Feynman-Kac formula

V a nice function (say bounded). u ∈ C 1,2 solves

∂u 1 ∂2u
∂t = 2 ∂x 2 + Vu, u(0, x) = u0 (x)

u0 (x) exp{−x 2 /2t}dx < ∞. Then

R
Rt
V (B(s))ds
u(t, x) = Ex [e 0 u0 (B(t))]

Proof.
Rs
For 0 ≤ s ≤ t let Z (s) = u(t − s, B(s))e 0 V (B(u))du . By Itô’s lemma
Z t
∂u 1 ∂ 2 u
 R
s
Z (t) − Z (0) = − + 2
+ Vu e 0 V (B(u))du ds =0
0 ∂s 2 ∂x
Z t
∂u Rs
+ (t − s, B(s))e 0 V (B(u))du dB(s) = mart
0 ∂x

⇒ Ex [Z (t)] = Ex [Z (0)]
() Stochastic Calculus March 2, 2007 1 / 23
1 In d > 1 if u ∈ C 1,2 solves
∂u 1
= ∆u + Vu, u(0, x) = u0 (x)
∂t 2
then h Rt i
u(t, x) = Ex e 0 V (B(s))ds u0 (B(t))
where B(t) is d-dimensional Brownian motion
2 If V = V (t, x)
h Rt i
u(t, x) = Ex e 0 V (t−s,B(s))ds u0 (B(t))
1∂ u 2
3 Feynman: solution u of it Schrödinger equation ∂u ∂t = i[ 2 ∂x 2 + Vu]
has ”representation”
Z Rt Rt 0 2
i
u(x, t) = ei 0 V (fs )ds− 2 0 |f | ds u0 (ft )dµ
f :f (0)=x
where µ is translation invariant measure on space of functions
BUT, No such measure µ
Kac pointed out that it is rigorous if i 7→ 1 because
1
Rt
|f 0 |2 ds formally
e− 2 0 dµ = Brownian motion
() Stochastic Calculus March 2, 2007 2 / 23
The Monte Carlo Method
Problem. Given V compute
1
λ0 = sup spectrum( ∆ + V )
2
Idea (Ulam, Fermi, von Neumann, Metropolis)

∂u 1
= ∆u + Vu, u(0, x) = 1
∂t 2

1 1 Rt
λ0 = lim log u(0, t) = lim log E0 [e 0 V (B(s))ds ]
t→∞ t t→∞ t

Simulate N Brownian paths B1 , . . . BN , N large

Take t large
N
1 1 X R t V (Bi (s))ds
λ0 ≈ log e0
t N
i=1

() Stochastic Calculus March 2, 2007 3 / 23

Arcsin law
Rt
ξ(t) = 1t 0 1[0,∞) (B(s))ds = the fraction of time that Brownian motion
is positive up to time t

0

√
a < 0;
2
P(ξ(t) ≤ a) = π arcsin a 0 ≤ a ≤ 1;

1 a > 1.


Simple explanation why distribution of ξ(t) indep of t

Z 1 Z 1 Z 1
1
ξ(t) = 1[0,∞) (B(ts))ds = 1[0,∞) ( √ B(ts))ds = 1[0,∞) (B̃(s))ds
0 0 t 0

d
ξ(t) = ξ(1)

() Stochastic Calculus March 2, 2007 4 / 23

Proof
By Feynman-Kac if we can find a nice solution of

∂u 1 ∂2u
= − u1[0,∞) u(0, x) = 1
∂t 2 ∂x 2

Then Rt
u(t, x) = Ex [e− 0 1[0,∞) (B(s))ds
]
and Z 1
u(t, 0) = e−at dP(ξ ≤ a)
0
R∞
α > 0, φα (x) = α 0 u(t, x)e−αt dt −→ − 12 φ00α + (α + 1[0,∞) )φα = α
( √ √
α
α+1 + Aex 2(α+1)
+ Be−x 2(α+1)
, x ≥ 0,
φα (x) = √ √
1+ Cex 2α + De−x 2α , x ≤ 0.

u ≤ 1 ⇒ φα ≤ 1 ⇒ A = D = 0
() Stochastic Calculus March 2, 2007 5 / 23
Proof.
φα (0− ) = φα (0+ ), φ0α (0− ) = φ0α (0+ )
α1/2 √ √1 √
⇒B= √
(1+α)( α+ α+1)
,C = (1+α)1/2 ( α+ α+1)1/2
r Z ∞
α
φα (0) = = E[e−tξ αe−αt ]dt
α+1 0
h i q
α α
By Fubini’s theorem this reads E α+ξ = α+1 or

Z 1
1 1
dP(ξ ≤ a) = √
0 1 + γa 1+γ

Looking up a table of transforms we find

2 1
dP(ξ ≤ a) = p da 0≤a≤1
π a(1 − a)

which is the density of the arcsin distribution

() Stochastic Calculus March 2, 2007 6 / 23
Stochastic differential equations

σ(x, t), b(x, t) mble

Definition
A stochastic process Xt is a solution of a stochastic differential
equation
dXt = b(Xt , t)dt + σ(Xt , t)dBt , X0 = x0
on [0, T ] if Xt is progressively measurable with respect to Ft ,
RT RT 2
0 |b(Xt , t)|dt < ∞, 0 |σ(Xt , t)| dt < ∞ a.s. and
Z t Z T
Xt = x0 + b(Xs , s)ds + σ(Xs , s)dBs 0≤t ≤T
0 0

The main point is that σ(ω, t) = σ(Xt , t), b(ω, t) = b(Xt , t)

Under reasonable conditions the solution Xt exists, is unique, and is a
Markov process

() Stochastic Calculus March 2, 2007 7 / 23

Ornstein-Uhlenbeck Process
Xt , t ≥ 0 is the solution of the Langevin equation
dXt = −αXt dt + σdBt , α>0
To solve it
deαt Xt = αeαt Xt dt + eαt (−αXt dt + σdBt ) = σeαt dBt
so Z t
−αt
Xt = X0 e +σ e−α(t−s) dBs
0
If X0 ∼ N (m, V ) indep of Bt , t ≥ 0 ⇒ Xt Gaussian process
m(t) = E[Xt ] = me−αt
σ 2 2α min(t,s)
c(s, t) = Cov(Xs , Xt ) = [V + (e − 1)]e−α(t+s)
2α
σ2 2
m = 0, V = 2α
σ
⇒ Xt stationary Gaussian c(s, t) = 2α e−α(t−s)
Rt
Yt = 0 Xs ds ”Physical” Brownian motion
() Stochastic Calculus March 2, 2007 8 / 23
Geometric Brownian motion

dSt = µSt dt + σSt dBt µ = drift σ = volatility

By Ito’s formula
σ2
St = S0 e(µ− 2
)t+σBt

is the solution

St ≥ 0 so it is (Samuelson) a better model of stock prices than Bt

(Bachelier)

Sometimes people write

dSt
= µdt + σdBt
St
dSt
but note that St 6= d log St
() Stochastic Calculus March 2, 2007 9 / 23
Bessel process (d = 2)

Let Bt = (Bt1 , Bt2 ) be 2d Brownian motion starting at 0,

q
rt = |Bt | = (Bt1 )2 + (Bt2 )2 .

By Ito’s lemma,

B1 1 B2 2 1 1
drt = dB + dB + dt.
|B| |B| 2 |B|

This is not a stochastic differential equation.

() Stochastic Calculus March 2, 2007 10 / 23

 t Z t 2
B1 1
Z
B
Y (t) = dB + dB 2
0 |B| 0 |B|
Let f (t, y ) be a smooth function Use Itô’s lemma. Intuitively
1
df (t, Yt ) = ∂t fdt + ∂y fdY + ∂y2 f (dY )2
2

B1 1 B2 2 2
(dY )2 = ( dB + dB )
|B| |B|
 1 2  2 2
B 1 2 B1B2 1 2 B
= (dB ) + 2 2
dB dB + (dB 2 )2
|B| |B| |B|
= dt
Z t
1
f (t, Yt ) = f (0, Y0 ) + (∂t f + ∂y2 f )(s, Ys )ds
0 2
Z t 1 Z t
B B2 2
+ ∂y f dB 1 + ∂y f dB
0 |B| 0 |B|
() Stochastic Calculus March 2, 2007 11 / 23
Itô’s lemma
dXt = σ(t, Xt )dBt + b(t, Xt )dt

Z tn o
f (t, Xt ) = f (0, X0 ) + ∂s f (s, Xs ) + Lf (s, Xs ) ds
0
Z t d
∂
f (s, Xs )dBsj
X
+ σij (s, Xs )
0 i,j=1 ∂xi

d d
1X ∂2f X ∂f
Lf (t, x) = aij (t, x) (t, x) + bi (t, x) (t, x)
2 ∂xi ∂xj ∂xi
i,j=1 i=1

d
X
aij = σik σjk a = σσ T
k=1

() Stochastic Calculus March 2, 2007 12 / 23

 B1 1 B2 2 1 1 1
drt = dB + dB + dt = dYt + rt−1 dt
|B| |B| 2 |B| 2

t
Z
1
f (t, Yt ) = f (0, Y0 ) + (∂t f + ∂y2 f )(s, Ys )ds
0 2
Z t 1 Z t
B B2 2
+ ∂y f dB 1 + ∂y f dB
0 |B| 0 |B|

2 t/2
In particular eλYt −λ is a martingale

So Yt is a Brownian motion.

Therefore
1
drt = dYt + rt−1 dt
2
is a stochastic differential equation for the new Brownian motion Yt
() Stochastic Calculus March 2, 2007 13 / 23
Itô’s lemma
Z tn o Z t
f (t, Xt ) − f (0, X0 ) = ∂s + L f (s, Xs )ds + ∇f (s, Xs ) · σdBs
0 0

Proof
X
= f (ti+1 , Xti+1 ) − f (ti , Xti )
i
X ∂f
= (ti , Xti )(ti+1 − ti ) + ∇f (ti , Xti ) · (Xti+1 − Xti )
∂t
i
d
1 X ∂2f
+ (ti , Xti )(Xtji+1 − Xtji )(Xtki+1 − Xtki )
2 ∂xj ∂xk
j,k=1
+ higher order terms

() Stochastic Calculus March 2, 2007 14 / 23

Proof continued
X ∂f Z t
∂f
(ti , Xti )(ti+1 − ti )→ (s, Xs )ds
∂t 0 ∂t
i

X
∇f (ti , Xti ) · (Xti+1 − Xti )
i
X Z ti+1 Z t
= ∇f (ti , Xti ) · ( σ(s, Xs )dBs ) → ∇f · σdB
i ti 0
X Z ti+1 Z t
+ ∇f (ti , Xti ) · ( b(s, Xs )ds) → ∇f · b ds
i ti 0

t
∂2f ∂2f
Z
(ti , Xti )(Xtji+1 −Xtji )(Xtki+1 −Xtki ) →
X
(s, Xs )ajk (s, Xs )ds
∂xj ∂xk 0 ∂xj ∂xk
i

() Stochastic Calculus March 2, 2007 15 / 23

Proof continued
To show the last convergence, ie
Z t
g(ti , Xti )(Xtji+1 − Xtji )(Xtki+1 − Xtki ) →
X
g(s, Xs )ajk (s, Xs )ds
i 0

Z ti+1 X Z ti+1 X
Z (ti , ti+1 ) = ( σjl (s, Xs )dBsl )( σkm (s, Xs )dBsm )
ti l ti m
Z ti+1 X
− σjl σkl (s, Xs )ds
ti l

E[|Z (ti , ti+1 )|2 ] = O((ti+1 − ti )2 )

() Stochastic Calculus March 2, 2007 16 / 23

Proof continued
X Z ti+1 Z t
g(ti , Xti )E[ aij (s, Xs )ds] → g(s, Xs )aij (s, Xs )ds
i ti 0

X X
E[( g(ti , Xti )Z (ti , ti+1 ))2 ] = E[g(ti , Xti )Z (ti , ti+1 )g(tj , Xtj )Z (tj , tj+1 )]
i i,j

i <j E[E[g(ti , Xti )Z (ti , ti+1 )g(tj , Xtj )Z (tj , tj+1 ) | Ftj ]] = 0

i =j E[E[g 2 (ti , Xti )Z 2 (ti , ti+1 ) | Fti ]]

= E[g 2 (ti , Xti )E[Z 2 (ti , ti+1 ) | Fti ]]
= O((ti+1 − ti )2 )

() Stochastic Calculus March 2, 2007 17 / 23

 Z tn o Z t
f (t, Xt ) − f (0, X0 ) − ∂s + L f (s, Xs )ds = ∇f (s, Xs ) · σdBs
0 0

d d
1X ∂2 X ∂
L= aij (t, x) + bi (t, x) = generator
2 ∂xi ∂xj ∂xi
i,j=1 i=1
Rt n o
Mt = f (t, Xt ) −
0 ∂ s + L f (s, Xs )ds is a martingale
Z tn o
0 = E[f (t, Xt ) − f (s, Xs ) − ∂u + L f (u, Xu )du | Fs ]
s
Z
= f (t, y)p(s, x, t, y )dy − f (s, x)
Z tZ n o
− ∂u + L f (u, y)p(s, x, u, y)dydu, Xs = x
s

() Stochastic Calculus March 2, 2007 18 / 23

For any f ,
Z
0 = f (t, y )p(s, x, t, y)dy − f (s, x)
Z tZ n o
− ∂u + L f (u, y)p(s, x, u, y)dydu
s

Fokker-Planck (Forward) Equation

d
∂ 1 X ∂2
p(s, x, t, y ) = (ai,j (t, y )p(s, x, t, y))
∂t 2 ∂yi ∂yj
i,j=1
d
X ∂
− (bi (t, y )p(s, x, t, y))
∂yi
i=1
∗
= Ly p(s, x, t, y)

lim p(s, x, t, y ) = δ(y − x).

t↓s
() Stochastic Calculus March 2, 2007 19 / 23
Kolmogorov (Backward) Equation

d
∂ 1X ∂ 2 p(s, x, t, y)
− p(s, x, t, y) = aij (s, x)
∂s 2 ∂xi ∂xj
i,j=1
d
X ∂p(s, x, t, y )
+ bi (s, x)
∂xi
i=1
= Lx p(s, x, t, y)

lim p(s, x, t, y ) = δ(y − x).

s↑t

() Stochastic Calculus March 2, 2007 20 / 23

Proof.
f (x) smooth

∂
− u = Ls u 0≤s<t u(t, x) = f (x)
∂s
Ito’s formula: u(s, X (s)) martingale up to time t
Z
u(s, x) = Es,x [u(s, X (s))] = Es,x [u(t, X (t))] = f (z)p(s, x, t, z)dz

Let fn (z) smooth functions tending to δ(y − z). We get in the limit that
p satisfy the backward equations.

() Stochastic Calculus March 2, 2007 21 / 23

Example. Brownian motion d = 1
1 ∂2
L=
2 ∂x 2

∂p(s, x, t, y 1 ∂ 2 p(s, x, t, y)
Forward = , t >s
∂t 2 ∂y 2

p(s, x, s, y ) = δ(y − x)

∂p(s, x, t, y 1 ∂ 2 p(s, x, t, y)
Backward − = , s < t,
∂s 2 ∂x 2

p(t, x, t, y) = δ(y − x)

() Stochastic Calculus March 2, 2007 22 / 23

Example. Ornstein-Uhlenbeck Process
σ2 ∂ 2 ∂
L= 2
− αx
2 ∂x ∂x

∂p(s, x, t, y 1 ∂ 2 p(s, x, t, y ) ∂
Forward = 2
+ (αyp(s, x, t, y), t > s,
∂t 2 ∂y ∂y

p(s, x, s, y ) = δ(y − x)

∂p(s, x, t, y 1 ∂ 2 p(s, x, t, y ) ∂p(s, x, t, y)

Backward − = −αx , s < t,
∂s 2 ∂x 2 ∂x

p(t, x, t, y) = δ(y − x)

() Stochastic Calculus March 2, 2007 23 / 23