Stochastic Calculus: An Introduction with

Applications

Gregory F. Lawler

c 2014, Gregory F. Lawler
All rights reserved
ii
Contents

1 Martingales in discrete time 3

1.1 Conditional expectation . . . . . . . . . . . . . . . . . . . . . 3
1.2 Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Optional sampling theorem . . . . . . . . . . . . . . . . . . . 13
1.4 Martingale convergence theorem and Polyas urn . . . . . . . 18
1.5 Square integrable martingales . . . . . . . . . . . . . . . . . . 23
1.6 Integrals with respect to random walk . . . . . . . . . . . . . 25
1.7 A maximal inequality . . . . . . . . . . . . . . . . . . . . . . 26
1.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2 Brownian motion 33
2.1 Limits of sums of independent variables . . . . . . . . . . . . 33
2.2 Multivariate normal distribution . . . . . . . . . . . . . . . . 36
2.3 Limits of random walks . . . . . . . . . . . . . . . . . . . . . 40
2.4 Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.5 Construction of Brownian motion . . . . . . . . . . . . . . . . 43
2.6 Understanding Brownian motion . . . . . . . . . . . . . . . . 48
2.6.1 Brownian motion as a continuous martingale . . . . . 51
2.6.2 Brownian motion as a Markov process . . . . . . . . . 53
2.6.3 Brownian motion as a Gaussian process . . . . . . . . 54
2.6.4 Brownian motion as a self-similar process . . . . . . . 54
2.7 Computations for Brownian motion . . . . . . . . . . . . . . . 55
2.8 Quadratic variation . . . . . . . . . . . . . . . . . . . . . . . . 59
2.9 Multidimensional Brownian motion . . . . . . . . . . . . . . . 63
2.10 Heat equation and generator . . . . . . . . . . . . . . . . . . 65
2.10.1 One dimension . . . . . . . . . . . . . . . . . . . . . . 65
2.10.2 Expected value at a future time . . . . . . . . . . . . . 70
2.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

iii
iv CONTENTS

3 Stochastic integration 79
3.1 What is stochastic calculus? . . . . . . . . . . . . . . . . . . . 79
3.2 Stochastic integral . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.1 Review of Riemann integration . . . . . . . . . . . . . 81
3.2.2 Integration of simple processes . . . . . . . . . . . . . 82
3.2.3 Integration of continuous processes . . . . . . . . . . . 85
3.3 Itos formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.4 More versions of Itos formula . . . . . . . . . . . . . . . . . . 100
3.5 Diffusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.6 Covariation and the product rule . . . . . . . . . . . . . . . . 111
3.7 Several Brownian motions . . . . . . . . . . . . . . . . . . . . 112
3.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

4 More stochastic calculus 119

4.1 Martingales and local martingales . . . . . . . . . . . . . . . . 119
4.2 An example: the Bessel process . . . . . . . . . . . . . . . . . 124
4.3 Feynman-Kac formula . . . . . . . . . . . . . . . . . . . . . . 127
4.4 Binomial approximations . . . . . . . . . . . . . . . . . . . . 131
4.5 Continuous martingales . . . . . . . . . . . . . . . . . . . . . 135
4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

5 Change of measure and Girsanov theorem 139

5.1 Absolutely continuous measures . . . . . . . . . . . . . . . . . 139
5.2 Giving drift to a Brownian motion . . . . . . . . . . . . . . . 143
5.3 Girsanov theorem . . . . . . . . . . . . . . . . . . . . . . . . . 146
5.4 Black-Scholes formula . . . . . . . . . . . . . . . . . . . . . . 155
5.5 Martingale approach to Black-Scholes equation . . . . . . . . 159
5.6 Martingale approach to pricing . . . . . . . . . . . . . . . . . 162
5.7 Martingale representation theorem . . . . . . . . . . . . . . . 170
5.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

6 Jump processes 177

6.1 Levy processes . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6.2 Poisson process . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6.3 Compound Poisson process . . . . . . . . . . . . . . . . . . . 184
6.4 Integration with respect to compound Poisson processes . . . 192
6.5 Change of measure . . . . . . . . . . . . . . . . . . . . . . . . 197
6.6 Generalized Poisson processes I . . . . . . . . . . . . . . . . . 198
6.7 Generalized Poisson processes II . . . . . . . . . . . . . . . . 203
6.8 The Levy-Khinchin characterization . . . . . . . . . . . . . . 207
CONTENTS v

6.9 Integration with respect to Levy processes . . . . . . . . . . . 211

6.10 Symmetric stable process . . . . . . . . . . . . . . . . . . . . 213
6.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

7 Fractional Brownian motion 223

7.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
7.2 Stochastic integral representation . . . . . . . . . . . . . . . . 225
7.3 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

8 Harmonic functions 229

8.1 Dirichlet problem . . . . . . . . . . . . . . . . . . . . . . . . . 229
8.2 h-processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
8.3 Time changes . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
8.4 Complex Brownian motion . . . . . . . . . . . . . . . . . . . 238
8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
vi CONTENTS
Introductory comments

This is an introduction to stochastic calculus. I will assume that the reader

has had a post-calculus course in probability or statistics. For much of these
notes this is all that is needed, but to have a deep understanding of the
subject, one needs to know measure theory and probability from that per-
spective. My goal is to include discussion for readers with that background
as well. I also assume that the reader can write simple computer programs
either using a language like C++ or by using software such as Matlab or
Mathematica.

More advanced mathematical comments that can be skipped by

the reader will be indented with a different font. Comments here will
assume that the reader knows that language of measure-theoretic
probability theory.

We will discuss some of the applications to finance but our main focus
will be on the mathematics. Financial mathematics is a kind of applied
mathematics, and I will start by making some comments about the use of
mathematics in the real world. The general paradigm is as follows.
A mathematical model is made of some real world phenomenon. Usu-
ally this model requires simplification and does not precisely describe
the real situation. One hopes that models are robust in the sense that
if the model is not very far from reality then its predictions will also
be close to accurate.

The model consists of mathematical assumptions about the real world.

Given these assumptions, one does mathematical analysis to see what

they imply. The analysis can be of various types:

Rigorous derivations of consequences.

1
2 CONTENTS

Derivations that are plausible but are not mathematically rigor-

ous.
Approximations of the mathematical model which lead to
tractable calculations.
Numerical calculations on a computer.
For models that include randomness, Monte Carlo simulations
using a (pseudo) random number generator.

If the mathematical analysis is successful it will make predictions about

the real world. These are then checked.

If the predictions are bad, there are two possible reasons:

The mathematical analysis was faulty.

The model does not sufficiently reflect reality.

The user of mathematics does not always need to know the details of
the mathematical analysis, but it is critical to understand the assumptions
in the model. No matter how precise or sophisticated the analysis is, if the
assumptions are bad, one cannot expect a good answer.
Chapter 1

Martingales in discrete time

A martingale is a mathematical model of a fair game. To understand the def-

inition, we need to define conditional expectation. The concept of conditional
expectation will permeate this book.

1.1 Conditional expectation

If X is a random variable, then its expectation, E[X] can be thought of as
the best guess for X given no information about the result of the trial. A
conditional expectation can be considered as the best guess given some but
not total information.
Let X1 , X2 , . . . be random variables which we think of as a time series
with the data arriving one at a time. At time n we have viewed the values
X1 , . . . , Xn . If Y is another random variable, then E(Y | X1 , . . . , Xn ) is the
best guess for Y given X1 , . . . , Xn . We will assume that Y is an integrable
random variable which means E[|Y |] < . To save some space we will
write Fn for the information contained in X1 , . . . , Xn and E[Y | Fn ] for
E[Y | X1 , . . . , Xn ]. We view F0 as no information. The best guess should
satisfy the following properties.
If we have no information, then the best guess is the expected value.
In other words, E[Y | F0 ] = E[Y ].

The conditional expectation E[Y | Fn ] should only use the informa-

tion available at time n. In other words, it should be a function of
X1 , . . . , Xn ,
E[Y | Fn ] = (X1 , . . . , Xn ).
We say that E[Y | Fn ] is Fn -measurable.

3
4 CHAPTER 1. MARTINGALES IN DISCRETE TIME

The definitions in the last paragraph are certainly vague. We

can use measure theory to be precise. We assume that the random
variables Y, X1 , X2 , . . . are defined on a probability space (, F, P).
Here F is a -algebra or -field of subsets of , that is, a collection
of subsets satisfying
F;
A F implies that \ A F;
A1 , A2 , . . . F implies that
S
n=1 An F.
The information Fn is the smallest sub -algebra G of F such that
X1 , . . . , Xn are G-measurable. The latter statement means that for
all t R, the event {Xj t} Fn . The no information -
algebra F0 is the trivial -algebra containing only and .

The definition of conditional expectation is a little tricky, so let us try

to motivate it by considering an example from undergraduate probability
courses. Suppose that (X, Y ) have a joint density

f (x, y), 0 < x, y < ,

with marginal densities

Z Z
f (x) = f (x, y) dy, g(y) = f (x, y) dx.

The conditional density f (y|x) is defined by

f (x, y)
f (y|x) = .
f (x)

This is well defined provided that f (x) > 0, and if f (x) = 0, then x is an
impossible value for X to take. We can write
Z
E[Y | X = x] = y f (y | x) dy.

We can use this as the definition of conditional expectation in this case,

Z R
y f (X, y) dy
E[Y | X] = y f (y | X) dy = .
f (X)
1.1. CONDITIONAL EXPECTATION 5

Note that E[Y | X] is a random variable which is determined by the value

of the random variable X. Since it is a random variable, we can take its
expectation
Z
E [E[Y | X]] = E[Y | X = x] f (x) dx

Z Z 
= y f (y | x) dy f (x) dx
Z
Z

= y f (x, y) dy dx

= E[Y ].

This calculation illustrates a basic property of conditional expectation.

Suppose we are interested in the value of a random variable Y and we are
going to be given data X1 , . . . , Xn . Once we observe the data, we make
our best prediction E[Y | X1 , . . . , Xn ]. If we average our best prediction
given X1 , . . . , Xn over all the possible values of X1 , . . . , Xn , we get the best
prediction of Y . In other words,

E[Y ] = E [E[Y | Fn ]] .

More generally, suppose that A is an Fn -measurable event, that is to say,

if we observe the data X1 , . . . , Xn , then we know whether or not A has
occurred. An example of an F4 -measurable event would be

A = {X1 X2 , X4 < 4}.

Let 1A denote the indicator function (or indicator random variable) associ-
ated to the event A,

1 if A occurs
1A = .
0 if A does not occur

Using similar reasoning, we can see that if A is Fn -measurable, then

E [Y 1A ] = E [E[Y | Fn ] 1A ] .

At this point, we have not derived this relation mathematically; in fact, we

have not even defined the conditional expectation. Instead, we will use this
reasoning to motivate the following definition.

Definition The conditional expectation E[Y | Fn ] is the unique random

variable satisfying the following.
6 CHAPTER 1. MARTINGALES IN DISCRETE TIME

E[Y | Fn ] is Fn -measurable.
For every Fn -measurable event A,

E [E[Y | Fn ] 1A ] = E [Y 1A ] .

We have used different fonts for the E of conditional expectation and the
E of usual expectation in order to emphasize that the conditional expectation
is a random variable. However, most authors use the same font for both
leaving it up to the reader to determine which is being referred to.

Suppose (, F, P) is a probability space and Y is an integrable

random variable. Suppose G is a sub -algebra of F. Then E[Y | G]
is defined to be the unique (up to an event of measure zero) G-
measurable random variable such that if A G,

E [Y 1A ] = E [E[Y | G] 1A ] .

Uniqueness follows from the fact that if Z1 , Z2 are G-measurable

random variables with

E [Z1 1A ] = E [Z2 1A ]

for all A G, then P{Z1 6= Z2 } = 0. Existence of the condi-

tional expectation can be deduced from the Radon-Nikodym theo-
rem. Let (A) = E [Y 1A ]. Then is a (signed) measure on (, G, P)
with  P, and hence there exists an L1 random variable Z with
(A) = E [Z1A ] for all A G.

Although the definition does not give an immediate way to calculate the
conditional expectation, in many cases one can compute it. We will give a
number of properties of the conditional expectation most of which follow
quickly from the definition.
Proposition 1.1.1. Suppose X1 , X2 , . . . is a sequence of random variables
and Fn denotes the information at time n. The conditional expectation E[Y |
Fn ] satisfies the following properties.
If Y is Fn -measurable, then E[Y | Fn ] = Y .
If A is an Fn -measurable event, then E [E[Y | Fn ] 1A ] = E [Y 1A ]. In
particular,
E [E[Y | Fn ]] = E[Y ].
1.1. CONDITIONAL EXPECTATION 7

Suppose X1 , . . . , Xn are independent of Y . Then Fn contains no useful

information about Y and hence

E[Y | Fn ] = E[Y ].

Linearity. If Y, Z are random variables and a, b are constants, then

E[aY + bZ | Fn ] = a E[Y | Fn ] + b E[Z | Fn ]. (1.1)

Projection or Tower Property. If m < n, then

E [ E[Y | Fn ] | Fm ] = E[Y | Fm ]. (1.2)

If Z is an Fn -measurable random variable, then when conditioning

with respect to Fn , Z acts like a constant,

E[Y Z | Fn ] = Z E[Y | Fn ]. (1.3)

The proof of this proposition is not very difficult given our choice
of definition for the conditional expectation. We will discuss only a
couple of cases here, leaving the rest for the reader. To prove the
linearity property, we know that a E[Y | Fn ] + b E[Z | Fn ] is an
Fn -measurable random variable. Also if A Fn , then linearity of
expectation implies that

E [1A (a E[Y | Fn ] + b E[Z | Fn ])]

= aE [1A E[Y | Fn ]] + b E [1A E[Z | Fn ]]
= aE [1A Y ] + bE [1A Z]
= E [1A (aY + bZ)] .

Uniqueness of the conditional expectation then implies (1.1).

We first show the constants rule (1.3) for Z = 1A , A Fn ,
as follows. If B Fn , then A B Fn and

E [1B E(Y Z | Fn )] = E [1B E(1A Y | Fn )]

= E [1B 1A Y ] = E [1AB Y ] = E [1AB E(Y | Fn )]

= E [1B 1A E(Y | Fn )] = E [1B Z E(Y | Fn )] .
8 CHAPTER 1. MARTINGALES IN DISCRETE TIME

Hence E(Y Z | Fn ) = Z E(Y | Fn ) by definition. Using linearity,

the rule holds for simple random variables of the form
n
X
Z= aj 1Aj , Aj Fn .
j=1

We can then prove it for nonnegative Y by approximating from

below by nonnegative simple random variables and using the mono-
tone convergence theorem, and then for general Y by writing
Y = Y + Y . These are standard techniques in Lebesgue in-
tegration theory.

Example 1.1.1. Suppose that X1 , X2 , . . . are independent random variables

with E[Xj ] = for each j. Let Sn = X1 + + Xn , and let Fn denote the
information contained in X1 , . . . , Xn . Then if m < n,

E[Sn | Fm ] = E[Sm | Fm ] + E[Sn Sm | Fm ]

= Sm + E[Sn Sm ]
= Sm + (n m).

Example 1.1.2. In the same setup as Example 1.1.1 suppose that = 0

and E[Xj2 ] = 2 for each j. Then if m < n,

E[Sn2 | Fm ] = E([Sm + (Sn Sm )]2 | Fm )

2
= E[Sm | Fm ] + 2 E[Sm (Sn Sm ) | Fm ]
+E[(Sn Sm )2 | Fm ].

Since Sm is Fm -measurable and Sn Sm is independent of Fm ,

2 2
E[Sm | Fm ] = Sm ,

E[Sm (Sn Sm ) | Fm ] = Sm E[Sn Sm | Fm ] = Sm E[Sn Sm ] = 0,

E[(Sn Sm )2 | Fm ] = E[(Sn Sm )2 ] = Var(Sn Sm ) = 2 (n m),

and hence,
E[Sn2 | Fm ] = Sm
2
+ 2 (n m).
1.2. MARTINGALES 9

Example 1.1.3. In the same setup as Example 1.1.1, let us also assume
that X1 , X2 , . . . are identically distributed. We will compute E[X1 | Sn ].
Note that the information contained in the one data point Sn is less than the
information contained in X1 , . . . , Xn . However, since the random variables
are identically distributed, it must be the case that

E[X1 | Sn ] = E[X2 | Sn ] = = E[Xn | Sn ].

Linearity implies that

n
X
n E[X1 | Sn ] = E[Xj | Sn ] = E[X1 + + Xn | Sn ] = E[Sn | Sn ] = Sn .
j=1

Therefore,
Sn
E[X1 | Sn ] =.
n
It may be at first surprising that the answer does not depend on E[X1 ].

Definition If X1 , X2 , . . . is a sequence of random variables, then the asso-

ciated (discrete time) filtration is the collection {Fn } where Fn denotes the
information in X1 , . . . , Xn .

One assumption in the definition of a filtration, which may sometimes

not reflect reality, is that information is never lost. If m < n, then everything
known at time m is still known at time n. Sometimes a filtration is given
starting at time n = 1 and sometimes starting at n = 0. If it starts at time
n = 1, we define F0 to be no information.

More generally, a (discrete time) filtration {Fn } is an increasing

sequence of -algebras.

1.2 Martingales
A martingale is a model of a fair game. Suppose X1 , X2 , . . . is a sequence of
random variables to which we associate the filtration {Fn } where Fn is the
information contained in X1 , . . . , Xn .

Definition A sequence of random variables M0 , M1 , . . . is called a martin-

gale with respect to the filtration {Fn } if:
10 CHAPTER 1. MARTINGALES IN DISCRETE TIME

For each n, Mn is an Fn -measurable random variable with E[|Mn |] <

.
If m < n, then
E[Mn | Fm ] = Mm . (1.4)
We can also write (1.4) as
E[Mn Mm | Fm ] = 0.
If we think of Mn as the winnings of a game, then this implies that no
matter what has happened up to time m, the expected winnings in the next
n m games is 0. Sometimes one just says M0 , M1 , . . . is a martingale
without reference to the filtration. In this case, the assumed filtration is Fn ,
the information in M0 , . . . , Mn . In order to establish (1.4) it suffices to show
for all n,
E[Mn+1 | Fn ] = Mn . (1.5)
In order to see this, we can use the tower property (1.2) for conditional
expectation to see that
E[Mn+2 | Fn ] = E [ E[Mn+2 | Fn+1 ] | Fn ] = E[Mn+1 | Fn ] = Mn ,
and so forth. Also note that if Mn is a martingale, then
E[Mn ] = E [E[Mn | F0 ]] = E[M0 ].
Example 1.2.1. Suppose X1 , X2 , . . . are independent random variables
with E[Xj ] = 0 for each j. Let S0 = 0 and Sn = X1 + + Xn . In the
last section we showed that if m < n, then E[Sn | Fm ] = Sm . Hence, Sn is
a martingale with respect to Fn , the information in X1 , . . . , Xn .
Example 1.2.2. Suppose Xn , Sn , Fn are as in Example 1.2.1 and also as-
sume Var[Xj ] = E[Xj2 ] = j2 < . Let
An = 12 + + n2 ,
Mn = Sn2 An ,
where M0 = 0. Then Mn is a martingale with respect to Fn . To see this, we
compute as in Example 1.1.2,
2
E[Sn+1 | Fn ] = E[(Sn + Xn+1 )2 | Fn ]
= E[Sn2 | Fn ] + 2E[Sn Xn+1 | Fn ] + E[Xn+1
2
| Fn ]
= Sn2 + 2Sn E[Xn+1 | Fn ] + E[Xn+1
2
]
= Sn2 + 2Sn E[Xn+1 ] + E[Xn+1
2
]
= Sn2 + n+1
2
.
1.2. MARTINGALES 11

Therefore,
2
E[Mn+1 | Fn ] = E[Sn+1 An+1 | Fn ]
= Sn2 + n+1
2 2
(An + n+1 ) = Mn .

There are various ways to view a martingale. One can consider Mn as

the price of an asset (although we allow negative values of Mn ) or as the
winnings in a game. We can also consider

Mn = Mn Mn1

as either the change in the asset price or as the amount won in the game
at time n. Negative values indicate drops in price or money lost in the
game. The basic idea of stochastic integration is to allow one to change
ones portfolio (in the asset viewpoint) or change ones bet (in the game
viewpoint). However, we are not allowed to see the outcome before betting.
We make this precise in the next example.

Example 1.2.3. Discrete stochastic integral. Suppose that M0 , M1 , . . .

is a martingale with respect to the filtration Fn . For n 1, let Mn =
Mn Mn1 . Let Bj denote the bet on the jth game. We allow negative
values of Bj which indicate betting that the price will go down or the game
will be lost. Let Wn denote the winnings in this strategy: W0 = 0 and for
n 1,
Xn Xn
Wn = Bj [Mj Mj1 ] = Bj Mj .
j=1 j=1

Let us assume that for each n there is a number Kn < such that |Bn |
Kn . We also assume that we cannot see the result of nth game before betting.
This last assumption can be expressed mathematically by saying that Bn
is Fn1 -measurable. In other words, we can adjust our bet based on how
well we have been doing. We claim that under these assumptions, Wn is a
martingale with respect to Fn . It is clear that Wn is measurable with respect
to Fn , and integrability follows from the estimate
n
X
E[|Wn |] E[|Bj ||Mj Mj1 |]
j=1
n
X
Kj (E[|Mj |] + E[|Mj1 |]) < .
j=1
12 CHAPTER 1. MARTINGALES IN DISCRETE TIME

Also,

E[Wn+1 | Fn ] = E[Wn + Bn+1 (Mn+1 Mn ) | Fn ]

= E[Wn | Fn ] + E[Bn+1 (Mn+1 Mn ) | Fn ].

Since Wn is Fn -measurable, E[Wn | Fn ] = Wn . Also, since Bn+1 is Fn -

measurable and M is a martingale,

E[Bn+1 (Mn+1 Mn ) | Fn ] = Bn+1 E[Mn+1 Mn | Fn ] = 0.

Therefore,
E[Wn+1 | Fn ] = Wn .

Example 1.2.3 demonstrates an important aspect of martingales. One

cannot change a discrete-time martingale to a game in ones favor with a
betting strategy in a finite amount of time. However, the next example shows
that if we are allowed an infinite amount of time we can beat a fair game.

Example 1.2.4. Martingale betting strategy. Let X1 , X2 , . . . be inde-

pendent random variables with
1
P{Xj = 1} = P{Xj = 1} = . (1.6)
2
We will refer to such random variables as coin-tossing random variables
where 1 corresponds to heads and 1 corresponds to tails. Let M0 = 0, Mn =
X1 + + Xn . We have seen that Mn is a martingale. We will consider
the following betting strategy. We start by betting $1. If we win, we quit;
otherwise, we bet $2 on the next game. If we win the second game, we quit;
otherwise we double our bet to $4 and play. Each time we lose, we double
our bet. At the time that we win, we will be ahead $1. With probability
one, we will eventually win the game, so this strategy is a way to beat a fair
game. The winnings in this game can be written as
n
X n
X
Wn = Bj Mj = Bj Xj ,
j=1 j=1

where the bet B1 = 1 and for j > 1,

Bj = 2j1 if X1 = X2 = = Xj1 = 1,

and otherwise Bj = 0. This is an example of a discrete stochastic integral as

in the previous example, and hence, we know that Wn must be a martingale.
1.3. OPTIONAL SAMPLING THEOREM 13

In particular, for each n, E[Wn ] = 0. We can check this directly by noting

that Wn = 1 unless X1 = X2 = = Xn = 1 in which case

Wn = 1 21 22 2n1 = [2n 1].

This last event happens with probability (1/2)n , and hence

E[Wn ] = 1 [1 2n ] [2n 1] 2n = 0.

However, we will eventually win which means that with probability one

W = lim Wn = 1,
n

and
1 = E[W ] > E[W0 ] = 0.
We have beaten the game (but it takes an infinite amount of time to guar-
antee it).

If the condition (1.4) is replaced with

E[Mn | Fm ] Mm ,

then the process is called a submartingale. If it is replaced with

E[Mn | Fm ] Mm ,

then it is called a supermartingale. In other words, games that are always

in ones favor are submartingales and games that are always against one
are supermartingales. (At most games in Las Vegas, ones winnings give a
supermartingale.) Under this definition, a martingale is both a submartin-
gale and a supermartingale. The terminology may seem backwards at first:
submartingales get bigger and supermartingales get smaller. The terminol-
ogy was set to be consistent with the related notion of subharmonic and
superharmonic functions. Martingales are related to harmonic functions.

1.3 Optional sampling theorem

Suppose M0 , M1 , M2 , . . . is a martingale with respect to the filtration {Fn }.
In the last section we discussed the discrete stochastic integral. Here we will
consider a particular case of a betting strategy where one bets 1 up to some
14 CHAPTER 1. MARTINGALES IN DISCRETE TIME

time and then one bets 0 afterwards. Let T be the stopping time for the
strategy. Then the winnings at time t is
n
X
M0 + Bj [Mj Mj1 ],
j=1

where Bj = 1 if j T and Bj = 0 if j > T . We can write this as

MnT ,

where n T is shorthand for min{n, T }. The time T is random, but it must

satisfy the condition that the betting rule is allowable.

Definition A nonnegative integer-valued random variable T is a stopping

time with respect to the filtration {Fn } if for each n the event {T = n} is
Fn -measurable.

The following theorem is a special case of the discrete stochastic integral.

It restates the fact that one cannot beat a martingale in finite time. We call
this the optional sampling theorem; it is also called the optional stopping
theorem.

Theorem 1.3.1 (Optional Sampling Theorem I). Suppose T is a stopping

time and Mn is a martingale with respect to {Fn }. Then Yn = MnT is a
martingale. In particular, for each n,

E [MnT ] = E [M0 ] .

If T is bounded, that is, if there exists k < such that P{T k} = 1, then

E [MT ] = E [M0 ] . (1.7)

The final conclusion (1.7) of the theorem holds since E[MnT ] = E[MT ]
for n k. What if the stopping time T is not bounded but P{T < } = 1?
Then, we cannot conclude (1.7) without further assumptions. To see this we
need only consider the martingale betting strategy of the previous section.
If we define
T = min{n : Xn = 1} = min{n : Wn = 1},
then with probability one T < and WT = 1. Hence,

1 = E [WT ] > E [W0 ] = 0.

1.3. OPTIONAL SAMPLING THEOREM 15

Often one does want to conclude (1.7) for unbounded stopping times, so
it is useful to give conditions under which it holds. Let us try to derive the
equality and see what conditions we need to impose. First, we will assume
that we stop, P{T < } = 1, so that MT makes sense. For every n < ,
we know that

E[M0 ] = E[MnT ] = E[MT ] + E[MnT MT ].

If we can show that

lim E [|MnT MT |] = 0,
n

then we have (1.7). The random variable MnT MT is zero if n T = T ,

and
MnT MT = 1{T > n} [Mn MT ].
If E[|MT |] < , then one can show that

lim E [|MT | 1{T > n}] = 0.

n

In the martingale betting strategy example, this term did not cause a prob-
lem since WT = 1 and hence E[|WT |] < .

If P{T < } = 1, then the random variables Xn = |MT | 1{T >

n} converge to zero with probability one. If E [|MT |] < , then we
can use the dominated convergence theorem to conclude that

lim E[Xn ] = 0.
n

Finally, in order to conclude (1.7) we will make the hypothesis that the
other term acts nicely.

Theorem 1.3.2 (Optional Sampling Theorem II). Suppose T is a stopping

time and Mn is a martingale with respect to {Fn }. Suppose that P{T <
} = 1, E [|MT |] < , and for each n,

lim E [|Mn | 1{T > n}] = 0. (1.8)

n

Then,
E [MT ] = E [M0 ] .
16 CHAPTER 1. MARTINGALES IN DISCRETE TIME

Let us check that the martingale betting strategy does not satisfy the
conditions of the theorem (it better not since it does not satisfy the conclu-
sion!) In fact, it does not satisfy (1.8). For this strategy, if T > n, then we
have lost n times and Wn = 1 2n . Also, P{T > n} = 2n . Therefore,
lim E [|Wn | 1{T > n}] = lim (2n 1) 2n = 1 6= 0.
n n

Checking condition (1.8) can be difficult in general. We will give one

criterion which is useful.
Theorem 1.3.3 (Optional Sampling Theorem III). Suppose T is a stopping
time and Mn is a martingale with respect to {Fn }. Suppose that P{T <
} = 1, E [|MT |] < , and that there exists C < such that for each n,
E |MnT |2 C.
 
(1.9)
Then,
E [MT ] = E [M0 ] .

To prove this theorem, first note that with probability one,

|MT |2 = lim |MT n |2 1{T n},
n

and hence by the Holder inequality and the monotone convergence

theorem,
E [|MT |]2 E |MT |2 = lim E |MT n |2 | 1{T n} C.
   
n

We need to show that (1.9) implies (1.8). If b > 0, then for every
n,
E[|MnT |2 ] C
E [|Mn | 1{|Mn | b, T > n}] .
b b
Therefore,
E [|Mn | 1{T > n}] = E [|Mn | 1{T > n, |Mn | b}]
+E [|Mn | 1{T > n, |Mn | < b}]
C
+ b P{T > n}.
b
Hence,
C C
lim sup E [|Mn | 1{T > n}] + b lim P{T > n} = .
n b n b
Since this holds for every b > 0 we get (1.8).
1.3. OPTIONAL SAMPLING THEOREM 17

Example 1.3.1. Gamblers ruin for random walk. Let X1 , X2 , . . . be

independent, coin-tosses as in (1.6) and let

Sn = 1 + X1 + + Xn .

Sn is called simple (symmetric) random walk starting at 1. We have shown

that Sn is a martingale. Let K > 1 be a positive integer and let T denote the
first time n such that Sn = 0 or Sn = K. Then Mn = SnT is a martingale.
Also 0 Mn K for all n, so (1.9) is satisfied. We can apply the optional
sampling theorem to deduce that

1 = M0 = E[MT ] = 0 P{MT = 0} + K P{MT = K}.

By solving, we get
1
P{MT = K} = .
K
This relation is sometimes called the gamblers ruin estimate for the random
walk. Note that
lim P{MT = K} = 0.
K

If we consider 1 to be the starting stake of a gambler and K to be the

amount held by a casino, this shows that with a fair game, the gambler will
almost surely lose. If = min{n : Sn = 0}, then the last equality implies
that P{ < } = 1. The property that the walk always returns to the origin
is called recurrence.

Example 1.3.2. Let Sn = X1 + + Xn be simple random walk starting

at 0. We have seen that
Mn = Sn2 n
is a martingale. Let J, K be positive integers and let

T = min{n : Sn = J or Sn = K}.

As in Example 1.3.1, we have

0 = E[S0 ] = E[ST ] = [1 P{ST = K}] (J) + P{ST = K} K,

and solving gives

J
P{ST = K} = .
J +K
18 CHAPTER 1. MARTINGALES IN DISCRETE TIME

In Exercise 1.13 it is shown that there exists C < such that for all n
2
E[MnT ] C. Hence we can use Theorem 1.3.3 to conclude that
0 = E[M0 ] = E [MT ] = E ST2 E [T ] .
 

Moreover,
E ST2 = J 2 P{ST = J} + K 2 P{ST = K}
 

K J
= J2 + K2 = JK.
J +K J +K
Therefore,
E[T ] = E ST2 = JK.
 

In particular, the expected amount of time for the random walker starting
at the origin to get distance K from the origin is K 2 .
Example 1.3.3. As in Example 1.3.2, let Sn = X1 + + Xn be simple
random walk starting at 0. Let
T = min{n : Sn = 1}, TJ = min{n : Sn = 1 or Sn = J}.
Note that T = limJ TJ and
1
P{T = } = lim P {STJ = J} = lim = 0.
J J J + 1

Therefore, P{T < } = 1, although Example 1.3.2 shows that for every J,
E[T ] E [TJ ] = J,
and hence E[T ] = . Also, ST = 1, so we do not have E[S0 ] = E[ST ]. From
this we can see that (1.8) and (1.9) are not satisfied by this example.

1.4 Martingale convergence theorem and Polyas

urn
The martingale convergence theorem describes the behavior of a martingale
Mn as n .
Theorem 1.4.1 (Martingale Convergence Theorem). Suppose Mn is a mar-
tingale with respect to {Fn } and there exists C < such that E [|Mn |] C
for all n. Then there exists a random variable M such that with probability
one
lim Mn = M .
n
1.4. MARTINGALE CONVERGENCE THEOREM AND POLYAS URN19

It does not follow from the theorem that E[M ] = E[M0 ]. For example,
the martingale betting strategy satisfies the conditions of the theorem since

E [|Wn |] = (1 2n ) 1 + (2n 1) 2n 2.

However, W = 1 and W0 = 0.

We will prove the martingale convergence theorem. The proof

uses a well-known financial strategy buy low, sell high. Suppose
M0 , M1 , . . . is a martingale such that

E [|Mn |] C < ,

for all n. Suppose a < b are real numbers. We will show that it is
impossible for the martingale to fluctuate infinitely often below a
and above b. Define a sequence of stopping times by

S1 = min{n : Mn a}, T1 = min{n > S1 : Mn b},

and for j > 1,

Sj = min{n > Tj1 : Mn a},

Tj = min{n > Sj : Mn b}.

We set up the discrete stochastic integral
n
X
Wn = Bk [Mk Mk1 ] ,
k=0

with Bn = 0 if n 1 < S1 and

Bn = 1 if Sj n 1 < Tj ,

Bn = 0 if Tj n 1 < Sj+1 .
In other words, every time the price drops below a we buy a unit
of the asset and hold onto it until the price goes above b at which
time we sell. Let Un denote the number of times by time n that we
have seen a fluctuation; that is,

Un = j if Tj < n Tj+1 .
20 CHAPTER 1. MARTINGALES IN DISCRETE TIME

We call Un the number of upcrossings by time n. Every upcrossing

results in a profit of at least b a. From this we see that

Wn Un (b a) + (Mn a).

The term a Mn represents a possible loss caused by holding a

share of the asset at the current time. Since Wn is a martingale, we
know that E[Wn ] = E[W0 ] = 0, and hence
E[a Mn ] |a| + E[|Mn |] |a| + C
E[Un ] .
ba ba ba
This holds for every n, and hence
|a| + C
E[U ] < .
ba
In particular with probability one, U < , and hence there are
only a finite number of fluctuations. We now allow a, b to run over
all rational numbers to see that with probability one,

lim inf Mn = lim sup Mn .

n n

Therefore, the limit

M = lim Mn
n
exists. We have not yet ruled out the possibility that M is , but
it is not difficult to see that if this occurred with positive probability,
then E[|Mn |] would not be uniformly bounded.

To illustrate the martingale convergence theorem, we will consider an-

other example of a martingale called Polyas urn. Suppose we have an urn
with red and green balls. At time n = 0, we start with one red ball and one
green ball. At each positive integer time we choose a ball at random from
the urn (with each ball equally likely to be chosen), look at the color of the
ball, and then put the ball back in with another ball of the same color. Let
Rn , Gn denote the number of red and green balls in the urn after the draw
at time n so that

R0 = G0 = 1, Rn + Gn = n + 2,

and let
Rn Rn
Mn = =
Rn + Gn n+2
1.4. MARTINGALE CONVERGENCE THEOREM AND POLYAS URN21

be the fraction of red balls at this time. Let Fn denote the information in
the data M1 , . . . , Mn , which one can check is the same as the information in
R1 , R2 , . . . , Rn . Note that the probability that a red ball is chosen at time
n depends only on the number (or fraction) of red balls in the urn before
choosing. It does not depend on what order the red and green balls were put
in. This is an example of the Markov property. This concept will appear a
number of times for us, so let us define it.

Definition A discrete time process Y0 , Y1 , Y2 , . . . is called Markov if for each

n, the conditional distribution of

Yn+1 , Yn+2 , . . .

given Y0 , Y1 , . . . , Yn is the same as the conditional distribution given Yn . In

other words, the only thing about the past and present that is useful for
predicting the future is the current value of the process.

We can describe the rule of Polyas urn by

P{Rn+1 = Rn + 1 | Fn } = 1 P{Rn+1 = Rn | Fn } =

Rn
P{Rn+1 = Rn + 1 | Mn } = = Mn .
n+2
We claim that Mn is a martingale with respect to Fn . To check this,

E [Mn+1 | Fn ] = E [Mn+1 | Mn ]
Rn + 1 Rn
= Mn + [1 Mn ]
n+3 n+3
Rn (Rn + 1) (n + 2 Rn )Rn
= +
(n + 2)(n + 3) (n + 2)(n + 3)
Rn (n + 3)
= = Mn .
(n + 2)(n + 3)

Since E[|Mn |] = E[Mn ] = E[M0 ] = 1/2, this martingale satisfies the condi-
tions of the martingale convergence theorem. (In fact, the same argument
shows that every martingale that stays nonnegative satisfies the conditions.)
Hence, there exists a random variable M such that with probability one,

lim Mn = M .
n
22 CHAPTER 1. MARTINGALES IN DISCRETE TIME

It turns out that the random variable Mn is really random in the sense
that it has a nontrivial distribution. In Exercise 1.11 you will show that for
each n, the distribution of Mn is uniform on
 
1 2 n+1
, ,..., ,
n+2 n+2 n+2
and from this it is not hard to see that M has a uniform distribution on
[0, 1]. You will also be asked to simulate this process to see what happens.
There is a lot of randomness in the first few draws to see what fraction of
red balls the urn will settle down to. However, for large n this ratio changes
very little; for example, the ratio after 2000 draws is very close to the ratio
after 4000 draws.
While Polyas urn seems like a toy model, it arises in a number of places.
We will give an example from Bayesian statistics. Suppose that we perform
independent trials of an experiment where the probability of success for each
experiment is (such trials are called Bernoulli trials). Suppose that we do
not know the value of , but want to try to deduce it by observing trials.
Let X1 , X2 , . . . be independent random variables with

P{Xj = 1} = 1 P{Xj = 0} = .

The (strong) law of large numbers implies that with probability one,
X1 + + Xn
lim = . (1.10)
n n
Hence, if were able to observe infinitely many trials, we could deduce
exactly.
Clearly, we cannot deduce with 100% assurance if we see only a finite
number of trials. Indeed, if 0 < < 1, there is always a chance that the first
n trials will all be failures and there is a chance they will all be successes.
The Bayesian approach to statistics is to assume that is a random variable
with a certain prior distribution. As we observe the data we update to a
posterior distribution. We will assume we know nothing initially about the
value and choose the prior distribution to the uniform distribution on [0, 1]
with density
f0 () = 1, 0 < < 1.
Suppose that after observing n trials, we have had Sn = X1 + + Xn
successes. If we know , then the distribution of Sn is binomial,
 
n k
P{Sn = k | } = (1 )nk .
k
1.5. SQUARE INTEGRABLE MARTINGALES 23

We use a form of the Bayes rule to update the density

P{Sn = k | }
fn,k () := fn ( | Sn = k) = R 1 = Cn,k k (1 )nk ,
0 P{Sn = k | x} dx
where Cn,k is the appropriate constant so that fn,k is a probability density.
This is the beta density with parameters k + 1 and n k + 1. The probability
of a success on the (n + 1)st trial given that Sn = k is the conditional
expectation of given Sn = k. A little computation which we omit shows
that Z 1
k+1 Sn + 1
E[ | Sn = k] = fn,k () d = = .
0 n+2 n+2
These are exactly the transition probabilities for Polyas urn if we view Sn +1
as the number of red balls in the urn (Sn is the number of red balls added
to the urn). The martingale convergence theorem can now be viewed as the
law of large numbers (1.10) for . Even though we do not initially know
the value of (and hence treat it as a random variable) we know that the
conditional value of given Fn approaches .
Example 1.4.1. We end with a simple example where the conditions of
the martingale convergence theorem do not apply. Let Sn = X1 + + Xn
be simple symmetric random walk starting at the origin as in the previous
section. Then one can easily see that E[|Sn |] . For this example, with
probability one
lim sup Sn = ,
n
lim inf Sn = .
n

1.5 Square integrable martingales

Definition A martingale Mn is called square integrable if for each n,
E Mn2 < .
 

Note that this condition is not as strong as (1.9). We do not require that
2

there exists a C < such that E Mn C for each n. Random variables
X, Y are orthogonal if E[XY ] = E[X] E[Y ]. Independent random variables
are orthogonal, but orthogonal random variables need not be independent. If
X1 , . . . , Xn are pairwise orthogonal random variables with mean zero, then
E[Xj Xk ] = 0 for j 6= k and by expanding the square we can see that
n
X
2
E[Xj2 ].
 
E (X1 + + Xn ) =
j=1
24 CHAPTER 1. MARTINGALES IN DISCRETE TIME

This can be thought of as a generalization of the Pythagorean theorem a2 +

b2 = c2 for right triangles. The increments of a martingale are not necessarily
independent, but for square integrable martingales they are orthogonal as
we now show.

Proposition 1.5.1. Suppose that Mn is a square integrable martingale with

respect to {Fn }. Then if m < n,

E [(Mn+1 Mn ) (Mm+1 Mm )] = 0.

Moreover, for all n,

n
X
E[Mn2 ] = E[M02 ] + E (Mj Mj1 )2 .
 

j=1

Proof. If m < n, then Mm+1 Mm is Fn -measurable, and hence

E [(Mn+1 Mn ) (Mm+1 Mm ) | Fn ]

= (Mm+1 Mm ) E[Mn+1 Mn | Fn ] = 0.
Hence
E [(Mn+1 Mn ) (Mm+1 Mm )]
= E [E [(Mn+1 Mn ) (Mm+1 Mm ) | Fn ]] = 0.
Also, if we set M1 = 0,
2
Xn
Mn2 = M0 + (Mj Mj1 )
j=1
n
X X
= M02 + (Mj Mj1 )2 + (Mj Mj1 )(Mk Mk1 ).
j=1 j6=k

Taking expectations of both sides gives the second conclusion.

The natural place to discuss the role of orthogonality in the study

of square integrable martingales is L2 = L2 (, F, P), the space of
square integrable random variables. This is a (real) Hilbert space
under the inner product

(X, Y ) = E[XY ].
1.6. INTEGRALS WITH RESPECT TO RANDOM WALK 25

Two mean zero random variables are orthogonal if and only if

(X, Y ) = 0. The conditional expectation has a nice interpretation
in L2 . Suppose Y is a square integrable random variable and G is a
sub--algebra. Then L2 (, G, P) is a closed subspace of L2 (, F, P)
and the conditional expectation E[Y | G] is the same as the Hilbert
space projection onto the subspace. It can also be characterized
as the G-measurable random variable Z that minimizes the mean-
squared error
E[(Y Z)2 ].
The reason L2 rather than Lp for other values of p is so useful is
because of the inner product which gives the idea of orthogonality.

1.6 Integrals with respect to random walk

Suppose that X1 , X2 , . . . are independent, identically distributed random
variables with mean zero and variance 2 . The two main examples we will
use are:

Binomial or coin-tossing random variables,

1
P{Xj = 1} = P{Xj = 1} = ,
2
in which case 2 = 1.

Normal increments where Xj N (0, 2 ). We write Z N (, 2 ) if Z

has a normal distribution with mean and variance 2 .

Let Sn = X1 + + Xn and let {Fn } denote the filtration generated by

X1 , . . . , Xn . A sequence of random variables J1 , J2 , . . . is called predictable
(with respect to {Fn }) if for each n, Jn is Fn1 -measurable. Recall that this
is the condition that makes Jn allowable bets on the martingale in the
sense of the discrete stochastic integral.
Suppose J1 , J2 , . . . is a predictable sequence with E[Jn2 ] < for each n.
The integral of Jn with respect to Sn is defined by
n
X n
X
Zn = Jj Xj = Jj Sj .
j=1 j=1

There are three important properties that the integral satisfies.

26 CHAPTER 1. MARTINGALES IN DISCRETE TIME

Martingale property. The integral Zn is a martingale with respect

to {Fn }. We showed this in Section 1.2.

Linearity. If Jn , Kn are predictable sequences and a, b constants, then

aJn + bKn is a predictable sequence and
n
X n
X n
X
(aJj + bKj ) Sj = a Jj Sj + b Kj Sj .
j=1 j=1 j=1

This is immediate.

Variance rule

Xn Xn n
X
Jj Sj )2 = 2 E Jj2 .
 
Var Jj Sj = E (
j=1 j=1 j=1

To see this we first use the orthogonality of martingale increments to

write
X n X n
2
E Jj2 Xj2 .
 
E (
Jj Sj ) =
j=1 j=1

Since Jj is Fj1 -measurable and Xj is independent of Fj1 , we can

see that

E Jj2 Xj2 = E E[Jj2 Xj2 | Fj1 ]

   

= E Jj2 E[Xj2 | Fj1 ]

 

= E Jj2 E[Xj2 ] = 2 E[Jj2 ].

 

1.7 A maximal inequality

There is another result about martingales that we will use.
Theorem 1.7.1. Suppose Yn is a nonnegative submartingale
with respect to {Fn }, and

Y n = max{Y0 , Y1 , . . . , Yn }.

Then for every a > 0,

P{Y n a} a1 E[Yn ].
1.8. EXERCISES 27

Proof. Let T denote the smallest integer k such that Yk a.

Then
[n
{Y n a} = Ak , Ak = {T = k}.
k=0

Note that Ak is Fk -measurable. Since Yn is a submartingale,

E [Yn 1Ak ] = E [E(Yn | Fk ) 1Ak ] E [Yk 1Ak ] .

By summing over k, we see that

n
  X
E [Yn ] E Yn 1{Y n a} = E [Yn 1Ak ]
k=0

n
X  
E [Yk 1Ak ] = E YT 1{Y n a} a P{Y n a}.
k=0

Corollary 1.7.2. If Mn is a square integrable martingale with

respect to {Fn } and

M n = max {|M0 |, . . . , |Mn |} ,

then for every a > 0,

P M n a a2 E Mn2 .
  

Proof. In Exercise 1.15, it is shown that Mn2 is a submartingale,

and we can use the previous theorem.

1.8 Exercises
Exercise 1.1. Suppose we roll two dice, a red and a green one, and let X
be the value on the red die and Y the value on the green die. Let Z = XY .

1. Let W = E(Z | X). What are the possible values for W ? Give the
distribution of W .

2. Do the same exercise for U = E(X | Z).

3. Do the same exercise for V = E(Y | X, Z)

28 CHAPTER 1. MARTINGALES IN DISCRETE TIME

Exercise 1.2. Suppose we roll two dice, a red and a green one, and let X
be the value on the red die and Y the value on the green die. Let Z = X/Y .
1. Find E[(X + 2Y )2 | X].

2. Find E[(X + 2Y )2 | X, Z].

3. Find E[X + 2Y | Z].

4. Let W = E[Z | X]. What are the possible values for W ? Give the
distribution of W .
Exercise 1.3. Suppose X1 , X2 , . . . are independent random variables with
1
P{Xj = 2} = 1 P{Xj = 1} = .
3
Let Sn = X1 + + Xn and let Fn denote the information in X1 , . . . , Xn .
1. Find E[Sn ], E[Sn2 ], E[Sn3 ].

2. If m < n, find

E[Sn | Fm ], E[Sn2 | Fm ], E[Sn3 | Fm ].

3. If m < n, find E[Xm | Sn ].

Exercise 1.4. Repeat Exercise 1.3 assuming that
1
P{Xj = 3} = P{Xj = 1} = .
2
Exercise 1.5. Suppose X1 , X2 , . . . are independent random variables with
1
P{Xj = 1} = P{Xj = 1} = .
2
Let Sn = X1 + + Xn . Find

E sin Sn | Sn2 .

Exercise 1.6. In this exercise, we consider simple, nonsymmetric random

walk. Suppose 1/2 < q < 1 and X1 , X2 , . . . are independent random variables
with
P{Xj = 1} = 1 P{Xj = 1} = q.
Let S0 = 0 and Sn = X1 + +Xn . Let Fn denote the information contained
in X1 , . . . , Xn .
1.8. EXERCISES 29

1. Which of these is Sn : martingale, submartingale, supermartingale

(more than one answer is possible)?

2. For which values of r is Mn = Sn rn a martingale?

3. Let = (1 q)/q and let

Mn = S n .

Show that Mn is a martingale.

4. Let a, b be positive integers, and

Ta,b = min{j : Sj = b or Sj = a}.

Use the optional sampling theorem to determine


P STa,b = b .

5. Let Ta = Ta, . Find

P{Ta < }.

Exercise 1.7. Suppose two people want to play a game in which person A
has probability 2/3 of winning. However, the only thing that they have is a
fair coin which they can flip as many times as they want. They wish to find
a method that requires only a finite number of coin flips.

1. Give one method to use the coins to simulate an experiment with

probability 2/3 of success. The number of flips needed can be random,
but it must be finite with probability one.

2. Suppose K < . Explain why there is no method such that with

probability one we flip the coin at most K times.

Exercise 1.8. Repeat the last exercise with 2/3 replaced by 1/.

Exercise 1.9. Let X1 , X2 , . . . be independent, identically distributed ran-

dom variables with
1 1 2
P{Xj = 2} = , P{Xj = } = .
3 2 3
Let M0 = 1 and for n 1, Mn = X1 X2 Xn .

1. Show that Mn is a martingale.

30 CHAPTER 1. MARTINGALES IN DISCRETE TIME

2. Explain why Mn satisfies the conditions of the martingale convergence

theorem.
3. Let M = limn Mn . Explain why M = 0. (Hint: there are at least
two ways to show this. One is to consider log Mn and use the law of
large numbers. Another is to note that with probability one Mn+1 /Mn
does not converge.)
4. Use the optional sampling theorem to determine the probability that
Mn ever attains a value as large as 64.
5. Does there exist a C < such that E[Mn2 ] C for all n?
Exercise 1.10. Let X1 , X2 , . . . be independent, identically distributed ran-
dom variables with
P{Xj = 1} = q, P{Xj = 1} = 1 q.
Let S0 = 0 and for n 1, Sn = X1 + X2 + + Xn . Let Yn = eSn .
1. For which value of q is Yn a martingale?
2. For the remaining parts of this exercise assume q takes the value from
part 1. Explain why Yn satisfies the conditions of the martingale con-
vergence theorem.
3. Let Y = limn Yn . Explain why Y = 0. (Hint: there are at least two
ways to show this. One is to consider log Yn and use the law of large
numbers. Another is to note that with probability one Yn+1 /Yn does
not converge.)
4. Use the optional sampling theorem to determine the probability that
Yn ever attains a value greater than 100.
5. Does there exist a C < such that E[Yn2 ] C for all n?
Exercise 1.11. This exercise concerns Polyas urn and has a comput-
ing/simulation component. Let us start with one red and one green ball
as in the lecture and let Mn be the fraction of red balls at the nth stage.
1. Show that the distribution of Mn is uniform on the set
 
1 2 n+1
, ,..., .
n+2 n+2 n+2
(Use mathematical induction, that is, note that it is obviously true for
n = 0 and show that if it is true for n then it is true for n + 1.)
1.8. EXERCISES 31

2. Write a short program that will simulate this urn. Each time you
run the program note the fraction of red balls after 600 draws and
after 1200 draws. Compare the two fractions. Then, repeat this twenty
times.

Exercise 1.12. Consider the martingale betting strategy as discussed in

Section 1.2. Let Wn be the winnings at time n, which for positive n equals
either 1 or 1 2n .

1. Is Wn a square integrable martingale?

2. If n = Wn Wn1 what is E[2n ]?

3. What is E[Wn2 ]?

4. What is E(2n | Fn1 )?

Exercise 1.13. Suppose Sn = X1 + + Xn is simple random walk starting

at 0. For any K, let
T = min{n : |Sn | = K}.

Explain why for every j,

P{T j + K | T > j} 2K .

Show that there exists c < , > 0 such that for all j,

P{T > j} c ej .

Conclude that E[T r ] < for every r > 0.

Let Mn = Sn2 n. Show there exists C < such that for all n,
 2 
E MnT C.

Exercise 1.14. Suppose that X1 , X2 , . . . are independent random variables

with E[Xj ] = 0, Var[Xj ] = j2 , and suppose that

X
n2 < .
n=1

Let S0 = 0 and Sn = X1 + + Xn for n > 0. Let Fn denote the information

contained in X1 , . . . , Xn .
32 CHAPTER 1. MARTINGALES IN DISCRETE TIME

Show that Sn is a martingale with respect to {Fn }.

Show that there exists C < such that for all n, E[Sn2 ] C.

Show that with probability one the limit

S = lim Sn ,
n

exists.

Show that

X
E[S ] = 0, Var[S ] = n2 .
n=1

Exercise 1.15.
Suppose Y is a random variable and is a convex function,
that is, if 0 1,

(x + (1 ) y) (x) + (1 ) (y).

Suppose that E[|(Y )|] < . Show that E((Y ) | X)

(E(Y | X)). (Hint: you will need to review Jensens in-
equality.)
Show that if Mn is a martingale with respect to {Fn } and
r 1, then Yn = |Mn |r is a submartingale.
Chapter 2

Brownian motion

2.1 Limits of sums of independent variables

We will discuss two major results about sums of random variables that we
hope the reader has seen. They both discuss the limit distribution of

X1 + X2 + + Xn

where X1 , X2 , . . . , Xn are independent, identically distributed random vari-

ables.
Suppose X1 , X2 , . . . have mean and variance 2 < . Let

(X1 + + Xn ) n
Zn = .
n
We let denote the standard normal distribution function,
Z b
1 2
(b) = ex /2 dx.
2
While this function cannot be written down explicitly, the numerical values
are easily accessible in tables and computer software packages.

Theorem 2.1.1 (Central Limit Theorem). As n , the distribution of

Zn approaches a standard normal distribution. More precisely, if a < b, then

lim P{a Zn b} = (b) (a).

n

This amazing theorem states that no matter what distribution we choose

for the Xj , then as long as the distribution has a finite variance, the scaled

33
34 CHAPTER 2. BROWNIAN MOTION

random variables approach a normal distribution. This is what is referred to

in physics literature as a universality result and in the mathematics literature
as an invariance principle. The assumption that the random variables are
identically distributed can be relaxed somewhat. We will not go into details
here, but whenever a quantity can be written as a sum of independent (or
at least not too dependent) increments, all of which are small compared to
the sum, then the limiting distribution is normal. This is why it is often
reasonable to assume normal distributions in nature.

To see a complete proof of Theorem 2.1.1, consult any book on

measure-theoretic probability. We will sketch part of the proof which
shows how the normal distribution arises. Without loss of generality,
we can assume that = 0, 2 = 1 for otherwise we consider Yj =
(Xj )/. Let (t) = E[eitXj ] denote the characteristic function
of the increments. Then, can be expanded near the origin,

(i)2  2  2
(t) = 1 + i E [Xj ] t + E Xj t + o(t2 )
2
t2
= 1 + o(t2 ),
2
where o(t2 ) denotes a function such that |o(t2 )|/t2 0 as t 0.
Using the independence of the Xj , we see that the characteristic
function of Zn is
 2 n
n t2

t 2
Zn (t) = (t/ n) = 1 +o et /2 .
2n n
The right-hand side is the characteristic function of a standard nor-
mal random variable.

When one considers sums of independent random variables where a few

terms contribute the bulk of the sum, then one does not expect to get normal
limits. The prototypical example of a nonnormal limit is the Poisson distri-
(n) (n) (n)
bution. Suppose > 0 and X1 , X2 , . . . , Xn are independent random
variables each with
n o n o
(n) (n)
P Xj = 1 = 1 P Xj = 0 = .
n
Let
(n)
Yn = X1 + + Xn(n) ,
2.1. LIMITS OF SUMS OF INDEPENDENT VARIABLES 35

and note that for each n, E[Yn ] = . Recall that a random variable Y has a
Poisson distribution with mean if for each nonnegative integer k,

k
P{Y = k} = e .
k!
Theorem 2.1.2 (Convergence to the Poisson distribution). As n ,
the distribution of Yn approaches a Poisson distribution with mean . More
precisely, for every nonnegative integer k,

k
lim P {Yn = k} = e .
n k!
Proof. For each n, Yn has a binomial distribution with parameters n and
/n, and hence

lim P {Yn = k}
n
   k 
nk

n
= lim 1
n k n n
k  n
k

n(n 1) (n k + 1)
= lim 1 1
k! n nk n n

Since we are fixing k and letting n , one can see that

k
 
n(n 1) (n k + 1)
lim 1 = 1,
n nk n

and in calculus one learns the limit

n
 
lim 1 = e .
n n

We will be taking limits of processes as the time increment goes to zero.

The kind of limit we expect will depend on the assumptions. When we take
limits of random walks with finite variance, then we are in the regime of the
central limit theorem, and we will expect normal distributions. Also, because
all of the increments are small with respect to the sum, the limit process will
have paths that are continuous. Here the limit is Brownian motion which
we discuss in this section.
36 CHAPTER 2. BROWNIAN MOTION

In the Poisson case, the limit distribution will not have continuous paths
but rather will be a jump process. The prototypical case is the Poisson pro-
cess with intensity . In this case, Nt denotes the number of occurrences of
an event by time t. The function t 7 Nt takes on nonnegative integer values
and the jumps are always of size +1. It satisfies the following conditions.

For each s < t, the random variable Nt Ns has a Poisson distribution

with parameter (t s).

For each s < t, the random variable Nt Ns is independent of the

random variables {Nr : 0 r s}.

We discuss this further in Section 6.2.

2.2 Multivariate normal distribution

Although the normal or Gaussian distribution is a little inconvenient in the
sense that the distribution function cannot be computed exactly, there are
many other aspects that make the distribution very convenient. In partic-
ular, when dealing with many variables, assuming a joint or multivariate
normal distribution makes computations tractable. In this section we will
give the basic definitions. Roughly speaking, the basic assumption is that
if (X1 , . . . , Xn , Y ) have a joint normal distribution then not only does each
variable have a normal distribution but also, the conditional distribution of
Y given X1 , . . . , Xn is normal with mean E[Y |X1 , . . . , Xn ] and a variance
that depends on the joint distribution but not on the observed data points.
There are a number of equivalent ways to define a joint normal distribution.
We will use the following.

Definition A finite sequence of random variables (X1 , . . . , Xn ) has a joint

(or multivariate) normal (or Gaussian) distribution if they are linear com-
binations of independent standard normal random variables. In other words,
if there exist independent random variables (Z1 , . . . , Zm ), each N (0, 1), and
constants mj , ajk such that for j = 1, . . . , n,

Xj = mj + aj1 Z1 + aj2 Z2 + + ajm Zm .

Clearly E[Xj ] = mj . Let us consider the case of mean-zero (also called

centered) joint normals, in which case the equation above can be written in
matrix form
X = A Z,
2.2. MULTIVARIATE NORMAL DISTRIBUTION 37

where
X1 Z1
X2 Z2
X= , Z= ,

.. ..
. .
Xn Zm
and A is the n m matrix with entries ajk . Each Xj is a normal random
variable with mean zero and variance
E[Xj2 ] = a2j1 + + a2jm .
More generally, the covariance of Xj and Xk is given by
m
X
Cov(Xj , Xk ) = E[Xj Xk ] = ajl akl .
l=1

Let = AAT be the n n matrix whose entries are

jk = E[Xj Xk ].
Then is called the covariance matrix.
We list some important properties. Assume (X1 , . . . , Xn ) has a joint
normal distribution with mean zero and covariance matrix .
Each Xj has a normal distribution. In fact, if b1 , . . . , bn are constants,
then
b1 X1 + + bn Xn ,
has a normal distribution. We can see this easily since we can write
the sum above as a linear combination of the independent normals
Z1 , . . . , Z m .
The matrix is symmetric, jk = kj . Moreover, it is positive semi-
definite which means that if b = (b1 , . . . , bn ) is a vector in Rn , then
n X
X n
b b = jk bj bk 0. (2.1)
j=1 k=1

(If the is replaced with > 0 for all b = (b1 , . . . , bn ) 6= (0, . . . , 0),
then the matrix is called positive definite.) The inequality (2.1) can be
derived by noting that the left-hand side is the same as
E (b1 X1 + + bn Xn )2 ,
 

which is clearly nonnegative.

38 CHAPTER 2. BROWNIAN MOTION

If is a positive semidefinite, symmetric matrix, then it is the co-

variance matrix for a joint normal distribution. The proof of this fact,
which we omit, uses linear algebra to deduce that there exists an n n
matrix A with AAT = . (The A is not unique.)

The distribution of a mean-zero joint normal is determined by the

covariance matrix .

In order to show that the covariance matrix determines the

distribution of a mean-zero joint normal, we compute the character-
istic function. Suppose that = AAT where A is n n. Using the
independence of Z1 , . . . , Zn and the characteristic function of the
2
standard normal, E[eitZk ] = et /2 , we see that the characteristic
function of (X1 , . . . , Xn ) is

(1 , . . . , n ) = E [exp {i(1 X1 + . . . + n Xn )}]

X n X n
= E exp i j ajk Zk

j=1 k=1

X n X n
= E exp i Zk ( j ajk )

k=1 j=1

Yn X n
= E exp iZk ( j ajk )

k=1 j=1
2
1X n X n
= exp j ajk
2
k=1 j=1

1X n X n X n
= exp j i ajk alk
2
k=1 j=1 l=1
 
1 T T
= exp AA
2
 
1 T
= exp
2

where we write = (1 , . . . , n ). Even though we used A, which is

not unique, in our computation, the answer only involves . Since
2.2. MULTIVARIATE NORMAL DISTRIBUTION 39

the characteristic function determines the distribution, the distribu-

tion depends only on the covariance matrix.

If is invertible, then (X1 , . . . , Xn ) has a density. We write it in the

case that the random variables have mean m = (m1 , . . . , mn ),

f (x1 , . . . , xn ) = f (x) =

(x m) 1 (x m)T
 
1
exp .
(2)n/2 det 2
Sometimes this density is used as a definition of a joint normal. The
formula for the density looks messy, but note that if n = 1, m =
m, = [ 2 ], then the right-hand side is the density of a N (m, 2 )
random variable.

If (X1 , X2 ) have a mean-zero joint normal density, and E(X1 X2 ) = 0,

then X1 , X2 are independent random variables. To see this let j2 =
E[Xj2 ]. Then the covariance matrix of (X1 , X2 ) is the diagonal matrix
with diagonal entries j2 . If (Z1 , Z2 ) are independent N (0, 1) random
variables and Y1 = 1 Z1 , Y2 = 2 Z2 , then by our definition (Y1 , Y2 )
are joint normal with the same covariance matrix. Since the covariance
matrix determines the distribution, X1 , X2 must be independent,

It is a special property about joint normal random variables that orthog-

onality implies independence. In our construction of Brownian motion, we
will use a particular case, that we state as a lemma.

Proposition 2.2.1. Suppose X, Y are independent N (0, 1) random vari-

ables and
X +Y X Y
Z= , W = .
2 2
Then Z and W are independent N (0, 1) random variables.

Proof. By definition (Z, W ) has a joint normal distribution and Z, W clearly

have mean 0. Using E[X 2 ] = E[Y 2 ] = 1 and E[XY ] = 0, we get

E[Z 2 ] = 1, E[W 2 ] = 1, E[ZW ] = 0.

Hence the covariance matrix for (Z, W ) is the identity matrix and this is
the covariance matrix for independent N (0, 1) random variables.
40 CHAPTER 2. BROWNIAN MOTION

2.3 Limits of random walks

Brownian motion can be viewed as the limit of random walk as the time and
space increments tend to zero. It is necessary to be careful about how the
limit is taken. Suppose X1 , X2 , . . . are independent random variables with
P{Xj = 1} = P{Xj = 1} = 1/2 and let

Sn = X1 + + Xn

be the corresponding random walk. Here we have chosen time increment

t = 1 and space increment x = 1. Suppose we choose t = 1/N where
N is a large integer. Hence, we view the process at times

t, 2t, 3t, ,

and at each time we have a jump of x. At time 1 = N t, the value of

the process is
(N )
W1 = x (X1 + + XN ) .
(N )
We would like to choose x so that Var[W1 ] = 1, and since

Var [x (X1 + + XN )] = (x)2 [Var(X1 ) + + Var(XN )]

= (x)2 N,

we see that we need to choose

p
x = 1/N = t.

We also know from the central limit theorem, that if N is large, then the
distribution of
X1 + + XN
,
N
is approximately N (0, 1).
Brownian motion could be defined formally as the limit of random walks,
but there are subtleties in describing the kind of limit. In the next section, we
define it directly using the idea of continuous random motion. However,
the random walk intuition is useful to retain.

2.4 Brownian motion

Brownian motion or the Wiener process is a model of random continuous
motion. We will start by making the assumptions that underlie the phrase
2.4. BROWNIAN MOTION 41

random continuous motion. Let Bt = B(t) be the value at time t. For each
t, Bt is a random variable.1 A collection of random variables indexed by time
is called a stochastic process. We can view the process in two different ways:
For each t, there is a random variable Bt , and there are correlations
between the values at different times.
The function t 7 B(t) is a random function. In other words, it is a
random variable whose value is a function.
There are three major assumptions about the random variables Bt .
Stationary increments. If s < t, then the distribution of Bt Bs is
the same as that of Bts B0 .
Independent increments. If s < t, the random variable Bt Bs is
independent of the values Br for r s.
Continuous paths. The function t 7 Bt is a continuous function of
t.
We often assume B0 = 0 for convenience, but we can also take other initial
conditions. All of the assumptions above are very reasonable for a model of
random continuous motion. However, it is not obvious that these are enough
assumptions to characterize our process uniquely. It turns out that they do
up to two parameters. One can prove (see Theorem 6.8.3), that if Bt is a
process satisfying the three conditions above, then the distribution of Bt for
each t must be normal. Suppose Bt is such a process, and let m, 2 be the
mean and variance of B1 . If s < t, then independent, identically distributed
increments imply that
E[Bt ] = E[Bs ] + E[Bt Bs ] = E[Bs ] + E[Bts ],
Var[Bt ] = Var[Bs ] + Var[Bt Bs ] = Var[Bs ] + Var[Bts ].
Using this relation, we can see that E[Bt ] = tm, Var[Bt ] = t 2 . At this point,
we have only shown that if a process exists, then the increments must have
a normal distribution. We will show that such a process exists. It will be
convenient to put the normal distribution in the definition.
1
In this book and usually in the financial world, the terms Brownian motion and
Wiener process are synonymous. However, in the scientific world, the word Brownian
motion is often used for a physical process for which what we will describe is one possible
mathematical model. The term Wiener process always refers to the model we define. The
letter Wt is another standard notation for Brownian motion/Wiener process and is more
commonly used in financial literature. We will use both Bt and Wt later in the book when
we need two notations.
42 CHAPTER 2. BROWNIAN MOTION

Definition A stochastic process Bt is called a (one-dimensional) Brownian

motion with drift m and variance (parameter) 2 starting at the origin if it
satisfies the following.

B0 = 0.

For s < t, the distribution of Bt Bs is normal with mean m(t s)

and variance 2 (t s).

If s < t, the random variable Bt Bs is independent of the values Br

for r s.

With probability one, the function t 7 Bt is a continuous function of

t.

If m = 0, 2 = 1, then Bt is called a standard Brownian motion.

Recall that if Z has a N (0, 1) distribution and Y = Z + m, then Y has

a N (m, 2 ) distribution. Given that it is easy to show the following.

If Bt is a standard Brownian motion and

Yt = Bt + mt,

then Yt is a Brownian motion with drift m and variance 2 .

Indeed, one just checks that it satisfies the conditions above. Hence, in order
to establish the existence of Brownian motion, it suffices to construct a
standard Brownian motion.
There is a mathematical challenge in studying stochastic processes in-
dexed by continuous time. The problem is that the set of positive real num-
bers is uncountable, that is, the elements cannot be enumerated t1 , t2 , . . ..
The major axiom of probability theory is the fact that if A1 , A2 , . . . is a
countable sequence of disjoint events, then
" #
[ X
P An = P[An ].
n=1 n=1

This rule does not hold for uncountable unions. An example that we have
all had to deal with arises with continuous random variables. Suppose, for
instance, that Z has a N (0, 1) distribution. Then for each x R,

P{Z = x} = 0.
2.5. CONSTRUCTION OF BROWNIAN MOTION 43

However, " #
[
1 = P{Z R} = P Ax ,
xR

where Ax denotes the event {Z = x}. The events Ax are disjoint, each with
probability zero, but it is not the case that
" #
[ X
P Ax = P(Ax ) = 0.
xR xR

In constructing Brownian motion, we use the fact that if g : [0, ) R

is a continuous function and we know the value of g on a countable, dense
set, such as the dyadic rationals
 
k
: k = 0, 1, . . . ; n = 0, 1, . . . ,
2n

then we know the value at every t. Indeed, we need only find a sequence of
dyadic rationals tn that converge to t, and let

g(t) = lim g(tn ).

tn t

This is fine if a priori we know the function g is continuous. Our strategy

for constructing Brownian motion will be:

First define the process Bt when t is a dyadic rational.

Prove that with probability one, the function t 7 Bt is continuous on

the dyadics (this is the hardest step, and we need some care in the
definition of continuity).

Extend Bt to other t by continuity.

The next section shows that one can construct a Brownian motion. The
reader can skip this section and just have faith that such a process exists.

2.5 Construction of Brownian motion

We will show how to construct Brownian motion. For ease, we will
consider Bt , 0 t 1. Once we know how to construct this, we
can take a countable collection of such processes and connect them
44 CHAPTER 2. BROWNIAN MOTION

appropriately to get Bt , 0 t < . We start with a probability

space (, F, P) which is large enough so that it contains a countable
collection of independent standard normal random variables. Let
 
k n
Dn = : k = 0, 1, . . . , 2 ,
2n
denote the dyadic rationals in [0, 1] with denominator 2n and let
D = n D. Since D is a countable set, we can assume that our
collection of N (0, 1) random variables

{Zq : q D}

is indexed by D.
We will use the independent random variables {Zq } to define the
Brownian motion Bq , q D. We start by defining B0 , B1 , and then
B1/2 , and then B1/4 and B3/4 , and so forth, by always subdividing
our intervals. We start with B0 = 0 and we let B1 = Z1 which is
clearly N (0, 1). We then define
B1 Z1/2
B1/2 = + ,
2 2
and hence
B1 Z1/2
B1 B1/2 = .
2 2
We think of the definition of B1/2 as being E[B1/2 | B1 ] plus
some independent randomness. Using Proposition 2.2.1, we can see
that B1/2 and B1 B1/2 are independent random variables, each
N (0, 1/2). We continue this splitting. If q = (2k + 1)/2n+1
Dn+1 \ Dn , we define
B(k+1)2n Bk2n Zq
Bq = Bk2n + + (n+2)/2 .
2 2
This formula looks a little complicated, but we can again view this
as

Bq = E[Bq | Bk2n , B(k+1)2n ] + (independent randomness),

where the variance of the new randomness is chosen appropriately.

By repeated use of Proposition 2.2.1, we can see that for each n,
the random variables

Bk2n B(k1)2n : k = 1, . . . , 2n

2.5. CONSTRUCTION OF BROWNIAN MOTION 45

are independent, each with N (0, 2n ) distribution. From this we

can see that {Bq : q D} satisfies the first three properties in the
definition of a Brownian motion.
We will define Bt for other t by continuity. To do this we prove
a theorem that states that the Brownian motion restricted to D
has uniformly continuous paths. In other words, we show that with
probability one, for every  > 0, there exists > 0 such that if
|s t| and s, t D, then |Bs Bt | < . Equivalently, if we
define the random variable

Kn = sup |Bs Bt | : s, t D, |s t| 2n ,


then with probability one, Kn 0 as n . Note that Kn is the

supremum of a countable number of random variables and hence is
a well-defined random variable (perhaps taking the value ).
Proposition 2.5.1. With probability one, for all < 1/2,

lim 2n Kn = 0. (2.2)
n

In particular, Kn 0.
In order to prove this proposition, it is easier to consider another
sequence of random variables

Jn = max n Y (j, n)
j=1,...,2

where Y (j, n) equals

sup |Bq B(j1)2n | : q D, (j 1)2n q j2n .



A simple argument using the triangle inequality shows that Kn

3Jn . It turns out Jn is easier to analyze. For any  > 0,
2 n
X
P{Jn } P{Y (j, n) } = 2n P{Y (1, n) }.
j=1

Also the distribution of

Y (1, n) = sup |Bq | : q D, q 2n .



is the same as that of 2n/2 Y where

Y = Y (1, 0) = sup {|Bq | : q D} .

46 CHAPTER 2. BROWNIAN MOTION

Using this we see that

P{Jn C n 2n/2 } 2n P{Y C n}.

We will show below that if C > 2 log 2, then

X
2n P{Y C n} < . (2.3)
n=1

The Borel-Cantelli lemma then implies that with probability one,

the event {Jn 2n/2 n} happens for only finitely many values of
n. In particular,
lim 2n/2 n1 Jn = 0,
n

which implies (2.2). To get our estimate, we will need the following
lemma which is a form of the reflection principle for Brownian
motion.

Lemma 2.5.2. For every a 4,

2 /2
P{Y > a} 4 P{B1 a} ea .

To prove this lemma, let

Yn = max {|Bq | : q Dn } .

Then
P{Y > a} = lim P{Yn > a},
n

and hence it suffices to prove the inequality for each n. Fix n and
let Ak be the event that

|Bk2n | > a, |Bj2n | a, j = 1, . . . , k 1.

Then the event {Yn > a} can be written as a disjoint union

2 n
[
{Yn > a} = Ak .
k=1

The important observation is that if |Bk2n | > a, then with prob-

ability at least 1/2, we have |B1 | > a. Indeed, if Bk2n and
2.5. CONSTRUCTION OF BROWNIAN MOTION 47

B1 Bk2n have the same sign, which happens with probability

1/2, then |B1 | |Bk2n |. Therefore,
1
P [Ak {|B1 | > a}] P(Ak ),
2
and hence
2 n
X
P{|B1 | > a} = P [Ak {|B1 | > a}]
k=1
2n
1 X
P [Ak ]
2
k=1
1
= P{Yn > a}.
2
Here we use the fact that the events Ak are disjoint. Therefore,
since B1 has a N (0, 1) distribution,
Z
4 2
P{Yn > a} 2 P{|B1 | > a} = 4 P{B1 > a} = ex /2 dx.
2 a
To estimate the integral for a 4, we write
Z Z
4 2
ex /2 dx 2 eax/2 dx
2 a a
4 a2 /2 2
= e ea /2 .
a

This proves Lemma 2.5.2. If we apply the estimate with a = C n,
then we see that for large n,
2 n/2 C2
P{Y > C n} eC = 2n , = .
2 log 2
In particular, if > 1, then (2.3) holds and this gives (2.2) with
probability one.
The hard work is finished, with the proposition, we can now
define Bt for t [0, 1] by
Bt = lim Btn
tn

where tn t, tn D. One needs to check that this satisfies the

conditions for a Brownian motion. This is relatively straightforward
and we omit it.
48 CHAPTER 2. BROWNIAN MOTION

2.6 Understanding Brownian motion

Let Bt be a standard Brownian motion starting at the origin. By doing sim-
ulations, one can see that although the paths of the process are continuous
they are very rough. To do simulations, we must discretize time. If we let
t be chosen, then we will sample the values

B0 , Bt , B2t , B3t , . . .

The increment B(k+1)t Bkt is a normal random variable with mean 0 and
variance t. If N0 , N1 , N2 , . . . denote independent N (0, 1) random variables
(which can be generated on a computer), we set

B(k+1)t = Bkt + t Nk ,

which we can write as

Bkt = B(k+1)t Bkt = t Nk .

The relationship |Bt | = |Bt+t Bt
| t is fundamental for under-
standing Brownian motion. As t 0, t 0, which is consistent with
the continuity of the paths. However, we claim that the paths do not have
derivatives. To see why this should be true, recall that the definition of the
derivative at time t would be
Bt+t Bt
lim .
t0 t

As t 0, the absolute value of the numerator is of order t which is
much larger than t for small t. Hence, we do not expect the limit to
exist. In fact, the following stronger statement is true.

Theorem 2.6.1. With probability one, the function t 7 Bt is nowhere

differentiable.

Note carefully the statement of this theorem. It is a stronger statement

than the similar statement: for every t, with probability one, the derivative
does not exists at t. As a comparison, consider the following two statements:

With probability one, for every t, Bt 6= 1.

For every t, with probability one, Bt 6= 1.

2.6. UNDERSTANDING BROWNIAN MOTION 49

Since Bt has a N (0, t) distribution, the second of these statements is ob-

viously true. However, the first statement is false. To see this, note that
P{B1 > 1} is greater than 0, and if B0 = 0, B1 > 1, then continuity im-
plies that there exists t (0, 1) such that Bt = 1. Here again we have the
challenge of dealing with uncountably many events of probability 0. Even
though for each t, P{Bt = 1} = 0,

[
P {Bt = 1} > 0.
t[0,1]

While nondifferentiable paths may seem surprising, a little thought about

our assumptions will show why we could not expect to have differentiable
paths. Suppose that Bt were differentiable at t. Then, we could determine
the derivative by observing Bs , 0 s t. This would then tell us something
about the increment Bt+t Bt for t > 0. However, our assumptions tell
us that Bt+t Bt is independent of Bs , 0 s t.
We will makea more precise statement about the relationship |Bt | =
|Bt+t Bt | t. If > 0, then a function f : [0, 1] R is Holder
continuous of order if there exists C < such that for all 0 s, t 1,

|f (s) f (t)| C |s t| .

The smaller is the easier it is to be Holder continuous of order . Note

that differentiable functions are Holder continuous of order 1 since

|f (s) f (t)| |f 0 (t)| |s t|.

Theorem 2.6.2. With probability one, for all < 1/2, Bt is Holder con-
tinuous of order but it is not Holder continuous of order 1/2.

We will be using Brownian motion and functions of Brownian motions

to model prices of assets. In all of the Brownian models, the functions will
have Holder exponent 1/2.

We will prove Theorem 2.6.1 for 0 t 1. We start by making

an observation about differentiable functions. Suppose Bt were dif-
ferentiable at some 0 t 1 with derivative r. Then there would
exist such that for |t s| ,

|Bt Bs | 2|r(t s)|.

50 CHAPTER 2. BROWNIAN MOTION

Hence, we could find a positive integer M < such that for all
sufficiently large integers n, there exists k n such that Yk,n
M/n, where Yk,n is
    
 k+1 k 
max B
 B ,
n n 
         
B k + 2 B k + 1  , B k + 3 B k + 2  .
   
 n n   n n 

Let Yn = min{Yk,n : k = 0, 1, . . . , n 1} and let AM be the

event that for all n sufficiently large, Yn M/n. For each positive
integer M ,

P{Yk,n M/n} = [P{|B(1/n)| M/n}]3

h i3
= P{n1/2 |B1 | M/n}
"Z #3
1 y2 /2
=
e dy
|x|M/ n 2
2M 1 3 M3
 
3/2 ,
n 2 n

and hence,

n1
X M3
P{Yn M/n} P{Yk,n M/n} 0.
k=0
n1/2

This shows that P(AM ) = 0 for each M , and hence

" #
[
P AM = 0.
M =1

But our first remark shows that the event that Bt is differentiable
at some point is contained in M AM .
Theorem 2.6.2 is a restatement of (2.2).
2.6. UNDERSTANDING BROWNIAN MOTION 51

2.6.1 Brownian motion as a continuous martingale

The definition of a martingale in continuous time is essentially the same
as in discrete time. Suppose we have an increasing filtration {Ft } of in-
formation and integrable random variables Mt such that for each t, Mt is
Ft -measurable. (We say that Mt is adapted to the filtration if Mt is Ft -
measurable for each t.) Then, Mt is a martingale with respect to {Ft } if for
each s < t,
E [Mt | Fs ] = Ms .

When one writes an equality as above there is an implicit up to an event of

probability zero. In discrete time this presents no problem because there are
only a countable number of pairs of times (s, t) and hence there can be only
a countable number of sets of measure zero. For continuous time, there are
instances where some care is needed but we will not worry about this at the
moment. As in the discrete case, if the filtration is not mentioned explicitly
then one assumes that Ft is the information contained in {Ms : s t}. In
that case, if Bt is a standard Brownian motion and s < t,

E[Bt | Fs ] = E[Bs | Fs ] + E[Bt Bs | Fs ] = Bs + E[Bt Bs ] = Bs . (2.4)

Often we will have more information at time t than the values of the
Brownian motion so it is useful to extend our definition of Brownian motion.
We say that Bt is Brownian motion with respect to the filtration {Ft } if each
Bt is Ft -measurable and Bt satisfies the conditions to be a Brownian motion
with the third condition being replaced by

If s < t, the random variable Bt Bs is independent of Fs .

In other words, although we may have more information at time s than

the value of the Brownian motion, there is nothing useful for predicting the
future increments. Under these conditions, (2.4) holds and Bt is a martingale
with respect to {Ft }.
If Ms , 0 s t is a martingale, then by definition, for each s t,

E(Y | Fs ) = Ms

where Y = Mt . Conversely, if Y is an integrable random variable that is

measurable with respect to Ft , we can define a martingale Ms , 0 s t by

Ms = E(Y | Fs ).
52 CHAPTER 2. BROWNIAN MOTION

Indeed, if we define Ms as above and r < s, then the tower property for
conditional expectation implies that

E(Ms | Fr ) = E(E(Y | Fs ) | Fr ) = E(Y | Fr ) = Mr .

A martingale Mt is called a continuous martingale if with probability

one the function t 7 Mt is a continuous function. The word continuous in
continuous martingale refers not just to the fact that time is continuous but
also to the fact that the paths are continuous functions of t. One can have
martingales in continuous time that are not continuous martingales. One
example is to let Nt be a Poisson process with rate as in Section 2.1 and

Mt = Nt t.

Then using the fact that the increments are independent we see that for
s < t,

E[Mt | Fs ] = E[Ms | Fs ] + E[Nt Ns | Fs ] (t s)

= Ms + E [Nt Ns ] (t s) = Ms .

Brownian motion with zero drift (m = 0) is a continuous martingale with

one parameter . As we will see, the only continuous martingales are essen-
tially Brownian motion where one allows the to vary with time. The factor
will be the analogue of the bet from the discrete stochastic integral.

A continuous time filtration on a probability space (, F, P) is

a collection of sub--algebras {Ft } of F such that if s < t, then
Fs Ft . It has been found that it is useful to make some additional
technical assumptions. First, we assume that F is complete under
P, that is, it contains all the nulls sets. A set A0 is a null
set (for P) if A0 A for some A F with P(A) = 0. This is
convenient for then one can prove that an event has probability zero
by showing that it is contained in an event of probability zero. If F
is not complete, one can complete it by considering the collection
of subsets A A0 where A F and A0 is a null set. In the case of
Lebesgue measure on the Borel subsets of R, the completion is the
set of Lebesgue measurable sets.
When using filtrations {Ft } one often makes further assump-
tions:
2.6. UNDERSTANDING BROWNIAN MOTION 53

Right continuity. For each t

\
Ft = Fs .
s>t

It is easy to check that the right-hand side is a -algebra

containing Ft . and it is often denoted by Ft+ . Right continuity
states that Ft = Ft+ . If the original filtration was not right
continuous, we can replace it with the filtration {Ft+ }.
(Strong) completeness. We assume that Ft contains all the
null sets of F (note that this is stronger than saying that Ft
is complete). If this does not hold initially, we enlarge our -
algebra to include all events of the form A A0 where A Ft
and A0 is a null set.
If we start with Brownian motion and let Ft be the -algebra
generated by {Bs : s t}, then it is not true that Ft = Ft+ .
However, it is almost true in the sense that every set in Ft+ can
be written as A A0 where A Ft and A0 is a null set. This is
a consequence of the Blumenthal 0 1 law that states that the
-algebra F0+ contains only events of probability zero or one. We
will not prove this here but it is related to the Kolmogorov 0 1
law.
At this point one may not appreciate why one wants to make
these assumptions, but we will not try to motivate them further.

2.6.2 Brownian motion as a Markov process

A continuous time process Xt is called Markov if for every t, the conditional
distribution of {Xs : s t} given {Xr : r t} is the same as the con-
ditional distribution given Xt . In other words, the future of the process is
conditionally independent of the past given the present value.
Brownian motion is a Markov process. Indeed, if Bt is a Brownian motion
with parameters (m, 2 ), and

Ys = Bt+s , 0 s < ,

then the conditional distribution of {Ys } given Ft is that of a Brownian

motion with initial condition Y0 = Bt . Indeed, if

Bs = Bt+s Bt ,
54 CHAPTER 2. BROWNIAN MOTION

then Bs is a Brownian motion that is independent of Ft . There is a stronger

notion of this called the strong Markov property that we will discuss in
Section 2.7.

2.6.3 Brownian motion as a Gaussian process

A process {Xt } is called a Gaussian process if each finite subcollection

(Xt1 , . . . , Xtn )

has a joint normal distribution. Recall that to describe a joint normal dis-
tribution one needs only give the means and the covariances. Hence the
finite-dimensional distributions of a Gaussian process are determined by the
numbers
mt = E[Xt ], st = Cov(Xs , Xt ).
If Bt is a standard Brownian motion and t1 < t2 < < tn , then we
can write Bt1 , . . . , Btn as linear combinations of the independent standard
normal random variables
Bt Btj1
Zj = j , j = 1, . . . , n.
tj tj1
Hence Bt is a Gaussian process with mean zero. If s < t,

E[Bs Bt ] = E [Bs (Bs + Bt Bs )]

= E[Bs2 ] + E[Bs (Bt Bs )]
= s + E[Bs ] E[Bt Bs ] = s,

which gives the general rule

Cov(Bs , Bt ) = min{s, t}.

The description of Brownian motion as a Gaussian process describes only

the finite-dimensional distributions but our definition includes some aspects
that depend on more than finite-dimensional distributions. In particular,
one cannot tell from the finite-dimensional distributions alone whether or
not the paths are continuous.

2.6.4 Brownian motion as a self-similar process

If one looks at a small piece of a Brownian motion and blows it up, then the
blown-up picture looks like a Brownian motion provided that the dilation
uses the appropriate scaling. We leave the derivation as Exercise 2.12.
2.7. COMPUTATIONS FOR BROWNIAN MOTION 55

Theorem 2.6.3. Suppose Bt is a standard Brownian motion and a > 0.

Let
Bat
Yt = .
a
Then Yt is a standard Brownian motion.
The key here is that if time is scaled by a factor of a, then space must

be scaled by a factor of 1/ a. The proof of this theorem is not difficult; one
needs only show that Yt satisfies the conditions to be a standard Brownian
motion. One can check that the scaling is right by computing

Var[Yt ] = Var[Bat / a] = a1 Var[Bat ] = a1 (at) = t.

2.7 Computations for Brownian motion

We will discuss some methods for computing probabilities for Brownian
motions. For ease, we will assume that Bt is a standard Brownian motion
starting at the origin with respect to a filtration {Ft }. If we are interested
in probabilities about the Brownian motion at one time t, we need only use
the normal distribution. Often, it is easier to scale to the standard normal.
For example,
Z
h i t1/2 2
E [|Bt |] = E t1/2 |B1 | = |x| ex /2 dx
2
r Z
2t 2
= x ex /2 dx
0
r
2t
= ,

and

P{Bt r} = P{ t B1 r} = P{B1 r/ t}

= 1 (r/ t)
Z
1 x2 /2
= e dx,
r/ t 2
where denotes the distribution function for a standard normal.
If we are considering probabilities for a finite number of times, we can use
the joint normal distribution. Often it is easier to use the Markov property
as we now illustrate. We will compute
P{B1 > 0, B2 > 0}.
56 CHAPTER 2. BROWNIAN MOTION

The events {B1 > 0} and {B2 > 0} are not independent; we would expect
them to be positively correlated. We compute by considering the possibilities
at time 1,
Z
P{B1 > 0, B2 > 0} = P{B2 > 0 | B1 = x} dP{B1 = x}
Z0
1 2
= P{B2 B1 > x} ex /2 dx
2
Z0 Z
1 (x2 +y2 )/2
= e dy dx
0 x 2
Z Z /2 r2 /2
e 3
= d r dr = .
0 /4 2 8

One needs to review polar coordinates to do the fourth equality. Note that

P{B1 > 0, B2 > 0} 3

P{B2 > 0 | B1 > 0} = = ,
P{B1 > 0} 4

which confirms our intuition that the events are positively correlated.
For more complicated calculations, we need to use the strong Markov
property. We say that a random variable T taking values in [0, ] is a stop-
ping time (with respect to the filtration {Ft }) if for each t, the event P{T t}
is Ft -measurable. In other words, the decision to stop can use the informa-
tion up to time t but cannot use information about the future values of the
Brownian motion.

If x R and
T = min{t : Bt = x},
then T is a stopping time.

Constants are stopping times.

If S, T are stopping times then

S T = min{S, T }

and
S T = max{S, T }

are stopping times.

2.7. COMPUTATIONS FOR BROWNIAN MOTION 57

Theorem 2.7.1 (Strong Markov Property). If T is a stopping time with

P{T < } = 1 and
Yt = BT +t BT ,
then Yt is a standard Brownian motion. Moreover, Y is independent of

{Bt : 0 t T }.

Let us apply this theorem, to prove a very useful tool for computing
probabilities.
Proposition 2.7.2 (Reflection Principle). If Bt is a standard Brownian
motion with B0 = 0, then for every a > 0,
i
  h
P max Bs a = 2P{Bt > a} = 2 1 (a/ t) .
0st

To derive the reflection principle, let

Ta = min {s : Bs a} = min {s : Bs = a} .

The second equality holds because Bs is a continuous function of s. Then

 
P max Bs a = P{Ta t} = P{Ta < t}.
0st

The second equality uses the fact that P{Ta = t} P{Bt = a} = 0. Since
BTa = a,

P{Bt > a} = P{Ta < t, Bt > a}

= P{Ta < t} P{Bt BTa > 0 | Ta < t}.

We now appeal to the Strong Markov Property to say that

P{Bt BTa > 0 | Ta < t} = 1/2.

This gives the first equality of the proposition and the second follows from

P{Bt > a} = P{B1 > a/ t} = 1 (a/ t).

Example 2.7.1. Let a > 0 and let Ta = inf{t : Bt = a}. The random
variable Ta is called a passage time. We will find the density of Ta . To do
this, we first find its distribution function
i
  h
F (t) = P{Ta t} = P max Bs a = 2 1 (a/ t) .
0st
58 CHAPTER 2. BROWNIAN MOTION

The density is obtained by differentiating

 
0 0 a a  a a2
f (t) = F (t) = 2 3/2 = e 2t , 0 < t < .
t 2t t3/2 2
Example 2.7.2. We will compute
q(r, t) = P {Bs = 0 for some r s t} .
The scaling rule for Brownian motion (Theorem 2.6.3) shows that q(r, t) =
q(1, t/r), so it suffices to compute q(t) = q(1, 1 + t). Let A be the event that
Bs = 0 for some 1 s 1 + t. The Markov property for Brownian motion
and symmetry imply
Z
q(t) = P[A | B1 = r] dP{Bt = r}

Z  
1 2
= P[A | B1 = r] er /2 dr
2
r Z
2 2
= P[A | B1 = r] er /2 dr.
0
The reflection principle and symmetry imply that
 
P[A | B1 = r] = P min Bs 0 | B1 = r =
1s1+t

 
P max Bs r = 2P{Bt r} = 2 [1 (r/ t)].
0st

Combining this we get

Z 1 2
q(t) = 2 [1 (r/ t)] er /2 dr.
2
This integral can be computed with polar coordinates. We just give the
answer
2 1
q(t) = 1 arctan .
t

If the filtration {Ft } is right continuous then the condition {T

t} Ft for all t is equivalent to the condition that {T < t} Ft
for all t. Indeed,
 
\ 1
{T t} = T <t+ Ft+ ,
n
n=1
2.8. QUADRATIC VARIATION 59

 
[ 1
{T < t} = T t Ft .
n
n=1
If T is a stopping time, the -algebra FT is defined to be the
set of events A F such that for each t, A {T t} Ft . (It is
not hard to show that this is a -algebra.) We think of FT as the
information available up to time T . If {T < } let

Yt = BT +t BT ,

and let GT denote the -algebra generated by {Yt : t 0}. We can

state the strong Markov property as:
{Yt : t 0} is a standard Brownian motion.
The -algebras FT and GT are independent, that is, if A
FT , A0 GT , then P(A A0 ) = P(A) P(A0 ).
To prove this, one first considers the case where T takes values in
{k2n : k = 0, 1, 2, . . .}. By splitting up separately into the events
{T = k2n }, one can use the usual Markov property to prove the
result. For more general stopping times T we approximate T by Tn
where Tn = k2n on the event {(k 1)2n < T k2n }. It is
important that we approximate from above in order to guarantee
that the Tn are stopping times. The continuity of the Brownian
paths implies that

Yt = lim [BTn +t BTn ] ,

n

and this can be used to conclude the independence of FT and GT .

2.8 Quadratic variation

In the next chapter we will come across quantities such as
n   
j1 2
 
X j
Qn = B B ,
n n
j=1

where Bt is a standard Brownian motion. We can write Qn as

n
1X
Yj ,
n
j=1
60 CHAPTER 2. BROWNIAN MOTION

where     2

j
B n B j1
n
Yj = Yj,n = .
1/ n

The random variables Y1 , . . . , Yn are independent, each with the distribution

of Z 2 where Z is a standard normal. In particular,

E [Yj ] = E Z 2 = 1, E Yj2 = E Z 4 = 3.
     

 4
(One can use integration by parts to calculate
h i E Z or one could just look
it up somewhere.) Hence Var[Yj ] = E Yj2 E [Yj ]2 = 2, and

n n
1X 1 X 2
E [Qn ] = E [Yj ] = 1, Var [Qn ] = 2
Var [Yj ] = .
n n n
j=1 j=1

As n , the variance of Qn tends to zero and the random variable

approaches a constant random variable of 1.
Similarly, for any time t, we let

X  j  j1 2
 
Qn (t) = B B .
n n
jtn

As n , the variance of Qn (t) goes to zero and it approaches a constant

random variable with value t.

Definition If Xt is a process, the quadratic variation is defined by

X j1 2
   
j
hXit = lim X X ,
n n n
jtn

where the sum is over all j with j/n t.

The calculation above shows that the quadratic variation of a standard

Brownian motion is the constant process hBit = t. Suppose Wt is a Brownian
motion with drift m and variance 2 . Then we can write Wt = Bt + mt
where Bt is a standard Brownian motion. Fix t and write

X j  j1 2 j1 2
  X j   
2
W W = B B
n n n n
2.8. QUADRATIC VARIATION 61
     X 2
2m X j j1 m
+ B B + ,
n n n n2
where in each case the sum is over j tn. As n ,

X j  j1 2
 
2
B B 2 hBit = 2 t,
n n
    
2m X j j1 2m
B B Bt 0,
n n n n
X m2 tnm2
0.
n2 n2
We have established the following.

Theorem 2.8.1. If Wt is a Brownian motion with drift m and variance 2 ,

then hW it = 2 t.

The important facts are that the quadratic variation is not random and
that it depends on the variance but not on the mean. It may seem silly at
this point to give a name and notation to a quantity which is almost trivial
for Brownian motion, but in the next chapter we will deal with processes
for which the quadratic variation is not just a linear function of time.

If 0 = t0 < t1 < < tn = t, we call the times a partition of

[0, t]. We will write for partitions and write

kk = max {tj tj1 }.

j=1,...,n

For any partition of [0, t], we define

n
X
Q(t; ) = [B(tj ) B(tj1 )]2 .
j=1

Computing as above, we see that

n
X
E[Q(t; )] = (tj tj1 ) = t,
j=1
62 CHAPTER 2. BROWNIAN MOTION

n
X  
Var[Q(t; )] = Var [B(tj ) B(tj1 )]2
j=1
Xn
= 2 (tj tj1 )2
j=1
n
X
2kk (tj tj1 ) = 2kkt.
j=1

Theorem 2.8.2. Suppose B is a standard Brownian motion,

t > 0, and n is a sequence of partitions of the form

0 = t0,n < t1,n < < tln ,n = t,

with kn k 0. Then Q(t; n ) t in probability. Moreover, if

X
kn k < , (2.5)
n=1

then with probability one Q(t; n ) t.

Proof. Using Chebyshevs inequality, we see that for each integer

k,
 
1 Var[Q(t; n )]
P |Q(t; n ) t| > 2k 2 kn kt,
k (1/k)2

and the right-hand side goes to zero as n . This gives conver-

gence in probability. If (2.5) holds as well, then
 
X 1
P |Q(t; n ) t| > < ,
k
n=1

and hence by the Borel-Cantelli lemma, with probability one, for all
n sufficiently large,
1
|Q(t; n ) t| .
k
It is important to note the order of the quantifiers in this theo-
rem. Let t = 1. The theorem states that for every sequence {n }
satisfying (2.5), Q(1; n ) 1. The event of measure zero on which
convergence does not hold depends on the sequence of partitions.
2.9. MULTIDIMENSIONAL BROWNIAN MOTION 63

Because there are an uncountable number of such sequences we

cannot conclude from the theorem that with probability one, for all
sequences {n } satisfying (2.5), that Q(1; n ) 1. In fact, the
latter statement is false. Let us give an example. Let us start with
the dyadic partition n of [0, 1] with 2n intervals of length 2n .
We will now subdivide some, but not all, of these intervals into two
equal pieces. Suppose [s, t] is one of these intervals and let r denote
the midpoint. We will subdivide [s, t] into [s, r] and [r, t] if

(Br Bs )2 + (Bt Bs )2 > (Bt Bs )2 ,

and we will retain the interval [s, t] otherwise. This defines a

partition n with kn k kn k. It has been chosen so that
Q(1; n )Q(1; n ), and one can show (see Exercise 2.5) that
h i
2
lim Q(1; n ) = E max{B1/2 + (B1 B1/2 )2 , B12 } > 1.
n

This does not contradict the theorem because the partitions n

depend on the realization of the Brownian motion.
Sometimes it is convenient to fix a sequence of partitions, say
partitions with dyadic intervals. Let

X  j + 1  2
j
Qn (t) = B n
B .
n
2 2n
j<t2

Since the dyadic rationals are countable, the theorem implies that
with probability one for every dyadic rational t, Qn (t) t. However,
since t 7 Qn (t) is increasing, we can conclude that with probability
one, for every t0, Qn (t) t.

2.9 Multidimensional Brownian motion

In finance one is often interested in considering the value of many assets at
the same time. Multidimensional Brownian motion Bt is random continu-
ous motion in d-dimensional space. It can be viewed as d one-dimensional
Brownian motions  
Bt = Bt1 , . . . , Btd ,
64 CHAPTER 2. BROWNIAN MOTION

with perhaps some correlations. If we make the same assumptions that we

made in the one-dimensional case (independent, identically distributed in-
crements and continuous paths), then one can show that the increments must
be multivariate normal random variables. We use this fact in our definition.

Definition The d-dimensional process

 
Bt = Bt1 , . . . , Btd ,

is called a d-dimensional Brownian motion starting at the origin with drift

m = (m1 , . . . , md ) Rd and d d covariance matrix with respect to the
filtration {Ft } if each Bt is Ft -measurable and the following holds.

B0 = 0.

If s < t, the distribution of Bt Bs is joint normal with mean (t s)m

and covariance matrix (t s).

If s < t, the random vector Bt Bs is independent of Fs .

With probability one, the function t 7 Bt is continuous.

In particular, each Btk is a Brownian motion with drift mk and variance

kk with respect to the filtration {Ft }. If m = 0 and = I, then Bt1 , . . . , Btd
are independent standard Brownian motions and Bt is called standard d-
dimensional Brownian motion.

Definition If Xt , Yt are processes defined on the same probability space,

the covariation (process) is defined by
X  j  
j1
   
j

j1

hX, Y it = lim X X Y Y ,
n n n n n
jtn

where the sum is over all j with j/n t.

If X = Y , then the covariation is the same as the quadratic variation,

that is
hX, Xit = hXit .
If Bt = (Bt1 , . . . , Btd ) is a Brownian motion with drift 0 and covariance
matrix , then
h i
E (Bti Bsi )(Btk Bsk ) = (t s) ik .
2.10. HEAT EQUATION AND GENERATOR 65

As in the quadratic variation, the drift does not contribute to the covariation.
We state the following result which is proved in the same way as for quadratic
variation.
Theorem 2.9.1. If Bt is a d-dimensional Brownian motion with drift m
and covariance matrix , then

hB i , B k it = ik t.

In particular, if Bt = (Bt1 , . . . , Btd ) is a standard Brownian motion in Rd ,

then the components are independent and

hB i , B k it = 0, i 6= k.

2.10 Heat equation and generator

Even if ones interest in Brownian motion comes from other applications such
as finance, it is useful view Brownian motion as a model for the diffusion
of heat. Imagine heat as consisting of a large number of independent heat
particles each doing random continuous motion. This persepective leads
to a deterministic partial differential equation (PDE) that describes the
evolution of the temperature. Suppose for the moment that the temperature
on the line is determined by the density of heat particles. Let pt (x) denote the
temperature at x at time t. If the heat particles are moving independently
and randomly then we can assume that they are doing Brownian motions.
If we also assume that Z
pt (x) dx = 1,

then it is reasonable to see pt (x) as the probability density for Brownian
motion.

2.10.1 One dimension

We start by taking advantage of what we already know. Assume Bt is a
standard Brownian motion starting at the origin and let pt (x) denote the
density of Bt . Since Bt N (0, t), we know that
1 x2
pt (x) = e 2t . (2.6)
2t
We view this as a function of two variables t, x. If we are interested in the
position at time s + t, we can use the Markovian nature of Brownian motion
66 CHAPTER 2. BROWNIAN MOTION

to first observe the position at time s and then to consider what happens in
the next time interval of length t. This leads to the Chapman-Kolmogorov
equation Z
ps+t (x) = ps (y) pt (x y) dy. (2.7)

A straightforward calculation, which we omit, shows that if pt (x) is given by

(2.6), then pt (x) satisfies (2.7). We emphasize that (2.7) uses the Markovian
property: if we are interested in the position at time t + s and are given
the information up to time s, the only information that is relevant is the
position at time s.
We will give a heuristic derivation of the PDE that describes the evolu-
tion of pt (x). For a number of our heuristic derivations we will use a binomial
approximation where we view the Brownian motion as satisfying
1
P {Bt+t = Bt + x} = P {Bt+t = Bt x} = ,
2

where x = t. In the approximation, to be at x at time t + t one must
be at x x at time t which gives
1 1
pt+t (x) pt (x x) + pt (x + x).
2 2
Using t = (x)2 , this implies

pt+t (x) pt (x) pt (x + x) + pt (x x) 2 pt (x)

= . (2.8)
t 2 (x)2

We now let t 0 in (2.8). The definition of the partial derivative implies

that
pt+t (x) pt (x)
lim = t pt (x).
t0 t
The right-hand side of (2.8) is a little more complicated. Since such limits
are fundamental in our study, let us give two ways to compute the limit. One
way is to write the right-hand side as
 
1 pt (x + x) pt (x) pt (x) pt (x x)
.
2x x x

Waving our hands, we say this is about

1 x pt (x) x pt (x x)
,
2 x
2.10. HEAT EQUATION AND GENERATOR 67

and now we have a difference quotient for the first derivative in x in which
case the limit should be
1
xx pt (x).
2
Another, essentially equivalent, method to evaluate the limit is to write
f (x) = pt (x) and expand in a Taylor series about x,
1 00
f (x + ) = f (x) + f 0 (x)  + f (x) 2 + o(2 ),
2
where o(2 ) denotes a term such that

o(2 )
lim = 0.
0 2

Then we see that

f (x + ) + f (x ) 2f (x) = f 00 (x) 2 + o(2 ),

and hence the limit of the right-hand side of (2.8) is xx pt (x)/2. We have
derived the heat equation
1
t pt (x) = xx pt (x).
2
While we have been a bit sketchy on details, one could start with this equa-
tion and note that pt as defined in (2.6) satisfies this. This is the solution
given that B0 = 0, that is, when the initial density p0 (x) is the delta
function at 0. (The delta function, written () is the probability density
of the probability distribution that gives measure one to the point 0. This
is not really a density, but informally we write

(0) = , (x) = 0, x 6= 0,
Z
(x) dx = 1.

These last equations do not make mathematical sense, but they give a work-
able heuristic definition.)
If the Brownian motion has variance 2 , one binomial approximation is
1
P {Bt+t = Bt + x} = P {Bt+t = Bt x} = ,
2

where x = t. The factor is put in so that

Var [B(t + t) B(t)] = 2 t.

68 CHAPTER 2. BROWNIAN MOTION

We can use the same argument to derive that this density should satisfy the
heat equation
2
t pt (x) = xx pt (x).
2
The coefficient 2 (or in some texts ( 2 /2)) is referred to as the diffusion
coefficient. One can check that a solution to this equation is given by

x2
 
1
pt (x) = exp 2 .
2 2 t 2 t
When the Brownian motion has drift m, the equation gets another term.
To see what the term should look like, let us first consider the case of deter-
ministic linear motion, that is, motion with drift m but no variance. Then
if pt (x) denotes the density at x at time t, we get the relationship

pt+t (x) = pt (x m t),

since particles at x at time t + t must have been at x mt at time t.

This gives
pt+t (x) pt (x) pt (x m t) pt (x)
=
t t
pt (x) pt (x m t)
= m .
mt
Letting t 0, we get the equation for linear motion at rate m,

t pt (x) = m x pt (x).

This is a first-order equation which means it contains only first derivatives.

If Bt is Brownian motion with drift m and variance 2 , then we can do a
similar argument and show that the density pt (x) satisfies the second order
PDE
2
t pt (x) = m x pt (x) + xx pt (x).
2
As a check, one can consider the appropriate density,

(x mt)2
 
1
pt (x) = exp ,
2 2 t 2 2 t
and show that it satisfies this equation.
Before summarizing, we will change the notation slightly. Let pt (x, y)
denote the density of Bt given B0 = x. Under this notation pt (x) = pt (0, x).
2.10. HEAT EQUATION AND GENERATOR 69

We will define an operator on functions. A (linear) operator is a function L

from (a linear subspace of) functions to functions satisfying L(af + bg) =
aL(f ) + bL(g) where a, b are constants and f, g are functions. We say that
a function f : R R is C 2 if it is twice differentiable and the derivatives
are continuous functions of x. For any m, 2 , we define an operator on C 2
functions by
2 00
L f (x) = m f 0 (x) + f (x).
2
If f is a function of more variables, such as ft (x, y) we write Lx to indicate
the operator acting on the variable x,

2
Lx ft (x, y) = m x ft (x, y) + xx ft (x, y).
2
Theorem 2.10.1. Suppose Bt is a standard Brownian motion with drift
m and variance 2 . Then the transition density pt (x, y) satisfies the heat
equation
t pt (x, y) = Ly pt (x, y)
with initial condition p0 (x, ) = x (). Here L is the operator on functions

2 00
L f (y) = m f 0 (y) + f (y).
2
We think of pt (x, y) as the probability of being at y at time t given that
B0 = x. For driftless Brownian motion, this is the same as the probability
of being at x given that one was at y. However, the reversal of a Brownian
motion with drift m should be a Brownian motion with drift m. This gives
the following.

Theorem 2.10.2. Suppose Bt is a standard Brownian motion with drift

m and variance 2 . Then the transition density pt (x, y) satisfies the heat
equation
t pt (x, y) = Lx pt (x, y)
with initial condition p0 (, y) = y (). Here L is the operator on functions

2 00
Lf (x) = m f 0 (x) + f (x).
2
The operator L will be more important to us than the operator L which
is why we give it the simpler notation.
70 CHAPTER 2. BROWNIAN MOTION

Let H denote the real Hilbert space L2 (R) with inner product
Z
(f, g) = f (x) g(x)dx.

The operators L, Lare not defined on all of H but they can be

defined on a dense subspace, for example, the set of C 2 functions
all of whose derivatives decay rapidly at infinity. The operator L is
the adjoint of L which means,

(L f, g) = (f, Lg).

One can verify this using the following relations that are obtained
by integration by parts:
Z Z
f (x) g 0 (x) dx = f 0 (x) g(x) dx,

Z Z
f (x) g 00 (x) dx = f 00 (x) g(x) dx.

Suppose B0 has an initial density f . Then the density of Bt is
given by
Z

ft (y) = Pt f (y) := f (x) pt (x, y) dx.

This is the solution to the equation

t ft (y) = Ly ft (y)

with initial condition f0 (y) = f (y). We can write the heat equation
as a derivative for operators,

t Pt = L Pt .

2.10.2 Expected value at a future time

Suppose Bt is a Brownian motion with drift m and variance 2 , and let f
be a function on R. For example, if we consider Bt to be the price of a stock
and f to be the worth of a call option at strike price S at time t, then

x S if x S
f (x) = (x S)+ = . (2.9)
0 if x < S
2.10. HEAT EQUATION AND GENERATOR 71

Let (t, x) be the expected value of f (Bt ) given that B0 = x. We will write
this as
(t, x) = Ex [f (Bt )] = E [f (Bt ) | B0 = x] .

Then Z
(t, x) = f (y) pt (x, y) dy.

The time derivative can be computed by interchanging differentiation and

integration and using the rule from the previous section.
Z
t (t, x) = t f (y) pt (x, y) dy
Z
= f (y) t pt (x, y) dy

Z
= f (y) Lx pt (x, y) dy

Z
= Lx f (y) pt (x, y) dy = Lx (t, x).

We will give another derivation of this equation, and while we are at it we

will do the multivariate case. Suppose Bt = (Bt1 , . . . , Btd ) is a d-dimensional
Brownian motion with drift m = (m1 , . . . , md ) and covariance matrix . We
assume that the Brownian motion is truly d-dimensional, or, in other words,
that is a nonsingular matrix. Let pt (x, y) be the transition probability. If
f : Rd R is a bounded function, define Pt f by
Z
x
Pt f (x) = E [f (Bt )] = f (y) pt (x, y) dy.

Here x = (x1 , . . . , xd ), y = (y1 , . . . , yd ) Rd and dy = dy1 dyd .

Definition The (infinitesimal) generator of a Markov process Xt is the

operator L on functions defined by

Pt f (x) f (x)
Lf (x) = lim ,
t0 t

where
Pt f (x) = Ex [f (Xt )] .
72 CHAPTER 2. BROWNIAN MOTION

Theorem 2.10.3. If Bt is a d-dimensional Brownian motion with drift

m = (m1 , . . . , md ) and covariance matrix , then the infinitesimal generator
is given by
d d
1 XX
Lf (x) = m f (x) + jk jk f (x).
2
j=1 k=1

In particular, if f is a function and

(t, x) = Pt f (x) = Ex [f (Bt )],

then satisfies
t (t, x) = Lx (t, x), t > 0,

with initial condition (0, x) = f (x).

We will give a quick derivation, assuming f is C 2 and t = x = 0. The

second-order Taylor approximation of f around the origin is

d d d
X 1 XX
f () = f (0) + bj j + ajk j k + o(||2 ),
2
j=1 j=1 k=1

where  = (1 , . . . , d ) and

bj = j f (0), ajk = jk f (0).

In particular,

d d d
1 XX
bj Btj ajk Btj Btk + o(|Bt |2 ).
X
f (Bt ) f (B0 ) = +
2
j=1 j=1 k=1

We know that h i  
E Btj = mj t, E Btj Btk = jk t,

and the expectation of the error term o(|Bt |2 ) is o(t). Hence,

d E[f (Bt ) f (B0 )]

E [f (Bt )] |t=0+ = lim = Lf (0).
dt t0 t
2.10. HEAT EQUATION AND GENERATOR 73

Suppose f is a measurable function that does not grow too fast

at infinity. For example, we assume that
2
e|x| f (x) 0, |x| ,

for every > 0. Then

Z
|f (y)| pt (x, y) dy < ,

and hence
Z
(t, x) := Pt f (x) = f (y) pt (x, y) dy <

is well defined for all t. Moreover, for t > 0, the interchange of

integration
Z Z
t f (y) pt (x, y) dy = f (y) t pt (x, y) dy

can be justified (say by the dominated convergence theorem). Sim-

ilarly, integrals with respect to x can be taken to show that (t, )
is C in x and
Z
Lx (t, x) = f (y) Lx pt (x, y) dy

Therefore, for t > 0, x 7 (t, x) is C 2 and we can take the

right derivative with respect to time,
(t + s, x) (t, x)
lim = Lx (t, x),
s0 s
using the argument as above. Although this argument only com-
putes the right derivative, since for fixed x, (t, x) and Lx (t, x)
are continuous functions of t, we can conclude that

t (t, x) = Lx (t, x), t > 0.

We cannot expect this to hold at t = 0, but we can state that if f

is continuous at x, then

lim (t, x) = f (x).

t0
74 CHAPTER 2. BROWNIAN MOTION

A simple call option works a little differently. Suppose that Bt is a one-

dimensional Brownian motion with parameters m, 2 . A simple call option
at time T > 0 with strike price S allows the owner to buy a share of the
stock at time T for price S. If the price at time T is BT , then the value of
the option is f (BT ) = (BT S)+ as in (2.9). We are specifying the value
of the function at time T rather than at time 0. However, we can use our
work to give a PDE for the expected value of the option. If t < T , then the
expected value of this option, given that Bt = x is

F (t, x) = E [F (BT ) | Bt = x] = Ex [F (BT t )] = (T t, x).

Since

t F (t, x) = t (T t, x) = Lx (T t, x) = Lx F (t, x),

we get the following.

If f is a function, T > 0 and

F (t, x) = E[f (BT ) | Bt = x],

then for t < T , F satisfies the backwards heat equation

t F (t, x) = Lx F (t, x),

with terminal condition F (T, x) = f (x).

As in the one-dimensional case, we can find the operator associated to

the transition density. If we run a Brownian motion with drift m and covari-
ance matrix backwards, we get the same covariance matrix but the drift
becomes m. Therefore

For t > 0, the transition probability pt (x, y) satisfies the equation

t pt (x, y) = Ly pt (x, y),

where
d d
1 XX
L f (y) = m f (y) + jk jk f (y).
2
j=1 k=1
2.11. EXERCISES 75

The equations

t pt (x, y) = Lx pt (x, y), t pt (x, y) = Ly pt (x, y)

are sometimes called the Kolmogorov backwards and forwards equations, re-
spectively. The name comes from the fact that they can be derived from the
Chapman-Kolmogorov equations by writing
Z
pt+t (x, y) = pt (x, z) pt (z, y) dz,
Z
pt+t (x, y) = pt (x, z) pt (z, y) dz,

respectively. The forward equation is also known as the Fokker-Planck equa-

tion. We will make more use of the backwards equation.

2.11 Exercises
Exercise 2.1. Let Z1 , Z2 , Z3 be independent N (0, 1) random variables. Let

X1 = Z1 + Z3 , X2 = Z2 + 4Z3 , X3 = 2Z1 2Z2 + rZ3

where r is a real number.

1. Explain why X = (X1 , X2 , X3 ) has a joint normal distribution.

2. Find the covariance matrix for X (in terms of r).

3. For what values of r are X1 and X3 independent random variables?

4. For what values of r does the random vector X have a density in R3 ?

Exercise 2.2. Let Bt be a standard Brownian motion. Find the following

probabilities. If you cannot give the answer precisely give it up to at least
three decimal places using a table of the normal distribution.

1. P{B3 1/2}

2. P{B1 1/2, B3 > B1 + 2}

3. P(E) where E is the event that the path stays below the line y = 6 up
to time t = 10.

4. P{B4 0 | B2 0}.
76 CHAPTER 2. BROWNIAN MOTION

Exercise 2.3. Let Bt be a standard Brownian motion. For each positive

integer n, let
n   
j1 2
 
X j
Jn = B B .
n n
j=1

1. Find the mean and variance of the random variable Jn .

2. Prove the following weak law of large numbers: if  > 0, then

lim P {|Jn 1| > } = 0.

n

Hint: this uses Chebyshevs inequality look it up if this is not fa-

miliar to you.
In the next exercise, you can use the following computation. If X
N (0, 1), then the moment generating function of X is given by
2
m(t) = E etX = et /2 .
 

Exercise 2.4. Suppose Bt is a standard Brownian motion and let Ft be its

corresponding filtration. Let
2 t
Mt = eBt 2 .

Show that Mt is a martingale with respect to Ft . In other words, show that

if s < t, then
E(Mt | Fs ) = Ms .
Exercise 2.5. Let Bt be a standard Brownian motion and let

Y = max B12 + (B2 B1 )2 , B22 .



Show that

Y = B22 + B1 (B2 B1 ) 1{B1 (B2 B1 ) 0}.

Find E[Y ].

Show that for every k < , E Y k < .

 

Exercise 2.6. Let Bt be a standard Brownian motion and let {Ft } denote
the usual filtration. Suppose s < t. Compute the following.
1. E[Bt2 | Fs ]
2.11. EXERCISES 77

2. E[Bt3 | Fs ]

3. E[Bt4 | Fs ]

4. E[e4Bt 2 | Fs ]
Exercise 2.7. Let Bt be a standard Brownian motion and let

Y (t) = t B(1/t).

1. Is Y (t) a Gaussian process?

2. Compute Cov(Y (s), Y (t)).

3. Does Y (t) have the distribution of a standard Brownian motion?

Exercise 2.8. If f (t), 0 t 1 is a continuous function, define the (3/2)-
variation up to time one to be
n     3/2
X  j j 1 
Q = lim f f  .
n  n n 
j=1

What is Q if
1. f is a nonconstant, continuously differentiable function on R?

2. f is a Brownian motion?
Exercise 2.9. Suppose Bt is a standard Brownian motion. For the functions
(t, x), 0 < t < 1, < x < , defined below, give a PDE satisfied by the
function.
1. (t, x) = P{Bt 0 | B0 = x}.

2. (t, x) = E[B12 | Bt = x].

3. Repeat the two examples above if Bt is a Brownian motion with drift

m and variance 2 .
Exercise 2.10. Suppose Bt is a standard Brownian motion and

Mt = max Bs .
0st

1. Explain why Mt has the same distribution as t M1 .

2. Find the density of M1 .

78 CHAPTER 2. BROWNIAN MOTION

3. Find E[Mt ].

4. If a > 0, find E[M1 1{M1 a}].

5. Let Ft be the information in {Bs : 0 s t}. Find E[M2 | F1 ].

Exercise 2.11. Write a program that will simulate a standard Brownian

motion using step size t = .01.

1. Graph at least one realization of the Brownian motion to produce a

pretty picture of a Brownian path.

2. Use simulations to estimate the following probability:

P{Bt 2 for all t 1}.

Run the simulation enough times to get a good estimate for the prob-
ability. Use the reflection principle to calculate the actual probability
and compare the result.

Exercise 2.12. Prove Theorem 2.6.3.

Chapter 3

Stochastic integration

3.1 What is stochastic calculus?

Before venturing into stochastic calculus, it will be useful to review the basic
ideas behind the usual differential and integral calculus. The main deep idea
of calculus is that one can find values of a function knowing the rate of
change of the function. For example, suppose f (t) denotes the position of a
(one-dimensional) particle at time t, and we are given the rate of change
df (t) = C(t, f (t)) dt,
or as it is usually written,
df
= f 0 (t) = C(t, f (t)).
dt
In other words, at time t the graph of f moves for an infinitesimal amount
of time along a straight line with slope C(t, f (t)). This is an example of a
differential equation, where the rate depends both on time and position. If
we are given an initial condition f (0) = x0 , then (under some hypotheses
on the rate function C, see the end of Section 3.5), the function is defined
and can be given by
Z t
f (t) = x0 + C(s, f (s)) ds.
0
If one is lucky, then one can do the integration and find the function exactly.
If one cannot do this, one can still use a computer to approximate the
solution. The simplest, but still useful, technique is Eulers method where
one chooses a small increment t and then writes
f ((k + 1)t) = f (kt) + t C(kt, f (kt)).

79
80 CHAPTER 3. STOCHASTIC INTEGRATION

Stochastic calculus is similar, except that one adds randomness to the

change. We will make sense of equations such as
dXt = m(t, Xt ) dt + (t, Xt ) dBt , (3.1)
where Bt is a standard Brownian motion. This is an example of a stochastic
differential equation (SDE). We should read this equation as stating that at
time t, Xt is evolving like a Brownian motion with drift m(t, Xt ) and vari-
ance (t, Xt )2 . Solving such equations exactly is much harder than for usual
calculus, and we often have to resort to numerical methods. One technique is
the stochastic Euler method which allows us to do Monte Carlo simulations
of the process. This is easy enough to describe so we give the formula now
X((k + 1)t) =

X(kt) + t m(kt, X(kt)) + t (kt , X(kt)) Nk ,
where Nk is a N (0, 1) random variable.
The standard approach to calculus is to define the derivative, then define
the integral, and then give the relationship between the two. One could also
define the integral first. In stochastic calculus, the derivative is not so easily
defined, so in order to give a mathematical formulation we concentrate on
defining the stochastic integral. We will say that Xt is a solution to (3.1) if
Z t Z t
Xt = X0 + m(s, Xs ) ds + (s, Xs ) dBs .
0 0

The ds integral is the usual integral from calculus; the integrand m(s, Xs )
is random, but that does not give any problem in defining the integral. The
main task will be to give precise meaning to the second term, and more
generally to Z t
As dBs .
0
There are several approaches to stochastic integration. The approach we
give, which is most commonly used in mathematical finance, is that of the
Ito integral.

3.2 Stochastic integral

Let Bt be a standard (one-dimensional) Brownian motion with respect to a
filtration {Ft }. We want to define the process
Z t
Zt = As dBs .
0
3.2. STOCHASTIC INTEGRAL 81

We think of Zt as a Brownian motion which at time s has variance A2s . In

analogy to the discrete stochastic integral, we think of As as the amount
bet or invested at time s with negative bets indicating bets that the
process will go down. Also, as in the discrete case, we want to restrict our
betting strategies to those that cannot look into the future. For continuous
time processes for which changes are happening infinitesimally it is some-
what trickier to make the last condition precise. We will start by doing some
easy cases and then define more complicated cases by taking limits.

3.2.1 Review of Riemann integration

Let us review the definition of the usual Riemann integral from calculus.
Suppose f (t) is a continuous function and we wish to define
Z 1
f (t) dt.
0

We partition [0, 1] into small intervals, say

0 = t0 < t1 < < tn = 1,

and approximate f (t) by a step function

fn (t) = f (sj ), tj1 < t tj ,

where sj is some point chosen in [tj1 , tj ]. We define

Z 1 n
X
fn (t) dt = f (sj ) (tj tj1 ).
0 j=1

It is a theorem that if we take a sequence of partitions such that the maxi-

mum length of the intervals [tj1 , tj ] goes to zero, then the limit
Z 1 Z 1
f (t) dt = lim fn (t) dt
0 n 0

exists and is independent of the choice of sequence of partitions or the choice

of sj . This is the definition of the left-hand side. One later proves the im-
portant facts, for example, the fundamental theorem of calculus,
Z b
f 0 (t) dt = f (b) f (a).
a
82 CHAPTER 3. STOCHASTIC INTEGRATION

3.2.2 Integration of simple processes

The analogue of a step function for the stochastic integral is a simple process.
This corresponds to a betting strategy that allows one to change bets only
at a prescribed finite set of times. To be more precise, a process At is a
simple process if there exist times
0 = t0 < t1 < < tn <
and random variables Yj , j = 0, 1, . . . , n that are Ftj -measurable such that
At = Yj , tj t < tj+1 .
Here we set tn+1 = . Since Yj is Ftj -measurable, At is Ft -measurable. We
also assume that E[Yj2 ] < for each j. If At is a simple process we define
Z t
Zt = As dBs ,
0

by
j1
X
Ztj = Yi [Bti+1 Bti ],
i=0
and, more generally,
Zt = Ztj + Yj [Bt Btj ] if tj t tj+1 ,
Z t
As dBs = Zt Zr .
r
There are four important properties of the stochastic integral of simple
processes which we give in the next proposition. The reader should compare
these with the properties of integration with respect to random walk in
Section 1.6.
Proposition 3.2.1. Suppose Bt is a standard Brownian motion with respect
to a filtration {Ft }, and At , Ct are simple processes.
Linearity. If a, b are constants, then aAt +bCt is also a simple process
and Z t Z t Z t
(aAs + bCs ) dBs = a As dBs + b Cs dBs .
0 0 0
Moreover, if 0 < r < t,
Z t Z r Z t
As ds = As dBs + As dBs .
0 0 r
3.2. STOCHASTIC INTEGRAL 83

Martingale property. The process

Z t
Zt = As dBs
0

is a martingale with respect to {Ft }.

Variance rule. Zt is square integrable and

Z t
Var[Zt ] = E Zt2 = E[A2s ] ds.
 
0

Continuity. With probability one, the function t 7 Zt is a continuous

function.

Let us discuss the proof of the proposition. Linearity follows immediately

from the definition, and continuity follows from the continuity of Brownian
motion. To show that Zt is a martingale, we need to show that

E(Zt | Fs ) = Zs if s < t. (3.2)

We will do this in the case t = tj , s = tk for some j > k and leave the other
cases for the reader. In this case
k1
X
Zs = Yi [Bti+1 Bti ],
i=0

and
j1
X
Zt = Zs + Yi [Bti+1 Bti ].
i=k

Since E(Zs | Fs ) = Zs , we see that

j1
X  
E(Zt | Fs ) = Zs + E Yi [Bti+1 Bti ] | Fs .
i=k

For k i j 1, we have ti s and hence

  

E Yi [Bti+1 Bti ] | Fs = E E Yi [Bti+1 Bti ] | Fti | Fs .

Since Yi is Fti -measurable and Bti+1 Bti is independent of Fti ,

E Yi [Bti+1 Bti ] | Fti = Yi E Bti+1 Bti | Fti = Yi E[Bti+1 Bti ] = 0.
84 CHAPTER 3. STOCHASTIC INTEGRATION

This gives (3.2).

We will prove the variance rule for t = tj in which case
j1 X
X j1
Zt2 = Yi [Bti+1 Bti ]Yk [Btk+1 Btk ].
i=0 k=0

If i < k,  
E Yi [Bti+1 Bti ]Yk [Btk+1 Btk ]


= E E Yi [Bti+1 Bti ]Yk [Btk+1 Btk ] | Ftk .
The random variables Yi , Yk , Bti+1 Bti are all Ftk -measurable while Btk+1
Btk is independent of Ftk , and hence


E Yi [Bti+1 Bti ]Yk [Btk+1 Btk ] | Ftk


= Yi [Bti+1 Bti ]Yk E Btk+1 Btk | Ftk
 
= Yi [Bti+1 Bti ]Yk E Btk+1 Btk = 0.

Arguing similarly for i > k, we see that

j1
X
E[Zt2 ] E Yi2 (Bti+1 Bti )2 .
 
=
i=0

(We have just reproduced the argument showing orthogonality of martingale

increments.) Since Yi2 is Fti -measurable and (Bti+1 Bti )2 is independent
of Fti , we get

E Yi2 (Bti+1 Bti )2 | Fti = Yi2 E (Bti+1 Bti )2 | Fti

   

= Yi2 E (Bti+1 Bti )2

 

= Yi2 (ti+1 ti ),

and hence,

E Yi2 (Bti+1 Bti )2 = E E Yi2 (Bti+1 Bti )2 | Fti

  


= E[Yi2 ] (ti+1 ti ).

The function s 7 E[A2s ] is a step function that takes on the value E[Yi2 ] for
ti s < ti+1 . Therefore,
j1
X Z t
E[Zt2 ] = E[Yi2 ] (ti+1 ti ) = E[A2s ] ds.
i=0 0
3.2. STOCHASTIC INTEGRAL 85

3.2.3 Integration of continuous processes

In this section we will assume that At is a process with continuous paths that
is adapted to the filtration {Ft }. Recall that this implies that each At is Ft -
measurable. To integrate these processes we use the following approximation
result.
Lemma 3.2.2. Suppose At is a process with continuous paths, adapted to
the filtration {Ft }. Suppose also that there exists C < such that with
probability one |At | C for all t. Then there exists a sequence of simple
(n)
processes At such that for all t,
Z t h i
lim E |As A(n)
s |
2
ds = 0. (3.3)
n 0

(n)
Moreover, for all n, t, |At | C.

We will prove this lemma for t = 1. let

(n) j j+1
At = A(j, n), t< ,
n n
where A(0, n) = A0 and for j 1,
Z j/n
A(j, n) = n As ds.
(j1)/n

It suffices to prove (3.3) for each fixed value of t and for ease we
(n)
will choose t = 1. By construction, the At are simple processes
(n)
satisfying |At | C. Since (with probability one) the function
t 7 At is continuous, we have
(n)
At At ,
and hence by the bounded convergence theorem applied to Lebesgue
measure,
lim Yn = 0,
n
where Z 1
(n)
Yn = [At At ]2 dt. (3.4)
0
Since the random variables {Yn } are uniformly bounded, this implies
Z 1 
(n) 2
lim E [At At ] dt = lim E[Yn ] = 0.
n 0 n
86 CHAPTER 3. STOCHASTIC INTEGRATION

Given the lemma, we can define

Z t
As dBs ,
0

for a bounded, continuous process As as follows. Find a sequence of simple

(n)
process As satisfying (3.3). Then it can be shown that for each t there is
an square integrable random variable Zt such that
Z t
Zt = lim A(n)
s dBs .
n 0

We define Z t
As dBs = Zt .
0
The integral satisfies four properties which should start becoming familiar.

Proposition 3.2.3. Suppose Bt is a standard Brownian motion with re-

spect to a filtration {Ft }, and At , Ct are bounded, adapted processes with
continuous paths.

Linearity. If a, b are constants, then

Z t Z t Z t
(aAs + bCs ) dBs = a As dBs + b Cs dBs .
0 0 0

Also, if r < t,
Z t Z r Z t
As dBs = As dBs + As dBs .
0 0 r

Martingale property. The process

Z t
Zt = As dBs
0

is a martingale with respect to {Ft }.

Variance rule. Zt is square integrable and

Z t
 2
Var[Zt ] = E Zt = E[A2s ] ds.
0
3.2. STOCHASTIC INTEGRAL 87

Continuity. With probability one, t 7 Zt is a continuous function.

For fixed t, the existence of Zt follows from the estimate (3.3)

and the completeness of L2 . In this case, the convergence is in L2 .
However, there are issues dealing with the fact that there are an
uncountable number of times. This can be handled similarly to the
construction of Brownian motion. For convenience, we restrict to
0 t 1.
We first consider t D and define Zt , t D as the L2 limit
which is defined for t D up to a single event of probability zero.
Suppose At is a simple process with |At | C. Then (Exercise 3.1)
one can show that

E[(Zt Zs )4 ] 3 C 4 |t s|2 .

By Fatous lemma, this bound will also hold for the limit process.
This estimate and an argument first due to Kolmogorov suffice to
give continuity. We leave the argument as Exercise 3.11.
(n) (n)
Let At be a sequence of simple processes and let Zt , Zt
denote the corresponding stochastic integrals. Let
h i Z 1 h i
(n) (n)
kA(n) Ak2 = E (Z1 Z1 )2 = E (At At )2 dt,
0

and let
(n)
Qn = max |Zt Zt |.
0t1

If we view Z (n) , Z as random variables taking values in C[0, 1],

the set of continuous functions on [0, 1] with the supremum norm,
then Qn is the distance between Z (n) and Z. The next proposition
establishes convergence with probability one in this space.

Proposition 3.2.4. If

X
kA(n) Ak2 < , (3.5)
n=1

then with probability one, Qn 0.

88 CHAPTER 3. STOCHASTIC INTEGRATION

Proof. By the Borel-Cantelli lemma, it suffices to show that for

every  > 0,
X
P{Qn > } = 0.
n=1
(n)
Continuity of Zt and Zt implies that
 
(n)
P{Qn > } = lim P max |Zt Zt | .
m tDm

(n)
For fixed m, the process Mt = Zt Zt , t Dm , is a discrete time
martingale and Corollary 1.7.2 implies that
E[M12 ]
 
P max |Mt | >  .
tDm 2
Therefore,
E[M12 ]
P{Qn > } = 2 kA(n) At k2 .
2
Let us emphasize the order of quantifiers in the proposition.
Given a sequence A(n) of approximating simple processes satisfying
(3.5), we get convergence with probability one. It is not true that
with probability one, we get convergence for every sequence. See an
analogous discussion in Section 2.8 about the quadratic variation.

If At is adapted with continuous paths, but not necessarily bounded, we

can still define Z t
Zt = As dBs
0
(n)
as follows. For each n < , let Tn = min{t : |At | = n} and let As = AsTn .
(n)
Then As is a bounded, continuous process and
Z t
(n)
Zt = A(n)
s dBs .
0

is well defined. We define

(n)
Zt = lim Zt .
n
The existence of this limit is easy to establish. Indeed, let

Kt = max |As |.
0st
3.2. STOCHASTIC INTEGRAL 89

(n)
Then for n Kt , As = As for 0 s t, and hence
(n)
Zt = Zt , t Kt .

There is a subtlety here. Since Kt is a random variable, the n that one

needs to choose may depend on the path. Although Zt is well defined in the
limit and satisfies linearity and continuity, we will see in Section 4.1 that it
might not satisfy the martingale property. Let us not worry about this at
the moment. We will state the following.
Proposition 3.2.5. Suppose Bt is a standard Brownian motion with respect
to a filtration {Ft }, and At , Ct are adapted processes with continuous paths.
Linearity. If a, b are constants and r < t, then
Z t Z t Z t
(aAs + bCs ) dBs = a As dBs + b Cs dBs ,
0 0 0
Z t Z r Z t
As dBs = As dBs + As dBs .
0 0 r

Variance rule. The variance of Zt satisfies

Z t
 2
Var[Zt ] = E Zt = E[A2s ] ds.
0

It is possible for both sides to equal infinity. However, if Var[Zt ] <

for all t, then Zt is a square integrable martingale.

Continuity With probability one, t 7 Zt is a continuous function.

We can relax the condition of continuous paths. We say that the paths
are piecewise continuous if with probability one on each finite interval the
paths are continuous except a finite number of points, say 0 < t1 < t2 <
< tn = t. For a piecewise continuous function with discontinuities at
times t1 , t2 , . . ., we define
Z t Z t1 Z tn
As dBs = As dBs + + As dBs . (3.6)
0 0 tn1

The value of this integral does not depend on how we define At at the
discontinuity. In this book, the process At will have continuous or piecewise
continuous paths although the integral can be extended to more general
processes. Note that the simple processes have piecewise continuous paths.
90 CHAPTER 3. STOCHASTIC INTEGRATION

One important case that arises comes from a stopping time. Suppose T is
a stopping time with respect to {Ft }. Then if At is a continuous, adapted
process and Z t
Zt = As dBs ,
0
then Z tT Z t
ZtT = As dBs = As,T dBs ,
0 0
where As,T denotes the piecewise continuous process,

As s < T
As,T = .
0 sT

In other words, stopping a stochastic integral is the same as changing the

bet to zero.

Continuity of paths is much more than is needed for the stochas-

tic integral to exist. If At is a uniformly bounded, adapted process,
and we let Z t
(n)
At = n As ds, (3.7)
1
t n
(n)
then At is an adapted continuous process. This requires some
assumptions on the process At so that integrals as in (3.4) and (3.7)
exists, at least as Lebesgue integrals, and so that we can use Fubinis
theorem to interchange expectation and integrals on the line. The
condition to guarantee this is called progressive measurablility, but
we will not go into details here. If the At are uniformly bounded, then
(n)
the Lebesgue density theory implies that At At for Lebesgue
almost every t. Hence, with probability one,
Z 1
(n)
[At At ]2 dt 0,
0

and since the processes are bounded the convergence is also in L2 .

For such processes we can define the stochastic integral as above
as an L2 -limit.

We will write the stochastic differential

dXt = At dBt ,
3.2. STOCHASTIC INTEGRAL 91

to mean that Xt satisfies

Z t
Xt = X0 + As dBs .
0

This has been made mathematically precise by the definition of the integral.
Intuitively, we think of Xt as a process that at time t evolves like a Brownian
motion with zero drift and variance A2t . This is well defined for any adapted,
continuous process At , and Xt is a continuous function of t. In particular, if
is a bounded continuous function, then we can hope to solve the equation

dXt = (Xt ) dBt .

Solving such an equation can be difficult (see the end of Section 3.5), but
simulating such a process is straightforward using the stochastic Euler rule:

Xt+t = Xt + (Xt ) t N,

where N N (0, 1).

The rules of stochastic calculus are not the same as those of usual cal-
culus. As an example, consider the integral
Z t
Zt = Bs dBs .
0

Although Bt is not a bounded process, it is continuous, adapted and

t t
t2
Z Z
E[Bs2 ] ds = s ds = < ,
0 0 2

so Zt is a square integrable martingale. If one naively followed the rules of

calculus, one might hope that

1  2  B2
Zt = Bt B02 = t .
2 2
However, a quick check shows that this cannot be correct. The left-hand side
is a martingale with Z0 = 0 and hence

E[Zt ] = 0.

However,
E Bt2 /2 = t/2 6= 0.
 
92 CHAPTER 3. STOCHASTIC INTEGRATION

In the next section we will derive the main tool for calculating integrals,
Itos formula or lemma. Using this we will show that, in fact,
1 2
Zt = [B t]. (3.8)
2 t
This is a very special case. In general, one must do more than just sub-
tract the expectation. One thing to note about the solution for t > 0,
the random variable on the right-hand side of (3.8) does not have a normal
distribution. Even though stochastic integrals are defined as limits of nor-
mal increments, the betting factor At can depend on the past and this
allows one to get non-normal random variables. If the integrand At = f (t)
is nonrandom, then the integral
Z t
f (s) dBs ,
0

is really a limit of normal random variables and hence has a normal distri-
bution (see Exercise 3.8).
If Z t
Zt = As dBs ,
0
then the quadratic variation of Z is defined by
X  j  j1 2
 
hZit = lim Z Z .
n n n
jnt

Theorem 3.2.6. If At is an adapted process with continuous or piecewise

continuous paths and Z t
Zt = As dBs ,
0
then Z t
hZit = A2s ds.
0
Rt
Note that if > 0 is a constant and Zt = 0 dBs , then Z is a Brownian
motion with variance parameter 2 for which we have already seen that
hZit = 2 t. The quadratic variation hZit should be viewed as the total
amount of randomness or the total amount of betting that has been done
up to time t. For Brownian motion, the betting rate stays constant and hence
the quadratic variation grows linearly. For more general stochastic integrals,
the current bet depends on the results of the games up to that point and
3.2. STOCHASTIC INTEGRAL 93

hence is a random variable. The quadratic variation at time t is a random

variable with mean
Z t  Z t
2
E A2s ds.
 
E [hZit ] = E As ds =
0 0

An alternative way to define the quadratic variation hZit for

square integrable martingales is as the unique increasing process
such that
Mt = Zt2 hZit

is a martingale. If At is a simple process, it is not hard (Exercise 3.4)

to show that Mt as defined above is a martingale. More generally,
one can establish this by taking limits. To show that this charac-
terization defines hZi uniquely, we need the following proposition.
The total variation of a function f : [0, ) R is defined as
n
X
Vt (f ) = sup |f (tj1 ) f (tj )| K,
j=1

where the supremum is over all partitions

0 = t0 < t1 < t2 < < tn = t.

We say f has bounded variation (locally) if Vt (f ) < for all t.

Increasing functions clearly have bounded variation, and it is not too
hard to see that the difference of increasing functions has bounded
variation. Also, if f is continuous with bounded variation,

lim Vt (f ) = 0.
t0

The next proposition shows that nontrivial continuous martingales

never have paths of bounded variation.

Proposition 3.2.7. Suppose Mt is a square integrable martin-

gale with respect to {Ft } with M0 = 0 such that with probability
one, the paths of Mt are continuous with bounded variation. Then
Mt = 0 for every t.
94 CHAPTER 3. STOCHASTIC INTEGRATION

Proof. Since Mt has continuous paths, it suffices to show that

P{Mt = 0} = 1 for every rational t. The argument is the same for
all t, so let us assume t = 1. We first make the stronger assumption
that V1 (M ) K < . Let
n 
j1 2
   
X j
Qn = M M ,
n n
j=1

    
M j M j 1  ,
 
n = max
j=1,...,n  n n 

and note that

n     
X  j j 1 
Qn n M n M

 n K.
n
j=1

Orthogonality of martingale increments implies that E[M12 ] =

E[Qn ]. Continuity of the paths implies that with probability one
n 0 and hence Qn 0. Since Qn n K K 2 , we can use the
bounded convergence theorem to conclude that

E M12 = lim E[Qn ] = 0.

 
n

If V1 (M ) is not bounded, we can consider the stopping times

TK = inf{t : Vt (M ) = K}.
h i
The argument above gives E M1T 2 = 0 for each K, and
K
hence with probability one for each K, M1TK = 0. But M1 =
limK M1TK .
Using the proposition, we see that if Yt is an increasing adapted
process such that Zt2 Yt is a martingale, then

Mt (Zt2 Yt ) = hZit Yt ,

is a continuous martingale with paths of bounded variation. There-

fore, Yt = hZit .
3.3. ITOS FORMULA 95

3.3 Itos formula

Itos formula is the fundamental theorem of stochastic calculus. Before stat-
ing it, let us review the fundamental theorem of ordinary calculus. Suppose
f is a C 1 function. (We recall that a function if C k if it has k continuous
derivatives.) Then we can expand f in a Taylor approximation,

f (t + s) = f (t) + f 0 (t) s + o(s),

where o(s) denotes a term such that o(s)/s 0 as s 0. We write f as a

telescoping sum:
n     
X j j1
f (1) = f (0) + f f .
n n
j=1

The Taylor approximation gives

       
j j1 0 j1 1 1
f f =f +o .
n n n n n

Hence,
n   n  
X j1 1 X 1
f (1) = f (0) + lim f0 + lim o .
n n n n n
j=1 j=1

The first limit on the right-hand side is the Riemann sum approximation of
the definite integral and the second limit equals zero since n o(1/n) 0.
Therefore, Z 1
f (1) = f (0) + f 0 (t) dt.
0
Itos formula is similar but it requires considering both first and second
derivatives.

Theorem 3.3.1 (Itos formula I). Suppose f is a C 2 function and Bt is a

standard Brownian motion. Then for every t,
Z t
1 t 00
Z
0
f (Bt ) = f (B0 ) + f (Bs ) dBs + f (Bs ) ds.
0 2 0
The theorem is often written in the differential form
1 00
df (Bt ) = f 0 (Bt ) dBt + f (Bt ) dt.
2
96 CHAPTER 3. STOCHASTIC INTEGRATION

In other words, the process Yt = f (Bt ) at time t evolves like a Brown-

ian motion with drift f 00 (Bt )/2 and variance f 0 (Bt )2 . Note that f 0 (Bt ) is a
continuous adapted process so the stochastic integral is well defined.
To see where the formula comes from, let us assume for ease that t = 1.
Let us expand f in a second-order Taylor approximation,
1 00
f (x + y) = f (x) + f 0 (x) y + f (x) y 2 + o(y 2 ),
2
where o(y 2 )/y 2 0 as y 0. Similarly to above, we write
n
X 


f (B1 ) f (B0 ) = f Bj/n f B(j1)/n .
j=1

Using the Taylor approximation, we write

f Bj/n f B(j1)/n

1
= f 0 B(j1)/n j,n + f 00 B(j1)/n 2j,n + o(2j,n ),



2
where
j,n = Bj/n B(j1)/n .
Hence f (B1 ) f (B0 ) is equal to the sum of the following three limits:
n
X
f 0 B(j1)/n

 
lim Bj/n B(j1)/n , (3.9)
n
j=1

n
1 X 00
 2
lim f B(j1)/n Bj/n B(j1)/n , (3.10)
n 2
j=1
n
X  2
lim o( Bj/n B(j1)/n ). (3.11)
n
j=1
 2
The increment of the Brownian motion satisfies Bj/n B(j1)/n 1/n.
Since the sum in (3.11) looks like n terms of smaller order than 1/n the
limit equals zero. The limit in (3.9) is a simple process approximation to a
stochastic integral and hence we see that the limit is
Z 1
f 0 (Bt ) dBt .
0
3.3. ITOS FORMULA 97

If f 00 were constant, say b, then the limit in (3.10) would be

n
b X 2 b b
lim Bj/n B(j1)/n = hBi1 = .
n 2 2 2
j=1

This tells us what the limit should be in general. Let h(t) = f 00 (Bt ) which
is a continuous function. For every  > 0, there exists a step function h (t)
such that |h(t) h (t)| <  for every t. For fixed , we can consider each
interval on which h is constant to see that
n Z 1
X  2
lim h (t) Bj/n B(j1)/n = h (t) dt.
n 0
j=1

Also,
 
 n  n
 [h(t) h (t)] Bj/n B(j1)/n 2  
X    X  2
  Bj/n B(j1)/n .
 j=1  j=1

Therefore, the limit in (3.10) is the same as

1 1 1 1 1 1 00
Z Z Z
lim h (t) dt = h(t) dt = f (Bt ) dt.
0 2 0 2 0 2 0
Example 3.3.1. Let f (x) = x2 . Then f 0 (x) = 2x, f 00 (x) = 2, and hence
Z t
1 t 00
Z Z t
2 2 0
Bt = B0 + f (Bs ) dBs + f (Bs ) ds = 2 Bs dBs + t.
0 2 0 0

This gives the equation

Z t
1 2
Bs dBs = [B t],
0 2 t
which we mentioned in the last section. This example is particularly easy be-
cause f 00 is constant. If f 00 is not constant, then f 00 (Bt ) is a random variable,
and the dt integral is not so easy to compute.
Example 3.3.2. Let f (x) = ex where > 0. Then f 0 (x) = ex =
f (x), f 00 (x) = 2 ex = 2 f (x). Let Xt = f (Bt ) = eBt . Then
1 00 2
dXt = f 0 (Bt ) dBt + f (Bt ) dt = Xt dBt + Xt dt.
2 2
The process Xt is an example of geometric Brownian motion which we dis-
cuss in more detail in the next section.
98 CHAPTER 3. STOCHASTIC INTEGRATION

The derivation of Itos formula given above is essentially correct

but ignores some technical details. For ease assume t = 1 and
assume we have a partition of the form

0 = t0 < t1 < < tn = 1,

with mesh kk = max{tj tj1 }. Then we write

n
X  
f (B1 ) = f (B0 ) + f (Btj ) f (Btj1 ) .
j=1

Taylors theorem implies that

m(j, n)  2
Btj Btj1
2
f (Btj ) f (Btj1 ) f 0 (Btj1 ) Btj Btj1
 

M (j, n)  2
Btj Btj1 ,
2
where m(j, n), M (j, n) denote the minimum and maximum of
f 00 (x) for x on the interval with endpoints Btj1 and Btj . Hence if
we let
Xn
1
f 0 (Btj1 ) Btj Btj1 ,
 
Q () =
j=1

n
X M (j, n)  2
Q2+ () = Btj Btj1 ,
2
j=1

n
X m(j, n)  2
Q2 () = Btj Btj1 ,
2
j=1

we have

Q2 () f (B1 ) f (B0 ) Q1 () Q2+ ().

We now suppose that we have a sequence of partitions n of

the form
0 = t0,n < t1,n < < tkn ,n = 1,
3.3. ITOS FORMULA 99

P
with kn k < . Then we have seen that with probability one,
for all 0 < s < t < 1,
X  2
lim Btj,n Btj1,n = t s.
n
stj,n <t

Using the continuity of Bt and f 00 , we can see that on the event

that this is true, we also have

1 1 00
Z
2 2
lim Q (n ) = lim Q+ (n ) = f (Bs ) dBs .
n n 2 0

We now assume for the moment that there exists K < such
that |f 00 (x)| K for all x. This happens, for example, if f has
compact support. Then

|f 0 (Bs ) f 0 (Btj1,n )| K |Bs Btj1,n |.

(n)
Let At = f 0 (Bt ) and let At be the simple process that equals
f 0 (Btj1 ,n ) for tj1,n t < tj,n . For tj1,n s < tj,n ,

(n)
E([At At ]2 ) K 2 E([Bs Btj1,n ]2 ) = K 2 [stj1,n ] K 2 kn k.

Therefore, Z 1
(n)
E([At At ]2 ) dt K 2 kn k.
0
In particular, we get the following.
00
Proposition 3.3.2. Suppose P that f is bounded and {n } is a
sequence of partitions with kn k < . Let
kn
(n)
X
Yt = B0 + f 0 (Btj1,n ) [Btj,n Btj1,n ]
j=1

kn
X f 00 (Btj1,n )
+ [Btj,n Btj1,n ]2 .
2
j=1

Then with probability one

(n)
lim max |f (Bt ) Yt | = 0.
n 0t1
100 CHAPTER 3. STOCHASTIC INTEGRATION

If f 00 is not bounded, we can still prove Itos formula by using

stopping times. This general procedure is often called localization.
For each K, let

TK = inf{t : |f 00 (Bt )| K}.

Then, our argument shows that with probability one for each K,
Z tTK
1 tTK 00
Z
0
f (BtTK ) = f (Bs ) dBs + f (Bs ) ds.
0 2 0
However, with probability one TK as K , and hence this
gives us a formula for f (Bt ).
Suppose that f is defined only on an interval I = (a, b) and
B0 I. Let

T = inf{t : Bt a +  or Bt b }, T = T0 .

We can apply Itos formula to conclude for all t and all  > 0.
Z tT
1 tT 00
Z
f (BtT ) = f (B0 ) + f 0 (Bs ) dBs + f (s) ds.
0 2 0
This is sometimes written shorthand as
1 00
df (Bt ) = f 0 (Bt ) dBt + f (Bt ) dt, 0 t < T.
2
The general strategy for proving the generalizations of Itos for-
mula that we give in the next couple of sections is the same, and
we will not give the details.

3.4 More versions of Itos formula

We first extend Itos formula to functions that depend on time as well as
position.

Theorem 3.4.1 (Itos Formula II). Suppose f (t, x) is a function that is C 1

in t and C 2 in x. If Bt is a standard Brownian motion, then
Z t
f (t, Bt ) = f (0, B0 ) + x f (s, Bs ) dBs
0
3.4. MORE VERSIONS OF ITOS FORMULA 101
Z t 
1
+ s f (s, Bs ) + xx f (s, Bs ) ds,
0 2
or in differential form,
 
1
df (t, Bt ) = x f (t, Bt ) dBt + t f (t, Bt ) + xx f (t, Bt ) dt.
2

This is derived similarly to the first version except when we expand with
a Taylor polynomial around x we get another term:

f (t + t, x + x) f (t, x) =

1
t f (t, x) t + o(t) + x f (t, x) x + xx f (t, x) (x)2 + o((x)2 ).
2
If we set t = 1/n and write a telescoping sum for f (1, B1 ) f (0, B0 ) we
get terms as (3.9) (3.11) as well as two more terms:
n
X
lim t f ((j 1)/n, B(j1)/n ) (1/n), (3.12)
n
j=1

n
X
lim o(1/n). (3.13)
n
j=1

The limit in (3.13) equals zero, and the sum in (3.12) is a Riemann sum
approximation of a integral and hence the limit is
Z 1
t f (t, Bt ) dt.
0

Example 3.4.1. Let f (t, x) = eat+bx . Then

t f (t, x) = a f (t, x), x f (t, x) = b f (t, x), xx f (t, x) = b2 f (t, x).

Therefore, if Xt = f (t, Bt ) = exp {at + bBt },

 
1
dXt = t f (t, Bt ) + xx f (t, Bt ) dt + x f (t, Bt ) dBt
2
2
 
b
= a+ Xt dt + b Xt dBt .
2
102 CHAPTER 3. STOCHASTIC INTEGRATION

Definition A process Xt is a geometric Brownian motion with drift m and

volatility if it satisfies the SDE

dXt = m Xt dt + Xt dBt = Xt [m dt + dBt ] , (3.14)

where Bt is a standard Brownian motion.

Example 3.4.1 shows that if Bt is a standard Brownian motion and

2
  
Xt = X0 exp m t + Bt , (3.15)
2
then Xt is a geometric Brownian motion with parameters (m, ). Even
though we have an exact expression (3.15) for geometric Brownian motion,
it is generally more useful to think of it in terms of its SDE (3.14).
Geometric Brownian motion is more natural than usual Brownian mo-
tion for modeling prices of assets such as stock. It measures changes in terms
of fractions or percentages of the current price rather than the listed price
per share. In particular, the latter quantity includes a rather arbitrary unit
share which does not appear if one models with geometric Brownian mo-
tion.
The geometric Brownian motion (3.15) is what is sometimes called a
strong solution to the SDE (3.14). (All of the solutions to SDEs that
we discuss in this book are strong solutions.) We will not give the exact
definition, but roughly speaking, if one uses the same Brownian motion Bt
in both places, one gets the same function. Let us explain this in terms
of simulation. Suppose a small t is chosen. Then we can define Bkt by
B0 = 0 and
Bkt = B(k1)t + t Nk , (3.16)
where N1 , N2 , . . . is a sequence of independent N (0, 1) random variables.
Using the same sequence, we could define an approximate solution to (3.14)
by choosing X0 = e0 = 1 and
h i
Xkt = X(k1)t + X(k1)t m t + t Nj . (3.17)

If t is small, this should be very close to

2
  
Ykt = exp m (kt) + Bkt . (3.18)
2
In Exercise 3.9 you are asked to do a simulation to compare (3.17) and
(3.18).
3.4. MORE VERSIONS OF ITOS FORMULA 103

To do formal1 calculations in usual calculus, one writes down differentials

and throws away all terms that are of smaller order than dt. In stochastic
calculus, one can go far computing similarly using the rules

(dBt )2 = dt, (dBt ) (dt) = 0, (dt)2 = 0.

In some of our derivations below, we will use this kind of argument. For
example, a formal derivation of Itos formula II can be given as
1
df (t, Bt ) = t f (t, Bt ) dt + x f (t, Bt ) dBt + xx f (t, Bt )2 (dBt )2
2
+o(dt) + o((dt)(dBt )) + o((dBt )2 ).
By setting (dBt )2 = dt and setting the last three terms equal to zero we get
the formula.
Suppose that Xt satisfies

dXt = Rt dt + At dBt , (3.19)

which we recall is equivalent to

Z t Z t
Xt = X0 + Rs ds + As dBs .
0 0

Whenever we write such equations we will assume implicitly that Rt , At are

adapted processes with piecewise continuous paths. As before, we define the
quadratic variation by
X  j  j1 2
 
hXit = lim X X .
n n n
jtn

As in the case for Brownian motion, the drift term does not contribute to
the quadratic variation,
Z Z t
hXit = h A dBit = A2s ds.
0

This is often written in the differential form

d hXit = A2t dt.

1
Mathematicians use the word formal to refer to calculations that look correct in
form, but for which not all the steps have been justified. When they construct proofs,
they often start with formal calculations and then go back and justify the steps.
104 CHAPTER 3. STOCHASTIC INTEGRATION

To see that Rt does not appear, we can do the formal calculation

(dXt )2 = (Rt dt + At dBt )2
= Rt2 (dt)2 + 2 Rt At (dt) (dBt ) + A2t (dBt )2 = A2t dt.
If Xt satisfies (3.19) and Ht is another adapted process we define
Z t Z t Z t
Hs dXs = Hs Rs ds + Hs As dBs .
0 0 0
If we view this in terms of simulation then
Z t
Yt = Hs dXs
0
can be simulated by
h i
Yt = Ht Xt = Ht [Xt+t Xt ] = Ht Rt t + At t N ,

where N N (0, 1).

Theorem 3.4.2 (Itos formula III). Suppose Xt satisfies (3.19) and f (t, x)
is C 1 in t and C 2 in x. Then
1
df (t, Xt ) = t f (t, Xt ) dt + x f (t, Xt ) dXt + xx f (t, Xt ) dhXit
2
A2t
 
= t f (t, Xt ) + Rt x f (t, Xt ) + xx f (t, Xt ) dt
2
+At x f (t, Xt ) dBt .
Example 3.4.2. Suppose X is a geometric Brownian motion satisfying
dXt = m Xt dt + Xt dBt .
Let f (t, x) = et x3 . Then
t f (t, x) = et x3 = f (t, x),
3 6
x f (t, x) = 3x2 et = f (t, x), xx f (t, x) = 6x et = f (t, x),
x x2
and
1
df (t, Xt ) = t f (t, Xt ) dt + x f (t, Xt ) dXt + xx f (t, Xt ) dhXit
2
2 Xt2
 
= t f (t, Xt ) + m Xt x f (t, Xt ) + xx f (t, Xt ) dt
2
+ Xt x f (t, Xt ) dBt
6 2
 
= 1 + 3m + f (t, Xt ) dt + 3 f (t, Xt ) dBt .
2
3.4. MORE VERSIONS OF ITOS FORMULA 105
  
1
d[et Xt3 ] = 3et Xt3 + m + 2 dt + dBt .
3
Example 3.4.3. The exponential SDE is

dXt = At Xt dBt X0 = x0 .

We claim that the solution is

Z t Z t 
1
Xt = x0 exp As dBs A2s ds . (3.20)
0 2 0

To see this, let Z t Z t

1
Yt = As dBs A2s ds,
0 2 0
which satisfies
A2t
dYt = dt + At dBt , dhY it = A2t dt.
2
If f (x) = x0 ex , then f (x) = f 0 (x) = f 00 (x) and
1
df (Yt ) = f 0 (Yt ) dYt + f 00 (Yt ) dhY it = At Xt dBt .
2
We can compare (3.20) to the solution to the deterministic exponential equa-
tion
f 0 (t) = r(t) f (t) dt, f (0) = x0 ,
which is Z t 
f (t) = x0 exp r(s) ds .
0

Itos formula requires that the function f (t, x) be C 1 in t and C 2 in x.

In applications, one often has functions that are C 2 only for x in an interval
of the reals. In this case, we can use Itos formula until the (random) time
that the process leave the interval.
Theorem 3.4.3 (Itos formula III, local form). Suppose Xt satisfies (3.19)
with a < X0 < b, and f (t, x) is C 1 in t and C 2 in x for a < x < b. Let
T = inf{t : Xt = a or Xt = b}. Then if t < T ,
Z t
f (t, Xt ) = f (0, X0 ) + As x f (s, Xs ) dBs
0
Z t
A2s

+ s f (s, Xs ) + Rs x f (s, Xs ) + xx f (s, Xs ) ds.
0 2
106 CHAPTER 3. STOCHASTIC INTEGRATION

We write the conclusion of the theorem in differential form by

1 00
df (t, Xt ) = f(t, Xt ) dt + f 0 (t, Xt ) dXt + f (t, Xt ) dhXit t < T.
2
Example 3.4.4. Suppose that Bt is a standard Brownian motion, and Yt =
t/Bt2 . Let T = inf{t : Bt = 0}. Then we can use Itos formula to see that for
t < T,
1
dYt = f(t, Bt ) dt + f 0 (t, Bt ) dBt + f 00 (t, Bt ) dt
2
2 4 3
= [Bt + 3tBt ] dt 2tBt dBt .

To prove the local form of Itos formula, we let  > 0 and

T = inf{t : Xt a + 2 or Xt b 2}. We can find a C
function  such that  (x) = 0 if x < a +  or x > b  and such
that (x) = 1 for a + 2 x b 2. Let f (t, x) =  (x) f (t, x).
Then f satisfies the conditions of Itos formula and
Z t
f (t, Xt ) f (0, X0 ) = As f0 (s, Xs ) dBs
0
Z t
A2s 00

0
+ f (s, Xs ) + Rs f (s, Xs ) + f (s, Xs ) ds.
0 2 
If t < T , then both sides of this equation are not changed if we
replace f with f .

3.5 Diffusions
Geometric Brownian motion is an example of a time-homogeneous diffusion.
We say that Xt is a diffusion (process) if it is a solution to an SDE of the
form
dXt = m(t, Xt ) dt + (t, Xt ) dBt , (3.21)
where m(t, x), (t, x) are functions. It is called time-homogeneous if the func-
tions do not depend on t,

dXt = m(Xt ) dt + (Xt ) dBt .

3.5. DIFFUSIONS 107

Diffusion processes are Markov processes. Given all the information in Ft ,

the only important information needed to give the conditional distribution
of Xs , s t is Xt . Simulations of diffusions can be done using the stochastic
Euler rule:
Xt+t = Xt + m(t, Xt ) t + (t, Xt ) t N,
where N N (0, 1). The study of diffusions leads to partial differential
equations.
We define the generator L = L0 of a Markov process Xt by
Ex [f (Xt )] f (x)
Lf (x) = lim .
t0 t
We will use Itos formula to understand the generator of the diffusion Xt .
For ease, we will assume that m and are bounded continuous functions. If
f : R R is a C 2 function with bounded first and second derivatives, then
Itos formula gives
1
df (Xt ) = f 0 (Xt ) dXt + f 00 (Xt ) dhXit
2
2 (t, Xt ) 00
 
= m(t, Xt ) f 0 (Xt ) + f (Xt ) dt
2
0
+f (Xt ) (t, Xt ) dBt .

That is,
Z t
2 (s, Xs ) 00

0
f (Xt ) f (X0 ) = m(s, Xs ) f (Xs ) + f (Xs ) ds
0 2
Z t
+ f 0 (Xs ) (s, Xs ) dBs .
0
The second term on the right-hand side is a martingale (since the integrand
is bounded) and has expectation zero, so the expectation of the right-hand
side is
t E[Yt ],
where Z t
2 (s, Xs ) 00

1 0
Yt = m(s, Xs ) f (Xs ) + f (Xs ) ds.
t 0 2
The fundamental theorem of calculus implies that

2 (0, X0 ) 00
lim Yt = m(0, X0 ) f 0 (X0 ) + f (X0 ).
t0 2
108 CHAPTER 3. STOCHASTIC INTEGRATION

Moreover, the random variables Yt are uniformly bounded which allows us

to take the limit inside the expectation,
Ex [f (Xt )] f (x) 2 (0, x) 00
Lf (x) = lim = m(0, x) f 0 (x) + f (x).
t0 t 2
Similarly, if we define
E[f (Xt+s ) f (Xt ) | Xt = x]
Lt f (x) = lim ,
s0 s
we get
2 (t, x) 00
Lt f (x) = m(t, x) f 0 (x) + f (x).
2
The assumption that m, are bounded is stronger than we would like (for
example, the differential equation (3.14) for geometric Brownian motion does
not satisfy this), but one can ease this by stopping the path before it gets
too large. We will not worry about this now.

In this section we assumed we had a solution to the SDE (3.21).

This leads naturally to the question: given functions m(t, x), (t, x),
can we find a process Xt satisfying (3.21)?
Consider first the deterministic equation

y 0 (t) = F (t, y(t)), y(0) = y0 . (3.22)

We will assume that F is uniformly Lipschitz, that is, there exists

< such that for all s, t, x, y,

|F (s, x) F (t, y)| [|s t| + |x y|] . (3.23)

We can construct a solution to (3.22) valid for 0 t t0 by

using the following method called Picard iteration. We start with
the initial function
y0 (t) = y0 ,
and recursively define
Z t
yk (t) = y0 + F (s, yk1 (s)) ds.
0

Note that
Z t
|y1 (t) y0 (t)| |F (s, y0 )| ds Ct,
0
3.5. DIFFUSIONS 109

where C = C(y0 , t0 ) = max0st0 |F (s, y0 )|, and for k 1,

Z t
|yk+1 (t) yk (t)| |F (s, yk (s)) F (s, yk1 (s))| ds
0
Z t
|yk (s) yk1 (s)| ds
0

Using this estimate and induction, we see that

k C tk+1
|yk+1 (t) yk (t)| , 0 t t0 .
(k + 1)!

In particular, the limit

y(t) = lim yk (t),

k

exists and

X j tj+1
|yk (t) y(t)| C .
(j + 1)!
j=k

If we let Z t
y(t) = y0 + F (s, y(s)) ds,
0
then for each k,
Z t
|y(t) yk+1 (t)| |F (s, y(s)) F (s, yk (s))| ds
0
Z t
|y(s) yk (s)| ds,
0

and hence y = y, and y satisfies (3.22). This argument assumed

that (3.23) holds for all s, t, x, y, and we put no restriction on t0 .
If instead we had such an estimate only for times near zero and
x, y near y0 , we could establish the same result for some (but not
necessarily all) positive values of t0 .
Let us now consider (3.21) where the functions m(t, x) and
(t, x) both satisfy (3.23). For ease, choose t0 = 1 and define
recursively
Xt0 = y0 , 0 t 1,
Z t Z t
(k+1) (k)
Xt = y0 + m(s, Xs ) ds + (s, Xs(k) ) dBs ,
0 0
110 CHAPTER 3. STOCHASTIC INTEGRATION

so that
"Z 2 #
h i t
(k+1) k 2 (k) (k1)
E |Xt Xt | 2E |Xs Xs | ds
0
"Z 2 #
t
+2 E [(s, Xs(k) ) (s, Xs(k1) )] dBs .
0

Using the Holder inequality on the ds integral we see that for t 1,

"Z 2 #
t
E |Xs(k) Xs(k1) | ds
0
 Z t 
2
E t |Xs(k) Xs(k1) |2 ds
0
Z t h i
2 E |Xs(k) Xs(k1) |2 ds.
0

The variance rule for the stochastic integral gives

"Z 2 #
t
(k) (k1)
E [(s, Xs ) (s, Xs )] dBs
0
Z t 
= E [(s, Xs(k) ) (s, Xs(k1) )]2 ds
0
Z t h i
2
E |Xs(k) Xs(k1) |2 ds.
0

Combining these estimates, we see that there exists < such

that for all 0 t 1 and positive integers k,
h i k+1 tk+1
(k+1)
E |Xt Xtk |2 .
(k + 1)!
(k)
We then define Xt = limk Xt ; as before, we do this for dyadic t
as an L2 limit and extend it to other t by continuity. This gives a
solution.
If the Lipschitz condition holds only locally, then we can use
this argument to show that there exists a stopping time T with
P{T > 0} = 1 such that Xt , 0 t < T satisfies the equation.
In the random case the time T for which the solution is valid is a
random variable.
3.6. COVARIATION AND THE PRODUCT RULE 111

3.6 Covariation and the product rule

Suppose Xt , Yt satisfy

dXt = Ht dt + At dBt , dYt = Kt dt + Ct dBt . (3.24)

Here Bt is the same standard Brownian motion in both equations. As before,

the covariation process is defined by
X  j  
j1
   
j

j1

hX, Y it = lim X X Y Y .
n n n n n
jtn

Using the formal rules for stochastic calculus we get

[dXt ] [dYt ] = [Ht dt + At dBt ] [Kt dt + Ct dBt ]

= At Ct dt + O((dt)2 ) + O((dt)(dBt )).

This indeed shows what happens and we get

Z t
hX, Y it = As Cs ds,
0

or in differential form,
dhX, Y it = At Ct dt.
The product rule for the usual derivative can be written in differential
form as
d(f g) = f dg + g df = f g 0 dt + f 0 g dt.
It can be obtained formally by writing

d(f g) = f (t + dt) g(t + dt) f (t) g(t)

= [f (t + dt) f (t)] g(t + dt) + f (t) [g(t + dt) g(t)]
= (df )g + (dg)f + (df )(dg).

Since df = f 0 dt, dg = g 0 dt and df dg = O((dt)2 ), we get the usual product

formula.
If Xt , Yt are process as above, then when we take the stochastic differ-
ential d(Xt Yt ) we get in the same way

d(Xt Yt ) = Xt dYt + Yt dXt + (dXt ) (dYt ).

However, the (dXt ) (dYt ) term does not vanish, but rather equals dhX, Y it .
This gives the stochastic product rule.
112 CHAPTER 3. STOCHASTIC INTEGRATION

Theorem 3.6.1 (Stochastic product rule). If Xt , Yt satisfy (3.24), then

d(Xt Yt ) = Xt dYt + Yt dXt + dhX, Y it .

In other words,
Z t Z t Z t
Xt Yt = X0 Y0 + Xs dYs + Ys dXs + dhXY is
0 0 0
Z t
= X0 Y0 + [Xs Ks + Ys Hs + As Cs ] ds
0
Z t
+ [Xs Cs + Ys As ] dBs .
0

Example 3.6.1. Suppose Bt is a standard Brownian motion and Xt is the

geometric Brownian motion satisfying

dXt = m Xt dt + Xt dBt .

Then in the notation of (3.24), Ht = m Xt and At = Xt . If we set Yt = Bt ,

then Kt = 0, Ct = 1. Therefore,

d(Bt Xt ) = Bt dXt + Xt dBt + dhB, Xit

= Bt [m Xt dt + Xt dBt ] + Xt dBt + Xt dt
= Xt [(m Bt + ) dt + ( Bt + 1) dBt ] .

3.7 Several Brownian motions

Up to this point, we have integrated with respect to a single Brownian
motion. Extending the definitions to several Brownian motions is straight-
forward. Suppose that Bt1 , . . . , Btd are independent Brownian motions with
respect to the filtration {Ft }, and that A1t , . . . , Adt are adapted processes.
We write
d
Ajt dBtj ,
X
dXt = Ht dt +
j=1

to mean
Z t d Z
X t
Xt = X0 + Hs ds + Ajs dBsj .
0 j=1 0

The covariation process can be computed if we remember the rule

hB i , B j i = 0, i 6= j.
3.7. SEVERAL BROWNIAN MOTIONS 113

In particular, if
Z t d Z
X t
Yt = Y0 + Ks ds + Csj dBsj ,
0 j=1 0

then
d
Ajt Ctj dt.
X
dhX, Y it =
j=1

Itos formula for several processes can be obtained by expanding in a Taylor

approximation. As before, we need the first order term in the time variable
and all second order terms in the space variables. We state the theorem; the
proof is essentially the same as for the previous version. The reader may
note that this version includes all the previous versions as special cases, so
we call this the final form. If f : [0, ) Rd R is a function, we write

f(t, x) = t f (t, x), j f (t, x) = xj f (t, x), jk f (t, x) = xj xk f (t, x).

If Xt = (Xt1 , . . . , Xtn ) we write

n
X
f (t, Xt ) dXt = k f (t, Xt ) dXtk .
k=1

Theorem 3.7.1 (Itos formula, final form). Suppose Bt1 , . . . , Btd are inde-
pendent standard Brownian motions, and Xt1 , . . . , Xtn are processes satisfy-
ing
d
Ai,k
X
k k i
dXt = Ht dt + t dBt .
i=1

Suppose f (t, x), t 0, x Rn ,

is a function that is C 1 in t and C 2 in
x = (x1 , . . . , xn ). Then, if Xt = (Xt1 , . . . , Xtn ),

df (t, Xt ) = f(t, Xt ) dt + f (t, Xt ) dXt

n n
1 XX
+ jk f (t, Xt ) dhX j , X k it .
2
j=1 k=1

In other words,
d Z n
" #
X t X
f (t, Xt ) f (0, X0 ) = k f (s, Xs ) Ai,k
s dBsi
i=1 0 k=1
114 CHAPTER 3. STOCHASTIC INTEGRATION

Z t" n
!
X
+ f(s, Xs ) + k f (s, Xs ) Hsk
0 k=1

d n n
1 XXX
+ jk f (s, Xs ) Ai,j
s As
i,k
dt
2
i=1 j=1 k=1

If f (t, x) is C 2 only in an open subset U Rn , we can give a local form

of this theorem as in Theorem 3.4.3 to see that the conclusion holds until
the first time that Xt leaves U . Since it comes up often, we state a particular
case of this theorem. The Laplacian of a function f : Rd R is defined by
d
X
2 f (x) = jj f (x).
j=1

Other standard notations for 2 are and . In the statement below the
gradient and the Laplacian 2 are taken with respect to the x variable
only.

Theorem 3.7.2. Suppose Bt = (Bt1 , . . . , Btd ) is a standard Brownian mo-

tion in Rd . Then if f : [0, ) Rd is C 1 in t and C 2 in x Rd , then
 
1
df (t, Bt ) = f (t, Bt ) dBt + f(t, Bt ) + 2 f (t, Bt ) dt
2

A function f is harmonic if 2 f = 0. There is a close relationship be-

tween Brownian motion, martingales, and harmonic functions that we dis-
cuss in Chapter 8.

3.8 Exercises
Exercise 3.1. Suppose At is a simple process with |At | C for all t. Let
Z t
Zt = As dBs .
0

Show that
E Zt4 3 C 4 t2 .
 

(Hint: Write Zt out in step function form, expand the fourth power, and
use rules for conditional expectation to evaluate the different terms.)
3.8. EXERCISES 115

Exercise 3.2. Use Itos formula to find the stochastic differential f (t, Bt )
where Bt is a standard Brownian motion and

1. f (t, x) = sin x;

2. f (t, x) = et (x/t)2

3. Repeat these exercises for f (t, Xt ) where

dXt = Xt [m dt + dBt ].

Exercise 3.3. Suppose an asset follows the following geometric SDE,

dXt = 4 Xt dt + Xt dBt .

1. Write the exact solution of this equation. In other words, find Xt as a

function of Bt .

2. Suppose X0 = 2. What is the probability that X1 > 8?

3. Suppose X0 = 1. What is the probability that X2 < 6?

4. Suppose X0 = 4.4565. What is the probability that Xt < 0 for some

2 < t < 5?.

Exercise 3.4. Show that if Bt is a standard Brownian motion, At is a simple

process, and Z t
Zt = As dBs ,
0
then
Mt = Zt2 hZit
is a martingale.

Exercise 3.5. Suppose that two assets Xt , Yt follow the SDEs

dXt = Xt [1 dt + 1 dBt ],

dYt = Yt [2 dt + 2 dBt ],
where Bt is a standard Brownian motion. Suppose also that X0 = Y0 = 1.

1. Let Zt = Xt Yt . Give the SDE satisfied by Zt ; in other words write an

expression for dZt in terms of Zt , 1 , 2 , 1 , 1 .
116 CHAPTER 3. STOCHASTIC INTEGRATION

2. Does there exist a function f : R R such that f (Xt ) = Bt for all t?

3. Does there exist a function g : R R such that g(Zt ) = Bt for all t?

Exercise 3.6. Suppose Bt is a standard Brownian motion and Xt satisfies

dXt = Xt2 dt + Xt dBt .

For each of the following find At , Ct such that

dhY it = At dt, dhY, Xit = Ct dt.

1. Yt = Bt .

2. Yt = Xt3 .

3. Z t 
Yt = exp (Xs2 + 1) ds .
0

Exercise 3.7. Consider the geometric SDE

dXt = Xt [ 2 dt + 2 dBt ].

Use the stochastic Eulers method to do Monte Carlo simulations for Xt , 0

t 2 assuming X0 = 1. Use t = .01.

1. Graph at least one of your simulations.

2. Use the simulation to estimate the probability that X2 3. Do at

least 1000 simulations.

3. Compute the exact probability (up to at least three decimal places)

that X2 3 and compare this to your simulations.

Exercise 3.8. Suppose f (t) is a (nonrandom) continuously differentiable

function of t and Bt is a standard Brownian motion. Justify the integration
by parts formula
Z s Z s
f (t) dBt = f (s) Bs f (r) Br Bt f 0 (t) dt.
r r

(Hint: write the left-hand side as a limit of simple process approxima-

tions
R s and do summation by parts on the approximations.) Explain why
r f (t) dBt has a normal distribution.
3.8. EXERCISES 117

Exercise 3.9. Let m = 1, = 2, t = 1/1000 and simulate geometric

Brownian motion
dXt = Xt dt + 2 Xt dBt , X0 = 1,
using both (3.17) and (3.18). Be sure that the same sequence of normal
random variables N1 , N2 , . . . is used for (3.17) and (3.16). Run the simulation
at least 20 times and compare the values of X1/4 , X1/2 , X3/4 , X1 obtained
from the two formulas.
Exercise 3.10. Suppose that Xt satisfies the SDE
dXt = Xt [(1/2) dt + dBt ], X0 = 2
Let
M = max Xt .
0t1

1. Find the density of M .

2. Find the probability that M 4.
Hint: Write Xt = f (Bt ) for an appropriate function.

Exercise 3.11. Suppose Zt , t D are defined and there exists

c < , < , > 0 such that for all s, t D
E [|Zt Zs | ] c |t s|1+ .
Let  = /(2).
Show that
P{s Dn : |Zs+2n Zs | 2n } c 2n/2 .

Show that with probability one, for all n sufficiently large

and all s D,
|Zs+2n Zs | 2n .

Show that with probability one, there exists C such that

for all s, t D,
|Zs Zt | C |s t| .
118 CHAPTER 3. STOCHASTIC INTEGRATION
Chapter 4

More stochastic calculus

4.1 Martingales and local martingales

In the last chapter we defined
Z t
Zt = As dBs ,
0

where Bt is a standard Brownian motion and As is a continuous or piecewise

continuous process. If Z t
E[A2s ] ds <
0
for each t, then Zt is a square integrable martingale. However, our next
example will show that if this inequality does not hold, the stochastic integral
might not be a martingale. Our example is a modification of the martingale
betting strategy from Section 1.2. In that example, we beat a fair game by
doubling our bet until we won. We recall that if Wn denotes the winnings
after n plays, then E[Wn ] = 0, but with probability one W = 1. In order
to beat the game, we had to allow for arbitrarily large numbers of plays of
the game until we won. In continuous time, one can do the same idea, but
adapt it so that we guarantee to win by time 1.
Example 4.1.1. (Martingale betting strategy revisited) We will consider
Z t
Zt = As dBs
0

with a betting strategy As that changes only at times t0 < t1 < t2 < < 1
where
tn = 1 2n .

119
120 CHAPTER 4. MORE STOCHASTIC CALCULUS

We start by setting At = 1 for 0 t < 1/2. Then Z1/2 = B1/2 . Note that

P{Z1/2 1} = P{B1/2 1} = P{B1 2} = 1 ( 2) =: q > 0.

If Z1/2 1, we stop, that is, we set At = 0 for 1/2 t 1. If Z1/2 < 1, let
x = 1 Z1/2 > 0. Define a by the formula

P{a[B3/4 B1/2 ] x} = q.

We set At = a for 1/2 t < 3/4. Note that we only need to know Z1/2 to
determine a and hence a is F1/2 -measurable. Also, if a[B3/4 B1/2 ] x,
then Z 3/4
Z3/4 = As dBs = Z1/2 + a[B3/4 B1/2 ] 1.
0
Therefore,
P{Z3/4 1 | Z1/2 < 1} = q,
and hence
P{Z3/4 < 1} = (1 q)2 .
If Z3/4 1, we set At = 0 for 3/4 t 1. Otherwise, we proceed as above.
At each time tn we adjust the bet so that

P{Ztn+1 1 | Ztn < 1} = q,

and hence
P{Ztn < 1} (1 q)n .
Using this strategy, with probability one Z1 1, and hence, E[Z1 ] 1.
Therefore, Zt is not a martingale. Our choice of strategy used discontinu-
ous bets, but it is not difficult to adapt this example so that t 7 At is a
continuous function except at the one time at which the bet changes to zero.

Although the stochastic integral may not be a martingale, it is almost a

martingale in the sense that one needs to make the bets arbitrarily large to
find a way to change the expectation. The next definition makes precise the
idea of a process that is a martingale if it is stopped before the values get
big.

Definition A continuous process Mt adapted to the filtration {Ft } is called

a local martingale on [0, T ) if there exists an increasing sequence of stopping
times
1 2 3
4.1. MARTINGALES AND LOCAL MARTINGALES 121

such that with probability one

lim j = T,
j

and for each j,

(j)
Mt = Mtj ,
is a martingale.

In the case of the stochastic integral

Z t
Zt = As dBs ,
0

we let {j } be the stopping times,

 Z t 
2
j = inf t : hZit = As ds = j .
0

(j)
Then for each j, Mt is a square integrable martingale. Therefore, Zt is a
local martingale on [0, T ) where
 Z t 
2
T = inf t : As ds = .
0

The stochastic integral Zt is not defined after time T . Note that if s 7 As

is continuous for all s [0, ), then T = .
More generally, suppose that

dXt = Rt dt + At dBt .

If Rt 6= 0, then Xt is not a martingale. In other words, if Xt is to be a

martingale we need Rt 0. However, as we have just shown, it is possible
for Rt 0 but Xt to not be a martingale. It is a local martingale in this
case. In an abuse of terminology, it is standard to refer to At dBt as the
martingale part of dXt even though one should say the local martingale
part.
For continuous martingales, we have the optional sampling theorem
which states under certain circumstances that one cannot beat a fair game.
The next theorem includes two versions which are the most useful for appli-
cations.
Theorem 4.1.1 (Optional Sampling Theorem). Suppose Zt is a continuous
martingale and T is a stopping time, both with respect to the filtration {Ft }.
122 CHAPTER 4. MORE STOCHASTIC CALCULUS

If Mt = ZtT , then Mt is a continuous martingale with respect to {Ft }.

In particular, E[ZtT ] = E[Z0 ].
2 ] C. Then if
Suppose there exists C < such that for all t, E[ZtT
P{T < } = 1, E[ZT ] = E[Z0 ].

Example 4.1.2. Suppose Zt is a continuous martingale with Z0 = 0. Sup-

pose that a, b > 0 and let

T = inf{t : Zt = a or Zt = b}.

Suppose that P{T < } = 1, which happens, for example, if Zt is a standard

Brownian motion. Then ZtT is a bounded martingale and hence

0 = E[Z0 ] = E[ZT ] = a P{ZT = a} + b P{ZT = b}.

By solving, we get
a
P{ZT = b} = ,
a+b
which is the gamblers ruin estimate for continuous martingales.

Results about continuous martingales can be deduced from cor-

responding results about discrete-time martingales. If Zt is a con-
tinuous martingale and
 
k
Dn = : k = 0, 1, . . . ,
2n

then Zt , t Dn is a discrete-time martingale. If Tn is a stopping

time taking values in Dn , then ZtTn is also a martingale.
For more general T , we define Tn by
k k1 k
Tn = if T < n, k = 1, 2, . . . , n2n ,
2n 2n 2
Tn = n if T n,
Suppose s < t and let sn , tn denote the smallest element of Dn
greater than or equal to s, t, respectively. If A is an Fs -measurable
event, then it is also Fsn measurable, and since ZrTn , r Dn is a
discrete-time martingale,

E [1A Zsn Tn ] = E [1A Ztn Tn ] .

4.1. MARTINGALES AND LOCAL MARTINGALES 123

Since Z has continuous paths, we know that as n ,

Zsn Tn ZsT , Ztn Tn ZtT .
In general, if Xn is a sequence of random variables converging
to X, we cannot conclude that E[Xn ] E[X]. However, this will
hold if the sequence {Xn } is uniformly integrable, that is, if for
every  > 0, there exists K < such that for all n,
E [|Xn | 1{|Xn | K}] < .
(We leave this as an exercise or the reader can check a book on
measure-theoretic probability.)
Lemma 4.1.2. For every positive integer m, the collection of
random variables {ZTn } where n ranges over all positive integers,
T ranges over all stopping times with T m, and Tn is defined
as above, is uniformly integrable.
Proof. Let T be a stopping time with T m and n a positive
integer. Note that Nt = |Zt |, t Dn is a submartingale. Let
K > 0 and let n = n,K be the first t Dn such that Nt K.
If t Dn , let At = At,n be the event {n = t}. Since Nt is a
discrete-time submartingale, if t m,
E [Nm 1At ] E [NTn 1At ] .
By summing over t Dn , t m, we see that
E [NTn Jn,K ] E [Nm Jn,K ] .
where Jn,K = 1{n m}. As n . the random variables Jn,K
converge monotonically to
 
J,K = 1 max Ns > K .
0st

Therefore, for each K < ,

E [NTn 1{NTn > K}] E [NmTn J,K ] E [|Zm | J,K ] .
Note that the right-hand side is independent of T and n. As
K , J,K 0, and since Zm is integrable,
lim E [|Zm | J,K ] = 0.
K
124 CHAPTER 4. MORE STOCHASTIC CALCULUS

Using the lemma, we now conclude

E [1A ZsT ] = E [1A ZtT ] .

A sufficient condition for uniform integrability of {Xn } is the exis-

tence of C < such that E[Xn2 ] C for all n. Indeed,

E [|Xn | 1{|Xn | K}] K 1 E |Xn |2 1{|Xn | K} C/K.

 

2 ] C for all t and P{T < } = 1, then

In particular, if E[ZtT
ZnT ZT with probability one and uniform integrability implies
that
E[ZT ] = lim E[ZnT ] = lim E[Z0 ] = E[Z0 ].
n n

Two other theorems that extend almost immediately to the con-

tinuous martingale setting are the following. The proofs are essen-
tially the same as for their discrete counterparts.

Theorem 4.1.3. (Martingale Convergence Theorem) Suppose

Zt is a continuous martingale and there exists C < such that
E [|Zt |] C for all t. Then there exists a random variable Z
such that with probability one

lim Zt = Z .
t

Theorem 4.1.4. (Maximal inequality). Suppose Zt is a contin-

uous square integrable martingale, and let

Nt = max |Zt |.
0st

Then for every a > 0,

E[Zt2 ]
P{Nt a} .
a2

4.2 An example: the Bessel process

The Bessel process with parameter a is the solution to the SDE
a
dXt = dt + dBt , X0 = x0 > 0.
Xt
4.2. AN EXAMPLE: THE BESSEL PROCESS 125

If we choose a = (d1)/2, this is called the Bessel-d process and is related to

d-dimensional Brownian motion (see Exercise 4.4). Let T = inf{t : Xt }.
There is no problem finding a solution to this equation for t < T , and hence
it is well defined for t < T where

T = T0 = inf{t : Xt = 0}.

At time T the equation is ill-defined so we will not consider the process for
t > T . If a > 0, then when Xt gets close to 0, there is a strong drift away
from the origin. It is not obvious whether or not this is strong enough to
keep the diffusion from reaching the origin.
Suppose that 0 < r < x < R < and let (x) denote the probability
that the process starting at x reaches level R before reachng level r. In other
words, if
= (r, R) = inf{t : Xt = r or Xt = R},
then
(x) = P{X = R | X0 = x}.
We will use Itos formula to find a differential equation for and use this
to compute . Note that (r) = 0, (R) = 1. Let J denote the indicator
function of the event {X = R} and let

Mt = E[J | Ft ].

The tower property for conditional expectation implies that Mt is a martin-

gale; indeed, if s < t,

E[Mt | Fs ] = E[ E(J | Ft ) | Fs ] = E[J | Fs ] = Ms .

The Markovian nature of the diffusion X implies that

E[J | Ft ] = (Xt ).

In other words, if t, then we already know whether or not {X = R}.

However, if > t, the only useful information in Ft for predicting if X = R
is Xt , and the conditional probability is the probability that this occurs
given that the process started at Xt . The upshot of this reasoning is that
(Xt ) must be a martingale. Itos formula (see Theorem 3.4.3) gives for
t < T,
1
d(Xt ) = 0 (Xt ) dXt + 00 (Xt ) dhXit
 0 2
00 (Xt )

a (Xt )
= + dt + 0 (Xt ) dBt .
Xt 2
126 CHAPTER 4. MORE STOCHASTIC CALCULUS

If this is to be a martingale, the dt term must vanish at all times. The way
to guarantee this is to choose to satisfy the ordinary differential equation
(ODE)
x 00 (x) + 2a 0 (x) = 0.
Solving such equations is standard (and this one is particularly easy for one
can solve the first-order equation for g(x) = 0 (x)), and the solutions are of
the form
1
(x) = c1 + c2 x12a , a 6= ,
2
1
(x) = c1 + c2 log x, a= ,
2
where c1 , c2 are arbitrary constants. The boundary conditions (r) =
0, (R) = 1 determine the constants giving

x12a r12a 1
(x) = , a 6= , (4.1)
R12a r12a 2

log x log r 1
(x) = , a= . (4.2)
log R log r 2
We now answer the question that we posed.

Proposition 4.2.1. If a 1/2, then P{T = } = 1, that is, with

probability one the Bessel process never reaches zero. If a < 1/2, then
P{T < } = 1.

Proof. Assume X0 = x < R and let (r, R) be defined as above. If T < ,

then there must be R < such that X (r,R) = r for all r > 0. Using the
form of the probability in (4.1) and (4.2), we can see that

0 if a 1/2
lim P{X (r,R) = r} =
r0 1 (x/R)12a if a < 1/2

The alert reader will note that we cheated a little bit in our
derivation of because we assumed that was C 2 . After assuming
this, we obtained a differential equation and found what should
be. To finish a proof, we can start with as defined in (4.1) or
4.3. FEYNMAN-KAC FORMULA 127

(4.2), and use Itos formula to show that Mt = (Xt ) is a con-

tinuous martingale. Since it is also bounded, we can use the optional
sampling theorem to conclude that

(x0 ) = E[M0 ] = E[M ] = P{X = R}.

This kind of argument is often done in stochastic calculus where

one first assumes sufficient smoothness of a a function in order to
use Itos formula and obtain a differential equation for the function.
Once the differential equation is solved, one can go back and redo
the Itos formula calculation rigorously.

4.3 Feynman-Kac formula

Suppose that the price of a stock follows a geometric Brownian motion

dXt = m Xt dt + Xt dBt . (4.3)

Suppose that at some future time T we have an option to buy a share of the
stock at price S. Since we will exercise the option only if XT S, the value
of the option at time T is F (XT ) where

F (x) = (x S)+ = max {x S, 0} .

Suppose that there is an inflation rate of r so that x dollars at time t in

future is worth only ert x in current dollars. Let (t, x) be the expected
value of this option at time t, measured in dollars at time t, given that the
current price of the stock is x,
h i
(t, x) = E er(T t) F (XT ) | Xt = x . (4.4)

The Feynman-Kac formula gives a PDE for this quantity.

Since it is not any harder, we generalize and assume that Xt satisfies the
SDE
dXt = m(t, Xt ) dt + (t, Xt ) dBt , X0 = x0 ,
and that there is a payoff F (XT ) at some future time T . Suppose also that
there is an inflation rate r(t, x) so that if Rt denotes the value at time t of
R0 dollars at time 0,
dRt = r(t, Xt ) Rt dt,
128 CHAPTER 4. MORE STOCHASTIC CALCULUS
Z t 
Rt = R0 exp r(s, Xs ) ds .
0
If (t, x) denote the expected value of the payoff in time t dollars given
Xt = x, then
  Z T  
(t, x) = E exp r(s, Xs ) ds F (XT ) | Xt = x . (4.5)
t

We will use Itos formula to derive a PDE for under the assumption that
is C 1 in t and C 2 in x.
Let
Mt = E RT1 F (XT ) | Ft .
 

The tower property for conditional expectation implies that if s < t,

E[Mt | Fs ] = E[E(MT | Ft ) | Fs ] = E[MT | Fs ] = Ms .

In other words, Mt is a martingale. Since Rt is Ft -measurable, we see that

  Z T  
1
Mt = Rt E exp r(s, Xs ) ds F (XT ) | Ft .
t

However, since Xt is a Markov process, the only relevant information in Ft

is given by the value Xt . Hence, by the definition of ,

Mt = Rt1 (t, Xt ). (4.6)

We have shown the following.

If Mt is defined by (4.6), then Mt is a martingale.
We will use this fact to derive a partial differential equation for . If we
assume that is C 1 in t and C 2 in x, then Itos formula gives
1
d(t, Xt ) = t (t, Xt ) dt + x (t, Xt ) dXt + xx (t, Xt ) dhXit .
2
In particular,
d(t, Xt ) = Ht dt + At dBt ,
with
1
Ht = t (t, Xt ) + m(t, Xt ) x (t, Xt ) + (t, Xt )2 xx (t, Xt ),
2
At = (t, Xt ) x (t, Xt ).
4.3. FEYNMAN-KAC FORMULA 129

Since hRit = 0, the stochastic product rule implies that

d[Rt1 (t, Xt )] = Rt1 d(t, Xt ) + (t, Xt ) d[Rt1 ],

and hence the dt term of d[Rt1 (t, Xt )] is Rt1 times

r(t, Xt ) (t, Xt ) + t (t, Xt ) + m(t, Xt ) x (t, Xt )

1
+ (t, Xt )2 xx (t, Xt ).
2
Since Mt is a martingale, the dt term must be identically zero. This happens
only if for all t, x,

r(t, x) (t, x) + t (t, x) + m(t, x) x (t, x)

1
+ (t, x)2 xx (t, x) = 0.
2
We have derived the following.

Theorem 4.3.1 (Feynman-Kac formula). Suppose Xt satisfies

dXt = m(t, Xt ) dt + (t, Xt ) dBt , X0 = x0 ,

and r(t, x) 0 is a discounting rate. Suppose a payoff F at time T is given

with E[|F (XT )|] < . If (t, x), 0 t T is defined as in (4.5), and (t, x)
is C 1 in t and C 2 in x, then (t, x) satisfies the PDE

t (t, x) = m(t, x) x (t, x)

1
(t, x)2 xx (t, x) + r(t, x) (t, x),
2
for 0 t < T, with terminal condition (T, x) = F (x).

One of our assumptions in Theorem 4.3.1 is that is sufficiently

differentiable. One can give conditions on the coefficients and payoff
function F under which this holds, and they include the examples
that we give, but the general theory is beyond what we will discuss
in this book. See I. Karatzas and S. Shreve, Brownian Motion and
Stochastic Calculus, 2nd. ed, Springer (1991), Section 5.7.B for a
discussion of this.
130 CHAPTER 4. MORE STOCHASTIC CALCULUS

Example 4.3.1. Suppose Xt satisfies (4.3). Then m(t, x) = mx, (t, x) =

x and is as given in (4.4). The Feynman-Kac formula gives
2 x2
t = r mx x xx . (4.7)
2
This is a version of the Black-Scholes PDE. We will return to this later.
Let us give a somewhat different derivation of the Feynman-Kac formula
using the generator. Again, suppose that Xt satisfies

dXt = m(t, Xt ) dt + (t, Xt ) dBt ,

and that F is a function that does not grow too quickly. Let

f (t, x) = E [F (XT ) | Xt = x] .

Let r(t) 0 be a discount rate which for ease we will assume is a determin-
istic function of time, and
Z t 
Rt = R0 exp r(s) ds .
0

Let
 Z T 
(t, x) = E [(Rt /RT ) F (XT ) | Xt = x] = exp r(s) ds f (t, x)
t

Recall that
t f (t, x) = Ltx f (t, x),
where Lt is the generator
(t, x) 00
Lt h(x) = m(t, x) h0 (x) + h (x).
2
Therefore,
 Z T 
t (t, x) = r(t) exp r(s) ds f (t, x)
t
 Z T 
exp r(s) ds Ltx f (t, x)
t
= r(t) (t, x) Ltx (t, x)
2 (t, x)
= r(t) (t, x) m(t, x) x (t, x) xx (t.x),
2
which is the Feynman-Kac PDE.
4.4. BINOMIAL APPROXIMATIONS 131

4.4 Binomial approximations

In the next few sections we will be considering the SDE

dXt = m(t, Xt ) dt + (t, Xt ) dBt .

We will give some heuristic derivations based on discrete approximations.

Recall that the Euler method to sample from this equation uses the rule

X(t + t) = X(t) + m(t, X(t)) t + (t, X(t)) t Z

where Z N (0, 1). Here we wish to consider some binomial sampling

schemes, that is, schemes such that given X(t), X(t + t) takes one of
two values.
If Xt is Brownian motion with drift zero and variance 2 ,

dXt = dBt ,

then the binomial scheme is approximation by random walk,

P{X(t + t) X(t) = t | X(t)}
1
= P{X(t + t) X(t) = t | X(t)} = .
2
In this case the values of X(kt) lie in the lattice of points
n o
. . . , 2 t, t, 0, t, 2 t, . . . .

If m 0 and (t, x) is not constant, we can change this rule to

1
P{X(t + t) X(t) = (t, X(t)) t | X(t)} = ,
2
but then the values of X(t) do not stay in a lattice.
Suppose is constant, but m varies,

dXt = m(t, Xt ) dt + dBt .

There are two reasonable ways to incorporate drift in this binomial model.
One is to use a version of the Eulers method and set
1
P{X(t + t) X(t) = m(t, Xt ) t t | X(t)} = ,
2
132 CHAPTER 4. MORE STOCHASTIC CALCULUS

for which it is immediate that

E [X(t + t) X(t) | Ft ]

= E [X(t + t) X(t) | X(t)] = m(t, Xt ) t. (4.8)

This method has the disadvantage that we do not stay on a lattice of points.
Another approach is to always make the jumps t but to change the
probabilities based on the drift. In particular, we let
m(t, Xt )
 
1
P{X(t + t) X(t) = t | X(t)} = 1+ t ,
2
m(t, Xt )
 
1
P{X(t + t) X(t) = t | X(t)} = 1 t .
2
The probabilities are chosen so that (4.8) holds for this scheme as well. These
two binomial methods illustrate the two main ways to incorporate a drift
term (in other words, to turn a fair game into an unfair game):
Choose an increment with mean zero and then add a small amount to
it.

Change the probabilities of the increment so that it does not have

mean zero.
Example 4.4.1. We will use the latter rule to sample from Brownian motion
with constant drift m 6= 0 and 2 = 1. Our procedure for simulating uses
1 h i
P{X(t + t) X(t) = t | X(t)} = 1 m t . (4.9)
2
Suppose that t = 1/N for some large integer N , and let us consider the
distribution of X(1). There are 2N possible paths that the discrete process
can take which can be represented by

= (a1 , a2 , . . . , aN )

where aj = +1 or 1 depending on whether that step goes up or down.Let

J = J() denote the number of +1s, and define r by J = (N/2) + r N .
Note that the position at time 1 is

X(1) = t [a + + aN ]
1
= J t (N J) t

= 2r N t = 2r.
4.4. BINOMIAL APPROXIMATIONS 133

For each , the probability of its occurring is

 N h
1 iJ h iN J
q() = 1 + m t 1 m t .
2

If m were 0, then the probability would be be just (1/2)N . Therefore, the

ratio of the probabilities for m 6= 0 to m = 0 is

m J m N J
   
1+ 1
N N
N/2    
m2 m r N m r N

= 1 1+ 1 .
N N N

N
Using the relation 1 + Na ea , we see that the limit of the right-hand
side as N is
2 2
em /2 e2rm = emX(1) em /2 .
Given this, we see that in order to sample from a Brownian motion with
drift, we could first sample from a Brownian motion without drift and then
2
weight the samples by the factor emX(1) em /2 . We will show how to do this
directly in Section 5.2.
As one more application of binomial approximations, we will give a
heuristic argument for the following theorem.

Theorem 4.4.1. Suppose that

dXt = m(Xt ) dt + dBt ,

where m is continuously differentiable, and let p(t, x) denote the density of

Xt at time t. Then
t p(t, x) = Lx p(t, x),
where

2 00
L f (x) = [m(x) f (x)]0 + f (x)
2
2 00
= m0 (x) f (x) m(x) f 0 (x) + f (x).
2

134 CHAPTER 4. MORE STOCHASTIC CALCULUS

Recall that
2 00
Lf (x) = m(x) f 0 (x) + f (x).
2
If m is constant, then as we saw before, one obtains L from L by just
changing the sign of the drift. For varying m we get another term. We will
derive the expression for L by using the binomial approximation
m(Xt )
 
1
P{X(t + t) X(t) = t | X(t)} = 1 t .
2

Let  = t, 2 = t. To be at position x = k at time t + 2 , one must be
at position x  at time t. This gives the relation
 
2 1 m(x ) 
p(t +  , x) = p(t, x ) 1+
2
 
1 m(x + )
+p(t, x + ) 1  . (4.10)
2
We know that

p(t + 2 , x) = p(t, x) + t p(t, x) 2 + o(2 ). (4.11)

The definition of the derivative (see Section 2.10.1), implies that

2 2
p(t, x + ) + p(t, x ) = p(t, x) + xx p(t, x) + o(2 ).
2
p(t, x ) = p(t, x) x p(t, x)  + o(),
m(x ) = m(x) m0 (x)  + o(),
Plugging in, we see that the right-hand side of (4.10) equals
 2 
2 0
p(t, x) +  xx p(t, x) m (x) p(t, x) m(x) x p(t, x) + o(2 ).
2

Comparing this with (4.11) gives us the result.

One can also derive this result by considering L as the adjoint

of L. In other words, L is defined by
Z Z

L f (x) g(x) dx = f (x) Lg(x) dx.

4.5. CONTINUOUS MARTINGALES 135

If
2 00
Lg(x) = m(x) g 0 (x) + g (x),
2
then integration by parts gives

2 00
L f (x) = [m(x) f (x)]0 + f (x).
2

4.5 Continuous martingales

Earlier we made the comment that Brownian motion is the only continuous
martingale. We will make this explicit in this section.

Proposition 4.5.1. Suppose that Mt is a continuous martingale with respect

to a filtration {Ft } with M0 = 0, and suppose that the quadratic variation
of Mt is the same as that of standard Brownian motion,
X  j + 1  2
j
hM it = lim M n
M = t.
n
n
2 2n
j<2 t

Then every R,
2 t/2
E [exp{iMt }] = e .

Sketch of proof. Fix and let f (x) = eix . Note that the derivatives of f are
uniformly bounded in x. Following the proof of Itos formula we can show
that
1 t 00 2 t
Z Z
f (Mt ) f (M0 ) = Nt + f (Ms ) ds = Nt f (Ms ) ds,
2 0 2 0
where Nt is a martingale. In particular, if r < t,
Z t
2 t
 Z
1 00
E[f (Mt ) f (Mr )] = E f (Ms ) ds = E[f (Ms )] ds.
2 r 2 r

If we let G(t) = E[f (Mt )], we get the equation

2
G0 (t) = G(t), G(0) = 1,
2
2 t/2
which has solution G(t) = e .
136 CHAPTER 4. MORE STOCHASTIC CALCULUS

Theorem 4.5.2. Suppose that Mt is a continuous martingale with respect

to a filtration {Ft } with M0 = 0, and suppose that the quadratic variation
of Mt is the same as that of standard Brownian motion
X  j + 1  2
j
hM it = lim M n
M n
= t.
n
n
2 2
j<2 t

Then Mt is a standard Brownian motion.

Proof. We need to show that Mt satisfies the conditions of a standard Brow-
nian motion. We are given M0 = 0 and that t 7 Mt is continuous, so we
only need to establish independent, normal increments. One way to express
both of these conditions is to say that for every s < t, the conditional distri-
bution of Mt Ms given Fs is N (0, t s). But the proposition above shows
that for every ,
  2
E ei(Mt Ms ) | Fs = e (ts)/2 .

Since the characteristic function determines the distribution, we are finished.

This theorem can be extended to say that any continuous mar-

tingale is a time change of a Brownian motion. To be precise, sup-
pose that Mt is a continuous martingale. The quadratic variation
can be defined as the process hM it such that Mt2 hM it is a local
martingale. Let
s = inf{t : hM it = s}.
Then Ys = Ms is a continuous martingale whose quadratic varia-
tion is the same as standard Brownian motion. Hence Ys is a stan-
dard Brownian motion.

4.6 Exercises
Exercise 4.1. A process Xt satisfies the Ornstein-Uhlenbeck SDE if

dXt = mXt dt + dBt ,

where m R and Bt is a standard Brownian motion. Suppose that X0 = 1,

R > 1 and T = TR is the first time that Xt = R or Xt = 0.
4.6. EXERCISES 137

1. Find a function F with F (0) = 0 and F (x) > 0 for x > 0 such that
F (XtT ) is a martingale. (You may leave your answer in terms of a
definite integral.)

2. Find the probability that XT = R. You can write the answer in terms
of the function F .

3. For what values of m is true that

lim P{XT = R} = 0 ?
R

Exercise 4.2. Suppose Bt is a standard Brownian motion and Xt satisfies

the SDE
dXt = a cot(Xt ) dt + dBt , X0 = x0 (0, /2),
where a R. For  > 0, let

T = min{t : sin(Xt ) = sin },

T = T0+ = min {t : Xt {0, }} = min{t : sin(Xt ) = 0},

We will only consider the process Xt for t T . Let

= min{t : Xt = /2}.

1. Find a function F (x) that is positive for 0 < x < /2 with F (/2) = 0
and such that Mt = F (XtT ) is a local martingale for t < T . (It suffices
to write F in the form
Z /2
F (x) = g(y) dy,
x

where g is an explicit function.)

2. For which values of a is

F (0+) = lim F () < ?

0
138 CHAPTER 4. MORE STOCHASTIC CALCULUS

3. Assume that 0 <  < x0 < /2 and let

q(x0 , ) = P{T < }.

Write q(x0 , ) explicitly in terms of F .

4. For whch values of a is it the case that

lim q(x0 , ) = 0 ?
0

Exercise 4.3. Suppose Bt is a standard Brownian motion and let Xt =

eBt + 2. Let
h i
(t, x) = E e2(T t) (XT3 1) eXT | Xt = x , 0 < t < T.

Use the Feynman-Kac theorem to find a second-order PDE that satisfies.

Exercise 4.4. Suppose B11 , . . . , Btd are independent standard Brownian mo-
tions and let q
Xt = (Bt1 )2 + + (Btd )2 .

1. Use Itos formula to show that

a
dXt = dt + dMt ,
Xt
where a = (d 1)/2 and Mt is a continuous martingale satisfying
d
Ajt dBtj ,
X
dMt =
j=1

for suitable processes Ajt .

2. Show that hM it = t.

3. Explain why Mt is a standard Brownian motion. (Hint: use Proposition

4.5.1.)
Chapter 5

Change of measure and

Girsanov theorem

5.1 Absolutely continuous measures

In order to state the Girsanov theorem in the next section, we will need
to discuss absolute continuity and singularity of measures. For measures on
discrete (countable) spaces, this is an elementary idea, but it is more subtle
for uncountable spaces. We assume that we have a probability space (, F)
where is the space of outcomes (sometimes called the sample space) and
F is the collection of events. An event is a subset of , and we assume that
the set of events forms a -algebra (see Section 1.1). A (positive) measure is
a function : F [0, ] such that () = 0 and if A1 , A2 , . . . are disjoint,

!
[ X
An = (An ).
n=1 n=1

The measure is a probability (measure) if () = 1.

Definition Suppose , are measures on (, F).

is absolutely continuous with respect to , written  , if for every
E F, if (E) = 0, then (E) = 0.

and are mutually absolutely continuous or equivalent measures if

 and  .

and are singular measures, written , if there exists an event

E F with (E) = 0 and ( \ E) = 0.

139
140CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

Example 5.1.1. Suppose is a countable set and F is the collection of all

subsets of . If p : [0, ) is a function, then there is a corresponding
measure defined by X
(E) = p().
E
Suppose is another measure given by the function q. Let

A = { : p() > 0}, A = { : q() > 0}.

Then  if and only if A A , and

d
q() = () p(), ,
d
where d/d is defined on A by
d q()
() = .
d p()
Note that for any event E,
X d
(E) = () p().
d
E

In this case, and are equivalent if A = A , and if A A = .

Example 5.1.2. Suppose denotes Lebesgue measure on R,
Z Z
(A) = d = dx = length(A),
A A

and X is a continuous random variable with density f . Let PX denote the

distribution of X, that is, the probability measure on R given by
Z Z
PX (A) = P{X A} = f (x) dx = f d.
A A

Then PX  and we can write

Z
dPX dPX
PX (A) = d where = f.
A d d
If f (x) > 0 for all x, then PX (A) > 0 whenever (A) > 0 and hence  PX .
If Y is another continuous random variable with density g, let

AX = {x : f (x) > 0}, AY = {y : g(y) > 0}.

5.1. ABSOLUTELY CONTINUOUS MEASURES 141

If AY AX , then PY  PX and we can write

Z Z Z
g dPY
P{Y A} = g d = f d = dPX ,
A A f A dP X

where
dPY g
= .
dPX f
If AX AY = , then PX PY .

Example 5.1.3. Suppose X is a discrete random variable taking values in

the countable set A and Y is a continuous random variable with density g.
If PX , PY denote the distributions, then PX PY . Indeed,

PX (R \ A) = 0, PY (A) = 0.

These examples demonstrate the following theorem whose proof can be

found in any book on measure theory. A measure is -finite if we can write

[
= An ,
n=1

where (An ) < for each n.

Theorem 5.1.1 (Radon-Nikodym Theorem). Suppose , are -finite mea-

sures on (, F) with  . Then there exists a function f such that for
every E, Z
(E) = f d. (5.1)
E

The function f is called the Radon-Nikodym derivative of with respect

to and is denoted
d
f= .
d
Roughly speaking, the measure of a point x is (d/d)(x) times the
measure of x. This interpretation is precise in the case of discrete measures,
such as in Example 5.1.1, but does not make precise sense when the point x
gets zero measure in both measures. Example 5.1.2 shows that this subtlety
is something we are already familiar with. If denotes length, then the
density f = dPX /d of a continuous random variable can be interpreted as
saying that the probability of obtaining a value in [x, x + dx] is f (x) times
dx, the length of [x, x + dx].
142CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

If (, F, P ) is a probability space, and Q is a probability measure with

Q  P , then the Radon-Nikodym derivative
dQ
X=
dP
is a nonnegative random variable with E[X] = 1. If E is an event, then (5.1)
can be written as
Q(E) = EP [X 1E ] . (5.2)
Here EP denotes expectation using the probability measure P . (Up to now,
we have been considering a single probability measure P on a space and
using E to denote expectation with respect to it. Since we are now going to
consider different measures on the same space, we adopt the notation EP
to denote expectations using the measure P .) It is not hard to extend the
relation (5.2) to give  
dQ
EQ [Y ] = EP Y .
dP
Example 5.1.4. Suppose (, F, P) is a probability space, and G F is a
sub -algebra. As before, we think of G as partial information. Let X be
a nonnegative, integrable F-measurable random variable. Then

Q(A) = E [1A X] , A G,

defines a measure on (, G) that satisfies Q  P. Therefore, there exists a

G-measurable random variable Y such that for all A G,

Q(A) = E [1A Y ] .

This random variable Y is the conditional expectation E(X | G) as defined

in Section 1.1.
Example 5.1.5. Let the set of outcomes be the set of continuous func-
tions f : [0, 1] R. If Bt denotes a Brownian motion with drift zero and
variance 2 , there is a measure P on called Wiener measure (with vari-
ance 2 ). We think of P as the distribution of the function-valued random
variable t 7 Bt . If V is a subset of , then P (V ) is the probability that the
function t 7 Bt lies in V . We claim that if 6= 0 , then P P0 . In order
to show this, we need to find an event E such that P (E) = 1, P0 (E) = 0.
Let Er denote the set of functions f such that
2n  
j1 2
  
X j
lim f f = r2 .
n 2n 2n
j=1
5.2. GIVING DRIFT TO A BROWNIAN MOTION 143

Using what we know about the quadratic variation, we see that

P (E ) = 1, P0 (E0 ) = 1.

Let denote the set of continuous functions f : [0, 1] C

which is a Banach space (complete, normed, metric space) under
the norm
kf k = max{|f (s)| : 0 s 1}.

Let F denote the corresponding Borel -algebra, that is, the small-
est -algebra under which all the open sets under k k are measur-
able. The measures P are defined on (, F). It is easy to check
that the functions s (f ) = f (s) are measurable functions on this
space, and hence so are the functions

2n  
j1 2
  
X j
n (f ) = f f ,
2n 2n
j=1

and the sets n o

Er = f : lim n (f ) = r2 .
n

We proved that P (E ) = 1 in Theorem 2.8.2.

One rarely needs to deal with measures that are not -finite, but
the Radon-Nikodym derivative does not hold for such measures. For
example, if is counting measure and is Lebesgue measure, both
on R, then  , but there is no f such that (5.1) holds.

5.2 Giving drift to a Brownian motion

As we have already noted, there are two ways to take a fair game and make
it unfair or vice versa:

Play the game and then add a deterministic amount in one direction.

Change the probabilities of the outcome.

144CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

We will consider these methods in the case of a standard Brownian motion

Bt . For the first method, we can define a Brownian motion with drift m, by
setting
Wt = mt + Bt .
We will now consider the second way by changing the probabilities. We have
already motivated this with a binomial approximation in (4.9), but in this
section we will transform the game directly using the Brownian motion.
Suppose that Bt is defined on the probability space (, F, P) with a
filtration {Ft }. To change the probability is to consider a different measure
Q instead of P. Let
m2 t
Mt = emBt 2 . (5.3)
We have seen that Mt is a martingale. In fact (see (3.20)), Itos formula
shows that Mt satisfies

dMt = m Mt dBt , M0 = 1.

For each event V that is Ft measurable, we define

Qt (V ) = E [1V Mt ] .

In other words, on the probability space (, Ft , P),

dQt
= Mt .
dP
If s < t and V is Fs -measurable, then it is also Ft -measurable. In this case,
Qs (V ) = Qt (V ) as can be seen in the calculation

Qt (V ) = E [1V Mt ] = E [E(1V Mt | Fs )]

= E [1V E(Mt | Fs )] = E [1V Ms ] = Qs (V ). (5.4)

Hence we can write just Q for the measure.
We claim that

The process t 7 Bt , under the measure Q, is a Brownian motion with

drift m and 2 = 1.

We prove this by showing it satisfies the conditions to be a Brownian motion.

The continuity of paths is immediate as well as B0 = 0. To show the rest, it
suffices to show that if s, t 0, then Bt+s Bs is independent of Fs with a
5.2. GIVING DRIFT TO A BROWNIAN MOTION 145

normal distribution with mean mt and variance t. We can establish this by

showing it has the (conditional) moment generating function
2 t/2
EQ (exp{(Bt+s Bs )} | Fs ) = emt e . (5.5)

Here we are writing EQ to denote that the conditional expectation is taken

using the measure Q. To establish (5.5), we need to show that if V is Fs -
measurable, then
h 2
i
EQ [1V exp{(Bt+s Bs )}] = EQ 1V emt e t/2
2 t/2
= emt e Q(V ),

or equivalently, by the definition of EQ ,

2 t/2
E [1V exp{(Bt+s Bs )} Mt+s ] = emt e E[1V Ms ].

Note that if Y = Bt+s Bs , then Y is independent of Fs and

  2
 
E eY Mt+s | Fs = Ms em t/2 E eY emY | Fs
2
h i
= Ms em t/2 E e(+m)Y
2 t/2 2 t/2
= Ms em e(+m)
2 t/2
= Ms e emt .

Therefore, if V is Fs -measurable,

E [1V exp{(Bt+s Bs )} Mt+s ]

= E [E(1V exp{(Bt+s Bs )} Mt+s | Fs )]
= E [1V E(exp{(Bt+s Bs )} Mt+s | Fs )]
2 t/2
= e emt E[1V Ms ].

Example 5.2.1. Suppose Xt is a geometric Brownian motion satisfying

dXt = Xt [m dt + dBt ] ,

where Bt is a standard Brownian motion defined on the probability space

(, F, P). If r R, then we can find a new probability measure Q such that

dBt = r dt + dWt ,
146CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

where Wt is a Brownian motion with respect to Q. Then,

dXt = Xt [(m + r) dt + dWt ] .
Hence with respect to Q, Xt is a geometric Brownian motion with the same
volatility but a new drift. From this we can see that the measures for geo-
metric Brownian motions with the same are equivalent.
Example 5.2.2. We compute a quantity for Brownian motion with drift
by using a standard Brownian motion. Suppose Bt is a standard Brownian
motion, a > 0, m R, and Mt is the martingale in (5.3). Let Ta = inf{t :
Bt = a}. Then under the new measure Q, Bt is a Brownian motion with
drift m. Note that
Q{Ta < } = E [MTa 1{Ta < }] = E [MTa ] .
Here the expectation E is with respect to the original measure under which
B is a standard Brownian motion; the second equality follows from the fact
that P{Ta < } = 1. Therefore,
m2 Ta
  
Q{Ta < } = E exp mBTa
2
2
  
am m Ta
= e E exp .
2
In Exercise 2.7.1 we computed the density of Ta for Brownian motion. Given
this we could compute the expectation on the right-hand side. However we
will choose a different method. We know that driftless Brownian motion will
hit the line y = a with probability one, and hence so will Brownian motion
with m > 0. Therefore, if m > 0, Q{Ta < } = 1 and
m2 Ta
  
E exp = eam .
2
If the Brownian motion has drift m, then
m2 Ta
  
am
Q{Ta < } = e E exp = e2am .
2

5.3 Girsanov theorem

The Girsanov theorem describes the probability measure one obtains by
weighting or tilting the measure of a Brownian motion by a martingale.
One example was given in the previous section where the martingale was
2 t/2)
Mt = emBt (m , (5.6)
5.3. GIRSANOV THEOREM 147

and in the new measure, Bt was a Brownian motion with drift. We will
generalize this idea here.
Suppose Mt is a nonnegative martingale satisfying the exponential SDE
dMt = At Mt dBt , M0 = 1, (5.7)
where Bt is a standard Brownian motion. The solution to this equation was
given in (3.20),
Z t
1 t 2
Z
Yt
Mt = e where Yt = As dBs A ds. (5.8)
0 2 0 s
For many applications it suffices to consider the equation (5.7) and not worry
about the form of the solution (5.8). Solutions to (5.7) are local martingales,
but as we have seen, they might not be martingales. For now we will assume
that Mt is a martingale. In that case, we can define a probability measure
P by saying that if V is an Ft -measurable event, then
P (V ) = E [1V Mt ] . (5.9)
In other words, if we consider P, P as being defined on Ft -measurable events,
dP
= Mt .
dP
If s < t and V is Fs -measurable, then V is also Ft -measurable. Hence, in
order for the above definition to be consistent, we need that for such V ,
E [1V Ms ] = E [1V Mt ] .
Indeed, this holds by the computation (5.4) which only uses the fact that M
is a martingale and V is Fs -measurable. We write E for expectations with
respect to P . If X is Ft -measurable, then
E [X] = E [X Mt ] .
Theorem 5.3.1 (Girsanov Theorem). Suppose Mt is a nonnegative martin-
gale satisfying (5.7), and let P be the probability measure defined in (5.9).
If Z t
Wt = Bt As ds,
0
then with respect to the measure P , Wt is a standard Brownian motion. In
other words,
dBt = At dt + dWt ,
where W is a P -Brownian motion.
148CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

In other words, if we weight the probability measure P by the martingale,

then in the new measure P , the Brownian motion acquires a drift At . The
example of the previous section is the martingale (5.6) with At m and a
constant drift m is obtained.
Let us give a heuristic derivation of Girsanovs theorem using a binomial
approximation. Suppose t is given. In the binomial approximation to P,
we are equally likely to go up and down,
n o 1
P B(t + t) B(t) = t | B(t) = .
2
The binomial approximation to (5.7) is
n h i o 1
P M (t + t) = M (t) 1 A(t) t | B(t) = .
2


P the probability of a jump of
Therefore, in the weighted measure
should be proportional to 1A(t) t. Since the sum of the two probabilities
equals one, we see that
n o 1 h i
P B(t + t) B(t) = t | B(t) = 1 A(t) t .
2
As we saw in Section 4.4, this implies that

E [B(t + t) B(t) | B(t)] = A(t) t,

that is, in the probability measure P , the process obtains a drift of A(t).
The condition that Mt be a martingale (and not just a local martingale)
is necessary for Girsanovs theorem as we have stated it. Given only (5.7)
or (5.8), it may be hard to determine if Mt is a martingale, so it is useful
to have a version that applies for local martingales. If we do not know Mt
is a martingale, we can still use Theorem 5.3.1 if we are careful to stop the
process before anything bad happens. To be more precise, suppose Mt = eYt
satisfies (5.8), and note that
Z t
hY it = A2s ds.
0

Let
Tn = inf{t : Mt + hY it = n},
and let 
(n) At , t < Tn
At = .
0, t Tn
5.3. GIRSANOV THEOREM 149

Then
(n)
dMtTn = At MtTn dBt ,
which is a square integrable martingale since
Z t
E (MtTn 1)2 = E[A(n) 2
 
s MsTn ] ds
0
Z t
n E [A(n)
2 2 3
s ] ds n .
0
There is the corresponding measure, which we might denote by Pn , which
gives a drift of At up to time Tn and then proceeds with drift 0. If n < m,
then Pn and Pm are the same measure restricted to t Tn . Hence we can
write P for a measure on Bt , 0 t < T , where
T = lim Tn .
n
This shows how to tilt the measure up to time T , There are examples
such that P {T < }0. However, if for some fixed t0 , P {T > t0 } = 1, then
Mt , 0 t t0 is a martingale. In other words, what prevents a solution
to (5.7) from being a martingale is that with respect to the new measure
P , either Mt or |At | goes to infinity in finite time. We summarize with a
restatement of the Girsanov theorem.
Theorem 5.3.2 (Girsanov Theorem, local martingale form). Suppose Mt =
eYt satisfies (5.7)(5.8), and let
Tn = inf{t : Mt + hY it = n}, T = T = lim Tn .
n

Let P be the probability measure as above. If

Z t
Wt = Bt As ds, 0 t < T,
0
then with respect to the measure P , Wt , t < T is a standard Brownian
motion. In other words,
dBt = At dt + dWt , t < T,
where W is a P -Brownianmotion. If any one of the following three condi-
tions hold, then Ms , 0 s t, is actually a martingale:
P {Mt + hY it < } = 1,
E[Mt ] = 1
  
hY it
E exp < . (5.10)
2
150CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

It is not always easy to see whether or not the local martingale in (5.7)
will be a martingale. However, if any one of the three conditions at the end
of the theorem hold, then it is a martingale. The first condition uses P , the
new measure, while the expectation E in the other two conditions is with
respect to the original measure. The relation (5.10) is called the Novikov
condition.
Even if Mt is not a martingale, since
Mt = lim MtTn ,
n

Fatous lemma implies that

h i
E [Mt ] = E lim MtTn lim E[MtTn ] = E[M0 ] = 1.
n n

An extension of this argument shows that nonnegative local martingales are

supermartingales, that is, if s < t,
E[Mt | Fs ] Ms .
Example 5.3.1. Let Bt be a standard Brownian motion with B0 = 1. Let
= inf{t : Bt = 0}. Then Mt = Bt is a nonnegative martingale, satisfying
dMt = dBt = At Mt dBt , t < ,
where
1 1
At = = .
Mt Bt
If we tilt the measure using Mt , then
dt
dBt = At dt + dWt = + dWt ,
Bt
where Wt is a standard Brownian motion in the new measure P . Note that
this equation gives the Bessel process that we studied in Section 4.2 with
a = 1. Using properties of the Bessel process, we see that
P { < } = 0.
In other words, in the new measure the process avoids the origin.
Example 5.3.2. Let Bt be a standard Brownian motion with B0 = 1 and
r R. Let = inf{t : Bt = 0}. For t < , Itos formula gives
r(r 1) r2
dBtr = r Btr1 dBt + Bt dt
 2 
r r r(r 1)
= Bt dBt + dt .
Bt 2Bt2
5.3. GIRSANOV THEOREM 151

Let  Z  t
r(r 1)
Mt = exp 2
ds Btr .
0 2B s
The product rule shows that Mt satisfies the exponential SDE
r
dMt = Mt dBt , t < .
Bt
Therefore,
r
dBt =dt + dWt , t < ,
Bt
where Wt is a Brownian motion in the new measure. This equation is the
Bessel equation. In particular, if r 1/2, then P { = } = 1, and using
this we see that with P -probability one
Z t
Mt + A2s ds
0

is finite. Therefore, for r 1/2, Mt is a martingale.

Example 5.3.3. Suppose that Xt satisfies
dXt = Xt [m(t, Xt ) dt + (t, Xt ) dBt ] ,
where Bt is a standard Brownian motion with respect to the probability
measure P. Let us assume that (t, x) > 0 for all t, x. Suppose we want to
find a probability measure P that is mutually absolutely continuous with
respect to P and such that Xt is a martingale under P . Then we would want
m(t, Xt )
dBt = dt + dWt ,
(t, Xt )
where Wt is a standard Brownian motion with respect to P . This would
give
dXt = Xt (t, Xt ) dWt .
The local martingale that we need to consider is Mt satisfying
m(t, Xt )
dMt = Mt dBt , M0 = 1.
(t, Xt )
In other words,
Z t Z t 
1 m(t, Xt )
Mt = exp As dBs A2s ds , At = . (5.11)
0 2 0 (t, Xt )
While Mt is a local martingale, we cannot say it is a martingale without
verifying one of the conditions in Theorem 5.3.2.
152CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

Example 5.3.4. Suppose Xt is a Bessel process satisfying

1
dXt = dt + dBt , X0 = 1,
Xt
where Bt is a standard Browian motion. In Section 4.2 it was shown that
with probability one, the process Xt will never reach the origin and, in fact,
Xt . Let
1
Mt = .
Xt
Using Itos formula, we see that
1 1 1
dMt = 2 dXt + 3 dhXit = Mt dBt .
Xt Xt Xt

Therefore, Mt is a local martingale with M0 = 1 and we can apply the

second form of the Girsanov theorem. We will ask if Mt is a martingale.
Indeed, one can see that it is not by noting that Mt 0 and (with some
more work that we omit) E[Mt ] 0. If Mt were a martingale, we would
have E[Mt ] = 1 for all t. Another way to see that Mt is not a martingale is to
consider the measure P given by tilting by the martingale. In this measure,
1
dBt = dt + dWt ,
Xt
where Wt is a P -Brownian motion. Therefore, dXt = dWt which says that
the distribution of Xt in the measure P is that of a standard Brownian mo-
tion. Since the Brownian motion reaches zero, we can see that with respect
to P , the martingale Mt reaches infinity in finite time.

Suppose Mt is a continuous, nonnegative local martingale. To

show that Mt is a supermartingale, suppose that n is an increasing
sequence of stopping times such that Mtn is a martingale. Let s <
t. Suppose V is an Fs -measurable event and let Vk = V {k > s}.
Since Mtn is a martingale, if n k,

E [1Vk Mtn ] = E [1Vk Msn ] = E [1Vk Ms ] .

Hence Fatous lemma implies that

E [1Vk Mt ] lim inf E [1Vk Mtn ] = E [1Vk Ms ] .

n
5.3. GIRSANOV THEOREM 153

Since E [1V E(Mt | Fs )] = E [1V Mt ], by letting k and using

the monotone convergence theorem, we get

E [1V E(Mt | Fs )] = E [1V Mt ] E [1V Ms ] .

Since this holds for every Fs -measurable event V , we have E(Mt |

Fs ) Ms with probability one. Also, if V = {E(Mt | Fs ) < Ms }
and P(V ) > 0, then

E[Mt ] = E [1V Mt ]+E [1V c Mt ] < E [1V Ms ]+E [1V c Ms ] = E[Ms ].

Conversely, if E[Mt ] = E[M0 ] for all t, then Mt must be a martin-

gale. (We emphasize that we are using the fact that Mt 0.)
To prove the Girsanov theorem, let us first consider the case
when there exists K < such that with probability one hY it K
and Mt K. By Theorem 4.5.2, we need only show that in the
measure P , Wt is a continuous martingale with quadratic variation
hW it = t. Continuity and the quadratic variation calculation are
immediate (since they hold with probability one in the measure P),
so all we need to show is that Wt is a P -martingale, that is, if
s < t, then
EP (Wt | Fs ) = Ws .
In other words, if V is Fs -measurable, we need to show that

E [1V Ws ] = E [1V Wt ] ,

which by definition means

E [1V Ws Ms ] = E [1V Wt Mt ] .

In other words, we need to show that Zt = Wt Mt is a martingale

with respect to P. Since

dWt = At dt + dBt , dMt = At Mt dBt ,

the product rule gives

dZt = Wt dMt + Mt dWt + dhW, M it = (At + 1) Mt dBt ,

which shows that Zt is a local martingale, and since

Z t Z t
2 2 2
E[(As +1) Ms ] ds K E[2(A2s +1)]ds 2K 2 (K+t) < ,
0 0
154CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

Zt is a square-integrable martingale.
For more general At , Mt we use localization with the stopping
times Tn as above. Note that

E[Mt ] lim E [Mt 1{Tn > t}] = lim P {Tn > t} = P {T > t},
n n

and hence if P {T > t} = 1, then E[Mt ] = 1 and Ms , 0 s t is

a martingale.
We will now show that the Novikov condition (5.10) implies that
M is a martingale. For ease let t = 1. Let

r = inf{t : hY it = r}.

The process Z r
Xr = As dWs
0
is a standard Brownian motion with respect to the measure P .
Also,
1 r
Z
r
Yr = Xr + As ds = Xr + .
2 0 2
In particular,
 
t
max Ms = max exp Xt + .
0sr 0tr 2
In other words, T = lim Tn can be defined as

T = sup{t : hY it < }.

Let V denote the event that r 1 for all r < . We need to show
that P (V ) = 0.
Let
n = min {m n : Xm 0} .
Since Xm is a P -local martingale, for each n, P [V {n < }] =
P (V ). Also,
Mn = exp {Yn } en /2 .
Let us fix n, let = n , and note that

X
X
P (V ) E [Mm 1 { = m}] em/2 P{ = m}
m=n m=n
h i
E ehY i1 /2 1{ = m} .
5.4. BLACK-SCHOLES FORMULA 155

We therefore, get
h i
P (V ) E ehY i1 /2 1{r < 1} .

The events {r 1} shrink to a P-null set as r . The condition

(5.10) implies that ehY i1 /2 is integrable, and hence,
h i
lim E ehY i1 /2 1{r 1} = 0.
r

5.4 Black-Scholes formula

An arbitrage is a system that guarantees that a player (investor) will not lose
money while also giving a positive probability of making money. If P and
Q are equivalent probability measures, then an arbitrage under probability
P is the same as an arbitrage under probability Q. This holds since for
equivalent probability measures

P (V ) = 0 if and only if Q(V ) = 0,

P (V ) > 0 if and only if Q(V ) > 0.

This observation is the main tool for the pricing of options as we now show.
We will consider a simple (European) call option for a stock whose price
moves according to a geometric Brownian motion.
Suppose that the stock price St follows the geometric Brownian motion,

dSt = St [m dt + dBt ] , (5.12)

and there also exists a risk-free bound Rt satisfying

dRt = r Rt dt, (5.13)

that is, Rt = ert R0 . Let T be a time in the future and suppose we have the
option to buy a share of stock at time T for strike price K. The value of this
option at time T is

(ST K) if ST > K,
F (ST ) = (ST K)+ =
0 if ST K.

The goal is to find the price f (t, x) of the option at time t < T given St = x.
156CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

One possibility, which is not the Black-Scholes solution, is to price the

option by the expected value (in time t dollars),
h i
f(t, x) = E er(T t) F (ST ) | St = x .

In (4.7) we showed that this function satisfies the PDE

2 x2 00
t f(t, x) = r f(t, x) m x f0 (t, x) f (t, x). (5.14)
2
Here and throughout this section we use primes for x-derivatives. If one sells
an option at this price and uses the money to buy a bond at the current
interest rate, then there is a positive probability of losing money.
The Black-Scholes approach to pricing is to let f (t, x) be the value of
a portfolio at time t, given that St = x, that can be hedged in order to
guarantee a portfolio of value F (ST ) at time T . By a portfolio, we mean an
ordered pair (at , bt ) where at , bt denote the number of units of stocks and
bonds, respectively. Let Vt be the value of the portfolio at time t,

Vt = at St + bt Rt . (5.15)

We will manage the portfolio by switching between stocks and bonds so

that, no matter how the price of the stock moves, the value at time T will
be
VT = (ST K)+ .
We assume that the portfolio is self-financing, that is, one does not add
outside resources to the portfolio. The mathematical consequence of this
assumption is that the change of the value of the portfolio is given only by
the change of the price of the assets,

dVt = at dSt + bt dRt . (5.16)

It may appear at first that (5.16) is a consequence of (5.15). However, the

units of the assets at , bt vary with time and hence we need to use the product
rule. If Vt is defined by (5.15), the product rule implies that

d(at St ) = at dSt + St dat + dha, Sit ,

d(bt Rt ) = bt dRt + Rt dbt .

Here we use the fact that Rt is differentiable in t to see that hb, Rit = 0.
Hence, (5.16) is a strong assumption about the portfolio, and we will use it
to determine the price and the hedging strategy.
5.4. BLACK-SCHOLES FORMULA 157

If we assume (5.16) and plug in (5.12) and (5.13), we get

dVt = at St [m dt + dBt ] + bt r Rt dt
= at St [m dt + dBt ] + r [Vt at St ] dt
= [m at St + r (Vt at St )] dt + at St dBt . (5.17)

Under our definition, Vt = f (t, St ). Therefore, assuming f is sufficiently

differentiable, Itos formula shows that

dVt = df (t, St )
1
= t f (t, St ) dt + f 0 (t, St ) dSt + f 00 (t, St ) dhSit
2
2 St2 00
 
0
= t f (t, St ) + m St f (t, St ) + f (t, St ) dt
2
+ St f 0 (t, St ) dBt . (5.18)

By equating the dBt terms in (5.17) and (5.18), we see that that the portfolio
is given by
Vt at St
at = f 0 (t, St ), bt = , (5.19)
Rt
and then by equating the dt terms we get the Black-Scholes equation

2 x2 00
t f (t, x) = r f (t, x) r x f 0 (t, x) f (t, x).
2
There are two things to note about this equation.

The drift term m does not appear. If we think about this, we realize
why our assumptions should give us an equation independent of m.
Our price was based on being able to hedge our portfolio so that with
probability one the value at time T is (ST K)+ . Geometric Brown-
ian motions with the same but different m are mutually absolutely
continuous and hence have the same events of probability one.

The equation is exactly the same as (5.14) except that m has been
replaced with r. Therefore, we can write
h i
f (t, x) = E er(T t) F (St ) | St = x ,

where S satisfies
dSt = St [r dt + dBt ].
158CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

Using this, one can compute f (t, x) exactly; in the next section we do
this and derive the Black-Scholes formula
2
!
log(x/K) + (r + 2 )t
f (T t, x) = x
t

2
!
rt log(x/K) + (r 2 )t
K e , (5.20)
t

where is the standard normal distribution function. If one knows r

and (which from a practical perspective is a big assumption), one
can just plug into this formula.

We can easily generalize this to the case where the stock price satisfies

dSt = St [m(t, St ) dt + (t, St ) dBt ] ,

dRt = r(t, St ) Rt dt.

Under the assumption of a self-financing portfolio, we again get (5.17) and

(5.18), and by equating coefficients we get the Black-Scholes equation

(t, x)2 x2 00
t f (t, x) = r(t, x) f (t, x) r(t, x) x f 0 (t, x) f (t, x). (5.21)
2

As before, the drift term m(t, x) does not appear in the equation. The func-
tion f can be given by

f (t, x) = E [(Rt /RT ) F (ST ) | St = x] ,

where St , Rt satisfy

dSt = St [r(t, St ) dt + (t, St ) dBt ] ,

dRt = r(t, St ) Rt dt.

Mathematical justification of this arguments requires sufficient assumptions

on m and such that the equation (5.21) has a solution. Finding explicit
solutions to such equations is often impossible, but one can either solve the
PDE numerically or do Monte Carlo simulations of the associated diffusion
process St .
5.5. MARTINGALE APPROACH TO BLACK-SCHOLES EQUATION159

5.5 Martingale approach to Black-Scholes equa-

tion
We will give a different approach to deriving the Black-Scholes formula.
Suppose for the moment that the risk-free bond has rate r(t, x) and that
the volatility is given by (t, x). If Rt denotes the value of the bond at time
t, then
Z t 
dRt = r(t, St ) Rt dt, Rt = R0 exp r(s, Ss ) ds .
0

As explained in the previous section, if we want a strategy to hedge a port-

folio so that its value is determined at time T , then the strategy must be
independent of the drift coefficient m(t, x). For this reason, we may assume
that the stock price satisfies

dSt = St [r(t, St ) dt + (t, St ) dBt ] , (5.22)

and then the value of the portfolio at time t satisfies

Vt = f (t, St ) = EQ [(Rt /RT ) F (ST ) | St ] = EQ [(Rt /RT ) F (ST ) | Ft ] .

We write EQ , EQ to emphasize that the expectations are taken with respect

to the measure under which St satisfies (5.22).
Let St = St /Rt , Vt = Vt /Rt be the stock price and portfolio value, re-
spectively, discounted by the bond rate. The product rule shows that St
satisfies
dSt = (t, Rt ) St dBt .
In other words (under some growth restrictions on , for example, if is
uniformly bounded), St is a martingale. Also,

Vt = Vt /Rt = Rt1 EQ ((Rt /RT ) F (ST ) | Ft )

= EQ RT1 F (ST ) | Ft = EQ [VT | Ft ].

We have just demonstrated the following principle.

Theorem 5.5.1. Suppose St satisfies

dSt = St [m(t, St ) dt + (t, St ) dBt ] ,

and a risk-free bond Rt is available at rate r(t, St ),

dRt = r(t, St ) Rt dt.

160CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

Suppose that the Brownian motion is defined on a probability space (, F, P)

and that there exists a probability measure Q that is mutually absolutely
continuous with respect to P such that under Q, the discounted stock price
St = St /Rt is a martingale. Suppose there is an option at time T with value
F (ST ) such that EQ [RT1 |F (ST )|] < . Then the arbitrage-free price of the
option at time t is

Vt = Rt EQ RT1 F (ST ) | Ft .


(5.23)

A nice thing about this formulation is that VT = F (ST ) follows directly

from the formula. However, unlike the previous approach, we do not imme-
diately get the expression for the portfolio needed to hedge the option.

Example 5.5.1. We will derive the Black-Scholes formula leaving some of

the calculus calculations to Exercise 5.11. Suppose that r, are constants
and F (ST ) = (ST K)+ . The discounted values are St = ert St , Vt = ert Vt
and
VT = erT F (ST ) = erT (ST K)+ = (ST K)+ ,
where K = erT K. Under the measure Q, St satisfies

dSt = St dBt ,

which implies that

Z T Z T 
1 2
ST = St exp dBs ds
t 2 t
2
 
(T t)
= St exp (BT Bt ) .
2

In other words, the conditional distribution of ST given St is that of

Z = exp {aN + y} ,

where a = T t, N is a standard normal random variable, and

a2
y = log St .
2
Straightforward calculus shows that Z has a density
 
1 y + log z
g(z) = ,
az a
5.5. MARTINGALE APPROACH TO BLACK-SCHOLES EQUATION161

where is the standard normal density, and hence

Z
Vt = (z K) g(z) dz.
K

Another computation gives

a2 a2
! !
log(St /K) + 2 log(St /K) 2
Vt = St K ,
a a

which implies that

a2
!
rt log(St /K) + rs + 2
Vt = e Vt = St
a
a2
!
log(St /K) + rs
ers K 2
,
a

where s = T t. Plugging in a = s gives (5.20).
One of the hypotheses in Theorem 5.5.1 is that if St satisfies
dSt = St [m(t, St ) dt + (t, St ) dBt ] ,
then there exists a probability measure Q under which
dSt = St [r(t, St ) dt + (t, St ) dWt ] , (5.24)
where Wt is a Q-Brownian motion. Indeed, if St satisfies (5.24), then the
discounted price St = St /Rt satisfies
dSt = St (t, St ) dWt = St (t, Rt St ) dWt . (5.25)
The Girsanov theorem tells us that the way to obtain Q is to tilt by the
local martingale Mt where
r(t, St ) m(t, St )
dMt = Mt dBt ,
(t, St )
for then in the measure Q,
r(t, St ) m(t, St )
dBt = dt + dWt .
(t, St )
In order for the theorem to hold, we need that Mt is actually a martingale
and that St as given in (5.25) is a martingale. While one can give general
conditions when this is true, it is often just as easy to check this in each
case. If |r m|/ is uniformly bounded, these conditions are satisfied.
162CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

5.6 Martingale approach to pricing

We generalize the discussion in the previous section to the pricing of claims
that are functions of the entire history of the prices of an asset. Suppose St
denotes the price of an asset satisfying

dSt = St [mt dt + t dBt ] (5.26)

where Bt is a standard Brownian motion. Let Ft denote the information

in {Bs : 0 s t}, and as usual we assume that mt , t are processes
adapted to the filtration {Ft }. We also assume that there is a risk-free bond
Rt satisfying
Z t 
dRt = rt Rt dt, Rt = R0 exp rs ds ,
0

where rt is also adapted.

Let T be a fixed future time and assume that V is an FT -measurable
random variable. We call V a claim (at time T ). The examples we have seen
are of the form V = F (ST ), but other examples are

1 T
Z
V = max St , V = St dt.
0tT T 0
We will start with the following definition.
Definition, first try If V is a claim at time T , then the (arbitrage-free)
price Vt , 0 t T, of a claim VT is the minimum value of a self-financing
portfolio that can be hedged to guarantee that its value at time T is V .
Our goal is to determine the price Vt and the corresponding portfolio
(at , bt ), where at denotes the number of units of S and bt the number of
units of R. This will require some mathematical assumptions that we will
make as we need them. Recall that

Vt = at St + bt Rt ,

and (at , bt ) is self-financing if

dVt = at dSt + bt dRt .

We will start by giving two bad, but unrealistic, examples that show that we
need to take some care. The two examples are similar. In the first example,
we allow the stock price to fluctuate too much. In the second, we choose a
very risky portfolio similar to the martingale betting strategy.
5.6. MARTINGALE APPROACH TO PRICING 163

Example 5.6.1. Suppose for ease that rt 0 and R0 = 1 so that Rt 1.

Let St = eZt where Zt is the martingale betting strategy revisited in
Example 4.1.1. We recall that Zt satisfies
dZt = At dBt , Z0 = 0,
and the bets At are chosen so that with probability one Z1 1. Itos
formula shows that St satisfies
 2 
At
dSt = St dt + At dBt .
2
We cannot find an equivalent measure such that St is a martingale. Indeed,
we know that with probability one S1 > S0 , and hence this fact must be
true in any equivalent measure.
Example 5.6.2. Suppose that Rt 1 and St = eBt (t/2) which satisfies
dSt = St dBt , S0 = 1.
Let At be as in Example 5.6.1, and choose a portfolio with at = At /St . The
value of the portfolio is Vt = at St + bt Rt = At + bt . If the portfolio is to be
self-financing, we need that
At
dVt = at dSt + bt dRt = St dBt = At dBt .
St
We choose bt so that this holds (this may require choosing bt to be negative).
Since V0 = 1, we see that Vt = 1 + Zt where Zt is as in Example 5.6.1. In
particular, V1 = 2, and this portfolio hedges the claim V 2 at time T = 1.
Similarly for any constant J, we can find At such that with probability one
Zt J. Hence given any V0 (even negative values), we can find a portfolio
that hedges the claim V 2. One disadvantage of these portfolios, however,
is that they allow the value at times 0 < t < 1 to be negative. It will turn out
that we can eliminate examples like this by restricting to portfolios whose
value at all times is nonnegative.
With the bad examples in mind, we now proceed to develop the theory.
To start, we will consider discounted prices. Let
St = (R0 /Rt ) St , Vt = (R0 /Rt ) Vt
denote the discounted stock price and discounted value, respectively. The
portfolio (at , bt ) is the same whether or not we are using discounted units
and the discounted value Vt is given by
Vt = at St + bt R0 .
164CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

(Note that the discounted bond value is R0 .) Using the product formula,
we see that
dSt = St [(mt rt ) dt + t dBt ] .
Our goal is to find a self-financing portfolio (at , bt ) such that with probability
one
VT = aT ST + bT R0 = V .
Since this must happen with probability one, we may consider a mutually
absolutely continuous measure. We let Q be the probability measure (if it
exists) that is mutually absolutely continuous with respect to P such that
under Q the discounted stock price is a martingale. Recalling (5.11), we can
see that the Girsanov theorem tells us to choose

dQ = Mt dP,

where Mt satisfies
rt mt
dMt = Mt dBt , M0 = 1. (5.27)
t
The solution to this last equation is a local martingale, but it not necessarily
a martingale. If it is not a martingale, then some undesirable conclusions
may result as in our examples above. Our first assumption will be that it is
a martingale.
Assumption 1. The local martingale defined in (5.27) is actually a
martingale.
This assumption implies that Q is mutually absolutely continuous with re-
spect to P. Theorem 5.3.2 gives a number of ways to establish Q  P. If
Q  P, then we also get P  Q if P{Mt > 0} = 1. Let
Z t
rs ms
Wt = Bt ds,
0 s
which is a Brownian motion with respect to Q. Plugging in we see that

dSt = t St dWt . (5.28)

This shows that St is a local martingale with respect to Q. We will want this
to be a martingale, and we make this assumption.
Assumption 2. The Q-local martingale St satisfying (5.28) is actually
a Q-martingale.
5.6. MARTINGALE APPROACH TO PRICING 165

Again, Theorem 5.3.2 gives some sufficient conditions for establishing that
the solution to (5.28) is a Q-martingale. We write EQ and EQ for expecta-
tions (regular and conditional) with respect to Q.

Definition A claim V at time T is called a contingent claim if V 0 and

h i
EQ V 2 < .

The (arbitrage-free) price Vt , 0 t T, of a contingent claim VT is the

minimum value of a self-financing portfolio that can be hedged to guarantee
that its value never drops below zero and at time T equals V .

Given a contingent claim, we can set

h i
Vt = EQ V | Ft .

This is a square integrable martingale and VT = V . We would like to find a

portfolio (at , bt ) satisfying

Vt = at St + bt Rt ,

that is self-financing,
dVt = at dSt + bt dRt .

Recall that if At is an adapted process with

Z t
EQ [A2s ] ds < ,
0

then Z t
Zt = As dWs
0

is a square integrable martingale. Let us assume for the moment that there
exists such a process As such that
Z t
Vt = V0 + As dWs ,
0

that is
dVt = At dWt . (5.29)
166CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

We compute,

dVt = Rt dVt + Vt dRt

= Rt At dWt + Vt dRt
At
= Rt dSt + Vt dRt
t St
At
= [dSt St dRt ] + Vt dRt
t St
 
At At
= dSt + Vt dRt .
t St t
Therefore, if
At At
at = , bt = Vt , (5.30)
t St t
the portfolio is self-financing and
 
At At
at St + bt Rt = St Rt + Vt Rt = Rt Vt = Vt .
t St t

Along the way we made the assumption that we could write Vt as (5.29).
It turns out, as we discuss in the next section, that this can always be done
although we cannot guarantee that the process At is continuous or piecewise
continuous. Knowing existence of the process is not very useful if one cannot
find At . For now we just write as an assumption that the computations work
out.
Assumption 3. We can write Vt as (5.29), and if we define at , bt as
in (5.30), then the stochastic integral
Z t Z t
Vt = as dSs + bs dRs ,
0 0

is well defined.
Theorem 5.6.1. If V is a contingent claim and assumptions 1-3 hold, then
the arbitrage-free price is

Vt = Rt EQ (VT | Ft ).

We have done most of the work in proving this theorem. What remains
is to show that if (at , bt ) is a self-financing strategy with value

Vt = at St + bt Rt ,
5.6. MARTINGALE APPROACH TO PRICING 167

such that with probability one, Vt 0 for all t and VT V , then for all t,
with probability one Vt Vt . When we say with probability one this can
be with respect to either P or Q since one of our assumptions is that the
two measures are mutually absolutely continuous. Let Vt = Vt /Rt be the
discounted values. The product rule gives
h i
dVt = d(Rt Vt ) = Rt dVt + Vt dRt = Rt dVt + at St + bt dRt ,

and the self-financing assumptions implies that

h i
dVt = at dSt + bt dRt = at Rt dSt + St dRt + bt dRt .

By equating coefficients, we see that

dVt = at dSt = at t St dWt .

In particular, Vt is a nonnegative local martingale. We have seen that this

implies that Vt is a supermartingale and
h i
EQ VT | Ft Vt .

If V VT , then h i
EQ VT | Ft EQ [V | Ft ] = Vt .

Example 5.6.3. We will give another derivation of the Black-Scholes equa-

tion. Assume that the stock price is a diffusion satisfying

dSt = St [m(t, St ) dt + (t, St ) dBt ] ,

and the bond rate satisfies

dRt = r(t, St ) Rt dt.

The product rule implies that the discounted stock price satisfies

dSt = St [(m(t, St ) r(t, St )) dt + (t, St ) dBt ] .

If V is a claim of the form V = F (ST ), let be the function

(t, x) = EQ [(Rt /RT ) V | St = x] ,

and note that

Vt = Rt Vt = Rt EQ RT1 V | Ft = (t, St ).
 
168CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

Assuming sufficient smoothness, Itos formula gives

1 00
d(t, St ) = t (t, St ) dt + 0 (t, St ) dSt + (t, St ) dhSit .
2
Recalling that
dSt = St [r(t, St ) dt + (t, St ) dWt ] ,
we see that
dVt = d[Rt1 (t, St )] = Jt dt + At dWt ,
where h (t, St )2 St2 00
Jt = Rt1 t (t, St ) + (t, St )
2
i
+r(t, St ) St 0 (t, St ) r(t, St )(t, St ) ,

At = Rt1 St (t, St ) 0 (t, St ) = St (t, St ) 0 (t, St ).

Since Vt is a Q-martingale, Jt = 0, giving the Black-Scholes PDE again, and
dVt = At dWt . Plugging into (5.30), we recover the hedging portfolio,
At
at = = 0 (t, St ),
(t, St ) St
At
= Rt1 Vt St 0 (t, St ) .
 
bt = Vt
(t, St )
This can be compared to (5.19).
Example 5.6.4. Suppose St is a geometric Brownian motion

dSt = St [m dt + dBt ] ,

and the bond rate is constant r. Suppose that the claim is the average stock
price over the interval [0, T ],

1 T
Z
V = St dt.
T 0

In the new measure Q, the discounted stock price St = ert St satisfies

dSt = St dWt ,

where Wt is a Q-Brownian motion. The discounted value is

 rT Z T 
e
Vt = EQ Ss ds | Ft .
T 0
5.6. MARTINGALE APPROACH TO PRICING 169

Rt
Since 0 Ss ds is Ft -measurable, we get
Z t Z T 
rT
Te Vt = Ss ds + EQ Ss ds | Ft
0 t
Z t Z T
= Ss ds + EQ [Ss | Ft ] ds.
0 t

The second equality uses a form of Fubinis theorem that follows from the
linearity of conditional expectation. Since Ss is a Q-martingale, if s > t,
h i
EQ [Ss | Ft ] = ers EQ Ss | Ft = ers St = er(st) St .

Therefore,
T T
er(T t) 1
Z Z
EQ [Ss | Ft ] ds = St er(st) ds = St ,
t t r

and

erT t ert erT

Z
Vt = Ss ds + St
T 0 rT
erT t 1 er(T t)
Z
= Ss ds + St , (5.31)
T 0 rT

er(T t) t
1 er(T t)
Z
Vt = ert Vt = Ss ds + St .
T 0 rT
Note that VT = V which we needed, and the price at time 0 is

1 erT
V0 = S0 .
rT
The hedging portfolio can be worked out with a little thought. We will
start with all the money in stocks and as time progresses we move money
into bonds. Suppose that during time interval [t, t + t] we convert ut
units of stock into bonds. Then the value of these units of bonds at time
T will be about u er(T t) St t. If we choose u = er(tT ) /T , then the value
will be about St t/T and hence the value of all our bonds will be about
1 T
R
T 0 Ss ds. This gives us
dat er(tT )
=
dt T
170CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

and using the terminal condition aT = 0, we get

1 er(tT )
at = . (5.32)
rT
This is a special case where the hedging strategy does not depend on the
current stock price St . If we want to use the formula we derived, we use
(5.31) to give

1 er(tT ) 1 er(tT )
dVt = dSt = St dWt .
rT rT
Plugging this into (5.30) gives (5.32).

5.7 Martingale representation theorem

In the last section we assumed in (5.29) that a continuous martingale could
be represented as a stochastic integral. In fact, this is always that case.
Suppose Bt is a standard Brownian motion and {Ft } is the filtration given
by the Brownian motion. Suppose there is a claim V at time T and let

Vt = E[V | Ft ].

Then there exists an adapted process At such that

Z t
E[V | Ft ] = E[V ] + As dBs .
0

In order to be precise, we need to allow a wider class of processes than those

with continuous and piecewise continuous paths. We will not go into detalis
here, but any allowable process can be approximated by continuous and
piecewise continuous processes. From practical perspective, knowing that
such an As exists is not so useful unless one can give it explicitly. In the
examples we give the SDE for Vs explicitly.
This theorem states roughly that all of the randomness in the system
is given by the Brownian motion Bt and the only way to get martingales
with only this information is to vary the betting strategy on the Brownian
motion. To illustrate this idea, we will derive the representation theorem in
the easier case of random walk.
Suppose we have a binomial model with t = 1/N and x = t =
1/ N . We have independent random variables

Xt , X2t , . . .
5.7. MARTINGALE REPRESENTATION THEOREM 171

each with
1
P {Xkt = x} = P {Xkt = x} = .
2
Let Fkt denote the information in {Xt , X2t , , . . . , Xkt }, and assume
that Mkt is a martingale with respect to Fkt . The martingale property
implies that
 
E M(k+1)t | Fkt = Mkt . (5.33)

Since {M(k+1)t } is F(k+1)t -measurable, its value depends on the values of

Xt , X2t , . . . , X(k+1)t .

When we take a conditional expectation with respect to Fkt we are given

the value of the vector

Xk = (Xt , X2t , . . . , Xkt ) .

Given a particular value of Xkt , there are only two possible values for
M(k+1)t corresponding to the values when X(k+1)t = x and X(k+1)t =
x, respectively. Let us denote the two values by

[b + a] x, [b a] x,

where b x is the average of the two values. The two numbers a, b depend
on Xkt and hence are Fkt -measurable. The martingale property (5.33)
tells us that
b x = Mkt ,

and hence
M(k+1)t Mkt = a x = a X(k+1)t .

If we write J(k+1)t for the number a, then J(k+1)t is Fkt -measurable, and

k
X
Mkt = M0 + Jjt Xjt .
j=1

This is the form of the stochastic integral with respect to random walk as
in Section 1.6.
172CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

5.8 Exercises
Exercise 5.1. For each of the following random variables Xj on R, let j
be the distribution considered as a probability measure on R. For each pair
state whether or not j  k . Since there are six random variables, there
should be 30 answers.

X1 normal mean 2, variance 7

X2 binomial with parameters n = 7, p = .3

X3 Poisson with parameter 2
X4 = eX1
X5 uniform on [0, 1]
X6 = X1 X3 + X2 assuming X1 , X2 , X3 independent .

Exercise 5.2. Consider the martingale betting strategy from Chapter 1.

Let Wn denote the winnings at time n and let

M n = 1 Wn ,

so that at each time n, Mn equals 0 or 2n . Note that M0 = 1. Let T =

min{n : Mn = 0}. Let Fn be the information in M0 , M1 , . . . , Mn .

1. Explain why Mn is a nonnegative martingale.

2. Define a new probability measure Qn by saying that for any event V

that is Fn -measurable,

Qn (V ) = E [Mn 1V ] .

Show that if m < n and V is Fm -measurable, then Qm (V ) = Qn (V ).

3. Given the last part, we can write Q rather than just Qn . Find the
transition probability

Q Mn+1 = 2n+1 | Mn = 2n .


4. Find the Q-probability that T < .

5. Is Mn a martingale with respect to the measure Q?

5.8. EXERCISES 173

Exercise 5.3. Suppose Bt is a standard Brownian motion on (, P). For

each of the following choices of Xt , 0 t 1, state whether there is an
equivalent probability measure Q such that the Xt is a standard Brownian
motion in the new measure. If the answer is yes, give dQ/dP at t = 1. In all
cases assume that B0 = 0, X0 = 0.

dXt = 2 dt + dBt ,

dXt = 2 dt + 6 dBt ,
dXt = 2Bt dt + dBt .

Exercise 5.4. Let Bt be a standard Brownian motion with B0 = 0. Let

2
m R and let Xt = emBt .

1. Find a function g such that

Z t 
Mt := Xt exp g(Xs ) ds
0

is a local martingale.

2. What SDE does Mt satisfy?

3. Let Q be the probability measure obtained by tilting by Mt . Find the

SDE for Bt in terms of a Q-Brownian motion.

4. Explain why Mt is actually a martingale.

Exercise 5.5. Let Bt be a standard Brownian motion with B0 = 1. Let

T = min{t : Bt = 0} Let r > 0 and let Xt = Btr .

1. Find a function g such that

Z t 
Mt := Xt exp g(Xs ) ds
0

is a local martingale for t < T . (Do not worry about what happens
after time T .)

2. What SDE does Mt satisfy?

3. Let Q be the probability measure obtained by tilting by Mt . Find the

SDE for Bt in terms of a Q-Brownian motion.
174CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM

Exercise 5.6. Suppose a, m > 0 and Wt is a Brownian motion with drift

m and variance 1. Let
Ta = inf{t : Wt = a}.
Find the density of Ta . Hint: look at Examples 2.7.1 and 5.2.2.

Exercise 5.7. Suppose the price of a stock St follows a geometric Brownian

motion
dSt = St [3 dt + dBt ] ,
whee Bt is a standard Brownian motion with respect to a measure P.. Sup-
pose there is a risk-free bond with rate r = .05. Let Q be the measure, mu-
tually absolutely continuous with respect to P, under which the discounted
stock price St is a martingale.

1. Give the SDE for St in terms of Wt , a standard Brownian motion with

respect to Q.

2. Suppose there is a claim at time T = 2 of V = S22 . Is this a contingent

claim?

3. Find the discounted value Vt of the claim for 0 t 2.

4. Give the SDE for Vt in terms of the Brownian motion W .

5. Find the portfolio (at , bt ) that hedges the claim.

6. Find the value Vt .

Exercise 5.8. Repeat Exercise 5.7 with V = S23 .

Exercise 5.9. Repeat Exercise 5.7 with claim

Z 2
V = s Ss ds.
0

Exercise 5.10. Repeat Exercise 5.7 with claim

Z 2
V = Ss2 ds.
0

Exercise 5.11. Here we do the calculus needed to finish the derivation of

the Black-Scholes formula in Example 5.5.1.
5.8. EXERCISES 175

1. Show that if N has a standard normal distribution and a 6= 0, y are

constants, then the density of Z = exp{aN + y} is
 
1 y + log z
g(z) = ,
az a

where denotes the density of N .

2. Show that for all x,

Z
y log x + a2
   
y+(a2 /2) y log x
(z x) g(z) dz = e x .
x a a

3. Use these calculations to check the details of Example 5.5.1.

176CHAPTER 5. CHANGE OF MEASURE AND GIRSANOV THEOREM
Chapter 6

Jump processes

6.1 Levy processes

The assumptions that led to Brownian motion were:

Independent increments

Stationary increments

Continuous paths.

If we want to consider more general processes, we need to give up at least

one of these assumptions. In this chapter we will not assume continuity of
paths.

Definition A stochastic process Xt is called a Levy process if it has in-

dependent, stationary increments. That is to say, for every s, t > 0, the
random variable Xs+t Xs is independent of {Xr : r s} and has the same
distribution as Xt X0 .

Brownian motion is the only Levy process with continuous paths. We

have already discussed one example of a Levy process with discontinuous
paths, the Poisson process, and we will describe it in more detail in Section
6.2. Let us give two examples that can be derived from Brownian motion.

Example 6.1.1. Suppose Bt is a standard Brownian motion and Ts =

inf{t : Bt = s}. Using the strong Markov property, we can see that Ts , s 0
has independent, stationary increments. Therefore it is a Levy process. Note

177
178 CHAPTER 6. JUMP PROCESSES

that it is increasing, that is, with probability one if r < s, then Tr < Ts . We
calculated the density of Ts in Example 2.7.1,
s s2
fs (t) = e 2t , 0 < t < .
t3/2 2
In particular, the distribution of Ts is not normal. We claim that (with
probability one) the function s 7 Ts is not continuous. To see this, let
M = max0t1 Bt and let s0 (0, 1) be a time t with Bt = M . Then by
definition of M, Ts we see that TM s0 , but Ts > 1 for s > M , showing that
T is not continuous at s0 . The scaling property of Brownian motion implies
that Ts has the same distribution as s2 T1 . A tricky calculation which we
omit computes the characteristic function
Z
2
E[ei(r/s ) Ts ] = E[eirT1 ] = eirt f1 (t) dt = e(r) ,

where
|2r|1/2 (1 i)

if r 0
(r) = .
|2r|1/2 (1 + i) if r 0
Example 6.1.2. Let Bt = (Bt1 , Bt2 ) be a standard two-dimensional Brown-
ian motion. Let
Ts = inf{t : Bt1 = s},
and
Xs = B 2 (Ts ).
Using the strong Markov property, one can show that the increments are
independent, stationary and similarly to Example 6.1.1 that the paths as
discontinuous. The scaling property of Brownian motion implies that Xs has
the same distribution as s X1 . The density of X1 turns out to be Cauchy,
1
f (x) = , < x < ,
(x2 + 1)
with characteristic function
Z
eirx f (x) dx = e|r| .
 irXt 
E e =

Not every distribution can arise as the distribution of the increments of

a Levy process. Indeed, suppose that Xt is a Levy process. Then we can
write
n    
X j j1
X1 = Yj,n , where Yj,n = X X .
n n
j=1
6.1. LEVY PROCESSES 179

Definition A random variable X has an infinitely divisible distribution if

for each n we can find independent, identically distributed random variables
Y1,n , . . . , Yn,n such that X has the same distribution as

Y1,n + + Yn,n .

We just noted that if Xt is a Levy process, then X1 (and, in fact, Xt for

each t) has an infinitely divisible distribution. The converse is true and it not
too difficult to show. If F is the distribution function of an infinitely divisible
distribution, then there is a Levy process Xt such that X1 has distribution
F . Normal and Poisson random variables are infinitely divisible:

If X N (m, 2 ), then X has the same distribution as the sum of n

independent N (m/n, 2 /n) random variables.

If X has a Poisson distribution with mean , then X has the same

distribution as the sum of n independent Poisson random variables
with mean /n.

The goal of the next few sections is to show that every infinitely divisible
random variable is the sum of a normal random variable and a generalized
Poisson or jump random variable. The category generalized Poisson is
rather large and will include the distributions in Examples 6.1.1 and 6.1.2.

A precise definition of a Levy process needs another condition.

For Brownian motion, we put continuity of the paths into the as-
sumptions. To construct Brownian motion, we first defined Brown-
ian motion on dyadic rationals times, showed that the corresponding
process was uniformly continuous, and then extended to other times
by continuity. In the case of Levy processes, we similarly can first
define the process at dyadic times to satisfy independent, stationary
increments. As we discuss later, we then prove that with probability
one, for all times t, the limits

Xt = lim Xq , Xt+ = lim Xq ,

qt qt

exist where the limits are taken over the dyadic rationals. We then
define Xt to be Xt+ .
180 CHAPTER 6. JUMP PROCESSES

6.2 Poisson process

The Poisson process is the basic jump Levy process. It is obtained by taking
jumps of size one at a particular rate . Suppose we have such a process,
and let p(s) denote the probability that there is at least one jump in the
interval [t, t + s]. If the increments are to be stationary, this quantity must
not depend on t. Rate means that the probability that there is a jump
some time during the time interval [t, t + t] is about t; more precisely,

p(t) = t + o(t), t 0.

We can use this observation to construct the Poisson process.

We consider the waiting times between jumps. Let Xt denote the number
of jumps that have occurred by time t and let

T = inf{t : Xt = 1}

denote the amount of time until the first jump. Using independent, station-
ary increments, we see that
n   
Y (j 1)t jt
P{T > t} = P no jump during ,
n n
j=1
  n
t
= 1p .
n
Therefore,
 n

t
P{T > t} = lim 1 p
n n
  n
t t
= lim 1 +o = et .
n n n
Recall that a random variable T has an exponential distribution with rate
if it has density
f (t) = et , 0 < t < ,
and hence Z
P{T > t} = f (s) ds = et .
t
Our assumptions imply that the waiting times of a Poisson distribution with
parameter must be exponential with rate . Note that
Z
1
E[T ] = t f (t) dt = ,
0
6.2. POISSON PROCESS 181

so that the mean waiting time is (quite reasonably) the reciprocal of the
rate. This observation gives a way to construct a Poisson process. This con-
struction also gives a good way to simulate Poisson processes (see Exercise
6.1).

Let T1 , T2 , T3 , . . . be independent random variables each exponential

with rate .

Let 0 = 0 and for positive integer n,

n = T1 + + Tn .

In other words, n is the time at which the nth jump occurs.

Set
Xt = n for n t < n+1 .

Note that we have defined the process so that the paths are right-continuous,

Xt = Xt+ := lim Xs .
st

The paths also have limits from the left, that is, for every t the limit

Xt = lim Xs
st

exists. If t is a time that the process jumps, that is, if t = n for some n > 0,
then
Xt = Xt+ = Xt + 1.
At all other times the path is continuous, Xt = Xt+ .
The independent, stationary increments follow from the construction.
We can use the assumptions to show the following.

The random variable Xt+s Xs has a Poisson distribution with mean

t, that is,
(t)k
P{Xt+s Xs = k} = et .
k!

One way to derive this is to write a system of differential equations for the
functions
qk (t) = P{Xt = k}.
182 CHAPTER 6. JUMP PROCESSES

In the small time interval [t, t + t], the chance that there is more than one
jump is o(t) and the chance there is exactly one jump is t + o(t).
Therefore, up to errors that are o(t),

P{Xt+t = k} = P{Xt = k 1} (t) + P{Xt = k} [1 t] .

This gives

qk (t + t) qk (t) = t [qk1 (t) qk (t)] + o(t),

or
dqk (t)
= [qk1 (t) qk (t)].
dt
If we assume X0 = 0, we also have the initial conditions q0 (0) = 1 and
qk (0) = 0 for k > 0. We can solve this system of equations recursively, and
this yields the solutions
(t)k
qk (t) = et .
k!
(Although it takes some good guesswork to start with the equations and find
qk (t), it is easy to verify that qk (t) as given above satisfies the equations.)
When studying infinitely divisible distributions, it will be useful to con-
sider characteristic functions, and for notational ease, we will take loga-
rithms. Since the characteristic function is complex-valued, we take a little
care in defining the logarithm.

Definition If X is a random variable, then its characteristic exponent

(s) = X (s) is defined to be the continuous function satisfying

E[eisX ] = e(s) , (0) = 0.

If Xt is a Levy process, then the characteristic exponent of the process is

the characteristic exponent of X1 X0 .

Note that if X, Y are independent, then

X+Y = X + Y . (6.1)

If X N (m, 2 ), then

2 2
(s) = ims s .
2
6.2. POISSON PROCESS 183

If X is Poisson with mean , then

(s) = [eis 1].

This can be seen from the computation

X X (eis )n is
E[eisX ] = eisn P{X = n} = e = ee e .
n!
n=0 n=0

If Xt is a Levy process starting at the origin with characteristic exponent

, then independent, stationary increments and the relation (6.1), imply
that
Xt = t X1 = t .
An important property of Levy processes is the following.
Suppose Xt1 , Xt2 are independent Levy processes with X 1 =
1 , X 2 = 2 . Then Xt = Xt1 + Xt2 is a Levy process with Xt =
1 + 2 .
For example if Xt1 is a Brownian motion with drift m and variance 2 and
Xt2 is an independent Poisson process with rate , then Xt = Xt1 + Xt2 is a
Levy process with
2 s2
(s) = ims + [eis 1]. (6.2)
2
A Levy process is a Markov process so we can talk about its generator,
E[f (Xt ) | X0 = x] f (x)
Lf (x) = lim .
t0 t

Recall that for a Brownian motion Bt with drift m and variance 2 ,

2 00
Lf (x) = m f 0 (x) + f (x).
2
Moreover, if
f (t, x) = E [F (Bt ) | B0 = x] ,
then f satisfies the heat equation

t f (t, x) = Lx f (t, x). (6.3)

We now compute the generator for the Poisson process. Although we think
of the Poisson process as taking integer values, there is no problem extending
184 CHAPTER 6. JUMP PROCESSES

the definition so that X0 = x. In this case the values taken by the process are
x, x + 1, x + 2, . . . Up to terms that are o(t), P{Xt = X0 + 1} = 1 P{Xt =
X0 } = t. Therefore as t 0,

E[f (Xt ) | X0 = x] = t f (x + 1) + [1 t] f (x) + o(t),

and
Lf (x) = [f (x + 1) f (x)].
The same argument shows that if f (t, x) is defined by

f (t, x) = E [F (Xt ) | X0 = x] ,

then f satisfies the heat equation (6.3) with the generator L. We can view
the generator as the operator on functions f that makes (6.3) hold.
The generators satisfy the following linearity property: if Xt1 , Xt2 are
independent Levy processes with generators L1 , L2 , respectively, then Xt =
Xt1 + Xt2 has generator L1 + L2 . For example, if X is the Levy process with
as in (6.2),

2 00
Lf (x) = m f 0 (x) + f (x) + [f (x + 1) f (x)].
2

6.3 Compound Poisson process

In the Poisson process, the process jumps from x to x + 1 at rate . The
compound Poisson process generalizes this by allowing the jump size to be
random. For the moment we think of the process as having two parameters,
a jump rate and a distribution function F . We construct the process as
follows.

Let T1 , T2 , T3 , . . . be independent random variables each exponential

with rate .

Let Nt denote the corresponding Poisson process as in the previous

section,

Nt = n if T1 + + Tn t < T1 + + Tn+1 .

Let Y1 , Y2 , Y3 , . . . be independent random variables, independent of

T1 , T2 , . . ., with distribution function F , and let

Sn = Y1 + + Yn , S0 = 0.
6.3. COMPOUND POISSON PROCESS 185

Set Xt = SNt .

We call the process Xt a compound Poisson process (CPP) starting at the

origin.
Let # denote the distribution of the random variables Yj , that is, if
V R, then # (V ) = P{Y1 V }. We set

= #

which is a measure on R with (R) = . For the usual Poisson process # ,

are just point masses on the point 1,

# ({1}) = 1, ({1}) = , # (R \ {1}) = (R \ {1}) = 0.

The measure encodes both and # so we can consider as the parameter

for the compound Poisson process. For any measure with = (R) <
there is such a process. The measure is called the Levy measure for the
process. Without loss of generality we assume that ({0}) = 0 since jumps
of size zero do not affect the process Xt . The construction shows that
the increments of Xt are independent, stationary, and hence Xt is a Levy
process.

Proposition 6.3.1. Suppose Xt is a CPP with Levy measure with X0 = 0.

Then Z
(s) = [eisx 1] d(x),

Z
Lf (x) = [f (x + y) f (x)] d(y). (6.4)

Moreover, if Z
2
:= x2 d(x) < , (6.5)

and Z
m= x d(x),

then Mt = Xt tm is a square integrable martingale with Var[Mt ] = t 2 .

Proof. Let Z
isYj
(s) = E[e ]= eisx d# (x),

186 CHAPTER 6. JUMP PROCESSES

denote the characteristic function of a random variable with distribution

function F . Then,

X
isXt
E[e ]= P{Nt = n} E[eisXt | Nt = n].
n=0

Conditioned on Nt = n, the distribution of Xt is that of Y1 + + Yn , and

hence
E[eisXt | Nt = n] = (s)n .
Since Nt is Poisson with mean t,

X (t)n
E[e isXt
] = et (s)n
n!
n=0

t
X (t(s))n
= e
n!
n=0
= exp {t [(s) 1]}
 Z 
isx #
= exp t [e 1] d (x)

 Z 
isx
= exp t [e 1] d(x) .

The second-to-last equality uses the fact that # is a probability measure.

The computation for the generator L follows from the definition of the
CPP. In a small time t, the probability that there is a jump is t +
o(t), and given that there is a jump, the amount jumped is given by the
distribution # . In other words, the probability that there is a jump in time
t whose size is in (a, b) is given up to an error of size o(t) by

t # (a, b) = t (a, b).

Therefore,

E [f (Xt ) | X0 = x]
Z
= [1 t] f (x) + t f (x + y) d(y) + o(t)

Z
= f (x) + t [f (x + y) f (x)] d(y) + o(t),

which is a restatement of (6.4).

6.3. COMPOUND POISSON PROCESS 187

The moments of Xt can be computed by differentiating its characteristic

function,
0Xt (0) = i E[Xt ], 00Xt (0) = E[Xt2 ],
assuming that the moments exist. For any positive integer n, if
Z
|x|n d(x) < ,

then we can differentiate under the integral n times to see that

Z
(n) n
(s) = i xn eisx d(x).

In particular, if (6.5) holds,

Z Z
0 (0) = i x d(x), 00 (0) = x2 d(x),

and since Xt (s) = exp{t(s)} and (0) = 0, we get

Z
0 0
Xt (0) = t (0) = it x d(x) = itm,

00Xt (0) = t00 (0) + [t0 (0)]2 = [t 2 + (tm)2 ].

Therefore,

Var[Xt ] = E (Xt tm)2 = E Xt2 (E[Xt ])2 = t 2 .

   
E[Xt ] = tm,

If Fs denote the information contained in {Xr : r s} and s < t, then

Xt Xs is independent of Fs and

E[Xt | Fs ] = Xs + E[Xt Xs | Fs ] = Xs + E[Xt Xs ] = Xs + (t s)m,

and hence
E[Xt tm | Fs ] = Xs sm.
This shows that Mt is a martingale and

E[Mt2 ] = E[(Xt tm)2 ] = Var[Xt ] = t 2 .

188 CHAPTER 6. JUMP PROCESSES

We call Mt = Xt mt the compensated compound Poisson process (com-

pensated CPP) associated to Xt . It is a Levy process. If L denotes the
generator for Xt , then Mt has generator

LM f (x) = Lf (x) mf 0 (x)

Z
f (x + y) f (x) y f 0 (x) d(y).
 
= (6.6)

The quadratic variation of the CPP is defined as before

X  j  j1 2
 
[X]t = lim X X .
n n n
jtn

The notation hXit might seem more appropriate, but it is standard for jump
processes to change to this notation. Note that the terms in the sum are zero
unless there is a jump in the time interval [(j 1)/n, j/n]. Hence we see that
X
[X]t = [Xs Xs ]2 .
st

Unlike the case of Brownian motion, for fixed t the random variable hXit is
not constant. We can similarly find the quadratic variation of the martingale
Mt = Xt mt. By expanding the square, we see that this it is the limit as
n of three sums
X  j  j1 2
 
X X
n n
jtn
     X 2
2m X j j1 m
+ X X + .
n n n n2
jtn jtn

Since there are only finitely many jumps, we can see that the second and
third limits are zero and hence
X
[M ]t = [X]t = [Xs Xs ]2 .
st

This next proposition generalizes the last assertion of the previous propo-
sition and sheds light on the meaning of the generator L. In some sense, this
is an analogue of Itos formula for CPP. Recall that if Xt is a diffusion
satisfying
dXt = m(Xt ) dt + (Xt ) dBt ,
6.3. COMPOUND POISSON PROCESS 189

then the Markov process Xt has generator

2 (x) 00
Lf (x) = m(x) f 0 (x) + f (x),
2
and Itos formula gives

df (Xt ) = Lf (Xt ) dt + f 0 (Xt ) (Xt ) dBt .

In other words, if
Z t
Mt = f (Xt ) Lf (Xs ) ds,
0

then Mt a (local) martingale satisfying

dMt = f 0 (Xt ) (Xt ) dBt .

Proposition 6.3.2. Suppose Xt is a compound Poisson process with Levy

measure , and suppose that f is a continuous function such that for all x, t,

E f (Xt )2 | X0 = x < .
 

Then Z t
Mt = f (Xt ) Lf (Xs ) ds
0
is a square integrable martingale with
X
[M ]t = [f (X)]t = [f (Xs ) f (Xs )]2 .
st

The proof of this proposition is similar to the derivation of Itos formula.

We will first show that Mt is a martingale, that is, E[Mt | Fs ] = Ms . The
argument is essentially the same for all s so we will assume that s = 0 and
X0 = x. We need to show that
 Z t 
E f (Xt ) f (X0 ) Lf (Xs ) ds = 0.
0

This argument is the same for all t, so let us assume t = 1. Then, as in the
proof of Itos formula, we write
Z 1
f (X1 ) f (X0 ) Lf (Xs ) ds
0
190 CHAPTER 6. JUMP PROCESSES

n
" #
X     Z j/n
= f Xj f X (j1) Lf (Xs ) ds .
n n (j1)/n
j=1

We write the expectation of the right-hand side as the sum of two terms
Xn      1  
E f X j f X (j1) Lf X (j1) (6.7)
n n n n
j=1

n
" #
Z j/n
X 1  
E Lf X (j1) Lf (Xs ) ds , (6.8)
n n (j1)/n
j=1

The definition of L implies that

E [f (Xt+t ) | Ft ] = f (Xt ) + Lf (Xt ) t + o(t),

and hence
E [f (Xt+t ) f (Xt ) Lf (Xt ) t] = o(t).
This shows that the sum in (6.7) has n terms that are o(1/n) and hence
the limit is zero. The terms inside the expectation in (6.8) equal zero unless
there is a jump between time (j 1)/n and j/n. This occurs with probability
O(1/n) and in this case the value of the random variable is O(1/n). Hence
the expectations are O(1/n2 ) and the sum of n of them has limit zero.
The computation of the quadratic variation is essentially the same as in
Proposition 6.3.1.

Suppose that Xt = SNt where Nt is a Poisson process of rate .

Let En be the event that Nt+(1/n) Nt 2 for some t 1. We know
that n En has probability zero, and hence since E[|f (X1 )|] < ,

lim E [|f (X1 )| 1En ] = 0.

n

Therefore,
E[f (X1 )] = lim E[f (X1 )1Enc ].
n
However,
 
E [f (Xt+s ) f (Xt )] 1Enc = s E[Lf (Xt )][1 + O(s)].

We therefore get
n1
1X
E [f (X1 ) f (X0 )] = lim E[Lf (Xj/n )].
n n
j=0
6.3. COMPOUND POISSON PROCESS 191

Right-continuity implies that

n1 Z 1
1X
lim Lf (Xj/n ) = Lf (Xs ) ds,
n n 0
j=0

and the bound E[|f (X1 )|] < can be used to justify the inter-
change of limit and expectation.

For a CPP the paths t 7 Xt are piecewise constant and are discontinuous
at the jumps. As for the usual Poisson process, we have defined the path so
that is it right-continuous and has left-limits. We call a function cadlag (also
written cadlag), short for continue a droite, limite a gauche, if the paths are
right-continuous everywhere and have left-limits. That is, for every t, the
limits
Xt+ = lim Xs , Xt = lim Xs ,
st st

exist and Xt = Xt+ . The paths of a CPP are cadlag. We can write
X
Xt = X0 + [Xs Xs ] .
0st

Although as written this sum is over an uncountable number of times s, the

term Xs Xs is only nonzero at those s at which the path is discontinuous
and this is a finite set.

In the next section, we will use the following maximal lemma.

Lemma 6.3.3. If Xt is a CPP satisfying (6.5), Mt = Xt tm,
and
K = max |Mt |,
0t1

then
2
P{|K| a} .
a2
Proof. Let

Kn = max{|Mj/n | : j = 1, 2, . . . , n}.

Since Xt has piecewise constant paths, K = limn Kn and it

suffices to show for each n that
2
P{|Kn | a} .
a2
192 CHAPTER 6. JUMP PROCESSES

Fix n and let Zj = Mj/n , Z n = max{|Zj | : j = 1, . . . , n}. Then Zj

is a discrete-time martingale, and Corollary 1.7.2 gives

E[Zn2 ] E[M12 ] 2
P{Z n a} = = .
a2 a2 a2

6.4 Integration with respect to compound Poisson

processes
Defining the integral with respect to a compound Poisson process is easy. If
Xt is a CPP and At is another process, we could let
Z t X
As dXs = As [Xs Xs ] . (6.9)
0 0st

As noted before, there are only a finite number of nonzero terms in the sum
so the sum is well defined. This definition requires no assumptions on the
process As . However, if we want the integral to satisfy some of the properties
of the Ito integral, we will need to assume more.
Suppose that E[X1 ] = m, Var[X1 ] = 2 , and let Mt be the square in-
tegrable martingale Mt = Xt mt. Then if the paths of At are Riemann
integrable, and the integral
Z t Z t Z t
Zt = As dMs = As dXs m As ds
0 0 0

is well defined. Let Ft denote the information contained in {Ms : s t}

which is the same as the information contained in {Xs : s t}. In analogy
with the Ito integral, we might hope that if At is square integrable, piecewise
continuous, and adapted to {Ft }, then Zt would be a martingale. However,
this is not always the case as we now show.

Example 6.4.1. Suppose Xt is the CPP that takes jumps of rate 1 and
when it jumps it chooses 1 each with the same probability. In other words,
the Levy measure of X is the probability measure with ({1}) = ({1}) =
1/2. Then m = 0, 2 = 1, and At = Xt Xt is adapted to {Ft }. It is
6.4. INTEGRATION WITH RESPECT TO COMPOUND POISSON PROCESSES193

discontinuous only at the jumps of Xt and hence is piecewise continuous.

However, if we define the integral as in (6.9), then
Z t X
As dXs = [Xs Xs ]2 ,
0 0st

which has positive expectation (since it is nonnegative and has a positive

probability to be strictly positive).

The problem in our setup is that we allow a betting strategy that sees
a jump at time s and immediately changes the bet to take advantage of
this. In our frameweork, we will not allow these instantaneous changes by
restricting to strategies that are left-continuous.

Proposition 6.4.1. Suppose Xt is a CPP with 2 < and Mt = Xt mt.

Suppose that At is a process satisfying:

each At is measurable with respect to {Xs : 0 s t};

with probability one, the paths of At are left-continuous, that is, At =

At ;

for every t < ,

Z t
E[A2s ] ds < .
0

Then Z t
Zt = As dMs
0

is a square integrable martingale with

Z t
Zt2 2
E A2s ds.
   
E = (6.10)
0

Rather than assuming that As is left-continuous, we can assume that As

has cadlag paths, but then we change the definition in (6.9) to
Z t X
As dXs = As [Xs Xs ] .
0 0st
194 CHAPTER 6. JUMP PROCESSES

Proof of Proposition 6.4.1. If s < t, then arguing as before,

E[As (Mt Ms ) | Fs ] = As E[Mt Ms | Fs ] = 0, (6.11)

E[A2s (Mt Ms )2 ] = E E(A2s (Mt Ms )2 | Fs )

 

= E A2s E[(Mt Ms )2 ]
 

= 2 (t s) E[A2s ]. (6.12)

For each t, the probability that Mt 6= Mt = 0. Therefore,

with probability one, for all rational times t, Mt = Mt . Suppose
that As is a simple process that changes values only at times in
{0, 1/n, 2/n, 3/n, . . .}. Then with probability one, for all t, if j/n
t < (j + 1)/n,
Z t j1      
X i i+1 i
Zt = As dMs = A M M .
0 n n n
i=0

This expression uses the fact Mt = Mt for all rational t. Using

this expression and (6.11) and (6.12) we see that Zt is a square
integrable martingale satisfying (6.10).
We now assume that At is adapted, left-continuous, and uni-
formly bounded, that is, with probability one, |At | C for all t.
Define Ant by
j j+1
Ant = Aj/n if <t .
n n
Left-continuity implies that with probability one for all t

At = lim Ant ,
n

and hence Z t Z t
As dMs = lim Ans dMs .
0 n 0

Using uniform boundedness, one can interchange limits and expec-

tations to get (6.11) and (6.12).
Finally to remove the boundedness assumption, we let Tn =
inf{t : |At | = n} and let At,n = AtTn . As n , At,n At ,
and we can argue as before.
6.4. INTEGRATION WITH RESPECT TO COMPOUND POISSON PROCESSES195

Example 6.4.2. Suppose Xt is a CPP with Levy measure satisfying

Z
e2y d(y) < .

Let f (x) = ex and St = f (Xt ) = eXt which we can consider as a sim-

ple model of an asset price with jumps. This is an analogue of geometric
Brownian motion for jump processes. Note that
f (x + y) f (x) = ex h(y) where h(y) = ey 1,
and hence Z
Lf (x) = [f (x + y) f (x)] d(y) = rex ,

where Z
r= h(y) d(y) < .

Note that
St St = St h(Xt Xt ). (6.13)
In particular, the jump times for S are the same as the jump times for X.
Let X
Xt = h(Xt Xt )
st

be the CPP with X0 = 0 that takes a jump of size h(y) whenever Xt takes
a jump of size y. This process has Levy measure where
(h(V )) = (V ), (6.14)
with mean m = r. If Yt = Xt r t is the compensated CPP, we can write
(6.13) as
h i
dSt = St dXt = St dYt + r St dt = St dYt + r dt .

Since St is piecewise continuous, it does not matter in the dt integral whether

we write St or St . However, we must write St in the dXt and dYt integrals.
If r = 0, that is, if
Z
[ey 1] d(y) = 0,

then Xt and hence St are actually martingales. If r 6= 0, we let St = ert St =

eXt rt . Since t 7 ert is differentiable, we can use the product rule
dSt = St d[ert ] + ert dSt = St dYt .
196 CHAPTER 6. JUMP PROCESSES

Example 6.4.3. We will proceed backwards through the last example to

solve the exponential differential equation for compound Poisson processes.
Suppose Xt is a CPP whose Levy measure satisfies {x : x 1} = 0
and Z
x2 d(x) < .
1

Let Z
m = x d(x) < ,
1

and let Yt = Xt m t be the compensated CPP which is a square integrable

martingale. Let be the measure as defined in (6.14) and let Xt denote the
corresponding process as in the previous example. In other words, if Xt has
a jump of size y, then Xt has a jump of size h1 (y) = log[1 + y]. Then the
solutions to the exponential differential equation

dZt = Zt dXt , dMt = Mt dYt ,

are
Zt = Z0 eXt , Mt = Z0 eXt mt .

Note that we can write

X
Zt = Z0 exp log[1 + (Xs Xs )] .

0<st

Example 6.4.4. More generally, the equations

dZt = At Zt dXt , dMt = At Mt dYt

where X is a CPP with compensated process Yt = Xt mt, have solutions

X
Zt = Z0 exp log[1 + As (Xs Xs )] ,

0<st

 Z t   Z t 
Mt = Zt exp m As ds = Zt exp m As ds .
0 0
6.5. CHANGE OF MEASURE 197

6.5 Change of measure

Suppose Xt is a CPP defined on the probability space (, F, P) with Levy
measure and corresponding filtration {Ft }. We would like to do the ana-
logue of the Girsanov theorem to change the measure.
Suppose that strictly positive martingale Mt is a strictly positive mar-
tingale with respect to {Ft } with M0 = 1. As before, we can define a new
probability measure Q by saying that if V is Ft -measurable, then

Q[V ] = E [Mt 1V ] .

If we let Pt , Qt denote P, Q restricted to Ft , then Pt and Qt are mutually

absolutely continuous with
dQt dPt 1
= Mt , = .
dPt dQt Mt

If is a another measure on R, we can ask if we can find a martingale so

that under the measure Q, Xt is a CPP with Levy measure . A necessary
condition for this to hold is that and are mutually absolutely continuous
measures on R. Indeed, if V is a subset of R with (V ) > 0, (V ) = 0, then
Xt under P has positive probability of having Xs Xs V for some s 1,
while this has zero probability under Q.
Maybe a bit surprising is the fact that the converse holds. Suppose ,
are mutually absolutely continuous and let f = d/d which we assume is
strictly positive. With respect to P, Xt satisfies

P{Xt+t Xt V } = (V ) t + o(t),

and we would like

Q{Xt+t Xt V } = (V ) t + o(t),

which would give

Q{Xt+t Xt V } (V )
= + o(t).
P{Xt+t Xt V } (V )

To get this we choose

X
Mt = exp f (Xs Xs ) rt , r = (R) (R).

st
198 CHAPTER 6. JUMP PROCESSES

Suppose we model an asset price by St = eXt where Xt is a CPP, and

suppose that is any measure equivalent to with
Z
[ey 1] d(y) = 0.

Then Q is equivalent to P, and (see Example 6.4.2) S is a martingale with

respect to Q. In particular, there are many different measures Q that are
equivalent to P that make St a martingale. This is in contrast to the case
when Xt is a Brownian motion where such a measure is unique.

6.6 Generalized Poisson processes I

The generalized Poisson process is like the CPP except that the jump times
form a dense subset of the real line. Suppose that is a measure on R.
We want to consider the process Xt that, roughly speaking, satisfies the
condition that the probability of a jump whose size lies in [a, b] occurring in
time t is about [a, b] t. The CPP assumes that
Z
(R) = d(x) < ,

and this implies that there are only a finite number of jumps in a bounded
time interval. In this section we allow an infinite number of jumps ((R) =
), but require the expected sum of the absolute values of the jumps in an
interval to be finite. This translates to the condition
Z
|x| d(x) < . (6.15)

This implies that

{x : |x| > } < for every  > 0, (6.16)
but it is possible for {x : |x| > 0} to be infinite.
To be more specific, let  denote the measure restricted to jumps
of absolute value strictly greater than . For each , we have a compound
process with Levy measure  which can be written as
X
Xt = (Xs Xs

).
0st

Let us write X 
Vt = Xs Xs
 

,
0st
6.6. GENERALIZED POISSON PROCESSES I 199

which is a CPP with measure  defined by

 (a, b) =  (a, b) +  (b, a),  a < b < .

In other words Vt has the same jumps as Xt except that they all go in the
positive direction. As  0, Vt increases (since we are adding jumps and
they are all in the same direction) and using (6.15), we see that
Z

lim E [Vt ] = t |x| d(x) < .
0

Hence we can define

Vt = lim Vt
0

and E[Vt ] < . Similarly, we can define

Xt = lim Xt .
0

Since |Xt | Vt , the dominated convergence theorem is used to justify in-

terchanges of limits.
1/n
Let Tn be the set of times such that the process Xt is discontinuous,
and let

[
T = Tn .
n=1

Since each Tn is finite, the set T is countable, but if (R) = , T will be

infinite and, in fact, dense in every interval. Even though there are jumps in
every interval, the paths are cadlag, that is, they are right continuous and
the limits
Xt = lim Xs
st

exist for every s.

Calculating as in the previous section we can see the following.
Proposition 6.6.1. Suppose Xt is a generalized Poisson process with Levy
measure . Then Z
(s) = [eixs 1] d(x),

Z
E[Xt ] = tm, m= x d(x),

Z
Lf (x) = [f (x + y) f (x)] d(y),

200 CHAPTER 6. JUMP PROCESSES
Z
2
Var[Xt ] = E[(Xt mt) ] = t x2 d(x).

One may worry about the existence of the integrals above. Tay-
lors theorem implies that
x2 s2
eixs 1 = ixs + O(|xs|3 ).
2
Using this and (6.15) we can see that
Z
|eixs 1| d(x) < .
|x|1

Also, Z
|eixs 1| d(x) 2 {x : |x| > 1} < .
|x|>1

For the generator L we assume that f is C 1 and hence for y near

the origin
f (x + y) f (x) = O(|y|).
Then, Z
|f (x + y) f (x)| d(y) < .
|x|1
Finiteness of the integral
Z
|f (x + y) f (x)| d(y)
|x|>1

requires some assumptions on the growth of f at infinity.

We can relax the conditions somewhat by requiring (6.16) to hold but

replacing (6.15) with the weaker condition
Z 1
|x| d(x) < . (6.17)
1

Indeed, if (6.16) and (6.17) hold we can write

= 1 + 2 ,
where 1 is restricted to {|x| 1} and 2 is restricted to {|x| > 1}. To
1 we associate the generalized Poisson process Xt1 and to 2 we associate
an independent CPP, and then set Xt = Xt1 + Xt2 . In this case it is possible
that E[|Xt |] = . However, the formulas for and L are the same as above.
6.6. GENERALIZED POISSON PROCESSES I 201

Example 6.6.1 (Positive Stable Processes). Suppose that 0 < < 1 and
is defined by
d(x) = c x(1+) dx, 0 < x < ,
where c > 0. In other words, the probability of a jump of size between x and
x + x in time t, is approximately

c x(1+) (t) (x).

Note that Z
c
c x(1+) dx =  ,

Z 1 Z 1
c
x d(x) = c x dx = < ,
0 0 1
Z Z
x d(x) = c x dx = .
0 0

Therefore, satisfies (6.16) and (6.17), but not (6.15). The corresponding
generalized Poisson process has
Z
(s) = c [eisx 1] x(1+) dx.
0

With careful integration, this integral can be computed. We give the answer

in the case = 1/2, c = 2/ , in which case

(s) = |s|1/2 [1 is], s > 0.

We see that X1 has a Levy distribution as in Example 6.1.1.

Using the change of variables y = rx, we can see that if r > 0, then
Z Z
(1+)
(rs) = c [e irsx
1] x dx = cr
[eisy 1] y (1+) dy.
0 0

This implies that the distribution of rX1 is the same as that of Xr . In

particular, X1/n has the same distribution as n1/ X1 and hence we can
write
Z1 + + Zn
X1 = , Zj = n1/ [Xj/n X(j1)/n ],
n1/
where Z1 , . . . , Zn are independent with the same distribution as X1 . The
process Xt is called the positive stable process with exponent .
202 CHAPTER 6. JUMP PROCESSES

Example 6.6.2 (Gamma process). Suppose > 0 and is defined by

ex
d(x) = dx, 0 < x < .
x
Note that Z Z
d(x) = , x d(x) < .
0 0

Using a table of integrals, we can see that the characteristic exponent is

e(is)x ex
Z Z
isx
(s) = [e 1] d(x) = dx = log .
0 0 x is

In other words, the characteristic function of Xt is Xt (s) = [/( is)]t .

This is the characteristic function of a Gamma random variable with pa-
rameters and t, that is, the density of Xt is

ft (x) = (x)t1 ex , 0 < x < 1,
(t)

where Z
(t) = xt1 ex dx,
0

is the Gamma function. Important values of t are the following.

X1 has the distribution of an exponential random variable with rate

.

If n is a positive integer, then Xn has the distribution of the sum of

n independent random variables each exponential with rate . If Ns
is a Poisson process with rate with N0 = 0, then Xn has the same
distribution as inf{s : Ns = n}.

If = 1/2 and n is a positive integer, then Xn/2 has a 2 distribution

with n degrees of freedom, that is, it has the same distribution as

Z12 + + Zn2 ,

where Z1 , . . . , Zn are independent, standard normal random variables.

The process Xt is called the Gamma process with parameter .

6.7. GENERALIZED POISSON PROCESSES II 203

6.7 Generalized Poisson processes II

In the previous section, we assumed that the expected sum of the absolute
values of the jump was finite. There are jump Levy processes that do not
satisfy this condition. These processes have many small jumps. Although
the absolute values of the jumps are not summable, there is enough cancel-
lation between the positive jumps and negative jumps to give a nontrivial
process.
We will construct such a process associated to a Levy measure satisying
the following assumptions.

For every  > 0, the number of jumps of absolute value greater than
 in a finite time interval is finite. More precisely, if  denotes
restricted to {|x| > }, then

 (R) = {x : |x| > } < for every  > 0. (6.18)

All jumps are bounded in absolute value by 1, that is,

{x : |x| > 1} = 0. (6.19)

The second moment of the measure is finite,

Z
2
:= x2 d(x) < . (6.20)

Note that (6.20) can hold even if (6.15) does not hold.

Let Z Z
m = x d (x) = x d(x),
<|x|1

which is finite and well defined for all  > 0 by (6.18)(6.20). For each
 > 0, there is a CPP Xt associated to the measure  and a correponding
compensated Poisson process Mt = Xt m t.

Definition Suppose is a measure on R satisfying (6.18)(6.20). Then the

compensated generalized Poisson process (CGPP) with Levy measure is
given by
Yt = lim Mt = lim[Xt m t]. (6.21)
0 0

Two important cases are the following.

204 CHAPTER 6. JUMP PROCESSES

If Z
|x| d(x) < ,

then the (uncompensated) generalized Poisson process

Xt = lim Xt ,
0

as in the previous section exists. In this case

Yt = Xt mt,

where Z
m = lim m = x d(x).
0

If is symmetric about the origin, that is, if for every 0 < a < b the
rate of jumps with increments in [a, b] is the same as that of jumps in
[b, a], then m = 0 for every  and

Yt = lim Xt .
0

More generally, the last equation holds if m 0. In this case Yt is

called a pure jump process.

Proposition 6.7.1. If Yt is a CGPP with Levy measure , then

Z 1
(s) = [eisx 1 isx] d(x),
1
Z
Lf (s) = [f (x + y) f (x) yf 0 (x)] d(y).

The latter quantity is defined provided that f is a C 2 function. If also

E[f (Yt )2 ] < , then
Z t
Mt = f (Yt ) Lf (Ys ) ds
0

is a square integrable martingale with

X
[M ]t = [f (Ys ) f (Ys )]2 .
st
6.7. GENERALIZED POISSON PROCESSES II 205

Proof. Let X  = X1 and M  = M1 = X  m . Let  = M and note that

Z

(s) = X (s) im s = [eisx 1]d(x) im s.
|x|>

Since Z
m = x d(x),
|x|>

we can write Z
 (s) = [eisx 1 isx] d(x).
|x|>

Using Taylor series, we see that |eisx 1 isx| = O(s2 x2 ), and hence (6.20)
implies that for each s,
Z
|eisx 1 isx| d(x) < .

This allows us to use the dominated convergence theorem to conclude that

Z 1

lim (s) = [eisx 1 isx] d(x).
0 1

Since Y = lim0 M  , we can conclude that this limit equals (s).

The second equation is obtained by taking the limit of (6.6).

A particular case of the proposition is f (x) = x in which we conclude

that Yt is a square integrable martingale with
X
[Y ]t = [Ys Ys ]2 .
st

The same argument shows that the moment generating function also exists
and  Z 1 
sYt sx
E[e ] = exp t [e 1 sx] d(x) , s R.
1

In particular, E[esYt ] < for all s, t. (This is in contrast to CCPs for which
this expectation could be infinite.)
206 CHAPTER 6. JUMP PROCESSES

We will establish existence of the limit in (6.21) both in L2 and

with probability one. Let us first consider the case t = 1. Let

Y  = Y1 = X1 m .

It suffices to prove the existence of

Y = lim Y  .
0

n (n+1)
Let Zn = Y 2 Y2 , so that

X
Y = Zn .
n=0

The random variables Zn are independent, mean zero, and

X
Var[Zn ] < .
n=0

Zn converges to a random variable Y in both L2 and

P
Therefore
with probability one (see Exercise 1.14). If 2n  < 2n+1 , then
n
" # Z n+1
X 2
E (Y  Zk )2 x2 d(x),
k=0 2n

which goes to zero as n . Hence

lim E |Y  Y |2 = 0.
 
0

Convergence with probability one can be deduced from Lemma

6.3.3.
We now define Yq as in (6.21) for q D as the limit with
probability one. For other t we define

Yt+ = lim Yq , Yt = lim Yq . (6.22)

qt qt

Here we the shorthand that when q is the index, the limit is taken
over D only. Define the jump time T to be the set of times s such
that X  is discontinuous and T = j=1 T1/j . With probability one
each T is finite and hence T is countable.
6.8. THE LEVY-KHINCHIN CHARACTERIZATION 207

Proposition 6.7.2. With probability one, the limits in (6.22)

exists for all t. Moreover, they are equal unless t T .

Proof. Let
Z
S2 2
x2 d(x).
 
= E (Y Y ) =
|x|

Let
M  = sup |Yq Yq | : q D, q 1 .


Using Lemma 6.3.3 (first for Yt Yt and then letting 0), we
can see that
S2
P{M a} 2 .
a
We can find n 0 such that

X
n2 S2n < .
n=1

Then if Mn = M n , by the Borel-Cantelli lemma, with probability

one, for all n sufficiently large Mn 1/n and this inequality implies
that for all t,  
  1
n 
lim sup Yq+ lim Yq+  .

 qt qt  n

Similar estimates hold for the liminf and for limits from the left.
We now define Yt for all t by Yt = Yt+ , and then the paths of
Yt are cadlag.

6.8 The Levy-Khinchin characterization

The next theorem tells us that all Levy processes can be given as independent
sums of the processes we have described.

Theorem 6.8.1. Suppose Xt is a Levy process with X0 = 0. Then Xt can

be written as
Xt = m t + Bt + Ct + Yt , (6.23)
where m R, 0, and Bt , Ct , Yt are independent of the following types:
208 CHAPTER 6. JUMP PROCESSES

Bt is a standard Brownian motion,

Ct is a compound Poisson process with Levy measure C satisfying

C {|x| 1} = 0,

C (R) = C {|x| > 1} < ,

Yt is a compensated generalized Poisson process with Levy measure Y

satisfying
Y ({0}) = 0, Y {|x| > 1} = 0,
Z
x2 dY (x) < .

The process Ct +Yt is called a jump Levy process and if the compensating
term m in the definition of Yt is zero it is a pure jump process. We can
summarize the theorem by saying that every Levy process is the independent
sum of a Brownian motion (with mean m and variance 2 ) and a jump
process (with Levy measure = C + Y .)
We have not included the generalized Poisson process in our decompo-
sition (6.23). If Xt is such a process we can write it as

Xt = Ct + Xt

where Ct denotes the sum of jumps of absolute value greater than one. Then
we can write Xt = Yt + mt where Yt is a compensated generalized Poisson
process and m = E[X1 ], which by the assumptions is finite.
The conditions on C , Y can be summarized as follows:
Z
({0}) = 0, [x2 1] d(x) < . (6.24)

To specify the distribution of a Levy processes, one needs to give m R, >

0 and a measure satisfying (6.24). We then set C to be restricted to
{|x| > 1} and Y to be restricted to {|x| 1}. We have already shown
that (m, , ) determine a Levy process. To prove the theorem, therefore,
we need to show that any Levy process can be written like this.
We will not give the entire proof of this theorem, but let us sketch how
one derives it. Suppose Xt is a Levy process. Suppose that the process has
jumps. For every a > 0, we can write

Xt = Xt,a + Ct,a ,
6.8. THE LEVY-KHINCHIN CHARACTERIZATION 209

where Ct,a denotes the movement by jumps of absolute value greater than
a and Xt,a denotes a Levy process with all jumps bounded by a. For each a
one can show that Ct,a is a CPP independent of Xt,a . We let a go to zero,
and after careful analysis we see that

Xt = Zt + Ct,1 + Yt ,

where Yt is compensated generalized Poisson process and Zt is a Levy process

with continuous paths. We then show that Zt must be a Brownian motion.
The following theorem is related and proved similarly.
Theorem 6.8.2. (Levy-Khinchin) If X has an infinitely divisible distribu-
tion, then there exist m, , C , Y satisfying the conditions of Theorem 6.8.1
such that
2 2
Z
X (s) = ims s + [eisx 1] dC (x)
2 |x|>1
Z
+ [eisx 1 isx] dY (x).
|x|1

We will prove the following fact which is a key step in the proof
of Theorem 6.8.1.
Theorem 6.8.3. Suppose Xt is a Levy process with continuous
paths. Then Xt is a Brownian motion.
Proof. All we need to show is that X1 has a normal distribution.
Let
Xj,n = Xj/n X(j1)/n ,
Mn = max {|Xj,n | : j = 1, . . . , n} .
Continuity of the paths implies that Mn 0 and hence for every
a > 0, P{Mn < a} 1. Independence of the increments implies
that
 n
P{Mn < a} = 1 P{|X1/n | a}

exp nP{|X1/n | a} .

Therefore, for every a > 0,

lim n P{|X1/n | a} = 0. (6.25)

n
210 CHAPTER 6. JUMP PROCESSES

This implies that there exists a sequence an 0 with

lim n P{X1/n an } = 0. (6.26)

n

We claim that all the moments of X1 are finite. To see this let
J = max0t1 |Xt | and find k such that P{J k} 1/2. Then
using continuity of the paths, by stopping at the first time t that
|Xt | = nk, we can see that

P{J (n + 1)k | J nk} 1/2,

and hence
P{J nk} (1/2)n ,
from which finiteness of the moments follows. Let m = E[X1 ], 2 =
Var[X1 ], and note that E[X12 ] = m2 + 2 . Our goal is to show that
X1 N (m, 2 ).
Let
n
X
Xj,n = Xj,n 1{|Xj,n | an }, Zn = Xj,n .
j=1

Let n denote the characteristic function of Xj,n and hence [n ]n is

the characteristic function of Zn . It follows from (6.26) that Zn
X1 in distribution, so it suffices to prove that for every s
2s
 
n
lim [n (s)] = exp ims .
n 2

Also, it is easy to check that E[Zn ] m, Var[Zn ] 2 and hence

 2
2

m [1 + o(1)] 2 m
E[X1,n ] = , E[X1,n ] = + (1 + o(1)).
n n2 n

Also, since|X1,n | an ,

m2 2
 
3 2
E[|X1,n | ] an E[X1,n ] = an + [1 + o(1)].
n2 n
(This estimate uses (6.26) which in turn uses the fact that the paths
are continuous.) Using these estimates, we see that for fixed s,

ims 2 s
 
1
n (s) = 1 + +o ,
n 2n n
6.9. INTEGRATION WITH RESPECT TO LEVY PROCESSES 211

where the o() term depend on s. This implies that

 n
ims 2 s

n 1
lim n (s) = lim 1 + +o
n n n 2n n
2 2
 
s
= exp ims .
2

6.9 Integration with respect to Levy processes

Suppose Xt is a Levy process which by Theorem 6.8.1 we can write as
Xt = m t + Bt + Ct + Yt .
We will define the integral Z t
As Xs
0
by Z t Z t Z t Z t
m As ds + As dBs + As dCs + As dYs .
0 0 0 0
The first integral is a Riemann integral, the second is an Ito integral as
in Chapter 3, the third integral was defined in Section 6.4, and this leaves
us only the fourth integral to define. As we saw in Section (6.4), the third
integral does not have the properties we want unless we assume that At is
left continuous. The same problems exists for the fourth integral so we will
restrict to such processes.
Assume that Yt is a process with Levy measure satisfying (6.18)(6.20).
Then Yt is a square integrable martingale with E[Yt ] = 2 t (this is not the
same as in the previous paragraph). This will allow us to define
Z t
As dYs
0

as an Ito integral as in Chapter 3.

We start with simple processes. Suppose At is a simple process as in
Section 3.2.2. To be specific, suppose that times 0 t0 < t1 < < tn <
are given, and As = Atj t for tj s < tj+1 . Then, we define
Z t j1
X
As dYs = Ati [Yti+1 Yti ] + Atj [Yt Ytj ].
0 i=0
212 CHAPTER 6. JUMP PROCESSES

With probability one, the paths of Yt are continuous at the times t1 , . . . , tn

so the definition above does not care whether we choose the simple process
As to be right continuous or left continuous. The following proposition is
proved in the same way as Proposition 3.2.1.
Proposition 6.9.1. Suppose At , At are bounded simple processes adapted
to the filtration of the process Yt .
Linearity. If a, b are constants, then aAt +bAt is also a simple process
and Z t Z t Z t
(aAs + bAs ) dYs = a As dYs + b As dYs .
0 0 0
Moreover, if 0 < r < t,
Z t Z r Z t
As ds = As dYs + As dYs .
0 0 r

Martingale property. The process

Z t
Zt = As dYs
0

is a martingale with respect to {Ft }.

Variance rule. Zt is square integrable and

Z t
2
 2
Var[Zt ] = E Zt = E[A2s ] ds.
0

Cadlag paths. With probability one, the function t 7 Zt is cadlag.

We now suppose that As is an adapted process with left continuous paths.
If As is an adapted process with cadlag paths, we could take As = As . We
(n)
approximate As by a simple process As by choosing
Z j2n
j j+1
A(n)
s = 2n
Ar dr if n < r n .
(j1)2n 2 2

The integral is defined by

Z t Z t
As dYs = lim A(n)
s dYs .
0 n 0
6.10. SYMMETRIC STABLE PROCESS 213

Let us give be more precise about this definition. Suppose that

At is a process with left continuous paths with right limits, that is,
the left continuous version of a processes with cadlag paths. Assume
that exists K < such that with probability one. |At | K for
all t. Since with probability one, the paths are left continuous, we
know that
(n)
lim At = At .
n
(n)
SInce |At | K, we can used dominated convergence (for
Lebesgue measure) to show that with probability one,
Z t
lim |A(n) 2
s As | ds = 0,
n 0

and dominated convegence for P to see that

Z t  Z t h i
(n) 2
lim E |As As | ds = lim E |A(n)
s A s |2
ds = 0.
n 0 n 0

As in the case for Brownian motion, we use the proposition to see

that for fixed t, the sequence of random variables
Z t
(n)
Zt := A(n)
s dAs ,
0

is a Cauchy sequence in L2 and has a limit Zt in L2 .

We use this method to define the integral at dyadic times t, an
we define it at other times by

Zt = lim Ztn
tn t

where tn are dyadic rationals.

6.10 Symmetric stable process

An important example of a Levy process is the symmetric stable process.
A random variable X has a symmetric stable distribution with parameter
(0, 2] if X (s) = |cs| for some c > 0. If = 2, this means that X
has a normal distribution with mean zero and variance 2c2 . For 0 < < 2
these distributions arise in Levy processes that we call symmetric -stable
processes. The term stable comes from the following property.
214 CHAPTER 6. JUMP PROCESSES

Proposition 6.10.1. If X has a symmetric -stable distribution and

X1 , . . . , Xn are independent random variables with the same distribution as
X, then X has the same distribution as

X1 + + Xn
Y = .
n1/

Proof. Using properties of characteristic functions, we get

n
X
Y (s) = Xj (s/n1/ ) = nX (s/n1/ ) = X (s).
j=1

Definition If 0 < 2, then a Levy process Xt is a symmetric stable

process if
(s) = |cs|

for some c > 0.

Proposition 6.10.2. Suppose 0 < < 2, C > 0 and Xt is a Levy process

with Levy measure
C
d(x) = dx. (6.27)
|x|1+
Then Xt is a symmetric -stable process with (s) = Cb |s| where

1 cos y
Z
b = 2 dy < .
0 y 1+

The condition 0 < < 2 is needed to guarantee that as defined in

(6.27) satisfies (6.24). Since 1 cos(y) = O(y 2 ) as y 0, we can see that
b is finite. In fact, it can be calculated

b = ,
( + 1) sin(/2)

where denotes the Gamma function. The proof of the proposition is a

simple computation. We write
Z Z
isx dx dx
(s) = C [e 1 isx] 1+ + C [eisx 1] 1+ .
|x|1 |x| |x|>1 |x|
6.10. SYMMETRIC STABLE PROCESS 215

Using eisx = cos(sx) + i sin(sx) and the fact that is symmetric about the
origin, we see that

cos(sx) 1
Z
(s) = 2C dx
0 x1+
Z
cos(y) 1
= 2C|s| dy = Cb |s| .
0 y 1+

Suppose Xt has a symmetric -stable process with Levy measure as in

(6.27) with C = 1/b . It can be shown that X1 has a symmetric density g ,
Z b
P{a X b} = g (x) dx.
a

This density is bounded. Unfortunately, except for the case = 1, there

is no explicit form expression for the density. Despite this fact, it is known
that
1
g (x) , |x| ,
|x|1+
and hence Z
2
P{|X1 | K} g (x) dx = .
|x|K K

This comes from the fact that the easiest way for |X1 | to be unusually large
is for there to be a single very large jump, and the probability of a jump of
size at least K by time one is asymptotic to {|x| K}.
If = 1, then b = and the corresponding Levy measure is

1
d(x) = .
x2

The density of X1 is the Cauchy density

1
g1 (x) = .
(x2 + 1)

The easiest way to check this is to compute the characteristic function of gt

and obtain e|s| .
Symmetric stable distribution arise as limit distributions for sums of
symmetric random variables with heavy tails. The next proposition gives
one version.
216 CHAPTER 6. JUMP PROCESSES

Proposition 6.10.3. Suppose , c > 0 and X1 , X2 , . . . are independent,

identically distributed random variables with a bounded density f satisfying
f (x) = f (x) and
c
f (x) 1+ , x , (6.28)
x
where means the ratio of the two sides converges to one.

If 0 < < 2, and

X1 + + Xn
Zn = , (6.29)
n1/
then Zn converges in distribution to an -symmetic random variable.

If > 2, and
X1 + + Xn
Zn = ,
n
then Zn converges in distribution to a centered normal random vari-
able.

If = 2, and
X1 + + Xn
Zn = ,
n log n
then Zn converges in distribution to a centered normal random vari-
able.

If > 2, then Var[Xj ] < , and hence the result is a restatement of the
central limit theorem.

We will prove the proposition for 0 < < 2. Our first obser-
vation is that if X1 , X2 , . . . are independent, identically distributed
random variables whose characteristic exponent satisfies

(s) = r |s| + o(|s| ), s 0, (6.30)

then Zn as defined in (6.29) converges to an -stable distribution.

This follows since for each s,

lim Zn (s) = lim n (s/n1/ ) = r |s| .

n n

Therefore, we only need to show that (6.28) implies (6.30).

6.10. SYMMETRIC STABLE PROCESS 217

Let denote the characteristic function of X1 , and note that

(s) = log (s) (s) 1 as s 0. By changing variables,
we see that if s > 0,
Z
(s) 1 = [eisx 1] f (x) dx

Z
= 2 [cos(sx) 1] f (x) dx
0
Z
1
= 2s [cos y 1] f (y/s) dy
0
Z
= 2s [cos y 1] s(1+) f (y/s) dy
0

We claim that
Z
lim [cos y 1] s(1+) f (y/s) dy = cI,
s0 0

where
cos y 1
Z
I= dy,
0 y 1+
from which (6.30) follows with r = 2cI. To see this, let  = (2
)/6 > 0 and note that
Z 1  Z s1
 s 
(1+) (1+)
[cos y 1] s f (y/s) dy C s y 2 dy
 
 
 0  0

Cs(2)/2 0.

Also, for y s1 ,

f (y/s) = c (y/s)(1+) [1 + o(1)],

and hence
Z
lim [cos y 1] s(1+) f (y/s) dy
s0 s1
Z
= c lim [cos y 1] y (1+) dy = cI.
s0 s1

Proposition 6.10.3 only uses stable processes for 2. The next propo-
sition that shows that there are no nontrivial stable process for > 2.
218 CHAPTER 6. JUMP PROCESSES

Proposition 6.10.4. Suppose 0 < < 1/2, and X1 , X2 , . . . are indepen-

dent, identically distributed each having the same distribution as
X1 + + Xn
.
n
Then P{Xj = 0} = 1.

We use the following lemma to prove Proposition 6.10.4.

Lemma 6.10.5. Suppose X is a nonconstant random variable

with characteristic function . Then there exists  > 0 such that
if |s| , then
|(s)| 1  s2 .

Proof. Let us first assume that X is bounded with mean m and

variance 2 > 0. Then, the characteristic function satisfies

2 2
 
ims 3
(s) = e 1 s + O(|s| ) , s 0,
2

and hence there exists  > 0 such that

2 2 
 

|(s)| 1
 s , |s| .
4 

For general nonconstant X, we can find a bounded interval I such

that P{X I} = > 0 and such the variance 2 of the conditional
distribution given X I is positive. Then,

(s) = E[eisX | X I] + (1 ) E[eisX | X 6 I].

There exists  > 0 such that if |s| , |E[eisX | X I]|

1 ( 2 s2 /4) and hence

2 2 2 2
 
|(s)| 1 s +1=1 s .
4 4

To prove Proposition 6.10.4, let denote the characteristic func-

tion of Xj . Then
(s) = (n s)n .
6.11. EXERCISES 219

If the distribution of Xj is nontrivial, then there exists  > 0 such

that for |s| ,
|(s)| 1  s2 ,

and hence for every s 6= 0,

 n
|(s)| = lim |(n s)|n 1  n2 s2
lim
n n
n o
lim exp  s2 n12 = 0.
n

Since is continuous at 0 with (0) = 1, this is a contradiction.

Therefore, P{Xj = c} = 1 for some c and since c = (c+ +c)/n ,
c = 0.

6.11 Exercises
Exercise 6.1. Suppose F (t) = P{T t} is the distribution function of a
continuous random variable T . Define theinverse function G by

G(r) = sup{t : F (t) r}.

1. Suppose U is a random variable distributed uniformly on [0, 1]. Show

that G(U ) has distribution function F .

2. Find G in the case that T is exponential with rate .

3. Use the waiting time method as described in Section 6.2 to sample

from a Poisson process Xt with rate = 3/2. Run at least 3000 trials
to estimate the distributon of X3 .

4. Compare the simulations with the actual distribution for X3 .

Exercise 6.2. Suppose Yt is a Poisson process with parameter 2.

1. Find P{Y3 2}.

2. Find P{Y4 Y1 + 2 | Y1 = 4}.

3. Find P{Y1 = 1 | Y3 = 4}.

220 CHAPTER 6. JUMP PROCESSES

4. Find a nonrandom function a(t) such that

Xt = Yt a(t)

is a martingale, that is, if s t,

E[Xt | Fs ] = E[Xt | Ys ] = Xs .

Here Ft denotes the information in {Ys : s t}.

Exercise 6.3. Suppose Xt is a compound Poisson process with = 2, and
measure # is given by a N (0, 1) distribution. In other words, the process
jumps when a Poisson process of rate 2 jumps and when it jumps it chooses
a jump size from a N (0, 1) distribution.

Find the function f so that the Levy measure can be written as

d = f (x) dx.

Find E[Xt ].

Find E[Xt2 ]

What is the generator L of the process Xt ?

Does there exist a process St such that for each t, St is measurable

with respect to Ft , the information in {Xs : 0 s t}, and such that

With probability one, St is differentiable with respect to t,

Mt = Xt2 St is a martingale

If so find it.

Exercise 6.4. Let Xt be as in Exercise 6.3 and let

Zt = exp{Xt }.

1. Find E[Zt ].

2. Find a continuous process St , adapted to the information in {Ft }, with

St > 0 such that St is differentiable with respect to t (except at the t
at which Xt jumps) and

Mt = Zt St

is a martingale? Explain why St is not differentiable at the jump times

of Xt .
6.11. EXERCISES 221

3. Does there exist a process At , adapted to the information in {Ft },

with At > 0 such that At is differentiable with respect to t (except at
the t at which Xt jumps) and

Mt = Z t A t

is a martingale?
Exercise 6.5. Suppose Yt is a Cauchy process, that is, a Levy process such
that Y1 has a Cauchy distribution. Show why the following statement holds:
for every r > 0 and t > 0,
 
P max Ys > r () 2 P{Yt > r}.
0st

Here () is one of the following: >, =, <. Your task is to figure out which
of >, =, < should go into the statement and explain why the relation holds.
(Hint: go back to the derivation of the reflection principle for Brownian
motion. The only things about the Cauchy process that you should need to
use are that it is symmetric about the origin and has jumps. Indeed, the
same answer should be true for any symmetric Levy process with jumps.)
Exercise 6.6. Suppose Xt is a Poisson process with rate 1 and let r > 0
with filtration {Ft }, and
St = eXt rt .
1. Find a strictly positive martingale Mt with M0 = 1 such that St is a
martingale with respect to the tilted measure Q given by

Q(V ) = E [1V Mt ] , V Ft .

2. Is your choice of Mt and Q unique?

Exercise 6.7. Suppose f : R R is a function that is right continuous at
every point. Let
Kf = sup{|f (x)| : 0 x 1}.
1. Give an example to show that it is possible that Kf = .

2. Show that if f is a cadlag function, then Kf < .

Exercise 6.8. Suppose Xt is a generalized Poisson process with Levy mea-
sure given by
dx
d(x) = , 0 < x < 1.
x
222 CHAPTER 6. JUMP PROCESSES

1. Is this a generalized Poisson process of Type I?

2. Find E[Xt ], Var[Xt ].

3. What is the probability that there are no jumps of size greater than
1/2 by time t = 2?

4. Let
f (t, x) = E Xt4 | X0 = x .
 

Find the function g(x) such that

d
f (t, x) |t=0 = g(x).
dt

Exercise 6.9. Suppose Xt is as in Exercise 6.8 and Bt is an independent

standard Brownian motion. Let Yt = Xt + t + 2Bt and answer the same
questions for Yt .

Exercise 6.10. Do Exercise 6.8 where

dx
d(x) = , 0 < |x| < 1.
|x|5/2
Chapter 7

Fractional Brownian motion

7.1 Definition
The assumptions of independent, stationary increments along with continu-
ous paths give Brownian motion. In the last chapter we dropped the conti-
nuity assumption and obtained Levy processes. In this chapter we will retain
the assumptions of stationary increments and continuous paths, but will al-
low the increments to be dependent. The process Xt we construct is called
fractional Brownian motion and depends on a parameter H (0, 1) called
the Hurst index. It measures the correlation of the increments.

When H > 1/2, the increments are positively correlated, that is, if
the process has been increasing, then it is more likely to continue
increasing.

When H < 1/2, the increments are negatively correlated, that is, if
the process has been increasing, then it is more likely to decrease in
the future.

If H = 1/2, the increments are uncorrelated and the process is the

same as Brownian motion.

We will also assume that the process is self-similar.

If a > 0, then the distribution of

Yt := aH Xat

is the same as that of Xt . In particular, Xt has the same distribution

as tH X1 .

223
224 CHAPTER 7. FRACTIONAL BROWNIAN MOTION

Our final assumption is that the process is a centered Gaussian process.

For every t1 , . . . , tk , the random vector (Xt1 , . . . , Xtk ) has a joint nor-
mal distribution with mean zero.

We normalize so that Var(X1 ) = E[X12 ] = 1 and hence

Var(Xt ) = Var(tH X1 ) = t2H Var(X1 ) = t2H .

Definition If H (0, 1), the fractional Brownian motion with Hurst pa-
rameter H (f BmH ) is the centered (mean zero) Gaussian process Xt with
continuous paths such that for all s, t,

E (Xt Xs )2 = |t s|2H .
 

Since
E (Xt Xs )2 = E Xt2 + E Xs2 2E [Xs Xt ] ,
     

it follows that
1  2H
s + t2H |s t|2H .

Cov(Xs , Xt ) = E [Xs Xt ] = (7.1)
2
If H = 1/2, then fractional Brownian motion is the same as usual Brownian
motion.
As in the case of Brownian motion, we must show that such a process
exists. We will discuss this in the next section, but for now we assume it
does exist. If s < t, note that

E [Xs (Xt Xs )] = E [Xs Xt ] E Xs2

 

1  2H
t s2H (t s)2H

=
2

> 0, H > 1/2
= 0, H = 1/2 .
< 0, H < 1/2

Since Xt+ Xt N (0, 2H ), we can write roughly

|Xt+ Xt | H .

In other words, the Holder exponent of f BmH is given by the Hurst index H.
If H > 1/2, the paths are smoother than Brownian paths and if H < 1/2,
the paths are rougher.
7.2. STOCHASTIC INTEGRAL REPRESENTATION 225

To determine correlations for large t, suppose t is large. Then

E[(Xs+1 Xs )(Xs+t+1 Xs+t ]
= E[X1 (Xt+1 Xt )]
= E[X1 Xt+1 ] E[X1 Xt ]
1 
(t + 1)2H + (t 1)2H 2t2H

=
2
H (2H 1) t2H2 .
The coefficient is positive for H > 1/2 and negative for H < 1/2. As t goes
to infinity the size of the correlation goes to zero like a power law.

7.2 Stochastic integral representation

We will give an expression for the fractional Brownian motion in terms of a
stochastic integral. It will be useful to consider time going from to .
Suppose Bt1 , Bt2 are independent standard Brownian motions starting at the
origin. If  1
Bt , t 0
Bt = 2 , t0 ,
Bt
then Bt is standard Brownian motion from time to . The centering
B0 = 0 is rather arbitrary; we really think of this process as a collection of
increments {Bt Bs } for s < t. Viewed this way, this is sometimes called
white noise. We let Ft denote the information contained in the random
variables {Bs Br : r < s t}.
Suppose f (r, t) is a continuous (nonrandom) function, and let
Z t X k 
Yt = f (r, t) dBr = lim f , t B(k, n),
n n
where
B(k, n) = B(k+1)/n Bk/n ,
and the sum is over all integers (positive and negative) with (k + 1)/n t.
Then Yt is a limit of centered normal random variables and hence is mean
zero with variance Z t
E[Yt2 ] = f (r, t)2 dr.

We assume this is finite for each t. Moreover if s < t, we can write
Z s Z t
Yt Ys = [f (r, t) f (r, s)] dBr + f (r, t) dBr .
s
226 CHAPTER 7. FRACTIONAL BROWNIAN MOTION

The right-hand side is the sum of two independent normal random variables:
the first is Fs -measurable and the second is independent of Fs . Hence Yt
Ys has a normal distribution. More generally, one can check that Yt is a
Gaussian process whose covariance is given for s < t by
Z s
E [Ys Yt ] = f (r, s) f (r, t) dr.

If we make the further assumption that f (r, t) = (t r) for a one-

variable function , then the process Yt is stationary and has the form
Z t
Yt = (t r) dBr . (7.2)

Proposition 7.2.1. If
1
(s) = c sH 2 ,
and Yt is defined as in (7.2), then Xt = Yt Y0 is f BmH . Here
 Z 1/2
1 H 21 H 12 2
c = cH = + [(1 + r) r ] dr .
2H 0

Proof. By construction, Xt is a Gaussian process with continuous paths.

Since Yt is stationary, the distribution of Xt Xs = Yt Ys is the same as
that of Xts = Yts Y0 . In particular, E[(Xt Xs )2 ] = E[Xts
2 ]. Hence we
2 2H
only need to show that E[Xt ] = t . Note that
Z 0 Z t
Xt = [(t r) (r)] dBr + (t r) dBr .
0

The random variables on the right-hand side are independent and

Z 0 
2
c Var [(t r) (r)] dBr

Z 0 1 1
= [(t r)H 2 (r)H 2 ]2 dr
Z 1 1
= [(t + r)H 2 rH 2 ]2 dr
Z0
1 1
= [(t + st)H 2 (st)H 2 ]2 t ds
0
Z
1 1
2H
= t [(1 + s)H 2 sH 2 ]2 ds,
0
7.3. SIMULATION 227

t t
t2H
Z  Z
2
c Var (t r) dBr = (t r)2H1 dr = .
0 0 2H
If we choose c as in (7.3), we get E[Xt2 ] = t2H .

7.3 Simulation
Because the fractional Brownian motion has long range dependence it is not
obvious how to do simulations. The stochastic integral represetation (7.2)
is difficult to use because it uses the value of the Brownian motion for all
negative times. However, there is a way to do simulations that uses only the
fact that f BmH is a Gaussian process with continuous paths. Let us choose
a step size t = 1/N ; continuity tells us that it should suffice to sample
Y1 , Y2 , . . .
where Yj = Xj/N . For each n, the random vector (Y1 , . . . , Yn ) has a centered
Gaussian distribution with covariance matrix = [jk ]. Given we claim
that we can find numbers ajk with ajk = 0 if k > j, and independent
standard normal random variables Z1 , Z2 , . . . such that
Yn = an1 Z1 + + ann Zn . (7.3)
In matrix notation, A = [ajk ] is a lower triangular matrix such that =
AAT .
This decomposition = AAT is sometimes called the Cholesky decom-
position. We will now show that it exists by giving an algorithm for finding
the matrix. We start by setting
p
a11 = 11 .
Suppose we have found the first k 1 rows of A. This is a lower triangular
(k 1) (k 1) matrix. Suppose j < k. Then,
j
X
jk = E [Yj Yk ] = aji aki .
i=1

The coefficients ajk for j = 1, . . . , k 1 can be found by solving these k 1

linear equation in k 1 unknowns (one uses the fact that is nonnegatve
definite to show there is a unique solution), and then one sets
v
u
u k1
X
akk = kk
t a2ki .
i=1
228 CHAPTER 7. FRACTIONAL BROWNIAN MOTION

We see that this decomposition is unique if we impose the condition akk 0.

From a practical perspective, these computations are done with computer
packages. Note that one only needs to compute the coefficients ajk once and
store them. Then multiple simulations can be done using (7.3).

The Cholesky decomposition can be derived from the Gram-

Schmidt process. Consider the Hilbert space of mean-zero L2 ran-
dom variables and let Hn denote the finite dimensonal subspace
spanned by Y1 , . . . , Yn . Let Pn denote the projection onto Hn which
is the same as the conditional expectation,

Pn X = E[X | Y1 , . . . , Yn ].

Then we define
Yn Pn1 Yn
Zn = .
kYn Pn1 Yn k2
Since Hn is also the subspace spanned by Z1 , . . . , Zn , we can write
n1
X
Pn1 Yn = ajn Zj ,
j=1

for unique scalars ajn , and we set ann = kYn Pn1 Yn k2 .

Chapter 8

Harmonic functions

8.1 Dirichlet problem

Recall that the Laplacian of a function f on Rd is defined by
d
X
2 f (x) = jj f (x).
j=1

The Laplacian is closely related to mean values of functions. The (spherical)

mean value of f on the sphere of radius  about x is defined by
R
|yx|= f (y) d(y)
M V (x; , f ) = R ,
|yx|= 1 d(y)

where denotes (d 1)-dimensional surface measure. The mean value

can also be described in terms of Brownian motion. Suppose Bt is a d-
dimensional Brownian motion starting at x and let

 = inf{t : |Bt B0 | = }.

Since Brownian motion is rotationally invariant, the distribution of B is

uniform over the sphere {y : |y x| = }. Hence we can write

M V (x; , f ) = Ex [f (B )] . (8.1)

Proposition 8.1.1. Suppose f is a C 2 function in a neighborhood of x Rd .

Then
M V (x; , f ) f (x)
2 f (x) = 2d lim .
0 2

229
230 CHAPTER 8. HARMONIC FUNCTIONS

Proof. Without loss of generality, assume x = 0 and f (x) = 0, and write

M V () for M V (0; f, ). Let bj = j f (0), ajk = jk f (0). Taylors theorem
implies that

d d
X 1X
f (y) = bj yj + ajk yj yk + o(|y|2 ), y 0,
2
j=1 j=1

where we write y = (y1 , . . . , yd ). Therefore,

d d
X 1X
M V () = bj M V (0; yj , ) + ajk M V (0; yj yk , ) + o(2 ).
2
j=1 j=1

Note that M V (0; yj , ) = 0 since yj is an odd function. Similarly,

M V (0; yj yk , ) = 0 if j 6= k. Linearity of the integral implies that

d
X d
X
M V (0; yj2 , ) = M V (0; yj2 , ) = M V (0; 2 , ) = 2 .
j=1 j=1

Symmetriy implies that M V (0; yj2 , ) = M V (0; yk2 , ), and hence
M V (yj2 , ) = /d. We therefore have

d
1X 2 2
M V () = ajj (2 /d) + o(2 ) = f (0) + o(2 ).
2 2d
j=1

Definition Suppose f is a function defined on an open set D Rd .

f is harmonic in D if f is C 2 and 2 f (x) = 0 for all x D. f is

harmonic at x if it is harmonic in a neighborhood of x.

A function f satisfies the (spherical) mean value property on D if for

every x D and every  > 0 with {y : |x y| } D ,

f (x) = M V (f ; x, ).

Proposition 8.1.2.
8.1. DIRICHLET PROBLEM 231

Suppose f is a harmonic function defined on an open set D Rd . Let

Bt be a standard d-dimensional Brownian motion starting at x D
and let
= inf{t : Bt 6 D}.
Then Mt := f (Bt ) is a local martingale for t < satisfying

dMt = f (Bt ) dBt , t < .

Suppose also that f is defined on D so that f : D R is a bounded,

continuous function. Then Mt is a bounded continuous martingale. If
P{ < } = 1, then
f (x) = Ex [f (B )] . (8.2)

Proof. The first statement follows from Itos formula. In fact, we already
did this calculation in Theorem 3.7.2. Since bounded local martingales are
martingales, the second statement follows from the optional sampling theo-
rem.

The expression (8.2) is a generalization of the mean value property (8.1).

It states that we can take mean values over sets other than spheres as long
as integrate with respect to the correct the measure. This measure is often
called harmonic measure (in D starting at x) and is defined by

hmD (V ; x) = Px {B V }, V D.

Then (8.2) can be written

Z
f (x) = f (y) hmD (dy; x).
D

The next proposition shows that we could use the mean value propery
to define harmonic functions. In fact, this is the more natural definition.
Proposition 8.1.3. Suppose f is a continuous function on an open set
D Rd . Then f is harmonic in D if and only f satisfies the mean value
property on D.
Proof. If f is harmonic, then f restricted to a closed ball of radius  con-
tained in D is bounded. Therefore, the mean value property is a particular
case of (8.2).
If f is C 2 and satisfies the mean value property, then 2 f (x) = 0 by
Proposition 8.1.1. Hence we need only show that f is C 2 . We will, in fact,
show that f is C .
232 CHAPTER 8. HARMONIC FUNCTIONS

It is a standard exercise in advanced calculus to show that for every

> 0 we can find a nonnegative radially symmetric C function =
such that (y) = 0 for |y| and
Z
(y) dy = 1.

Here dy = dy1 dyd . If x D with {y : |x y| } D, and = /2 ,

then the mean value property implies that
Z Z
f (x) = f (x + y) (y) dy = f (z) (z x) dz. (8.3)

Since f is locally bounded and is bounded and C , we can differentiate

with respect to x by bringing the derivatives inside the integral.

The proof showed that we did not need to assume that f is continuous.
It suffices for f to be measurable and locally bounded so that derivatives
can be taken on the right-hand side of (8.3).
We will solve the classic Dirichlet problem for harmonic functions. Sup-
pose D is a bounded open set and F : D R is a continuous function.
The goal is to find a continuous function f : D R that is harmonic in D
with f (x) = F (x) for x D. Suppose that such a function f existed. Let
= D = inf{t 0 : Bt D}. Since D is compact, f must be bounded,
and hence
Mt = f (Bt )
is a continuous, bounded martingale. Arguing as in (8.2) we see that

f (x) = Ex [f (B )] = Ex [F (B )] . (8.4)

The right-hand side gives the only candidate for the solution. The strong
Markov property can be used to see that this candidate satisfies the mean
value property and the last proposition gives that f is harmonic in D.
It is a little more subtle to check if f is continuous on D. This requires
futher assumptions on D which can be described most easily in terms of
Brownian motion. Suppose z D and x D with x near z. Can we say
that f (x) is near F (z)? Since F is continuous, this will be true if B is near
z. To make this precise, one defines a point z D to be regular if for every
 > 0 there exists > 0 such that if x D with |x z| < , then

Px {|B z| } .
8.1. DIRICHLET PROBLEM 233

Theorem 8.1.4. Suppose D is a bounded open set such that each z D

is regular. Suppose F is a continuous function on D. Then there exists a
unique continuous function f : D R that is harmonic on D and agrees
with F on D given by (8.4).

Example 8.1.1. Let D be the annular region

D = {x Rd : r < |x| < R},

and let F (x) = 1 if |x| = R, F (x) = 0 if |x| = r. Then for x D,

f (x; r, R) = Ex [F (B )] = Px {|B | = R}.

We claim that
xr
f (x; r, R) = , d = 1,
Rr
log |x| log r
f (x; r, R) = , d = 2,
log R log r
rd2 |x|d2
f (x; r, R) = , d 3.
rd2 Rd2
One can check this by noting that 2 f (x) = 0 and f has the correct bound-
ary condition. The theorem implies that there is only one such function.
Note that 
0, d2
lim f (x; r, R) = ,
R 1 (r/|x|)2d , d 3

x/R d = 1
lim f (x; r, R) = . (8.5)
r0 1, d2
We have already seen the d = 1 case as the gamblers ruin estimate for
Brownian motion.

Example 8.1.2. Let d 2. It follows from (8.5) that if x 6= 0, then the

probability that a Brownian motion starting at x ever hits zero is zero. Let
D = {x Rd : 0 < |x| < 1}. Then 0 is not a regular point of D, since if
we start near the origin the Brownian motion will not exit D near 0. Let
F (0) = 0, F (y) = 1 if |y| = 1. Then, the only candidate for the solution of
the Dirichlet problem is

f (x) = Ex [F (B )] = 1, x D.

If f (0) = 0, this function is not continuous at the origin.

234 CHAPTER 8. HARMONIC FUNCTIONS

If D = Ur = {x : |x| < r} is the ball of radius r about the origin, then the
harmonic measure hmUr (; x) is known explicitly. It is absolutely continuous
with respect to (d 1)-dimensional surface measure s on Ur . Its density is
called the Poisson kernel,

r2 |x|2
Hr (x, y) = ,
r d1 |x y|d

where Z
d1 = ds(y)
|y|=1

denotes the surface area of the unit ball. If F : Ur R is continuous, then

the unique solution to the Dirichlet problem is
Z
x
 
f (x) = E F (BUr ) = F (y) Hr (x, y) ds(y). (8.6)
|y|=r

To verify this, one checks by direct computation the following facts:

If y Ur , the function h(x) = H(x, y) is harmonic on Ur ;

If x Ur , Z
Hr (x, y) ds(y) = 1;
|y|=r

For every  > 0, there exists > 0 such that if x Ur , y Ur and

|x y| < , then
Z
Hr (x, z) ds(z) .
|z|=r,|zy|

From these one concludes that f as defined by the right-hand side of (8.6)
is harmonic in Ur and continuous on U r .
The reader may note that we did not need the probabilistic interpreta-
tion of the solution in order to verify that (8.6) solves the Dirichlet problem.
Indeed, the solution using the Poisson kernel was discovered before the re-
lationship with Brownian motion. An important corollary of this explicit
solution is the following theorem; we leave the verification as Exercise 8.1.
The key part of the theorem is the fact that the same constant C works for
all harmonic functions.

Theorem 8.1.5.
8.2. H-PROCESSES 235

1. For every positive integer n, there exists C = C(d, n) < such that if
f is a harmonic function on an open set D Rd , x D, {y : |x y| <
r} D, and j1 , . . . , jn {1, . . . , d} then
 
xj1 xjn f (x) C rn sup |f (y)|.
 
|yx|<r

2. (Harnack inequality) For every 0 < u < 1, there exists C = C(d, u) <
such that if f is a positive harmonic function on an open set D
Rd , x D, {y : |x y| < r} D, then if |x z| ur,

C 1 f (x) f (z) C f (x).

8.2 h-processes
Suppose h is a positive harmonic function on an open set D Rd , and let
Bt be a standard Brownian motion starting at x D. Let = D be the
first time t with Bt 6 D. Then Mt = h(Bt ) is a positive local martingale for
t < satisfying
h(Bt )
dMt = Mt dBt , t < .
h(Bt )
Let n be the minimum of and the first time t with h(Bt ) n. Then Mtn
is a continuous bounded martingale.
We can use the Girsanov theorem to consider the measure obtained by
weighting by the local martingale Mt . To be more precise, if V is an event
that depends only on Bt , 0 t n , then

P (V ) = h(x)1 E [Mn 1V ] .

One can use the Girsanov theorem (more precisely, a simple generalization
of the theorem to d-dimensional Brownian motion), to see that
h(Bt )
dBt = dt + dWt ,
h(Bt )
where Wt is a standard Brownian motion with respect to P . The process
Bt in the new measure P is often called the (Doob) h-process associated to
the positive harmonic function h. It is defined for t < .
As an example, suppose that D is the unit ball, y = (1, 0, . . . , 0) D
and
1 |x|2
h(x) = d1 H1 (x, y) = ,
|x y|d
236 CHAPTER 8. HARMONIC FUNCTIONS

is the Poisson kernel. Then the h-process can be viewed as Brownian motion
conditioned so that B = y, where = D . This is not precise because the
conditioning is with respect to a event of probability zero. We claim that
the P -probability that B = y equals one. To see this, assume that B0 = 0
and let

Tn = inf{t : h(Bt ) = n3 }, Tn0 = inf{t > Tn : h(Bt ) = n}.

We claim that P {Tn < } = 1. Indeed, if we let r = inf{t : |Bt | = r}, then
we can check directly that
Z
3
lim P {h(Br ) n } = lim h(rx) d(x) = 1,
r1 r1 |x|=1,h(rx)n3

and hence
lim P {Tn < r } = 1.
r1

Also, since h(BTn0 ) = n2 h(BTn ),

P {Tn0 < } n2 ,

X
P {Tn0 < } < .
n=1

Therefore, by the Borel-Canelli lemma, we see that with P -probability one

for all n sufficiently large h(Bt ) n for t Tn .

8.3 Time changes

To prepare for the next section we will consider time changes of Brownian
motion. If Bt is a standard Brownian motion and Yt = Bat , then Yt is a
Brownian motion with variance parameter a. We can write
Z t
Yt = a dWs ,
0

where Wt = a1/2 Bat is a standard Brownian motion. We generalize this by

considering Yt = Ba(t) where a is a strictly increasing continuous function.
We allow a(t) to be random, but it must be adapted to the Brownian motion.
In other words, for each t, the random time a(t) must be a stopping time
8.3. TIME CHANGES 237

for the Brownian motion Bt . We will make the further assumption that a is
a C 1 function, that is, we can write
Z t
a(t) = a(s) ds,
0

where s 7 a(s) is a continuous nonnegative function. The assumption that

a(t) is strictly increasing implies that there are no open intervals on which
a(s) is identically zero. If Y (t) = Ba(t) , then Yt is a continuous local mar-
tingale and we can write
Z tp
Yt = a(s) dWs ,
0

where Wt is the standard Brownian motion

Z t
Wt = [a(s)]1/2 dBa(s) . (8.7)
0

One can verify this is a standard Brownian motion by checking that is a

martingale with hW it = t. More generally, if Xt satisfies the SDE

dXt = Rt dt + At dBt ,

and Yt = Xa(t) , then Yt satisfies the SDE

p
dYt = a(t) Ra(t) dt + a(t) Aa(t) dWt ,

where Wt is the standard Brownian motion (8.7).

Example 8.3.1. Suppose that Xt satisfies the Bessel equation
r
dXt = dt + dBt , X0 = 1,
Xt
where r R. This is defined for t < T = inf{s : Xs = 0}. For t < T , let
Lt = log Xt . Then Itos formula shows that
1 1
dLt = dXt dhXit
Xt 2 Xt2
 
1
= r e2Lt dt + eLt dBt .
2
Consider the time change with a(t) = e2Lt , and let Yt = La(t) . Then Yt
satisfies  
1
dYt = r dt + dWt ,
2
238 CHAPTER 8. HARMONIC FUNCTIONS

where Wt is a standard Brownian motion. In other words, Yt is a Brownian

motion with variance parameter 1 and drift m = r 12 . From this we can
see that Yt if r > 1/2, Yt if r < 1/2, and

= lim inf Yt < lim sup Yt = ,

t t

if r = 1/2. One can use this to give another proof of Proposition 4.2.1. Note
that if r < 1/2, then in the new time, it take an infinite amount of time for
L = log X to reach . However, in the original time, this happens at time
T < . In other words,
Z
a() = a(s) ds = T < .
0

8.4 Complex Brownian motion

If Bt1 , Bt2 are independent standard Brownian motions then

Bt = Bt1 + i Bt2

is called a (standard) complex Brownian motion. Note that this is the same
as a two-dimensional Brownian motion except that the point in the plane is
viewed as a complex number.
Suppose f : C C is a nonconstant holomorphic (complex differ-
entiable) function. We will consider Xt = f (Bt ). Near any point z with
f 0 (z) 6= 0, the function looks like f (w) = f (z) + f 0 (z) (w z). Multiplication
by f 0 (z) is the same as a dilation by |f 0 (z)| and a rotation by arg f 0 (z). A
rotation of a standard two-dimensional Brownian motion gives a standard
two-dimensional Brownian motion and a dilation gives a Brownian motion
with a different variance. This leads us to believe that Xt is a time-change
of standard Brownian motion.
Let us make this more precise. Let
Z t
a(t) = |f 0 (Bs )|2 ds.
0

We consider the time-change in which it takes time a(t) to
traverse

f (B[0, t]). Equivalently, in time t one has traversed f B[0, a1 (t)] .

Theorem 8.4.1. Let Yt = Xa1 (t) . Then Yt is a standard complex Brownian

motion.
8.4. COMPLEX BROWNIAN MOTION 239

Proof. Write f = u + iv, Xt = Ut + iVt , Yt = Ut + iVt . The Cauchy-Riemann

equations imply that x u = y v, y u = x v and
2 u = x x u + y y u = x y v + y (x v) = 0.
Similarly 2 v = 0. Since Ut = u(Bt1 , Bt2 ) , Vt = v(Bt1 , Bt2 ), then Itos formula
(see Theorem 3.7.2) gives
dUt = u(Bt ) dBt = x u(Bt ) dBt1 + y u(Bt ) dBt2 .

dVt = v(Bt ) dBt = x v(Bt ) dBt1 + y v(Bt ) dBt2

= y u(Bt ) dBt1 + x u(Bt ) dBt2 .
Note that
[x u(Bt )]2 + [y u(Bt )]2 dt


dhU it =
[x u(Bt )]2 + [x v(Bt )]2 dt = |f 0 (Bt )|2 dt,


=
dhV it = |f 0 (Bt )|2 dt,
dhU, V it = 0.
Therefore, Ut , Vt are martingales with quadratic variation t and hence are
standard Brownian motions. Since hU, V it = 0, we also have hU , V it = 0,
and hence Ut , Vt are independent.
The last theorem assumes that the function f is entire, that is, it is
defined and holomorphic on all of C. Suppose D is domain, that is, an open,
connected subset of C, and suppose that f is a nonconstant holomorphic
function on D. Let Bt be a complex Brownian motion, and let
D = inf{t 0 : Bt 6 D}.
Then the process Xt as above is defined for t < D , and the process Yt is
defined for t < a(D ) with the processes taking values in f (D).
Theorem 8.4.2. If f is a holomorphic function on a domain D, and Xt , Yt
are defined as above, then Yt , 0 t < a(D ) is a standard complex Brown-
ian motion.
One very important case is when f maps D one-to-one onto f (D). Such
functions are often called conformal transformations. In this case f 1 is a
conformal transformation from f (D) to D, and
a(D ) = inf{t : Yt 6 f (D)}.
In particular, this shows that harmonic measure is a conformal invariant,
hmD (V ; z) = hmf (D) (f (V ); f (z)).
240 CHAPTER 8. HARMONIC FUNCTIONS

8.5 Exercises
Exercise 8.1. Use (8.6) to prove Theorem 8.1.5.

Exercise 8.2. Suppose D C is a domain and is a harmonic function

on D (here we are also viewing D as a subset of R2 ). Let f : D f (D) be
a conformal transformation and let

(w) = (f 1 (w)).

Show that is a harmonic function on f (D) in two different ways:

Do a direct computation to show that 2 = 0.

Use the conformal invariance of harmonic measure and (8.2).