lOMoARcPSD|3446508

Stochastic analysis in Finance II

Stochastic Analysis in Finance (The University of Edinburgh)

StuDocu is not sponsored or endorsed by any college or university

Downloaded by jaslskd Afas ([email protected])
 lOMoARcPSD|3446508

Stochastic Analysis in Finance

Basic Stochastic Analysis
István Gyöngy
School of Mathematics, 5612 JCMB

1. Basic Notions and Notation

2. Introduction
Ordinary differential equations (ODEs)
dXt
= b(Xt )
dt
are often used to model the evolution of a quantity Xt , when the change ∆Xt := Xt+∆t − Xt ,
during every short time interval [t, t + ∆t] is approximately equal to b(Xt )∆t, where b is some
function. This differential equation is also written in the more intuitive differential form,
dXt = b(Xt ) dt,
or in the integral form Z t
Xt = X0 + b(Xs ) ds.
0
In many cases, however, ODEs are not suitable to describe the evolution of the quantity Xt ,
if its change is influenced also by some random phenomenon. In order to find some good
mathematical model in this situation, we want to make a precise mathematical meaning of the
heuristic expression
∆Xt ≈ b(Xt ) ∆t + “random change” .
First, depending on the nature of the random phenomenon, we have to find a suitable stochastic
process to model the “random change”. When Xt changes continuously, then we look for a
continuous stochastic process. In this case it is natural to model the “random change” by
σ(Xt ) ∆Wt := σ(Xt )(Wt+∆t − Wt ),
where W = {Wt : t ≥ 0} is a Brownian motion, and σ is a function. The role of σ is to magnify
the changes of the Brownian motion at each time t by a factor σ(Xt ), in order to model the
volatility of the random effect. Now our task is to define precisely the expression
∆Xt ≈ b(Xt ) ∆t + σ(Xt )∆Wt .
The main difficulty which arises is due to the irregularity of the Brownian trajectories. Namely,
by a well-known theorem of Erdős and Dvoretzky, the trajectories of the Brownian motion are
nowhere differentiable. This means, we cannot write the above expression as
dWt
∆Xt ≈ b(Xt ) ∆t + σ(Xt ) ∆t,
dt
and interpret it as an ODE, because dW dt
t
does not exist. Therefore we want to give a meaning
to the integral form Z t Z t
Xt = X0 + b(Xs ) ds + σ(Xs ) dWs .
0 0
There is no difficulty in understanding the first integral, but the integral
Z t
σ(Xs ) dWs
0
is not defined in classical analysis, because the Brownian trajectories are too irregular. Never-
theless, about half a century ago a famous mathematician, Itô gave a precise meaning of such
1

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

integrals and developed the theory of stochastic differential equations based on his notion of
stochastic integral. It turned out that Itô’s theory can be successfully applied in many areas
of sciences and engineering. It plays a fundamental role also in financial mathematics. Next
we give a brief presentation of Itô’s integral.

3. Wiener process
Notice that ({σ(Xt ))t≥0 are stochastic processes for functions σ and stochastic processes
(Xt )t≥0 . Thus to give a rigorousRmeaning of stochastic differential equations (SDEs) we need
t
to define the stochastic integral 0 fs dWs for a large class of stochastic processes (ft )t≥0 .
Let us first recall some definitions and notation. We fix a probability space (Ω, F, P ) and a
flow of information (called also filtration) F = (Ft )t≥0 , i.e., F is a family of σ-algebras Ft ⊂ F
for t ≥ 0, such that Fs ⊂ Ft for 0 ≤ s ≤ t. We assume that F0 contains all P -zero sets. We
say that A is a P -zero set if it is a subset of a set of probability zero.
Definition 3.1. A random process (Wt )t≥0 is a Wiener process if
(i) it has independent increments, i.e.
Wt1 − Wt0 , . . . , Wtn − Wtn−1
are independent random variables for any sequence of ‘time points’ for 0 ≤ t0 ≤
t1 , · · · , ≤ tn , for integers n ≥ 2;
(ii) Wt − Ws ∼ N (0, t − s) for all 0 ≤ s ≤ t;
(iii) it is a continuous process, i.e., its trajectories, W· (ω), ω ∈ Ω, are continuous functions
of t;
(iv) W0 = 0.
Exercise 3.1. Let W be a Wiener process. Show that for every constant c 6= 0 the process
(cWt/c2 )t≥0 is also a Wiener process.
Exercise 3.2. Assume that a stochastic process W = (Wt )t≥0 has normally distributed in-
crements with mean 0 and variance equal to the length of the time increment, i.e. (ii) of the
previous definition holds. Then W has independent increments if and only if
(i’) Wt − Ws is independent of G := σ(Wr : r ≤ s) for all 0 ≤ s < t.
By virtue of this exercise we can replace (i) in Definition 3.1 with (i’) to get and equivalent
definition of the Wiener process.
Definition 3.2. A random process (Wt )t≥0 is a Wiener process if
(i’) Wt − Ws is independent of G := σ(Wr : r ≤ s) for all 0 ≤ s < t;
(ii) Wt − Ws ∼ N (0, t − s) for all 0 ≤ s ≤ t;
(iii) it is a continuous process, i.e., its trajectories, W· (ω), ω ∈ Ω, are continuous functions
of t;
(iv) W0 = 0.
In applications, for example in financial mathematics, it often happens that the increments
Wt − Ws of the Wiener process are independent of Fs for all 0 ≤ s ≤ t for a flow of information
(Ft )t≥0 , larger than the history of the Wiener process (Gt )t≥0 , Gt := σ(Wr : r ≤ t).
We capture this situation by the following definition. Recall that a stochastic process (Xt )t≥0
is called Ft -adapted if Xt is Ft -measurable for each t ≥ 0.
Definition 3.3. A process (Wt )t≥0 is called an Ft -Wiener process (or Wiener martingale with
respect to (Ft )t≥0 ) if W is an Ft -adapted Wiener process and Wt − Ws is independent of Ft for
all 0 ≤ s ≤ t.

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

Definition 3.4. A stochastic process ((Mt )t∈I , I ⊂ R is called a martingale with respect to a
filtration (Ft )t∈I (or Ft -martingale) if
(i) Mt is Ft measurable for every t ∈ I
(ii) E|Mt | < ∞ for every t ∈ I, and
E(Mt |Fs ) = Ms for all s ≤ t.
Exercise 3.3. Show that an Ft -Wiener process is a martingale with respect to the filtration
(Ft )t≥0 . Since it is Ft adapted by definition, you need only prove that E|Wt | < ∞ and
E(Wt |Fs ) = Ws for all 0 ≤ s ≤ t.
Since an Ft -Wiener process is an Ft -martingale, by virtue of the previous definition, we call
it also an Ft -Wiener martingale, or a Wiener martingale with respect to (Ft )t≥0 .
Exercise 3.4. Let W be an Ft Wiener martingale. Show that Xt := Wt2 − t, t ≥ 0 and
2
Yt = eλWt −λ t/2 , t ≥ 0 are Ft -martingales, for every λ ∈ R.
The following property of the Wiener process is extremely important in stochastic analysis.
To formulate it we fix an interval [s, t] ⊂ [0, ∞) and a partition
s ≤ tn0 ≤ tn1 ≤ . . . tnk(n) ≤ t
of it for each integer n ≥ 1 such that
d(n) := max |tnj+1 − tnj | → 0 as n → ∞. (3.1)
j

Set
n(k)−1 n(k)−1
X X
2
Sn = |Wtnj+1 − Wtnj | and An = |Wtnj+1 − Wtnj |.
j=1 j=1

Theorem 3.1. If n → ∞ then Sn → t − s in mean square, i.e.,

lim E(Sn − (t − s))2 = 0.
n→∞

If (Xt )t≥0 is a stochastic process such that

n(k)−1
X
Sn = |Xtnj+1 − Xtnj |2
j=1

has a limit in probability for every sequence s ≤ tn0 ≤ tn1 ≤ . . . ≤ tnk(n) of partitions satisfying
(3.1), then clearly, this limit does not depend on the sequence of partitions and it is called
the quadratic variation of X over the interval [s, t]. The quadratic variation of X over [0, t] is
denoted by [X]t . Thus by virtue of the previous theorem
[W ]t = t for all t ≥ 0.
Corollary 3.2. Almost surely
lim An = ∞.
n→∞

This corollary tells us that almost surely the trajectories of a Wiener process has infinite
first variation over any interval [s, t]. Therefore we cannot use the notion of Lebesgue-Stieltjes
integral to define the stochastic integral pathwise, i.e., for each ω. Therefore Itô has chosen
another way to define it, which we will present below.
We have learnt above that an Ft -Wiener martingale W is a continuous martingale with
respect to (Ft )t≥0 and [W ]t = t for all t ≥ 0. These properties characterise the Wiener
martingales by virtue of the following theorem of P. Lévy.

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

Theorem 3.3. Let (Vt )t≥0 be a continuous local martingale with respect to a filtration (Gt )t≥0
and assume that [V ]t = t for all t ≥ 0. Then (Vt )t≥0 is a Wiener martingale with respect to
(Gt )t≥0 .
4. Stochastic Itô integral
We fix an Ft -Wiener martingale W , and start with a class of integrands, called simple
processes defined as follows.
Definition 4.1. A stochastic process (ft )t≥0 is called a simple process if there exist a finite
number of time points time points 0 ≤ t0 ≤ t1 ≤ . . . ≤ tN and an Ftj -measurable bounded
random variable Zj for each j = 0, . . . N − 1, such that
N
X −1
ft = Zj 1(tj ,tj+1 ] (t), for t ≥ 0, (4.1)
j=0

where 1A denotes the indicator of a set A, i.e., 1A (t) = 1 if t ∈ A and equals 0 otherwise. We
denote the set of all simple processes by H0 .
Notice that 1A f ∈ H0 for f ∈ H0 and intervals A = (a, b] ⊂ (0, ∞).
Exercise 4.1. If f ∈ H0 has the representation (4.1), then for any a ∈ (0, ∞] we have
N
X −1
1(0,a] f = Zi 1(ti ∧a,ti+1 ∧a] ,
i=0

where a ∧ b := min(a, b).

Definition 4.2. For f ∈ H0 given by (4.1) we define
Z ∞ N
X −1
fs dWs := Zj (Wtj+1 − Wtj ) =: I(f ),
0 j=0

and Z t Z ∞
fs dWs := 1(0,t] (s)fs dWs =: It (f )
0 0
for every t ≥ 0.
Proposition 4.1. For f ∈ H0 we have
(1) Itô’s identity Z ∞
2
EI (f ) = E fs2 ds.
0
(2) EI(f ) = 0
One can easily verify that I is a linear operator on H0 i.e.,
It (αf + βg) = αIt (f ) + βIt (g)
for all f, g ∈ H0 and α, β ∈ R.
Proposition 4.2. For f ∈ H0 he integral process (It (f ))t≥0 is an F-martingale, i.e., it is
Ft -adapted and
E(It (f )|Fs ) = Is (f ) (a.s.) for every 0 ≤ s ≤ t.
Definition 4.3. We denote by H the set of Ft -adapted stochastic processes (ft )t≥0 , such that
almost surely Z ∞
E ft2 dt < ∞.
0

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

Exercise 4.2. Show that 1A f ∈ H if f ∈ H and A = (a, b] ⊂ [0, ∞).

Lemma 4.3. For f ∈ H there exists f (n) ∈ H0 for integers n ≥ 1 such that
Z ∞
(n)
lim E (ft − ft )2 dt = 0. (4.2)
n→∞ 0

We say that f (n) ∈ H0 is an approximating sequence from H0 for f in H if (4.2) holds.

Lemma 4.4. If f (n) is an approximating sequence from H0 for f ∈ H then
lim E(|I(f (n) ) − I(f (m) )|2 ) = 0.
n,m→∞

Hence I(f (n) ) converges in mean square to a random variable, which depends only on f , i.e., if
f¯(n) is another approximating sequence from H0 for f , then the mean square limits of I(f (n) )
and I(f¯(n) ) coincide almost surely.
Proof. By the linearity of the stochastic integral and by Itô’s isometry we have
E(|I(f (n) ) − I(f (m) )|2 ) = EI 2 (f (n) − f (m) )
Z ∞ Z ∞ Z ∞
(n) (m) 2 (n) 2
=E |f − f | dt ≤ E |f − f | dt + E |f − f (m) |2 dt → 0,
0 0 0
(n)
i.e., I(f is a Cauchy sequence in mean square, which implies its mean square convergence.
Arranging both sequences f (n) and f¯(n) into one sequence g (n) , we get again an approximating
sequence from H0 and then I(g (n) ) converges in mean square. Therefore its sub-sequences
I(f (n) and I(f¯(n) have the same limit. 
R∞
Definition 4.4. For f ∈ H we define the stochastic integral 0 ft dWt as the limit in mean
square of I(f (n) ) for an approximating sequence f (n) from H0 for f . We denote this limit also
by I(f ). We define
It (f ) = I(1(0,t] f ) for t ≥ 0
Clearly, the stochastic integral is a linear operator on H. One can show that Itô’s isometry
and the martingale property remain valid.
Theorem 4.5. For f ∈ H we have
(1) Itô’s identity Z ∞
2
E(I (f )) = E ft2 dt, (4.3)
0
(2)
EI(f ) = 0 (4.4)
(3) It (f ) is Ft -measurable for every t ≥ 0, and
E(It (f )|Fs ) = Is (f ) (a.s.) for all 0 ≤ s ≤ t. (4.5)

Using properties (1) and (3) for f (n) ∈ H0 one can show that for f ∈ H the integral process
(It (f ))t≥0 has a continuous version, which means that there exists a continuous stochastic
process (Yt )t≥0 , such that P (Yt = It (f )) = 1 for every t ≥ 0.
We use the notation Z t
fs dWs , t≥0
0
for the continuous version of the integral process (It (f ))t≥0 .
One can extend the notion of stochastic integrals to a larger class of integrands.

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

Definition 4.5. We denote by S the set of Ft -adapted processes (ft )t≥0 satisfying
Z T
ft2 dt < ∞ (a.s.) for all T > 0.
0

It is useful to recall the notion of stopping times with respect to a filtration F = (Ft )t≥0 .
Definition 4.6. A mapping τ : Ω → R ∪ {∞} is called a stopping time (with respect to F if
[τ ≤ t] ∈ Ft for every t ≥ 0.
Exercise 4.3. Let τ be a stopping time with respect to (Ft )t≥0 . Show that Xt = 1(0,τ ] (t) is
Ft -measurable for every t ≥ 0.
Exercise 4.4. Let f ∈ S. Show the existence of a sequence of processes f (n) from H such
that Z T
(n)
(ft − ft )2 dt → 0 in probability for each T . (4.6)
0

By exercise 4.4 for f ∈ S there exists a sequence f (n) ∈ H, such that (4.6) holds. Then for
each n the integral process Z t
fs(n) dWs t ≥ 0
0
is defined, and one can show that there exists an Ft -adapted continuous stochastic process
Z = (Zt )t≥0 such that for n → ∞
Z t
sup | fs(n) dWs − Zt | → 0 in probability for each T > 0.
t∈[0,T ] 0

Moreover, Z depends only on f and not on the sequence f (n) .

Definition 4.7. We say that Zt , defined above is the Itô integral of f over [0, t] and we denote
it by Z t
fs dWs .
0
Remark 4.1. Notice that H0 ⊂ H ⊂ S. It is important to have in mind that the integral
process Z t
fs dWs t ≥ 0
0
is an F-martingale if f ∈ H. In particular,
Z t
E fs dWt = 0 for f ∈ H.
0
In general, the integral process is not a martingale, and its expectation is not 0, if f is only
from S. It is however still inherits something from the martingale property.
Definition 4.8. A continuous Ft -adapted process Z = (Zt )t≥0 is called local martingale with
respect to (Ft )t≥0 if there exists a sequence of stopping times (τn )∞
n=1 (with respect to (Ft )t≥0 ),
such that almost surely τn → ∞ as n → ∞, and
(n)
Zt := Zt∧τn , t≥0
is a (continuous) martingale (with respect to (Ft )t≥0 ) for each n.
Proposition 4.6. Let f ∈ S and let τ be a stopping time. Then for every t ≥ 0
Z t∧τ Z t
fs dWs = 1(0,τ ] fs dWs .
0 0

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

Proposition 4.7. If f ∈ S then the integral process is a continuous local martingale.

Let T be a fixed positive number. Then the restriction of the classes H and S onto [0, T ]
are denoted by H([0, T ]) and S([0, T ]), respectively, i.e., S([0, T ]) denotes the set of adapted
processes f = (ft )t∈[0,T ] such that
Z T
ft2 dt < ∞ (a.s.),
0

and H([0, T ]) denotes the set of adapted processes g = (gt )t∈[0,T ] such that
Z T
E ft2 dt < ∞.
0

Clearly, if f ∈ S([0, T ]) and g ∈ H([0, T ]) then f¯ ∈ S and ḡ ∈ H, where f¯ and ḡ are the
extensions of f and g respectively, defined as
f¯t = tt , ḡt = gt , for t ≤ T , and f¯t = 0, ḡt = 0 for t > T .
For f ∈ S([0, T ]) we set
Z t Z t
fs dWs = f¯s dWs for t ≤ T .
0 0
Summary: Rt
• The Itô integral 0 fs dWs is defined for each t ∈ [0, T ] when (ft )t∈[0,T ] is an Ft -adapted
process, satisfying
Z T
ft2 dt < ∞. (4.7)
0
Rt
In this case the integral process 0 fs dWs , t ∈ [0, T ], is a continuous local martingale (with
respect to (Ft )t≥0 .
• If (ft )t∈[0,T ] is an Ft -adapted process, satisfying
Z T
Eft2 dt < ∞, (4.8)
0
Rt
then the integral process 0 fs dWs , t ∈ [0, T ] is an F-martingale, which means that it is
Ft -adapted has finite expectation for each t ≥ 0, and
Z t  Z s
E fr dWr |Fs = fr dWr for all 0 ≤ s ≤ t ≤ T.
0 0

In particular,
Z t 
E fs dWs = 0 for all t ∈ [0, T ].
0
Moreover, Itô’s isometry holds:
Z t 2 Z t
E fs dWs = Efs2 ds < ∞ for all t ∈ [0, T ].
0 0

Remark 4.2. Itô’s isometry implies the identity

Z t Z t  Z t
E fs dWs gs dWs = E(fs gs ) ds for all t ∈ [0, T ].
0 0 0

valid for all f, g ∈ HT

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

Exercise 4.5. Let (ft )t∈[0,T ] be a deterministic process from H(0, T ]). (This means f does not
RT
depend on ω and 0 ft2 dt < ∞.) Then
Z t
fs dWs
0
RT
is a normal random variable with mean zero and variance 0
fs2 ds for every t ∈ [0, T ].
Remark 4.3. One can also show that if f is a deterministic process from H([0, T ]) then the
integral process Z t
Xt = fs dWs , t ∈ [0, T ]
0
is a Gaussian process, i.e., (Xt1 , Xt2 ...., Xtk ) is a normal random vector for every finite sequence
0 ≤ t1 < t2 < ... < tk ≤ T .
In applications we often need to work with multi-dimensional processes.
Definition 4.9. An m-dimensional process W = (Wt1 , . . . , Wtm )t≥0 is called an m-dimensional
Wiener martingale with respect to F = (Ft )t≥0 , if W 1 ,...,W m are independent Wiener martin-
gales w.r. to F.
We write σ ∈ S d×m if σ = (σtij )t≥0 is a stochastic process with values in Rd×m , the space
of d × m-matrices, such that each component σ ij belongs to S. For σ ∈ S d×m and an m-
dimensional Wiener martingale W = (W 1 , ..., W m ) the notation
Z t
σs dWs , t ≥ 0
0
means the d-dimensional stochastic process Y with co-ordinates
Xm Z t
i
Yt = σsij dW j , i = 1, 2, . . . , d.
j=1 0

We use the notation S d×m ([0, T ]) for the set of processes from S d×m restricted to [0, T ], i.e.,
S d×m ([0, T ]) denotes the set of Rd×m -valued processes f = (f ij )t∈[0,T ] such that f ik ∈ S([0, T ])
for each i = 1, ..., d and k = 1, 2, ..., m.

5. The Itô formula

Let A denote the set of Ft -adapted processes (gt )t≥0 such that
Z T
|gs | ds < ∞ (a.s.) for all T > 0.
0
For T > 0 fixed we use also the notation A([0, T ]) for the set of processes from A, restricted
to the interval [0, T ] in t, i.e., A([0, T ]) is the set of Ft -adapted processes g = (gt )t∈[0,T ] such
that Z T
|gt | < ∞ (a.s.).
0

Definition 5.1. We call a process X = (Xt )t∈[0,T ] an Itô process, if almost surely
Z t Z t
Xt = X + bs ds + σs dWs for all t ∈ [0, T ], (5.9)
0 0
where X is an F0 -measurable random variable, b ∈ A([0, T ]) and σ ∈ S([0, T ]). In this case
we say that X has the stochastic differential
dXt = bt dt + σt dWt on [0, T ]. (5.10)

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

The processes b and σ are often called the drift and diffusion parameters (of X), respectively.
Let C 1,2 ([0, T ] × R) denote the set of real valued continuous functions u = u(t, x) of (t, x) ∈
∂ ∂ ∂2
[0, T ] × R such that the partial derivatives ut := ∂t u, ux := ∂x and uxx := ∂x 2 exist and are

continuous in (t, x) ∈ [0, T ] × R. Then we have the following generalization of the classical
chain rule.
Theorem 5.1. (Itô’s formula) Let X be an Itô process with stochastic differential (5.10). Let
u ∈ C 1,2 ([0, T ] × R). Then Yt = (u(t, Xt ))t∈[0,T ] is an Itô process with stochastic differential
du(t, Xt ) = ut (t, Xt ) dt + ux (t, Xt ) dXt + 12 fxx (t, Xt ) (dXt )2 ,
where for multiplication and addition of stochastic differentials the usual rules of algebra (com-
mutativity, associativity and distributivity) are in force, together with the following ‘multipli-
cation table’
(dt)2 = 0, dt dWt = 0, (dWt )2 = dt.
Thus Itô’s formula reads as follows
 
du(t, Xt ) = ut (t, Xt ) + bt ux (t, Xt ) + 12 σt2 uxx (t, Xt ) dt + σt ux (t, Xt ) dWt .

Itô’s formula can be extended to multi-dimensional processes X and functions u of many

variables. To formulate it define the multi-dimensional Itô processes.
Let W = (W 1 , ..., W m ) be an m-dimensional Wiener martingale with respect to (Ft )t≥0 .
Recall that S d×m ([0, T ]) is the set of Rd×m -valued processes σ = (σtik ) such that σ ik ∈ S([0, T ])
for each i = 1, 2, ..., d and k = 1, ..., m. We use also the notation Ad ([0, T ]) for the set of
Rd -valued processes b = (bit )t∈[0,T ] , such that bi ∈ A([0, T ]) for each i = 1, ..., d.
Definition 5.2. We say that an Rd -valued process X = (Xti )t∈[0,T ] is a d-dimensional Itô
process with stochastic differential
dXt = bt dt + σt dWt , t ∈ [0, T ] (5.11)
if almost surely
Z t m Z
X t
Xti =X + i
bis ds + σsik dWsk
0 k=1 0
i
for all t ∈ [0, T ] and i = 1, ..., m, where X is an F0 -measurable random variable for each
i = 1, ..., d, b ∈ Ad ([0, T ]) and σ ∈ S d×m ([0, T ]).
Let C 1,2 ([0, T ] × Rd ) denote the set of continuous functions u = u(t, x) of t ∈ [0, T ] and
x = (x1 , x2 , ..., xd ) ∈ Rd , such that the partial derivatives ut , uxi and uxi xj exist and are
continuous in (t, x) ∈ [0, T ] × Rd for each i, j = 1, 2..., d.
Theorem 5.2. Let X be a d-dimensional Itô process with stochastic differential (5.11), and
let u ∈ C 1,2 ([0, T ] × Rd ). Then (u(t, Xt ))t∈[0,T ] is an Itô process with stochastic differential
d
X d
X
du(t, Xt ) = ut (t, Xt ) dt + uxi (t, Xt ) dXti + 1
2
uxi xj (t, Xt )dXti dXtj , (5.12)
i=1 i,j=1

where for the addition and multiplication of stochastic differentials the usual arithmetic rules
are in force, together with the multiplication table dtdWti = (dt)2 = 0, (dWti )2 = dt and
dWti dWtj = 0 for every i 6= j. Thus applying these rules we have
m
X
dXti dXtj = σtik σtjk dt = (σt σt∗ )ij dt,
k=1

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

10

and (5.12) has the form

d
X d
X
bit uxi (t, Xt ) 1
(σt σt∗ )ij uxi xj (t, Xt ) dt


du(t, Xt ) = ut (t, Xt ) + + 2
i=1 i,j=1
m X
X d
+ σtik uxi (t, Xt ) dWtk . (5.13)
k=1 i=1

Remark 5.1. Introducing the partial differential operators

d d
∂ X ∂2 X ∂ ∂
L σt ,bt
= + 1
2
(σt σt∗ )ij + bit i , Mσk t = σtik , (5.14)
∂t i,j=1
∂xi ∂xj i=1
∂x ∂xi

k = 1, 2, ..., k, we can write formula (5.13) as

m
X
du(t, Xt ) = L σt ,bt
u(t, Xt ) dt + Mσk t uxi (t, Xt ) dWtk . (5.15)
k=1

Remark 5.2. The drift and diffusion coefficients in (5.11), b = (bt )t∈[0,T ] and ((σσ ∗ )ij
t )t∈[0,T ] ,
i, j = 1, ..., d, respectively, are uniquely determined by the process X, i.e., if
dXt = bt dt + σt dWt = b̄t dt + σ̄t dWt , t ∈ [0, T ]
then almost surely bt = b̄t and σ = σ̄t for almost every t ∈ [0, T ].
The special case of Theorem 5.2, when u = u(x, y) = xy for (x, y) ∈ R2 is often used. It is
called Itô’s product rule, and we formulate this separately.
Theorem 5.3. (Product rule) Let X and Y be Itô processes. Then XY has the stochastic
differential
d(Xt Yt ) = Yt dXt + Xt dYt + dXt dYt .
In particular, if
dXt = bt dt + σt dWt , dYt = b̄t dt + σ̄t dWt . (5.16)
then
d(Xt Yt ) = (Yt bt + Xt b̄t + σt σ̄t ) dt + (Yt σt + Xt σ̄t ) dWt .
Example 5.1. The following examples are simple applications of Itô’s formula. Verify the
following equalities.
1. deXt = eXt (at + 12 b2t ) dt + eXt bt dWt ;
2. dWtp = pWtp−1 dWt + p(p−1) 2
Wtp−2 dt; for any real number p ≥ 2;
3. d sin(Wt ) = cos(Wt ) dWt − 21 sin(Wt ) dt;
4. d(et/2 sin(Wt )) = et/2 d sin(Wt ), + sin(Wt ) det/2 = et/2 cos(Wt ) dWt .
ExerciseR t5.1. Let f be a deterministic process from S. Show that for each
R t 2t the stochastic
integral 0 fs dWs is a normal random variable with mean 0 and variance 0 fs ds.
Remark 5.3. One can solve the above exercise also by noticing that the Itô integral of a
deterministic simple process is a sum of independent normal random variables, which is a
normal random variable. The Itô integral of a deterministic process from H is the limit in
mean square of a sequence of Itô integrals of deterministic simple processes, i.e., it is the limit
in mean square of normal random variables. Hence it is also a normal random variable.

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

11

6. Stochastic Differential Equations

Let f and g be real functions, defined on [0, T ] × R. The expression
dXt = f (t, Xt ) dt + g(t, Xt ) dWt t ∈ [0, T ] (6.17)
is called a stochastic differential equation (SDE) for (Xt )t∈[0,T ] .
Definition 6.1. Let X be an F0 -measurable random variable. Then a process (Xt )t∈[0,T ] is a
solution of SDE (6.17) on the interval [0, T ], with initial condition X0 = X, if almost surely
Z t Z t
Xt = X + f (s, Xs ) ds + g(s, Xs ) dWs for all t ∈ [0, T ].
0 0

Examples 6.1. We consider here some examples of SDEs that are often used in applications.
1. (Black-Scholes SDE) Let µ and σ be some constants. We look for a solution of
dXt = µXt dt + σXt dWt (6.18)
with initial value X0 = X, where X is an F0 -adapted random variable.
By Example 1 from Examples 5.1 we know that for any constants λ and σ the process St :=
eλt+σWt has the stochastic differential
1
dSt = St (λ + σ 2 ) dt + St σ dWt
2
1 2
Taking here λ := µ − 12 σ 2 we see that the process St := e(µ− 2 σ )t+σWt is a solution of the the
1 2
above SDE with initial value S0 = 1. Hence Xt := XSt = Xe(µ− 2 σ )t+σWt is a solution with
initial value X0 = X. This process is somes called the ‘the geometric Brownian motion’.
2. (Orstein-Uhlenbeck process) Let α > 0 and σ be constants. Consider
dXt = −αXt dt + σ dWt (6.19)
with initial condition X0 , where X is an F0 -measurable random variable.
We can solve this equation by the method of integrating factor to get
 Z t 
−αt
Xt = e X0 + σ eαs dWs .
0
Using Itô’s formula we can verify that this process is indeed a solution.
3. (Vasicek’s interest rate model) The instantaneous interest rate rt , in the Vasicek model for
interest rate satisfies the SDE
drt = b(a − rt ) dt + σ dWt ,
where a, b, σ and r0 , the initial value, are positive constants. Denoting rt − a by Xt , we see
that Xt satisfies the SDE (6.19) with b in place of α. Hence
Z t
−bt −bt −bt
rt = a(1 − e ) + r0 e + e σ ebs dWs .
0
By Exercise 5.1 we know that rt is a normal random variable with mean
Ert = a(1 − e−bt ) + r0 e−bt
and variance
Z t
−2bt 2
V ar(rt ) = e σ e2bs ds = σ 2 e−2bt (e2bt − 1)/2b = σ 2 (1 − e−2bt )/2b.
0
Notice that for n → ∞
Ert → a, V ar(rt ) → σ 2 /2b.

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

12

4.(Cox-Ingersoll-Ross model) The evolution of the instantaneous interest rate rt is modelled

by the SDE p
drt = b(a − rt ) dt + σ |rt | dWt ,
where a, b, σ are positive constants. One can show that for every given initial condition r0 > 0
this equation has a unique solution (rt )t∈[0,T ] , and it takes only non-negative values.
Let W = (W 1 , ..., W m ) be an m-dimensional Wiener process with respect to (Ft )t≥0 , and
let b = b(t, x) and σ = σ(t, x) be (Borel) functions defined on [0, T ] × Rd and taking values in
Rd and in Rd×m , respectively. Then
dXt = b(t, Xt ) dt + σ(t, Xt ) dWt , t ∈ [0, T ] (6.20)
is a multi-dimensional SDE (or a system of SDEs).
Definition 6.2. Let X = X i be an F0 -measurable random vector with values in Rd . Then a
d-dimensional process X = (Xti )t∈[0,T ] is a solution of (6.20) with initial condition X0 = X if
almost surely
Z t Xm Z t
i i i
Xt = X + b (s, Xs ) ds + σ ik (s, Xs ) dWsk
0 k=1 0

holds for all t ∈ [0, T ], i = 1, ..., d.

In order to present a classical result existence and uniqueness theorem for (6.20) with initial
condition X0 = X, we need some conditions. We use the notation
d
X d X
X m
|b|2 := |bi |2 , |σ|2 := |σ ik |2
i=1 i=1 k=1

for vectors b = (b , ..., b ) ∈ R and matrices σ = (σ ) ∈ Rd×m .

1 d d ik

Linear growth condition: There is a constant K such that

|b(t, x)|2 + |σ(t, x)|2 ≤ K(1 + |x|2 ) for t ∈ [0, T ] and x ∈ Rd , (6.21)
Local Lipschitz condition on b and σ: For every R > 0 there is a constant LR such that
(i)
|b(t, x) − b(t, y)|2 LR |x − y|2 (6.22)
(ii)
|σ(t, x) − σ(t, y)|2 ≤ LR |x − y|2 (6.23)
d
for all t ∈ [0, T ] and x, y ∈ R such that |x| ≤ R and |y| ≤ R.
Theorem 6.1. Let the above linear growth and local Lipschitz conditions hold. Then for
every F0 -measurable random vector X with values in Rd , equation (6.20) with initial condition
X0 = X has a unique solution on [0, T ].
In the one dimensional case, i.e., when d = m = 1, the previous theorem remains valid if
instead of the local Lipschitz condition on σ we assume only the following condition σ:
Local Hölder continuity with exponent 1/2 of σ :
For each R > 0 there exists a constant LR such that
|σ(t, x) − σ(t, y)|2 ≤ LR |x − y| (6.24)
for all t ∈ [0, T ], x, y ∈ R, |x| ≤ R, |y| ≤ R.
Theorem 6.2. Let d = m = 1 and let the linear growth condition, the local Lipschitz condition
(i) on b and the local Hölder condition with exponent 1/2 on σ hold. Then for each F0 -
measurable random variable X, equation (6.20) has a unique has a solution X on [0, T ].

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

13

Theorem 6.3. (Comparison Theorem) Assume that the differential equations

dXt = b(t, Xt ) dt + σ(t, Xt ) dWt , (6.25)
and
dYt = B(t, Yt ) dt + σ(t, Yt ) dWt (6.26)
with initial conditions X0 = X and Y0 = Y , respectively, satisfy the conditions of the previous
theorem. Assume, moreover, that
b(t, x) ≤ B(t, x) for all t ≥ 0 and x ∈ R,
X≤Y (a.s.).
Then almost surely Xt ≤ Yt for all t ∈ [0, T ].
7. Girsanov’s theorem
The theorems of this section are very important results on changing measures.
Exercise 7.1. Let γ be a non-negative random variable, and define the function Q on F by
Q(A) = E(1A γ), for A ∈ F.
Then Q is a measure on F. It is a probability measure if and only if Eγ = 1.
Before formulating Girsanov’s theorem it useful to discuss the following simple analogue of
it, as an exercise.
Exercise 7.2. Let Z be a standard normal random random variable. Set X := m + Z for
some number m. Show that the measure Q defined by
Q(A) = E(1A exp(−mZ − 12 m2 )), A∈F
is a probability measure, and that X is a standard normal random variable under Q.
Theorem 7.1. (Girsanov’s theorem) Let W = (Wt )t≥0 be a Wiener martingale with respect
to F = (Ft )t≥0 , and consider the process
Z t
X t = Wt + bs ds, t ∈ [0, T ],
0
for some fixed T > 0, where b = (bt )t∈[0,T ] ∈ S([0, T ]). Define the measure Q on F by
Q(A) = E(1A γT ), A ∈ F,
where  Z T Z T 
γT := exp − bt dWt − 1
2
b2t dt .
0 0
Assume that EγT = 1. Then Q is a probability measure, and under Q the process Y is a
Wiener martingale on [0, T ] with respect to F .
Theorem 7.2. (Multidimensional Girsanov theorem) Let W be a d-dimensional Wiener mar-
tingale with respect to F, and for a process b ∈ S d let X = (X i ) be a d-dimensional process
defined by Z t
Xti = bis ds + Wti , t ∈ [0, T ], i = 1, ..., d
0
for a terminal time T > 0. Define the measure Q on F by
Q(A) = E(1A γT ), A ∈ F,
where
d Z d
!
X T Z T X
γT := exp − bit dWti − 1
2
|bit |2 dt .
i=1 0 0 i

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

14

Then under Q the process (Xt )t∈[0,T ] is a d-dimensional Wiener martingale with respect to F.
Remark 7.1. A well-known sufficient condition for EγT = 1 to hold is the condition
1
RT
|bs |2 ds
Ee 2 0 < ∞,
Pd
where |bs |2 = i=1 |bis |2 . This condition is called Novikov’s condition.
Example 7.1. Let (Wt )t∈[0,T ] be a Wiener martingale w.r. to (Ft ) − t ∈ [0, T ]. Consider
Xt = λt + Wt , t ∈ [0, T ], and define the measure Q(A) = E(1A γT ), A ∈ F, where
γT = exp(−λWT − 21 λ2 T ).
Then Q is a probability measure and under Q the process (Xt )t∈[0,T ] is a Wiener martingale
w.r. to (Ft ) − t ∈ [0, T ].
Proposition 7.3. Let b ∈ S([0, T ]). Then for
Z t Z t
1
γt := exp( bs dWs − 2 b2s ds)
0 0
we have Eγt ≤ 1 for every t ∈ [0, T ].
Proof. By Itô’s formula we get dγt = bt γt dWt . Hence (γt )t∈[0,T ] is a local martingale, i.e., for a
suitable sequence of stopping times, (τn )∞
n=1 , converging to infinity, the stopped process (γt∧τn )
is a martingale. Consequently,
1 = Eγ0 = Eγt∧τn .
Hence by Fatou’s lemma we have
Eγt = E( lim γt∧τn ) ≤ lim Eγt∧τn = 1
n→∞ n→∞

Exercise 7.3. Let (bt )t∈[0,T ] be an Ft -adapted process such that |b| ≤ K for all t ∈ [0, T ] and
ω ∈ Ω, where K is a constant. Show that
Z t Z t
E exp( bs dWs − 2 1
b2s ds) = 1 for every t ∈ [0, T ].
0 0

Definition 7.1. If Q and P are measures on F such that for a random variable γ ≥ 0
Q(A) = E(1A γ), for all A ∈ F.
Then we say that Q is absolutely continuous with respect to P , and its Radon-Nikodym deriv-
ative with respect to P is γ. In this case the notation dQ = γ dP is used.
Proposition 7.4. Let P and Q be probability measures on F such that dQ = γdP with some
random variable γ. Then for every non-negative random variable Y we have
EQ Y = EP (Y γ), (7.27)
where EQ and EP denote expectation with respect to the measures Q and P , respectively.
Proposition 7.5. Let Q and P be probability measures on F such that dQ = γ dP for some
random variable γ. Assume that γ > 0. Then P is absolutely continuous with respect to Q,
and dP = γ1 dP .
Definition 7.2. If P and Q are measures on F such that Q is absolutely continuous with
respect to P , and P is absolutely continuous with respect to Q, then we say that they are
equivalent measures.
Remark 7.2. Notice that in Girsanov’s theorem P and Q are equivalent measures.

Downloaded by jaslskd Afas ([email protected])

 lOMoARcPSD|3446508

15

8. Martingale representation theorems

Let W be a Wiener process, fix T > 0 and set GT = σ(Wr : r ≤ T ). Then we know that for
every f ∈ H the integral Z T
fs dWs (8.28)
0
is a GT measurable random variable with mean 0 and finite second moment.
The reverse of this also holds.
Theorem 8.1. Let X be a GT -measurable random variable such that EX 2 < ∞. Then there
exists f ∈ H such that Z T
X = EX + fs dWs . (8.29)
0
This representation of X is unique, i.e., if it holds with f ′ ∈ H then almost surely f ′ = f for
almost every t ∈ [0, T ].
One can easily show that this theorem is equivalent to the following theorem on representa-
tion of Gt -martingales.
Theorem 8.2. Let M = (Mt )t∈[0,T ] be a martingale with respect to (Gt )t∈[0,T ] , such that EMT2 <
∞. Then there exists f ∈ H([0, T ]) such that
Z t
M t = M0 + fs dWs (a.s.) for every t ∈ [0, T ]. (8.30)
0
Moreover, this representation is unique, i.e., if the representation (8.30) holds with f ′ ∈
H([0, T ]), then almost surely f ′ = f for almost every t ∈ [0, T ].
Remark 8.1. The above theorems remains valid if we assume E|X| < ∞ and E|MT | < ∞,
instead of EX 2 < ∞ and EMT2 < ∞, respectively. Then (8.29) and respectively (8.30) hold
with f ∈ S in place of f ∈ H. Moreover, these theorems can be generalized to the case of
multidimensional Wiener processes W .
Proposition 8.3. Theorems 8.1 and 8.2 are equivalent.

The representation of a σ(Wr : r ≤ T ) as a stochastic integral with respect to a Wiener

process W is closely related to the solvability of the heat equation
ut (t, x) + 21 uxx (t, x) = 0, (t, x) ∈ [0, T ), x ∈ R, (8.31)
with terminal condition u(T, x) = ϕ(x), x ∈ R.
Exercise 8.1. Consider the random variable X = cos(WT ). Find f ∈ H([0, T ]) such that
Z T
X = EX + fs dWs .
0

Downloaded by jaslskd Afas ([email protected])