A Short Introduction to

Diffusion Processes and Ito Calculus

Cedric Archambeau
University College, London
Center for Computational Statistics and Machine Learning
[email protected]
January 24, 2007
Notes for the Reading Group on Stochastic Differential Equations (SDEs).
The text is largely based on the book Numerical Solution of Stochastic Differential Equations by E. Kloeden & E. Platen (1992) [3]. Other good references
are [4] and [2], as well as [1]

Elements of Measure and Probability Theory

In this section, we review some important concepts and definitions, which will
be extensively used in the next sections.
Definition 1. A collection A of subsets of is a -algebra if
A,
Ac A if A A,
[
An if A1 , A2 , . . . , An , . . . A.

(1)
(2)
(3)

This means that A is a collection of subsets of containing and which

is closed under the set of operations of complementation and countable unions.
Note that this implies that A is also closed under countable intersections.
Definition 2. Let (, A) be a measurable space, i.e. (, A) is an ordered pair
consisting of a non-empty set and a -algebra A of subsets of . A measure
on (, A) is a nonnegative valued set function on A satisfying
() = 0,

(4)
!

An

(An ),

for any sequence A1 , A2 , . . . , An , . . . A and Ai Aj = for i 6= j.

1

(5)

From (5) it follows that (A) (B) for all A B in A. The measure
is finite if 0 () . Hence, it can be normalized to obtain a probability
measure P with P (A) = (A)/() [0, 1] for all A A.
An important measure is the Borel measure B on the -algebra B of Borel
subsets1 of R, which assigns to each finite interval its length. However, the
measure space (R, B, B ) is not complete in the sense that there exist subsets B
of R with B
/ B, but B B for some B B with B (B) = 0. Therefore, we
enlarge the -algebra B to a -algebra L and extend the measure B uniquely
to the measure L on L so that (R, L, L ) is complete, that is L L with
L (L ) = 0 whenever L L for some L L with L = 0. We call L the
Lebesgue subsets of R and L the Lebesgue measure.
Definition 3. Let (1 , A1 ) and (2 , A2 ) be two measurable spaces. The function f : 1 2 is A1 : A2 -measurable if
f 1 (A2 ) = {1 1 : f (1 ) A2 } A1 ,

(6)

for all A2 A2 .
This means that the pre-image of any A2 A2 is in A1 .
Definition 4. Let be the sample space, the -algebra A a collection of
events and P the associated probability measure. We call a triplet (, A, P ) a
probability space if A and P satisfy the following properties:
Ac = \ A, A B, A B A if A, B A
0 P (A) 1, P (Ac ) = 1 P (A), P () = 0, P () = 1,
S
T
n An ,
n An A if {An , An A},
S
P
P ( n An ) = n P (An ) if {An , An A} and Ai Aj = for all i 6= j.

(7)
(8)
(9)
(10)

The last two properties hold for a countably (finite or infinite) number of
outcomes. For uncountably many basic outcomes we restrict attention to countable combinations of natural events, which are subintervals of and to which
non-zero probabilities are (possibly) assigned.
When (1 , A1 , P ) is a probability space and (2 , A2 ) is either (R, B) or
(R, L), we call the measurable function f : 1 R a random variable and
denote it usually by X.

Stochastic Processes

In this section, we review the general properties of standard stochastic processes

and discuss Markov chains and diffusion processes.
Definition 5. Let T denote the time set under consideration and let (, A, P )
be a common underlying probability space. A stochastic process X = {Xt , t
T } is a function of two variables X : T R, where
1 Any subset of R generated from countable unions, intersections or complements of the
semi-infinite intervals {x R : < x a} is a Borel subset of R.

 Xt = X(t, ) : R is a random variable for each t T ,

X(, ) : T R is a realization or sample path for each .
Depending on T being a discrete or a continuous time set, we call the stochastic process a discrete or a continuous time process.
Example. A Gaussian process is a stochastic process for which the probability
law is Gaussian, that is any joint distribution Fti1 ,ti2 ,...,tin (xi1 , xi2 , . . . , xin ) is
Gaussian for all tij T .
The time variability of a stochastic process is described by all its conditional
probabilities (see Section 2.1 and 2.2). However, substantial information can
already be gained from by the following quantities:
The means: t = E{Xt } for each t T .
The variances: t2 = E{(Xt t )2 } for each t T .
The (two-time) covariances: Cs,t = E{(Xs s )(Xt t )} for distinct
time instants s, t T .
Definition 6. A stochastic process X = {Xt , t T } for which the random
variables Xtj+1 Xtj with j = 1, . . . , n 1 are independent for any finite combination of time instants t1 < . . . < tn in T is a stochastic process with independent
increments.
Example. A Poisson process is a continuous time stochastic process X =
{Xt , t R+ } with (non-overlapping) independent increments for which
X0 = 0 w.p. 1,
E{Xt } = 0,
Xt Xs P((t s)),
for all 0 s t and where is the intensity parameter.
As a consequence, the means, the variances and the covariances of a Poisson
process are respectively given by t = t, t2 = t and Cs,t = min{s, t}.
Property 6.1. A stochastic process is strictly stationary if all its joint distributions are invariant under time displacement, that is Fti1 +h ,ti2 +h ,...,tin +h =
Fti1 ,ti2 ,...,tin for all tij , tij+1 T with h 0.
This constraint can be relaxed to stationarity with respect to the first and
the second moments only.
Property 6.2. A stochastic process X = {Xt , t R+ } is wide-sense stationary
if there exists a constant R and a function c : R+ R, such that
t = , t2 = c(0) and Cs,t = c(t s),
for all s, t T .
3

(11)

Example. The Ornstein-Uhlenbeck process with parameter > 0 is a widesense stationary Gaussian process X = {Xt , t R+ } for which
X0 N (0, 1),
E{Xt } = 0,
Cs,t = e|ts| ,
for all s, t R+ .
Definition 7. Let (, A, P ) be the probability space and {At , t 0} an increasing family of sub--algebras2 of A. The stochastic process X = {Xt , t R+ },
with Xt being At -measurable for each t 0, is a martingale if
E{Xt |As } = Xs , w.p. 1,

(12)

for all 0 s < t.

Equivalently, we can write E{Xt Xs |As } = 0. Thus, a stochastic process
is a martingale if the expectation of some future event given the past and the
present is always the same as if given only the present.
When the process Xt satisfies the Markov property (see Sections 2.1 and
2.2), we have E{Xt |As } = E{Xt |Xs }.

2.1

Markov chains

We first describe discrete time Markov chains and then generalize to their continuous time counterpart.
Definition 8. Let X = {x1 , . . . , xN } be the set of a finite number of discrete
states. The discrete time stochastic process X = {Xt , t T } is a discrete time
Markov chain if it satisfies the Markov property, that is
P (Xn+1 = xj |Xn = xin ) = P (Xn+1 = xj |X1 = xi1 , . . . , Xn = xin )

(13)

for all possible xj , xi1 , . . . , xin X with n = 1, 2, . . ..

This means that only the present value of Xn is needed to determine the
future value of Xn+1 .
The entries of the transition matrix Pn RN N of the Markov chain are
given by
p(i,j)
= P (Xn+1 = xj |Xn = xi )
n

(14)

for i, j = 1, . . . , N . We call them the transition probabilities. They satisfy

PN (i,j)
= 1 for each i, as Xn+1 can only attain states in X .
j pn
Let pn be the column vector of the marginal probabilities P (Xn = xi ) for
i = 1, . . . , N . The probability vector pn+1 is then given by pn+1 = PT
n pn .
2 The sequence {A , t 0} is called an increasing family of sub--algebras of A if A A
t
s
t
for any 0 s t. This means that more information becomes available with increasing time.

 for all
Property 8.1. A discrete time Markov chain is homogeneous if Pn = P
n = 1, 2, . . ..
As a consequence, the probability vector of a homogenous Markov chain


k T pn for any k = N \ {0}. The probability distributions
satisfies pn+k = P
depend only on the time that has elapsed. However, this does not mean that
the Markov chain is strictly stationary. In order to be so, it is also required that
for each n = 1, 2, . . ., which implies that the probability distributions
pn = p
Tp
=P
.
are equal for all times such that p
It can be shown that a homogenous Markov chain has at least one stationary
probability vector solution. Therefore, it is sufficient that the initial random
variable X1 is distributed according to one of its stationary probability vectors
for the Markov chain to be strictly stationary.
Property 8.2. Let X = {x1 , . . . , xN } be the set of discrete states and f : X
R. The discrete time homogeneous Markov chain X = {Xn , n = 1, 2, . . .} is
ergodic if
T
N
X
1 X
f (Xn ) =
f (xi )
p(i)
T T
n=1
i=1

lim

(15)

is a stationary probability vector and all xi X .

where p
This property is an extension of the Law of Large Numbers to stochastic
processes. A sufficient condition for a Markov chain to be ergodic is that all the
k are nonzero. The same result holds if all the
components of some k th power P
are nonzero.
entries of the unique p
Definition 9. Let X = {x1 , . . . , xN } be the set of a finite number of discrete
states. The stochastic process X = {Xt , t R+ } is a continuous time Markov
chain if it satisfies the following Markov property:
P (Xt = xj |Xs = xi ) = P (Xt = xj |Xr1 = xi1 , . . . , Xrn = xin , Xs = xi )

(16)

for 0 r1 . . . rn < s < t and all xi1 , . . . , xin , xi , xj X .

The entries of the transition matrix Ps,t RN N and the probability vectors
(i,j)
are now respectively given by ps,t = P (Xt = xj |Xs = xi ) and pt = PT
s,t ps
for any 0 s t. Furthermore, the transition matrices satisfy the relationship
Pr,t = Pr,s Ps,t for any 0 r s t.
Property 9.1. If all its transition matrices depend only on the time differences,
then the continuous time Markov chain is homogeneous and we write Ps,t =
P0,ts Pts for any 0 s t.
Hence, we have Pt+s = Pt Ps = Ps Pt for all s, t 0.

Example. The Poisson process is a continuous time homogenous Markov chain

on the countably infinite state space N. Its transition matrix is given by
Pt =

(t)m t
e ,
m!

(17)

where m N.
Indeed, invoking the independent increments of the Poisson process we get
P (Xs = ms , Xt = mt ) = P (Xs = ms , Xt Xs = mt ms )
= P (Xs = ms )P (Xt Xs = mt ms )
=

ms sms s mt ms (t s)mt ms (ts)

e
e
.
ms !
(mt ms )!

The second factor on the right hand side corresponds to the transition probability P (Xt = mt |Xs = ms ) for ms mt . Hence, the Poisson process is
homogeneous since
P (Xt+h = mt + m|Xt = mt ) =

(h)m h
e
,
m!

where h 0.
Property 9.2. Let f : X R. The continuous time homogeneous Markov
chain X = {Xt , t R+ } is ergodic if
1
lim
T T

f (Xt ) dt =
0

N
X

f (xi )
p(i)

(18)

i=1

is a stationary probability vector and all xi X .

where p
Definition 10. The intensity matrix A RN N of a homogeneous continuous
time Markov chain is defined as follows:

(i,j)
pt
lim
if i 6= j,
t0
(i,j)
t
a
=
(19)
(i,j)
pt
1
lim
if i = j.
t0

Theorem 1. The homogeneous continuous time Markov chain X = {Xt , t

R+ } is completely characterized by the initial probability vector p0 and its
intensity matrix A RN N . Furthermore, if all the diagonal elements of A are
finite, then the transition probabilities satisfy the Kolmogorov forward equation
dPt
Pt A = 0,
dt

(20)

and the Kolmogorov backward equation

dPt
AT Pt = 0.
dt
6

(21)

2.2

Diffusion processes

Diffusion processes are an important class of continuous time continuous state

Markov processes, that is X R.
Definition 11. The stochastic process X = {Xt , t R+ } is a (continuous time
continuous state) Markov process if it satisfies the following Markov property:
P (Xt B|Xs = x) = P (Xt B|Xr1 = x1 , . . . , Xrn = xn , Xs = x)

(22)

for all Borel subsets B R, time instants 0 r1 . . . rn s t and all

x1 , . . . , xn , x R for which the conditional probabilities are defined.
For fixed s, x and t the transition probability P (Xt B|Xs = x) is a probability measure on the -algebra B of Borel subsets of R such that
Z
P (Xt B|Xs = x) =
p(s, x; t, y) dy
(23)
B

for all B B. The quantity p(s, x; t, ) is the transition density. It plays a

similar role as the transition matrix in Markov chains.
From the Markov property, it follows that
Z
p(s, x; t, y) =
p(s, x; , z)p(, z; t, y) dz
(24)

for all s t and x, y R. This equation is known as the ChapmanKolmogorov equation.

Property 11.1. If all its transition densities depend only on the time differences, then the Markov process is homogeneous and we write p(s, x; t, y) =
p(0, x; t s, y) p(x; t s, y) for any 0 s t.
Example. The transition probability of the Ornstein-Uhlenbeck process with
parameter > 0 is given by
(

2 )
y xe(ts)
1


p(s, x; t, y) = p
exp
(25)
2 1 e2(ts)
2(1 e2(ts) )
for all 0 s t. Hence, it is a homogeneous Markov process.
Property 11.2. Let f : R R be a bounded measurable function. The
Markov process X = {Xt , t R+ } is ergodic if
Z
Z
1 T
lim
f (Xt ) dt =
f (y)
p(y) dy
(26)
T T 0

R
p(x)dx is a stationary probability density.
where p(y) = p(s, x; t, y)
This means that the time average limit coincide with the spatial average. In
practice, this is often difficult to verify.
7

Definition 12. A Markov process X = {Xt , t R+ } is a diffusion process if

the following limits exist for all  > 0, s 0 and x R:
Z
1
p(s, x; t, y) dy = 0,
(27)
lim
ts t s |yx|>
Z
1
lim
(y x)p(s, x; t, y) dy = (s, x),
(28)
ts t s |yx|<
Z
1
lim
(y x)2 p(s, x; t, y) dy = 2 (s, x),
(29)
ts t s |yx|<
where (s, x) is the drift and (s, x) is the diffusion coefficient at time s and
position x.
In other words, condition (27) prevents the diffusion process from having instantaneous jumps. From (28) and (29) one can see that (s, x) and 2 (s, x) are
respectively the instantaneous rate of change of the mean and the instantaneous
rate of change of the squared fluctuations of the process, given that Xs = x.
Example. The Ornstein-Uhlenbeck process is a diffusion process with drift
(s, x) = x and diffusion coefficient (s, x) = 2.


To proof this we first note that p(s, x; t, y) = N y|xe(ts) , (1 e2(ts) )
for any t s. Therefore, we have
E{y} x
1 e(ts)
= x lim
= x ,
ts
ts
ts
ts
E{y 2 } 2E{y}x + x2
2 (x, s) = lim
ts
ts


2(ts)
(1 e(ts) )2
1e
+ x2
= lim
ts
ts
ts
(x, s) = lim

= 2 + x2 0.
Diffusion processes are almost surely continuous functions of time, but they
need not to be differentiable. Without going into the mathematical details,
the continuity of a stochastic process can be defined in terms of continuity
with probability one, mean square continuity and continuity in probability or
distribution (see for example [3]).
Another interesting criterion is Kolmogorovs continuity criterion, which
states that a continuous time stochastic process X = {Xt , t T } has continuous sample paths if there exists a, b, c, h > 0 such that
E{|Xt Xs |a } c|t s|1+b

(30)

for all s, t T and |t s| h.

Theorem 2. Let the stochastic process X = {Xt , t R+ } be a diffusion
process for which and are moderately smooth. The forward evolution of its
8

transition density p(s, x; t, y) is given by the Kolmogorov forward equation (also

known as the Fokker-Planck equation)
p

1 2
{ 2 (t, y)p} = 0,
+
{(t, y)p}
t
y
2 y 2

(31)

for a fixed initial state (s, x). The backward evolution of the transition density
p(s, x; t, y) is given by the Kolmogorov backward equation
p
p 1 2
2p
+ (s, x)
+ (s, x) 2 = 0,
s
x 2
x

(32)

for a fixed final state (t, y).

A rough proof of (32) is the following. Consider the approximate time discrete continuous state
process with two equally probable jumps from (s, x) to
(s + s, x + s s), which is consistent with (28) and (29). The approximate transition probability is then given by

1
p(s, x; t, y) = p(s + s , x + s + s ; t , y)
2

1
+ p(s + s , x + s s ; t , y).
2
Taking Taylor expansions up to the first order in s about (s, x; t, y) leads to
0=



p
p
1 2 p
s + s + 2 2 s + O (s)3/2 .
s
x
2 x

Since the discrete time process converges in distribution to the diffusion process,
we obtain the backward Kolmogorov equation when s 0.
Example. The Kolmogorov foward and the backward equations for the OrnsteinUhlenbeck process with parameter > 0 are respectively given by
p

2p
{yp} 2 = 0,
t
y
y
2
p
p
p
x
+ 2 = 0.
s
x
x

2.3

(33)
(34)

Wiener processes

The Wiener process was proposed by Wiener as mathematical description of

Brownian motion. This physical process characterizes the erratic motion (i.e.
diffusion) of a grain pollen on a water surface due to the fact that is continually
bombarded by water molecules. The resulting motion can be viewed as a scaled
random walk on any finite time interval and is almost surely continuous, w.p. 1.

Definition 13. A standard Wiener process is a continuous time Gaussian

Markov process W = {Wt , t R+ } with (non-overlapping) independent increments for which
W0 = 0 w.p. 1,
E{Wt } = 0,
Wt Ws N (0, t s),
for all 0 s t.
Its covariance is given by Cs,t = min{s, t}. Indeed, if 0 s < t, then
Cs,t = E{(Wt t )(Ws s )}
= E{Wt Ws }
= E{(Wt Ws + Ws )Ws }
= E{Wt Ws }E{Ws } + E{Ws2 }
= 0 0 + s.
Hence, it is not a wide-sense stationary process. However, it is a homogeneous
Markov process since its transition probability is given by


1
(y x)2
p(s, x; t, y) = p
exp
(35)
2(t s)
2(t s)
Although the sample paths of Wiener processes are almost surely continuous
functions of time (the Kolmogorov continuity criterion (30) is satisfied for a = 4,
b = 1 and c = 3), they are almost surely nowhere differentiable. Consider the
(n) (n)
partition of a bounded time interval [s, t] into subintervals [k , k+1 ] of equal
(n)

length (t s)/2n , where k

be shown [3, p. 72] that
lim

n
2X
1 

= s + k(t s)/2n for k = 0, 1, . . . , 2n 1. It can

2
W (n) () W (n) () = t s, w.p. 1,
k+1

k=0

where W () is a sample path of the standard Wiener process W = {W ,

[s, t]} for any . Hence,


t s lim sup
maxn W (n) () W (n) ()
n

n
2X
1

k=0

0k2 1

(n)

k+1

k+1


() W (n) ().
k

(Note that lim sup is the limit superior or supremum limit, that is the supremum3 of all the limit points.) From the sample path continuity, we have


maxn W (n) () W (n) () 0, w.p. 1 as n
0k2 1

k+1

3 For S T , the supremum of S is the least element of T , which is greater or equal to all
elements of S.

10

and therefore
n
2X
1

k=0

(n)
k+1


() W (n) () , w.p. 1 as n .
k

As a consequence, the sample paths do, almost surely, not have bounded variation on [s, t] and cannot be differentiated.
The standard Wiener process is also a diffusion process with drift (s, x) = 0
and diffusion coefficient (s, x) = 1. Indeed, we have
E{y} x
= 0,
ts


ts
E{y 2 } 2E{y}x + x2
= lim
+ 0 = 1.
2 (x, s) = lim
ts
ts
ts
ts
(x, s) = lim
ts

Hence, the Kolmogorov forward and backward equations are given by

p 1 2 p

= 0,
(36)
t
2 y 2
p 1 2 p
+
= 0.
(37)
s 2 x2
Directly evaluating the partial derivatives of the transition density leads to the
same results.
Note finally that the Wiener process W = {Wt , t R+ } is a martingale.
Since E{Wt Ws |Ws } = 0, w.p. 1, and E{Ws |Ws } = Ws , w.p. 1, we have
E{Wt |Ws } = Ws , w.p. 1.

2.4

The Brownian Bridge

Definition 14. Let W : R+ R be a standard Wiener process. The

(0,x;T,y)
Brownian bridge Bt
is a stochastic process defined sample pathwise such
that
t
(0,x;T,y)
Bt
() = x + Wt () (WT () y + x)
(38)
T
for 0 t T .
This means that the Brownian bridge is a Wiener process for which the
(0,x;T,y)
sample paths all start at Bt
() = x R and pass through a given point
(0,x;T,y)
BT
() = y R at a later time T for all .
(0,x;T,y)

Property 14.1. A Brownian bridge Bt

and covariances respectively given by

t
(x y),
T
st
= min{s, t} ,
T

t = x
Cs,t

is a Gaussian process with means

for 0 s, t T .
11

(39)
(40)

2.5

White Noise

Definition 15. The power spectral density of a (wide-sense) stationary process

X = {Xt , t R} is defined as the Fourier transform of its covariance:
Z
Ct ejt dt,
(41)
S() =

where = 2f and C0,t Ct .

The spectral density measures the power per unit frequency at frequency
f and the variance of the process can be interpreted as the average power (or
energy):
Z
1
Var{Xt } = C0 =
S() d.
(42)
2
R jt
1
e S() d is the inverse Fourier transNote that the covariance Ct = 2

form of the spectral density.

Definition 16. Gaussian white noise is a zero-mean wide-sense stationary process with constant nonzero spectral density S() = S0 for all R.
Hence, its covariances satisfy Ct = S0 (t) for all t R+ . Without loss of
generality, if we assume that S0 = 1, it can be shown that Gaussian white noise
corresponds to the following limit process:
lim Xth =

h0

Wt+h Wt
,
h

(43)

where W = {Wt , t R+ }. Hence, it is the derivative of the Wiener process.

However, the sample paths of a Wiener process are not differentiable anywhere.
Therefore, Gaussian white noise cannot be realized physically, but can be approached by allowing a broad banded spectrum (that is let h grow).
A similar result holds for the Ornstein-Uhlenbeck process with .

2.6

Multivariate Diffusion Processes

See for example [3, p. 68].

Ito Stochastic Calculus

The ordinary differential equation dx/dt = (t, x) can be viewed as a degenerate

form a stochastic differential equation as no randomness is involved. It can be
written in symbolic differential form
dx = (t, x) dt

12

(44)

or as an integral equation
Z

(s, x(s)) ds

x(t) = x0 +

(45)

t0

where x(t; t0 , x0 ) is a solution satisfying the initial condition x0 = x(t0 ). For

some regularity conditions on , this solution is unique, which means that the
future is completely defined by the present given the initial condition.
The symbolic differential form of a stochastic differential equation is written
as follows:
dXt = (t, Xt ) dt + (t, Xt )t dt

(46)

where X = {Xt , t R+ } is a diffusion process and t N (0, 1) for each t, i.e.

it is a Gaussian process. In particular, (46) is called the Langevin equation if

(t, Xt ) =
Xt and (t, Xt ) = for constant
and .
This symbolic differential can be interpreted as the integral equation along
a sample path
Z t
Z t
Xt () = Xt0 () +
(s, Xs ()) ds +
(s, Xs ())s () ds
(47)
t0

t0

for each . Now, for the case where a 0 and 1, we see that t should
be the derivative of Wiener process Wt = Xt , i.e it is Gaussian white noise.
This suggests that (47) can be written as follows:
Z t
Z t
Xt () = Xt0 () +
(s, Xs ()) ds +
(s, Xs ()) dWs ().
(48)
t0

t0

The problem with this formulation is that the Wiener process Wt is (almost
surely) nowhere differentiable such that the white noise process t does not
exist as a conventional function of t. As a result, the second integral in (48)
cannot be understood as an ordinary (Riemann or Lebesgue) integral. Worse,
it is not a Riemann-Stieltjes integral since the continuous sample paths of a
Wiener process are not of bounded variation for each sample path. Hence, it is
at this point that Itos stochastic integral comes into play!

3.1

Ito Stochastic Integral

In this section, we consider a probability space (, A, P ), a Wiener process

W = {Wt , t R+ } and an increasing family {At , t 0} of sub--algebras of A
such that Wt is At -measurable for each t 0 and with
E{Wt |A0 } = 0 and E{Wt Ws |As } = 0, w.p. 1,
for 0 s t.
Itos starting point is the following. For constant (t, x) the second
t () Wt ()}.
integral in (48) is expected to be equal to {W
0
13

We consider the integral of the random function f : T R on the unit

time interval:
Z 1
I[f ]() =
f (s, ) dWs ().
(49)
0

First, if the function f is a nonrandom step function, that is f (t, ) = fj on

tj t < tj+1 for j = 1, 2, . . . , n 1 with 0 = t1 < t2 < . . . < tn = 1, then we
should obviously have
I[f ]() =

n1
X

fj {Wtj+1 () Wtj ()}, w.p. 1.

(50)

j=1

Note that this integral is a random variable with zero mean as it is a sum of
random variables with zero mean. Furthermore, we have the following result
E{I[f ]()} =

n1
X

fj2 (tj+1 tj ).

(51)

j=1

Second, if the function f is a random step function, that is f (t, ) = fj ()

on tj t < tj+1 for j = 1, 2, . . . , n 1 with t1 < t2 < . . . < tn is Atj -measurable
and mean square integrable over , that is E{fj2 } < for j = 1, 2, . . . , n. The
stochastic integral I[f ]() is defined as follows:
I[f ]() =

n1
X

fj (){Wtj+1 () Wtj ()}, w.p. 1.

(52)

j=1

Lemma. For any a, b R and any random step function f, g such that fj , gj
on tj t < tj+1 for j = 1, 2, . . . , n 1 with 0 = t1 < t2 < . . . < tn = 1 is Atj measurable and mean square integrable, the stochastic integral (52) satisfies the
following properties:
I[f ] is A1 measurable,
E{I[f ]} = 0,
P
E{I 2 [f ]} = j E{fj2 }(tj+1 tj ),

(53)
(54)

I[af + bg] = aI[f ] + bI[g], w.p. 1.

(56)

(55)

Since fj is Atj -measurable and {Wtj+1 Wtj } is Atj+1 -measurable, each term
fj {Wtj+1 Wtj } is Atj+1 -measurable and thus A1 -measurable. Hence, I[f ] is
A1 -measurable.
From the Cauchy-Schwarz inequality4 and the fact that each term in (52) is
mean-squre integrable, it follows that I[f ] is integrable. Hence, I[f ]() is again
4 The

R
2 R
R

Cauchy-Schwarz inequality states that ab f g dx ab f 2 dx ab g 2 dx.

14

a zero mean random variable:

E{I[f ]} =

n1
X

E{fj (Wtj+1 Wtj )} =

j=1

n1
X

E{fj E{Wtj+1 Wtj |Atj }} = 0.

j=1

Furthermore, I[f ] is mean square integrable:

E{I 2 [f ]} =

n1
X

E{fj2 }E{(Wtj+1 Wtj )2 |Atj } =

j=1

n1
X

E{fj2 }(tj+1 tj ).

j=1

Finally, af + bg is a step random step function for any a, b R. Therefore,

we obtain (56), w.p. 1, after algebraic rearrangement.
Third, if the (continuous) function f is a general integrand such that it
f (t, ) is At -measurable and mean square integrable, then we define the stochastic integral I[f ] as the limit of integrals I[f (n) ] of random step functions f (n)
converging to f . The problem is thus to characterize the limit of the following
finite sums:
I[f

(n)

]() =

n1
X

(n)

f (tj , ){Wtj+1 () Wtj ()}, w.p. 1.

(57)

j=1
(n)

where f (n) (t, ) = f (tj , ) on tj t tj+1 for j = 1, 2, . . . , n 1 with

t1 < t2 < . . . < tn . From (55), we get
E{I 2 [f (n) ]} =

n1
X

(n)

E{f 2 (tj , )}(tj+1 tj ).

j=1

R1
This converges to the Riemann integral 0 E{f 2 (s, )} ds for n . This
result, along with the well-behaved mean square property of the Wiener process,
i.e. E{(Wt Ws )2 } = t s, suggests defining the stochastic integral in terms
of mean square convergence.
Theorem 3. The Ito (stochastic) integral I[f ] of a function f : T R is
the (unique) mean square limit of sequences I[f (n) ] for any sequence of random
step functions f (n) converging5 to f :
I[f ]() = m.s. lim
n

5 In

n1
X

(n)

f (tj , ){Wtj+1 () Wtj ()}, w.p. 1.

(59)

j=1

particular, we call the sequence f (n) mean square convergent to f if

Z t
ff
E
(f (n) (, ) f (, ))2 d 0, for n .
s

15

(58)

R1
The properties (5356) still apply, but we write E{I 2 [f ]} = 0 E{f 2 (t, )} dt
for (56) and call it the Ito isometry (on the unit time interval).
Similarly, the time-dependent Ito integral is a random variable defined on
any interval [t0 , t]:
Z t
f (s, ) dWs (),
(60)
Xt () =
t0

which is At -measurable and mean square integrable. From the independence of

non-overlapping increments of a Wiener process, we have E{Xt Xs |As } = 0,
w.p. 1, for any t0 s t. Hence, the process Xt is a martingale.
As the Riemann and the Riemann-Stieltjes integrals, (60) satisfies conventional properties such as the linearity property and the additivity property.
However, it has also the unusual property that
Z t
1
1
(61)
Ws () dWs () = Wt2 () t, w.p. 1,
2
2
0
where W0 = 0, w.p. 1. Note that this expression follows from the fact that
P
P
1
1
2
2
(62)
j Wtj (Wtj+1 Wtj ) = 2 Wt 2
j (Wtj+1 Wtj ) ,
where the second term tends to t in mean square sense.
Rt
By contrast, standard non-stochastic calculus would give 0 w(s) dw(s) =
1 2
2 w (t) if w(0) = 0.

3.2

The Ito Formula

The main advantage of the Ito stochastic integral is the martingale property.
However, a consequence is that stochastic differentials, which are interpreted as
stochastic integrals, do not follow the chain rule of classical calculus! Roughly
speaking, an additional term is appearing due to the fact that dWt2 is equal to
dt in the mean square sense.
Consider the stochastic process Y = {Yt = U (t, Xt ), t 0} with U (t, x)
having continuous second order partial derivatives.
If Xt were continuously differentiable, the chain rule of classical calculus
would give the following expression:
dYt =

U
U
dt +
dXt .
t
x

(63)

This follows from a Taylor expansion of U in Yt and discarding the second

and higher order terms in t.
When Xt is a process of the form (60), we get (with equality interpreted in
the mean square sense)


U
1 2U
U
dYt =
+ f 2 2 dt +
dXt ,
(64)
t
2 x
x
16

where dXt = f dWt is the symbolic differential form of (60). The additional
term is due to the fact that E{dXt2 } = E{f 2 }dt gives rise to an additional term
of the first order in t of the Taylor expansion for U :




U
2U
U
1 2U 2
2U
2
Yt =
t
+
2
x
+ ....
t +
x +
tx
+
t
x
2 t2
tx
x2
Theorem 4. Consider the following general stochastic differential:
Z t
Z t
Xt () Xs () =
e(u, ) du +
f (u, ) dWu ().
s

(65)

U
Let Yt = U (t, Xt ) with U having continuous partial derivatives U
t , x and
2U
x2 . The Ito formula is the following stochastic chain rule:

Z t
Z t
1 2 2U
U
U
U
Yt Ys =
+ eu
+ fu 2 du +
dXu , w.p. 1. (66)
t
x
2
x
s
s x

The partial derivatives of U are evaluated at (u, Xu ).

From the Ito formula, one can recover (61). For Xt = Wt and u(x) = xm ,
we have
m(m 1) (m2)
(m1)
d(Wtm ) = mWt
dWt +
Wt
dt.
(67)
2
In the special case m = 2, this reads d(Wt2 ) = 2Wt dWt + dt, which leads to
Z t
Wt2 Ws2 = 2
Wt dWt + (t s).
(68)
s

Hence, we recover

3.3

Rt
0

Wt dWt =

1
2
2 Wt

12 t for s = 0.

Multivariate case

See [3, p. 97].

Probability Distributions

Definition 17. The probability density function of the Poisson distribution

with parameter > 0 is defined as follows:
P(n|) =

n
e ,
n!

(69)

for n N. Note that E{n} = and E{(n E{n})2 } = .

Definition 18. The probability density function of the multivariate Gaussian
distribution with mean vector and covariance matrix is given by
1

N (x|, ) = (2)D/2 ||1/2 e 2 (x)

where RDD is symmetric and positive definite.
17

1 (x)

(70)

References
[1] Lawrence E. Evans. An Introduction to Stochastic Differential Equations.
Lecture notes (Department of Mathematics, UC Berkeley), available from
http://math.berkeley.edu/evans/.
[2] Crispin W. Gardiner. Handbook of Stochastic Methods. Springer-Verlag,
Berlin, 2004 (3rd edition).
[3] Peter E. Kloeden and Eckhard Platen. Numerical Solution of Stochastic
Differential Equations. Springer-Verlag, Berlin, 1992.
[4] Bernt ksendael. Stochastic Differential Equations. Springer-Verlag, Berlin,
1985.

18