BROWNIAN MOTION

A tutorial

Krzysztof Burdzy
University of Washington
 A paradox

f : [0,1] → R, sup | f ′′(t ) | < ∞

t∈[ 0 ,1]

P ( f (t ) − ε < Bt < f (t ) + ε , 0 < t < 1)

⎛ 11 ⎞
≈ c(ε ) exp⎜⎜ − ∫ ( f ′(t )) dt ⎟⎟
2
(*)
⎝ 20 ⎠
 (*) is maximized by f(t) = 0, t>0
The most likely (?!?) shape of a
Brownian path:

Microsoft stock
- the last 5 years
 Definition of Brownian motion
Brownian motion is the unique process
with the following properties:
(i) No memory
(ii) Invariance
(iii) Continuity
(iv) B0 = 0, E ( Bt ) = 0, Var ( Bt ) = t
Memoryless process

t0 t1 t2 t3

Bt1 − Bt0 , Bt2 − Bt1 , Bt3 − Bt2 , K

are independent
 Invariance

The distribution of Bt + s − Bs
depends only on t.
 Path regularity
(i) t → Bt is continuous a.s.
(ii) t → Bt is nowhere differentiable a.s.
 Why Brownian motion?
Brownian motion belongs to several families
of well understood stochastic processes:
(i) Markov processes
(ii) Martingales
(iii) Gaussian processes
(iv) Levy processes
 Markov processes
L {Bt , t ≥ s | Bs } = L {Bt , t ≥ s | Bu , 0 ≤ u ≤ s}

The theory of Markov processes uses

tools from several branches of analysis:
(i) Functional analysis (transition semigroups)
(ii) Potential theory (harmonic, Green functions)
(iii) Spectral theory (eigenfunction expansion)
(iv) PDE’s (heat equation)
 Martingales
s < t ⇒ E ( Bt | Bs ) = Bs

Martingales are the only family of processes

for which the theory of stochastic integrals is
fully developed, successful and satisfactory.
t

∫X
0
s dBs
 Gaussian processes
Bt1 , Bt2 , K, Btn is multidimensional
normal (Gaussian)

(i) Excellent bounds for tails

(ii) Second moment calculations
(iii) Extensions to unordered parameter(s)
 The Ito formula
t nt

∫ X s dBs = lim ∑
n →∞
X k / n
k =0
( B( k +1) / n − Bk / n )
0

t t
1
f ( Bt ) = f ( B0 ) + ∫ f ′( Bs )dBs + ∫ f ′′( Bs )ds
0
20
 Random walk
Wt
t
Independent steps, P(up)=P(down)

{ aW , t ≥ 0}
ta ⎯⎯
⎯→a →∞
{Bt , t ≥ 0}
(in distribution)
 Scaling
Central Limit Theorem (CLT),
parabolic PDE’s
D
{Bt , 0 ≤ t ≤ 1} = { a Bta , 0 ≤ t ≤ 1}
Cameron-Martin-Girsanov formula
Multiply the probability of each Brownian path
{Bt , 0 ≤ t ≤ 1} by
⎛1 1
1
⎞
exp⎜⎜ ∫ f ′( s )dBs − ∫ ( f ′( s )) ds ⎟⎟
2

⎝0 20 ⎠

The effect is the same as replacing

{Bt , 0 ≤ t ≤ 1} with {Bt + f (t ), 0 ≤ t ≤ 1}
 Invariance (2)

Time reversal
D
{Bt , 0 ≤ t ≤ 1} = {B1−t − B1 , 0 ≤ t ≤ 1}

0 1
Brownian motion and the heat equation
u ( x, t ) – temperature at location x at time t
∂ 1
Heat equation: u ( x, t ) = Δ x u ( x, t )
∂t 2
μ (dx) = u ( x,0)dx
μ
Forward
representation
u ( y, t )dy = P ( Bt ∈ dy )
Backward representation
(Feynman-Kac formula) u ( y, t ) = Eu ( Bt − y,0)
y
μ

0 t
Multidimensional Brownian motion
1 2 3
B , B , B , K - independent 1-dimensional
t t t
Brownian motions

1 2 d
( B , B ,K, B )
t t t
- d-dimensional Brownian
motion
 Feynman-Kac formula (2)
Bτ x
f (x)

1
Δu ( x) − V ( x)u ( x) = 0
2
⎛ ⎡ τ
⎤ ⎞
u ( x) = E f ( Bτ ) exp ⎢− ∫ V ( Bs )ds ⎥ ⎟
x⎜
⎜ ⎟
⎝ ⎣ 0 ⎦⎠
 Invariance (3)
The d-dimensional Brownian motion is invariant
under isometries of the d-dimensional space.
It also inherits invariance properties of the
1-dimensional Brownian motion.
1 1
exp(− x1 / 2)
2
exp(− x2 / 2)
2

2π 2π
1
= exp(−( x1 + x2 ) / 2)
2 2

2π
 Conformal invariance

Bt
f
f ( Bt )
analytic

{ f ( Bt ) − f ( B0 ), t ≥ 0}
has the same distribution as
t
{Bc (t ) , t ≥ 0}, c(t ) = ∫ | f ′( Bs ) | ds
2

0
 The Ito formula
Disappearing terms (1)
t t
1
f ( Bt ) = f ( B0 ) + ∫ ∇f ( Bs )dBs + ∫ Δf ( Bs )ds
0
20

If Δf ≡ 0 then
t
f ( Bt ) = f ( B0 ) + ∫ ∇f ( Bs )dBs
0
 Brownian martingales
Theorem (Martingale representation theorem).
{Brownian martingales} = {stochastic integrals}
t
M t = ∫ X s dBs
0

E ( M t | M s ) = M s , M t ∈ Ft = σ {Bs , s ≤ t}
B
 The Ito formula
Disappearing terms (2)
∂
t
f (t , Bt ) − f (t , B0 ) = ∫ f ( s, Bs )dBs
0
∂x
∂ 1 ∂2
t t
+∫ f ( s, Bs )ds + ∫ 2 f ( s, Bs )ds
0
∂s 2 0 ∂x

Ef (t , Bt ) − Ef (t , B0 )
∂ ∂
t t 2
1
= E∫ f ( s, Bs )ds + E ∫ 2 f ( s, Bs )ds
0
∂s 2 0 ∂x
 Mild modifications of BM
Mild = the new process corresponds
to the Laplacian
(i) Killing – Dirichlet problem
(ii) Reflection – Neumann problem
Related models – diffusions
dX t = σ ( X t )dBt + μ ( X t )dt

(i) Markov property – yes

(ii) Martingale – only if μ ≡ 0
(iii) Gaussian – no, but Gaussian tails
Related models – stable processes

Brownian motion – dB = (dt ) 1/ 2

1/ α
Stable processes – dX = (dt )
(i) Markov property – yes
(ii) Martingale – yes and no
(iii) Gaussian – no

Price to pay: jumps, heavy tails, 0 <α ≤ 2

0<2≤2
Related models – fractional BM
1/ α
dX = (dt )
(i) Markov property – no
(ii) Martingale – no
(iii) Gaussian – yes
(iv) Continuous
1<α < ∞
1< 2 < ∞
 Related models – super BM

Super Brownian motion is related to

Δu = u 2

and to a stochastic PDE.

 Related models – SLE
Schramm-Loewner Evolution is a model
for non-crossing conformally invariant
2-dimensional paths.
 Path properties
(i) t → Bt is continuous a.s.
(ii) t → Bt is nowhere differentiable a.s.
(iii) t → Bt is Holder (1 / 2 − ε )
(iv) Local Law if Iterated Logarithm

Bt
lim sup =1
t ↓0 2t log | log t |
 Exceptional points
Bt
lim sup =1
t ↓0 2t log | log t |
For any fixed s>0, a.s.,
Bt − Bs
lim sup =1
t↓s 2(t − s ) log | log(t − s ) |
There exist s>0, a.s., such that
Bt − Bs
lim sup ∈ (0, ∞)
t↓s 2(t − s )
 Cut points
For any fixed t>0, a.s., the 2-dimensional
Brownian path contains a closed loop
around Bt in every interval (t , t + ε )
Almost surely, there exist t ∈ (0,1)
such that B ([0, t )) ∩ B ((t ,1]) = ∅
 Intersection properties
(d = 1) a.s., ∀t ∃s ≠ t Bs = Bt
(d = 2) ∀t a.s. ∀s ≠ t Bs ≠ Bt
−1
a.s., ∃x ∈ R Card ( B ( x)) = ∞
2

−1
(d = 3) a.s., ∃x ∈ R Card ( B ( x)) = 2
3

−1
a.s., ∀x ∈ R Card ( B ( x)) ≤ 2
3

−1
(d = 4) a.s., ∀x ∈ R Card ( B ( x)) ≤ 1
4
Intersections of random sets
dim( A) + dim( B) > d
c
A∩ B ≠ ∅
The dimension of Brownian trace is 2
in every dimension.
 Invariance principle
(i) Random walk converges to Brownian
motion (Donsker (1951))
(ii) Reflected random walk converges
to reflected Brownian motion
2
(Stroock and Varadhan (1971) - C domains,
B and Chen (2007) – uniform domains, not all
domains)
(iii) Self-avoiding random walk in 2 dimensions
converges to SLE (200?)
(open problem)
 Local time
t
1
Lt = lim ∫ 1{−ε < Bs <ε }ds
ε →0 2ε
0

M t = sup Bs
s ≤t
 Local time (2)
D
{Lt ,0 ≤ t ≤ 1} = {M t ,0 ≤ t ≤ 1}
D
{M t − Bt ,0 ≤ t ≤ 1} = {| B |t ,0 ≤ t ≤ 1}
 Local time (3)

σ t = inf {Ls ≥ t}
s >0

Inverse local time is a stable process

with index ½.
 References
„ R. Bass Probabilistic Techniques in
Analysis, Springer, 1995
„ F. Knight Essentials of Brownian Motion
and Diffusion, AMS, 1981
„ I. Karatzas and S. Shreve Brownian
Motion and Stochastic Calculus, Springer,
1988