Lecture 8

Stochastic Integration and Differentiation

随机积分和微分
Quadratic variation
Itô integrals
Wiener integrals
Itô multiplication table
Itô’s formula
Computing Itô integrals via Itô’s formula

181
8.1 Introduction
! b
Recall that the Riemann integral f (t)dt is defined as follows.
a
1. Partition the interval [a, b] into n subintervals
a = t0 < t1 < t2 < · · · < tn−1 < tn = b.
2. Choose a point τi in the subinterval [ti, ti+1] for i = 0, 1, . . . , n−1.
3. Form the Riemann sum
n−1
"
Sn = f (τi)(ti+1 − ti)
i=0
4. Define the integral as the limit
! b
f (t)dt := lim Sn
a n→∞

where the subintervals are chosen so that maxi{ti+1 − ti} → 0.

182
In this lecture, we introduce the following new kinds of integrals
(called Itô integrals)
! b ! b ! b
f (t) dBt f (t, Bt) dBt f (t, Bt) (dBt)2
a a a
where Bt (i.e. B(t)) is the standard Brownian motion process.
If B(t) were differentiable, then
! b ! b
f (t) dBt = f (t)B $(t)dt,
a a
which is a classical integral. It is because B(t) is non-differentiable
that new calculus rules are sought.

183
We will then introduce a chain rule for stochastic processes (called
Itô’s formula) which can be applied to evaluate Itô integrals.
In the next lecture, we will introduce stochastic differential equa-
tions (SDE) whose solutions may be obtained via Itô integrals.
Applications of SDE to the modeling of stock price, interest rate,
and bond price will be presented afterwards.

184
8.2 Non-anticipating Processes
Consider the standard Brownian motion Bt (i.e. B(t)).
A process Ft is called a non-anticipating process if Ft is independent
of any future increment of Bt.
E.g. Ft = Bt, eBt , Bt2 + t are non-anticipating.
E.g. Ft = Bt+1, B2t, (Bt+1 − Bt)2 are not.
Itô integrals apply only to non-anticipating processes.

185
8.3 Increments of Brownian Motions
Proposition 8.1 Let Bt be the standard Brownian motion. If s <
t, then we have Variance2
≥

1. E[(Bt − Bs)2] = t − s; Be
) BsrNio
-
t s ,
-

2. E[(Bt − Bs)4] = 3(t − s)2; σ )

# $ Z ~ N (0 ,

2
3. Var (Bt − Bs) = 2(t − s)2. 2

E ) 3
4
σ
( Z
=

Proof:
! ∞
1 x2
− 2(t−s)
2 2 2
E[(Bt − Bs) ] = E[Bt−s] = % x ·e dx = t − s
2π(t − s) −∞
! ∞
1 x2
− 2(t−s)
4 4 4
E[(Bt − Bs) ] = E[Bt−s] = % x ·e dx = 3(t − s)2
2π(t − s) −∞
# $ # $
Var (Bt − Bs)2 = Var Bt−s
2 4
= E[Bt−s 2
] − (E[Bt−s ])2 = 2(t − s)2.
=

3
~ 2
(t s) 2 ( t s)
( t s)
-

=
-
-
-

186
Mean Square Convergence
A random variable is a function of the outcome ω ∈ Ω. Thus, a
→
sequence of random variables S1, S2, . . . is a sequence of functions.
Convergence of a sequence of functions can be studied in various
modes, e.g. pointwise, uniform, and mean square.
For sequence of random variables, we use mean square convergence,
weighted by a probability density.
Definition 8.2 A sequence of random variables S1, S2, . . . is said to
converge to a random variable S in mean square sense if
& 2
'
lim E (Sn − S) = 0.
n→∞
The limit S is denoted by
S = ms-lim Sn.
n→∞

187
Quadratic Variation
Proposition 8.3 Let T > 0 and consider the equidistant partition
ti = iTn , for i = 0, 1, . . . , n, of the interval [0, T ]. Then
n−1
"
ms-lim (Bti+1 − Bti )2 = T. (8.1)
n→∞
i=0

The expression on the left is called the quadratic variation of Bt.

188
Proof: Consider the random variable
n−1
"
Sn = (Bti+1 − Bti )2.
i=0
By Proposition 8.1,
n−1
" n−1
"
& 2
'
E[Sn] = E (Bti+1 − Bti ) = (ti+1 − ti) = T.
i=0 i=0

Since the increments of {Bt, t ≥ 0} are independent, we have

n−1
" n−1
"
& 2
'
Var(Sn) = Var (Bti+1 − Bti ) = 2(ti+1 − ti)2
i=0 i=0
Å ã2
T 2T 2
= n·2 = .
n n
The second equality above is due to Proposition 8.1 and the fact that
ti+1 − ti = Tn for all i.
189
 Next, note that
2 2 2T 2
E[(Sn − T ) ] = E[(Sn − E[Sn]) ] = Var(Sn) = →0
n
as n → ∞. Thus,
n−1
"
ms-lim (Bti+1 − Bti )2 = T.
n→∞
i=0

190
8.4 The Itô Integral
Let Ft = f (t, Bt) be a non-anticipating process.
Without loss of generality, we assume the subintervals equidistant,
i.e. ti+1 − ti = b−a
n .

Consider the partial sum

n−1
"
Sn = Fti (Bti+1 − Bti ).
i=0
Here, the intermediate points τi are chosen to be the left endpoint ti.
The choice τi = ti makes Fti and Bti+1 − Bti independent.
The Itô integral is the mean square limit of the partial sums Sn
! b
Ft dBt := ms-lim Sn.
a n→∞

191
8.5 Examples of Itô Integrals
We provide two examples of computing Itô integrals from first prin-
ciples.
However, just like Riemann integrals, we seldom compute them from
scratch in practice.
Later, we shall deduce some rules that allow us to compute Itô inte-
grals more efficiently.

192
 (T
The case Ft = 1 (i.e. 0 dBt)
Example 8.4 Let Ft = 1 and [a, b] = [0, T ]. Then,
n−1
" n−1
"
Sn = Fti (Bti+1 − Bti ) = (Bti+1 − Bti ) = BT − B0 = BT .
i=0 i=0

It follows that ms-limn→∞ Sn = BT . Thus, we have the result

! T
dBt = BT .
0

193
 (T
The case Ft = Bt (i.e. 0 Bt dBt)
iT
Example 8.5 Let Ft = Bt and ti = n for i = 0, 1, . . . , n. Then,
n−1
" n−1
"
Sn = Fti (Bti+1 − Bti ) = Bti (Bti+1 − Bti ).
i=0 i=0

Since
1& 2 2 2
'
xy = (x + y) − x − y ,
2
letting x = Bti and y = Bti+1 − Bti yields
1 2 1 2 1
Bti (Bti+1 − Bti ) = Bti+1 − Bti − (Bti+1 − Bti )2.
2 2 2
By summing over i, we have
n−1 n−1 n−1
1" 2 1" 2 1"
Sn = Bti+1 − B ti − (Bti+1 − Bti )2.
2 i=0 2 i=0 2 i=0

194
Then, by telescoping, the sum becomes
n−1
1 2 1"
Sn = B T − (Bti+1 − Bti )2.
2 2 i=0
By Eq. (8.1), we obtain
" n−1
1 2 1 2 1 2 1
ms-lim Sn = BT − · ms-lim (Bti+1 − Bti ) = BT − T.
n→∞ 2 2 n→∞
i=0
2 2
Thus, we have the result
! T
BT2 T
Bt dBt = − .
0 2 2

(T f (T )2 f (0)2
Remark: For Riemann integrals, we have 0 f (x) df (x) = 2 − 2 ,
provided that f differentiable. The additional term − T2 appeared in
the Itô integral is an instance of Itô correction. It is due to the non-
differentiability of B(t).
195
8.6 Properties of the Itô Integral
Proposition 8.6 Let Ft = f (t, Bt) and Gt = g(t, Bt) be two non-
anticipating processes and c ∈ R. Then we have
1. Additivity:
! b ! b ! b
(Ft + Gt) dBt = Ft dBt + Gt dBt.
a a a

2. Homogeneity: ! !
b b
cFt dBt = c Ft dBt.
a a
3. Partition property:
! b ! c ! b
Ft dBt = Ft dBt + Ft dBt
a a c
for any a < c < b.

196
 As a random variable, Itô integrals have the following properties.
Proposition 8.7 Let Ft = f (t, Bt) and Gt = g(t, Bt) be two non-
anticipating processes. We have
1. Zero mean: ï! b ò
E Ft dBt = 0.
a

2. Isometry (variance):
ñÅ! ã2ô ï! b ò
b
E Ft dBt =E Ft2 dt .
a a

3. Covariance:
ï! b ! b ò ï! b ò
E Ft dBt · Gt dBt = E FtGt dt .
a a a

197
 By considering the upper integration limit as the time variable t, the
(t
integral It = 0 f (u, Bu) dBu is a continuous-time process.
Denote by Fs = {Bu, 0 ≤ u ≤ s} the historical values of the
Brownian motion available at time s.
(t
Proposition 8.8 The Itô integral It = 0 Fu dBu, where Fu =
f (u, Bu), is a martingale. That is, for any s < t, we have
E[It|Fs] = Is.
Proof: ï! ) ò ï! t ) ò
s ) )
E[It|Fs] = E )
Fu dBu) Fs + E Fu dBu)) Fs
! s 0 s

= Fu dBu + 0 = Is.
0
(t
Here, E[ s Fu dBu|Fs] = 0 because the increments {Bti+1 − Bti } are
independent of Fs and have mean 0. The details are skipped.
198
8.7 The Wiener Integral
The Wiener integral is a special case of the Itô integral in which
f (t, Bt) = f (t), i.e. a deterministic function.
( ( t
E.g. t dBt and e dBt.
Being an Itô integral, a Wiener integral inherits all properties of the
Itô integral.
But it has an additional property that it is normally distributed.

199
 (b
Proposition 8.9 The Wiener integral I(f ) = a f (t) dBt is nor-
0
mally distributed with mean 0 and variance
! b
Var(I(f )) = f (t)2dt.
a
Proof: The mean and variance are basic properties of Itô integrals,
followed directly from Proposition 8.7. To see that I(f ) is normally
distributed, note that
n−1
"
Sn = f (ti)(Bti+1 − Bti ).
i=0
Since the increments Bti+1 − Bti ∼ N (0, ti+1 − ti), it follows that
Ç n−1 å
"
Sn ∼ N 0, f (ti)2(ti+1 − ti) .
i=0
We claim without giving details that the mean square limit is also nor-
mally distributed (under some suitable assumptions on f ).
200
Example 8.10 The following are examples of Wiener integrals.
(T
1. The random variable I = √1 dBt is normally distributed with
1 t
mean 0 and variance ln T .
(T √
2. The random variable I = 1 t dBt is normally distributed with
mean 0 and variance (T 2 − 1)/2.
(T
3. The random variable I = 0 eT −t dBt is normally distributed with
mean 0 and variance (e2T − 1)/2.

201
8.8 Fundamental Relations of Differentials
By Eq. (8.1),
n−1
"
ms-lim (Bti+1 − Bti )2 = T
n→∞
i=0
(T
The right hand side can be regarded as a Riemann integral 0 dt.
(T
The left hand side can be regarded as a stochastic integral 0 (dBt)2.
Therefore, we have the integral equation
! T ! T
(dBt)2 = dt.
0 0

The above equation written in differential form is (dBt)2 = dt. The

differential form is a shorthand for the integral equation.

202
Likewise, we can show that
n−1
" n−1
"
ms-lim (Bti+1 − Bti )(ti+1 − ti) = ms-lim (ti+1 − ti)2 = 0.
n→∞ n→∞
i=0 i=0

This yields the integral equations

! T ! T
dBt dt = 0 and (dt)2 = 0,
0 0

whose differential forms are dBt dt = 0 and (dt)2 = 0.

The following are fundamental relations in stochastic calculus:
1. (dBt)2 = dt;
2. dBt dt = 0;
3. (dt)2 = 0.
They are also known as the Itô multiplication table.

203
8.9 Itô’s Formula
Itô Processes
A process Xt is called an Itô diffusion (a.k.a. Itô process) if
! t ! t
Xt = X0 + a(s, Bs) ds + b(s, Bs) dBs
0 0
for some functions a(t, x) and b(t, x).
This equation can also be expressed in differential form
dXt = a(t, Bt) dt + b(t, Bt) dBt.
This form is again a shorthand for the integral form above.
Examples of Itô processes:
dXt = µ dt + σ dBt (Brownian motion with drift)
dSt = µSt dt + σSt dBt (GBM model of stock prices)

204
Differential of Ft = f (t, Xt)
Consider a process Ft = f (t, Xt) where Xt is an Itô process. We
would like to devise the chain rule for evaluating dFt.
By Taylor expansion of f (t, x)
∂f (t, x) ∂f (t, x) 1 ∂ 2f (t, x) 2
df (t, x) = dt + dx + · 2
(dx) + ··· .
∂t ∂x 2 ∂x
Substituting x = Xt (an Itô diffusion) and yields
∂f (t, Xt) ∂f (t, Xt) 1 ∂ 2f (t, Xt) 2
df (t, Xt) = dt + dXt + · 2
(dX t ) .
∂t ∂x 2 ∂x
All higher terms, e.g. (dt)2, dt dXt, and (dXt)3, are 0 because of
the Itô multiplication table.

205
 By using dXt = a(t, Bt)dt + b(t, Bt)dBt and the Itô multiplication
table, we have
(dXt)2 = a2(dt)2 + 2ab dt dBt + b2(dBt)2 = b2dt.

Thus, we arrive at the Itô’s formula

ï 2 2
ò
∂f (t, Xt) ∂f (t, Xt) b(t, Bt) ∂ f (t, Xt)
dFt = + a(t, Bt) + · 2
dt
∂t ∂x 2 ∂x
∂f (t, Xt)
+ b(t, Bt) dBt. (8.2)
∂x
Remark: Note that the traditional chain rule for smooth functions is
∂f (t, Xt) ∂f (t, Xt)
df (t, Xt) = dt + dXt.
∂t ∂x
Thus, the Itô formula has an additional term, called Itô correction,
b(t, Bt)2 ∂ 2f (t, Xt)
· dt.
2 ∂x2
206
Example 8.11 Determine d(tBt2).
Solution: Let Ft = f (t, Xt) = tBt2. Therefore, we must have a = 0,
b = 1, and f (t, x) = tx2. Then,
2
∂f 2 ∂f ∂ f
=x , = 2tx, = 2t.
∂t ∂x ∂x2
By Itô’s formula (8.2) (with a = 0, b = 1),
* 1 +
d(tBt2) = Bt2 + · 2t dt + 2tBt dBt
2
= (Bt2 + t)dt + 2tBt dBt.
i 毙叫
dea
df (e ,
xt 1 =
( 花⼗和⼗
α

Bed
⼗ b
数
207
Example 8.12 If Xt is a process such that dXt = µdt + σdBt where
µ and σ are constants, determine d(e−tXt).
Solution: Let Ft = f (t, Xt) = e−tXt. Therefore, we have f (t, x) =
e−tx. Then,
∂f −t ∂f ∂ 2f
= −e x, = e−t, = 0.
∂t ∂x ∂x2
By Itô’s formula (8.2) (with a = µ, b = σ)
−t
# −t −t
$
d(e Xt) = −e Xt + µe dt + σe−t dBt
= e−t(µ − Xt) dt + σe−t dBt.

208
Example 8.13 Consider the stock price St generated by the GBM
model dSt = µSt dt + σSt dBt where µ and σ are constants, determine
d(ln St).
Solution: Let Ft = f (t, St) = ln St. Therefore, we have f (t, x) = ln x.
Then,
∂f ∂f 1 ∂ 2f 1
= 0, = , = − 2.
∂t ∂x x ∂x2 x
By Itô’s formula (8.2) (with a = µSt and b = σSt)
Å ã
1 1 1 1
d(ln St) = 0 + µSt · − · (σSt)2 · 2 dt + σSt · dBt
St 2 St St
Å 2
ã
σ
= µ− dt + σ dBt.
2

This example shows that the log price is a Brownian motion with a drift.

209
8.10 Stochastic Integration Techniques: Itô’s Formula
Let Xt = Bt, i.e. an Itô process with a = 0, b = 1. By Itô’s formula,
ï 2
ò
∂f (t, Bt) 1 ∂ f (t, Bt) ∂f (t, Bt)
df (t, Bt) = + · 2
dt + dBt.
∂t 2 ∂x ∂x
Integrating both sides, we have
! t
∂f (s,Bs ) ( t î ∂f (s,Bs) 1 ∂ 2f (s,Bs) ó
∂x dBs = f (t, Bt) − f (0, 0) − 0 ∂t + 2 · ∂x2 ds
0
(t
Suppose that we’d like to determine an Itô integral 0 g(s, Bs) dBs.
(t
We can thus determine 0 g(s, Bs) dBs by finding an f (t, x) such
that ∂f∂x
(t,x)
= g(t, x) and applying the above formula.

210
 (t
Example 8.14 Determine 0 Bs dBs.
Solution: Let ∂f∂x (t, Xt ) = Bt . Therefore, we have a = 0, b = 1, and ⼀

x2 Bt2
f (t, x) = 2 . Let Ft = f (t, Xt) = 2 . By Itô’s formula (with a = 0,
b = 1)
Å ã
Bt2 1
d = dt + Bt dBt.
2 2
Hence,
! t
Bt2 B02 t
= + + Bs dBs,
2 2 2 0
so that ! t
Bt2 t
Bs dBs = − .
0 2 2

211
 (t
Example 8.15 Determine 0 sBs dBs.
Solution: Let ∂f∂x (t, Xt ) = tBt . Therefore, we have a = 0, b = 1, and →
tx2 tBt2
f (t, x) = 2 . Let Ft = f (t, Xt) = 2 . By Itô’s formula (with a = 0,
b = 1)
Å ã Å ã
tBt2 Bt2 t
d = + dt + tBt dBt.
2 2 2
Hence,
! tÅ ã ! t
tBt2 0· B02 Bs2 s
= + + ds + sBs dBs,
2 2 0 2 2 0
so that ! t ! t
tBt2 1 t2
sBs dBs = − Bs2 ds − .
0 2 2 0 4

(t 2
Remark: B
0 s ds is a Riemann integral; it cannot be further simplified.
212