This is a brief techical note on Brownian Motion and SDE. For more exhaustive
and rigorous material, checkout the references.
Brownian Motion
To define Brownian motion, we first need to know that such disribution exists.
Theorem 1 (existence)∃ a probability distribution over the set of all
continuous function B:R≥0↦R s.t.
P(B(0)=0)=1
(Stationary) ∀0≤s≤t,B(t)−B(s)∼N(0,t−s)
(Independent Increment) If intervals {[si,ti]} are non-overlapping,
then B(ti)−B(si) are independent.
Now we define the Brownian motion:
Definition The distribution given by Theorem 1 is called Brownian motion.
One important property of Brownian motion is that it’s NOT differentiable!
We will not prove this. Since Brownian motion can’t be properly computed using
normal calculus, we will use Ito’s Calculus to derive all the equations. To see
how we can derive Ito’s Lemma, we first check an important theorem:
Theorem 2 (Quadractic Variation)∀T>0, we have:
n→∞limt=1∑n(B(ntT))−(B(nt−1T))2=T
This is a surprising fact about Brownian motion because for any continuous and
differentiable function f, we have that
i=1∑n(f(ti+1)−f(ti))2≤i=1∑nf′(si)(ti+1−ti)∵ Mean Value Thm. =(0≤s≤Tmaxf′(s)2)i=1∑n(ti+1−ti)2=(0≤s≤Tmaxf′(s)2)nT2
and as n→0, ∑i=1n(f(ti+1)−f(ti))2→0. Thus,
for normal functions, the quadratic variation is zero. We can denote the
quadractic variation of Brownian motion as (dB)2=dt. We say that square of
difference in Brownian motion goes to the difference in time with probability 1.
Intuitively, the quadratic variation simply denote the variance of the Brownian motion between
an infinitesimal step, which, by definition, equals to the time step. For
deterministic function, the variance is simply zero. This accumulation of the
second order term makes the Linear appoximation of Brownian motion much different
from ordinary functions.
Ito’s Calculus
Suppose we want to compute f(Bt) for some smooth function f. The Taylor
Expansion of the f is:
df=f′dBt+2f′′(dBt)2+3!f′′′(dBt)3+...
Normally, we would ignore terms after the first order. However, from Theorem 2,
we know that (dBt)2=dt, and thus we can no longer ignore the second order
term. Finally, we can write the (simple) Ito’s Lemma:
Lemma (Simple Ito)
df=f′dBt+21f′′dt
For a stronger Ito’s Lemma, consider the Taylor Expansion of the function
f(t,Bt):
Now, let’s consider a more general stochastic process Xt defined as
dXt=μdt+σdBt
Where μdt is called the drift term. We can use Ito Lemma to write the
df(t,Xt) as follow:
df=(∂t∂f+21∂Xt2∂2f)dt+∂Xt∂fdXt
Stochastic Differential Equation (SDE)
Definition
A SDE is a differential equation of the form
dXt=μ(t,Xt)dt+σ(t,Xt)dBt
Then, we have the following theorem:
Theorem
SDE has a solution and if given X0 then the solution is unique as long as μ,σ
satisfy certain condition†.
Let’s consider when μ and σ are proportional to Xt:
dXt=μXtdt+σXtdBt,σ,μ∈R,X0=x0
Using Ito Lemma, we can show that the solution for the given SDE is Xt=X0eσXt+(μ−21σ2)t.
Now, let’s consider when the stochastic term σ(t,Xt) is
independent of the current state Xt:
dXt=−αXtdt+σdBt,α>0,μ∈R,X0=x0
This is called
the Ornstein-Uhlenbeck (OU) Process.
And we have the following solution for the OU process:
Xt=e−αtX0+∫0tσeα(s−t)dBs
For more detailed derivation of the solution, check out the awesome lecture note
[2] on SDE.
Connection to the Heat Equation
For fun, let’s see the connection between Brownian Motion and the heat equation.
This profound connection was discovered by none other than Albert Einstein in
1905 when he was 26 and working on Special Relativity and the Photoelectric Effect :).
Let’s first define a Heat Equation:
∂t∂u=∂x2∂2u
with a given boundary condition u(0,x) and u(t,x) is the probability density
function at time t and position x. Due to the linearity property of the
heat equation (if u1,u2 are solutions then u1+u2 is also a solution), the
intergral ∫usds is a solution if all us are solutions to the heat equation.
So, for all initial condition, it’s sufficient to solve for
u(0,x)=δ(x).
Let denote the solution to this initial condition as uδ(t,x).
We can easily see this if we have some complex initial condition u0, we can
decompose u0 into delta functions:
u0(x)=∫−∞∞δ(x−s)u0(s)ds=δ(x)∗u0(x)
Then, the solution given u0 can also be decomposed into solutions of the u0=δ:
u(t,x)=∫−∞∞uδ(t,x−s)u0(s)ds=uδ(t,x)∗u0(x)
Finally, the solution uδ(t,x) is also known as the heat kernel or gaussian
kernel and is given as
uδ(t,x)=2πt1e−4tx2
To connect this to Brownian Motion, consider the following SDE:
dXt=2kdBtX0=0
then, the probability density function of Xt is given as