Stochastic Integration Rule Book - 2023

The idea is to collect up a summary of all the rules for Ito Calculus, sufficient for all our
example sheet questions and for the exam.

The set-up.
All processes are defined on a probability space (Ω, F, P).
There is a filtration (Ft ) of information: 0 ≤ s ≤ t implies Fs ⊆ Ft ⊆ F.

The integrands.
We always insist integrands G = (Gt : t ≥ 0) are (Ft ) adapted, that is Gt is (Ft ) measurable
for all t ≥ 0.
Of special interest are those that also in L2 (0, t) = {G satisfying E[ 0t G2s ds] < ∞} as for
R

these the mean zero and isometry properties hold.

The integral can be defined for those G that lie in M 2 (0, t) = {G satisfying 0t G2s ds < ∞ a.s.}.
R

In this module we only used integrands in the two cases (i) t → Gt is continuous (which
may or may not lie in L2 (0, T )) or, typically by stopping, (ii) t → Gt is left continuous and
bounded (which do lie in L2 (0, T )).

The integral.
For any (Ft ) Brownian motion W on R the integral 0t GdW is defined and satisfies:
R

Rt Rt Rt
1. Linearity: 0 (αG + βH)dW = α 0 GdW + β 0 HdW for all G, H, α, β.
Rt
2. Mean zero: for G ∈ L2 (0, t) we have E[ 0 GdW ] = 0.
Rt Rt Rt
3. Isometry: for G, H ∈ L2 (0, t) we have E[ 0 GdW 0 HdW ] = E[ 0 GHds].
Rt
4. Continuity: there is a continuous version for t → 0 GdW .
Rs Rt
5. Maximal inequality: E[sups≤t | 0 GdW |2 ] ≤ 4 E[ 0 G2 ds].
R t∧τ Rt
6. Stopping: for (Ft ) stopping time τ we have 0 GdW = 0 GIs≤τ dW .
Rt Rt
7. Time change: there exists another BM W̃ so that 0 GdW = W̃ ( 0 G2 ds) for all t.

Ito’s formula - d = 1 case.

Fix an (Ft ) Brownian motion W . A continuous (Ft ) adapted processes X is called an Ito
diffusion if there are integrands F, G so that
Z t Z t
Xt = X0 + Fs ds + Gs dWs for an F0 measurable initial condition X0 .
0 0

We write dX = F dt + GdB for short.

1
For f ∈ C 2 (R) the process Y = f (X) satisfies
Z 1 Z t
1 Z t 00
f (X) = f (X0 ) + f 0 (Xs )Fs ds + f 0 (Xs )Gs dBs + f (Xs )G2s ds.
0 0 2 0
We write for short dY = f 0 (X)dX + 12 f 00 (X)(dX)2 , using (dW )2 = dt, dW dt = 0 (dt)2 = 0.
For f : R × [0, ∞) → R with continuous derivatives ∂t f, ∂x f, ∂xx f we have
1
df (X, t) = ∂t f (X, t)dt + ∂x f (X, t)dX + ∂xx f (X, t)(dX)2
2

Ito’s formula - d > 1 case.

Take a d-dimensional (Ft ) Brownian motion W = (W 1 , . . . , W d ).
An adapted processes X is called an Ito diffusion if there are integrands F ∈ M 1 (0, T ), Gi ∈
M 2 (0, T ) for i = 1, . . . , M so that

dX = F dt + G1 dW 1 + . . . Gd W M

If all components of X = (X1 , . . . , XN ) are Ito diffusions and f ∈ C 2 (RN ) then

X 1X
df (X) = Di f (X)dXi + Dij f (X)dXi dXj
i 2 i,j

where dXi dXj can be calculated using

dW i dW j = 0 if i 6= j and dW i dW i = dt, and dW i dt = dtdt = 0.

If D ⊆ RN is open and f ∈ C 2 (D) then Ito’s formula holds for t < τD = inf{t : Xt ∈ ∂D}.

Levy’s Theorem
(d = 1) If X is an Ito process with zero drift (that is F ≡ 0), starting at X0 = 0, and
satisfying (dX)2 = dt then X is a Brownian motion.
(d > 1) If X 1 , . . . , X d are Ito processes with zero drift and satisfying (dX i )2 = dt and
dX i dX j = 0 then X − X0 is an d dimensional Brownian motion.

Diffusion - PDE dictionary

Suppose X solves dX = µ(X)dt + σ(X)dB on a bounded region D ⊆ Rd started at X = x.
00
The infinitesimal generator L for X is Lf = µf 0 + 21 σ 2 f in d = 1 and generally as

1X
(σσ T )ij Dij f
X
Lf = µi Di f +
i 2 i,j

2
Under suitable assumptions the following PDEs have the following probabilistic represen-
tation: (
Lu = 0 on D,
. . . u(x) = Ex [f (Xτ )].
u = f on ∂D.
( Z τ
Lu = g on D,

... u(x) = Ex g(Xt )dt .
u = 0 on ∂D. 0
(
Lu = hu on D,
 Rτ 
... u(x) = Ex f (Xτ ) exp− 0
h(Xs )ds
.
u = f on ∂D.


 ∂t u = Lu on [0, T ] × D,
u = g on [0, T ] × ∂D, ... u(t, x) = Ex [f (Xt I(t < τ ) + g(Xτ )I(t ≥ τ )] .


u = f at t = 0.
(
∂t u = Lu − hu on [0, T ] × Rd , h ≥ 0,
 Rt 
... u(t, x) = Ex f (Xt ) exp− 0
h(Xs )ds
.
u = f at t = 0.

Feller’s 1d exit Analysis

Set-up: X solves dX = µ(X)dt + σ(X)dB and X0 = x where σ(x) > 0 on (a, b).
Aim: describe the paths up to the exit from (a, b).
Step 1: Scale analysis. Find a scale function f : (a, b) → I so that Y = f (X) solves a
driftless equation. The scale function, which solves f 00 /f 0 = −2µ/σ 2 , can be taken to be
strictly increasing and the range I can be taken as one the four possibilities (−∞, ∞) or
(0, ∞) or (−∞, 0) or (0, 1).

Step 2. Speed analysis. After scale change we suppose dY = σ̂(y)dB and Y0 = y = f (x).
Set ν(dy) = 2/σ̂ 2 (y)dy.

1. Case 1: I = (∞, ∞). Exit in finite time is impossible; lim supt→∞ Yt = ∞ and
lim inf t→∞ Yt = −∞.

2. Case 2: I = (0, ∞). Exit at 0 with probability one; exit in finite time iff
R
0+ zν(dz) <
∞.

|z|ν(dz) <
R
3. Case 3: I = (−∞, 0). Exit at 0 with probability one; exit in finite time iff 0−
∞.

4. RCase 4: I = (0, 1). Exit at 0 with probability 1 − x, and then in finite Rtime iff
0+ zν(dz) < ∞; exit at 1 with probability x and then in finite time iff 1− (1 −
z)ν(dz) < ∞.

3
Invariant measure, ergodic theorem, CLT’s
The adjoint L∗ of the infinitesimal generator L∗ is given in d = 1 by L∗ φ = −(µφ)0 + 21 (σ 2 φ)00
and more generally by
1X
L∗ φ = − Dij ((σσ T )ij φ).
X
Di (µi φ) +
i 2 i,j

If φ(x) ≥ 0, φ(x) = 1 and L∗ φ = 0 then we expect φ(x)dx to be an invariant measure

R

(a.k.a. stationary distribution).

|g|dν < ∞ we expect an
R
When X has a unique invariant measure ν(dx) = φ(x)dx and
ergodic theorem
1Z t Z
g(Xs )ds → α = gdν
t 0
and typically also a central limit theorem (CLT)
1 Zt
√ (g(Xs ) − c0 )ds → N (0, σ 2 )
t 0
The Poisson equation method gives a method of establishing both results, summarisedR here.
Step (1): find a solution to the Poisson
Rt
equation Lu = g−α (you will need to take α = gdν;
Step (2): develop u(Xt ) − u(X0 ) = 0 (g(Xs ) − α)ds + stochastic integral; Step 3:√divide by
t and take limits to try and establish the ergodic theorem; Step (4): divide by t and try
to take limits to establish a CLT.
Existence and Uniqueness for dX = µ(X)dt + σ(X)dW
Pathwise Uniqueness means that whenever X and Y are solutions with respect to the
same brownian motion W and with initial conditions satisfying P [X0 = Y0 ] = 1 then
P [Xt = Yt , ∀t] = 1.
Uniqueness in law means that whenever X is a solution with respect to W and X is a
D D
solution with respect to W̃ so that X0 = Y0 then (Xt : t ≥ 0) = (Yt : t ≥ 0).
Theorem. When µ and σ are globally Lipschitz, that is

|µ(x) − µ(y)| ≤ C|x − y|, |σ(x) − σ(y)| ≤ C|x − y| for all x, y

then both pathwise uniqueness and uniqueness in law are true.

Improvements:
(i) In dimension d = 1 the same result holds under the assumption that |σ(x) − σ(y)| ≤
C|x − y|1/2 for all x, y, which covers Fisher-Wright and Feller examples.
(ii) If µ, σ are only locally Lipschitz (that is for each R > 0 there is CR where |µ(x)−µ(y)| ≤
CR |x − y| and |σ(x) − σ(y)| ≤ CR |x − y| for all |x|, |y| ≤ R) then existence and uniqueness
holds up to the explosion time τ = supn inf{t : |X|t ≥ n}. If there is a Lyapunov function
the explosion time is infinite - see Ex 3.12.

4
List of examples met during the module, and some of the information we gleaned.

1. Geometric Brownian Motion: dX = µXdt + σXdW with µ ∈ R and σ > 0 constant.

Explicit solution found. Moments studied in Ex 1.1. Inhomogeneous version in Ex
2.7.

2. Ornstein-Uhlenbeck SDE: dX = −µXdt+dW . Explicit solution found Ex 1.5. Gaus-

sian; mean and covariance found; stationary distribution found; shifted version Ex
2.6.
Xt
3. Brownian Bridge: dXt = − 1−t dt + dW . Explicit solution found Ex 1.6 Gaussian
(mean and covariance found). Bridges to general point in Ex 2.7.
√
4. Feller’s diffusion: dX = σ XdW . The transform E[exp(−θXt )] calculated, and
its limit P
√[Xt = 0]; the extension to the non-critical branching equation dX =
µXdt + σ XdW is in Ex. 2.9, 2.10.
q
5. Fisher Wright diffusion: dX = X(1 − X)dW . All moments are calculable Ex 2.8.
1
6. Polar decomposition: in d = 2 we get dR = 2R
dt + dW ; d = 3 explored in Ex 3.5.
c−1
7. Bessel SDE: dX = dt + dW ; stochastic Pythagoras Ex 3.4. Bessel squared: dX =
√ 2X
¸dt + 2 XdW .

8. Explosive SDE: dX = X 2 dt + X β dW ; scale analysis shows explodes only if β < 3/2.

9. Noisy gradient flow: dX = −∇V (X)dt + dW . Stationary distribution found. The

one dimensional case V 0 (x) = x−x3 used as an example for non-Lipschitz coefficients.

10. Krylov’s example ...

11. Stabilisation by noise example ...

12. Destabilisation by noise example ...

... to be completed ...