Chapter 7

The Ito’s Formula

We will derive Ito’s lemma, i.e., the chain rule in the Ito’s integrals. It is one of the most
fundamental results in Stochastic Calculus, and is a very useful rule in calculating Ito’s integrals.

Despite many similarities, calculus and stochastic calculus have fundamental differences.

• In calculus, we start with differentiation, and then move onto integration. The two
concepts are linked by the fundamental theorem of calculus, from which we derive chain
rule, integration by parts.

• In stochastic calculus, however, we have no differentiation theory, only integration theory.

We have Ito formula (i.e., fundamental theorem of stochastic calculus), from which we
derive corresponding chain rule, integration by parts.

1
7.1 Review on Riemann integrals

We don’t use the definition to calculate Riemann integrals. Instead, we use the fundamental
theorem of calculus, which relates the differentiation and integration, the two central concepts
in calculus.

Theorem 7.1 (Fundamental Theorem of Calculus) f is continuous on [a, b].

Rx
1. If F (x) = a f (t)dt, then F 0 (x) = f (x).
Rb
2. If f (x) = F 0 (x), then a f (x)dx = F (b) − F (a).

Z t
t2 x2
Example 7.1 xdx = F (t) − F (0) = , where f (x) = x and F (x) = .
0 2 2

2
 The chain rule and integration by parts formula follow easily from the Fundamental The-
orem of Calculus.

Theorem 7.2 (Chain rule or substitution rule)

dy dy du
1. [f (g(x))]0 = f 0 (g(x))g 0 (x). Or = , where y = f (u) and u = g(x).
dx du dx
2. Integrating the above and applying Theorem 7.1, we get
Z t Z t Z g(t) !
0 0 0 0
f (g(t)) − f (g(0)) = f (g(s))g (s)ds = f (g(s))dg(s) = f (u)du.
0 0 g(0)

Example
Rt
7.2 Take y = u2 and u = g(x). We have [g 2 (x)]0 = 2g(x)g 0 (x), and g 2 (x) − g 2 (0) =
2 0 g(s)dg(s). However, this chain rule fails for Bt as
Z t
2 Bs dBs = Bt2 − t 6= Bt2 .
0

(In fact, the LHS is a martingale, but the RHS is not (it is a submartingale.)

3
Theorem 7.3 (Integration by parts formula)
Z b Z b
0
f (x)g (x) dx = f (x)g(x)|ba − f 0 (x)g(x) dx
a a

Proof. By [uv]0 = u0 v + uv 0 and the Fundamental Theorem of Calculus, we have

Z b Z b Z b
f (b)g(b) − f (a)g(a) = [f (x)g(x)]0 dx = f 0 (x)g(x) dx + f (x)g 0 (x) dx.
a a a

Rx Rx
xd sin(x) = x sin(x) − sin(x)dx = x sin(x) + cos(x).
R
Example 7.3 0 x cos(x)dx = 0

Example 7.4 Integration by parts formula fails for RBrownian motion. If not, we
would have 0 Bs dBs = Bt − B0 − 0 Bs dBs or equivalently 0t Bs dBs = 12 Bt2 , which is clearly
Rt 2 2
Rt

false.

4
7.2 Variation and co-variation processes

Definition 7.1 Let Xt and Yt be two continuous stochastic processes. Let 0 = t0 < t1 <
... < tn = t, ∆tk = tk − tk−1 , and ∆k X = Xtk − Xtk−1 .

• The covariation process of X and Y is: hX, Y i0 = 0

X
hX, Y it = lim (∆k X)(∆k Y ) ( limit in probability).
∆tk →0
tk ≤t

• The quadratic variation process of Xt is: hXi0 = 0 and

(∆k X)2
X
hXit =: hX, Xit = lim ( limit in probability).
∆tk →0
tk ≤t

Remark 7.1 It is easy to see that

1 1
hX, Y it := (hX + Y it − hX − Y it ) = (hX + Y it − hXit − hY it ) ,
4 2
which is similar to relationship between variance and covariance:
1 1
cov(X, Y ) = (var(X + Y ) − var(X − Y )) = (var(X + Y ) − var(X) − var(Y )) .
4 2

5
Theorem 7.4 X, Y , and Z are continuous stochastic processes. Then,

1. hXit ↑.

2. hX, Y it = hY, Xit

3. hX + Y, Zit = hX, Zit + hY, Zit

4. hX − X0 , Zit = hX, Zit

5. haX, bY it = abhX, Y it , where a, b ∈ R

6. haXit = a2 hXit , where a ∈ R

Proof. Left as an exercise.

6
Theorem 7.5 At and Xt are both continuous, and At is of bounded variation. Then

(a) hAit = 0 a.s.

(b) hA, Xit = 0 a.s.
(c) hA + Xit = hXit a.s.

Proof.

• (a) follows from (c) by taking X = 0.

• (b) also follows from (c): hA, Xit = 14 (hX + Ait − hX − Ait ) = 41 (hXit − hXit ) = 0.

• We prove (c) below. Denote ∆k Y = Ytk − Ytk−1 , where either Y = A or Y = X. Then,

the key step is:
  ! 
 
X  X

 (∆k A)(∆k Y ) ≤ sup |∆k Y |  |∆k A| −→ 0, (2.1)
tk ≤t  k tk ≤t

as the continuity of Y implies supk |∆k Y | → 0 while the bounded variation of A implies
tk ≤t |∆k A| < ∞. Therefore,
P

|∆k (X + A)|2 = |∆k X + ∆k A|2

X X

tk ≤t tk ≤t

(∆k X)2 + (∆k A)2 + 2

X X X
= (∆k X)(∆k A)
tk ≤t tk ≤t tk ≤t
→ hXit + 0 + 0, in probability. (from (2.1))

On the other hand, the LHS → hX + Ait in probability by definition. The result then
follows from the uniqueness of the probability limit.

7
 Part (a) of Theorem 7.5 can be further strengthened.

Theorem 7.6 If At is a continuous function of bounded p-variation with 1 ≤ p < 2, then

we have hAit = 0.

Proof. We have

|∆k A|2 = |∆k A|p |∆k A|2−p

X X

tk ≤t tk ≤t
 

≤ (sup |∆k A|)2−p  |∆k A|p 

X

tk ≤t
2−p
≤ C (sup |∆k A|)
→ 0.

8
Variation processes of Ito’s integrals

Theorem 7.7 hBit = t.

The proof has been done before. This is used in Theorem 7.15 to characterize Brownian motion:
If Xt is a continuous martingale with hXit = t, Xt must be a Brownian motion.

Rt Rt
Theorem 7.8 Let Xt = 0 Cs dBs and Yt = 0 Ds dBs . Then
Z t Z t
hXit = Cs2 dhBis = Cs2 ds
0 0
Z t Z t
hX, Y it = Cs Ds dhBis = Cs Ds ds
0 0

Proof. Too long, omitted here.

9
7.3 Ito’s integration w.r.t. martingales and semi-martingales

In defining 0t Cs dBs , we have not used properties of Bt much except the Isometry property.
R

For more general continuous martingales Xt , we could Rt

establish the corresponding Isometry
property. Then the Ito integralof C w.r.t. X is a r.v. 0 Cs dXs : Ω → R, defined to be the L2
limit of ni=1 Cti−1 Xti − Xti−1 as the mesh of the partition 0 = t0 < t1 < ... < tn = t of [0, t]
P

tends to 0.

Definition 7.2 Xt is said to be a continuous semi-martingale if

Xt = Mt + At

where Mt and At are continuous martingale and adapted process, respectively. The Ito integral
of C w.r.t. X is defined to be
Z t Z t Z t
Cs dXs =: Cs dMs + Cs dAs ,
0 0 0

where the two integrals on the RHS are Ito’s and R-S integrals, respectively.

By Doob-Meyer decomposition, sub- and super-martingales are both semimartingales.

But semimartingales are much more general processes.

10
7.4 One-dimensional Ito’s formula

Theorem 7.9 If Xt is a continuous semi-martingale, and f 00 is continuous, then

Z t
1 Z t 00
f (Xt ) − f (X0 ) = f 0 (Xs )dXs + f (Xs )dhXis .
0 2 0
Formally, we could write this in “differential form”:
1
df (Xt ) = f 0 (Xt )dXt + f 00 (Xt )dhXit .
2

Remark: Intuitively, Ito’s formula is just a local two-term Taylor expansion in that
1
f (Xt+∆t ) − f (Xt ) ≈ f 0 (Xt ) (Xt+∆t − Xt ) + f 00 (Xt ) (Xt+∆t − Xt )2 .
2

11
 Proof. We will consider the case where Xt is a continuous martingale, from which
the case for continuous semi-martingale follows easily. First we have
1
f (b) − f (a) = (b − a)f 0 (a) + f 00 (c(a, b)) (b − a)2 .
2
where c = c(a, b) in between a and b. Let 0 = t0 < t1 < ... < tn = t. Therefore,
n
X
f (Xt ) − f (X0 ) = [f (Xti ) − f (Xti−1 )]
i=1
n n
  1X   2
f 0 (Xti−1 ) Xti − Xti−1 + f 00 c(Xti−1 , Xti ) Xti − Xti−1
X
=
i=1 2 i=1
1
=: In1 + In2 .
2

• ClearlyR In1 →p 0t f 0 (Xs )dXs as n → ∞. By taking subsequences we can suppose that

R

In1 → 0t f 0 (Xs )dXs , a.s.

12
 Z t
• For In2 , we can rewrite it as In2 = Gns dAns , where
0
 
Gns = f 00 c(Xti−1 , Xti ) , when s ∈ (ti−1 , ti ]
00
= f (Xt ) , when s ≥ t

and
X 2
Ans = Xti − Xti−1 .
tk ≤s

Now Ans −→p hXis . By taking subsequences we can suppose that, as n → ∞,

Ans −→ hXis , a.s. (4.2)

Next let us look at the integrand Gns . Observe that the uniform continuity of f 00 implies
Gns → f 00 (Xs∧t ) uniformly in s. Since Xs is continuous, it is bounded on [0, t], i.e.
|Xs | ≤ C for s ∈ [0, t]. Therefore, |c(Xti−1 , Xti )| ≤ C. Since f 00 is assumed continuous,
thus |f 00 (r)| ≤ C0 for |r| ∈ [0, C], implying that

|Gns | ≤ C0 . (4.3)

In view of (4.2) and (4.3), for almost all ω, we can apply the bounded convergence theorem
to obtain
Z t Z t
In2 = Gns dAns −→ f 00 (Xs )dhXis .
0 0

13
Remark 7.2

1. Martingales (semi-martingales)
Rt
are closed under the Ito’s integration. (If Xt is a martin-
gale (semi-martingale), so is 0 Ct dXt .)

2. By Ito’s lemma, martingales are NOT closed under smooth transformation, but semi-
martingales are.
Rt 0
3. If Xt = At (B.V.), then hAit = 0, we have f (At ) − f (A0 ) = 0f (A s )dAs . In fact, this
holds if f 0 is continuous.
 
4. Since ∆i B ∼ N (0, ∆i t), we have ∆i B = Op (∆i t)1/2 or (∆i B)2 = Op (∆i t) . The fol-
lowing heuristics are often helpful in formal derivations:

dBt = (dt)1/2 , dBt · dBt = dt,

dt · dBt = (dt)3/2 = o(dt) = 0, dt · dt = (dt)2 = o(dt) = 0.

14
Theorem 7.10 At is a continuous function of bounded variation, and f 0 is continuous, then
Z t
f (At ) − f (A0 ) = f 0 (As )dAs .
0

Proof.
n
X
f (At ) − f (A0 ) = [f (Ati ) − f (Ati−1 )]
i=1
n   
f 0 c(Ati−1 , Ati )
X
= Ati − Ati−1
i=1
n  
f 0 (Ayi ) Ati − Ati−1
X
=
i=1
(∃yi ∈ [ti−1 , ti ] s.t. Ayi is in between Ati−1 and Ati as At is continuous.)
Z t
−→ f 0 (As )dAs .
0

15
 Rt
Example 7.5 Find 0 Bsm dBs where m ≥ 1.

Solution. Take Xs = Bs and f 0 (x) = xm . Then f (x) = xm+1 /(m + 1), f 00 (x) = mxm−1 , and
hXis = hBis = s. Applying the Ito’s formula, we get

Btm+1 B m+1 Z t
1Z t
− 0 = Bsm dBs + mBsm−1 ds
m+1 m+1 0 2 0
Hence,
Z t
Btm+1 m Z t m−1
Bsm dBs = − B ds. (4.4)
0 m+1 2 0 s
Rt
In particular, if m = 1, we have 0 Bs dBs = (Bt2 − t)/2.

B m+1
Alternative solution: We guess that 0t Bsm dBs should be roughly like m+1
R t
, which we can
m+1
apply Ito’s lemma to find out the exact relationship. (Let f (x) = x /(m + 1) and take
x = Bt .)

16
7.5 Multi-dimensional Ito’s formula

Theorem 7.11 Let Xt = (Xt1 , ..., Xtd ) and f : Rd → R. Then,

d Z t
1 X Zt
fi (Xs )dXsi fij (Xs )dhX i , X j is
X
f (Xt ) − f (X0 ) = +
i=1 0 2 1≤i,j≤d 0

In differential form, this is

1X
fi dXti + fij dhX i , X j it .
X
df (Xt ) =
i 2 i<j

Proof. The proof is similar to that in one-dimensional case, and hence omitted.

17
Corollary 7.1 In Theorem 7.11, if Xt1 , ..., Xtc are continuous semi-martingales, and Xtc+1 , ..., Xtd
are of bounded variation, then we have
d Z t
1 X Zt
fi (Xs )dXsi fij (Xs )dhX i , X j is .
X
f (Xt ) − f (X0 ) = +
i=1 0 2 1≤i,j≤c 0

Corollary 7.2 If Xt is a continuous semi-martingale, then

Z t Z t
1Z t
f (t, Xt ) − f (0, X0 ) = f1 (s, Xs )ds + f2 (s, Xs )dXs + f22 (s, Xs )dhXis .
0 0 2 0
or
1
df (t, Xt ) = f1 (t, Xt )dt + f2 (t, Xt )dXt + f22 (t, Xt )dhXit .
2

18
 Rt
Example 7.6 Find 0 sdBs .

Solution 1. Take Xs = Bs and f2 (x, y) = x. Then f (x, y) = xy, f1 (x, y) = y and f22 (x, y) = 0,
and hXis = hBis = s. Applying the Ito’s formula, we get
Z t Z t
tBt − 0 = Bs ds + sdBs + 0.
0 0
Rt Rt
Hence, 0 sdBs = tBt − 0 Bs ds.

Solution 2. Applying Ito’s formula to f (t, Bt ) = tBt , where f (x, y) = xy, we have
1
d(tBt ) = df (t, Bt ) = f1 dt + f2 dBt + f22 dhBit = Bt dt + tdBt + 0.
2

19
7.6 Integration by parts formula

Theorem 7.12 Xt and Yt are continuous semi-martingales, then,

Z t Z t
Xt Yt − X0 Y0 = Ys dXs + Xs dYs + hX, Y it .
0 0

Proof. Let Xt = (Xt , Yt ), f (x, y) = xy. Hence

f1 = y, f2 = x, f11 = f22 = 0, f12 = 1.

Then applying the multivariate Ito’s formula, we get

d(Xt Yt ) = df (Xt , Yt )
= f1 (Xt , Yt )dXt + f2 (Xt , Yt )dYt
1 1
+ f11 (Xt , Yt )dhXit + f12 (Xt , Yt )dhX, Y it + f22 (Xt , Yt )dhY it
2 2
= Yt dXt + Xt dYt + dhX, Y it .

20
 Rt
Example 7.7 Find 0 g(s)dBs , where g is differentiable.

Solution. By integration by parts formula,

Z t Z t Z t
0
g(s)dBs = g(t)Bt − Bs g (s)ds − hg, Bit = g(t)Bt − Bs g 0 (s)ds.
0 0 0

In particular, when g(x) = x, this has been done in Example 7.6 by Ito’s formula. But it is
much simpler to use integration by parts formula here.

21
7.7 Ito exponential
Rt B
One can easily check by Ito’s formula that eBt − 1 6= 0 e s dB s (LHS is a submartingale while
RHS is a martingale). However, we have

Z t
Theorem 7.13 We have eBt −t/2 − 1 = eBs −s/2 dBs .
0

For a general martingale X, we have

Theorem 7.14 Xt is a continuous martingale. Then

Z t
Xt − 21 hXit 1
e −1= eXs − 2 hXis dXs . (7.5)
0

1
Consequently, eXt − 2 hXit is a martingale.

1
Proof. Let Yt = Xt − hXit , and f (x) = ex . Then, by the Ito’s lemma,
2
1 1 1
 
df (Yt ) = f 0 (Yt )dYt + f 00 (Yt )dhY it = f (Yt )d Xt − hXit + f (Yt )dhXit = f (Yt )dXt .
2 2 2

1
Now eXt − 2 hXit is called the Ito’s exponential w.r.t. Xt . The theorem will prove very useful
in the change of measure (Girsanov theorem). It is also very useful result in characterizing BM
as shown below.

22
7.8 Levy characterization of Brownian motion

Theorem 7.15 (Levy characterization of Brownian motion) Wt is a Brownian motion

if and only if

• Wt is a continuous martingale with W0 = 0;

• hW it = t.

Proof. The “=⇒” part is trivial. Let us check the “⇐=” part. We must check

(a) Wt − Ws ∼ N (0, t − s),

(b) Wt − Ws ⊥ Fs ,
(c) Wt is continuous.

23
 We only need to check (a) and (b) as (c) is assumed to be true. From Theorem 7.14,
1 1
   
Mt =: exp θWt − hθWt i = exp θWt − θ2 t
2 2
is a martingale. That is

E(Mt |Fs ) = Ms
1 1
     
⇐⇒ E exp θWt − θ2 t |Fs = exp θWs − θ2 s
2 2
1 2
 
⇐⇒ E (exp (θ[Wt − Ws ]) |Fs ) = exp θ (t − s)
2
1
 
⇐⇒ E (exp (θ[Wt − Ws ])) = E (exp (θ[Wt − Ws ]) |Fs ) = exp θ2 (t − s) ,
2
which implies (a) and (b).

24
7.9 Geometric Brownian motion (GBM)

Bachelier (1900) used Brownian motion to model stock markets, but the shortcoming of this
approach is that Brownian motion could be negative. To avoid this problem, Samuelson (1964)
introduced geometric Brownian motion (GBM) to model stock markets.

2 /2)t+σB
Example 7.8 The GBM Xt = e(µ−σ t
satisfies
Z t Z t
Xt = X0 + µXs ds + σXs dBs .
0 0

Proof. Let Xt = eYt = f (Yt ), where Yt = (µ − σ 2 /2)t + σBt and f (x) = ex . Then,

dXt = df (Yt )
1
= f 0 (Yt )dYt + f 00 (Yt )dhY it
2
  1
= f (Yt )d (µ − σ 2 /2)t + σBt + f (Yt )d(σ 2 t)
2
= f (Yt )d (µt + σBt )
= µXt dt + σXt dBt .

25
Is GBM appropriate?

Empirical evidence shows and most economists agree that the relative returns (or log returns)
of stock markets for the same time length can be regarded as constants and the relative returns
over non-overlapping intervals remain roughly uncorrelated. As a first approximation, we
could describe this phenomenon by
Xt+∆t − Xt
≈ µ∆t + σ(Bt+∆t − Bt ), (9.6)
Xt
Note that
Xt+∆t − Xt Xt+∆t − Xt
   
E = µ∆t, var = σ 2 ∆t,
Xt Xt

hence µ and σ 2 represent the average rate and variance (or volatility) of relative return per
unit. Letting ∆t → 0 in (9.6), we formally get dX Xt
t
= µdt + σdBt , or equivalently, dXt =
(µ−σ 2 /2)t+σBt
µXt dt + σXt dBt . From the last example, Xt = e .

26
 Empirical evidence also shows that, although the relative returns over non-overlapping
intervals are roughly uncorrelated, the square of the relative returns over non-overlapping in-
tervals are correlated. This means that the model (9.6) is not really appropriate to describe
stock markets. This implies that the stock price is a GBM in practice. Despite this, GBM
offers a very good starting point.

One way to improve the GBM is to introduce the so-called general Ito process:

dXt = µ(t, Xt )dt + σ(t, Xt )dBt .

The shortcoming is that we no longer have simple interpretations for µ(t, Xt ) and σ(t, Xt )
anymore.

27
7.10 Exercises
1. f is Lipschitz-continuous on [0, T ] if there exists a constant C > 0 such that for any
t, t0 ∈ [0, T ], |f (t) − f (t0 )| < C|t − t0 |. Show that a Lipschitz-continuous function is of
bounded variation and its quadratic variation over [0, T ] is zero, i.e. , hf iT = 0.

2. Derive dXt and check which ones are martingales, where Xt is

(a) et/2 cos(Bt ),

Z t
2
(b) t Bt − 2 sBs ds,
0
 3
(c) a1/3 + Bt /3 ,
(d) exp{σBt },
(e) Bt4 − 6tBt2 + 3t2 ,
Rt
(f) tBt − 0 Bs ds,
(g) (Bt + t) exp{−Bt − 12 t},
(h) tBt e−Bt ,
Z t
dBs
(i) (1 − t) , where 0 ≤ t < 1,
0 1−s

(j) Φ (Yt ), where Yt = √BT t−t , 0 ≤ t ≤ T , and Φ(x) is the standard Normal d.f.

3. Find f (x) such that f (Bt + t) is a martingale.

Rt
4. Show that Xt = 0 sign(Bs )dBs is a Brownian motion, where

sign(x) = −1 if x ≤ 0,
1 if x > 0.

5. Let dXt = a(Xt )dt + σ(Xt )dBt , X0 = 0, where σ(t) ≥ δ > 0. Find an increasing
function f (·) such that Yt = f (Xt ) satisfies

dYt = b(Yt )dt + dBt .

R Xt
6. Let Xt = eBt and Yt = 1 ln zdz.

(a) Find dXt and dhXit .

(b) Find dYt and hence show that Yt can be expressed as dYt = a(Bt )dt + b(Bt )dBt .
Specify these two functions a(x) and b(x).
Z t
7. Let Yt := Su du, where dSt = σSt dBt with S0 = 1. Find EYt and V ar(Yt ).
0

8. (The Ornstein-Uhlenbeck process) The Ornstein-Uhlenbeck process, {Xt , t ≥ 0}, is the

unique solution to Langevin’s equation,

dXt = −αXt dt + dBt , X0 = x.

Show that Z t
−αt −αt
Xt = e x+e eαs dBs ,
0

28
 and use this expression to calculate the mean and variance of Xt .
(This equation was originally introduced as a simple idealized model for the velocity of
a particle suspended in a liquid. In finance it is a special case of the Vasicek model of
interest rates.)

9. The Cox-Ingersoll-Ross model of interest rates assumes that the interest rate, r, is not
deterministic, but satisfies the stochastic differential equation
√
drt = (α − βrt )dt + σ rt dBt .

(This process is known as a squared Bessel process.) Taking α = 0 above, find the
√
stochastic differential equation followed by { rt , t ≥ 0}.
(k) (k) (k)
10. Let dXt = Xt Vt dBt , where k = 1, 2. Using Ito’s lemma, it can be shown that this
is equivalent to
Z t
(k) (k) 1 Z t (k) 2 (k) 1
Xt = X0 exp{ Vs(k) dBs − Vs ds} =: X0 exp{Zs(k) − hZ (k) is }
0 2 0 2
(1) (2)
Is Yt =: Xt Xt a martingale? If not, how do you turn it into one?

11. For a continuous semimartingale Z, denote E(Zt ) as the stochastic exponential of Z, i.e.,
E(Z) satisfies
dE(Zt ) = E(Zt )dZt , and E(Z0 ) = 1.
For any two continuous semimartingales X and Y , show that

(a) E(Xt )E(Yt ) = E (Xt + Yt + hX, Y it ) .

(b) (E(Xt ))−1 = E (−Xt + hXit ) .

29