Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
Scaling Limits of StochasticProcesses
Amanda Georgina TurnerSt John’s College and Statistical Laboratory
University of Cambridge
A dissertation submitted for the
degree of Doctor of Philosophy
September 2006
Abstract
In this thesis we analyse two classes of stochastic processes, both of which exhibit
unusual scaling limits.
The first class is of sequences of Markov processes in two dimensions whose fluid
limit is a stable solution of an ordinary differential equation with a saddle fixed
point. In order to investigate the most interesting behaviour of these processes,
we establish a fluid limit which is valid for large times. The limit is shown to be
inherently random and its distribution is obtained. This is then used to derive
surprising scaling limits for the points where these processes hit straight lines
through the origin, and the minimum distance from the origin that the processes
can attain.
We apply the above results to study the accumulation of numerical rounding
errors incurred by some deterministic solvers for systems of ordinary differential
equations. We show that the trajectory of the numerical solution exhibits random-
like behaviour, and calculate the theoretical distribution of the trajectory. Numer-
ical experiments are then performed and the results are fitted to the predicted
distributions with good agreement.
The second class of processes that we consider are stochastic flows on the circle,
constructed by iteratively composing specific maps at times of a Poisson process.
These maps occur naturally in a simplified version of the Hastings-Levitov model
for planar diffusion-limited aggregation on the circle, known as the Eden model.
We define a metric space on which to realize our stochastic flows and show that,
under a specified scaling, they converge to the Brownian web with respect to this
metric.
i
Acknowledgments
This thesis would not have been achievable without the help and support of a large
number of people.
First of all I would like to thank my supervisor, Professor James Norris.
Throughout my PhD he has been generous with his time, never ceasing to provide
me with encouragement and inspiration. It has been a great pleasure to work
under his guidance.
I am privileged to have been part of the Cambridge University Statistical Lab-
oratory. I would like to thank everyone for making it such an enjoyable place to
work, and in particular the secretaries and computer officer for being so accom-
modating. Special thanks go to my long-suffering officemate Teresa Barata, and
to Christina Goldschmidt and Richard Samworth for their continual willingness to
share their time and knowledge. I am also indebted to Christina for proofreading
my thesis and offering many helpful comments.
This work was made possible by a studentship from the Engineering and Phys-
ical Sciences Research Council. I am grateful to the Department of Pure Math-
ematics and Mathematical Statistics and to St John’s College for the financial
support they have given to enable me to travel to conferences, and to St John’s
College for providing a pleasant environment in which to live.
Special mention should go to Danielle, Alan and all the friends I have made over
the years in Cambridge, and to the Cambridge University Canoe Club, for never
failing to provide me with distractions when I needed them, and sometimes when
I didn’t! My parents have also been an endless source of help and encouragement,
from painstakingly reading through drafts of my thesis to pick up the odd missing
bracket, to regularly enquiring if I was ever going to finish!
ii
iii
I would like to acknowledge the anonymous referees at the Annals of Probability
for their helpful comments in the preparation of the paper [32], which have been
incorporated into Chapter 2. The numerical work in Chapter 3 of my thesis was
done in collaboration with Sebastian Mosbach and has formed the basis for a joint
paper [28]. Chapter 4, and in particular Section 4.2, of my thesis contains work
which was done in collaboration with my supervisor, James Norris. It is intended
that this will contribute towards a joint paper in the future.
This dissertation is my own work and contains nothing which is the
outcome of work done in collaboration with others, except where specif-
ically indicated in these acknowledgments and in the text. This disser-
tation has not been submitted in whole or in part for any other degree
or qualification at any other university.
Amanda Turner
Cambridge
September 2006
I would like to add my thanks to my examiners Geoffrey Grimmett and Terry
Lyons for their thorough reading of my thesis and insightful comments.
Amanda Turner
February 2007
Contents
Abstract i
Acknowledgments ii
1 Introduction 1
2 Convergence of Markov processes near saddle fixed points 4
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 The linear case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Linearization of the limit process . . . . . . . . . . . . . . . . . . . 18
2.4 Convergence of the fluctuations . . . . . . . . . . . . . . . . . . . . 24
2.5 A fluid limit for jump Markov processes . . . . . . . . . . . . . . . 27
2.6 Continuous diffusion Markov processes . . . . . . . . . . . . . . . . 32
2.7 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.7.1 Hitting lines through the origin . . . . . . . . . . . . . . . . 36
2.7.2 Minimum distance from the origin . . . . . . . . . . . . . . . 37
3 Accumulation of rounding errors in the numerical solution of
ODEs 42
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
iv
v
3.2 Theoretical background . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2.1 Accumulation of rounding errors . . . . . . . . . . . . . . . . 45
3.2.2 Explicit calculation of the variance . . . . . . . . . . . . . . 48
3.3 Numerical experiments . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3.1 The system . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3.2 Theoretical hitting distribution . . . . . . . . . . . . . . . . 51
3.3.3 Choice of parameters . . . . . . . . . . . . . . . . . . . . . . 52
3.3.4 Results and observations for explicit methods . . . . . . . . 54
3.3.5 Adaptive solvers . . . . . . . . . . . . . . . . . . . . . . . . 57
4 Stochastic flows, planar aggregation and the Brownian web 59
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 A Levy flow on the circle . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2.1 Some generalities for functions on the circle . . . . . . . . . 61
4.2.2 Construction of the flow . . . . . . . . . . . . . . . . . . . . 64
4.2.3 Convergence to the Arratia flow . . . . . . . . . . . . . . . . 65
4.3 Hastings–Levitov DLA . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.4 The Brownian web . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.4.1 A description of the flow space . . . . . . . . . . . . . . . . . 77
4.4.2 Existence and uniqueness of the Brownian web . . . . . . . . 79
4.4.3 Convergence to the Brownian web . . . . . . . . . . . . . . . 80
4.5 Some properties of (D, dD) . . . . . . . . . . . . . . . . . . . . . . . 81
4.6 An equivalent space for the Brownian web . . . . . . . . . . . . . . 89
4.6.1 Compact sets of functions . . . . . . . . . . . . . . . . . . . 91
4.6.2 The isomorphism between the spaces . . . . . . . . . . . . . 92
Bibliography 97
Chapter 1
Introduction
Many fundamental results in probability theory stem from considering the limit of a
random process whose jump sizes tend to zero whilst the jump rate tends to infinity.
The simplest example of this is the law of large numbers for random variables which
states that the average value of a sequence (Xn)n∈N of independent identically
distributed random variables with finite mean µ converges to the deterministic
limit µ. By viewing these random variables as the jump sizes in a random walk,
the value ofX1 + · · ·+XN
N
can be regarded as the position of a random walk at time 1 with jump sizes of
order N−1 and jump rate N . The central limit theorem asserts that if, in addition,
the random variables have finite variance σ2, then
(X1 − µ) + · · ·+ (XN − µ)√N
converges in distribution to an N(0, σ2) random variable. As above, this value
can be regarded as the position of a (mean zero) random walk at time 1 with
jump sizes of order N− 12 and jump rate N . The fluid limit theorem and diffusion
approximation for stochastic processes generalize the law of large numbers and
central limit theorem for random variables by proving the existence of a limit
along the entire trajectory of a process, scaled as above, on a compact time-set.
Scaling limits of stochastic processes arise in a variety of contexts. Where the
1
Chapter 1. Introduction 2
random objects originate from geometrical or physical settings, the jump sizes can
be proportional to lattice spacing or particle sizes, or inversely proportional to the
number of particles. As in the case of the law of large numbers and central limit
theorem, limit theorems are often obtained by scaling time by the order of the
jump sizes or the square root of the order of the jump sizes. However, in some
situations limit results can be proved by scaling by more unusual powers of the
jump size, and are generally accompanied by interesting behaviour. In this thesis
we analyse two classes of processes both of which exhibit unexpected scaling limits.
The first class is of sequences (XNt )t>0 of Markov processes in two dimensions
whose fluid limit is a stable solution of an ordinary differential equation of the
form xt = b(xt), where
b(x) =(−µ 0
0 λ
)
x+ τ(x)
for some λ, µ > 0 and τ(x) = O(|x|2). Here the processes are indexed so that the
variance of the fluctuations of XNt is inversely proportional to N . The simplest
example arises from the OK Corral gunfight model which was formulated in 1998 by
Williams and McIlroy [34] and studied by Kingman [25] in 1999. These processes
exhibit their most interesting behaviour at times of order logN , so it is necessary
to establish a fluid limit that is valid for large times. We find that this limit is
inherently random and obtain its distribution. Using this, it is possible to derive
scaling limits for the points where these processes hit straight lines through the
origin, and the minimum distance from the origin that the processes can attain.
The power of N that gives the appropriate scaling is surprising. For example, if T
is the time that XNt first hits one of the lines y = x or y = −x, then
Nµ
2(λ+µ) |XNT | ⇒ |Z|
µλ+µ ,
for some zero mean Gaussian random variable Z.
Numerical rounding errors incurred by some deterministic solvers for systems of
ordinary differential equations can be modelled as a special case of these processes.
The above results can then be applied to acquire theoretical predictions about the
accumulation of these rounding errors. It is shown that the trajectory of the nu-
merical solution exhibits random-like behaviour, and the theoretical distribution
of the trajectory is obtained as a function of time, the step size and the numer-
Chapter 1. Introduction 3
ical precision of the computer. By performing multiple repetitions with different
values of the time step size, the random distributions predicted theoretically can
be observed numerically. We mainly focus on the explicit Euler and fourth order
Runge-Kutta (RK4) methods, but also briefly consider more complex algorithms
such as the implicit solvers VODE [5] and RADAU5 [14].
The second class of processes is motivated by Hastings-Levitov diffusion-limited
aggregation (DLA). DLA is a random growth model which was originally intro-
duced in 1981 by Witten and Sander [35]. In this model, particles perform Brow-
nian motions in the plane until they collide with a cluster at the origin, at which
point they stick to the cluster. In 1998 Hastings and Levitov [17] formulated a
model of DLA in which the cluster is represented by a sequence of iterated con-
formal maps. We construct a family of stochastic flows on the circle by iteratively
applying small localized perturbations to the circle at uniformly distributed points
and show that a case of simplified Hastings-Levitov DLA, where the incoming par-
ticles are slits of length N−1 sticking to the unit disc, falls under this scheme. This
model is known as the Eden model [8], and describes the growth of bacterial cells
or tissue cultures of cells that are constrained from moving. If time is scaled in
such a way that particles arrive as a Poisson process of rate proportional to N 3, the
resulting flow map (restricted to points on the unit circle) converges to a random
object known as the Brownian web. This object can be defined loosely as a family
of coalescing Brownian motions starting at all possible points in continuous space-
time. It was first studied in 1979 by Arratia [1] as a limit for discrete coalescing
random walks. Once again, the power of N by which time is scaled is curious and
may give rise to the fractal behaviour which can be observed in simulations of the
model.
The Markov processes near saddle points are investigated in Chapter 2, and
our results are applied to rounding errors in Chapter 3. Chapter 4 concerns the
stochastic flows arising from Hastings-Levitov DLA and shows that they converge
to the Brownian web.
Chapter 2
Convergence of Markov processes
near saddle fixed points
2.1 Introduction
The fluid limit theorem is a powerful result which shows that, under certain condi-
tions, sequences of Markov processes converge to solutions of ordinary differential
equations. We are interested in situations where the differential equation can be
written in the form
xt = Bxt + τ(xt), (2.1)
for some matrix B, where τ(x) = O(|x|2) is twice continuously differentiable. These
differential equations have been studied extensively in the dynamical systems lit-
erature, with the aim of finding precise relationships between their solutions and
solutions of the corresponding linear differential equations
yt = Byt. (2.2)
We restrict ourselves to the two dimensional case where the origin is a saddle
fixed point of the system i.e. B has eigenvalues λ,−µ, with λ, µ > 0. The phase
portrait of (2.1) in the neighbourhood of the origin is shown in Figure 2.1.
In particular, there exists some x0 6= 0 such that φt(x0) → 0 as t → ∞, where
4
Chapter 2. Convergence near saddle points 5
0
Uns
tabl
e M
anif
old
Stable Manifold Stable Manifold
Uns
tabl
e M
anif
old
Figure 2.1: The phase portrait of an ordinary differential equation having a saddlefixed point at the origin.
φ is the flow associated with the ordinary differential equation (2.1). The set of
such x0 is the stable manifold. There also exists some x∞ such that φ−1t (x∞) → 0
as t → ∞. The set of such x∞ is the unstable manifold. The saddle point case is
interesting in this setting as it is the only case in two dimensions where there is
both a stable and an unstable manifold.
Fix an x0 in the stable manifold and consider sequences of Markov processes
with initial condition XN0 = x0, where the processes are indexed so that the vari-
ance of the fluctuations of XNt is inversely proportional to N . The fluid limit
theorem tells us that for fixed values of t, XNt → φt(x0) as N → ∞. However, if
we allow the value of t to grow with N as N → ∞, we shall see that XNt deviates
from the stable solution to a limit which is inherently random, before converging
to an unstable solution (see Figure 2.2).
More precisely, we observe three different types of behaviour depending on the
time scale:
A. On compact time intervals [0, R], XNt converges to the stable solution of
(2.1), the fluctuations around this limit being of order N− 12 .
Chapter 2. Convergence near saddle points 6
0x
Markov Process
Stable Solution
Unstable Solution
A
C
B
0
Figure 2.2: Diagram showing how the Markov process XNt deviates from the stable
solution φt(x0) for large values of t.
B. There exists some x0 6= 0, depending only on x0, and a Gaussian random
variable Z∞ such that if t lies in the interval [R, 12λ
logN −R], then
XNt = x0e
−µt(e1 + ε1) +N− 12Z∞e
λt(e2 + ε2)
for some εi(t, N) → 0 uniformly in t in probability as R,N → ∞, where
e1, e2 is the standard basis for R2. In other words, XNt can be approximated
by the solution to the linear ordinary differential equation (2.2) starting from
the random point(
x0
N−12 Z∞
)
.
C. On time intervals of a fixed length around 12λ
logN , XNt converges to the
unstable solution of (2.1).
The most interesting behaviour occurs on time intervals of fixed lengths around1
2(λ+µ)logN , as for these values of t the two terms x0e
−µt and N− 12Z∞e
λt are of
the same order. By considering
x0e−µte1 +N− 1
2Z∞eλte2,
Chapter 2. Convergence near saddle points 7
we show in Section 2.7 that it is at these times that XNt crosses all the straight
lines passing through 0, and also that |XNt | attains its minimum value when t
is in this range. The distance from the origin of XNt for these values of t is of
order N− µ2(λ+µ) , which gives us surprising scaling limits for the points at which XN
t
intersects various straight lines, and for inf |XNt |.
In order to study the Markov processes at times of order logN , it is necessary to
establish a strong form of the fluid limit theorem that is valid for large times. The
key idea is to show that for N and t0 sufficiently large, the process (XNt )t>t0 is close
to (φt−t0(XNt0
))t>t0 . This is done in Section 2.2 in the case when (2.1) is linear and
XNt is a pure jump Markov process, in Section 2.5 for pure jump Markov processes
where (2.1) is non linear, and in Section 2.6 for continuous diffusion processes. In
Sections 2.3 and 2.4 we look at the process (φt−t0(XNt0 ))t>t0 for large values of N
and t0, which then enables us to obtain scaling limits for the process XNt . The same
idea can be used to obtain fluid limit theorems for arbitrary matrices B in (2.1)
e.g. with eigenvalues having the same sign, or in higher dimensions. However, an
analysis of the solutions of the underlying differential equation is required, which
we do not go into here.
The simplest example of this type of behaviour arises from the OK Corral
gunfight model which was formulated by Williams and McIlroy [34] and studied
by Kingman [25] and Kingman and Volkov [26]. Two lines of gunmen face each
other, there initially being N on each side. Each gunman fires lethal gunshots at
times of a Poisson process with rate 1 until either there is no one left on the other
side or he is killed. The process terminates when all the gunmen on one side are
dead. It is shown by Kingman that if SN is the number of survivors when the
process terminates, then
N− 34SN ⇒ 2
34 |Z| 12 ,
where Z ∼ N(0, 13). It is the occurrence of the unexpected power of N that
interested the above authors in the problem. By using our scaling limits we re-
derive this result in Section 2.2.1 and show that it is a special case of a much more
general phenomenon, and that in fact by a suitable choice of B, every number in
the interval (12, 1) may be obtained as a power of N in this way. An application of
the nonlinear case to a model of two competing species is given in Section 2.7.
Chapter 2. Convergence near saddle points 8
2.2 The linear case
In this section we restrict ourselves to sequences of Markov processes in the special
case where equation (2.1) is linear. We begin by describing the conditions under
which a limit theorem exists for large times and then establish the exact limit
by means of an appropriate martingale inequality. In Section 2.2.1 this result is
used to derive scaling limits for the points where these processes hit straight lines
through the origin and we use this to obtain a solution to the OK Corral problem.
The fluid limit theorem that we state below is widely known and has been the
subject of many works. We use the formulation found in Darling and Norris [6].
Let (XNt )t>0 be a sequence of pure jump Markov processes, starting from x0
and taking values in some subsets IN of R2, with Levy kernels KN(x, dy). Let S
be an open subset of R2 with x0 ∈ S, and set SN = IN ∩ S. For x ∈ SN and
θ ∈ (R2)∗, define the Laplace transform corresponding to Levy kernel KN(x, dy)
by
mN (x, θ) =
∫
R2
e〈θ,y〉KN(x, dy).
We assume that there is a limit kernel K(x, dy) defined for x ∈ S, with corre-
sponding Laplace transform m(x, θ), with the following properties.
(a) There exists a constant η0 > 0 such that m(x, θ) is uniformly bounded for
all x ∈ S and |θ| 6 η0.
(b) As N → ∞,
supx∈SN
sup|θ|6η0
∣
∣
∣
∣
mN(x,Nθ)
N−m(x, θ)
∣
∣
∣
∣
→ 0.
Set b(x) = m′(x, 0) where ′ denotes differentiation in θ. Suppose that b is Lipschitz
on S so that b has an extension to a Lipschitz vector field b on R2. Then there is
a unique solution (xt)t>0 to the ordinary differential equation xt = b(xt) starting
from x0. Suppose that S contains a neighbourhood of the path (xt)t>0. By stopping
XNt at the first time it leaves S if necessary, we may assume that XN
t remains in
S for all t > 0. Under these assumptions, for all t0 > 0 and δ > 0,
lim supN→∞
N−1 log P(supt6t0
|XNt − xt| > δ) < 0.
Chapter 2. Convergence near saddle points 9
Suppose additionally that
(c) b is C1 on S and
supx∈SN
N12 |bN(x) − b(x)| → 0,
where bN (x) = mN ′(x, 0).
(d) a, defined by a(x) = m′′(x, 0), is Lipschitz on S.
It follows from the above that for any η < η0 there exists a constant A such that
supx∈SN
sup|θ|6η
N |mN ′′(x,Nθ)| 6 A, (2.3)
where | · | is the operator norm.
Let γNt = N
12
(
XNt − xt
)
. Then for any t > 0, γNt ⇒ γt as N → ∞, where
(γt)t>0 is the unique solution to the linear stochastic differential equation
dγt = σ(xt)dWt + ∇b(xt)γtdt (2.4)
starting from 0, W a Brownian motion in R2, and σ ∈ R2 ⊗ (R2)∗ satisfying
σ(x)σ(x)∗ = a(x). The distribution of (γt)t>0 does not depend on the choice of σ.
We are interested in the case where b(x) = Bx for some matrix B =(−µ 0
0 λ
)
,
µ, λ > 0.
Let φt(x) be the solution to the ordinary differential equation
φt(x) = b (φt(x)) , φ0(x) = x. (2.5)
In the linear case we can solve (2.5) explicitly to get φt(x) = eBtx. We concen-
trate on processes where the initial condition is chosen to be x0 = (x0,1, 0) with
x0,1 6= 0, so that xt = φt(x0) → 0 as t → ∞. We shall show that for sufficiently
large values of N and t0, XNt is in some sense close to φt−t0(X
Nt0
) for t > t0.
Introduce random measures µN and νN on (0,∞) × R2, given by
µN =∑
∆XNt 6=0
δ(t,∆XNt ),
Chapter 2. Convergence near saddle points 10
νN (dt, dy) = KN (XNt−, dy)dt,
where δ(t,y) denotes the unit mass at (t, y) and ∆XNt = XN
t −XNt−.
Let f(t, x) = e−Bt(
x− φt−t0(XNt0
))
, for t > t0. By Ito’s formula,
f(t, XNt ) = f(t0, X
Nt0
) +MB,Nt −MB,N
t0 +
∫ t
t0
(
∂f
∂t+KNf
)
(s,XNs−)ds,
where∂f
∂t= −Be−Btx,
KNf(s, x) =
∫
R2
(f(s, x+ y) − f(s, x))KN(x, dy)
=
∫
R2
e−BsyKN(x, dy)
= e−BsbN(x),
and
MB,Nt =
∫
(0,t]×R2
(
f(s,XNs− + y) − f(s,XN
s−))
(µN − νN)(ds, dy)
=
∫
(0,t]×R2
e−Bsy(µN − νN )(ds, dy).
So if t > t0, then
e−Bt(XNt − φt−t0(X
Nt0
)) = MB,Nt −MB,N
t0 +
∫ t
t0
e−Bs(bN(XNs−) − b(XN
s−))ds. (2.6)
Lemma 2.1. There exists some constant C such that
E
(
supt>t0
e−λt|eBt(MB,Nt −MB,N
t0 )|)
6 CN− 12 e−λt0 .
Chapter 2. Convergence near saddle points 11
Proof. By the product rule,
e(B−λI)t(MB,Nt −MB,N
t0 ) =
∫ t
t0
(B − λI)e(B−λI)s(MB,Ns −MB,N
t0 )ds
+
∫ t
t0
∫
R2
e−λsy(µN − νN)(dy, ds)
and hence,
E
(
supt>t0
e−λt|eBt(MB,Nt −MB,N
t0 )|)
6 E
(
supt>t0
∫ t
t0
(λ+ µ)e−(λ+µ)s|(MB,Ns −MB,N
t0 )1|ds)
+ E
(
supt>t0
∣
∣
∣
∣
∫ t
t0
∫
R2
e−λsy(µN − νN )(dy, ds)
∣
∣
∣
∣
)
6
∫ ∞
t0
(λ+ µ)e−(λ+µ)s(
E(MB,Ns −MB,N
t0 )21
)12ds
+ E
(
supt>t0
∣
∣
∣
∣
∫ t
t0
∫
R2
e−λsy(µN − νN)(dy, ds)
∣
∣
∣
∣
2)
12
.
Since
E
∫ t
0
∫
R2
|e−λsy|νN(dy, ds) <∞
for all t > 0, the process
(∫ t
0
∫
R2
e−λsy(µN − νN)(dy, ds)
)
t>0
is a martingale, and hence, by Doob’s L2 inequality
E
(
supt>t0
∣
∣
∣
∣
∫ t
t0
∫
R2
e−λsy(µN − νN )(dy, ds)
∣
∣
∣
∣
2)
6 4 supt>t0
E
(
∣
∣
∣
∣
∫ t
t0
∫
R2
e−λsy(µN − νN)(dy, ds)
∣
∣
∣
∣
2)
.
Chapter 2. Convergence near saddle points 12
Now
E
(
(MB,Nt −MB,N
t0 )21
)
= E
∫ t
t0
∫
R2
e2µsy21ν
N (dy, ds)
6 E
∫ t
t0
e2µs|mN ′′(XNs−, 0)|ds
6e2µtA
2µN,
where A is defined in (2.3). Similarly
E
(
∣
∣
∣
∣
∫ t
t0
∫
R2
e−λsy(µN − νN )(dy, ds)
∣
∣
∣
∣
2)
6e−2λt0A
2λN.
Hence,
E
(
supt>t0
e−λt|eBt(MB,Nt −MB,N
t0 )|)
6
∫ ∞
t0
(λ+ µ)e−λs
(
A
2µN
) 12
ds+ e−λt0
(
2A
λN
) 12
6A
12 (λ+ µ+ 2(λµ)
12 )
λ(2µ)12
N− 12 e−λt0 .
Theorem 2.2. For all ε > 0,
limt0→∞
lim supN→∞
P
(
supt>t0
e−λt|XNt − φt−t0(X
Nt0 )| > N− 1
2 ε
)
= 0.
Proof. Let N0 be sufficiently large that supN>N0N
12 ‖bN − b‖ < λε/2, where ‖bN −
b‖ = supx∈SN |bN(x) − b(x)|, and set
ΩN,t0 =
supt>t0
e−λt|eBt(MB,Nt −MB,N
t0 )| 6 N− 12ε
2
.
Chapter 2. Convergence near saddle points 13
By (2.6), on the set ΩN,t0 with N > N0,
supt>t0
e−λt∣
∣XNt − φt−t0(X
Nt0
)∣
∣ 6 supt>t0
∣
∣
∣e−λteBt(MB,N
t −MB,Nt0 )
∣
∣
∣
+ supt>t0
e−λt
∫ t
t0
|eB(t−s)| ‖bN − b‖ds
6 N− 12 ε.
Hence,
lim supN→∞
P
(
supt>t0
e−λt|XNt − φt−t0(X
Nt0 )| > N− 1
2 ε
)
6 lim supN→∞
P(ΩcN,t0)
62Ce−λt0
ε→ 0
as t0 → ∞, where the second inequality follows by Markov’s inequality and Lemma
2.1.
Let Z∞ ∼ N(0, σ2∞), where
σ2∞ =
∫ ∞
0
e−2λsa(xs)2,2ds.
Theorem 2.3. The following converge in probability as N → ∞.
(i)
supt6tN
|eµtXNt,1 − x0,1| → 0
for any sequence tN → ∞ with e(λ+µ)tN = O(N12 );
(ii)
supt>tN
N12 e−λt|XN
t,1| → 0
for any sequence tN with e(λ+µ)tN = ω(N12 );
(iii)
supt1,t2>tN
N12 |e−λt1XN
t1,2 − e−λt2XNt2,2| → 0
Chapter 2. Convergence near saddle points 14
for any sequence tN → ∞.
Furthermore, if σ∞ 6= 0, then
N12 e−λtXN
t,2 ⇒ Z∞
as t, N → ∞.
Remark 2.4. Given any sequence of times tN → ∞ as N → ∞, by the Skorohod
Representation Theorem, it is possible to choose a sample space in which ZN∞ =
N12 e−λtNXN
tN ,2 → Z∞ almost surely as N → ∞. In this case the above result can
be expressed as
XNt = x0,1e
−µt(e1 + ε1) +N− 12Z∞e
λt(e2 + ε2) (2.7)
where εi = εi(N, t) → 0, uniformly in t, in probability as N → ∞.
Proof. For any fixed t0, supt6t0 |eµtXNt,1 − x0,1| → 0 in probability as an immediate
consequence of the fluid limit theorem. For (i), it is therefore sufficient to show
that for any ε > 0, limt0→∞ lim supN→∞ P(supt06tN|eµtXN
t,1 − x0,1| > ε) = 0. Now
if t > t0, then φt−t0(XNt0 ) = eB(t−t0)XN
t0 = eB(t−t0)(xt0 +N− 12γN
t0 ). Since x0 = x0,1e1,
we have that eB(t−t0)xt0 = e−µtx0. Hence,
φt−t0(XNt0
) = e−µt(x0 +N− 12 eµt0γN
t0,1e1) +N− 12 eλte−λt0γN
t0,2e2,
and so
XNt,1 = e−µt
(
x0,1 +N− 12 eµt0γN
t0,1 + e(λ+µ)te−λt(XNt − φt−t0(X
Nt0
))1
)
and
XNt,2 = N− 1
2 eλt(
e−λt0γNt0,2 +N
12 e−λt(XN
t − φt−t0(XNt0 ))2
)
. (2.8)
Let N → ∞ and then t0 → ∞. Statements (i)-(iii) follow by Theorem 2.2 and the
fact that γNt0
⇒ γt0, a Gaussian random variable.
For the last part, note that by (2.4),
e−λt0γNt0,2 ⇒ e−λt0γt0,2 =
∫ t0
0
e−λs〈e2, σ(xs)dWs〉
Chapter 2. Convergence near saddle points 15
as N → ∞. Since
∫ ∞
0
|e−λse∗2σ(xs)|2ds 6
∫ ∞
0
e−2λs|a(xs)|ds 6A
2λ,
where e∗i is the transpose of ei, e−λt0γt0,2 → Z∞ almost surely as t0 → ∞, for
Z∞ =
(∫ ∞
0
e−λtσ(xt)dWt
)
2
∼ N(0, σ2∞).
The result follows by (2.8) and Theorem 2.2.
2.2.1 Applications
Applications will be dealt with more fully in Section 2.7. However, we illustrate
here how the above result can be used to study the first time that XNt hits lθ or l−θ,
the straight lines passing through the origin at angles θ and −θ, where θ ∈ (0, π2),
as N → ∞. As XNt is not continuous, we define the time that XN
t first crosses
one of the lines l±θ as
TNθ = inf
t > 0 :∣
∣
∣
XNt−,2
XNt−,1
∣
∣
∣6 | tan θ| and
∣
∣
∣
XNt,2
XNt,1
∣
∣
∣> | tan θ|
.
Let
tN =1
2(λ+ µ)logN,
and
cθ =1
λ+ µlog
∣
∣
∣
∣
x0,1 tan θ
Z∞
∣
∣
∣
∣
.
Theorem 2.5. Under the assumptions listed at the beginning of Section 2.2,
TNθ − tN ⇒ cθ (2.9)
and
Nµ
2(λ+µ) |XNT N
θ| ⇒ | sec θ|| tan θ|−
µλ+µ |x0|
λλ+µ |Z∞|
µλ+µ (2.10)
as N → ∞.
Proof. For simplicity, we work in a sample space in which ZN∞ → Z∞ almost surely.
Chapter 2. Convergence near saddle points 16
Define εi as in Remark 2.4. By observing that
x0,1e−µte1 +N− 1
2Z∞eλte2
first intersects one of the lines l±θ at time t = tN + cθ, given any ε > 0,
P(
TNθ 6 tN + cθ − ε
)
6 P
(
supt6tN +cθ−ε
∣
∣
∣
∣
∣
XNt,2
XNt,1
∣
∣
∣
∣
∣
> | tan θ|)
= P
(
supt6tN +cθ−ε
∣
∣
∣
∣
∣
x0,1e−µtε1,2 +N− 1
2Z∞eλt(1 + ε2,2)
x0,1e−µt(1 + ε1,1) +N− 12Z∞eλtε2,1
∣
∣
∣
∣
∣
> | tan θ|)
→ 0
as N → ∞, where εi,j is the jth coordinate of εi. Similarly,
P(
TNθ > tN + cθ + ε
)
6 P
(
inft>tN +cθ+ε
∣
∣
∣
∣
∣
XNt,2
XNt,1
∣
∣
∣
∣
∣
6 | tan θ|)
→ 0.
The result follows immediately.
Remark 2.6. The sign of Z∞ determines whether XNt hits lθ or l−θ at time TN
θ .
Since Z∞ is a Gaussian random variable with mean 0, each event occurs with
probability 12.
Example 2.7 (The OK Corral Problem). The OK Corral process is a Z2-
valued process (UNt , V
Nt ) used to model the famous gunfight. Here UN
t and V Nt
are the number of gunmen on each side and UN0 = V N
0 = N . Each gunman fires
lethal gunshots at times of a Poisson process with rate 1 until either there is no-one
left on the other side or he is killed. The transition rates are
(u, v) →
(u− 1, v) at rate v
(u, v − 1) at rate u
until uv = 0.
The process terminates when all the gunmen on one side are dead. We are
interested in the number of gunmen surviving when the process terminates, for
Chapter 2. Convergence near saddle points 17
large values of N .
This model was formulated by Williams and McIlroy [34] and later studied by
Kingman [25] and subsequently Kingman and Volkov [26].
Let XNt =
(
UNt , V
Nt
)
/N . This gives a sequence of pure jump Markov processes,
starting from x0 = (1, 1), with Levy kernels
KN (x, dy) = Nx2δ(−1/N,0) +Nx1δ(0,−1/N).
If we let
K(x, dy) = x2δ(−1,0) + x1δ(0,−1),
then
m(x, θ) = x2e−θ1 + x1e
−θ2 =mN (x,Nθ)
N,
b(x) =(
0 −1−1 0
)
x = bN(x),
and
a(x) =(
x2 00 x1
)
.
So, under a rotation by π4, the conditions required for Theorem 2.5 are satisfied,
with λ = µ = 1. In the original coordinates, the process terminates when XNt
hits the x or y axes. Under the rotation, this corresponds to hitting l±π4. Hence,
if the OK Corral process terminates at time TN and there are SN survivors, then
TN = TNπ/4 and SN = N |XN
T Nπ/4
|, and so
TN − 1
4logN ⇒ 1
4log 2 − 1
2log |Z∞|
and
N− 34SN ⇒ 2
34 |Z∞| 12 ,
where Z∞ ∼ N(0, 13). The limiting distribution of N− 3
4SN is the one obtained by
Kingman in [25].
Remark 2.8. It is remarked by Kingman [25] that it is the occurrence of the sur-
prising power of N that makes the OK Corral process of interest. Theorem 2.5
shows that this is a special case of a more general phenomenon and, in fact, by
a suitable choice of λµ, every number in the interval ( 1
2, 1) may be obtained as a
Chapter 2. Convergence near saddle points 18
power of N in this way.
2.3 Linearization of the limit process
We now turn to the general case where b(x) = Bx+τ(x) for B =(−µ 0
0 λ
)
, µ, λ > 0,
and τ : R2 → R2 twice continuously differentiable, with τ(0) = ∇τ(0) = 0. Let
φt(x) be the solution to the ordinary differential equation
φt(x) = b (φt(x)) , φ0(x) = x. (2.11)
This section consists of a technical calculation which expresses φt(x) in a linear
form.
We are interested in the behaviour of solutions starting near the stable mani-
fold. Lemma 2.10 proves the existence of the stable manifold and establishes the
limiting behaviour of a stable solution. First order behaviour is investigated in
Lemma 2.11, and these results are then used in Theorem 2.12 to express solutions
near the stable manifold in the required linear form. Theorem 2.13 shows that
over large time periods, solutions starting near the stable manifold approach the
unstable manifold.
Throughout this section we use the following classical planar linearization the-
orem due to Hartman [16].
Theorem 2.9. There exists a C1 diffeomorphism h : U → V = h(U), defined on
an open neighbourhood U of the origin, with uniformly Holder continuous partial
derivatives and having the form h(x) = x+ o(x) such that
h(φt(x)) = eBth(x)
for all (t, x) with φt(x) ∈ U .
Pick 0 < δ < 1 sufficiently small that the ball of radius δ centered at the origin
is contained in U ∩ V . Since h−1(x) = x + o(x), and ∇h(x) = I + o(1) we can
Chapter 2. Convergence near saddle points 19
further ensure that δ is sufficiently small that
sup0<|x|<δ
(
|h(x)/x| ∨∣
∣h−1(x)/x∣
∣
)
< 2
and
sup|x|<δ
(|∇h(x) − I| ∨ |∇h−1(x) − I|) < 1/2.
Lemma 2.10. There exists an x0 with 0 < |x0| < δ/8 such that φt(x0) → 0 as
t → ∞. Furthermore, for any such x0, there exists some x0 with 0 < |x0| < δ/4
such that
eµtφt(x0) → ( x00 )
as t→ ∞, and
|φt(x0)| 6 2|x0|e−µt < δe−µt/2
for all t > 0.
Proof. Pick some x0 ∈ R with 0 < |x0| < δ/16 and define x0 = h−1(x0, 0). Then
0 < |x0| 6 sup0<|x|<δ
|h−1(x)/x||x0| <δ
8,
and
φt(x0) = h−1(
eBt ( x00 ))
= h−1(
e−µtx00
)
→ 0
as t→ ∞.
Conversely, given x0 satisfying the above conditions, define x0 = h(x0)1. Note
that because of the form of h(x), x0 has the same sign as x0,1. Since eBth(x0) =
h(φt(x0)) → 0 as t→ ∞, h(x0)2 = 0, and so
0 < |x0| = |h(x0)| 6 2|x0| < δ/4.
Also
eµtφt(x0) = eµth−1(
eBt ( x00 ))
= eµt((
e−µtx00
)
+ o(e−µtx0))
→ ( x00 )
Chapter 2. Convergence near saddle points 20
as t→ ∞, and
|φt(x0)| =∣
∣h−1(
eBt ( x00 ))∣
∣ =∣
∣h−1(
e−µtx00
)∣
∣ 6 2|x0|e−µt <δ
2e−µt
for all t > 0.
Lemma 2.11. (i) There exists some D0 ∈ (R2)∗ \ 0, where 0 = (0 0), such
that
e−λt∇φt(x0) →(
0D0
)
as t→ ∞.
(ii) If |x| < δ and |φt(x)| < δ/2, then |∇φt(x)| < 4eλt.
(iii) If |x| + |y| < δ and sup06θ61 |φt(x + θy)| < δ/2, then there exist constants
K ∈ R and 0 < α 6 1 such that
|∇φt(x+ y) −∇φt(x)| 6 Keλt(1+α)|y|α.
Proof. (i) Let D0 = ∇h2(x0) ∈ (R2)∗ \ 0. Then
e−λt∇φt(x0) = ∇h−1(
e−µtx00
)
e(B−λI)t∇h(x0)
→ ( 0 00 1 )∇h(x0)
=(
0D0
)
as t→ ∞.
(ii) If |φt(x)| < δ/2, then |eBth(x)| = |h(φt(x))| < δ and so
|∇φt(x)| = |∇h−1(eBth(x))eBt∇h(x)|6 sup
|y|<δ
|∇h−1(y)| sup|y|<δ
|∇h(y)|eλt
< 4eλt.
(iii) Since h and h−1 have uniformly Holder continuous partial derivatives, there
exists some K0 ∈ R and 0 < α < 1 such that
|∇h(w) −∇h(z)| 6 K0|w − z|α
Chapter 2. Convergence near saddle points 21
and
|∇h−1(w) −∇h−1(z)| 6 K0|w − z|α.
Therefore
|∇φt(x+ y) −∇φt(x)|=
∣
∣∇h−1(eBth(x+ y))eBt∇h(x+ y) −∇h−1(eBth(x))eBt∇h(x)∣
∣
6∣
∣∇h−1(eBth(x+ y))eBt (∇h(x + y) −∇h(x))∣
∣
+∣
∣
(
∇h−1(eBth(x + y)) −∇h−1(eBth(x)))
eBt∇h(x)∣
∣
6 2eλtK0|y|α + 2eλtK0|eλt(h(x + y) − h(x))|α
6 8K0eλt(1+α)|y|α.
Suppose that z ∈ R2, with 0 < |z| < 1, and xz = x0 + z.
Theorem 2.12. Fix C and consider the limit z → 0 with∣
∣
∣
zD0z
∣
∣
∣< C, where D0
is defined in Lemma 2.11. There exist wi, i = 1, 2 (not necessarily unique) with
wi(t, z) → 0 uniformly in t ∈ [R,− 1λ
log |z| −R] as z → 0 and R→ ∞ such that
φt(xz) = x0e−µt(e1 + w1) +D0ze
λt(e2 + w2).
Proof. Suppose that R > 1λ
log 8δ−4|x0| . If |x− x0| 6 |z| and
0 6 t 6
(
inf|x−x0|6|z|
inf
s > 0 : |φs(x)| >δ
2
)
∧(
−1
λlog |z| − R
)
,
then
|φt(x)| 6 |φt(x0)| + |φt(x) − φt(x0)|6 2|x0|e−µt + |∇φt(x0 + θ′(x− x0))| |x− x0|6 2|x0|e−µt + 4|z|eλt
6 2|x0| + 4e−λR
<δ
2
Chapter 2. Convergence near saddle points 22
where θ′ ∈ (0, 1). Hence, |φt(x)| < δ/2 for all |x−x0| 6 |z| and t 6 − 1λ
log |z|−R.
Now
φt(xz) = φt(x0) + ∇φt(x0)z + (∇φt(x0 + θz) −∇φt(x0)) z
for some θ ∈ (0, 1) and so, defining
w1(t, z) = x−10
(
eµtφt(x0) − x0e1)
and
w2(t, z) = (D0z)−1(
e−λt∇φt(x0)z −D0ze2 + e−λt(∇φt(x0 + θz) −∇φt(x0))z)
,
we have
φt(xz) = x0e−µt(e1 + w1) +D0ze
λt(e2 + w2).
Then |w1| → 0 uniformly in t > R as R → ∞ by Lemma 2.10, and
|w2| 6|z|
|D0z|(∣
∣e−λt∇φt(x0) −(
0D0
)∣
∣+Keλαt|z|α)
6 C(∣
∣e−λt∇φt(x0) −(
0D0
)∣
∣ +Ke−λαR)
→ 0
uniformly in t ∈ [R,− 1λ
log |z| −R] as R → ∞ and z → 0, by Lemma 2.11.
Since φ−1t (x) satisfies (2.11) with b replaced by −b, we may apply Lemma
2.10 and Lemma 2.11 to deduce the existence of x∞ with 0 < |x∞| < δ/8 such
that eλtφ−1t (x∞) → ( 0
x∞) for some x∞ ∈ R as t → ∞, and D∞ such that
e−µt∇φ−1t (x∞) →
(
D∞
0
)
as t → ∞. Suppose that as z → 0, the sign of D0z
is eventually constant and non-zero. As x∞ has the same sign as x∞,2 (see the
proof of Lemma 2.10), we may choose x∞ such that D0zx∞
> 0.
There exists some t∞ > 0 such that φt(x0) does not intersect the line l(r) =
x∞ + rD∗∞ for any t > t∞. Let
sz = inft > t∞ : φt(xz) = x∞ + rD∗∞ for some r ∈ R.
Chapter 2. Convergence near saddle points 23
Theorem 2.13. Fix C > 0 and consider the limit z → 0 with∣
∣
∣
zD0z
∣
∣
∣6 C. Then
sz −1
λlog
x∞D0z
→ 0
and(
x∞D0z
)µλ
(φsz(xz) − x∞) → x0D∗
∞|D∞|2
as z → 0.
Proof. We shall prove this theorem in the case where for z sufficiently small D0z,
x0 > 0. The other cases are similar.
Sinceφt(x0)2
φt(x0)1=eµtφt(x0)2
eµtφt(x0)1→ 0
x0= 0
as t→ ∞, there exists some T > 0 such that∣
∣
∣
φt(x0)2φt(x0)1
∣
∣
∣< 1 for all t > T . Let
tz = inft > T : |φt(xz)1| = |φt(xz)2|.
By expressing φt(xz) in the form derived in Theorem 2.12, we may use a similar
argument to that in Theorem 2.5 to show
tz −1
λ+ µlog
x0
D0z→ 0
as z → 0. Let f : B(0, 1) → R be defined by f(z) = φtz(xz)1. Again as in Theorem
2.5,
(D0z)− µ
λ+µf(z) → xλ
λ+µ
0
as z → 0.
Define g : R+ → R by
g(y) = φ−1t′y
(
x∞ + y D∗
∞
|D∞|2
)
1,
where t′y is defined in the same way as tz except for φ−1 instead of φ. (The
scaling factor of |D∞|2 is chosen so that D∞
(
y D∗
∞
|D∞|2
)
= y). Note that φsz(xz) =
x∞ + g−1(f(z)) D∗
∞
|D∞|2 .
Chapter 2. Convergence near saddle points 24
By a similar argument to above, y−λ
λ+µg(y) → xµ
λ+µ∞ as y → 0. But then
∣
∣
∣
∣
∣
(
x∞D0z
)µλ
g−1(f(z)) − x0
∣
∣
∣
∣
∣
6 (D0z)−µ
λ
∣
∣
∣x
µλ∞g
−1(f(z)) − f(z)λ+µ
λ
∣
∣
∣+
∣
∣
∣
∣
(
(D0z)− µ
λ+µf(z))
λ+µλ − x0
∣
∣
∣
∣
=
(
(D0z)− µ
λ+µf(z)
y−λ
λ+µ g(y)
)λ+µ
λ ∣
∣
∣
∣
xµλ∞ −
(
y−λ
λ+µ g(y))
λ+µλ
∣
∣
∣
∣
+
∣
∣
∣
∣
(
(D0z)− µ
λ+µf(z))
λ+µλ − x0
∣
∣
∣
∣
→ 0
as z → 0, where y = g−1(f(z)) → 0 as z → 0. So
(
x∞D0z
)µλ
(φsz(xz) − x∞) =
(
x∞D0z
)µλ
g−1(f(z))D∗
∞|D∞|2 → x0
D∗∞
|D∞|2 .
Also, since t′y = sz − tz, and t′y − 1λ+µ
log x∞
y→ 0 as y → 0,
(sz − tz) −1
λ + µlog
x∞(
D0zx∞
)µλx0
→ 0
i.e.
sz −1
λlog
x∞D0z
→ 0.
2.4 Convergence of the fluctuations
Now suppose that XNt is a pure jump Markov process satisfying all the conditions
in Section 2.2, except with b(x) = Bx + τ(x), B and τ defined as in Section 2.3.
In this section we express φt−t0(XNt0
) in a linear form for large values of N and t0.
Recall from Section 2.2 (page 9) that γNt = N
12
(
XNt − xt
)
and γNt ⇒ γt for
each t as N → ∞, where (γt)t>0 is the unique solution to the linear stochastic
Chapter 2. Convergence near saddle points 25
differential equation (2.4).
Fix some t0 > 0. Then φt−t0(XNt0 ) = φt(φ
−1t0 (XN
t0 )) and using the same notation
as in Section 2.2, there exists some θ ∈ (0, 1) such that
φ−1t0
(XNt0
) = φ−1t0
(xt0) +N− 12∇φ−1
t0(xt0)γ
Nt0
+N− 12 (∇φ−1
t0(xt0 + θN− 1
2γNt0
) −∇φ−1t0
(xt0))γNt0
= x0 +N− 12ZN
t0 ,
where ZNt0 ⇒ Zt0 = ∇φ−1
t0 (xt0)γt0 as N → ∞. Now
D0Zt0 = limt→∞
e∗2e−λt∇φt(x0)
∫ t0
0
∇φ−1s (xs)σ(xs)dWs
= limt→∞
e∗2e−λt
∫ t0
0
∇φt−s(xs)σ(xs)dWs,
and
lim inft→∞
e−2λt
∫ ∞
0
|e∗2∇φt−s(xs)σ(xs)|2ds
6 lim inft→∞
e−2λt
∫ ∞
0
|∇φt−s(xs)2|2|a(xs)|ds
6 lim inft→∞
e−2λt
∫ ∞
0
16|Ds|2e2λ(t−s)Ads
632A
λ,
where A is defined in (2.3) and the modulus of Ds = limt→∞ e−λt∇φt(xs)2 is
bounded above by 2, by the same argument used to show existence of D0 in
Lemma 2.11. Hence, if we define
σ2∞ =
∫ ∞
0
limt→∞
e−2λt∇φt−s(xs)2a(xs)∇φt−s(xs)∗2ds
=
∫ ∞
0
e−2λsDsa(xs)D∗sds, (2.12)
then D0Zt0 → Z∞ almost surely as t0 → ∞, where Z∞ ∼ N(0, σ2∞).
Choose x+∞ and x−∞, with 0 < |x±∞| < δ/2 and x−∞,2 < 0 < x+
∞,2, such that
Chapter 2. Convergence near saddle points 26
φ−1t (x±∞) → 0 as t→ ∞. Define a random variable X∞ on the same sample space
as Z∞ by
X∞ =
x+∞ if Z∞ > 0
0 if Z∞ = 0
x−∞ if Z∞ < 0
and define X∞ similarly, except replacing x±∞ by x±∞.
By the Skorohod Representation Theorem, we may assume we are working
in a sample space in which ZNt0
→ Zt0 almost surely for all t0 ∈ N. Without
this assumption, analogous results about weak convergence hold, however this
assumption simplifies the formulation. Let
SN,t0 = infs > t∞ : φs−t0(XNt0
) = X∞ + rD∗∞ for some r ∈ R (2.13)
and
SN =1
2λlogN +
1
λlog
X∞Z∞
, (2.14)
where we interpret 00
= 1.
Theorem 2.14. Suppose that σ∞ 6= 0.
(i) As N → ∞ and then t0 → ∞,
eµt|φt−t0(XNt0
) − φt(x0)| → 0
in probability, uniformly in t on compacts.
(ii) If R 6 t 612λ
logN − R, then there exist ε′i(N, t0, t) → 0, uniformly in t, in
probability as R,N → ∞ and then t0 → ∞, such that
φt−t0(XNt0
) = x0e−µt(e1 + ε′1) +N− 1
2Z∞eλt(e2 + ε′2).
(iii) As N → ∞ and then t0 → ∞, SN,t0 − SN → 0 in probability. Furthermore,
if t = SN,t0 − s for some s, then
eλs|φt−t0(XNt0
) − φ−1s (X∞)| → 0
Chapter 2. Convergence near saddle points 27
uniformly in s on compacts, in probability as N → ∞ and then t0 → ∞.
Proof. (i) By Lemma 2.11, for some θ ∈ (0, 1)
eµt|φt−t0(XNt0
) − φt(x0)| = eµt∣
∣
∣∇φt
(
x0 + θN− 12ZN
t0
)∣
∣
∣N− 1
2 |ZNt0|
6 4e(λ+µ)tN− 12 |ZN
t0 |→ 0
uniformly in t on compacts, in probability.
(ii) We apply Theorem 2.12 with z = N− 12ZN
t0and use the fact that D0Z
Nt0
→ Z∞
almost surely as N → ∞ and then t0 → ∞. A potential problem arises when
Z∞ is close to 0. However, as it is a Gaussian random variable, the probability
of this occurring can be made arbitrarily small.
(iii) The first result follows from Theorem 2.13 by a similar argument to (ii). For
the second result apply a similar argument to the proof of (i) to φ−1t .
2.5 A fluid limit for jump Markov processes
We now show that for large values of N and t, XNt is in some sense close to
φt−t0(XNt0 ) as t0 → ∞, and combine this with results from Section 2.3 to obtain
results analogous to those in the linear case in Section 2.2.
Let f(t, x) = e−Bt(
x− φt−t0(XNt0
))
. By Ito’s formula,
f(t, XNt ) = f(0, XN
0 ) +MB,Nt +
∫ t
0
(
∂f
∂t+Kf
)
(s,XNs−)ds,
where∂f
∂t= −Be−Btx− e−Btτ(φt−t0(X
Nt0 )),
Chapter 2. Convergence near saddle points 28
Kf(s,XNs−) =
∫
R2
(
f(s,XNs− + y) − f(s,XN
s−))
KN (XNs−, dy)
=
∫
R2
e−BsyKN(XNs−, dy)
= e−BsbN (XNs−),
and
MB,Nt =
∫
(0,t]×R2
(
f(s,XNs− + y) − f(s,XN
s−))
(µN − νN)(ds, dy)
=
∫
(0,t]×R2
e−Bsy(µN − νN )(ds, dy).
So if t > t0, then
e−Bt(
XNt − φt−t0(X
Nt0 ))
= MB,Nt −MB,N
t0
+
∫ t
t0
e−Bs(
bN(XNs−) − b(XN
s−))
ds (2.15)
+
∫ t
t0
e−Bs(
τ(XNs−) − τ(φs−t0(X
Nt0
)))
ds.
Since τ ∈ C2, ∇τ is Lipschitz continuous on the unit disc with Lipschitz
constant denoted by K0. In addition to the restrictions on δ from Section 2.3,
suppose that δ < λµ9K0(λ+µ)
.
Theorem 2.15. For all ε > 0,
limt0→0
lim supN→∞
P
(
supt06t6SN,t0
e−λt|XNt − φt−t0(X
Nt0
)| > εN− 12
)
= 0.
Proof. Let
RN,t0 = inf
t > t0 : e−λt|XNt − φt−t0(X
Nt0
)| > N− 12 ε
∧ SN,t0 .
We shall show that RN,t0 = SN,t0 by bounding the terms on the right hand side of
(2.15).
Fix c > 0. Since increasing ε decreases the above probability, we may assume
Chapter 2. Convergence near saddle points 29
0 < ε < η0 ∧ λe−λc
9K0, where η0 is defined at the start of section 2.2. Suppose that
C > 4 and pick R >1λ
log(
8CK0eλc
λ
)
. Define
Ω1N,t0
=
supt>t0
e−λt|eBt(MB,Nt −MB,N
t0 )| < N− 12ε
3
,
Ω2N,t0 ,R =
sup06t6R
eµt|φt−t0(XNt0 ) − φt(x0)| <
δ
2
∩
supR<t<SN,t0
−R|ε′1(N, t0, t)| ∨ |ε′2(N, t0, t)| < 1
∩
supSN,t0
−R6t6SN,t0
eλ(SN,t0−t)|φt−t0(X
Nt0 ) − φ−1
SN,t0−t(X∞)| < δ
2
,
where ε′1 and ε′2 are defined in Theorem 2.14, and
Ω3N,t0,c =
St0,N 61
2λlogN + c
.
Let N0 be sufficiently large that supN>N0N
12 ‖bN − b‖ < λε/3.
On the set Ω1N,t0
∩ Ω2N,t0 ,R ∩ Ω3
N,t0,c ∩ C−1 < |Z∞| < C with N > N0, if
t0 6 t < R, then
|φt−t0(XNt0 )| 6 δe−µt,
if R 6 t 6 SN,t0 − R, then
|φt−t0(XNt0
)| 6 |x0|e−µt(1 + |ε′1|) +N− 12 |Z∞|eλt(1 + |ε′2|)
6δ
2e−µt +N− 1
2 2Ceλt,
and if SN,t0 − R 6 t 6 SN,t0 , then
|φt−t0(XNt0 )| < δe−λ(SN,t0
−t).
Chapter 2. Convergence near saddle points 30
From (2.15), for some θ ∈ (0, 1),
e−λt∣
∣XNt − φt−t0(X
Nt0
)∣
∣
6e−λt|eBt(MB,Nt −MB,N
t0 )| + e−λt
∫ t
t0
|eB(t−s)|∣
∣bN(XNs−) − b(XN
s−)∣
∣ ds
+ e−λt
∫ t
t0
|eB(t−s)|∣
∣τ(XNs−) − τ(φs−t0(X
Nt0
))∣
∣ ds
6e−λt|eBt(MB,Nt −MB,N
t0 )| + 1
λ‖bN − b‖
+
∫ t
t0
e−λs∣
∣∇τ(
φs−t0(XNt0 ) + θ(XN
s− − φs−t0(XNt0 )))∣
∣ |XNs− − φs−t0(X
Nt0 )|ds
6e−λt|eBt(MB,Nt −MB,N
t0 )| + 1
λ‖bN − b‖
+K0
∫ t
t0
(
|φs−t0(XNt0
)| + |XNs− − φs−t0(X
Nt0
)|)
e−λs|XNs− − φs−t0(X
Nt0
)|ds.
Hence, on Ω1N,t0
∩ Ω2N,t0 ,R ∩ Ω3
N,t0,c ∩ C−1 < |Z∞| < C with N > N0,
supt06t6RN,t0
e−λt∣
∣XNt − φt−t0(X
Nt0 )∣
∣
6N− 12ε
3+N− 1
2ε
3+K0
∫ RN,t0
t0
(|φs−t0(XNt0 )| + |XN
s− − φs−t0(XNt0 )|)N− 1
2 εds
6N− 12 ε
(
2
3+K0
(
∫ Rt0,N
t0
(
δ(e−µt + e−λ(SN,t0−t)) +N− 1
2 εeλt)
dt
+
∫ SN,t0−R
t0
N− 12 2Ceλtdt
))
6N− 12 ε
(
2
3+K0
(
δ(λ+ µ)
λµ+εeλc
λ+
2Ceλc
λe−λR
))
<N− 12 ε.
Since XNt is right continuous, this means RN,t0 = SN,t0 and so
P
(
supt06t6SN,t0
e−λt|XNt − φt−t0(X
Nt0 )| > N− 1
2 ε
)
6 P((Ω1N,t0
)c) + P((Ω2N,t0,R)c) + P((Ω3
N,t0,c)c) + P
(
|Z∞| 6∈ (C−1, C))
.
Chapter 2. Convergence near saddle points 31
Letting N, t0, R, C, c→ ∞ in that order, and using Lemma 2.1 and Theorem 2.14
gives
limt0→∞
lim supN→∞
P
(
supt6SN,t0
e−λt|XNt − φt−t0(X
Nt0 )| > N− 1
2 ε
)
= 0.
Remark 2.16. The same idea can be used to obtain convergence results for arbitrary
matrices B e.g. with eigenvalues having the same sign or in higher dimensions. The
rate of convergence and the time up to which convergence is valid will depend on
the eigenvalues of B and bounds on |φt(x)|.
Combining the above result with Theorem 2.14 we get the following.
Theorem 2.17. (i) For all N ∈ N,
N12 |XN
t − φt(x0)|
is bounded uniformly in t on compacts, in probability. (This follows directly
from the fluid limit theorem and diffusion approximation stated in Section
2.2).
(ii) Suppose that R 6 t 612λ
logN −R. Then, provided that σ∞ 6= 0, for i = 1, 2
there exist εi(N, t) → 0 uniformly in t in probability as R,N → ∞ such that
XNt = x0e
−µt(e1 + ε1) +N− 12Z∞e
λt(e2 + ε2),
(cf. (2.7)).
(iii) As N → ∞,
XNSN−s → φ−1
s (X∞),
uniformly on compacts in s > 0, in probability.
Remark 2.18. These results can be reformulated as results about weak convergence
which are true independently of the choice of sample space, in a manner analogous
to Theorem 2.3. In particular, for any sequence tN → ∞ as N → ∞, ZN∞ =
N12 e−λtNXN
tN ,2 ⇒ Z∞. Working on a space in which this sequence converges almost
surely is sufficient for Theorem 2.17.
Chapter 2. Convergence near saddle points 32
2.6 Continuous diffusion Markov processes
Our interest in this problem arose through looking at the OK Corral problem. It
was therefore natural to prove results for pure jump Markov processes. However
the proof of the analogous result in the case of continuous diffusion processes is
similar and we give it below. The pure jump and continuous cases can be combined
to obtain results for more general Markov processes.
Let (XNt )t>0 be a sequence of diffusion processes, starting from x0 and taking
values in some open subset S ⊂ R2, that satisfy the stochastic differential equations
dXNt = σN(XN
t )dWt + bN(XNt )dt
with σN , bN Lipschitz.
Suppose that there exist limit functions b(x) = Bx+ τ(x), with B and τ as in
Section 2.3 and σ, bounded, satisfying
(a)
supx∈S
N12 |bN(x) − b(x)| → 0.
(b)
supx∈S
|N 12σN(x) − σ(x)| → 0.
It follows that there exists a constant A such that for all N
‖σN‖ 6 (A/N)12 . (2.16)
Let γNt = N
12 (XN
t − xt), where xt is defined as before. It is straightforward,
using Gronwall’s Lemma, to show that γNt → γt as N → ∞, where (γt)t>0 is the
unique solution to the linear stochastic differential equation
dγt = σ(xt)dWt + ∇b(xt)γtdt (2.17)
starting from 0, where W is a Brownian motion.
Chapter 2. Convergence near saddle points 33
Consider the function f(t, x) = e−Bt(
x− φt−t0(XNt0
))
for t > t0. By Ito’s
formula,
f(t, XNt ) = f(t0, X
Nt0
) +MB,Nt −MB,N
t0 +
∫ t
t0
(
∂f
∂s(s,XN
s ) + e−BsbN(XNs )
)
ds,
where∂f
∂t= −Be−Btx− e−Btτ(φt−t0(Xt0)),
and
MB,Nt =
∫ t
0
e−BsσN(XNs )dWs.
So if t > t0,
e−Bt(
XNt − φt−t0(X
Nt0
))
= MB,Nt −MB,N
t0
+
∫ t
t0
e−Bs(
bN(XNs−) − b(XN
s−))
ds (2.18)
+
∫ t
t0
e−Bs(
τ(XNs−) − τ(φs−t0(X
Nt0
)))
ds.
By comparison with (2.15), in order for the conclusion of Theorem 2.17 to hold
for diffusion processes, it is sufficient to prove an analogue of Lemma 2.1.
Lemma 2.19. There exists some constant C such that
E
(
supt>t0
e−λt|eBt(MB,Nt −MB,N
t0 )|)
6 CN− 12 e−λt0 .
Proof. By the product rule,
e(B−λI)t(MB,Nt −MB,N
t0 ) =
∫ t
t0
(B − λI)e(B−λI)s(MB,Ns −MB,N
t0 )ds
+
∫ t
t0
e−λsσN(XNs )dWs
Chapter 2. Convergence near saddle points 34
and, hence,
E
(
supt>t0
e−λt|eBt(MB,Nt −MB,N
t0 )|)
6 E
(
supt>t0
∫ t
t0
(λ+ µ)e−(λ+µ)s|(MB,Ns −MB,N
t0 )1|ds)
+ E
(
supt>t0
∣
∣
∣
∣
∫ t
t0
e−λsσN(XNs )dWs
∣
∣
∣
∣
)
6
∫ ∞
t0
(λ+ µ)e−(λ+µ)s(
E(MB,Ns −MB,N
t0 )21
)12ds
+ E
(
supt>t0
∣
∣
∣
∣
∫ t
t0
e−λsσN(XNs )dWs
∣
∣
∣
∣
2)
12
.
Since
E
∫ t
0
|e−λsσN(XNs )|2ds <∞
for all t > 0, the process
(∫ t
0
∫
R2
e−λsσN(XNs )dWs
)
t>0
is a martingale, and hence, by Doob’s L2 inequality
E
(
supt>t0
∣
∣
∣
∣
∫ t
t0
e−λsσN(XNs )dWs
∣
∣
∣
∣
2)
6 4 supt>t0
E
(
∣
∣
∣
∣
∫ t
t0
e−λsσN (XNs )dWs
∣
∣
∣
∣
2)
.
Now
E
(
(MB,Nt −MB,N
t0 )21
)
= E
∫ t
t0
e2µsaN(XNs )1,1ds
6 E
∫ t
t0
e2µs A
Nds
6e2µtA
2µN,
Chapter 2. Convergence near saddle points 35
where A is defined in (2.16). Similarly,
E
(
∣
∣
∣
∣
∫ t
t0
e−λsσN(XNs )dWs
∣
∣
∣
∣
2)
6e−2λt0A
2λN.
Hence,
E
(
supt>t0
e−λt|eBt(MB,Nt −MB,N
t0 )|)
6
∫ ∞
t0
(λ+ µ)e−λs
(
A
2µN
)12
ds+ e−λt0
(
2A
λN
)12
6A
12 (λ+ µ+ 2(λµ)
12 )
λ(2µ)12
N− 12 e−λt0 .
Define σ∞, Z∞, X∞, X∞ as in Section 2.4 and let
SN =1
2λlogN +
1
λlog
X∞Z∞
.
The following analogue of Theorem 2.17 for diffusion processes holds.
Theorem 2.20. (i) For all N ∈ N,
N12 |XN
t − φt(x0)|
is bounded uniformly in t on compacts, in probability.
(ii) Suppose that R 6 t 612λ
logN −R. Then, provided that σ∞ 6= 0, for i = 1, 2
there exist εi(N, t) → 0 uniformly in t in probability as R,N → ∞ such that
XNt = x0e
−µt(e1 + ε1) +N− 12Z∞e
λt(e2 + ε2).
(iii) As N → ∞,
XNSN−s → φ−1
s (X∞),
uniformly on compacts in s > 0, in probability.
Chapter 2. Convergence near saddle points 36
2.7 Applications
Throughout this section we work in a sample space on which ZN∞ → Z∞ almost
surely so that, in particular, the statement of Theorem 2.17 holds.
2.7.1 Hitting lines through the origin
As in the linear case, Theorems 2.17 and 2.20 may be used to study the first time
that XNt hits lθ or l−θ, the straight lines passing through the origin at angles θ and
−θ, where θ ∈ (0, π2), as N → ∞. As in Section 2.2, we define the time that XN
t
first crosses one of the lines l±θ by
TNθ = inf
t > 0 :∣
∣
∣
XNt−,2
XNt−,1
∣
∣
∣6 | tan θ| and
∣
∣
∣
XNt,2
XNt,1
∣
∣
∣> | tan θ|
.
First note that by Lemma 2.10,
φt(x0)2
φt(x0)1=eµtφt(x0)2
eµtφt(x0)1→ 0
x0= 0
as t → ∞. In particular, since tan θ 6= 0, there exists some sθ > 0 such that∣
∣
∣
φt(x0)2φt(x0)1
∣
∣
∣< | tan θ| for all t > sθ. To rule out the trivial case where TN
θ converges
to the first time that φt(x0) hits l±θ, we shall assume that x0 is chosen sufficiently
close to the origin that sθ = 0.
We prove the following result in the case where XNt is a pure jump process. The
proof for continuous diffusion processes is identical, except that it uses Theorem
2.20 in place of Theorem 2.17.
Theorem 2.21. Under the conditions required for Theorem 2.17
TNθ − tN ⇒ cθ
and
Nµ
2(λ+µ) |XNT N
θ| ⇒ | sec θ|| tan θ|−
µλ+µ |x0|
λλ+µ |Z∞|
µλ+µ
Chapter 2. Convergence near saddle points 37
as N → ∞, where
tN =1
2(λ+ µ)logN and cθ =
1
λ + µlog
∣
∣
∣
∣
x0 tan θ
Z∞
∣
∣
∣
∣
.
Proof. By the fluid limit theorem and diffusion approximation, for any constant
R > 0,
P(
TNθ 6 R
)
6 P
(
supt6R
∣
∣
∣
∣
∣
XNt,2
XNt,1
∣
∣
∣
∣
∣
> | tan θ|)
= P
(
supt6R
∣
∣
∣
∣
∣
φt(x0)2 +N− 12γN
t,2
φt(x0)1 +N− 12γN
t,1
∣
∣
∣
∣
∣
> | tan θ|)
→ 0
as N → ∞.
By an identical argument to that in the proof of Theorem 2.5,
P(
R 6 TNθ 6 tN + cθ − ε
)
→ 0
and
P(
tN + cθ + ε 6 TNθ 6 SN − R
)
→ 0
as R,N → ∞. The result follows immediately.
Remark 2.22. As in the linear case, the sign of Z∞ determines whether XNt hits lθ
or l−θ at time TNθ . Since Z∞ is a Gaussian random variable with mean 0, each event
occurs with probability 12. Furthermore, provided that x∞ is chosen sufficiently
close to the origin that φ−1t (x∞) does not intersect l±θ, if XN
t hits one of the two
lines then the probability of it hitting either line again before SN converges to 0,
as N → ∞.
2.7.2 Minimum distance from the origin
Our second application is to investigate the minimum distance from the origin that
XNt can attain for large values of N .
Chapter 2. Convergence near saddle points 38
Theorem 2.23. Under the conditions required for Theorem 2.17,
Nµ
2(λ+µ) inft6SN
|XNt | ⇒
(µ
λ
) λ2(λ+µ)
(
λ
µ+ 1
)12
|x0|λ
λ+µ |Z∞|µ
λ+µ
as N → ∞.
Proof. By the fluid limit theorem and diffusion approximation, for any constant
R > 0,
inft6R
Nµ
2(λ+µ)∣
∣XNt
∣
∣ > inft6R
Nµ
2(λ+µ)
(
|φt(x0)| −N− 12 |γN
t |)
→ ∞,
as N → ∞.
By Theorem 2.17,
infR6t6tN−R
Nµ
2(λ+µ)∣
∣XNt
∣
∣
> infR6t6tN−R
(
eµ(tN−t)|x0|(1 − |ε1|) − eλ(t−tN )|Z∞|(1 + |ε2|))
→ ∞
in probability as R,N → ∞, where
tN =1
2(λ+ µ)logN.
For each c > 0, there exists some ε = ε(N) → 0 in probability as N → ∞ such
that
infSN−c6t6SN
Nµ
2(λ+µ)∣
∣XNt
∣
∣ > inf06s6c
Nµ
2(λ+µ)(
|φ−1s (X∞)| − ε
)
→ ∞.
Also,
inftN +R6t6 1
2λlog N−R
Nµ
2(λ+µ)∣
∣XNt
∣
∣
> inftN +R6t6 1
2λlog N−R
eλ(t−tN )|Z∞|(1 − |ε2|) − eµ(tN−t)|x0|(1 + |ε1|)
→ ∞
in probability as R,N → ∞.
Chapter 2. Convergence near saddle points 39
Finally if t = tN + c, then
Nµ
2(λ+µ)∣
∣XNt
∣
∣ = Nµ
2(λ+µ)
(
(
e−µtx0(1 + ε1,1) +N− 12Z∞e
λtε2,1
)2
+(
e−µtx0ε1,2 +N− 12Z∞e
λt(1 + ε2,2))2)
12
→(
(e−µcx0)2 + (eλcZ∞)2
)12
in probability uniformly in c on compact intervals. The right hand side is minimised
when
c =1
2(λ+ µ)log
µx20
Z2∞λ
.
Therefore,
Nµ
2(λ+µ) inft6SN
|XNt | ⇒
(µ
λ
)λ
2(λ+µ)
(
λ
µ+ 1
)12
|x0|λ
λ+µ |Z∞|µ
λ+µ
as N → ∞.
Example 2.24. Let (UNt , V
Nt ) be a Z2-valued process modelling the sizes of two
populations of the same species with UN0 = V N
0 = N . The environment that
they occupy is assumed to be closed. Each individual reproduces at rate 1. Ad-
ditionally, the individuals are in competition with each other, a death occurring
due to competition over resources at rate α, and due to aggression between the
populations at rate β. Hence, the transition rates are
(u, v) →
(u+ 1, v) at rate u
(u− 1, v) at rate αu(u+ v − 1)/N + βuv/N
(u, v + 1) at rate v
(u, v − 1) at rate αv(u+ v − 1)/N + βuv/N.
Let XNt =
(
UNt , V
Nt
)
/N . This gives a sequence of pure jump Markov processes,
Chapter 2. Convergence near saddle points 40
starting from x0 = (1, 1), with Levy kernels
KN (x, dy) = Nx1δ(1/N,0) +N (αx1(x1 + x2 − 1/N) + βx1x2) δ(−1/N,0)
+Nx2δ(0,1/N) +N (αx2(x1 + x2 − 1/N) + βx1x2) δ(0,−1/N).
If we let
K(x, dy) = x1δ(1,0) +(αx21 +(α+β)x1x2)δ(−1,0) +x2δ(0,1) +(αx2
2 +(α+β)x1x2)δ(0,−1)
then for S = (0, 2)2 and η0 = 1,
m(x, θ) = x1eθ1 + (αx2
1 + (α + β)x1x2)e−θ1 + x2e
θ2 + (αx22 + (α + β)x1x2)e
−θ2
satisfies
supx∈SN
sup|θ|6η0
∣
∣
∣
∣
mN (x,Nθ)
N−m(x, θ)
∣
∣
∣
∣
→ 0
as N → ∞. Therefore,
b(x) = m′(x, 0) =(
x1(1−αx1−(α+β)x2)x2(1−αx2−(α+β)x1)
)
.
The deterministic differential equation
φt(x) = b(φt(x)), φ0(x) = x
is a special case of the Lotka-Volterra model for two-species competition. See
Brown and Rothery [4] for a detailed interpretation of the parameters α and β.
Further generalizations are discussed in Durrett [7].
It is straightforward to check that b(x) is C1 on S and satisfies
supx∈SN
N12 |bN (x) − b(x)| → 0
as N → 0, and that
a(x) =(
x1(1+αx1+(α+β)x2) 00 x2(1+αx2+(α+β)x1)
)
is Lipschitz on S.
Chapter 2. Convergence near saddle points 41
Now b(x) has a saddle fixed point at(
12α+β
, 12α+β
)
and, by symmetry, any
point x on the line x1 = x2 satisfies φt(x) →(
12α+β
, 12α+β
)
as t → ∞. So under
an appropriate translation and rotation, the conditions required for Theorem 2.17
are satisfied, with λ = 1 and µ = β2α+β
. (Note that σ2∞ > 0 since a(x) is positive
definite on S). Hence, for times t satisfying t tN , where tN = 2α+β4(α+β)
logN ,
the two populations co-exist with the sizes of both being equal. However, at
time tN +O(1), the deterministic approximation breaks down and one side begins
to dominate. Our results give a quantitative description of the behaviour of the
processes in this region; however, we do not go into this here. At time t = SN +s =12logN +O(1), XN
t → φ−1s (X∞) in probability as N → ∞, where SN is defined in
Theorem 2.15 and X∞ is defined in Section 2.4. Now b(x) has stable fixed points at
(α−1, 0) and (0, α−1) and hence φ−1s (X∞) converges to one of these two fixed points
as s→ ∞. For any ε ∈ (0, 1) we say that a population is ε-extinct if the proportion
of the original population that remains is less than ε. Thus, for arbitrarily small
ε, one of the populations will become ε-extinct at time 12logN +O(1).
Chapter 3
Accumulation of rounding errors
in the numerical solution of ODEs
3.1 Introduction
In this chapter we examine the rounding errors incurred by deterministic solvers
for systems of ordinary differential equations (ODEs). We show, by the application
of ideas from Chapter 2, that the accumulation of rounding errors results in a ‘so-
lution’ to the ODE which exhibits random behaviour. The theoretical distribution
of the solution is obtained as a function of time, the step size and the numerical
precision of the computer. We consider, in particular, systems which amplify the
effect of the rounding errors so that over long time periods the solutions exhibit
divergent behaviour. The distributions predicted theoretically are then observed
numerically by performing multiple repetitions with different values of the time
step size.
Consider ordinary differential equations of the form
xt = b(xt).
These can be solved numerically using iteration methods of the type
xt+h = xt + β(h, xt),
42
Chapter 3. Rounding errors 43
where β(h, x)/h→ b(x) as h→ 0.
The simplest example is the Euler method, where β(h, x) = hb(x). This method
is generally not used in practice as it is relatively inaccurate and unstable compared
to other methods. However, more useful methods, such as the fourth order Runge-
Kutta formula (RK4), fall into this scheme.
When solving an ordinary differential equation numerically, each time an iter-
ation is performed an error ε is incurred due to rounding i.e.
Xht+h = Xh
t + β(h,Xht ) + ε. (3.1)
Rounding errors in numerical computations are an inevitable consequence of fi-
nite precision arithmetic. The first work thoroughly analysing the effects of round-
ing errors on numerical algorithms is the classical textbook by Wilkinson [33]. A
recent comprehensive treatment of the behaviour of numerical algorithms in finite
precision, including an extensive list of references, can be found in Higham [21].
Although rounding errors are not random in the sense that the exact error incurred
in any given calculation is fully determined (see Higham [21] or Forsythe [12]), in
many situations probabilistic models have been shown to adequately describe their
behaviour. In fact, statistical analysis of rounding errors can be traced back to
one of the first works on rounding error analysis by Goldstine and von Neumann
[13].
Henrici [18, 19, 20] proposes a probabilistic model for individual rounding errors
whereby they are assumed to be independent and uniform, the exact distribution
depending on the specific finite precision arithmetic being used. Using the central
limit theorem, he shows that the theoretical distribution of the error accumulated
after a fixed number of steps in the numerical solution of an ODE is asymptotically
normal with variance proportional to h−1. By varying the initial conditions, he
obtains numerical distributions for the accumulated errors with good agreement.
Hull and Swenson [22] test the validity of the above model by adding a randomly
generated error with the same distribution at each stage of the calculation, and
comparing the distribution of the accumulated errors with those obtained purely
by rounding. They observe that, although rounding is neither a random process
nor are successive errors independent, probabilistic models appear to provide a
Chapter 3. Rounding errors 44
good description of what actually happens.
We shall concentrate on floating point arithmetic, as used by modern com-
puters. However, our methods can be used equally well for any finite precision
arithmetic. We use the model, discussed and tested by the authors cited above,
whereby under generic conditions the errors in (3.1) can be viewed as independent,
zero mean, uniform random variables,
εi ∼ U [−|Xht,i|2−p, |Xh
t,i|2−p],
p being a constant determined by the precision of the computer.
In the first half of the chapter we analyse the cumulative effect of these rounding
errors as the step size h tends to 0. Where previous authors have considered the
accumulated error at a particular point, we derive a theoretical model for the entire
trajectory. Cases in R2 where the ordinary differential equation has a saddle fixed
point at the origin demonstrate the most interesting behaviour, as the structure
of the ODE system amplifies the effect of the rounding errors and causes the
numerical solution to diverge from the actual solution. In this case, the solution
Xht exhibits random behaviour and its theoretical distribution can be obtained as
an explicit function of time, the step size and the precision of the computer. As
the step size h tends to 0, the numerical solution exhibits three different types
of behaviour, depending on the time. More precisely, there exists a constant c,
determined by the ODE system, such that for times much smaller than −c log h
the numerical solution converges to the actual solution; for times close to −c log h
the solution undergoes a transition, determined by a Gaussian random variable
whose distribution is obtained; for times much larger than −c log h the numerical
solution diverges from the actual solution.
In the second half of the chapter, we perform numerical simulations which
illustrate this behaviour. By performing multiple repetitions with different values
of the time step size, the random distributions predicted theoretically are observed.
Where previous authors have obtained their numerical distributions by varying
the initial conditions, we do so by introducing small variations in the step size h.
During the transition period described in the previous paragraph, the numerical
solution intersects straight lines through the origin and we compare the theoretical
Chapter 3. Rounding errors 45
and numerical distributions for the points at which these intersections occur. Both
the mean and the standard deviation of these distributions are of the form ahγ ,
where γ ∈ (0, 1/2] is a constant determined by the ODE system, and a can be
found explicitly in terms of the precision of the computer, i.e. the number of bits
used internally by the computer to represent floating point numbers. We mainly
focus on the explicit Euler and RK4 methods, but show that the same behaviour
is also observable for more complex algorithms such as the adaptive solvers VODE
[5] and RADAU5 [14].
3.2 Theoretical background
In Chapter 2, limiting results are established for sequences of Markov processes
that approximate solutions of ordinary differential equations with saddle fixed
points. By modelling the rounding errors as random variables, we show that
the solutions obtained when performing numerical schemes for solving ordinary
differential equations can be viewed as a special case of this. This enables us to
quantify how the rounding errors accumulate. The resulting numerical solutions
exhibit random behaviour, the exact distribution of which is obtained.
In Section 3.2.1 we describe how rounding errors can be modelled as random
variables with specified distributions. The results of Chapter 2 are applied to
obtain a qualitative description of the accumulation of the rounding errors. The
distribution is calculated explicitly in Section 3.2.2.
3.2.1 Accumulation of rounding errors
We are interested in numerically solving ordinary differential equations of the form
xt = b(xt). (3.2)
In particular we consider using iteration methods of the type
xt+h = xt + β(h, xt) (3.3)
Chapter 3. Rounding errors 46
where β(h, x)/h→ b(x) as h→ 0.
Each time an iteration is performed, an error ε = ε(h, t) is incurred due to
rounding. The process (Xht )t∈hN is obtained iteratively by
Xht+h = Xh
t + β(h,Xht ) + ε. (3.4)
Modern computers store real numbers by expressing them in binary as x = m2n
for some 1 6 |m| < 2 and n ∈ Z. They allocate a fixed number of bits to store
the mantissa m and a (different) fixed number of bits to store the exponent n [23].
When adding to x a number of smaller order, the size of the rounding error incurred
is between 0 and 2n−p = 2blog2 |x|c−p, where p is the number of bits allocated to store
the mantissa. Although it is possible to carry out the calculations below using the
exact value of 2blog2 |x|c−p, the calculations are greatly simplified by approximating
it by |x|2−p. This results in the ‘effective’ value of p differing from the actual value
of p by some number between 0 and 1. Provided β(h,Xht ) is sufficiently small
compared with Xht , the errors ε can therefore be viewed as independent, mean
zero, uniform random variables with approximate distribution
εi ∼ U [−|Xht,i|2−p, |Xh
t,i|2−p]
(see Henrici [18, 19, 20]). The assumption that the εi are independent is in general
not true. In fact, in certain pathological cases, for example where there is a lot
of symmetry in the components, the εi can be strongly correlated. Nevertheless,
under generic conditions one would expect any correlations to be weak and so this
is a reasonable assumption to make. We shall see by the agreement of our numerical
and theoretical results that the effect of making this assumption is indeed small.
Although the above iterations are carried out at discrete time intervals, it is
convenient to embed the processes in continuous time by performing the iterations
at times of a Poisson process with rate h−1. As β(h, x) does not depend on t,
this does not affect the shape of the resulting trajectories. In this way Markov
processes Xht are obtained that approximate the stable solution of (3.2) for small
Chapter 3. Rounding errors 47
values of h. If, in addition, the assumption is made that
h−12
(
β(h, x)
h− b(x)
)
→ 0
as h → 0 (note that both the Euler and Runge-Kutta methods satisfy this con-
dition), then under the correspondence N ∼ h−1, the conditions needed to apply
the results in Chapter 2 are satisfied.
We focus on R2 in the case where the origin is a saddle fixed point of the
system i.e. b(xt) = Bxt + τ(xt), where B is a matrix with eigenvalues λ,−µ,
with λ, µ > 0 and corresponding eigenvectors v1, v2, and τ(x) = O(|x|2) is twice
continuously differentiable. This case is of particular interest as the structure of
the system amplifies the effect of the rounding errors and causes the numerical
solution to diverge from the actual solution over large times. Similar behaviour
can be observed in higher dimensions where the matrix B has at least one positive
and one negative eigenvalue, although the corresponding quantitative analysis is
much harder and we do not go into it here.
As shown in Chapter 2, our numerical solution exhibits the following random
behaviour:
A. For times of order much smaller than − log h, Xht approximates the stable
solution of (3.2), the fluctuations around this limit being of order h12 .
B. There exists some x0 6= 0, depending only on x0, and a Gaussian random
variable Z∞, such that if t lies in the interval [−c log h,− 12λ
log h + c log h]
for some c > 0, then Xht is asymptotic to
x0e−µtv1 + h
12Z∞e
λtv2, (3.5)
the solution to the linear ordinary differential equation
yt = Byt
starting from the random point x0v1 + h12Z∞v2.
C. Provided Z∞ 6= 0, in time intervals around − 12λ
log h whose length is of
Chapter 3. Rounding errors 48
much smaller order than − log h, Xht approximates one of the two unstable
trajectories of (3.2), each with probability 12, depending on the sign of Z∞.
The random behaviour resulting from the accumulation of rounding errors is
most noticeable on time intervals of fixed lengths around − 12(λ+µ)
log h, as for these
values of t the two terms x0e−µt and h
12Z∞e
λt in (3.5) are of the same order. During
these time intervals, the numerical solution undergoes a transition from converging
to the actual solution to diverging from it. During this transition, for each value of
θ ∈ (0, π/2), Xht crosses one of the straight lines passing through 0 in the direction
v1 cos θ ± v2 sin θ. These intersections are important as they indicate the onset of
divergent behaviour. The distribution of the point at which Xht intersects one of
the lines in the direction v1 cos θ ± v2 sin θ is asymptotic to
hµ
2(λ+µ) |Z∞|µ
λ+µ |x0|λ
λ+µ | tan θ|µ
λ+µ (v1 cos θ ± v2 sin θ). (3.6)
In Section 3.2.2 we show how to evaluate the variance of Z∞, doing so explicitly
in the linear case and obtaining bounds in the non-linear case. In Section 3.3 these
results are verified by numerically obtaining the predicted distribution for hitting
a line through the origin.
3.2.2 Explicit calculation of the variance
Consider a numerical scheme that satisfies the above conditions, applied to obtain
a solution to the ordinary differential equation (3.2), starting from x0 for some
x0 in the stable manifold. In the non-linear case we require that x0 is sufficiently
close to the origin such that τ(x0) is small. In general, for simplicity, we assume
that |x0| 6 1.
Define the flow φ associated with this system by
φt(x) = b(φt(x)), φ0(x) = x
and let xt = φt(x0).
Suppose that v1, v2 ∈ R2 are the unit right-eigenvectors of B corresponding to
−µ, λ respectively, and that v′1, v′2 ∈ (R2)∗ are the corresponding left-eigenvectors
Chapter 3. Rounding errors 49
(i.e. v′ivj = δij).
Define
x0 = limt→∞
eµtv′1φt(x0)
and
Ds = limt→∞
e−λtv′2∇φt(xs).
It is shown in the proofs of Lemmas 2.10 and 2.11 that these limits exist and that
|x0| 6 2|x0| 6 2 and |Ds| 6 2.
Finally, let
a(x) =1
32−2p
(
x21 0
0 x22
)
be the covariance matrix of the multivariate uniform random variable ε, defined
in equation (3.4), when Xht = x. Then Z∞ ∼ N(0, σ2
∞) where, by (2.12),
σ2∞ =
∫ ∞
0
e−2λsDsa(xs)D∗sds.
Note that σ2∞ 6
23λ
2−2p.
In the general non-linear case, evaluating σ2∞ explicitly is not possible as it
involves solving (3.2). It is possible to obtain better approximations than that
above, although the important observation is that σ2∞ is proportional to 2−2p.
In the linear case, φt(x) = eBtx and x0 = |x0|v1. Hence xt = |x0|e−µtv1,
x0 = |x0|, and Ds = v′2, and so
σ2∞ =
1
3(λ+ µ)2−2p|x0|2(v1,1v
′2,1)
2.
Note that the directions of v1 and v′2, relative to the standard basis, are critical.
For example, if either v1 or v′2 is parallel to one of the standard basis vectors, then
σ2∞ = 0.
Chapter 3. Rounding errors 50
3.3 Numerical experiments
In this section we solve ODEs numerically using deterministic solvers and observe
the predicted random distributions arising as a consequence of the accumulation
of rounding errors. For simplicity, and in order to observe the desired effects as
clearly as possible, we mainly focus on the most elementary of all numerical ODE
solution methods, the standard explicit Euler algorithm with constant time step
size. However, we observe similar behaviour for RK4 and also briefly mention
results obtained with more complex solvers, such as VODE [5].
3.3.1 The system
For x : [0,∞) → R2, consider the linear ODE
x(t) = Bx(t), (3.7)
where
B =
(
−µ 0
0 λ
)
for fixed λ, µ > 0. Introduce new coordinates
x(t) = R(ϕ)x(t)
by rotating about the origin by a fixed angle ϕ ∈ [0, π/2), i.e.
R(ϕ) =
(
cosϕ − sinϕ
sinϕ cosϕ
)
.
We arrive at the transformed system
˙x(t) = B(ϕ)x(t) (3.8)
with
B(ϕ) = R(ϕ)BR(ϕ)∗,
Chapter 3. Rounding errors 51
which will be the system under consideration in the following. Throughout, the
initial value
x(0) = R(ϕ)
(
1
0
)
=
(
cosϕ
sinϕ
)
(3.9)
is used. The phase space evolution is sketched in Figure 3.1.
Figure 3.1: Phase space for the saddle point ODE system (3.8) with sample tra-jectories and lines where hitting distributions are recorded (dashed lines).
3.3.2 Theoretical hitting distribution
As discussed in Section 3.2.1, the numerical solution to the above ODE system
undergoes a transition from converging to the actual solution to diverging from
it. During this transition, the numerical trajectory crosses one of the straight
lines passing through 0 at an angle ϕ ± θ for each value of θ ∈ (0, π/2). These
intersections are important as they indicate the onset of divergent behaviour. The
hitting distributions also provide a means of measuring the random variable Z∞,
which determines the random variations in our solutions, and hence of verifying
the theoretical results.
Equation (3.6) gives the asymptotic distribution of the magnitude of the point
at which the numerical solution hits the line through the origin at angle ϕ± π4
as
Chapter 3. Rounding errors 52
|Z|µ
λ+µ , where Z is a Gaussian random variable with mean 0 and variance
σ2 = hσ2∞ =
1
3(λ+ µ)h2−2p(cosϕ sinϕ)2 (3.10)
i.e. Z ∼ N(0, σ2). We obtain an explicit formula for the asymptotic distribution
by starting from the N(0, σ2) distribution
p(x)dx =1√2πσ
exp
(
− 1
2σ2x2
)
dx
and performing a change of variable given by y = |x|µ
λ+µ . The result is
p(y)dy =2(λ+ µ)√
2πσµy
λµ exp
(
− 1
2σ2y
2(λ+µ)µ
)
dy.
In the case λ = µ = 1, which is considered below, setting a = 4√2πσ
produces
the family of distributions
f(y)dy = ay exp(
− π
16a2y4
)
dy, y ∈ (0,∞), (3.11)
which will be fitted to the numerical data to confirm the theoretical value of a.
3.3.3 Choice of parameters
Rounding errors are deterministic in the sense that any given number of iterations
of a particular numerical scheme will generate the same solution. In order to
obtain a distribution from the numerical solutions to (3.7), for each repetition it
is necessary to vary at least one parameter by a small amount. In this section we
discuss this issue as well as the choice of the fixed parameters of the system such
as the eigenvalues.
The possible parameters that can be varied are the initial value x0, and the time
step size h. As x0 is constrained by being on the stable manifold, any variation
is required to be in the direction of the eigenvector corresponding to eigenvalue
−µ. Varying the initial value in this way did not yield any interesting results
as the chosen distribution of initial values was reproduced exactly in the hitting
Chapter 3. Rounding errors 53
distribution. In terms of the system it is also preferable to vary the time step
size as this parameter is internal to the algorithm, whereas the initial value is a
physical parameter of the system. We varied the time step size as follows. Given
a user-supplied value of h, define the step size hi for the ith repetition by
hi = h+ ∆h(i− 1 − k), i = 1, . . . , L,
where the number of repetitions L = 2k + 1 and 0 < ∆h h are user-supplied.
For all simulations, we set k = 104.
Reasonable choices of h and ∆h are limited by several factors. The hitting
distribution predicted theoretically in Section 3.3.2 is asymptotic as h → 0 and
hence, if h is too large (in the considered case, if h > 10−1 for both single and
double precision), the observed hitting distribution differs substantially from the
theoretical one. The onset of such effects can be seen for large values of h in
Figure 3.4. Lower bounds on h are imposed by computational cost and by the
numerical precision of the computer. In practice, computational expense becomes
prohibitive for values of h much larger than the smallest values permitted by
numerical accuracy. Our particular choice of step size distribution requires that
k∆h should be (much) smaller than h. The lower limit for ∆h is determined solely
by the numerical precision, i.e. ∆h/h must not be smaller than the numerical
precision.
We did not investigate in detail the dependence of our observations on the
distribution of step sizes. However, preliminary experiments with varying ∆h
and even with non-uniform step size distributions suggest that this dependence
is very weak for a wide range of conditions. Figure 3.2 shows that the shape of
the distribution exhibits no discernible systematic dependence on ∆h over at least
nine orders of magnitude. The deviations seen for values of ∆h smaller than about
10−19 are due to the fact that ∆h/h approaches the limits of numerical precision.
The remaining parameters that we need to choose are the eigenvalues λ,−µand the rotation angle ϕ. Since the limit distribution is given by |Z|
µλ+µ , for some
Gaussian random variable Z, if the values of λ and µ differ significantly then the
distribution is hard to observe in a numerical experiment. This suggests choosing
λ and µ of the same order of magnitude, and we therefore take λ = µ = 1 for all
Chapter 3. Rounding errors 54
0
1
2
3
4
5
10-20
10-18
10-16
10-14
10-12
10-10
Pa
ram
ete
r a
[1
01
4]
Step size variation ∆h
Figure 3.2: Step size variation for Euler’s algorithm (double precision, step sizeh = 10−4, L = 20001 repetitions each).
simulations.
There is some subtlety in the choice of the rotation angle ϕ. For certain values,
trivial trajectories or symmetry effects can occur which conceal the desired accu-
mulation of rounding errors. For instance, for ϕ = 0 the second component x2 of
the solution is always zero, and therefore the trajectory stays on the line x2 = 0
(or equivalently x2 = 0) with no fluctuations. Note that this is in agreement with
σ2 = 0 in equation (3.10). For ϕ = π/4, any rounding error that appears in one
component also appears in the other one, which implies that, again, the trajectory
always stays on the line x2 = 0 (or equivalently x1 = x2). This case is pathological
as it consistently violates our assumption that the rounding errors for the different
components are independent. For these reasons, we chose ϕ = π/5 throughout.
3.3.4 Results and observations for explicit methods
Using the values of the parameters discussed above, we carried out multiple repe-
titions of Euler’s algorithm and RK4. In each run we noted the point at which the
trajectory given by the numerical solution intersected one of the lines x1 = ±x2
Chapter 3. Rounding errors 55
(the dashed lines in Figure 3.1). Histograms were then produced by partition-
ing the interval [0, 1] into a given fixed number of subintervals of equal length and
counting how many times y fell into each subinterval, where y denotes the distance
of the point of intersection from the origin. The empirical distributions shown in
Figure 3.3 were obtained. The theoretical distribution (3.11) was fitted to the
empirical distributions with very good agreement.
0
0.5
1
1.5
2
2.5
3
3.5
0 2 4 6 8 10
h=1.0x10-4
h=3.2x10-4
h=1.0x10-3
Rel
ativ
e fr
eque
ncy
dens
ity p
(y)
[107 ]
y [10-8]
Figure 3.3: Observed hitting distributions with theoretical fits for Euler’s algorithm(∆h = 10−10, L = 20001 repetitions each).
For each value of h, we obtained a value for the parameter a by fitting a
distribution of the form (3.11) to our numerical data. In Figure 3.4 the parameter
a is plotted as a function of the time step size h, both for single (Figure 3.4(a)) and
double (Figure 3.4(b)) precision (4 and 8 bytes internal representation of floating
point numbers respectively). Error bars due to the fit are only about 1% and hence
insignificant. In both cases, the dependence between a and h is well described by
a ∝√h.
Chapter 3. Rounding errors 56
(a) Single precision (∆h = 10−8).
1014
1015
1016
10-5 10-4 10-3 10-2
Euler
4th order Runge-Kutta
4.9563x1016*h1/2
Par
amet
er a
Time step size h
(b) Double precision (∆h = 10−10).
Figure 3.4: Parameter a in equation (3.11) as function of the time step size h forsimple explicit methods (Euler and 4th order Runge-Kutta).
Equation (3.10) predicts the value of ah− 12 to be
ah−12 =
4√
3√π cos π
5sin π
5
× 2p = 8.220 × 2p.
For Euler’s method, the above data give ah− 12 = 9.411×107 for single precision and
ah−12 = 4.956× 1016 for double precision. For the 4th order Runge-Kutta method,
the values are ah−12 = 9.27 × 107 (with a relatively large error of ±0.12 × 107)
for single precision and ah−12 = 4.746 × 1016 for double precision. Using the
approximation discussed in Section 3.2.1, the actual value of p is between 23 and 24,
when working in single precision, and between 52 and 53 when working in double
precision. The particular value depends on the exact number being computed. Our
theoretical results therefore predict ah− 12 lies between 6.895× 107 and 1.379× 108
for single precision and between 3.702×1016 and 7.404×1016 for double precision.
There are three possible sources of error in our calculations. The first is the
error in fitting the numerical data to the theoretical model, the second is that
our theoretical models are based on asymptotic results as h → 0, whereas we are
Chapter 3. Rounding errors 57
applying them to values of h which are necessarily larger than the precision of the
computer. The third source of error arises from the assumption that at each stage
the rounding error can be viewed as an independent uniform random variable,
depending on a fixed value of p. The above results show that these errors are all
small and that our theoretical model provides a very good fit.
3.3.5 Adaptive solvers
Our theoretical results cover ODE solvers which use algorithms of the form (3.3).
In practice, more sophisticated adaptive solvers are used, such as VODE [5] and
RADAU5 [14]. For these solvers, the user inputs the error tolerances RTOL (relative)
and ATOL (absolute) and the global time step hg (the time interval after which
the user requests solution output from the solver). However, the user has no
immediate control over the size of the actual steps taken. These are determined
algorithmically as a function of the error tolerance parameters RTOL and ATOL,
generally by trial-and-error methods using heuristics, rather than by an explicit
formula.
0
1
2
3
4
0 1 2 3 4
hg=1.0x10-4
hg=3.2x10-4
Rel
ativ
e fr
eque
ncy
dens
ity p
(y)
[107 ]
y [10-8]
Figure 3.5: Hitting distributions for VODE.
Chapter 3. Rounding errors 58
Although it is not possible to analyse such adaptive solvers in the way that we
have analysed explicit solvers above, it is still of interest to see whether they exhibit
the same qualitative random behaviour. We performed numerical experiments
similar to those discussed above and obtained the distributions shown in Figure
3.5 in the case where RTOL=0.
Experiments do not readily suggest a simple relationship between the param-
eter a in equation (3.11) and any of the parameters ATOL, RTOL, and hg. This
is possibly not surprising given the lack of direct control over the time step size.
However, the fact that the results are qualitatively similar supports the assertion
that the observed phenomena are not specific to a particular algorithm, but rather
are general effects.
Chapter 4
Stochastic flows, planar
aggregation and the Brownian
web
4.1 Introduction
In this chapter we change course and consider a class of stochastic flows on the
circle which, under a certain scaling, converge to the Brownian web.
The Brownian web can loosely be defined as a family of coalescing Brownian
motions, starting at all possible points in continuous space-time. Arratia [1] first
considered this object in 1979 as a limit for discrete coalescing random walks, and
since then it has been studied by Toth and Werner [30] and Fontes, Isopi, Newman
and Ravishankar [10] amongst others.
Our motivation for looking at the Brownian web arises from a surprising connec-
tion with Hastings-Levitov diffusion-limited aggregation (DLA). DLA is a random
growth model which was originally introduced in 1981 by Witten and Sander [35].
In this model particles diffuse in from “infinity” and perform Brownian motions in
the plane until they collide with a cluster at the origin, at which point they stick to
the cluster. In 1998 Hastings and Levitov [17] formulated a model of DLA in which
the cluster is represented by a sequence of iterated conformal maps. We study a
59
Chapter 4. The Brownian web 60
simplified version of this model, known as the Eden model [8], and show that the
boundary values of the associated process of conformal mappings converge to the
Brownian web.
We begin by defining a class of stochastic flows which result from iteratively
applying small localized perturbations to the circle at uniformly distributed points.
By scaling the rate at which we apply these perturbations appropriately, the flows
converge to the Brownian web as the size of the individual perturbations ap-
proaches zero. We consider the case of simplified Hastings-Levitov DLA where
the incoming particles are slits of length N−1 sticking to the unit disc. If time is
scaled in such a way that particles arrive as a Poisson process of rate proportional
to N3, the resulting flow map, restricted to points on the unit circle, evolves as
an element of our class of stochastic flows and so converges to the Brownian web.
The power of N by which time is scaled is curious, and may give rise to the fractal
behaviour which can be observed in simulations of the model.
The paper [10] of Fontes, Isopi, Newman and Ravishankar, the original work
characterizing the Brownian web, constructs it as a random element of the space of
compact collections of paths with specified starting points. In this chapter we take
an alternative approach and formulate the Brownian web as an element of a space
of flows. We believe that working in a space whose structure inherently contains
the restrictions imposed by the Brownian web is more natural and find that this
simplifies characterization and convergence results. In particular, we prove that
there is a unique measure on our space of flows with respect to which the finite
dimensional distributions of the flows are those of coalescing Brownian motions.
This contrasts with the results of Fontes, Isopi, Newman and Ravishankar, where
they find that there are other natural measures on their space with this property.
We also show that any sequence of flows, whose finite dimensional distributions
converge to those of the Brownian web, converges to the Brownian web. It is
interesting to note that tightness is automatically satisfied.
This chapter is organized as follows. In Section 4.2 we construct the class of
flows on the circle and show that the finite dimensional distributions converge to
those of coalescing Brownian motions. The simplified Hastings-Levitov DLA model
is discussed in Section 4.3 and scaling limits are established for the boundary of
the associated process. A metric space of flows is constructed in Section 4.4, and
Chapter 4. The Brownian web 61
characterization and weak convergence results are formulated for the Brownian web
as an element of this space. Section 4.5 deals with some technical issues pertaining
to the space of flows, and Section 4.6 looks more closely at the correspondence
between our results and those of Fontes, Isopi, Newman and Ravishankar. In
particular, we show that the space of flows on which we construct the Brownian
web is isomorphic to a subspace of the space of compact sets of functions on which
they construct the Brownian web.
4.2 A Levy flow on the circle
In this section we construct a family of stochastic flows on the circle and show that
under certain conditions they converge to Arratia’s flow of coalescing Brownian
motions. In Section 4.2.1 we describe the class of functions on the circle which will
form the “building blocks” of our flows. In Section 4.2.2 we construct the flows
themselves, and in Section 4.2.3 we define Arratia’s flow of coalescing Brownian
motions and show that the finite dimensional distributions of our flows converge
to coalescing Brownian motions.
4.2.1 Some generalities for functions on the circle
We identify the unit circle S1 with the set R/Z ∼ [0, 1). We say that a nonde-
creasing function f : R → R is of degree 1 on the circle if
f(x + n) = f(x) + n for every n ∈ Z. (4.1)
Such maps correspond to functions S1 → S1 in an obvious way. This correspon-
dence is not strictly bijective, however, as for any m ∈ Z, both f and f +m give
rise to the same map.
Let D0 denote the set of all nondecreasing maps of degree 1 on the circle. We
can define an equivalence relation on this set by f ∼ g if f(x) = g(x) for all
points x at which f is continuous. Let [D0] be the set of all equivalence classes of
nondecreasing degree 1 maps on the circle. Every element of [D0] has a unique right
Chapter 4. The Brownian web 62
(and left) continuous version. Let C0 denote the set of all contractions g : R → R
which are periodic with period 1. Define a map Φ : [D0] → C0 by
Φ(f)(t) = t− x where1
2(x+ f(x−)) 6 t 6
1
2(x + f(x+)).
A geometrical interpretation of the function Φ(f) is that it is the map obtained
from f by rotating the axes by π/4 and scaling appropriately (see Figure 4.1).
10−1
−1
1
x
f(x) t
tΦ( )( )f
Figure 4.1: The map Φ(f) obtained from f by rotating the axes by π4.
Since f ∈ D0 is nondecreasing, Φ(f) is well defined and, by (4.1), is periodic
with period 1. Suppose that t > s, and x > y are such that Φ(f)(t) = t − x and
Φ(f)(s) = s− y. If x = y, then Φ(f)(t) − Φ(f)(s) = t− s. Otherwise,
Φ(f)(t) − Φ(f)(s) = t− s− (x− y) < t− s.
Also, if x 6= y,
Φ(f)(t) − Φ(f)(s) = −(t− s) + (2t− x) − (2s− y)
> −(t− s) + f(x−) − f(y+)
> −(t− s).
Chapter 4. The Brownian web 63
Hence,
|Φ(f)(t) − Φ(f)(s)| 6 |t− s|,
and so Φ(f) ∈ C0. The map Φ is invertible with the (right continuous representative
of the) inverse Φ−1 : C0 → [D0] given by
Φ−1(g)(x) = supt + g(t) : x = t− g(t).
Define a metric on [D0] by
dD0(f, g) = ‖Φ(f) − Φ(g)‖ = supt∈[0,1)
|Φ(f)(t) − Φ(g)(t)|.
The condition dD0(f, g) < ε is equivalent to
f(x− ε) − ε < g(x−) 6 g(x+) < f(x+ ε) + ε
for all x ∈ R and so, if dD0(fn, f) → 0 as n → ∞, then fn(x) → f(x) for every
point x at which f is continuous. Lemma 4.11 shows that the converse is also
true. Since (C0, ‖ ·‖) is a complete separable metric space, ([D0], dD0) is a complete
separable metric space.
Maps in D0 can be thought of as being perturbations of the identity map. For
each non-zero f ∈ D0 define f : R → R by f(x) = f(x) − x. The function f is
periodic with period 1. Let η = η(f) be the positive real number that satisfies
η
∫
[0,1)
f(x)2dx = 1. (4.2)
Suppose that |f(x)| is maximized at some xm ∈ [0, 1). Since f is increasing,
|f(x)| >12‖f‖ for all x ∈ [xm, xm + ‖f/2‖] if f(xm) > 0, or x ∈ [xm − ‖f /2‖, xm]
if f(xm) 6 0. Hence,
‖f‖ ∧ 1 6 2η−13 , (4.3)
so large values of η mean that the perturbation of f about the identity map is
small.
Chapter 4. The Brownian web 64
Define the drift b = b(f) by
b = η
∫
[0,1)
f(x)dx. (4.4)
Let δ = δ(f) > 0 be the smallest positive real number that satisfies
supδ6a61−δ
η
∫
[0,1)
|f(x + a)f(x)|dx 6 δ. (4.5)
Small values of δ mean that the perturbation of f about the identity map is well
localized on the circle.
Note that since the set of discontinuities of any f ∈ D0 has Lebesgue measure
zero, these values are well defined for elements of [D0].
4.2.2 Construction of the flow
Given f ∈ D0, we construct a compensated Poisson process which results from
applying f , ‘centered’ at uniformly distributed points, at rate η.
Let Ω denote the set of integer-valued measures on R × [0, 1) which are finite
on bounded sets, excluding those measures with two or more atoms at the same
time point i.e. for all ω ∈ Ω, ω(t× [0, 1)) ∈ 0, 1 for all t ∈ R. Write F o for the
σ-algebra on Ω generated by evaluations on open sets.
We first construct the process for f ∈ D0 with b = 0, where b is defined in
(4.4). Given ω ∈ Ω, for each t ∈ R define Ft ∈ D0 by
Ft(x) =
u+ f(x− u) if ωt×[0,1) = δ(t,u)
x otherwise.
Write Xts(x) = X(·, s, t, x), where X = Xf is the map X : Ω × (s, t) : s 6
t × R → R defined as follows. For each s 6 t, construct Xts ∈ D0 by
Xts = FTn · · · FT1 id,
Chapter 4. The Brownian web 65
where T1 < · · · < Tn are the times of the atoms of ω in (s, t]× [0, 1). Equivalently,
define Xts for t > s recursively at jump times by Xss = id and Xts = Ft Xt−,s for
all t > s.
There is a unique probability measure P = Pη on (Ω,F o) making the identity
map µ(ω, dt, du) = ω(dt, du), ω ∈ Ω, into a Poisson random measure on R × [0, 1)
with intensity ν(dt, du) = ηdtdu, where η = η(f) is as in (4.2). We write F for the
completion of F o with respect to P, extending P to F as usual. For s, t ∈ R with
s 6 t, let Fst denote the completion with respect to P of the σ-algebra generated
by µ(U) for open sets U ⊆ [s, t] × [0, 1).
With respect to P, for each x ∈ R, Xts(x) is an Fts-martingale starting from
x, obtained by applying f , ‘centered’ at uniformly distributed points, at times of
a Poisson process with rate η.
We now extend this to f ∈ D0 with b taking any real value. Set
ωb(dt, du) = ω(dt, d(u+ bt))
and apply the preceding construction with ω replaced by ωb in the definition of Ft
to obtain the map Xb. Define
Xts(x) = Xbts(x+ bs) − bt.
As above, with respect to P, for each x ∈ R, Xts(x) is an Fts-martingale starting
from x, obtained by applying f , ‘centered’ at uniformly distributed points, at times
of a Poisson process with rate η.
4.2.3 Convergence to the Arratia flow
The Brownian web is a continuous family of coalescing one dimensional Brownian
motions. We give a full characterization in Section 4.4, where we prove that as
η → ∞ and δ → 0, the distribution of X converges to that of the Brownian web.
In this section we define what we mean by a finite dimensional flow of coalescing
Brownian motions and the Arratia flow on the circle. We show that if f ∈ D0 with
η sufficiently large and δ sufficiently small, then X approximates the Arratia flow.
Chapter 4. The Brownian web 66
Definition 4.1. An n-dimensional system of coalescing Brownian motions on the
circle, starting at x, is an n-dimensional random process B = (B1, . . . , Bn) with
the following properties.
(i) For each 1 6 i 6 n, the continuous process (Bi(t))t>0 is a standard Brownian
motion with respect to FB, where FB denotes the filtration generated by B,
with Bi(0) = xi.
(ii) For each pair 1 6 i < j 6 n, the process (Bj(t) − Bi(t))t>0 is a Brownian
motion (with diffusivity σ2 = 2) with respect to FB, stopped at the first
time it hits Z.
The intuitive interpretation is of a family of Brownian motions on the circle,
the paths of any two of which are independent until they meet, at which point
they coalesce.
Definition 4.2. An n-dimensional flow of coalescing Brownian motions on the
circle, starting at (x1, s1), . . . , (xn, sn), is an n-dimensional random process B =
(B1, . . . , Bn) with the following properties.
(i) For each 1 6 i 6 n, the continuous process (Bi(t))t>siis a standard Brownian
motion with respect to FB, where FB denotes the filtration generated by B,
with Bi(si) = xi.
(ii) For each pair 1 6 i < j 6 n, the process (Bj(t)−Bi(t))t>si∨sjis a Brownian
motion (with diffusivity σ2 = 2) with respect to FB, stopped at the first
time it hits Z.
Definition 4.3. The Arratia flow on the circle is a family of random processes
(Bsx(t))t>s : s, x ∈ R, with Bsx(s) = x, such that for any deterministic n ∈ N any
(s1, x1), . . . , (sn, xn) ∈ R2, Bs1x1 , . . . , Bsnxn is an n-dimensional flow of coalescing
Brownian motions on the circle.
The construction of the above objects is discussed in Arratia [2].
In the following theorem we show that the finite dimensional distributions of
the flows constructed in Section 4.2.2 converge to those of the Arratia flow in the
Chapter 4. The Brownian web 67
following sense. Fix n ∈ N and (s1, x1), . . . , (sn, xn) ∈ R2. Suppose that for each
f ∈ D0, µf is the law of ((Xts1(x1))t>s1 , . . . , (Xtsn(xn))t>sn), and µB is the law of
(Bs1x1, . . . , Bsnxn). Then
((Xts1(x1))t>s1, . . . , (Xtsn(xn))t>sn) ⇒ (Bs1x1, . . . , Bsnxn)
as η → ∞ and δ → 0 means
supd(µf , µB) : f ∈ D0, η(f) > N, δ(f) 6 N−1 → 0
as N → ∞, where d is the Prohorov metric on the space of probability measures
(see Ethier and Kurtz [9]).
Throughout the rest of this chapter we use interchangeably the notation Xts
and Xt,s and, similarly, Bsx and Bs,x.
Theorem 4.4. Let X be constructed as above, and η, δ be defined as in (4.2) and
(4.5). Then the following hold.
(i) As η → ∞,
(Xts(x))t>s ⇒ (Bsx(t))t>s,
where (Bsx(t))t>s is a standard Brownian motion, starting from x at time s.
(ii) Given any n ∈ N and (s1, x1), . . . , (sn, xn) ∈ R2,
((Xts1(x1))t>s1 , . . . , (Xtsn(xn))t>sn) ⇒ B
as η → ∞ and δ → 0, where B is a flow of coalescing Brownian motions
starting at (s1, x1), . . . , (sn, xn).
Proof. (i) Denote Xt = Xts(x), t > s. By Ito’s formula (see, for example,
Kallenberg [24]),
(
exp
(
iθXt −(∫
(s,t]×[0,1)
(
eiθf(Xr−u) − 1 − iθf(Xr − u))
ηdrdu
)))
t>s
Chapter 4. The Brownian web 68
is a martingale and hence, for any θ ∈ R and t2 > t1 > s,
E(eiθ(Xt2−Xt1 ))
= E
(
exp
(∫
(t1,t2]×[0,1)
(
eiθf(Xr−u) − 1 − iθf(Xr − u))
ηdrdu
))
= E
(
exp
(∫
(t1,t2]×[0,1)
∫ 1
0
η(iθf(Xr − u))2(1 − z)eizθf(Xr−u)dzdrdu
))
= exp
(
−θ2(t2 − t1)
∫ 1
0
(1 − z)η
∫
[0,1)
f(u)2eizθf(u)dudz
)
,
where the second equality follows by Taylor’s theorem and the third follows
by Fubini’s theorem and substituting u for Xr − u.
By (4.2) and (4.3),
∣
∣
∣
∣
η
∫
[0,1)
f 2(u)eizθf(u)du− 1
∣
∣
∣
∣
6 z|θ|η∫
[0,1)
|f(u)|3du
6 z|θ| ‖f‖→ 0
as η → ∞. Hence,
∣
∣
∣
∣
E(eiθ(Xt2−Xt1 )) − exp
(
−1
2θ2(t2 − t1)
)∣
∣
∣
∣
61
2|θ|3(t2 − t1)‖f‖ → 0
as η → ∞.
As Xt has independent increments, the finite dimensional distributions con-
verge to those of Brownian motion.
It remains to show that the family of laws of (Xt)t>s for all f ∈ D0 is tight.
We prove the more general result that for any n ∈ N, (s1, x1), . . . , (sn, xn) ∈R2 the family of laws of ((Xts1(x1))t>s1 , . . . , (Xtsn(xn))t>sn), f ∈ D0 is tight.
Chapter 4. The Brownian web 69
Suppose that (s, x) ∈ R2 and s 6 r 6 u 6 t. Then
E(|Xus(x) −Xrs(x)|2|Xts(x) −Xus(x)|2)= E(|Xus(x) −Xrs(x)|2)E(|Xts(x) −Xus(x)|2)
= E
(∫
(r,u]×[0,1)
f(Xrs(x) − u)2ηdrdu
)
×E
(∫
(u,t]×[0,1)
f(Xrs(x) − u)2ηdrdu
)
= (u− r)(t− u)
6 (t− r)2.
Hence (see Billingsley [3], page 143), (Xts(x))t>s ⇒ (Bsx(t))t>s as η → ∞.
(ii) We first show the following. For any x < y ∈ [0, 1), let (Wts)t>s be a
Brownian motion on the circle with diffusivity σ2 = 2, starting from y − x.
Given any ε > 0, let Sε = inft > s : Xts(y) − Xts(x) /∈ (ε, 1 − ε) and
Tε = inft > s : Wts /∈ (ε, 1 − ε). Then as η → ∞ and δ → 0,
(XSεts (y) −XSε
ts (x))t>s ⇒ (W Tεts )t>s.
Suppose that X ′ is constructed on the same probability space as X, with X
and X ′ independent and identically distributed. Define
Zt(x) = Xts(x) t > s,
Zt(y) =
Xts(y) s 6 t 6 Sε,
X ′tSε
(XSεs(y)) t > Sε.
Chapter 4. The Brownian web 70
For any θ ∈ R and t > s, as in the proof of (i),
E(eiθ(Zt(x)−Zt(y)))
= E(
eiθ(Zt∧Sε(x)−Zt∧Sε(y))E(eiθ((Zt(x)−Zt∧Sε(x))−(Zt(y)−Zt∧Sε(y)))|FSε))
= E
(
exp
(
− θ2
∫
(s,t∧Sε]×[0,1)
∫ 1
0
η(f(Zr(x) − u) − f(Zr(y) − u))2
(1 − z)eizθ(f(Xr(x)−u)−f(Zr(y)−u))dzdrdu
)
× E(eiθ(Zt(x)−Zt∧Sε(x))|FSε)E(eiθ(Zt(y)−Zt∧Sε(y))|FSε)
)
.
Now provided δ < ε, by (4.5),
∣
∣
∣
∣
η
∫
[0,1)
f(Zr(x) − u)f(Zr(y) − u)eizθ(f(Zr(x)−u)−f(Zr(y)−u))du
∣
∣
∣
∣
6 δ,
for all s 6 r 6 Sε, and so, by a similar argument to (i),
∣
∣E(eiθ(Zt(x)−Zt(y))) − exp(
−θ2(t− s))∣
∣
=∣
∣
∣E(eiθ(Zt(x)−Zt(y))) − E
(
e−θ2(t∧Sε−s)e−12θ2(t−t∧Sε)e−
12θ2(t−t∧Sε)
)∣
∣
∣
→ 0
as η → ∞ and δ → 0.
Convergence of the finite dimensional distributions and tightness can be
shown by similar arguments to (i). Hence, (Zt(x) − Zt(y))t>s ⇒ (Wts)t>s
and so (XSεts (x) −XSε
ts (y))t>s ⇒ (W Tεts )t>s, as η → ∞ and δ → 0.
We now prove the general result. Suppose that for each f ∈ D0 the law of
((Xts1(x1))t>s1 , . . . , (Xtsn(xn))t>sn) is given by µf . By (i) we know that the
family of laws µf : f ∈ D0 is tight. Hence, every sequence with η → ∞and δ → 0 has a subsequence which converges weakly. We prove that any
such weak limit law is equal to µB, the law of (Bs1x1, . . . , Bsnxn).
It is enough to show that for all x < y ∈ [0, 1), if (Xts(y) − Xts(x))t>s ⇒(Yts)t>s, then (Yts)t>s ∼ (BT
ts)t>s, where (Bts)t>s is a Brownian motion on the
Chapter 4. The Brownian web 71
circle with diffusivity σ2 = 2, starting from y − x and T = inft > s : Bts ∈0, 1. Since our initial result is true for all values of ε > 0, and (Yts)t>s is
almost surely continuous, (Y Sts )t>s ∼ (BT
ts)t>s, where S = inft > s : Yts ∈0, 1. It remains to show that Yt+S,s
S<∞ = YSs
S<∞ almost surely for all
t > 0.
We show that (Yts)t>s is a martingale (with respect to the natural filtration)
by the following standard argument. Suppose that s 6 t1 < · · · < tn 6 t < u.
Since Xts has independent increments of mean 0, and Xts(y) −Xts(x) takes
values in the interval [0, 1], if g : Rn → R is any bounded continuous function
E((Xus(y) −Xus(x))g(Xt1s(y) −Xt1s(x), . . . , Xtns(y) −Xtns(x)))
= E((Xts(y) −Xts(x))g(Xt1s(y) −Xt1s(x), . . . , Xtns(y) −Xtns(x))).
But
E((Xus(y) −Xus(x))g(Xt1s(y) −Xt1s(x), . . . , Xtns(y) −Xtns(x)))
→ E(Yusg(Yt1s, . . . , Ytns))
and
E((Xts(y) −Xts(x))g(Xt1s(y) −Xt1s(x), . . . , Xtns(y) −Xtns(x)))
→ E(Ytsg(Yt1s, . . . , Ytns)).
Therefore (Yts)t>s is a martingale taking values in [0, 1]. Hence, by the op-
tional stopping theorem,
E(Yt+S,s
S<∞,YSs=0) = E(YSs
S<∞,YSs=0) = 0
and so Yt+S,s
S<∞,YSs=0 = 0 almost surely. Similarly Yt+S,s
S<∞,YSs=1 =
S<∞ almost surely, as required.
Chapter 4. The Brownian web 72
4.3 Hastings–Levitov DLA
In this section we describe diffusion-limited aggregation (DLA) and show how the
processes defined in Section 4.2.2 occur naturally in a simplified version of the
Hastings-Levitov model for planar DLA [17].
Diffusion-limited aggregation is a random growth model which was originally
introduced in 1981 by Witten and Sander [35]. Consider the unit disc at the origin
of the plane R2. A particle is introduced at “infinity” and performs a Brownian
motion until it contacts the unit disc, at which point it sticks. A second particle
is introduced and performs a Brownian motion until it contacts either the disc or
the first particle at which point it sticks. Further particles are introduced creating
a large tree-like structure, the shape of which is strongly dependent on the size of
the incoming particles. Of particular interest is the limiting case where the particle
size tends to zero as the rate of arrivals tends to infinity. Simulations suggest that
an incoming particle is most likely to attach to the tips of the cluster; the resulting
structure is therefore highly branched and fractal.
Hastings and Levitov formulated a model of DLA in which the cluster is repre-
sented by a sequence of iterated conformal maps. We describe a simplified version
of their construction below.
Suppose that D is the open unit disc, and let A be a compact subset of C,
of diameter r0 such that K = A ∪ D is simply connected. Set D0 = C \ D and
D = C \K. There is a unique conformal map g : D → D0 and a unique constant
κ ∈ [0,∞) such that g(z) ∼ e−κz as z → ∞ (see, for example, Lawler [27]).
For θ ∈ [0, 2π) and z ∈ eiθD, set gθ(z) = eiθg(e−iθz). Let (Θn)n∈N be a sequence
of independent random variables, distributed uniformly on [0, 2π) and let (νt)t>0
be a Poisson process of rate 1, with jump times 0 = T0 < T1 < · · · .
For each z ∈ D0, define a C-valued jump process (Zt(z))t<ζ(z) as follows. Set
Z0(z) = z, and initially set ζ(z) = ∞. While ζ(z) = ∞, recursively for n > 0,
check whether ZTn(z) ∈ eiΘn+1K. If so, set ζ(z) = Tn+1; otherwise, set ZTn+1(z) =
gΘn+1(ZTn(z)). Then, for Tn 6 t < Tn+1 ∧ ζ(z), let Zt(z) = ZTn(z).
We define set-valued jump processes (Kt)t>0, (Dt)t>0 by Dt = z ∈ D0 : ζ(z) >
t and Kt = C \ Dt. Note that Zt is the unique conformal map Dt → D0 with
Chapter 4. The Brownian web 73
(a) The cluster after a few arrivalswith N = 1.
(b) The cluster after 100 arrivalswith N = 1.
(c) The cluster after 800 arrivalswith N = 10.
(d) The cluster after 5000 arrivalswith N = 25.
(e) The cluster after 20000 arrivalswith N = 50.
(f) The stochastic flow (Xt0)t∈[0,1]
with N = 50.
Figure 4.2: The slit model case of simplified Hastings-Levitov DLA
Chapter 4. The Brownian web 74
Zt(z) ∼ e−κνtz as z → ∞. The set Kt represents the cluster formed by particles
with shape A \ D and hence studying the process Zt gives an insight into the
evolving shape of the cluster.
This model is a simplification of Hastings-Levitov DLA. At each iteration we
are scaling the size of the cluster boundary, whilst keeping the size and shape of
the set A unchanged, whereas for Hastings-Levitov DLA it is necessary to scale
A appropriately as well. As the scaling depends on the position on the cluster
boundary at which the particle attaches itself, this complicates the problem signif-
icantly and, in particular, the resulting process is no longer Markov. As such our
model is a toy model, and differs in structure from DLA. For example, it is easy
to show that the cluster formed by the above model has only one infinite branch,
whereas simulations of actual DLA appear to have several. This simplification is
known as the Eden model [8], and describes the growth of bacterial cells or tissue
cultures of cells that are constrained from moving. Figure 4.2(b) illustrates how,
in the simplified model, the particles arriving later tend to be larger than those
arriving earlier.
The advantage of this model, however, is that the location at which each arriv-
ing particle attaches itself is correctly distributed. A particle performing Brownian
motion will meet the cluster in a location whose distribution is determined by the
harmonic measure along the boundary of the cluster. As conformal mappings pre-
serve harmonic measure, the location is precisely that obtained by ‘centering’ the
map g at a uniformly distributed point on the boundary of the unit disc.
We are interested in the asymptotics of the processes Z (and the corresponding
clusters K) in the limit as r0 (and hence κ) tends to 0. In particular, we study the
case where A is the slit y : 1 6 y 6 1+N−1 as N → ∞ (see Figure 4.2). In this
case g(z) = m−1 r h 14(
11+2N )
2 m(z), where
m(z) =iz − i
z + 1
is the Mobius transformation taking the unit circle to the real line,
r(z) =z
√
1 −(
11+2N
)2
Chapter 4. The Brownian web 75
is a linear scaling, and
ht(z)2 = 4t+ z2
is the Loewner transform taking H \ z = si : s ∈ (0, t] → H, the branch of the
square root to be used being determined accordingly. This gives
g(z) =
(
1 −(
1
1 + 2N
)2)
z + 1
2z
z + 1 +
√
√
√
√z2 + 1 − 2z1 +
(
11+2N
)2
1 −(
11+2N
)2
− 1.
The process Z exhibits some interesting and unexpected behaviour. In particu-
lar, let us consider how the boundary of the disc evolves under appropriate scaling.
Let Xts(x) be the position at time t of the point on the boundary that was at x
at time s, under the identification of the boundary with the interval [0, 1).
Scaling time by η (to be determined),
Xts(x) = Ft(Xt−,s(x)),
where Ft is constructed as in Section 4.2.2 from f ∈ D0 given by
f(x+ n) = n +
π−1 tan−1√
tan2 πx+(1+2N)−2
1−(1+2N)−2 x ∈ [0, 12)
−π−1 tan−1√
tan2 πx+(1+2N)−2
1−(1+2N)−2 x ∈ (12, 1),
for each n ∈ N. Then ‖f‖ = Θ(N−1) and b(f) = 0. By considering separately the
contributions when x = O(N−1) and x = ω(N−1), and using appropriate Taylor
expansions, it can be shown that
6N3π3
∫
[0,1)
f(x)2dx→ 1
as N → ∞ and hence η(f) = 6N 3π3 + o(N3) → ∞ as N → ∞.
Also, by considering separately the cases x = O(N−1) with x + a = ω(N−1),
and x = ω(N−1), it can be shown that
N14 sup
N−14 6a61−N−
14
η
∫
[0,1)
|f(x + a)f(x)|dx→ 0
Chapter 4. The Brownian web 76
as N → ∞. Therefore δ(f) = O(N− 14 ) and so, in particular, δ → 0 as N → ∞.
Hence, X satisfies the conditions of Theorem 4.4 and so its finite dimensional
distributions converge to those of the Arratia flow as N → ∞ (see Figure 4.2(f)).
We shall prove in the next section the stronger result that X converges to the
Brownian web.
In future work, we intend to investigate scaling limits for the evolution of the
conformal map in the above model away from the boundary and, therefore, to
obtain a limiting structure for the cluster Kt.
4.4 The Brownian web
The Brownian web is the collection of graphs of coalescing one-dimensional Brown-
ian motions (with unit diffusion constant and zero drift) starting from all possible
points in continuous space-time. This object was originally studied in 1979 by
Arratia [1]. Further work has been carried out by many people including Harris
[15], who was interested in coalescing stochastic flows, and Piterbarg [29], who
showed that the Arratia flow arises as a weak limit of rescaled isotropic stochastic
flows. In 1998 Toth and Werner [30] took an alternative view of the Brownian web,
showing that it could be used to construct a continuous true self-repelling motion.
Fontes, Isopi, Newman and Ravishankar [10] extended this work in 2004 by pro-
viding a new characterization and obtaining convergence results. In 2005, Fontes
and Newman [11] studied a closely related object, the so-called full Brownian web.
In both [10] and [11], the authors obtained results by regarding the Brownian
web as an object in the space of compact sets of functions with specified starting
points. By instead regarding the Brownian web as an element of a space of flows,
we are able to simplify these authors’ characterization and convergence results.
The exact correspondence between our work and that in [10] and [11] is discussed
in Section 4.6. The viewpoint that we take of the Brownian web being an element
of an equivalence class on a space of stochastic flows, is used by Tsirelson [31].
However, he does not explicitly construct a metric space on which to realize the
flows in the way that we do below.
In Section 4.4.1 we define the space of flows. In Section 4.4.2 we construct
Chapter 4. The Brownian web 77
the Brownian web as a random element of this space and prove uniqueness of the
distribution. Section 4.4.3 establishes criteria for convergence to the Brownian web,
and shows that the flows constructed in Section 4.2.2 converge to the Brownian
web as elements of the space of flows. The proofs of the various technical results
referred to in this section can be found in Section 4.5.
4.4.1 A description of the flow space
Throughout the rest of this chapter we limit ourselves to time taking values in the
compact interval [−T, T ] for some fixed T > 0. An alternative approach (used in
[10] and [11]) is to introduce a metric under which R ∪ ∞ is compact. Both
approaches result in identical topologies, but we use the former as it avoids the
technical difficulties resulting from compactification.
Let D0 be the set of nondecreasing degree 1 maps on the circle, and ([D0], dD0)
be the metric space of equivalence classes on D0, defined in Section 4.2.1.
Definition 4.5. We say that φ : (s, t) ∈ [−T, T ] × [−T, T ] : s 6 t × R → R,
denoted by φ(s, t, x) = φts(x), is a cadlag flow on the circle if it satisfies the
following properties.
(a) For all s 6 t, φts ∈ D0.
(b) For all s 6 t 6 u, φus = φut φts.
(c) For all s ∈ [−T, T ], φss = id.
(d) For all s ∈ [−T, T ] and x ∈ R, φ(s, ·, x) : [s, T ] → R is right continuous with
regular left limits, where we say φt−,s(x) is regular if φt−,s(x) → x as s ↑ t.
We define an equivalence relation on this set of flows by φ ∼ φ′ if φts(x) = φ′ts(x)
at all points (s, t, x) for which φts is continuous at x. Let D be the space of all
equivalence classes of cadlag flows on the circle, together with the metric
dD(φ1, φ2) = infλ∈Λ
[γ(λ) ∨ sups6t
dD0(φ1ts, φ
2λ(t)λ(s))],
Chapter 4. The Brownian web 78
where Λ is the set of strictly increasing functions λ mapping [−T, T ] onto itself for
which
γ(λ) = sup−T6t<u6T
logλ(u) − λ(t)
u− t<∞
(cf. the Skorohod metric (see Billingsley [3]) on the space DR[−T, T ] of cadlag
functions from [−T, T ] to R. The difference here is that the supremum is taken
over the set (s, t) ∈ [−T, T ] × [−T, T ] : s 6 t). (D, dD) is a complete separable
metric space (see Theorem 4.15). Define FD to be the Borel σ-algebra on D.
Definition 4.6. We now give an alternative formulation of D which only depends
on the value of φ at continuity points. A map φ : (s, t) ∈ [−T, T ] × [−T, T ] : s 6
t → [D0], denoted by φ(s, t) = φts, is a cadlag flow on the circle if it satisfies the
following properties.
(a′) For all s 6 t, φts ∈ [D0].
(b′) For all s 6 t 6 u, φus(x) = φut(φts(x)) at every point x for which φts is
continuous at x and φut is continuous at φts(x).
(c′) For all s ∈ [−T, T ], φss = id.
(d′) Given ε > 0, for each s ∈ [−T, T ] there exists δ > 0 such that ‖φut − id‖ < ε
for all s 6 t 6 u < s+ δ and all s− δ < t 6 u < s.
It is shown in Lemma 4.14 that if φ satisfies the above conditions, then there exists
some φ′ ∼ φ that satisfies the conditions in Definition 4.5.
In what follows, we use interchangeably the same notation to mean a class
representative of an element of D, an element of D defined as in Definition 4.5,
and an element of D defined as in Definition 4.6. Lemma 4.14 shows that it is
consistent to view these three objects as the same thing.
Definition 4.7. We say that φ : (s, t) ∈ [−T, T ] × [−T, T ] : s 6 t × R → R,
denoted by φ(s, t, x) = φts(x), is a compact flow on the circle if conditions (c) and
(d) in Definition 4.5 are replaced by the following condition.
(e) For every ε > 0, there exists some δ > 0 such that for all s 6 t 6 s + δ,
‖φts − id‖ < ε.
Chapter 4. The Brownian web 79
Let C ⊂ D be the space of all equivalence classes of compact flows on the circle.
The metric dD restricted to C simplifies to
dC(φ1, φ2) = sup
s6tdD0(φ
1ts, φ
2ts).
Define FC to be the Borel σ-algebra on C.
4.4.2 Existence and uniqueness of the Brownian web
The following theorem characterizes the Brownian web.
Theorem 4.8. There exists a (C,FC)-valued random variable W whose distri-
bution is uniquely determined by the following property. If E is any determinis-
tic countable dense subset of [−T, T ] × [0, 1) (dense on −T × [0, 1)), then for
any deterministic m ∈ N and (s1, x1), . . . , (sm, xm) ∈ E , the joint distribution of
Wts1(x1), . . . ,Wtsm(xm) is that of coalescing Brownian motions on the circle (with
unit diffusion constant).
Proof. We first prove that the distribution determined by the condition in Theorem
4.8 is unique. Suppose that W 1,W 2 are two such C-valued random variables.
Define S to be the d-system given by
S = C ∈ FC : P(W 1 ∈ C) = P(W 2 ∈ C).
Now S contains the π-system consisting of all finite intersections of open cylinders,
defined in Definition 4.16 in the following section. By Lemma 4.17, the open
cylinders of C generate the σ-algebra FC and so, by Dynkin’s π-system Lemma,
S = FC. Hence, W 1 and W 2 have the same distribution.
We construct a class representative of a C-valued random variable with the
required property as follows. Let (Ω,F ,P) be a probability space on which (Bj)j∈N,
an independent identically distributed family of standard Brownian motions, is
defined. Suppose that E = (sj, xj) : j ∈ N. For each j ∈ N, n ∈ Z, we define
W nj to be the Brownian path starting at position xj + n at time sj given by
W nj (t) = xj + n+Bj(t− sj), t > sj.
Chapter 4. The Brownian web 80
Using the method of Arratia [2], we construct coalescing Brownian paths out of
the family of paths (W nj )j∈N,n∈Z by specifying coalescing rules. When two paths
meet for the first time, they coalesce into a single path, which is that of the
Brownian motion with the lower index. Denote the coalescing Brownian paths by
(W nj )j∈N,n∈Z. Define a random variable W : Ω × (s, t) ∈ [−T, T ] × [−T, T ] : s 6
t × R → R by
Wts(x) = infW nj (t) : sj 6 s, W n
j (s) > x.
By definition, for all (sj, xj) ∈ E , (Wtsj(xj))t>sj
= (W 0j (t))t>sj
and so W has the
required finite dimensional distributions. It is proved by Arratia in Section 5 of
[2] that almost surely W satisfies the conditions to be a class representative of a
C-valued random variable.
In what follows, we fix E to be a deterministic countable dense subset of
[−T, T ] × [0, 1) (dense on −T × [0, 1)). By the uniqueness theorem proved
above, all results are independent of the choice of E .
4.4.3 Convergence to the Brownian web
In this section, we establish criteria under which D-valued random variables con-
verge to the Brownian web, and show that, under specified conditions, the stochas-
tic flows constructed in Section 4.2.2 converge to the Brownian web.
Theorem 4.9. Suppose that (Xn)n∈N is a sequence of (D,FD)-valued random
variables. If, for any deterministic m ∈ N and (s1, x1), . . . , (sm, xm) ∈ E , the joint
distribution of Xnts1
(x1), . . . , Xntsm
(xm) converges as n → ∞ to that of coalescing
Brownian motions on the circle (with unit diffusion constant), then the distribution
µn of Xn converges to the distribution µW of the Brownian web.
Proof. Let E = (si, xi) : i ∈ N and define µm = limn→∞ µmn , where µm
n is the
law of (Xnts1
(x1), . . . , Xntsm
(xm)). By Kolmogorov’s existence theorem, there exists
a measure µ on the space of families of continuous functions fi : [si, T ] → R :
i ∈ N : fi(si) = xi such that µ (πm)−1 = µm where πm is the natural projection
map from fi : [si, T ] → R : i ∈ N : fi(si) = xi to fi : [si, T ] → R : i 6
m : fi(si) = xi. Suppose that Bsixi: i ∈ N is distributed with law µ. As
Chapter 4. The Brownian web 81
in the proof of existence in Theorem 4.8, the random variable W : Ω × (s, t) ∈[−T, T ] × [−T, T ] : s 6 t × R → R, defined by
W (·, s, t, x+m) = Wts(x +m) = m + infBsjxj(t) : sj 6 s, Bsjxj
(s) > x
for all s 6 t, x ∈ [0, 1), m ∈ Z, is almost surely a class representative of a C-valued
random variable, whose distribution is that of the Brownian web.
It remains to show that the distributions of the Xn converge to that of W as
elements of the metric space (D, dD). But this is an immediate consequence of
Proposition 4.18 in the next section.
Theorem 4.10. Let X be constructed as in Section 4.2.2, and define η, δ as in
(4.2) and (4.5) respectively. Then as η → ∞ and δ → 0, the distribution of X
converges to that of the Brownian web.
Proof. Suppose that f ∈ D0 and that X = Xf is constructed as in Section 4.2.2.
Since Xts = Ft Xt−,s almost surely for some Ft ∈ D0, X (with time restricted to
the compact set [−T, T ]) satisfies the conditions in Definition 4.5 and so is a (class
representative of a) D-valued random variable.
By Theorem 4.4, the conditions of Theorem 4.9 are satisfied and hence the
distribution of X converges to that of the Brownian web.
4.5 Some properties of (D, dD)
This section contains the proofs of various technical results, pertaining to the space
(D, dD), which are referred to elsewhere in this chapter.
Lemma 4.11. Suppose that (fn)n>1 is a sequence in D0, with fn(x) → f(x) for
every x at which f is continuous. Then dD0(fn, f) → 0 as n→ ∞.
Proof. It is enough to show that there exists a subsequence nr for which fnr → f .
Suppose that tn ∈ [0, 1] are chosen such that
|Φ(fn)(tn) − Φ(f)(tn)| = dD0(fn, f).
Chapter 4. The Brownian web 82
Since [0, 1] is a compact set, there exists a subsequence nr for which tnr → t. Let
xn ∈ R be such that
t ∈ [xn + fn(xn−), xn + fn(xn+)].
By restricting to a further subsequence if necessary, there exists some x ∈ R such
that xnr → x. Given ε > 0, there exist x − ε < y1 < x < y2 < x + ε such that
f is continuous at y1 and y2. Pick N ∈ N sufficiently large that for all nr > N ,
|tnr − t| < ε, y1 < xnr < y2 and |fnr(yi) − f(yi)| < ε for i = 1, 2. Then
f(x− 2ε) + x− 2ε 6 f(y1) + x− 2ε < fnr(y1) + x− ε 6 fnr(xnr−) + xnr
and, similarly,
f(x + 2ε) + x + 2ε > fnr(xnr+) + xnr .
Hence,
t ∈ [x− 2ε+ f(x− 2ε), x+ 2ε+ f(x + 2ε)].
Therefore, if nr > N ,
dD0(fnr , f) = |Φ(fnr)(tnr) − Φ(f)(tnr)|6 |Φ(fnr)(tnr) − Φ(fnr)(t)| + |Φ(fnr)(t) − Φ(f)(t)|
+|Φ(f)(t) − Φ(f)(tnr)|6 2|tnr − t| + |xnr − (x− 2ε)| ∨ |xnr − (x+ 2ε)|< 5ε,
and so fnr → f as required.
Lemma 4.12. Suppose that (fn)n>1 and (gn)n>1 are sequences in D0, with fn → f
and gn → g as n → ∞. Then for every x ∈ R for which g is continuous at x and
f is continuous at g(x), fn gn(x) → f g(x) as n→ ∞.
Proof. Given ε > 0, there exists 0 < δ < ε such that
f(g(x+ δ) + 2δ) + δ < f(g(x)) + ε
Chapter 4. The Brownian web 83
and
f(g(x− δ) − 2δ) − δ > f(g(x)) − ε.
Pick N ∈ N sufficiently large that for all nr > N , dD0(fnr , f) < δ and dD0(gnr , g) <
δ. Then for all y ∈ R,
f(y − δ) − δ < fnr(y−) 6 fnr(y) < f(y + δ) + δ
and
g(y − δ) − δ < gnr(y−) 6 gnr(y) < g(y + δ) + δ.
Hence,
f g(x) − ε < f(g(x− δ) − 2δ) − δ
6 f(gnr(x) − δ) − δ
< fnr(gnr(x)).
Similarly
f g(x) + ε > fnr gnr(x),
and so fnr gnr(x) → f g(x), as required.
Lemma 4.13. Suppose that φ : (s, t) ∈ [−T, T ] × [−T, T ] : s 6 t × R → R,
denoted by φ(s, t, x) = φts(x), satisfies conditions (a), (b) and (c) in Definition
4.5. Then the following are equivalent.
(i) For all s ∈ [−T, T ] and x ∈ R, φ(s, ·, x) : [s, T ] → R is right continuous.
(ii) For all s ∈ [−T, T ] and x ∈ R, φts(x) → x as t ↓ s.
(iii) Given ε > 0, for each s ∈ [−T, T ] there exists δ > 0 such that ‖φut − id‖ < ε
for all s 6 t 6 u < s+ δ.
Similarly, the following are equivalent.
(i′) For all s ∈ [−T, T ] and x ∈ R, φ(s, ·, x) : [s, T ] → R has regular left limits.
(ii′) Given ε > 0, for each s ∈ [−T, T ] and x ∈ R, there exists δ > 0 such that
|φtu(x) − x| < ε for all s− δ < u 6 t < s.
Chapter 4. The Brownian web 84
(iii′) Given ε > 0, for each s ∈ [−T, T ] there exists δ > 0 such that ‖φut − id‖ < ε
for all s− δ < t 6 u < s.
Proof. It is immediate that (i) implies (ii). Suppose that (ii) holds. Then given
s ∈ [−T, T ], for each m ∈ N, there exists some δ > 0 such that s 6 t < s + δ
implies that |φts(i/m) − i/m| < m−1, i = 0, . . . , m− 1. But then if t 6 u < s+ δ,
‖φut − id‖6 sup
i(|φts((i + 2)/m) − (i− 1)/m| ∨ |φts((i− 1)/m) − (i + 2)/m|)
< 4m−1,
and so (iii) holds. To see that (iii) implies (i), suppose that s 6 t ∈ [−T, T ], x ∈ R
and tn ↓ t. Then |φtns(x) − φts(x)| 6 ‖φtnt − id‖ → 0.
The proof of the conditions for left limits is similar.
Lemma 4.14. Suppose that φ satisfies the conditions in Definition 4.6. Then
there exists some φ′ with φ′ts ∼ φts for all s 6 t that satisfies the conditions in
Definition 4.5.
Proof. We first show that, for any φ satisfying the conditions in Definition 4.6,
there are only countably many values of t ∈ [−T, T ] for which φ(s, ·) : [s, T ] → [D0]
has a discontinuity at t for some s 6 t. Let An = t ∈ [−T, T ] : ‖φts − φt−,s‖ >n−1 for some s 6 t. Suppose that (tm)m>1 is a sequence of distinct points in
An. Since [−T, T ] is compact, by restricting to a subsequence if necessary, we
may assume tm ↑ t or tm ↓ t for some t. But then there exists δ > 0 such that
if t 6 u < tm 6 t + δ, then ‖φtms − φus‖ 6 ‖φtmt − id‖ + ‖φut − id‖ < n−1,
and similarly if t − δ < u < tm < t, then ‖φtms − φus‖ 6 ‖φtmu − id‖ < n−1,
contradicting tm ∈ An.
For every s ∈ ([−T, T ] ∩ Q) ∪ (⋃
n>1An) ∪ T, there exists a countable dense
set consisting of points x at which φts is continuous for all rationals t > s. Let Ebe the countable dense set consisting of such pairs (s, x). Define fsx : [s, T ] → R
by fsx(t) = φts(x). The family fsx : (s, x) ∈ E consists of noncrossing paths. Set
φ′ts(x) = inffry(t) : r 6 s, fry(s) > x, (r, y) ∈ E.
Chapter 4. The Brownian web 85
It is straightforward to check that φ′ satisfies conditions (a), (b) and (c) in Def-
inition 4.5, and that φ′ts ∼ φts for all s 6 t. Condition (d) follows by Lemma
4.13.
Theorem 4.15. The spaces (D, dD) and (C, dC), defined in Section 4.4.1, are
complete separable metric spaces.
Proof. To prove completeness, it is enough to prove that every Cauchy sequence
in (D, dD) contains a convergent subsequence. Suppose that (ψn)n>1 is a Cauchy
sequence in (D, dD). There exists a subsequence (φr)r>1 = (ψnr)r>1 such that
dD(φn, φn+1) < 2−n. Then Λ contains a sequence (µn)n>1 for which γ(µn) < 2−n
and
sups6t
dD0(φnts, φ
n+1µn(t)µn(s)) = sup
s6tdD0(φ
nµ−1
n (t)µ−1n (s)
, φn+1ts ) < 2−n.
As in the proof of the completeness of the Skorohod space DR[−T, T ] (see Billings-
ley [3]),
µn+m · · · µn → λn
as m→ ∞ for some λn ∈ Λ, with γ(λn) 6 2−(n−1) and
sups6t
dD0(φnλ−1
n (t)λ−1n (s)
, φn+1
λ−1n+1(t)λ
−1n+1(s)
) < 2−n.
Hence, for all s 6 t, (φnλ−1
n (t)λ−1n (s)
)n>1 is a Cauchy sequence in [D0] and, as [D0] is
complete, there exist φts ∈ [D0] for which φnλ−1
n (t)λ−1n (s)
→ φts. Define φ : (s, t) ∈[−T, T ] × [−T, T ] : s 6 t → [D0], by φ(s, t, x) = φts(x). Conditions (a′) and
(c′) in Definition 4.6 are immediate. Condition (b′) follows from Lemma 4.12, and
condition (d′) follows from the property that
‖f − id‖ = 2dD0(f, id)
for all f ∈ D0. Hence, by Lemma 4.14, there exists some φ′ ∈ D with φ′ ∼ φ.
Then, as for all m > n
γ(λn) ∨ sups6t
dD0(φnλ−1
n (t)λ−1n (s)
, φmλ−1
m (t)λ−1m (s)
) < 2−(n−1),
letting m→ ∞ gives φn → φ′ as n→ ∞.
Chapter 4. The Brownian web 86
To show that (C, dC) is complete, it is enough to check that if (φn)n>1 is a
sequence in C with φn → φ for some φ ∈ D, then φ satisfies condition (e) in
Definition 4.7. This follows by an identical argument to the proof of condition (d′)
above.
To prove separability we first observe that since [D0] is separable, it has a
countable dense subset A0 = fi : i ∈ N. For each n ∈ N, im,k ∈ N, where
0 6 m 6 k 6 n, and −T = t0 < t1 < · · · < tn = T < tn+1, define α : (s, t) ∈[−T, T ] × [−T, T ] : s 6 t → [D0] by
α(s, t) = αts = fim,k, s ∈ [tm, tm+1), t ∈ [tk, tk+1). (4.6)
Let A be the countable collection of all such elements α, where −T = t0 < t1 <
· · · < tn = T < tn+1 are of the form qT for some rational q. Note that in general
such elements α do not lie in D. We shall show that there is a dense subset of D,
indexed by A.
Suppose that φ ∈ D and ε > 0. Recall from the proof of Lemma 4.14 that the
map t 7→ φts has only finitely many discontinuities in the interval (−T, T ) (for all
s 6 t) of size > ε, at s1, . . . , sm, say, where −T = s0 < s1 < · · · < sm < sm+1 = T .
By Lemma 4.13, for each j = 0, . . . , m there exists some 0 < δj < sj+1 − sj such
that if sj 6 s 6 t < sj+1∧(s+δj), then ‖φts−id‖ < ε. Pick some integers M > 1eε−1
and n > 3M(T ∧ 1)(δ0 ∧ · · · ∧ δm)−1 and let tk = −T + knT for k = 0, . . . , n + 1.
Since A0 is dense in [D0], there exist positive integers ik,m, 0 6 m 6 k 6 n such
that
dD0(φtktm , fim,k) < ε
for all 0 6 m 6 k 6 n. Define α ∈ A as in (4.6). For each j = 1, . . . , m withnsj
T/∈ N, there exists some kj such that sj ∈ (tkj
, tkj+1). Let
t′kj=sj − e−εtkj+1
1 − e−ε∈ (tkj−M , sj),
and
t′kj+1=sj − eεtkj+1
1 − eε∈ (tkj+1, tkj+M+1).
Note that by the definition of n, |kj − ki| > 3M for all i 6= j and so the t′kjare
Chapter 4. The Brownian web 87
strictly increasing. Let λ ∈ Λ be the strictly increasing piecewise linear function
which joins (t′kj, t′kj
) to (tkj+1, sj) to (t′kj+1, t′kj+1) for those j = 1, . . . , m for which
nsj
T/∈ N, and which has gradient 1 otherwise. Then
γ(λ) 6 minj
(
log
∣
∣
∣
∣
∣
t′kj+1 − sj
t′kj+1 − tkj+1
∣
∣
∣
∣
∣
∧ log
∣
∣
∣
∣
∣
sj − t′kj
tkj+1 − t′kj
∣
∣
∣
∣
∣
)
= ε.
Also, for any s 6 t, there exist m 6 k with 0 6 t− tk, s− tm < δ0 ∧ · · · ∧ δm such
that
dD0(φts, αλ(t)λ(s)) 6 dD0(φts, φtktm) + dD0(φtktm , αtktm)
< ‖φttk − id‖ + dD0(φtks, φtks φstm) + ε
< 3ε,
where in the last inequality we have used the fact that for any f, g ∈ D0,
dD0(f, f g) 6 ‖g − id‖.
Hence, dD(φ, α) 6 3ε. Observe that although A 6⊂ D, the metric dD extends to
general functions α : (s, t) ∈ [−T, T ]× [−T, T ] : s 6 t → [D0] in an obvious way.
Now, for each α ∈ A, pick
φα ∈ φ ∈ D : dD(φ, α) = infφ′∈D
dD(φ′, α).
Note that φα exists since D is closed. Given ε > 0 and φ ∈ D, there exists α ∈ Asuch that dD(φ, α) < ε. But then dD(φ, φα) 6 dD(φ, α) + dD(α, φα) < 2ε and
hence, the countable set φα : α ∈ A is dense in D.
By an identical argument, it is possible to construct a countable dense subset
of C. Hence, C is separable.
Definition 4.16. Suppose that the sets Ii, i = 0, . . . , n are intervals in R. Let
t0 < · · · < tn ∈ [−T, T ], and define
Ct0,...,tnI0,...,In
= φ ∈ C : there exists x ∈ I0 such that φtit0(x±) ∈ Ii, i = 1, . . . , n.
Chapter 4. The Brownian web 88
We call such sets the open cylinders of C if the Ii are all open, and closed cylinders
of C if the Ii are all closed. It is straightforward to see that the closed cylinders
can be generated by the open cylinders, using countable set operations.
Lemma 4.17. The σ-algebra FC is generated by the open cylinders of C.
Proof. We first show that the open cylinders are open subsets of C. Suppose that
φ ∈ Ct0,...,tnI0,...,In
and that x ∈ I0 is such that φtit0(x±) ∈ Ii for i = 1, . . . , n. Since the
Ii are open intervals, there exists some ε > 0 such that (φtit0(x− ε) − ε, φtit0(x +
ε) + ε) ⊂ Ii for i = 0, . . . , n. Then if φ′ ∈ BC(φ, ε), we have φ′ ∈ Ct0,...,tnI0,...,In
. Hence,
Ct0,...,tnI0,...,In
is open.
We now show that it is possible to generate the closed ε-balls in C with cylinders.
Suppose that φ ∈ C. For s ∈ [−T, T ], n ∈ N, let ti = ti(s) = s+ in(T −s). For each
x ∈ [0, 1), let Ii = Ii(x, s) = [φtis((x − ε)−) − ε, φtis((x + ε)+) + ε], i = 1, . . . , n,
and Cεφ(s, x, n) = Ct0,...,tn
I0,...,In. Let
Cεφ =
⋂
(s,x)∈E
⋂
n∈N
Cεφ(s, x, n).
The closed ball BC(φ, ε) is contained in Cεφ. Conversely, if φ′ ∈ Cε
φ, by considering
sequences tn → t, and sequences (sn, xn) ∈ E with φssn(xn) → x, it can be shown
that for every s 6 t, dD0(φ′ts, φts) 6 ε. Hence, Cε
φ ⊂ BC(φ, ε), as required.
Proposition 4.18. Suppose that (φn)n∈N is a sequence in D and that for each
(s, x) ∈ E the process (φnts(x))t>s converges (uniformly) to some continuous R-
valued process (fsx(t))t>s. If the flow map φ : (s, t) ∈ [−T, T ] × [−T, T ] : s 6
t × R → R defined by
φ(s, t, x+m) = φts(x +m) = m + inffsjxj(t) : sj 6 s, fsjxj
(s) > x
for all s 6 t, x ∈ [0, 1), m ∈ Z, is a class representative of a C-valued random
variable, then dD(φn, φ) → 0.
Proof. Given ε > 0, pick some m ∈ N with m > 5/ε. There exist −T = a0 <
· · · < ak = T such that ‖φsai− id‖ < m−1 for all s ∈ [ai, ai+1), i = 0, . . . , k − 1.
Chapter 4. The Brownian web 89
Let yj = j/m for j = 0, . . .m. For each 0 6 i 6 k and 0 6 j 6 m, there exists
some (si,j, xi,j) ∈ E for which
fsi,jxi,j(t) −m−1 < φtai
(yj) 6 fsi,jxi,j(t),
for all t > ai. Pick N ∈ N sufficiently large that
supt>si,j
|φntsi,j
(xi,j) − fsi,jxi,j(t)| < m−1.
Suppose s 6 t with s ∈ [ai, ai+1). For any x ∈ [0, 1), there exists some j such that
x ∈ [yj, yj+1). Then if n > N
x > φsai(yj−3) + 2m−1 > fsi,j−3xi,j−3
(s) +m−1 > φnssi,j−3
(xi,j−3)
and hence, φnssi,j−3
(xi,j−3) < x. Similarly φnssi,j+4
(xi,j+4) > x and so
φntsi,j−3
(xi,j−3) 6 φnts(x−) 6 φn
ts(x+) 6 φntsi,j+4
(xi,j+4).
But then
φts(x− 5m−1) − 5m−1 < φtai(yj−3) − 2m−1
< fsi,j−3xi,j−3(t) −m−1
< φntsi,j−3
(xi,j−3)
6 φnts(x−),
and similarly
φnts(x+) < φts(x + 5m−1) + 5m−1.
Therefore dD0(φnts, φts) < 5m−1 and so dD(φ, φn) < ε, as required.
4.6 An equivalent space for the Brownian web
The paper [10] of Fontes, Isopi, Newman and Ravishankar, the original work char-
acterizing the Brownian web, constructs it as a random element of the space H
Chapter 4. The Brownian web 90
of compact collections of R-valued paths with specified starting points. In this
chapter, we have chosen to formulate the Brownian web as an element of the
space C of flows defined in Section 4.4.1. We note that H is a considerably larger
space than is required to support the Brownian web and believe that using a space
whose structure inherently contains the noncrossing and space filling restrictions
imposed by the Brownian web is more natural and simplifies characterization and
convergence results.
In [10], the authors make the comment that there is more than one natural
H-valued random variable that satisfies the following two conditions.
(i) From any deterministic point (x, t) in space-time, there is almost surely a
unique path Wx,t starting from (x, t).
(ii) For any deterministic n and (x1, t1), . . . , (xn, tn), the joint distribution of
Wx1,t1 , . . . ,Wxn,tn is that of coalescing Brownian motions (with unit diffusion
constant).
The standard Brownian web is the minimal collection of paths in H that satisfies
(i) and (ii). In [11], Fontes and Newman describe the forward full Brownian web,
which is the maximal collection of (noncrossing) paths that satisfies (i) and (ii).
In [11], the authors characterize a third object, the full Brownian web, which is a
random variable on the space HF of compact collections of paths from R → R, and
show that there is a one-to-one correspondence between the full Brownian web and
the forward full Brownian web. It can be similarly shown that there is a one-to-one
correspondence between the standard Brownian web and the full Brownian web.
In this section we shall show that C is in some sense isomorphic to a subset of
H′ ⊂ HF and that the full Brownian web almost surely lives on H′. Therefore, the
Brownian web that we construct on C is equivalent to the full Brownian web and,
hence, is also equivalent to both the standard Brownian web and the forward full
Brownian web. By working in the space C, we have the advantage of the existence
of a unique natural random variable that satisfies conditions (i) and (ii) above.
In Section 4.6.1 we describe the space HF on which the full Brownian web is
constructed and give its characterization. In Section 4.6.2 we discuss the sense in
which C is isomorphic to H′.
Chapter 4. The Brownian web 91
4.6.1 Compact sets of functions
In Fontes and Newman [11], HF is constructed from space-time points in (R2, ρ),
the compactification of R2 under a specified metric ρ. Instead, we shall take our
space points on the circle and our time points from [−T, T ] for some fixed T .
This avoids any technical issues resulting from compactification, but is essentially
equivalent. We give an outline of the results in [11] below.
Construct the two spaces (ΠF , dF ) and (HF , dHF ) as follows. Let ΠF denote
the set of continuous functions f : [−T, T ] → R and let
dF (f1, f2) = sup−T6t6T
|f1(t) − f2(t)|.
The space (ΠF , dF ) is complete and separable.
Let HF denote the set of all subsets K of (ΠF , dF ), for which f ∈ K if and
only if f +m ∈ K for all m ∈ Z and for which K|[0,1] = f ∈ K : f(−T ) ∈ [0, 1]is compact. Define the induced Hausdorff metric dHF by
dHF (K1, K2) = supg1∈K1
infg2∈K2
dF (g1, g2) ∨ supg2∈K2
infg1∈K1
dF (g1, g2).
The space (HF , dHF ) is also complete and separable.
Definition 4.19. Let H′ ⊂ HF consist of those K ∈ HF which satisfy the follow-
ing two conditions.
(a) The paths of K are noncrossing (although they may touch; in particular they
may coalesce or bifurcate).
(b) For any point (x, t) ∈ R × [−T, T ], there exists some f ∈ K with f(t) = x.
Fontes and Newman [11] give the following characterization of the full Brownian
web.
Definition 4.20. A full Brownian web WF is any (HF , dHF )-valued random vari-
able whose distribution has the following properties.
(a) Almost surely the paths of WF are noncrossing (although they may touch,
including coalescing and bifurcating).
Chapter 4. The Brownian web 92
(b1) From any deterministic point (x, t) ∈ R × [−T, T ], there is almost surely a
unique path W Fx,t passing through x at time t.
(b2) For any deterministic n, (x1, t1), . . . , (xn, tn), the joint distribution of the
semipaths W Fxj ,tj
(t), t > tj, j = 1, . . . , n is that of a flow of coalescing
Brownian motions on the circle (with unit diffusion constant).
They show (Theorem 3.2 of [11]) that any two Brownian webs have the same
distribution.
We would like to show that (H′, dHF ) is isomorphic to (C, dC). However, we
need the following additional regularity condition.
Definition 4.21. Let H1 ⊂ H′ consist of those K ∈ HF which satisfy the following
additional condition.
(c) For each point (x, s) ∈ Q2 with s ∈ [−T, T ], there is a unique path fx,s
passing through x at time s.
Let C1 ⊂ C consist of those flows φ ∈ C which satisfy the following condition.
(c′) For each (x, s) ∈ Q2 with s ∈ [−T, T ], φts is continuous at x for all t > s.
We shall show that (H1, dHF ) is isomorphic to (C1, dC). In both settings the
full Brownian web (respectively Brownian web) is almost surely H1 (respectively
C1) valued.
4.6.2 The isomorphism between the spaces
We define an isometry θ : C1 → H1 as follows.
For each t0 ∈ [−T, T ], let C[t0] be the set of all continuous functions f :
[t0, T ] → R. Let
Π =⋃
t0∈[−T,T ]
C[t0].
Chapter 4. The Brownian web 93
Let H be the set containing all subsets of Π with noncrossing paths. For each
φ ∈ C1, define θ(φ) ∈ H by
θ(φ) =⋃
t0∈[−T,T ]∩Q
f ∈ C[t0] : f = φ(t0, ·, x) for some x ∈ Q.
We say that a set U ∈ H is maximal if, for any f ∈ Π, U ∪ f ∈ H implies
f ∈ U . There is a unique maximal set, which we denote θ(φ), that contains θ(φ).
Define θ(φ) ∈ H1 by θ(φ) = θ(φ) ∩ C[−T ].
Proposition 4.22. θ(φ) ∈ H1.
Proof. By definition, θ(φ) consists of continuous noncrossing functions and, since
φts ∈ D0, by (4.1), if f ∈ θ(φ), then f+m ∈ θ(φ) for allm ∈ Z. It is straightforward
to check that condition (c′) in Definition 4.21 implies condition (c). The property
that for every (x, t) ∈ R × [−T, T ] there exists some f ∈ θ(φ) with f(t) = x is an
immediate consequence of maximality. It remains to check that θ(φ)|[0,1] = f ∈θ(φ) : f(−T ) ∈ [0, 1] is compact.
Suppose that f1, f2, . . . ∈ θ(φ)|[0,1]. Since paths in θ(φ) are noncrossing, there
exists a subsequence nr such that fnr(t) is monotone for all t ∈ [−T, T ]. Further-
more, for each t, fnr(t) lies in an interval of length 1 for all nr and so there exists
some f : [−T, T ] → R such that fnr → f pointwise.
Given ε > 0, there exist −T = a0 < · · · < aM = T such that ‖φsak− id‖ < ε
4
for all ak 6 s 6 ak+1, k = 0, . . . ,M − 1. Pick N sufficiently large that if nr > N ,
then |fnr(ak) − f(ak)| < ε4
for k = 1, . . . ,M − 1. But then, for ak 6 s < ak+1,
|fnr(s) − f(s)| 6 |φsak(f(ak) + ε
4) − φsak
(f(ak) − ε4)| < ε.
Hence, fnr → f uniformly. Therefore, f is continuous and θ(φ)∪f is noncross-
ing and so f ∈ θ(φ) with f(−T ) ∈ [0, 1], proving compactness.
Proposition 4.23. The function θ : C1 → H1 is bijective.
Proof. To see that θ is injective, suppose that θ(φ) = θ(φ′) for some φ, φ′ ∈ C1.
Then θ(φ) ∪ θ(φ′) ∈ H. Now for each s ∈ [−T, T ], x ∈ R, there exists fsx :
Chapter 4. The Brownian web 94
[s, T ] → R such that fsx ∈ θ(φ) and f ′sx : [s, T ] → R such that f ′
sx ∈ θ(φ′), with
fsx(t) = φts(x±) and f ′sx(t) = φ′
ts(x±) for all t > s. By the noncrossing property,
for each n ∈ N, f ′sx(t) ∈ [fs,x− 1
n(t), fs,x+ 1
n(t)]. Letting n→ ∞ gives φ′
ts(x) = φts(x)
at every point (s, t, x) for which φts is continuous at x. Hence, φ = φ′.
To see that θ is surjective, for K ∈ H1, let
φts(x) = inffur(t) : u 6 s, r ∈ Q, fur(s) > x,
where, for each (u, r) ∈ Q, fur is the unique element of K with f(u) = r. We first
show that φ ∈ C1. Since K consists of noncrossing functions and f ∈ K implies
f +m ∈ K for all m ∈ Z, φts ∈ D0. It is straightforward to check that condition
(c) in Definition 4.21 implies condition (c′).
Now suppose that r 6 s 6 t. For every x ∈ R there exists a sequence of
functions fn ∈ K with fn(r) ↓ x and fn(s) ↓ φsr(x) as n → ∞. Since K|[y,y+1] =
f ∈ K : f(−T ) ∈ [y, y + 1] is compact for all y (and the functions fn are
monotone and eventually lie in K|[y,y+1] for some y), there exists some f ∈ K such
that fn → f . Then φtr(x) = f(t) for all t > r and φts(φsr(x)) = f(t) for all t > s.
Hence, φts φsr = φtr.
Since K|[0,1] is compact, given ε > 0, there exist f1, . . . , fN ∈ K such that
‖fi − fi+1‖ < ε2
for i = 1, . . . , N (where we take fN+1 = f1 + 1). Since the fi are
continuous, there exists δ > 0 such that if |s− t| < δ, then |fi(s) − fi(t)| < ε2
for
i = 1, . . . N . Then for any s 6 t < s+ δ and any x ∈ R at which φts is continuous,
there exists some i,m for which x ∈ [fi(s) +m, fi+1(s) +m). But then
|φts(x) − x| 6 |fi(t) − fi+1(s)| ∨ |fi+1(t) − fi(s)|6 |fi(t) − fi(s)| ∨ |fi+1(t) − fi+1(s)| + ‖fi − fi+1‖< ε.
Finally, observe that by the argument used to show that φtsφsr = φtr, θ(φ) ⊂
⋃
s∈[−T,T ]f ∈ C[s] : there exists g ∈ K such that f(t) = g(t) for all t > s and
hence θ(φ) = K, as required.
Proposition 4.24. The function θ : C1 → H1 is an isometry.
Chapter 4. The Brownian web 95
Proof. Since for all K ∈ H1, K =⋃
m∈Z K|[m,m+1] and K|[m,m+1] is compact for
all m, given K1, K2 ∈ H1, there exist f ∈ K1, g ∈ K2 such that ‖f − g‖ =
dHF (K1, K2). Without loss of generality, suppose that g is chosen so that ‖f −g‖ = infh∈K2 ‖f − h‖. Because K2 satisfies property (b) of Definition 4.19, g can
be chosen so that there exist s, t ∈ [−T, T ] for which f(t) − g(t) = ‖f − g‖ =
−(f(s) − g(s)). In the case where Ki = θ(φi) for some φ1, φ2 ∈ C1, if f ∈ K1 with
f(s) = x and g ∈ K2 with g(s) = y, then
f(t) ∈ [φ1ts(x−), φ1
ts(x+)],
g(t) ∈ [φ2ts(y−), φ2
ts(y+)],
and so
φ2ts(y−) − φ1
ts(x+) 6 x− y 6 φ2ts(y+) − φ1
ts(x−).
Hence there exists some
u ∈[
1
2(x+ φ1
ts(x−)),1
2(x+ φ1
ts(x+))
]
∩[
1
2(y + φ2
ts(y−)),1
2(y + φ2
ts(y+))
]
.
Therefore,
dHF (θ(φ1), θ(φ2)) = |x− y|= |Φ(φ1
ts)(u) − Φ(φ2ts)(u)|
6 ‖Φ(φ1ts) − Φ(φ2
ts)‖6 dC(φ
1, φ2).
Conversely, suppose that s, t ∈ [−T, T ] and u ∈ [0, 1) are such that
dC(φ1, φ2) = |Φ(φ1
ts)(u) − Φ(φ2ts)(u)|.
There exist x, y ∈ [0, 1) such that
u ∈[
1
2(x+ φ1
ts(x−)),1
2(x+ φ1
ts(x+))
]
∩[
1
2(y + φ2
ts(y−)),1
2(y + φ2
ts(y+))
]
.
Without loss of generality suppose that x > y. There exists f ∈ θ(φ1) such that
Chapter 4. The Brownian web 96
f(r) = φ1rs(x−) for all r > s. Since x− y 6 φ2
ts(y+)− φ1ts(x−), if g ∈ θ(φ2), either
|g(s) − f(s)| > |x− y| or |g(t) − f(t)| > |φ2ts(y+) − φ1
ts(x−)| > |x− y|. Hence,
dHF (θ(φ1), θ(φ2)) > |x− y| = |Φ(φ1ts)(u) − Φ(φ2
ts)(u)| = dC(φ1, φ2),
and so
dHF (θ(φ1), θ(φ2)) = dC(φ1, φ2),
as required.
Bibliography
[1] Arratia, R. A. (1979). Coalescing Brownian motions on the line. Ph.D.
thesis, University of Wisconsin, Madison.
[2] Arratia, R. A. (1981). Coalescing Brownian motions and the voter model on
Z. Unpublished partial manuscript. Available from [email protected].
[3] Billingsley, P. (1999). Convergence of probability measures. Wiley Series in
Probability and Statistics: Probability and Statistics. John Wiley & Sons Inc.,
New York.
[4] Brown, D. and Rothery, P. (1993). Models in Biology: Mathematics,
Statistics and Computing. Wiley, Chichester.
[5] Brown, P. N., Byrne, G. D., and Hindmarsh, A. C. (1989). VODE:
a variable-coefficient ODE solver. SIAM J. Sci. Statist. Comput. 10, 5, 1038–
1051.
[6] Darling, R. W. R. and Norris, J. R. (2005). Structure of large random
hypergraphs. Ann. Appl. Probab. 15, 1A, 125–152.
[7] Durrett, R. (1999). Stochastic spatial models. SIAM Rev. 41, 4, 677–718
(electronic).
[8] Eden, M. (1961). A two-dimensional growth process. In Proc. 4th Berkeley
Sympos. Math. Statist. and Prob., Vol. IV. Univ. California Press, Berkeley,
Calif., 223–239.
97
Bibliography 98
[9] Ethier, S. N. and Kurtz, T. G. (1986). Markov processes. Wiley Se-
ries in Probability and Mathematical Statistics: Probability and Mathematical
Statistics. John Wiley & Sons Inc., New York.
[10] Fontes, L. R. G., Isopi, M., Newman, C. M., and Ravishankar,
K. (2004). The Brownian web: characterization and convergence. Ann.
Probab. 32, 4, 2857–2883.
[11] Fontes, L. R. G. and Newman, C. M. (2005). The full Brownian web as
scaling limit of stochastic flows. arXiv:math.PR/0511029.
[12] Forsythe, G. E. (1959). Reprint of a note on rounding-off errors. SIAM
Rev. 1, 66–67.
[13] Goldstine, H. H. and von Neumann, J. (1951). Numerical inverting of
matrices of high order. II. Proc. Amer. Math. Soc. 2, 188–202.
[14] Hairer, E. and Wanner, G. (1996). Solving ordinary differential equa-
tions. II. Springer Series in Computational Mathematics, Vol. 14. Springer-
Verlag, Berlin.
[15] Harris, T. E. (1984). Coalescing and noncoalescing stochastic flows in R1.
Stochastic Process. Appl. 17, 2, 187–210.
[16] Hartman, P. (1960). On local homeomorphisms of Euclidean spaces. Bol.
Soc. Mat. Mexicana (2) 5, 220–241.
[17] Hastings, M. B. and Levitov, L. S. (1998). Laplacian growth as one-
dimensional turbulence. Physica D 116 (1-2), 244.
[18] Henrici, P. (1962). Discrete variable methods in ordinary differential equa-
tions. John Wiley & Sons Inc., New York.
[19] Henrici, P. (1963). Error propagation for difference method. John Wiley
and Sons, Inc., New York.
[20] Henrici, P. (1964). Elements of numerical analysis. John Wiley & Sons
Inc., New York.
Bibliography 99
[21] Higham, N. J. (1996). Accuracy and stability of numerical algorithms. So-
ciety for Industrial and Applied Mathematics (SIAM), Philadelphia, PA.
[22] Hull, T. E. and Swenson, J. R. (1966). Tests of probabilistic models for
the propagation of roundoff errors. Comm. ACM 9, 108–113.
[23] IEEE754 (1985). IEEE standard for binary floating-point arithmetic,
ANSI/IEEE Standard 754-1985. Reprinted in SIGPLAN Notices, 22(2):9-25,
1987.
[24] Kallenberg, O. (2002). Foundations of modern probability. Probability
and its Applications (New York). Springer-Verlag, New York.
[25] Kingman, J. F. C. (1999). Martingales in the OK Corral. Bull. London
Math. Soc. 31, 5, 601–606.
[26] Kingman, J. F. C. and Volkov, S. E. (2003). Solution to the OK Corral
model via decoupling of Friedman’s urn. J. Theoret. Probab. 16, 1, 267–276.
[27] Lawler, G. F. (2005). Conformally invariant processes in the plane. Math-
ematical Surveys and Monographs, Vol. 114. American Mathematical Society,
Providence, RI.
[28] Mosbach, S. and Turner, A. G. (2005). A quantitative investigation
into the accumulation of rounding errors in numerical ODE solution. Technical
Report 36, c4e-Preprint Series, Cambridge.
[29] Piterbarg, V. V. (1998). Expansions and contractions of isotropic stochas-
tic flows of homeomorphisms. Ann. Probab. 26, 2, 479–499.
[30] Toth, B. and Werner, W. (1998). The true self-repelling motion. Probab.
Theory Related Fields 111, 3, 375–452.
[31] Tsirelson, B. (2004). Nonclassical stochastic flows and continuous products.
Probab. Surv. 1, 173–298 (electronic).
[32] Turner, A. G. (2006). Convergence of Markov processes near saddle fixed
points. To appear in Ann. Probab.
Bibliography 100
[33] Wilkinson, J. H. (1994). Rounding errors in algebraic processes. Dover
Publications Inc., New York.
[34] Williams, D. and McIlroy, P. (1998). The OK Corral and the power
of the law (a curious Poisson-kernel formula for a parabolic equation). Bull.
London Math. Soc. 30, 2, 166–170.
[35] Witten, T. A. and Sander, L. M. (1981). Diffusion-limited aggregation,
a kinetic critical phenomenon. Phys. Rev. Lett. 47, 19, 1400–1403.