NUMERICAL SOLUTION OF

STOCHASTIC DIFFERENTIAL
EQUATIONS
A thesis submitted to university of edinburgh and heriot-watt
university
for the degree of Master of Science
2005
By
Seokhyun Han
1
Contents
Abstract 4
Acknowledgments 5
1 Introduction 6
2 Part I - Review 8
2.1 Ito formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Uniqueness and Existence of SDEs . . . . . . . . . . . . . . . . . 9
2.3 Vector SDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Ito-Taylor expansion . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Euler Approximation . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Pathwise Approximation and Strong Convergence . . . . . . . . . 16
2.7 Approximation of Moments and Weak Convergence . . . . . . . . 16
2.8 Strong Approximations . . . . . . . . . . . . . . . . . . . . . . . . 18
2.8.1 Euler scheme . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.8.2 Milstein scheme . . . . . . . . . . . . . . . . . . . . . . . . 20
2.9 Weak Approximation . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.9.1 Weak Euler Scheme . . . . . . . . . . . . . . . . . . . . . . 22
2.9.2 Order 2.0 Weak Taylor scheme . . . . . . . . . . . . . . . . 22
3 Part II - Improving Accuracy 24
3.1 Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 Part III - Implementation for Strong Approximation 38
5 Conclusion 41
2
Bibliography 44
3
Abstract
The aim of this paper is to provide a systematic treatment of time discretized
numerical schemes for stochastic dierential equations and to demonstrate that
it is possible to improve the accuracy of the weak error of Eulers approximation
by implementing Richardsons idea.
4
Acknowledgments
I would like to express my deep gratitude to many teaching sta who have taught
and helped me in course of Graduate Certicate in Mathematical Science and MSc
in Financial Mathematics. I am greatly indebted to Professor Istvan Gyongy, MSc
project supervisor, who have given many interesting ideas and helpful advice. I
also owe many thanks to my colleague, Tomas Lagland, for his assistance.
5
Chapter 1
Introduction
This paper aims to provide a systematic framework for an understanding of the
basic concepts and tools needed for the development and implementation of nu-
merical methods for stochastic dierential equations, primarily time discretization
methods for initial value problems of stochastic dierential equations with Ito dif-
fusions as their solutions. Recent years have witnessed that the most ecient and
widely applicable approach to solving SDEs seems to be the simulation of sample
paths of time discrete approximations on digital computers. This is based on a
nite discretization of the time interval [0, T] under consideration and generates
approximate values of the sample paths step by step at the discretization times.
Ito process X = {X
t
, t 0} has the form
X
t
= X
0
+
_
t
0
a(X
s
)ds +
_
t
0
b(X
s
)dW
s
for t 0 with an initial value X
0
which may be random, the drift a(X
t
), and the
diusion b(X
t
). This integral equation is often written in the dierential form
dX
t
= a(X
t
)dt + b(X
t
)dW
t
which is called an Ito stochastic dierential equation. Unfortunately explicitly
solvable SDEs are rare in practical applications. However, there are increasing
number of numerical methods for the solution of SDEs. In connection with this is-
sue, my intent here is to review the systematic development of ecient numerical
6
CHAPTER 1. INTRODUCTION 7
methods for SDEs by using [2] [3]. Obviously such methods should be imple-
mentable on digital computers. We shall survey various time discrete numerical
methods which are appropriate for the simulation of sample paths or functionals
of Ito processes in order to estimate various statistical features of the desired so-
lution. In dealing with this issue, we should also speak of how to judge the quality
of a discrete time approximation, i.e. a criterion needs to be specied. Such a
criterion should reect the main goal of practical simulations. There are two basic
types of tasks connected with the simulation of solutions of stochastic dierential
equations. The rst occurs in situation where a good pathwise approximation is
required, for instance in direct simulations, ltering or testing statistical estima-
tors. In the second interest focuses on approximating expectations of functionals
of the Ito process, such as its probability distribution and its moments. This is
relevant in many practical problems because such functionals can not often be
determined analytically. Furthermore, the ultimate goal of this paper will be to
investigate a general idea of improving accuracy of numerical methods which is
due to Richardson. He suggested to consider approximations obtained by linear
combinations of approximation corresponding to dierent step sizes.
Before we undertake this task, it is worth taking a quick look at the main
parts. Part I on Review will survey Ito formula, vector SDEs, stochastic Taylor
expansions which provide a universally applicable tool for SDEs which is analo-
gous to the deterministic Taylor formula in ordinary calculus, and Euler scheme
to highlight the basic issues, types of problems, and objectives that arise when
SDEs are solved numerically. In particular, we distinguish between strong and
weak approximations, depending on whether good pathwise or good probability
distributional approximations are sought. The object of Part II is to revisit the
themes originally explored in D.Talay and L.Tubaros paper Expansion of the
global error for numerical schemes solving stochastic dierential equation [1] in
order to show how to improve the weak error of the Euler scheme.
Chapter 2
Part I - Review
In the following I will give a brief general survey of numerical solution of stochastic
dierential equations using P.E.Kloeden and E.Platens two books [2] [3].
2.1 Ito formula
Let a and b be two functions from [0, T] into R with measurability and
integrability properties that the ordinary and stochastic integrals appearing in the
following formula make sense. By a stochastic dierential we mean an expression
(2.1) dX
t
() = a(t, )dt + b(t, )dW
t
()
which is just a symbolical way of writing
X
t
() X
s
() =
_
t
s
a(u, )du +
_
t
s
b(u, )dW
u
()
w.p.1, for any 0 s t T. The rst integral in the above is an ordinary
(Riemann or Lebesgue) integral for each and the second is an Ito integral.
Let f : [0, T] R R have continuous partial derivatives
f
t
,
f
x
and

2
f
x
2
. A
scalar transformation by f of the stochastic dierential (2.1) results after some
non trivial analysis in the Ito formula,
df(t, X
t
) =
f(t, X
t
)
t
dt +
f(t, X
t
)
x
dX
t
+
1
2

2
f(t, X
t
)
x
2
(dX
t
)
2
8
CHAPTER 2. PART I - REVIEW 9
and in other way,
f(t, X
t
) f(s, X
s
) =
_
t
s
_
f
t
+ a
f
x
+
1
2
b
2

2
f
x
2
_
du +
_
t
s
b
f
x
dW
u
w.p.1, for any 0 s t T, where the integrands are all evaluated at (u, X
u
).
2.2 Uniqueness and Existence of SDEs
When we have no explicit solution of an Ito stochastic equation
(2.2) X
t
= X
t
0
+
_
t
t
0
a(X
s
)ds +
_
t
t
0
b(X
s
)dW
s
we need somehow to ensure the existence and uniqueness of a process X =
{X
t
, t [t
0
, T]} which satises it. We call a(X
s
) the drift coecient and b(X
s
)
the diusion coecient.We shall say that solutions of (2.2) are pathwise unique
if any two such solutions X = {X
t
, t [t
0
, T]} and

X = {

X
t
, t [t
0
, T]} have,
almost surely, the same sample paths on [t
0
, T], that is if
P
_
sup
t
0
tT
| X
t

X
t
|> 0
_
= 0
In this case we call X a unique strong solution of (2.2). From a basic existence and
uniqueness theorem, it follows that equation (2.2) has a unique strong solution
X = {X
t
, t [t
0
, T]} on [t
0
, T] with
sup
t
0
tT
E(X
2
t
) <
provided X
t
0
is independent of W = {W
t
, t [t
0
, T]} with E(X
t
0
)
2
< and
the coecients a, b satisfy Lipschitz conditions. We often call X an Ito diusion
process. Sometimes equation (2.2) may have solutions which are unique in the
weaker sense that only their probability laws coincide, but not necessarily their
sample paths. We shall say then that we have a unique weak solution. Existence
and uniqueness of weak solutions follow if the coecients a and b are bounded
and continuous and b is nondegenerate |b| > 0. Of course, any unique strong
solution is always a unique weak solution.
CHAPTER 2. PART I - REVIEW 10
2.3 Vector SDEs
We shall interpret a vector as a column vector and its transpose as a row vector
and consider an m-dimensional Wiener process W = {W
t
, t 0} with compo-
nents W
1
t
, W
2
t
, . . . , W
m
t
, which are independent independent scalar Wiener pro-
cesses. Then, we take a d-dimensional vector valued function a : [t
0
, T]R
d
R
d
,
the drift coecient, and a d m-matrix valued function b : [t
0
, T] R
d
R
dm
,
the diusion coecient, t
0
[0, T], to form a d-dimensional vector stochastic
dierential equation
(2.3) dX
t
= a(t, X
t
)dt + b(t, X
t
)dW
t
We interpret this as a stochastic integral equation
(2.4) X
t
= X
t
0
+
_
t
t
0
a(s, X
s
)ds +
_
t
t
0
b(s, X
s
)dW
s
with initial value X
t
0
R
d
, where the Lebesgue and Ito integrals determined
component by component, with the ith component of (2.4) being
X
i
t
= X
i
t
0
+
_
t
t
0
a
i
(s, X
s
)ds +
m

j=1
_
t
t
0
b
i,j
(s, X
s
)dW
j
s
If the drift and diusion coecients do not depend on the time variable, that
is if a(t, x) a(x) and b(t, x) b(x), then we say that stochastic equation is
autonomous. We can always write a nonautonomous equation as a vector au-
tonomous equation of one dimension more by setting in the rst component the
drift coecient equal to 1 and the diusion coecient as 0 to obtain as the rst
component of X
t
the time variable X
1
t
= t.
There is a vector version of the Ito formula. For a suciently smooth transfor-
mation f = [t
0
, T] R
d
R
k
of the solution X = {X
t
, t
0
t T} of (2.3) we
obtain a k-dimensional process Y = {Y
t
= f(t, X
t
), t
0
t T} with the vector
stochastic dierential in component form
dY
p
t
=
_
f
p
t
+
d

i=1
a
i
f
p
x
i
+
1
2
d

i,j=1
m

l=1
b
i,l
b
j,l

2
f
p
x
i
x
j
_
dt +
m

l=1
d

i=1
b
i,l
f
p
x
i
dW
l
t
CHAPTER 2. PART I - REVIEW 11
for p = 1, 2, . . . , k, where the terms are all evaluated at (t, X
t
). We can sometimes
use this formula to determine the solutions of certain vector stochastic dierential
equations in terms of known solutions of the other equations, for example linear
equations.
2.4 Ito-Taylor expansion
To begin we consider the equation X = {X
t
, t [t
0
, T]} of 1-dimensional ordinary
dierential equation
d
dt
X
t
= a(X
t
)
with initial value X
t
0
, for t [t
0
, T] where 0 t
0
< T, which we can write in the
equivalent integral equation form as
(2.5) X
t
= X
t
0
+
_
t
t
0
a(X
s
)ds
To justify the following constructions we require that the function a satises
appropriate properties, for instance to be suciently smooth with a linear growth
bound. let f : R R be a continuously dierentiable function. Then by the
chain rule, we have
(2.6)
d
dt
f(X
t
) = a(X
t
)f

(X
t
)
which using the operator
Lf = af

where

denotes dierentiation with respect to x, we can express (2.6) as the
integral relation
(2.7) f(X
t
) = f(X
t
0
) +
_
t
t
0
Lf(X
s
)ds
for all t [t
0
, T]. When f(x) x we have Lf = a, L
2
f = La,. . . and (2.7)
reduces to
(2.8) X
t
= X
t
0
+
_
t
t
0
a(X
s
)ds
CHAPTER 2. PART I - REVIEW 12
that is, to equation (2.5). If we apply the relation (2.7) to the function f = a in
the integral in (2.8), we obtain
(2.9) X
t
= X
t
0
+
_
t
t
0
(a(X
t
0
) +
_
s
t
0
La(X
z
)dz)ds
= X
t
0
+ a(X
t
0
) + a(X
t
0
)
_
t
t
0
ds +
_
t
t
0
_
s
t
0
La(X
z
)dzds
which is the simplest nontrivial Taylor expansion for X
t
. We can apply (2.7)
again to the function f = La in the double integral of (2.9) to derive
X
t
= X
t
0
+ a(X
t
0
)
_
t
t
0
ds + La(X
t
0
)
_
t
t
0
_
s
t
0
dzds + R
3
with remainder
R
3
=
_
t
t
0
_
s
t
0
_
z
t
0
L
2
a(X
u
)dudzds
for t [t
0
, T]. For a general r + 1 times continuously dierentiable function
f : R R, this method gives the classical Taylor formula in integral form
(2.10) f(X
t
) = f(X
0
) +
r

i=1
(t t
0
)
l
l!
L
l
f(X
t
0
) +
_
t
t
0

_
s
2
t
0
L
r+1
f(X
s
1
)ds
1
for t [t
0
, T] and r = 1, 2, 3, . . . since
_
t
t
0
_
s
1
t
0

_
s
l1
t
0
ds
1
ds
l
=
1
l!
(t t
0
)
l
for l = 1, 2, . . .. The Taylor formula (2.10) has proven to be a very useful tool in
both theoretical and practical investigations, particularly in numerical analysis.
It allows the approximation of a suciently smooth function in a neighborhood
of a given point to any desired order of accuracy. The expansion depends on
the values of the function and some of its higher derivatives at the expansion
point, weighted by corresponding multiple time integrals. In addition, there is
a remainder term which contains the next following multiple time integral, but
with a time dependent integrand.
A stochastic counterpart of the deterministic Taylor formula for the expansion
of smooth functions of an Ito process about a given value has many potential
CHAPTER 2. PART I - REVIEW 13
applications in stochastic analysis, for instance in the derivation of numerical
methods for stochastic dierential equations. There are several possibilities for
such a stochastic Taylor formula. One is based on the iterated application of the
Ito formula, which we shall call the Ito-Taylor expansion. We shall indicate it
here for the solution X
t
of the 1-dimensional Ito stochastic dierential equation
in integral form
X
t
= X
t
0
+
_
t
t
0
a(X
s
)ds +
_
t
t
0
b(X
s
)dW
s
for t [t
0
, T], where the second integral is an Ito stochastic integral and the
coecients a and b are suciently smooth real valued functions satisfying a linear
growth bound. For any twice continuously dierentiable function f : R R, the
Ito formula then gives
(2.11) f(X
t
) = f(X
0
) +
_
t
t
0
_
a(X
s
)f

(X
s
) +
1
2
b
2
(X
s
)f

(X
s
)
_
ds
+
_
t
t
0
b(X
s
)f

(X
s
)dW
s
= f(X
t
0
) +
_
t
t
0
L
0
f(X
s
)ds +
_
t
t
0
L
1
f(X
s
)dW
s
for t [t
0
, T]. Here we have introduced the operators
L
0
f = af

+
1
2
b
2
f

and
L
1
f = bf

Obviously for f(x) x we have L

0
f = a and L
1
f = b, in which case (2.11)
reduces to the original Ito equation for X
t
, that is to
(2.12) X
t
= X
t
0
+
_
t
t
0
a(X
s
)ds +
_
t
t
0
b(X
s
)dW
s
In analogy with the deterministic expansion above, if we apply the Ito formula
(2.11) to the functions f = a and f = b in (2.12) we obtain
(2.13) X
t
= X
t
0
+
_
t
t
0
_
a(X
t
0
) +
_
s
t
0
L
0
a(X
z
)dz +
_
s
t
0
L
1
a(X
z
)dW
z
_
ds
CHAPTER 2. PART I - REVIEW 14
+
_
t
t
0
_
b(X
t
0
) +
_
s
t
0
L
0
b(X
z
)dz +
_
s
t
0
L
1
b(X
z
)dW
z
_
dW
s
= X
t
0
+ a(X
t
0
)
_
t
t
0
ds + b(X
t
0
)
_
t
t
0
dW
s
+ R
with remainder
R =
_
t
t
0
_
s
t
0
L
0
a(X
z
)dzds +
_
t
t
0
_
s
t
0
L
1
a(X
z
)dW
z
ds
+
_
t
t
0
_
s
t
0
L
0
b(X
z
)dzdW
s
+
_
t
t
0
_
t
t
0
L
1
b(X
z
)dW
z
dW
s
This is the simplest nontrivial Ito-Taylor expansion of X
t
. It involves integrals
with respect to both the time variable and the Wiener process, with multiple
integrals with respect to both in the remainder. We can apply repeat the above
procedure, for instance by applying the Ito formula (2.11) to f = L
1
b in (2.13),
in which case we get
X
t
= X
t
0
+ a(X
t
0
)
_
t
t
0
ds + b(X
t
0
)
_
t
t
0
dW
s
+ L
1
b(X
t
0
)
_
t
t
0
_
s
t
0
dW
z
dW
s
+

R
with remainder

R =
_
t
t
0
_
s
t
0
L
0
a(X
Z
)dzds +
_
t
t
0
_
s
t
0
L
1
a(X
z
)dW
z
ds
+
_
t
t
0
_
s
t
0
L
0
b(X
Z
)dzdW
s
+
_
t
t
0
_
s
t
0
_
z
t
0
L
0
L
1
b(X
u
)dudW
z
dW
s
+
_
t
t
0
_
s
t
0
_
z
t
0
L
1
L
1
b(X
u
)dW
u
dW
z
dW
s
It is possible to express the Ito-Taylor expansion for a general function f and arbi-
trarily many expansion terms in a succinct way. Nevertheless, its main properties
are already apparent in the preceding example, with the multiple Ito integrals
_
t
t
0
ds,
_
t
t
0
dW
s
,
_
t
t
0
_
s
t
0
dW
z
dW
s
multiplied by certain constants and a remainder term involving the next following
multiple Ito integrals, but with nonconstant integrands. The Ito-Taylor expan-
sion can thus be considered as a generalization of both the Ito formula and the
CHAPTER 2. PART I - REVIEW 15
deterministic Taylor formula.
2.5 Euler Approximation
One of the simplest time discrete approximations of an Ito process is the Euler
approximation. We shall consider an Ito process X = {X
t
, t
0
t T} satisfying
the scalar stochastic dierential equation
dX
t
= a(t, X
t
)dt + b(t, X
t
)dW
t
on t
0
t T with the initial value X
t
0
= X
0
. For a given discretization
t
0
=
0
<
1
< <
n
= T of the time interval [t
0
, T], an Euler approximation
is a continuous time stochastic process Y = {Y (t), t
0
t T} satisfying the
iterative scheme
Y
n+1
= Y
n
+ a(
n
, Y
n
)(
n+1

n
) + b(
n
, Y
n
)(W

n+1
W

n
)
for n = 0, 1, 2, . . . , N 1 with initial value Y
0
= X
0
, where we have written
Y
n
= Y (
n
) for the value of the approximation at the discretization time
n
. We
shall consider equidistant discretization times

n
= t
0
+ n
with stepsize = (T t
0
)/N for some integer N. Now, we need to generate the
random increments
W
n
= W

n+1
W

n
for n = 0, 1, 2, . . . , N 1, of the Wiener process W = {W
t
, t 0}. We notice
that these increments are independent Gaussian random variables with mean
E(W
n
) = 0 and variance E((W
n
)
2
) = . We shall apply the Euler scheme
Y
n+1
= Y
n
+ a + bW
n
for n = 0, 1, . . . N 1 to approximate specic stochastic dierential equation.
CHAPTER 2. PART I - REVIEW 16
2.6 Pathwise Approximation and Strong Con-
vergence
We shall introduce the absolute error criterion which is appropriate for a situation
where a pathwise approximation is required. The absolute error criterion is just
the expectation of the absolute value of the dierence between the approximation
and the Ito process at the time T, that is
= (E|X
T
Y (T)|
q
)
1
q
for some q 1, and gives a measure of the pathwise closeness at the end of
the time interval [0, T]. We shall say that a discrete time approximation Y with
maximum time step size converges strongly to X at time T if
lim
0
E(|X
T
Y (T)|) = 0
In order to assess and compare dierent discrete time approximations, we need
to know their rates of strong convergence. We shall say that a discrete time
approximation Y converges strongly with order > 0 at time T if there exists a
positive constant C, which does not depend on and a
0
> 0 such that
() = E(|X
T
Y (T)|) C

for each (0,

0
). We shall investigate the strong convergence of a number of
dierent discrete time approximations experimentally. We shall see in particular
that the Euler approximation has strong order of convergence = 0.5.
2.7 Approximation of Moments and Weak Con-
vergence
To introduce this weak type of convergence, we shall carry out some computer
experiments to investigate the mean error
= EY (T) EX
T
CHAPTER 2. PART I - REVIEW 17
for the linear stochastic dierential equation
dX
t
= aX
t
dt + bX
t
dW
t
and its Euler approximation
Y
n+1
= Y
n
+ aY
n

n
+ bY
n
W
n
for n = 0, 1, 2, . . . , N 1, where W
n
=
n+1

n
denotes the step size and
W
n
= W

n+1
W

n
the increment of the Wiener process. We see that this
criterion diers in its properties from the strong convergence criterion. To some
extent, the above mean error is special and not appropriate for applications where
the approximation of some higher moment
E(|X
T
|
q
)
with q = 2, 3, . . . or of some functional
E(g(X
T
))
is of interest. These do not require a good pathwise approximation of the Ito
process, but only an approximation of the probability distribution of X
T
. We
shall say that a general discrete time approximation Y with maximum time step
size converges weakly to X at time T as 0 with respect to a class C of test
functions g : R
d
R if we have
lim
0
|E(g(X
T
)) E(g(Y (T)))| = 0
for all g C. If C contains all polynomials this denition implies the convergence
of all moments, so investigations involving it will require the existence of all
moments. We shall say that a time discrete approximation Y converges weakly
with order > 0 to X at time T as 0 if for each polynomial g there exists a
positive constant C, which does not depend on and a nite
0
> 0 such that
|E(g(X
T
)) E(g(Y (T))))| C

CHAPTER 2. PART I - REVIEW 18

for each (0,
0
). we shall see that the Euler approximation usually converges
with weak order = 1 in contrast with the strong order = 0.5.
2.8 Strong Approximations
We shall consider discrete time approximations of various strong orders that have
been derived from stochastic Taylor expansions by including appropriately many
terms. To describe these schemes succinctly for a general d-dimensional Ito pro-
cess satisfying the stochastic dierential equation
X
t
= X
t
0
+
_
t
t
0
a(s, X
s
)ds +
m

j=1
_
t
t
0
b
j
(s, X
s
)dW
j
s
in Ito form, for t [t
0
, T], we shall use the following generalizations of the oper-
ators:
L
0
=

t
+
d

k=1
a
k

x
k
+
1
2
d

k,l=1
m

j=1
b
k,j
b
l,j

2
x
k
x
l
L
0
=

t
+
d

k=1
a
k

x
k
L
j
= L
j
=
d

k=1
b
k,j

x
k
for j = 1, 2, . . . , m, with the corrected drift
a
k
= a
k

1
2
m

j=1
L
j
b
k,j
for k = 1, 2, . . . , d. In addition, we shall abbreviate multiple Ito integrals by
I(j
1
, . . . , j
l
) =
_

n+1

n

_
s
2

n
dW
j
1
s
1
dW
j
l
s
l
for j
1
, . . . , j
l
{0, 1, . . . , m}, l = 1, 2, . . ., and n = 0, 1, 2, . . . with the convention
that
W
0
t
= t
CHAPTER 2. PART I - REVIEW 19
for all t R
+
. We shall also use the abbreviation
f = f(
n
, Y
n
)
for n = 0, 1, 2, . . . in the schemes for any given function f dened on R
+
R
d
and
usually not explicitly mention the initial value Y
0
or the step indices n = 0, 1, . . ..
2.8.1 Euler scheme
The Euler scheme is simplest strong Taylor approximation, containing only the
time and Wiener integrals of multiplicity one from the Ito-Taylor expansion, and
usually attains the order of strong convergence = 0.5. In the 1-dimensional
case d = m = 1, the Euler scheme has the form
Y
n+1
= Y
n
+ a
n
+ bW
n
where

n
=
_

n+1

n
dt =
n+1

n
is the length of the time discretization subinterval [
n
,
n+1
] and
W
n
=
_

n+1

n
dW
t
= W

n+1
W

n
is the N(0,
n
) distributed increment of the Wiener process W on [
n
,
n+1
]. For
the general multiple dimensional case with d, m = 1, 2, . . . the kth componenet
of the Euler scheme has the form
Y
k
n+1
= Y
k
n
+ a
k

n
+
m

j=1
b
k,j
W
j
n
where
W
j
n
=
_

n+1

n
dW
j
t
= W
j

n+1
W
j

n
is the N(0,
n
) distributed increment of the jth componenet of the m-dimensional
standard Wiener process W on [
n
,
n+1
]; thus W
j
1
n
and W
j
2
n
are independent
for j
1
= j
2
.
Theorem 2.1 Suppose that
CHAPTER 2. PART I - REVIEW 20
(i) E(|X
0
|
2
) < ,
(ii) E(|X
0
Y

0
|
2
)
1/2
K
1

1/2
,
(iii) |a(t, x) a(t, y)| +|b(t, x) b(t, y)| K
2
|x y|,
(iv) |a(t, x)| +|b(t, x)| K
3
(1 +|x|),
(v) |a(s, x) a(t, x)| +|b(s, x) b(t, x)| K
4
(1 +|x|)|s t|
1/2
for all s, t [0, T] and x, y R
d
, where the constants K
1
, K
2
, K
3
, K
4
do not
depend on . Then, for the Euler approximation Y

, the estimate
E(|X
T
Y

(T)|) K
5

1/2
holds, where the constant K
5
does not depend on .
Proof. refer to [2].
In special cases, the Euler scheme may actually achieve a higher order of
strong convergence. For example, when the noise is additive, that is when the
diusion coecient has the form
b(t, x) b(t)
for all (t, x) R
+
R
d
, it turns out that the Euler scheme has order = 1.0
of strong convergence under appropriate smoothness assumptions on a and b.
We remark that additive noise is sometimes understood to have b(t) as constant.
Usually the Euler scheme gives good numerical results when the drift and diu-
sion coecients are nearly constant. In general, however, it is not particularly
satisfactory and the use of higher order schemes is recommended.
2.8.2 Milstein scheme
If, in the 1-dimensional case with d = m = 1, we add to the Euler scheme the
term
bb

I(1, 1) =
1
2
bb

{(W
n
)
2

n
}
from the Ito-Taylor expansion, then we obtain the Milstein scheme
Y
n+1
= Y
n
+ a
n
+ bW
n
+
1
2
bb

{(W
n
)
2

n
}
CHAPTER 2. PART I - REVIEW 21
We can rewrite this as
Y
n+1
= Y
n
+ a
n
+ bW
n
+
1
2
bb

(W
n
)
2
since
a = a
1
2
bb

Theorem 2.2 Suppose that

(i) E(|X
0
|
2
) < ,
(ii) E(|X
0
Y

0
|
2
)
1/2
K
1

1/2
,
(iii)
|a(t, x) a(t, y)| K
2
|x y|
|b
j
1
(t, x) b
j
1
(t, y)| K
2
|x y|
|L
j
1
b
j
2
(t, x) L
j
1
b
j
2
(t, y)| K
2
|x y|
(iv)
|a(t, x)| +|L
j
a(t, x))| K
3
(1 +|x|)
|b
j
1
(t, x)| +|L
j
b
j
2
(t, x)| K
3
(1 +|x|)
|L
j
L
j
1
b
j
2
(t, x))| K
3
(1 +|x|)
(v)
|a(s, x) a(t, x)| K
4
(1 +|x|)|s t|
1/2
|b
j
1
(s x) b
j
1
(t x)| K
4
(1 +|x|)|s t|
1/2
|L
j
1
b
j
2
(s x) L
j
1
b
j
2
(t x)| K
4
(1 +|x|)|s t|
1/2
for all s, t [0, T], x, y R
d
, j = 0, . . . , m and j
1
, j
2
= 1, . . . , m, where the con-
stants K
1
, K
2
, K
3
, K
4
do not depend on . Then for the Milstein approximation
Y

, the estimate
E(|X
T
Y

(T)|) K
5

holds, where the constant K

5
does not depend on .
Proof. refer to [2].
CHAPTER 2. PART I - REVIEW 22
2.9 Weak Approximation
2.9.1 Weak Euler Scheme
We recall that for the general multi-dimensional case d, m = 1, 2, . . ., the kth
component of the Euler scheme has the form
Y
k
n+1
= Y
k
n
+ a
k

n
+
m

j=1
b
k,j
W
j
n
with initial value Y
0
= X
0
, where
n
=
n+1

n
and W
j
n
= W
j

n+1
W
j

n
. If,
amongst other assumptions, a and b are suciently smooth, then the Euler ap-
proximation has order of weak convergence = 1.0. With the weak convergence
criterion
|E(g(X
T
)) E(g(Y (T))))| C

we have much freedom to choose other simpler random variables instead of using
the Gaussian increments W
j
n
. Such random variables have only to coincide in
their lower order moments with those of W
1
n
and W
2
n
to provide a sucient
accurate approximation of the probability law of the Ito diusion. For instance,
we could use two-point distributed random variables

W
j
n
with
P
_

W
j
n
=
_

n
_
=
1
2
in which case becomes the simplied Euler scheme
(2.14) Y
k
n+1
= Y
k
n
+ a
k

n
+
m

j=1
b
k,j

W
j
n
2.9.2 Order 2.0 Weak Taylor scheme
More accurate weak Taylor schemes can be derived by including further multiple
stochastic integrals from the stochastic Taylor expansion. Since the objective is
to obtain more information about the probability measure of the underlying Ito
process rather than about its sample space, we also have the freedom to replace
the multiple stochastic integrals by much simpler random variables as in (2.14).
We shall consider the weak Taylor scheme obtained by adding all of the double
stochastic integrals from the Ito-Taylor expansion to the Euler scheme. In the
CHAPTER 2. PART I - REVIEW 23
autonomous 1-dimensional case d = m = 1 we obtain the order 2.0 weak Taylor
scheme
Y
n+1
= Y
n
+ a
n
+ bW
n
+
1
2
bb

{(W
n
)
2

n
}
+a

bZ
n
+
1
2
_
aa

+
1
2
a

b
2
_

2
n
_
ab

+
1
2
b

b
2
_
{W
n

n
Z
n
}
where Z
n
represents the double Ito integral
Z
n
=
_

n+1

n
_
s
2

n
dW
s
1
dW
s
2
The pair of correlated Gaussian random variables (W
n
, Z
n
) can be generated
from a pair of independent standard Gaussian random variables.
Chapter 3
Part II - Improving Accuracy
In the discussion of themes below, works will be mainly taken from D.Talay and
L.Tubaro [1].
3.1 Description
Let us consider the following Ito stochastic dierential equation
(3.1) dX
t
= b(t, X
t
)dt + (t, X
t
)dW
t
where X
t
is a stochastic process in R
d
, W
t
is a Wiener process in R
l
, b(t, x) is a
d-vector and (t, x) is a d l-matrix. We denote the solution of (3.1) with the
deterministic initial condition x at time s by X
s,x
t
or X
s
= x. We suppose that
the coecients b and are smooth so that the existence and uniqueness of this
solution are ensured. Let Y be a random variable independent of any increment
W
t
W
s
of the Wiener process, and having moments of any order. The solution
of (3.1) has the initial condition X
0,Y
t
or X
0
= Y at time 0. Let us consider the
numerical evaluation of the quantity
Ef(X
T
)
where T is some xed time, f a given smooth function from R
d
R by a Monte-
Carlo method based on the simulation of a piecewise constant approximating
process

X
h
p
, p N. We divide the interval [0, T] in n subintervals with the same
length h = T/n. According to our notation,

X
h
n
=

X
h
T
. We want to study the
24
CHAPTER 3. PART II - IMPROVING ACCURACY 25
global error
Err(T, h) = Ef(X
T
) Ef(

X
h
T
)
We recall that the Euler and Milshtein discretization schemes are of rst order:
there exists a positive constant C(T), independent of h, such that
|Err(T, h)| C(T)h
Let us dene the following two numerical schemes:
Euler scheme

X
h
0
= X
0

X
h
p+1
=

X
h
p
+ b(ph,

X
h
p
)h + (ph,

X
h
p
)
h
p+1
W
where
h
p+1
W = W
(p+1)h
W
ph
.
Milshtein scheme

X
h
p+1
=

X
h
p
+
r

j=1

j
(ph,

X
h
p
)
h
p+1
W
j
+ b(ph,

X
h
p
)h
+
r

j,k=1

j
(ph,

X
h
p
)
k
(ph,

X
h
p
)Z
kj
p+1
where

j
denotes the jth column of ,

j
denotes the matrix whose element of the ith row and kth column
is
k

i
j
,

U
kj
p+1
is dened as the family of iid random variables satisfying common
law
P
_

U
kj
p
=
1
2
_
= P
_

U
kj
p
=
1
2
_
=
1
2
(

U
kj
q+1
,
h
p+1
W
l
)
(p,q,k,j,l)
s are mutually independent,
CHAPTER 3. PART II - IMPROVING ACCURACY 26

Z
kj
p+1
=
1
2

h
p+1
W
k

h
p+1
W
j
+

U
kj
p+1
h , k < j
Z
kj
p+1
=
1
2

h
p+1
W
k

h
p+1
W
j

U
kj
p+1
h , k > j
Z
jj
p+1
=
1
2
_
(
h
p+1
W
j
)
2
h
_
Lemma 3.1 Suppose that the b
i
s and the
ij
s are Lipschitz continuous func-
tions. Then, for any integer k, there exists a strictly positive constant C
k
such
that
E|

X
h
p
|
k
e
C
k
T
for any p between 0 and n.
Let us now consider the class F
T
of functions : [0, T] R
d
R with the
following properties: is of class C

, and for some positive integer s and positive

C(T)
[0, T] , x R
d
: |(, x)| C(T)(1 +|x|
s
)
A function of F
T
will be called homogeneous if it does not depend on the time
variable: (, x) = (x). We will denote by L the dierential operator associated
to (3.1):
(3.2) L =
1
2
d

i,j=1
a
i
j
(t, x)
ij
+
d

i=1
b
i
(t, x)
i
where a(t, x) = (t, x)

(t, x). If F
T
, then the function u(; t, x) dened by
u(; t, x) = E(, X
t,x
T
) = E
t,x
(, X
T
)
veries the following equation
_
u
t
+Lu = 0
u(; T, x) = (, x)
Lemma 3.2 Let us suppose that the function b and are C

functions, whose
derivatives of any order are bounded. For any multi-index , there exist strictly
CHAPTER 3. PART II - IMPROVING ACCURACY 27
positive constants K

(T), C

(T) such that

[0, T] : |

u(; t, x)| C

(T)(1 +|x|
k

(T)
)
Here by

we mean the mixed partial derivative of order

||

1
1

2
2

r
r
where = (
1
,
2
, . . . ,
r
), and || =
1
+ +
r
.
In all the sequel, we suppose that the functions b and are C

functions,
whose derivatives of any order are bounded. Consider a homogeneous function f
of F
T
and u(t, x) = Ef(X
t,x
T
) which solves
(3.3)
u
t
+Lu = 0
(3.4) u(T, x) = f(x)
then we have
(3.5) Err(T, h) = Eu(T,

X
h
n
) Eu(0, Y )
Let us rst consider the Euler scheme. We compute Eu(T,

X
h
n
) Eu((n
1)h,

X
h
n1
) by performing a Taylor expansion at the point ((n 1)h,

X
h
n1
) of
the form
u(t + t, x + x) = u(t, x) + t

t
u(t, x) +
1
2
(t)
2

2
t
2
u(t, x)
+t

||=1
x

t
u(t, x)

u(t, x)
+
1
2
t

||=2
x

t
u(t, x)

u(t, x)
+
1
6
t

||=3
x

t
u(t, x)

u(t, x)
+
1
||!
5

||=1
x

u(t, x) +
CHAPTER 3. PART II - IMPROVING ACCURACY 28
where = (
1
, . . . ,
r
) and x

means x

= (x
1
)

1
. . . (x
r
)

r
We need to do the following easy and tedious computation using (3.2), (3.3), (3.4)
with t = h and x =

X
h
n
=

X
h
n


X
h
n1
u(T,

X
h
n
) = u((n 1)h,

X
h
n1
) + h

t
u((n 1)h,

X
h
n1
)
+
1
2
h
2

2
t
u((n 1)h,

X
h
n1
)
+h

t
_
d

i=1
hb
i
((n 1)h,

X
h
n1
) +
l

j=1

i
j
((n 1)h,

X
h
n1
)
h
n
W
j
_
+
We get
(3.6) Eu(T,

X
h
n
) = Eu((n 1)h,

X
h
n1
) + h
2
E
e
((n 1)h,

X
h
n1
) + h
3
R
h
n
where

e
(t, x) =
1
2
d

i,j=1
b
i
(t, x)b
j
(t, x)
ij
u(t, x)
+
1
2
d

i,j,k=1
b
i
(t, x)a
j
k
(t, x)
ijk
u(t, x)
+
1
8
d

i,j,k,l=1
a
i
j
(t, x)a
k
l
(t, x)
ijkl
u(t, x) +
1
2

2
t
2
u(t, x)
+
d

i=1
b
i
(t, x)

t

i
u(t, x) +
1
2
d

i,j=1
a
i
j
(t, x)

t

ij
u(t, x)
and there exists a constant C(T) independent of h such that
|R
h
n
| C(T)
We can use the same expansion for u((n 1)h,

X
h
n1
), continuing in this way n
times, and we arrive to
(3.7) Eu(T,

X
h
n
) = Eu(0, Y ) + h
2
n1

j=0
E
e
(jh,

X
h
j
) + h
2
R
h
n
CHAPTER 3. PART II - IMPROVING ACCURACY 29
(3.8) Eu(T,

X
h
n
) Eu(0, Y ) = h
2
n1

j=0
E
e
(jh,

X
h
j
) + h
2
R
h
n
(3.9) Err(T, h) = h
2
n1

j=0
E
e
(jh,

X
h
j
) + h
2
R
h
n
by (3.5)
with
|R
h
n
| C(T)
where R
h
n
=

n
j=1
R
h
j
.
Proposition 3.3 There exists a real number C(T), independent on h, such that
(3.10) h
n1

j=0
E|
e
(jh,

X
h
j
)| C(T)
Proof. By Lemma 3.2,
E|
e
(jh,

X
h
j
)| C(1 + E|

X
h
j
|
K
)
and by Lemma 3.1,
E|
e
(jh,

X
h
j
)| K
where C and K are constants independent of h and j. Hence,
h
n1

j=0
E|
e
(jh,

X
h
j
)| hnK = TK since h = T/n
Proposition 3.4 For any function of F
T
, there exists a real number C(T),
independent on h, such that
E(,

X
h
T
) = E(, X
T
)R
T
(h)
with |R
T
(h)| C(T)h.
Proposition 3.5
(3.11)

h
n1

j=1
E
e
(jh,

X
h
j
)
_
T
0
E
e
(s, X
s
)ds

= O(h)
CHAPTER 3. PART II - IMPROVING ACCURACY 30
Remark 3.6 For the Milshtein scheme,
m
can be derived by performing a Tay-
lor expansion:

m
(t, x) =
e
(t, x)
+
1
4

i
1
,i
2
,j,k,l
a
l
k
(t, x)
l

i
1
j
(t, x)
k

i
2
j
(t, x)
i
1
,i
2
u(t, x)
+
1
2

i
1
,i
2
,i
3
,j
1
,j
2
,k

i
1
j
1
(t, x)
i
2
j
2
(t, x)
k
j
1
(t, x)
k

i
3
j
2
(t, x)
i
1
,i
2
,i
3
u(t, x)
and
m
can be substituted into (3.6),(3.7),(3.10), and (3.11) instead of
e
.
3.2 Main results
Theorem 3.7 For the Euler scheme, the error is given by
(3.12) Err
e
(T, h) = h
_
T
0
E
e
(s, X
s
)ds +O(h
2
)
Proof.
Err
e
(T, h) = h
_
h
n1

j=0
E
e
(jh,

X
h
j
) +O(h)
_
by (3.7) (3.8) and (3.9)
Err
e
(T, h) = h
_

_
T
0
E
e
(s, X
s
)ds +O(h)
_
by Proposition 3.5
Err
e
(T, h) = h
_
T
0
E
e
(s, X
s
)ds +O(h
2
) as required
Theorem 3.8 For the Milshtein scheme, the error is given by
Err
m
(T, h) = h
_
T
0
E
m
(s, X
s
)ds +O(h
2
)
Proof. goes in the same way as that of
e
in Theorem 3.7 with only replacing

e
by
m
.
Remark 3.9 We have proved that for the Euler approximation

X
h
T
and smooth
CHAPTER 3. PART II - IMPROVING ACCURACY 31
functions f
Ef(

X
h
T
) Ef(X
T
) = h
_
T
0
E
e
(s, X
s
)ds + g
h
(T)
holds for all h, h/2, h/3, h/4, . . . with some functions
e
and g
h
, where
e
is in-
dependent of the step size h = T/n and g
h
is a function depending on h such
that
|g
h
| Ch
2
with some constant C independent of h.
Hence, the following important consequence holds.
Theorem 3.10 (Richardson extrapolation) Consider that the following new
approximation
(3.13) Z
h
T
= 2Ef(

X
h/2
T
) Ef(

X
h
T
)
where

X
h/2
T
means that we perform an approximation with the step size h/2. Then
Err(T, h) = Ef(X
T
) Z
h
T
Ch
2
where C is a constant indepedent of h.
Proof. Using (3.12) with h and h/2
Ef(

X
h
T
) = Ef(X
T
) + e
1
(T)h + g
h
(T))
Ef(

X
h/2
T
) = Ef(X
T
) + e
1
(T)
h
2
+ g
h/2
(T)
where
e
1
(T) =
_
T
0
E
e
(s, X
s
)ds
Hence, by dierence
Z
h
T
= 2Ef(

X
h/2
T
) Ef(

X
h
T
) = Ef(X
T
) + 2g
h/2
(T) g
h
(T)
Z
h
T
Ef(X
T
) = 2g
h/2
(T) g
h
(T))
CHAPTER 3. PART II - IMPROVING ACCURACY 32
Notice that,
|2g
h/2
(T) g
h
(T))| |2g
h/2
(T)| +|g
h
(T))|
2
Ch
2
4
+ Ch
2
=
3
2
Ch
2
Therefore, Err(T, h)
3
2
Ch
2
as required.
We have concluded from the above Theorem 3.10 that it is possible to get a
result of precision of second order from results given by a rst order scheme. We
realize that extrapolation is an elegant and simple way to obtain weak higher
order methods. The construction of such methods is based on the existence of an
asymptotic expansion of the error with respect to powers of the time step size.
Theorem 3.11 For the second order scheme, the global expansion error can be
written
Err(T, h) = h
2
_
T
0
E(s, X
s
)ds +O(h
3
)
for some smooth function .
Theorem 3.12 For all these schemes, it is possible to obtain an expression of
the form
Err(T, h) = e
r
(T)h
r
+ e
r+1
(T)h
r+1
+ + e
m
(T)h
m
+O(h
m+1
)
where e
1
, e
2
, . . . , e
m
are independent of h, and O(h
m+1
) means a function g of h
such that
|g
h
| Ch
m+1
for all h with a constant C.
Denition 3.13 For step sizes h, h/2, h/4, h/8, . . ., Richardson extrapolation method
is formulated recursively as follows, by using
Z
h
T,,
=
1
2
1
1
(2
1
Z
h/2
T,1,1
Z
h
T,1,1
)
CHAPTER 3. PART II - IMPROVING ACCURACY 33
Z
h
T,1,2
= 2Ef(

X
h/2
T
) Ef(

X
h
T
) (3.14)
Z
h
T,2,3
=
1
3
(4Z
h/2
T,1,2
Z
h
T,1,2
) (3.15)
Z
h
T,3,4
=
1
7
(8Z
h/2
T,2,3
Z
h
T,2,3
) (3.16)
Z
h
T,4,5
=
1
15
(16Z
h/2
T,3,4
Z
h
T,3,4
) (3.17)
Remark 3.14 Theorem 3.12 implies that there exist universal constants
0
,
1
, . . . ,
m
such that for
(3.18) Z
h
T
=
0
Ef(

X
h
T
) +
1
Ef(

X
h/2
T
) +
2
Ef(

X
h/4
T
) + +
m
Ef(

X
h/2
m
T
)
where

m
i=1

i
= 1, we have
(3.19) |Z
h
T
Ef(X
T
)| Ch
m+1
Continuing with Theorem 3.10, we could apply the second order extraploation
(3.14) in (3.15) using the Euler scheme for step sizes h/4, h/2, h to obtain the
third order method
(3.20) Z
h
T,1,3
=
1
3
(8

X
h/4
T
6

X
h/2
T
+

X
h
T
)
where Z
h
T,1,3
implies that we could get a result of precision of third order from re-
sults given by a rst order scheme. Actually, it is possible to extend the argument
if the following is veried by Theorem 3.12.
(3.21) Err(T, h) = e
1
(T)h
1
+ e
2
(T)h
2
+O(h
3
)
where
e
1
(T) =
_
T
0
E
e
(s, X
s
)ds
and e
2
(T) =
_
T
0
E(s, X
s
)ds from Theorem 3.11
CHAPTER 3. PART II - IMPROVING ACCURACY 34
Using (3.21) and (3.20), we get
8Ef(

X
h/4
T
) = 8Ef(X
T
) + 8e
1
(T)
h
4
+ 8e
2
(T)
h
2
4
2
+ 8g
h/4
(T)
6Ef(

X
h/2
T
) = 6Ef(X
T
) 6e
1
(T)
h
2
6e
2
(T)
h
2
2
2
6g
h/2
(T)
Ef(

X
h
T
) = Ef(X
T
) + e
1
(T)h + e
2
(T)h
2
+ g
h
(T)
Hence,
Z
h
T,1,3
= Ef(X
T
) + 8g
h/4
(T) 6g
h/2
(T) + g
h
(T)
Err(T, h) = 8g
h/4
(T) 6g
h/2
(T) + g
h
(T)
Notice that
|8g
h/4
(T) 6g
h/2
(T) + g
h
(T)| 8
h
3
4
3
+ 6
h
3
2
3
+ h
3
= Ch
3
Consequently, we have got a result of precision of order h
3
from the result given
by a second order scheme. Therefore, if we use recursively Richardson extrapo-
lation methods dened in Denition 3.13 with Theorem 3.12 in sequence, we can
eventually arrive to (3.19). The choice of weights (
0
,
1
, . . . ,
m
) in the linear
combination (3.18) of the outcomes from the Euler approximations with step size
h, h/2, . . . , h/2
m
causes the error term of order h, h
2
, . . . , h
m
in the leading error
expansion to cancel out asymptotically. The remaining error term involve the
power h
m+1
indicating we have a (m + 1) order weak scheme.
3.3 Demonstration
Consider the Ito process X satisfying the linear stochastic dierential equation
dX
t
= bX
t
dt + X
t
dW
t
X
t
= X
0
exp
__
b
1
2

2
_
t + W
t
_
CHAPTER 3. PART II - IMPROVING ACCURACY 35
6 5.5 5 4.5 4 3.5 3
13
12
11
10
9
8
7
6
5
4
log2(h)
l
o
g
2
|
E
r
r
(
T
,
h
)
|
Euler scheme of first order
Extrapolation method of second order
Euler
Extrapolation
Figure 3.1: log
2
of the mean error versus log
2
h
CHAPTER 3. PART II - IMPROVING ACCURACY 36
5 4.8 4.6 4.4 4.2 4 3.8 3.6 3.4 3.2 3
13
12
11
10
9
8
7
log2(h)
l
o
g
2
|
E
r
r
(
T
,
h
)
|
Extrapolation of second order
Extrapolation of third order
second order
third order
Figure 3.2: log
2
of the mean error versus log
2
h
CHAPTER 3. PART II - IMPROVING ACCURACY 37
with X
0
= 0.1, b = 1.5 and = 0.01 on the time interval [0, T] where T = 1. Use
the following Euler scheme

X
h
0
= X
0

X
h
p+1
=

X
h
p
+ b(ph,

X
h
p
)h + (ph,

X
h
p
)
h
p+1
W
where
h
p+1
W = W
(p+1)h
W
ph
, in order to simulate the order 2.0 weak extrap-
olation
Z
h
T,1,2
= 2Ef(

X
h/3
T
) Ef(

X
h
T
)
and the order 3.0 weak extrapolation
Z
h
T,1,3
=
1
3
(8

X
h/4
T
6

X
h/2
T
+

X
h
T
)
for f(x) = x and h = 2
3
. We generate 1000 trajectories for
Err(T, h) = Z
h
T
Ef(X
T
)
We repeat the calculation for step size h = 2
4
, 2
5
, 2
6
and 2
7
, and plot the
results on log
2
|Err(T, h)| versus log
2
h. From Figure 3.1 we see that the exponen-
tial mean error for the extrapolation method follows a steeper line than that of the
Euler scheme and that the accuracy of the weak error of Eulers approximation by
the extrapolation method of second order is improved. Furthermore, from Figure
3.2, we nd the accuracy of the weak error become smaller by implementing the
extrapolation method of third order.
Chapter 4
Part III - Implementation for
Strong Approximation
Consider the Ito process X satisfying the linear stochastic dierential equation
dX
t
= bX
t
dt + X
t
dW
t
X
t
= X
0
exp
__
b
1
2

2
_
t + W
t
_
with X
0
= 0.1, b = 1.5 and = 0.2 on the time interval [0, T] where T = 1. Use
the following Euler scheme

X
h
0
= X
0

X
h
p+1
=

X
h
p
+ b(ph,

X
h
p
)h + (ph,

X
h
p
)
h
p+1
W
where
h
p+1
W = W
(p+1)h
W
ph
, in order to simulate the order 1.0 Milshtein
scheme

Z
h
0
= X
0

Z
h
p+1
=

Z
h
p
+ b(ph,

Z
h
p
)h + (ph,

Z
h
p
)
h
p+1
W
+
1
2
(ph,

Z
h
p
)(ph,

Z
h
p
)

{(
h
p+1
W)
2
h}
We generate 100 trajectories for
Err(T, h) = E|

Z
h
T
f(X
T
)|
38
CHAPTER 4. PART III - IMPLEMENTATIONFORSTRONGAPPROXIMATION39
9 8 7 6 5 4 3
10
9
8
7
6
5
4
log2(h)
l
o
g
2
|
E
r
r
(
T
,
h
)
|
Milshtein scheme of order 1.0
Euler scheme of order 0.5
duplicate line
Figure 4.1: log
2
of the pathwise error versus log
2
h
CHAPTER 4. PART III - IMPLEMENTATIONFORSTRONGAPPROXIMATION40
We repeat the calculation for step size h = 2
4
, 2
5
, 2
6
, 2
7
, 2
8
and 2
9
, and
plot the results on log
2
|Err(T, h)| versus log
2
h.
From Figure 4.1, regrettably, we see that the pathwise error of Milstein scheme
is almost the same as that of Euler scheme despite of the two theoretical re-
sults(Theorem 2.1 and Theorem 2.2). Unexpectedly, Milstein scheme does not
make a signicant improvement on the pathwise error in case of dimension 1.
Chapter 5
Conclusion
The conclusion which can be drawn from this study of numerical solution of
stochastic dierential equations are these:
I. It is important that the trajectories, that is the sample paths, of the approx-
imation should be close to those of the Ito process in some problems. We
need to consider the absolute error at the nal time instant T, that is
() = (E|X
T
Y
N
|
q
)
1/q
for k 1. The absolute error is certainly a criterion for the closeness of
the sample paths of the Ito process X and the approximation Y at time
T. We shall say that an approximation process Y converges in the strong
sense with order (0, ] if there exists a nite constant K and a positive
constant
0
such that
E|X
T
Y
N
| K
r
for any time discretization with maximum step size (0,
0
). The Euler
aprroximation for SDEs
Y
n+1
= Y
n
+ a(Y
n
)
n
+ b(Y
n
)W
n
has strong order = 0.5.
II. We may be interested only in some function of the value of the Ito process
at a given nal time T such as one of the rst two moments EX
T
and
E(X
T
)
2
or, more generally, the expectation E(g(X
T
)) for some function g.
In simulating such a functional it suces to have a good approximation
41
CHAPTER 5. CONCLUSION 42
of the probability distribution of the random variable X
T
rather than a
close approximation of sample paths. we shall say that a time discrete
approximation Y converges in the weak sense with order (0, ] if for
any polynomial g there exists a nite constant K and a positive constant

0
such that
E(g(X
T
)) E(g(Y
N
)) K

for any time discretization with maximum step size (0,

0
). We see
that an Euler approximation of an Ito process converges with weak order
= 1.0, which is greater than its strong order of convergence = 0.5.
III. we obtain the Milstein scheme by adding the additional term
Y
n+1
= Y
n
+ a
n
+ bW
n
+
1
2
bb

{(W
n
)
2

n
}
we shall see that the Milstein scheme converges with strong order = 1.0
under the assumption that E(X
0
)
2
< , that a and b are twice continu-
ously dierentiable, and that a, a

, b, b

and b

satisfy a uniform Lipschitz

condition.
IV. It is the purpose of this paper to attempt to represent Err(T, h) as a function
of h
(5.1) Err(T, h) = e
1
(T)h + e
2
(T)h
2
+ + e
m
(T)h
m
+O(h
m+1
)
If the above result is veried, higher order approximations of functionals
can be obtained with lower order weak schemes by extrapolation methods.
D.Talay and L.Tubaro [1] proposed an order 2.0 weak extrapolation method
Z
h
T,1,2
= 2Ef(

X
h/2
T
) Ef(

X
h
T
)
where

X
h/2
T
and

X
h
T
are the Euler approximation at time T for the step size
h/2 and h respectively. Generally, if we consider
Z
h
T
=
0
Ef(

X
h
T
) +
1
Ef(

X
h/2
T
) +
2
Ef(

X
h/4
T
) + +
m
Ef(

X
h/2
m
T
)
for a sequence
h > h/2 > h/4 > . . . > h/2
m
CHAPTER 5. CONCLUSION 43
we can get eventually from (5.1)
|Z
h
T
Ef(X
T
)| O(h
m+1
)
Bibliography
[1] D.Talay and L.Tubaro, Expansion of the global error for numerical schemes
solving stochastic dierential equation, Stochastic Analysis and Applications
8(4), 483-509, 1990.
[2] Peter E Kloeden and Eckhard Platen, Numerical Solution of Stochastic Dif-
ferential Equations, Springer third edition, 1999.
[3] Peter E Kloeden, Eckhard Platen and Henri Schurz, Numerical Solution of
SDE Through Computer Experiments, Springer second edition, 1997.
44