Section 10. Connections with PDE.

We have a progressively measurable stochastic process x(t, ) on (, F

t
, P) such that
the paths are continuous with probability 1. We have bounded progressively measurable
functions b(t, ) and a(t, ) with a(t, ) 0. Moreover
(1) y(t) = x(t) x(0)

t
0
b(s, )ds
and
(2) y
2
(t)

t
0
a(s, )ds
are martingales with respect to (, F
t
, P). It follows that
exp

[x(t) x(0)

t
0
b(s, )ds]

2
2

t
0
a(s, )ds]

is a martingale for all real . We proved it for the special case of b = 0 and a = 1. If we are
in d dimensions x(t, ) and b(t, ) would be R
d
valued and a = {a
i,j
} would be a positive
semi-denite matrix. The conclusion would then be
(3) exp

, x(t) x(0)

t
0
, b(s, )ds]
1
2

t
0
, a(s, )ds

are martingales with respect to (, F

t
, P). From this it would then follow that
(4) exp

f(t, x(t)) f(0, x(0))

t
0
[e
f(s,x(s))
(

t
+ L
s,
e
f
)(s, x(s))]ds

is again a martingale with respect to (, F

t
, P) for any smooth f. Replacing f by , x +
f(t, x) yields more martingales.
exp

[f(t, x(t)) f(0, x(0))] +, x(t) x(0)

t
0
[

b(s, )ds
1
2

t
0

, a(s, ),

]ds

Here

= [, ],

b(s, ) = [f
s
+ (L
s,
f)(s, x(s)), b(s, )],
a(s, ) =

a(s, )f(s, x(s)), f(s, x(s)) [a(s, )f(s, x(s))]

[a(s, )f(s, x(s))]
t
a(s, )

and L
s,
f is the operator
(L
s,
f)(s, x) =
1
2

i,j
a
i,j
(s, )(D
x
i
D
x
j
f)(s, x) +

j
b
j
(s, )(D
x
j
f)(s, x)
1
With x
0
(t) = f(t, x(t)), we dene the stochastic integral
z(t) =

t
0
dx
0
(s)

t
0
f(s, x(s)), dx(s)
Since a applied to [1, f(s, x(s))] is 0, z(t) is of bounded variation and
z(t) = z(0) +

t
0
f
s
(s, x(s))ds +

t
0
(L
s,
f)(s, x(s))

t
0
b(s, ), (f)(s, x(s))ds
This yields Itos formula
df(t, x(t)) = f
t
(t, x(t))dt + (f)(t, x(t))dx(t) +
1
2

i,j
a
i,j
(s, )(D
x
i
D
x
j
f)(t, x(t))dt
Why does (3) imply (4) ? To see this let us, for simplicity, suppose that d = 1 and f does
not depend on t. we can replace by i. This gives us the martingales
M

(t) = exp

i[x(t) x(0)

t
0
b(s, )ds] +

2
2

t
0
a(s, )ds]

We can take
A(t) = exp

t
0
b(s, )ds]

2
2

t
0
a(s, )ds]

then the martingale M(t)A(t)

t
0
M(s)dA(s) which reduces to
f(x(t)) f(x(0))

t
0
[(L
s,
f)(s, x(s))]ds
with f(x) = e
ix
is again a martingale. By Fourier integral representation any smooth
function is a super position of exponentials e
ix
. The martingale property extends by
linearity. Therefore
f(x(t)) f(x(0))

t
0
[(L
s,
f)(s, x(s))]ds
are martingales. Taking e
f
instead of f we will get
N(t) = e
f(x(t))
e
f(x(0))

t
0
(L
s,
e
f
)(x(s))ds
are martingales. Now take
A(t) = exp[f(x(0))

t
0
[(e
f
L
s,
e
f
)(x(s))ds]
2
and
N(t)A(t)

t
0
N(s)dA(s)
reduces to what we want. The important thing here is that a continuous process x(t, )
with b(t, ) and a(t, ) representing the conditional innitesimal mean and covariance in
the sense described above is connected very closely to the operator L
s,
. Of course for
L
s,
to be really an operator it is important to have b(s, ) = b(s, x(s, )) and a(s, ) =
a(s, x(s, ). Then the process is expected to be a Markov process and could have arisen
as a solution of a stochastic dierential equation
dx(t) = (t, x(t)) d(t) + b(t, x(t))dt
where

= a.
There are a few simple formal rules that summarize Itos formula. Suppose (t) is a
Brownian motion then
d(t)
2
= dt
{
i
()} are independent Brownian Motions
d
i
(t)d
j
(t) =
i,j
dt
and (dt)
2
= d(t)dt = 0. Consequently, if
dx(t) = a(t, )dt +

i
(t, )d
i
and
dy(t) = b(t, )dt +

i
c
i
(t, )d
i
then
dx(t)dy(t) = [

i
(t, )c
i
(t, )]dt
Finally
df(x(t)) = (f)(x(t)) dx(t) +
1
2

i,j
(D
x
i
D
x
j
f)(dx
i
(t)dx
j
(t))
Given a(s, x), and b(s, x), let us dene for each s the dierential operator
L
s
=
1
2

i,j
a
i,j
(s, x)D
x
i
D
x
j
+

j
b
j
(s, x)D
x
j
Let u(s, x) be a solution of the partial dierential equation
u
s
+ (L
s
u)(s, x) + g(s, x) = 0; u(T, x) = f(x)
3
Then if x(t, ) is any almost surely continuous process satisfying (1), (2) and P[x(0) =
x] = 1, then
u(0, x) = E
P
[

T
0
g(s, x(s))ds + f(x(T))]
Proof is elementary.
du(t, x(t)) = u
t
(t, x(t))dt + (u)(x(t)) dx(t) +
1
2

i,j
(D
x
i
D
x
j
f)(dx
i
(t)dx
j
(t))
= g(t, x(t))dt +u,

(t, x(t))d(t)
Therefore
u(t, x(t)) u(0, x(0)) +

t
0
g(s, x(s))
is a martingale. Equate expectations at t = 0 and t = T. There are other relations. If
u
s
+ (L
s
u)(s, x) + V (s, x)u(s, x) + g(s, x) = 0; u(T, x) = f(x)
then
u(0, x) = E
P

T
0
g(s, x(s))exp[

s
0
V (, x())d]ds + exp[

T
0
V (, x())d]f(x(T))

Ex: Work it out. Enough to show that

M(t) = u(t, x(t))A(t) + B(t)
is a martingale where
A(t) = exp[

t
0
V (s, x(s))ds]
and
B(t) =

t
0
exp[

s
0
V (, x())d]g(s, x(s))ds
Calculate dM(t) and keep only the dt terms.
dM(t) = A(t)du(t, x(t)) + u(t, x(t))dA(t) + dB(t)
= A(t)(u
t
+ L
t
u)dt + u(t, x(t))A(t)V (t, x(t)) + A(t)g(t, x(t))
= A(t)[u
t
(t, x(t)) + (L
t
u)(t, x(t)) + u(t, x(t))V (t, x(t)) + g(t, x(t))]dt
= 0
Black and Scholes: If u(t, x) solves
u
t
+

2
x
2
2
u
xx
= 0
4
and x(t) is the solution of
dx(t) = x(t)d(t) + b(t, x(t))dt
then
u(t, x(t)) u(s, x(s)) =

t
s
u
x
(, x())dx()
Modify to take care of interest rate r. We need, in prices discounted to current value,
e
rt
u(t, x(t)) e
rs
u(s, x(s)) =

t
s
e
r
u
x
(, x())[dx() r x()d]
to keep the hedge. In other words we need
d[e
rt
u(t, x(t))] = e
rt
[u
t
ru + u
x
dx +

2
x
2
2
u
xx
dt] = e
rt
[u
x
dx r xu
x
dt]
or a solution of
u
t
+

2
x
2
2
u
xx
dt + r xu
x
ru = 0
with u(T, x) = f(x).
We can solve equations in domains as well. The solution of
Lu =
1
2

i,j
a
i,j
(x)D
x
i
D
x
j
u +

b
j
(x)D
x
j
u = 0 for x D
with u = f on D, is represented as
u(x) = E
x
[f(x()]
where is the stopping time
= inf[t : x(t) / D]
and E
x
[ ] refers to expectation relative to the diusion process corresponding to L starting
from the point x D.
There is not that much conceptual dierence between the time independent and the time
dependent cases. We can always add an extra coordinate x
0
and take b
0
1 and a
0,j
0
for all j. Then we changed
u
t
+ L
t
u = u
t
+
1
2

i,j
a
i,j
(t, x)D
x
i
D
x
j
u +

b
j
(t, x)D
x
j
u
to

Lu =
1
2
d

i,j=1
a
i,j
(x
0
, x)D
x
i
D
x
j
u +
d

j=1
b
j
(x
0
, x)D
x
j
u + D
x
0
u
5
The matrix a is now degenerate.
The process starting form L can be dened through PDE. Solve
(5) u
s
+ L
s
u = 0 for s t and u(t, x) = f(x)
Represent
u(s, x) =

f(y)p(s, x, t, y)dy
Show p 0, satises Chapman-Kolmogorov equations and is nice enough to be the tran-
sition probabilities of a process with continuous paths. This will work if a, b are bounded
and Holder continuous and a is uniformly elliptic.

j
|b
j
(t, x)| C
for some C < . For some C and > 0,

i,j
|a
i,j
(t, x) a
i,j
(t, y)| +

j
|b
j
(t, x) b
j
(t, y)| C|x y|

,
and

i,j
|a
i,j
(t, x) a
i,j
(s, x)| +

j
|b
j
(t, x) b
j
(s, x)| C|s t|

,
Finally for some 0 < c C < ,
c

2
j

i,j
a
i,j
(t, x)
i

j
C

2
j
.
SDE may not work here unless = 1. But any process related to [a, b] by (1) and (2) will
be unique and be the same as the one coming from PDE. Because the PDE solution u(t, x)
of (5) will still have the property that u(t, x(t)) is a martingale with respect to any such
process.
6