2 For (iii) first define the function h(z) = kσ2(z,ξ)−σ2(z0,ξ)k dξ. Then note that h(z0) = 0 and that this is a minimum for h. Therefore, since the first partials
33 of h vanish at z0
2 2 h(z1) ≤ |∇zh(z0) · (z1 − z0)| + csupk∇z h(z)k|z1 − z0| z 2 ≤ C|z1 − z0| by the boundedness of the second derivatives of h (which follows from the assump- tions on σ2). This gives us the result in this case. For case (iv), consider
σ2(µ1,z1,ξ) − σ2(µ2,z2,ξ) Z Z n n n = f (z1,x,ξ)d(µ1 − µ2 )(x) − ( f (z2,x,ξ) − f (z1,x,ξ))dµ2 (x) Rnd Rnd and hence through applications of Cauchy-Schwartz and Fubini, we get Z 2 kσ2(µ1,z1,ξ) − σ2(µ2,z2,ξ)k dξ Z 2 n n 2 ≤ 2 c(ξ) d(µ1 , µ2 ) dξ ZZ 2n 2 n + 2µ2(1) k f (z2,x,ξ) − f (z1,x,ξ)k dξdµ2 (x) since f (z,·,ξ)/c(ξ) ∈ Lip1(nd) for each pair z,ξ by the assumptions on f . Then by using the same proof as in part (iii) to handle the second term, the fact that c(ξ) 2 is L , and the definition of Lip1(nd) we get Z 2 n 2n 2 kσ2(µ1,z1,ξ) − σ2(µ2,z2,ξ)k dξ ≤ Cd(µ1, µ2) +C(µ1(1)µ2(1)) |z1 − z2|
2 2 ≤ L(µ1(1) ∨ µ2(1))[d(µ1, µ2) + |z1 − z2| ] by choosing L(x) = Cxn−1 +Cx2n. This gives (Lip) (b). The proof for case (v) follows easily from the proofs of either (ii) or (iii) and the definition of the Vasserstein metric.
34 2.4 A martingale problem We will now show that X satisfies a certain martingale problem; a generalized ver- sion of the martingale problem satisfied by the process constructed in [2] and an ex- tension to the one considered in [13]. First we need to show that any solution (X,Z) d to (SE)Z0,K satisfies the following martingale problem (MP): X0 = µ ∈ MF (R ) and 2 d for any φ ∈ Cb(R ),
Z t X Mt (φ) ≡ Xt (φ) − µ(φ) − Xs(LXs φ)ds (2.19) 0 is a continuous martingale with square function
X hM (φ)it (2.20) Z t Z t ZZ 2 ∗ = Xs(φ )ds + ∇φ(z1)ρXs (z1,z2)∇φ (z2)dXs(z1)dXs(z2)ds 0 0
d where for ν ∈ MF (R ),
d d Z i 1 i j ∗ ∗ Lν φ(z) = ∑ b (ν,z)φi(z) + ∑ a (ν,z)φi j(z) with a = σ1σ1 + σ2σ2 (ξ)dξ i=1 2 i, j=1 (2.21) and ρν is the d × d matrix given by Z ∗ ρν (z1,z2) = σ2(ν,z1,ξ)σ2 (ν,z2,ξ)dξ. Rm R Lemma 2.4.1. Any solution (X,Z) of (SE)Z0,K satisfies (MP) with X0(·) = 1(Z0 ∈ ·)K0(dy).
Proof. The proof follows directly from Theorem 2.2.11 and (SE)Z0,K. First we d note that for φ : R → R bounded and measurable, (SE)Z0,K implies that Xt (φ) = R 2 d φ(Zt (y))Kt (dy). Then from Ito’sˆ Lemma (Theorem 2.2.11), for ψ ∈ Cb(R ) we
35 get
Z Z ψ(Zt )Kt (dy) = ψ(Z0)dK0(y) + Nt
Z t n ! 1 i j + Ks ∇ψ(Zs) · b(Xs,Zs) + ∑ ψi, j(Zs)a (Xs,Zs) ds, 0 2 i, j=1 where Nt is a martingale. This becomes, by the boundedness of ψ and its partials,
Z t Z Xt (ψ) = X0(ψ) + Nt + LXs ψ(z)Xs(dz)ds. 0 Since
Z t Z Z t Z Nt = ψ(Zs)dM(s,y) + Ks(∇ψ(Zs)σ2(Xs,Zs,ξ))dW(s,ξ), 0 0 by the orthogonality of M and W we have
Z t 2 hNit = Xs(ψ )ds 0 Z t Z Z Z ∗ + ∇ψσ2(Xs,z,ξ)dXs(z) ∇ψσ2(Xs,z,ξ)dXs(z) dξds, 0 which can be rewritten in the form of (2.20).
nd d Definition. For a function f : R → R, define Fn, f : MF (R ) → R by Fn, f (µ) = R f dµn.
Nn It can be shown, by induction and using Lemma 2.4.1, that for f = i=1 φi, 2 d where each φi ∈ Cb(R ), that
Z t Nt (n, f ) = Fn, f (Xt ) − Fn, f (X0) − LFˆ n, f (Xs)ds (2.22) 0
36 is a martingale, where
n ! ˆ LFn, f (µ) = ∑ µ(Lφi) ∏ µ(φ j) (2.23) i=1 j6=i ! 1 n Z + µ(φk) µ(φiφ j) + µ(∇φiσ2(ξ)) · µ(∇φ jσ2(ξ))dξ . ∑ ∏ m 2 i, j=1 k6=i, j R i6= j
We have suppressed the measure dependence of the generator L here and will do so below as well.
nd Definition. If f ∈ B(R ) (i.e. f a real valued, bounded Borel measurable function nd n nd (n−1)d on R ) then we can define the operator Φi, j : B(R ) → B(R ) acting on f as: n Φi, j f (z1,...,zn−1) = f (z1,...,zn−1,...,zn−1,...,zn−2) where zn−1 is inserted into the vector (z1,...,zn−2) such that it appears at the ith and jth spots.
We can now rewrite (2.23) in terms of f as follows:
n Z n Z ˆ i n 1 n n−1 LFn, f (µ) = ∑ L f dµ + ∑ Φi, j( f )dµ (2.24) i=1 2 i, j=1 i6= j n d Z 2 1 k,l ∂ n + ∑ ∑ ρµ (zi,z j) f (z)dµ (z). 2 i, j=1 k,l=1 ∂zik ∂z jl i6= j
i Here z = (z1,...,zn) and L f (z1,...,zn) = L fz1,...,zi−1,zi+1,...,zn (zi) (ie, applying L to a function of just zi and fixing the other coordinates) where L is as in (2.21). Note that we have suppressed the dependence of Li on ν, as with L. By Lemma 2.4.1
37 and an inductive argument, one can see that
Z t Z n ! Nt (n, f ) = ∑ ∏Xs(φ j) Xs(∇φiσ2(ξ))dW(s,ξ) (2.25) 0 i=1 j6=i Z t Z n ! + ∑ ∏Xs(φ j) φi(Zs)dM(s,y) 0 i=1 j6=i n Z t Z Z n = ∑ ∇zi f (z)σ2(zi,ξ)Xs (dz) dW(s,ξ) (2.26) i=1 0 n Z t Z Z n−1 i + ∑ f (Zs(y1),...,Zs(yn))Ks (dy ) dM(s,yi) i=1 0
i where y = (y1,...,yi−1,yi+1,...,yn). Then a calculation gives
n Z t Z ∗ 2n hN(n, f )it = ∑ ∇i f (z1)ρXs (zi,zn+ j)(∇n+ j f (z2)) Xs (dz) ds (2.27) i, j=1 0 n Z t Z Z 2n 2n−1 2n + ∑ Φi, j+n( f ⊗ f )Xs (dz ) ds i, j=1 0 where z = (z1,z2) with z1 = (z1,...,zn),z2 = (zn+1,...,z2n) and where 2n z = (z1,...,z2n−1).
2 nd Theorem 2.4.2. For any function f ∈ Cb(R ),
Z t Nt (n, f ) = Fn, f (Xt ) − Fn, f (X0) − LFˆ n, f (Xs)ds (2.28) 0 is a martingale, where Lˆ is given by (2.24).
Proof. Let
M nd O ∞ Hn = ∑ ai φi j : φi j ∈ Cc (R),ai ∈ R,M ∈ N . i=1 j=1
∞ nd We will first use functions in Hn to prove that for f ∈ Cc (R ), (2.28) is a mar- 2 nd tingale and then boost that up to include functions in Cc (R ) and finally those in 2 nd ∞ nd Cb(R ). By the Stone-Weierstrass theorem, for any f ∈ Cc (R ), there is a se- α quence gk ∈ Hn such that gk → ∂ f uniformly, for α = (2,...,2). Then, by inte-
38 α α grating in the various components, we obtain a sequence fk such that ∂ fk → ∂ f uniformly for |α| ≤ 2. ˆ ˆ This in turn gives point-wise convergence of Fn, fk to Fn, f and LFn, fk to LFn, f d on MF (R ) (using the Dominated Convergence Theorem). Now let TN = inf{t : Kt (1) ≥ N} ∧ N as above. Then for each k,l, 2 P sup |Ns(n, fk) − Ns(n, fl)| ≤ cP(hN(n, fk) − N(n, fl)it∧TN ) s≤t∧TN and hence by sending l → ∞ and using the Dominated Convergence Theorem, we conclude that 2 P sup |Ns(n, fk) − Ns(n, f )| ≤ cP(hN(n, fk) − N(n, f )it∧TN ) s≤t∧TN → 0 as k → ∞, by the Dominated Convergence Theorem and Lemma 2.3.1. Hence
Nt (n, f ) is a continuous local martingale. Since f and its derivatives up to order 2 are bounded and (Lip) (c) holds, we have by (2.27) that
Z t 2n 2n−1 P(hN(n, f )it ) ≤ cP Xs(1) + Xs(1) ds . 0
The right side is finite by the boundedness of moments of the mass process of super-Brownian motion. Therefore Nt (n, f ) is a martingale. 2 nd Now let f ∈ Cc (R ). Using approximate identities we can construct functions ∞ nd α α fk ∈ Cc (R ) such that ∂ fk → ∂ f uniformly for all α with |α| ≤ 2. Repeating the above argument again and using the fact that each Nt ( fk) is a martingale shows that Nt ( f ) is a continuous martingale. 2 nd ∞ nd Finally, suppose f ∈ Cb(R ). Let ηk ∈ Cc (R ) such that ηk = 1 on B(0,k), c α ηk = 0 on B(0,k + 1) and k∂ ηkk < 2 for each multi-index with |α| ≤ 2. Letting α α fk = f ηk, we see that ∂ fk → ∂ f pointwise for each |α| ≤ 2. Here B(0,k) denotes the open ball of radius k about the origin. ˆ ˆ This gives LFn, fk → LFn, f pointwise (by the Dominated Convergence Theo- rem). Then by noting that each Nt ( fk) is a continuous martingale and using the
39 method in the first part of this proof shows that Nt ( f ) is a continuous martin- gale.
To see that the above martingale problem reduces to the one found in [2] in their setting, let m = d = 1, b = 0, σ2(µ,z,ξ) = h(ξ − z) where h is square-integrable and ( ,z) = (z) is Lipschitz.. Additionally, suppose that (z) ≡ R h( − σ1 µ σ1 η R ξ z)h(ξ)dξ is twice continuously differentiable with η0 and η00 bounded. Then one can easily check that these coefficients satisfy (Lip) (use Taylor’s theorem on η to show (Lip)(b)). To get the form of the infinitesimal generator, we use (2.24) :
n Z n Z ˆ 1 2 n 1 n LFn, f (µ) = ∑ (σ1 (zi)+η(0)) fii(z)dµ (z)+ ∑ ρ(zi,z j) fi j(z)dµ (z) 2 i=1 2 i, j=1 i6= j n Z 1 n n−1 + ∑ Φi, j( f )dµ 2 i, j=1 i6= j which is equal to the sum of the generators (2.1) and (2.2) of [2], since ρ(zi,z j) =
η(zi − z j), if we allow their branching rate σ to be constant at 1. In [2], uniqueness in law of the martingale problem is established. Hence in this particular setting, our solution is both unique in law and pathwise unique and therefore is a version of the process constructed in [2].
2.5 Additional properties of solutions
We will start off with a stability result for solutions of (SE)Z0,K, and then use it to prove the strong Markov property. We conclude with a proof of the compact support property for the solutions of (SE)Z0,K. Recall that d denotes the Vasserstein d metric on MF (R ).
1 2 Theorem 2.5.1. Assume (Lip) holds. Let K ≤ K ≡ K be (Ft )-historical super- i d i d d Brownian motions with mi(·) ≡ E(K0(·)) ∈ MF (R ), and let Z0 : R → R be Borel maps, for i = 1,2. Let TN = inf{t : Kt (1) ≥ N} ∧ N. There are universal constants
40 i i {cN,N ∈ } so that if (X ,Z ) is the unique solution of (SE) i i , then N Z0,K
Z TN 1 2 2 P supd(X (r),X (r)) ds 0 r≤s Z 1 2 2 ≤ cN |Z0 − Z0 | ∧ 1dm2 + m2(1) − m1(1)
The proof of this theorem essentially follows from the proof of Theorem V.4.2 of [14]. Note that before we can follow that proof, we will need to show that the condition (2.17) of Lemma 2.3.2 above holds under (Lip), where (X˜ i,Z˜i) = Φi(Xi,Zi) where each Φi depends on different historical super-Brownian motions i 1 2 dK1 K ,i = 1,2 and K ≤ K (ie, the Radon-Nikodym derivative dK2 ≤ 1). The compu- tation showing this is almost exactly like the one in the proof of Theorem 2.3.3. We now construct a canonical space upon which we can define a white noise and an independent historical super-Brownian motion. Let K be an (Ft )-adapted ¯ historical super-Brownian motion on Ω = (Ω,F,(Ft )t≥0,Qν ). Let (X,Z) be the ˆ ∗ unique solution to (SE)Z0,K. Recall that X is (Ft )-predictable, and Z is (Ft )- d predictable. Denote by ΩX the space of continuous paths in MF (R ) and FX its Borel σ-field. t t Let ΩK = {H· ∈ C([0,∞),MF (C)) : Ht ∈ MF (C)∀t ≥ 0} where MF (C) = {µ ∈ t MF (C) : µ(A) = µ (A ∩ {y ∈ C : y = y })∀A ∈ C}. The historical super-Brownian motion K has sample paths in ΩK. Let FK be the Borel σ-field of ΩK and let
∞ K \ Ft = σ(Hr : 0 ≤ r ≤ t + 1/n). n=1
To treat W we will follow the method of [12]. Let QW be the law of the white d noise W. For functions f ∈ C(R ), define
−λ|x| k f kλ = sup | f (x)|e , x∈Rd
m and let Ctem = { f ∈ C(R ) : k f kλ < ∞ for all λ > 0}. We let Ctem be endowed with the topology generated by the family of norms k · kλ ,λ > 0. Hence, Ctem can ∞ −k be thought of as a metric space with metric d given by d( f ,g) = ∑k=1 2 (k f − 1 d gk1/k ∧ 1). It turns out that (Ctem,d) is a Polish space. Let W = (W ,...,W ) be
41 a d-dimensional white noise, as in Section 2.2. We can now identify each white noise W i with its associated Brownian sheet and say each W i is a process with sample paths in C([0,∞),Ctem). Hence we will say that W has sample paths in d ΩW ≡ C([0,∞),Ctem). Endow ΩW with the topology of uniform convergence on compacts. d Let We : [0,∞) × ΩW → Ctem be the coordinate projection map, and let
∞ W \ Ft = σ(Wes : s ≤ t + 1/n). n=1
Then (H,We ) is a Borel strong Markov process on (ΩK,W ,H,Ht ,Qν,W ) where ΩK,W = ΩK × ΩW ,Qν,W = Qν × QW and for t ≥ 0, Ht is given by σ((H,We )(s) : s ≤ t) with the Qν,W -null sets thrown in. We will now treat H and We as random processes on the space ΩK,W with H(t,ω,α) = Ht (ω) and We (t,ω,α) = αt . Under Qν,W , H and We are independent by definition and have laws Qν and QW respectively. Recall from the definition in Section 2.2 that Ω[K,W = ΩK,W ×C and similarly ˆ ˆ ∗ Ht = Ht × Ct with universal completion Ht .
Statement (a) below shows that (SE)Z0,K has a strong solution, whereas (b) shows that there is continuity of the laws of the solutions in the initial condition. ¯ Theorem 2.5.2. Let (X,Z) be the unique solution of (SE)Z0,K on Ω. Then the following holds: ˆ ∗ ˜ (a) There are (Ht )-predictable and (Ht )-predictable maps X : R+ × ΩK,W → d d MF (R ) and Z˜ : R+ × Ω[K,W → R , respectively, depending only on (Z0,ν), and such that
X(t,ω),Z(t,ω,y) = X˜ t,K(ω),W(ω),Z˜ t,K(ω),W(ω),y (2.29)
gives the unique solution of (SE)Z0,K.
(b) There is a continuous map from X 7→ 0 from M ( d) to M (Ω ) such that 0 PX0 F R 1 X ¯ if (X,Z) is a solution of (SE)Z0,K on some filtered space Ω then Z (X ∈ ·) = 0 (·)d (ω). (2.30) P PX0(ω) P
42 (c) If T is an a.s. finite (Ft )-stopping time, then for all A ∈ FX
(X(T + ·) ∈ A|F )( ) = 0 (A), -a.s. P T ω PXT (ω) P
Proof. The proof of (a) is an analogue of the proof of Theorem V.4.1 (b) of [14]. ˜ ˜ For (a), let (X,Z) be the unique solution of (SE)Z0,K where (H,We ) is the canon- ¯ 0 ical process on Ω ≡ (ΩK,W ,H,Ht ,Qν,W ). It is clear from Theorem 2.3.3 that (X˜ ,Z˜) depends only on Z0 and ν, and satisfies the desired predictability condi- tions. To show (2.29), we need only show that the latter process solves (SE)Z0,K on ¯ 0 0 0 0 Ω. Let I ( f ,t,H,W,y),J (g,t,H,W,y) and f ∈ D1(d,d),g ∈ D2(m,d,d) denote the sample path stochastic integral and the white noise integral respectively from Sec- 0 tion 2.2 on Ω¯ . Similarly define I( f ,t),J(g,t), f ∈ D1,g ∈ D2 to be the stochastic ¯ 0 integrals in Section 2.2 on Ω. We claim that if f ∈ D1, then f ◦ (K,W)(t,ω,y) ≡ f (t,K(ω),W(ω),y) ∈ D1 and
I( f ◦ (K,W)) = I0( f ) ◦ (K,W),K-a.e. (2.31)
0 By the definition of D1 and D1 and since K and H have the same laws, we get the first implication. Equation (2.31) clearly holds for simple functions. Using 0 Proposition V.3.2 (d) of [14] allows us to show that it holds for all f ∈ D1. The 0 situation with the white noise integral J is similar in that if g ∈ D2, then J ◦(K,W) ∈ D2 and J(g ◦ (K,W)) = J0(g) ◦ (K,W),K-a.e..
This can again be shown by first starting at simple functions and then bootstrap- ping up using Theorem 2.2.8. To show the result, we can now simply replace ˜ ˜ (H,We ) with (K(ω),W(ω)) in (SE)Z0,H and get that (X ◦(K,W),Z ◦(K,W)) solves ¯ (SE)Z0,K on Ω. The proof of (b) is also a straight-forward analogue of the proof of Theorem V.4.1 (c) in [14]. The continuity condition in (b) depends on the stability result above (which is the analogue of the stability result in [14]). The proof of (c) is more involved, but follows exactly the same route as the proof of Theorem V.4.1 (d) in [14]. It uses the continuity proven in (b) to reduce the problem to proving the Markov property (via finite-valued stopping times) which
43 is then established by using the uniqueness of solutions of the strong equation. Again, the only additional detail to worry about is the interpretation of the white noise integral J.
We now prove the compact support property for solutions of (SE)Z0,K. Definition. Let f : [0,∞) × Ωˆ → (E,k · k) (a normed linear space). A bounded
(Ft )-stopping time T is called a reducing time for f if and only if 1(0 < t ≤
T)k f (t,ω,y)k is uniformly bounded. We say {Tn} reduces f if and only if each Tn reduces f and Tn ↑ ∞ P-a.s. If such a sequence exists, we say f is locally bounded. ˆ n Theorem 2.5.3. Assume that f ∈ D1(n,d), g ∈ D2(m,n,d) and b : [0,∞)×Ω → R ˆ ∗ is (Ft )-predictable, and all three are locally bounded. Let Z(t) = Z0 + I( f ,t) + R t ˆ t J(g,t) + V(b,t) where V(b,t) = 0 b(s)ds. Define K(·) = Kt ({y : Z (ω,y) ∈ ·}). Then for a.a. ω for each k ∈ N there is a compact set Sk(ω) ⊂ C such that −1 supp(Kˆt ) ⊂ Sk ∀t ∈ [k ,k], and furthermore, supp(Kˆt ) ⊂ Sk ∀t ∈ [0,k] on {ω : supp(Kˆ0) is compact}. The proof of this theorem is given by a small modification of the proofs of Corollaries 3.3 and 3.4 of [13]
Corollary 2.5.4. Assume (Lip). Let (X,Z) denote the solution of (SE)Z0,K. Then 0 d for a.a. ω for each k ∈ N there is a compact set Sk(ω) ⊂ R such that supp(Xt ) ⊂ 0 −1 0 Sk ∀t ∈ [k ,k], and on {ω : supp(X0) is compact}, supp(Xt ) ⊂ Sk ∀t ∈ [0,k].
Proof. Note that under (Lip), the coefficients σ1,σ2 and b are locally bounded and R −1 −1 that in the above, Xt (A) = 1(Z(t) ∈ A)Kt (dy) = Kˆt (πt A) where πt A = {y : Zt (ω,y) ∈ A} for A bounded measurable. Apply Theorem 2.5.3 and use the fact d that πt : C → R is a continuous mapping. −1 Let Sk be as in Theorem 2.5.3. Then for t ∈ [k ,k],
c ˆ −1 c Xt ((πt Sk) ) = Kt (πt ((πt Sk) )) ˆ c ≤ Kt (Sk) = 0.
0 Hence setting Sk = πkSk gives the result in this case. Proceed similarly for the extension for {ω : S(X0) is compact}.
44 Chapter 3
A super Ornstein-Uhlenbeck process interacting with its center of mass
3.1 Introduction The existence and uniqueness of a self-interacting measure-valued diffusion that is either attracted or repelled from its centre of mass is shown below. It seems natural to consider a super Ornstein-Uhlenbeck (SOU) process with attractor (repeller) given by the centre of mass of the process initially as it is the simplest diffusion of this type. This type of model first appeared in a recent paper of Englander¨ [3] where a d- dimensional binary Brownian motion, with each parent giving birth to exactly two offspring and branching occurring at integral times, is used to construct a binary branching Ornstein-Uhlenbeck process where each particle is attracted (repelled) by the center of mass (COM). This is done by solving the appropriate SDE along each branch of the particle system and then stitching these solutions together. This model can be generalized such that the underlying process is a branching Brownian motion (BBM), T, (i.e. with a general offspring distribution). We might
45 then solve an SDE on each branch of T:
Z t Z t n n−1 ¯ n n Yi (t) = Yp(i) (n − 1) + γ Yn(s) −Yi (s)ds + dBi (s), (3.1) n−1 n−1
n for n − 1 < t ≤ n, where Bi labels the ith particle of T alive from time n − 1 to n, p(i) is the parent of i and τn ¯ 1 n Yn(s) = ∑Yi (s) τn i=1 is the center of mass. Here τn is the population of particles alive from time n−1 to n. This constructs a branching OU system with attraction to the COM when γ > 0 and repulsion when γ < 0. It seems reasonable then, to take a scaling limit of branching particle systems of this form and expect it to converge in distribution to a measure-valued process where the representative particles behave like an OU process attracting to (repelling from) the COM of the process. Though viable, this approach will be avoided in lieu of a second method utilizing the historical stochastic calculus of Perkins [14] which is more convenient for both constructing the SOU process interacting with its COM and for proving various properties. The idea is to use a supercritical historical Brownian motion to construct the interactive SOU process by solving a certain stochastic equation. This approach for constructing interacting measure- valued diffusions was pioneered in [13] and utilized in, for example, [8]. A supercritical historical Brownian motion, K, is a stochastic process taking d values in the space of measures over the space of paths in R . One can think of K as a supercritical superprocess which has a path-valued Brownian motion as the underlying process. That is, if Bt is a d-dimensional Brownian motion, then
Bˆt = B·∧t is the underlying process of K. More information about K is provided in Section 3.2. d It can be shown that if a path y : [0,∞) → R is chosen according to Kt (loosely speaking – this is made rigorous below), then y(s) is a Brownian motion stopped at t. Then projecting down gives Z K Xt (·) = 1(yt ∈ ·)Kt (dy),
46 a (supercritical) super Brownian motion. One can sensibly define the
Ornstein-Uhlenbeck SDE driven by y according to Kt :
dZs = −γZsds + dys where γ > 0 implies Zt is attracted to the origin, and γ < 0 implies repulsion. Defining Z Xt (·) = 1(Zt (y) ∈ ·)Kt (dy) will give an ordinary super Ornstein-Uhlenbeck process. The intuition here is that
K keeps track of the underlying branching structure, and Zt is a function transform- ing a typical Brownian path into a typical Ornstein-Uhlenbeck path. Similarly, we can define a function of the path y that is an OU process with attraction to (repelling from) the COM, according to Kt and the SOU process with attraction (repulsion) to COM as follows:
dYs = γ(Y¯s −Ys)ds + dys Z 0 Xt (·) = 1(Yt (y) ∈ ·)Kt (dy) where the COM is defined as
R xX0 dx ¯ s( ) Ys = R 0 . 1Xs(dx)
The details are given in Section 3.3 and it is shown that such a process X0 can be constructed by a correspondence with the ordinary SOU process, X. In Section 3.4 we prove various properties of the ordinary SOU process, X, and the interacting SOU process, X0. It is shown in Theorem 3.4.1 that the COM converges in the case where X0 goes extinct in finite time, regardless of the value of γ. Theorem 3.4.2 shows that on the set where X0 survives indefinitely, the COM
Y¯t converges in the attractive case (where γ > 0). Theorem 3.4.3 then shows that in 0 Xt the extinction case the mass normalized process 0 converges a.s. to a point as t Xt (1) approaches the extinction time, which is a version of a result of Tribe [20]. With the convergence of the COM on the survival set, Theorem 3.4.9 gives that
47 in the attractive case, Xt (·) a.s. −→ P∞ (·) as t → ∞ Xt (1) where P∞ (·) is the stationary distribution of the OU process, which is an extension of Proposition 20 of Englander¨ and Turaev [4] and Theorem 1 of Englander¨ and Winter [5] for the SOU process. The correspondence between X and X0 given in Section 3.3 is used in Theorem 3.4.11 to prove in the attractive case,
X0 (·) t a.s. Y¯∞ 0 −→ P∞ (·) as t → ∞, Xt (1)
Y¯∞ where the distribution P∞ (·) is equal to P∞ (·) shifted to the limiting value of the COM Y¯t . In final section, in Theorem 3.5.1, we prove that the COM of the SOU process repelling from its COM converges (i.e. when γ < 0), provided the repulsion is not too strong. It is shown that the corresponding result for the SOU process with repulsion from the origin fails. This is one of the most interesting distinctions between the interactive SOU process and the non-interactive SOU process. We conclude with some conjectures, as well as a comparison with what is known for the SOU process repelling from the origin.
3.2 Background We will take K to be a supercritical historical Brownian motion. Specifically, we let K be a (∆/2,β,1)-historical superprocess (here ∆ is the d-dimensional Laplacian), where β > 0 constant, on the probability space Ω,F,(Ft )t≥0 ,P . The β in the definition corresponds to the branching bias in the offspring distribution, and the 1 to the variance of the offspring distribution. A martingale problem characterizing K is given below. For a more thorough explanation of historical Brownian motion than found here, see Section V.2 of [14]. d t t Let C = C(R+,R ) with C its Borel σ-field and define C = {y : y ∈ C} where t s y = y·∧t and Ct = σ(y ,s ≤ t,y ∈ C). 0 0 0 For metric spaces E and E , take Cc(E,E ) and Cb(E,E ) to be the space of compactly supported continuous functions from E to E0 and the space of bounded
48 continuous functions from E to E0, respectively. For a measure space (E,E), let bE denote the set of bounded E-measurable real valued functions.
Notation. For a measure µ on a space E and a measurable function f : E → R, let R n µ( f ) = f dµ. Furthermore, if f : E → R , define
µ( f ) = (µ( f1),··· , µ( fn)).
d For a point p ∈ R , let |p| denote Euclidean norm of p, and for a bounded function d d f : E → R , denote k f k = supx∈E ∑i=1 | fi(x)|.
t It turns out that Kt is supported on C ⊂ C a.s. and typically, Kt puts mass on those paths that are “Brownian” (until time t and fixed thereafter). As K takes values in MF (C), Kt (·) will denote integration over the y variable.
Let Bˆt = B(·∧t) be the path-valued process associated with B, taking values in t s,y C . Then for φ ∈ bC, if s ≤ t let Ps,t φ(y) = E (φ(Bˆt )), where the right hand side denotes expectation at time t given that until time s, Bˆ follows the path y. The weak generator, Aˆ, of Bˆ is as follows. If φ : R+ ×C → R we say φ ∈ D(Aˆ) if and only if φ is bounded, continuous and (Ct )-predictable, and for some Aˆsφ(y) with the same properties as φ,
Z t φ(t,Bˆ) − φ(s,Bˆ) − Aˆrφ(Bˆ)dr,t ≥ s, s
s is a (Ct )-martingale under Ps,y for all s ≥ 0,y ∈ C . d If m ∈ MF R , we will say K satisfies the historical martingale problem, (HMP)m, if and only if K0 = m a.s. and
Z t Z t ∀φ ∈ D(Aˆ), Mt (φ) ≡ Kt (φt ) − K0(φ0) − Ks(Aˆsφ)ds − β Ks(φs)ds 0 0
(HMP)m is a continuous (Ft )-martingale with Z t 2 hM(φ)it = Ks(φs )ds ∀t ≥ 0, a.s. 0
Using the martingale problem (HMP)m, one can construct an orthogonal mar- tingale measure Mt (·) with the method of Walsh [22]. Denote by P, the σ-field of
49 (Ft )-predictable sets in R+ × Ω. If ψ : R+ × Ω ×C → R is P × C-measurable and
Z t 2 Ks(ψs )ds < ∞ ∀ t ≥ 0, (3.2) 0 then there exists a continuous local martingale Mt (ψ) with quadratic variation R t 2 hM(ψ)it = 0 Ks(ψs )ds. If the expectation of the term in (3.2) is finite, then Mt (ψ) is an L2-martingale. ˆ ˆ ˆ ˆ ∗ Definition. Let (Ω,F,Ft ) = (Ω ×C,F × C,Ft × Ct ). Let Ft denote the univer- sal completion of Fˆ t . If T is a bounded (Ft )-stopping time, then the normalized ˆ ˆ ˆ Campbell measure associated with T is the measure PT on (Ω,F) given by
ˆ P(1AKT (B)) PT (A × B) = for A ∈ F,B ∈ C. mT (1)
ˆ where mT (1) = P(KT (1)). We denote sample points in Ω by (ω,y). Therefore, un- ˆ −1 der PT , ω has law KT (1)dP·mT (1) and conditional on ω, y has law KT (·)/KT (1). Notation. Henceforth we denote the super Ornstein-Uhlenbeck process by X and the interacting super Ornstein-Uhlenbeck process by X0. The symbol X˚ will be used as a catch-all and will vary according to the setting, but will be a measure- d valued process taking values in MF R . Occasionally the symbol K˚ will appear and will denote a historical process different from K.
3.3 Existence and preliminary results ˆ ∗ 1 2 1 Definition. For two (Ft )-measurable processes X and X , we will say that X = X2,K-a.e. if 1 2 X (s,ω,y) = X (s,ω,y),∀s ≤ t,Kt -a.a. y for all fixed times t ≥ 0.
(X,Z) 1 We say that is a solution to the strong equation (SE)Z0,K if it satisfies
Z t 1 Z ( ,y) = Z ( ,y ) + y − y − Z ( ,y)ds K − (SE)Z0,K (a) t ω 0 ω 0 t 0 γ s ω a.e. 0 Z d (b) Xt (A) = 1(Zt ∈ A)Kt (dy) ∀A ∈ B(R ),∀t ≥ 0,
50 ˆ ∗ where Zt is an (Ft )-predictable process and Xt is an (Ft )-predictable process. That this equation has a pathwise unique solution does not follow immediately from Theorem 4.10 of Perkins [13]. There it is shown that equations of the above type (but with more general, interactive, drift and diffusion terms in (a)) have solutions if K is a critical historical Brownian motion. 1 2 Note that in the definitions of (SE)Z0,K and (SE)Y0,K given below, part (b) is unnecessary to solve the equations. It has been included to provide an easy com- parison to the strong equation of chapter V.1 of [14].
There is a pathwise unique solution (X,Z) to (SE)1 . That is, X Theorem 3.3.1. Z0,K is unique P-a.s. and Z K-a.e. unique. Furthermore, the map t → Xt is continuous and X is a β-super-critical super Ornstein-Uhlenbeck process.
−βt −βt Proof. Note that K˚t = e Kt defines a (∆/2,0,e )-Historical superprocess. Let ˆ 1 ˚ PT be the Campbell measure associated with K (note that if T is taken to be a ˆ random stopping time, this measure differs from PT ). The proof of Theorem V.4.1 of [14] with minor modifications shows that (SE)1 has a pathwise unique so- Z0,K˚ ˆ 1 lution. This is because (K3) of Theorem 2.6 of [13] shows that under PT , yt is a Brownian motion stopped at time T and Proposition 2.7 of the same memoir can be used to replace Proposition 2.4 and Remark 2.5 (c) of [14] for the setting where the branching variance depends on time. Once this is established, it is simple to deduce that if (X˚ ,Z) is the solution of (SE)1 , and we let Z0,K˚ Z βt Xt (·) ≡ e X˚t (·) = 1(Zt ∈ ·)Kt (dy)
(X,Z) 1 then is the pathwise unique solution of (SE)Z0,K. The only thing to check R t this is that Zt (ω,y) = Z0(ω,y0) + yt − y0 − γ 0 Zs(ω,y)ds K-a.e., but this follows from the fact that K˚t Kt ,∀t. It can be shown by using by Theorem 2.14 of [13] that X satisfies the following 2 d martingale problem: For φ ∈ Cb(R ),
Z t Z ∆ Z t Mt (φ) ≡ Xt (φ) − X0(φ) − −γx · ∇φ(x) + φ(x)Xs(dx)ds − β Xs(φ)ds 0 2 0
51 R t 2 is a martingale where hM(φ)it = 0 Xs(φ )ds. Then by Theorem II.5.1 of [14] this implies that X is a version of a super Ornstein-Uhlenbeck process, with initial −1 distribution given by K0 Z0 (·) .
Remark 3.3.2. (a) Under the Lipschitz assumptions of Section V.1 of 1 K [14], one can in fact uniquely solve (SE)Z0,K where is a supercritical his- torical Brownian motion. The proof above can be extended with minor mod- ifications.
ˆ (b) The proof of Theorem 3.3.1 essentially shows that under PT , T fixed, the + ˆ d path process y : R × Ω → R such that (t,(ω,y)) 7→ yt is a d-dimensional Brownian motion stopped at T.
ˆ (c) Under PT , T fixed, Zt can be written explicitly as a function of the driving path: For t ≤ T,
Z t Z t γt γs γs e Zt = Z0 + e dZs + Zs (γe )ds (3.3) 0 0 Z t Z t Z t γs γs γ = Z0 + e dys + e (−γZs)ds + Zs(γe s)ds. 0 0 0 Hence, Z t −γt −γ(t−s) Zt (y) = e Z0 + e dys. 0 (X0,Y) 2 We say is the solution to the stochastic equation (SE)Y0,K if:
Z t 2 (a) Y ( ,y) = Y ( ,y) + y − y + Y¯ ( ) −Y ( ,y)ds,K y (SE)Y0,K t ω 0 ω t 0 γ s ω s ω -a.a. 0 Z 0 d (b) Xt (A) = 1(Yt ∈ A)Kt (dy),∀A ∈ B(R ),∀t ≥ 0.
ˆ ∗ 0 where Yt is an (Ft )-predictable process, Xt is an (Ft )-predictable process and
R xX0(dx) K (Y ) ¯ t t t Yt ≡ 0 = , Xt (1) Kt (1) is the center of mass of X0. Intuitively, X0 is an interactive measure-valued diffusion where the represen- tative particles are attracted or repelled by the centre of mass of the process (de-
52 pending on the sign of γ). We will call this the SOU process with attraction (or repulsion) to the COM. It will be shown below that questions regarding (X0,Y) can be naturally reformulated as questions about the ordinary SOU process.
(X,Z) 1 Notation. Unless stated otherwise, will refer to a solution of (SE)Z0,K and (X0,Y) 2 to the solution of (SE)Y0,K. ˆ ∗ d Definition. For an arbitrary Ft -adapted, R -valued process zt , define the centre of mass (COM),z ¯t , with respect to Kt as follows:
Kt (zt ) z¯t = . Kt (1)
We also let
X K Z X0 K Y K ˜ t (·) t ( t ∈ ·) ˜ 0 t (·) t ( t ∈ ·) ˜ t (·) Xt (·) ≡ = , Xt (·) ≡ 0 = and Kt (·) = . Xt (1) Kt (1) Xt (1) Kt (1) Kt (1)
Note that as K is supercritical, Kt survives indefinitely on a set of positive probability S, and goes extinct on the set of positive probability Sc. Hence we can only make sense ofz ¯t for t < η where η is the extinction time. 2 Next we show that there exists a solution to (SE)Y0,K. We cannot adapt the proof of Theorem V.4.1 of [14] to accomplish this as the Lipschitz assumptions on the drift coefficients fail to be satisfied by our model, due to the normalizing K ( ) Y¯ 1 factor t 1 in the definition of t . Instead we will utilize the solution of (SE)Y0,K 2 to define a unique solution for (SE)Y0,K.
There is a pathwise unique solution to (SE)2 . Theorem 3.3.3. Y0,K
Proof. Y 2 ˆ Y Suppose there exists a solution satisfying (SE)Y0,K. Then under PT , t can be written as a function of the driving path y and Y¯ . Using integration by parts gives
Z t Z t γt γs γs e Yt = Y0 + e dYs + γe Ysds 0 0 Z t Z t γs γs = Y0 + e dys + γe Y¯sds 0 0
53 and hence Z t Z t −γt γ(s−t) γ(s−t) Yt = e Y0 + e dys + γe Y¯sds. (3.4) 0 0 (X,Z) 1 Z = Y If is the solution to (SE)Z0,K where 0 0, then note that by Remark 3.3.2 (c), Z t γ(s−t) Yt = Zt + γe Y¯sds. 0
By taking the normalized measure K˜t of both sides of the above equation, we get
Z t −γ(t−s) Y¯t = Z¯t + γ e Y¯sds (3.5) 0
Hence Y¯t is seen to satisfy an integral equation. This is a Volterra Integral Equation of the second kind (see Equation 2.2.1 of [16]) and therefore can be solved pathwise to give Z t Y¯t = Z¯t + γ Z¯sds, 0 ¯ 1 which is easily verified using integration by parts. Also, if Yt is a second process which solves (3.5), then
Z t ¯ ¯ 1 −γ(t−s) ¯ ¯ 1 |Yt −Yt | = γ e (Ys −Ys )ds 0 Z t −γ(t−s) ¯ ¯ 1 ≤ |γ| e |Ys −Ys |ds. 0
¯ ¯ 1 By Gronwall’s inequality, this implies Yt =Yt , for all t and ω. Pathwise uniqueness X0 1 of follows from the uniqueness of the solution to (SE)Z0,K and the uniqueness of the process Y¯t solving (3.5). 2 We have shown that if there exists a solution to (SE)Y0,K, then it is necessarily pathwise unique. Turning now to existence to complete the proof, we work in the opposite order and define Y and X0 as functions of the pathwise unique solution to
54 1 Z = Y (SE)Z0,K where 0 0:
Z t Yt = Zt + γ Z¯sds 0 0 Xt (·) = Kt (Yt ∈ ·).
Then Y¯t satisfies the integral equation (3.5), and hence
Z t Z t −γ(t−s) Z¯sds = e Y¯sds. 0 0 Therefore
Z t −γ(t−s) Yt = Zt + γ e Y¯sds 0 Z t Z t −γt −γ(t−s) −γ(t−s) = e Y0 + e dys + γ e Y¯sds, 0 0 and so Z t Z t γt γs γs e Yt = Y0 + e dys + γ e Y¯sds. 0 0 Multiplying by e−γt and using integration by parts shows
Z t Yt = Y0 + yt − y0 + γ (Y¯s −Ys)ds 0
K y (X0,Y) 2 which holds for -a.e. , thereby showing satisfies (SE)Y0,K.
Remark 3.3.4. Some useful equivalences in the above proof are collected below.
If Y0 = Z0, then for t < η,
Z t (a) Yt = Zt + γ Z¯sds 0 Z t (b) Y¯t = Z¯t + γ Z¯sds 0
(c) Yt −Y¯t = Zt − Z¯t Z t Z t −γ(t−s) (d) e Y¯sds = Z¯sds. 0 0
55 The significance of these equations is that they intimately tie the behaviour of the SOU process with attraction to its center of mass to the SOU process with attraction to the origin. Part (a) says that the SOU process with attraction to the center of mass is the same as the ordinary SOU process pushed by the position of its center of mass.
Definition. Let the process Mt be defined for t ∈ [0,ζ) where ζ ≤ ∞ possibly random. M is called a local martingale on its lifetime if there exist stopping times
TN ↑ ζ such that MTN ∧· is a martingale for all N. The interval [0,ζ) is called the lifetime of M.
0 d We now consider the martingale problem for X . For φ : R → R, let φ¯t ≡ Kt (φ(Yt ))/Kt (1). Note that the lifetime of the process φ¯ is [0,η). Then the follow- ing theorem holds:
2 d Theorem 3.3.5. For φ ∈ Cb(R ,R), and t < η,
Z t φ¯t = φ¯0 + Mt + b¯sds 0 where 1 b = γ∇φ(Y ) · (Y¯ −Y ) + ∆φ(Y ) s s s s 2 s and Mt is a continuous local martingale on its lifetime such that
t Z Z φ(Ys) − φ¯s Mt = dM(s,y) 0 Ks(1) and hence has quadratic variation given by
Z t 2 ¯ 2 φs − (φs) [M]t = ds. 0 Ks(1)
Proof. The proof is not very difficult; one need only use Ito’sˆ Lemma followed by some slight modifications of theorems in Chapter V of [14] to deal with the drift introduced in the historical martingale problem due to the supercritical branching. ˆ Let T be a fixed time and t ≤ T. Recall that under PT , y is a stopped Brownian motion by Remark 3.3.2 (b), and hence Yt (y) is a stopped Ornstein-Uhlenbeck
56 ˆ process (attracting to Y¯t ). Hence, by the classical Ito’sˆ Lemma we have, under PT ,
Z t Z t 1 i j φ(Yt ) = φ(Y0) + ∇φ(Ys) · dYs + ∑ φi j(Ys)d[Y ,Y ]s 0 2 i, j≤d 0 Z t Z t 1 = φ(Y0) + ∇φ(Ys) · dys + γ∇φ(Ys) · (Y¯s −Ys) + ∆φ(Ys)ds. 0 0 2 Then
Z t Z t Kt (φ(Yt )) = Kt (φ(Y0)) + Kt ∇φ(Ys) · dys + Kt bsds 0 0 Z t Z Z t = K0(φ(Y0)) + φ(Y0)dM(s,y) + β Ks (φ(Y0))ds 0 0 Z t Z Z s Z t Z s + ∇φ(Yr) · dyr dM(s,y) + β Ks ∇φ(Yr) · dyr ds 0 0 0 0 Z t Z Z t Z t + bsdM(s,y) + β Ks(bs)ds + Ks(bs)ds 0 0 0 Z t Z = K0(φ(Y0)) + φ(Ys)dM(s,y) 0 Z t Z t + β Ks (φ(Ys))ds + Ks(bs)ds. 0 0