You are on page 1of 20

Chapter 8

Stochastic Differential Equations

Stochastic Differential Equations (SDE) do not usually have explicit solutions except for linear
ones. So we first look at the linear SDEs. For others, we will resort to numerical methods, such
as Euler scheme, or Milstein scheme.

Linear SDEs are given by


Z t Z t
X t = X0 + [c1 (s)Xs + c2 (s)]ds + [σ1 (s)Xs + σ2 (s)]dBs , t ∈ [0, T ].
0 0

In differential form, this is

dXt = [c1 (t)Xt + c2 (t)]dt + [σ1 (t)Xt + σ2 (t)]dBt .

That is, the coefficients of dt and dBt are linear in Xt . We shall look at these one by one.

1
8.1 Simple linear SDEs

8.1.1 Geometric Brownian motion (simple linear SDE with multi-


plicative noise)

Consider the linear SDE


Z t Z t
Xt = X0 + c Xs ds + σ Xs dBs , t ∈ [0, T ].
0 0

In differential form, this is

dXt = cXt dt + σXt dBt . (1.1)

2
Approach I: Assume Xt = f (t, Bt ) for some smooth function f (t, x). By the Ito lemma,
1
 
dXt = f1 (t, Bt ) + f22 (t, Bt ) dt + f2 (t, Bt )dBt .
2
By (1.1), dXt = cf (t, Bt )dt + σf (t, Bt )dBt . Equating the coefficients of dt and dBt of the two
equations, we have
1
f1 (t, x) + f22 (t, x) = cf (t, x) (1.2)
2
f2 (t, x) = σf (t, x) (1.3)

We first turn second-order differentiable equations into first-order ones. To do this, differenti-
ating (1.3) w.r.t. x, we get

f22 (t, x) = σf2 (t, x) = σ 2 f (t, x) (1.4)

Putting this into (1.2), then (1.2) and (1.3) reduce to

f1 (t, x) = rf (t, x) (where r = c − 0.5σ 2 ) (1.5)


f2 (t, x) = σf (t, x) (1.6)

Note that (1.5) admits solution of the form f (t, x) = C1 (x)ert while (1.6) admits solution of
the form f (t, x) = D1 (t)eσx . Therefore, both equations are satisfied by

f (t, x) = Cert eσx = f (0, 0)ert+σx .

Therefore,   
Xt = f (t, Bt ) = X0 ert+σBt = X0 exp c − σ 2 /2 t + σBt .

3
Approach II: From dXt = cXt dt + σXt dBt , we have
dXt
= cdt + σdBt .
Xt
The LHS looks like d ln(Xt ), but not quite the same since Xt is not differentiable. Using Ito’s
Lemma, we have
! !
dXt 1 −1 dXt 1 −1 dXt 1 2
d ln(Xt ) = + 2
dhXit = + 2
σ 2 Xt2 dt = − σ dt
Xt 2 Xt Xt 2 Xt Xt 2

Therefore,
dXt 1
= d ln(Xt ) + σ 2 dt = cdt + σdBt ,
Xt 2
2
!
σ
=⇒ d ln(Xt ) = c − dt + σdBt
2
σ2
Z t ! Z t
=⇒ ln(Xt ) = ln(X0 ) + c− ds + σdBs
0 2 0

σ2
!
= ln(X0 ) + c − t + σBt
2
  
=⇒ Xt = X0 exp c − σ 2 /2 t + σBt .

Remark 8.1 The solution is in fact the Geometric Brownian motion. Given the initial starting
point X0 = x, ln Xt processes a normal distribution as
 
ln Xt = ln x + c − σ 2 /2 t + σBt

4
8.1.2 Ornstein-Uhlenbeck process (simple linear SDE with additive
noise)

Consider the linear SDE


Z t
Xt = X 0 + c Xs ds + σBt , t ∈ [0, T ],
0

for given constants c and σ > 0. In differential form, this is

dXt = cXt dt + σdBt .

The solution is called the Ornstein-Uhlenbeck process.

5
Approach I: Suppose Xt = f (t, Bt ) for some smooth function f (t, x). By Ito lemma,
1
 
dXt = f1 (t, Bt ) + f22 (t, Bt ) dt + f2 (t, Bt )dBt .
2
On the other hand, dXt = cf (t, Bt )dt + σdBt . Compare the above two equations, we have
1
f1 (t, x) + f22 (t, x) = cf (t, x) (1.7)
2
f2 (t, x) = σ (1.8)

We now turn second-order differentiable equations into first-order ones. Differentiating (1.3)
w.r.t. x, we get

f22 (t, x) = 0 (1.9)

Putting this into (1.2), then (1.2) and (1.3) reduce to

f1 (t, x) = cf (t, x) (1.10)


f2 (t, x) = σ (1.11)

From (1.10), we guess that f might have a component ect . So let f (t, x) = ect g(t, x), i.e.,
Xt = f (t, Bt ) = ect g(t, Bt ) = ect Yt . By integration by parts,
 
dYt = dg(t, Bt ) = d e−ct Xt = e−ct dXt − ce−ct Xt dt
= e−ct (cXt dt + σdBt ) − ce−ct Xt dt
= e−ct σdBt
R t −cs R t −cs
That is, Yt = Y0 + 0 e σdB s = X0 + 0 e σdB s. Therefore,
Z t
ct ct
Xt = e Yt = e X 0 + e ct
e−cs σdBs .
0

6
Approach II:

dXt = cXt dt + σdBt ,


=⇒ dXt − cXt dt = σdBt ,
=⇒ e−ct

dXt − ce−ct Xt dt = σe−ct dBt ,
=⇒ d e−ct Xt = σe−ct dBt ,
Z t
−ct
=⇒ e Xt − X0 = σe−cs dBs ,
0
Z t
=⇒ Xt = e X0 + σe ct ct
e−cs dBs .
0

 R t −2cs 
Remark 8.2 From the next theorem, 0t e−cs dBs ∼ N 0,
R
0 e ds . Then, given the initial
starting point X0 = x, Xt has a normal distribution as
Z t
Xt = ect x + σect e−cs dBs ∼ N (µx , σx2 ),
0

where

µx = ect x,
e2ct − 1
Z t !
−2cs 1  
σx2 2 2ct
= σ e e ds = σ e 2 2ct
1 − e−2ct = σ2.
0 2c 2c

In particular, if c < 0, we have µx → 0, and σx2 → σ 2 /(−2c).

7
Rt
Theorem 8.1 Let Yt = 0 h(s)dBs .

 Rt 2 
(a) If h is non-random (deterministic), then Yt ∼ N 0, h (s)ds .
0

(b) RIf h is random, then Yt may not be normal. But we still have EYt = 0 and V ar (Yt ) =
t 2
0 Eh (s)ds.

Proof.

Pn Rt
(a) By definition, Zn =: i=1 h(ti−1 )∆i B −→ 0 h(s)dBs =: Z in L2 , and hence in distribu-
tion. However, since
n
!
h2 (ti−1 )∆i t ,
X
Zn ∼ N 0,
i=1

the characteristic function (c.f.) of Zn


n
!
1 2X 1 2Z t 2
 
iλZn 2
ψZn (λ) = Ee = exp − λ h (ti−1 )∆i t −→ exp − λ h (s)ds ,
2 i=1 2 0

Rt 2
which is the c.f. of a normal r.v. N (0, 0h (s)ds).
Rt
(b) Let h(s) = Bs , then Yt = 0 Bs dBs = 21 (Bt2 − t) is centered χ21 , but not normal.

8
Remark 8.3 OU process in discrete time is related to AR(1) model in time series analysis.

Formally setting dt = 1, we have

∆Xt = Xt+1 − Xt = cXt + σ∆Bt = cXt + σ(Bt+1 − Bt ) =: cXt + t+1 ,

then
Xt+1 = (c + 1)Xt + t+1 ,
where t ∼i.i.d. N (0, σ 2 ), and t+1 ⊥ Xt . This is in fact AR(1) model in time series. Typically,
|1 + c| < 1 (i.e., −1 < c < 0).

9
Remark 8.4 (O-U process) Another form of the Ornstein-Uhlenbeck process is (a.k.a. the
mean-reverting process):
dXt = θ(µ − Xt )dt + σdBt .
This can be used to model interest rates, currency exchange rates, and commodity prices, where
µ represents the equilibrium or mean value supported by fundamentals; σ the volatility, and θ
the rate the variable reverts towards the mean.

Letting Yt = Xt − µ, we can rewrite the above process:

dYt = −θYt dt + σdBt .

From earlier discussion, Yt = Xt − µ → 0, i.e., Xt → µ as t → ∞. Furthermore, if Xt < µ,


the drift will be positive; if Xt > µ, the drift will be negative. In other words, the mean acts as
an equilibrium level for the process.

Example 8.1 (The Vasicek interest model) See the textbook, page 152.

• mean reverting property

• conditions are needed to ensure that interest rates are non-negative.

10
8.2 General linear SDEs

Consider the general linear SDE


Z t Z t
X t = X0 + [c1 (s)Xs + c2 (s)]ds + [σ1 (s)Xs + σ2 (s)]dBs , t ∈ [0, T ].
0 0

In differential form, dXt = [c1 (t)Xt + c2 (t)]dt + [σ1 (t)Xt + σ2 (t)]dBt .

8.2.1 Linear SDE with additive noise

Setting σ1 (t) = 0, we have

dXt = [c1 (t)Xt + c2 (t)]dt + σ2 (t)dBt ,


=⇒ dXt − c1 (t)Xt dt = c2 (t)dt + σ2 (t)dBt ,
Rt Rt
=⇒ e− 0
c1 (s)ds
dXt − c1 (t)e− 0
c1 (s)ds
Xt dt
Rt
= e− 0
c1 (s)ds
[c2 (t)dt + σ2 (t)dBt ],
 Rt  Rt
=⇒ d e− 0
c1 (s)ds
Xt = e− 0
c1 (s)ds
[c2 (t)dt + σ2 (t)dBt ],
Rt Z t Rt
− c1 (s)ds − c1 (s)ds
=⇒ e 0 Xt − X0 = e 0 [c2 (t)dt + σ2 (t)dBt ],
0
Rt  Z t Rt 
c1 (s)ds − c1 (s)ds
=⇒ Xt = e 0 X0 + e 0 [c2 (t)dt + σ2 (t)dBt ] .
0

11
8.2.2 Homogeneous linear SDE

Setting c2 (t) = σ2 (t) = 0, we have

dXt = c1 (t)Xt dt + σ1 (t)Xt dBt .

Dividing Xt on both sides, we get


dXt
= c1 (t)dt + σ1 (t)dBt .
Xt
The LHS looks like d ln(Xt ), but not quite the same. By Ito’s Lemma,
! !
dXt 1 −1 dXt 1 −1
d ln(Xt ) = + 2
dhXit = + σ12 (t)Xt2 dt
Xt 2 Xt Xt 2 Xt2
dXt 1 2
= − σ1 (t)dt.
Xt 2
Therefore,
dXt 1
= d ln(Xt ) + σ12 (t)dt = c1 (t)dt + σ1 (t)dBt ,
Xt 2
σ12 (t)
!
=⇒ d ln(Xt ) = c1 (t) − dt + σ1 (t)dBt
2
σ12 (s)
Z t ! Z t
=⇒ ln(Xt ) = ln(X0 ) + c1 (s) − ds + σ1 (s)dBs
0 2 0
Z t   Z t 
=⇒ Xt = X0 exp c1 (s) − σ12 (s)/2 ds + σ1 (s)dBs .
0 0

12
8.2.3 General linear SDE

We will now consider solving general linear SDEs:

dXt = [c1 (t)Xt + c2 (t)]dt + [σ1 (t)Xt + σ2 (t)]dBt

Step 1: Solving the SDE only involving Xt ’s

Rewrite

dXt = [c1 (t)Xt dt + σ1 (t)Xt dBt ] + [c2 (t)dt + σ2 (t)dBt ]

If we set c2 (t) = σ2 (t) = 0, we get a homogeneous linear SDE. Denote its solution by Yt , i.e.,
Yt satisfies

dYt = c1 (t)Yt dt + σ1 (t)Yt dBt .

This has been solved before:


Z t   Z t 
Yt = Y0 exp c1 (s) − σ12 (s)/2 ds + σ1 (s)dBs .
0 0

13
Step 2: Solving the original SDE

Assume that the solution Xt is of the form

Xt = Yt Zt .

Then, it suffices to solve for Zt . Now


     
dZt = d Xt Yt−1 = Xt d Yt−1 + Yt−1 dXt + dhX, Y −1 it ,

But by Ito’s Lemma,


    1  −3 
d Yt−1 = − Yt−2 dYt + 2Yt dhY it
2
  1  −3  2
= − Yt−2 [c1 (t)Yt dt + σ1 (t)Yt dBt ] + 2Yt σ1 (t)Yt2 dt
2
= −[c1 (t)Yt−1 dt + σ1 (t)Yt−1 dBt ] + σ12 (t)Yt−1 dt
= [σ12 (t) − c1 (t)]Yt−1 dt − σ1 (t)Yt−1 dBt

14
Therefore,
   
dZt = Xt d Yt−1 + Yt−1 dXt + dhX, Y −1 it
 
= Xt [σ12 (t) − c1 (t)]Yt−1 dt − σ1 (t)Yt−1 dBt
 
+ Yt−1 ([c1 (t)Xt + c2 (t)]dt + [σ1 (t)Xt + σ2 (t)]dBt )
 
− (σ1 (t)Xt + σ2 (t)) σ1 (t)Yt−1 dt
 
= [σ12 (t) − c1 (t)]Zt dt − σ1 (t)Zt dBt
 
+ [c1 (t)Zt + c2 (t)Yt−1 ]dt + [σ1 (t)Zt + σ2 (t)Yt−1 ]dBt
 
− σ12 (t)Zt + σ2 (t)σ1 (t)Yt−1 dt
= [c2 (t) − σ2 (t)σ1 (t)]Yt−1 dt + σ2 (t)Yt−1 dBt .

That is,
Z t Z t
Zt = Z0 + [c2 (s) − σ2 (s)σ1 (s)]Ys−1 ds + σ2 (s)Yt−1 dBs .
0 0

Then, noting that X0 = Y0 Z0 ,

Xt = Yt Zt = (Yt /Y0 )(Y0 Zt )


Z t
= (Yt /Y0 )(X0 + [c2 (s) − σ2 (s)σ1 (s)](Y0 /Ys )ds
0
Z t
+ σ2 (s)(Y0 /Ys )dBs )
0
Z t
= Ỹt (X0 + [c2 (s) − σ2 (s)σ1 (s)]Ỹs−1 ds
0
Z t
+ σ2 (s)Ỹs−1 dBs ),
0

where
Z t   Z t 
Ỹt = exp c1 (s) − σ12 (s)/2 ds + σ1 (s)dBs , t ∈ [0, T ].
0 0

15
8.3 Euler and Milstein approximations

SDE’s do not usually admit explicit solutions except for linear ones. In those cases, we have to
resort to numerical methods. We introduce two such methods below.

8.3.1 Euler approximations

Look at the simplest SDE


Z t Z t
Xt = X0 + a(Xs )ds + b(Xs )dBs
0 0

Then,
Z t+∆t Z t+∆t
Xt+∆t = Xt + a(Xs )ds + b(Xs )dBs
t t
Z t+∆t Z t+∆t
' Xt + a(Xt )ds + b(Xt )dBs
t t
Z t+∆t Z t+∆t
= Xt + a(Xt ) ds + b(Xt ) dBs
t t
= Xt + a(Xt )∆t + b(Xt ) (Bt+∆t − Bt )
√ (Bt+∆t − Bt )
= Xt + a(Xt )∆t + b(Xt ) ∆t √
∆t

:= Xt + a(Xt )∆t + b(Xt ) ∆t t ,

where
(Bt+∆t − Bt )
t := √ ∼iid N (0, 1).
∆t

16
8.3.2 Milstein approximations

In the
√ Euler approximation: X t+∆t = X t + a(X t )∆t + b(X t ) ∆t t , the last term is of order
Op ( ∆t) which is much larger than the second last term which is of size Op (∆t). To improve
the approximation, we may consider
R t+∆t
expanding one more term in the volatility term. In other
words, expand b(Xs ) further in t b(Xs )dBs . So for

t ≤ s ≤ t + ∆t,

we have

b(Xs ) = b(Xt ) + b0 (Xt )(Xs − Xt ) + ...


Z s Z s 
0
= b(Xt ) + b (Xt ) a(Xr )dr + b(Xr )dBr + ...
t t
= b(Xt ) + b0 (Xt )a(Xt )(s − t) + b0 (Xt )b(Xt )(Bs − Bt ) + ...

Then,
Z t+∆t
b(Xs )dBs
t
Z t+∆t
≈ [b(Xt ) + b0 (Xt )a(Xt )(s − t) + b0 (Xt )b(Xt )(Bs − Bt )]dBs
t
Z t+∆t Z t+∆t
= b(Xt ) (Bt+∆t − Bt ) + b0 (Xt )a(Xt ) (s − t)dBs + b0 (Xt )b(Xt ) (Bs − Bt )dBs
t t
=: b(Xt ) (Bt+∆t − Bt ) + b0 (Xt )a(Xt )Mt1 + b0 (Xt )b(Xt )Mt2 ,

where t is defined as before and


Z t+∆t Z t+∆t
Mt1 := (s − t)dBs , Mt2 := (Bs − Bt )dBs
t t

are both martingales.

17
First for Mt2 , we have
Z t+∆t
Mt2 = (Bs − Bt )dBs
t
Z t+∆t Z t+∆t
= Bs dBs − Bt dBs
t t
1 2 
= Bs − s |tt+∆t − Bt (Bt+∆t − Bt )
2
1 2  1
= Bt+∆t − Bt2 − ∆t − Bt (Bt+∆t − Bt )
2 2
1 2  1
= Bt+∆t + Bt2 − 2Bt Bt+∆t − ∆t
2 2
1 2 1
= (Bt+∆t − Bt ) − ∆t
2 2
1 2
= ∆t(t − 1),
2
where
(Bt+∆t − Bt )
t := √ ∼iid N (0, 1).
∆t

18
Secondly, noting that Mt1 is a martingale, by Theorem 8.1, we have

(∆t)3
Z t+∆t !
Mt1 ∼ N 0, (s − t)2 ds = .
t 3
 
Therefore, Mt1 = Op (∆t)3/2 .
 
So we can keep Mt2 = Op ((∆t)) and throw away Mt1 = Op (∆t)3/2 . Finally, we have
Z t+∆t Z t+∆t
Xt+∆t = Xt + a(Xs )ds + b(Xs )dBs
t
√t
' Xt + a(Xt )∆t + b(Xt ) ∆t t + b0 (Xt )a(Xt )Mt1 + b0 (Xt )b(Xt )Mt2

' Xt + a(Xt )∆t + b(Xt ) ∆t t + b0 (Xt )b(Xt )Mt2
√ 1
= Xt + a(Xt )∆t + b(Xt ) ∆t t + b0 (Xt )b(Xt )∆t(2t − 1).
2
This is so called Milstein approximations.

19
8.4 Exercises
1. RLet I = 01 B(r)2 dr with B(r) = 0r f (s)dW (s) and f (s) is deterministic such that
R R
1 2
0 f (s) ds < ∞ and W (s)
 is Ra standard Brownian motion. Find E(I 2 ). (Hint: E(I 2 ) =
E 0 B(r)2 dr 01 B(s)2 ds = 01 01 E[B(r)2 B(s)2 ]drds).
R1 R R

2. Solve the following SDEs:

(a) dXt = Bt Xt dt + Bt Xt dBt with X0 = 1;


(b) dXt = Xt dt + Bt dBt with X0 = 1;
(c) dXt = Xt3 dt − Xt2 dBt with X0 = 1. (Note that this is NOT a linear SDE.)
(d) dXt = rdt + αXt dBt with X0 = 1 where r, α are real constants.
1 1/3 2/3
(e) dXt = Xt dt + Xt dBt ; X0 = a.
3
1 1
(f) dXt = − Xt dt + dBt , where X0 = 0.
1+t 1+t
(g) dXt = c(b − Xt )dt + σdBt , X0 = x0 .

3. Assume that the interest rate rt follows

drt = a(b − rt )dt + crt dBt ,

where a, b, c > 0. Solve for rt by the following steps.


 
c2
(a) Write St = exp (−a − 2
)t + cBt . Derive the SDE for St−1 .
 
(b) Show that d rt St−1 = abSt−1 dt.
(c) Hence find the explicit representation of rt .

[Remark: This is just a special case of the general procedure given in the lecture note:
(1) solve the SDE only involving rt : dSt = −aSt dt + cSt dBt , (2) let rt = St Zt , and then
derive a SDE for Zt ]

4. Let Yt = 0t Bs dBs and Wt = 0t sign(Bs )dBs . Show that dYt = t + 2Yt dWt .
R R

20

You might also like